iommu/arm-smmu: Optimize ->tlb_flush_walk() for qcom implementation
Subsystems (4)
PATCH v4 1/1 iommu/arm-smmu: Optimize ->tlb_flush_walk() for qcom implementation
Currently for iommu_unmap() of large scatter-gather list with page size elements, the majority of time is spent in flushing of partial walks in __arm_lpae_unmap() which is a VA based TLB invalidation invalidating page-by-page on iommus like arm-smmu-v2 (TLBIVA).
For example: to unmap a 32MB scatter-gather list with page size elements (8192 entries), there are 16->2MB buffer unmaps based on the pgsize (2MB for 4K granule) and each of 2MB will further result in 512 TLBIVAs (2MB/4K) resulting in a total of 8192 TLBIVAs (512*16) for 16->2MB causing a huge overhead.
On qcom implementation, there are several performance improvements for TLB cache invalidations in HW like wait-for-safe (for realtime clients such as camera and display) and few others to allow for cache lookups/updates when TLBI is in progress for the same context bank. So the cost of over-invalidation is less compared to the unmap latency on several usecases like camera which deals with large buffers. So, ASID based TLB invalidations (TLBIASID) can be used to invalidate the entire context for partial walk flush thereby improving the unmap latency.
Non-strict mode can use this by default for all platforms given its all about over-invalidation saving time on individual unmaps and non-deterministic generally.
For this example of 32MB scatter-gather list unmap, this change results in just 16 ASID based TLB invalidations (TLBIASIDs) as opposed to 8192 TLBIVAs thereby increasing the performance of unmaps drastically.
Test on QTI SM8150 SoC for 10 iterations of iommu_{map_sg}/unmap:
(average over 10 iterations)
Before this optimization:
size iommu_map_sg iommu_unmap
4K 2.067 us 1.854 us
64K 9.598 us 8.802 us
1M 148.890 us 130.718 us
2M 305.864 us 67.291 us
12M 1793.604 us 390.838 us
16M 2386.848 us 518.187 us
24M 3563.296 us 775.989 us
32M 4747.171 us 1033.364 usAfter this optimization:
size iommu_map_sg iommu_unmap
4K 1.723 us 1.765 us
64K 9.880 us 8.869 us
1M 155.364 us 135.223 us
2M 303.906 us 5.385 us
12M 1786.557 us 21.250 us
16M 2391.890 us 27.437 us
24M 3570.895 us 39.937 us
32M 4755.234 us 51.797 usReal world data also shows big difference in unmap performance as below:
There were reports of camera frame drops because of high overhead in iommu unmap without this optimization because of frequent unmaps issued by camera of about 100MB/s taking more than 100ms thereby causing frame drops.
---
Changes in v4:
* Use a flag in struct arm_smmu_cfg to prefer TLBIASID (Will)
Changes in v3:
* Move the logic to arm-smmu driver from io-pgtable (Robin)
* Use a new set of iommu_flush_ops->arm_smmu_s1_tlb_impl_ops and use it for qcom impl
Changes in v2:
* Add a quirk to choose tlb_flush_all in partial walk flush
* Set the quirk for QTI SoC implementation
--- drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c | 11 +++++++++++ drivers/iommu/arm/arm-smmu/arm-smmu.c | 17 +++++++++++++---- drivers/iommu/arm/arm-smmu/arm-smmu.h | 1 + 3 files changed, 25 insertions(+), 4 deletions(-)
drivers/iommu/arm/arm-smmu/arm-smmu-qcom.c +11 -0
@@ -193,6 +193,8 @@ static int qcom_adreno_smmu_init_context(struct arm_smmu_domain *smmu_domain, {struct adreno_smmu_priv *priv; + smmu_domain->cfg.flush_walk_prefer_tlbiasid = true; +/* Only enable split pagetables for the GPU device (SID 0) */ if (!qcom_adreno_smmu_is_gpu_device (dev))return 0 ;
@@ -235,6 +237,14 @@ static const struct of_device_id qcom_smmu_client_of_match[] __maybe_unused = { { } }; +static int qcom_smmu_init_context (struct arm_smmu_domain *smmu_domain, +struct io_pgtable_cfg *pgtbl_cfg,struct device *dev) +{ + smmu_domain->cfg.flush_walk_prefer_tlbiasid = true; + +return 0 ; +} +static int qcom_smmu_cfg_probe (struct arm_smmu_device *smmu) {unsigned int last_s2cr =ARM_SMMU_GR0_S2CR (smmu->num_mapping_groups -1 );
@@ -358,6 +368,7 @@ static int qcom_smmu500_reset(struct arm_smmu_device *smmu) }static const struct arm_smmu_impl qcom_smmu_impl = { + .init_context = qcom_smmu_init_context, .cfg_probe = qcom_smmu_cfg_probe, .def_domain_type = qcom_smmu_def_domain_type, .reset = qcom_smmu500_reset,
drivers/iommu/arm/arm-smmu/arm-smmu.c +13 -4
@@ -327,9 +327,16 @@ static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size,static void arm_smmu_tlb_inv_walk_s1 (unsigned long iova,size_t size,size_t granule,void *cookie) { -arm_smmu_tlb_inv_range_s1 (iova, size, granule, cookie, -ARM_SMMU_CB_S1_TLBIVA ); -arm_smmu_tlb_sync_context (cookie); +struct arm_smmu_domain *smmu_domain = cookie; +struct arm_smmu_cfg *cfg = &smmu_domain->cfg; + +if (cfg->flush_walk_prefer_tlbiasid) { +arm_smmu_tlb_inv_context_s1 (cookie); + }else { +arm_smmu_tlb_inv_range_s1 (iova, size, granule, cookie, +ARM_SMMU_CB_S1_TLBIVA ); +arm_smmu_tlb_sync_context (cookie); + } } static void arm_smmu_tlb_add_page_s1(struct iommu_iotlb_gather *gather,
@@ -765,8 +772,10 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain, .iommu_dev = smmu->dev, }; -if (!iommu_get_dma_strict (domain)) +if (!iommu_get_dma_strict (domain)) { pgtbl_cfg.quirks |=IO_PGTABLE_QUIRK_NON_STRICT ; + cfg->flush_walk_prefer_tlbiasid = true; + }if (smmu->impl && smmu->impl->init_context) { ret = smmu->impl->init_context (smmu_domain, &pgtbl_cfg, dev);
drivers/iommu/arm/arm-smmu/arm-smmu.h +1 -0
@@ -346,6 +346,7 @@ struct arm_smmu_cfg { };enum arm_smmu_cbar_type cbar;enum arm_smmu_context_fmt fmt; +bool flush_walk_prefer_tlbiasid; };#define ARM_SMMU_INVALID_IRPTNDX 0xff
Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation
Will Deacon
maintainerQuoted 6 lines · Sai Prakash Ranjan
drivers/iommu/arm/arm-smmu/arm-smmu.c +0
-0@@ -327,9 +327,16 @@ static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size, static void arm_smmu_tlb_inv_walk_s1(unsigned long iova, size_t size,
drivers/iommu/arm/arm-smmu/arm-smmu.c static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size, · line 334
Quoted 10 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +5
-3size_t granule, void *cookie) { -arm_smmu_tlb_inv_range_s1 (iova ,size ,granule ,cookie , -ARM_SMMU_CB_S1_TLBIVA ); -arm_smmu_tlb_sync_context (cookie); + struct arm_smmu_domain *smmu_domain = cookie; +struct arm_smmu_cfg *cfg = &smmu_domain->cfg; + +if (cfg->flush_walk_prefer_tlbiasid) { +arm_smmu_tlb_inv_context_s1 (cookie);
Hmm, this introduces an unconditional wmb() if tlbiasid is preferred. I think that should be predicated on ARM_SMMU_FEAT_COHERENT_WALK like it is for the by-VA ops. Worth doing as a separate patch.
Quoted 9 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +5
-0+ }else { +arm_smmu_tlb_inv_range_s1 (iova, size, granule, cookie, +ARM_SMMU_CB_S1_TLBIVA ); +arm_smmu_tlb_sync_context (cookie); + } } static void arm_smmu_tlb_add_page_s1(struct iommu_iotlb_gather *gather,
drivers/iommu/arm/arm-smmu/arm-smmu.c static int arm_smmu_init_domain_context(struct iommu_domain *domain, · line 777
Quoted 7 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +2
-1@@ -765,8 +772,10 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain, .iommu_dev = smmu->dev, }; -if (!iommu_get_dma_strict (domain)) +if (!iommu_get_dma_strict (domain)) { pgtbl_cfg.quirks |=IO_PGTABLE_QUIRK_NON_STRICT ; + cfg->flush_walk_prefer_tlbiasid = true;
This is going to interact badly with Robin's series to allow dynamic transition to non-strict mode, as we don't have a mechanism to switch over to the by-ASID behaviour. Yes, it should _work_, but it's ugly having different TLBI behaviour just because of the how the domain became non-strict.
Robin -- I think this originated from your idea at [1]. Any idea how to make it work with your other series, or shall we drop this part for now and leave the TLB invalidation behaviour the same for now?
Will
Quoted 14 lines · Will Deacon
Quoted 13 lines · Sai Prakash Ranjan
drivers/iommu/arm/arm-smmu/arm-smmu.c +1
-3@@ -327,9 +327,16 @@ static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size, static void arm_smmu_tlb_inv_walk_s1(unsigned long iova, size_t size, size_t granule, void *cookie) { -arm_smmu_tlb_inv_range_s1 (iova ,size ,granule ,cookie , -ARM_SMMU_CB_S1_TLBIVA ); -arm_smmu_tlb_sync_context (cookie); + struct arm_smmu_domain *smmu_domain = cookie;
Quoted 9 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size, · line 334
Quoted 4 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +4
-0+struct arm_smmu_cfg *cfg = &smmu_domain->cfg; + +if (cfg->flush_walk_prefer_tlbiasid) { +arm_smmu_tlb_inv_context_s1 (cookie);Hmm, this introduces an unconditional
wmb()if tlbiasid is preferred. I think that should be predicated onARM_SMMU_FEAT_COHERENT_WALKlike it is for the by-VA ops. Worth doing as a separate patch.
Ok I will keep this as-is for now then.
Quoted 25 lines
Quoted 10 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +5
-0+ }else { +arm_smmu_tlb_inv_range_s1 (iova, size, granule, cookie, +ARM_SMMU_CB_S1_TLBIVA ); +arm_smmu_tlb_sync_context (cookie); + } }static void arm_smmu_tlb_add_page_s1(struct
iommu_iotlb_gather
*gather,drivers/iommu/arm/arm-smmu/arm-smmu.c static int arm_smmu_init_domain_context(struct · line 778
Quoted 8 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +2
-1@@ -765,8 +772,10 @@ static int arm_smmu_init_domain_context(structiommu_domain *domain, .iommu_dev = smmu->dev, }; -if (!iommu_get_dma_strict (domain)) +if (!iommu_get_dma_strict (domain)) { pgtbl_cfg.quirks |=IO_PGTABLE_QUIRK_NON_STRICT ; + cfg->flush_walk_prefer_tlbiasid = true;This is going to interact badly with Robin's series to allow dynamic transition to non-strict mode, as we don't have a mechanism to switch over to the by-ASID behaviour. Yes, it should _work_, but it's ugly having different TLBI behaviour just because of the how the domain became non-strict.
Quoted 7 lines
Robin -- I think this originated from your idea at [1]. Any idea how to make it work with your other series, or shall we drop this part for now and leave the TLB invalidation behaviour the same for now?
Will
Right, I think we can drop this non-strict change for now because it also makes it a pain to backport it to 5.4/5.10 kernels because of large number of changes in dma apis in recent kernels. I will let you and Robin decide if it's ok to drop this change and introduce it later with a different patch.
Thanks,
Sai
Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation
Robin Murphy
reviewerQuoted 45 lines · Will Deacon
Quoted 6 lines · Sai Prakash Ranjan
drivers/iommu/arm/arm-smmu/arm-smmu.c +0
-0@@ -327,9 +327,16 @@ static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size, static void arm_smmu_tlb_inv_walk_s1(unsigned long iova, size_t size,drivers/iommu/arm/arm-smmu/arm-smmu.c static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size, · line 334
Quoted 10 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +5
-3size_t granule, void *cookie) { -arm_smmu_tlb_inv_range_s1 (iova ,size ,granule ,cookie , -ARM_SMMU_CB_S1_TLBIVA ); -arm_smmu_tlb_sync_context (cookie); + struct arm_smmu_domain *smmu_domain = cookie; +struct arm_smmu_cfg *cfg = &smmu_domain->cfg; + +if (cfg->flush_walk_prefer_tlbiasid) { +arm_smmu_tlb_inv_context_s1 (cookie);Hmm, this introduces an unconditional
wmb()if tlbiasid is preferred. I think that should be predicated onARM_SMMU_FEAT_COHERENT_WALKlike it is for the by-VA ops. Worth doing as a separate patch.Quoted 9 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +5
-0+ }else { +arm_smmu_tlb_inv_range_s1 (iova, size, granule, cookie, +ARM_SMMU_CB_S1_TLBIVA ); +arm_smmu_tlb_sync_context (cookie); + } } static void arm_smmu_tlb_add_page_s1(struct iommu_iotlb_gather *gather,drivers/iommu/arm/arm-smmu/arm-smmu.c static int arm_smmu_init_domain_context(struct iommu_domain *domain, · line 777
Quoted 7 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +2
-1@@ -765,8 +772,10 @@ static int arm_smmu_init_domain_context(struct iommu_domain *domain, .iommu_dev = smmu->dev, }; -if (!iommu_get_dma_strict (domain)) +if (!iommu_get_dma_strict (domain)) { pgtbl_cfg.quirks |=IO_PGTABLE_QUIRK_NON_STRICT ; + cfg->flush_walk_prefer_tlbiasid = true;This is going to interact badly with Robin's series to allow dynamic transition to non-strict mode, as we don't have a mechanism to switch over to the by-ASID behaviour. Yes, it should _work_, but it's ugly having different TLBI behaviour just because of the how the domain became non-strict.
Quoted 3 lines
Robin -- I think this originated from your idea at [1]. Any idea how to make it work with your other series, or shall we drop this part for now and leave the TLB invalidation behaviour the same for now?
Yeah, I'd say drop it - I'm currently half an hour into a first attempt at removing io_pgtable_tlb_flush_walk() entirely, which would make it moot for non-strict anyway.
Robin.
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu
Quoted 48 lines · Robin Murphy
Quoted 47 lines · Will Deacon
Quoted 8 lines · Sai Prakash Ranjan
drivers/iommu/arm/arm-smmu/arm-smmu.c +0
-0@@ -327,9 +327,16 @@ static void arm_smmu_tlb_inv_range_s2(unsigned long iova, size_t size, static void arm_smmu_tlb_inv_walk_s1(unsigned long iova, size_t size,drivers/iommu/arm/arm-smmu/arm-smmu.c static void arm_smmu_tlb_inv_range_s2(unsigned · line 336
Quoted 10 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +5
-3size_t granule, void *cookie) { -arm_smmu_tlb_inv_range_s1 (iova ,size ,granule ,cookie , -ARM_SMMU_CB_S1_TLBIVA ); -arm_smmu_tlb_sync_context (cookie); + struct arm_smmu_domain *smmu_domain = cookie; +struct arm_smmu_cfg *cfg = &smmu_domain->cfg; + +if (cfg->flush_walk_prefer_tlbiasid) { +arm_smmu_tlb_inv_context_s1 (cookie);Hmm, this introduces an unconditional
wmb()if tlbiasid is preferred. I think that should be predicated onARM_SMMU_FEAT_COHERENT_WALKlike it is for the by-VA ops. Worth doing as a separate patch.Quoted 9 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +5
-0+ }else { +arm_smmu_tlb_inv_range_s1 (iova, size, granule, cookie, +ARM_SMMU_CB_S1_TLBIVA ); +arm_smmu_tlb_sync_context (cookie); + } }static void arm_smmu_tlb_add_page_s1(struct
iommu_iotlb_gather
*gather,drivers/iommu/arm/arm-smmu/arm-smmu.c static int arm_smmu_init_domain_context(struct · line 778
Quoted 7 lines
drivers/iommu/arm/arm-smmu/arm-smmu.c +2
-0@@ -765,8 +772,10 @@ static int arm_smmu_init_domain_context(structiommu_domain *domain, .iommu_dev = smmu->dev, }; - if (!iommu_get_dma_strict(domain)) +if (!iommu_get_dma_strict (domain)) { pgtbl_cfg.quirks |=IO_PGTABLE_QUIRK_NON_STRICT ; + cfg->flush_walk_prefer_tlbiasid = true;This is going to interact badly with Robin's series to allow dynamic transition to non-strict mode, as we don't have a mechanism to switch over to the by-ASID behaviour. Yes, it should _work_, but it's ugly having different TLBI behaviour just because of the how the domain became non-strict.
Quoted 8 lines
Quoted 3 lines
Robin -- I think this originated from your idea at [1]. Any idea how to make it work with your other series, or shall we drop this part for now and leave the TLB invalidation behaviour the same for now?
Yeah, I'd say drop it - I'm currently half an hour into a first attempt at removing
io_pgtable_tlb_flush_walk()entirely, which would make it moot for non-strict anyway.
I have dropped it and sent a v5.
Thanks,
Sai
Signature · 2 lines
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation