| ef738d19 | 21-Jun-2024 |
Manish Pandey <manish.pandey2@arm.com> |
feat(psci): remove cpu context init by index
Currently, the calling core (meaning the core which received the call to
CPU_ON or the powerdown path of CPU_SUSPEND on the same core) is in
charge of in
feat(psci): remove cpu context init by index
Currently, the calling core (meaning the core which received the call to
CPU_ON or the powerdown path of CPU_SUSPEND on the same core) is in
charge of initialising the context for the waking core (the warmboot
entrypoint for both). This is convenient because the calling core can
write the context while in coherency and the waking core will only need
the context after its entered coherency. This avoids any cache
maintenance and makes communication simple.
However, this has 3 main problems:
a) asymmetric feature support is problematic - the calling core has no
way of knowing the feature set of the waking core. If the two
diverge, the architectural feature discovery via ID registers breaks
down. We've thus far "fixed" this on a case by case basis which
doesn't scale and introduces redundancy.
b) powerdown abandon (pabandon) introduces a contradiction - the calling
core has to initialise the context for when the core wakes up, but
should the core not powerdown it needs its old context intact. The only
way to work around this is by keeping two copies of context which
incurs a runtime and memory overhead.
c) cm_prepare_el3_exit[_ns]() doesn't have access to the entrypoint but needs
it to make initialisation decisions. We can infer some of this from
registers that have already been written but this is awkwardly
limiting for what we can do. This also necessitates the split from
the context initialisation.
We can solve all three by a making a core be in full ownership of its
own context. The calling core then only writes entrypoint information
and nothing else. The waking core then initialises its own context as it
sees fit with full knowledge of the whole picture.
The only tricky bit is cache coherency - the waking core has to be able
to coherently observe its new entrypoint. Calling cores will write to
the shared region with coherent caches on. If we make sure to read the
context only after the waking core has entered coherency, then we can
avoid cache operations and let hardware handle everything.
We can skip the spsr check for FEAT_TCR2 as it doesn't make a
difference. We can also skip enabling it twice from generic code.
Signed-off-by: Boyan Karatotev <boyan.karatotev@arm.com>
Signed-off-by: Manish Pandey <manish.pandey2@arm.com>
Change-Id: I86e7fe8b698191fc3b469e5ced1fd010f8754b0e
show more ...
|
| 4a7916a5 | 26-Mar-2025 |
Sudeep Holla <sudeep.holla@arm.com> |
docs: clarify multiple UUID support in ffa manifest
If a partition supports multiple UUID, the UUID property in the list
of partition properties in the FF-A manifest must be a list or array.
Update
docs: clarify multiple UUID support in ffa manifest
If a partition supports multiple UUID, the UUID property in the list
of partition properties in the FF-A manifest must be a list or array.
Update the document to clarify the same.
Change-Id: Id13fe8dbf57a00e5cd186158270b716a4a9aedf7
Signed-off-by: Sudeep Holla <sudeep.holla@arm.com>
show more ...
|
| 0a580b51 | 15-Nov-2024 |
Boyan Karatotev <boyan.karatotev@arm.com> |
perf(cm): drop ZCR_EL3 saving and some ISBs and replace them with root context
SVE and SME aren't enabled symmetrically for all worlds, but EL3 needs
to context switch them nonetheless. Previously,
perf(cm): drop ZCR_EL3 saving and some ISBs and replace them with root context
SVE and SME aren't enabled symmetrically for all worlds, but EL3 needs
to context switch them nonetheless. Previously, this had to happen by
writing the enable bits just before reading/writing the relevant
context. But since the introduction of root context, this need not be
the case. We can have these enables always be present for EL3 and save
on some work (and ISBs!) on every context switch.
We can also hoist ZCR_EL3 to a never changing register, as we set its
value to be identical for every world, which happens to be the one we
want for EL3 too.
Change-Id: I3d950e72049a298008205ba32f230d5a5c02f8b0
Signed-off-by: Boyan Karatotev <boyan.karatotev@arm.com>
show more ...
|
| 83ec7e45 | 06-Nov-2024 |
Boyan Karatotev <boyan.karatotev@arm.com> |
perf(amu): greatly simplify AMU context management
The current code is incredibly resilient to updates to the spec and
has worked quite well so far. However, recent implementations expose a
weakness
perf(amu): greatly simplify AMU context management
The current code is incredibly resilient to updates to the spec and
has worked quite well so far. However, recent implementations expose a
weakness in that this is rather slow. A large part of it is written in
assembly, making it opaque to the compiler for optimisations. The
future proofness requires reading registers that are effectively
`volatile`, making it even harder for the compiler, as well as adding
lots of implicit barriers, making it hard for the microarchitecutre to
optimise as well.
We can make a few assumptions, checked by a few well placed asserts, and
remove a lot of this burden. For a start, at the moment there are 4
group 0 counters with static assignments. Contexting them is a trivial
affair that doesn't need a loop. Similarly, there can only be up to 16
group 1 counters. Contexting them is a bit harder, but we can do with a
single branch with a falling through switch. If/when both of these
change, we have a pair of asserts and the feature detection mechanism to
guard us against pretending that we support something we don't.
We can drop contexting of the offset registers. They are fully
accessible by EL2 and as such are its responsibility to preserve on
powerdown.
Another small thing we can do, is pass the core_pos into the hook.
The caller already knows which core we're running on, we don't need to
call this non-trivial function again.
Finally, knowing this, we don't really need the auxiliary AMUs to be
described by the device tree. Linux doesn't care at the moment, and any
information we need for EL3 can be neatly placed in a simple array.
All of this, combined with lifting the actual saving out of assembly,
reduces the instructions to save the context from 180 to 40, including a
lot fewer branches. The code is also much shorter and easier to read.
Also propagate to aarch32 so that the two don't diverge too much.
Change-Id: Ib62e6e9ba5be7fb9fb8965c8eee148d5598a5361
Signed-off-by: Boyan Karatotev <boyan.karatotev@arm.com>
show more ...
|
| 3406ff00 | 18-Sep-2024 |
Madhukar Pappireddy <madhukar.pappireddy@arm.com> |
docs: fix ff-a manifest binding document
The support for runtime-model has never been implemented by any SPMC.
Hence, remove the corresponding field from binding document.
Also, fix the incorrect d
docs: fix ff-a manifest binding document
The support for runtime-model has never been implemented by any SPMC.
Hence, remove the corresponding field from binding document.
Also, fix the incorrect description of the `managed-exit-virq` property.
Signed-off-by: Madhukar Pappireddy <madhukar.pappireddy@arm.com>
Change-Id: I0a5ef3f08202a8c76edd9a6e1ac680ac3a38ca60
show more ...
|
| 42cf6026 | 10-Jul-2024 |
Juan Pablo Conde <juanpablo.conde@arm.com> |
refactor(rmmd): plat token requests in pieces
Until now, the attestation token size was limited by the size of the
shared buffer between RMM and TF-A. With this change, RMM can now
request the token
refactor(rmmd): plat token requests in pieces
Until now, the attestation token size was limited by the size of the
shared buffer between RMM and TF-A. With this change, RMM can now
request the token in pieces, so they fit in the shared buffer. A new
output parameter was added to the SMC call, which will return (along
with the size of bytes copied into the buffer) the number of bytes
of the token that remain to be retrieved.
TF-A will keep an offset variable that will indicate the position in
the token where the next call will retrieve bytes from. This offset
will be increased on every call by adding the number number of bytes
copied. If the received hash size is not 0, TF-A will reset the
offset to 0 and copy from that position on.
The SMC call will now return at most the size of the shared buffer
in bytes on every call. Therefore, from now on, multiple SMC calls
may be needed to be issued if the token size exceeds the shared
buffer size.
Change-Id: I591f7013d06f64e98afaf9535dbea6f815799723
Signed-off-by: Juan Pablo Conde <juanpablo.conde@arm.com>
show more ...
|