Merge branch 'akpm'
[linux-2.6/next.git] / arch / x86 / mm / numa_emulation.c
blobd0ed086b6247d35fb986d98dfe35a1711d901876
1 /*
2 * NUMA emulation
3 */
4 #include <linux/kernel.h>
5 #include <linux/errno.h>
6 #include <linux/topology.h>
7 #include <linux/memblock.h>
8 #include <linux/bootmem.h>
9 #include <asm/dma.h>
11 #include "numa_internal.h"
13 static int emu_nid_to_phys[MAX_NUMNODES] __cpuinitdata;
14 static char *emu_cmdline __initdata;
16 void __init numa_emu_cmdline(char *str)
18 emu_cmdline = str;
21 static int __init emu_find_memblk_by_nid(int nid, const struct numa_meminfo *mi)
23 int i;
25 for (i = 0; i < mi->nr_blks; i++)
26 if (mi->blk[i].nid == nid)
27 return i;
28 return -ENOENT;
32 * Sets up nid to range from @start to @end. The return value is -errno if
33 * something went wrong, 0 otherwise.
35 static int __init emu_setup_memblk(struct numa_meminfo *ei,
36 struct numa_meminfo *pi,
37 int nid, int phys_blk, u64 size)
39 struct numa_memblk *eb = &ei->blk[ei->nr_blks];
40 struct numa_memblk *pb = &pi->blk[phys_blk];
42 if (ei->nr_blks >= NR_NODE_MEMBLKS) {
43 pr_err("NUMA: Too many emulated memblks, failing emulation\n");
44 return -EINVAL;
47 ei->nr_blks++;
48 eb->start = pb->start;
49 eb->end = pb->start + size;
50 eb->nid = nid;
52 if (emu_nid_to_phys[nid] == NUMA_NO_NODE)
53 emu_nid_to_phys[nid] = pb->nid;
55 pb->start += size;
56 if (pb->start >= pb->end) {
57 WARN_ON_ONCE(pb->start > pb->end);
58 numa_remove_memblk_from(phys_blk, pi);
61 printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
62 eb->start, eb->end, (eb->end - eb->start) >> 20);
63 return 0;
67 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
68 * to max_addr. The return value is the number of nodes allocated.
70 static int __init split_nodes_interleave(struct numa_meminfo *ei,
71 struct numa_meminfo *pi,
72 u64 addr, u64 max_addr, int nr_nodes)
74 nodemask_t physnode_mask = NODE_MASK_NONE;
75 u64 size;
76 int big;
77 int nid = 0;
78 int i, ret;
80 if (nr_nodes <= 0)
81 return -1;
82 if (nr_nodes > MAX_NUMNODES) {
83 pr_info("numa=fake=%d too large, reducing to %d\n",
84 nr_nodes, MAX_NUMNODES);
85 nr_nodes = MAX_NUMNODES;
89 * Calculate target node size. x86_32 freaks on __udivdi3() so do
90 * the division in ulong number of pages and convert back.
92 size = max_addr - addr - memblock_x86_hole_size(addr, max_addr);
93 size = PFN_PHYS((unsigned long)(size >> PAGE_SHIFT) / nr_nodes);
96 * Calculate the number of big nodes that can be allocated as a result
97 * of consolidating the remainder.
99 big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
100 FAKE_NODE_MIN_SIZE;
102 size &= FAKE_NODE_MIN_HASH_MASK;
103 if (!size) {
104 pr_err("Not enough memory for each node. "
105 "NUMA emulation disabled.\n");
106 return -1;
109 for (i = 0; i < pi->nr_blks; i++)
110 node_set(pi->blk[i].nid, physnode_mask);
113 * Continue to fill physical nodes with fake nodes until there is no
114 * memory left on any of them.
116 while (nodes_weight(physnode_mask)) {
117 for_each_node_mask(i, physnode_mask) {
118 u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
119 u64 start, limit, end;
120 int phys_blk;
122 phys_blk = emu_find_memblk_by_nid(i, pi);
123 if (phys_blk < 0) {
124 node_clear(i, physnode_mask);
125 continue;
127 start = pi->blk[phys_blk].start;
128 limit = pi->blk[phys_blk].end;
129 end = start + size;
131 if (nid < big)
132 end += FAKE_NODE_MIN_SIZE;
135 * Continue to add memory to this fake node if its
136 * non-reserved memory is less than the per-node size.
138 while (end - start -
139 memblock_x86_hole_size(start, end) < size) {
140 end += FAKE_NODE_MIN_SIZE;
141 if (end > limit) {
142 end = limit;
143 break;
148 * If there won't be at least FAKE_NODE_MIN_SIZE of
149 * non-reserved memory in ZONE_DMA32 for the next node,
150 * this one must extend to the boundary.
152 if (end < dma32_end && dma32_end - end -
153 memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
154 end = dma32_end;
157 * If there won't be enough non-reserved memory for the
158 * next node, this one must extend to the end of the
159 * physical node.
161 if (limit - end -
162 memblock_x86_hole_size(end, limit) < size)
163 end = limit;
165 ret = emu_setup_memblk(ei, pi, nid++ % nr_nodes,
166 phys_blk,
167 min(end, limit) - start);
168 if (ret < 0)
169 return ret;
172 return 0;
176 * Returns the end address of a node so that there is at least `size' amount of
177 * non-reserved memory or `max_addr' is reached.
179 static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
181 u64 end = start + size;
183 while (end - start - memblock_x86_hole_size(start, end) < size) {
184 end += FAKE_NODE_MIN_SIZE;
185 if (end > max_addr) {
186 end = max_addr;
187 break;
190 return end;
194 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
195 * `addr' to `max_addr'. The return value is the number of nodes allocated.
197 static int __init split_nodes_size_interleave(struct numa_meminfo *ei,
198 struct numa_meminfo *pi,
199 u64 addr, u64 max_addr, u64 size)
201 nodemask_t physnode_mask = NODE_MASK_NONE;
202 u64 min_size;
203 int nid = 0;
204 int i, ret;
206 if (!size)
207 return -1;
209 * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
210 * increased accordingly if the requested size is too small. This
211 * creates a uniform distribution of node sizes across the entire
212 * machine (but not necessarily over physical nodes).
214 min_size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) /
215 MAX_NUMNODES;
216 min_size = max(min_size, FAKE_NODE_MIN_SIZE);
217 if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size)
218 min_size = (min_size + FAKE_NODE_MIN_SIZE) &
219 FAKE_NODE_MIN_HASH_MASK;
220 if (size < min_size) {
221 pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
222 size >> 20, min_size >> 20);
223 size = min_size;
225 size &= FAKE_NODE_MIN_HASH_MASK;
227 for (i = 0; i < pi->nr_blks; i++)
228 node_set(pi->blk[i].nid, physnode_mask);
231 * Fill physical nodes with fake nodes of size until there is no memory
232 * left on any of them.
234 while (nodes_weight(physnode_mask)) {
235 for_each_node_mask(i, physnode_mask) {
236 u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
237 u64 start, limit, end;
238 int phys_blk;
240 phys_blk = emu_find_memblk_by_nid(i, pi);
241 if (phys_blk < 0) {
242 node_clear(i, physnode_mask);
243 continue;
245 start = pi->blk[phys_blk].start;
246 limit = pi->blk[phys_blk].end;
248 end = find_end_of_node(start, limit, size);
250 * If there won't be at least FAKE_NODE_MIN_SIZE of
251 * non-reserved memory in ZONE_DMA32 for the next node,
252 * this one must extend to the boundary.
254 if (end < dma32_end && dma32_end - end -
255 memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
256 end = dma32_end;
259 * If there won't be enough non-reserved memory for the
260 * next node, this one must extend to the end of the
261 * physical node.
263 if (limit - end -
264 memblock_x86_hole_size(end, limit) < size)
265 end = limit;
267 ret = emu_setup_memblk(ei, pi, nid++ % MAX_NUMNODES,
268 phys_blk,
269 min(end, limit) - start);
270 if (ret < 0)
271 return ret;
274 return 0;
278 * numa_emulation - Emulate NUMA nodes
279 * @numa_meminfo: NUMA configuration to massage
280 * @numa_dist_cnt: The size of the physical NUMA distance table
282 * Emulate NUMA nodes according to the numa=fake kernel parameter.
283 * @numa_meminfo contains the physical memory configuration and is modified
284 * to reflect the emulated configuration on success. @numa_dist_cnt is
285 * used to determine the size of the physical distance table.
287 * On success, the following modifications are made.
289 * - @numa_meminfo is updated to reflect the emulated nodes.
291 * - __apicid_to_node[] is updated such that APIC IDs are mapped to the
292 * emulated nodes.
294 * - NUMA distance table is rebuilt to represent distances between emulated
295 * nodes. The distances are determined considering how emulated nodes
296 * are mapped to physical nodes and match the actual distances.
298 * - emu_nid_to_phys[] reflects how emulated nodes are mapped to physical
299 * nodes. This is used by numa_add_cpu() and numa_remove_cpu().
301 * If emulation is not enabled or fails, emu_nid_to_phys[] is filled with
302 * identity mapping and no other modification is made.
304 void __init numa_emulation(struct numa_meminfo *numa_meminfo, int numa_dist_cnt)
306 static struct numa_meminfo ei __initdata;
307 static struct numa_meminfo pi __initdata;
308 const u64 max_addr = PFN_PHYS(max_pfn);
309 u8 *phys_dist = NULL;
310 size_t phys_size = numa_dist_cnt * numa_dist_cnt * sizeof(phys_dist[0]);
311 int max_emu_nid, dfl_phys_nid;
312 int i, j, ret;
314 if (!emu_cmdline)
315 goto no_emu;
317 memset(&ei, 0, sizeof(ei));
318 pi = *numa_meminfo;
320 for (i = 0; i < MAX_NUMNODES; i++)
321 emu_nid_to_phys[i] = NUMA_NO_NODE;
324 * If the numa=fake command-line contains a 'M' or 'G', it represents
325 * the fixed node size. Otherwise, if it is just a single number N,
326 * split the system RAM into N fake nodes.
328 if (strchr(emu_cmdline, 'M') || strchr(emu_cmdline, 'G')) {
329 u64 size;
331 size = memparse(emu_cmdline, &emu_cmdline);
332 ret = split_nodes_size_interleave(&ei, &pi, 0, max_addr, size);
333 } else {
334 unsigned long n;
336 n = simple_strtoul(emu_cmdline, NULL, 0);
337 ret = split_nodes_interleave(&ei, &pi, 0, max_addr, n);
340 if (ret < 0)
341 goto no_emu;
343 if (numa_cleanup_meminfo(&ei) < 0) {
344 pr_warning("NUMA: Warning: constructed meminfo invalid, disabling emulation\n");
345 goto no_emu;
348 /* copy the physical distance table */
349 if (numa_dist_cnt) {
350 u64 phys;
352 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
353 phys_size, PAGE_SIZE);
354 if (phys == MEMBLOCK_ERROR) {
355 pr_warning("NUMA: Warning: can't allocate copy of distance table, disabling emulation\n");
356 goto no_emu;
358 memblock_x86_reserve_range(phys, phys + phys_size, "TMP NUMA DIST");
359 phys_dist = __va(phys);
361 for (i = 0; i < numa_dist_cnt; i++)
362 for (j = 0; j < numa_dist_cnt; j++)
363 phys_dist[i * numa_dist_cnt + j] =
364 node_distance(i, j);
368 * Determine the max emulated nid and the default phys nid to use
369 * for unmapped nodes.
371 max_emu_nid = 0;
372 dfl_phys_nid = NUMA_NO_NODE;
373 for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++) {
374 if (emu_nid_to_phys[i] != NUMA_NO_NODE) {
375 max_emu_nid = i;
376 if (dfl_phys_nid == NUMA_NO_NODE)
377 dfl_phys_nid = emu_nid_to_phys[i];
380 if (dfl_phys_nid == NUMA_NO_NODE) {
381 pr_warning("NUMA: Warning: can't determine default physical node, disabling emulation\n");
382 goto no_emu;
385 /* commit */
386 *numa_meminfo = ei;
389 * Transform __apicid_to_node table to use emulated nids by
390 * reverse-mapping phys_nid. The maps should always exist but fall
391 * back to zero just in case.
393 for (i = 0; i < ARRAY_SIZE(__apicid_to_node); i++) {
394 if (__apicid_to_node[i] == NUMA_NO_NODE)
395 continue;
396 for (j = 0; j < ARRAY_SIZE(emu_nid_to_phys); j++)
397 if (__apicid_to_node[i] == emu_nid_to_phys[j])
398 break;
399 __apicid_to_node[i] = j < ARRAY_SIZE(emu_nid_to_phys) ? j : 0;
402 /* make sure all emulated nodes are mapped to a physical node */
403 for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
404 if (emu_nid_to_phys[i] == NUMA_NO_NODE)
405 emu_nid_to_phys[i] = dfl_phys_nid;
407 /* transform distance table */
408 numa_reset_distance();
409 for (i = 0; i < max_emu_nid + 1; i++) {
410 for (j = 0; j < max_emu_nid + 1; j++) {
411 int physi = emu_nid_to_phys[i];
412 int physj = emu_nid_to_phys[j];
413 int dist;
415 if (physi >= numa_dist_cnt || physj >= numa_dist_cnt)
416 dist = physi == physj ?
417 LOCAL_DISTANCE : REMOTE_DISTANCE;
418 else
419 dist = phys_dist[physi * numa_dist_cnt + physj];
421 numa_set_distance(i, j, dist);
425 /* free the copied physical distance table */
426 if (phys_dist)
427 memblock_x86_free_range(__pa(phys_dist), __pa(phys_dist) + phys_size);
428 return;
430 no_emu:
431 /* No emulation. Build identity emu_nid_to_phys[] for numa_add_cpu() */
432 for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
433 emu_nid_to_phys[i] = i;
436 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
437 void __cpuinit numa_add_cpu(int cpu)
439 int physnid, nid;
441 nid = early_cpu_to_node(cpu);
442 BUG_ON(nid == NUMA_NO_NODE || !node_online(nid));
444 physnid = emu_nid_to_phys[nid];
447 * Map the cpu to each emulated node that is allocated on the physical
448 * node of the cpu's apic id.
450 for_each_online_node(nid)
451 if (emu_nid_to_phys[nid] == physnid)
452 cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
455 void __cpuinit numa_remove_cpu(int cpu)
457 int i;
459 for_each_online_node(i)
460 cpumask_clear_cpu(cpu, node_to_cpumask_map[i]);
462 #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
463 static void __cpuinit numa_set_cpumask(int cpu, bool enable)
465 int nid, physnid;
467 nid = early_cpu_to_node(cpu);
468 if (nid == NUMA_NO_NODE) {
469 /* early_cpu_to_node() already emits a warning and trace */
470 return;
473 physnid = emu_nid_to_phys[nid];
475 for_each_online_node(nid) {
476 if (emu_nid_to_phys[nid] != physnid)
477 continue;
479 debug_cpumask_set_cpu(cpu, nid, enable);
483 void __cpuinit numa_add_cpu(int cpu)
485 numa_set_cpumask(cpu, true);
488 void __cpuinit numa_remove_cpu(int cpu)
490 numa_set_cpumask(cpu, false);
492 #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */