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);
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 64Mb break (each element of the array will
63 * represent 64Mb 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[0-15] = 0;
68 * physnode_map[16-31] = 1;
69 * physnode_map[32- ] = -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: %lx, end_pfn: %lx\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;
84 printk(KERN_CONT "%lx ", pfn);
86 printk(KERN_CONT "\n");
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 unsigned long highend_pfn, highstart_pfn;
104 #define LARGE_PAGE_BYTES (PTRS_PER_PTE * PAGE_SIZE)
106 unsigned long node_remap_size[MAX_NUMNODES];
107 static void *node_remap_start_vaddr[MAX_NUMNODES];
108 void set_pmd_pfn(unsigned long vaddr, unsigned long pfn, pgprot_t flags);
110 static unsigned long kva_start_pfn;
111 static unsigned long kva_pages;
113 * FLAT - support for basic PC memory model with discontig enabled, essentially
114 * a single node with all available processors in it with a flat
117 int __init get_memcfg_numa_flat(void)
119 printk(KERN_DEBUG "NUMA - single node, flat memory mode\n");
121 node_start_pfn[0] = 0;
122 node_end_pfn[0] = max_pfn;
123 e820_register_active_regions(0, 0, max_pfn);
124 memory_present(0, 0, max_pfn);
125 node_remap_size[0] = node_memmap_size_bytes(0, 0, max_pfn);
127 /* Indicate there is one node available. */
128 nodes_clear(node_online_map);
134 * Find the highest page frame number we have available for the node
136 static void __init propagate_e820_map_node(int nid)
138 if (node_end_pfn[nid] > max_pfn)
139 node_end_pfn[nid] = max_pfn;
141 * if a user has given mem=XXXX, then we need to make sure
142 * that the node _starts_ before that, too, not just ends
144 if (node_start_pfn[nid] > max_pfn)
145 node_start_pfn[nid] = max_pfn;
146 BUG_ON(node_start_pfn[nid] > node_end_pfn[nid]);
150 * Allocate memory for the pg_data_t for this node via a crude pre-bootmem
151 * method. For node zero take this from the bottom of memory, for
152 * subsequent nodes place them at node_remap_start_vaddr which contains
153 * node local data in physically node local memory. See setup_memory()
156 static void __init allocate_pgdat(int nid)
160 if (node_has_online_mem(nid) && node_remap_start_vaddr[nid])
161 NODE_DATA(nid) = (pg_data_t *)node_remap_start_vaddr[nid];
163 unsigned long pgdat_phys;
164 pgdat_phys = find_e820_area(min_low_pfn<<PAGE_SHIFT,
165 max_pfn_mapped<<PAGE_SHIFT,
168 NODE_DATA(nid) = (pg_data_t *)(pfn_to_kaddr(pgdat_phys>>PAGE_SHIFT));
169 memset(buf, 0, sizeof(buf));
170 sprintf(buf, "NODE_DATA %d", nid);
171 reserve_early(pgdat_phys, pgdat_phys + sizeof(pg_data_t), buf);
173 printk(KERN_DEBUG "allocate_pgdat: node %d NODE_DATA %08lx\n",
174 nid, (unsigned long)NODE_DATA(nid));
178 * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
179 * virtual address space (KVA) is reserved and portions of nodes are mapped
180 * using it. This is to allow node-local memory to be allocated for
181 * structures that would normally require ZONE_NORMAL. The memory is
182 * allocated with alloc_remap() and callers should be prepared to allocate
183 * from the bootmem allocator instead.
185 static unsigned long node_remap_start_pfn[MAX_NUMNODES];
186 static void *node_remap_end_vaddr[MAX_NUMNODES];
187 static void *node_remap_alloc_vaddr[MAX_NUMNODES];
188 static unsigned long node_remap_offset[MAX_NUMNODES];
190 void *alloc_remap(int nid, unsigned long size)
192 void *allocation = node_remap_alloc_vaddr[nid];
194 size = ALIGN(size, L1_CACHE_BYTES);
196 if (!allocation || (allocation + size) >= node_remap_end_vaddr[nid])
199 node_remap_alloc_vaddr[nid] += size;
200 memset(allocation, 0, size);
205 static void __init remap_numa_kva(void)
211 for_each_online_node(node) {
212 printk(KERN_DEBUG "remap_numa_kva: node %d\n", node);
213 for (pfn=0; pfn < node_remap_size[node]; pfn += PTRS_PER_PTE) {
214 vaddr = node_remap_start_vaddr[node]+(pfn<<PAGE_SHIFT);
215 printk(KERN_DEBUG "remap_numa_kva: %08lx to pfn %08lx\n",
216 (unsigned long)vaddr,
217 node_remap_start_pfn[node] + pfn);
218 set_pmd_pfn((ulong) vaddr,
219 node_remap_start_pfn[node] + pfn,
225 #ifdef CONFIG_HIBERNATION
227 * resume_map_numa_kva - add KVA mapping to the temporary page tables created
228 * during resume from hibernation
229 * @pgd_base - temporary resume page directory
231 void resume_map_numa_kva(pgd_t *pgd_base)
235 for_each_online_node(node) {
236 unsigned long start_va, start_pfn, size, pfn;
238 start_va = (unsigned long)node_remap_start_vaddr[node];
239 start_pfn = node_remap_start_pfn[node];
240 size = node_remap_size[node];
242 printk(KERN_DEBUG "%s: node %d\n", __func__, node);
244 for (pfn = 0; pfn < size; pfn += PTRS_PER_PTE) {
245 unsigned long vaddr = start_va + (pfn << PAGE_SHIFT);
246 pgd_t *pgd = pgd_base + pgd_index(vaddr);
247 pud_t *pud = pud_offset(pgd, vaddr);
248 pmd_t *pmd = pmd_offset(pud, vaddr);
250 set_pmd(pmd, pfn_pmd(start_pfn + pfn,
251 PAGE_KERNEL_LARGE_EXEC));
253 printk(KERN_DEBUG "%s: %08lx -> pfn %08lx\n",
254 __func__, vaddr, start_pfn + pfn);
260 static unsigned long calculate_numa_remap_pages(void)
263 unsigned long size, reserve_pages = 0;
265 for_each_online_node(nid) {
270 * The acpi/srat node info can show hot-add memroy zones
271 * where memory could be added but not currently present.
273 printk(KERN_DEBUG "node %d pfn: [%lx - %lx]\n",
274 nid, node_start_pfn[nid], node_end_pfn[nid]);
275 if (node_start_pfn[nid] > max_pfn)
277 if (!node_end_pfn[nid])
279 if (node_end_pfn[nid] > max_pfn)
280 node_end_pfn[nid] = max_pfn;
282 /* ensure the remap includes space for the pgdat. */
283 size = node_remap_size[nid] + sizeof(pg_data_t);
285 /* convert size to large (pmd size) pages, rounding up */
286 size = (size + LARGE_PAGE_BYTES - 1) / LARGE_PAGE_BYTES;
287 /* now the roundup is correct, convert to PAGE_SIZE pages */
288 size = size * PTRS_PER_PTE;
290 node_kva_target = round_down(node_end_pfn[nid] - size,
292 node_kva_target <<= PAGE_SHIFT;
294 node_kva_final = find_e820_area(node_kva_target,
295 ((u64)node_end_pfn[nid])<<PAGE_SHIFT,
296 ((u64)size)<<PAGE_SHIFT,
298 node_kva_target -= LARGE_PAGE_BYTES;
299 } while (node_kva_final == -1ULL &&
300 (node_kva_target>>PAGE_SHIFT) > (node_start_pfn[nid]));
302 if (node_kva_final == -1ULL)
303 panic("Can not get kva ram\n");
305 node_remap_size[nid] = size;
306 node_remap_offset[nid] = reserve_pages;
307 reserve_pages += size;
308 printk(KERN_DEBUG "Reserving %ld pages of KVA for lmem_map of"
309 " node %d at %llx\n",
310 size, nid, node_kva_final>>PAGE_SHIFT);
313 * prevent kva address below max_low_pfn want it on system
314 * with less memory later.
315 * layout will be: KVA address , KVA RAM
317 * we are supposed to only record the one less then max_low_pfn
318 * but we could have some hole in high memory, and it will only
319 * check page_is_ram(pfn) && !page_is_reserved_early(pfn) to decide
321 * So reserve_early here, hope we don't run out of that array
323 reserve_early(node_kva_final,
324 node_kva_final+(((u64)size)<<PAGE_SHIFT),
327 node_remap_start_pfn[nid] = node_kva_final>>PAGE_SHIFT;
328 remove_active_range(nid, node_remap_start_pfn[nid],
329 node_remap_start_pfn[nid] + size);
331 printk(KERN_INFO "Reserving total of %lx pages for numa KVA remap\n",
333 return reserve_pages;
336 static void init_remap_allocator(int nid)
338 node_remap_start_vaddr[nid] = pfn_to_kaddr(
339 kva_start_pfn + node_remap_offset[nid]);
340 node_remap_end_vaddr[nid] = node_remap_start_vaddr[nid] +
341 (node_remap_size[nid] * PAGE_SIZE);
342 node_remap_alloc_vaddr[nid] = node_remap_start_vaddr[nid] +
343 ALIGN(sizeof(pg_data_t), PAGE_SIZE);
345 printk(KERN_DEBUG "node %d will remap to vaddr %08lx - %08lx\n", nid,
346 (ulong) node_remap_start_vaddr[nid],
347 (ulong) node_remap_end_vaddr[nid]);
350 void __init initmem_init(unsigned long start_pfn,
351 unsigned long end_pfn)
357 * When mapping a NUMA machine we allocate the node_mem_map arrays
358 * from node local memory. They are then mapped directly into KVA
359 * between zone normal and vmalloc space. Calculate the size of
360 * this space and use it to adjust the boundary between ZONE_NORMAL
366 kva_pages = roundup(calculate_numa_remap_pages(), PTRS_PER_PTE);
368 kva_target_pfn = round_down(max_low_pfn - kva_pages, PTRS_PER_PTE);
370 kva_start_pfn = find_e820_area(kva_target_pfn<<PAGE_SHIFT,
371 max_low_pfn<<PAGE_SHIFT,
372 kva_pages<<PAGE_SHIFT,
373 PTRS_PER_PTE<<PAGE_SHIFT) >> PAGE_SHIFT;
374 kva_target_pfn -= PTRS_PER_PTE;
375 } while (kva_start_pfn == -1UL && kva_target_pfn > min_low_pfn);
377 if (kva_start_pfn == -1UL)
378 panic("Can not get kva space\n");
380 printk(KERN_INFO "kva_start_pfn ~ %lx max_low_pfn ~ %lx\n",
381 kva_start_pfn, max_low_pfn);
382 printk(KERN_INFO "max_pfn = %lx\n", max_pfn);
384 /* avoid clash with initrd */
385 reserve_early(kva_start_pfn<<PAGE_SHIFT,
386 (kva_start_pfn + kva_pages)<<PAGE_SHIFT,
388 #ifdef CONFIG_HIGHMEM
389 highstart_pfn = highend_pfn = max_pfn;
390 if (max_pfn > max_low_pfn)
391 highstart_pfn = max_low_pfn;
392 printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
393 pages_to_mb(highend_pfn - highstart_pfn));
394 num_physpages = highend_pfn;
395 high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
397 num_physpages = max_low_pfn;
398 high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
400 printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
401 pages_to_mb(max_low_pfn));
402 printk(KERN_DEBUG "max_low_pfn = %lx, highstart_pfn = %lx\n",
403 max_low_pfn, highstart_pfn);
405 printk(KERN_DEBUG "Low memory ends at vaddr %08lx\n",
406 (ulong) pfn_to_kaddr(max_low_pfn));
407 for_each_online_node(nid) {
408 init_remap_allocator(nid);
414 printk(KERN_DEBUG "High memory starts at vaddr %08lx\n",
415 (ulong) pfn_to_kaddr(highstart_pfn));
416 for_each_online_node(nid)
417 propagate_e820_map_node(nid);
419 for_each_online_node(nid)
420 memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
422 NODE_DATA(0)->bdata = &bootmem_node_data[0];
423 setup_bootmem_allocator();
426 #ifdef CONFIG_MEMORY_HOTPLUG
427 static int paddr_to_nid(u64 addr)
430 unsigned long pfn = PFN_DOWN(addr);
433 if (node_start_pfn[nid] <= pfn &&
434 pfn < node_end_pfn[nid])
441 * This function is used to ask node id BEFORE memmap and mem_section's
442 * initialization (pfn_to_nid() can't be used yet).
443 * If _PXM is not defined on ACPI's DSDT, node id must be found by this.
445 int memory_add_physaddr_to_nid(u64 addr)
447 int nid = paddr_to_nid(addr);
448 return (nid >= 0) ? nid : 0;
451 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);