On 4/8/24 8:17 PM, Christoph Lameter (Ampere) wrote: > On Mon, 8 Apr 2024, Vlastimil Babka wrote: > >> On 4/5/24 6:50 PM, Christoph Lameter (Ampere) wrote: >>> On Sat, 30 Mar 2024, Chen Jun wrote: >>> >>>> When kmalloc_node() is called without __GFP_THISNODE and the target node >>>> lacks sufficient memory, SLUB allocates a folio from a different node >>>> other than the requested node, instead of taking a partial slab from it. >>> >>> Hmmm... This would mean that we do not consult the partial lists of the >>> other nodes. That is something to be fixed in the allocator. >> >> Which allocator? If you mean SLUB, this patch fixes it. If you mean page >> allocator, I don't see how. > > > The SLUB allocator of course. And the patch does not fix it. It tries to > convince the page allocator to give us a folio from the right node. The patch primarily makes slub use its partial lists before going to the page allocator. The "give us a folio from the right node" is not "convincing" but an opportunistic request. > That kind of activity can be controlled within the page allocator via the > node reclaim setting. No point in doing multiple calls into the page > allocator. That's assuming there's something to reclaim on that overloaded and requested node in the first place. But yeah, such unbalanced system will likely have multiple issues and slub wouldn't be the only one where it manifests. But if we can remove the slub pathological behavior on such system easily, we should. >>>> However, since the allocated folio does not belong to the requested >>>> node, it is deactivated and added to the partial slab list of the node >>>> it belongs to. >>> >>> That should only occur if a request for an object for node X follows a >>> request for an object from node Y. >> >> Are you sure? I think it's a stream of requests for node X happening on a >> cpu of node Y, AFAICS the first attempt will allocate the slab page from >> node different than X (possibly node Y because it's local and has pages >> available unlike node X which is full). It does get installed as the cpu >> slab, but then the next request is also for node X, so the node matching >> checks make the slab deactivate and allocate a new one. > > Then there is something broken in the cpuslab logic. > > The first request of CPU C for memory from node X should lead to: > > 1. deactivation of current cpu slab if it is not from node X > 2. retrieval of a slab from node X and activation of that slab as cpuslab > 3. Return of an object from that slab and therefore from node X. > > Further allocation should be caught by the hotpatch where we realize that > there is a request from node X and the current cpuslab is from node X and > therefore fastpath logic can be used to retrieve the next object. Yes and that logic AFAIK works. But here we are addressing a situation where we won't get a slab from node X because it's just full. >>> get_any_partial() should do that. Maybe it is not called in the >>> kmalloc_node case. >> >> Yes, get_any_partial() is currently skipped for requests of numa node >> different from NUMA_NO_NODE. > > Maybe we can use that function after checking that the page allocator is > over the watermark on the node that we were wanting to allocate from. That > check should be fast. A GFP_NOWAIT | __GFP_THISNODE attempt is basically that check. Maybe it could be made a bit faster but either we duplicate code or not handle some corner case the full attempt does. This is not a fast path and not a common case (kmalloc_node() vs kmalloc()) so to me it seems better to call the page allocator. >>> I think it's a useful tradeof to first try satisfy the node preference with >> a GFP_NOWAIT allocation. If it succeeds, the target node is not overloaded, >> we get the page from the desired node and further allocations will of the >> same node will not deactivate it. If it doesn't succeed then we indeed >> fallback to slabs on partial list from other nodes before wastefully >> allocating new pages from the other nodes, which addresses the scenario that >> motivated this patch. > > There are also the memory policies etc to consider. F.e. for the > interleave policy the pages must come from different nodes in sequence to > properly balance the allocations over multiple NUMA nodes. There are cases > in which the allocations are forced to specific sets of nodes or where a > node is preferred but fallback to local should occur. > > If you now do multiple page allocator calls then the NUMA interleave > policy etc etc may no longer work. I have not looked to deep into those. Yeah there are policies and there are kmalloc_node() calls with a specific node. If they are incompatible, what should happen? Arguably kmalloc_node() should win as it's a more specific call than a per-process policy? I think as long as there's memory available on all nodes, things will continue working fine and respecting policies. In the corner case we are addressing, where a node is overloaded, observing policies becomes infeasible anyway. Let's look at the steps again: + * 1) try to get a partial slab from target node with GPF_NOWAIT | + * __GFP_THISNODE opportunistically. This was always the first step in get_partial() anyway. + * 2) if 1) failed, try to allocate a new slab from target node with + * GPF_NOWAIT | __GFP_THISNODE opportunistically too. This will try to satisfy the kmalloc_node() preferred node, which should be more important than a policy. + * 3) if 2) failed, retry 1) and 2) with original gfpflags. The step where we attempt to allocate memory from any node will use original gfpflags This is where we reuse existing slabs on other node partials list, as we do for all allocations before resorting to a page allocator. So that's unchanged. If that fails, we go to the page allocator with the original flags so if there are any policies, the page allocator will satisfy them.
