Re: Fwd: [PATCH 1/1] RDMA/umem: add back hugepage sg list
From: Jason Gunthorpe <jgg@nvidia.com>
Date: 2021-03-12 00:26:33
On Thu, Mar 11, 2021 at 06:41:43PM +0800, Zhu Yanjun wrote:
On Mon, Mar 8, 2021 at 8:16 PM Jason Gunthorpe [off-list ref] wrote:quoted
On Mon, Mar 08, 2021 at 06:13:52PM +0800, Zhu Yanjun wrote:quoted
And I delved into the source code of __sg_alloc_table_from_pages. I found that this function is related with ib_dma_max_seg_size. So when ib_dma_max_seg_size is set to UINT_MAX, the sg dma address is 4K (one page). When ib_dma_max_seg_size is set to SZ_2M, the sg dma address is 2M now.That seems like a bug, you should fix itHi, Jason && Leon I compared the function __sg_alloc_table_from_pages with ib_umem_add_sg_table. In __sg_alloc_table_from_pages: " 449 if (prv) { 450 unsigned long paddr = (page_to_pfn(sg_page(prv)) * PAGE_SIZE + 451 prv->offset + prv->length) / 452 PAGE_SIZE; 453 454 if (WARN_ON(offset)) 455 return ERR_PTR(-EINVAL); 456 457 /* Merge contiguous pages into the last SG */ 458 prv_len = prv->length; 459 while (n_pages && page_to_pfn(pages[0]) == paddr) { 460 if (prv->length + PAGE_SIZE > max_segment) 461 break; 462 prv->length += PAGE_SIZE; 463 paddr++; 464 pages++; 465 n_pages--; 466 } 467 if (!n_pages) 468 goto out; 469 } " if prv->length + PAGE_SIZE > max_segment, then set another sg. In the commit "RDMA/umem: Move to allocate SG table from pages", max_segment is dma_get_max_seg_size. Normally it is UINT_MAX. So in my host, prv->length + PAGE_SIZE is usually less than max_segment since length is unsigned int.
I don't understand what you are trying to say 460 if (prv->length + PAGE_SIZE > max_segment) max_segment should be a very big number and "prv->length + PAGE_SIZE" should always be < max_segment so it should always be increasing the size of prv->length and 'rpv' here is the sgl. The other loops are the same. Jason