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.
25 #include <linux/config.h>
27 #include <linux/bootmem.h>
28 #include <linux/mmzone.h>
29 #include <linux/highmem.h>
30 #include <linux/initrd.h>
31 #include <linux/nodemask.h>
33 #include <asm/setup.h>
34 #include <asm/mmzone.h>
35 #include <asm/crash_dump.h>
36 #include <bios_ebda.h>
38 struct pglist_data *node_data[MAX_NUMNODES];
39 bootmem_data_t node0_bdata;
42 * numa interface - we expect the numa architecture specfic code to have
43 * populated the following initialisation.
45 * 1) numnodes - the total number of nodes configured in the system
46 * 2) physnode_map - the mapping between a pfn and owning node
47 * 3) node_start_pfn - the starting page frame number for a node
48 * 3) node_end_pfn - the ending page fram number for a node
52 * physnode_map keeps track of the physical memory layout of a generic
53 * numa node on a 256Mb break (each element of the array will
54 * represent 256Mb of memory and will be marked by the node id. so,
55 * if the first gig is on node 0, and the second gig is on node 1
56 * physnode_map will contain:
58 * physnode_map[0-3] = 0;
59 * physnode_map[4-7] = 1;
60 * physnode_map[8- ] = -1;
62 s8 physnode_map[MAX_ELEMENTS] = { [0 ... (MAX_ELEMENTS - 1)] = -1};
64 unsigned long node_start_pfn[MAX_NUMNODES];
65 unsigned long node_end_pfn[MAX_NUMNODES];
67 extern unsigned long find_max_low_pfn(void);
68 extern void find_max_pfn(void);
69 extern void one_highpage_init(struct page *, int, int);
71 extern struct e820map e820;
72 extern unsigned long init_pg_tables_end;
73 extern unsigned long highend_pfn, highstart_pfn;
74 extern unsigned long max_low_pfn;
75 extern unsigned long totalram_pages;
76 extern unsigned long totalhigh_pages;
78 #define LARGE_PAGE_BYTES (PTRS_PER_PTE * PAGE_SIZE)
80 unsigned long node_remap_start_pfn[MAX_NUMNODES];
81 unsigned long node_remap_size[MAX_NUMNODES];
82 unsigned long node_remap_offset[MAX_NUMNODES];
83 void *node_remap_start_vaddr[MAX_NUMNODES];
84 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags);
87 * FLAT - support for basic PC memory model with discontig enabled, essentially
88 * a single node with all available processors in it with a flat
91 int __init get_memcfg_numa_flat(void)
93 printk("NUMA - single node, flat memory mode\n");
95 /* Run the memory configuration and find the top of memory. */
97 node_start_pfn[0] = 0;
98 node_end_pfn[0] = max_pfn;
100 /* Indicate there is one node available. */
107 * Find the highest page frame number we have available for the node
109 static void __init find_max_pfn_node(int nid)
111 if (node_end_pfn[nid] > max_pfn)
112 node_end_pfn[nid] = max_pfn;
114 * if a user has given mem=XXXX, then we need to make sure
115 * that the node _starts_ before that, too, not just ends
117 if (node_start_pfn[nid] > max_pfn)
118 node_start_pfn[nid] = max_pfn;
119 if (node_start_pfn[nid] > node_end_pfn[nid])
124 * Allocate memory for the pg_data_t for this node via a crude pre-bootmem
125 * method. For node zero take this from the bottom of memory, for
126 * subsequent nodes place them at node_remap_start_vaddr which contains
127 * node local data in physically node local memory. See setup_memory()
130 static void __init allocate_pgdat(int nid)
133 NODE_DATA(nid) = (pg_data_t *)node_remap_start_vaddr[nid];
135 NODE_DATA(nid) = (pg_data_t *)(__va(min_low_pfn << PAGE_SHIFT));
136 min_low_pfn += PFN_UP(sizeof(pg_data_t));
137 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
142 * Register fully available low RAM pages with the bootmem allocator.
144 static void __init register_bootmem_low_pages(unsigned long system_max_low_pfn)
148 for (i = 0; i < e820.nr_map; i++) {
149 unsigned long curr_pfn, last_pfn, size;
151 * Reserve usable low memory
153 if (e820.map[i].type != E820_RAM)
156 * We are rounding up the start address of usable memory:
158 curr_pfn = PFN_UP(e820.map[i].addr);
159 if (curr_pfn >= system_max_low_pfn)
162 * ... and at the end of the usable range downwards:
164 last_pfn = PFN_DOWN(e820.map[i].addr + e820.map[i].size);
166 if (last_pfn > system_max_low_pfn)
167 last_pfn = system_max_low_pfn;
170 * .. finally, did all the rounding and playing
171 * around just make the area go away?
173 if (last_pfn <= curr_pfn)
176 size = last_pfn - curr_pfn;
177 free_bootmem_node(NODE_DATA(0), PFN_PHYS(curr_pfn), PFN_PHYS(size));
181 void __init remap_numa_kva(void)
187 for (node = 1; node < numnodes; ++node) {
188 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
189 vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
190 set_pmd_pfn((ulong) vaddr,
191 node_remap_start_pfn[node] + pfn,
197 static unsigned long calculate_numa_remap_pages(void)
200 unsigned long size, reserve_pages = 0;
202 for (nid = 1; nid < numnodes; nid++) {
203 /* calculate the size of the mem_map needed in bytes */
204 size = (node_end_pfn[nid] - node_start_pfn[nid] + 1)
205 * sizeof(struct page) + sizeof(pg_data_t);
206 /* convert size to large (pmd size) pages, rounding up */
207 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
208 /* now the roundup is correct, convert to PAGE_SIZE pages */
209 size = size * PTRS_PER_PTE;
210 printk("Reserving %ld pages of KVA for lmem_map of node %d\n",
212 node_remap_size[nid] = size;
213 reserve_pages += size;
214 node_remap_offset[nid] = reserve_pages;
215 printk("Shrinking node %d from %ld pages to %ld pages\n",
216 nid, node_end_pfn[nid], node_end_pfn[nid] - size);
217 node_end_pfn[nid] -= size;
218 node_remap_start_pfn[nid] = node_end_pfn[nid];
220 printk("Reserving total of %ld pages for numa KVA remap\n",
222 return reserve_pages;
226 * workaround for Dell systems that neglect to reserve EBDA
228 static void __init reserve_ebda_region_node(void)
231 addr = get_bios_ebda();
233 reserve_bootmem_node(NODE_DATA(0), addr, PAGE_SIZE);
236 unsigned long __init setup_memory(void)
239 unsigned long bootmap_size, system_start_pfn, system_max_low_pfn;
240 unsigned long reserve_pages, pfn;
243 * When mapping a NUMA machine we allocate the node_mem_map arrays
244 * from node local memory. They are then mapped directly into KVA
245 * between zone normal and vmalloc space. Calculate the size of
246 * this space and use it to adjust the boundry between ZONE_NORMAL
251 /* Fill in the physnode_map */
252 for (nid = 0; nid < numnodes; nid++) {
253 printk("Node: %d, start_pfn: %ld, end_pfn: %ld\n",
254 nid, node_start_pfn[nid], node_end_pfn[nid]);
255 printk(" Setting physnode_map array to node %d for pfns:\n ",
257 for (pfn = node_start_pfn[nid]; pfn < node_end_pfn[nid];
258 pfn += PAGES_PER_ELEMENT) {
259 physnode_map[pfn / PAGES_PER_ELEMENT] = nid;
265 reserve_pages = calculate_numa_remap_pages();
267 /* partially used pages are not usable - thus round upwards */
268 system_start_pfn = min_low_pfn = PFN_UP(init_pg_tables_end);
271 system_max_low_pfn = max_low_pfn = find_max_low_pfn() - reserve_pages;
272 printk("reserve_pages = %ld find_max_low_pfn() ~ %ld\n",
273 reserve_pages, max_low_pfn + reserve_pages);
274 printk("max_pfn = %ld\n", max_pfn);
275 #ifdef CONFIG_HIGHMEM
276 highstart_pfn = highend_pfn = max_pfn;
277 if (max_pfn > system_max_low_pfn)
278 highstart_pfn = system_max_low_pfn;
279 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
280 pages_to_mb(highend_pfn - highstart_pfn));
282 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
283 pages_to_mb(system_max_low_pfn));
284 printk("min_low_pfn = %ld, max_low_pfn = %ld, highstart_pfn = %ld\n",
285 min_low_pfn, max_low_pfn, highstart_pfn);
287 printk("Low memory ends at vaddr %08lx\n",
288 (ulong) pfn_to_kaddr(max_low_pfn));
289 for (nid = 0; nid < numnodes; nid++) {
290 node_remap_start_vaddr[nid] = pfn_to_kaddr(
291 (highstart_pfn + reserve_pages) - node_remap_offset[nid]);
293 printk ("node %d will remap to vaddr %08lx - %08lx\n", nid,
294 (ulong) node_remap_start_vaddr[nid],
295 (ulong) pfn_to_kaddr(highstart_pfn + reserve_pages
296 - node_remap_offset[nid] + node_remap_size[nid]));
298 printk("High memory starts at vaddr %08lx\n",
299 (ulong) pfn_to_kaddr(highstart_pfn));
300 vmalloc_earlyreserve = reserve_pages * PAGE_SIZE;
301 for (nid = 0; nid < numnodes; nid++)
302 find_max_pfn_node(nid);
304 NODE_DATA(0)->bdata = &node0_bdata;
307 * Initialize the boot-time allocator (with low memory only):
309 bootmap_size = init_bootmem_node(NODE_DATA(0), min_low_pfn, 0, system_max_low_pfn);
311 register_bootmem_low_pages(system_max_low_pfn);
314 * Reserve the bootmem bitmap itself as well. We do this in two
315 * steps (first step was init_bootmem()) because this catches
316 * the (very unlikely) case of us accidentally initializing the
317 * bootmem allocator with an invalid RAM area.
319 reserve_bootmem_node(NODE_DATA(0), HIGH_MEMORY, (PFN_PHYS(min_low_pfn) +
320 bootmap_size + PAGE_SIZE-1) - (HIGH_MEMORY));
323 * reserve physical page 0 - it's a special BIOS page on many boxes,
324 * enabling clean reboots, SMP operation, laptop functions.
326 reserve_bootmem_node(NODE_DATA(0), 0, PAGE_SIZE);
329 * But first pinch a few for the stack/trampoline stuff
330 * FIXME: Don't need the extra page at 4K, but need to fix
331 * trampoline before removing it. (see the GDT stuff)
333 reserve_bootmem_node(NODE_DATA(0), PAGE_SIZE, PAGE_SIZE);
335 /* reserve EBDA region, it's a 4K region */
336 reserve_ebda_region_node();
338 #ifdef CONFIG_ACPI_SLEEP
340 * Reserve low memory region for sleep support.
342 acpi_reserve_bootmem();
346 * Find and reserve possible boot-time SMP configuration:
350 #ifdef CONFIG_BLK_DEV_INITRD
351 if (LOADER_TYPE && INITRD_START) {
352 if (INITRD_START + INITRD_SIZE <= (system_max_low_pfn << PAGE_SHIFT)) {
353 reserve_bootmem_node(NODE_DATA(0), INITRD_START, INITRD_SIZE);
355 INITRD_START ? INITRD_START + PAGE_OFFSET : 0;
356 initrd_end = initrd_start+INITRD_SIZE;
359 printk(KERN_ERR "initrd extends beyond end of memory "
360 "(0x%08lx > 0x%08lx)\ndisabling initrd\n",
361 INITRD_START + INITRD_SIZE,
362 system_max_low_pfn << PAGE_SHIFT);
368 crash_reserve_bootmem();
370 return system_max_low_pfn;
373 void __init zone_sizes_init(void)
378 * Insert nodes into pgdat_list backward so they appear in order.
379 * Clobber node 0's links and NULL out pgdat_list before starting.
382 for (nid = numnodes - 1; nid >= 0; nid--) {
384 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
385 NODE_DATA(nid)->pgdat_next = pgdat_list;
386 pgdat_list = NODE_DATA(nid);
389 for (nid = 0; nid < numnodes; nid++) {
390 unsigned long zones_size[MAX_NR_ZONES] = {0, 0, 0};
391 unsigned long *zholes_size;
392 unsigned int max_dma;
394 unsigned long low = max_low_pfn;
395 unsigned long start = node_start_pfn[nid];
396 unsigned long high = node_end_pfn[nid];
398 max_dma = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
401 #ifdef CONFIG_HIGHMEM
402 BUG_ON(start > high);
403 zones_size[ZONE_HIGHMEM] = high - start;
407 zones_size[ZONE_DMA] = low;
409 BUG_ON(max_dma > low);
411 zones_size[ZONE_DMA] = max_dma;
412 zones_size[ZONE_NORMAL] = low - max_dma;
413 #ifdef CONFIG_HIGHMEM
414 zones_size[ZONE_HIGHMEM] = high - low;
418 zholes_size = get_zholes_size(nid);
420 * We let the lmem_map for node 0 be allocated from the
421 * normal bootmem allocator, but other nodes come from the
422 * remapped KVA area - mbligh
425 free_area_init_node(nid, NODE_DATA(nid),
426 zones_size, start, zholes_size);
428 unsigned long lmem_map;
429 lmem_map = (unsigned long)node_remap_start_vaddr[nid];
430 lmem_map += sizeof(pg_data_t) + PAGE_SIZE - 1;
431 lmem_map &= PAGE_MASK;
432 NODE_DATA(nid)->node_mem_map = (struct page *)lmem_map;
433 free_area_init_node(nid, NODE_DATA(nid), zones_size,
440 void __init set_highmem_pages_init(int bad_ppro)
442 #ifdef CONFIG_HIGHMEM
445 for_each_zone(zone) {
446 unsigned long node_pfn, node_high_size, zone_start_pfn;
447 struct page * zone_mem_map;
449 if (!is_highmem(zone))
452 printk("Initializing %s for node %d\n", zone->name,
453 zone->zone_pgdat->node_id);
455 node_high_size = zone->spanned_pages;
456 zone_mem_map = zone->zone_mem_map;
457 zone_start_pfn = zone->zone_start_pfn;
459 for (node_pfn = 0; node_pfn < node_high_size; node_pfn++) {
460 one_highpage_init((struct page *)(zone_mem_map + node_pfn),
461 zone_start_pfn + node_pfn, bad_ppro);
464 totalram_pages += totalhigh_pages;
468 void __init set_max_mapnr_init(void)
470 #ifdef CONFIG_HIGHMEM
471 struct zone *high0 = &NODE_DATA(0)->node_zones[ZONE_HIGHMEM];
472 if (high0->spanned_pages > 0)
473 highmem_start_page = high0->zone_mem_map;
475 highmem_start_page = pfn_to_page(max_low_pfn - 1) + 1;
476 num_physpages = highend_pfn;
478 num_physpages = max_low_pfn;