Linux Audio

Check our new training course

Loading...
   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 NUMA_NO_NODE 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*/
  67
  68#include <linux/mempolicy.h>
  69#include <linux/mm.h>
  70#include <linux/highmem.h>
  71#include <linux/hugetlb.h>
  72#include <linux/kernel.h>
  73#include <linux/sched.h>
  74#include <linux/nodemask.h>
  75#include <linux/cpuset.h>
  76#include <linux/slab.h>
  77#include <linux/string.h>
  78#include <linux/export.h>
  79#include <linux/nsproxy.h>
  80#include <linux/interrupt.h>
  81#include <linux/init.h>
  82#include <linux/compat.h>
  83#include <linux/swap.h>
  84#include <linux/seq_file.h>
  85#include <linux/proc_fs.h>
  86#include <linux/migrate.h>
  87#include <linux/ksm.h>
  88#include <linux/rmap.h>
  89#include <linux/security.h>
  90#include <linux/syscalls.h>
  91#include <linux/ctype.h>
  92#include <linux/mm_inline.h>
  93#include <linux/mmu_notifier.h>
  94
  95#include <asm/tlbflush.h>
  96#include <asm/uaccess.h>
  97#include <linux/random.h>
  98
  99#include "internal.h"
 100
 101/* Internal flags */
 102#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)	/* Skip checks for continuous vmas */
 103#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)		/* Invert check for nodemask */
 104
 105static struct kmem_cache *policy_cache;
 106static struct kmem_cache *sn_cache;
 107
 108/* Highest zone. An specific allocation for a zone below that is not
 109   policied. */
 110enum zone_type policy_zone = 0;
 111
 112/*
 113 * run-time system-wide default policy => local allocation
 114 */
 115static struct mempolicy default_policy = {
 116	.refcnt = ATOMIC_INIT(1), /* never free it */
 117	.mode = MPOL_PREFERRED,
 118	.flags = MPOL_F_LOCAL,
 119};
 120
 121static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 122
 123static struct mempolicy *get_task_policy(struct task_struct *p)
 124{
 125	struct mempolicy *pol = p->mempolicy;
 126
 127	if (!pol) {
 128		int node = numa_node_id();
 129
 130		if (node != NUMA_NO_NODE) {
 131			pol = &preferred_node_policy[node];
 132			/*
 133			 * preferred_node_policy is not initialised early in
 134			 * boot
 135			 */
 136			if (!pol->mode)
 137				pol = NULL;
 138		}
 139	}
 140
 141	return pol;
 142}
 143
 144static const struct mempolicy_operations {
 145	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 146	/*
 147	 * If read-side task has no lock to protect task->mempolicy, write-side
 148	 * task will rebind the task->mempolicy by two step. The first step is
 149	 * setting all the newly nodes, and the second step is cleaning all the
 150	 * disallowed nodes. In this way, we can avoid finding no node to alloc
 151	 * page.
 152	 * If we have a lock to protect task->mempolicy in read-side, we do
 153	 * rebind directly.
 154	 *
 155	 * step:
 156	 * 	MPOL_REBIND_ONCE - do rebind work at once
 157	 * 	MPOL_REBIND_STEP1 - set all the newly nodes
 158	 * 	MPOL_REBIND_STEP2 - clean all the disallowed nodes
 159	 */
 160	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
 161			enum mpol_rebind_step step);
 162} mpol_ops[MPOL_MAX];
 163
 164/* Check that the nodemask contains at least one populated zone */
 165static int is_valid_nodemask(const nodemask_t *nodemask)
 166{
 167	return nodes_intersects(*nodemask, node_states[N_MEMORY]);
 168}
 169
 170static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 171{
 172	return pol->flags & MPOL_MODE_FLAGS;
 173}
 174
 175static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 176				   const nodemask_t *rel)
 177{
 178	nodemask_t tmp;
 179	nodes_fold(tmp, *orig, nodes_weight(*rel));
 180	nodes_onto(*ret, tmp, *rel);
 181}
 182
 183static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
 184{
 185	if (nodes_empty(*nodes))
 186		return -EINVAL;
 187	pol->v.nodes = *nodes;
 188	return 0;
 189}
 190
 191static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 192{
 193	if (!nodes)
 194		pol->flags |= MPOL_F_LOCAL;	/* local allocation */
 195	else if (nodes_empty(*nodes))
 196		return -EINVAL;			/*  no allowed nodes */
 197	else
 198		pol->v.preferred_node = first_node(*nodes);
 199	return 0;
 200}
 201
 202static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
 203{
 204	if (!is_valid_nodemask(nodes))
 205		return -EINVAL;
 206	pol->v.nodes = *nodes;
 207	return 0;
 208}
 209
 210/*
 211 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 212 * any, for the new policy.  mpol_new() has already validated the nodes
 213 * parameter with respect to the policy mode and flags.  But, we need to
 214 * handle an empty nodemask with MPOL_PREFERRED here.
 215 *
 216 * Must be called holding task's alloc_lock to protect task's mems_allowed
 217 * and mempolicy.  May also be called holding the mmap_semaphore for write.
 218 */
 219static int mpol_set_nodemask(struct mempolicy *pol,
 220		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 221{
 222	int ret;
 223
 224	/* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
 225	if (pol == NULL)
 226		return 0;
 227	/* Check N_MEMORY */
 228	nodes_and(nsc->mask1,
 229		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 230
 231	VM_BUG_ON(!nodes);
 232	if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
 233		nodes = NULL;	/* explicit local allocation */
 234	else {
 235		if (pol->flags & MPOL_F_RELATIVE_NODES)
 236			mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
 237		else
 238			nodes_and(nsc->mask2, *nodes, nsc->mask1);
 239
 240		if (mpol_store_user_nodemask(pol))
 241			pol->w.user_nodemask = *nodes;
 242		else
 243			pol->w.cpuset_mems_allowed =
 244						cpuset_current_mems_allowed;
 245	}
 246
 247	if (nodes)
 248		ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 249	else
 250		ret = mpol_ops[pol->mode].create(pol, NULL);
 251	return ret;
 252}
 253
 254/*
 255 * This function just creates a new policy, does some check and simple
 256 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 257 */
 258static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 259				  nodemask_t *nodes)
 260{
 261	struct mempolicy *policy;
 262
 263	pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 264		 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
 265
 266	if (mode == MPOL_DEFAULT) {
 267		if (nodes && !nodes_empty(*nodes))
 268			return ERR_PTR(-EINVAL);
 269		return NULL;
 270	}
 271	VM_BUG_ON(!nodes);
 272
 273	/*
 274	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 275	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 276	 * All other modes require a valid pointer to a non-empty nodemask.
 277	 */
 278	if (mode == MPOL_PREFERRED) {
 279		if (nodes_empty(*nodes)) {
 280			if (((flags & MPOL_F_STATIC_NODES) ||
 281			     (flags & MPOL_F_RELATIVE_NODES)))
 282				return ERR_PTR(-EINVAL);
 283		}
 284	} else if (mode == MPOL_LOCAL) {
 285		if (!nodes_empty(*nodes))
 286			return ERR_PTR(-EINVAL);
 287		mode = MPOL_PREFERRED;
 288	} else if (nodes_empty(*nodes))
 289		return ERR_PTR(-EINVAL);
 290	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 291	if (!policy)
 292		return ERR_PTR(-ENOMEM);
 293	atomic_set(&policy->refcnt, 1);
 294	policy->mode = mode;
 295	policy->flags = flags;
 296
 297	return policy;
 298}
 299
 300/* Slow path of a mpol destructor. */
 301void __mpol_put(struct mempolicy *p)
 302{
 303	if (!atomic_dec_and_test(&p->refcnt))
 304		return;
 305	kmem_cache_free(policy_cache, p);
 306}
 307
 308static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
 309				enum mpol_rebind_step step)
 310{
 311}
 312
 313/*
 314 * step:
 315 * 	MPOL_REBIND_ONCE  - do rebind work at once
 316 * 	MPOL_REBIND_STEP1 - set all the newly nodes
 317 * 	MPOL_REBIND_STEP2 - clean all the disallowed nodes
 318 */
 319static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
 320				 enum mpol_rebind_step step)
 321{
 322	nodemask_t tmp;
 323
 324	if (pol->flags & MPOL_F_STATIC_NODES)
 325		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 326	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 327		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 328	else {
 329		/*
 330		 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
 331		 * result
 332		 */
 333		if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
 334			nodes_remap(tmp, pol->v.nodes,
 335					pol->w.cpuset_mems_allowed, *nodes);
 336			pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
 337		} else if (step == MPOL_REBIND_STEP2) {
 338			tmp = pol->w.cpuset_mems_allowed;
 339			pol->w.cpuset_mems_allowed = *nodes;
 340		} else
 341			BUG();
 342	}
 343
 344	if (nodes_empty(tmp))
 345		tmp = *nodes;
 346
 347	if (step == MPOL_REBIND_STEP1)
 348		nodes_or(pol->v.nodes, pol->v.nodes, tmp);
 349	else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
 350		pol->v.nodes = tmp;
 351	else
 352		BUG();
 353
 354	if (!node_isset(current->il_next, tmp)) {
 355		current->il_next = next_node(current->il_next, tmp);
 356		if (current->il_next >= MAX_NUMNODES)
 357			current->il_next = first_node(tmp);
 358		if (current->il_next >= MAX_NUMNODES)
 359			current->il_next = numa_node_id();
 360	}
 361}
 362
 363static void mpol_rebind_preferred(struct mempolicy *pol,
 364				  const nodemask_t *nodes,
 365				  enum mpol_rebind_step step)
 366{
 367	nodemask_t tmp;
 368
 369	if (pol->flags & MPOL_F_STATIC_NODES) {
 370		int node = first_node(pol->w.user_nodemask);
 371
 372		if (node_isset(node, *nodes)) {
 373			pol->v.preferred_node = node;
 374			pol->flags &= ~MPOL_F_LOCAL;
 375		} else
 376			pol->flags |= MPOL_F_LOCAL;
 377	} else if (pol->flags & MPOL_F_RELATIVE_NODES) {
 378		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 379		pol->v.preferred_node = first_node(tmp);
 380	} else if (!(pol->flags & MPOL_F_LOCAL)) {
 381		pol->v.preferred_node = node_remap(pol->v.preferred_node,
 382						   pol->w.cpuset_mems_allowed,
 383						   *nodes);
 384		pol->w.cpuset_mems_allowed = *nodes;
 385	}
 386}
 387
 388/*
 389 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 390 *
 391 * If read-side task has no lock to protect task->mempolicy, write-side
 392 * task will rebind the task->mempolicy by two step. The first step is
 393 * setting all the newly nodes, and the second step is cleaning all the
 394 * disallowed nodes. In this way, we can avoid finding no node to alloc
 395 * page.
 396 * If we have a lock to protect task->mempolicy in read-side, we do
 397 * rebind directly.
 398 *
 399 * step:
 400 * 	MPOL_REBIND_ONCE  - do rebind work at once
 401 * 	MPOL_REBIND_STEP1 - set all the newly nodes
 402 * 	MPOL_REBIND_STEP2 - clean all the disallowed nodes
 403 */
 404static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
 405				enum mpol_rebind_step step)
 406{
 407	if (!pol)
 408		return;
 409	if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
 410	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 411		return;
 412
 413	if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
 414		return;
 415
 416	if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
 417		BUG();
 418
 419	if (step == MPOL_REBIND_STEP1)
 420		pol->flags |= MPOL_F_REBINDING;
 421	else if (step == MPOL_REBIND_STEP2)
 422		pol->flags &= ~MPOL_F_REBINDING;
 423	else if (step >= MPOL_REBIND_NSTEP)
 424		BUG();
 425
 426	mpol_ops[pol->mode].rebind(pol, newmask, step);
 427}
 428
 429/*
 430 * Wrapper for mpol_rebind_policy() that just requires task
 431 * pointer, and updates task mempolicy.
 432 *
 433 * Called with task's alloc_lock held.
 434 */
 435
 436void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
 437			enum mpol_rebind_step step)
 438{
 439	mpol_rebind_policy(tsk->mempolicy, new, step);
 440}
 441
 442/*
 443 * Rebind each vma in mm to new nodemask.
 444 *
 445 * Call holding a reference to mm.  Takes mm->mmap_sem during call.
 446 */
 447
 448void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 449{
 450	struct vm_area_struct *vma;
 451
 452	down_write(&mm->mmap_sem);
 453	for (vma = mm->mmap; vma; vma = vma->vm_next)
 454		mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
 455	up_write(&mm->mmap_sem);
 456}
 457
 458static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 459	[MPOL_DEFAULT] = {
 460		.rebind = mpol_rebind_default,
 461	},
 462	[MPOL_INTERLEAVE] = {
 463		.create = mpol_new_interleave,
 464		.rebind = mpol_rebind_nodemask,
 465	},
 466	[MPOL_PREFERRED] = {
 467		.create = mpol_new_preferred,
 468		.rebind = mpol_rebind_preferred,
 469	},
 470	[MPOL_BIND] = {
 471		.create = mpol_new_bind,
 472		.rebind = mpol_rebind_nodemask,
 473	},
 474};
 475
 476static void migrate_page_add(struct page *page, struct list_head *pagelist,
 477				unsigned long flags);
 478
 479/*
 480 * Scan through pages checking if pages follow certain conditions,
 481 * and move them to the pagelist if they do.
 482 */
 483static int queue_pages_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 484		unsigned long addr, unsigned long end,
 485		const nodemask_t *nodes, unsigned long flags,
 486		void *private)
 487{
 488	pte_t *orig_pte;
 489	pte_t *pte;
 490	spinlock_t *ptl;
 491
 492	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 493	do {
 494		struct page *page;
 495		int nid;
 496
 497		if (!pte_present(*pte))
 498			continue;
 499		page = vm_normal_page(vma, addr, *pte);
 500		if (!page)
 501			continue;
 502		/*
 503		 * vm_normal_page() filters out zero pages, but there might
 504		 * still be PageReserved pages to skip, perhaps in a VDSO.
 505		 */
 506		if (PageReserved(page))
 507			continue;
 508		nid = page_to_nid(page);
 509		if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
 510			continue;
 511
 512		if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
 513			migrate_page_add(page, private, flags);
 514		else
 515			break;
 516	} while (pte++, addr += PAGE_SIZE, addr != end);
 517	pte_unmap_unlock(orig_pte, ptl);
 518	return addr != end;
 519}
 520
 521static void queue_pages_hugetlb_pmd_range(struct vm_area_struct *vma,
 522		pmd_t *pmd, const nodemask_t *nodes, unsigned long flags,
 523				    void *private)
 524{
 525#ifdef CONFIG_HUGETLB_PAGE
 526	int nid;
 527	struct page *page;
 528	spinlock_t *ptl;
 529	pte_t entry;
 530
 531	ptl = huge_pte_lock(hstate_vma(vma), vma->vm_mm, (pte_t *)pmd);
 532	entry = huge_ptep_get((pte_t *)pmd);
 533	if (!pte_present(entry))
 534		goto unlock;
 535	page = pte_page(entry);
 536	nid = page_to_nid(page);
 537	if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
 538		goto unlock;
 539	/* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
 540	if (flags & (MPOL_MF_MOVE_ALL) ||
 541	    (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
 542		isolate_huge_page(page, private);
 543unlock:
 544	spin_unlock(ptl);
 545#else
 546	BUG();
 547#endif
 548}
 549
 550static inline int queue_pages_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 551		unsigned long addr, unsigned long end,
 552		const nodemask_t *nodes, unsigned long flags,
 553		void *private)
 554{
 555	pmd_t *pmd;
 556	unsigned long next;
 557
 558	pmd = pmd_offset(pud, addr);
 559	do {
 560		next = pmd_addr_end(addr, end);
 561		if (!pmd_present(*pmd))
 562			continue;
 563		if (pmd_huge(*pmd) && is_vm_hugetlb_page(vma)) {
 564			queue_pages_hugetlb_pmd_range(vma, pmd, nodes,
 565						flags, private);
 566			continue;
 567		}
 568		split_huge_page_pmd(vma, addr, pmd);
 569		if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 570			continue;
 571		if (queue_pages_pte_range(vma, pmd, addr, next, nodes,
 572				    flags, private))
 573			return -EIO;
 574	} while (pmd++, addr = next, addr != end);
 575	return 0;
 576}
 577
 578static inline int queue_pages_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 579		unsigned long addr, unsigned long end,
 580		const nodemask_t *nodes, unsigned long flags,
 581		void *private)
 582{
 583	pud_t *pud;
 584	unsigned long next;
 585
 586	pud = pud_offset(pgd, addr);
 587	do {
 588		next = pud_addr_end(addr, end);
 589		if (pud_huge(*pud) && is_vm_hugetlb_page(vma))
 590			continue;
 591		if (pud_none_or_clear_bad(pud))
 592			continue;
 593		if (queue_pages_pmd_range(vma, pud, addr, next, nodes,
 594				    flags, private))
 595			return -EIO;
 596	} while (pud++, addr = next, addr != end);
 597	return 0;
 598}
 599
 600static inline int queue_pages_pgd_range(struct vm_area_struct *vma,
 601		unsigned long addr, unsigned long end,
 602		const nodemask_t *nodes, unsigned long flags,
 603		void *private)
 604{
 605	pgd_t *pgd;
 606	unsigned long next;
 607
 608	pgd = pgd_offset(vma->vm_mm, addr);
 609	do {
 610		next = pgd_addr_end(addr, end);
 611		if (pgd_none_or_clear_bad(pgd))
 612			continue;
 613		if (queue_pages_pud_range(vma, pgd, addr, next, nodes,
 614				    flags, private))
 615			return -EIO;
 616	} while (pgd++, addr = next, addr != end);
 617	return 0;
 618}
 619
 620#ifdef CONFIG_NUMA_BALANCING
 621/*
 622 * This is used to mark a range of virtual addresses to be inaccessible.
 623 * These are later cleared by a NUMA hinting fault. Depending on these
 624 * faults, pages may be migrated for better NUMA placement.
 625 *
 626 * This is assuming that NUMA faults are handled using PROT_NONE. If
 627 * an architecture makes a different choice, it will need further
 628 * changes to the core.
 629 */
 630unsigned long change_prot_numa(struct vm_area_struct *vma,
 631			unsigned long addr, unsigned long end)
 632{
 633	int nr_updated;
 634
 635	nr_updated = change_protection(vma, addr, end, vma->vm_page_prot, 0, 1);
 636	if (nr_updated)
 637		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 638
 639	return nr_updated;
 640}
 641#else
 642static unsigned long change_prot_numa(struct vm_area_struct *vma,
 643			unsigned long addr, unsigned long end)
 644{
 645	return 0;
 646}
 647#endif /* CONFIG_NUMA_BALANCING */
 648
 649/*
 650 * Walk through page tables and collect pages to be migrated.
 651 *
 652 * If pages found in a given range are on a set of nodes (determined by
 653 * @nodes and @flags,) it's isolated and queued to the pagelist which is
 654 * passed via @private.)
 655 */
 656static struct vm_area_struct *
 657queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 658		const nodemask_t *nodes, unsigned long flags, void *private)
 659{
 660	int err;
 661	struct vm_area_struct *first, *vma, *prev;
 662
 663
 664	first = find_vma(mm, start);
 665	if (!first)
 666		return ERR_PTR(-EFAULT);
 667	prev = NULL;
 668	for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
 669		unsigned long endvma = vma->vm_end;
 670
 671		if (endvma > end)
 672			endvma = end;
 673		if (vma->vm_start > start)
 674			start = vma->vm_start;
 675
 676		if (!(flags & MPOL_MF_DISCONTIG_OK)) {
 677			if (!vma->vm_next && vma->vm_end < end)
 678				return ERR_PTR(-EFAULT);
 679			if (prev && prev->vm_end < vma->vm_start)
 680				return ERR_PTR(-EFAULT);
 681		}
 682
 683		if (flags & MPOL_MF_LAZY) {
 684			change_prot_numa(vma, start, endvma);
 685			goto next;
 686		}
 687
 688		if ((flags & MPOL_MF_STRICT) ||
 689		     ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
 690		      vma_migratable(vma))) {
 691
 692			err = queue_pages_pgd_range(vma, start, endvma, nodes,
 693						flags, private);
 694			if (err) {
 695				first = ERR_PTR(err);
 696				break;
 697			}
 698		}
 699next:
 700		prev = vma;
 701	}
 702	return first;
 703}
 704
 705/*
 706 * Apply policy to a single VMA
 707 * This must be called with the mmap_sem held for writing.
 708 */
 709static int vma_replace_policy(struct vm_area_struct *vma,
 710						struct mempolicy *pol)
 711{
 712	int err;
 713	struct mempolicy *old;
 714	struct mempolicy *new;
 715
 716	pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 717		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 718		 vma->vm_ops, vma->vm_file,
 719		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 720
 721	new = mpol_dup(pol);
 722	if (IS_ERR(new))
 723		return PTR_ERR(new);
 724
 725	if (vma->vm_ops && vma->vm_ops->set_policy) {
 726		err = vma->vm_ops->set_policy(vma, new);
 727		if (err)
 728			goto err_out;
 729	}
 730
 731	old = vma->vm_policy;
 732	vma->vm_policy = new; /* protected by mmap_sem */
 733	mpol_put(old);
 734
 735	return 0;
 736 err_out:
 737	mpol_put(new);
 738	return err;
 739}
 740
 741/* Step 2: apply policy to a range and do splits. */
 742static int mbind_range(struct mm_struct *mm, unsigned long start,
 743		       unsigned long end, struct mempolicy *new_pol)
 744{
 745	struct vm_area_struct *next;
 746	struct vm_area_struct *prev;
 747	struct vm_area_struct *vma;
 748	int err = 0;
 749	pgoff_t pgoff;
 750	unsigned long vmstart;
 751	unsigned long vmend;
 752
 753	vma = find_vma(mm, start);
 754	if (!vma || vma->vm_start > start)
 755		return -EFAULT;
 756
 757	prev = vma->vm_prev;
 758	if (start > vma->vm_start)
 759		prev = vma;
 760
 761	for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 762		next = vma->vm_next;
 763		vmstart = max(start, vma->vm_start);
 764		vmend   = min(end, vma->vm_end);
 765
 766		if (mpol_equal(vma_policy(vma), new_pol))
 767			continue;
 768
 769		pgoff = vma->vm_pgoff +
 770			((vmstart - vma->vm_start) >> PAGE_SHIFT);
 771		prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 772				  vma->anon_vma, vma->vm_file, pgoff,
 773				  new_pol);
 774		if (prev) {
 775			vma = prev;
 776			next = vma->vm_next;
 777			if (mpol_equal(vma_policy(vma), new_pol))
 778				continue;
 779			/* vma_merge() joined vma && vma->next, case 8 */
 780			goto replace;
 781		}
 782		if (vma->vm_start != vmstart) {
 783			err = split_vma(vma->vm_mm, vma, vmstart, 1);
 784			if (err)
 785				goto out;
 786		}
 787		if (vma->vm_end != vmend) {
 788			err = split_vma(vma->vm_mm, vma, vmend, 0);
 789			if (err)
 790				goto out;
 791		}
 792 replace:
 793		err = vma_replace_policy(vma, new_pol);
 794		if (err)
 795			goto out;
 796	}
 797
 798 out:
 799	return err;
 800}
 801
 802/* Set the process memory policy */
 803static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 804			     nodemask_t *nodes)
 805{
 806	struct mempolicy *new, *old;
 807	struct mm_struct *mm = current->mm;
 808	NODEMASK_SCRATCH(scratch);
 809	int ret;
 810
 811	if (!scratch)
 812		return -ENOMEM;
 813
 814	new = mpol_new(mode, flags, nodes);
 815	if (IS_ERR(new)) {
 816		ret = PTR_ERR(new);
 817		goto out;
 818	}
 819	/*
 820	 * prevent changing our mempolicy while show_numa_maps()
 821	 * is using it.
 822	 * Note:  do_set_mempolicy() can be called at init time
 823	 * with no 'mm'.
 824	 */
 825	if (mm)
 826		down_write(&mm->mmap_sem);
 827	task_lock(current);
 828	ret = mpol_set_nodemask(new, nodes, scratch);
 829	if (ret) {
 830		task_unlock(current);
 831		if (mm)
 832			up_write(&mm->mmap_sem);
 833		mpol_put(new);
 834		goto out;
 835	}
 836	old = current->mempolicy;
 837	current->mempolicy = new;
 838	if (new && new->mode == MPOL_INTERLEAVE &&
 839	    nodes_weight(new->v.nodes))
 840		current->il_next = first_node(new->v.nodes);
 841	task_unlock(current);
 842	if (mm)
 843		up_write(&mm->mmap_sem);
 844
 845	mpol_put(old);
 846	ret = 0;
 847out:
 848	NODEMASK_SCRATCH_FREE(scratch);
 849	return ret;
 850}
 851
 852/*
 853 * Return nodemask for policy for get_mempolicy() query
 854 *
 855 * Called with task's alloc_lock held
 856 */
 857static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 858{
 859	nodes_clear(*nodes);
 860	if (p == &default_policy)
 861		return;
 862
 863	switch (p->mode) {
 864	case MPOL_BIND:
 865		/* Fall through */
 866	case MPOL_INTERLEAVE:
 867		*nodes = p->v.nodes;
 868		break;
 869	case MPOL_PREFERRED:
 870		if (!(p->flags & MPOL_F_LOCAL))
 871			node_set(p->v.preferred_node, *nodes);
 872		/* else return empty node mask for local allocation */
 873		break;
 874	default:
 875		BUG();
 876	}
 877}
 878
 879static int lookup_node(struct mm_struct *mm, unsigned long addr)
 880{
 881	struct page *p;
 882	int err;
 883
 884	err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
 885	if (err >= 0) {
 886		err = page_to_nid(p);
 887		put_page(p);
 888	}
 889	return err;
 890}
 891
 892/* Retrieve NUMA policy */
 893static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 894			     unsigned long addr, unsigned long flags)
 895{
 896	int err;
 897	struct mm_struct *mm = current->mm;
 898	struct vm_area_struct *vma = NULL;
 899	struct mempolicy *pol = current->mempolicy;
 900
 901	if (flags &
 902		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 903		return -EINVAL;
 904
 905	if (flags & MPOL_F_MEMS_ALLOWED) {
 906		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 907			return -EINVAL;
 908		*policy = 0;	/* just so it's initialized */
 909		task_lock(current);
 910		*nmask  = cpuset_current_mems_allowed;
 911		task_unlock(current);
 912		return 0;
 913	}
 914
 915	if (flags & MPOL_F_ADDR) {
 916		/*
 917		 * Do NOT fall back to task policy if the
 918		 * vma/shared policy at addr is NULL.  We
 919		 * want to return MPOL_DEFAULT in this case.
 920		 */
 921		down_read(&mm->mmap_sem);
 922		vma = find_vma_intersection(mm, addr, addr+1);
 923		if (!vma) {
 924			up_read(&mm->mmap_sem);
 925			return -EFAULT;
 926		}
 927		if (vma->vm_ops && vma->vm_ops->get_policy)
 928			pol = vma->vm_ops->get_policy(vma, addr);
 929		else
 930			pol = vma->vm_policy;
 931	} else if (addr)
 932		return -EINVAL;
 933
 934	if (!pol)
 935		pol = &default_policy;	/* indicates default behavior */
 936
 937	if (flags & MPOL_F_NODE) {
 938		if (flags & MPOL_F_ADDR) {
 939			err = lookup_node(mm, addr);
 940			if (err < 0)
 941				goto out;
 942			*policy = err;
 943		} else if (pol == current->mempolicy &&
 944				pol->mode == MPOL_INTERLEAVE) {
 945			*policy = current->il_next;
 946		} else {
 947			err = -EINVAL;
 948			goto out;
 949		}
 950	} else {
 951		*policy = pol == &default_policy ? MPOL_DEFAULT :
 952						pol->mode;
 953		/*
 954		 * Internal mempolicy flags must be masked off before exposing
 955		 * the policy to userspace.
 956		 */
 957		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 958	}
 959
 960	if (vma) {
 961		up_read(&current->mm->mmap_sem);
 962		vma = NULL;
 963	}
 964
 965	err = 0;
 966	if (nmask) {
 967		if (mpol_store_user_nodemask(pol)) {
 968			*nmask = pol->w.user_nodemask;
 969		} else {
 970			task_lock(current);
 971			get_policy_nodemask(pol, nmask);
 972			task_unlock(current);
 973		}
 974	}
 975
 976 out:
 977	mpol_cond_put(pol);
 978	if (vma)
 979		up_read(&current->mm->mmap_sem);
 980	return err;
 981}
 982
 983#ifdef CONFIG_MIGRATION
 984/*
 985 * page migration
 986 */
 987static void migrate_page_add(struct page *page, struct list_head *pagelist,
 988				unsigned long flags)
 989{
 990	/*
 991	 * Avoid migrating a page that is shared with others.
 992	 */
 993	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
 994		if (!isolate_lru_page(page)) {
 995			list_add_tail(&page->lru, pagelist);
 996			inc_zone_page_state(page, NR_ISOLATED_ANON +
 997					    page_is_file_cache(page));
 998		}
 999	}
1000}
1001
1002static struct page *new_node_page(struct page *page, unsigned long node, int **x)
1003{
1004	if (PageHuge(page))
1005		return alloc_huge_page_node(page_hstate(compound_head(page)),
1006					node);
1007	else
1008		return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
1009}
1010
1011/*
1012 * Migrate pages from one node to a target node.
1013 * Returns error or the number of pages not migrated.
1014 */
1015static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1016			   int flags)
1017{
1018	nodemask_t nmask;
1019	LIST_HEAD(pagelist);
1020	int err = 0;
1021
1022	nodes_clear(nmask);
1023	node_set(source, nmask);
1024
1025	/*
1026	 * This does not "check" the range but isolates all pages that
1027	 * need migration.  Between passing in the full user address
1028	 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1029	 */
1030	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1031	queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1032			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1033
1034	if (!list_empty(&pagelist)) {
1035		err = migrate_pages(&pagelist, new_node_page, dest,
1036					MIGRATE_SYNC, MR_SYSCALL);
1037		if (err)
1038			putback_movable_pages(&pagelist);
1039	}
1040
1041	return err;
1042}
1043
1044/*
1045 * Move pages between the two nodesets so as to preserve the physical
1046 * layout as much as possible.
1047 *
1048 * Returns the number of page that could not be moved.
1049 */
1050int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1051		     const nodemask_t *to, int flags)
1052{
1053	int busy = 0;
1054	int err;
1055	nodemask_t tmp;
1056
1057	err = migrate_prep();
1058	if (err)
1059		return err;
1060
1061	down_read(&mm->mmap_sem);
1062
1063	err = migrate_vmas(mm, from, to, flags);
1064	if (err)
1065		goto out;
1066
1067	/*
1068	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1069	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
1070	 * bit in 'tmp', and return that <source, dest> pair for migration.
1071	 * The pair of nodemasks 'to' and 'from' define the map.
1072	 *
1073	 * If no pair of bits is found that way, fallback to picking some
1074	 * pair of 'source' and 'dest' bits that are not the same.  If the
1075	 * 'source' and 'dest' bits are the same, this represents a node
1076	 * that will be migrating to itself, so no pages need move.
1077	 *
1078	 * If no bits are left in 'tmp', or if all remaining bits left
1079	 * in 'tmp' correspond to the same bit in 'to', return false
1080	 * (nothing left to migrate).
1081	 *
1082	 * This lets us pick a pair of nodes to migrate between, such that
1083	 * if possible the dest node is not already occupied by some other
1084	 * source node, minimizing the risk of overloading the memory on a
1085	 * node that would happen if we migrated incoming memory to a node
1086	 * before migrating outgoing memory source that same node.
1087	 *
1088	 * A single scan of tmp is sufficient.  As we go, we remember the
1089	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1090	 * that not only moved, but what's better, moved to an empty slot
1091	 * (d is not set in tmp), then we break out then, with that pair.
1092	 * Otherwise when we finish scanning from_tmp, we at least have the
1093	 * most recent <s, d> pair that moved.  If we get all the way through
1094	 * the scan of tmp without finding any node that moved, much less
1095	 * moved to an empty node, then there is nothing left worth migrating.
1096	 */
1097
1098	tmp = *from;
1099	while (!nodes_empty(tmp)) {
1100		int s,d;
1101		int source = NUMA_NO_NODE;
1102		int dest = 0;
1103
1104		for_each_node_mask(s, tmp) {
1105
1106			/*
1107			 * do_migrate_pages() tries to maintain the relative
1108			 * node relationship of the pages established between
1109			 * threads and memory areas.
1110                         *
1111			 * However if the number of source nodes is not equal to
1112			 * the number of destination nodes we can not preserve
1113			 * this node relative relationship.  In that case, skip
1114			 * copying memory from a node that is in the destination
1115			 * mask.
1116			 *
1117			 * Example: [2,3,4] -> [3,4,5] moves everything.
1118			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1119			 */
1120
1121			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1122						(node_isset(s, *to)))
1123				continue;
1124
1125			d = node_remap(s, *from, *to);
1126			if (s == d)
1127				continue;
1128
1129			source = s;	/* Node moved. Memorize */
1130			dest = d;
1131
1132			/* dest not in remaining from nodes? */
1133			if (!node_isset(dest, tmp))
1134				break;
1135		}
1136		if (source == NUMA_NO_NODE)
1137			break;
1138
1139		node_clear(source, tmp);
1140		err = migrate_to_node(mm, source, dest, flags);
1141		if (err > 0)
1142			busy += err;
1143		if (err < 0)
1144			break;
1145	}
1146out:
1147	up_read(&mm->mmap_sem);
1148	if (err < 0)
1149		return err;
1150	return busy;
1151
1152}
1153
1154/*
1155 * Allocate a new page for page migration based on vma policy.
1156 * Start assuming that page is mapped by vma pointed to by @private.
1157 * Search forward from there, if not.  N.B., this assumes that the
1158 * list of pages handed to migrate_pages()--which is how we get here--
1159 * is in virtual address order.
1160 */
1161static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1162{
1163	struct vm_area_struct *vma = (struct vm_area_struct *)private;
1164	unsigned long uninitialized_var(address);
1165
1166	while (vma) {
1167		address = page_address_in_vma(page, vma);
1168		if (address != -EFAULT)
1169			break;
1170		vma = vma->vm_next;
1171	}
1172
1173	if (PageHuge(page)) {
1174		BUG_ON(!vma);
1175		return alloc_huge_page_noerr(vma, address, 1);
1176	}
1177	/*
1178	 * if !vma, alloc_page_vma() will use task or system default policy
1179	 */
1180	return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1181}
1182#else
1183
1184static void migrate_page_add(struct page *page, struct list_head *pagelist,
1185				unsigned long flags)
1186{
1187}
1188
1189int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1190		     const nodemask_t *to, int flags)
1191{
1192	return -ENOSYS;
1193}
1194
1195static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1196{
1197	return NULL;
1198}
1199#endif
1200
1201static long do_mbind(unsigned long start, unsigned long len,
1202		     unsigned short mode, unsigned short mode_flags,
1203		     nodemask_t *nmask, unsigned long flags)
1204{
1205	struct vm_area_struct *vma;
1206	struct mm_struct *mm = current->mm;
1207	struct mempolicy *new;
1208	unsigned long end;
1209	int err;
1210	LIST_HEAD(pagelist);
1211
1212	if (flags & ~(unsigned long)MPOL_MF_VALID)
1213		return -EINVAL;
1214	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1215		return -EPERM;
1216
1217	if (start & ~PAGE_MASK)
1218		return -EINVAL;
1219
1220	if (mode == MPOL_DEFAULT)
1221		flags &= ~MPOL_MF_STRICT;
1222
1223	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1224	end = start + len;
1225
1226	if (end < start)
1227		return -EINVAL;
1228	if (end == start)
1229		return 0;
1230
1231	new = mpol_new(mode, mode_flags, nmask);
1232	if (IS_ERR(new))
1233		return PTR_ERR(new);
1234
1235	if (flags & MPOL_MF_LAZY)
1236		new->flags |= MPOL_F_MOF;
1237
1238	/*
1239	 * If we are using the default policy then operation
1240	 * on discontinuous address spaces is okay after all
1241	 */
1242	if (!new)
1243		flags |= MPOL_MF_DISCONTIG_OK;
1244
1245	pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1246		 start, start + len, mode, mode_flags,
1247		 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1248
1249	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1250
1251		err = migrate_prep();
1252		if (err)
1253			goto mpol_out;
1254	}
1255	{
1256		NODEMASK_SCRATCH(scratch);
1257		if (scratch) {
1258			down_write(&mm->mmap_sem);
1259			task_lock(current);
1260			err = mpol_set_nodemask(new, nmask, scratch);
1261			task_unlock(current);
1262			if (err)
1263				up_write(&mm->mmap_sem);
1264		} else
1265			err = -ENOMEM;
1266		NODEMASK_SCRATCH_FREE(scratch);
1267	}
1268	if (err)
1269		goto mpol_out;
1270
1271	vma = queue_pages_range(mm, start, end, nmask,
1272			  flags | MPOL_MF_INVERT, &pagelist);
1273
1274	err = PTR_ERR(vma);	/* maybe ... */
1275	if (!IS_ERR(vma))
1276		err = mbind_range(mm, start, end, new);
1277
1278	if (!err) {
1279		int nr_failed = 0;
1280
1281		if (!list_empty(&pagelist)) {
1282			WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1283			nr_failed = migrate_pages(&pagelist, new_vma_page,
1284					(unsigned long)vma,
1285					MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1286			if (nr_failed)
1287				putback_movable_pages(&pagelist);
1288		}
1289
1290		if (nr_failed && (flags & MPOL_MF_STRICT))
1291			err = -EIO;
1292	} else
1293		putback_movable_pages(&pagelist);
1294
1295	up_write(&mm->mmap_sem);
1296 mpol_out:
1297	mpol_put(new);
1298	return err;
1299}
1300
1301/*
1302 * User space interface with variable sized bitmaps for nodelists.
1303 */
1304
1305/* Copy a node mask from user space. */
1306static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1307		     unsigned long maxnode)
1308{
1309	unsigned long k;
1310	unsigned long nlongs;
1311	unsigned long endmask;
1312
1313	--maxnode;
1314	nodes_clear(*nodes);
1315	if (maxnode == 0 || !nmask)
1316		return 0;
1317	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1318		return -EINVAL;
1319
1320	nlongs = BITS_TO_LONGS(maxnode);
1321	if ((maxnode % BITS_PER_LONG) == 0)
1322		endmask = ~0UL;
1323	else
1324		endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1325
1326	/* When the user specified more nodes than supported just check
1327	   if the non supported part is all zero. */
1328	if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1329		if (nlongs > PAGE_SIZE/sizeof(long))
1330			return -EINVAL;
1331		for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1332			unsigned long t;
1333			if (get_user(t, nmask + k))
1334				return -EFAULT;
1335			if (k == nlongs - 1) {
1336				if (t & endmask)
1337					return -EINVAL;
1338			} else if (t)
1339				return -EINVAL;
1340		}
1341		nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1342		endmask = ~0UL;
1343	}
1344
1345	if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1346		return -EFAULT;
1347	nodes_addr(*nodes)[nlongs-1] &= endmask;
1348	return 0;
1349}
1350
1351/* Copy a kernel node mask to user space */
1352static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1353			      nodemask_t *nodes)
1354{
1355	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1356	const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1357
1358	if (copy > nbytes) {
1359		if (copy > PAGE_SIZE)
1360			return -EINVAL;
1361		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1362			return -EFAULT;
1363		copy = nbytes;
1364	}
1365	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1366}
1367
1368SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1369		unsigned long, mode, unsigned long __user *, nmask,
1370		unsigned long, maxnode, unsigned, flags)
1371{
1372	nodemask_t nodes;
1373	int err;
1374	unsigned short mode_flags;
1375
1376	mode_flags = mode & MPOL_MODE_FLAGS;
1377	mode &= ~MPOL_MODE_FLAGS;
1378	if (mode >= MPOL_MAX)
1379		return -EINVAL;
1380	if ((mode_flags & MPOL_F_STATIC_NODES) &&
1381	    (mode_flags & MPOL_F_RELATIVE_NODES))
1382		return -EINVAL;
1383	err = get_nodes(&nodes, nmask, maxnode);
1384	if (err)
1385		return err;
1386	return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1387}
1388
1389/* Set the process memory policy */
1390SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1391		unsigned long, maxnode)
1392{
1393	int err;
1394	nodemask_t nodes;
1395	unsigned short flags;
1396
1397	flags = mode & MPOL_MODE_FLAGS;
1398	mode &= ~MPOL_MODE_FLAGS;
1399	if ((unsigned int)mode >= MPOL_MAX)
1400		return -EINVAL;
1401	if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1402		return -EINVAL;
1403	err = get_nodes(&nodes, nmask, maxnode);
1404	if (err)
1405		return err;
1406	return do_set_mempolicy(mode, flags, &nodes);
1407}
1408
1409SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1410		const unsigned long __user *, old_nodes,
1411		const unsigned long __user *, new_nodes)
1412{
1413	const struct cred *cred = current_cred(), *tcred;
1414	struct mm_struct *mm = NULL;
1415	struct task_struct *task;
1416	nodemask_t task_nodes;
1417	int err;
1418	nodemask_t *old;
1419	nodemask_t *new;
1420	NODEMASK_SCRATCH(scratch);
1421
1422	if (!scratch)
1423		return -ENOMEM;
1424
1425	old = &scratch->mask1;
1426	new = &scratch->mask2;
1427
1428	err = get_nodes(old, old_nodes, maxnode);
1429	if (err)
1430		goto out;
1431
1432	err = get_nodes(new, new_nodes, maxnode);
1433	if (err)
1434		goto out;
1435
1436	/* Find the mm_struct */
1437	rcu_read_lock();
1438	task = pid ? find_task_by_vpid(pid) : current;
1439	if (!task) {
1440		rcu_read_unlock();
1441		err = -ESRCH;
1442		goto out;
1443	}
1444	get_task_struct(task);
1445
1446	err = -EINVAL;
1447
1448	/*
1449	 * Check if this process has the right to modify the specified
1450	 * process. The right exists if the process has administrative
1451	 * capabilities, superuser privileges or the same
1452	 * userid as the target process.
1453	 */
1454	tcred = __task_cred(task);
1455	if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1456	    !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1457	    !capable(CAP_SYS_NICE)) {
1458		rcu_read_unlock();
1459		err = -EPERM;
1460		goto out_put;
1461	}
1462	rcu_read_unlock();
1463
1464	task_nodes = cpuset_mems_allowed(task);
1465	/* Is the user allowed to access the target nodes? */
1466	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1467		err = -EPERM;
1468		goto out_put;
1469	}
1470
1471	if (!nodes_subset(*new, node_states[N_MEMORY])) {
1472		err = -EINVAL;
1473		goto out_put;
1474	}
1475
1476	err = security_task_movememory(task);
1477	if (err)
1478		goto out_put;
1479
1480	mm = get_task_mm(task);
1481	put_task_struct(task);
1482
1483	if (!mm) {
1484		err = -EINVAL;
1485		goto out;
1486	}
1487
1488	err = do_migrate_pages(mm, old, new,
1489		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1490
1491	mmput(mm);
1492out:
1493	NODEMASK_SCRATCH_FREE(scratch);
1494
1495	return err;
1496
1497out_put:
1498	put_task_struct(task);
1499	goto out;
1500
1501}
1502
1503
1504/* Retrieve NUMA policy */
1505SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1506		unsigned long __user *, nmask, unsigned long, maxnode,
1507		unsigned long, addr, unsigned long, flags)
1508{
1509	int err;
1510	int uninitialized_var(pval);
1511	nodemask_t nodes;
1512
1513	if (nmask != NULL && maxnode < MAX_NUMNODES)
1514		return -EINVAL;
1515
1516	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1517
1518	if (err)
1519		return err;
1520
1521	if (policy && put_user(pval, policy))
1522		return -EFAULT;
1523
1524	if (nmask)
1525		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1526
1527	return err;
1528}
1529
1530#ifdef CONFIG_COMPAT
1531
1532COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1533		       compat_ulong_t __user *, nmask,
1534		       compat_ulong_t, maxnode,
1535		       compat_ulong_t, addr, compat_ulong_t, flags)
1536{
1537	long err;
1538	unsigned long __user *nm = NULL;
1539	unsigned long nr_bits, alloc_size;
1540	DECLARE_BITMAP(bm, MAX_NUMNODES);
1541
1542	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1543	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1544
1545	if (nmask)
1546		nm = compat_alloc_user_space(alloc_size);
1547
1548	err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1549
1550	if (!err && nmask) {
1551		unsigned long copy_size;
1552		copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1553		err = copy_from_user(bm, nm, copy_size);
1554		/* ensure entire bitmap is zeroed */
1555		err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1556		err |= compat_put_bitmap(nmask, bm, nr_bits);
1557	}
1558
1559	return err;
1560}
1561
1562COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1563		       compat_ulong_t, maxnode)
1564{
1565	long err = 0;
1566	unsigned long __user *nm = NULL;
1567	unsigned long nr_bits, alloc_size;
1568	DECLARE_BITMAP(bm, MAX_NUMNODES);
1569
1570	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1571	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1572
1573	if (nmask) {
1574		err = compat_get_bitmap(bm, nmask, nr_bits);
1575		nm = compat_alloc_user_space(alloc_size);
1576		err |= copy_to_user(nm, bm, alloc_size);
1577	}
1578
1579	if (err)
1580		return -EFAULT;
1581
1582	return sys_set_mempolicy(mode, nm, nr_bits+1);
1583}
1584
1585COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1586		       compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1587		       compat_ulong_t, maxnode, compat_ulong_t, flags)
1588{
1589	long err = 0;
1590	unsigned long __user *nm = NULL;
1591	unsigned long nr_bits, alloc_size;
1592	nodemask_t bm;
1593
1594	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1595	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1596
1597	if (nmask) {
1598		err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1599		nm = compat_alloc_user_space(alloc_size);
1600		err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1601	}
1602
1603	if (err)
1604		return -EFAULT;
1605
1606	return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1607}
1608
1609#endif
1610
1611/*
1612 * get_vma_policy(@task, @vma, @addr)
1613 * @task - task for fallback if vma policy == default
1614 * @vma   - virtual memory area whose policy is sought
1615 * @addr  - address in @vma for shared policy lookup
1616 *
1617 * Returns effective policy for a VMA at specified address.
1618 * Falls back to @task or system default policy, as necessary.
1619 * Current or other task's task mempolicy and non-shared vma policies must be
1620 * protected by task_lock(task) by the caller.
1621 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1622 * count--added by the get_policy() vm_op, as appropriate--to protect against
1623 * freeing by another task.  It is the caller's responsibility to free the
1624 * extra reference for shared policies.
1625 */
1626struct mempolicy *get_vma_policy(struct task_struct *task,
1627		struct vm_area_struct *vma, unsigned long addr)
1628{
1629	struct mempolicy *pol = get_task_policy(task);
1630
1631	if (vma) {
1632		if (vma->vm_ops && vma->vm_ops->get_policy) {
1633			struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1634									addr);
1635			if (vpol)
1636				pol = vpol;
1637		} else if (vma->vm_policy) {
1638			pol = vma->vm_policy;
1639
1640			/*
1641			 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1642			 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1643			 * count on these policies which will be dropped by
1644			 * mpol_cond_put() later
1645			 */
1646			if (mpol_needs_cond_ref(pol))
1647				mpol_get(pol);
1648		}
1649	}
1650	if (!pol)
1651		pol = &default_policy;
1652	return pol;
1653}
1654
1655bool vma_policy_mof(struct task_struct *task, struct vm_area_struct *vma)
1656{
1657	struct mempolicy *pol = get_task_policy(task);
1658	if (vma) {
1659		if (vma->vm_ops && vma->vm_ops->get_policy) {
1660			bool ret = false;
1661
1662			pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1663			if (pol && (pol->flags & MPOL_F_MOF))
1664				ret = true;
1665			mpol_cond_put(pol);
1666
1667			return ret;
1668		} else if (vma->vm_policy) {
1669			pol = vma->vm_policy;
1670		}
1671	}
1672
1673	if (!pol)
1674		return default_policy.flags & MPOL_F_MOF;
1675
1676	return pol->flags & MPOL_F_MOF;
1677}
1678
1679static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1680{
1681	enum zone_type dynamic_policy_zone = policy_zone;
1682
1683	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1684
1685	/*
1686	 * if policy->v.nodes has movable memory only,
1687	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1688	 *
1689	 * policy->v.nodes is intersect with node_states[N_MEMORY].
1690	 * so if the following test faile, it implies
1691	 * policy->v.nodes has movable memory only.
1692	 */
1693	if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1694		dynamic_policy_zone = ZONE_MOVABLE;
1695
1696	return zone >= dynamic_policy_zone;
1697}
1698
1699/*
1700 * Return a nodemask representing a mempolicy for filtering nodes for
1701 * page allocation
1702 */
1703static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1704{
1705	/* Lower zones don't get a nodemask applied for MPOL_BIND */
1706	if (unlikely(policy->mode == MPOL_BIND) &&
1707			apply_policy_zone(policy, gfp_zone(gfp)) &&
1708			cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1709		return &policy->v.nodes;
1710
1711	return NULL;
1712}
1713
1714/* Return a zonelist indicated by gfp for node representing a mempolicy */
1715static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1716	int nd)
1717{
1718	switch (policy->mode) {
1719	case MPOL_PREFERRED:
1720		if (!(policy->flags & MPOL_F_LOCAL))
1721			nd = policy->v.preferred_node;
1722		break;
1723	case MPOL_BIND:
1724		/*
1725		 * Normally, MPOL_BIND allocations are node-local within the
1726		 * allowed nodemask.  However, if __GFP_THISNODE is set and the
1727		 * current node isn't part of the mask, we use the zonelist for
1728		 * the first node in the mask instead.
1729		 */
1730		if (unlikely(gfp & __GFP_THISNODE) &&
1731				unlikely(!node_isset(nd, policy->v.nodes)))
1732			nd = first_node(policy->v.nodes);
1733		break;
1734	default:
1735		BUG();
1736	}
1737	return node_zonelist(nd, gfp);
1738}
1739
1740/* Do dynamic interleaving for a process */
1741static unsigned interleave_nodes(struct mempolicy *policy)
1742{
1743	unsigned nid, next;
1744	struct task_struct *me = current;
1745
1746	nid = me->il_next;
1747	next = next_node(nid, policy->v.nodes);
1748	if (next >= MAX_NUMNODES)
1749		next = first_node(policy->v.nodes);
1750	if (next < MAX_NUMNODES)
1751		me->il_next = next;
1752	return nid;
1753}
1754
1755/*
1756 * Depending on the memory policy provide a node from which to allocate the
1757 * next slab entry.
1758 */
1759unsigned int mempolicy_slab_node(void)
1760{
1761	struct mempolicy *policy;
1762	int node = numa_mem_id();
1763
1764	if (in_interrupt())
1765		return node;
1766
1767	policy = current->mempolicy;
1768	if (!policy || policy->flags & MPOL_F_LOCAL)
1769		return node;
1770
1771	switch (policy->mode) {
1772	case MPOL_PREFERRED:
1773		/*
1774		 * handled MPOL_F_LOCAL above
1775		 */
1776		return policy->v.preferred_node;
1777
1778	case MPOL_INTERLEAVE:
1779		return interleave_nodes(policy);
1780
1781	case MPOL_BIND: {
1782		/*
1783		 * Follow bind policy behavior and start allocation at the
1784		 * first node.
1785		 */
1786		struct zonelist *zonelist;
1787		struct zone *zone;
1788		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1789		zonelist = &NODE_DATA(node)->node_zonelists[0];
1790		(void)first_zones_zonelist(zonelist, highest_zoneidx,
1791							&policy->v.nodes,
1792							&zone);
1793		return zone ? zone->node : node;
1794	}
1795
1796	default:
1797		BUG();
1798	}
1799}
1800
1801/* Do static interleaving for a VMA with known offset. */
1802static unsigned offset_il_node(struct mempolicy *pol,
1803		struct vm_area_struct *vma, unsigned long off)
1804{
1805	unsigned nnodes = nodes_weight(pol->v.nodes);
1806	unsigned target;
1807	int c;
1808	int nid = NUMA_NO_NODE;
1809
1810	if (!nnodes)
1811		return numa_node_id();
1812	target = (unsigned int)off % nnodes;
1813	c = 0;
1814	do {
1815		nid = next_node(nid, pol->v.nodes);
1816		c++;
1817	} while (c <= target);
1818	return nid;
1819}
1820
1821/* Determine a node number for interleave */
1822static inline unsigned interleave_nid(struct mempolicy *pol,
1823		 struct vm_area_struct *vma, unsigned long addr, int shift)
1824{
1825	if (vma) {
1826		unsigned long off;
1827
1828		/*
1829		 * for small pages, there is no difference between
1830		 * shift and PAGE_SHIFT, so the bit-shift is safe.
1831		 * for huge pages, since vm_pgoff is in units of small
1832		 * pages, we need to shift off the always 0 bits to get
1833		 * a useful offset.
1834		 */
1835		BUG_ON(shift < PAGE_SHIFT);
1836		off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1837		off += (addr - vma->vm_start) >> shift;
1838		return offset_il_node(pol, vma, off);
1839	} else
1840		return interleave_nodes(pol);
1841}
1842
1843/*
1844 * Return the bit number of a random bit set in the nodemask.
1845 * (returns NUMA_NO_NODE if nodemask is empty)
1846 */
1847int node_random(const nodemask_t *maskp)
1848{
1849	int w, bit = NUMA_NO_NODE;
1850
1851	w = nodes_weight(*maskp);
1852	if (w)
1853		bit = bitmap_ord_to_pos(maskp->bits,
1854			get_random_int() % w, MAX_NUMNODES);
1855	return bit;
1856}
1857
1858#ifdef CONFIG_HUGETLBFS
1859/*
1860 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1861 * @vma = virtual memory area whose policy is sought
1862 * @addr = address in @vma for shared policy lookup and interleave policy
1863 * @gfp_flags = for requested zone
1864 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1865 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1866 *
1867 * Returns a zonelist suitable for a huge page allocation and a pointer
1868 * to the struct mempolicy for conditional unref after allocation.
1869 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1870 * @nodemask for filtering the zonelist.
1871 *
1872 * Must be protected by read_mems_allowed_begin()
1873 */
1874struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1875				gfp_t gfp_flags, struct mempolicy **mpol,
1876				nodemask_t **nodemask)
1877{
1878	struct zonelist *zl;
1879
1880	*mpol = get_vma_policy(current, vma, addr);
1881	*nodemask = NULL;	/* assume !MPOL_BIND */
1882
1883	if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1884		zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1885				huge_page_shift(hstate_vma(vma))), gfp_flags);
1886	} else {
1887		zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1888		if ((*mpol)->mode == MPOL_BIND)
1889			*nodemask = &(*mpol)->v.nodes;
1890	}
1891	return zl;
1892}
1893
1894/*
1895 * init_nodemask_of_mempolicy
1896 *
1897 * If the current task's mempolicy is "default" [NULL], return 'false'
1898 * to indicate default policy.  Otherwise, extract the policy nodemask
1899 * for 'bind' or 'interleave' policy into the argument nodemask, or
1900 * initialize the argument nodemask to contain the single node for
1901 * 'preferred' or 'local' policy and return 'true' to indicate presence
1902 * of non-default mempolicy.
1903 *
1904 * We don't bother with reference counting the mempolicy [mpol_get/put]
1905 * because the current task is examining it's own mempolicy and a task's
1906 * mempolicy is only ever changed by the task itself.
1907 *
1908 * N.B., it is the caller's responsibility to free a returned nodemask.
1909 */
1910bool init_nodemask_of_mempolicy(nodemask_t *mask)
1911{
1912	struct mempolicy *mempolicy;
1913	int nid;
1914
1915	if (!(mask && current->mempolicy))
1916		return false;
1917
1918	task_lock(current);
1919	mempolicy = current->mempolicy;
1920	switch (mempolicy->mode) {
1921	case MPOL_PREFERRED:
1922		if (mempolicy->flags & MPOL_F_LOCAL)
1923			nid = numa_node_id();
1924		else
1925			nid = mempolicy->v.preferred_node;
1926		init_nodemask_of_node(mask, nid);
1927		break;
1928
1929	case MPOL_BIND:
1930		/* Fall through */
1931	case MPOL_INTERLEAVE:
1932		*mask =  mempolicy->v.nodes;
1933		break;
1934
1935	default:
1936		BUG();
1937	}
1938	task_unlock(current);
1939
1940	return true;
1941}
1942#endif
1943
1944/*
1945 * mempolicy_nodemask_intersects
1946 *
1947 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1948 * policy.  Otherwise, check for intersection between mask and the policy
1949 * nodemask for 'bind' or 'interleave' policy.  For 'perferred' or 'local'
1950 * policy, always return true since it may allocate elsewhere on fallback.
1951 *
1952 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1953 */
1954bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1955					const nodemask_t *mask)
1956{
1957	struct mempolicy *mempolicy;
1958	bool ret = true;
1959
1960	if (!mask)
1961		return ret;
1962	task_lock(tsk);
1963	mempolicy = tsk->mempolicy;
1964	if (!mempolicy)
1965		goto out;
1966
1967	switch (mempolicy->mode) {
1968	case MPOL_PREFERRED:
1969		/*
1970		 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1971		 * allocate from, they may fallback to other nodes when oom.
1972		 * Thus, it's possible for tsk to have allocated memory from
1973		 * nodes in mask.
1974		 */
1975		break;
1976	case MPOL_BIND:
1977	case MPOL_INTERLEAVE:
1978		ret = nodes_intersects(mempolicy->v.nodes, *mask);
1979		break;
1980	default:
1981		BUG();
1982	}
1983out:
1984	task_unlock(tsk);
1985	return ret;
1986}
1987
1988/* Allocate a page in interleaved policy.
1989   Own path because it needs to do special accounting. */
1990static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1991					unsigned nid)
1992{
1993	struct zonelist *zl;
1994	struct page *page;
1995
1996	zl = node_zonelist(nid, gfp);
1997	page = __alloc_pages(gfp, order, zl);
1998	if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1999		inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
2000	return page;
2001}
2002
2003/**
2004 * 	alloc_pages_vma	- Allocate a page for a VMA.
2005 *
2006 * 	@gfp:
2007 *      %GFP_USER    user allocation.
2008 *      %GFP_KERNEL  kernel allocations,
2009 *      %GFP_HIGHMEM highmem/user allocations,
2010 *      %GFP_FS      allocation should not call back into a file system.
2011 *      %GFP_ATOMIC  don't sleep.
2012 *
2013 *	@order:Order of the GFP allocation.
2014 * 	@vma:  Pointer to VMA or NULL if not available.
2015 *	@addr: Virtual Address of the allocation. Must be inside the VMA.
2016 *
2017 * 	This function allocates a page from the kernel page pool and applies
2018 *	a NUMA policy associated with the VMA or the current process.
2019 *	When VMA is not NULL caller must hold down_read on the mmap_sem of the
2020 *	mm_struct of the VMA to prevent it from going away. Should be used for
2021 *	all allocations for pages that will be mapped into
2022 * 	user space. Returns NULL when no page can be allocated.
2023 *
2024 *	Should be called with the mm_sem of the vma hold.
2025 */
2026struct page *
2027alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2028		unsigned long addr, int node)
2029{
2030	struct mempolicy *pol;
2031	struct page *page;
2032	unsigned int cpuset_mems_cookie;
2033
2034retry_cpuset:
2035	pol = get_vma_policy(current, vma, addr);
2036	cpuset_mems_cookie = read_mems_allowed_begin();
2037
2038	if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
2039		unsigned nid;
2040
2041		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2042		mpol_cond_put(pol);
2043		page = alloc_page_interleave(gfp, order, nid);
2044		if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2045			goto retry_cpuset;
2046
2047		return page;
2048	}
2049	page = __alloc_pages_nodemask(gfp, order,
2050				      policy_zonelist(gfp, pol, node),
2051				      policy_nodemask(gfp, pol));
2052	if (unlikely(mpol_needs_cond_ref(pol)))
2053		__mpol_put(pol);
2054	if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2055		goto retry_cpuset;
2056	return page;
2057}
2058
2059/**
2060 * 	alloc_pages_current - Allocate pages.
2061 *
2062 *	@gfp:
2063 *		%GFP_USER   user allocation,
2064 *      	%GFP_KERNEL kernel allocation,
2065 *      	%GFP_HIGHMEM highmem allocation,
2066 *      	%GFP_FS     don't call back into a file system.
2067 *      	%GFP_ATOMIC don't sleep.
2068 *	@order: Power of two of allocation size in pages. 0 is a single page.
2069 *
2070 *	Allocate a page from the kernel page pool.  When not in
2071 *	interrupt context and apply the current process NUMA policy.
2072 *	Returns NULL when no page can be allocated.
2073 *
2074 *	Don't call cpuset_update_task_memory_state() unless
2075 *	1) it's ok to take cpuset_sem (can WAIT), and
2076 *	2) allocating for current task (not interrupt).
2077 */
2078struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2079{
2080	struct mempolicy *pol = get_task_policy(current);
2081	struct page *page;
2082	unsigned int cpuset_mems_cookie;
2083
2084	if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
2085		pol = &default_policy;
2086
2087retry_cpuset:
2088	cpuset_mems_cookie = read_mems_allowed_begin();
2089
2090	/*
2091	 * No reference counting needed for current->mempolicy
2092	 * nor system default_policy
2093	 */
2094	if (pol->mode == MPOL_INTERLEAVE)
2095		page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2096	else
2097		page = __alloc_pages_nodemask(gfp, order,
2098				policy_zonelist(gfp, pol, numa_node_id()),
2099				policy_nodemask(gfp, pol));
2100
2101	if (unlikely(!page && read_mems_allowed_retry(cpuset_mems_cookie)))
2102		goto retry_cpuset;
2103
2104	return page;
2105}
2106EXPORT_SYMBOL(alloc_pages_current);
2107
2108int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2109{
2110	struct mempolicy *pol = mpol_dup(vma_policy(src));
2111
2112	if (IS_ERR(pol))
2113		return PTR_ERR(pol);
2114	dst->vm_policy = pol;
2115	return 0;
2116}
2117
2118/*
2119 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2120 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2121 * with the mems_allowed returned by cpuset_mems_allowed().  This
2122 * keeps mempolicies cpuset relative after its cpuset moves.  See
2123 * further kernel/cpuset.c update_nodemask().
2124 *
2125 * current's mempolicy may be rebinded by the other task(the task that changes
2126 * cpuset's mems), so we needn't do rebind work for current task.
2127 */
2128
2129/* Slow path of a mempolicy duplicate */
2130struct mempolicy *__mpol_dup(struct mempolicy *old)
2131{
2132	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2133
2134	if (!new)
2135		return ERR_PTR(-ENOMEM);
2136
2137	/* task's mempolicy is protected by alloc_lock */
2138	if (old == current->mempolicy) {
2139		task_lock(current);
2140		*new = *old;
2141		task_unlock(current);
2142	} else
2143		*new = *old;
2144
2145	rcu_read_lock();
2146	if (current_cpuset_is_being_rebound()) {
2147		nodemask_t mems = cpuset_mems_allowed(current);
2148		if (new->flags & MPOL_F_REBINDING)
2149			mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2150		else
2151			mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2152	}
2153	rcu_read_unlock();
2154	atomic_set(&new->refcnt, 1);
2155	return new;
2156}
2157
2158/* Slow path of a mempolicy comparison */
2159bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2160{
2161	if (!a || !b)
2162		return false;
2163	if (a->mode != b->mode)
2164		return false;
2165	if (a->flags != b->flags)
2166		return false;
2167	if (mpol_store_user_nodemask(a))
2168		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2169			return false;
2170
2171	switch (a->mode) {
2172	case MPOL_BIND:
2173		/* Fall through */
2174	case MPOL_INTERLEAVE:
2175		return !!nodes_equal(a->v.nodes, b->v.nodes);
2176	case MPOL_PREFERRED:
2177		return a->v.preferred_node == b->v.preferred_node;
2178	default:
2179		BUG();
2180		return false;
2181	}
2182}
2183
2184/*
2185 * Shared memory backing store policy support.
2186 *
2187 * Remember policies even when nobody has shared memory mapped.
2188 * The policies are kept in Red-Black tree linked from the inode.
2189 * They are protected by the sp->lock spinlock, which should be held
2190 * for any accesses to the tree.
2191 */
2192
2193/* lookup first element intersecting start-end */
2194/* Caller holds sp->lock */
2195static struct sp_node *
2196sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2197{
2198	struct rb_node *n = sp->root.rb_node;
2199
2200	while (n) {
2201		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2202
2203		if (start >= p->end)
2204			n = n->rb_right;
2205		else if (end <= p->start)
2206			n = n->rb_left;
2207		else
2208			break;
2209	}
2210	if (!n)
2211		return NULL;
2212	for (;;) {
2213		struct sp_node *w = NULL;
2214		struct rb_node *prev = rb_prev(n);
2215		if (!prev)
2216			break;
2217		w = rb_entry(prev, struct sp_node, nd);
2218		if (w->end <= start)
2219			break;
2220		n = prev;
2221	}
2222	return rb_entry(n, struct sp_node, nd);
2223}
2224
2225/* Insert a new shared policy into the list. */
2226/* Caller holds sp->lock */
2227static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2228{
2229	struct rb_node **p = &sp->root.rb_node;
2230	struct rb_node *parent = NULL;
2231	struct sp_node *nd;
2232
2233	while (*p) {
2234		parent = *p;
2235		nd = rb_entry(parent, struct sp_node, nd);
2236		if (new->start < nd->start)
2237			p = &(*p)->rb_left;
2238		else if (new->end > nd->end)
2239			p = &(*p)->rb_right;
2240		else
2241			BUG();
2242	}
2243	rb_link_node(&new->nd, parent, p);
2244	rb_insert_color(&new->nd, &sp->root);
2245	pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2246		 new->policy ? new->policy->mode : 0);
2247}
2248
2249/* Find shared policy intersecting idx */
2250struct mempolicy *
2251mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2252{
2253	struct mempolicy *pol = NULL;
2254	struct sp_node *sn;
2255
2256	if (!sp->root.rb_node)
2257		return NULL;
2258	spin_lock(&sp->lock);
2259	sn = sp_lookup(sp, idx, idx+1);
2260	if (sn) {
2261		mpol_get(sn->policy);
2262		pol = sn->policy;
2263	}
2264	spin_unlock(&sp->lock);
2265	return pol;
2266}
2267
2268static void sp_free(struct sp_node *n)
2269{
2270	mpol_put(n->policy);
2271	kmem_cache_free(sn_cache, n);
2272}
2273
2274/**
2275 * mpol_misplaced - check whether current page node is valid in policy
2276 *
2277 * @page   - page to be checked
2278 * @vma    - vm area where page mapped
2279 * @addr   - virtual address where page mapped
2280 *
2281 * Lookup current policy node id for vma,addr and "compare to" page's
2282 * node id.
2283 *
2284 * Returns:
2285 *	-1	- not misplaced, page is in the right node
2286 *	node	- node id where the page should be
2287 *
2288 * Policy determination "mimics" alloc_page_vma().
2289 * Called from fault path where we know the vma and faulting address.
2290 */
2291int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2292{
2293	struct mempolicy *pol;
2294	struct zone *zone;
2295	int curnid = page_to_nid(page);
2296	unsigned long pgoff;
2297	int thiscpu = raw_smp_processor_id();
2298	int thisnid = cpu_to_node(thiscpu);
2299	int polnid = -1;
2300	int ret = -1;
2301
2302	BUG_ON(!vma);
2303
2304	pol = get_vma_policy(current, vma, addr);
2305	if (!(pol->flags & MPOL_F_MOF))
2306		goto out;
2307
2308	switch (pol->mode) {
2309	case MPOL_INTERLEAVE:
2310		BUG_ON(addr >= vma->vm_end);
2311		BUG_ON(addr < vma->vm_start);
2312
2313		pgoff = vma->vm_pgoff;
2314		pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2315		polnid = offset_il_node(pol, vma, pgoff);
2316		break;
2317
2318	case MPOL_PREFERRED:
2319		if (pol->flags & MPOL_F_LOCAL)
2320			polnid = numa_node_id();
2321		else
2322			polnid = pol->v.preferred_node;
2323		break;
2324
2325	case MPOL_BIND:
2326		/*
2327		 * allows binding to multiple nodes.
2328		 * use current page if in policy nodemask,
2329		 * else select nearest allowed node, if any.
2330		 * If no allowed nodes, use current [!misplaced].
2331		 */
2332		if (node_isset(curnid, pol->v.nodes))
2333			goto out;
2334		(void)first_zones_zonelist(
2335				node_zonelist(numa_node_id(), GFP_HIGHUSER),
2336				gfp_zone(GFP_HIGHUSER),
2337				&pol->v.nodes, &zone);
2338		polnid = zone->node;
2339		break;
2340
2341	default:
2342		BUG();
2343	}
2344
2345	/* Migrate the page towards the node whose CPU is referencing it */
2346	if (pol->flags & MPOL_F_MORON) {
2347		polnid = thisnid;
2348
2349		if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2350			goto out;
2351	}
2352
2353	if (curnid != polnid)
2354		ret = polnid;
2355out:
2356	mpol_cond_put(pol);
2357
2358	return ret;
2359}
2360
2361static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2362{
2363	pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2364	rb_erase(&n->nd, &sp->root);
2365	sp_free(n);
2366}
2367
2368static void sp_node_init(struct sp_node *node, unsigned long start,
2369			unsigned long end, struct mempolicy *pol)
2370{
2371	node->start = start;
2372	node->end = end;
2373	node->policy = pol;
2374}
2375
2376static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2377				struct mempolicy *pol)
2378{
2379	struct sp_node *n;
2380	struct mempolicy *newpol;
2381
2382	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2383	if (!n)
2384		return NULL;
2385
2386	newpol = mpol_dup(pol);
2387	if (IS_ERR(newpol)) {
2388		kmem_cache_free(sn_cache, n);
2389		return NULL;
2390	}
2391	newpol->flags |= MPOL_F_SHARED;
2392	sp_node_init(n, start, end, newpol);
2393
2394	return n;
2395}
2396
2397/* Replace a policy range. */
2398static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2399				 unsigned long end, struct sp_node *new)
2400{
2401	struct sp_node *n;
2402	struct sp_node *n_new = NULL;
2403	struct mempolicy *mpol_new = NULL;
2404	int ret = 0;
2405
2406restart:
2407	spin_lock(&sp->lock);
2408	n = sp_lookup(sp, start, end);
2409	/* Take care of old policies in the same range. */
2410	while (n && n->start < end) {
2411		struct rb_node *next = rb_next(&n->nd);
2412		if (n->start >= start) {
2413			if (n->end <= end)
2414				sp_delete(sp, n);
2415			else
2416				n->start = end;
2417		} else {
2418			/* Old policy spanning whole new range. */
2419			if (n->end > end) {
2420				if (!n_new)
2421					goto alloc_new;
2422
2423				*mpol_new = *n->policy;
2424				atomic_set(&mpol_new->refcnt, 1);
2425				sp_node_init(n_new, end, n->end, mpol_new);
2426				n->end = start;
2427				sp_insert(sp, n_new);
2428				n_new = NULL;
2429				mpol_new = NULL;
2430				break;
2431			} else
2432				n->end = start;
2433		}
2434		if (!next)
2435			break;
2436		n = rb_entry(next, struct sp_node, nd);
2437	}
2438	if (new)
2439		sp_insert(sp, new);
2440	spin_unlock(&sp->lock);
2441	ret = 0;
2442
2443err_out:
2444	if (mpol_new)
2445		mpol_put(mpol_new);
2446	if (n_new)
2447		kmem_cache_free(sn_cache, n_new);
2448
2449	return ret;
2450
2451alloc_new:
2452	spin_unlock(&sp->lock);
2453	ret = -ENOMEM;
2454	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2455	if (!n_new)
2456		goto err_out;
2457	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2458	if (!mpol_new)
2459		goto err_out;
2460	goto restart;
2461}
2462
2463/**
2464 * mpol_shared_policy_init - initialize shared policy for inode
2465 * @sp: pointer to inode shared policy
2466 * @mpol:  struct mempolicy to install
2467 *
2468 * Install non-NULL @mpol in inode's shared policy rb-tree.
2469 * On entry, the current task has a reference on a non-NULL @mpol.
2470 * This must be released on exit.
2471 * This is called at get_inode() calls and we can use GFP_KERNEL.
2472 */
2473void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2474{
2475	int ret;
2476
2477	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2478	spin_lock_init(&sp->lock);
2479
2480	if (mpol) {
2481		struct vm_area_struct pvma;
2482		struct mempolicy *new;
2483		NODEMASK_SCRATCH(scratch);
2484
2485		if (!scratch)
2486			goto put_mpol;
2487		/* contextualize the tmpfs mount point mempolicy */
2488		new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2489		if (IS_ERR(new))
2490			goto free_scratch; /* no valid nodemask intersection */
2491
2492		task_lock(current);
2493		ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2494		task_unlock(current);
2495		if (ret)
2496			goto put_new;
2497
2498		/* Create pseudo-vma that contains just the policy */
2499		memset(&pvma, 0, sizeof(struct vm_area_struct));
2500		pvma.vm_end = TASK_SIZE;	/* policy covers entire file */
2501		mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2502
2503put_new:
2504		mpol_put(new);			/* drop initial ref */
2505free_scratch:
2506		NODEMASK_SCRATCH_FREE(scratch);
2507put_mpol:
2508		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
2509	}
2510}
2511
2512int mpol_set_shared_policy(struct shared_policy *info,
2513			struct vm_area_struct *vma, struct mempolicy *npol)
2514{
2515	int err;
2516	struct sp_node *new = NULL;
2517	unsigned long sz = vma_pages(vma);
2518
2519	pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2520		 vma->vm_pgoff,
2521		 sz, npol ? npol->mode : -1,
2522		 npol ? npol->flags : -1,
2523		 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2524
2525	if (npol) {
2526		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2527		if (!new)
2528			return -ENOMEM;
2529	}
2530	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2531	if (err && new)
2532		sp_free(new);
2533	return err;
2534}
2535
2536/* Free a backing policy store on inode delete. */
2537void mpol_free_shared_policy(struct shared_policy *p)
2538{
2539	struct sp_node *n;
2540	struct rb_node *next;
2541
2542	if (!p->root.rb_node)
2543		return;
2544	spin_lock(&p->lock);
2545	next = rb_first(&p->root);
2546	while (next) {
2547		n = rb_entry(next, struct sp_node, nd);
2548		next = rb_next(&n->nd);
2549		sp_delete(p, n);
2550	}
2551	spin_unlock(&p->lock);
2552}
2553
2554#ifdef CONFIG_NUMA_BALANCING
2555static int __initdata numabalancing_override;
2556
2557static void __init check_numabalancing_enable(void)
2558{
2559	bool numabalancing_default = false;
2560
2561	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2562		numabalancing_default = true;
2563
2564	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2565	if (numabalancing_override)
2566		set_numabalancing_state(numabalancing_override == 1);
2567
2568	if (nr_node_ids > 1 && !numabalancing_override) {
2569		pr_info("%s automatic NUMA balancing. "
2570			"Configure with numa_balancing= or the "
2571			"kernel.numa_balancing sysctl",
2572			numabalancing_default ? "Enabling" : "Disabling");
2573		set_numabalancing_state(numabalancing_default);
2574	}
2575}
2576
2577static int __init setup_numabalancing(char *str)
2578{
2579	int ret = 0;
2580	if (!str)
2581		goto out;
2582
2583	if (!strcmp(str, "enable")) {
2584		numabalancing_override = 1;
2585		ret = 1;
2586	} else if (!strcmp(str, "disable")) {
2587		numabalancing_override = -1;
2588		ret = 1;
2589	}
2590out:
2591	if (!ret)
2592		pr_warn("Unable to parse numa_balancing=\n");
2593
2594	return ret;
2595}
2596__setup("numa_balancing=", setup_numabalancing);
2597#else
2598static inline void __init check_numabalancing_enable(void)
2599{
2600}
2601#endif /* CONFIG_NUMA_BALANCING */
2602
2603/* assumes fs == KERNEL_DS */
2604void __init numa_policy_init(void)
2605{
2606	nodemask_t interleave_nodes;
2607	unsigned long largest = 0;
2608	int nid, prefer = 0;
2609
2610	policy_cache = kmem_cache_create("numa_policy",
2611					 sizeof(struct mempolicy),
2612					 0, SLAB_PANIC, NULL);
2613
2614	sn_cache = kmem_cache_create("shared_policy_node",
2615				     sizeof(struct sp_node),
2616				     0, SLAB_PANIC, NULL);
2617
2618	for_each_node(nid) {
2619		preferred_node_policy[nid] = (struct mempolicy) {
2620			.refcnt = ATOMIC_INIT(1),
2621			.mode = MPOL_PREFERRED,
2622			.flags = MPOL_F_MOF | MPOL_F_MORON,
2623			.v = { .preferred_node = nid, },
2624		};
2625	}
2626
2627	/*
2628	 * Set interleaving policy for system init. Interleaving is only
2629	 * enabled across suitably sized nodes (default is >= 16MB), or
2630	 * fall back to the largest node if they're all smaller.
2631	 */
2632	nodes_clear(interleave_nodes);
2633	for_each_node_state(nid, N_MEMORY) {
2634		unsigned long total_pages = node_present_pages(nid);
2635
2636		/* Preserve the largest node */
2637		if (largest < total_pages) {
2638			largest = total_pages;
2639			prefer = nid;
2640		}
2641
2642		/* Interleave this node? */
2643		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2644			node_set(nid, interleave_nodes);
2645	}
2646
2647	/* All too small, use the largest */
2648	if (unlikely(nodes_empty(interleave_nodes)))
2649		node_set(prefer, interleave_nodes);
2650
2651	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2652		printk("numa_policy_init: interleaving failed\n");
2653
2654	check_numabalancing_enable();
2655}
2656
2657/* Reset policy of current process to default */
2658void numa_default_policy(void)
2659{
2660	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2661}
2662
2663/*
2664 * Parse and format mempolicy from/to strings
2665 */
2666
2667/*
2668 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2669 */
2670static const char * const policy_modes[] =
2671{
2672	[MPOL_DEFAULT]    = "default",
2673	[MPOL_PREFERRED]  = "prefer",
2674	[MPOL_BIND]       = "bind",
2675	[MPOL_INTERLEAVE] = "interleave",
2676	[MPOL_LOCAL]      = "local",
2677};
2678
2679
2680#ifdef CONFIG_TMPFS
2681/**
2682 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2683 * @str:  string containing mempolicy to parse
2684 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2685 *
2686 * Format of input:
2687 *	<mode>[=<flags>][:<nodelist>]
2688 *
2689 * On success, returns 0, else 1
2690 */
2691int mpol_parse_str(char *str, struct mempolicy **mpol)
2692{
2693	struct mempolicy *new = NULL;
2694	unsigned short mode;
2695	unsigned short mode_flags;
2696	nodemask_t nodes;
2697	char *nodelist = strchr(str, ':');
2698	char *flags = strchr(str, '=');
2699	int err = 1;
2700
2701	if (nodelist) {
2702		/* NUL-terminate mode or flags string */
2703		*nodelist++ = '\0';
2704		if (nodelist_parse(nodelist, nodes))
2705			goto out;
2706		if (!nodes_subset(nodes, node_states[N_MEMORY]))
2707			goto out;
2708	} else
2709		nodes_clear(nodes);
2710
2711	if (flags)
2712		*flags++ = '\0';	/* terminate mode string */
2713
2714	for (mode = 0; mode < MPOL_MAX; mode++) {
2715		if (!strcmp(str, policy_modes[mode])) {
2716			break;
2717		}
2718	}
2719	if (mode >= MPOL_MAX)
2720		goto out;
2721
2722	switch (mode) {
2723	case MPOL_PREFERRED:
2724		/*
2725		 * Insist on a nodelist of one node only
2726		 */
2727		if (nodelist) {
2728			char *rest = nodelist;
2729			while (isdigit(*rest))
2730				rest++;
2731			if (*rest)
2732				goto out;
2733		}
2734		break;
2735	case MPOL_INTERLEAVE:
2736		/*
2737		 * Default to online nodes with memory if no nodelist
2738		 */
2739		if (!nodelist)
2740			nodes = node_states[N_MEMORY];
2741		break;
2742	case MPOL_LOCAL:
2743		/*
2744		 * Don't allow a nodelist;  mpol_new() checks flags
2745		 */
2746		if (nodelist)
2747			goto out;
2748		mode = MPOL_PREFERRED;
2749		break;
2750	case MPOL_DEFAULT:
2751		/*
2752		 * Insist on a empty nodelist
2753		 */
2754		if (!nodelist)
2755			err = 0;
2756		goto out;
2757	case MPOL_BIND:
2758		/*
2759		 * Insist on a nodelist
2760		 */
2761		if (!nodelist)
2762			goto out;
2763	}
2764
2765	mode_flags = 0;
2766	if (flags) {
2767		/*
2768		 * Currently, we only support two mutually exclusive
2769		 * mode flags.
2770		 */
2771		if (!strcmp(flags, "static"))
2772			mode_flags |= MPOL_F_STATIC_NODES;
2773		else if (!strcmp(flags, "relative"))
2774			mode_flags |= MPOL_F_RELATIVE_NODES;
2775		else
2776			goto out;
2777	}
2778
2779	new = mpol_new(mode, mode_flags, &nodes);
2780	if (IS_ERR(new))
2781		goto out;
2782
2783	/*
2784	 * Save nodes for mpol_to_str() to show the tmpfs mount options
2785	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2786	 */
2787	if (mode != MPOL_PREFERRED)
2788		new->v.nodes = nodes;
2789	else if (nodelist)
2790		new->v.preferred_node = first_node(nodes);
2791	else
2792		new->flags |= MPOL_F_LOCAL;
2793
2794	/*
2795	 * Save nodes for contextualization: this will be used to "clone"
2796	 * the mempolicy in a specific context [cpuset] at a later time.
2797	 */
2798	new->w.user_nodemask = nodes;
2799
2800	err = 0;
2801
2802out:
2803	/* Restore string for error message */
2804	if (nodelist)
2805		*--nodelist = ':';
2806	if (flags)
2807		*--flags = '=';
2808	if (!err)
2809		*mpol = new;
2810	return err;
2811}
2812#endif /* CONFIG_TMPFS */
2813
2814/**
2815 * mpol_to_str - format a mempolicy structure for printing
2816 * @buffer:  to contain formatted mempolicy string
2817 * @maxlen:  length of @buffer
2818 * @pol:  pointer to mempolicy to be formatted
2819 *
2820 * Convert @pol into a string.  If @buffer is too short, truncate the string.
2821 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2822 * longest flag, "relative", and to display at least a few node ids.
2823 */
2824void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2825{
2826	char *p = buffer;
2827	nodemask_t nodes = NODE_MASK_NONE;
2828	unsigned short mode = MPOL_DEFAULT;
2829	unsigned short flags = 0;
2830
2831	if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2832		mode = pol->mode;
2833		flags = pol->flags;
2834	}
2835
2836	switch (mode) {
2837	case MPOL_DEFAULT:
2838		break;
2839	case MPOL_PREFERRED:
2840		if (flags & MPOL_F_LOCAL)
2841			mode = MPOL_LOCAL;
2842		else
2843			node_set(pol->v.preferred_node, nodes);
2844		break;
2845	case MPOL_BIND:
2846	case MPOL_INTERLEAVE:
2847		nodes = pol->v.nodes;
2848		break;
2849	default:
2850		WARN_ON_ONCE(1);
2851		snprintf(p, maxlen, "unknown");
2852		return;
2853	}
2854
2855	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2856
2857	if (flags & MPOL_MODE_FLAGS) {
2858		p += snprintf(p, buffer + maxlen - p, "=");
2859
2860		/*
2861		 * Currently, the only defined flags are mutually exclusive
2862		 */
2863		if (flags & MPOL_F_STATIC_NODES)
2864			p += snprintf(p, buffer + maxlen - p, "static");
2865		else if (flags & MPOL_F_RELATIVE_NODES)
2866			p += snprintf(p, buffer + maxlen - p, "relative");
2867	}
2868
2869	if (!nodes_empty(nodes)) {
2870		p += snprintf(p, buffer + maxlen - p, ":");
2871	 	p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2872	}
2873}
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Simple NUMA memory policy for the Linux kernel.
   4 *
   5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
   6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
   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 NUMA_NO_NODE 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*/
  67
  68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  69
  70#include <linux/mempolicy.h>
  71#include <linux/pagewalk.h>
  72#include <linux/highmem.h>
  73#include <linux/hugetlb.h>
  74#include <linux/kernel.h>
  75#include <linux/sched.h>
  76#include <linux/sched/mm.h>
  77#include <linux/sched/numa_balancing.h>
  78#include <linux/sched/task.h>
  79#include <linux/nodemask.h>
  80#include <linux/cpuset.h>
  81#include <linux/slab.h>
  82#include <linux/string.h>
  83#include <linux/export.h>
  84#include <linux/nsproxy.h>
  85#include <linux/interrupt.h>
  86#include <linux/init.h>
  87#include <linux/compat.h>
  88#include <linux/ptrace.h>
  89#include <linux/swap.h>
  90#include <linux/seq_file.h>
  91#include <linux/proc_fs.h>
  92#include <linux/migrate.h>
  93#include <linux/ksm.h>
  94#include <linux/rmap.h>
  95#include <linux/security.h>
  96#include <linux/syscalls.h>
  97#include <linux/ctype.h>
  98#include <linux/mm_inline.h>
  99#include <linux/mmu_notifier.h>
 100#include <linux/printk.h>
 101#include <linux/swapops.h>
 102
 103#include <asm/tlbflush.h>
 104#include <linux/uaccess.h>
 105
 106#include "internal.h"
 107
 108/* Internal flags */
 109#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)	/* Skip checks for continuous vmas */
 110#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)		/* Invert check for nodemask */
 111
 112static struct kmem_cache *policy_cache;
 113static struct kmem_cache *sn_cache;
 114
 115/* Highest zone. An specific allocation for a zone below that is not
 116   policied. */
 117enum zone_type policy_zone = 0;
 118
 119/*
 120 * run-time system-wide default policy => local allocation
 121 */
 122static struct mempolicy default_policy = {
 123	.refcnt = ATOMIC_INIT(1), /* never free it */
 124	.mode = MPOL_LOCAL,
 125};
 126
 127static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 128
 129/**
 130 * numa_map_to_online_node - Find closest online node
 131 * @node: Node id to start the search
 132 *
 133 * Lookup the next closest node by distance if @nid is not online.
 134 */
 135int numa_map_to_online_node(int node)
 136{
 137	int min_dist = INT_MAX, dist, n, min_node;
 138
 139	if (node == NUMA_NO_NODE || node_online(node))
 140		return node;
 141
 142	min_node = node;
 143	for_each_online_node(n) {
 144		dist = node_distance(node, n);
 145		if (dist < min_dist) {
 146			min_dist = dist;
 147			min_node = n;
 148		}
 149	}
 150
 151	return min_node;
 152}
 153EXPORT_SYMBOL_GPL(numa_map_to_online_node);
 154
 155struct mempolicy *get_task_policy(struct task_struct *p)
 156{
 157	struct mempolicy *pol = p->mempolicy;
 158	int node;
 159
 160	if (pol)
 161		return pol;
 162
 163	node = numa_node_id();
 164	if (node != NUMA_NO_NODE) {
 165		pol = &preferred_node_policy[node];
 166		/* preferred_node_policy is not initialised early in boot */
 167		if (pol->mode)
 168			return pol;
 169	}
 170
 171	return &default_policy;
 172}
 173
 174static const struct mempolicy_operations {
 175	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 176	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
 177} mpol_ops[MPOL_MAX];
 178
 179static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 180{
 181	return pol->flags & MPOL_MODE_FLAGS;
 182}
 183
 184static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 185				   const nodemask_t *rel)
 186{
 187	nodemask_t tmp;
 188	nodes_fold(tmp, *orig, nodes_weight(*rel));
 189	nodes_onto(*ret, tmp, *rel);
 190}
 191
 192static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
 193{
 194	if (nodes_empty(*nodes))
 195		return -EINVAL;
 196	pol->nodes = *nodes;
 197	return 0;
 198}
 199
 200static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 201{
 202	if (nodes_empty(*nodes))
 203		return -EINVAL;
 204
 205	nodes_clear(pol->nodes);
 206	node_set(first_node(*nodes), pol->nodes);
 207	return 0;
 208}
 209
 210static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
 211{
 212	if (nodes_empty(*nodes))
 213		return -EINVAL;
 214	pol->nodes = *nodes;
 215	return 0;
 216}
 217
 218/*
 219 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 220 * any, for the new policy.  mpol_new() has already validated the nodes
 221 * parameter with respect to the policy mode and flags.
 222 *
 223 * Must be called holding task's alloc_lock to protect task's mems_allowed
 224 * and mempolicy.  May also be called holding the mmap_lock for write.
 225 */
 226static int mpol_set_nodemask(struct mempolicy *pol,
 227		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 228{
 229	int ret;
 230
 231	/*
 232	 * Default (pol==NULL) resp. local memory policies are not a
 233	 * subject of any remapping. They also do not need any special
 234	 * constructor.
 235	 */
 236	if (!pol || pol->mode == MPOL_LOCAL)
 237		return 0;
 238
 239	/* Check N_MEMORY */
 240	nodes_and(nsc->mask1,
 241		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 242
 243	VM_BUG_ON(!nodes);
 244
 245	if (pol->flags & MPOL_F_RELATIVE_NODES)
 246		mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
 247	else
 248		nodes_and(nsc->mask2, *nodes, nsc->mask1);
 249
 250	if (mpol_store_user_nodemask(pol))
 251		pol->w.user_nodemask = *nodes;
 252	else
 253		pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
 254
 255	ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 256	return ret;
 257}
 258
 259/*
 260 * This function just creates a new policy, does some check and simple
 261 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 262 */
 263static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 264				  nodemask_t *nodes)
 265{
 266	struct mempolicy *policy;
 267
 268	pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 269		 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
 270
 271	if (mode == MPOL_DEFAULT) {
 272		if (nodes && !nodes_empty(*nodes))
 273			return ERR_PTR(-EINVAL);
 274		return NULL;
 275	}
 276	VM_BUG_ON(!nodes);
 277
 278	/*
 279	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 280	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 281	 * All other modes require a valid pointer to a non-empty nodemask.
 282	 */
 283	if (mode == MPOL_PREFERRED) {
 284		if (nodes_empty(*nodes)) {
 285			if (((flags & MPOL_F_STATIC_NODES) ||
 286			     (flags & MPOL_F_RELATIVE_NODES)))
 287				return ERR_PTR(-EINVAL);
 288
 289			mode = MPOL_LOCAL;
 290		}
 291	} else if (mode == MPOL_LOCAL) {
 292		if (!nodes_empty(*nodes) ||
 293		    (flags & MPOL_F_STATIC_NODES) ||
 294		    (flags & MPOL_F_RELATIVE_NODES))
 295			return ERR_PTR(-EINVAL);
 296	} else if (nodes_empty(*nodes))
 297		return ERR_PTR(-EINVAL);
 298	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 299	if (!policy)
 300		return ERR_PTR(-ENOMEM);
 301	atomic_set(&policy->refcnt, 1);
 302	policy->mode = mode;
 303	policy->flags = flags;
 304
 305	return policy;
 306}
 307
 308/* Slow path of a mpol destructor. */
 309void __mpol_put(struct mempolicy *p)
 310{
 311	if (!atomic_dec_and_test(&p->refcnt))
 312		return;
 313	kmem_cache_free(policy_cache, p);
 314}
 315
 316static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
 317{
 318}
 319
 320static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 321{
 322	nodemask_t tmp;
 323
 324	if (pol->flags & MPOL_F_STATIC_NODES)
 325		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 326	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 327		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 328	else {
 329		nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
 330								*nodes);
 331		pol->w.cpuset_mems_allowed = *nodes;
 332	}
 333
 334	if (nodes_empty(tmp))
 335		tmp = *nodes;
 336
 337	pol->nodes = tmp;
 338}
 339
 340static void mpol_rebind_preferred(struct mempolicy *pol,
 341						const nodemask_t *nodes)
 342{
 343	pol->w.cpuset_mems_allowed = *nodes;
 344}
 345
 346/*
 347 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 348 *
 349 * Per-vma policies are protected by mmap_lock. Allocations using per-task
 350 * policies are protected by task->mems_allowed_seq to prevent a premature
 351 * OOM/allocation failure due to parallel nodemask modification.
 352 */
 353static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
 354{
 355	if (!pol)
 356		return;
 357	if (!mpol_store_user_nodemask(pol) &&
 358	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 359		return;
 360
 361	mpol_ops[pol->mode].rebind(pol, newmask);
 362}
 363
 364/*
 365 * Wrapper for mpol_rebind_policy() that just requires task
 366 * pointer, and updates task mempolicy.
 367 *
 368 * Called with task's alloc_lock held.
 369 */
 370
 371void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
 372{
 373	mpol_rebind_policy(tsk->mempolicy, new);
 374}
 375
 376/*
 377 * Rebind each vma in mm to new nodemask.
 378 *
 379 * Call holding a reference to mm.  Takes mm->mmap_lock during call.
 380 */
 381
 382void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 383{
 384	struct vm_area_struct *vma;
 385
 386	mmap_write_lock(mm);
 387	for (vma = mm->mmap; vma; vma = vma->vm_next)
 388		mpol_rebind_policy(vma->vm_policy, new);
 389	mmap_write_unlock(mm);
 390}
 391
 392static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 393	[MPOL_DEFAULT] = {
 394		.rebind = mpol_rebind_default,
 395	},
 396	[MPOL_INTERLEAVE] = {
 397		.create = mpol_new_interleave,
 398		.rebind = mpol_rebind_nodemask,
 399	},
 400	[MPOL_PREFERRED] = {
 401		.create = mpol_new_preferred,
 402		.rebind = mpol_rebind_preferred,
 403	},
 404	[MPOL_BIND] = {
 405		.create = mpol_new_bind,
 406		.rebind = mpol_rebind_nodemask,
 407	},
 408	[MPOL_LOCAL] = {
 409		.rebind = mpol_rebind_default,
 410	},
 411};
 412
 413static int migrate_page_add(struct page *page, struct list_head *pagelist,
 414				unsigned long flags);
 415
 416struct queue_pages {
 417	struct list_head *pagelist;
 418	unsigned long flags;
 419	nodemask_t *nmask;
 420	unsigned long start;
 421	unsigned long end;
 422	struct vm_area_struct *first;
 423};
 424
 425/*
 426 * Check if the page's nid is in qp->nmask.
 427 *
 428 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
 429 * in the invert of qp->nmask.
 430 */
 431static inline bool queue_pages_required(struct page *page,
 432					struct queue_pages *qp)
 433{
 434	int nid = page_to_nid(page);
 435	unsigned long flags = qp->flags;
 436
 437	return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
 438}
 439
 440/*
 441 * queue_pages_pmd() has four possible return values:
 442 * 0 - pages are placed on the right node or queued successfully, or
 443 *     special page is met, i.e. huge zero page.
 444 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 445 *     specified.
 446 * 2 - THP was split.
 447 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
 448 *        existing page was already on a node that does not follow the
 449 *        policy.
 450 */
 451static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
 452				unsigned long end, struct mm_walk *walk)
 453	__releases(ptl)
 454{
 455	int ret = 0;
 456	struct page *page;
 457	struct queue_pages *qp = walk->private;
 458	unsigned long flags;
 459
 460	if (unlikely(is_pmd_migration_entry(*pmd))) {
 461		ret = -EIO;
 462		goto unlock;
 463	}
 464	page = pmd_page(*pmd);
 465	if (is_huge_zero_page(page)) {
 466		spin_unlock(ptl);
 467		walk->action = ACTION_CONTINUE;
 468		goto out;
 469	}
 470	if (!queue_pages_required(page, qp))
 471		goto unlock;
 472
 473	flags = qp->flags;
 474	/* go to thp migration */
 475	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 476		if (!vma_migratable(walk->vma) ||
 477		    migrate_page_add(page, qp->pagelist, flags)) {
 478			ret = 1;
 479			goto unlock;
 480		}
 481	} else
 482		ret = -EIO;
 483unlock:
 484	spin_unlock(ptl);
 485out:
 486	return ret;
 487}
 488
 489/*
 490 * Scan through pages checking if pages follow certain conditions,
 491 * and move them to the pagelist if they do.
 492 *
 493 * queue_pages_pte_range() has three possible return values:
 494 * 0 - pages are placed on the right node or queued successfully, or
 495 *     special page is met, i.e. zero page.
 496 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 497 *     specified.
 498 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
 499 *        on a node that does not follow the policy.
 500 */
 501static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
 502			unsigned long end, struct mm_walk *walk)
 503{
 504	struct vm_area_struct *vma = walk->vma;
 505	struct page *page;
 506	struct queue_pages *qp = walk->private;
 507	unsigned long flags = qp->flags;
 508	int ret;
 509	bool has_unmovable = false;
 510	pte_t *pte, *mapped_pte;
 511	spinlock_t *ptl;
 512
 513	ptl = pmd_trans_huge_lock(pmd, vma);
 514	if (ptl) {
 515		ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
 516		if (ret != 2)
 517			return ret;
 518	}
 519	/* THP was split, fall through to pte walk */
 520
 521	if (pmd_trans_unstable(pmd))
 522		return 0;
 523
 524	mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
 525	for (; addr != end; pte++, addr += PAGE_SIZE) {
 526		if (!pte_present(*pte))
 527			continue;
 528		page = vm_normal_page(vma, addr, *pte);
 529		if (!page)
 530			continue;
 531		/*
 532		 * vm_normal_page() filters out zero pages, but there might
 533		 * still be PageReserved pages to skip, perhaps in a VDSO.
 534		 */
 535		if (PageReserved(page))
 536			continue;
 537		if (!queue_pages_required(page, qp))
 538			continue;
 539		if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 540			/* MPOL_MF_STRICT must be specified if we get here */
 541			if (!vma_migratable(vma)) {
 542				has_unmovable = true;
 543				break;
 544			}
 545
 546			/*
 547			 * Do not abort immediately since there may be
 548			 * temporary off LRU pages in the range.  Still
 549			 * need migrate other LRU pages.
 550			 */
 551			if (migrate_page_add(page, qp->pagelist, flags))
 552				has_unmovable = true;
 553		} else
 554			break;
 555	}
 556	pte_unmap_unlock(mapped_pte, ptl);
 557	cond_resched();
 558
 559	if (has_unmovable)
 560		return 1;
 561
 562	return addr != end ? -EIO : 0;
 563}
 564
 565static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
 566			       unsigned long addr, unsigned long end,
 567			       struct mm_walk *walk)
 568{
 569	int ret = 0;
 570#ifdef CONFIG_HUGETLB_PAGE
 571	struct queue_pages *qp = walk->private;
 572	unsigned long flags = (qp->flags & MPOL_MF_VALID);
 573	struct page *page;
 574	spinlock_t *ptl;
 575	pte_t entry;
 576
 577	ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 578	entry = huge_ptep_get(pte);
 579	if (!pte_present(entry))
 580		goto unlock;
 581	page = pte_page(entry);
 582	if (!queue_pages_required(page, qp))
 583		goto unlock;
 584
 585	if (flags == MPOL_MF_STRICT) {
 586		/*
 587		 * STRICT alone means only detecting misplaced page and no
 588		 * need to further check other vma.
 589		 */
 590		ret = -EIO;
 591		goto unlock;
 592	}
 593
 594	if (!vma_migratable(walk->vma)) {
 595		/*
 596		 * Must be STRICT with MOVE*, otherwise .test_walk() have
 597		 * stopped walking current vma.
 598		 * Detecting misplaced page but allow migrating pages which
 599		 * have been queued.
 600		 */
 601		ret = 1;
 602		goto unlock;
 603	}
 604
 605	/* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
 606	if (flags & (MPOL_MF_MOVE_ALL) ||
 607	    (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
 608		if (!isolate_huge_page(page, qp->pagelist) &&
 609			(flags & MPOL_MF_STRICT))
 610			/*
 611			 * Failed to isolate page but allow migrating pages
 612			 * which have been queued.
 613			 */
 614			ret = 1;
 615	}
 616unlock:
 617	spin_unlock(ptl);
 618#else
 619	BUG();
 620#endif
 621	return ret;
 622}
 623
 624#ifdef CONFIG_NUMA_BALANCING
 625/*
 626 * This is used to mark a range of virtual addresses to be inaccessible.
 627 * These are later cleared by a NUMA hinting fault. Depending on these
 628 * faults, pages may be migrated for better NUMA placement.
 629 *
 630 * This is assuming that NUMA faults are handled using PROT_NONE. If
 631 * an architecture makes a different choice, it will need further
 632 * changes to the core.
 633 */
 634unsigned long change_prot_numa(struct vm_area_struct *vma,
 635			unsigned long addr, unsigned long end)
 636{
 637	int nr_updated;
 638
 639	nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
 640	if (nr_updated)
 641		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 642
 643	return nr_updated;
 644}
 645#else
 646static unsigned long change_prot_numa(struct vm_area_struct *vma,
 647			unsigned long addr, unsigned long end)
 648{
 649	return 0;
 650}
 651#endif /* CONFIG_NUMA_BALANCING */
 652
 653static int queue_pages_test_walk(unsigned long start, unsigned long end,
 654				struct mm_walk *walk)
 655{
 656	struct vm_area_struct *vma = walk->vma;
 657	struct queue_pages *qp = walk->private;
 658	unsigned long endvma = vma->vm_end;
 659	unsigned long flags = qp->flags;
 660
 661	/* range check first */
 662	VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
 663
 664	if (!qp->first) {
 665		qp->first = vma;
 666		if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 667			(qp->start < vma->vm_start))
 668			/* hole at head side of range */
 669			return -EFAULT;
 670	}
 671	if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 672		((vma->vm_end < qp->end) &&
 673		(!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
 674		/* hole at middle or tail of range */
 675		return -EFAULT;
 676
 677	/*
 678	 * Need check MPOL_MF_STRICT to return -EIO if possible
 679	 * regardless of vma_migratable
 680	 */
 681	if (!vma_migratable(vma) &&
 682	    !(flags & MPOL_MF_STRICT))
 683		return 1;
 684
 685	if (endvma > end)
 686		endvma = end;
 687
 688	if (flags & MPOL_MF_LAZY) {
 689		/* Similar to task_numa_work, skip inaccessible VMAs */
 690		if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
 691			!(vma->vm_flags & VM_MIXEDMAP))
 692			change_prot_numa(vma, start, endvma);
 693		return 1;
 694	}
 695
 696	/* queue pages from current vma */
 697	if (flags & MPOL_MF_VALID)
 698		return 0;
 699	return 1;
 700}
 701
 702static const struct mm_walk_ops queue_pages_walk_ops = {
 703	.hugetlb_entry		= queue_pages_hugetlb,
 704	.pmd_entry		= queue_pages_pte_range,
 705	.test_walk		= queue_pages_test_walk,
 706};
 707
 708/*
 709 * Walk through page tables and collect pages to be migrated.
 710 *
 711 * If pages found in a given range are on a set of nodes (determined by
 712 * @nodes and @flags,) it's isolated and queued to the pagelist which is
 713 * passed via @private.
 714 *
 715 * queue_pages_range() has three possible return values:
 716 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
 717 *     specified.
 718 * 0 - queue pages successfully or no misplaced page.
 719 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
 720 *         memory range specified by nodemask and maxnode points outside
 721 *         your accessible address space (-EFAULT)
 722 */
 723static int
 724queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 725		nodemask_t *nodes, unsigned long flags,
 726		struct list_head *pagelist)
 727{
 728	int err;
 729	struct queue_pages qp = {
 730		.pagelist = pagelist,
 731		.flags = flags,
 732		.nmask = nodes,
 733		.start = start,
 734		.end = end,
 735		.first = NULL,
 736	};
 737
 738	err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
 739
 740	if (!qp.first)
 741		/* whole range in hole */
 742		err = -EFAULT;
 743
 744	return err;
 745}
 746
 747/*
 748 * Apply policy to a single VMA
 749 * This must be called with the mmap_lock held for writing.
 750 */
 751static int vma_replace_policy(struct vm_area_struct *vma,
 752						struct mempolicy *pol)
 753{
 754	int err;
 755	struct mempolicy *old;
 756	struct mempolicy *new;
 757
 758	pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 759		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 760		 vma->vm_ops, vma->vm_file,
 761		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 762
 763	new = mpol_dup(pol);
 764	if (IS_ERR(new))
 765		return PTR_ERR(new);
 766
 767	if (vma->vm_ops && vma->vm_ops->set_policy) {
 768		err = vma->vm_ops->set_policy(vma, new);
 769		if (err)
 770			goto err_out;
 771	}
 772
 773	old = vma->vm_policy;
 774	vma->vm_policy = new; /* protected by mmap_lock */
 775	mpol_put(old);
 776
 777	return 0;
 778 err_out:
 779	mpol_put(new);
 780	return err;
 781}
 782
 783/* Step 2: apply policy to a range and do splits. */
 784static int mbind_range(struct mm_struct *mm, unsigned long start,
 785		       unsigned long end, struct mempolicy *new_pol)
 786{
 787	struct vm_area_struct *next;
 788	struct vm_area_struct *prev;
 789	struct vm_area_struct *vma;
 790	int err = 0;
 791	pgoff_t pgoff;
 792	unsigned long vmstart;
 793	unsigned long vmend;
 794
 795	vma = find_vma(mm, start);
 796	VM_BUG_ON(!vma);
 797
 798	prev = vma->vm_prev;
 799	if (start > vma->vm_start)
 800		prev = vma;
 801
 802	for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 803		next = vma->vm_next;
 804		vmstart = max(start, vma->vm_start);
 805		vmend   = min(end, vma->vm_end);
 806
 807		if (mpol_equal(vma_policy(vma), new_pol))
 808			continue;
 809
 810		pgoff = vma->vm_pgoff +
 811			((vmstart - vma->vm_start) >> PAGE_SHIFT);
 812		prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 813				 vma->anon_vma, vma->vm_file, pgoff,
 814				 new_pol, vma->vm_userfaultfd_ctx);
 815		if (prev) {
 816			vma = prev;
 817			next = vma->vm_next;
 818			if (mpol_equal(vma_policy(vma), new_pol))
 819				continue;
 820			/* vma_merge() joined vma && vma->next, case 8 */
 821			goto replace;
 822		}
 823		if (vma->vm_start != vmstart) {
 824			err = split_vma(vma->vm_mm, vma, vmstart, 1);
 825			if (err)
 826				goto out;
 827		}
 828		if (vma->vm_end != vmend) {
 829			err = split_vma(vma->vm_mm, vma, vmend, 0);
 830			if (err)
 831				goto out;
 832		}
 833 replace:
 834		err = vma_replace_policy(vma, new_pol);
 835		if (err)
 836			goto out;
 837	}
 838
 839 out:
 840	return err;
 841}
 842
 843/* Set the process memory policy */
 844static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 845			     nodemask_t *nodes)
 846{
 847	struct mempolicy *new, *old;
 848	NODEMASK_SCRATCH(scratch);
 849	int ret;
 850
 851	if (!scratch)
 852		return -ENOMEM;
 853
 854	new = mpol_new(mode, flags, nodes);
 855	if (IS_ERR(new)) {
 856		ret = PTR_ERR(new);
 857		goto out;
 858	}
 859
 860	ret = mpol_set_nodemask(new, nodes, scratch);
 861	if (ret) {
 862		mpol_put(new);
 863		goto out;
 864	}
 865	task_lock(current);
 866	old = current->mempolicy;
 867	current->mempolicy = new;
 868	if (new && new->mode == MPOL_INTERLEAVE)
 869		current->il_prev = MAX_NUMNODES-1;
 870	task_unlock(current);
 871	mpol_put(old);
 872	ret = 0;
 873out:
 874	NODEMASK_SCRATCH_FREE(scratch);
 875	return ret;
 876}
 877
 878/*
 879 * Return nodemask for policy for get_mempolicy() query
 880 *
 881 * Called with task's alloc_lock held
 882 */
 883static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 884{
 885	nodes_clear(*nodes);
 886	if (p == &default_policy)
 887		return;
 888
 889	switch (p->mode) {
 890	case MPOL_BIND:
 891	case MPOL_INTERLEAVE:
 892	case MPOL_PREFERRED:
 893		*nodes = p->nodes;
 894		break;
 895	case MPOL_LOCAL:
 896		/* return empty node mask for local allocation */
 897		break;
 898	default:
 899		BUG();
 900	}
 901}
 902
 903static int lookup_node(struct mm_struct *mm, unsigned long addr)
 904{
 905	struct page *p = NULL;
 906	int err;
 907
 908	int locked = 1;
 909	err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
 910	if (err > 0) {
 911		err = page_to_nid(p);
 912		put_page(p);
 913	}
 914	if (locked)
 915		mmap_read_unlock(mm);
 916	return err;
 917}
 918
 919/* Retrieve NUMA policy */
 920static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 921			     unsigned long addr, unsigned long flags)
 922{
 923	int err;
 924	struct mm_struct *mm = current->mm;
 925	struct vm_area_struct *vma = NULL;
 926	struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
 927
 928	if (flags &
 929		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 930		return -EINVAL;
 931
 932	if (flags & MPOL_F_MEMS_ALLOWED) {
 933		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 934			return -EINVAL;
 935		*policy = 0;	/* just so it's initialized */
 936		task_lock(current);
 937		*nmask  = cpuset_current_mems_allowed;
 938		task_unlock(current);
 939		return 0;
 940	}
 941
 942	if (flags & MPOL_F_ADDR) {
 943		/*
 944		 * Do NOT fall back to task policy if the
 945		 * vma/shared policy at addr is NULL.  We
 946		 * want to return MPOL_DEFAULT in this case.
 947		 */
 948		mmap_read_lock(mm);
 949		vma = vma_lookup(mm, addr);
 950		if (!vma) {
 951			mmap_read_unlock(mm);
 952			return -EFAULT;
 953		}
 954		if (vma->vm_ops && vma->vm_ops->get_policy)
 955			pol = vma->vm_ops->get_policy(vma, addr);
 956		else
 957			pol = vma->vm_policy;
 958	} else if (addr)
 959		return -EINVAL;
 960
 961	if (!pol)
 962		pol = &default_policy;	/* indicates default behavior */
 963
 964	if (flags & MPOL_F_NODE) {
 965		if (flags & MPOL_F_ADDR) {
 966			/*
 967			 * Take a refcount on the mpol, lookup_node()
 968			 * will drop the mmap_lock, so after calling
 969			 * lookup_node() only "pol" remains valid, "vma"
 970			 * is stale.
 971			 */
 972			pol_refcount = pol;
 973			vma = NULL;
 974			mpol_get(pol);
 975			err = lookup_node(mm, addr);
 976			if (err < 0)
 977				goto out;
 978			*policy = err;
 979		} else if (pol == current->mempolicy &&
 980				pol->mode == MPOL_INTERLEAVE) {
 981			*policy = next_node_in(current->il_prev, pol->nodes);
 982		} else {
 983			err = -EINVAL;
 984			goto out;
 985		}
 986	} else {
 987		*policy = pol == &default_policy ? MPOL_DEFAULT :
 988						pol->mode;
 989		/*
 990		 * Internal mempolicy flags must be masked off before exposing
 991		 * the policy to userspace.
 992		 */
 993		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 994	}
 995
 996	err = 0;
 997	if (nmask) {
 998		if (mpol_store_user_nodemask(pol)) {
 999			*nmask = pol->w.user_nodemask;
1000		} else {
1001			task_lock(current);
1002			get_policy_nodemask(pol, nmask);
1003			task_unlock(current);
1004		}
1005	}
1006
1007 out:
1008	mpol_cond_put(pol);
1009	if (vma)
1010		mmap_read_unlock(mm);
1011	if (pol_refcount)
1012		mpol_put(pol_refcount);
1013	return err;
1014}
1015
1016#ifdef CONFIG_MIGRATION
1017/*
1018 * page migration, thp tail pages can be passed.
1019 */
1020static int migrate_page_add(struct page *page, struct list_head *pagelist,
1021				unsigned long flags)
1022{
1023	struct page *head = compound_head(page);
1024	/*
1025	 * Avoid migrating a page that is shared with others.
1026	 */
1027	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1028		if (!isolate_lru_page(head)) {
1029			list_add_tail(&head->lru, pagelist);
1030			mod_node_page_state(page_pgdat(head),
1031				NR_ISOLATED_ANON + page_is_file_lru(head),
1032				thp_nr_pages(head));
1033		} else if (flags & MPOL_MF_STRICT) {
1034			/*
1035			 * Non-movable page may reach here.  And, there may be
1036			 * temporary off LRU pages or non-LRU movable pages.
1037			 * Treat them as unmovable pages since they can't be
1038			 * isolated, so they can't be moved at the moment.  It
1039			 * should return -EIO for this case too.
1040			 */
1041			return -EIO;
1042		}
1043	}
1044
1045	return 0;
1046}
1047
1048/*
1049 * Migrate pages from one node to a target node.
1050 * Returns error or the number of pages not migrated.
1051 */
1052static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1053			   int flags)
1054{
1055	nodemask_t nmask;
1056	LIST_HEAD(pagelist);
1057	int err = 0;
1058	struct migration_target_control mtc = {
1059		.nid = dest,
1060		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1061	};
1062
1063	nodes_clear(nmask);
1064	node_set(source, nmask);
1065
1066	/*
1067	 * This does not "check" the range but isolates all pages that
1068	 * need migration.  Between passing in the full user address
1069	 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1070	 */
1071	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1072	queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1073			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1074
1075	if (!list_empty(&pagelist)) {
1076		err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1077				(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1078		if (err)
1079			putback_movable_pages(&pagelist);
1080	}
1081
1082	return err;
1083}
1084
1085/*
1086 * Move pages between the two nodesets so as to preserve the physical
1087 * layout as much as possible.
1088 *
1089 * Returns the number of page that could not be moved.
1090 */
1091int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1092		     const nodemask_t *to, int flags)
1093{
1094	int busy = 0;
1095	int err = 0;
1096	nodemask_t tmp;
1097
1098	lru_cache_disable();
1099
1100	mmap_read_lock(mm);
1101
1102	/*
1103	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1104	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
1105	 * bit in 'tmp', and return that <source, dest> pair for migration.
1106	 * The pair of nodemasks 'to' and 'from' define the map.
1107	 *
1108	 * If no pair of bits is found that way, fallback to picking some
1109	 * pair of 'source' and 'dest' bits that are not the same.  If the
1110	 * 'source' and 'dest' bits are the same, this represents a node
1111	 * that will be migrating to itself, so no pages need move.
1112	 *
1113	 * If no bits are left in 'tmp', or if all remaining bits left
1114	 * in 'tmp' correspond to the same bit in 'to', return false
1115	 * (nothing left to migrate).
1116	 *
1117	 * This lets us pick a pair of nodes to migrate between, such that
1118	 * if possible the dest node is not already occupied by some other
1119	 * source node, minimizing the risk of overloading the memory on a
1120	 * node that would happen if we migrated incoming memory to a node
1121	 * before migrating outgoing memory source that same node.
1122	 *
1123	 * A single scan of tmp is sufficient.  As we go, we remember the
1124	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1125	 * that not only moved, but what's better, moved to an empty slot
1126	 * (d is not set in tmp), then we break out then, with that pair.
1127	 * Otherwise when we finish scanning from_tmp, we at least have the
1128	 * most recent <s, d> pair that moved.  If we get all the way through
1129	 * the scan of tmp without finding any node that moved, much less
1130	 * moved to an empty node, then there is nothing left worth migrating.
1131	 */
1132
1133	tmp = *from;
1134	while (!nodes_empty(tmp)) {
1135		int s, d;
1136		int source = NUMA_NO_NODE;
1137		int dest = 0;
1138
1139		for_each_node_mask(s, tmp) {
1140
1141			/*
1142			 * do_migrate_pages() tries to maintain the relative
1143			 * node relationship of the pages established between
1144			 * threads and memory areas.
1145                         *
1146			 * However if the number of source nodes is not equal to
1147			 * the number of destination nodes we can not preserve
1148			 * this node relative relationship.  In that case, skip
1149			 * copying memory from a node that is in the destination
1150			 * mask.
1151			 *
1152			 * Example: [2,3,4] -> [3,4,5] moves everything.
1153			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1154			 */
1155
1156			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1157						(node_isset(s, *to)))
1158				continue;
1159
1160			d = node_remap(s, *from, *to);
1161			if (s == d)
1162				continue;
1163
1164			source = s;	/* Node moved. Memorize */
1165			dest = d;
1166
1167			/* dest not in remaining from nodes? */
1168			if (!node_isset(dest, tmp))
1169				break;
1170		}
1171		if (source == NUMA_NO_NODE)
1172			break;
1173
1174		node_clear(source, tmp);
1175		err = migrate_to_node(mm, source, dest, flags);
1176		if (err > 0)
1177			busy += err;
1178		if (err < 0)
1179			break;
1180	}
1181	mmap_read_unlock(mm);
1182
1183	lru_cache_enable();
1184	if (err < 0)
1185		return err;
1186	return busy;
1187
1188}
1189
1190/*
1191 * Allocate a new page for page migration based on vma policy.
1192 * Start by assuming the page is mapped by the same vma as contains @start.
1193 * Search forward from there, if not.  N.B., this assumes that the
1194 * list of pages handed to migrate_pages()--which is how we get here--
1195 * is in virtual address order.
1196 */
1197static struct page *new_page(struct page *page, unsigned long start)
1198{
1199	struct vm_area_struct *vma;
1200	unsigned long address;
1201
1202	vma = find_vma(current->mm, start);
1203	while (vma) {
1204		address = page_address_in_vma(page, vma);
1205		if (address != -EFAULT)
1206			break;
1207		vma = vma->vm_next;
1208	}
1209
1210	if (PageHuge(page)) {
1211		return alloc_huge_page_vma(page_hstate(compound_head(page)),
1212				vma, address);
1213	} else if (PageTransHuge(page)) {
1214		struct page *thp;
1215
1216		thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1217					 HPAGE_PMD_ORDER);
1218		if (!thp)
1219			return NULL;
1220		prep_transhuge_page(thp);
1221		return thp;
1222	}
1223	/*
1224	 * if !vma, alloc_page_vma() will use task or system default policy
1225	 */
1226	return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1227			vma, address);
1228}
1229#else
1230
1231static int migrate_page_add(struct page *page, struct list_head *pagelist,
1232				unsigned long flags)
1233{
1234	return -EIO;
1235}
1236
1237int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1238		     const nodemask_t *to, int flags)
1239{
1240	return -ENOSYS;
1241}
1242
1243static struct page *new_page(struct page *page, unsigned long start)
1244{
1245	return NULL;
1246}
1247#endif
1248
1249static long do_mbind(unsigned long start, unsigned long len,
1250		     unsigned short mode, unsigned short mode_flags,
1251		     nodemask_t *nmask, unsigned long flags)
1252{
1253	struct mm_struct *mm = current->mm;
1254	struct mempolicy *new;
1255	unsigned long end;
1256	int err;
1257	int ret;
1258	LIST_HEAD(pagelist);
1259
1260	if (flags & ~(unsigned long)MPOL_MF_VALID)
1261		return -EINVAL;
1262	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1263		return -EPERM;
1264
1265	if (start & ~PAGE_MASK)
1266		return -EINVAL;
1267
1268	if (mode == MPOL_DEFAULT)
1269		flags &= ~MPOL_MF_STRICT;
1270
1271	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1272	end = start + len;
1273
1274	if (end < start)
1275		return -EINVAL;
1276	if (end == start)
1277		return 0;
1278
1279	new = mpol_new(mode, mode_flags, nmask);
1280	if (IS_ERR(new))
1281		return PTR_ERR(new);
1282
1283	if (flags & MPOL_MF_LAZY)
1284		new->flags |= MPOL_F_MOF;
1285
1286	/*
1287	 * If we are using the default policy then operation
1288	 * on discontinuous address spaces is okay after all
1289	 */
1290	if (!new)
1291		flags |= MPOL_MF_DISCONTIG_OK;
1292
1293	pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1294		 start, start + len, mode, mode_flags,
1295		 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1296
1297	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1298
1299		lru_cache_disable();
1300	}
1301	{
1302		NODEMASK_SCRATCH(scratch);
1303		if (scratch) {
1304			mmap_write_lock(mm);
1305			err = mpol_set_nodemask(new, nmask, scratch);
1306			if (err)
1307				mmap_write_unlock(mm);
1308		} else
1309			err = -ENOMEM;
1310		NODEMASK_SCRATCH_FREE(scratch);
1311	}
1312	if (err)
1313		goto mpol_out;
1314
1315	ret = queue_pages_range(mm, start, end, nmask,
1316			  flags | MPOL_MF_INVERT, &pagelist);
1317
1318	if (ret < 0) {
1319		err = ret;
1320		goto up_out;
1321	}
1322
1323	err = mbind_range(mm, start, end, new);
1324
1325	if (!err) {
1326		int nr_failed = 0;
1327
1328		if (!list_empty(&pagelist)) {
1329			WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1330			nr_failed = migrate_pages(&pagelist, new_page, NULL,
1331				start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1332			if (nr_failed)
1333				putback_movable_pages(&pagelist);
1334		}
1335
1336		if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1337			err = -EIO;
1338	} else {
1339up_out:
1340		if (!list_empty(&pagelist))
1341			putback_movable_pages(&pagelist);
1342	}
1343
1344	mmap_write_unlock(mm);
1345mpol_out:
1346	mpol_put(new);
1347	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1348		lru_cache_enable();
1349	return err;
1350}
1351
1352/*
1353 * User space interface with variable sized bitmaps for nodelists.
1354 */
1355
1356/* Copy a node mask from user space. */
1357static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1358		     unsigned long maxnode)
1359{
1360	unsigned long k;
1361	unsigned long t;
1362	unsigned long nlongs;
1363	unsigned long endmask;
1364
1365	--maxnode;
1366	nodes_clear(*nodes);
1367	if (maxnode == 0 || !nmask)
1368		return 0;
1369	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1370		return -EINVAL;
1371
1372	nlongs = BITS_TO_LONGS(maxnode);
1373	if ((maxnode % BITS_PER_LONG) == 0)
1374		endmask = ~0UL;
1375	else
1376		endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1377
1378	/*
1379	 * When the user specified more nodes than supported just check
1380	 * if the non supported part is all zero.
1381	 *
1382	 * If maxnode have more longs than MAX_NUMNODES, check
1383	 * the bits in that area first. And then go through to
1384	 * check the rest bits which equal or bigger than MAX_NUMNODES.
1385	 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1386	 */
1387	if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1388		for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1389			if (get_user(t, nmask + k))
1390				return -EFAULT;
1391			if (k == nlongs - 1) {
1392				if (t & endmask)
1393					return -EINVAL;
1394			} else if (t)
1395				return -EINVAL;
1396		}
1397		nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1398		endmask = ~0UL;
1399	}
1400
1401	if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1402		unsigned long valid_mask = endmask;
1403
1404		valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1405		if (get_user(t, nmask + nlongs - 1))
1406			return -EFAULT;
1407		if (t & valid_mask)
1408			return -EINVAL;
1409	}
1410
1411	if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1412		return -EFAULT;
1413	nodes_addr(*nodes)[nlongs-1] &= endmask;
1414	return 0;
1415}
1416
1417/* Copy a kernel node mask to user space */
1418static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1419			      nodemask_t *nodes)
1420{
1421	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1422	unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1423
1424	if (copy > nbytes) {
1425		if (copy > PAGE_SIZE)
1426			return -EINVAL;
1427		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1428			return -EFAULT;
1429		copy = nbytes;
1430	}
1431	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1432}
1433
1434/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1435static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1436{
1437	*flags = *mode & MPOL_MODE_FLAGS;
1438	*mode &= ~MPOL_MODE_FLAGS;
1439	if ((unsigned int)(*mode) >= MPOL_MAX)
1440		return -EINVAL;
1441	if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1442		return -EINVAL;
1443	if (*flags & MPOL_F_NUMA_BALANCING) {
1444		if (*mode != MPOL_BIND)
1445			return -EINVAL;
1446		*flags |= (MPOL_F_MOF | MPOL_F_MORON);
1447	}
1448	return 0;
1449}
1450
1451static long kernel_mbind(unsigned long start, unsigned long len,
1452			 unsigned long mode, const unsigned long __user *nmask,
1453			 unsigned long maxnode, unsigned int flags)
1454{
1455	unsigned short mode_flags;
1456	nodemask_t nodes;
1457	int lmode = mode;
1458	int err;
1459
1460	start = untagged_addr(start);
1461	err = sanitize_mpol_flags(&lmode, &mode_flags);
1462	if (err)
1463		return err;
1464
1465	err = get_nodes(&nodes, nmask, maxnode);
1466	if (err)
1467		return err;
1468
1469	return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1470}
1471
1472SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1473		unsigned long, mode, const unsigned long __user *, nmask,
1474		unsigned long, maxnode, unsigned int, flags)
1475{
1476	return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1477}
1478
1479/* Set the process memory policy */
1480static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1481				 unsigned long maxnode)
1482{
1483	unsigned short mode_flags;
1484	nodemask_t nodes;
1485	int lmode = mode;
1486	int err;
1487
1488	err = sanitize_mpol_flags(&lmode, &mode_flags);
1489	if (err)
1490		return err;
1491
1492	err = get_nodes(&nodes, nmask, maxnode);
1493	if (err)
1494		return err;
1495
1496	return do_set_mempolicy(lmode, mode_flags, &nodes);
1497}
1498
1499SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1500		unsigned long, maxnode)
1501{
1502	return kernel_set_mempolicy(mode, nmask, maxnode);
1503}
1504
1505static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1506				const unsigned long __user *old_nodes,
1507				const unsigned long __user *new_nodes)
1508{
1509	struct mm_struct *mm = NULL;
1510	struct task_struct *task;
1511	nodemask_t task_nodes;
1512	int err;
1513	nodemask_t *old;
1514	nodemask_t *new;
1515	NODEMASK_SCRATCH(scratch);
1516
1517	if (!scratch)
1518		return -ENOMEM;
1519
1520	old = &scratch->mask1;
1521	new = &scratch->mask2;
1522
1523	err = get_nodes(old, old_nodes, maxnode);
1524	if (err)
1525		goto out;
1526
1527	err = get_nodes(new, new_nodes, maxnode);
1528	if (err)
1529		goto out;
1530
1531	/* Find the mm_struct */
1532	rcu_read_lock();
1533	task = pid ? find_task_by_vpid(pid) : current;
1534	if (!task) {
1535		rcu_read_unlock();
1536		err = -ESRCH;
1537		goto out;
1538	}
1539	get_task_struct(task);
1540
1541	err = -EINVAL;
1542
1543	/*
1544	 * Check if this process has the right to modify the specified process.
1545	 * Use the regular "ptrace_may_access()" checks.
1546	 */
1547	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1548		rcu_read_unlock();
1549		err = -EPERM;
1550		goto out_put;
1551	}
1552	rcu_read_unlock();
1553
1554	task_nodes = cpuset_mems_allowed(task);
1555	/* Is the user allowed to access the target nodes? */
1556	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1557		err = -EPERM;
1558		goto out_put;
1559	}
1560
1561	task_nodes = cpuset_mems_allowed(current);
1562	nodes_and(*new, *new, task_nodes);
1563	if (nodes_empty(*new))
1564		goto out_put;
1565
1566	err = security_task_movememory(task);
1567	if (err)
1568		goto out_put;
1569
1570	mm = get_task_mm(task);
1571	put_task_struct(task);
1572
1573	if (!mm) {
1574		err = -EINVAL;
1575		goto out;
1576	}
1577
1578	err = do_migrate_pages(mm, old, new,
1579		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1580
1581	mmput(mm);
1582out:
1583	NODEMASK_SCRATCH_FREE(scratch);
1584
1585	return err;
1586
1587out_put:
1588	put_task_struct(task);
1589	goto out;
1590
1591}
1592
1593SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1594		const unsigned long __user *, old_nodes,
1595		const unsigned long __user *, new_nodes)
1596{
1597	return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1598}
1599
1600
1601/* Retrieve NUMA policy */
1602static int kernel_get_mempolicy(int __user *policy,
1603				unsigned long __user *nmask,
1604				unsigned long maxnode,
1605				unsigned long addr,
1606				unsigned long flags)
1607{
1608	int err;
1609	int pval;
1610	nodemask_t nodes;
1611
1612	if (nmask != NULL && maxnode < nr_node_ids)
1613		return -EINVAL;
1614
1615	addr = untagged_addr(addr);
1616
1617	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1618
1619	if (err)
1620		return err;
1621
1622	if (policy && put_user(pval, policy))
1623		return -EFAULT;
1624
1625	if (nmask)
1626		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1627
1628	return err;
1629}
1630
1631SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1632		unsigned long __user *, nmask, unsigned long, maxnode,
1633		unsigned long, addr, unsigned long, flags)
1634{
1635	return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1636}
1637
1638#ifdef CONFIG_COMPAT
1639
1640COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1641		       compat_ulong_t __user *, nmask,
1642		       compat_ulong_t, maxnode,
1643		       compat_ulong_t, addr, compat_ulong_t, flags)
1644{
1645	long err;
1646	unsigned long __user *nm = NULL;
1647	unsigned long nr_bits, alloc_size;
1648	DECLARE_BITMAP(bm, MAX_NUMNODES);
1649
1650	nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1651	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1652
1653	if (nmask)
1654		nm = compat_alloc_user_space(alloc_size);
1655
1656	err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1657
1658	if (!err && nmask) {
1659		unsigned long copy_size;
1660		copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1661		err = copy_from_user(bm, nm, copy_size);
1662		/* ensure entire bitmap is zeroed */
1663		err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1664		err |= compat_put_bitmap(nmask, bm, nr_bits);
1665	}
1666
1667	return err;
1668}
1669
1670COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1671		       compat_ulong_t, maxnode)
1672{
1673	unsigned long __user *nm = NULL;
1674	unsigned long nr_bits, alloc_size;
1675	DECLARE_BITMAP(bm, MAX_NUMNODES);
1676
1677	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1678	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1679
1680	if (nmask) {
1681		if (compat_get_bitmap(bm, nmask, nr_bits))
1682			return -EFAULT;
1683		nm = compat_alloc_user_space(alloc_size);
1684		if (copy_to_user(nm, bm, alloc_size))
1685			return -EFAULT;
1686	}
1687
1688	return kernel_set_mempolicy(mode, nm, nr_bits+1);
1689}
1690
1691COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1692		       compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1693		       compat_ulong_t, maxnode, compat_ulong_t, flags)
1694{
1695	unsigned long __user *nm = NULL;
1696	unsigned long nr_bits, alloc_size;
1697	nodemask_t bm;
1698
1699	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1700	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1701
1702	if (nmask) {
1703		if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1704			return -EFAULT;
1705		nm = compat_alloc_user_space(alloc_size);
1706		if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1707			return -EFAULT;
1708	}
1709
1710	return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1711}
1712
1713COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1714		       compat_ulong_t, maxnode,
1715		       const compat_ulong_t __user *, old_nodes,
1716		       const compat_ulong_t __user *, new_nodes)
1717{
1718	unsigned long __user *old = NULL;
1719	unsigned long __user *new = NULL;
1720	nodemask_t tmp_mask;
1721	unsigned long nr_bits;
1722	unsigned long size;
1723
1724	nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1725	size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1726	if (old_nodes) {
1727		if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1728			return -EFAULT;
1729		old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1730		if (new_nodes)
1731			new = old + size / sizeof(unsigned long);
1732		if (copy_to_user(old, nodes_addr(tmp_mask), size))
1733			return -EFAULT;
1734	}
1735	if (new_nodes) {
1736		if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1737			return -EFAULT;
1738		if (new == NULL)
1739			new = compat_alloc_user_space(size);
1740		if (copy_to_user(new, nodes_addr(tmp_mask), size))
1741			return -EFAULT;
1742	}
1743	return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1744}
1745
1746#endif /* CONFIG_COMPAT */
1747
1748bool vma_migratable(struct vm_area_struct *vma)
1749{
1750	if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1751		return false;
1752
1753	/*
1754	 * DAX device mappings require predictable access latency, so avoid
1755	 * incurring periodic faults.
1756	 */
1757	if (vma_is_dax(vma))
1758		return false;
1759
1760	if (is_vm_hugetlb_page(vma) &&
1761		!hugepage_migration_supported(hstate_vma(vma)))
1762		return false;
1763
1764	/*
1765	 * Migration allocates pages in the highest zone. If we cannot
1766	 * do so then migration (at least from node to node) is not
1767	 * possible.
1768	 */
1769	if (vma->vm_file &&
1770		gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1771			< policy_zone)
1772		return false;
1773	return true;
1774}
1775
1776struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1777						unsigned long addr)
1778{
1779	struct mempolicy *pol = NULL;
1780
1781	if (vma) {
1782		if (vma->vm_ops && vma->vm_ops->get_policy) {
1783			pol = vma->vm_ops->get_policy(vma, addr);
1784		} else if (vma->vm_policy) {
1785			pol = vma->vm_policy;
1786
1787			/*
1788			 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1789			 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1790			 * count on these policies which will be dropped by
1791			 * mpol_cond_put() later
1792			 */
1793			if (mpol_needs_cond_ref(pol))
1794				mpol_get(pol);
1795		}
1796	}
1797
1798	return pol;
1799}
1800
1801/*
1802 * get_vma_policy(@vma, @addr)
1803 * @vma: virtual memory area whose policy is sought
1804 * @addr: address in @vma for shared policy lookup
1805 *
1806 * Returns effective policy for a VMA at specified address.
1807 * Falls back to current->mempolicy or system default policy, as necessary.
1808 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1809 * count--added by the get_policy() vm_op, as appropriate--to protect against
1810 * freeing by another task.  It is the caller's responsibility to free the
1811 * extra reference for shared policies.
1812 */
1813static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1814						unsigned long addr)
1815{
1816	struct mempolicy *pol = __get_vma_policy(vma, addr);
1817
1818	if (!pol)
1819		pol = get_task_policy(current);
1820
1821	return pol;
1822}
1823
1824bool vma_policy_mof(struct vm_area_struct *vma)
1825{
1826	struct mempolicy *pol;
1827
1828	if (vma->vm_ops && vma->vm_ops->get_policy) {
1829		bool ret = false;
1830
1831		pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1832		if (pol && (pol->flags & MPOL_F_MOF))
1833			ret = true;
1834		mpol_cond_put(pol);
1835
1836		return ret;
1837	}
1838
1839	pol = vma->vm_policy;
1840	if (!pol)
1841		pol = get_task_policy(current);
1842
1843	return pol->flags & MPOL_F_MOF;
1844}
1845
1846static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1847{
1848	enum zone_type dynamic_policy_zone = policy_zone;
1849
1850	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1851
1852	/*
1853	 * if policy->nodes has movable memory only,
1854	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1855	 *
1856	 * policy->nodes is intersect with node_states[N_MEMORY].
1857	 * so if the following test fails, it implies
1858	 * policy->nodes has movable memory only.
1859	 */
1860	if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1861		dynamic_policy_zone = ZONE_MOVABLE;
1862
1863	return zone >= dynamic_policy_zone;
1864}
1865
1866/*
1867 * Return a nodemask representing a mempolicy for filtering nodes for
1868 * page allocation
1869 */
1870nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1871{
1872	/* Lower zones don't get a nodemask applied for MPOL_BIND */
1873	if (unlikely(policy->mode == MPOL_BIND) &&
1874			apply_policy_zone(policy, gfp_zone(gfp)) &&
1875			cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1876		return &policy->nodes;
1877
1878	return NULL;
1879}
1880
1881/* Return the node id preferred by the given mempolicy, or the given id */
1882static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1883{
1884	if (policy->mode == MPOL_PREFERRED) {
1885		nd = first_node(policy->nodes);
1886	} else {
1887		/*
1888		 * __GFP_THISNODE shouldn't even be used with the bind policy
1889		 * because we might easily break the expectation to stay on the
1890		 * requested node and not break the policy.
1891		 */
1892		WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1893	}
1894
1895	return nd;
1896}
1897
1898/* Do dynamic interleaving for a process */
1899static unsigned interleave_nodes(struct mempolicy *policy)
1900{
1901	unsigned next;
1902	struct task_struct *me = current;
1903
1904	next = next_node_in(me->il_prev, policy->nodes);
1905	if (next < MAX_NUMNODES)
1906		me->il_prev = next;
1907	return next;
1908}
1909
1910/*
1911 * Depending on the memory policy provide a node from which to allocate the
1912 * next slab entry.
1913 */
1914unsigned int mempolicy_slab_node(void)
1915{
1916	struct mempolicy *policy;
1917	int node = numa_mem_id();
1918
1919	if (in_interrupt())
1920		return node;
1921
1922	policy = current->mempolicy;
1923	if (!policy)
1924		return node;
1925
1926	switch (policy->mode) {
1927	case MPOL_PREFERRED:
1928		return first_node(policy->nodes);
1929
1930	case MPOL_INTERLEAVE:
1931		return interleave_nodes(policy);
1932
1933	case MPOL_BIND: {
1934		struct zoneref *z;
1935
1936		/*
1937		 * Follow bind policy behavior and start allocation at the
1938		 * first node.
1939		 */
1940		struct zonelist *zonelist;
1941		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1942		zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1943		z = first_zones_zonelist(zonelist, highest_zoneidx,
1944							&policy->nodes);
1945		return z->zone ? zone_to_nid(z->zone) : node;
1946	}
1947	case MPOL_LOCAL:
1948		return node;
1949
1950	default:
1951		BUG();
1952	}
1953}
1954
1955/*
1956 * Do static interleaving for a VMA with known offset @n.  Returns the n'th
1957 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1958 * number of present nodes.
1959 */
1960static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1961{
1962	nodemask_t nodemask = pol->nodes;
1963	unsigned int target, nnodes;
1964	int i;
1965	int nid;
1966	/*
1967	 * The barrier will stabilize the nodemask in a register or on
1968	 * the stack so that it will stop changing under the code.
1969	 *
1970	 * Between first_node() and next_node(), pol->nodes could be changed
1971	 * by other threads. So we put pol->nodes in a local stack.
1972	 */
1973	barrier();
1974
1975	nnodes = nodes_weight(nodemask);
1976	if (!nnodes)
1977		return numa_node_id();
1978	target = (unsigned int)n % nnodes;
1979	nid = first_node(nodemask);
1980	for (i = 0; i < target; i++)
1981		nid = next_node(nid, nodemask);
1982	return nid;
1983}
1984
1985/* Determine a node number for interleave */
1986static inline unsigned interleave_nid(struct mempolicy *pol,
1987		 struct vm_area_struct *vma, unsigned long addr, int shift)
1988{
1989	if (vma) {
1990		unsigned long off;
1991
1992		/*
1993		 * for small pages, there is no difference between
1994		 * shift and PAGE_SHIFT, so the bit-shift is safe.
1995		 * for huge pages, since vm_pgoff is in units of small
1996		 * pages, we need to shift off the always 0 bits to get
1997		 * a useful offset.
1998		 */
1999		BUG_ON(shift < PAGE_SHIFT);
2000		off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
2001		off += (addr - vma->vm_start) >> shift;
2002		return offset_il_node(pol, off);
2003	} else
2004		return interleave_nodes(pol);
2005}
2006
2007#ifdef CONFIG_HUGETLBFS
2008/*
2009 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2010 * @vma: virtual memory area whose policy is sought
2011 * @addr: address in @vma for shared policy lookup and interleave policy
2012 * @gfp_flags: for requested zone
2013 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2014 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
2015 *
2016 * Returns a nid suitable for a huge page allocation and a pointer
2017 * to the struct mempolicy for conditional unref after allocation.
2018 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
2019 * @nodemask for filtering the zonelist.
2020 *
2021 * Must be protected by read_mems_allowed_begin()
2022 */
2023int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2024				struct mempolicy **mpol, nodemask_t **nodemask)
2025{
2026	int nid;
2027
2028	*mpol = get_vma_policy(vma, addr);
2029	*nodemask = NULL;	/* assume !MPOL_BIND */
2030
2031	if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
2032		nid = interleave_nid(*mpol, vma, addr,
2033					huge_page_shift(hstate_vma(vma)));
2034	} else {
2035		nid = policy_node(gfp_flags, *mpol, numa_node_id());
2036		if ((*mpol)->mode == MPOL_BIND)
2037			*nodemask = &(*mpol)->nodes;
2038	}
2039	return nid;
2040}
2041
2042/*
2043 * init_nodemask_of_mempolicy
2044 *
2045 * If the current task's mempolicy is "default" [NULL], return 'false'
2046 * to indicate default policy.  Otherwise, extract the policy nodemask
2047 * for 'bind' or 'interleave' policy into the argument nodemask, or
2048 * initialize the argument nodemask to contain the single node for
2049 * 'preferred' or 'local' policy and return 'true' to indicate presence
2050 * of non-default mempolicy.
2051 *
2052 * We don't bother with reference counting the mempolicy [mpol_get/put]
2053 * because the current task is examining it's own mempolicy and a task's
2054 * mempolicy is only ever changed by the task itself.
2055 *
2056 * N.B., it is the caller's responsibility to free a returned nodemask.
2057 */
2058bool init_nodemask_of_mempolicy(nodemask_t *mask)
2059{
2060	struct mempolicy *mempolicy;
2061
2062	if (!(mask && current->mempolicy))
2063		return false;
2064
2065	task_lock(current);
2066	mempolicy = current->mempolicy;
2067	switch (mempolicy->mode) {
2068	case MPOL_PREFERRED:
2069	case MPOL_BIND:
2070	case MPOL_INTERLEAVE:
2071		*mask = mempolicy->nodes;
2072		break;
2073
2074	case MPOL_LOCAL:
2075		init_nodemask_of_node(mask, numa_node_id());
2076		break;
2077
2078	default:
2079		BUG();
2080	}
2081	task_unlock(current);
2082
2083	return true;
2084}
2085#endif
2086
2087/*
2088 * mempolicy_in_oom_domain
2089 *
2090 * If tsk's mempolicy is "bind", check for intersection between mask and
2091 * the policy nodemask. Otherwise, return true for all other policies
2092 * including "interleave", as a tsk with "interleave" policy may have
2093 * memory allocated from all nodes in system.
2094 *
2095 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2096 */
2097bool mempolicy_in_oom_domain(struct task_struct *tsk,
2098					const nodemask_t *mask)
2099{
2100	struct mempolicy *mempolicy;
2101	bool ret = true;
2102
2103	if (!mask)
2104		return ret;
2105
2106	task_lock(tsk);
2107	mempolicy = tsk->mempolicy;
2108	if (mempolicy && mempolicy->mode == MPOL_BIND)
2109		ret = nodes_intersects(mempolicy->nodes, *mask);
2110	task_unlock(tsk);
2111
2112	return ret;
2113}
2114
2115/* Allocate a page in interleaved policy.
2116   Own path because it needs to do special accounting. */
2117static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2118					unsigned nid)
2119{
2120	struct page *page;
2121
2122	page = __alloc_pages(gfp, order, nid, NULL);
2123	/* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2124	if (!static_branch_likely(&vm_numa_stat_key))
2125		return page;
2126	if (page && page_to_nid(page) == nid) {
2127		preempt_disable();
2128		__count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2129		preempt_enable();
2130	}
2131	return page;
2132}
2133
2134/**
2135 * alloc_pages_vma - Allocate a page for a VMA.
2136 * @gfp: GFP flags.
2137 * @order: Order of the GFP allocation.
2138 * @vma: Pointer to VMA or NULL if not available.
2139 * @addr: Virtual address of the allocation.  Must be inside @vma.
2140 * @node: Which node to prefer for allocation (modulo policy).
2141 * @hugepage: For hugepages try only the preferred node if possible.
2142 *
2143 * Allocate a page for a specific address in @vma, using the appropriate
2144 * NUMA policy.  When @vma is not NULL the caller must hold the mmap_lock
2145 * of the mm_struct of the VMA to prevent it from going away.  Should be
2146 * used for all allocations for pages that will be mapped into user space.
2147 *
2148 * Return: The page on success or NULL if allocation fails.
2149 */
2150struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2151		unsigned long addr, int node, bool hugepage)
2152{
2153	struct mempolicy *pol;
2154	struct page *page;
2155	int preferred_nid;
2156	nodemask_t *nmask;
2157
2158	pol = get_vma_policy(vma, addr);
2159
2160	if (pol->mode == MPOL_INTERLEAVE) {
2161		unsigned nid;
2162
2163		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2164		mpol_cond_put(pol);
2165		page = alloc_page_interleave(gfp, order, nid);
2166		goto out;
2167	}
2168
2169	if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2170		int hpage_node = node;
2171
2172		/*
2173		 * For hugepage allocation and non-interleave policy which
2174		 * allows the current node (or other explicitly preferred
2175		 * node) we only try to allocate from the current/preferred
2176		 * node and don't fall back to other nodes, as the cost of
2177		 * remote accesses would likely offset THP benefits.
2178		 *
2179		 * If the policy is interleave, or does not allow the current
2180		 * node in its nodemask, we allocate the standard way.
2181		 */
2182		if (pol->mode == MPOL_PREFERRED)
2183			hpage_node = first_node(pol->nodes);
2184
2185		nmask = policy_nodemask(gfp, pol);
2186		if (!nmask || node_isset(hpage_node, *nmask)) {
2187			mpol_cond_put(pol);
2188			/*
2189			 * First, try to allocate THP only on local node, but
2190			 * don't reclaim unnecessarily, just compact.
2191			 */
2192			page = __alloc_pages_node(hpage_node,
2193				gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2194
2195			/*
2196			 * If hugepage allocations are configured to always
2197			 * synchronous compact or the vma has been madvised
2198			 * to prefer hugepage backing, retry allowing remote
2199			 * memory with both reclaim and compact as well.
2200			 */
2201			if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2202				page = __alloc_pages_node(hpage_node,
2203								gfp, order);
2204
2205			goto out;
2206		}
2207	}
2208
2209	nmask = policy_nodemask(gfp, pol);
2210	preferred_nid = policy_node(gfp, pol, node);
2211	page = __alloc_pages(gfp, order, preferred_nid, nmask);
2212	mpol_cond_put(pol);
2213out:
2214	return page;
2215}
2216EXPORT_SYMBOL(alloc_pages_vma);
2217
2218/**
2219 * alloc_pages - Allocate pages.
2220 * @gfp: GFP flags.
2221 * @order: Power of two of number of pages to allocate.
2222 *
2223 * Allocate 1 << @order contiguous pages.  The physical address of the
2224 * first page is naturally aligned (eg an order-3 allocation will be aligned
2225 * to a multiple of 8 * PAGE_SIZE bytes).  The NUMA policy of the current
2226 * process is honoured when in process context.
2227 *
2228 * Context: Can be called from any context, providing the appropriate GFP
2229 * flags are used.
2230 * Return: The page on success or NULL if allocation fails.
2231 */
2232struct page *alloc_pages(gfp_t gfp, unsigned order)
2233{
2234	struct mempolicy *pol = &default_policy;
2235	struct page *page;
2236
2237	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2238		pol = get_task_policy(current);
2239
2240	/*
2241	 * No reference counting needed for current->mempolicy
2242	 * nor system default_policy
2243	 */
2244	if (pol->mode == MPOL_INTERLEAVE)
2245		page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2246	else
2247		page = __alloc_pages(gfp, order,
2248				policy_node(gfp, pol, numa_node_id()),
2249				policy_nodemask(gfp, pol));
2250
2251	return page;
2252}
2253EXPORT_SYMBOL(alloc_pages);
2254
2255int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2256{
2257	struct mempolicy *pol = mpol_dup(vma_policy(src));
2258
2259	if (IS_ERR(pol))
2260		return PTR_ERR(pol);
2261	dst->vm_policy = pol;
2262	return 0;
2263}
2264
2265/*
2266 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2267 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2268 * with the mems_allowed returned by cpuset_mems_allowed().  This
2269 * keeps mempolicies cpuset relative after its cpuset moves.  See
2270 * further kernel/cpuset.c update_nodemask().
2271 *
2272 * current's mempolicy may be rebinded by the other task(the task that changes
2273 * cpuset's mems), so we needn't do rebind work for current task.
2274 */
2275
2276/* Slow path of a mempolicy duplicate */
2277struct mempolicy *__mpol_dup(struct mempolicy *old)
2278{
2279	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2280
2281	if (!new)
2282		return ERR_PTR(-ENOMEM);
2283
2284	/* task's mempolicy is protected by alloc_lock */
2285	if (old == current->mempolicy) {
2286		task_lock(current);
2287		*new = *old;
2288		task_unlock(current);
2289	} else
2290		*new = *old;
2291
2292	if (current_cpuset_is_being_rebound()) {
2293		nodemask_t mems = cpuset_mems_allowed(current);
2294		mpol_rebind_policy(new, &mems);
2295	}
2296	atomic_set(&new->refcnt, 1);
2297	return new;
2298}
2299
2300/* Slow path of a mempolicy comparison */
2301bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2302{
2303	if (!a || !b)
2304		return false;
2305	if (a->mode != b->mode)
2306		return false;
2307	if (a->flags != b->flags)
2308		return false;
2309	if (mpol_store_user_nodemask(a))
2310		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2311			return false;
2312
2313	switch (a->mode) {
2314	case MPOL_BIND:
2315	case MPOL_INTERLEAVE:
2316	case MPOL_PREFERRED:
2317		return !!nodes_equal(a->nodes, b->nodes);
2318	case MPOL_LOCAL:
2319		return true;
2320	default:
2321		BUG();
2322		return false;
2323	}
2324}
2325
2326/*
2327 * Shared memory backing store policy support.
2328 *
2329 * Remember policies even when nobody has shared memory mapped.
2330 * The policies are kept in Red-Black tree linked from the inode.
2331 * They are protected by the sp->lock rwlock, which should be held
2332 * for any accesses to the tree.
2333 */
2334
2335/*
2336 * lookup first element intersecting start-end.  Caller holds sp->lock for
2337 * reading or for writing
2338 */
2339static struct sp_node *
2340sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2341{
2342	struct rb_node *n = sp->root.rb_node;
2343
2344	while (n) {
2345		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2346
2347		if (start >= p->end)
2348			n = n->rb_right;
2349		else if (end <= p->start)
2350			n = n->rb_left;
2351		else
2352			break;
2353	}
2354	if (!n)
2355		return NULL;
2356	for (;;) {
2357		struct sp_node *w = NULL;
2358		struct rb_node *prev = rb_prev(n);
2359		if (!prev)
2360			break;
2361		w = rb_entry(prev, struct sp_node, nd);
2362		if (w->end <= start)
2363			break;
2364		n = prev;
2365	}
2366	return rb_entry(n, struct sp_node, nd);
2367}
2368
2369/*
2370 * Insert a new shared policy into the list.  Caller holds sp->lock for
2371 * writing.
2372 */
2373static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2374{
2375	struct rb_node **p = &sp->root.rb_node;
2376	struct rb_node *parent = NULL;
2377	struct sp_node *nd;
2378
2379	while (*p) {
2380		parent = *p;
2381		nd = rb_entry(parent, struct sp_node, nd);
2382		if (new->start < nd->start)
2383			p = &(*p)->rb_left;
2384		else if (new->end > nd->end)
2385			p = &(*p)->rb_right;
2386		else
2387			BUG();
2388	}
2389	rb_link_node(&new->nd, parent, p);
2390	rb_insert_color(&new->nd, &sp->root);
2391	pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2392		 new->policy ? new->policy->mode : 0);
2393}
2394
2395/* Find shared policy intersecting idx */
2396struct mempolicy *
2397mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2398{
2399	struct mempolicy *pol = NULL;
2400	struct sp_node *sn;
2401
2402	if (!sp->root.rb_node)
2403		return NULL;
2404	read_lock(&sp->lock);
2405	sn = sp_lookup(sp, idx, idx+1);
2406	if (sn) {
2407		mpol_get(sn->policy);
2408		pol = sn->policy;
2409	}
2410	read_unlock(&sp->lock);
2411	return pol;
2412}
2413
2414static void sp_free(struct sp_node *n)
2415{
2416	mpol_put(n->policy);
2417	kmem_cache_free(sn_cache, n);
2418}
2419
2420/**
2421 * mpol_misplaced - check whether current page node is valid in policy
2422 *
2423 * @page: page to be checked
2424 * @vma: vm area where page mapped
2425 * @addr: virtual address where page mapped
2426 *
2427 * Lookup current policy node id for vma,addr and "compare to" page's
2428 * node id.  Policy determination "mimics" alloc_page_vma().
2429 * Called from fault path where we know the vma and faulting address.
2430 *
2431 * Return: -1 if the page is in a node that is valid for this policy, or a
2432 * suitable node ID to allocate a replacement page from.
2433 */
2434int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2435{
2436	struct mempolicy *pol;
2437	struct zoneref *z;
2438	int curnid = page_to_nid(page);
2439	unsigned long pgoff;
2440	int thiscpu = raw_smp_processor_id();
2441	int thisnid = cpu_to_node(thiscpu);
2442	int polnid = NUMA_NO_NODE;
2443	int ret = -1;
2444
2445	pol = get_vma_policy(vma, addr);
2446	if (!(pol->flags & MPOL_F_MOF))
2447		goto out;
2448
2449	switch (pol->mode) {
2450	case MPOL_INTERLEAVE:
2451		pgoff = vma->vm_pgoff;
2452		pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2453		polnid = offset_il_node(pol, pgoff);
2454		break;
2455
2456	case MPOL_PREFERRED:
2457		polnid = first_node(pol->nodes);
2458		break;
2459
2460	case MPOL_LOCAL:
2461		polnid = numa_node_id();
2462		break;
2463
2464	case MPOL_BIND:
2465		/* Optimize placement among multiple nodes via NUMA balancing */
2466		if (pol->flags & MPOL_F_MORON) {
2467			if (node_isset(thisnid, pol->nodes))
2468				break;
2469			goto out;
2470		}
2471
2472		/*
2473		 * allows binding to multiple nodes.
2474		 * use current page if in policy nodemask,
2475		 * else select nearest allowed node, if any.
2476		 * If no allowed nodes, use current [!misplaced].
2477		 */
2478		if (node_isset(curnid, pol->nodes))
2479			goto out;
2480		z = first_zones_zonelist(
2481				node_zonelist(numa_node_id(), GFP_HIGHUSER),
2482				gfp_zone(GFP_HIGHUSER),
2483				&pol->nodes);
2484		polnid = zone_to_nid(z->zone);
2485		break;
2486
2487	default:
2488		BUG();
2489	}
2490
2491	/* Migrate the page towards the node whose CPU is referencing it */
2492	if (pol->flags & MPOL_F_MORON) {
2493		polnid = thisnid;
2494
2495		if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2496			goto out;
2497	}
2498
2499	if (curnid != polnid)
2500		ret = polnid;
2501out:
2502	mpol_cond_put(pol);
2503
2504	return ret;
2505}
2506
2507/*
2508 * Drop the (possibly final) reference to task->mempolicy.  It needs to be
2509 * dropped after task->mempolicy is set to NULL so that any allocation done as
2510 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2511 * policy.
2512 */
2513void mpol_put_task_policy(struct task_struct *task)
2514{
2515	struct mempolicy *pol;
2516
2517	task_lock(task);
2518	pol = task->mempolicy;
2519	task->mempolicy = NULL;
2520	task_unlock(task);
2521	mpol_put(pol);
2522}
2523
2524static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2525{
2526	pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2527	rb_erase(&n->nd, &sp->root);
2528	sp_free(n);
2529}
2530
2531static void sp_node_init(struct sp_node *node, unsigned long start,
2532			unsigned long end, struct mempolicy *pol)
2533{
2534	node->start = start;
2535	node->end = end;
2536	node->policy = pol;
2537}
2538
2539static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2540				struct mempolicy *pol)
2541{
2542	struct sp_node *n;
2543	struct mempolicy *newpol;
2544
2545	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2546	if (!n)
2547		return NULL;
2548
2549	newpol = mpol_dup(pol);
2550	if (IS_ERR(newpol)) {
2551		kmem_cache_free(sn_cache, n);
2552		return NULL;
2553	}
2554	newpol->flags |= MPOL_F_SHARED;
2555	sp_node_init(n, start, end, newpol);
2556
2557	return n;
2558}
2559
2560/* Replace a policy range. */
2561static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2562				 unsigned long end, struct sp_node *new)
2563{
2564	struct sp_node *n;
2565	struct sp_node *n_new = NULL;
2566	struct mempolicy *mpol_new = NULL;
2567	int ret = 0;
2568
2569restart:
2570	write_lock(&sp->lock);
2571	n = sp_lookup(sp, start, end);
2572	/* Take care of old policies in the same range. */
2573	while (n && n->start < end) {
2574		struct rb_node *next = rb_next(&n->nd);
2575		if (n->start >= start) {
2576			if (n->end <= end)
2577				sp_delete(sp, n);
2578			else
2579				n->start = end;
2580		} else {
2581			/* Old policy spanning whole new range. */
2582			if (n->end > end) {
2583				if (!n_new)
2584					goto alloc_new;
2585
2586				*mpol_new = *n->policy;
2587				atomic_set(&mpol_new->refcnt, 1);
2588				sp_node_init(n_new, end, n->end, mpol_new);
2589				n->end = start;
2590				sp_insert(sp, n_new);
2591				n_new = NULL;
2592				mpol_new = NULL;
2593				break;
2594			} else
2595				n->end = start;
2596		}
2597		if (!next)
2598			break;
2599		n = rb_entry(next, struct sp_node, nd);
2600	}
2601	if (new)
2602		sp_insert(sp, new);
2603	write_unlock(&sp->lock);
2604	ret = 0;
2605
2606err_out:
2607	if (mpol_new)
2608		mpol_put(mpol_new);
2609	if (n_new)
2610		kmem_cache_free(sn_cache, n_new);
2611
2612	return ret;
2613
2614alloc_new:
2615	write_unlock(&sp->lock);
2616	ret = -ENOMEM;
2617	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2618	if (!n_new)
2619		goto err_out;
2620	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2621	if (!mpol_new)
2622		goto err_out;
2623	goto restart;
2624}
2625
2626/**
2627 * mpol_shared_policy_init - initialize shared policy for inode
2628 * @sp: pointer to inode shared policy
2629 * @mpol:  struct mempolicy to install
2630 *
2631 * Install non-NULL @mpol in inode's shared policy rb-tree.
2632 * On entry, the current task has a reference on a non-NULL @mpol.
2633 * This must be released on exit.
2634 * This is called at get_inode() calls and we can use GFP_KERNEL.
2635 */
2636void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2637{
2638	int ret;
2639
2640	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2641	rwlock_init(&sp->lock);
2642
2643	if (mpol) {
2644		struct vm_area_struct pvma;
2645		struct mempolicy *new;
2646		NODEMASK_SCRATCH(scratch);
2647
2648		if (!scratch)
2649			goto put_mpol;
2650		/* contextualize the tmpfs mount point mempolicy */
2651		new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2652		if (IS_ERR(new))
2653			goto free_scratch; /* no valid nodemask intersection */
2654
2655		task_lock(current);
2656		ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2657		task_unlock(current);
2658		if (ret)
2659			goto put_new;
2660
2661		/* Create pseudo-vma that contains just the policy */
2662		vma_init(&pvma, NULL);
2663		pvma.vm_end = TASK_SIZE;	/* policy covers entire file */
2664		mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2665
2666put_new:
2667		mpol_put(new);			/* drop initial ref */
2668free_scratch:
2669		NODEMASK_SCRATCH_FREE(scratch);
2670put_mpol:
2671		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
2672	}
2673}
2674
2675int mpol_set_shared_policy(struct shared_policy *info,
2676			struct vm_area_struct *vma, struct mempolicy *npol)
2677{
2678	int err;
2679	struct sp_node *new = NULL;
2680	unsigned long sz = vma_pages(vma);
2681
2682	pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2683		 vma->vm_pgoff,
2684		 sz, npol ? npol->mode : -1,
2685		 npol ? npol->flags : -1,
2686		 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2687
2688	if (npol) {
2689		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2690		if (!new)
2691			return -ENOMEM;
2692	}
2693	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2694	if (err && new)
2695		sp_free(new);
2696	return err;
2697}
2698
2699/* Free a backing policy store on inode delete. */
2700void mpol_free_shared_policy(struct shared_policy *p)
2701{
2702	struct sp_node *n;
2703	struct rb_node *next;
2704
2705	if (!p->root.rb_node)
2706		return;
2707	write_lock(&p->lock);
2708	next = rb_first(&p->root);
2709	while (next) {
2710		n = rb_entry(next, struct sp_node, nd);
2711		next = rb_next(&n->nd);
2712		sp_delete(p, n);
2713	}
2714	write_unlock(&p->lock);
2715}
2716
2717#ifdef CONFIG_NUMA_BALANCING
2718static int __initdata numabalancing_override;
2719
2720static void __init check_numabalancing_enable(void)
2721{
2722	bool numabalancing_default = false;
2723
2724	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2725		numabalancing_default = true;
2726
2727	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2728	if (numabalancing_override)
2729		set_numabalancing_state(numabalancing_override == 1);
2730
2731	if (num_online_nodes() > 1 && !numabalancing_override) {
2732		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2733			numabalancing_default ? "Enabling" : "Disabling");
2734		set_numabalancing_state(numabalancing_default);
2735	}
2736}
2737
2738static int __init setup_numabalancing(char *str)
2739{
2740	int ret = 0;
2741	if (!str)
2742		goto out;
2743
2744	if (!strcmp(str, "enable")) {
2745		numabalancing_override = 1;
2746		ret = 1;
2747	} else if (!strcmp(str, "disable")) {
2748		numabalancing_override = -1;
2749		ret = 1;
2750	}
2751out:
2752	if (!ret)
2753		pr_warn("Unable to parse numa_balancing=\n");
2754
2755	return ret;
2756}
2757__setup("numa_balancing=", setup_numabalancing);
2758#else
2759static inline void __init check_numabalancing_enable(void)
2760{
2761}
2762#endif /* CONFIG_NUMA_BALANCING */
2763
2764/* assumes fs == KERNEL_DS */
2765void __init numa_policy_init(void)
2766{
2767	nodemask_t interleave_nodes;
2768	unsigned long largest = 0;
2769	int nid, prefer = 0;
2770
2771	policy_cache = kmem_cache_create("numa_policy",
2772					 sizeof(struct mempolicy),
2773					 0, SLAB_PANIC, NULL);
2774
2775	sn_cache = kmem_cache_create("shared_policy_node",
2776				     sizeof(struct sp_node),
2777				     0, SLAB_PANIC, NULL);
2778
2779	for_each_node(nid) {
2780		preferred_node_policy[nid] = (struct mempolicy) {
2781			.refcnt = ATOMIC_INIT(1),
2782			.mode = MPOL_PREFERRED,
2783			.flags = MPOL_F_MOF | MPOL_F_MORON,
2784			.nodes = nodemask_of_node(nid),
2785		};
2786	}
2787
2788	/*
2789	 * Set interleaving policy for system init. Interleaving is only
2790	 * enabled across suitably sized nodes (default is >= 16MB), or
2791	 * fall back to the largest node if they're all smaller.
2792	 */
2793	nodes_clear(interleave_nodes);
2794	for_each_node_state(nid, N_MEMORY) {
2795		unsigned long total_pages = node_present_pages(nid);
2796
2797		/* Preserve the largest node */
2798		if (largest < total_pages) {
2799			largest = total_pages;
2800			prefer = nid;
2801		}
2802
2803		/* Interleave this node? */
2804		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2805			node_set(nid, interleave_nodes);
2806	}
2807
2808	/* All too small, use the largest */
2809	if (unlikely(nodes_empty(interleave_nodes)))
2810		node_set(prefer, interleave_nodes);
2811
2812	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2813		pr_err("%s: interleaving failed\n", __func__);
2814
2815	check_numabalancing_enable();
2816}
2817
2818/* Reset policy of current process to default */
2819void numa_default_policy(void)
2820{
2821	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2822}
2823
2824/*
2825 * Parse and format mempolicy from/to strings
2826 */
2827
2828static const char * const policy_modes[] =
2829{
2830	[MPOL_DEFAULT]    = "default",
2831	[MPOL_PREFERRED]  = "prefer",
2832	[MPOL_BIND]       = "bind",
2833	[MPOL_INTERLEAVE] = "interleave",
2834	[MPOL_LOCAL]      = "local",
2835};
2836
2837
2838#ifdef CONFIG_TMPFS
2839/**
2840 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2841 * @str:  string containing mempolicy to parse
2842 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2843 *
2844 * Format of input:
2845 *	<mode>[=<flags>][:<nodelist>]
2846 *
2847 * On success, returns 0, else 1
2848 */
2849int mpol_parse_str(char *str, struct mempolicy **mpol)
2850{
2851	struct mempolicy *new = NULL;
2852	unsigned short mode_flags;
2853	nodemask_t nodes;
2854	char *nodelist = strchr(str, ':');
2855	char *flags = strchr(str, '=');
2856	int err = 1, mode;
2857
2858	if (flags)
2859		*flags++ = '\0';	/* terminate mode string */
2860
2861	if (nodelist) {
2862		/* NUL-terminate mode or flags string */
2863		*nodelist++ = '\0';
2864		if (nodelist_parse(nodelist, nodes))
2865			goto out;
2866		if (!nodes_subset(nodes, node_states[N_MEMORY]))
2867			goto out;
2868	} else
2869		nodes_clear(nodes);
2870
2871	mode = match_string(policy_modes, MPOL_MAX, str);
2872	if (mode < 0)
2873		goto out;
2874
2875	switch (mode) {
2876	case MPOL_PREFERRED:
2877		/*
2878		 * Insist on a nodelist of one node only, although later
2879		 * we use first_node(nodes) to grab a single node, so here
2880		 * nodelist (or nodes) cannot be empty.
2881		 */
2882		if (nodelist) {
2883			char *rest = nodelist;
2884			while (isdigit(*rest))
2885				rest++;
2886			if (*rest)
2887				goto out;
2888			if (nodes_empty(nodes))
2889				goto out;
2890		}
2891		break;
2892	case MPOL_INTERLEAVE:
2893		/*
2894		 * Default to online nodes with memory if no nodelist
2895		 */
2896		if (!nodelist)
2897			nodes = node_states[N_MEMORY];
2898		break;
2899	case MPOL_LOCAL:
2900		/*
2901		 * Don't allow a nodelist;  mpol_new() checks flags
2902		 */
2903		if (nodelist)
2904			goto out;
2905		break;
2906	case MPOL_DEFAULT:
2907		/*
2908		 * Insist on a empty nodelist
2909		 */
2910		if (!nodelist)
2911			err = 0;
2912		goto out;
2913	case MPOL_BIND:
2914		/*
2915		 * Insist on a nodelist
2916		 */
2917		if (!nodelist)
2918			goto out;
2919	}
2920
2921	mode_flags = 0;
2922	if (flags) {
2923		/*
2924		 * Currently, we only support two mutually exclusive
2925		 * mode flags.
2926		 */
2927		if (!strcmp(flags, "static"))
2928			mode_flags |= MPOL_F_STATIC_NODES;
2929		else if (!strcmp(flags, "relative"))
2930			mode_flags |= MPOL_F_RELATIVE_NODES;
2931		else
2932			goto out;
2933	}
2934
2935	new = mpol_new(mode, mode_flags, &nodes);
2936	if (IS_ERR(new))
2937		goto out;
2938
2939	/*
2940	 * Save nodes for mpol_to_str() to show the tmpfs mount options
2941	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2942	 */
2943	if (mode != MPOL_PREFERRED) {
2944		new->nodes = nodes;
2945	} else if (nodelist) {
2946		nodes_clear(new->nodes);
2947		node_set(first_node(nodes), new->nodes);
2948	} else {
2949		new->mode = MPOL_LOCAL;
2950	}
2951
2952	/*
2953	 * Save nodes for contextualization: this will be used to "clone"
2954	 * the mempolicy in a specific context [cpuset] at a later time.
2955	 */
2956	new->w.user_nodemask = nodes;
2957
2958	err = 0;
2959
2960out:
2961	/* Restore string for error message */
2962	if (nodelist)
2963		*--nodelist = ':';
2964	if (flags)
2965		*--flags = '=';
2966	if (!err)
2967		*mpol = new;
2968	return err;
2969}
2970#endif /* CONFIG_TMPFS */
2971
2972/**
2973 * mpol_to_str - format a mempolicy structure for printing
2974 * @buffer:  to contain formatted mempolicy string
2975 * @maxlen:  length of @buffer
2976 * @pol:  pointer to mempolicy to be formatted
2977 *
2978 * Convert @pol into a string.  If @buffer is too short, truncate the string.
2979 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2980 * longest flag, "relative", and to display at least a few node ids.
2981 */
2982void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2983{
2984	char *p = buffer;
2985	nodemask_t nodes = NODE_MASK_NONE;
2986	unsigned short mode = MPOL_DEFAULT;
2987	unsigned short flags = 0;
2988
2989	if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2990		mode = pol->mode;
2991		flags = pol->flags;
2992	}
2993
2994	switch (mode) {
2995	case MPOL_DEFAULT:
2996	case MPOL_LOCAL:
2997		break;
2998	case MPOL_PREFERRED:
2999	case MPOL_BIND:
3000	case MPOL_INTERLEAVE:
3001		nodes = pol->nodes;
3002		break;
3003	default:
3004		WARN_ON_ONCE(1);
3005		snprintf(p, maxlen, "unknown");
3006		return;
3007	}
3008
3009	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3010
3011	if (flags & MPOL_MODE_FLAGS) {
3012		p += snprintf(p, buffer + maxlen - p, "=");
3013
3014		/*
3015		 * Currently, the only defined flags are mutually exclusive
3016		 */
3017		if (flags & MPOL_F_STATIC_NODES)
3018			p += snprintf(p, buffer + maxlen - p, "static");
3019		else if (flags & MPOL_F_RELATIVE_NODES)
3020			p += snprintf(p, buffer + maxlen - p, "relative");
3021	}
3022
3023	if (!nodes_empty(nodes))
3024		p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3025			       nodemask_pr_args(&nodes));
3026}