Hi Marek,

Could you elaborate a bit on the target use case and describe some examples?

Sure. The goal is to achieve "zero copy" from a block device (like a SATA drive) into a contiguous buffer, which is mmap'ed from user-space with the O_DIRECT flag. Skipping the buffer cache and DMA'ing directly into a contiguous buffer (which would be initially carved out via CMA) is essential for resource-constrained embedded systems.

What problem are you trying to solve?

On our platform (an UltraHD-capable set-top box), we would like to ensure efficient operation of the system as a whole by ensuring video stream data gets DMA'ed directly into a user-space accessible buffer, with zero copies. Introducing extra CPU memory copies will introduce latency, leading to issues like dropped frames and lower system performance.

I have completely no idea how this patch is related to O_DIRECT and how it does.

When user-space invokes mmap(), the kernel uses get_user_pages() to check for the existence of pages for a particular VMA. In our system, we leverage CMA for contiguous buffers. The missing link was determining whether or not a particular page in the VMA is CMA, and if it has been allocated.

Thanks, Marc

On Tue, Jul 1, 2014 at 3:14 AM, Marek Szyprowski m.szyprowski@samsung.com wrote:

Hello,

On 2014-06-24 11:07, Marc Carino wrote:

...

It is desirable to perform "zero copy" operations on the numerous embedded systems that run the Linux kernel, to enable efficient performance on resource-constrained systems. Usually such embedded systems leverage mechanisms like CMA to carve-out regions of contiguous physical memory for use by bus-mastering peripherals.

This patch adds some logic which determines if user-space is attempting to perform direct I/O mapping on CMA pages, and if so, ensures the mapping succeeds.

This patch is a 1st attempt to allow O_DIRECT on the memory carved-out by CMA. Any suggestions or comments are appreciated!

Signed-off-by: Marc Carino marc.ceeeee@gmail.com

mm/gup.c | 65 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 65 insertions(+)

diff --git a/mm/gup.c b/mm/gup.c index cc5a9e7..5fd1bfa 100644 --- a/mm/gup.c +++ b/mm/gup.c @@ -353,6 +353,61 @@ static int check_vma_flags(struct vm_area_struct *vma, unsigned long gup_flags) return 0; } +static int cma_get_page(struct mm_struct *mm, struct vm_area_struct *vma,

•                   unsigned long start, struct page **page)
  

+{ +#ifdef CONFIG_CMA

•   const unsigned long pg = start & PAGE_MASK;
  

•   unsigned long pfn;
  

•   int ret = -EFAULT;
  

•   pgd_t *pgd;
  

•   pud_t *pud;
  

•   pmd_t *pmd;
  

•   pte_t *pte;
  

•   struct page *tmp_page;
  

•   pgd = pgd_offset(mm, pg);
  

•   BUG_ON(pgd_none(*pgd));
  

•   pud = pud_offset(pgd, pg);
  

•   BUG_ON(pud_none(*pud));
  

•   pmd = pmd_offset(pud, pg);
  

•   if (pmd_none(*pmd))
  

•           return ret;
  

•   pte = pte_offset_map(pmd, pg);
  

•   if (!pte)
  

•           return ret;
  

•   if (pte_none(*pte))
  

•           goto out;
  

•   tmp_page = pte_page(*pte);
  

•   if (!tmp_page)
  

•           goto out;
  

•   if (!page_count(tmp_page))
  

•           goto out;
  

•   if (page_mapped(tmp_page))
  

•           goto out;
  

•   if (get_pageblock_migratetype(tmp_page) != MIGRATE_CMA)
  

•           goto out;
  

•   if (page) {
  

•           *page = tmp_page;
  

•           get_page(*page);
  

•   }
  

•   ret = 0;
  

+out:

•   pte_unmap(pte);
  

•   return ret;
  

+#else

•   return 0;
  

+#endif +}

• /**
  • __get_user_pages() - pin user pages in memory
  • @tsk: task_struct of target task

@@ -443,6 +498,16 @@ long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm, goto next_page; }

•                 if (!cma_get_page(mm, vma, start,
  

•                                     pages ? &pages[i] : NULL)) {
  

•                           if (vmas)
  

•                                   vmas[i] = vma;
  

•                           i++;
  

•                           start += PAGE_SIZE;
  

•                           nr_pages--;
  

•                           continue;
  

•                   }
  

•                   if (!vma || check_vma_flags(vma, gup_flags))
                            return i ? : -EFAULT;
                    if (is_vm_hugetlb_page(vma)) {
  

Frankly, I have completely no idea how this patch is related to O_DIRECT and how it does. Could you elaborate a bit on the target use case and describe some examples? What problem are you trying to solve?

Best regards

Marek Szyprowski, PhD Samsung R&D Institute Poland