eal: remove unnecessary hugepage zero-filling
authorZhihong Wang <zhihong.wang@intel.com>
Sun, 22 Nov 2015 19:13:35 +0000 (14:13 -0500)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Thu, 21 Jan 2016 14:57:50 +0000 (15:57 +0100)
The kernel fills new allocated (huge) pages with zeros.
DPDK just has to populate page tables to trigger the allocation.

Signed-off-by: Zhihong Wang <zhihong.wang@intel.com>
Acked-by: Stephen Hemminger <stephen@networkplumber.org>
lib/librte_eal/linuxapp/eal/eal_memory.c

index 846fd31..48c6175 100644 (file)
@@ -399,8 +399,10 @@ map_all_hugepages(struct hugepage_file *hugepg_tbl,
                        return -1;
                }
 
+               /* map the segment, and populate page tables,
+                * the kernel fills this segment with zeros */
                virtaddr = mmap(vma_addr, hugepage_sz, PROT_READ | PROT_WRITE,
-                               MAP_SHARED, fd, 0);
+                               MAP_SHARED | MAP_POPULATE, fd, 0);
                if (virtaddr == MAP_FAILED) {
                        RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__,
                                        strerror(errno));
@@ -410,7 +412,6 @@ map_all_hugepages(struct hugepage_file *hugepg_tbl,
 
                if (orig) {
                        hugepg_tbl[i].orig_va = virtaddr;
-                       memset(virtaddr, 0, hugepage_sz);
                }
                else {
                        hugepg_tbl[i].final_va = virtaddr;
@@ -529,22 +530,16 @@ remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 
                        old_addr = vma_addr;
 
-                       /* map new, bigger segment */
+                       /* map new, bigger segment, and populate page tables,
+                        * the kernel fills this segment with zeros */
                        vma_addr = mmap(vma_addr, total_size,
-                                       PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+                                       PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, fd, 0);
 
                        if (vma_addr == MAP_FAILED || vma_addr != old_addr) {
                                RTE_LOG(ERR, EAL, "%s(): mmap failed: %s\n", __func__, strerror(errno));
                                close(fd);
                                return -1;
                        }
-
-                       /* touch the page. this is needed because kernel postpones mapping
-                        * creation until the first page fault. with this, we pin down
-                        * the page and it is marked as used and gets into process' pagemap.
-                        */
-                       for (offset = 0; offset < total_size; offset += hugepage_sz)
-                               *((volatile uint8_t*) RTE_PTR_ADD(vma_addr, offset));
                }
 
                /* set shared flock on the file. */
@@ -592,9 +587,6 @@ remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
                        }
                }
 
-               /* zero out the whole segment */
-               memset(hugepg_tbl[page_idx].final_va, 0, total_size);
-
                page_idx++;
        }