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>
42 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
43 EXPORT_SYMBOL(node_data);
44 static bootmem_data_t node0_bdata;
47 * numa interface - we expect the numa architecture specific code to have
48 * populated the following initialisation.
50 * 1) node_online_map - the map of all nodes configured (online) in the system
51 * 2) node_start_pfn - the starting page frame number for a node
52 * 3) node_end_pfn - the ending page fram number for a node
54 unsigned long node_start_pfn[MAX_NUMNODES] __read_mostly;
55 unsigned long node_end_pfn[MAX_NUMNODES] __read_mostly;
58 #ifdef CONFIG_DISCONTIGMEM
60 * 4) physnode_map - the mapping between a pfn and owning node
61 * physnode_map keeps track of the physical memory layout of a generic
62 * numa node on a 256Mb break (each element of the array will
63 * represent 256Mb of memory and will be marked by the node id. so,
64 * if the first gig is on node 0, and the second gig is on node 1
65 * physnode_map will contain:
67 * physnode_map[0-3] = 0;
68 * physnode_map[4-7] = 1;
69 * physnode_map[8- ] = -1;
71 s8 physnode_map[MAX_ELEMENTS] __read_mostly = { [0 ... (MAX_ELEMENTS - 1)] = -1};
72 EXPORT_SYMBOL(physnode_map);
74 void memory_present(int nid, unsigned long start, unsigned long end)
78 printk(KERN_INFO "Node: %d, start_pfn: %ld, end_pfn: %ld\n",
80 printk(KERN_DEBUG " Setting physnode_map array to node %d for pfns:\n", nid);
81 printk(KERN_DEBUG " ");
82 for (pfn = start; pfn < end; pfn += PAGES_PER_ELEMENT) {
83 physnode_map[pfn / PAGES_PER_ELEMENT] = nid;
89 unsigned long node_memmap_size_bytes(int nid, unsigned long start_pfn,
90 unsigned long end_pfn)
92 unsigned long nr_pages = end_pfn - start_pfn;
97 return (nr_pages + 1) * sizeof(struct page);
101 extern unsigned long find_max_low_pfn(void);
102 extern void add_one_highpage_init(struct page *, int, int);
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 /* Run the memory configuration and find the top of memory. */
123 propagate_e820_map();
124 node_start_pfn[0] = 0;
125 node_end_pfn[0] = max_pfn;
126 memory_present(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 NODE_DATA(nid) = (pg_data_t *)(pfn_to_kaddr(min_low_pfn));
163 min_low_pfn += PFN_UP(sizeof(pg_data_t));
168 * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
169 * virtual address space (KVA) is reserved and portions of nodes are mapped
170 * using it. This is to allow node-local memory to be allocated for
171 * structures that would normally require ZONE_NORMAL. The memory is
172 * allocated with alloc_remap() and callers should be prepared to allocate
173 * from the bootmem allocator instead.
175 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
176 static void *node_remap_end_vaddr[MAX_NUMNODES];
177 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
178 static unsigned long node_remap_offset[MAX_NUMNODES];
180 void *alloc_remap(int nid, unsigned long size)
182 void *allocation = node_remap_alloc_vaddr[nid];
184 size = ALIGN(size, L1_CACHE_BYTES);
186 if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
189 node_remap_alloc_vaddr[nid] += size;
190 memset(allocation, 0, size);
195 void __init remap_numa_kva(void)
201 for_each_online_node(node) {
202 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
203 vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
204 set_pmd_pfn((ulong) vaddr,
205 node_remap_start_pfn[node] + pfn,
211 static unsigned long calculate_numa_remap_pages(void)
214 unsigned long size, reserve_pages = 0;
217 for_each_online_node(nid) {
218 unsigned old_end_pfn = node_end_pfn[nid];
221 * The acpi/srat node info can show hot-add memroy zones
222 * where memory could be added but not currently present.
224 if (node_start_pfn[nid] > max_pfn)
226 if (node_end_pfn[nid] > max_pfn)
227 node_end_pfn[nid] = max_pfn;
229 /* ensure the remap includes space for the pgdat. */
230 size = node_remap_size[nid] + sizeof(pg_data_t);
232 /* convert size to large (pmd size) pages, rounding up */
233 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
234 /* now the roundup is correct, convert to PAGE_SIZE pages */
235 size = size * PTRS_PER_PTE;
238 * Validate the region we are allocating only contains valid
241 for (pfn = node_end_pfn[nid] - size;
242 pfn < node_end_pfn[nid]; pfn++)
243 if (!page_is_ram(pfn))
246 if (pfn != node_end_pfn[nid])
249 printk("Reserving %ld pages of KVA for lmem_map of node %d\n",
251 node_remap_size[nid] = size;
252 node_remap_offset[nid] = reserve_pages;
253 reserve_pages += size;
254 printk("Shrinking node %d from %ld pages to %ld pages\n",
255 nid, node_end_pfn[nid], node_end_pfn[nid] - size);
257 if (node_end_pfn[nid] & (PTRS_PER_PTE-1)) {
259 * Align node_end_pfn[] and node_remap_start_pfn[] to
260 * pmd boundary. remap_numa_kva will barf otherwise.
262 printk("Shrinking node %d further by %ld pages for proper alignment\n",
263 nid, node_end_pfn[nid] & (PTRS_PER_PTE-1));
264 size += node_end_pfn[nid] & (PTRS_PER_PTE-1);
267 node_end_pfn[nid] -= size;
268 node_remap_start_pfn[nid] = node_end_pfn[nid];
269 shrink_active_range(nid, old_end_pfn, node_end_pfn[nid]);
271 printk("Reserving total of %ld pages for numa KVA remap\n",
273 return reserve_pages;
276 static void init_remap_allocator(int nid)
278 node_remap_start_vaddr[nid] = pfn_to_kaddr(
279 kva_start_pfn + node_remap_offset[nid]);
280 node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
281 (node_remap_size[nid] * PAGE_SIZE);
282 node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
283 ALIGN(sizeof(pg_data_t), PAGE_SIZE);
285 printk ("node %d will remap to vaddr %08lx - %08lx\n", nid,
286 (ulong) node_remap_start_vaddr[nid],
287 (ulong) pfn_to_kaddr(highstart_pfn
288 + node_remap_offset[nid] + node_remap_size[nid]));
291 extern void setup_bootmem_allocator(void);
292 unsigned long __init setup_memory(void)
295 unsigned long system_start_pfn, system_max_low_pfn;
296 unsigned long wasted_pages;
299 * When mapping a NUMA machine we allocate the node_mem_map arrays
300 * from node local memory. They are then mapped directly into KVA
301 * between zone normal and vmalloc space. Calculate the size of
302 * this space and use it to adjust the boundary between ZONE_NORMAL
307 kva_pages = calculate_numa_remap_pages();
309 /* partially used pages are not usable - thus round upwards */
310 system_start_pfn = min_low_pfn = PFN_UP(init_pg_tables_end);
312 kva_start_pfn = find_max_low_pfn() - kva_pages;
314 #ifdef CONFIG_BLK_DEV_INITRD
315 /* Numa kva area is below the initrd */
317 kva_start_pfn = PFN_DOWN(initrd_start - PAGE_OFFSET)
322 * We waste pages past at the end of the KVA for no good reason other
323 * than how it is located. This is bad.
325 wasted_pages = kva_start_pfn & (PTRS_PER_PTE-1);
326 kva_start_pfn -= wasted_pages;
327 kva_pages += wasted_pages;
329 system_max_low_pfn = max_low_pfn = find_max_low_pfn();
330 printk("kva_start_pfn ~ %ld find_max_low_pfn() ~ %ld\n",
331 kva_start_pfn, max_low_pfn);
332 printk("max_pfn = %ld\n", max_pfn);
333 #ifdef CONFIG_HIGHMEM
334 highstart_pfn = highend_pfn = max_pfn;
335 if (max_pfn > system_max_low_pfn)
336 highstart_pfn = system_max_low_pfn;
337 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
338 pages_to_mb(highend_pfn - highstart_pfn));
339 num_physpages = highend_pfn;
340 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
342 num_physpages = system_max_low_pfn;
343 high_memory = (void *) __va(system_max_low_pfn * PAGE_SIZE - 1) + 1;
345 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
346 pages_to_mb(system_max_low_pfn));
347 printk("min_low_pfn = %ld, max_low_pfn = %ld, highstart_pfn = %ld\n",
348 min_low_pfn, max_low_pfn, highstart_pfn);
350 printk("Low memory ends at vaddr %08lx\n",
351 (ulong) pfn_to_kaddr(max_low_pfn));
352 for_each_online_node(nid) {
353 init_remap_allocator(nid);
357 printk("High memory starts at vaddr %08lx\n",
358 (ulong) pfn_to_kaddr(highstart_pfn));
359 for_each_online_node(nid)
360 propagate_e820_map_node(nid);
362 memset(NODE_DATA(0), 0, sizeof(struct pglist_data));
363 NODE_DATA(0)->bdata = &node0_bdata;
364 setup_bootmem_allocator();
368 void __init numa_kva_reserve(void)
371 reserve_bootmem(PFN_PHYS(kva_start_pfn), PFN_PHYS(kva_pages),
375 void __init zone_sizes_init(void)
378 unsigned long max_zone_pfns[MAX_NR_ZONES];
379 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
380 max_zone_pfns[ZONE_DMA] =
381 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
382 max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
383 #ifdef CONFIG_HIGHMEM
384 max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
387 /* If SRAT has not registered memory, register it now */
388 if (find_max_pfn_with_active_regions() == 0) {
389 for_each_online_node(nid) {
390 if (node_has_online_mem(nid))
391 add_active_range(nid, node_start_pfn[nid],
396 free_area_init_nodes(max_zone_pfns);
400 void __init set_highmem_pages_init(int bad_ppro)
402 #ifdef CONFIG_HIGHMEM
406 for_each_zone(zone) {
407 unsigned long node_pfn, zone_start_pfn, zone_end_pfn;
409 if (!is_highmem(zone))
412 zone_start_pfn = zone->zone_start_pfn;
413 zone_end_pfn = zone_start_pfn + zone->spanned_pages;
415 printk("Initializing %s for node %d (%08lx:%08lx)\n",
416 zone->name, zone_to_nid(zone),
417 zone_start_pfn, zone_end_pfn);
419 for (node_pfn = zone_start_pfn; node_pfn < zone_end_pfn; node_pfn++) {
420 if (!pfn_valid(node_pfn))
422 page = pfn_to_page(node_pfn);
423 add_one_highpage_init(page, node_pfn, bad_ppro);
426 totalram_pages += totalhigh_pages;
430 #ifdef CONFIG_MEMORY_HOTPLUG
431 static int paddr_to_nid(u64 addr)
434 unsigned long pfn = PFN_DOWN(addr);
437 if (node_start_pfn[nid] <= pfn &&
438 pfn < node_end_pfn[nid])
445 * This function is used to ask node id BEFORE memmap and mem_section's
446 * initialization (pfn_to_nid() can't be used yet).
447 * If _PXM is not defined on ACPI's DSDT, node id must be found by this.
449 int memory_add_physaddr_to_nid(u64 addr)
451 int nid = paddr_to_nid(addr);
452 return (nid >= 0) ? nid : 0;
455 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
458 #if defined(CONFIG_ACPI_NUMA) && !defined(CONFIG_HAVE_ARCH_PARSE_SRAT)
460 * Dummy on 32-bit, for now:
462 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
467 acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
471 void __init acpi_numa_arch_fixup(void)