2 * ACPI 3.0 based NUMA setup
3 * Copyright 2004 Andi Kleen, SuSE Labs.
5 * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs.
7 * Called from acpi_numa_init while reading the SRAT and SLIT tables.
8 * Assumes all memory regions belonging to a single proximity domain
9 * are in one chunk. Holes between them will be included in the node.
12 #include <linux/kernel.h>
13 #include <linux/acpi.h>
14 #include <linux/mmzone.h>
15 #include <linux/bitmap.h>
16 #include <linux/module.h>
17 #include <linux/topology.h>
18 #include <linux/bootmem.h>
20 #include <asm/proto.h>
24 int acpi_numa __initdata;
26 #if (defined(CONFIG_ACPI_HOTPLUG_MEMORY) || \
27 defined(CONFIG_ACPI_HOTPLUG_MEMORY_MODULE)) \
28 && !defined(CONFIG_MEMORY_HOTPLUG)
29 #define RESERVE_HOTADD 1
32 static struct acpi_table_slit *acpi_slit;
34 static nodemask_t nodes_parsed __initdata;
35 static struct bootnode nodes[MAX_NUMNODES] __initdata;
36 static struct bootnode nodes_add[MAX_NUMNODES] __initdata;
37 static int found_add_area __initdata;
38 int hotadd_percent __initdata = 0;
39 #ifndef RESERVE_HOTADD
40 #define hotadd_percent 0 /* Ignore all settings */
43 /* Too small nodes confuse the VM badly. Usually they result
45 #define NODE_MIN_SIZE (4*1024*1024)
47 static __init int setup_node(int pxm)
49 return acpi_map_pxm_to_node(pxm);
52 static __init int conflicting_nodes(unsigned long start, unsigned long end)
55 for_each_node_mask(i, nodes_parsed) {
56 struct bootnode *nd = &nodes[i];
57 if (nd->start == nd->end)
59 if (nd->end > start && nd->start < end)
61 if (nd->end == end && nd->start == start)
67 static __init void cutoff_node(int i, unsigned long start, unsigned long end)
69 struct bootnode *nd = &nodes[i];
74 if (nd->start < start) {
76 if (nd->end < nd->start)
81 if (nd->start > nd->end)
86 static __init void bad_srat(void)
89 printk(KERN_ERR "SRAT: SRAT not used.\n");
92 for (i = 0; i < MAX_LOCAL_APIC; i++)
93 apicid_to_node[i] = NUMA_NO_NODE;
94 for (i = 0; i < MAX_NUMNODES; i++)
95 nodes_add[i].start = nodes[i].end = 0;
96 remove_all_active_ranges();
99 static __init inline int srat_disabled(void)
101 return numa_off || acpi_numa < 0;
105 * A lot of BIOS fill in 10 (= no distance) everywhere. This messes
106 * up the NUMA heuristics which wants the local node to have a smaller
107 * distance than the others.
108 * Do some quick checks here and only use the SLIT if it passes.
110 static __init int slit_valid(struct acpi_table_slit *slit)
113 int d = slit->localities;
114 for (i = 0; i < d; i++) {
115 for (j = 0; j < d; j++) {
116 u8 val = slit->entry[d*i + j];
120 } else if (val <= 10)
127 /* Callback for SLIT parsing */
128 void __init acpi_numa_slit_init(struct acpi_table_slit *slit)
130 if (!slit_valid(slit)) {
131 printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n");
137 /* Callback for Proximity Domain -> LAPIC mapping */
139 acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa)
144 if (pa->header.length != sizeof(struct acpi_table_processor_affinity)) {
148 if (pa->flags.enabled == 0)
150 pxm = pa->proximity_domain;
151 node = setup_node(pxm);
153 printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm);
157 apicid_to_node[pa->apic_id] = node;
159 printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n",
160 pxm, pa->apic_id, node);
163 #ifdef RESERVE_HOTADD
165 * Protect against too large hotadd areas that would fill up memory.
167 static int hotadd_enough_memory(struct bootnode *nd)
169 static unsigned long allocated;
170 static unsigned long last_area_end;
171 unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT;
172 long mem = pages * sizeof(struct page);
174 unsigned long allowed;
175 unsigned long oldpages = pages;
179 allowed = (end_pfn - absent_pages_in_range(0, end_pfn)) * PAGE_SIZE;
180 allowed = (allowed / 100) * hotadd_percent;
181 if (allocated + mem > allowed) {
183 /* Give them at least part of their hotadd memory upto hotadd_percent
184 It would be better to spread the limit out
185 over multiple hotplug areas, but that is too complicated
187 if (allocated >= allowed)
189 range = allowed - allocated;
190 pages = (range / PAGE_SIZE);
191 mem = pages * sizeof(struct page);
192 nd->end = nd->start + range;
194 /* Not completely fool proof, but a good sanity check */
195 addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem);
198 if (pages != oldpages)
199 printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n",
200 pages << PAGE_SHIFT);
201 last_area_end = addr + mem;
207 * It is fine to add this area to the nodes data it will be used later
208 * This code supports one contigious hot add area per node.
210 static int reserve_hotadd(int node, unsigned long start, unsigned long end)
212 unsigned long s_pfn = start >> PAGE_SHIFT;
213 unsigned long e_pfn = end >> PAGE_SHIFT;
215 struct bootnode *nd = &nodes_add[node];
217 /* I had some trouble with strange memory hotadd regions breaking
218 the boot. Be very strict here and reject anything unexpected.
219 If you want working memory hotadd write correct SRATs.
221 The node size check is a basic sanity check to guard against
223 if ((signed long)(end - start) < NODE_MIN_SIZE) {
224 printk(KERN_ERR "SRAT: Hotplug area too small\n");
228 /* This check might be a bit too strict, but I'm keeping it for now. */
229 if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) {
230 printk(KERN_ERR "SRAT: Hotplug area has existing memory\n");
234 if (!hotadd_enough_memory(&nodes_add[node])) {
235 printk(KERN_ERR "SRAT: Hotplug area too large\n");
242 if (nd->start == nd->end) {
247 if (nd->start == end) {
251 if (nd->end == start) {
256 printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n");
259 if ((nd->end >> PAGE_SHIFT) > end_pfn)
260 end_pfn = nd->end >> PAGE_SHIFT;
263 printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end);
268 /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */
270 acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma)
272 struct bootnode *nd, oldnode;
273 unsigned long start, end;
279 if (ma->header.length != sizeof(struct acpi_table_memory_affinity)) {
283 if (ma->flags.enabled == 0)
285 if (ma->flags.hot_pluggable && hotadd_percent == 0)
287 start = ma->base_addr_lo | ((u64)ma->base_addr_hi << 32);
288 end = start + (ma->length_lo | ((u64)ma->length_hi << 32));
289 pxm = ma->proximity_domain;
290 node = setup_node(pxm);
292 printk(KERN_ERR "SRAT: Too many proximity domains.\n");
296 i = conflicting_nodes(start, end);
299 "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n",
300 pxm, start, end, nodes[i].start, nodes[i].end);
303 "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n",
304 pxm, start, end, node_to_pxm(i),
305 nodes[i].start, nodes[i].end);
311 if (!node_test_and_set(node, nodes_parsed)) {
315 if (start < nd->start)
321 printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm,
323 e820_register_active_regions(node, nd->start >> PAGE_SHIFT,
324 nd->end >> PAGE_SHIFT);
326 #ifdef RESERVE_HOTADD
327 if (ma->flags.hot_pluggable && reserve_hotadd(node, start, end) < 0) {
328 /* Ignore hotadd region. Undo damage */
329 printk(KERN_NOTICE "SRAT: Hotplug region ignored\n");
331 if ((nd->start | nd->end) == 0)
332 node_clear(node, nodes_parsed);
337 /* Sanity check to catch more bad SRATs (they are amazingly common).
338 Make sure the PXMs cover all memory. */
339 static int nodes_cover_memory(void)
342 unsigned long pxmram, e820ram;
345 for_each_node_mask(i, nodes_parsed) {
346 unsigned long s = nodes[i].start >> PAGE_SHIFT;
347 unsigned long e = nodes[i].end >> PAGE_SHIFT;
349 pxmram -= absent_pages_in_range(s, e);
350 pxmram -= nodes_add[i].end - nodes_add[i].start;
351 if ((long)pxmram < 0)
355 e820ram = end_pfn - absent_pages_in_range(0, end_pfn);
356 /* We seem to lose 3 pages somewhere. Allow a bit of slack. */
357 if ((long)(e820ram - pxmram) >= 1*1024*1024) {
359 "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n",
360 (pxmram << PAGE_SHIFT) >> 20,
361 (e820ram << PAGE_SHIFT) >> 20);
367 static void unparse_node(int node)
370 node_clear(node, nodes_parsed);
371 for (i = 0; i < MAX_LOCAL_APIC; i++) {
372 if (apicid_to_node[i] == node)
373 apicid_to_node[i] = NUMA_NO_NODE;
377 void __init acpi_numa_arch_fixup(void) {}
379 /* Use the information discovered above to actually set up the nodes. */
380 int __init acpi_scan_nodes(unsigned long start, unsigned long end)
384 /* First clean up the node list */
385 for (i = 0; i < MAX_NUMNODES; i++) {
386 cutoff_node(i, start, end);
387 if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) {
396 if (!nodes_cover_memory()) {
401 memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES);
402 if (memnode_shift < 0) {
404 "SRAT: No NUMA node hash function found. Contact maintainer\n");
409 /* Finally register nodes */
410 for_each_node_mask(i, nodes_parsed)
411 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
412 /* Try again in case setup_node_bootmem missed one due
413 to missing bootmem */
414 for_each_node_mask(i, nodes_parsed)
416 setup_node_bootmem(i, nodes[i].start, nodes[i].end);
418 for (i = 0; i < NR_CPUS; i++) {
419 if (cpu_to_node[i] == NUMA_NO_NODE)
421 if (!node_isset(cpu_to_node[i], nodes_parsed))
422 numa_set_node(i, NUMA_NO_NODE);
428 void __init srat_reserve_add_area(int nodeid)
430 if (found_add_area && nodes_add[nodeid].end) {
433 printk(KERN_INFO "SRAT: Reserving hot-add memory space "
434 "for node %d at %Lx-%Lx\n",
435 nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end);
436 total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start)
438 total_mb *= sizeof(struct page);
440 printk(KERN_INFO "SRAT: This will cost you %Lu MB of "
441 "pre-allocated memory.\n", (unsigned long long)total_mb);
442 reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start,
443 nodes_add[nodeid].end - nodes_add[nodeid].start);
447 int __node_distance(int a, int b)
452 return a == b ? 10 : 20;
453 index = acpi_slit->localities * node_to_pxm(a);
454 return acpi_slit->entry[index + node_to_pxm(b)];
457 EXPORT_SYMBOL(__node_distance);