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 void add_one_highpage_init(struct page *, int, int);
104 extern unsigned long highend_pfn, highstart_pfn;
106 #define LARGE_PAGE_BYTES (PTRS_PER_PTE * PAGE_SIZE)
108 unsigned long node_remap_size[MAX_NUMNODES];
109 static void *node_remap_start_vaddr[MAX_NUMNODES];
110 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags);
112 static unsigned long kva_start_pfn;
113 static unsigned long kva_pages;
115 * FLAT - support for basic PC memory model with discontig enabled, essentially
116 * a single node with all available processors in it with a flat
119 int __init get_memcfg_numa_flat(void)
121 printk("NUMA - single node, flat memory mode\n");
123 /* Run the memory configuration and find the top of memory. */
125 node_start_pfn[0] = 0;
126 node_end_pfn[0] = max_pfn;
127 memory_present(0, 0, max_pfn);
128 node_remap_size[0] = node_memmap_size_bytes(0, 0, max_pfn);
130 /* Indicate there is one node available. */
131 nodes_clear(node_online_map);
137 * Find the highest page frame number we have available for the node
139 static void __init propagate_e820_map_node(int nid)
141 if (node_end_pfn[nid] > max_pfn)
142 node_end_pfn[nid] = max_pfn;
144 * if a user has given mem=XXXX, then we need to make sure
145 * that the node _starts_ before that, too, not just ends
147 if (node_start_pfn[nid] > max_pfn)
148 node_start_pfn[nid] = max_pfn;
149 BUG_ON(node_start_pfn[nid] > node_end_pfn[nid]);
153 * Allocate memory for the pg_data_t for this node via a crude pre-bootmem
154 * method. For node zero take this from the bottom of memory, for
155 * subsequent nodes place them at node_remap_start_vaddr which contains
156 * node local data in physically node local memory. See setup_memory()
159 static void __init allocate_pgdat(int nid)
161 if (nid && node_has_online_mem(nid))
162 NODE_DATA(nid) = (pg_data_t *)node_remap_start_vaddr[nid];
164 unsigned long pgdat_phys;
165 pgdat_phys = find_e820_area(min_low_pfn<<PAGE_SHIFT,
166 (nid ? max_low_pfn:max_pfn_mapped)<<PAGE_SHIFT,
169 NODE_DATA(nid) = (pg_data_t *)(pfn_to_kaddr(pgdat_phys>>PAGE_SHIFT));
170 reserve_early(pgdat_phys, pgdat_phys + sizeof(pg_data_t),
173 printk(KERN_DEBUG "allocate_pgdat: node %d NODE_DATA %08lx\n",
174 nid, (unsigned long)NODE_DATA(nid));
177 #ifdef CONFIG_DISCONTIGMEM
179 * In the discontig memory model, a portion of the kernel virtual area (KVA)
180 * is reserved and portions of nodes are mapped using it. This is to allow
181 * node-local memory to be allocated for structures that would normally require
182 * ZONE_NORMAL. The memory is allocated with alloc_remap() and callers
183 * should be prepared to allocate from the bootmem allocator instead. This KVA
184 * mechanism is incompatible with SPARSEMEM as it makes assumptions about the
185 * layout of memory that are broken if alloc_remap() succeeds for some of the
186 * map and fails for others
188 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
189 static void *node_remap_end_vaddr[MAX_NUMNODES];
190 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
191 static unsigned long node_remap_offset[MAX_NUMNODES];
193 void *alloc_remap(int nid, unsigned long size)
195 void *allocation = node_remap_alloc_vaddr[nid];
197 size = ALIGN(size, L1_CACHE_BYTES);
199 if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
202 node_remap_alloc_vaddr[nid] += size;
203 memset(allocation, 0, size);
208 void __init remap_numa_kva(void)
214 for_each_online_node(node) {
215 printk(KERN_DEBUG "remap_numa_kva: node %d\n", node);
216 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
217 vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
218 printk(KERN_DEBUG "remap_numa_kva: %08lx to pfn %08lx\n",
219 (unsigned long)vaddr,
220 node_remap_start_pfn[node] + pfn);
221 set_pmd_pfn((ulong) vaddr,
222 node_remap_start_pfn[node] + pfn,
228 static unsigned long calculate_numa_remap_pages(void)
231 unsigned long size, reserve_pages = 0;
234 for_each_online_node(nid) {
235 unsigned old_end_pfn = node_end_pfn[nid];
238 * The acpi/srat node info can show hot-add memroy zones
239 * where memory could be added but not currently present.
241 if (node_start_pfn[nid] > max_pfn)
243 if (node_end_pfn[nid] > max_pfn)
244 node_end_pfn[nid] = max_pfn;
246 /* ensure the remap includes space for the pgdat. */
247 size = node_remap_size[nid] + sizeof(pg_data_t);
249 /* convert size to large (pmd size) pages, rounding up */
250 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
251 /* now the roundup is correct, convert to PAGE_SIZE pages */
252 size = size * PTRS_PER_PTE;
255 * Validate the region we are allocating only contains valid
258 for (pfn = node_end_pfn[nid] - size;
259 pfn < node_end_pfn[nid]; pfn++)
260 if (!page_is_ram(pfn))
263 if (pfn != node_end_pfn[nid])
266 printk("Reserving %ld pages of KVA for lmem_map of node %d\n",
268 node_remap_size[nid] = size;
269 node_remap_offset[nid] = reserve_pages;
270 reserve_pages += size;
271 printk("Shrinking node %d from %ld pages to %ld pages\n",
272 nid, node_end_pfn[nid], node_end_pfn[nid] - size);
274 if (node_end_pfn[nid] & (PTRS_PER_PTE-1)) {
276 * Align node_end_pfn[] and node_remap_start_pfn[] to
277 * pmd boundary. remap_numa_kva will barf otherwise.
279 printk("Shrinking node %d further by %ld pages for proper alignment\n",
280 nid, node_end_pfn[nid] & (PTRS_PER_PTE-1));
281 size += node_end_pfn[nid] & (PTRS_PER_PTE-1);
284 node_end_pfn[nid] -= size;
285 node_remap_start_pfn[nid] = node_end_pfn[nid];
286 shrink_active_range(nid, old_end_pfn, node_end_pfn[nid]);
288 printk("Reserving total of %ld pages for numa KVA remap\n",
290 return reserve_pages;
293 static void init_remap_allocator(int nid)
295 node_remap_start_vaddr[nid] = pfn_to_kaddr(
296 kva_start_pfn + node_remap_offset[nid]);
297 node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
298 (node_remap_size[nid] * PAGE_SIZE);
299 node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
300 ALIGN(sizeof(pg_data_t), PAGE_SIZE);
302 printk ("node %d will remap to vaddr %08lx - %08lx\n", nid,
303 (ulong) node_remap_start_vaddr[nid],
304 (ulong) node_remap_end_vaddr[nid]);
307 void *alloc_remap(int nid, unsigned long size)
312 static unsigned long calculate_numa_remap_pages(void)
317 static void init_remap_allocator(int nid)
321 void __init remap_numa_kva(void)
324 #endif /* CONFIG_DISCONTIGMEM */
326 extern void setup_bootmem_allocator(void);
327 unsigned long __init setup_memory(void)
330 unsigned long system_start_pfn, system_max_low_pfn;
333 * When mapping a NUMA machine we allocate the node_mem_map arrays
334 * from node local memory. They are then mapped directly into KVA
335 * between zone normal and vmalloc space. Calculate the size of
336 * this space and use it to adjust the boundary between ZONE_NORMAL
341 kva_pages = round_up(calculate_numa_remap_pages(), PTRS_PER_PTE);
343 /* partially used pages are not usable - thus round upwards */
344 system_start_pfn = min_low_pfn = PFN_UP(init_pg_tables_end);
346 system_max_low_pfn = max_low_pfn = find_max_low_pfn();
347 kva_start_pfn = round_down(max_low_pfn - kva_pages, PTRS_PER_PTE);
348 kva_start_pfn = find_e820_area(kva_start_pfn<<PAGE_SHIFT,
349 max_low_pfn<<PAGE_SHIFT,
350 kva_pages<<PAGE_SHIFT,
351 PTRS_PER_PTE<<PAGE_SHIFT) >> PAGE_SHIFT;
353 printk("kva_start_pfn ~ %ld find_max_low_pfn() ~ %ld\n",
354 kva_start_pfn, max_low_pfn);
355 printk("max_pfn = %ld\n", max_pfn);
357 /* avoid clash with initrd */
358 reserve_early(kva_start_pfn<<PAGE_SHIFT,
359 (kva_start_pfn + kva_pages)<<PAGE_SHIFT,
361 #ifdef CONFIG_HIGHMEM
362 highstart_pfn = highend_pfn = max_pfn;
363 if (max_pfn > system_max_low_pfn)
364 highstart_pfn = system_max_low_pfn;
365 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
366 pages_to_mb(highend_pfn - highstart_pfn));
367 num_physpages = highend_pfn;
368 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
370 num_physpages = system_max_low_pfn;
371 high_memory = (void *) __va(system_max_low_pfn * PAGE_SIZE - 1) + 1;
373 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
374 pages_to_mb(system_max_low_pfn));
375 printk("min_low_pfn = %ld, max_low_pfn = %ld, highstart_pfn = %ld\n",
376 min_low_pfn, max_low_pfn, highstart_pfn);
378 printk("Low memory ends at vaddr %08lx\n",
379 (ulong) pfn_to_kaddr(max_low_pfn));
380 for_each_online_node(nid) {
381 init_remap_allocator(nid);
385 printk("High memory starts at vaddr %08lx\n",
386 (ulong) pfn_to_kaddr(highstart_pfn));
387 for_each_online_node(nid)
388 propagate_e820_map_node(nid);
390 memset(NODE_DATA(0), 0, sizeof(struct pglist_data));
391 NODE_DATA(0)->bdata = &node0_bdata;
392 setup_bootmem_allocator();
396 void __init zone_sizes_init(void)
399 unsigned long max_zone_pfns[MAX_NR_ZONES];
400 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
401 max_zone_pfns[ZONE_DMA] =
402 virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
403 max_zone_pfns[ZONE_NORMAL] = max_low_pfn;
404 #ifdef CONFIG_HIGHMEM
405 max_zone_pfns[ZONE_HIGHMEM] = highend_pfn;
408 /* If SRAT has not registered memory, register it now */
409 if (find_max_pfn_with_active_regions() == 0) {
410 for_each_online_node(nid) {
411 if (node_has_online_mem(nid))
412 add_active_range(nid, node_start_pfn[nid],
417 free_area_init_nodes(max_zone_pfns);
421 void __init set_highmem_pages_init(int bad_ppro)
423 #ifdef CONFIG_HIGHMEM
427 for_each_zone(zone) {
428 unsigned long node_pfn, zone_start_pfn, zone_end_pfn;
430 if (!is_highmem(zone))
433 zone_start_pfn = zone->zone_start_pfn;
434 zone_end_pfn = zone_start_pfn + zone->spanned_pages;
436 printk("Initializing %s for node %d (%08lx:%08lx)\n",
437 zone->name, zone_to_nid(zone),
438 zone_start_pfn, zone_end_pfn);
440 for (node_pfn = zone_start_pfn; node_pfn < zone_end_pfn; node_pfn++) {
441 if (!pfn_valid(node_pfn))
443 page = pfn_to_page(node_pfn);
444 add_one_highpage_init(page, node_pfn, bad_ppro);
447 totalram_pages += totalhigh_pages;
451 #ifdef CONFIG_MEMORY_HOTPLUG
452 static int paddr_to_nid(u64 addr)
455 unsigned long pfn = PFN_DOWN(addr);
458 if (node_start_pfn[nid] <= pfn &&
459 pfn < node_end_pfn[nid])
466 * This function is used to ask node id BEFORE memmap and mem_section's
467 * initialization (pfn_to_nid() can't be used yet).
468 * If _PXM is not defined on ACPI's DSDT, node id must be found by this.
470 int memory_add_physaddr_to_nid(u64 addr)
472 int nid = paddr_to_nid(addr);
473 return (nid >= 0) ? nid : 0;
476 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);