On Fri, Jun 5, 2020 at 6:32 AM Rafael J. Wysocki <rafael@xxxxxxxxxx> wrote:
>
> On Fri, May 8, 2020 at 1:55 AM Dan Williams <dan.j.williams@xxxxxxxxx> wrote:
> >
> > Recently a performance problem was reported for a process invoking a
> > non-trival ASL program. The method call in this case ends up
> > repetitively triggering a call path like:
> >
> > acpi_ex_store
> > acpi_ex_store_object_to_node
> > acpi_ex_write_data_to_field
> > acpi_ex_insert_into_field
> > acpi_ex_write_with_update_rule
> > acpi_ex_field_datum_io
> > acpi_ex_access_region
> > acpi_ev_address_space_dispatch
> > acpi_ex_system_memory_space_handler
> > acpi_os_map_cleanup.part.14
> > _synchronize_rcu_expedited.constprop.89
> > schedule
> >
> > The end result of frequent synchronize_rcu_expedited() invocation is
> > tiny sub-millisecond spurts of execution where the scheduler freely
> > migrates this apparently sleepy task. The overhead of frequent scheduler
> > invocation multiplies the execution time by a factor of 2-3X.
> >
> > For example, performance improves from 16 minutes to 7 minutes for a
> > firmware update procedure across 24 devices.
> >
> > Perhaps the rcu usage was intended to allow for not taking a sleeping
> > lock in the acpi_os_{read,write}_memory() path which ostensibly could be
> > called from an APEI NMI error interrupt?
>
> Not really.
>
> acpi_os_{read|write}_memory() end up being called from non-NMI
> interrupt context via acpi_hw_{read|write}(), respectively, and quite
> obviously ioremap() cannot be run from there, but in those cases the
> mappings in question are there in the list already in all cases and so
> the ioremap() isn't used then.
>
> RCU is there to protect these users from walking the list while it is
> being updated.
>
> > Neither rcu_read_lock() nor ioremap() are interrupt safe, so add a WARN_ONCE() to validate that rcu
> > was not serving as a mechanism to avoid direct calls to ioremap().
>
> But it would produce false-positives if the IRQ context was not NMI,
> wouldn't it?
>
> > Even the original implementation had a spin_lock_irqsave(), but that is not
> > NMI safe.
>
> Which is not a problem (see above).
>
> > APEI itself already has some concept of avoiding ioremap() from
> > interrupt context (see erst_exec_move_data()), if the new warning
> > triggers it means that APEI either needs more instrumentation like that
> > to pre-emptively fail, or more infrastructure to arrange for pre-mapping
> > the resources it needs in NMI context.
>
> Well, I'm not sure about that.
Right, this patch set is about 2-3 generations behind the architecture
of the fix we are discussing internally, you might mention that.
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.
