xref: /openbmc/linux/mm/mempolicy.c (revision 87c2ce3b)
1 /*
2  * Simple NUMA memory policy for the Linux kernel.
3  *
4  * Copyright 2003,2004 Andi Kleen, SuSE Labs.
5  * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
6  * Subject to the GNU Public License, version 2.
7  *
8  * NUMA policy allows the user to give hints in which node(s) memory should
9  * be allocated.
10  *
11  * Support four policies per VMA and per process:
12  *
13  * The VMA policy has priority over the process policy for a page fault.
14  *
15  * interleave     Allocate memory interleaved over a set of nodes,
16  *                with normal fallback if it fails.
17  *                For VMA based allocations this interleaves based on the
18  *                offset into the backing object or offset into the mapping
19  *                for anonymous memory. For process policy an process counter
20  *                is used.
21  *
22  * bind           Only allocate memory on a specific set of nodes,
23  *                no fallback.
24  *                FIXME: memory is allocated starting with the first node
25  *                to the last. It would be better if bind would truly restrict
26  *                the allocation to memory nodes instead
27  *
28  * preferred       Try a specific node first before normal fallback.
29  *                As a special case node -1 here means do the allocation
30  *                on the local CPU. This is normally identical to default,
31  *                but useful to set in a VMA when you have a non default
32  *                process policy.
33  *
34  * default        Allocate on the local node first, or when on a VMA
35  *                use the process policy. This is what Linux always did
36  *		  in a NUMA aware kernel and still does by, ahem, default.
37  *
38  * The process policy is applied for most non interrupt memory allocations
39  * in that process' context. Interrupts ignore the policies and always
40  * try to allocate on the local CPU. The VMA policy is only applied for memory
41  * allocations for a VMA in the VM.
42  *
43  * Currently there are a few corner cases in swapping where the policy
44  * is not applied, but the majority should be handled. When process policy
45  * is used it is not remembered over swap outs/swap ins.
46  *
47  * Only the highest zone in the zone hierarchy gets policied. Allocations
48  * requesting a lower zone just use default policy. This implies that
49  * on systems with highmem kernel lowmem allocation don't get policied.
50  * Same with GFP_DMA allocations.
51  *
52  * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
53  * all users and remembered even when nobody has memory mapped.
54  */
55 
56 /* Notebook:
57    fix mmap readahead to honour policy and enable policy for any page cache
58    object
59    statistics for bigpages
60    global policy for page cache? currently it uses process policy. Requires
61    first item above.
62    handle mremap for shared memory (currently ignored for the policy)
63    grows down?
64    make bind policy root only? It can trigger oom much faster and the
65    kernel is not always grateful with that.
66    could replace all the switch()es with a mempolicy_ops structure.
67 */
68 
69 #include <linux/mempolicy.h>
70 #include <linux/mm.h>
71 #include <linux/highmem.h>
72 #include <linux/hugetlb.h>
73 #include <linux/kernel.h>
74 #include <linux/sched.h>
75 #include <linux/mm.h>
76 #include <linux/nodemask.h>
77 #include <linux/cpuset.h>
78 #include <linux/gfp.h>
79 #include <linux/slab.h>
80 #include <linux/string.h>
81 #include <linux/module.h>
82 #include <linux/interrupt.h>
83 #include <linux/init.h>
84 #include <linux/compat.h>
85 #include <linux/mempolicy.h>
86 #include <linux/swap.h>
87 #include <linux/seq_file.h>
88 #include <linux/proc_fs.h>
89 
90 #include <asm/tlbflush.h>
91 #include <asm/uaccess.h>
92 
93 /* Internal flags */
94 #define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)	/* Skip checks for continuous vmas */
95 #define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)		/* Invert check for nodemask */
96 #define MPOL_MF_STATS (MPOL_MF_INTERNAL << 2)		/* Gather statistics */
97 
98 static kmem_cache_t *policy_cache;
99 static kmem_cache_t *sn_cache;
100 
101 #define PDprintk(fmt...)
102 
103 /* Highest zone. An specific allocation for a zone below that is not
104    policied. */
105 int policy_zone = ZONE_DMA;
106 
107 struct mempolicy default_policy = {
108 	.refcnt = ATOMIC_INIT(1), /* never free it */
109 	.policy = MPOL_DEFAULT,
110 };
111 
112 /* Do sanity checking on a policy */
113 static int mpol_check_policy(int mode, nodemask_t *nodes)
114 {
115 	int empty = nodes_empty(*nodes);
116 
117 	switch (mode) {
118 	case MPOL_DEFAULT:
119 		if (!empty)
120 			return -EINVAL;
121 		break;
122 	case MPOL_BIND:
123 	case MPOL_INTERLEAVE:
124 		/* Preferred will only use the first bit, but allow
125 		   more for now. */
126 		if (empty)
127 			return -EINVAL;
128 		break;
129 	}
130 	return nodes_subset(*nodes, node_online_map) ? 0 : -EINVAL;
131 }
132 /* Generate a custom zonelist for the BIND policy. */
133 static struct zonelist *bind_zonelist(nodemask_t *nodes)
134 {
135 	struct zonelist *zl;
136 	int num, max, nd;
137 
138 	max = 1 + MAX_NR_ZONES * nodes_weight(*nodes);
139 	zl = kmalloc(sizeof(void *) * max, GFP_KERNEL);
140 	if (!zl)
141 		return NULL;
142 	num = 0;
143 	for_each_node_mask(nd, *nodes)
144 		zl->zones[num++] = &NODE_DATA(nd)->node_zones[policy_zone];
145 	zl->zones[num] = NULL;
146 	return zl;
147 }
148 
149 /* Create a new policy */
150 static struct mempolicy *mpol_new(int mode, nodemask_t *nodes)
151 {
152 	struct mempolicy *policy;
153 
154 	PDprintk("setting mode %d nodes[0] %lx\n", mode, nodes_addr(*nodes)[0]);
155 	if (mode == MPOL_DEFAULT)
156 		return NULL;
157 	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
158 	if (!policy)
159 		return ERR_PTR(-ENOMEM);
160 	atomic_set(&policy->refcnt, 1);
161 	switch (mode) {
162 	case MPOL_INTERLEAVE:
163 		policy->v.nodes = *nodes;
164 		if (nodes_weight(*nodes) == 0) {
165 			kmem_cache_free(policy_cache, policy);
166 			return ERR_PTR(-EINVAL);
167 		}
168 		break;
169 	case MPOL_PREFERRED:
170 		policy->v.preferred_node = first_node(*nodes);
171 		if (policy->v.preferred_node >= MAX_NUMNODES)
172 			policy->v.preferred_node = -1;
173 		break;
174 	case MPOL_BIND:
175 		policy->v.zonelist = bind_zonelist(nodes);
176 		if (policy->v.zonelist == NULL) {
177 			kmem_cache_free(policy_cache, policy);
178 			return ERR_PTR(-ENOMEM);
179 		}
180 		break;
181 	}
182 	policy->policy = mode;
183 	policy->cpuset_mems_allowed = cpuset_mems_allowed(current);
184 	return policy;
185 }
186 
187 static void gather_stats(struct page *, void *);
188 static void migrate_page_add(struct vm_area_struct *vma,
189 	struct page *page, struct list_head *pagelist, unsigned long flags);
190 
191 /* Scan through pages checking if pages follow certain conditions. */
192 static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
193 		unsigned long addr, unsigned long end,
194 		const nodemask_t *nodes, unsigned long flags,
195 		void *private)
196 {
197 	pte_t *orig_pte;
198 	pte_t *pte;
199 	spinlock_t *ptl;
200 
201 	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
202 	do {
203 		struct page *page;
204 		unsigned int nid;
205 
206 		if (!pte_present(*pte))
207 			continue;
208 		page = vm_normal_page(vma, addr, *pte);
209 		if (!page)
210 			continue;
211 		nid = page_to_nid(page);
212 		if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
213 			continue;
214 
215 		if (flags & MPOL_MF_STATS)
216 			gather_stats(page, private);
217 		else if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
218 			spin_unlock(ptl);
219 			migrate_page_add(vma, page, private, flags);
220 			spin_lock(ptl);
221 		}
222 		else
223 			break;
224 	} while (pte++, addr += PAGE_SIZE, addr != end);
225 	pte_unmap_unlock(orig_pte, ptl);
226 	return addr != end;
227 }
228 
229 static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
230 		unsigned long addr, unsigned long end,
231 		const nodemask_t *nodes, unsigned long flags,
232 		void *private)
233 {
234 	pmd_t *pmd;
235 	unsigned long next;
236 
237 	pmd = pmd_offset(pud, addr);
238 	do {
239 		next = pmd_addr_end(addr, end);
240 		if (pmd_none_or_clear_bad(pmd))
241 			continue;
242 		if (check_pte_range(vma, pmd, addr, next, nodes,
243 				    flags, private))
244 			return -EIO;
245 	} while (pmd++, addr = next, addr != end);
246 	return 0;
247 }
248 
249 static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
250 		unsigned long addr, unsigned long end,
251 		const nodemask_t *nodes, unsigned long flags,
252 		void *private)
253 {
254 	pud_t *pud;
255 	unsigned long next;
256 
257 	pud = pud_offset(pgd, addr);
258 	do {
259 		next = pud_addr_end(addr, end);
260 		if (pud_none_or_clear_bad(pud))
261 			continue;
262 		if (check_pmd_range(vma, pud, addr, next, nodes,
263 				    flags, private))
264 			return -EIO;
265 	} while (pud++, addr = next, addr != end);
266 	return 0;
267 }
268 
269 static inline int check_pgd_range(struct vm_area_struct *vma,
270 		unsigned long addr, unsigned long end,
271 		const nodemask_t *nodes, unsigned long flags,
272 		void *private)
273 {
274 	pgd_t *pgd;
275 	unsigned long next;
276 
277 	pgd = pgd_offset(vma->vm_mm, addr);
278 	do {
279 		next = pgd_addr_end(addr, end);
280 		if (pgd_none_or_clear_bad(pgd))
281 			continue;
282 		if (check_pud_range(vma, pgd, addr, next, nodes,
283 				    flags, private))
284 			return -EIO;
285 	} while (pgd++, addr = next, addr != end);
286 	return 0;
287 }
288 
289 /* Check if a vma is migratable */
290 static inline int vma_migratable(struct vm_area_struct *vma)
291 {
292 	if (vma->vm_flags & (
293 		VM_LOCKED|VM_IO|VM_HUGETLB|VM_PFNMAP))
294 		return 0;
295 	return 1;
296 }
297 
298 /*
299  * Check if all pages in a range are on a set of nodes.
300  * If pagelist != NULL then isolate pages from the LRU and
301  * put them on the pagelist.
302  */
303 static struct vm_area_struct *
304 check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
305 		const nodemask_t *nodes, unsigned long flags, void *private)
306 {
307 	int err;
308 	struct vm_area_struct *first, *vma, *prev;
309 
310 	first = find_vma(mm, start);
311 	if (!first)
312 		return ERR_PTR(-EFAULT);
313 	prev = NULL;
314 	for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
315 		if (!(flags & MPOL_MF_DISCONTIG_OK)) {
316 			if (!vma->vm_next && vma->vm_end < end)
317 				return ERR_PTR(-EFAULT);
318 			if (prev && prev->vm_end < vma->vm_start)
319 				return ERR_PTR(-EFAULT);
320 		}
321 		if (!is_vm_hugetlb_page(vma) &&
322 		    ((flags & MPOL_MF_STRICT) ||
323 		     ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
324 				vma_migratable(vma)))) {
325 			unsigned long endvma = vma->vm_end;
326 
327 			if (endvma > end)
328 				endvma = end;
329 			if (vma->vm_start > start)
330 				start = vma->vm_start;
331 			err = check_pgd_range(vma, start, endvma, nodes,
332 						flags, private);
333 			if (err) {
334 				first = ERR_PTR(err);
335 				break;
336 			}
337 		}
338 		prev = vma;
339 	}
340 	return first;
341 }
342 
343 /* Apply policy to a single VMA */
344 static int policy_vma(struct vm_area_struct *vma, struct mempolicy *new)
345 {
346 	int err = 0;
347 	struct mempolicy *old = vma->vm_policy;
348 
349 	PDprintk("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
350 		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
351 		 vma->vm_ops, vma->vm_file,
352 		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
353 
354 	if (vma->vm_ops && vma->vm_ops->set_policy)
355 		err = vma->vm_ops->set_policy(vma, new);
356 	if (!err) {
357 		mpol_get(new);
358 		vma->vm_policy = new;
359 		mpol_free(old);
360 	}
361 	return err;
362 }
363 
364 /* Step 2: apply policy to a range and do splits. */
365 static int mbind_range(struct vm_area_struct *vma, unsigned long start,
366 		       unsigned long end, struct mempolicy *new)
367 {
368 	struct vm_area_struct *next;
369 	int err;
370 
371 	err = 0;
372 	for (; vma && vma->vm_start < end; vma = next) {
373 		next = vma->vm_next;
374 		if (vma->vm_start < start)
375 			err = split_vma(vma->vm_mm, vma, start, 1);
376 		if (!err && vma->vm_end > end)
377 			err = split_vma(vma->vm_mm, vma, end, 0);
378 		if (!err)
379 			err = policy_vma(vma, new);
380 		if (err)
381 			break;
382 	}
383 	return err;
384 }
385 
386 static int contextualize_policy(int mode, nodemask_t *nodes)
387 {
388 	if (!nodes)
389 		return 0;
390 
391 	cpuset_update_task_memory_state();
392 	if (!cpuset_nodes_subset_current_mems_allowed(*nodes))
393 		return -EINVAL;
394 	return mpol_check_policy(mode, nodes);
395 }
396 
397 /* Set the process memory policy */
398 long do_set_mempolicy(int mode, nodemask_t *nodes)
399 {
400 	struct mempolicy *new;
401 
402 	if (contextualize_policy(mode, nodes))
403 		return -EINVAL;
404 	new = mpol_new(mode, nodes);
405 	if (IS_ERR(new))
406 		return PTR_ERR(new);
407 	mpol_free(current->mempolicy);
408 	current->mempolicy = new;
409 	if (new && new->policy == MPOL_INTERLEAVE)
410 		current->il_next = first_node(new->v.nodes);
411 	return 0;
412 }
413 
414 /* Fill a zone bitmap for a policy */
415 static void get_zonemask(struct mempolicy *p, nodemask_t *nodes)
416 {
417 	int i;
418 
419 	nodes_clear(*nodes);
420 	switch (p->policy) {
421 	case MPOL_BIND:
422 		for (i = 0; p->v.zonelist->zones[i]; i++)
423 			node_set(p->v.zonelist->zones[i]->zone_pgdat->node_id,
424 				*nodes);
425 		break;
426 	case MPOL_DEFAULT:
427 		break;
428 	case MPOL_INTERLEAVE:
429 		*nodes = p->v.nodes;
430 		break;
431 	case MPOL_PREFERRED:
432 		/* or use current node instead of online map? */
433 		if (p->v.preferred_node < 0)
434 			*nodes = node_online_map;
435 		else
436 			node_set(p->v.preferred_node, *nodes);
437 		break;
438 	default:
439 		BUG();
440 	}
441 }
442 
443 static int lookup_node(struct mm_struct *mm, unsigned long addr)
444 {
445 	struct page *p;
446 	int err;
447 
448 	err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
449 	if (err >= 0) {
450 		err = page_to_nid(p);
451 		put_page(p);
452 	}
453 	return err;
454 }
455 
456 /* Retrieve NUMA policy */
457 long do_get_mempolicy(int *policy, nodemask_t *nmask,
458 			unsigned long addr, unsigned long flags)
459 {
460 	int err;
461 	struct mm_struct *mm = current->mm;
462 	struct vm_area_struct *vma = NULL;
463 	struct mempolicy *pol = current->mempolicy;
464 
465 	cpuset_update_task_memory_state();
466 	if (flags & ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR))
467 		return -EINVAL;
468 	if (flags & MPOL_F_ADDR) {
469 		down_read(&mm->mmap_sem);
470 		vma = find_vma_intersection(mm, addr, addr+1);
471 		if (!vma) {
472 			up_read(&mm->mmap_sem);
473 			return -EFAULT;
474 		}
475 		if (vma->vm_ops && vma->vm_ops->get_policy)
476 			pol = vma->vm_ops->get_policy(vma, addr);
477 		else
478 			pol = vma->vm_policy;
479 	} else if (addr)
480 		return -EINVAL;
481 
482 	if (!pol)
483 		pol = &default_policy;
484 
485 	if (flags & MPOL_F_NODE) {
486 		if (flags & MPOL_F_ADDR) {
487 			err = lookup_node(mm, addr);
488 			if (err < 0)
489 				goto out;
490 			*policy = err;
491 		} else if (pol == current->mempolicy &&
492 				pol->policy == MPOL_INTERLEAVE) {
493 			*policy = current->il_next;
494 		} else {
495 			err = -EINVAL;
496 			goto out;
497 		}
498 	} else
499 		*policy = pol->policy;
500 
501 	if (vma) {
502 		up_read(&current->mm->mmap_sem);
503 		vma = NULL;
504 	}
505 
506 	err = 0;
507 	if (nmask)
508 		get_zonemask(pol, nmask);
509 
510  out:
511 	if (vma)
512 		up_read(&current->mm->mmap_sem);
513 	return err;
514 }
515 
516 /*
517  * page migration
518  */
519 
520 /* Check if we are the only process mapping the page in question */
521 static inline int single_mm_mapping(struct mm_struct *mm,
522 			struct address_space *mapping)
523 {
524 	struct vm_area_struct *vma;
525 	struct prio_tree_iter iter;
526 	int rc = 1;
527 
528 	spin_lock(&mapping->i_mmap_lock);
529 	vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
530 		if (mm != vma->vm_mm) {
531 			rc = 0;
532 			goto out;
533 		}
534 	list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
535 		if (mm != vma->vm_mm) {
536 			rc = 0;
537 			goto out;
538 		}
539 out:
540 	spin_unlock(&mapping->i_mmap_lock);
541 	return rc;
542 }
543 
544 /*
545  * Add a page to be migrated to the pagelist
546  */
547 static void migrate_page_add(struct vm_area_struct *vma,
548 	struct page *page, struct list_head *pagelist, unsigned long flags)
549 {
550 	/*
551 	 * Avoid migrating a page that is shared by others and not writable.
552 	 */
553 	if ((flags & MPOL_MF_MOVE_ALL) || !page->mapping || PageAnon(page) ||
554 	    mapping_writably_mapped(page->mapping) ||
555 	    single_mm_mapping(vma->vm_mm, page->mapping)) {
556 		int rc = isolate_lru_page(page);
557 
558 		if (rc == 1)
559 			list_add(&page->lru, pagelist);
560 		/*
561 		 * If the isolate attempt was not successful then we just
562 		 * encountered an unswappable page. Something must be wrong.
563 	 	 */
564 		WARN_ON(rc == 0);
565 	}
566 }
567 
568 static int swap_pages(struct list_head *pagelist)
569 {
570 	LIST_HEAD(moved);
571 	LIST_HEAD(failed);
572 	int n;
573 
574 	n = migrate_pages(pagelist, NULL, &moved, &failed);
575 	putback_lru_pages(&failed);
576 	putback_lru_pages(&moved);
577 
578 	return n;
579 }
580 
581 /*
582  * For now migrate_pages simply swaps out the pages from nodes that are in
583  * the source set but not in the target set. In the future, we would
584  * want a function that moves pages between the two nodesets in such
585  * a way as to preserve the physical layout as much as possible.
586  *
587  * Returns the number of page that could not be moved.
588  */
589 int do_migrate_pages(struct mm_struct *mm,
590 	const nodemask_t *from_nodes, const nodemask_t *to_nodes, int flags)
591 {
592 	LIST_HEAD(pagelist);
593 	int count = 0;
594 	nodemask_t nodes;
595 
596 	nodes_andnot(nodes, *from_nodes, *to_nodes);
597 
598 	down_read(&mm->mmap_sem);
599 	check_range(mm, mm->mmap->vm_start, TASK_SIZE, &nodes,
600 			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
601 
602 	if (!list_empty(&pagelist)) {
603 		count = swap_pages(&pagelist);
604 		putback_lru_pages(&pagelist);
605 	}
606 
607 	up_read(&mm->mmap_sem);
608 	return count;
609 }
610 
611 long do_mbind(unsigned long start, unsigned long len,
612 		unsigned long mode, nodemask_t *nmask, unsigned long flags)
613 {
614 	struct vm_area_struct *vma;
615 	struct mm_struct *mm = current->mm;
616 	struct mempolicy *new;
617 	unsigned long end;
618 	int err;
619 	LIST_HEAD(pagelist);
620 
621 	if ((flags & ~(unsigned long)(MPOL_MF_STRICT |
622 				      MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
623 	    || mode > MPOL_MAX)
624 		return -EINVAL;
625 	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_RESOURCE))
626 		return -EPERM;
627 
628 	if (start & ~PAGE_MASK)
629 		return -EINVAL;
630 
631 	if (mode == MPOL_DEFAULT)
632 		flags &= ~MPOL_MF_STRICT;
633 
634 	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
635 	end = start + len;
636 
637 	if (end < start)
638 		return -EINVAL;
639 	if (end == start)
640 		return 0;
641 
642 	if (mpol_check_policy(mode, nmask))
643 		return -EINVAL;
644 
645 	new = mpol_new(mode, nmask);
646 	if (IS_ERR(new))
647 		return PTR_ERR(new);
648 
649 	/*
650 	 * If we are using the default policy then operation
651 	 * on discontinuous address spaces is okay after all
652 	 */
653 	if (!new)
654 		flags |= MPOL_MF_DISCONTIG_OK;
655 
656 	PDprintk("mbind %lx-%lx mode:%ld nodes:%lx\n",start,start+len,
657 			mode,nodes_addr(nodes)[0]);
658 
659 	down_write(&mm->mmap_sem);
660 	vma = check_range(mm, start, end, nmask,
661 			  flags | MPOL_MF_INVERT, &pagelist);
662 
663 	err = PTR_ERR(vma);
664 	if (!IS_ERR(vma)) {
665 		int nr_failed = 0;
666 
667 		err = mbind_range(vma, start, end, new);
668 		if (!list_empty(&pagelist))
669 			nr_failed = swap_pages(&pagelist);
670 
671 		if (!err && nr_failed && (flags & MPOL_MF_STRICT))
672 			err = -EIO;
673 	}
674 	if (!list_empty(&pagelist))
675 		putback_lru_pages(&pagelist);
676 
677 	up_write(&mm->mmap_sem);
678 	mpol_free(new);
679 	return err;
680 }
681 
682 /*
683  * User space interface with variable sized bitmaps for nodelists.
684  */
685 
686 /* Copy a node mask from user space. */
687 static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
688 		     unsigned long maxnode)
689 {
690 	unsigned long k;
691 	unsigned long nlongs;
692 	unsigned long endmask;
693 
694 	--maxnode;
695 	nodes_clear(*nodes);
696 	if (maxnode == 0 || !nmask)
697 		return 0;
698 
699 	nlongs = BITS_TO_LONGS(maxnode);
700 	if ((maxnode % BITS_PER_LONG) == 0)
701 		endmask = ~0UL;
702 	else
703 		endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
704 
705 	/* When the user specified more nodes than supported just check
706 	   if the non supported part is all zero. */
707 	if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
708 		if (nlongs > PAGE_SIZE/sizeof(long))
709 			return -EINVAL;
710 		for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
711 			unsigned long t;
712 			if (get_user(t, nmask + k))
713 				return -EFAULT;
714 			if (k == nlongs - 1) {
715 				if (t & endmask)
716 					return -EINVAL;
717 			} else if (t)
718 				return -EINVAL;
719 		}
720 		nlongs = BITS_TO_LONGS(MAX_NUMNODES);
721 		endmask = ~0UL;
722 	}
723 
724 	if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
725 		return -EFAULT;
726 	nodes_addr(*nodes)[nlongs-1] &= endmask;
727 	return 0;
728 }
729 
730 /* Copy a kernel node mask to user space */
731 static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
732 			      nodemask_t *nodes)
733 {
734 	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
735 	const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
736 
737 	if (copy > nbytes) {
738 		if (copy > PAGE_SIZE)
739 			return -EINVAL;
740 		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
741 			return -EFAULT;
742 		copy = nbytes;
743 	}
744 	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
745 }
746 
747 asmlinkage long sys_mbind(unsigned long start, unsigned long len,
748 			unsigned long mode,
749 			unsigned long __user *nmask, unsigned long maxnode,
750 			unsigned flags)
751 {
752 	nodemask_t nodes;
753 	int err;
754 
755 	err = get_nodes(&nodes, nmask, maxnode);
756 	if (err)
757 		return err;
758 	return do_mbind(start, len, mode, &nodes, flags);
759 }
760 
761 /* Set the process memory policy */
762 asmlinkage long sys_set_mempolicy(int mode, unsigned long __user *nmask,
763 		unsigned long maxnode)
764 {
765 	int err;
766 	nodemask_t nodes;
767 
768 	if (mode < 0 || mode > MPOL_MAX)
769 		return -EINVAL;
770 	err = get_nodes(&nodes, nmask, maxnode);
771 	if (err)
772 		return err;
773 	return do_set_mempolicy(mode, &nodes);
774 }
775 
776 asmlinkage long sys_migrate_pages(pid_t pid, unsigned long maxnode,
777 		const unsigned long __user *old_nodes,
778 		const unsigned long __user *new_nodes)
779 {
780 	struct mm_struct *mm;
781 	struct task_struct *task;
782 	nodemask_t old;
783 	nodemask_t new;
784 	nodemask_t task_nodes;
785 	int err;
786 
787 	err = get_nodes(&old, old_nodes, maxnode);
788 	if (err)
789 		return err;
790 
791 	err = get_nodes(&new, new_nodes, maxnode);
792 	if (err)
793 		return err;
794 
795 	/* Find the mm_struct */
796 	read_lock(&tasklist_lock);
797 	task = pid ? find_task_by_pid(pid) : current;
798 	if (!task) {
799 		read_unlock(&tasklist_lock);
800 		return -ESRCH;
801 	}
802 	mm = get_task_mm(task);
803 	read_unlock(&tasklist_lock);
804 
805 	if (!mm)
806 		return -EINVAL;
807 
808 	/*
809 	 * Check if this process has the right to modify the specified
810 	 * process. The right exists if the process has administrative
811 	 * capabilities, superuser priviledges or the same
812 	 * userid as the target process.
813 	 */
814 	if ((current->euid != task->suid) && (current->euid != task->uid) &&
815 	    (current->uid != task->suid) && (current->uid != task->uid) &&
816 	    !capable(CAP_SYS_ADMIN)) {
817 		err = -EPERM;
818 		goto out;
819 	}
820 
821 	task_nodes = cpuset_mems_allowed(task);
822 	/* Is the user allowed to access the target nodes? */
823 	if (!nodes_subset(new, task_nodes) && !capable(CAP_SYS_ADMIN)) {
824 		err = -EPERM;
825 		goto out;
826 	}
827 
828 	err = do_migrate_pages(mm, &old, &new, MPOL_MF_MOVE);
829 out:
830 	mmput(mm);
831 	return err;
832 }
833 
834 
835 /* Retrieve NUMA policy */
836 asmlinkage long sys_get_mempolicy(int __user *policy,
837 				unsigned long __user *nmask,
838 				unsigned long maxnode,
839 				unsigned long addr, unsigned long flags)
840 {
841 	int err, pval;
842 	nodemask_t nodes;
843 
844 	if (nmask != NULL && maxnode < MAX_NUMNODES)
845 		return -EINVAL;
846 
847 	err = do_get_mempolicy(&pval, &nodes, addr, flags);
848 
849 	if (err)
850 		return err;
851 
852 	if (policy && put_user(pval, policy))
853 		return -EFAULT;
854 
855 	if (nmask)
856 		err = copy_nodes_to_user(nmask, maxnode, &nodes);
857 
858 	return err;
859 }
860 
861 #ifdef CONFIG_COMPAT
862 
863 asmlinkage long compat_sys_get_mempolicy(int __user *policy,
864 				     compat_ulong_t __user *nmask,
865 				     compat_ulong_t maxnode,
866 				     compat_ulong_t addr, compat_ulong_t flags)
867 {
868 	long err;
869 	unsigned long __user *nm = NULL;
870 	unsigned long nr_bits, alloc_size;
871 	DECLARE_BITMAP(bm, MAX_NUMNODES);
872 
873 	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
874 	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
875 
876 	if (nmask)
877 		nm = compat_alloc_user_space(alloc_size);
878 
879 	err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
880 
881 	if (!err && nmask) {
882 		err = copy_from_user(bm, nm, alloc_size);
883 		/* ensure entire bitmap is zeroed */
884 		err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
885 		err |= compat_put_bitmap(nmask, bm, nr_bits);
886 	}
887 
888 	return err;
889 }
890 
891 asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
892 				     compat_ulong_t maxnode)
893 {
894 	long err = 0;
895 	unsigned long __user *nm = NULL;
896 	unsigned long nr_bits, alloc_size;
897 	DECLARE_BITMAP(bm, MAX_NUMNODES);
898 
899 	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
900 	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
901 
902 	if (nmask) {
903 		err = compat_get_bitmap(bm, nmask, nr_bits);
904 		nm = compat_alloc_user_space(alloc_size);
905 		err |= copy_to_user(nm, bm, alloc_size);
906 	}
907 
908 	if (err)
909 		return -EFAULT;
910 
911 	return sys_set_mempolicy(mode, nm, nr_bits+1);
912 }
913 
914 asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
915 			     compat_ulong_t mode, compat_ulong_t __user *nmask,
916 			     compat_ulong_t maxnode, compat_ulong_t flags)
917 {
918 	long err = 0;
919 	unsigned long __user *nm = NULL;
920 	unsigned long nr_bits, alloc_size;
921 	nodemask_t bm;
922 
923 	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
924 	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
925 
926 	if (nmask) {
927 		err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
928 		nm = compat_alloc_user_space(alloc_size);
929 		err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
930 	}
931 
932 	if (err)
933 		return -EFAULT;
934 
935 	return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
936 }
937 
938 #endif
939 
940 /* Return effective policy for a VMA */
941 static struct mempolicy * get_vma_policy(struct task_struct *task,
942 		struct vm_area_struct *vma, unsigned long addr)
943 {
944 	struct mempolicy *pol = task->mempolicy;
945 
946 	if (vma) {
947 		if (vma->vm_ops && vma->vm_ops->get_policy)
948 			pol = vma->vm_ops->get_policy(vma, addr);
949 		else if (vma->vm_policy &&
950 				vma->vm_policy->policy != MPOL_DEFAULT)
951 			pol = vma->vm_policy;
952 	}
953 	if (!pol)
954 		pol = &default_policy;
955 	return pol;
956 }
957 
958 /* Return a zonelist representing a mempolicy */
959 static struct zonelist *zonelist_policy(gfp_t gfp, struct mempolicy *policy)
960 {
961 	int nd;
962 
963 	switch (policy->policy) {
964 	case MPOL_PREFERRED:
965 		nd = policy->v.preferred_node;
966 		if (nd < 0)
967 			nd = numa_node_id();
968 		break;
969 	case MPOL_BIND:
970 		/* Lower zones don't get a policy applied */
971 		/* Careful: current->mems_allowed might have moved */
972 		if (gfp_zone(gfp) >= policy_zone)
973 			if (cpuset_zonelist_valid_mems_allowed(policy->v.zonelist))
974 				return policy->v.zonelist;
975 		/*FALL THROUGH*/
976 	case MPOL_INTERLEAVE: /* should not happen */
977 	case MPOL_DEFAULT:
978 		nd = numa_node_id();
979 		break;
980 	default:
981 		nd = 0;
982 		BUG();
983 	}
984 	return NODE_DATA(nd)->node_zonelists + gfp_zone(gfp);
985 }
986 
987 /* Do dynamic interleaving for a process */
988 static unsigned interleave_nodes(struct mempolicy *policy)
989 {
990 	unsigned nid, next;
991 	struct task_struct *me = current;
992 
993 	nid = me->il_next;
994 	next = next_node(nid, policy->v.nodes);
995 	if (next >= MAX_NUMNODES)
996 		next = first_node(policy->v.nodes);
997 	me->il_next = next;
998 	return nid;
999 }
1000 
1001 /* Do static interleaving for a VMA with known offset. */
1002 static unsigned offset_il_node(struct mempolicy *pol,
1003 		struct vm_area_struct *vma, unsigned long off)
1004 {
1005 	unsigned nnodes = nodes_weight(pol->v.nodes);
1006 	unsigned target = (unsigned)off % nnodes;
1007 	int c;
1008 	int nid = -1;
1009 
1010 	c = 0;
1011 	do {
1012 		nid = next_node(nid, pol->v.nodes);
1013 		c++;
1014 	} while (c <= target);
1015 	return nid;
1016 }
1017 
1018 /* Determine a node number for interleave */
1019 static inline unsigned interleave_nid(struct mempolicy *pol,
1020 		 struct vm_area_struct *vma, unsigned long addr, int shift)
1021 {
1022 	if (vma) {
1023 		unsigned long off;
1024 
1025 		off = vma->vm_pgoff;
1026 		off += (addr - vma->vm_start) >> shift;
1027 		return offset_il_node(pol, vma, off);
1028 	} else
1029 		return interleave_nodes(pol);
1030 }
1031 
1032 /* Return a zonelist suitable for a huge page allocation. */
1033 struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr)
1034 {
1035 	struct mempolicy *pol = get_vma_policy(current, vma, addr);
1036 
1037 	if (pol->policy == MPOL_INTERLEAVE) {
1038 		unsigned nid;
1039 
1040 		nid = interleave_nid(pol, vma, addr, HPAGE_SHIFT);
1041 		return NODE_DATA(nid)->node_zonelists + gfp_zone(GFP_HIGHUSER);
1042 	}
1043 	return zonelist_policy(GFP_HIGHUSER, pol);
1044 }
1045 
1046 /* Allocate a page in interleaved policy.
1047    Own path because it needs to do special accounting. */
1048 static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1049 					unsigned nid)
1050 {
1051 	struct zonelist *zl;
1052 	struct page *page;
1053 
1054 	zl = NODE_DATA(nid)->node_zonelists + gfp_zone(gfp);
1055 	page = __alloc_pages(gfp, order, zl);
1056 	if (page && page_zone(page) == zl->zones[0]) {
1057 		zone_pcp(zl->zones[0],get_cpu())->interleave_hit++;
1058 		put_cpu();
1059 	}
1060 	return page;
1061 }
1062 
1063 /**
1064  * 	alloc_page_vma	- Allocate a page for a VMA.
1065  *
1066  * 	@gfp:
1067  *      %GFP_USER    user allocation.
1068  *      %GFP_KERNEL  kernel allocations,
1069  *      %GFP_HIGHMEM highmem/user allocations,
1070  *      %GFP_FS      allocation should not call back into a file system.
1071  *      %GFP_ATOMIC  don't sleep.
1072  *
1073  * 	@vma:  Pointer to VMA or NULL if not available.
1074  *	@addr: Virtual Address of the allocation. Must be inside the VMA.
1075  *
1076  * 	This function allocates a page from the kernel page pool and applies
1077  *	a NUMA policy associated with the VMA or the current process.
1078  *	When VMA is not NULL caller must hold down_read on the mmap_sem of the
1079  *	mm_struct of the VMA to prevent it from going away. Should be used for
1080  *	all allocations for pages that will be mapped into
1081  * 	user space. Returns NULL when no page can be allocated.
1082  *
1083  *	Should be called with the mm_sem of the vma hold.
1084  */
1085 struct page *
1086 alloc_page_vma(gfp_t gfp, struct vm_area_struct *vma, unsigned long addr)
1087 {
1088 	struct mempolicy *pol = get_vma_policy(current, vma, addr);
1089 
1090 	cpuset_update_task_memory_state();
1091 
1092 	if (unlikely(pol->policy == MPOL_INTERLEAVE)) {
1093 		unsigned nid;
1094 
1095 		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT);
1096 		return alloc_page_interleave(gfp, 0, nid);
1097 	}
1098 	return __alloc_pages(gfp, 0, zonelist_policy(gfp, pol));
1099 }
1100 
1101 /**
1102  * 	alloc_pages_current - Allocate pages.
1103  *
1104  *	@gfp:
1105  *		%GFP_USER   user allocation,
1106  *      	%GFP_KERNEL kernel allocation,
1107  *      	%GFP_HIGHMEM highmem allocation,
1108  *      	%GFP_FS     don't call back into a file system.
1109  *      	%GFP_ATOMIC don't sleep.
1110  *	@order: Power of two of allocation size in pages. 0 is a single page.
1111  *
1112  *	Allocate a page from the kernel page pool.  When not in
1113  *	interrupt context and apply the current process NUMA policy.
1114  *	Returns NULL when no page can be allocated.
1115  *
1116  *	Don't call cpuset_update_task_memory_state() unless
1117  *	1) it's ok to take cpuset_sem (can WAIT), and
1118  *	2) allocating for current task (not interrupt).
1119  */
1120 struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1121 {
1122 	struct mempolicy *pol = current->mempolicy;
1123 
1124 	if ((gfp & __GFP_WAIT) && !in_interrupt())
1125 		cpuset_update_task_memory_state();
1126 	if (!pol || in_interrupt())
1127 		pol = &default_policy;
1128 	if (pol->policy == MPOL_INTERLEAVE)
1129 		return alloc_page_interleave(gfp, order, interleave_nodes(pol));
1130 	return __alloc_pages(gfp, order, zonelist_policy(gfp, pol));
1131 }
1132 EXPORT_SYMBOL(alloc_pages_current);
1133 
1134 /*
1135  * If mpol_copy() sees current->cpuset == cpuset_being_rebound, then it
1136  * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1137  * with the mems_allowed returned by cpuset_mems_allowed().  This
1138  * keeps mempolicies cpuset relative after its cpuset moves.  See
1139  * further kernel/cpuset.c update_nodemask().
1140  */
1141 void *cpuset_being_rebound;
1142 
1143 /* Slow path of a mempolicy copy */
1144 struct mempolicy *__mpol_copy(struct mempolicy *old)
1145 {
1146 	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1147 
1148 	if (!new)
1149 		return ERR_PTR(-ENOMEM);
1150 	if (current_cpuset_is_being_rebound()) {
1151 		nodemask_t mems = cpuset_mems_allowed(current);
1152 		mpol_rebind_policy(old, &mems);
1153 	}
1154 	*new = *old;
1155 	atomic_set(&new->refcnt, 1);
1156 	if (new->policy == MPOL_BIND) {
1157 		int sz = ksize(old->v.zonelist);
1158 		new->v.zonelist = kmalloc(sz, SLAB_KERNEL);
1159 		if (!new->v.zonelist) {
1160 			kmem_cache_free(policy_cache, new);
1161 			return ERR_PTR(-ENOMEM);
1162 		}
1163 		memcpy(new->v.zonelist, old->v.zonelist, sz);
1164 	}
1165 	return new;
1166 }
1167 
1168 /* Slow path of a mempolicy comparison */
1169 int __mpol_equal(struct mempolicy *a, struct mempolicy *b)
1170 {
1171 	if (!a || !b)
1172 		return 0;
1173 	if (a->policy != b->policy)
1174 		return 0;
1175 	switch (a->policy) {
1176 	case MPOL_DEFAULT:
1177 		return 1;
1178 	case MPOL_INTERLEAVE:
1179 		return nodes_equal(a->v.nodes, b->v.nodes);
1180 	case MPOL_PREFERRED:
1181 		return a->v.preferred_node == b->v.preferred_node;
1182 	case MPOL_BIND: {
1183 		int i;
1184 		for (i = 0; a->v.zonelist->zones[i]; i++)
1185 			if (a->v.zonelist->zones[i] != b->v.zonelist->zones[i])
1186 				return 0;
1187 		return b->v.zonelist->zones[i] == NULL;
1188 	}
1189 	default:
1190 		BUG();
1191 		return 0;
1192 	}
1193 }
1194 
1195 /* Slow path of a mpol destructor. */
1196 void __mpol_free(struct mempolicy *p)
1197 {
1198 	if (!atomic_dec_and_test(&p->refcnt))
1199 		return;
1200 	if (p->policy == MPOL_BIND)
1201 		kfree(p->v.zonelist);
1202 	p->policy = MPOL_DEFAULT;
1203 	kmem_cache_free(policy_cache, p);
1204 }
1205 
1206 /*
1207  * Shared memory backing store policy support.
1208  *
1209  * Remember policies even when nobody has shared memory mapped.
1210  * The policies are kept in Red-Black tree linked from the inode.
1211  * They are protected by the sp->lock spinlock, which should be held
1212  * for any accesses to the tree.
1213  */
1214 
1215 /* lookup first element intersecting start-end */
1216 /* Caller holds sp->lock */
1217 static struct sp_node *
1218 sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
1219 {
1220 	struct rb_node *n = sp->root.rb_node;
1221 
1222 	while (n) {
1223 		struct sp_node *p = rb_entry(n, struct sp_node, nd);
1224 
1225 		if (start >= p->end)
1226 			n = n->rb_right;
1227 		else if (end <= p->start)
1228 			n = n->rb_left;
1229 		else
1230 			break;
1231 	}
1232 	if (!n)
1233 		return NULL;
1234 	for (;;) {
1235 		struct sp_node *w = NULL;
1236 		struct rb_node *prev = rb_prev(n);
1237 		if (!prev)
1238 			break;
1239 		w = rb_entry(prev, struct sp_node, nd);
1240 		if (w->end <= start)
1241 			break;
1242 		n = prev;
1243 	}
1244 	return rb_entry(n, struct sp_node, nd);
1245 }
1246 
1247 /* Insert a new shared policy into the list. */
1248 /* Caller holds sp->lock */
1249 static void sp_insert(struct shared_policy *sp, struct sp_node *new)
1250 {
1251 	struct rb_node **p = &sp->root.rb_node;
1252 	struct rb_node *parent = NULL;
1253 	struct sp_node *nd;
1254 
1255 	while (*p) {
1256 		parent = *p;
1257 		nd = rb_entry(parent, struct sp_node, nd);
1258 		if (new->start < nd->start)
1259 			p = &(*p)->rb_left;
1260 		else if (new->end > nd->end)
1261 			p = &(*p)->rb_right;
1262 		else
1263 			BUG();
1264 	}
1265 	rb_link_node(&new->nd, parent, p);
1266 	rb_insert_color(&new->nd, &sp->root);
1267 	PDprintk("inserting %lx-%lx: %d\n", new->start, new->end,
1268 		 new->policy ? new->policy->policy : 0);
1269 }
1270 
1271 /* Find shared policy intersecting idx */
1272 struct mempolicy *
1273 mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
1274 {
1275 	struct mempolicy *pol = NULL;
1276 	struct sp_node *sn;
1277 
1278 	if (!sp->root.rb_node)
1279 		return NULL;
1280 	spin_lock(&sp->lock);
1281 	sn = sp_lookup(sp, idx, idx+1);
1282 	if (sn) {
1283 		mpol_get(sn->policy);
1284 		pol = sn->policy;
1285 	}
1286 	spin_unlock(&sp->lock);
1287 	return pol;
1288 }
1289 
1290 static void sp_delete(struct shared_policy *sp, struct sp_node *n)
1291 {
1292 	PDprintk("deleting %lx-l%x\n", n->start, n->end);
1293 	rb_erase(&n->nd, &sp->root);
1294 	mpol_free(n->policy);
1295 	kmem_cache_free(sn_cache, n);
1296 }
1297 
1298 struct sp_node *
1299 sp_alloc(unsigned long start, unsigned long end, struct mempolicy *pol)
1300 {
1301 	struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
1302 
1303 	if (!n)
1304 		return NULL;
1305 	n->start = start;
1306 	n->end = end;
1307 	mpol_get(pol);
1308 	n->policy = pol;
1309 	return n;
1310 }
1311 
1312 /* Replace a policy range. */
1313 static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
1314 				 unsigned long end, struct sp_node *new)
1315 {
1316 	struct sp_node *n, *new2 = NULL;
1317 
1318 restart:
1319 	spin_lock(&sp->lock);
1320 	n = sp_lookup(sp, start, end);
1321 	/* Take care of old policies in the same range. */
1322 	while (n && n->start < end) {
1323 		struct rb_node *next = rb_next(&n->nd);
1324 		if (n->start >= start) {
1325 			if (n->end <= end)
1326 				sp_delete(sp, n);
1327 			else
1328 				n->start = end;
1329 		} else {
1330 			/* Old policy spanning whole new range. */
1331 			if (n->end > end) {
1332 				if (!new2) {
1333 					spin_unlock(&sp->lock);
1334 					new2 = sp_alloc(end, n->end, n->policy);
1335 					if (!new2)
1336 						return -ENOMEM;
1337 					goto restart;
1338 				}
1339 				n->end = start;
1340 				sp_insert(sp, new2);
1341 				new2 = NULL;
1342 				break;
1343 			} else
1344 				n->end = start;
1345 		}
1346 		if (!next)
1347 			break;
1348 		n = rb_entry(next, struct sp_node, nd);
1349 	}
1350 	if (new)
1351 		sp_insert(sp, new);
1352 	spin_unlock(&sp->lock);
1353 	if (new2) {
1354 		mpol_free(new2->policy);
1355 		kmem_cache_free(sn_cache, new2);
1356 	}
1357 	return 0;
1358 }
1359 
1360 int mpol_set_shared_policy(struct shared_policy *info,
1361 			struct vm_area_struct *vma, struct mempolicy *npol)
1362 {
1363 	int err;
1364 	struct sp_node *new = NULL;
1365 	unsigned long sz = vma_pages(vma);
1366 
1367 	PDprintk("set_shared_policy %lx sz %lu %d %lx\n",
1368 		 vma->vm_pgoff,
1369 		 sz, npol? npol->policy : -1,
1370 		npol ? nodes_addr(npol->v.nodes)[0] : -1);
1371 
1372 	if (npol) {
1373 		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
1374 		if (!new)
1375 			return -ENOMEM;
1376 	}
1377 	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
1378 	if (err && new)
1379 		kmem_cache_free(sn_cache, new);
1380 	return err;
1381 }
1382 
1383 /* Free a backing policy store on inode delete. */
1384 void mpol_free_shared_policy(struct shared_policy *p)
1385 {
1386 	struct sp_node *n;
1387 	struct rb_node *next;
1388 
1389 	if (!p->root.rb_node)
1390 		return;
1391 	spin_lock(&p->lock);
1392 	next = rb_first(&p->root);
1393 	while (next) {
1394 		n = rb_entry(next, struct sp_node, nd);
1395 		next = rb_next(&n->nd);
1396 		rb_erase(&n->nd, &p->root);
1397 		mpol_free(n->policy);
1398 		kmem_cache_free(sn_cache, n);
1399 	}
1400 	spin_unlock(&p->lock);
1401 }
1402 
1403 /* assumes fs == KERNEL_DS */
1404 void __init numa_policy_init(void)
1405 {
1406 	policy_cache = kmem_cache_create("numa_policy",
1407 					 sizeof(struct mempolicy),
1408 					 0, SLAB_PANIC, NULL, NULL);
1409 
1410 	sn_cache = kmem_cache_create("shared_policy_node",
1411 				     sizeof(struct sp_node),
1412 				     0, SLAB_PANIC, NULL, NULL);
1413 
1414 	/* Set interleaving policy for system init. This way not all
1415 	   the data structures allocated at system boot end up in node zero. */
1416 
1417 	if (do_set_mempolicy(MPOL_INTERLEAVE, &node_online_map))
1418 		printk("numa_policy_init: interleaving failed\n");
1419 }
1420 
1421 /* Reset policy of current process to default */
1422 void numa_default_policy(void)
1423 {
1424 	do_set_mempolicy(MPOL_DEFAULT, NULL);
1425 }
1426 
1427 /* Migrate a policy to a different set of nodes */
1428 void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
1429 {
1430 	nodemask_t *mpolmask;
1431 	nodemask_t tmp;
1432 
1433 	if (!pol)
1434 		return;
1435 	mpolmask = &pol->cpuset_mems_allowed;
1436 	if (nodes_equal(*mpolmask, *newmask))
1437 		return;
1438 
1439 	switch (pol->policy) {
1440 	case MPOL_DEFAULT:
1441 		break;
1442 	case MPOL_INTERLEAVE:
1443 		nodes_remap(tmp, pol->v.nodes, *mpolmask, *newmask);
1444 		pol->v.nodes = tmp;
1445 		*mpolmask = *newmask;
1446 		current->il_next = node_remap(current->il_next,
1447 						*mpolmask, *newmask);
1448 		break;
1449 	case MPOL_PREFERRED:
1450 		pol->v.preferred_node = node_remap(pol->v.preferred_node,
1451 						*mpolmask, *newmask);
1452 		*mpolmask = *newmask;
1453 		break;
1454 	case MPOL_BIND: {
1455 		nodemask_t nodes;
1456 		struct zone **z;
1457 		struct zonelist *zonelist;
1458 
1459 		nodes_clear(nodes);
1460 		for (z = pol->v.zonelist->zones; *z; z++)
1461 			node_set((*z)->zone_pgdat->node_id, nodes);
1462 		nodes_remap(tmp, nodes, *mpolmask, *newmask);
1463 		nodes = tmp;
1464 
1465 		zonelist = bind_zonelist(&nodes);
1466 
1467 		/* If no mem, then zonelist is NULL and we keep old zonelist.
1468 		 * If that old zonelist has no remaining mems_allowed nodes,
1469 		 * then zonelist_policy() will "FALL THROUGH" to MPOL_DEFAULT.
1470 		 */
1471 
1472 		if (zonelist) {
1473 			/* Good - got mem - substitute new zonelist */
1474 			kfree(pol->v.zonelist);
1475 			pol->v.zonelist = zonelist;
1476 		}
1477 		*mpolmask = *newmask;
1478 		break;
1479 	}
1480 	default:
1481 		BUG();
1482 		break;
1483 	}
1484 }
1485 
1486 /*
1487  * Wrapper for mpol_rebind_policy() that just requires task
1488  * pointer, and updates task mempolicy.
1489  */
1490 
1491 void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
1492 {
1493 	mpol_rebind_policy(tsk->mempolicy, new);
1494 }
1495 
1496 /*
1497  * Rebind each vma in mm to new nodemask.
1498  *
1499  * Call holding a reference to mm.  Takes mm->mmap_sem during call.
1500  */
1501 
1502 void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
1503 {
1504 	struct vm_area_struct *vma;
1505 
1506 	down_write(&mm->mmap_sem);
1507 	for (vma = mm->mmap; vma; vma = vma->vm_next)
1508 		mpol_rebind_policy(vma->vm_policy, new);
1509 	up_write(&mm->mmap_sem);
1510 }
1511 
1512 /*
1513  * Display pages allocated per node and memory policy via /proc.
1514  */
1515 
1516 static const char *policy_types[] = { "default", "prefer", "bind",
1517 				      "interleave" };
1518 
1519 /*
1520  * Convert a mempolicy into a string.
1521  * Returns the number of characters in buffer (if positive)
1522  * or an error (negative)
1523  */
1524 static inline int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
1525 {
1526 	char *p = buffer;
1527 	int l;
1528 	nodemask_t nodes;
1529 	int mode = pol ? pol->policy : MPOL_DEFAULT;
1530 
1531 	switch (mode) {
1532 	case MPOL_DEFAULT:
1533 		nodes_clear(nodes);
1534 		break;
1535 
1536 	case MPOL_PREFERRED:
1537 		nodes_clear(nodes);
1538 		node_set(pol->v.preferred_node, nodes);
1539 		break;
1540 
1541 	case MPOL_BIND:
1542 		get_zonemask(pol, &nodes);
1543 		break;
1544 
1545 	case MPOL_INTERLEAVE:
1546 		nodes = pol->v.nodes;
1547 		break;
1548 
1549 	default:
1550 		BUG();
1551 		return -EFAULT;
1552 	}
1553 
1554 	l = strlen(policy_types[mode]);
1555  	if (buffer + maxlen < p + l + 1)
1556  		return -ENOSPC;
1557 
1558 	strcpy(p, policy_types[mode]);
1559 	p += l;
1560 
1561 	if (!nodes_empty(nodes)) {
1562 		if (buffer + maxlen < p + 2)
1563 			return -ENOSPC;
1564 		*p++ = '=';
1565 	 	p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
1566 	}
1567 	return p - buffer;
1568 }
1569 
1570 struct numa_maps {
1571 	unsigned long pages;
1572 	unsigned long anon;
1573 	unsigned long mapped;
1574 	unsigned long mapcount_max;
1575 	unsigned long node[MAX_NUMNODES];
1576 };
1577 
1578 static void gather_stats(struct page *page, void *private)
1579 {
1580 	struct numa_maps *md = private;
1581 	int count = page_mapcount(page);
1582 
1583 	if (count)
1584 		md->mapped++;
1585 
1586 	if (count > md->mapcount_max)
1587 		md->mapcount_max = count;
1588 
1589 	md->pages++;
1590 
1591 	if (PageAnon(page))
1592 		md->anon++;
1593 
1594 	md->node[page_to_nid(page)]++;
1595 	cond_resched();
1596 }
1597 
1598 int show_numa_map(struct seq_file *m, void *v)
1599 {
1600 	struct task_struct *task = m->private;
1601 	struct vm_area_struct *vma = v;
1602 	struct numa_maps *md;
1603 	int n;
1604 	char buffer[50];
1605 
1606 	if (!vma->vm_mm)
1607 		return 0;
1608 
1609 	md = kzalloc(sizeof(struct numa_maps), GFP_KERNEL);
1610 	if (!md)
1611 		return 0;
1612 
1613 	check_pgd_range(vma, vma->vm_start, vma->vm_end,
1614 		    &node_online_map, MPOL_MF_STATS, md);
1615 
1616 	if (md->pages) {
1617 		mpol_to_str(buffer, sizeof(buffer),
1618 			    get_vma_policy(task, vma, vma->vm_start));
1619 
1620 		seq_printf(m, "%08lx %s pages=%lu mapped=%lu maxref=%lu",
1621 			   vma->vm_start, buffer, md->pages,
1622 			   md->mapped, md->mapcount_max);
1623 
1624 		if (md->anon)
1625 			seq_printf(m," anon=%lu",md->anon);
1626 
1627 		for_each_online_node(n)
1628 			if (md->node[n])
1629 				seq_printf(m, " N%d=%lu", n, md->node[n]);
1630 
1631 		seq_putc(m, '\n');
1632 	}
1633 	kfree(md);
1634 
1635 	if (m->count < m->size)
1636 		m->version = (vma != get_gate_vma(task)) ? vma->vm_start : 0;
1637 	return 0;
1638 }
1639 
1640