Date:   Thu, 16 Nov 2017 02:04:17 +0900
From:   Jaewon Kim <jaewon31.kim@...sung.com>
To:     Marek Szyprowski <m.szyprowski@...sung.com>, hch@....de,
        robin.murphy@....com, gregkh@...uxfoundation.org,
        iommu@...ts.linux-foundation.org
Cc:     akpm@...ux-foundation.org, mhocko@...e.com, vbabka@...e.cz,
        linux-mm@...ck.org, linux-kernel@...r.kernel.org,
        jaewon31.kim@...il.com
Subject: Re: [RFC PATCH] drivers: base: dma-coherent: find free region
 without alignment

Hello Marek

On 2017년 11월 14일 20:07, Marek Szyprowski wrote:
> Hi Jaewon,
>
> On 2017-11-14 09:42, Jaewon Kim wrote:
>> dma-coherent uses bitmap API which internally consider align based on the
>> requested size. Depending on some usage pattern, using align, I think, may
>> be good for fast search and anti-fragmentation. But with the align, an
>> allocation may be failed.
>>
>> This is a example, total size is 30MB, only few memory at front is being
>> used, and 9MB is being requsted. Then 9MB will be aligned to 16MB. The
>> first try on offset 0MB will be failed because of others already using. The
>> second try on offset 16MB will be failed because of ouf of bound.
>>
>> So if the align is not necessary on dma-coherent, this patch removes the
>> align policy to allow allocation without increasing the total size.
>
> You are right that keeping strict alignment is waste of memory for large
> allocations. However for the smaller ones, typically under 1MiB, it helps
> to reduce memory fragmentation. The alignment of the allocated buffers is
> de-facto guaranteed by the memory management framework in Linux kernel
> and there are drivers that depends on this feature.
>
> Maybe it would make sense to keep alignment for buffers smaller than some
> predefined value (like 1MiB), something similar to config
> ARM_DMA_IOMMU_ALIGNMENT in arch/arm/Kconfig. Otherwise I would expect that
> some drivers will be broken by this patch.
Thank you for your comment.

I looked ARM_DMA_IOMMU_ALIGNMENT in ARM, it looks similar but it is using
bitmap_find_next_zero_area rather than bitmap_find_free_region. bitmap_find_next_zero_area
apply aligning only onto offset but not onto size. So I think ARM_DMA_IOMMU_ALIGNMENT way
is not perfect on this dma-coherent APIs which tries to align even on size.

Let me say another way where each reserved_mem from device tree can decide if it wants aligning.
This could be implemented like below. I need to change other dma-coherent APIs though.
I will wait for your comment on this.

--- a/drivers/base/dma-coherent.c
+++ b/drivers/base/dma-coherent.c
@@ -17,6 +17,7 @@ struct dma_coherent_mem {
        unsigned long   *bitmap;
        spinlock_t      spinlock;
        bool            use_dev_dma_pfn_offset;
+       bool            no_align;
 };
 
 static struct dma_coherent_mem *dma_coherent_default_memory __ro_after_init;
@@ -162,7 +163,6 @@ EXPORT_SYMBOL(dma_mark_declared_memory_occupied);
 static void *__dma_alloc_from_coherent(struct dma_coherent_mem *mem,
                ssize_t size, dma_addr_t *dma_handle)
 {
-       int order = get_order(size);
        unsigned long flags;
        int pageno;
        void *ret;
@@ -172,9 +172,21 @@ static void *__dma_alloc_from_coherent(struct dma_coherent_mem *mem,
        if (unlikely(size > (mem->size << PAGE_SHIFT)))
                goto err;
 
-       pageno = bitmap_find_free_region(mem->bitmap, mem->size, order);
-       if (unlikely(pageno < 0))
-               goto err;
+       if (mem->no_align) {
+               int nr_page = PAGE_ALIGN(size) >> PAGE_SHIFT;
+
+               pageno = bitmap_find_next_zero_area(mem->bitmap, mem->size, 0,
+                                                   nr_page, 0);
+               if (unlikely(pageno >= mem->size))
+                       goto err;
+               bitmap_set(mem->bitmap, pageno, nr_page);
+       } else {
+               int order = get_order(size);
+
+               pageno = bitmap_find_free_region(mem->bitmap, mem->size, order);
+               if (unlikely(pageno < 0))
+                       goto err;
+       }
 
        /*
         * Memory was found in the coherent area.
@@ -346,6 +358,7 @@ static struct reserved_mem *dma_reserved_default_memory __initdata;
 static int rmem_dma_device_init(struct reserved_mem *rmem, struct device *dev)
 {
        struct dma_coherent_mem *mem = rmem->priv;
+       unsigned long node = rmem->fdt_node;
        int ret;
 
        if (!mem) {
@@ -360,6 +373,8 @@ static int rmem_dma_device_init(struct reserved_mem *rmem, struct device *dev)
        }
        mem->use_dev_dma_pfn_offset = true;
        rmem->priv = mem;
+       if (of_get_flat_dt_prop(node, "no-align", NULL))
+               mem->no_align = true;
        dma_assign_coherent_memory(dev, mem);
        return 0;
 }


>
>> Signed-off-by: Jaewon Kim <jaewon31.kim@...sung.com>
>> ---
>>   drivers/base/dma-coherent.c | 8 +++++---
>>   1 file changed, 5 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/base/dma-coherent.c b/drivers/base/dma-coherent.c
>> index 744f64f43454..b86a96d0cd07 100644
>> --- a/drivers/base/dma-coherent.c
>> +++ b/drivers/base/dma-coherent.c
>> @@ -162,7 +162,7 @@ EXPORT_SYMBOL(dma_mark_declared_memory_occupied);
>>   static void *__dma_alloc_from_coherent(struct dma_coherent_mem *mem,
>>           ssize_t size, dma_addr_t *dma_handle)
>>   {
>> -    int order = get_order(size);
>> +    int nr_page = PAGE_ALIGN(size) >> PAGE_SHIFT;
>>       unsigned long flags;
>>       int pageno;
>>       void *ret;
>> @@ -172,9 +172,11 @@ static void *__dma_alloc_from_coherent(struct dma_coherent_mem *mem,
>>       if (unlikely(size > (mem->size << PAGE_SHIFT)))
>>           goto err;
>>   -    pageno = bitmap_find_free_region(mem->bitmap, mem->size, order);
>> -    if (unlikely(pageno < 0))
>> +    pageno = bitmap_find_next_zero_area(mem->bitmap, mem->size, 0,
>> +                        nr_page, 0);
>> +    if (unlikely(pageno >= mem->size)) {
>>           goto err;
>> +    bitmap_set(mem->bitmap, pageno, nr_page);
>>         /*
>>        * Memory was found in the coherent area.
>
> Best regards

Powered by blists - more mailing lists