Re: [RFC PATCH v4 7/7] mm/demotion: Demote pages according to allocation fallback order

[Ying Huang]

On Mon, 2022-06-06 at 09:37 +0530, Aneesh Kumar K V wrote:
> On 6/6/22 6:13 AM, Ying Huang wrote:
> > On Fri, 2022-06-03 at 20:39 +0530, Aneesh Kumar K V wrote:
> > > On 6/2/22 1:05 PM, Ying Huang wrote:
> > > > On Fri, 2022-05-27 at 17:55 +0530, Aneesh Kumar K.V wrote:
> > > > > From: Jagdish Gediya <jvgediya@xxxxxxxxxxxxx>
> > > > > 
> > > > > currently, a higher tier node can only be demoted to selected
> > > > > nodes on the next lower tier as defined by the demotion path,
> > > > > not any other node from any lower tier.  This strict, hard-coded
> > > > > demotion order does not work in all use cases (e.g. some use cases
> > > > > may want to allow cross-socket demotion to another node in the same
> > > > > demotion tier as a fallback when the preferred demotion node is out
> > > > > of space). This demotion order is also inconsistent with the page
> > > > > allocation fallback order when all the nodes in a higher tier are
> > > > > out of space: The page allocation can fall back to any node from any
> > > > > lower tier, whereas the demotion order doesn't allow that currently.
> > > > > 
> > > > > This patch adds support to get all the allowed demotion targets mask
> > > > > for node, also demote_page_list() function is modified to utilize this
> > > > > allowed node mask by filling it in migration_target_control structure
> > > > > before passing it to migrate_pages().
> > > > 
> > > 
> > > ...
> > > 
> > > > >     * Take pages on @demote_list and attempt to demote them to
> > > > >     * another node.  Pages which are not demoted are left on
> > > > > @@ -1481,6 +1464,19 @@ static unsigned int demote_page_list(struct list_head *demote_pages,
> > > > >    {
> > > > >    	int target_nid = next_demotion_node(pgdat->node_id);
> > > > >    	unsigned int nr_succeeded;
> > > > > +	nodemask_t allowed_mask;
> > > > > +
> > > > > +	struct migration_target_control mtc = {
> > > > > +		/*
> > > > > +		 * Allocate from 'node', or fail quickly and quietly.
> > > > > +		 * When this happens, 'page' will likely just be discarded
> > > > > +		 * instead of migrated.
> > > > > +		 */
> > > > > +		.gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN |
> > > > > +			__GFP_NOMEMALLOC | GFP_NOWAIT,
> > > > > +		.nid = target_nid,
> > > > > +		.nmask = &allowed_mask
> > > > > +	};
> > > > 
> > > > IMHO, we should try to allocate from preferred node firstly (which will
> > > > kick kswapd of the preferred node if necessary).  If failed, we will
> > > > fallback to all allowed node.
> > > > 
> > > > As we discussed as follows,
> > > > 
> > > > https://lore.kernel.org/lkml/69f2d063a15f8c4afb4688af7b7890f32af55391.camel@xxxxxxxxx/
> > > > 
> > > > That is, something like below,
> > > > 
> > > > static struct page *alloc_demote_page(struct page *page, unsigned long node)
> > > > {
> > > > 	struct page *page;
> > > > 	nodemask_t allowed_mask;
> > > > 	struct migration_target_control mtc = {
> > > > 		/*
> > > > 		 * Allocate from 'node', or fail quickly and quietly.
> > > > 		 * When this happens, 'page' will likely just be discarded
> > > > 		 * instead of migrated.
> > > > 		 */
> > > > 		.gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) |
> > > > 			    __GFP_THISNODE  | __GFP_NOWARN |
> > > > 			    __GFP_NOMEMALLOC | GFP_NOWAIT,
> > > > 		.nid = node
> > > > 	};
> > > > 
> > > > 	page = alloc_migration_target(page, (unsigned long)&mtc);
> > > > 	if (page)
> > > > 		return page;
> > > > 
> > > > 	mtc.gfp_mask &= ~__GFP_THISNODE;
> > > > 	mtc.nmask = &allowed_mask;
> > > > 
> > > > 	return alloc_migration_target(page, (unsigned long)&mtc);
> > > > }
> > > 
> > > I skipped doing this in v5 because I was not sure this is really what we
> > > want.
> > 
> > I think so.  And this is the original behavior.  We should keep the
> > original behavior as much as possible, then make changes if necessary.
> > 
> 
> That is the reason I split the new page allocation as a separate patch. 
> Previous discussion on this topic didn't conclude on whether we really 
> need to do the above or not
> https://lore.kernel.org/lkml/CAAPL-u9endrWf_aOnPENDPdvT-2-YhCAeJ7ONGckGnXErTLOfQ@xxxxxxxxxxxxxx/

Please check the later email in the thread you referenced.  Both Wei and
me agree that the use case needs to be supported.  We just didn't reach
concensus about how to implement it.  If you think Wei's solution is
better (referenced as below), you can try to do that too.  Although I
think my proposed implementation is much simpler.

"
This is true with the current allocation code. But I think we can make
some changes for demotion allocations. For example, we can add a
GFP_DEMOTE flag and update the allocation function to wake up kswapd
when this flag is set and we need to fall back to another node.
"

> Based on the above I looked at avoiding GFP_THISNODE allocation. If you 
> have experiment results that suggest otherwise can you share? I could 
> summarize that in the commit message for better description of why 
> GFP_THISNODE enforcing is needed.

Why?  GFP_THISNODE is just the first step.  We will fallback to
allocation without it if necessary.

Best Regards,
Huang, Ying

> > > I guess we can do this as part of the change that is going to
> > > introduce the usage of memory policy for the allocation?
> > 
> > Like the memory allocation policy, the default policy should be local
> > preferred.  We shouldn't force users to use explicit memory policy for
> > that.
> > 
> > And the added code isn't complex.
> > 
> 
> -aneesh