[PATCH V2 1/2] iommu: Optimize IOMMU UnMap

[PATCH V2 1/2] iommu: Optimize IOMMU UnMap

[Ashish Mhetre]

The current __arm_lpae_unmap() function calls dma_sync() on individual
PTEs after clearing them. Overall unmap performance can be improved by
around 25% for large buffer sizes by combining the syncs for adjacent
leaf entries.
Optimize the unmap time by clearing all the leaf entries and issuing a
single dma_sync() for them.
Below is detailed analysis of average unmap latency(in us) with and
without this optimization obtained by running dma_map_benchmark for
different buffer sizes.

		UnMap Latency(us)
Size	Without		With		% gain with
	optimiztion	optimization	optimization

4KB	3		3		0
8KB	4		3.8		5
16KB	6.1		5.4		11.48
32KB	10.2		8.5		16.67
64KB	18.5		14.9		19.46
128KB	35		27.5		21.43
256KB	67.5		52.2		22.67
512KB	127.9		97.2		24.00
1MB	248.6		187.4		24.62
2MB	65.5		65.5		0
4MB	119.2		119		0.17

Signed-off-by: Ashish Mhetre <amhetre@nvidia.com>
---
Changes in V2:
- Updated the commit message to be imperative.
- Fixed ptep at incorrect index getting cleared for non-leaf entries.
---
 drivers/iommu/io-pgtable-arm.c | 34 +++++++++++++++++++++-------------
 1 file changed, 21 insertions(+), 13 deletions(-)

diff --git a/drivers/iommu/io-pgtable-arm.c b/drivers/iommu/io-pgtable-arm.c
index f5d9fd1f45bf..32401948b980 100644
--- a/drivers/iommu/io-pgtable-arm.c
+++ b/drivers/iommu/io-pgtable-arm.c
@@ -274,13 +274,15 @@ static void __arm_lpae_sync_pte(arm_lpae_iopte *ptep, int num_entries,
 				   sizeof(*ptep) * num_entries, DMA_TO_DEVICE);
 }
 
-static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct io_pgtable_cfg *cfg)
+static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct io_pgtable_cfg *cfg, int num_entries)
 {
+	int i;
 
-	*ptep = 0;
+	for (i = 0; i < num_entries; i++)
+		ptep[i] = 0;
 
 	if (!cfg->coherent_walk)
-		__arm_lpae_sync_pte(ptep, 1, cfg);
+		__arm_lpae_sync_pte(ptep, num_entries, cfg);
 }
 
 static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
@@ -635,9 +637,10 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
 			       unsigned long iova, size_t size, size_t pgcount,
 			       int lvl, arm_lpae_iopte *ptep)
 {
+	bool gather_queued;
 	arm_lpae_iopte pte;
 	struct io_pgtable *iop = &data->iop;
-	int i = 0, num_entries, max_entries, unmap_idx_start;
+	int i = 0, j = 0, num_entries, max_entries, unmap_idx_start;
 
 	/* Something went horribly wrong and we ran out of page table */
 	if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
@@ -652,28 +655,33 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
 	/* If the size matches this level, we're in the right place */
 	if (size == ARM_LPAE_BLOCK_SIZE(lvl, data)) {
 		max_entries = ARM_LPAE_PTES_PER_TABLE(data) - unmap_idx_start;
+		gather_queued = iommu_iotlb_gather_queued(gather);
 		num_entries = min_t(int, pgcount, max_entries);
 
-		while (i < num_entries) {
-			pte = READ_ONCE(*ptep);
+		/* Find and handle non-leaf entries */
+		for (i = 0; i < num_entries; i++) {
+			pte = READ_ONCE(ptep[i]);
 			if (WARN_ON(!pte))
 				break;
 
-			__arm_lpae_clear_pte(ptep, &iop->cfg);
-
 			if (!iopte_leaf(pte, lvl, iop->fmt)) {
+				__arm_lpae_clear_pte(&ptep[i], &iop->cfg, 1);
+
 				/* Also flush any partial walks */
 				io_pgtable_tlb_flush_walk(iop, iova + i * size, size,
 							  ARM_LPAE_GRANULE(data));
 				__arm_lpae_free_pgtable(data, lvl + 1, iopte_deref(pte, data));
-			} else if (!iommu_iotlb_gather_queued(gather)) {
-				io_pgtable_tlb_add_page(iop, gather, iova + i * size, size);
 			}
-
-			ptep++;
-			i++;
 		}
 
+		/* Clear the remaining entries */
+		if (i)
+			__arm_lpae_clear_pte(ptep, &iop->cfg, i);
+
+		if (!gather_queued)
+			for (j = 0; j < i; j++)
+				io_pgtable_tlb_add_page(iop, gather, iova + j * size, size);
+
 		return i * size;
 	} else if (iopte_leaf(pte, lvl, iop->fmt)) {
 		/*
-- 
2.25.1

[PATCH V2 2/2] include: linux: Update gather only if it's not NULL

[Ashish Mhetre]

Gather can be NULL when unmap is called for freeing old table while
mapping. If it's NULL then there is no need to add page for syncing
the TLB.

Signed-off-by: Ashish Mhetre <amhetre@nvidia.com>
---
 include/linux/iommu.h | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/include/linux/iommu.h b/include/linux/iommu.h
index 4d47f2c33311..2a28c1ef8517 100644
--- a/include/linux/iommu.h
+++ b/include/linux/iommu.h
@@ -928,6 +928,9 @@ static inline void iommu_iotlb_gather_add_page(struct iommu_domain *domain,
 					       struct iommu_iotlb_gather *gather,
 					       unsigned long iova, size_t size)
 {
+	if (!gather)
+		return;
+
 	/*
 	 * If the new page is disjoint from the current range or is mapped at
 	 * a different granularity, then sync the TLB so that the gather
-- 
2.25.1

Re: [PATCH V2 2/2] include: linux: Update gather only if it's not NULL

[Robin Murphy]

On 01/08/2024 4:34 am, Ashish Mhetre wrote:
> Gather can be NULL when unmap is called for freeing old table while
> mapping. If it's NULL then there is no need to add page for syncing
> the TLB.

But that's only because __arm_lpae_unmap() is now choosing to 
over-invalidate the table entries for simplicity. I think it would make 
more sense to handle that at the callsite, e.g. "if (gather && 
!iommu_iotlb_gather_queued(gather))".

Also doesn't this mean that bisection is broken as-is since patch #1 on 
its own is going to blow up dereferencing NULL->pgsize when it gets here?

Thanks,
Robin.

> Signed-off-by: Ashish Mhetre <amhetre@nvidia.com>
> ---
>   include/linux/iommu.h | 3 +++
>   1 file changed, 3 insertions(+)
> 
> diff --git a/include/linux/iommu.h b/include/linux/iommu.h
> index 4d47f2c33311..2a28c1ef8517 100644
> --- a/include/linux/iommu.h
> +++ b/include/linux/iommu.h
> @@ -928,6 +928,9 @@ static inline void iommu_iotlb_gather_add_page(struct iommu_domain *domain,
>   					       struct iommu_iotlb_gather *gather,
>   					       unsigned long iova, size_t size)
>   {
> +	if (!gather)
> +		return;
> +
>   	/*
>   	 * If the new page is disjoint from the current range or is mapped at
>   	 * a different granularity, then sync the TLB so that the gather

Re: [PATCH V2 2/2] include: linux: Update gather only if it's not NULL

[Ashish Mhetre]

On 8/1/2024 10:12 PM, Robin Murphy wrote:
> External email: Use caution opening links or attachments
>
>
> On 01/08/2024 4:34 am, Ashish Mhetre wrote:
>> Gather can be NULL when unmap is called for freeing old table while
>> mapping. If it's NULL then there is no need to add page for syncing
>> the TLB.
>
> But that's only because __arm_lpae_unmap() is now choosing to
> over-invalidate the table entries for simplicity. I think it would make
> more sense to handle that at the callsite, e.g. "if (gather &&
> !iommu_iotlb_gather_queued(gather))".
>
> Also doesn't this mean that bisection is broken as-is since patch #1 on
> its own is going to blow up dereferencing NULL->pgsize when it gets here?
>
> Thanks,
> Robin.
>
Yes, gather can be NULL because of patch #1 only. I will fix it at 
callsite only
with a single patch rather than splitting.

>> Signed-off-by: Ashish Mhetre <amhetre@nvidia.com>
>> ---
>>   include/linux/iommu.h | 3 +++
>>   1 file changed, 3 insertions(+)
>>
>> diff --git a/include/linux/iommu.h b/include/linux/iommu.h
>> index 4d47f2c33311..2a28c1ef8517 100644
>> --- a/include/linux/iommu.h
>> +++ b/include/linux/iommu.h
>> @@ -928,6 +928,9 @@ static inline void 
>> iommu_iotlb_gather_add_page(struct iommu_domain *domain,
>>                                              struct 
>> iommu_iotlb_gather *gather,
>>                                              unsigned long iova, 
>> size_t size)
>>   {
>> +     if (!gather)
>> +             return;
>> +
>>       /*
>>        * If the new page is disjoint from the current range or is 
>> mapped at
>>        * a different granularity, then sync the TLB so that the gather

Re: [PATCH V2 1/2] iommu: Optimize IOMMU UnMap

[Robin Murphy]

On 01/08/2024 4:34 am, Ashish Mhetre wrote:
> The current __arm_lpae_unmap() function calls dma_sync() on individual
> PTEs after clearing them. Overall unmap performance can be improved by
> around 25% for large buffer sizes by combining the syncs for adjacent
> leaf entries.
> Optimize the unmap time by clearing all the leaf entries and issuing a
> single dma_sync() for them.
> Below is detailed analysis of average unmap latency(in us) with and
> without this optimization obtained by running dma_map_benchmark for
> different buffer sizes.
> 
> 		UnMap Latency(us)
> Size	Without		With		% gain with
> 	optimiztion	optimization	optimization
> 
> 4KB	3		3		0
> 8KB	4		3.8		5
> 16KB	6.1		5.4		11.48
> 32KB	10.2		8.5		16.67
> 64KB	18.5		14.9		19.46
> 128KB	35		27.5		21.43
> 256KB	67.5		52.2		22.67
> 512KB	127.9		97.2		24.00
> 1MB	248.6		187.4		24.62
> 2MB	65.5		65.5		0
> 4MB	119.2		119		0.17
> 
> Signed-off-by: Ashish Mhetre <amhetre@nvidia.com>
> ---
> Changes in V2:
> - Updated the commit message to be imperative.
> - Fixed ptep at incorrect index getting cleared for non-leaf entries.
> ---
>   drivers/iommu/io-pgtable-arm.c | 34 +++++++++++++++++++++-------------
>   1 file changed, 21 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/iommu/io-pgtable-arm.c b/drivers/iommu/io-pgtable-arm.c
> index f5d9fd1f45bf..32401948b980 100644
> --- a/drivers/iommu/io-pgtable-arm.c
> +++ b/drivers/iommu/io-pgtable-arm.c
> @@ -274,13 +274,15 @@ static void __arm_lpae_sync_pte(arm_lpae_iopte *ptep, int num_entries,
>   				   sizeof(*ptep) * num_entries, DMA_TO_DEVICE);
>   }
>   
> -static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct io_pgtable_cfg *cfg)
> +static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct io_pgtable_cfg *cfg, int num_entries)
>   {
> +	int i;

You can make this a nice tidy loop-local declaration now.

> -	*ptep = 0;
> +	for (i = 0; i < num_entries; i++)
> +		ptep[i] = 0;
>   
>   	if (!cfg->coherent_walk)
> -		__arm_lpae_sync_pte(ptep, 1, cfg);
> +		__arm_lpae_sync_pte(ptep, num_entries, cfg);
>   }
>   
>   static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
> @@ -635,9 +637,10 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
>   			       unsigned long iova, size_t size, size_t pgcount,
>   			       int lvl, arm_lpae_iopte *ptep)
>   {
> +	bool gather_queued;
>   	arm_lpae_iopte pte;
>   	struct io_pgtable *iop = &data->iop;
> -	int i = 0, num_entries, max_entries, unmap_idx_start;
> +	int i = 0, j = 0, num_entries, max_entries, unmap_idx_start;

Similarly there's no need to initialise j here, but I'd make it 
loop-local anyway.

>   	/* Something went horribly wrong and we ran out of page table */
>   	if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
> @@ -652,28 +655,33 @@ static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
>   	/* If the size matches this level, we're in the right place */
>   	if (size == ARM_LPAE_BLOCK_SIZE(lvl, data)) {
>   		max_entries = ARM_LPAE_PTES_PER_TABLE(data) - unmap_idx_start;
> +		gather_queued = iommu_iotlb_gather_queued(gather);

This is used exactly once, do we really need to introduce the variable?

>   		num_entries = min_t(int, pgcount, max_entries);
>   
> -		while (i < num_entries) {
> -			pte = READ_ONCE(*ptep);
> +		/* Find and handle non-leaf entries */
> +		for (i = 0; i < num_entries; i++) {
> +			pte = READ_ONCE(ptep[i]);
>   			if (WARN_ON(!pte))
>   				break;
>   
> -			__arm_lpae_clear_pte(ptep, &iop->cfg);
> -
>   			if (!iopte_leaf(pte, lvl, iop->fmt)) {
> +				__arm_lpae_clear_pte(&ptep[i], &iop->cfg, 1);
> +
>   				/* Also flush any partial walks */
>   				io_pgtable_tlb_flush_walk(iop, iova + i * size, size,
>   							  ARM_LPAE_GRANULE(data));
>   				__arm_lpae_free_pgtable(data, lvl + 1, iopte_deref(pte, data));
> -			} else if (!iommu_iotlb_gather_queued(gather)) {
> -				io_pgtable_tlb_add_page(iop, gather, iova + i * size, size);
>   			}
> -
> -			ptep++;
> -			i++;
>   		}
>   
> +		/* Clear the remaining entries */
> +		if (i)

It seems a little non-obvious to optimise for one specific corner of 
unexpected failures here. I'd hope a zero-sized dma_sync wouldn't blow 
up in that case, but if you want to be safe then I'd cover it by 
tweaking the condition in __arm_lpae_clear_pte() to "if 
(!cfg->coherent_walk && num_entries)".

Thanks,
Robin.

> +			__arm_lpae_clear_pte(ptep, &iop->cfg, i);
> +
> +		if (!gather_queued)
> +			for (j = 0; j < i; j++)
> +				io_pgtable_tlb_add_page(iop, gather, iova + j * size, size);
> +
>   		return i * size;
>   	} else if (iopte_leaf(pte, lvl, iop->fmt)) {
>   		/*

Re: [PATCH V2 1/2] iommu: Optimize IOMMU UnMap

[Ashish Mhetre]

On 8/1/2024 9:53 PM, Robin Murphy wrote:
> External email: Use caution opening links or attachments
>
>
> On 01/08/2024 4:34 am, Ashish Mhetre wrote:
>> The current __arm_lpae_unmap() function calls dma_sync() on individual
>> PTEs after clearing them. Overall unmap performance can be improved by
>> around 25% for large buffer sizes by combining the syncs for adjacent
>> leaf entries.
>> Optimize the unmap time by clearing all the leaf entries and issuing a
>> single dma_sync() for them.
>> Below is detailed analysis of average unmap latency(in us) with and
>> without this optimization obtained by running dma_map_benchmark for
>> different buffer sizes.
>>
>>               UnMap Latency(us)
>> Size  Without         With            % gain with
>>       optimiztion     optimization    optimization
>>
>> 4KB   3               3               0
>> 8KB   4               3.8             5
>> 16KB  6.1             5.4             11.48
>> 32KB  10.2            8.5             16.67
>> 64KB  18.5            14.9            19.46
>> 128KB 35              27.5            21.43
>> 256KB 67.5            52.2            22.67
>> 512KB 127.9           97.2            24.00
>> 1MB   248.6           187.4           24.62
>> 2MB   65.5            65.5            0
>> 4MB   119.2           119             0.17
>>
>> Signed-off-by: Ashish Mhetre <amhetre@nvidia.com>
>> ---
>> Changes in V2:
>> - Updated the commit message to be imperative.
>> - Fixed ptep at incorrect index getting cleared for non-leaf entries.
>> ---
>>   drivers/iommu/io-pgtable-arm.c | 34 +++++++++++++++++++++-------------
>>   1 file changed, 21 insertions(+), 13 deletions(-)
>>
>> diff --git a/drivers/iommu/io-pgtable-arm.c 
>> b/drivers/iommu/io-pgtable-arm.c
>> index f5d9fd1f45bf..32401948b980 100644
>> --- a/drivers/iommu/io-pgtable-arm.c
>> +++ b/drivers/iommu/io-pgtable-arm.c
>> @@ -274,13 +274,15 @@ static void __arm_lpae_sync_pte(arm_lpae_iopte 
>> *ptep, int num_entries,
>>                                  sizeof(*ptep) * num_entries, 
>> DMA_TO_DEVICE);
>>   }
>>
>> -static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct 
>> io_pgtable_cfg *cfg)
>> +static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct 
>> io_pgtable_cfg *cfg, int num_entries)
>>   {
>> +     int i;
>
> You can make this a nice tidy loop-local declaration now.
>
Sure, will update this in new version.

>> -     *ptep = 0;
>> +     for (i = 0; i < num_entries; i++)
>> +             ptep[i] = 0;
>>
>>       if (!cfg->coherent_walk)
>> -             __arm_lpae_sync_pte(ptep, 1, cfg);
>> +             __arm_lpae_sync_pte(ptep, num_entries, cfg);
>>   }
>>
>>   static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
>> @@ -635,9 +637,10 @@ static size_t __arm_lpae_unmap(struct 
>> arm_lpae_io_pgtable *data,
>>                              unsigned long iova, size_t size, size_t 
>> pgcount,
>>                              int lvl, arm_lpae_iopte *ptep)
>>   {
>> +     bool gather_queued;
>>       arm_lpae_iopte pte;
>>       struct io_pgtable *iop = &data->iop;
>> -     int i = 0, num_entries, max_entries, unmap_idx_start;
>> +     int i = 0, j = 0, num_entries, max_entries, unmap_idx_start;
>
> Similarly there's no need to initialise j here, but I'd make it
> loop-local anyway.
>
Ack, I will make j loop-local.

>>       /* Something went horribly wrong and we ran out of page table */
>>       if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
>> @@ -652,28 +655,33 @@ static size_t __arm_lpae_unmap(struct 
>> arm_lpae_io_pgtable *data,
>>       /* If the size matches this level, we're in the right place */
>>       if (size == ARM_LPAE_BLOCK_SIZE(lvl, data)) {
>>               max_entries = ARM_LPAE_PTES_PER_TABLE(data) - 
>> unmap_idx_start;
>> +             gather_queued = iommu_iotlb_gather_queued(gather);
>
> This is used exactly once, do we really need to introduce the variable?
>
Ack, I will remove the unnecessary variable.

>>               num_entries = min_t(int, pgcount, max_entries);
>>
>> -             while (i < num_entries) {
>> -                     pte = READ_ONCE(*ptep);
>> +             /* Find and handle non-leaf entries */
>> +             for (i = 0; i < num_entries; i++) {
>> +                     pte = READ_ONCE(ptep[i]);
>>                       if (WARN_ON(!pte))
>>                               break;
>>
>> -                     __arm_lpae_clear_pte(ptep, &iop->cfg);
>> -
>>                       if (!iopte_leaf(pte, lvl, iop->fmt)) {
>> +                             __arm_lpae_clear_pte(&ptep[i], 
>> &iop->cfg, 1);
>> +
>>                               /* Also flush any partial walks */
>>                               io_pgtable_tlb_flush_walk(iop, iova + i 
>> * size, size,
>> ARM_LPAE_GRANULE(data));
>>                               __arm_lpae_free_pgtable(data, lvl + 1, 
>> iopte_deref(pte, data));
>> -                     } else if (!iommu_iotlb_gather_queued(gather)) {
>> -                             io_pgtable_tlb_add_page(iop, gather, 
>> iova + i * size, size);
>>                       }
>> -
>> -                     ptep++;
>> -                     i++;
>>               }
>>
>> +             /* Clear the remaining entries */
>> +             if (i)
>
> It seems a little non-obvious to optimise for one specific corner of
> unexpected failures here. I'd hope a zero-sized dma_sync wouldn't blow
> up in that case, but if you want to be safe then I'd cover it by
> tweaking the condition in __arm_lpae_clear_pte() to "if
> (!cfg->coherent_walk && num_entries)".
>
> Thanks,
> Robin.
>
Yes, makes sense to cover this in __arm_lpae_clear_pte().
I will update this in V3.

>> + __arm_lpae_clear_pte(ptep, &iop->cfg, i);
>> +
>> +             if (!gather_queued)
>> +                     for (j = 0; j < i; j++)
>> +                             io_pgtable_tlb_add_page(iop, gather, 
>> iova + j * size, size);
>> +
>>               return i * size;
>>       } else if (iopte_leaf(pte, lvl, iop->fmt)) {
>>               /*