On 13/02/18 12:54, Tomasz Figa wrote:
On Tue, Feb 13, 2018 at 9:00 PM, Robin Murphy <robin.mur...@arm.com> wrote:
On 13/02/18 07:44, Tomasz Figa wrote:
On Wed, Feb 7, 2018 at 7:31 PM, Vivek Gautam
The device link allows the pm framework to tie the supplier and
consumer. So, whenever the consumer is powered-on the supplier
is powered-on first.
There are however cases in which the consumer wants to power-on
the supplier, but not itself.
E.g., A Graphics or multimedia driver wants to power-on the SMMU
to unmap a buffer and finish the TLB operations without powering
This sounds strange to me. If the SMMU is powered down, wouldn't the
TLB lose its contents as well (and so no flushing needed)?
Depends on implementation details - if runtime PM is actually implemented
via external clock gating (in the absence of fine-grained power domains),
then "suspended" TLBs might both retain state and not receive invalidation
requests, which is really the worst case.
Agreed. That's why in "[PATCH v7 3/6] iommu/arm-smmu: Invoke
pm_runtime during probe, add/remove device" I actually suggested
managing clocks separately from runtime PM. At least until runtime PM
framework arrives at a state, where multiple power states can be
managed, i.e. full power state, clock-gated state, domain-off state.
(I think I might have seen some ongoing work on this on LWN though...)
Other than that, what kind of hardware operations would be needed
besides just updating the page tables from the CPU?
Domain attach/detach also require updating SMMU hardware state (and possibly
TLB maintenance), but don't logically require the master device itself to be
active at the time.
Wouldn't this hardware state need to be reinitialized anyway after
respective power domain power cycles? (In other words, hardware would
only need programming if it's powered on at the moment.)
Yes, if the entire SMMU was fully powered down because all masters were
inactive, then all that should need to be done is to update the software
shadow state in the expectation that arm_smmu_reset() would re-sync it
upon TCU powerup. If at least some part of the internal logic remains
active, though, then you may or may not need to fiddle with zero or more
clocks and/or power domains (depending on microarchitecture and
integration) in order to be sure that everything from the programming
slave interface through to wherever that state is kept works correctly
so that it can be changed.
The main motivation here is that the Qualcomm SMMU microarchitecture
apparently allows the programming interface to be shut down separately
from the TCU core (context banks, page table walker, etc.), and they get
an appreciable power saving from doing so. This is different from, say,
the Arm Ltd. implementations, where the entire TCU is a single
clock/power domain internally (although you could maybe still gate the
external APB interface clock).
As the previous discussions have shown, this is really, really hard to
do properly in a generic manner.
Freedreno mailing list