Re: [PATCH 4/9] mailbox: tegra-hsp: Add support for shared mailboxes
From: Pekka Pessi
Date: Tue Oct 30 2018 - 12:16:15 EST
I guess it being an example doesn't make it strictly a recommendation, but
I wonder if we should avoid giving examples that use mappings which we
discourage.
Well, yes, that is example for top0 â but I agree, giving examples that
do not make much sense is not very productive.
So currently
we list all available interrupts in DT and then we could pick just the
first. But what if somebody else already picked the first and "owns" it
We have rather static hardware allocation in device trees, if an another
VM instance or another aux cpu owns interrupt, we just remove it from
the DTS or DTSI. So the native Linux would see all the possible
interrupts (that are not used by SPE or RCE, for instance), but a
virtualized one would see only one.
For mailboxes that the CCPLEX writes to, we can reconfigure them as
producers by disabling the FULL interrupt and enabling the EMPTY
interrupt instead. We can do that the first time somebody calls the
mbox_send_message() on the mailbox.
The problem here is "disabling the FULL interrupt". The main problem are
the mailbox-specific IE registers, they are not really shared, but the
consumer owns the full_ie and producer empty_ie. We can not enable or
disable interrupts without interfering with the interrupt settings of
the remote end, when the interrupt is enabled, we either modify
full_ie/empty_ie register or run with a risk that it has incorrect value.
Can we just leave the mbox channel without interrupts? How the consumer
indicates that it is interested in receiving messages? mbox_peek_data()?
--Pekka
On 10/29/2018 03:16 PM, Thierry Reding wrote:
On Mon, Oct 29, 2018 at 02:25:42PM +0200, Pekka Pessi wrote:
Hi Thierry,
From the 0/9:
Are you aware of any others that we need to take into account?
We would like to use upstream driver for RCE (and probably AON and SCE)
mailbox handling, too. Eventually.
This is a bit
of a problem because the mailbox driver doesn't really know anything
about the direction when it starts up, so how would it make the decision
about how to program the registers?
I'm afraid that information must be stored in the device tree.
What's somewhat surprising is that you're saying that both FULL and
EMPTY interrupts should be handled by the same shared interrupt. That's
the opposite of what the recommended programming sequence is that the
TRM specifies.
Which TRM you mean? Something here?
https://p4viewer.nvidia.com/get///hw/ar/doc/Tech_Pubs/Xavier/TRM/
I browsed through he Xavier_TRM_DP09253001v1.pdf HSP section and I could not
find any recommendations? But see below.
I'm referring to Xavier_TRM_Public.pdf in that location. If you look at
page 4422, "Shared Interrupts Configuration", the example has aggregated
EMPTY and FULL interrupts on shared interrupts 0 and 1, respectively. I
guess it being an example doesn't make it strictly a recommendation, but
I wonder if we should avoid giving examples that use mappings which we
discourage.
Why is it better to handle both FULL and EMPTY interrupts
with the same shared interrupt?
Only top0 has 8 interrupts, rest of the HSP blocks have only 4 (or 5 in case
of top1) interrupts per HSP available through LIC. Hogging two of them means
that only two VMs can access a HSP.
Virtualization isn't something that we're very concerned about upstream,
but I'll take that under consideration.
Would it be safe to clear all of the IE registers to 0 on driver probe?
Nope, the driver should clear only the IE register for the shared interrupt
that the driver uses. Other IEs are used by other entities.
Right, that makes sense. But within that IE register it can consider all
bits fair game, right? One thing I wonder, though, is whether there
should be some external mechanism to set the shared interrupt to use. If
we go purely by convention we'll eventually get this wrong. So currently
we list all available interrupts in DT and then we could pick just the
first. But what if somebody else already picked the first and "owns" it?
I'm not sure we have any practical mechanism to rewrite the DTB, but
perhaps something to keep in mind if ever we need to support other
entities down the road.
If they are indeed separate
for each processor, it should be fairly easy to keep track of the
mailboxes used by the kernel and process only those.
Yes, that is what we should do. Again the directionality should be specified
in DT.
Currently we encode the shared mailboxes in DT like this:
mboxes = <&hsp_top0 TEGRA_HSP_MBOX_TYPE_SM 0>,
<&hsp_aon TEGRA_HSP_MBOX_TYPE_SM 1>;
mbox-names = "rx", "tx";
I suppose we could change that to something like:
mboxes = <&hsp_top0 TEGRA_HSP_MBOX_TYPE_SM (0 << 31 | 0)>,
<&hsp_aon TEGRA_HSP_MBOX_TYPE_SM (1 << 31 | 0)>;
Where the MSB of the mailbox index would indicate the direction. We
could maybe add some eye-candy to make it easier to read:
mboxes = <&hsp_top0 TEGRA_HSP_MBOX_TYPE_SM TEGRA_HSP_MBOX_SM_RX(0)>,
<&hsp_aon TEGRA_HSP_MBOX_TYPE_SM TEGRA_HSP_MBOX_SM_TX(1)>;
That said, I wonder if perhaps it is safe to treat all mailboxes as
consumers by default and omit and direction in DT. If a mailbox is used
as consumer, then the CCPLEX is only interested in FULL interrupts, so
we enable those. The entity that shares the mailbox will write data to
it and is only interested in the EMPTY interrupt, which we don't touch
from the CCPLEX.
For mailboxes that the CCPLEX writes to, we can reconfigure them as
producers by disabling the FULL interrupt and enabling the EMPTY
interrupt instead. We can do that the first time somebody calls the
mbox_send_message() on the mailbox.
Do you see any problems with that approach?
I'm not entirely clear on what the advantages are of using the per-
mailbox registers, or how they are supposed to be used. The existing
documentation doesn't really explain how these are supposed to be used
either, so I was mostly just going by trial and error.
On virtualized configs, multiple VMs can use same HSP block but each VM must
use its own interrupt. Because all the IE registers are on the same 64 KB
page, the writes to the page are trapped by hypervisor, which means that
enabling or disabling an interrupt via the shared IE register is pretty
heavy operation. The per-mailbox IE registers are on a page accessed freely
by the IE, no need to trap.
Do the per-mailbox IE registers act as a second level enable, then? The
HSP driver would still have to set the IE register to aggregate FULL and
EMPTY interrupts as needed, but could then use the per-mailbox IE
registers to actually enable and disable (that is, unmask and mask) the
interrupts?
If a VM uses only one empty mailbox interrupt, using a dedicated "shared"
interrupt and disabling that in GIC could be a lighter operation, we used to
do that in Parker.
Okay, that seems like it would somewhat complicate things and we don't
really support VM uses yet, so I think it best to leave that out for
now.
Thierry
On 10/29/2018 12:39 PM, Thierry Reding wrote:
On Mon, Oct 29, 2018 at 12:04:22PM +0200, Pekka Pessi wrote:
Hi Thierry,
There is typically one entity (aux cpu or a VM running on CCPLEX) owning the
"empty" or producer side of mailbox (iow, waking up on empty) and another
entity owning the "full" or consumer side of mailbox (waking up on full). An
entity should not muck with the interrupts used by the opposite side.
Okay, that explains some of my observations. I was initially trying to
program interrupt enables for both FULL and EMPTY interrupts for all
mailboxes, but then I'd usually get timeouts because the consumer wasn't
responding (i.e. the SPE wasn't getting FULL interrupts for the CCPLEX's
TX mailbox).
If I understand correctly, you're saying that the CPU should only be
using the EMPTY interrupts for it's TX mailbox (while leaving the FULL
interrupts completely untouched) and only the FULL interrupt for it's
RX mailbox (while leaving the EMPTY interrupts untouched). This is a bit
of a problem because the mailbox driver doesn't really know anything
about the direction when it starts up, so how would it make the decision
about how to program the registers?
One entity typically owns one shared interrupt only. For the
BPMP/SCE/RCE/SPE HSP blocks the shared interrupt 0 is owned by the auxiliary
processor itself, the shared interrupts 1..4 are connected to LIC and are
available to other entities. The convention is to go through the interrupts
0..4 and then using the first available shared interrupt for both full and
empty.
That partially matches another of my observations. It seems like we
can't use the shared interrupt 0 at all on at least the AON HSP. That's
fine because that HSP instance contains the TX mailbox for TCU and by
the current convention in the HSP driver, shared interrupt 0 would be
aggregating the FULL interrupts, which according to the above we don't
need for TX mailboxes.
What's somewhat surprising is that you're saying that both FULL and
EMPTY interrupts should be handled by the same shared interrupt. That's
the opposite of what the recommended programming sequence is that the
TRM specifies. Why is it better to handle both FULL and EMPTY interrupts
with the same shared interrupt?
The interrupt functions should use a mask for mailboxes owned by kernel (in
essence what the IE register should be for the HSP shared interrupt owned by
the kernel) and serve only those mailboxes owned by kernel. Note that there
is no reset for HSP in Xavier, and the IE register contents may be stale.
Would it be safe to clear all of the IE registers to 0 on driver probe?
I seem to remember trying to do that and getting similar behaviour to
what I describe above, namely that interrupts on the SPE weren't working
anymore. I concluded that the IE register must be shared between the
various processors, even though that's somewhat suprising given that
there is no way to synchronize accesses to those registers, so their
programming would be somewhat up to chance.
Do you know any more about these registers? If they are indeed separate
for each processor, it should be fairly easy to keep track of the
mailboxes used by the kernel and process only those. Again I don't know
how exactly to distinguish between TX and RX mailboxes because they all
start out the same and only their use defines which direction they go.
Currently this works because we program them as consumers by default.
That means we enable the FULL interrupts but keep EMPTY interrupts
disabled until a message in transmitted on the mailbox, at which point
we enable the EMPTY interrupt. I suppose at that point we should also
disable the FULL interrupt, given the above discussion.
And lastly, if we want to support only Xavier and later, perhaps we should
be more clear in the bindings? There are no mailbox-specific interrupt
enable registers available on Parker and your design relies on them.
That was certainly not the intention. I thought I had seen the per-
mailbox interrupt enable registers also in Tegra186 documentation, but
after double-checking they're indeed not there. I don't think the driver
currently "relies" on them because it uses them in addition to the
HSP_IE registers. I suppose that accessing them might cause aborts on
Tegra186 if they don't exist, though.
I'm not entirely clear on what the advantages are of using the per-
mailbox registers, or how they are supposed to be used. The existing
documentation doesn't really explain how these are supposed to be used
either, so I was mostly just going by trial and error.
Do you know anything more on how to use these registers? I can easily
make them Tegra194 specific in the code, but if they're aren't any clear
advantages, it might just be easier to stick with HSP_IE programming
only.
Thierry