Question about arm smmu

[Will Deacon]

Hi Will,

Hello,

I am trying to enable SMMU(mmu500) on an platform.
My platform support only 32 SIDs, but it have more than 100 masters.
So I need to let different masters share one SID. I read current
arm-smmu.c, but it needs each master has unique SID. Do you have
some suggestions about how to let different masters sharing one
SID?

We can achieve that using iommu_groups, but then we need a way to
describe those groups in the device-tree, as opposed to putting
each device into its own group like we do at present. Robin (CC'd) had
some work-in-progress for this, iirc.

On my platform, SID can be dynamically programmed. So I can program
DMA0 and DMA1 using one SID, saying 0x5. But I do not have a good
idea how to support this use case in arm-smmu.c driver.

I think we've have the firmware allocating the SIDs, then describing
the grouping to Linux in the device-tree.

Will

[Robin Murphy]

Post by Will Deacon

Hi Will,

Hello,

I am trying to enable SMMU(mmu500) on an platform.
My platform support only 32 SIDs, but it have more than 100 masters.

Do you mean the MMU-500 configuration really only has 5 bits of stream
ID wired up (which seems a bit unlikely, especially given that in such a
configuration you should gain extra bits from the prepended TBU number),
or that you have a sufficient number of stream ID bits but only have 32
stream matching groups?

Post by Will Deacon

So I need to let different masters share one SID. I read current
arm-smmu.c, but it needs each master has unique SID. Do you have
some suggestions about how to let different masters sharing one
SID?

We can achieve that using iommu_groups, but then we need a way to
describe those groups in the device-tree, as opposed to putting
each device into its own group like we do at present. Robin (CC'd) had
some work-in-progress for this, iirc.

I spent enough time on it that I reached the point of thinking there's
no feasible DT binding that is sufficiently hardware-agnostic, and isn't
just trying to paint a Linux implementation detail as hardware
description - the master IDs are already there, so it might as well be
up to the thing parsing them to work out what, if anything, it wants to
do with duplicates.

What I have now is a simple little lookup table solution inside the
SMMUv2 driver to automatically track groups by stream ID for platform
devices. I've not yet tried pulling in Joerg's generic group series, but
I think it's going to slot together really neatly with that, so I might
try breaking it out of the big stream ID probing rework I'm currently
half-way through.

Post by Will Deacon

On my platform, SID can be dynamically programmed. So I can program
DMA0 and DMA1 using one SID, saying 0x5. But I do not have a good
idea how to support this use case in arm-smmu.c driver.

I think we've have the firmware allocating the SIDs, then describing
the grouping to Linux in the device-tree.

Yeah, regardless of how we handle groups we need the SIDs to be in the
DT and not change at runtime, so the programming is going to have to be
done before the kernel even runs.

Robin.

Post by Will Deacon
Will

[Peng Fan]

Hi Robin, Will,

Post by Robin Murphy

Post by Will Deacon

Hi Will,

Hello,

I am trying to enable SMMU(mmu500) on an platform.
My platform support only 32 SIDs, but it have more than 100 masters.

Do you mean the MMU-500 configuration really only has 5 bits of
stream ID wired up (which seems a bit unlikely, especially given that
in such a configuration you should gain extra bits from the prepended
TBU number), or that you have a sufficient number of stream ID bits
but only have 32 stream matching groups?

There are 32 Stream matching registers, 32 Contexts. The stream id width
seems to be 10 bits.

Post by Robin Murphy

Post by Will Deacon

So I need to let different masters share one SID. I read current
arm-smmu.c, but it needs each master has unique SID. Do you have
some suggestions about how to let different masters sharing one
SID?

We can achieve that using iommu_groups, but then we need a way to
describe those groups in the device-tree, as opposed to putting
each device into its own group like we do at present. Robin (CC'd) had
some work-in-progress for this, iirc.

I spent enough time on it that I reached the point of thinking
there's no feasible DT binding that is sufficiently
hardware-agnostic, and isn't just trying to paint a Linux
implementation detail as hardware description - the master IDs are
already there, so it might as well be up to the thing parsing them to
work out what, if anything, it wants to do with duplicates.
What I have now is a simple little lookup table solution inside the
SMMUv2 driver to automatically track groups by stream ID for platform
devices. I've not yet tried pulling in Joerg's generic group series,
but I think it's going to slot together really neatly with that, so I
might try breaking it out of the big stream ID probing rework I'm
currently half-way through.

Would you mind share me you code? I'd like to try on my platform.

Post by Robin Murphy

Post by Will Deacon

On my platform, SID can be dynamically programmed. So I can program
DMA0 and DMA1 using one SID, saying 0x5. But I do not have a good
idea how to support this use case in arm-smmu.c driver.

I think we've have the firmware allocating the SIDs, then describing
the grouping to Linux in the device-tree.

Yeah, regardless of how we handle groups we need the SIDs to be in
the DT and not change at runtime, so the programming is going to have
to be done before the kernel even runs.

Yeah. I programmed the SID in bootloader. When kernel runs, the SID will
not change.


Thanks,
Peng.

Post by Robin Murphy
Robin.

Post by Will Deacon
Will