2 * Written by: Patricia Gaughen <gone@us.ibm.com>, IBM Corporation
3 * August 2002: added remote node KVA remap - Martin J. Bligh
5 * Copyright (C) 2002, IBM Corp.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
14 * This program is distributed in the hope that it will be useful, but
15 * WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
17 * NON INFRINGEMENT. See the GNU General Public License for more
20 * You should have received a copy of the GNU General Public License
21 * along with this program; if not, write to the Free Software
22 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
26 #include <linux/bootmem.h>
27 #include <linux/mmzone.h>
28 #include <linux/highmem.h>
29 #include <linux/initrd.h>
30 #include <linux/nodemask.h>
31 #include <linux/module.h>
32 #include <linux/kexec.h>
33 #include <linux/pfn.h>
34 #include <linux/swap.h>
35 #include <linux/acpi.h>
38 #include <asm/setup.h>
39 #include <asm/mmzone.h>
40 #include <asm/bios_ebda.h>
41 #include <asm/proto.h>
43 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
44 EXPORT_SYMBOL(node_data);
45 static bootmem_data_t node0_bdata;
48 * numa interface - we expect the numa architecture specific code to have
49 * populated the following initialisation.
51 * 1) node_online_map - the map of all nodes configured (online) in the system
52 * 2) node_start_pfn - the starting page frame number for a node
53 * 3) node_end_pfn - the ending page fram number for a node
55 unsigned long node_start_pfn[MAX_NUMNODES] __read_mostly;
56 unsigned long node_end_pfn[MAX_NUMNODES] __read_mostly;
59 #ifdef CONFIG_DISCONTIGMEM
61 * 4) physnode_map - the mapping between a pfn and owning node
62 * physnode_map keeps track of the physical memory layout of a generic
63 * numa node on a 64Mb break (each element of the array will
64 * represent 64Mb of memory and will be marked by the node id. so,
65 * if the first gig is on node 0, and the second gig is on node 1
66 * physnode_map will contain:
68 * physnode_map[0-15] = 0;
69 * physnode_map[16-31] = 1;
70 * physnode_map[32- ] = -1;
72 s8 physnode_map[MAX_ELEMENTS] __read_mostly = { [0 ... (MAX_ELEMENTS - 1)] = -1};
73 EXPORT_SYMBOL(physnode_map);
75 void memory_present(int nid, unsigned long start, unsigned long end)
79 printk(KERN_INFO "Node: %d, start_pfn: %ld, end_pfn: %ld\n",
81 printk(KERN_DEBUG " Setting physnode_map array to node %d for pfns:\n", nid);
82 printk(KERN_DEBUG " ");
83 for (pfn = start; pfn < end; pfn += PAGES_PER_ELEMENT) {
84 physnode_map[pfn / PAGES_PER_ELEMENT] = nid;
85 printk(KERN_CONT "%ld ", pfn);
87 printk(KERN_CONT "\n");
90 unsigned long node_memmap_size_bytes(int nid, unsigned long start_pfn,
91 unsigned long end_pfn)
93 unsigned long nr_pages = end_pfn - start_pfn;
98 return (nr_pages + 1) * sizeof(struct page);
102 extern unsigned long find_max_low_pfn(void);
103 extern unsigned long highend_pfn, highstart_pfn;
105 #define LARGE_PAGE_BYTES (PTRS_PER_PTE * PAGE_SIZE)
107 unsigned long node_remap_size[MAX_NUMNODES];
108 static void *node_remap_start_vaddr[MAX_NUMNODES];
109 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags);
111 static unsigned long kva_start_pfn;
112 static unsigned long kva_pages;
114 * FLAT - support for basic PC memory model with discontig enabled, essentially
115 * a single node with all available processors in it with a flat
118 int __init get_memcfg_numa_flat(void)
120 printk("NUMA - single node, flat memory mode\n");
122 node_start_pfn[0] = 0;
123 node_end_pfn[0] = max_pfn;
124 e820_register_active_regions(0, 0, max_pfn);
125 memory_present(0, 0, max_pfn);
126 node_remap_size[0] = node_memmap_size_bytes(0, 0, max_pfn);
128 /* Indicate there is one node available. */
129 nodes_clear(node_online_map);
135 * Find the highest page frame number we have available for the node
137 static void __init propagate_e820_map_node(int nid)
139 if (node_end_pfn[nid] > max_pfn)
140 node_end_pfn[nid] = max_pfn;
142 * if a user has given mem=XXXX, then we need to make sure
143 * that the node _starts_ before that, too, not just ends
145 if (node_start_pfn[nid] > max_pfn)
146 node_start_pfn[nid] = max_pfn;
147 BUG_ON(node_start_pfn[nid] > node_end_pfn[nid]);
151 * Allocate memory for the pg_data_t for this node via a crude pre-bootmem
152 * method. For node zero take this from the bottom of memory, for
153 * subsequent nodes place them at node_remap_start_vaddr which contains
154 * node local data in physically node local memory. See setup_memory()
157 static void __init allocate_pgdat(int nid)
159 if (nid && node_has_online_mem(nid) && node_remap_start_vaddr[nid])
160 NODE_DATA(nid) = (pg_data_t *)node_remap_start_vaddr[nid];
162 unsigned long pgdat_phys;
163 pgdat_phys = find_e820_area(min_low_pfn<<PAGE_SHIFT,
164 (nid ? max_low_pfn:max_pfn_mapped)<<PAGE_SHIFT,
167 NODE_DATA(nid) = (pg_data_t *)(pfn_to_kaddr(pgdat_phys>>PAGE_SHIFT));
168 reserve_early(pgdat_phys, pgdat_phys + sizeof(pg_data_t),
171 printk(KERN_DEBUG "allocate_pgdat: node %d NODE_DATA %08lx\n",
172 nid, (unsigned long)NODE_DATA(nid));
176 * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
177 * virtual address space (KVA) is reserved and portions of nodes are mapped
178 * using it. This is to allow node-local memory to be allocated for
179 * structures that would normally require ZONE_NORMAL. The memory is
180 * allocated with alloc_remap() and callers should be prepared to allocate
181 * from the bootmem allocator instead.
183 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
184 static void *node_remap_end_vaddr[MAX_NUMNODES];
185 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
186 static unsigned long node_remap_offset[MAX_NUMNODES];
188 void *alloc_remap(int nid, unsigned long size)
190 void *allocation = node_remap_alloc_vaddr[nid];
192 size = ALIGN(size, L1_CACHE_BYTES);
194 if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
197 node_remap_alloc_vaddr[nid] += size;
198 memset(allocation, 0, size);
203 void __init remap_numa_kva(void)
209 for_each_online_node(node) {
210 printk(KERN_DEBUG "remap_numa_kva: node %d\n", node);
211 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
212 vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
213 printk(KERN_DEBUG "remap_numa_kva: %08lx to pfn %08lx\n",
214 (unsigned long)vaddr,
215 node_remap_start_pfn[node] + pfn);
216 set_pmd_pfn((ulong) vaddr,
217 node_remap_start_pfn[node] + pfn,
223 static unsigned long calculate_numa_remap_pages(void)
226 unsigned long size, reserve_pages = 0;
228 for_each_online_node(nid) {
233 * The acpi/srat node info can show hot-add memroy zones
234 * where memory could be added but not currently present.
236 printk("node %d pfn: [%lx - %lx]\n",
237 nid, node_start_pfn[nid], node_end_pfn[nid]);
238 if (node_start_pfn[nid] > max_pfn)
240 if (!node_end_pfn[nid])
242 if (node_end_pfn[nid] > max_pfn)
243 node_end_pfn[nid] = max_pfn;
245 /* ensure the remap includes space for the pgdat. */
246 size = node_remap_size[nid] + sizeof(pg_data_t);
248 /* convert size to large (pmd size) pages, rounding up */
249 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
250 /* now the roundup is correct, convert to PAGE_SIZE pages */
251 size = size * PTRS_PER_PTE;
253 node_kva_target = round_down(node_end_pfn[nid] - size,
255 node_kva_target <<= PAGE_SHIFT;
257 node_kva_final = find_e820_area(node_kva_target,
258 ((u64)node_end_pfn[nid])<<PAGE_SHIFT,
259 ((u64)size)<<PAGE_SHIFT,
261 node_kva_target -= LARGE_PAGE_BYTES;
262 } while (node_kva_final == -1ULL &&
263 (node_kva_target>>PAGE_SHIFT) > (node_start_pfn[nid]));
265 if (node_kva_final == -1ULL)
266 panic("Can not get kva ram\n");
268 node_remap_size[nid] = size;
269 node_remap_offset[nid] = reserve_pages;
270 reserve_pages += size;
271 printk("Reserving %ld pages of KVA for lmem_map of node %d at %llx\n",
272 size, nid, node_kva_final>>PAGE_SHIFT);
275 * prevent kva address below max_low_pfn want it on system
276 * with less memory later.
277 * layout will be: KVA address , KVA RAM
279 * we are supposed to only record the one less then max_low_pfn
280 * but we could have some hole in high memory, and it will only
281 * check page_is_ram(pfn) && !page_is_reserved_early(pfn) to decide
283 * So reserve_early here, hope we don't run out of that array
285 reserve_early(node_kva_final,
286 node_kva_final+(((u64)size)<<PAGE_SHIFT),
289 node_remap_start_pfn[nid] = node_kva_final>>PAGE_SHIFT;
290 remove_active_range(nid, node_remap_start_pfn[nid],
291 node_remap_start_pfn[nid] + size);
293 printk("Reserving total of %ld pages for numa KVA remap\n",
295 return reserve_pages;
298 static void init_remap_allocator(int nid)
300 node_remap_start_vaddr[nid] = pfn_to_kaddr(
301 kva_start_pfn + node_remap_offset[nid]);
302 node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
303 (node_remap_size[nid] * PAGE_SIZE);
304 node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
305 ALIGN(sizeof(pg_data_t), PAGE_SIZE);
307 printk ("node %d will remap to vaddr %08lx - %08lx\n", nid,
308 (ulong) node_remap_start_vaddr[nid],
309 (ulong) node_remap_end_vaddr[nid]);
312 extern void setup_bootmem_allocator(void);
313 unsigned long __init setup_memory(void)
316 unsigned long system_start_pfn, system_max_low_pfn;
320 * When mapping a NUMA machine we allocate the node_mem_map arrays
321 * from node local memory. They are then mapped directly into KVA
322 * between zone normal and vmalloc space. Calculate the size of
323 * this space and use it to adjust the boundary between ZONE_NORMAL
327 /* call find_max_low_pfn at first, it could update max_pfn */
328 system_max_low_pfn = max_low_pfn = find_max_low_pfn();
330 remove_all_active_ranges();
333 kva_pages = round_up(calculate_numa_remap_pages(), PTRS_PER_PTE);
335 /* partially used pages are not usable - thus round upwards */
336 system_start_pfn = min_low_pfn = PFN_UP(init_pg_tables_end);
338 kva_target_pfn = round_down(max_low_pfn - kva_pages, PTRS_PER_PTE);
340 kva_start_pfn = find_e820_area(kva_target_pfn<<PAGE_SHIFT,
341 max_low_pfn<<PAGE_SHIFT,
342 kva_pages<<PAGE_SHIFT,
343 PTRS_PER_PTE<<PAGE_SHIFT) >> PAGE_SHIFT;
344 kva_target_pfn -= PTRS_PER_PTE;
345 } while (kva_start_pfn == -1UL && kva_target_pfn > min_low_pfn);
347 if (kva_start_pfn == -1UL)
348 panic("Can not get kva space\n");
350 printk("kva_start_pfn ~ %ld find_max_low_pfn() ~ %ld\n",
351 kva_start_pfn, max_low_pfn);
352 printk("max_pfn = %ld\n", max_pfn);
354 /* avoid clash with initrd */
355 reserve_early(kva_start_pfn<<PAGE_SHIFT,
356 (kva_start_pfn + kva_pages)<<PAGE_SHIFT,
358 #ifdef CONFIG_HIGHMEM
359 highstart_pfn = highend_pfn = max_pfn;
360 if (max_pfn > system_max_low_pfn)
361 highstart_pfn = system_max_low_pfn;
362 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
363 pages_to_mb(highend_pfn - highstart_pfn));
364 num_physpages = highend_pfn;
365 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
367 num_physpages = system_max_low_pfn;
368 high_memory = (void *) __va(system_max_low_pfn * PAGE_SIZE - 1) + 1;
370 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
371 pages_to_mb(system_max_low_pfn));
372 printk("min_low_pfn = %ld, max_low_pfn = %ld, highstart_pfn = %ld\n",
373 min_low_pfn, max_low_pfn, highstart_pfn);
375 printk("Low memory ends at vaddr %08lx\n",
376 (ulong) pfn_to_kaddr(max_low_pfn));
377 for_each_online_node(nid) {
378 init_remap_allocator(nid);
382 printk("High memory starts at vaddr %08lx\n",
383 (ulong) pfn_to_kaddr(highstart_pfn));
384 for_each_online_node(nid)
385 propagate_e820_map_node(nid);
387 memset(NODE_DATA(0), 0, sizeof(struct pglist_data));
388 NODE_DATA(0)->bdata = &node0_bdata;
389 setup_bootmem_allocator();
393 void __init zone_sizes_init(void)
395 unsigned long max_zone_pfns[MAX_NR_ZONES];
396 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
397 max_zone_pfns[ZONE_DMA] =
398 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
399 max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
400 #ifdef CONFIG_HIGHMEM
401 max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
404 free_area_init_nodes(max_zone_pfns);
408 void __init set_highmem_pages_init(void)
410 #ifdef CONFIG_HIGHMEM
414 for_each_zone(zone) {
415 unsigned long zone_start_pfn, zone_end_pfn;
417 if (!is_highmem(zone))
420 zone_start_pfn = zone->zone_start_pfn;
421 zone_end_pfn = zone_start_pfn + zone->spanned_pages;
423 nid = zone_to_nid(zone);
424 printk("Initializing %s for node %d (%08lx:%08lx)\n",
425 zone->name, nid, zone_start_pfn, zone_end_pfn);
427 add_highpages_with_active_regions(nid, zone_start_pfn,
430 totalram_pages += totalhigh_pages;
434 #ifdef CONFIG_MEMORY_HOTPLUG
435 static int paddr_to_nid(u64 addr)
438 unsigned long pfn = PFN_DOWN(addr);
441 if (node_start_pfn[nid] <= pfn &&
442 pfn < node_end_pfn[nid])
449 * This function is used to ask node id BEFORE memmap and mem_section's
450 * initialization (pfn_to_nid() can't be used yet).
451 * If _PXM is not defined on ACPI's DSDT, node id must be found by this.
453 int memory_add_physaddr_to_nid(u64 addr)
455 int nid = paddr_to_nid(addr);
456 return (nid >= 0) ? nid : 0;
459 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
462 #if defined(CONFIG_ACPI_NUMA) && !defined(CONFIG_HAVE_ARCH_PARSE_SRAT)
464 * Dummy on 32-bit, for now:
466 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
471 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
475 void __init acpi_numa_arch_fixup(void)