On 12/4/2024 8:51 PM, Rob Clark wrote:
On Wed, Dec 4, 2024 at 3:27 AM Bibek Kumar Patro
<quic_bibekkum@xxxxxxxxxxx> wrote:
On 11/22/2024 10:33 PM, Rob Clark wrote:
On Fri, Nov 22, 2024 at 8:19 AM Bibek Kumar Patro
<quic_bibekkum@xxxxxxxxxxx> wrote:
On 11/20/2024 10:47 PM, Rob Clark wrote:
On Thu, Nov 14, 2024 at 8:10 AM Bibek Kumar Patro
<quic_bibekkum@xxxxxxxxxxx> wrote:
Add an adreno-smmu-priv interface for drm/msm to call
into arm-smmu-qcom and initiate the PRR bit setup or reset
sequence as per request.
This will be used by GPU to setup the PRR bit and related
configuration registers through adreno-smmu private
interface instead of directly poking the smmu hardware.
Suggested-by: Rob Clark <robdclark@xxxxxxxxx>
Signed-off-by: Bibek Kumar Patro <quic_bibekkum@xxxxxxxxxxx>
---
drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c | 37 ++++++++++++++++++++++
drivers/iommu/arm/arm-smmu/arm-smmu.h | 2 ++
include/linux/adreno-smmu-priv.h | 14 ++++++++
3 files changed, 53 insertions(+)
diff --git a/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c b/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c
index d26f5aea248e..0e4f3fbda961 100644
--- a/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c
+++ b/drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c
@@ -16,6 +16,8 @@
#define QCOM_DUMMY_VAL -1
+#define GFX_ACTLR_PRR (1 << 5)
+
static struct qcom_smmu *to_qcom_smmu(struct arm_smmu_device *smmu)
{
return container_of(smmu, struct qcom_smmu, smmu);
@@ -99,6 +101,32 @@ static void qcom_adreno_smmu_resume_translation(const void *cookie, bool termina
arm_smmu_cb_write(smmu, cfg->cbndx, ARM_SMMU_CB_RESUME, reg);
}
+static void qcom_adreno_smmu_set_prr_bit(const void *cookie, bool set)
+{
+ struct arm_smmu_domain *smmu_domain = (void *)cookie;
+ struct arm_smmu_device *smmu = smmu_domain->smmu;
+ struct arm_smmu_cfg *cfg = &smmu_domain->cfg;
+ u32 reg = 0;
+
+ reg = arm_smmu_cb_read(smmu, cfg->cbndx, ARM_SMMU_CB_ACTLR);
+ reg &= ~GFX_ACTLR_PRR;
+ if (set)
+ reg |= FIELD_PREP(GFX_ACTLR_PRR, 1);
+ arm_smmu_cb_write(smmu, cfg->cbndx, ARM_SMMU_CB_ACTLR, reg);
+}
+
+static void qcom_adreno_smmu_set_prr_addr(const void *cookie, phys_addr_t page_addr)
+{
+ struct arm_smmu_domain *smmu_domain = (void *)cookie;
+ struct arm_smmu_device *smmu = smmu_domain->smmu;
+
+ writel_relaxed(lower_32_bits(page_addr),
+ smmu->base + ARM_SMMU_GFX_PRR_CFG_LADDR);
+
+ writel_relaxed(upper_32_bits(page_addr),
+ smmu->base + ARM_SMMU_GFX_PRR_CFG_UADDR);
+}
+
#define QCOM_ADRENO_SMMU_GPU_SID 0
static bool qcom_adreno_smmu_is_gpu_device(struct device *dev)
@@ -210,6 +238,7 @@ static bool qcom_adreno_can_do_ttbr1(struct arm_smmu_device *smmu)
static int qcom_adreno_smmu_init_context(struct arm_smmu_domain *smmu_domain,
struct io_pgtable_cfg *pgtbl_cfg, struct device *dev)
{
+ const struct device_node *np = smmu_domain->smmu->dev->of_node;
struct adreno_smmu_priv *priv;
smmu_domain->cfg.flush_walk_prefer_tlbiasid = true;
@@ -239,6 +268,14 @@ static int qcom_adreno_smmu_init_context(struct arm_smmu_domain *smmu_domain,
priv->get_fault_info = qcom_adreno_smmu_get_fault_info;
priv->set_stall = qcom_adreno_smmu_set_stall;
priv->resume_translation = qcom_adreno_smmu_resume_translation;
+ priv->set_prr_bit = NULL;
+ priv->set_prr_addr = NULL;
+
+ if (of_device_is_compatible(np, "qcom,smmu-500") &&
+ of_device_is_compatible(np, "qcom,adreno-smmu")) {
fwiw, it seems like PRR actually works on sc7180, which is _not_
mmu-500, so I guess the support actually goes back further.
As I checked sc7180 was on previous variant of SMMU,
so targets on this variant have different steps to set PRR bit.
<Do not have both prr bit and PRR page registers>.
Experimentally, I have to set both the PRR LADDR/UADDR regs and
ACTLR.PRR bit on sc7180 to get the sparse-residency tests passing. So
the underlying hw seems to work in the same way as mmu-500. _But_
this is on a sc7180 chromebook, the situation might be different
depending on fw on things that have QC hyp.
I checked on sc7180 chipset which is based on the smmu-v2,
this time by looking for these offsets specifically. I can see the
nomenclature of the PRR related registers are a bit different
compared to MMU-500 variant.
Also the implementation register is 64 bit instead of
dual 32 bit as in case of MMU-500. and PRR bit is not marked in
ACTLR register offset.
Interesting.. in my experiments it needed both the ACTLR.PRR bit set
and the LADDR/UADDR. Maybe it was just happy coincidence that two 32b
writes worked?
I think it's similar to case we discussed [1] in v11
on why not good idea to use writeq on 32bit regs.
So some corner cases for unaligned 32b addresses might get truncated.
[1]:
https://lore.kernel.org/all/ae35bf9b-4401-4a99-abd7-c0d9d399a46b@xxxxxxxxxxx/#:~:text=So%20I%20think-,writeq,-for%2064bit%20write
So turns out PRR is supported but with some modification and
can be carried out with same compatible based approach only - as per
my understanding.
In current patch plan was to provide support for MMU-500 based targets
and won't break any legacy targets, so we can take the PRR support
for legacy targets in different series once our evaluation is done on
smmu-v2 ?
I guess it wouldn't be the end of the world, but it would mean that
drm would need to expose PRR support to userspace separately from
sparse binding support. Maybe we need to do that anyways. (I'm not
really sure how many different a6xx+smmu-v2 devices are out there, but
I guess they are all based on the same generation of snapdragon?)
We would need to audit on all the smmu-v2 based devices using a6xx
version before I can comment on this, but mostly it should be in the
same generation of snapdragon (scxx/smxx).
Thanks & regards,
bibek
BR,
-R
We would explore more on this PRR feature for smmu-v2 based targets,
before supporting it.
Thanks & regards,
Bibek
It's MMU-500 targets only where PRR support is with both PRR bit
and PRR page addr registers. As I was re-visiting our discussions on v13
regarding this - I remember that's why we started using the SMMU-
compatible string based PRR procedure selection instead of the reserved-
memory node. [1] i.e Based on SMMU variant (as selected by compatible
string), specific sequence will be selected.
So for now only MMU-500 based selection has been supported as part of
this series and will add subsequent support for other SMMU-variants
thereafter.
I'm curious if we can just rely on this being supported by any hw that
has a6xx or newer?
I'd need to check on targets which will be based on a6xx onwards, on
what will be the procedure planned to support PRR feature. I'll update
the information over this space.
One of the open questions about the drm/msm sparse-memory patchset is
whether we need to expose to userspace whether PRR is supported, or if
we can just rely on sparse-binding support implying sparse residency
(ie, PRR) support. All a6xx devices support per-process pgtables,
which is the only requirement for basic sparseBinding support. But
PRR is required to additionally expose
sparseResidencyBuffer/sparseResidencyImage2D.
I think, long term, turnip probably will want to drop support for
older kernels and remove support for legacy buffer mapping. But if we
have some a6xx devices without PRR, then to do that we'd need to
decouple sparse binding and sparse residency. (Vulkan allows a driver
to expose the former without the latter.)
BR,
-R
[1]:
https://lore.kernel.org/all/5790afa3-f9c0-4720-9804-8a7ff3d91854@xxxxxxxxxxx/#:~:text=%3E%20I%20guess%20if,part%20as%20well.
Thanks & regards,
Bibek
BR,
-R
+ priv->set_prr_bit = qcom_adreno_smmu_set_prr_bit;
+ priv->set_prr_addr = qcom_adreno_smmu_set_prr_addr;
+ }
return 0;
}
diff --git a/drivers/iommu/arm/arm-smmu/arm-smmu.h b/drivers/iommu/arm/arm-smmu/arm-smmu.h
index e2aeb511ae90..2dbf3243b5ad 100644
--- a/drivers/iommu/arm/arm-smmu/arm-smmu.h
+++ b/drivers/iommu/arm/arm-smmu/arm-smmu.h
@@ -154,6 +154,8 @@ enum arm_smmu_cbar_type {
#define ARM_SMMU_SCTLR_M BIT(0)
#define ARM_SMMU_CB_ACTLR 0x4
+#define ARM_SMMU_GFX_PRR_CFG_LADDR 0x6008
+#define ARM_SMMU_GFX_PRR_CFG_UADDR 0x600C
#define ARM_SMMU_CB_RESUME 0x8
#define ARM_SMMU_RESUME_TERMINATE BIT(0)
diff --git a/include/linux/adreno-smmu-priv.h b/include/linux/adreno-smmu-priv.h
index c637e0997f6d..614665153b3e 100644
--- a/include/linux/adreno-smmu-priv.h
+++ b/include/linux/adreno-smmu-priv.h
@@ -50,6 +50,18 @@ struct adreno_smmu_fault_info {
* the GPU driver must call resume_translation()
* @resume_translation: Resume translation after a fault
*
+ * *CAUTION* : PRR callbacks (set_prr_bit/set_prr_addr) are NULL terminated for
+ * targets without PRR support. Exercise caution and verify target
+ * capabilities before invoking these callbacks to prevent potential
+ * runtime errors or unexpected behavior.
+ *
+ * @set_prr_bit: Extendible interface to be used by GPU to modify the
+ * ACTLR register bits, currently used to configure
+ * Partially-Resident-Region (PRR) bit for feature's
+ * setup and reset sequence as requested.
+ * @set_prr_addr: Configure the PRR_CFG_*ADDR register with the
+ * physical address of PRR page passed from
+ * GPU driver.
*
* The GPU driver (drm/msm) and adreno-smmu work together for controlling
* the GPU's SMMU instance. This is by necessity, as the GPU is directly
@@ -67,6 +79,8 @@ struct adreno_smmu_priv {
void (*get_fault_info)(const void *cookie, struct adreno_smmu_fault_info *info);
void (*set_stall)(const void *cookie, bool enabled);
void (*resume_translation)(const void *cookie, bool terminate);
+ void (*set_prr_bit)(const void *cookie, bool set);
+ void (*set_prr_addr)(const void *cookie, phys_addr_t page_addr);
};
#endif /* __ADRENO_SMMU_PRIV_H */
--
2.34.1
