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
;
140 if (node_start_pfn
[nid
] > max_pfn
)
141 node_start_pfn
[nid
] = max_pfn
;
142 BUG_ON(node_start_pfn
[nid
] > node_end_pfn
[nid
]);
146 * Allocate memory for the pg_data_t for this node via a crude pre-bootmem
147 * method. For node zero take this from the bottom of memory, for
148 * subsequent nodes place them at node_remap_start_vaddr which contains
149 * node local data in physically node local memory. See setup_memory()
152 static void __init
allocate_pgdat(int nid
)
156 if (node_has_online_mem(nid
) && node_remap_start_vaddr
[nid
])
157 NODE_DATA(nid
) = (pg_data_t
*)node_remap_start_vaddr
[nid
];
159 unsigned long pgdat_phys
;
160 pgdat_phys
= find_e820_area(min_low_pfn
<<PAGE_SHIFT
,
161 max_pfn_mapped
<<PAGE_SHIFT
,
164 NODE_DATA(nid
) = (pg_data_t
*)(pfn_to_kaddr(pgdat_phys
>>PAGE_SHIFT
));
165 memset(buf
, 0, sizeof(buf
));
166 sprintf(buf
, "NODE_DATA %d", nid
);
167 reserve_early(pgdat_phys
, pgdat_phys
+ sizeof(pg_data_t
), buf
);
169 printk(KERN_DEBUG
"allocate_pgdat: node %d NODE_DATA %08lx\n",
170 nid
, (unsigned long)NODE_DATA(nid
));
174 * In the DISCONTIGMEM and SPARSEMEM memory model, a portion of the kernel
175 * virtual address space (KVA) is reserved and portions of nodes are mapped
176 * using it. This is to allow node-local memory to be allocated for
177 * structures that would normally require ZONE_NORMAL. The memory is
178 * allocated with alloc_remap() and callers should be prepared to allocate
179 * from the bootmem allocator instead.
181 static unsigned long node_remap_start_pfn
[MAX_NUMNODES
];
182 static void *node_remap_end_vaddr
[MAX_NUMNODES
];
183 static void *node_remap_alloc_vaddr
[MAX_NUMNODES
];
184 static unsigned long node_remap_offset
[MAX_NUMNODES
];
186 void *alloc_remap(int nid
, unsigned long size
)
188 void *allocation
= node_remap_alloc_vaddr
[nid
];
190 size
= ALIGN(size
, L1_CACHE_BYTES
);
192 if (!allocation
|| (allocation
+ size
) >= node_remap_end_vaddr
[nid
])
195 node_remap_alloc_vaddr
[nid
] += size
;
196 memset(allocation
, 0, size
);
201 static void __init
remap_numa_kva(void)
207 for_each_online_node(node
) {
208 printk(KERN_DEBUG
"remap_numa_kva: node %d\n", node
);
209 for (pfn
=0; pfn
< node_remap_size
[node
]; pfn
+= PTRS_PER_PTE
) {
210 vaddr
= node_remap_start_vaddr
[node
]+(pfn
<<PAGE_SHIFT
);
211 printk(KERN_DEBUG
"remap_numa_kva: %08lx to pfn %08lx\n",
212 (unsigned long)vaddr
,
213 node_remap_start_pfn
[node
] + pfn
);
214 set_pmd_pfn((ulong
) vaddr
,
215 node_remap_start_pfn
[node
] + pfn
,
221 #ifdef CONFIG_HIBERNATION
223 * resume_map_numa_kva - add KVA mapping to the temporary page tables created
224 * during resume from hibernation
225 * @pgd_base - temporary resume page directory
227 void resume_map_numa_kva(pgd_t
*pgd_base
)
231 for_each_online_node(node
) {
232 unsigned long start_va
, start_pfn
, size
, pfn
;
234 start_va
= (unsigned long)node_remap_start_vaddr
[node
];
235 start_pfn
= node_remap_start_pfn
[node
];
236 size
= node_remap_size
[node
];
238 printk(KERN_DEBUG
"%s: node %d\n", __func__
, node
);
240 for (pfn
= 0; pfn
< size
; pfn
+= PTRS_PER_PTE
) {
241 unsigned long vaddr
= start_va
+ (pfn
<< PAGE_SHIFT
);
242 pgd_t
*pgd
= pgd_base
+ pgd_index(vaddr
);
243 pud_t
*pud
= pud_offset(pgd
, vaddr
);
244 pmd_t
*pmd
= pmd_offset(pud
, vaddr
);
246 set_pmd(pmd
, pfn_pmd(start_pfn
+ pfn
,
247 PAGE_KERNEL_LARGE_EXEC
));
249 printk(KERN_DEBUG
"%s: %08lx -> pfn %08lx\n",
250 __func__
, vaddr
, start_pfn
+ pfn
);
256 static __init
unsigned long calculate_numa_remap_pages(void)
259 unsigned long size
, reserve_pages
= 0;
261 for_each_online_node(nid
) {
266 * The acpi/srat node info can show hot-add memroy zones
267 * where memory could be added but not currently present.
269 printk(KERN_DEBUG
"node %d pfn: [%lx - %lx]\n",
270 nid
, node_start_pfn
[nid
], node_end_pfn
[nid
]);
271 if (node_start_pfn
[nid
] > max_pfn
)
273 if (!node_end_pfn
[nid
])
275 if (node_end_pfn
[nid
] > max_pfn
)
276 node_end_pfn
[nid
] = max_pfn
;
278 /* ensure the remap includes space for the pgdat. */
279 size
= node_remap_size
[nid
] + sizeof(pg_data_t
);
281 /* convert size to large (pmd size) pages, rounding up */
282 size
= (size
+ LARGE_PAGE_BYTES
- 1) / LARGE_PAGE_BYTES
;
283 /* now the roundup is correct, convert to PAGE_SIZE pages */
284 size
= size
* PTRS_PER_PTE
;
286 node_kva_target
= round_down(node_end_pfn
[nid
] - size
,
288 node_kva_target
<<= PAGE_SHIFT
;
290 node_kva_final
= find_e820_area(node_kva_target
,
291 ((u64
)node_end_pfn
[nid
])<<PAGE_SHIFT
,
292 ((u64
)size
)<<PAGE_SHIFT
,
294 node_kva_target
-= LARGE_PAGE_BYTES
;
295 } while (node_kva_final
== -1ULL &&
296 (node_kva_target
>>PAGE_SHIFT
) > (node_start_pfn
[nid
]));
298 if (node_kva_final
== -1ULL)
299 panic("Can not get kva ram\n");
301 node_remap_size
[nid
] = size
;
302 node_remap_offset
[nid
] = reserve_pages
;
303 reserve_pages
+= size
;
304 printk(KERN_DEBUG
"Reserving %ld pages of KVA for lmem_map of"
305 " node %d at %llx\n",
306 size
, nid
, node_kva_final
>>PAGE_SHIFT
);
309 * prevent kva address below max_low_pfn want it on system
310 * with less memory later.
311 * layout will be: KVA address , KVA RAM
313 * we are supposed to only record the one less then max_low_pfn
314 * but we could have some hole in high memory, and it will only
315 * check page_is_ram(pfn) && !page_is_reserved_early(pfn) to decide
317 * So reserve_early here, hope we don't run out of that array
319 reserve_early(node_kva_final
,
320 node_kva_final
+(((u64
)size
)<<PAGE_SHIFT
),
323 node_remap_start_pfn
[nid
] = node_kva_final
>>PAGE_SHIFT
;
324 remove_active_range(nid
, node_remap_start_pfn
[nid
],
325 node_remap_start_pfn
[nid
] + size
);
327 printk(KERN_INFO
"Reserving total of %lx pages for numa KVA remap\n",
329 return reserve_pages
;
332 static void init_remap_allocator(int nid
)
334 node_remap_start_vaddr
[nid
] = pfn_to_kaddr(
335 kva_start_pfn
+ node_remap_offset
[nid
]);
336 node_remap_end_vaddr
[nid
] = node_remap_start_vaddr
[nid
] +
337 (node_remap_size
[nid
] * PAGE_SIZE
);
338 node_remap_alloc_vaddr
[nid
] = node_remap_start_vaddr
[nid
] +
339 ALIGN(sizeof(pg_data_t
), PAGE_SIZE
);
341 printk(KERN_DEBUG
"node %d will remap to vaddr %08lx - %08lx\n", nid
,
342 (ulong
) node_remap_start_vaddr
[nid
],
343 (ulong
) node_remap_end_vaddr
[nid
]);
346 void __init
initmem_init(unsigned long start_pfn
, unsigned long end_pfn
,
353 * When mapping a NUMA machine we allocate the node_mem_map arrays
354 * from node local memory. They are then mapped directly into KVA
355 * between zone normal and vmalloc space. Calculate the size of
356 * this space and use it to adjust the boundary between ZONE_NORMAL
362 kva_pages
= roundup(calculate_numa_remap_pages(), PTRS_PER_PTE
);
364 kva_target_pfn
= round_down(max_low_pfn
- kva_pages
, PTRS_PER_PTE
);
366 kva_start_pfn
= find_e820_area(kva_target_pfn
<<PAGE_SHIFT
,
367 max_low_pfn
<<PAGE_SHIFT
,
368 kva_pages
<<PAGE_SHIFT
,
369 PTRS_PER_PTE
<<PAGE_SHIFT
) >> PAGE_SHIFT
;
370 kva_target_pfn
-= PTRS_PER_PTE
;
371 } while (kva_start_pfn
== -1UL && kva_target_pfn
> min_low_pfn
);
373 if (kva_start_pfn
== -1UL)
374 panic("Can not get kva space\n");
376 printk(KERN_INFO
"kva_start_pfn ~ %lx max_low_pfn ~ %lx\n",
377 kva_start_pfn
, max_low_pfn
);
378 printk(KERN_INFO
"max_pfn = %lx\n", max_pfn
);
380 /* avoid clash with initrd */
381 reserve_early(kva_start_pfn
<<PAGE_SHIFT
,
382 (kva_start_pfn
+ kva_pages
)<<PAGE_SHIFT
,
384 #ifdef CONFIG_HIGHMEM
385 highstart_pfn
= highend_pfn
= max_pfn
;
386 if (max_pfn
> max_low_pfn
)
387 highstart_pfn
= max_low_pfn
;
388 printk(KERN_NOTICE
"%ldMB HIGHMEM available.\n",
389 pages_to_mb(highend_pfn
- highstart_pfn
));
390 num_physpages
= highend_pfn
;
391 high_memory
= (void *) __va(highstart_pfn
* PAGE_SIZE
- 1) + 1;
393 num_physpages
= max_low_pfn
;
394 high_memory
= (void *) __va(max_low_pfn
* PAGE_SIZE
- 1) + 1;
396 printk(KERN_NOTICE
"%ldMB LOWMEM available.\n",
397 pages_to_mb(max_low_pfn
));
398 printk(KERN_DEBUG
"max_low_pfn = %lx, highstart_pfn = %lx\n",
399 max_low_pfn
, highstart_pfn
);
401 printk(KERN_DEBUG
"Low memory ends at vaddr %08lx\n",
402 (ulong
) pfn_to_kaddr(max_low_pfn
));
403 for_each_online_node(nid
) {
404 init_remap_allocator(nid
);
410 printk(KERN_DEBUG
"High memory starts at vaddr %08lx\n",
411 (ulong
) pfn_to_kaddr(highstart_pfn
));
412 for_each_online_node(nid
)
413 propagate_e820_map_node(nid
);
415 for_each_online_node(nid
) {
416 memset(NODE_DATA(nid
), 0, sizeof(struct pglist_data
));
417 NODE_DATA(nid
)->node_id
= nid
;
418 #ifndef CONFIG_NO_BOOTMEM
419 NODE_DATA(nid
)->bdata
= &bootmem_node_data
[nid
];
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
);