Merge branch 'x86/numa' into x86-mm

Merge reason: Pick up x86-32 remap allocator cleanup changes - 14
commits, 3fe14ab541^..993ba1585c.

  3fe14ab541: x86-32, numa: Fix failure condition check in alloc_remap()
  993ba1585c: x86-32, numa: Update remap allocator comments

Scheduled NUMA init 32/64bit unification changes depend on them.

Signed-off-by: Tejun Heo <tj@kernel.org>

2 files changed, 111 insertions, 158 deletions

diff --git a/arch/x86/mm/numa_32.c b/arch/x86/mm/numa_32.c
index bde3906420df..c757c0a3b529 100644
--- a/arch/x86/mm/numa_32.c
+++ b/arch/x86/mm/numa_32.c
@@ -104,13 +104,9 @@ extern unsigned long highend_pfn, highstart_pfn;
 
 #define LARGE_PAGE_BYTES (PTRS_PER_PTE * PAGE_SIZE)
 
-unsigned long node_remap_size[MAX_NUMNODES];
 static void *node_remap_start_vaddr[MAX_NUMNODES];
 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags);
 
-static unsigned long kva_start_pfn;
-static unsigned long kva_pages;
-
 int __cpuinit numa_cpu_node(int cpu)
 {
         return apic->x86_32_numa_cpu_node(cpu);
@@ -129,7 +125,6 @@ int __init get_memcfg_numa_flat(void)
         node_end_pfn[0] = max_pfn;
         memblock_x86_register_active_regions(0, 0, max_pfn);
         memory_present(0, 0, max_pfn);
-        node_remap_size[0] = node_memmap_size_bytes(0, 0, max_pfn);
 
         /* Indicate there is one node available. */
         nodes_clear(node_online_map);
@@ -164,9 +159,8 @@ static void __init allocate_pgdat(int nid)
 {
         char buf[16];
 
-        if (node_has_online_mem(nid) && node_remap_start_vaddr[nid])
-                NODE_DATA(nid) = (pg_data_t *)node_remap_start_vaddr[nid];
-        else {
+        NODE_DATA(nid) = alloc_remap(nid, ALIGN(sizeof(pg_data_t), PAGE_SIZE));
+        if (!NODE_DATA(nid)) {
                 unsigned long pgdat_phys;
                 pgdat_phys = memblock_find_in_range(min_low_pfn<<PAGE_SHIFT,
                                  max_pfn_mapped<<PAGE_SHIFT,
@@ -182,25 +176,38 @@ static void __init allocate_pgdat(int nid)
 }
 
 /*
- * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
- * virtual address space (KVA) is reserved and portions of nodes are mapped
- * using it. This is to allow node-local memory to be allocated for
- * structures that would normally require ZONE_NORMAL. The memory is
- * allocated with alloc_remap() and callers should be prepared to allocate
- * from the bootmem allocator instead.
+ * Remap memory allocator
  */
 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
 static void *node_remap_end_vaddr[MAX_NUMNODES];
 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
-static unsigned long node_remap_offset[MAX_NUMNODES];
 
+/**
+ * alloc_remap - Allocate remapped memory
+ * @nid: NUMA node to allocate memory from
+ * @size: The size of allocation
+ *
+ * Allocate @size bytes from the remap area of NUMA node @nid.  The
+ * size of the remap area is predetermined by init_alloc_remap() and
+ * only the callers considered there should call this function.  For
+ * more info, please read the comment on top of init_alloc_remap().
+ *
+ * The caller must be ready to handle allocation failure from this
+ * function and fall back to regular memory allocator in such cases.
+ *
+ * CONTEXT:
+ * Single CPU early boot context.
+ *
+ * RETURNS:
+ * Pointer to the allocated memory on success, %NULL on failure.
+ */
 void *alloc_remap(int nid, unsigned long size)
 {
         void *allocation = node_remap_alloc_vaddr[nid];
 
         size = ALIGN(size, L1_CACHE_BYTES);
 
-        if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
+        if (!allocation || (allocation + size) > node_remap_end_vaddr[nid])
                 return NULL;
 
         node_remap_alloc_vaddr[nid] += size;
@@ -209,26 +216,6 @@ void *alloc_remap(int nid, unsigned long size)
         return allocation;
 }
 
-static void __init remap_numa_kva(void)
-{
-        void *vaddr;
-        unsigned long pfn;
-        int node;
-
-        for_each_online_node(node) {
-                printk(KERN_DEBUG "remap_numa_kva: node %d\n", node);
-                for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
-                        vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
-                        printk(KERN_DEBUG "remap_numa_kva: %08lx to pfn %08lx\n",
-                                (unsigned long)vaddr,
-                                node_remap_start_pfn[node] + pfn);
-                        set_pmd_pfn((ulong) vaddr, 
-                                node_remap_start_pfn[node] + pfn, 
-                                PAGE_KERNEL_LARGE);
-                }
-        }
-}
-
 #ifdef CONFIG_HIBERNATION
 /**
  * resume_map_numa_kva - add KVA mapping to the temporary page tables created
@@ -240,15 +227,16 @@ void resume_map_numa_kva(pgd_t *pgd_base)
         int node;
 
         for_each_online_node(node) {
-                unsigned long start_va, start_pfn, size, pfn;
+                unsigned long start_va, start_pfn, nr_pages, pfn;
 
                 start_va = (unsigned long)node_remap_start_vaddr[node];
                 start_pfn = node_remap_start_pfn[node];
-                size = node_remap_size[node];
+                nr_pages = (node_remap_end_vaddr[node] -
+                            node_remap_start_vaddr[node]) >> PAGE_SHIFT;
 
                 printk(KERN_DEBUG "%s: node %d\n", __func__, node);
 
-                for (pfn = 0; pfn < size; pfn += PTRS_PER_PTE) {
+                for (pfn = 0; pfn < nr_pages; pfn += PTRS_PER_PTE) {
                         unsigned long vaddr = start_va + (pfn << PAGE_SHIFT);
                         pgd_t *pgd = pgd_base + pgd_index(vaddr);
                         pud_t *pud = pud_offset(pgd, vaddr);
@@ -264,132 +252,102 @@ void resume_map_numa_kva(pgd_t *pgd_base)
 }
 #endif
 
-static __init unsigned long calculate_numa_remap_pages(void)
+/**
+ * init_alloc_remap - Initialize remap allocator for a NUMA node
+ * @nid: NUMA node to initizlie remap allocator for
+ *
+ * NUMA nodes may end up without any lowmem.  As allocating pgdat and
+ * memmap on a different node with lowmem is inefficient, a special
+ * remap allocator is implemented which can be used by alloc_remap().
+ *
+ * For each node, the amount of memory which will be necessary for
+ * pgdat and memmap is calculated and two memory areas of the size are
+ * allocated - one in the node and the other in lowmem; then, the area
+ * in the node is remapped to the lowmem area.
+ *
+ * As pgdat and memmap must be allocated in lowmem anyway, this
+ * doesn't waste lowmem address space; however, the actual lowmem
+ * which gets remapped over is wasted.  The amount shouldn't be
+ * problematic on machines this feature will be used.
+ *
+ * Initialization failure isn't fatal.  alloc_remap() is used
+ * opportunistically and the callers will fall back to other memory
+ * allocation mechanisms on failure.
+ */
+static __init void init_alloc_remap(int nid)
 {
-        int nid;
-        unsigned long size, reserve_pages = 0;
+        unsigned long size, pfn;
+        u64 node_pa, remap_pa;
+        void *remap_va;
 
-        for_each_online_node(nid) {
-                u64 node_kva_target;
-                u64 node_kva_final;
-
-                /*
-                 * The acpi/srat node info can show hot-add memroy zones
-                 * where memory could be added but not currently present.
-                 */
-                printk(KERN_DEBUG "node %d pfn: [%lx - %lx]\n",
-                        nid, node_start_pfn[nid], node_end_pfn[nid]);
-                if (node_start_pfn[nid] > max_pfn)
-                        continue;
-                if (!node_end_pfn[nid])
-                        continue;
-                if (node_end_pfn[nid] > max_pfn)
-                        node_end_pfn[nid] = max_pfn;
-
-                /* ensure the remap includes space for the pgdat. */
-                size = node_remap_size[nid] + sizeof(pg_data_t);
-
-                /* convert size to large (pmd size) pages, rounding up */
-                size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
-                /* now the roundup is correct, convert to PAGE_SIZE pages */
-                size = size * PTRS_PER_PTE;
-
-                node_kva_target = round_down(node_end_pfn[nid] - size,
-                                                 PTRS_PER_PTE);
-                node_kva_target <<= PAGE_SHIFT;
-                do {
-                        node_kva_final = memblock_find_in_range(node_kva_target,
-                                        ((u64)node_end_pfn[nid])<<PAGE_SHIFT,
-                                                ((u64)size)<<PAGE_SHIFT,
-                                                LARGE_PAGE_BYTES);
-                        node_kva_target -= LARGE_PAGE_BYTES;
-                } while (node_kva_final == MEMBLOCK_ERROR &&
-                         (node_kva_target>>PAGE_SHIFT) > (node_start_pfn[nid]));
-
-                if (node_kva_final == MEMBLOCK_ERROR)
-                        panic("Can not get kva ram\n");
-
-                node_remap_size[nid] = size;
-                node_remap_offset[nid] = reserve_pages;
-                reserve_pages += size;
-                printk(KERN_DEBUG "Reserving %ld pages of KVA for lmem_map of"
-                                  " node %d at %llx\n",
-                                size, nid, node_kva_final>>PAGE_SHIFT);
-
-                /*
-                 *  prevent kva address below max_low_pfn want it on system
-                 *  with less memory later.
-                 *  layout will be: KVA address , KVA RAM
-                 *
-                 *  we are supposed to only record the one less then max_low_pfn
-                 *  but we could have some hole in high memory, and it will only
-                 *  check page_is_ram(pfn) && !page_is_reserved_early(pfn) to decide
-                 *  to use it as free.
-                 *  So memblock_x86_reserve_range here, hope we don't run out of that array
-                 */
-                memblock_x86_reserve_range(node_kva_final,
-                              node_kva_final+(((u64)size)<<PAGE_SHIFT),
-                              "KVA RAM");
-
-                node_remap_start_pfn[nid] = node_kva_final>>PAGE_SHIFT;
-        }
-        printk(KERN_INFO "Reserving total of %lx pages for numa KVA remap\n",
-                        reserve_pages);
-        return reserve_pages;
-}
+        /*
+         * The acpi/srat node info can show hot-add memroy zones where
+         * memory could be added but not currently present.
+         */
+        printk(KERN_DEBUG "node %d pfn: [%lx - %lx]\n",
+               nid, node_start_pfn[nid], node_end_pfn[nid]);
+        if (node_start_pfn[nid] > max_pfn)
+                return;
+        if (!node_end_pfn[nid])
+                return;
+        if (node_end_pfn[nid] > max_pfn)
+                node_end_pfn[nid] = max_pfn;
 
-static void init_remap_allocator(int nid)
-{
-        node_remap_start_vaddr[nid] = pfn_to_kaddr(
-                        kva_start_pfn + node_remap_offset[nid]);
-        node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
-                (node_remap_size[nid] * PAGE_SIZE);
-        node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
-                ALIGN(sizeof(pg_data_t), PAGE_SIZE);
-
-        printk(KERN_DEBUG "node %d will remap to vaddr %08lx - %08lx\n", nid,
-                (ulong) node_remap_start_vaddr[nid],
-                (ulong) node_remap_end_vaddr[nid]);
+        /* calculate the necessary space aligned to large page size */
+        size = node_memmap_size_bytes(nid, node_start_pfn[nid],
+                                      min(node_end_pfn[nid], max_pfn));
+        size += ALIGN(sizeof(pg_data_t), PAGE_SIZE);
+        size = ALIGN(size, LARGE_PAGE_BYTES);
+
+        /* allocate node memory and the lowmem remap area */
+        node_pa = memblock_find_in_range(node_start_pfn[nid] << PAGE_SHIFT,
+                                         (u64)node_end_pfn[nid] << PAGE_SHIFT,
+                                         size, LARGE_PAGE_BYTES);
+        if (node_pa == MEMBLOCK_ERROR) {
+                pr_warning("remap_alloc: failed to allocate %lu bytes for node %d\n",
+                           size, nid);
+                return;
+        }
+        memblock_x86_reserve_range(node_pa, node_pa + size, "KVA RAM");
+
+        remap_pa = memblock_find_in_range(min_low_pfn << PAGE_SHIFT,
+                                          max_low_pfn << PAGE_SHIFT,
+                                          size, LARGE_PAGE_BYTES);
+        if (remap_pa == MEMBLOCK_ERROR) {
+                pr_warning("remap_alloc: failed to allocate %lu bytes remap area for node %d\n",
+                           size, nid);
+                memblock_x86_free_range(node_pa, node_pa + size);
+                return;
+        }
+        memblock_x86_reserve_range(remap_pa, remap_pa + size, "KVA PG");
+        remap_va = phys_to_virt(remap_pa);
+
+        /* perform actual remap */
+        for (pfn = 0; pfn < size >> PAGE_SHIFT; pfn += PTRS_PER_PTE)
+                set_pmd_pfn((unsigned long)remap_va + (pfn << PAGE_SHIFT),
+                            (node_pa >> PAGE_SHIFT) + pfn,
+                            PAGE_KERNEL_LARGE);
+
+        /* initialize remap allocator parameters */
+        node_remap_start_pfn[nid] = node_pa >> PAGE_SHIFT;
+        node_remap_start_vaddr[nid] = remap_va;
+        node_remap_end_vaddr[nid] = remap_va + size;
+        node_remap_alloc_vaddr[nid] = remap_va;
+
+        printk(KERN_DEBUG "remap_alloc: node %d [%08llx-%08llx) -> [%p-%p)\n",
+               nid, node_pa, node_pa + size, remap_va, remap_va + size);
 }
 
 void __init initmem_init(void)
 {
         int nid;
-        long kva_target_pfn;
-
-        /*
-         * When mapping a NUMA machine we allocate the node_mem_map arrays
-         * from node local memory.  They are then mapped directly into KVA
-         * between zone normal and vmalloc space.  Calculate the size of
-         * this space and use it to adjust the boundary between ZONE_NORMAL
-         * and ZONE_HIGHMEM.
-         */
 
         get_memcfg_numa();
         numa_init_array();
 
-        kva_pages = roundup(calculate_numa_remap_pages(), PTRS_PER_PTE);
-
-        kva_target_pfn = round_down(max_low_pfn - kva_pages, PTRS_PER_PTE);
-        do {
-                kva_start_pfn = memblock_find_in_range(kva_target_pfn<<PAGE_SHIFT,
-                                        max_low_pfn<<PAGE_SHIFT,
-                                        kva_pages<<PAGE_SHIFT,
-                                        PTRS_PER_PTE<<PAGE_SHIFT) >> PAGE_SHIFT;
-                kva_target_pfn -= PTRS_PER_PTE;
-        } while (kva_start_pfn == MEMBLOCK_ERROR && kva_target_pfn > min_low_pfn);
-
-        if (kva_start_pfn == MEMBLOCK_ERROR)
-                panic("Can not get kva space\n");
-
-        printk(KERN_INFO "kva_start_pfn ~ %lx max_low_pfn ~ %lx\n",
-                kva_start_pfn, max_low_pfn);
-        printk(KERN_INFO "max_pfn = %lx\n", max_pfn);
+        for_each_online_node(nid)
+                init_alloc_remap(nid);
 
-        /* avoid clash with initrd */
-        memblock_x86_reserve_range(kva_start_pfn<<PAGE_SHIFT,
-                      (kva_start_pfn + kva_pages)<<PAGE_SHIFT,
-                     "KVA PG");
 #ifdef CONFIG_HIGHMEM
         highstart_pfn = highend_pfn = max_pfn;
         if (max_pfn > max_low_pfn)
@@ -409,12 +367,8 @@ void __init initmem_init(void)
 
         printk(KERN_DEBUG "Low memory ends at vaddr %08lx\n",
                         (ulong) pfn_to_kaddr(max_low_pfn));
-        for_each_online_node(nid) {
-                init_remap_allocator(nid);
-
+        for_each_online_node(nid)
                 allocate_pgdat(nid);
-        }
-        remap_numa_kva();
 
         printk(KERN_DEBUG "High memory starts at vaddr %08lx\n",
                         (ulong) pfn_to_kaddr(highstart_pfn));
diff --git a/arch/x86/mm/srat_32.c b/arch/x86/mm/srat_32.c
index 48651c6f657d..1b9e82c96dc5 100644
--- a/arch/x86/mm/srat_32.c
+++ b/arch/x86/mm/srat_32.c
@@ -276,7 +276,6 @@ int __init get_memcfg_from_srat(void)
                 unsigned long end = min(node_end_pfn[nid], max_pfn);
 
                 memory_present(nid, start, end);
-                node_remap_size[nid] = node_memmap_size_bytes(nid, start, end);
         }
         return 1;
 out_fail: