2023-04-18 06:37:19

by Zhen Lei

Subject: [PATCH] iommu/iova: Optimize iova_magazine_alloc()

Only the member 'size' needs to be initialized to 0. Clearing the array
pfns[], which is about 1 KiB in size, not only wastes time, but also
causes cache pollution.

Signed-off-by: Zhen Lei <[email protected]>
---
drivers/iommu/iova.c | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
index fe452ce466429a7..c970b9a2819d7bb 100644
--- a/drivers/iommu/iova.c
+++ b/drivers/iommu/iova.c
@@ -647,7 +647,13 @@ struct iova_rcache {

static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
{
- return kzalloc(sizeof(struct iova_magazine), flags);
+ struct iova_magazine *mag;
+
+ mag = kmalloc(sizeof(struct iova_magazine), flags);
+ if (mag)
+ mag->size = 0;
+
+ return mag;
}

static void iova_magazine_free(struct iova_magazine *mag)
--
2.25.1

2023-04-20 10:46:25

by Robin Murphy

Subject: Re: [PATCH] iommu/iova: Optimize iova_magazine_alloc()

On 2023-04-18 07:25, Zhen Lei wrote:
> Only the member 'size' needs to be initialized to 0. Clearing the array
> pfns[], which is about 1 KiB in size, not only wastes time, but also
> causes cache pollution.

Makes sense to me.

Acked-by: Robin Murphy <[email protected]>

> Signed-off-by: Zhen Lei <[email protected]>
> ---
> drivers/iommu/iova.c | 8 +++++++-
> 1 file changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
> index fe452ce466429a7..c970b9a2819d7bb 100644
> --- a/drivers/iommu/iova.c
> +++ b/drivers/iommu/iova.c
> @@ -647,7 +647,13 @@ struct iova_rcache {
>
> static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
> {
> - return kzalloc(sizeof(struct iova_magazine), flags);
> + struct iova_magazine *mag;
> +
> + mag = kmalloc(sizeof(struct iova_magazine), flags);

Nit: maybe sizeof(*mag) is preferable?

> + if (mag)
> + mag->size = 0;
> +
> + return mag;
> }
>
> static void iova_magazine_free(struct iova_magazine *mag)

2023-04-20 12:07:06

by Zhen Lei

Subject: Re: [PATCH] iommu/iova: Optimize iova_magazine_alloc()

On 2023/4/20 18:38, Robin Murphy wrote:
> On 2023-04-18 07:25, Zhen Lei wrote:
>> Only the member 'size' needs to be initialized to 0. Clearing the array
>> pfns[], which is about 1 KiB in size, not only wastes time, but also
>> causes cache pollution.
>
> Makes sense to me.
>
> Acked-by: Robin Murphy <[email protected]>
>
>> Signed-off-by: Zhen Lei <[email protected]>
>> ---
>>   drivers/iommu/iova.c | 8 +++++++-
>>   1 file changed, 7 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
>> index fe452ce466429a7..c970b9a2819d7bb 100644
>> --- a/drivers/iommu/iova.c
>> +++ b/drivers/iommu/iova.c
>> @@ -647,7 +647,13 @@ struct iova_rcache {
>>     static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
>>   {
>> -    return kzalloc(sizeof(struct iova_magazine), flags);
>> +    struct iova_magazine *mag;
>> +
>> +    mag = kmalloc(sizeof(struct iova_magazine), flags);
>
> Nit: maybe sizeof(*mag) is preferable?

Yes, sizeof(*mag) is the more recommended usage, I will post v2 tomorrow.

>
>> +    if (mag)
>> +        mag->size = 0;
>> +
>> +    return mag;
>>   }
>>     static void iova_magazine_free(struct iova_magazine *mag)
>
> .
>

--
Regards,
Zhen Lei

2023-05-22 15:22:35

by Joerg Roedel

Subject: Re: [PATCH] iommu/iova: Optimize iova_magazine_alloc()

On Thu, Apr 20, 2023 at 07:52:14PM +0800, Leizhen (ThunderTown) wrote:
> Yes, sizeof(*mag) is the more recommended usage, I will post v2 tomorrow.

Somehow v2 didn't make it to my inbox, but I found it on lore. This is
applied now, thanks.

Regards,

Joerg