On Thu 03-11-22 14:36:40, Yang Shi wrote:
[...]
> So use nodemask to record the nodes which have the same hit record, the
> hugepage allocation could fallback to those nodes. And remove
> __GFP_THISNODE since it does disallow fallback. And if nodemask is
> empty (no node is set), it means there is one single node has the most
> hist record, the nodemask approach actually behaves like __GFP_THISNODE.
>
> Reported-by: syzbot+0044b22d177870ee974f@xxxxxxxxxxxxxxxxxxxxxxxxx
> Suggested-by: Zach O'Keefe <zokeefe@xxxxxxxxxx>
> Suggested-by: Michal Hocko <mhocko@xxxxxxxx>
> Signed-off-by: Yang Shi <shy828301@xxxxxxxxx>
> ---
> mm/khugepaged.c | 32 ++++++++++++++------------------
> 1 file changed, 14 insertions(+), 18 deletions(-)
>
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index ea0d186bc9d4..572ce7dbf4b0 100644
> --- a/mm/khugepaged.c
> +++ b/mm/khugepaged.c
> @@ -97,8 +97,8 @@ struct collapse_control {
> /* Num pages scanned per node */
> u32 node_load[MAX_NUMNODES];
>
> - /* Last target selected in hpage_collapse_find_target_node() */
> - int last_target_node;
> + /* nodemask for allocation fallback */
> + nodemask_t alloc_nmask;
This will eat another 1k on the stack on most configurations
(NODE_SHIFT=10). Along with 4k of node_load this is quite a lot even
on shallow call chains like madvise resp. khugepaged. I would just
add a follow up patch which changes both node_load and alloc_nmask to
dynamically allocated objects.
Other than that LGTM. I thought we want to keep __GFP_THISNODE but after
a closer look it seems that this flag is not really compatible with
nodemask after all. node_zonelist() will simply return a trivial zone
list for a single (preferred node) so no fallback to other nodes is
possible. My bad to not realize it earlier.
--
Michal Hocko
SUSE Labs
