On Fri, Jun 5, 2020 at 6:39 PM Dan Williams dan.j.williams@intel.com wrote:
On Fri, Jun 5, 2020 at 9:22 AM Rafael J. Wysocki rafael@kernel.org wrote: [..]
The fix we are looking at now is to pre-map operation regions in a similar manner as the way APEI resources are pre-mapped. The pre-mapping would arrange for synchronize_rcu_expedited() to be elided on each dynamic mapping attempt. The other piece is to arrange for operation-regions to be mapped at their full size at once rather than a page at a time.
However, if the RCU usage in ACPI OSL can be replaced with an rwlock, some of the ACPICA changes above may not be necessary anymore (even though some of them may still be worth making).
I don't think you can replace the RCU usage in ACPI OSL and still maintain NMI lookups in a dynamic list.
I'm not sure what NMI lookups have to do with the issue at hand.
If acpi_os_{read|write}_memory() is used from NMI, that is a bug already in there which is unrelated to the performance problem with opregions.
However, there are 3 solutions I see:
- Prevent acpi_os_map_cleanup() from triggering at high frequency by
pre-mapping and never unmapping operation-regions resources (internal discussion in progress)
Yes, that can be done, if necessary.
- Prevent walks of the 'acpi_ioremaps' list (acpi_map_lookup_virt())
from NMI context by re-writing the physical addresses in the APEI tables with pre-mapped virtual address, i.e. remove rcu_read_lock() and list_for_each_entry_rcu() from NMI context.
That sounds a bit convoluted to me.
- Split operation-region resources into a separate mapping mechanism
than APEI resources so that typical locking can be used for the sleepable resources and let the NMI accessible resources be managed separately.
That last one is one we have not discussed internally, but it occurred to me when you mentioned replacing RCU.
So NMI cannot use acpi_os_{read|write}_memory() safely which you have pointed out for a few times.
But even if NMI resources are managed separately, the others will still not be sleepable (at least not all of them).
Cheers!