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: %lx, end_pfn: %lx\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 "%lx ", 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(KERN_DEBUG "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)
161 if (node_has_online_mem(nid) && node_remap_start_vaddr[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 max_pfn_mapped<<PAGE_SHIFT,
169 NODE_DATA(nid) = (pg_data_t *)(pfn_to_kaddr(pgdat_phys>>PAGE_SHIFT));
170 memset(buf, 0, sizeof(buf));
171 sprintf(buf, "NODE_DATA %d", nid);
172 reserve_early(pgdat_phys, pgdat_phys + sizeof(pg_data_t), buf);
174 printk(KERN_DEBUG "allocate_pgdat: node %d NODE_DATA %08lx\n",
175 nid, (unsigned long)NODE_DATA(nid));
179 * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
180 * virtual address space (KVA) is reserved and portions of nodes are mapped
181 * using it. This is to allow node-local memory to be allocated for
182 * structures that would normally require ZONE_NORMAL. The memory is
183 * allocated with alloc_remap() and callers should be prepared to allocate
184 * from the bootmem allocator instead.
186 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
187 static void *node_remap_end_vaddr[MAX_NUMNODES];
188 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
189 static unsigned long node_remap_offset[MAX_NUMNODES];
191 void *alloc_remap(int nid, unsigned long size)
193 void *allocation = node_remap_alloc_vaddr[nid];
195 size = ALIGN(size, L1_CACHE_BYTES);
197 if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
200 node_remap_alloc_vaddr[nid] += size;
201 memset(allocation, 0, size);
206 static void __init remap_numa_kva(void)
212 for_each_online_node(node) {
213 printk(KERN_DEBUG "remap_numa_kva: node %d\n", node);
214 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
215 vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
216 printk(KERN_DEBUG "remap_numa_kva: %08lx to pfn %08lx\n",
217 (unsigned long)vaddr,
218 node_remap_start_pfn[node] + pfn);
219 set_pmd_pfn((ulong) vaddr,
220 node_remap_start_pfn[node] + pfn,
226 static unsigned long calculate_numa_remap_pages(void)
229 unsigned long size, reserve_pages = 0;
231 for_each_online_node(nid) {
236 * The acpi/srat node info can show hot-add memroy zones
237 * where memory could be added but not currently present.
239 printk(KERN_DEBUG "node %d pfn: [%lx - %lx]\n",
240 nid, node_start_pfn[nid], node_end_pfn[nid]);
241 if (node_start_pfn[nid] > max_pfn)
243 if (!node_end_pfn[nid])
245 if (node_end_pfn[nid] > max_pfn)
246 node_end_pfn[nid] = max_pfn;
248 /* ensure the remap includes space for the pgdat. */
249 size = node_remap_size[nid] + sizeof(pg_data_t);
251 /* convert size to large (pmd size) pages, rounding up */
252 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
253 /* now the roundup is correct, convert to PAGE_SIZE pages */
254 size = size * PTRS_PER_PTE;
256 node_kva_target = round_down(node_end_pfn[nid] - size,
258 node_kva_target <<= PAGE_SHIFT;
260 node_kva_final = find_e820_area(node_kva_target,
261 ((u64)node_end_pfn[nid])<<PAGE_SHIFT,
262 ((u64)size)<<PAGE_SHIFT,
264 node_kva_target -= LARGE_PAGE_BYTES;
265 } while (node_kva_final == -1ULL &&
266 (node_kva_target>>PAGE_SHIFT) > (node_start_pfn[nid]));
268 if (node_kva_final == -1ULL)
269 panic("Can not get kva ram\n");
271 node_remap_size[nid] = size;
272 node_remap_offset[nid] = reserve_pages;
273 reserve_pages += size;
274 printk(KERN_DEBUG "Reserving %ld pages of KVA for lmem_map of"
275 " node %d at %llx\n",
276 size, nid, node_kva_final>>PAGE_SHIFT);
279 * prevent kva address below max_low_pfn want it on system
280 * with less memory later.
281 * layout will be: KVA address , KVA RAM
283 * we are supposed to only record the one less then max_low_pfn
284 * but we could have some hole in high memory, and it will only
285 * check page_is_ram(pfn) && !page_is_reserved_early(pfn) to decide
287 * So reserve_early here, hope we don't run out of that array
289 reserve_early(node_kva_final,
290 node_kva_final+(((u64)size)<<PAGE_SHIFT),
293 node_remap_start_pfn[nid] = node_kva_final>>PAGE_SHIFT;
294 remove_active_range(nid, node_remap_start_pfn[nid],
295 node_remap_start_pfn[nid] + size);
297 printk(KERN_INFO "Reserving total of %lx pages for numa KVA remap\n",
299 return reserve_pages;
302 static void init_remap_allocator(int nid)
304 node_remap_start_vaddr[nid] = pfn_to_kaddr(
305 kva_start_pfn + node_remap_offset[nid]);
306 node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
307 (node_remap_size[nid] * PAGE_SIZE);
308 node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
309 ALIGN(sizeof(pg_data_t), PAGE_SIZE);
311 printk(KERN_DEBUG "node %d will remap to vaddr %08lx - %08lx\n", nid,
312 (ulong) node_remap_start_vaddr[nid],
313 (ulong) node_remap_end_vaddr[nid]);
316 void __init initmem_init(unsigned long start_pfn,
317 unsigned long end_pfn)
323 * When mapping a NUMA machine we allocate the node_mem_map arrays
324 * from node local memory. They are then mapped directly into KVA
325 * between zone normal and vmalloc space. Calculate the size of
326 * this space and use it to adjust the boundary between ZONE_NORMAL
332 kva_pages = round_up(calculate_numa_remap_pages(), PTRS_PER_PTE);
334 kva_target_pfn = round_down(max_low_pfn - kva_pages, PTRS_PER_PTE);
336 kva_start_pfn = find_e820_area(kva_target_pfn<<PAGE_SHIFT,
337 max_low_pfn<<PAGE_SHIFT,
338 kva_pages<<PAGE_SHIFT,
339 PTRS_PER_PTE<<PAGE_SHIFT) >> PAGE_SHIFT;
340 kva_target_pfn -= PTRS_PER_PTE;
341 } while (kva_start_pfn == -1UL && kva_target_pfn > min_low_pfn);
343 if (kva_start_pfn == -1UL)
344 panic("Can not get kva space\n");
346 printk(KERN_INFO "kva_start_pfn ~ %lx max_low_pfn ~ %lx\n",
347 kva_start_pfn, max_low_pfn);
348 printk(KERN_INFO "max_pfn = %lx\n", max_pfn);
350 /* avoid clash with initrd */
351 reserve_early(kva_start_pfn<<PAGE_SHIFT,
352 (kva_start_pfn + kva_pages)<<PAGE_SHIFT,
354 #ifdef CONFIG_HIGHMEM
355 highstart_pfn = highend_pfn = max_pfn;
356 if (max_pfn > max_low_pfn)
357 highstart_pfn = max_low_pfn;
358 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
359 pages_to_mb(highend_pfn - highstart_pfn));
360 num_physpages = highend_pfn;
361 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
363 num_physpages = max_low_pfn;
364 high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
366 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
367 pages_to_mb(max_low_pfn));
368 printk(KERN_DEBUG "max_low_pfn = %lx, highstart_pfn = %lx\n",
369 max_low_pfn, highstart_pfn);
371 printk(KERN_DEBUG "Low memory ends at vaddr %08lx\n",
372 (ulong) pfn_to_kaddr(max_low_pfn));
373 for_each_online_node(nid) {
374 init_remap_allocator(nid);
380 printk(KERN_DEBUG "High memory starts at vaddr %08lx\n",
381 (ulong) pfn_to_kaddr(highstart_pfn));
382 for_each_online_node(nid)
383 propagate_e820_map_node(nid);
385 for_each_online_node(nid)
386 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
388 NODE_DATA(0)->bdata = &node0_bdata;
389 setup_bootmem_allocator();
392 void __init set_highmem_pages_init(void)
394 #ifdef CONFIG_HIGHMEM
398 for_each_zone(zone) {
399 unsigned long zone_start_pfn, zone_end_pfn;
401 if (!is_highmem(zone))
404 zone_start_pfn = zone->zone_start_pfn;
405 zone_end_pfn = zone_start_pfn + zone->spanned_pages;
407 nid = zone_to_nid(zone);
408 printk(KERN_INFO "Initializing %s for node %d (%08lx:%08lx)\n",
409 zone->name, nid, zone_start_pfn, zone_end_pfn);
411 add_highpages_with_active_regions(nid, zone_start_pfn,
414 totalram_pages += totalhigh_pages;
418 #ifdef CONFIG_MEMORY_HOTPLUG
419 static int paddr_to_nid(u64 addr)
422 unsigned long pfn = PFN_DOWN(addr);
425 if (node_start_pfn[nid] <= pfn &&
426 pfn < node_end_pfn[nid])
433 * This function is used to ask node id BEFORE memmap and mem_section's
434 * initialization (pfn_to_nid() can't be used yet).
435 * If _PXM is not defined on ACPI's DSDT, node id must be found by this.
437 int memory_add_physaddr_to_nid(u64 addr)
439 int nid = paddr_to_nid(addr);
440 return (nid >= 0) ? nid : 0;
443 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);