MPMM is a core-specific microarchitectural feature. It has been present
in every Arm core since the Cortex-A510 and has been implemented in
exactly the same way. Despite that, it is enabled more like an
architectural feature with a top level enable flag. This utilised the
identical implementation.
This duality has left MPMM in an awkward place, where its enablement
should be generic, like an architectural feature, but since it is not,
it should also be core-specific if it ever changes. One choice to do
this has been through the device tree.
This has worked just fine so far, however, recent implementations expose
a weakness in that this is rather slow - the device tree has to be read,
there's a long call stack of functions with many branches, and system
registers are read. In the hot path of PSCI CPU powerdown, this has a
significant and measurable impact. Besides it being a rather large
amount of code that is difficult to understand.
Since MPMM is a microarchitectural feature, its correct placement is in
the reset function. The essence of the current enablement is to write
CPUPPMCR_EL3.MPMM_EN if CPUPPMCR_EL3.MPMMPINCTL == 0. Replacing the C
enablement with an assembly macro in each CPU's reset function achieves
the same effect with just a single close branch and a grand total of 6
instructions (versus the old 2 branches and 32 instructions).
Having done this, the device tree entry becomes redundant. Should a core
that doesn't support MPMM arise, this can cleanly be handled in the
reset function. As such, the whole ENABLE_MPMM_FCONF and platform hooks
mechanisms become obsolete and are removed.
Change-Id: I1d0475b21a1625bb3519f513ba109284f973ffdf
Signed-off-by: Boyan Karatotev <boyan.karatotev@arm.com>
Errata application is painful for performance. For a start, it's done
when the core has just come out of reset, which means branch predictors
and caches will be empty so a branch to a workaround function must be
fetched from memory and that round trip is very slow. Then it also runs
with the I-cache off, which means that the loop to iterate over the
workarounds must also be fetched from memory on each iteration.
We can remove both branches. First, we can simply apply every erratum
directly instead of defining a workaround function and jumping to it.
Currently, no errata that need to be applied at both reset and runtime,
with the same workaround function, exist. If the need arose in future,
this should be achievable with a reset + runtime wrapper combo.
Then, we can construct a function that applies each erratum linearly
instead of looping over the list. If this function is part of the reset
function, then the only "far" branches at reset will be for the checker
functions. Importantly, this mitigates the slowdown even when an erratum
is disabled.
The result is ~50% speedup on N1SDP and ~20% on AArch64 Juno on wakeup
from PSCI calls that end in powerdown. This is roughly back to the
baseline of v2.9, before the errata framework regressed on performance
(or a little better). It is important to note that there are other
slowdowns since then that remain unknown.
Change-Id: Ie4d5288a331b11fd648e5c4a0b652b74160b07b9
Signed-off-by: Boyan Karatotev <boyan.karatotev@arm.com>
Travis' and Gelas' TRMs tell us to disable SME (set PSTATE.{ZA, SM} to
0) when we're attempting to power down. What they don't tell us is that
if this isn't done, the powerdown request will be rejected. On the
CPU_OFF path that's not a problem - we can force SVCR to 0 and be
certain the core will power off.
On the suspend to powerdown path, however, we cannot do this. The TRM
also tells us that the sequence could also be aborted on eg. GIC
interrupts. If this were to happen when we have overwritten SVCR to 0,
upon a return to the caller they would experience a loss of context. We
know that at least Linux may call into PSCI with SVCR != 0. One option
is to save the entire SME context which would be quite expensive just to
work around. Another option is to downgrade the request to a normal
suspend when SME was left on. This option is better as this is expected
to happen rarely enough to ignore the wasted power and we don't want to
burden the generic (correct) path with needless context management.
Signed-off-by: Boyan Karatotev <boyan.karatotev@arm.com>
Change-Id: I698fa8490ebf51461f6aa8bba84f9827c5c46ad4
Add basic CPU library code to support the Alto CPU.
Change-Id: I45958be99c4a350a32a9e511d3705fb568b97236
Signed-off-by: Igor Podgainõi <igor.podgainoi@arm.com>
