2 * Generic VM initialization for x86-64 NUMA setups.
3 * Copyright 2002,2003 Andi Kleen, SuSE Labs.
5 #include <linux/kernel.h>
7 #include <linux/string.h>
8 #include <linux/init.h>
9 #include <linux/bootmem.h>
10 #include <linux/memblock.h>
11 #include <linux/mmzone.h>
12 #include <linux/ctype.h>
13 #include <linux/module.h>
14 #include <linux/nodemask.h>
15 #include <linux/sched.h>
16 #include <linux/acpi.h>
19 #include <asm/proto.h>
23 #include <asm/amd_nb.h>
33 struct numa_memblk blk[NR_NODE_MEMBLKS];
36 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
37 EXPORT_SYMBOL(node_data);
39 nodemask_t numa_nodes_parsed __initdata;
41 struct memnode memnode;
43 static unsigned long __initdata nodemap_addr;
44 static unsigned long __initdata nodemap_size;
46 static struct numa_meminfo numa_meminfo __initdata;
48 static int numa_distance_cnt;
49 static u8 *numa_distance;
51 #ifdef CONFIG_NUMA_EMU
52 static bool numa_emu_dist;
56 * Given a shift value, try to populate memnodemap[]
59 * 0 if memnodmap[] too small (of shift too small)
60 * -1 if node overlap or lost ram (shift too big)
62 static int __init populate_memnodemap(const struct numa_meminfo *mi, int shift)
64 unsigned long addr, end;
67 memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
68 for (i = 0; i < mi->nr_blks; i++) {
69 addr = mi->blk[i].start;
73 if ((end >> shift) >= memnodemapsize)
76 if (memnodemap[addr >> shift] != NUMA_NO_NODE)
78 memnodemap[addr >> shift] = mi->blk[i].nid;
79 addr += (1UL << shift);
86 static int __init allocate_cachealigned_memnodemap(void)
90 memnodemap = memnode.embedded_map;
91 if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
95 nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
96 nodemap_addr = memblock_find_in_range(addr, get_max_mapped(),
97 nodemap_size, L1_CACHE_BYTES);
98 if (nodemap_addr == MEMBLOCK_ERROR) {
100 "NUMA: Unable to allocate Memory to Node hash map\n");
101 nodemap_addr = nodemap_size = 0;
104 memnodemap = phys_to_virt(nodemap_addr);
105 memblock_x86_reserve_range(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
107 printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
108 nodemap_addr, nodemap_addr + nodemap_size);
113 * The LSB of all start and end addresses in the node map is the value of the
114 * maximum possible shift.
116 static int __init extract_lsb_from_nodes(const struct numa_meminfo *mi)
118 int i, nodes_used = 0;
119 unsigned long start, end;
120 unsigned long bitfield = 0, memtop = 0;
122 for (i = 0; i < mi->nr_blks; i++) {
123 start = mi->blk[i].start;
124 end = mi->blk[i].end;
135 i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
136 memnodemapsize = (memtop >> i)+1;
140 static int __init compute_hash_shift(const struct numa_meminfo *mi)
144 shift = extract_lsb_from_nodes(mi);
145 if (allocate_cachealigned_memnodemap())
147 printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
150 if (populate_memnodemap(mi, shift) != 1) {
151 printk(KERN_INFO "Your memory is not aligned you need to "
152 "rebuild your kernel with a bigger NODEMAPSIZE "
153 "shift=%d\n", shift);
159 int __meminit __early_pfn_to_nid(unsigned long pfn)
161 return phys_to_nid(pfn << PAGE_SHIFT);
164 static void * __init early_node_mem(int nodeid, unsigned long start,
165 unsigned long end, unsigned long size,
171 * put it on high as possible
172 * something will go with NODE_DATA
174 if (start < (MAX_DMA_PFN<<PAGE_SHIFT))
175 start = MAX_DMA_PFN<<PAGE_SHIFT;
176 if (start < (MAX_DMA32_PFN<<PAGE_SHIFT) &&
177 end > (MAX_DMA32_PFN<<PAGE_SHIFT))
178 start = MAX_DMA32_PFN<<PAGE_SHIFT;
179 mem = memblock_x86_find_in_range_node(nodeid, start, end, size, align);
180 if (mem != MEMBLOCK_ERROR)
183 /* extend the search scope */
184 end = max_pfn_mapped << PAGE_SHIFT;
185 start = MAX_DMA_PFN << PAGE_SHIFT;
186 mem = memblock_find_in_range(start, end, size, align);
187 if (mem != MEMBLOCK_ERROR)
190 printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
196 int __init numa_add_memblk(int nid, u64 start, u64 end)
198 struct numa_meminfo *mi = &numa_meminfo;
200 /* ignore zero length blks */
204 /* whine about and ignore invalid blks */
205 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
206 pr_warning("NUMA: Warning: invalid memblk node %d (%Lx-%Lx)\n",
211 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
212 pr_err("NUMA: too many memblk ranges\n");
216 mi->blk[mi->nr_blks].start = start;
217 mi->blk[mi->nr_blks].end = end;
218 mi->blk[mi->nr_blks].nid = nid;
223 static void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
226 memmove(&mi->blk[idx], &mi->blk[idx + 1],
227 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
230 /* Initialize bootmem allocator for a node */
232 setup_node_bootmem(int nodeid, unsigned long start, unsigned long end)
234 unsigned long start_pfn, last_pfn, nodedata_phys;
235 const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
242 * Don't confuse VM with a node that doesn't have the
243 * minimum amount of memory:
245 if (end && (end - start) < NODE_MIN_SIZE)
248 start = roundup(start, ZONE_ALIGN);
250 printk(KERN_INFO "Initmem setup node %d %016lx-%016lx\n", nodeid,
253 start_pfn = start >> PAGE_SHIFT;
254 last_pfn = end >> PAGE_SHIFT;
256 node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
258 if (node_data[nodeid] == NULL)
260 nodedata_phys = __pa(node_data[nodeid]);
261 memblock_x86_reserve_range(nodedata_phys, nodedata_phys + pgdat_size, "NODE_DATA");
262 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
263 nodedata_phys + pgdat_size - 1);
264 nid = phys_to_nid(nodedata_phys);
266 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid);
268 memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
269 NODE_DATA(nodeid)->node_id = nodeid;
270 NODE_DATA(nodeid)->node_start_pfn = start_pfn;
271 NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn;
273 node_set_online(nodeid);
276 static int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
279 const u64 high = (u64)max_pfn << PAGE_SHIFT;
282 for (i = 0; i < mi->nr_blks; i++) {
283 struct numa_memblk *bi = &mi->blk[i];
285 /* make sure all blocks are inside the limits */
286 bi->start = max(bi->start, low);
287 bi->end = min(bi->end, high);
289 /* and there's no empty block */
290 if (bi->start == bi->end) {
291 numa_remove_memblk_from(i--, mi);
295 for (j = i + 1; j < mi->nr_blks; j++) {
296 struct numa_memblk *bj = &mi->blk[j];
297 unsigned long start, end;
300 * See whether there are overlapping blocks. Whine
301 * about but allow overlaps of the same nid. They
302 * will be merged below.
304 if (bi->end > bj->start && bi->start < bj->end) {
305 if (bi->nid != bj->nid) {
306 pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
307 bi->nid, bi->start, bi->end,
308 bj->nid, bj->start, bj->end);
311 pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
312 bi->nid, bi->start, bi->end,
317 * Join together blocks on the same node, holes
318 * between which don't overlap with memory on other
321 if (bi->nid != bj->nid)
323 start = max(min(bi->start, bj->start), low);
324 end = min(max(bi->end, bj->end), high);
325 for (k = 0; k < mi->nr_blks; k++) {
326 struct numa_memblk *bk = &mi->blk[k];
328 if (bi->nid == bk->nid)
330 if (start < bk->end && end > bk->start)
335 printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%lx,%lx)\n",
336 bi->nid, bi->start, bi->end, bj->start, bj->end,
340 numa_remove_memblk_from(j--, mi);
344 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
345 mi->blk[i].start = mi->blk[i].end = 0;
346 mi->blk[i].nid = NUMA_NO_NODE;
353 * Set nodes, which have memory in @mi, in *@nodemask.
355 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
356 const struct numa_meminfo *mi)
360 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
361 if (mi->blk[i].start != mi->blk[i].end &&
362 mi->blk[i].nid != NUMA_NO_NODE)
363 node_set(mi->blk[i].nid, *nodemask);
367 * Reset distance table. The current table is freed. The next
368 * numa_set_distance() call will create a new one.
370 static void __init numa_reset_distance(void)
374 size = numa_distance_cnt * sizeof(numa_distance[0]);
375 memblock_x86_free_range(__pa(numa_distance),
376 __pa(numa_distance) + size);
377 numa_distance = NULL;
378 numa_distance_cnt = 0;
382 * Set the distance between node @from to @to to @distance. If distance
383 * table doesn't exist, one which is large enough to accomodate all the
384 * currently known nodes will be created.
386 void __init numa_set_distance(int from, int to, int distance)
388 if (!numa_distance) {
389 nodemask_t nodes_parsed;
394 /* size the new table and allocate it */
395 nodes_parsed = numa_nodes_parsed;
396 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
398 for_each_node_mask(i, nodes_parsed)
400 size = ++cnt * sizeof(numa_distance[0]);
402 phys = memblock_find_in_range(0,
403 (u64)max_pfn_mapped << PAGE_SHIFT,
405 if (phys == MEMBLOCK_ERROR) {
406 pr_warning("NUMA: Warning: can't allocate distance table!\n");
407 /* don't retry until explicitly reset */
408 numa_distance = (void *)1LU;
411 memblock_x86_reserve_range(phys, phys + size, "NUMA DIST");
413 numa_distance = __va(phys);
414 numa_distance_cnt = cnt;
416 /* fill with the default distances */
417 for (i = 0; i < cnt; i++)
418 for (j = 0; j < cnt; j++)
419 numa_distance[i * cnt + j] = i == j ?
420 LOCAL_DISTANCE : REMOTE_DISTANCE;
421 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
424 if (from >= numa_distance_cnt || to >= numa_distance_cnt) {
425 printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n",
430 if ((u8)distance != distance ||
431 (from == to && distance != LOCAL_DISTANCE)) {
432 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
437 numa_distance[from * numa_distance_cnt + to] = distance;
440 int __node_distance(int from, int to)
442 #if defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA_EMU)
444 return acpi_emu_node_distance(from, to);
446 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
447 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
448 return numa_distance[from * numa_distance_cnt + to];
450 EXPORT_SYMBOL(__node_distance);
453 * Sanity check to catch more bad NUMA configurations (they are amazingly
454 * common). Make sure the nodes cover all memory.
456 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
458 unsigned long numaram, e820ram;
462 for (i = 0; i < mi->nr_blks; i++) {
463 unsigned long s = mi->blk[i].start >> PAGE_SHIFT;
464 unsigned long e = mi->blk[i].end >> PAGE_SHIFT;
466 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
467 if ((long)numaram < 0)
471 e820ram = max_pfn - (memblock_x86_hole_size(0,
472 max_pfn << PAGE_SHIFT) >> PAGE_SHIFT);
473 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
474 if ((long)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
475 printk(KERN_ERR "NUMA: nodes only cover %luMB of your %luMB e820 RAM. Not used.\n",
476 (numaram << PAGE_SHIFT) >> 20,
477 (e820ram << PAGE_SHIFT) >> 20);
483 static int __init numa_register_memblks(struct numa_meminfo *mi)
487 /* Account for nodes with cpus and no memory */
488 node_possible_map = numa_nodes_parsed;
489 numa_nodemask_from_meminfo(&node_possible_map, mi);
490 if (WARN_ON(nodes_empty(node_possible_map)))
493 memnode_shift = compute_hash_shift(mi);
494 if (memnode_shift < 0) {
495 printk(KERN_ERR "NUMA: No NUMA node hash function found. Contact maintainer\n");
499 for (i = 0; i < mi->nr_blks; i++)
500 memblock_x86_register_active_regions(mi->blk[i].nid,
501 mi->blk[i].start >> PAGE_SHIFT,
502 mi->blk[i].end >> PAGE_SHIFT);
504 /* for out of order entries */
506 if (!numa_meminfo_cover_memory(mi))
509 init_memory_mapping_high();
512 * Finally register nodes. Do it twice in case setup_node_bootmem
513 * missed one due to missing bootmem.
515 for (i = 0; i < 2; i++) {
516 for_each_node_mask(nid, node_possible_map) {
517 u64 start = (u64)max_pfn << PAGE_SHIFT;
520 if (node_online(nid))
523 for (j = 0; j < mi->nr_blks; j++) {
524 if (nid != mi->blk[j].nid)
526 start = min(mi->blk[j].start, start);
527 end = max(mi->blk[j].end, end);
531 setup_node_bootmem(nid, start, end);
538 #ifdef CONFIG_NUMA_EMU
540 static struct bootnode nodes[MAX_NUMNODES] __initdata;
541 static struct bootnode physnodes[MAX_NUMNODES] __cpuinitdata;
542 static char *cmdline __initdata;
544 void __init numa_emu_cmdline(char *str)
549 int __init find_node_by_addr(unsigned long addr)
551 const struct numa_meminfo *mi = &numa_meminfo;
554 for (i = 0; i < mi->nr_blks; i++) {
556 * Find the real node that this emulated node appears on. For
557 * the sake of simplicity, we only use a real node's starting
558 * address to determine which emulated node it appears on.
560 if (addr >= mi->blk[i].start && addr < mi->blk[i].end)
561 return mi->blk[i].nid;
566 static int __init setup_physnodes(unsigned long start, unsigned long end)
568 const struct numa_meminfo *mi = &numa_meminfo;
572 memset(physnodes, 0, sizeof(physnodes));
574 for (i = 0; i < mi->nr_blks; i++) {
575 int nid = mi->blk[i].nid;
577 if (physnodes[nid].start == physnodes[nid].end) {
578 physnodes[nid].start = mi->blk[i].start;
579 physnodes[nid].end = mi->blk[i].end;
581 physnodes[nid].start = min(physnodes[nid].start,
583 physnodes[nid].end = max(physnodes[nid].end,
589 * Basic sanity checking on the physical node map: there may be errors
590 * if the SRAT or AMD code incorrectly reported the topology or the mem=
591 * kernel parameter is used.
593 for (i = 0; i < MAX_NUMNODES; i++) {
594 if (physnodes[i].start == physnodes[i].end)
596 if (physnodes[i].start > end) {
597 physnodes[i].end = physnodes[i].start;
600 if (physnodes[i].end < start) {
601 physnodes[i].start = physnodes[i].end;
604 if (physnodes[i].start < start)
605 physnodes[i].start = start;
606 if (physnodes[i].end > end)
607 physnodes[i].end = end;
612 * If no physical topology was detected, a single node is faked to cover
613 * the entire address space.
616 physnodes[ret].start = start;
617 physnodes[ret].end = end;
623 static void __init fake_physnodes(int acpi, int amd, int nr_nodes)
628 #ifdef CONFIG_ACPI_NUMA
630 acpi_fake_nodes(nodes, nr_nodes);
632 #ifdef CONFIG_AMD_NUMA
634 amd_fake_nodes(nodes, nr_nodes);
637 for (i = 0; i < nr_cpu_ids; i++)
642 * Setups up nid to range from addr to addr + size. If the end
643 * boundary is greater than max_addr, then max_addr is used instead.
644 * The return value is 0 if there is additional memory left for
645 * allocation past addr and -1 otherwise. addr is adjusted to be at
646 * the end of the node.
648 static int __init setup_node_range(int nid, u64 *addr, u64 size, u64 max_addr)
651 nodes[nid].start = *addr;
653 if (*addr >= max_addr) {
657 nodes[nid].end = *addr;
658 node_set(nid, node_possible_map);
659 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
660 nodes[nid].start, nodes[nid].end,
661 (nodes[nid].end - nodes[nid].start) >> 20);
666 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
667 * to max_addr. The return value is the number of nodes allocated.
669 static int __init split_nodes_interleave(u64 addr, u64 max_addr, int nr_nodes)
671 nodemask_t physnode_mask = NODE_MASK_NONE;
679 if (nr_nodes > MAX_NUMNODES) {
680 pr_info("numa=fake=%d too large, reducing to %d\n",
681 nr_nodes, MAX_NUMNODES);
682 nr_nodes = MAX_NUMNODES;
685 size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) / nr_nodes;
687 * Calculate the number of big nodes that can be allocated as a result
688 * of consolidating the remainder.
690 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
693 size &= FAKE_NODE_MIN_HASH_MASK;
695 pr_err("Not enough memory for each node. "
696 "NUMA emulation disabled.\n");
700 for (i = 0; i < MAX_NUMNODES; i++)
701 if (physnodes[i].start != physnodes[i].end)
702 node_set(i, physnode_mask);
705 * Continue to fill physical nodes with fake nodes until there is no
706 * memory left on any of them.
708 while (nodes_weight(physnode_mask)) {
709 for_each_node_mask(i, physnode_mask) {
710 u64 end = physnodes[i].start + size;
711 u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
714 end += FAKE_NODE_MIN_SIZE;
717 * Continue to add memory to this fake node if its
718 * non-reserved memory is less than the per-node size.
720 while (end - physnodes[i].start -
721 memblock_x86_hole_size(physnodes[i].start, end) < size) {
722 end += FAKE_NODE_MIN_SIZE;
723 if (end > physnodes[i].end) {
724 end = physnodes[i].end;
730 * If there won't be at least FAKE_NODE_MIN_SIZE of
731 * non-reserved memory in ZONE_DMA32 for the next node,
732 * this one must extend to the boundary.
734 if (end < dma32_end && dma32_end - end -
735 memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
739 * If there won't be enough non-reserved memory for the
740 * next node, this one must extend to the end of the
743 if (physnodes[i].end - end -
744 memblock_x86_hole_size(end, physnodes[i].end) < size)
745 end = physnodes[i].end;
748 * Avoid allocating more nodes than requested, which can
749 * happen as a result of rounding down each node's size
750 * to FAKE_NODE_MIN_SIZE.
752 if (nodes_weight(physnode_mask) + ret >= nr_nodes)
753 end = physnodes[i].end;
755 if (setup_node_range(ret++, &physnodes[i].start,
756 end - physnodes[i].start,
757 physnodes[i].end) < 0)
758 node_clear(i, physnode_mask);
765 * Returns the end address of a node so that there is at least `size' amount of
766 * non-reserved memory or `max_addr' is reached.
768 static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
770 u64 end = start + size;
772 while (end - start - memblock_x86_hole_size(start, end) < size) {
773 end += FAKE_NODE_MIN_SIZE;
774 if (end > max_addr) {
783 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
784 * `addr' to `max_addr'. The return value is the number of nodes allocated.
786 static int __init split_nodes_size_interleave(u64 addr, u64 max_addr, u64 size)
788 nodemask_t physnode_mask = NODE_MASK_NONE;
796 * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
797 * increased accordingly if the requested size is too small. This
798 * creates a uniform distribution of node sizes across the entire
799 * machine (but not necessarily over physical nodes).
801 min_size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) /
803 min_size = max(min_size, FAKE_NODE_MIN_SIZE);
804 if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size)
805 min_size = (min_size + FAKE_NODE_MIN_SIZE) &
806 FAKE_NODE_MIN_HASH_MASK;
807 if (size < min_size) {
808 pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
809 size >> 20, min_size >> 20);
812 size &= FAKE_NODE_MIN_HASH_MASK;
814 for (i = 0; i < MAX_NUMNODES; i++)
815 if (physnodes[i].start != physnodes[i].end)
816 node_set(i, physnode_mask);
818 * Fill physical nodes with fake nodes of size until there is no memory
819 * left on any of them.
821 while (nodes_weight(physnode_mask)) {
822 for_each_node_mask(i, physnode_mask) {
823 u64 dma32_end = MAX_DMA32_PFN << PAGE_SHIFT;
826 end = find_end_of_node(physnodes[i].start,
827 physnodes[i].end, size);
829 * If there won't be at least FAKE_NODE_MIN_SIZE of
830 * non-reserved memory in ZONE_DMA32 for the next node,
831 * this one must extend to the boundary.
833 if (end < dma32_end && dma32_end - end -
834 memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
838 * If there won't be enough non-reserved memory for the
839 * next node, this one must extend to the end of the
842 if (physnodes[i].end - end -
843 memblock_x86_hole_size(end, physnodes[i].end) < size)
844 end = physnodes[i].end;
847 * Setup the fake node that will be allocated as bootmem
848 * later. If setup_node_range() returns non-zero, there
849 * is no more memory available on this physical node.
851 if (setup_node_range(ret++, &physnodes[i].start,
852 end - physnodes[i].start,
853 physnodes[i].end) < 0)
854 node_clear(i, physnode_mask);
861 * Sets up the system RAM area from start_pfn to last_pfn according to the
862 * numa=fake command-line option.
864 static int __init numa_emulation(unsigned long start_pfn,
865 unsigned long last_pfn, int acpi, int amd)
867 static struct numa_meminfo ei __initdata;
868 u64 addr = start_pfn << PAGE_SHIFT;
869 u64 max_addr = last_pfn << PAGE_SHIFT;
874 * If the numa=fake command-line contains a 'M' or 'G', it represents
875 * the fixed node size. Otherwise, if it is just a single number N,
876 * split the system RAM into N fake nodes.
878 if (strchr(cmdline, 'M') || strchr(cmdline, 'G')) {
881 size = memparse(cmdline, &cmdline);
882 num_nodes = split_nodes_size_interleave(addr, max_addr, size);
886 n = simple_strtoul(cmdline, NULL, 0);
887 num_nodes = split_nodes_interleave(addr, max_addr, n);
893 ei.nr_blks = num_nodes;
894 for (i = 0; i < ei.nr_blks; i++) {
895 ei.blk[i].start = nodes[i].start;
896 ei.blk[i].end = nodes[i].end;
900 memnode_shift = compute_hash_shift(&ei);
901 if (memnode_shift < 0) {
903 printk(KERN_ERR "No NUMA hash function found. NUMA emulation "
909 * We need to vacate all active ranges that may have been registered for
910 * the e820 memory map.
912 remove_all_active_ranges();
913 for_each_node_mask(i, node_possible_map)
914 memblock_x86_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
915 nodes[i].end >> PAGE_SHIFT);
916 init_memory_mapping_high();
917 for_each_node_mask(i, node_possible_map)
918 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
919 setup_physnodes(addr, max_addr);
920 fake_physnodes(acpi, amd, num_nodes);
922 numa_emu_dist = true;
925 #endif /* CONFIG_NUMA_EMU */
927 static int dummy_numa_init(void)
929 printk(KERN_INFO "%s\n",
930 numa_off ? "NUMA turned off" : "No NUMA configuration found");
931 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
932 0LU, max_pfn << PAGE_SHIFT);
934 node_set(0, numa_nodes_parsed);
935 numa_add_memblk(0, 0, (u64)max_pfn << PAGE_SHIFT);
940 void __init initmem_init(void)
942 int (*numa_init[])(void) = { [2] = dummy_numa_init };
946 #ifdef CONFIG_ACPI_NUMA
947 numa_init[0] = x86_acpi_numa_init;
949 #ifdef CONFIG_AMD_NUMA
950 numa_init[1] = amd_numa_init;
954 for (i = 0; i < ARRAY_SIZE(numa_init); i++) {
958 for (j = 0; j < MAX_LOCAL_APIC; j++)
959 set_apicid_to_node(j, NUMA_NO_NODE);
961 nodes_clear(numa_nodes_parsed);
962 nodes_clear(node_possible_map);
963 nodes_clear(node_online_map);
964 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
965 remove_all_active_ranges();
966 numa_reset_distance();
968 if (numa_init[i]() < 0)
971 if (numa_cleanup_meminfo(&numa_meminfo) < 0)
973 #ifdef CONFIG_NUMA_EMU
974 setup_physnodes(0, max_pfn << PAGE_SHIFT);
975 if (cmdline && !numa_emulation(0, max_pfn, i == 0, i == 1))
977 setup_physnodes(0, max_pfn << PAGE_SHIFT);
978 nodes_clear(node_possible_map);
979 nodes_clear(node_online_map);
981 if (numa_register_memblks(&numa_meminfo) < 0)
984 for (j = 0; j < nr_cpu_ids; j++) {
985 int nid = early_cpu_to_node(j);
987 if (nid == NUMA_NO_NODE)
989 if (!node_online(nid))
998 unsigned long __init numa_free_all_bootmem(void)
1000 unsigned long pages = 0;
1003 for_each_online_node(i)
1004 pages += free_all_bootmem_node(NODE_DATA(i));
1006 pages += free_all_memory_core_early(MAX_NUMNODES);
1011 int __cpuinit numa_cpu_node(int cpu)
1013 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
1015 if (apicid != BAD_APICID)
1016 return __apicid_to_node[apicid];
1017 return NUMA_NO_NODE;
1021 * UGLINESS AHEAD: Currently, CONFIG_NUMA_EMU is 64bit only and makes use
1022 * of 64bit specific data structures. The distinction is artificial and
1023 * should be removed. numa_{add|remove}_cpu() are implemented in numa.c
1024 * for both 32 and 64bit when CONFIG_NUMA_EMU is disabled but here when
1027 * NUMA emulation is planned to be made generic and the following and other
1028 * related code should be moved to numa.c.
1030 #ifdef CONFIG_NUMA_EMU
1031 # ifndef CONFIG_DEBUG_PER_CPU_MAPS
1032 void __cpuinit numa_add_cpu(int cpu)
1037 nid = numa_cpu_node(cpu);
1038 if (nid == NUMA_NO_NODE)
1039 nid = early_cpu_to_node(cpu);
1040 BUG_ON(nid == NUMA_NO_NODE || !node_online(nid));
1043 * Use the starting address of the emulated node to find which physical
1044 * node it is allocated on.
1046 addr = node_start_pfn(nid) << PAGE_SHIFT;
1047 for (physnid = 0; physnid < MAX_NUMNODES; physnid++)
1048 if (addr >= physnodes[physnid].start &&
1049 addr < physnodes[physnid].end)
1053 * Map the cpu to each emulated node that is allocated on the physical
1054 * node of the cpu's apic id.
1056 for_each_online_node(nid) {
1057 addr = node_start_pfn(nid) << PAGE_SHIFT;
1058 if (addr >= physnodes[physnid].start &&
1059 addr < physnodes[physnid].end)
1060 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
1064 void __cpuinit numa_remove_cpu(int cpu)
1068 for_each_online_node(i)
1069 cpumask_clear_cpu(cpu, node_to_cpumask_map[i]);
1071 # else /* !CONFIG_DEBUG_PER_CPU_MAPS */
1072 static void __cpuinit numa_set_cpumask(int cpu, int enable)
1074 int node = early_cpu_to_node(cpu);
1075 struct cpumask *mask;
1078 if (node == NUMA_NO_NODE) {
1079 /* early_cpu_to_node() already emits a warning and trace */
1082 for_each_online_node(i) {
1085 addr = node_start_pfn(i) << PAGE_SHIFT;
1086 if (addr < physnodes[node].start ||
1087 addr >= physnodes[node].end)
1089 mask = debug_cpumask_set_cpu(cpu, enable);
1094 cpumask_set_cpu(cpu, mask);
1096 cpumask_clear_cpu(cpu, mask);
1100 void __cpuinit numa_add_cpu(int cpu)
1102 numa_set_cpumask(cpu, 1);
1105 void __cpuinit numa_remove_cpu(int cpu)
1107 numa_set_cpumask(cpu, 0);
1109 # endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
1110 #endif /* CONFIG_NUMA_EMU */