Messages in this thread | | | Date | Tue, 26 Apr 2022 14:25:11 +0100 | From | Sudeep Holla <> | Subject | Re: [PATCH V2] arm64: add SCHED_CLUSTER's dependency on ACPI |
| |
On Tue, Apr 26, 2022 at 06:52:34AM +0000, 王擎 wrote: > > >> > >> >> > >> >> >> From: Wang Qing <wangqing@vivo.com> > >> >> >> > >> >> >> cluster sched_domain configured by cpu_topology[].cluster_sibling, > >> >> >> which is set by cluster_id, cluster_id can only get from ACPI. > >> >> >> > >> >> >> If the system does not enable ACPI, cluster_id is always -1, even enable > >> >> >> SCHED_CLUSTER is invalid, this is misleading. > >> >> >> > >> >> >> So we add SCHED_CLUSTER's dependency on ACPI here. > >> >> >> > >> >> > > >> >> >Any reason why this can't be extended to support DT based systems via > >> >> >cpu-map in the device tree. IMO we almost have everything w.r.t topology > >> >> >in DT and no reason to deviate this feature between ACPI and DT. > >> >> > > >> >> That's the problem, we parse out "cluster" info according to the > >> >> description in cpu-map, but do assign it to package_id, which used to > >> >> configure the MC sched domain, not cluster sched domain. > >> >> > >> > > >> >Right, we haven't updated the code after updating the bindings to match > >> >ACPI sockets which are the physical package boundaries. Clusters are not > >> >the physical boundaries and the current topology code is not 100% aligned > >> >with the bindings after Commit 849b384f92bc ("Documentation: DT: arm: add > >> >support for sockets defining package boundaries") > >> > >> I see, but this commit is a long time ago, why hasn't it been used widely. > >> Maybe I can help about it if you need. > >> > > > >I assume no one cared or had a requirement for the same. > > It took me a while to find the root cause why enabling SCHED_CLUSTER > didn't work. > > We should add SCHED_CLUSTER's dependency before implementation. > Otherwise, everyone who doesn't have ACPI but use SCHED_CLUSTER > will have this problem. >
I am fine with that or mark it broken for DT, but ideally I wouldn't want to create unnecessary dependency on ACPI or DT when both supports the feature.
-- Regards, Sudeep
| |