Re: [PATCH 2/3] dma-mapping: benchmark: add support for dma_map_sg

[Qinxin Xia <xiaqinxin@huawei.com>]

在 2025/2/22 14:36, Barry Song 写道:
> On Fri, Feb 21, 2025 at 4:16 PM xiaqinxin <xiaqinxin@huawei.com> wrote:
>>
>>
>> -----邮件原件-----
>> 发件人: Barry Song <21cnbao@gmail.com>
>> 发送时间: 2025年2月18日 4:59
>> 收件人: xiaqinxin <xiaqinxin@huawei.com>
>> 抄送: chenxiang66@hisilicon.com; yangyicong <yangyicong@huawei.com>; hch@lst.de; iommu@lists.linux.dev; Jonathan Cameron <jonathan.cameron@huawei.com>; Zengtao (B) <prime.zeng@hisilicon.com>; fanghao (A) <fanghao11@huawei.com>; linux-kernel@vger.kernel.org
>> 主题: Re: [PATCH 2/3] dma-mapping: benchmark: add support for dma_map_sg
>>
>> On Wed, Feb 12, 2025 at 3:27 PM Qinxin Xia <xiaqinxin@huawei.com> wrote:
>>> Support for dma scatter-gather mapping and is intended for testing
>>> mapping performance. It achieves by introducing the dma_sg_map_param
>>> structure and related functions, which enable the implementation of
>>> scatter-gather mapping preparation, mapping, and unmapping operations.
>>> Additionally, the dma_map_benchmark_ops array is updated to include
>>> operations for scatter-gather mapping. This commit aims to provide a
>>> wider range of mapping performance test  to cater to different scenarios.
>> This benchmark is mainly designed to debug contention issues, such as IOMMU TLB flushes or IOMMU driver bottlenecks. I don't fully understand how SG or single will impact the evaluation of the IOMMU driver, making it unclear if the added complexity is justified.
>>
>> Can you add some explanation on why single mode is not sufficient for profiling and improving IOMMU drivers?
>>
>> Hello Barry ! 😊
>> Currently, the HiSilicon accelerator service uses the dma_map_sg interface. We want to evaluate the performance of the entire DMA map process. (including not only the iommu, but also the map framework). In addition, for scatterlist, __iommu_map is executed for each nent. This increases the complexity and time overhead of mapping. The effect of this fragmentation is not obvious in dma_map_single, which only handles a single contiguous block of memory.
>>
> Thanks!
> Please update your editor to ensure it doesn't respond with such long sentences
> without line breaks in the future :-)

Hello Barry !

Thank you for your advice, I will I'll pay attention. Leon

> Can you provide concrete examples or data showing how the newly added mode
> improves profiling of the entire DMA map process? For instance, what limitations
> exist without this mode? What performance issues cannot be identified without
> it?

You can see this patch 
:https://lore.kernel.org/all/cover.1738765879.git.leonro@nvidia.com/

Leon provides new interface for scatterlist scenarios to improve 
performance and gives some

application instance in rdma and vfio. Users can use dma_map_sg 
benchmark to measure

the performance improvement of the optimized interface compared with the 
previous one.

Thanks!

>
>>> Signed-off-by: Qinxin Xia <xiaqinxin@huawei.com>
>>> ---
>>>   include/linux/map_benchmark.h |   1 +
>>>   kernel/dma/map_benchmark.c    | 102 ++++++++++++++++++++++++++++++++++
>>>   2 files changed, 103 insertions(+)
>>>
>>> diff --git a/include/linux/map_benchmark.h
>>> b/include/linux/map_benchmark.h index 054db02a03a7..a9c1a104ba4f
>>> 100644
>>> --- a/include/linux/map_benchmark.h
>>> +++ b/include/linux/map_benchmark.h
>>> @@ -17,6 +17,7 @@
>>>
>>>   enum {
>>>          DMA_MAP_SINGLE_MODE,
>>> +       DMA_MAP_SG_MODE,
>>>          DMA_MAP_MODE_MAX
>>>   };
>>>
>>> diff --git a/kernel/dma/map_benchmark.c b/kernel/dma/map_benchmark.c
>>> index d8ec0ce058d8..b5828eeb3db7 100644
>>> --- a/kernel/dma/map_benchmark.c
>>> +++ b/kernel/dma/map_benchmark.c
>>> @@ -17,6 +17,7 @@
>>>   #include <linux/module.h>
>>>   #include <linux/pci.h>
>>>   #include <linux/platform_device.h>
>>> +#include <linux/scatterlist.h>
>>>   #include <linux/slab.h>
>>>   #include <linux/timekeeping.h>
>>>
>>> @@ -111,8 +112,109 @@ static struct map_benchmark_ops dma_single_map_benchmark_ops = {
>>>          .do_unmap = dma_single_map_benchmark_do_unmap,
>>>   };
>>>
>>> +struct dma_sg_map_param {
>>> +       struct sg_table sgt;
>>> +       struct device *dev;
>>> +       void **buf;
>>> +       u32 npages;
>>> +       u32 dma_dir;
>>> +};
>>> +
>>> +static void *dma_sg_map_benchmark_prepare(struct map_benchmark_data
>>> +*map) {
>>> +       struct scatterlist *sg;
>>> +       int i = 0;
>>> +
>>> +       struct dma_sg_map_param *mparam __free(kfree) = kzalloc(sizeof(*mparam), GFP_KERNEL);
>>> +       if (!mparam)
>>> +               return NULL;
>>> +
>>> +       mparam->npages = map->bparam.granule;
>>> +       mparam->dma_dir = map->bparam.dma_dir;
>>> +       mparam->dev = map->dev;
>>> +       mparam->buf = kmalloc_array(mparam->npages, sizeof(*mparam->buf),
>>> +                                   GFP_KERNEL);
>>> +       if (!mparam->buf)
>>> +               goto err1;
>>> +
>>> +       if (sg_alloc_table(&mparam->sgt, mparam->npages, GFP_KERNEL))
>>> +               goto err2;
>>> +
>>> +       for_each_sgtable_sg(&mparam->sgt, sg, i) {
>>> +               mparam->buf[i] = (void *)__get_free_page(GFP_KERNEL);
>>> +               if (!mparam->buf[i])
>>> +                       goto err3;
>>> +
>>> +               if (mparam->dma_dir != DMA_FROM_DEVICE)
>>> +                       memset(mparam->buf[i], 0x66, PAGE_SIZE);
>>> +
>>> +               sg_set_buf(sg, mparam->buf[i], PAGE_SIZE);
>>> +       }
>>> +
>>> +       return_ptr(mparam);
>>> +
>>> +err3:
>>> +       while (i-- > 0)
>>> +               free_page((unsigned long)mparam->buf[i]);
>>> +
>>> +       pr_err("dma_map_sg failed get free page on %s\n", dev_name(mparam->dev));
>>> +       sg_free_table(&mparam->sgt);
>>> +err2:
>>> +       pr_err("dma_map_sg failed alloc sg table on %s\n", dev_name(mparam->dev));
>>> +       kfree(mparam->buf);
>>> +err1:
>>> +       pr_err("dma_map_sg failed alloc mparam buf on %s\n", dev_name(mparam->dev));
>>> +       return NULL;
>>> +}
>>> +
>>> +static void dma_sg_map_benchmark_unprepare(void *arg) {
>>> +       struct dma_sg_map_param *mparam = arg;
>>> +       int i;
>>> +
>>> +       for (i = 0; i < mparam->npages; i++)
>>> +               free_page((unsigned long)mparam->buf[i]);
>>> +
>>> +       sg_free_table(&mparam->sgt);
>>> +
>>> +       kfree(mparam->buf);
>>> +       kfree(mparam);
>>> +}
>>> +
>>> +static int dma_sg_map_benchmark_do_map(void *arg) {
>>> +       struct dma_sg_map_param *mparam = arg;
>>> +
>>> +       int sg_mapped = dma_map_sg(mparam->dev, mparam->sgt.sgl,
>>> +                                  mparam->npages, mparam->dma_dir);
>>> +       if (!sg_mapped) {
>>> +               pr_err("dma_map_sg failed on %s\n", dev_name(mparam->dev));
>>> +               return -ENOMEM;
>>> +       }
>>> +
>>> +       return 0;
>>> +}
>>> +
>>> +static int dma_sg_map_benchmark_do_unmap(void *arg) {
>>> +       struct dma_sg_map_param *mparam = arg;
>>> +
>>> +       dma_unmap_sg(mparam->dev, mparam->sgt.sgl, mparam->npages,
>>> +                    mparam->dma_dir);
>>> +
>>> +       return 0;
>>> +}
>>> +
>>> +static struct map_benchmark_ops dma_sg_map_benchmark_ops = {
>>> +       .prepare = dma_sg_map_benchmark_prepare,
>>> +       .unprepare = dma_sg_map_benchmark_unprepare,
>>> +       .do_map = dma_sg_map_benchmark_do_map,
>>> +       .do_unmap = dma_sg_map_benchmark_do_unmap, };
>>> +
>>>   static struct map_benchmark_ops *dma_map_benchmark_ops[DMA_MAP_MODE_MAX] = {
>>>          [DMA_MAP_SINGLE_MODE] = &dma_single_map_benchmark_ops,
>>> +       [DMA_MAP_SG_MODE] = &dma_sg_map_benchmark_ops,
>>>   };
>>>
>>>   static int map_benchmark_thread(void *data)
>>> --
>>> 2.33.0
>>>
> Thanks
> Barry