Re: [PATCH v7 3/4] iommu: rockchip: Add internal ops to handle variants

From: Robin Murphy
Date: Fri Jul 30 2021 - 09:29:18 EST

On 2021-07-30 13:52, Dafna Hirschfeld wrote:


On 29.07.21 18:58, Robin Murphy wrote:
On 2021-07-29 17:08, Heiko Stübner wrote:
Hi Dafna,

Am Donnerstag, 29. Juli 2021, 17:59:26 CEST schrieb Dafna Hirschfeld:
On 25.05.21 14:15, Benjamin Gaignard wrote:
@@ -879,7 +895,7 @@ static int rk_iommu_enable(struct rk_iommu *iommu)
       for (i = 0; i < iommu->num_mmu; i++) {
           rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
-                   rk_domain->dt_dma);
+                   rk_ops->dma_addr_dte(rk_domain->dt_dma));

Hi,
This is not related to that patch, I was wondring why are all mmu devices initialized
with the same dt_dma?
I see for example that the isp0_mmu in rk3399.dtsi has two resources. Can't each resource
be initialized with different dt_dma and this way there are two dt tables instead of the two mmus pointing
to the same dt table.

maybe
git log -1 cd6438c5f8446691afa4829fe1a9d7b656204f11

"iommu/rockchip: Reconstruct to support multi slaves
There are some IPs, such as video encoder/decoder, contains 2 slave iommus,
one for reading and the other for writing. They share the same irq and
clock with master.
This patch reconstructs to support this case by making them share the same
Page Directory, Page Tables and even the register operations.
That means every instruction to the reading MMU registers would be
duplicated to the writing MMU and vice versa."

Right. In theory we *could* maintain a separate pagetable for each IOMMU instance, but it would just lead to a load of complexity and overhead. For a map request, we'd have to do extra work to decide which table(s) need modifying, and duplicate all the work of the actual mapping if it's more than one. For an unmap request, we'd have no choice but to walk *all* the tables backing that domain to figure out which (if any) actually had it mapped in the first place.

Given that we already have distinct read and write permissions for mappings within a single table, there's not even any functional benefit that could be gained in this case (and in the more general case where the device might emit all kinds of transactions from all its interfaces you'd have to maintain identical mappings for all its IOMMUs anyway). Saving memory and code complexity by physically sharing one pagetable and not worrying about trying to do selective TLB maintenance is a bigger win than anything else could be.

Robin.

Hi, I just try to understand how this iommu hardware/software works. I have two questions,

1. So we currently miss a potential mapping of the hardware right? I mean , each mmu can map 1024*1024*4K = 4G addresses, so two mmus can potentially map 8G. But since
we set them to identical values, we can map only up to 4G.

Not quite. We have 4GB of address space in which read transaction operate, and 4GB of address space in which write transactions operate, but it's hopefully obvious why those are not interchangeable. Technically you *could* map a piece of physical memory to be read and written via different virtual addresses, but you could do that with permissions in a single table anyway, and mostly it would just break any device which expects a single buffer address to both read and write.

2. What is the benefit of setting all mmus if they are all set to the same values? Can't we just work with the first mmu like it was done before that patch
cd6438c5f8446691afa4829fe1a9d7b656204f11

The hardware has two physical interfaces through which it issues transactions - if we only program the IOMMU on one of those interfaces, different transactions will have inconsistent views of memory as above, and the device will probably not function correctly.

Before that patch, the cases where just the "first" MMU was programmed were the ones where it was also the only MMU, as they still are. The IOMMUs for these multi-interface devices weren't enabled at all because it would have just broken things.

Robin.