On 10/12/2024 4:26 pm, Matthew Rosato wrote:
On 12/9/24 9:57 PM, Baolu Lu wrote:
On 12/10/24 03:24, Matthew Rosato wrote:
In addition to IOMMU_DOMAIN_DMA, def_domain_type can also return
IOMMU_DOMAIN_DMA_FQ when applicable, else flush queues will never be
used.
Signed-off-by: Matthew Rosato<mjrosato@xxxxxxxxxxxxx>
---
include/linux/iommu.h | 1 +
1 file changed, 1 insertion(+)
diff --git a/include/linux/iommu.h b/include/linux/iommu.h
index 05279109c732..d0da1918d2de 100644
--- a/include/linux/iommu.h
+++ b/include/linux/iommu.h
@@ -585,6 +585,7 @@ iommu_copy_struct_from_full_user_array(void *kdst, size_t kdst_entry_size,
* @def_domain_type: device default domain type, return value:
* - IOMMU_DOMAIN_IDENTITY: must use an identity domain
* - IOMMU_DOMAIN_DMA: must use a dma domain
+ * - IOMMU_DOMAIN_DMA_FQ: dma domain with batch invalidation
In which case must an iommu driver return IOMMU_DOMAIN_DMA_FQ?
The flush queue is a policy of "when and how to synchronize the IOTLB"
in dma-iommu.c. The iommu driver actually has no need to understand such
policy.
If you look ahead to the next patch where I implement def_domain_type for s390, I found that if I only ever return IOMMU_DOMAIN_DMA from ops->def_domain_type then when go through iommu_dma_init_domain() we will never call iommu_dma_init_fq() regardless of IOMMU_CAP_DEFERRED_FLUSH because of the if (domain->type == IOMMU_DOMAIN_DMA_FQ) check. So something isn't right here.
Conceptually I don't think it ever makes sense for a driver to *require*
a device to use deferred invalidation. Furthermore it's been the whole
design for a while now that drivers should never see nor have to
acknowledge IOMMU_DOMAIN_DMA_FQ, it's now just an internal type which
exists largely for the sake of making the sysfs interface work really
neatly. Also beware that a major reason for overriding
iommu_def_domain_type with a paging domain is for untrusted devices, so
massaging the result based on iommu_dma_strict is still not necessarily
appropriate anyway.
It appears the real underlying issue is that, like everyone else in the
same situation, you're doing def_domain_type wrong. If and when you
can't support IOMMU_DOMAIN_IDENTITY, the expectation is that you make
__iommu_alloc_identity_domain() fail, such that if iommu_def_domain_type
is then ever set to passthrough, iommu_group_alloc_default_domain()
falls back to IOMMU_DOMAIN_DMA by itself, and the user gets told they
did a silly thing.
What you see apple-dart doing is a hack around the old bus-based
iommu_domain_alloc() API where there wasn't enough information at the
right point to necessarily do the right thing.
Thanks,
Robin.
It looks to me like the following is happening:
We first have the iommu_def_domain_type set in iommu_subsys_init or via one of the set_default routines, e.g.:
if (!iommu_default_passthrough() && !iommu_dma_strict)
iommu_def_domain_type = IOMMU_DOMAIN_DMA_FQ;
But when we arrive at iommu_group_alloc_default_domain()...
if we have no ops->def_domain_type() defined we will call __iommu_group_alloc_default_domain using what is in iommu_def_domain_type, which could be IOMMU_DOMAIN_DMA, IOMMU_DOMAIN_DMA_FQ or IOMMU_DOMAIN_IDENTITY based on strict/passthrough settings. Testing an s390 scenario today without this series applied, we will call __iommu_group_alloc_default_domain with IOMMU_DOMAIN_DMA_FQ, as long as iommu.strict/passthrough is not specified, so then later in dma-iommu:iommu_dma_init_domain() we can use FQ based on IOMMU_CAP_DEFERRED_FLUSH.
but once we add ops->def_domain_type() then we end up calling iommu_group_alloc_default_domain() with a req_type == the return value from ops->def_domain_type(), which by the current definition can only be IOMMU_DOMAIN_DMA or IOMMU_DOMAIN_IDENTITY. We will then call __iommu_group_alloc_default_domain with that req_type; so without this patch + the DMA_FQ path in patch 6 we would always end up allocating IOMMU_DOMAIN_DMA instead of IOMMU_DOMAIN_DMA_FQ by default, so when we arrive at dma:iommu_dma_init_domain() we won't check for IOMMU_CAP_DEFERRED_FLUSH because of the type.
So unless I'm missing something I think either we have to
1) be more flexible in what ops->default_domain_type() is allowed to return as this patch does
or
2) iommu core needs to look at the return from ops->default_domain_type() and decide whether it's OK to convert IOMMU_DOMAIN_DMA->IOMMU_DOMAIN_DMA_FQ based on strict setting. This removes the decision from the individual drivers and dma-iommu can later decide whether or not to use it or not based on IOMMU_CAP_DEFERRED_FLUSH? But would also affect other users of def_domain_type() today that perhaps did not want DMA_FQ? Unsure. What I mean is something like (untested):
diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
index 6bdede4177ff..275daa7f819d 100644
--- a/drivers/iommu/iommu.c
+++ b/drivers/iommu/iommu.c
@@ -1744,9 +1744,11 @@ static int iommu_get_def_domain_type(struct iommu_group *group,
*/
type = ops->default_domain->type;
} else {
- if (ops->def_domain_type)
+ if (ops->def_domain_type) {
type = ops->def_domain_type(dev);
- else
+ if (type == IOMMU_DOMAIN_DMA && !iommu_dma_strict)
+ type = IOMMU_DOMAIN_DMA_FQ;
+ } else
return cur_type;
}
if (!type || cur_type == type)
