bonding: fix broken multicast with round-robin mode
[linux-2.6.git] / mm / sparse-vmemmap.c
index d3b718b..392b9bb 100644 (file)
@@ -1,7 +1,7 @@
 /*
  * Virtual Memory Map support
  *
- * (C) 2007 sgi. Christoph Lameter <clameter@sgi.com>.
+ * (C) 2007 sgi. Christoph Lameter.
  *
  * Virtual memory maps allow VM primitives pfn_to_page, page_to_pfn,
  * virt_to_page, page_address() to be implemented as a base offset
@@ -24,6 +24,7 @@
 #include <linux/module.h>
 #include <linux/spinlock.h>
 #include <linux/vmalloc.h>
+#include <linux/sched.h>
 #include <asm/dma.h>
 #include <asm/pgalloc.h>
 #include <asm/pgtable.h>
  * or to back the page tables that are used to create the mapping.
  * Uses the main allocators if they are available, else bootmem.
  */
+
+static void * __init_refok __earlyonly_bootmem_alloc(int node,
+                               unsigned long size,
+                               unsigned long align,
+                               unsigned long goal)
+{
+       return __alloc_bootmem_node_high(NODE_DATA(node), size, align, goal);
+}
+
+static void *vmemmap_buf;
+static void *vmemmap_buf_end;
+
 void * __meminit vmemmap_alloc_block(unsigned long size, int node)
 {
        /* If the main allocator is up use that, fallback to bootmem. */
        if (slab_is_available()) {
-               struct page *page = alloc_pages_node(node,
+               struct page *page;
+
+               if (node_state(node, N_HIGH_MEMORY))
+                       page = alloc_pages_node(node,
                                GFP_KERNEL | __GFP_ZERO, get_order(size));
+               else
+                       page = alloc_pages(GFP_KERNEL | __GFP_ZERO,
+                               get_order(size));
                if (page)
                        return page_address(page);
                return NULL;
        } else
-               return __alloc_bootmem_node(NODE_DATA(node), size, size,
+               return __earlyonly_bootmem_alloc(node, size, size,
                                __pa(MAX_DMA_ADDRESS));
 }
 
+/* need to make sure size is all the same during early stage */
+void * __meminit vmemmap_alloc_block_buf(unsigned long size, int node)
+{
+       void *ptr;
+
+       if (!vmemmap_buf)
+               return vmemmap_alloc_block(size, node);
+
+       /* take the from buf */
+       ptr = (void *)ALIGN((unsigned long)vmemmap_buf, size);
+       if (ptr + size > vmemmap_buf_end)
+               return vmemmap_alloc_block(size, node);
+
+       vmemmap_buf = ptr + size;
+
+       return ptr;
+}
+
 void __meminit vmemmap_verify(pte_t *pte, int node,
                                unsigned long start, unsigned long end)
 {
        unsigned long pfn = pte_pfn(*pte);
        int actual_node = early_pfn_to_nid(pfn);
 
-       if (actual_node != node)
+       if (node_distance(actual_node, node) > LOCAL_DISTANCE)
                printk(KERN_WARNING "[%lx-%lx] potential offnode "
                        "page_structs\n", start, end - 1);
 }
@@ -63,9 +100,9 @@ pte_t * __meminit vmemmap_pte_populate(pmd_t *pmd, unsigned long addr, int node)
        pte_t *pte = pte_offset_kernel(pmd, addr);
        if (pte_none(*pte)) {
                pte_t entry;
-               void *p = vmemmap_alloc_block(PAGE_SIZE, node);
+               void *p = vmemmap_alloc_block_buf(PAGE_SIZE, node);
                if (!p)
-                       return 0;
+                       return NULL;
                entry = pfn_pte(__pa(p) >> PAGE_SHIFT, PAGE_KERNEL);
                set_pte_at(&init_mm, addr, pte, entry);
        }
@@ -78,7 +115,7 @@ pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, int node)
        if (pmd_none(*pmd)) {
                void *p = vmemmap_alloc_block(PAGE_SIZE, node);
                if (!p)
-                       return 0;
+                       return NULL;
                pmd_populate_kernel(&init_mm, pmd, p);
        }
        return pmd;
@@ -90,7 +127,7 @@ pud_t * __meminit vmemmap_pud_populate(pgd_t *pgd, unsigned long addr, int node)
        if (pud_none(*pud)) {
                void *p = vmemmap_alloc_block(PAGE_SIZE, node);
                if (!p)
-                       return 0;
+                       return NULL;
                pud_populate(&init_mm, pud, p);
        }
        return pud;
@@ -102,7 +139,7 @@ pgd_t * __meminit vmemmap_pgd_populate(unsigned long addr, int node)
        if (pgd_none(*pgd)) {
                void *p = vmemmap_alloc_block(PAGE_SIZE, node);
                if (!p)
-                       return 0;
+                       return NULL;
                pgd_populate(&init_mm, pgd, p);
        }
        return pgd;
@@ -146,3 +183,55 @@ struct page * __meminit sparse_mem_map_populate(unsigned long pnum, int nid)
 
        return map;
 }
+
+void __init sparse_mem_maps_populate_node(struct page **map_map,
+                                         unsigned long pnum_begin,
+                                         unsigned long pnum_end,
+                                         unsigned long map_count, int nodeid)
+{
+       unsigned long pnum;
+       unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
+       void *vmemmap_buf_start;
+
+       size = ALIGN(size, PMD_SIZE);
+       vmemmap_buf_start = __earlyonly_bootmem_alloc(nodeid, size * map_count,
+                        PMD_SIZE, __pa(MAX_DMA_ADDRESS));
+
+       if (vmemmap_buf_start) {
+               vmemmap_buf = vmemmap_buf_start;
+               vmemmap_buf_end = vmemmap_buf_start + size * map_count;
+       }
+
+       for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
+               struct mem_section *ms;
+
+               if (!present_section_nr(pnum))
+                       continue;
+
+               map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
+               if (map_map[pnum])
+                       continue;
+               ms = __nr_to_section(pnum);
+               printk(KERN_ERR "%s: sparsemem memory map backing failed "
+                       "some memory will not be available.\n", __func__);
+               ms->section_mem_map = 0;
+       }
+
+       if (vmemmap_buf_start) {
+               /* need to free left buf */
+#ifdef CONFIG_NO_BOOTMEM
+               free_early(__pa(vmemmap_buf_start), __pa(vmemmap_buf_end));
+               if (vmemmap_buf_start < vmemmap_buf) {
+                       char name[15];
+
+                       snprintf(name, sizeof(name), "MEMMAP %d", nodeid);
+                       reserve_early_without_check(__pa(vmemmap_buf_start),
+                                                   __pa(vmemmap_buf), name);
+               }
+#else
+               free_bootmem(__pa(vmemmap_buf), vmemmap_buf_end - vmemmap_buf);
+#endif
+               vmemmap_buf = NULL;
+               vmemmap_buf_end = NULL;
+       }
+}