mem: support layout of IBM Power
authorChao Zhu <chaozhu@linux.vnet.ibm.com>
Tue, 25 Nov 2014 22:17:15 +0000 (17:17 -0500)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Wed, 26 Nov 2014 20:50:10 +0000 (21:50 +0100)
The mmap of hugepage files on IBM Power starts from high address to low
address. This is different from x86. This patch modified the memory
segment detection code to get the correct memory segment layout on Power
architecture. This patch also added a commond ARCH_PPC_64 definition for
64 bit systems.

Signed-off-by: Chao Zhu <chaozhu@linux.vnet.ibm.com>
Acked-by: David Marchand <david.marchand@6wind.com>
config/defconfig_ppc_64-power8-linuxapp-gcc
config/defconfig_x86_64-native-linuxapp-clang
config/defconfig_x86_64-native-linuxapp-gcc
config/defconfig_x86_64-native-linuxapp-icc
lib/librte_eal/linuxapp/eal/eal_memory.c

index ffcb095..cefb3be 100644 (file)
@@ -35,6 +35,7 @@ CONFIG_RTE_MACHINE="power8"
 CONFIG_RTE_ARCH="ppc_64"
 CONFIG_RTE_ARCH_PPC_64=y
 CONFIG_RTE_ARCH_BIG_ENDIAN=y
+CONFIG_RTE_ARCH_64=y
 
 CONFIG_RTE_TOOLCHAIN="gcc"
 CONFIG_RTE_TOOLCHAIN_GCC=y
index bbda080..5f3074e 100644 (file)
@@ -36,6 +36,7 @@ CONFIG_RTE_MACHINE="native"
 
 CONFIG_RTE_ARCH="x86_64"
 CONFIG_RTE_ARCH_X86_64=y
+CONFIG_RTE_ARCH_64=y
 
 CONFIG_RTE_TOOLCHAIN="clang"
 CONFIG_RTE_TOOLCHAIN_CLANG=y
index 3de818a..60baf5b 100644 (file)
@@ -36,6 +36,7 @@ CONFIG_RTE_MACHINE="native"
 
 CONFIG_RTE_ARCH="x86_64"
 CONFIG_RTE_ARCH_X86_64=y
+CONFIG_RTE_ARCH_64=y
 
 CONFIG_RTE_TOOLCHAIN="gcc"
 CONFIG_RTE_TOOLCHAIN_GCC=y
index 795333b..71d1e28 100644 (file)
@@ -36,6 +36,7 @@ CONFIG_RTE_MACHINE="native"
 
 CONFIG_RTE_ARCH="x86_64"
 CONFIG_RTE_ARCH_X86_64=y
+CONFIG_RTE_ARCH_64=y
 
 CONFIG_RTE_TOOLCHAIN="icc"
 CONFIG_RTE_TOOLCHAIN_ICC=y
index f2454f4..e6cb919 100644 (file)
@@ -316,11 +316,12 @@ map_all_hugepages(struct hugepage_file *hugepg_tbl,
 #endif
                        hugepg_tbl[i].filepath[sizeof(hugepg_tbl[i].filepath) - 1] = '\0';
                }
-#ifndef RTE_ARCH_X86_64
-               /* for 32-bit systems, don't remap 1G pages, just reuse original
-                * map address as final map address.
+#ifndef RTE_ARCH_64
+               /* for 32-bit systems, don't remap 1G and 16G pages, just reuse
+                * original map address as final map address.
                 */
-               else if (hugepage_sz == RTE_PGSIZE_1G){
+               else if ((hugepage_sz == RTE_PGSIZE_1G)
+                       || (hugepage_sz == RTE_PGSIZE_16G)) {
                        hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
                        hugepg_tbl[i].orig_va = NULL;
                        continue;
@@ -335,9 +336,17 @@ map_all_hugepages(struct hugepage_file *hugepg_tbl,
                         * physical block: count the number of
                         * contiguous physical pages. */
                        for (j = i+1; j < hpi->num_pages[0] ; j++) {
+#ifdef RTE_ARCH_PPC_64
+                               /* The physical addresses are sorted in
+                                * descending order on PPC64 */
+                               if (hugepg_tbl[j].physaddr !=
+                                   hugepg_tbl[j-1].physaddr - hugepage_sz)
+                                       break;
+#else
                                if (hugepg_tbl[j].physaddr !=
                                    hugepg_tbl[j-1].physaddr + hugepage_sz)
                                        break;
+#endif
                        }
                        num_pages = j - i;
                        vma_len = num_pages * hugepage_sz;
@@ -412,11 +421,12 @@ remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 
        while (i < hpi->num_pages[0]) {
 
-#ifndef RTE_ARCH_X86_64
-               /* for 32-bit systems, don't remap 1G pages, just reuse original
-                * map address as final map address.
+#ifndef RTE_ARCH_64
+               /* for 32-bit systems, don't remap 1G pages and 16G pages,
+                * just reuse original map address as final map address.
                 */
-               if (hugepage_sz == RTE_PGSIZE_1G){
+               if ((hugepage_sz == RTE_PGSIZE_1G)
+                       || (hugepage_sz == RTE_PGSIZE_16G)) {
                        hugepg_tbl[i].final_va = hugepg_tbl[i].orig_va;
                        hugepg_tbl[i].orig_va = NULL;
                        i++;
@@ -428,8 +438,17 @@ remap_all_hugepages(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
                 * physical block: count the number of
                 * contiguous physical pages. */
                for (j = i+1; j < hpi->num_pages[0] ; j++) {
-                       if (hugepg_tbl[j].physaddr != hugepg_tbl[j-1].physaddr + hugepage_sz)
+#ifdef RTE_ARCH_PPC_64
+                       /* The physical addresses are sorted in descending
+                        * order on PPC64 */
+                       if (hugepg_tbl[j].physaddr !=
+                               hugepg_tbl[j-1].physaddr - hugepage_sz)
                                break;
+#else
+                       if (hugepg_tbl[j].physaddr !=
+                               hugepg_tbl[j-1].physaddr + hugepage_sz)
+                               break;
+#endif
                }
                num_pages = j - i;
                vma_len = num_pages * hugepage_sz;
@@ -652,21 +671,21 @@ error:
 }
 
 /*
- * Sort the hugepg_tbl by physical address (lower addresses first). We
- * use a slow algorithm, but we won't have millions of pages, and this
- * is only done at init time.
+ * Sort the hugepg_tbl by physical address (lower addresses first on x86,
+ * higher address first on powerpc). We use a slow algorithm, but we won't
+ * have millions of pages, and this is only done at init time.
  */
 static int
 sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
 {
        unsigned i, j;
-       int smallest_idx;
-       uint64_t smallest_addr;
+       int compare_idx;
+       uint64_t compare_addr;
        struct hugepage_file tmp;
 
        for (i = 0; i < hpi->num_pages[0]; i++) {
-               smallest_addr = 0;
-               smallest_idx = -1;
+               compare_addr = 0;
+               compare_idx = -1;
 
                /*
                 * browse all entries starting at 'i', and find the
@@ -674,23 +693,28 @@ sort_by_physaddr(struct hugepage_file *hugepg_tbl, struct hugepage_info *hpi)
                 */
                for (j=i; j< hpi->num_pages[0]; j++) {
 
-                       if (smallest_addr == 0 ||
-                           hugepg_tbl[j].physaddr < smallest_addr) {
-                               smallest_addr = hugepg_tbl[j].physaddr;
-                               smallest_idx = j;
+                       if (compare_addr == 0 ||
+#ifdef RTE_ARCH_PPC_64
+                               hugepg_tbl[j].physaddr > compare_addr) {
+#else
+                               hugepg_tbl[j].physaddr < compare_addr) {
+#endif
+                               compare_addr = hugepg_tbl[j].physaddr;
+                               compare_idx = j;
                        }
                }
 
                /* should not happen */
-               if (smallest_idx == -1) {
+               if (compare_idx == -1) {
                        RTE_LOG(ERR, EAL, "%s(): error in physaddr sorting\n", __func__);
                        return -1;
                }
 
                /* swap the 2 entries in the table */
-               memcpy(&tmp, &hugepg_tbl[smallest_idx], sizeof(struct hugepage_file));
-               memcpy(&hugepg_tbl[smallest_idx], &hugepg_tbl[i],
-                               sizeof(struct hugepage_file));
+               memcpy(&tmp, &hugepg_tbl[compare_idx],
+                       sizeof(struct hugepage_file));
+               memcpy(&hugepg_tbl[compare_idx], &hugepg_tbl[i],
+                       sizeof(struct hugepage_file));
                memcpy(&hugepg_tbl[i], &tmp, sizeof(struct hugepage_file));
        }
        return 0;
@@ -1260,12 +1284,25 @@ rte_eal_hugepage_init(void)
                        new_memseg = 1;
                else if (hugepage[i].size != hugepage[i-1].size)
                        new_memseg = 1;
+
+#ifdef RTE_ARCH_PPC_64
+               /* On PPC64 architecture, the mmap always start from higher
+                * virtual address to lower address. Here, both the physical
+                * address and virtual address are in descending order */
+               else if ((hugepage[i-1].physaddr - hugepage[i].physaddr) !=
+                   hugepage[i].size)
+                       new_memseg = 1;
+               else if (((unsigned long)hugepage[i-1].final_va -
+                   (unsigned long)hugepage[i].final_va) != hugepage[i].size)
+                       new_memseg = 1;
+#else
                else if ((hugepage[i].physaddr - hugepage[i-1].physaddr) !=
                    hugepage[i].size)
                        new_memseg = 1;
                else if (((unsigned long)hugepage[i].final_va -
                    (unsigned long)hugepage[i-1].final_va) != hugepage[i].size)
                        new_memseg = 1;
+#endif
 
                if (new_memseg) {
                        j += 1;
@@ -1284,6 +1321,12 @@ rte_eal_hugepage_init(void)
                }
                /* continuation of previous memseg */
                else {
+#ifdef RTE_ARCH_PPC_64
+               /* Use the phy and virt address of the last page as segment
+                * address for IBM Power architecture */
+                       mcfg->memseg[j].phys_addr = hugepage[i].physaddr;
+                       mcfg->memseg[j].addr = hugepage[i].final_va;
+#endif
                        mcfg->memseg[j].len += mcfg->memseg[j].hugepage_sz;
                }
                hugepage[i].memseg_id = j;