Linux Audio

Check our new training course

Loading...
v5.4
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * NUMA emulation
  4 */
  5#include <linux/kernel.h>
  6#include <linux/errno.h>
  7#include <linux/topology.h>
  8#include <linux/memblock.h>
 
  9#include <asm/dma.h>
 10
 11#include "numa_internal.h"
 12
 13static int emu_nid_to_phys[MAX_NUMNODES];
 14static char *emu_cmdline __initdata;
 15
 16void __init numa_emu_cmdline(char *str)
 17{
 18	emu_cmdline = str;
 19}
 20
 21static int __init emu_find_memblk_by_nid(int nid, const struct numa_meminfo *mi)
 22{
 23	int i;
 24
 25	for (i = 0; i < mi->nr_blks; i++)
 26		if (mi->blk[i].nid == nid)
 27			return i;
 28	return -ENOENT;
 29}
 30
 31static u64 __init mem_hole_size(u64 start, u64 end)
 32{
 33	unsigned long start_pfn = PFN_UP(start);
 34	unsigned long end_pfn = PFN_DOWN(end);
 35
 36	if (start_pfn < end_pfn)
 37		return PFN_PHYS(absent_pages_in_range(start_pfn, end_pfn));
 38	return 0;
 39}
 40
 41/*
 42 * Sets up nid to range from @start to @end.  The return value is -errno if
 43 * something went wrong, 0 otherwise.
 44 */
 45static int __init emu_setup_memblk(struct numa_meminfo *ei,
 46				   struct numa_meminfo *pi,
 47				   int nid, int phys_blk, u64 size)
 48{
 49	struct numa_memblk *eb = &ei->blk[ei->nr_blks];
 50	struct numa_memblk *pb = &pi->blk[phys_blk];
 51
 52	if (ei->nr_blks >= NR_NODE_MEMBLKS) {
 53		pr_err("NUMA: Too many emulated memblks, failing emulation\n");
 54		return -EINVAL;
 55	}
 56
 57	ei->nr_blks++;
 58	eb->start = pb->start;
 59	eb->end = pb->start + size;
 60	eb->nid = nid;
 61
 62	if (emu_nid_to_phys[nid] == NUMA_NO_NODE)
 63		emu_nid_to_phys[nid] = pb->nid;
 64
 65	pb->start += size;
 66	if (pb->start >= pb->end) {
 67		WARN_ON_ONCE(pb->start > pb->end);
 68		numa_remove_memblk_from(phys_blk, pi);
 69	}
 70
 71	printk(KERN_INFO "Faking node %d at [mem %#018Lx-%#018Lx] (%LuMB)\n",
 72	       nid, eb->start, eb->end - 1, (eb->end - eb->start) >> 20);
 73	return 0;
 74}
 75
 76/*
 77 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
 78 * to max_addr.
 79 *
 80 * Returns zero on success or negative on error.
 81 */
 82static int __init split_nodes_interleave(struct numa_meminfo *ei,
 83					 struct numa_meminfo *pi,
 84					 u64 addr, u64 max_addr, int nr_nodes)
 85{
 86	nodemask_t physnode_mask = numa_nodes_parsed;
 87	u64 size;
 88	int big;
 89	int nid = 0;
 90	int i, ret;
 91
 92	if (nr_nodes <= 0)
 93		return -1;
 94	if (nr_nodes > MAX_NUMNODES) {
 95		pr_info("numa=fake=%d too large, reducing to %d\n",
 96			nr_nodes, MAX_NUMNODES);
 97		nr_nodes = MAX_NUMNODES;
 98	}
 99
100	/*
101	 * Calculate target node size.  x86_32 freaks on __udivdi3() so do
102	 * the division in ulong number of pages and convert back.
103	 */
104	size = max_addr - addr - mem_hole_size(addr, max_addr);
105	size = PFN_PHYS((unsigned long)(size >> PAGE_SHIFT) / nr_nodes);
106
107	/*
108	 * Calculate the number of big nodes that can be allocated as a result
109	 * of consolidating the remainder.
110	 */
111	big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
112		FAKE_NODE_MIN_SIZE;
113
114	size &= FAKE_NODE_MIN_HASH_MASK;
115	if (!size) {
116		pr_err("Not enough memory for each node.  "
117			"NUMA emulation disabled.\n");
118		return -1;
119	}
120
 
 
 
121	/*
122	 * Continue to fill physical nodes with fake nodes until there is no
123	 * memory left on any of them.
124	 */
125	while (nodes_weight(physnode_mask)) {
126		for_each_node_mask(i, physnode_mask) {
127			u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
128			u64 start, limit, end;
129			int phys_blk;
130
131			phys_blk = emu_find_memblk_by_nid(i, pi);
132			if (phys_blk < 0) {
133				node_clear(i, physnode_mask);
134				continue;
135			}
136			start = pi->blk[phys_blk].start;
137			limit = pi->blk[phys_blk].end;
138			end = start + size;
139
140			if (nid < big)
141				end += FAKE_NODE_MIN_SIZE;
142
143			/*
144			 * Continue to add memory to this fake node if its
145			 * non-reserved memory is less than the per-node size.
146			 */
147			while (end - start - mem_hole_size(start, end) < size) {
148				end += FAKE_NODE_MIN_SIZE;
149				if (end > limit) {
150					end = limit;
151					break;
152				}
153			}
154
155			/*
156			 * If there won't be at least FAKE_NODE_MIN_SIZE of
157			 * non-reserved memory in ZONE_DMA32 for the next node,
158			 * this one must extend to the boundary.
159			 */
160			if (end < dma32_end && dma32_end - end -
161			    mem_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
162				end = dma32_end;
163
164			/*
165			 * If there won't be enough non-reserved memory for the
166			 * next node, this one must extend to the end of the
167			 * physical node.
168			 */
169			if (limit - end - mem_hole_size(end, limit) < size)
170				end = limit;
171
172			ret = emu_setup_memblk(ei, pi, nid++ % nr_nodes,
173					       phys_blk,
174					       min(end, limit) - start);
175			if (ret < 0)
176				return ret;
177		}
178	}
179	return 0;
180}
181
182/*
183 * Returns the end address of a node so that there is at least `size' amount of
184 * non-reserved memory or `max_addr' is reached.
185 */
186static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
187{
188	u64 end = start + size;
189
190	while (end - start - mem_hole_size(start, end) < size) {
191		end += FAKE_NODE_MIN_SIZE;
192		if (end > max_addr) {
193			end = max_addr;
194			break;
195		}
196	}
197	return end;
198}
199
200static u64 uniform_size(u64 max_addr, u64 base, u64 hole, int nr_nodes)
201{
202	unsigned long max_pfn = PHYS_PFN(max_addr);
203	unsigned long base_pfn = PHYS_PFN(base);
204	unsigned long hole_pfns = PHYS_PFN(hole);
205
206	return PFN_PHYS((max_pfn - base_pfn - hole_pfns) / nr_nodes);
207}
208
209/*
210 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
211 * `addr' to `max_addr'.
212 *
213 * Returns zero on success or negative on error.
214 */
215static int __init split_nodes_size_interleave_uniform(struct numa_meminfo *ei,
216					      struct numa_meminfo *pi,
217					      u64 addr, u64 max_addr, u64 size,
218					      int nr_nodes, struct numa_memblk *pblk,
219					      int nid)
220{
221	nodemask_t physnode_mask = numa_nodes_parsed;
222	int i, ret, uniform = 0;
223	u64 min_size;
 
 
224
225	if ((!size && !nr_nodes) || (nr_nodes && !pblk))
226		return -1;
227
228	/*
229	 * In the 'uniform' case split the passed in physical node by
230	 * nr_nodes, in the non-uniform case, ignore the passed in
231	 * physical block and try to create nodes of at least size
232	 * @size.
233	 *
234	 * In the uniform case, split the nodes strictly by physical
235	 * capacity, i.e. ignore holes. In the non-uniform case account
236	 * for holes and treat @size as a minimum floor.
237	 */
238	if (!nr_nodes)
239		nr_nodes = MAX_NUMNODES;
240	else {
241		nodes_clear(physnode_mask);
242		node_set(pblk->nid, physnode_mask);
243		uniform = 1;
244	}
245
246	if (uniform) {
247		min_size = uniform_size(max_addr, addr, 0, nr_nodes);
248		size = min_size;
249	} else {
250		/*
251		 * The limit on emulated nodes is MAX_NUMNODES, so the
252		 * size per node is increased accordingly if the
253		 * requested size is too small.  This creates a uniform
254		 * distribution of node sizes across the entire machine
255		 * (but not necessarily over physical nodes).
256		 */
257		min_size = uniform_size(max_addr, addr,
258				mem_hole_size(addr, max_addr), nr_nodes);
259	}
260	min_size = ALIGN(max(min_size, FAKE_NODE_MIN_SIZE), FAKE_NODE_MIN_SIZE);
261	if (size < min_size) {
262		pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
263			size >> 20, min_size >> 20);
264		size = min_size;
265	}
266	size = ALIGN_DOWN(size, FAKE_NODE_MIN_SIZE);
 
 
 
267
268	/*
269	 * Fill physical nodes with fake nodes of size until there is no memory
270	 * left on any of them.
271	 */
272	while (nodes_weight(physnode_mask)) {
273		for_each_node_mask(i, physnode_mask) {
274			u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
275			u64 start, limit, end;
276			int phys_blk;
277
278			phys_blk = emu_find_memblk_by_nid(i, pi);
279			if (phys_blk < 0) {
280				node_clear(i, physnode_mask);
281				continue;
282			}
283
284			start = pi->blk[phys_blk].start;
285			limit = pi->blk[phys_blk].end;
286
287			if (uniform)
288				end = start + size;
289			else
290				end = find_end_of_node(start, limit, size);
291			/*
292			 * If there won't be at least FAKE_NODE_MIN_SIZE of
293			 * non-reserved memory in ZONE_DMA32 for the next node,
294			 * this one must extend to the boundary.
295			 */
296			if (end < dma32_end && dma32_end - end -
297			    mem_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
298				end = dma32_end;
299
300			/*
301			 * If there won't be enough non-reserved memory for the
302			 * next node, this one must extend to the end of the
303			 * physical node.
304			 */
305			if ((limit - end - mem_hole_size(end, limit) < size)
306					&& !uniform)
307				end = limit;
308
309			ret = emu_setup_memblk(ei, pi, nid++ % MAX_NUMNODES,
310					       phys_blk,
311					       min(end, limit) - start);
312			if (ret < 0)
313				return ret;
314		}
315	}
316	return nid;
317}
318
319static int __init split_nodes_size_interleave(struct numa_meminfo *ei,
320					      struct numa_meminfo *pi,
321					      u64 addr, u64 max_addr, u64 size)
322{
323	return split_nodes_size_interleave_uniform(ei, pi, addr, max_addr, size,
324			0, NULL, NUMA_NO_NODE);
325}
326
327int __init setup_emu2phys_nid(int *dfl_phys_nid)
328{
329	int i, max_emu_nid = 0;
330
331	*dfl_phys_nid = NUMA_NO_NODE;
332	for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++) {
333		if (emu_nid_to_phys[i] != NUMA_NO_NODE) {
334			max_emu_nid = i;
335			if (*dfl_phys_nid == NUMA_NO_NODE)
336				*dfl_phys_nid = emu_nid_to_phys[i];
337		}
338	}
339
340	return max_emu_nid;
341}
342
343/**
344 * numa_emulation - Emulate NUMA nodes
345 * @numa_meminfo: NUMA configuration to massage
346 * @numa_dist_cnt: The size of the physical NUMA distance table
347 *
348 * Emulate NUMA nodes according to the numa=fake kernel parameter.
349 * @numa_meminfo contains the physical memory configuration and is modified
350 * to reflect the emulated configuration on success.  @numa_dist_cnt is
351 * used to determine the size of the physical distance table.
352 *
353 * On success, the following modifications are made.
354 *
355 * - @numa_meminfo is updated to reflect the emulated nodes.
356 *
357 * - __apicid_to_node[] is updated such that APIC IDs are mapped to the
358 *   emulated nodes.
359 *
360 * - NUMA distance table is rebuilt to represent distances between emulated
361 *   nodes.  The distances are determined considering how emulated nodes
362 *   are mapped to physical nodes and match the actual distances.
363 *
364 * - emu_nid_to_phys[] reflects how emulated nodes are mapped to physical
365 *   nodes.  This is used by numa_add_cpu() and numa_remove_cpu().
366 *
367 * If emulation is not enabled or fails, emu_nid_to_phys[] is filled with
368 * identity mapping and no other modification is made.
369 */
370void __init numa_emulation(struct numa_meminfo *numa_meminfo, int numa_dist_cnt)
371{
372	static struct numa_meminfo ei __initdata;
373	static struct numa_meminfo pi __initdata;
374	const u64 max_addr = PFN_PHYS(max_pfn);
375	u8 *phys_dist = NULL;
376	size_t phys_size = numa_dist_cnt * numa_dist_cnt * sizeof(phys_dist[0]);
377	int max_emu_nid, dfl_phys_nid;
378	int i, j, ret;
379
380	if (!emu_cmdline)
381		goto no_emu;
382
383	memset(&ei, 0, sizeof(ei));
384	pi = *numa_meminfo;
385
386	for (i = 0; i < MAX_NUMNODES; i++)
387		emu_nid_to_phys[i] = NUMA_NO_NODE;
388
389	/*
390	 * If the numa=fake command-line contains a 'M' or 'G', it represents
391	 * the fixed node size.  Otherwise, if it is just a single number N,
392	 * split the system RAM into N fake nodes.
393	 */
394	if (strchr(emu_cmdline, 'U')) {
395		nodemask_t physnode_mask = numa_nodes_parsed;
396		unsigned long n;
397		int nid = 0;
398
399		n = simple_strtoul(emu_cmdline, &emu_cmdline, 0);
400		ret = -1;
401		for_each_node_mask(i, physnode_mask) {
402			/*
403			 * The reason we pass in blk[0] is due to
404			 * numa_remove_memblk_from() called by
405			 * emu_setup_memblk() will delete entry 0
406			 * and then move everything else up in the pi.blk
407			 * array. Therefore we should always be looking
408			 * at blk[0].
409			 */
410			ret = split_nodes_size_interleave_uniform(&ei, &pi,
411					pi.blk[0].start, pi.blk[0].end, 0,
412					n, &pi.blk[0], nid);
413			if (ret < 0)
414				break;
415			if (ret < n) {
416				pr_info("%s: phys: %d only got %d of %ld nodes, failing\n",
417						__func__, i, ret, n);
418				ret = -1;
419				break;
420			}
421			nid = ret;
422		}
423	} else if (strchr(emu_cmdline, 'M') || strchr(emu_cmdline, 'G')) {
424		u64 size;
425
426		size = memparse(emu_cmdline, &emu_cmdline);
427		ret = split_nodes_size_interleave(&ei, &pi, 0, max_addr, size);
428	} else {
429		unsigned long n;
430
431		n = simple_strtoul(emu_cmdline, &emu_cmdline, 0);
432		ret = split_nodes_interleave(&ei, &pi, 0, max_addr, n);
433	}
434	if (*emu_cmdline == ':')
435		emu_cmdline++;
436
437	if (ret < 0)
438		goto no_emu;
439
440	if (numa_cleanup_meminfo(&ei) < 0) {
441		pr_warning("NUMA: Warning: constructed meminfo invalid, disabling emulation\n");
442		goto no_emu;
443	}
444
445	/* copy the physical distance table */
446	if (numa_dist_cnt) {
447		u64 phys;
448
449		phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
450					      phys_size, PAGE_SIZE);
451		if (!phys) {
452			pr_warning("NUMA: Warning: can't allocate copy of distance table, disabling emulation\n");
453			goto no_emu;
454		}
455		memblock_reserve(phys, phys_size);
456		phys_dist = __va(phys);
457
458		for (i = 0; i < numa_dist_cnt; i++)
459			for (j = 0; j < numa_dist_cnt; j++)
460				phys_dist[i * numa_dist_cnt + j] =
461					node_distance(i, j);
462	}
463
464	/*
465	 * Determine the max emulated nid and the default phys nid to use
466	 * for unmapped nodes.
467	 */
468	max_emu_nid = setup_emu2phys_nid(&dfl_phys_nid);
 
 
 
 
 
 
 
 
 
 
 
 
469
470	/* commit */
471	*numa_meminfo = ei;
472
473	/* Make sure numa_nodes_parsed only contains emulated nodes */
474	nodes_clear(numa_nodes_parsed);
475	for (i = 0; i < ARRAY_SIZE(ei.blk); i++)
476		if (ei.blk[i].start != ei.blk[i].end &&
477		    ei.blk[i].nid != NUMA_NO_NODE)
478			node_set(ei.blk[i].nid, numa_nodes_parsed);
479
480	/*
481	 * Transform __apicid_to_node table to use emulated nids by
482	 * reverse-mapping phys_nid.  The maps should always exist but fall
483	 * back to zero just in case.
484	 */
485	for (i = 0; i < ARRAY_SIZE(__apicid_to_node); i++) {
486		if (__apicid_to_node[i] == NUMA_NO_NODE)
487			continue;
488		for (j = 0; j < ARRAY_SIZE(emu_nid_to_phys); j++)
489			if (__apicid_to_node[i] == emu_nid_to_phys[j])
490				break;
491		__apicid_to_node[i] = j < ARRAY_SIZE(emu_nid_to_phys) ? j : 0;
492	}
493
494	/* make sure all emulated nodes are mapped to a physical node */
495	for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
496		if (emu_nid_to_phys[i] == NUMA_NO_NODE)
497			emu_nid_to_phys[i] = dfl_phys_nid;
498
499	/* transform distance table */
500	numa_reset_distance();
501	for (i = 0; i < max_emu_nid + 1; i++) {
502		for (j = 0; j < max_emu_nid + 1; j++) {
503			int physi = emu_nid_to_phys[i];
504			int physj = emu_nid_to_phys[j];
505			int dist;
506
507			if (get_option(&emu_cmdline, &dist) == 2)
508				;
509			else if (physi >= numa_dist_cnt || physj >= numa_dist_cnt)
510				dist = physi == physj ?
511					LOCAL_DISTANCE : REMOTE_DISTANCE;
512			else
513				dist = phys_dist[physi * numa_dist_cnt + physj];
514
515			numa_set_distance(i, j, dist);
516		}
517	}
518
519	/* free the copied physical distance table */
520	if (phys_dist)
521		memblock_free(__pa(phys_dist), phys_size);
522	return;
523
524no_emu:
525	/* No emulation.  Build identity emu_nid_to_phys[] for numa_add_cpu() */
526	for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
527		emu_nid_to_phys[i] = i;
528}
529
530#ifndef CONFIG_DEBUG_PER_CPU_MAPS
531void numa_add_cpu(int cpu)
532{
533	int physnid, nid;
534
535	nid = early_cpu_to_node(cpu);
536	BUG_ON(nid == NUMA_NO_NODE || !node_online(nid));
537
538	physnid = emu_nid_to_phys[nid];
539
540	/*
541	 * Map the cpu to each emulated node that is allocated on the physical
542	 * node of the cpu's apic id.
543	 */
544	for_each_online_node(nid)
545		if (emu_nid_to_phys[nid] == physnid)
546			cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
547}
548
549void numa_remove_cpu(int cpu)
550{
551	int i;
552
553	for_each_online_node(i)
554		cpumask_clear_cpu(cpu, node_to_cpumask_map[i]);
555}
556#else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
557static void numa_set_cpumask(int cpu, bool enable)
558{
559	int nid, physnid;
560
561	nid = early_cpu_to_node(cpu);
562	if (nid == NUMA_NO_NODE) {
563		/* early_cpu_to_node() already emits a warning and trace */
564		return;
565	}
566
567	physnid = emu_nid_to_phys[nid];
568
569	for_each_online_node(nid) {
570		if (emu_nid_to_phys[nid] != physnid)
571			continue;
572
573		debug_cpumask_set_cpu(cpu, nid, enable);
574	}
575}
576
577void numa_add_cpu(int cpu)
578{
579	numa_set_cpumask(cpu, true);
580}
581
582void numa_remove_cpu(int cpu)
583{
584	numa_set_cpumask(cpu, false);
585}
586#endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */
v3.15
 
  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>
 10
 11#include "numa_internal.h"
 12
 13static int emu_nid_to_phys[MAX_NUMNODES];
 14static char *emu_cmdline __initdata;
 15
 16void __init numa_emu_cmdline(char *str)
 17{
 18	emu_cmdline = str;
 19}
 20
 21static int __init emu_find_memblk_by_nid(int nid, const struct numa_meminfo *mi)
 22{
 23	int i;
 24
 25	for (i = 0; i < mi->nr_blks; i++)
 26		if (mi->blk[i].nid == nid)
 27			return i;
 28	return -ENOENT;
 29}
 30
 31static u64 __init mem_hole_size(u64 start, u64 end)
 32{
 33	unsigned long start_pfn = PFN_UP(start);
 34	unsigned long end_pfn = PFN_DOWN(end);
 35
 36	if (start_pfn < end_pfn)
 37		return PFN_PHYS(absent_pages_in_range(start_pfn, end_pfn));
 38	return 0;
 39}
 40
 41/*
 42 * Sets up nid to range from @start to @end.  The return value is -errno if
 43 * something went wrong, 0 otherwise.
 44 */
 45static int __init emu_setup_memblk(struct numa_meminfo *ei,
 46				   struct numa_meminfo *pi,
 47				   int nid, int phys_blk, u64 size)
 48{
 49	struct numa_memblk *eb = &ei->blk[ei->nr_blks];
 50	struct numa_memblk *pb = &pi->blk[phys_blk];
 51
 52	if (ei->nr_blks >= NR_NODE_MEMBLKS) {
 53		pr_err("NUMA: Too many emulated memblks, failing emulation\n");
 54		return -EINVAL;
 55	}
 56
 57	ei->nr_blks++;
 58	eb->start = pb->start;
 59	eb->end = pb->start + size;
 60	eb->nid = nid;
 61
 62	if (emu_nid_to_phys[nid] == NUMA_NO_NODE)
 63		emu_nid_to_phys[nid] = nid;
 64
 65	pb->start += size;
 66	if (pb->start >= pb->end) {
 67		WARN_ON_ONCE(pb->start > pb->end);
 68		numa_remove_memblk_from(phys_blk, pi);
 69	}
 70
 71	printk(KERN_INFO "Faking node %d at [mem %#018Lx-%#018Lx] (%LuMB)\n",
 72	       nid, eb->start, eb->end - 1, (eb->end - eb->start) >> 20);
 73	return 0;
 74}
 75
 76/*
 77 * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr
 78 * to max_addr.  The return value is the number of nodes allocated.
 
 
 79 */
 80static int __init split_nodes_interleave(struct numa_meminfo *ei,
 81					 struct numa_meminfo *pi,
 82					 u64 addr, u64 max_addr, int nr_nodes)
 83{
 84	nodemask_t physnode_mask = NODE_MASK_NONE;
 85	u64 size;
 86	int big;
 87	int nid = 0;
 88	int i, ret;
 89
 90	if (nr_nodes <= 0)
 91		return -1;
 92	if (nr_nodes > MAX_NUMNODES) {
 93		pr_info("numa=fake=%d too large, reducing to %d\n",
 94			nr_nodes, MAX_NUMNODES);
 95		nr_nodes = MAX_NUMNODES;
 96	}
 97
 98	/*
 99	 * Calculate target node size.  x86_32 freaks on __udivdi3() so do
100	 * the division in ulong number of pages and convert back.
101	 */
102	size = max_addr - addr - mem_hole_size(addr, max_addr);
103	size = PFN_PHYS((unsigned long)(size >> PAGE_SHIFT) / nr_nodes);
104
105	/*
106	 * Calculate the number of big nodes that can be allocated as a result
107	 * of consolidating the remainder.
108	 */
109	big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) /
110		FAKE_NODE_MIN_SIZE;
111
112	size &= FAKE_NODE_MIN_HASH_MASK;
113	if (!size) {
114		pr_err("Not enough memory for each node.  "
115			"NUMA emulation disabled.\n");
116		return -1;
117	}
118
119	for (i = 0; i < pi->nr_blks; i++)
120		node_set(pi->blk[i].nid, physnode_mask);
121
122	/*
123	 * Continue to fill physical nodes with fake nodes until there is no
124	 * memory left on any of them.
125	 */
126	while (nodes_weight(physnode_mask)) {
127		for_each_node_mask(i, physnode_mask) {
128			u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
129			u64 start, limit, end;
130			int phys_blk;
131
132			phys_blk = emu_find_memblk_by_nid(i, pi);
133			if (phys_blk < 0) {
134				node_clear(i, physnode_mask);
135				continue;
136			}
137			start = pi->blk[phys_blk].start;
138			limit = pi->blk[phys_blk].end;
139			end = start + size;
140
141			if (nid < big)
142				end += FAKE_NODE_MIN_SIZE;
143
144			/*
145			 * Continue to add memory to this fake node if its
146			 * non-reserved memory is less than the per-node size.
147			 */
148			while (end - start - mem_hole_size(start, end) < size) {
149				end += FAKE_NODE_MIN_SIZE;
150				if (end > limit) {
151					end = limit;
152					break;
153				}
154			}
155
156			/*
157			 * If there won't be at least FAKE_NODE_MIN_SIZE of
158			 * non-reserved memory in ZONE_DMA32 for the next node,
159			 * this one must extend to the boundary.
160			 */
161			if (end < dma32_end && dma32_end - end -
162			    mem_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
163				end = dma32_end;
164
165			/*
166			 * If there won't be enough non-reserved memory for the
167			 * next node, this one must extend to the end of the
168			 * physical node.
169			 */
170			if (limit - end - mem_hole_size(end, limit) < size)
171				end = limit;
172
173			ret = emu_setup_memblk(ei, pi, nid++ % nr_nodes,
174					       phys_blk,
175					       min(end, limit) - start);
176			if (ret < 0)
177				return ret;
178		}
179	}
180	return 0;
181}
182
183/*
184 * Returns the end address of a node so that there is at least `size' amount of
185 * non-reserved memory or `max_addr' is reached.
186 */
187static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size)
188{
189	u64 end = start + size;
190
191	while (end - start - mem_hole_size(start, end) < size) {
192		end += FAKE_NODE_MIN_SIZE;
193		if (end > max_addr) {
194			end = max_addr;
195			break;
196		}
197	}
198	return end;
199}
200
 
 
 
 
 
 
 
 
 
201/*
202 * Sets up fake nodes of `size' interleaved over physical nodes ranging from
203 * `addr' to `max_addr'.  The return value is the number of nodes allocated.
 
 
204 */
205static int __init split_nodes_size_interleave(struct numa_meminfo *ei,
206					      struct numa_meminfo *pi,
207					      u64 addr, u64 max_addr, u64 size)
 
 
208{
209	nodemask_t physnode_mask = NODE_MASK_NONE;
 
210	u64 min_size;
211	int nid = 0;
212	int i, ret;
213
214	if (!size)
215		return -1;
 
216	/*
217	 * The limit on emulated nodes is MAX_NUMNODES, so the size per node is
218	 * increased accordingly if the requested size is too small.  This
219	 * creates a uniform distribution of node sizes across the entire
220	 * machine (but not necessarily over physical nodes).
 
 
 
 
221	 */
222	min_size = (max_addr - addr - mem_hole_size(addr, max_addr)) / MAX_NUMNODES;
223	min_size = max(min_size, FAKE_NODE_MIN_SIZE);
224	if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size)
225		min_size = (min_size + FAKE_NODE_MIN_SIZE) &
226						FAKE_NODE_MIN_HASH_MASK;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
227	if (size < min_size) {
228		pr_err("Fake node size %LuMB too small, increasing to %LuMB\n",
229			size >> 20, min_size >> 20);
230		size = min_size;
231	}
232	size &= FAKE_NODE_MIN_HASH_MASK;
233
234	for (i = 0; i < pi->nr_blks; i++)
235		node_set(pi->blk[i].nid, physnode_mask);
236
237	/*
238	 * Fill physical nodes with fake nodes of size until there is no memory
239	 * left on any of them.
240	 */
241	while (nodes_weight(physnode_mask)) {
242		for_each_node_mask(i, physnode_mask) {
243			u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN);
244			u64 start, limit, end;
245			int phys_blk;
246
247			phys_blk = emu_find_memblk_by_nid(i, pi);
248			if (phys_blk < 0) {
249				node_clear(i, physnode_mask);
250				continue;
251			}
 
252			start = pi->blk[phys_blk].start;
253			limit = pi->blk[phys_blk].end;
254
255			end = find_end_of_node(start, limit, size);
 
 
 
256			/*
257			 * If there won't be at least FAKE_NODE_MIN_SIZE of
258			 * non-reserved memory in ZONE_DMA32 for the next node,
259			 * this one must extend to the boundary.
260			 */
261			if (end < dma32_end && dma32_end - end -
262			    mem_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE)
263				end = dma32_end;
264
265			/*
266			 * If there won't be enough non-reserved memory for the
267			 * next node, this one must extend to the end of the
268			 * physical node.
269			 */
270			if (limit - end - mem_hole_size(end, limit) < size)
 
271				end = limit;
272
273			ret = emu_setup_memblk(ei, pi, nid++ % MAX_NUMNODES,
274					       phys_blk,
275					       min(end, limit) - start);
276			if (ret < 0)
277				return ret;
278		}
279	}
280	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
281}
282
283/**
284 * numa_emulation - Emulate NUMA nodes
285 * @numa_meminfo: NUMA configuration to massage
286 * @numa_dist_cnt: The size of the physical NUMA distance table
287 *
288 * Emulate NUMA nodes according to the numa=fake kernel parameter.
289 * @numa_meminfo contains the physical memory configuration and is modified
290 * to reflect the emulated configuration on success.  @numa_dist_cnt is
291 * used to determine the size of the physical distance table.
292 *
293 * On success, the following modifications are made.
294 *
295 * - @numa_meminfo is updated to reflect the emulated nodes.
296 *
297 * - __apicid_to_node[] is updated such that APIC IDs are mapped to the
298 *   emulated nodes.
299 *
300 * - NUMA distance table is rebuilt to represent distances between emulated
301 *   nodes.  The distances are determined considering how emulated nodes
302 *   are mapped to physical nodes and match the actual distances.
303 *
304 * - emu_nid_to_phys[] reflects how emulated nodes are mapped to physical
305 *   nodes.  This is used by numa_add_cpu() and numa_remove_cpu().
306 *
307 * If emulation is not enabled or fails, emu_nid_to_phys[] is filled with
308 * identity mapping and no other modification is made.
309 */
310void __init numa_emulation(struct numa_meminfo *numa_meminfo, int numa_dist_cnt)
311{
312	static struct numa_meminfo ei __initdata;
313	static struct numa_meminfo pi __initdata;
314	const u64 max_addr = PFN_PHYS(max_pfn);
315	u8 *phys_dist = NULL;
316	size_t phys_size = numa_dist_cnt * numa_dist_cnt * sizeof(phys_dist[0]);
317	int max_emu_nid, dfl_phys_nid;
318	int i, j, ret;
319
320	if (!emu_cmdline)
321		goto no_emu;
322
323	memset(&ei, 0, sizeof(ei));
324	pi = *numa_meminfo;
325
326	for (i = 0; i < MAX_NUMNODES; i++)
327		emu_nid_to_phys[i] = NUMA_NO_NODE;
328
329	/*
330	 * If the numa=fake command-line contains a 'M' or 'G', it represents
331	 * the fixed node size.  Otherwise, if it is just a single number N,
332	 * split the system RAM into N fake nodes.
333	 */
334	if (strchr(emu_cmdline, 'M') || strchr(emu_cmdline, 'G')) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
335		u64 size;
336
337		size = memparse(emu_cmdline, &emu_cmdline);
338		ret = split_nodes_size_interleave(&ei, &pi, 0, max_addr, size);
339	} else {
340		unsigned long n;
341
342		n = simple_strtoul(emu_cmdline, &emu_cmdline, 0);
343		ret = split_nodes_interleave(&ei, &pi, 0, max_addr, n);
344	}
345	if (*emu_cmdline == ':')
346		emu_cmdline++;
347
348	if (ret < 0)
349		goto no_emu;
350
351	if (numa_cleanup_meminfo(&ei) < 0) {
352		pr_warning("NUMA: Warning: constructed meminfo invalid, disabling emulation\n");
353		goto no_emu;
354	}
355
356	/* copy the physical distance table */
357	if (numa_dist_cnt) {
358		u64 phys;
359
360		phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
361					      phys_size, PAGE_SIZE);
362		if (!phys) {
363			pr_warning("NUMA: Warning: can't allocate copy of distance table, disabling emulation\n");
364			goto no_emu;
365		}
366		memblock_reserve(phys, phys_size);
367		phys_dist = __va(phys);
368
369		for (i = 0; i < numa_dist_cnt; i++)
370			for (j = 0; j < numa_dist_cnt; j++)
371				phys_dist[i * numa_dist_cnt + j] =
372					node_distance(i, j);
373	}
374
375	/*
376	 * Determine the max emulated nid and the default phys nid to use
377	 * for unmapped nodes.
378	 */
379	max_emu_nid = 0;
380	dfl_phys_nid = NUMA_NO_NODE;
381	for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++) {
382		if (emu_nid_to_phys[i] != NUMA_NO_NODE) {
383			max_emu_nid = i;
384			if (dfl_phys_nid == NUMA_NO_NODE)
385				dfl_phys_nid = emu_nid_to_phys[i];
386		}
387	}
388	if (dfl_phys_nid == NUMA_NO_NODE) {
389		pr_warning("NUMA: Warning: can't determine default physical node, disabling emulation\n");
390		goto no_emu;
391	}
392
393	/* commit */
394	*numa_meminfo = ei;
 
 
 
 
 
 
 
395
396	/*
397	 * Transform __apicid_to_node table to use emulated nids by
398	 * reverse-mapping phys_nid.  The maps should always exist but fall
399	 * back to zero just in case.
400	 */
401	for (i = 0; i < ARRAY_SIZE(__apicid_to_node); i++) {
402		if (__apicid_to_node[i] == NUMA_NO_NODE)
403			continue;
404		for (j = 0; j < ARRAY_SIZE(emu_nid_to_phys); j++)
405			if (__apicid_to_node[i] == emu_nid_to_phys[j])
406				break;
407		__apicid_to_node[i] = j < ARRAY_SIZE(emu_nid_to_phys) ? j : 0;
408	}
409
410	/* make sure all emulated nodes are mapped to a physical node */
411	for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
412		if (emu_nid_to_phys[i] == NUMA_NO_NODE)
413			emu_nid_to_phys[i] = dfl_phys_nid;
414
415	/* transform distance table */
416	numa_reset_distance();
417	for (i = 0; i < max_emu_nid + 1; i++) {
418		for (j = 0; j < max_emu_nid + 1; j++) {
419			int physi = emu_nid_to_phys[i];
420			int physj = emu_nid_to_phys[j];
421			int dist;
422
423			if (get_option(&emu_cmdline, &dist) == 2)
424				;
425			else if (physi >= numa_dist_cnt || physj >= numa_dist_cnt)
426				dist = physi == physj ?
427					LOCAL_DISTANCE : REMOTE_DISTANCE;
428			else
429				dist = phys_dist[physi * numa_dist_cnt + physj];
430
431			numa_set_distance(i, j, dist);
432		}
433	}
434
435	/* free the copied physical distance table */
436	if (phys_dist)
437		memblock_free(__pa(phys_dist), phys_size);
438	return;
439
440no_emu:
441	/* No emulation.  Build identity emu_nid_to_phys[] for numa_add_cpu() */
442	for (i = 0; i < ARRAY_SIZE(emu_nid_to_phys); i++)
443		emu_nid_to_phys[i] = i;
444}
445
446#ifndef CONFIG_DEBUG_PER_CPU_MAPS
447void numa_add_cpu(int cpu)
448{
449	int physnid, nid;
450
451	nid = early_cpu_to_node(cpu);
452	BUG_ON(nid == NUMA_NO_NODE || !node_online(nid));
453
454	physnid = emu_nid_to_phys[nid];
455
456	/*
457	 * Map the cpu to each emulated node that is allocated on the physical
458	 * node of the cpu's apic id.
459	 */
460	for_each_online_node(nid)
461		if (emu_nid_to_phys[nid] == physnid)
462			cpumask_set_cpu(cpu, node_to_cpumask_map[nid]);
463}
464
465void numa_remove_cpu(int cpu)
466{
467	int i;
468
469	for_each_online_node(i)
470		cpumask_clear_cpu(cpu, node_to_cpumask_map[i]);
471}
472#else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
473static void numa_set_cpumask(int cpu, bool enable)
474{
475	int nid, physnid;
476
477	nid = early_cpu_to_node(cpu);
478	if (nid == NUMA_NO_NODE) {
479		/* early_cpu_to_node() already emits a warning and trace */
480		return;
481	}
482
483	physnid = emu_nid_to_phys[nid];
484
485	for_each_online_node(nid) {
486		if (emu_nid_to_phys[nid] != physnid)
487			continue;
488
489		debug_cpumask_set_cpu(cpu, nid, enable);
490	}
491}
492
493void numa_add_cpu(int cpu)
494{
495	numa_set_cpumask(cpu, true);
496}
497
498void numa_remove_cpu(int cpu)
499{
500	numa_set_cpumask(cpu, false);
501}
502#endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */