Linux Audio

Check our new training course

Loading...
v5.4
   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_PREFERRED,
 125	.flags = MPOL_F_LOCAL,
 126};
 127
 128static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 129
 130struct mempolicy *get_task_policy(struct task_struct *p)
 131{
 132	struct mempolicy *pol = p->mempolicy;
 133	int node;
 134
 135	if (pol)
 136		return pol;
 137
 138	node = numa_node_id();
 139	if (node != NUMA_NO_NODE) {
 140		pol = &preferred_node_policy[node];
 141		/* preferred_node_policy is not initialised early in boot */
 142		if (pol->mode)
 143			return pol;
 144	}
 145
 146	return &default_policy;
 147}
 148
 149static const struct mempolicy_operations {
 150	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 151	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
 152} mpol_ops[MPOL_MAX];
 153
 154static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 155{
 156	return pol->flags & MPOL_MODE_FLAGS;
 157}
 158
 159static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 160				   const nodemask_t *rel)
 161{
 162	nodemask_t tmp;
 163	nodes_fold(tmp, *orig, nodes_weight(*rel));
 164	nodes_onto(*ret, tmp, *rel);
 165}
 166
 167static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
 168{
 169	if (nodes_empty(*nodes))
 170		return -EINVAL;
 171	pol->v.nodes = *nodes;
 172	return 0;
 173}
 174
 175static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 176{
 177	if (!nodes)
 178		pol->flags |= MPOL_F_LOCAL;	/* local allocation */
 179	else if (nodes_empty(*nodes))
 180		return -EINVAL;			/*  no allowed nodes */
 181	else
 182		pol->v.preferred_node = first_node(*nodes);
 183	return 0;
 184}
 185
 186static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
 187{
 188	if (nodes_empty(*nodes))
 189		return -EINVAL;
 190	pol->v.nodes = *nodes;
 191	return 0;
 192}
 193
 194/*
 195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 196 * any, for the new policy.  mpol_new() has already validated the nodes
 197 * parameter with respect to the policy mode and flags.  But, we need to
 198 * handle an empty nodemask with MPOL_PREFERRED here.
 199 *
 200 * Must be called holding task's alloc_lock to protect task's mems_allowed
 201 * and mempolicy.  May also be called holding the mmap_semaphore for write.
 202 */
 203static int mpol_set_nodemask(struct mempolicy *pol,
 204		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 205{
 206	int ret;
 207
 208	/* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
 209	if (pol == NULL)
 210		return 0;
 211	/* Check N_MEMORY */
 212	nodes_and(nsc->mask1,
 213		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 214
 215	VM_BUG_ON(!nodes);
 216	if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
 217		nodes = NULL;	/* explicit local allocation */
 218	else {
 219		if (pol->flags & MPOL_F_RELATIVE_NODES)
 220			mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
 221		else
 222			nodes_and(nsc->mask2, *nodes, nsc->mask1);
 223
 224		if (mpol_store_user_nodemask(pol))
 225			pol->w.user_nodemask = *nodes;
 226		else
 227			pol->w.cpuset_mems_allowed =
 228						cpuset_current_mems_allowed;
 229	}
 230
 231	if (nodes)
 232		ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 233	else
 234		ret = mpol_ops[pol->mode].create(pol, NULL);
 235	return ret;
 236}
 237
 238/*
 239 * This function just creates a new policy, does some check and simple
 240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 241 */
 242static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 243				  nodemask_t *nodes)
 244{
 245	struct mempolicy *policy;
 246
 247	pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 248		 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
 249
 250	if (mode == MPOL_DEFAULT) {
 251		if (nodes && !nodes_empty(*nodes))
 252			return ERR_PTR(-EINVAL);
 253		return NULL;
 254	}
 255	VM_BUG_ON(!nodes);
 256
 257	/*
 258	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 259	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 260	 * All other modes require a valid pointer to a non-empty nodemask.
 261	 */
 262	if (mode == MPOL_PREFERRED) {
 263		if (nodes_empty(*nodes)) {
 264			if (((flags & MPOL_F_STATIC_NODES) ||
 265			     (flags & MPOL_F_RELATIVE_NODES)))
 266				return ERR_PTR(-EINVAL);
 267		}
 268	} else if (mode == MPOL_LOCAL) {
 269		if (!nodes_empty(*nodes) ||
 270		    (flags & MPOL_F_STATIC_NODES) ||
 271		    (flags & MPOL_F_RELATIVE_NODES))
 272			return ERR_PTR(-EINVAL);
 273		mode = MPOL_PREFERRED;
 274	} else if (nodes_empty(*nodes))
 275		return ERR_PTR(-EINVAL);
 276	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 277	if (!policy)
 278		return ERR_PTR(-ENOMEM);
 279	atomic_set(&policy->refcnt, 1);
 280	policy->mode = mode;
 281	policy->flags = flags;
 282
 283	return policy;
 284}
 285
 286/* Slow path of a mpol destructor. */
 287void __mpol_put(struct mempolicy *p)
 288{
 289	if (!atomic_dec_and_test(&p->refcnt))
 290		return;
 291	kmem_cache_free(policy_cache, p);
 292}
 293
 294static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
 295{
 296}
 297
 298static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 299{
 300	nodemask_t tmp;
 301
 302	if (pol->flags & MPOL_F_STATIC_NODES)
 303		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 304	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 305		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 306	else {
 307		nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
 308								*nodes);
 309		pol->w.cpuset_mems_allowed = *nodes;
 310	}
 311
 312	if (nodes_empty(tmp))
 313		tmp = *nodes;
 314
 315	pol->v.nodes = tmp;
 316}
 317
 318static void mpol_rebind_preferred(struct mempolicy *pol,
 319						const nodemask_t *nodes)
 320{
 321	nodemask_t tmp;
 322
 323	if (pol->flags & MPOL_F_STATIC_NODES) {
 324		int node = first_node(pol->w.user_nodemask);
 325
 326		if (node_isset(node, *nodes)) {
 327			pol->v.preferred_node = node;
 328			pol->flags &= ~MPOL_F_LOCAL;
 329		} else
 330			pol->flags |= MPOL_F_LOCAL;
 331	} else if (pol->flags & MPOL_F_RELATIVE_NODES) {
 332		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 333		pol->v.preferred_node = first_node(tmp);
 334	} else if (!(pol->flags & MPOL_F_LOCAL)) {
 335		pol->v.preferred_node = node_remap(pol->v.preferred_node,
 336						   pol->w.cpuset_mems_allowed,
 337						   *nodes);
 338		pol->w.cpuset_mems_allowed = *nodes;
 339	}
 340}
 341
 342/*
 343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 344 *
 345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
 346 * policies are protected by task->mems_allowed_seq to prevent a premature
 347 * OOM/allocation failure due to parallel nodemask modification.
 348 */
 349static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
 350{
 351	if (!pol)
 352		return;
 353	if (!mpol_store_user_nodemask(pol) && !(pol->flags & MPOL_F_LOCAL) &&
 354	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 355		return;
 356
 357	mpol_ops[pol->mode].rebind(pol, newmask);
 358}
 359
 360/*
 361 * Wrapper for mpol_rebind_policy() that just requires task
 362 * pointer, and updates task mempolicy.
 363 *
 364 * Called with task's alloc_lock held.
 365 */
 366
 367void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
 368{
 369	mpol_rebind_policy(tsk->mempolicy, new);
 370}
 371
 372/*
 373 * Rebind each vma in mm to new nodemask.
 374 *
 375 * Call holding a reference to mm.  Takes mm->mmap_sem during call.
 376 */
 377
 378void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 379{
 380	struct vm_area_struct *vma;
 381
 382	down_write(&mm->mmap_sem);
 383	for (vma = mm->mmap; vma; vma = vma->vm_next)
 384		mpol_rebind_policy(vma->vm_policy, new);
 385	up_write(&mm->mmap_sem);
 386}
 387
 388static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 389	[MPOL_DEFAULT] = {
 390		.rebind = mpol_rebind_default,
 391	},
 392	[MPOL_INTERLEAVE] = {
 393		.create = mpol_new_interleave,
 394		.rebind = mpol_rebind_nodemask,
 395	},
 396	[MPOL_PREFERRED] = {
 397		.create = mpol_new_preferred,
 398		.rebind = mpol_rebind_preferred,
 399	},
 400	[MPOL_BIND] = {
 401		.create = mpol_new_bind,
 402		.rebind = mpol_rebind_nodemask,
 403	},
 404};
 405
 406static int migrate_page_add(struct page *page, struct list_head *pagelist,
 407				unsigned long flags);
 408
 409struct queue_pages {
 410	struct list_head *pagelist;
 411	unsigned long flags;
 412	nodemask_t *nmask;
 413	struct vm_area_struct *prev;
 414};
 415
 416/*
 417 * Check if the page's nid is in qp->nmask.
 418 *
 419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
 420 * in the invert of qp->nmask.
 421 */
 422static inline bool queue_pages_required(struct page *page,
 423					struct queue_pages *qp)
 424{
 425	int nid = page_to_nid(page);
 426	unsigned long flags = qp->flags;
 427
 428	return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
 429}
 430
 431/*
 432 * queue_pages_pmd() has four possible return values:
 433 * 0 - pages are placed on the right node or queued successfully.
 434 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 435 *     specified.
 436 * 2 - THP was split.
 437 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
 438 *        existing page was already on a node that does not follow the
 439 *        policy.
 440 */
 441static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
 442				unsigned long end, struct mm_walk *walk)
 443{
 444	int ret = 0;
 445	struct page *page;
 446	struct queue_pages *qp = walk->private;
 447	unsigned long flags;
 448
 449	if (unlikely(is_pmd_migration_entry(*pmd))) {
 450		ret = -EIO;
 451		goto unlock;
 452	}
 453	page = pmd_page(*pmd);
 454	if (is_huge_zero_page(page)) {
 455		spin_unlock(ptl);
 456		__split_huge_pmd(walk->vma, pmd, addr, false, NULL);
 457		ret = 2;
 458		goto out;
 459	}
 460	if (!queue_pages_required(page, qp))
 
 461		goto unlock;
 
 462
 
 463	flags = qp->flags;
 464	/* go to thp migration */
 465	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 466		if (!vma_migratable(walk->vma) ||
 467		    migrate_page_add(page, qp->pagelist, flags)) {
 468			ret = 1;
 469			goto unlock;
 470		}
 471	} else
 472		ret = -EIO;
 473unlock:
 474	spin_unlock(ptl);
 475out:
 476	return ret;
 477}
 478
 479/*
 480 * Scan through pages checking if pages follow certain conditions,
 481 * and move them to the pagelist if they do.
 482 *
 483 * queue_pages_pte_range() has three possible return values:
 484 * 0 - pages are placed on the right node or queued successfully.
 485 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 486 *     specified.
 487 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
 488 *        on a node that does not follow the policy.
 489 */
 490static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
 491			unsigned long end, struct mm_walk *walk)
 492{
 493	struct vm_area_struct *vma = walk->vma;
 494	struct page *page;
 495	struct queue_pages *qp = walk->private;
 496	unsigned long flags = qp->flags;
 497	int ret;
 498	bool has_unmovable = false;
 499	pte_t *pte;
 500	spinlock_t *ptl;
 501
 502	ptl = pmd_trans_huge_lock(pmd, vma);
 503	if (ptl) {
 504		ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
 505		if (ret != 2)
 506			return ret;
 507	}
 508	/* THP was split, fall through to pte walk */
 509
 510	if (pmd_trans_unstable(pmd))
 511		return 0;
 512
 513	pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
 514	for (; addr != end; pte++, addr += PAGE_SIZE) {
 515		if (!pte_present(*pte))
 516			continue;
 517		page = vm_normal_page(vma, addr, *pte);
 518		if (!page)
 519			continue;
 520		/*
 521		 * vm_normal_page() filters out zero pages, but there might
 522		 * still be PageReserved pages to skip, perhaps in a VDSO.
 523		 */
 524		if (PageReserved(page))
 525			continue;
 526		if (!queue_pages_required(page, qp))
 527			continue;
 528		if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 529			/* MPOL_MF_STRICT must be specified if we get here */
 530			if (!vma_migratable(vma)) {
 531				has_unmovable = true;
 532				break;
 533			}
 534
 535			/*
 536			 * Do not abort immediately since there may be
 537			 * temporary off LRU pages in the range.  Still
 538			 * need migrate other LRU pages.
 539			 */
 540			if (migrate_page_add(page, qp->pagelist, flags))
 541				has_unmovable = true;
 542		} else
 543			break;
 544	}
 545	pte_unmap_unlock(pte - 1, ptl);
 546	cond_resched();
 547
 548	if (has_unmovable)
 549		return 1;
 550
 551	return addr != end ? -EIO : 0;
 552}
 553
 554static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
 555			       unsigned long addr, unsigned long end,
 556			       struct mm_walk *walk)
 557{
 558#ifdef CONFIG_HUGETLB_PAGE
 559	struct queue_pages *qp = walk->private;
 560	unsigned long flags = qp->flags;
 561	struct page *page;
 562	spinlock_t *ptl;
 563	pte_t entry;
 564
 565	ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 566	entry = huge_ptep_get(pte);
 567	if (!pte_present(entry))
 568		goto unlock;
 569	page = pte_page(entry);
 570	if (!queue_pages_required(page, qp))
 571		goto unlock;
 572	/* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
 573	if (flags & (MPOL_MF_MOVE_ALL) ||
 574	    (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
 575		isolate_huge_page(page, qp->pagelist);
 576unlock:
 577	spin_unlock(ptl);
 578#else
 579	BUG();
 580#endif
 581	return 0;
 582}
 583
 584#ifdef CONFIG_NUMA_BALANCING
 585/*
 586 * This is used to mark a range of virtual addresses to be inaccessible.
 587 * These are later cleared by a NUMA hinting fault. Depending on these
 588 * faults, pages may be migrated for better NUMA placement.
 589 *
 590 * This is assuming that NUMA faults are handled using PROT_NONE. If
 591 * an architecture makes a different choice, it will need further
 592 * changes to the core.
 593 */
 594unsigned long change_prot_numa(struct vm_area_struct *vma,
 595			unsigned long addr, unsigned long end)
 596{
 597	int nr_updated;
 598
 599	nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
 600	if (nr_updated)
 601		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 602
 603	return nr_updated;
 604}
 605#else
 606static unsigned long change_prot_numa(struct vm_area_struct *vma,
 607			unsigned long addr, unsigned long end)
 608{
 609	return 0;
 610}
 611#endif /* CONFIG_NUMA_BALANCING */
 612
 613static int queue_pages_test_walk(unsigned long start, unsigned long end,
 614				struct mm_walk *walk)
 615{
 616	struct vm_area_struct *vma = walk->vma;
 617	struct queue_pages *qp = walk->private;
 618	unsigned long endvma = vma->vm_end;
 619	unsigned long flags = qp->flags;
 620
 621	/*
 622	 * Need check MPOL_MF_STRICT to return -EIO if possible
 623	 * regardless of vma_migratable
 624	 */
 625	if (!vma_migratable(vma) &&
 626	    !(flags & MPOL_MF_STRICT))
 627		return 1;
 628
 629	if (endvma > end)
 630		endvma = end;
 631	if (vma->vm_start > start)
 632		start = vma->vm_start;
 633
 634	if (!(flags & MPOL_MF_DISCONTIG_OK)) {
 635		if (!vma->vm_next && vma->vm_end < end)
 636			return -EFAULT;
 637		if (qp->prev && qp->prev->vm_end < vma->vm_start)
 638			return -EFAULT;
 639	}
 640
 641	qp->prev = vma;
 642
 643	if (flags & MPOL_MF_LAZY) {
 644		/* Similar to task_numa_work, skip inaccessible VMAs */
 645		if (!is_vm_hugetlb_page(vma) &&
 646			(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
 647			!(vma->vm_flags & VM_MIXEDMAP))
 648			change_prot_numa(vma, start, endvma);
 649		return 1;
 650	}
 651
 652	/* queue pages from current vma */
 653	if (flags & MPOL_MF_VALID)
 654		return 0;
 655	return 1;
 656}
 657
 658static const struct mm_walk_ops queue_pages_walk_ops = {
 659	.hugetlb_entry		= queue_pages_hugetlb,
 660	.pmd_entry		= queue_pages_pte_range,
 661	.test_walk		= queue_pages_test_walk,
 662};
 663
 664/*
 665 * Walk through page tables and collect pages to be migrated.
 666 *
 667 * If pages found in a given range are on a set of nodes (determined by
 668 * @nodes and @flags,) it's isolated and queued to the pagelist which is
 669 * passed via @private.
 670 *
 671 * queue_pages_range() has three possible return values:
 672 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
 673 *     specified.
 674 * 0 - queue pages successfully or no misplaced page.
 675 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
 676 *         memory range specified by nodemask and maxnode points outside
 677 *         your accessible address space (-EFAULT)
 678 */
 679static int
 680queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 681		nodemask_t *nodes, unsigned long flags,
 682		struct list_head *pagelist)
 683{
 684	struct queue_pages qp = {
 685		.pagelist = pagelist,
 686		.flags = flags,
 687		.nmask = nodes,
 688		.prev = NULL,
 689	};
 
 
 
 
 
 
 
 690
 691	return walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
 692}
 693
 694/*
 695 * Apply policy to a single VMA
 696 * This must be called with the mmap_sem held for writing.
 697 */
 698static int vma_replace_policy(struct vm_area_struct *vma,
 699						struct mempolicy *pol)
 700{
 701	int err;
 702	struct mempolicy *old;
 703	struct mempolicy *new;
 704
 705	pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 706		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 707		 vma->vm_ops, vma->vm_file,
 708		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 709
 710	new = mpol_dup(pol);
 711	if (IS_ERR(new))
 712		return PTR_ERR(new);
 713
 714	if (vma->vm_ops && vma->vm_ops->set_policy) {
 715		err = vma->vm_ops->set_policy(vma, new);
 716		if (err)
 717			goto err_out;
 718	}
 719
 720	old = vma->vm_policy;
 721	vma->vm_policy = new; /* protected by mmap_sem */
 722	mpol_put(old);
 723
 724	return 0;
 725 err_out:
 726	mpol_put(new);
 727	return err;
 728}
 729
 730/* Step 2: apply policy to a range and do splits. */
 731static int mbind_range(struct mm_struct *mm, unsigned long start,
 732		       unsigned long end, struct mempolicy *new_pol)
 733{
 734	struct vm_area_struct *next;
 735	struct vm_area_struct *prev;
 736	struct vm_area_struct *vma;
 737	int err = 0;
 738	pgoff_t pgoff;
 739	unsigned long vmstart;
 740	unsigned long vmend;
 741
 742	vma = find_vma(mm, start);
 743	if (!vma || vma->vm_start > start)
 744		return -EFAULT;
 745
 746	prev = vma->vm_prev;
 747	if (start > vma->vm_start)
 748		prev = vma;
 749
 750	for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 751		next = vma->vm_next;
 752		vmstart = max(start, vma->vm_start);
 753		vmend   = min(end, vma->vm_end);
 754
 755		if (mpol_equal(vma_policy(vma), new_pol))
 756			continue;
 757
 758		pgoff = vma->vm_pgoff +
 759			((vmstart - vma->vm_start) >> PAGE_SHIFT);
 760		prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 761				 vma->anon_vma, vma->vm_file, pgoff,
 762				 new_pol, vma->vm_userfaultfd_ctx);
 763		if (prev) {
 764			vma = prev;
 765			next = vma->vm_next;
 766			if (mpol_equal(vma_policy(vma), new_pol))
 767				continue;
 768			/* vma_merge() joined vma && vma->next, case 8 */
 769			goto replace;
 770		}
 771		if (vma->vm_start != vmstart) {
 772			err = split_vma(vma->vm_mm, vma, vmstart, 1);
 773			if (err)
 774				goto out;
 775		}
 776		if (vma->vm_end != vmend) {
 777			err = split_vma(vma->vm_mm, vma, vmend, 0);
 778			if (err)
 779				goto out;
 780		}
 781 replace:
 782		err = vma_replace_policy(vma, new_pol);
 783		if (err)
 784			goto out;
 785	}
 786
 787 out:
 788	return err;
 789}
 790
 791/* Set the process memory policy */
 792static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 793			     nodemask_t *nodes)
 794{
 795	struct mempolicy *new, *old;
 796	NODEMASK_SCRATCH(scratch);
 797	int ret;
 798
 799	if (!scratch)
 800		return -ENOMEM;
 801
 802	new = mpol_new(mode, flags, nodes);
 803	if (IS_ERR(new)) {
 804		ret = PTR_ERR(new);
 805		goto out;
 806	}
 807
 808	task_lock(current);
 809	ret = mpol_set_nodemask(new, nodes, scratch);
 810	if (ret) {
 811		task_unlock(current);
 812		mpol_put(new);
 813		goto out;
 814	}
 815	old = current->mempolicy;
 816	current->mempolicy = new;
 817	if (new && new->mode == MPOL_INTERLEAVE)
 818		current->il_prev = MAX_NUMNODES-1;
 819	task_unlock(current);
 820	mpol_put(old);
 821	ret = 0;
 822out:
 823	NODEMASK_SCRATCH_FREE(scratch);
 824	return ret;
 825}
 826
 827/*
 828 * Return nodemask for policy for get_mempolicy() query
 829 *
 830 * Called with task's alloc_lock held
 831 */
 832static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 833{
 834	nodes_clear(*nodes);
 835	if (p == &default_policy)
 836		return;
 837
 838	switch (p->mode) {
 839	case MPOL_BIND:
 840		/* Fall through */
 841	case MPOL_INTERLEAVE:
 842		*nodes = p->v.nodes;
 843		break;
 844	case MPOL_PREFERRED:
 845		if (!(p->flags & MPOL_F_LOCAL))
 846			node_set(p->v.preferred_node, *nodes);
 847		/* else return empty node mask for local allocation */
 848		break;
 849	default:
 850		BUG();
 851	}
 852}
 853
 854static int lookup_node(struct mm_struct *mm, unsigned long addr)
 855{
 856	struct page *p;
 857	int err;
 858
 859	int locked = 1;
 860	err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
 861	if (err >= 0) {
 862		err = page_to_nid(p);
 863		put_page(p);
 864	}
 865	if (locked)
 866		up_read(&mm->mmap_sem);
 867	return err;
 868}
 869
 870/* Retrieve NUMA policy */
 871static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 872			     unsigned long addr, unsigned long flags)
 873{
 874	int err;
 875	struct mm_struct *mm = current->mm;
 876	struct vm_area_struct *vma = NULL;
 877	struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
 878
 879	if (flags &
 880		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 881		return -EINVAL;
 882
 883	if (flags & MPOL_F_MEMS_ALLOWED) {
 884		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 885			return -EINVAL;
 886		*policy = 0;	/* just so it's initialized */
 887		task_lock(current);
 888		*nmask  = cpuset_current_mems_allowed;
 889		task_unlock(current);
 890		return 0;
 891	}
 892
 893	if (flags & MPOL_F_ADDR) {
 894		/*
 895		 * Do NOT fall back to task policy if the
 896		 * vma/shared policy at addr is NULL.  We
 897		 * want to return MPOL_DEFAULT in this case.
 898		 */
 899		down_read(&mm->mmap_sem);
 900		vma = find_vma_intersection(mm, addr, addr+1);
 901		if (!vma) {
 902			up_read(&mm->mmap_sem);
 903			return -EFAULT;
 904		}
 905		if (vma->vm_ops && vma->vm_ops->get_policy)
 906			pol = vma->vm_ops->get_policy(vma, addr);
 907		else
 908			pol = vma->vm_policy;
 909	} else if (addr)
 910		return -EINVAL;
 911
 912	if (!pol)
 913		pol = &default_policy;	/* indicates default behavior */
 914
 915	if (flags & MPOL_F_NODE) {
 916		if (flags & MPOL_F_ADDR) {
 917			/*
 918			 * Take a refcount on the mpol, lookup_node()
 919			 * wil drop the mmap_sem, so after calling
 920			 * lookup_node() only "pol" remains valid, "vma"
 921			 * is stale.
 922			 */
 923			pol_refcount = pol;
 924			vma = NULL;
 925			mpol_get(pol);
 926			err = lookup_node(mm, addr);
 927			if (err < 0)
 928				goto out;
 929			*policy = err;
 930		} else if (pol == current->mempolicy &&
 931				pol->mode == MPOL_INTERLEAVE) {
 932			*policy = next_node_in(current->il_prev, pol->v.nodes);
 933		} else {
 934			err = -EINVAL;
 935			goto out;
 936		}
 937	} else {
 938		*policy = pol == &default_policy ? MPOL_DEFAULT :
 939						pol->mode;
 940		/*
 941		 * Internal mempolicy flags must be masked off before exposing
 942		 * the policy to userspace.
 943		 */
 944		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 945	}
 946
 947	err = 0;
 948	if (nmask) {
 949		if (mpol_store_user_nodemask(pol)) {
 950			*nmask = pol->w.user_nodemask;
 951		} else {
 952			task_lock(current);
 953			get_policy_nodemask(pol, nmask);
 954			task_unlock(current);
 955		}
 956	}
 957
 958 out:
 959	mpol_cond_put(pol);
 960	if (vma)
 961		up_read(&mm->mmap_sem);
 962	if (pol_refcount)
 963		mpol_put(pol_refcount);
 964	return err;
 965}
 966
 967#ifdef CONFIG_MIGRATION
 968/*
 969 * page migration, thp tail pages can be passed.
 970 */
 971static int migrate_page_add(struct page *page, struct list_head *pagelist,
 972				unsigned long flags)
 973{
 974	struct page *head = compound_head(page);
 975	/*
 976	 * Avoid migrating a page that is shared with others.
 977	 */
 978	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
 979		if (!isolate_lru_page(head)) {
 980			list_add_tail(&head->lru, pagelist);
 981			mod_node_page_state(page_pgdat(head),
 982				NR_ISOLATED_ANON + page_is_file_cache(head),
 983				hpage_nr_pages(head));
 984		} else if (flags & MPOL_MF_STRICT) {
 985			/*
 986			 * Non-movable page may reach here.  And, there may be
 987			 * temporary off LRU pages or non-LRU movable pages.
 988			 * Treat them as unmovable pages since they can't be
 989			 * isolated, so they can't be moved at the moment.  It
 990			 * should return -EIO for this case too.
 991			 */
 992			return -EIO;
 993		}
 994	}
 995
 996	return 0;
 997}
 998
 999/* page allocation callback for NUMA node migration */
1000struct page *alloc_new_node_page(struct page *page, unsigned long node)
1001{
1002	if (PageHuge(page))
1003		return alloc_huge_page_node(page_hstate(compound_head(page)),
1004					node);
1005	else if (PageTransHuge(page)) {
1006		struct page *thp;
1007
1008		thp = alloc_pages_node(node,
1009			(GFP_TRANSHUGE | __GFP_THISNODE),
1010			HPAGE_PMD_ORDER);
1011		if (!thp)
1012			return NULL;
1013		prep_transhuge_page(thp);
1014		return thp;
1015	} else
1016		return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
1017						    __GFP_THISNODE, 0);
1018}
1019
1020/*
1021 * Migrate pages from one node to a target node.
1022 * Returns error or the number of pages not migrated.
1023 */
1024static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1025			   int flags)
1026{
1027	nodemask_t nmask;
1028	LIST_HEAD(pagelist);
1029	int err = 0;
1030
1031	nodes_clear(nmask);
1032	node_set(source, nmask);
1033
1034	/*
1035	 * This does not "check" the range but isolates all pages that
1036	 * need migration.  Between passing in the full user address
1037	 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1038	 */
1039	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1040	queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1041			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1042
1043	if (!list_empty(&pagelist)) {
1044		err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
1045					MIGRATE_SYNC, MR_SYSCALL);
1046		if (err)
1047			putback_movable_pages(&pagelist);
1048	}
1049
1050	return err;
1051}
1052
1053/*
1054 * Move pages between the two nodesets so as to preserve the physical
1055 * layout as much as possible.
1056 *
1057 * Returns the number of page that could not be moved.
1058 */
1059int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1060		     const nodemask_t *to, int flags)
1061{
1062	int busy = 0;
1063	int err;
1064	nodemask_t tmp;
1065
1066	err = migrate_prep();
1067	if (err)
1068		return err;
1069
1070	down_read(&mm->mmap_sem);
1071
1072	/*
1073	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1074	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
1075	 * bit in 'tmp', and return that <source, dest> pair for migration.
1076	 * The pair of nodemasks 'to' and 'from' define the map.
1077	 *
1078	 * If no pair of bits is found that way, fallback to picking some
1079	 * pair of 'source' and 'dest' bits that are not the same.  If the
1080	 * 'source' and 'dest' bits are the same, this represents a node
1081	 * that will be migrating to itself, so no pages need move.
1082	 *
1083	 * If no bits are left in 'tmp', or if all remaining bits left
1084	 * in 'tmp' correspond to the same bit in 'to', return false
1085	 * (nothing left to migrate).
1086	 *
1087	 * This lets us pick a pair of nodes to migrate between, such that
1088	 * if possible the dest node is not already occupied by some other
1089	 * source node, minimizing the risk of overloading the memory on a
1090	 * node that would happen if we migrated incoming memory to a node
1091	 * before migrating outgoing memory source that same node.
1092	 *
1093	 * A single scan of tmp is sufficient.  As we go, we remember the
1094	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1095	 * that not only moved, but what's better, moved to an empty slot
1096	 * (d is not set in tmp), then we break out then, with that pair.
1097	 * Otherwise when we finish scanning from_tmp, we at least have the
1098	 * most recent <s, d> pair that moved.  If we get all the way through
1099	 * the scan of tmp without finding any node that moved, much less
1100	 * moved to an empty node, then there is nothing left worth migrating.
1101	 */
1102
1103	tmp = *from;
1104	while (!nodes_empty(tmp)) {
1105		int s,d;
1106		int source = NUMA_NO_NODE;
1107		int dest = 0;
1108
1109		for_each_node_mask(s, tmp) {
1110
1111			/*
1112			 * do_migrate_pages() tries to maintain the relative
1113			 * node relationship of the pages established between
1114			 * threads and memory areas.
1115                         *
1116			 * However if the number of source nodes is not equal to
1117			 * the number of destination nodes we can not preserve
1118			 * this node relative relationship.  In that case, skip
1119			 * copying memory from a node that is in the destination
1120			 * mask.
1121			 *
1122			 * Example: [2,3,4] -> [3,4,5] moves everything.
1123			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1124			 */
1125
1126			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1127						(node_isset(s, *to)))
1128				continue;
1129
1130			d = node_remap(s, *from, *to);
1131			if (s == d)
1132				continue;
1133
1134			source = s;	/* Node moved. Memorize */
1135			dest = d;
1136
1137			/* dest not in remaining from nodes? */
1138			if (!node_isset(dest, tmp))
1139				break;
1140		}
1141		if (source == NUMA_NO_NODE)
1142			break;
1143
1144		node_clear(source, tmp);
1145		err = migrate_to_node(mm, source, dest, flags);
1146		if (err > 0)
1147			busy += err;
1148		if (err < 0)
1149			break;
1150	}
1151	up_read(&mm->mmap_sem);
1152	if (err < 0)
1153		return err;
1154	return busy;
1155
1156}
1157
1158/*
1159 * Allocate a new page for page migration based on vma policy.
1160 * Start by assuming the page is mapped by the same vma as contains @start.
1161 * Search forward from there, if not.  N.B., this assumes that the
1162 * list of pages handed to migrate_pages()--which is how we get here--
1163 * is in virtual address order.
1164 */
1165static struct page *new_page(struct page *page, unsigned long start)
1166{
1167	struct vm_area_struct *vma;
1168	unsigned long uninitialized_var(address);
1169
1170	vma = find_vma(current->mm, start);
1171	while (vma) {
1172		address = page_address_in_vma(page, vma);
1173		if (address != -EFAULT)
1174			break;
1175		vma = vma->vm_next;
1176	}
1177
1178	if (PageHuge(page)) {
1179		return alloc_huge_page_vma(page_hstate(compound_head(page)),
1180				vma, address);
1181	} else if (PageTransHuge(page)) {
1182		struct page *thp;
1183
1184		thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1185					 HPAGE_PMD_ORDER);
1186		if (!thp)
1187			return NULL;
1188		prep_transhuge_page(thp);
1189		return thp;
1190	}
1191	/*
1192	 * if !vma, alloc_page_vma() will use task or system default policy
1193	 */
1194	return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1195			vma, address);
1196}
1197#else
1198
1199static int migrate_page_add(struct page *page, struct list_head *pagelist,
1200				unsigned long flags)
1201{
1202	return -EIO;
1203}
1204
1205int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1206		     const nodemask_t *to, int flags)
1207{
1208	return -ENOSYS;
1209}
1210
1211static struct page *new_page(struct page *page, unsigned long start)
1212{
1213	return NULL;
1214}
1215#endif
1216
1217static long do_mbind(unsigned long start, unsigned long len,
1218		     unsigned short mode, unsigned short mode_flags,
1219		     nodemask_t *nmask, unsigned long flags)
1220{
1221	struct mm_struct *mm = current->mm;
1222	struct mempolicy *new;
1223	unsigned long end;
1224	int err;
1225	int ret;
1226	LIST_HEAD(pagelist);
1227
1228	if (flags & ~(unsigned long)MPOL_MF_VALID)
1229		return -EINVAL;
1230	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1231		return -EPERM;
1232
1233	if (start & ~PAGE_MASK)
1234		return -EINVAL;
1235
1236	if (mode == MPOL_DEFAULT)
1237		flags &= ~MPOL_MF_STRICT;
1238
1239	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1240	end = start + len;
1241
1242	if (end < start)
1243		return -EINVAL;
1244	if (end == start)
1245		return 0;
1246
1247	new = mpol_new(mode, mode_flags, nmask);
1248	if (IS_ERR(new))
1249		return PTR_ERR(new);
1250
1251	if (flags & MPOL_MF_LAZY)
1252		new->flags |= MPOL_F_MOF;
1253
1254	/*
1255	 * If we are using the default policy then operation
1256	 * on discontinuous address spaces is okay after all
1257	 */
1258	if (!new)
1259		flags |= MPOL_MF_DISCONTIG_OK;
1260
1261	pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1262		 start, start + len, mode, mode_flags,
1263		 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1264
1265	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1266
1267		err = migrate_prep();
1268		if (err)
1269			goto mpol_out;
1270	}
1271	{
1272		NODEMASK_SCRATCH(scratch);
1273		if (scratch) {
1274			down_write(&mm->mmap_sem);
1275			task_lock(current);
1276			err = mpol_set_nodemask(new, nmask, scratch);
1277			task_unlock(current);
1278			if (err)
1279				up_write(&mm->mmap_sem);
1280		} else
1281			err = -ENOMEM;
1282		NODEMASK_SCRATCH_FREE(scratch);
1283	}
1284	if (err)
1285		goto mpol_out;
1286
1287	ret = queue_pages_range(mm, start, end, nmask,
1288			  flags | MPOL_MF_INVERT, &pagelist);
1289
1290	if (ret < 0) {
1291		err = ret;
1292		goto up_out;
1293	}
1294
1295	err = mbind_range(mm, start, end, new);
1296
1297	if (!err) {
1298		int nr_failed = 0;
1299
1300		if (!list_empty(&pagelist)) {
1301			WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1302			nr_failed = migrate_pages(&pagelist, new_page, NULL,
1303				start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1304			if (nr_failed)
1305				putback_movable_pages(&pagelist);
1306		}
1307
1308		if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1309			err = -EIO;
1310	} else {
1311up_out:
1312		if (!list_empty(&pagelist))
1313			putback_movable_pages(&pagelist);
1314	}
1315
1316	up_write(&mm->mmap_sem);
1317mpol_out:
1318	mpol_put(new);
1319	return err;
1320}
1321
1322/*
1323 * User space interface with variable sized bitmaps for nodelists.
1324 */
1325
1326/* Copy a node mask from user space. */
1327static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1328		     unsigned long maxnode)
1329{
1330	unsigned long k;
1331	unsigned long t;
1332	unsigned long nlongs;
1333	unsigned long endmask;
1334
1335	--maxnode;
1336	nodes_clear(*nodes);
1337	if (maxnode == 0 || !nmask)
1338		return 0;
1339	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1340		return -EINVAL;
1341
1342	nlongs = BITS_TO_LONGS(maxnode);
1343	if ((maxnode % BITS_PER_LONG) == 0)
1344		endmask = ~0UL;
1345	else
1346		endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1347
1348	/*
1349	 * When the user specified more nodes than supported just check
1350	 * if the non supported part is all zero.
1351	 *
1352	 * If maxnode have more longs than MAX_NUMNODES, check
1353	 * the bits in that area first. And then go through to
1354	 * check the rest bits which equal or bigger than MAX_NUMNODES.
1355	 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1356	 */
1357	if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1358		for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1359			if (get_user(t, nmask + k))
1360				return -EFAULT;
1361			if (k == nlongs - 1) {
1362				if (t & endmask)
1363					return -EINVAL;
1364			} else if (t)
1365				return -EINVAL;
1366		}
1367		nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1368		endmask = ~0UL;
1369	}
1370
1371	if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1372		unsigned long valid_mask = endmask;
1373
1374		valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1375		if (get_user(t, nmask + nlongs - 1))
1376			return -EFAULT;
1377		if (t & valid_mask)
1378			return -EINVAL;
1379	}
1380
1381	if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1382		return -EFAULT;
1383	nodes_addr(*nodes)[nlongs-1] &= endmask;
1384	return 0;
1385}
1386
1387/* Copy a kernel node mask to user space */
1388static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1389			      nodemask_t *nodes)
1390{
1391	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1392	unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1393
1394	if (copy > nbytes) {
1395		if (copy > PAGE_SIZE)
1396			return -EINVAL;
1397		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1398			return -EFAULT;
1399		copy = nbytes;
1400	}
1401	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1402}
1403
1404static long kernel_mbind(unsigned long start, unsigned long len,
1405			 unsigned long mode, const unsigned long __user *nmask,
1406			 unsigned long maxnode, unsigned int flags)
1407{
1408	nodemask_t nodes;
1409	int err;
1410	unsigned short mode_flags;
1411
1412	start = untagged_addr(start);
1413	mode_flags = mode & MPOL_MODE_FLAGS;
1414	mode &= ~MPOL_MODE_FLAGS;
1415	if (mode >= MPOL_MAX)
1416		return -EINVAL;
1417	if ((mode_flags & MPOL_F_STATIC_NODES) &&
1418	    (mode_flags & MPOL_F_RELATIVE_NODES))
1419		return -EINVAL;
1420	err = get_nodes(&nodes, nmask, maxnode);
1421	if (err)
1422		return err;
1423	return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1424}
1425
1426SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1427		unsigned long, mode, const unsigned long __user *, nmask,
1428		unsigned long, maxnode, unsigned int, flags)
1429{
1430	return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1431}
1432
1433/* Set the process memory policy */
1434static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1435				 unsigned long maxnode)
1436{
1437	int err;
1438	nodemask_t nodes;
1439	unsigned short flags;
1440
1441	flags = mode & MPOL_MODE_FLAGS;
1442	mode &= ~MPOL_MODE_FLAGS;
1443	if ((unsigned int)mode >= MPOL_MAX)
1444		return -EINVAL;
1445	if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1446		return -EINVAL;
1447	err = get_nodes(&nodes, nmask, maxnode);
1448	if (err)
1449		return err;
1450	return do_set_mempolicy(mode, flags, &nodes);
1451}
1452
1453SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1454		unsigned long, maxnode)
1455{
1456	return kernel_set_mempolicy(mode, nmask, maxnode);
1457}
1458
1459static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1460				const unsigned long __user *old_nodes,
1461				const unsigned long __user *new_nodes)
1462{
1463	struct mm_struct *mm = NULL;
1464	struct task_struct *task;
1465	nodemask_t task_nodes;
1466	int err;
1467	nodemask_t *old;
1468	nodemask_t *new;
1469	NODEMASK_SCRATCH(scratch);
1470
1471	if (!scratch)
1472		return -ENOMEM;
1473
1474	old = &scratch->mask1;
1475	new = &scratch->mask2;
1476
1477	err = get_nodes(old, old_nodes, maxnode);
1478	if (err)
1479		goto out;
1480
1481	err = get_nodes(new, new_nodes, maxnode);
1482	if (err)
1483		goto out;
1484
1485	/* Find the mm_struct */
1486	rcu_read_lock();
1487	task = pid ? find_task_by_vpid(pid) : current;
1488	if (!task) {
1489		rcu_read_unlock();
1490		err = -ESRCH;
1491		goto out;
1492	}
1493	get_task_struct(task);
1494
1495	err = -EINVAL;
1496
1497	/*
1498	 * Check if this process has the right to modify the specified process.
1499	 * Use the regular "ptrace_may_access()" checks.
1500	 */
1501	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1502		rcu_read_unlock();
1503		err = -EPERM;
1504		goto out_put;
1505	}
1506	rcu_read_unlock();
1507
1508	task_nodes = cpuset_mems_allowed(task);
1509	/* Is the user allowed to access the target nodes? */
1510	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1511		err = -EPERM;
1512		goto out_put;
1513	}
1514
1515	task_nodes = cpuset_mems_allowed(current);
1516	nodes_and(*new, *new, task_nodes);
1517	if (nodes_empty(*new))
1518		goto out_put;
1519
 
 
 
 
1520	err = security_task_movememory(task);
1521	if (err)
1522		goto out_put;
1523
1524	mm = get_task_mm(task);
1525	put_task_struct(task);
1526
1527	if (!mm) {
1528		err = -EINVAL;
1529		goto out;
1530	}
1531
1532	err = do_migrate_pages(mm, old, new,
1533		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1534
1535	mmput(mm);
1536out:
1537	NODEMASK_SCRATCH_FREE(scratch);
1538
1539	return err;
1540
1541out_put:
1542	put_task_struct(task);
1543	goto out;
1544
1545}
1546
1547SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1548		const unsigned long __user *, old_nodes,
1549		const unsigned long __user *, new_nodes)
1550{
1551	return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1552}
1553
1554
1555/* Retrieve NUMA policy */
1556static int kernel_get_mempolicy(int __user *policy,
1557				unsigned long __user *nmask,
1558				unsigned long maxnode,
1559				unsigned long addr,
1560				unsigned long flags)
1561{
1562	int err;
1563	int uninitialized_var(pval);
1564	nodemask_t nodes;
1565
1566	addr = untagged_addr(addr);
1567
1568	if (nmask != NULL && maxnode < nr_node_ids)
1569		return -EINVAL;
1570
1571	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1572
1573	if (err)
1574		return err;
1575
1576	if (policy && put_user(pval, policy))
1577		return -EFAULT;
1578
1579	if (nmask)
1580		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1581
1582	return err;
1583}
1584
1585SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1586		unsigned long __user *, nmask, unsigned long, maxnode,
1587		unsigned long, addr, unsigned long, flags)
1588{
1589	return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1590}
1591
1592#ifdef CONFIG_COMPAT
1593
1594COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1595		       compat_ulong_t __user *, nmask,
1596		       compat_ulong_t, maxnode,
1597		       compat_ulong_t, addr, compat_ulong_t, flags)
1598{
1599	long err;
1600	unsigned long __user *nm = NULL;
1601	unsigned long nr_bits, alloc_size;
1602	DECLARE_BITMAP(bm, MAX_NUMNODES);
1603
1604	nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1605	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1606
1607	if (nmask)
1608		nm = compat_alloc_user_space(alloc_size);
1609
1610	err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1611
1612	if (!err && nmask) {
1613		unsigned long copy_size;
1614		copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1615		err = copy_from_user(bm, nm, copy_size);
1616		/* ensure entire bitmap is zeroed */
1617		err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1618		err |= compat_put_bitmap(nmask, bm, nr_bits);
1619	}
1620
1621	return err;
1622}
1623
1624COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1625		       compat_ulong_t, maxnode)
1626{
1627	unsigned long __user *nm = NULL;
1628	unsigned long nr_bits, alloc_size;
1629	DECLARE_BITMAP(bm, MAX_NUMNODES);
1630
1631	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1632	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1633
1634	if (nmask) {
1635		if (compat_get_bitmap(bm, nmask, nr_bits))
1636			return -EFAULT;
1637		nm = compat_alloc_user_space(alloc_size);
1638		if (copy_to_user(nm, bm, alloc_size))
1639			return -EFAULT;
1640	}
1641
1642	return kernel_set_mempolicy(mode, nm, nr_bits+1);
1643}
1644
1645COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1646		       compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1647		       compat_ulong_t, maxnode, compat_ulong_t, flags)
1648{
1649	unsigned long __user *nm = NULL;
1650	unsigned long nr_bits, alloc_size;
1651	nodemask_t bm;
1652
1653	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1654	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1655
1656	if (nmask) {
1657		if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1658			return -EFAULT;
1659		nm = compat_alloc_user_space(alloc_size);
1660		if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1661			return -EFAULT;
1662	}
1663
1664	return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1665}
1666
1667COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1668		       compat_ulong_t, maxnode,
1669		       const compat_ulong_t __user *, old_nodes,
1670		       const compat_ulong_t __user *, new_nodes)
1671{
1672	unsigned long __user *old = NULL;
1673	unsigned long __user *new = NULL;
1674	nodemask_t tmp_mask;
1675	unsigned long nr_bits;
1676	unsigned long size;
1677
1678	nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1679	size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1680	if (old_nodes) {
1681		if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1682			return -EFAULT;
1683		old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1684		if (new_nodes)
1685			new = old + size / sizeof(unsigned long);
1686		if (copy_to_user(old, nodes_addr(tmp_mask), size))
1687			return -EFAULT;
1688	}
1689	if (new_nodes) {
1690		if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1691			return -EFAULT;
1692		if (new == NULL)
1693			new = compat_alloc_user_space(size);
1694		if (copy_to_user(new, nodes_addr(tmp_mask), size))
1695			return -EFAULT;
1696	}
1697	return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1698}
1699
1700#endif /* CONFIG_COMPAT */
1701
1702struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1703						unsigned long addr)
1704{
1705	struct mempolicy *pol = NULL;
1706
1707	if (vma) {
1708		if (vma->vm_ops && vma->vm_ops->get_policy) {
1709			pol = vma->vm_ops->get_policy(vma, addr);
1710		} else if (vma->vm_policy) {
1711			pol = vma->vm_policy;
1712
1713			/*
1714			 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1715			 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1716			 * count on these policies which will be dropped by
1717			 * mpol_cond_put() later
1718			 */
1719			if (mpol_needs_cond_ref(pol))
1720				mpol_get(pol);
1721		}
1722	}
1723
1724	return pol;
1725}
1726
1727/*
1728 * get_vma_policy(@vma, @addr)
1729 * @vma: virtual memory area whose policy is sought
1730 * @addr: address in @vma for shared policy lookup
1731 *
1732 * Returns effective policy for a VMA at specified address.
1733 * Falls back to current->mempolicy or system default policy, as necessary.
1734 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1735 * count--added by the get_policy() vm_op, as appropriate--to protect against
1736 * freeing by another task.  It is the caller's responsibility to free the
1737 * extra reference for shared policies.
1738 */
1739static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1740						unsigned long addr)
1741{
1742	struct mempolicy *pol = __get_vma_policy(vma, addr);
1743
1744	if (!pol)
1745		pol = get_task_policy(current);
1746
1747	return pol;
1748}
1749
1750bool vma_policy_mof(struct vm_area_struct *vma)
1751{
1752	struct mempolicy *pol;
1753
1754	if (vma->vm_ops && vma->vm_ops->get_policy) {
1755		bool ret = false;
1756
1757		pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1758		if (pol && (pol->flags & MPOL_F_MOF))
1759			ret = true;
1760		mpol_cond_put(pol);
1761
1762		return ret;
1763	}
1764
1765	pol = vma->vm_policy;
1766	if (!pol)
1767		pol = get_task_policy(current);
1768
1769	return pol->flags & MPOL_F_MOF;
1770}
1771
1772static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1773{
1774	enum zone_type dynamic_policy_zone = policy_zone;
1775
1776	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1777
1778	/*
1779	 * if policy->v.nodes has movable memory only,
1780	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1781	 *
1782	 * policy->v.nodes is intersect with node_states[N_MEMORY].
1783	 * so if the following test faile, it implies
1784	 * policy->v.nodes has movable memory only.
1785	 */
1786	if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1787		dynamic_policy_zone = ZONE_MOVABLE;
1788
1789	return zone >= dynamic_policy_zone;
1790}
1791
1792/*
1793 * Return a nodemask representing a mempolicy for filtering nodes for
1794 * page allocation
1795 */
1796static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1797{
1798	/* Lower zones don't get a nodemask applied for MPOL_BIND */
1799	if (unlikely(policy->mode == MPOL_BIND) &&
1800			apply_policy_zone(policy, gfp_zone(gfp)) &&
1801			cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1802		return &policy->v.nodes;
1803
1804	return NULL;
1805}
1806
1807/* Return the node id preferred by the given mempolicy, or the given id */
1808static int policy_node(gfp_t gfp, struct mempolicy *policy,
1809								int nd)
1810{
1811	if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1812		nd = policy->v.preferred_node;
1813	else {
1814		/*
1815		 * __GFP_THISNODE shouldn't even be used with the bind policy
1816		 * because we might easily break the expectation to stay on the
1817		 * requested node and not break the policy.
1818		 */
1819		WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1820	}
1821
1822	return nd;
1823}
1824
1825/* Do dynamic interleaving for a process */
1826static unsigned interleave_nodes(struct mempolicy *policy)
1827{
1828	unsigned next;
1829	struct task_struct *me = current;
1830
1831	next = next_node_in(me->il_prev, policy->v.nodes);
1832	if (next < MAX_NUMNODES)
1833		me->il_prev = next;
1834	return next;
1835}
1836
1837/*
1838 * Depending on the memory policy provide a node from which to allocate the
1839 * next slab entry.
1840 */
1841unsigned int mempolicy_slab_node(void)
1842{
1843	struct mempolicy *policy;
1844	int node = numa_mem_id();
1845
1846	if (in_interrupt())
1847		return node;
1848
1849	policy = current->mempolicy;
1850	if (!policy || policy->flags & MPOL_F_LOCAL)
1851		return node;
1852
1853	switch (policy->mode) {
1854	case MPOL_PREFERRED:
1855		/*
1856		 * handled MPOL_F_LOCAL above
1857		 */
1858		return policy->v.preferred_node;
1859
1860	case MPOL_INTERLEAVE:
1861		return interleave_nodes(policy);
1862
1863	case MPOL_BIND: {
1864		struct zoneref *z;
1865
1866		/*
1867		 * Follow bind policy behavior and start allocation at the
1868		 * first node.
1869		 */
1870		struct zonelist *zonelist;
1871		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1872		zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1873		z = first_zones_zonelist(zonelist, highest_zoneidx,
1874							&policy->v.nodes);
1875		return z->zone ? zone_to_nid(z->zone) : node;
1876	}
1877
1878	default:
1879		BUG();
1880	}
1881}
1882
1883/*
1884 * Do static interleaving for a VMA with known offset @n.  Returns the n'th
1885 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1886 * number of present nodes.
1887 */
1888static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1889{
1890	unsigned nnodes = nodes_weight(pol->v.nodes);
1891	unsigned target;
1892	int i;
1893	int nid;
1894
1895	if (!nnodes)
1896		return numa_node_id();
1897	target = (unsigned int)n % nnodes;
1898	nid = first_node(pol->v.nodes);
1899	for (i = 0; i < target; i++)
1900		nid = next_node(nid, pol->v.nodes);
1901	return nid;
1902}
1903
1904/* Determine a node number for interleave */
1905static inline unsigned interleave_nid(struct mempolicy *pol,
1906		 struct vm_area_struct *vma, unsigned long addr, int shift)
1907{
1908	if (vma) {
1909		unsigned long off;
1910
1911		/*
1912		 * for small pages, there is no difference between
1913		 * shift and PAGE_SHIFT, so the bit-shift is safe.
1914		 * for huge pages, since vm_pgoff is in units of small
1915		 * pages, we need to shift off the always 0 bits to get
1916		 * a useful offset.
1917		 */
1918		BUG_ON(shift < PAGE_SHIFT);
1919		off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1920		off += (addr - vma->vm_start) >> shift;
1921		return offset_il_node(pol, off);
1922	} else
1923		return interleave_nodes(pol);
1924}
1925
1926#ifdef CONFIG_HUGETLBFS
1927/*
1928 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1929 * @vma: virtual memory area whose policy is sought
1930 * @addr: address in @vma for shared policy lookup and interleave policy
1931 * @gfp_flags: for requested zone
1932 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1933 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1934 *
1935 * Returns a nid suitable for a huge page allocation and a pointer
1936 * to the struct mempolicy for conditional unref after allocation.
1937 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1938 * @nodemask for filtering the zonelist.
1939 *
1940 * Must be protected by read_mems_allowed_begin()
1941 */
1942int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1943				struct mempolicy **mpol, nodemask_t **nodemask)
1944{
1945	int nid;
1946
1947	*mpol = get_vma_policy(vma, addr);
1948	*nodemask = NULL;	/* assume !MPOL_BIND */
1949
1950	if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1951		nid = interleave_nid(*mpol, vma, addr,
1952					huge_page_shift(hstate_vma(vma)));
1953	} else {
1954		nid = policy_node(gfp_flags, *mpol, numa_node_id());
1955		if ((*mpol)->mode == MPOL_BIND)
1956			*nodemask = &(*mpol)->v.nodes;
1957	}
1958	return nid;
1959}
1960
1961/*
1962 * init_nodemask_of_mempolicy
1963 *
1964 * If the current task's mempolicy is "default" [NULL], return 'false'
1965 * to indicate default policy.  Otherwise, extract the policy nodemask
1966 * for 'bind' or 'interleave' policy into the argument nodemask, or
1967 * initialize the argument nodemask to contain the single node for
1968 * 'preferred' or 'local' policy and return 'true' to indicate presence
1969 * of non-default mempolicy.
1970 *
1971 * We don't bother with reference counting the mempolicy [mpol_get/put]
1972 * because the current task is examining it's own mempolicy and a task's
1973 * mempolicy is only ever changed by the task itself.
1974 *
1975 * N.B., it is the caller's responsibility to free a returned nodemask.
1976 */
1977bool init_nodemask_of_mempolicy(nodemask_t *mask)
1978{
1979	struct mempolicy *mempolicy;
1980	int nid;
1981
1982	if (!(mask && current->mempolicy))
1983		return false;
1984
1985	task_lock(current);
1986	mempolicy = current->mempolicy;
1987	switch (mempolicy->mode) {
1988	case MPOL_PREFERRED:
1989		if (mempolicy->flags & MPOL_F_LOCAL)
1990			nid = numa_node_id();
1991		else
1992			nid = mempolicy->v.preferred_node;
1993		init_nodemask_of_node(mask, nid);
1994		break;
1995
1996	case MPOL_BIND:
1997		/* Fall through */
1998	case MPOL_INTERLEAVE:
1999		*mask =  mempolicy->v.nodes;
2000		break;
2001
2002	default:
2003		BUG();
2004	}
2005	task_unlock(current);
2006
2007	return true;
2008}
2009#endif
2010
2011/*
2012 * mempolicy_nodemask_intersects
2013 *
2014 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
2015 * policy.  Otherwise, check for intersection between mask and the policy
2016 * nodemask for 'bind' or 'interleave' policy.  For 'perferred' or 'local'
2017 * policy, always return true since it may allocate elsewhere on fallback.
2018 *
2019 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2020 */
2021bool mempolicy_nodemask_intersects(struct task_struct *tsk,
2022					const nodemask_t *mask)
2023{
2024	struct mempolicy *mempolicy;
2025	bool ret = true;
2026
2027	if (!mask)
2028		return ret;
2029	task_lock(tsk);
2030	mempolicy = tsk->mempolicy;
2031	if (!mempolicy)
2032		goto out;
2033
2034	switch (mempolicy->mode) {
2035	case MPOL_PREFERRED:
2036		/*
2037		 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
2038		 * allocate from, they may fallback to other nodes when oom.
2039		 * Thus, it's possible for tsk to have allocated memory from
2040		 * nodes in mask.
2041		 */
2042		break;
2043	case MPOL_BIND:
2044	case MPOL_INTERLEAVE:
2045		ret = nodes_intersects(mempolicy->v.nodes, *mask);
2046		break;
2047	default:
2048		BUG();
2049	}
2050out:
2051	task_unlock(tsk);
2052	return ret;
2053}
2054
2055/* Allocate a page in interleaved policy.
2056   Own path because it needs to do special accounting. */
2057static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2058					unsigned nid)
2059{
2060	struct page *page;
2061
2062	page = __alloc_pages(gfp, order, nid);
2063	/* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2064	if (!static_branch_likely(&vm_numa_stat_key))
2065		return page;
2066	if (page && page_to_nid(page) == nid) {
2067		preempt_disable();
2068		__inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
2069		preempt_enable();
2070	}
2071	return page;
2072}
2073
2074/**
2075 * 	alloc_pages_vma	- Allocate a page for a VMA.
2076 *
2077 * 	@gfp:
2078 *      %GFP_USER    user allocation.
2079 *      %GFP_KERNEL  kernel allocations,
2080 *      %GFP_HIGHMEM highmem/user allocations,
2081 *      %GFP_FS      allocation should not call back into a file system.
2082 *      %GFP_ATOMIC  don't sleep.
2083 *
2084 *	@order:Order of the GFP allocation.
2085 * 	@vma:  Pointer to VMA or NULL if not available.
2086 *	@addr: Virtual Address of the allocation. Must be inside the VMA.
2087 *	@node: Which node to prefer for allocation (modulo policy).
2088 *	@hugepage: for hugepages try only the preferred node if possible
2089 *
2090 * 	This function allocates a page from the kernel page pool and applies
2091 *	a NUMA policy associated with the VMA or the current process.
2092 *	When VMA is not NULL caller must hold down_read on the mmap_sem of the
2093 *	mm_struct of the VMA to prevent it from going away. Should be used for
2094 *	all allocations for pages that will be mapped into user space. Returns
2095 *	NULL when no page can be allocated.
2096 */
2097struct page *
2098alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2099		unsigned long addr, int node, bool hugepage)
2100{
2101	struct mempolicy *pol;
2102	struct page *page;
2103	int preferred_nid;
2104	nodemask_t *nmask;
2105
2106	pol = get_vma_policy(vma, addr);
2107
2108	if (pol->mode == MPOL_INTERLEAVE) {
2109		unsigned nid;
2110
2111		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2112		mpol_cond_put(pol);
2113		page = alloc_page_interleave(gfp, order, nid);
2114		goto out;
2115	}
2116
2117	if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2118		int hpage_node = node;
2119
2120		/*
2121		 * For hugepage allocation and non-interleave policy which
2122		 * allows the current node (or other explicitly preferred
2123		 * node) we only try to allocate from the current/preferred
2124		 * node and don't fall back to other nodes, as the cost of
2125		 * remote accesses would likely offset THP benefits.
2126		 *
2127		 * If the policy is interleave, or does not allow the current
2128		 * node in its nodemask, we allocate the standard way.
2129		 */
2130		if (pol->mode == MPOL_PREFERRED && !(pol->flags & MPOL_F_LOCAL))
 
2131			hpage_node = pol->v.preferred_node;
2132
2133		nmask = policy_nodemask(gfp, pol);
2134		if (!nmask || node_isset(hpage_node, *nmask)) {
2135			mpol_cond_put(pol);
2136			page = __alloc_pages_node(hpage_node,
2137						gfp | __GFP_THISNODE, order);
2138
2139			/*
2140			 * If hugepage allocations are configured to always
2141			 * synchronous compact or the vma has been madvised
2142			 * to prefer hugepage backing, retry allowing remote
2143			 * memory as well.
2144			 */
2145			if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2146				page = __alloc_pages_node(hpage_node,
2147						gfp | __GFP_NORETRY, order);
2148
2149			goto out;
2150		}
2151	}
2152
2153	nmask = policy_nodemask(gfp, pol);
2154	preferred_nid = policy_node(gfp, pol, node);
2155	page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2156	mpol_cond_put(pol);
2157out:
2158	return page;
2159}
2160EXPORT_SYMBOL(alloc_pages_vma);
2161
2162/**
2163 * 	alloc_pages_current - Allocate pages.
2164 *
2165 *	@gfp:
2166 *		%GFP_USER   user allocation,
2167 *      	%GFP_KERNEL kernel allocation,
2168 *      	%GFP_HIGHMEM highmem allocation,
2169 *      	%GFP_FS     don't call back into a file system.
2170 *      	%GFP_ATOMIC don't sleep.
2171 *	@order: Power of two of allocation size in pages. 0 is a single page.
2172 *
2173 *	Allocate a page from the kernel page pool.  When not in
2174 *	interrupt context and apply the current process NUMA policy.
2175 *	Returns NULL when no page can be allocated.
2176 */
2177struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2178{
2179	struct mempolicy *pol = &default_policy;
2180	struct page *page;
2181
2182	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2183		pol = get_task_policy(current);
2184
2185	/*
2186	 * No reference counting needed for current->mempolicy
2187	 * nor system default_policy
2188	 */
2189	if (pol->mode == MPOL_INTERLEAVE)
2190		page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2191	else
2192		page = __alloc_pages_nodemask(gfp, order,
2193				policy_node(gfp, pol, numa_node_id()),
2194				policy_nodemask(gfp, pol));
2195
2196	return page;
2197}
2198EXPORT_SYMBOL(alloc_pages_current);
2199
2200int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2201{
2202	struct mempolicy *pol = mpol_dup(vma_policy(src));
2203
2204	if (IS_ERR(pol))
2205		return PTR_ERR(pol);
2206	dst->vm_policy = pol;
2207	return 0;
2208}
2209
2210/*
2211 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2212 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2213 * with the mems_allowed returned by cpuset_mems_allowed().  This
2214 * keeps mempolicies cpuset relative after its cpuset moves.  See
2215 * further kernel/cpuset.c update_nodemask().
2216 *
2217 * current's mempolicy may be rebinded by the other task(the task that changes
2218 * cpuset's mems), so we needn't do rebind work for current task.
2219 */
2220
2221/* Slow path of a mempolicy duplicate */
2222struct mempolicy *__mpol_dup(struct mempolicy *old)
2223{
2224	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2225
2226	if (!new)
2227		return ERR_PTR(-ENOMEM);
2228
2229	/* task's mempolicy is protected by alloc_lock */
2230	if (old == current->mempolicy) {
2231		task_lock(current);
2232		*new = *old;
2233		task_unlock(current);
2234	} else
2235		*new = *old;
2236
2237	if (current_cpuset_is_being_rebound()) {
2238		nodemask_t mems = cpuset_mems_allowed(current);
2239		mpol_rebind_policy(new, &mems);
2240	}
2241	atomic_set(&new->refcnt, 1);
2242	return new;
2243}
2244
2245/* Slow path of a mempolicy comparison */
2246bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2247{
2248	if (!a || !b)
2249		return false;
2250	if (a->mode != b->mode)
2251		return false;
2252	if (a->flags != b->flags)
2253		return false;
2254	if (mpol_store_user_nodemask(a))
2255		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2256			return false;
2257
2258	switch (a->mode) {
2259	case MPOL_BIND:
2260		/* Fall through */
2261	case MPOL_INTERLEAVE:
2262		return !!nodes_equal(a->v.nodes, b->v.nodes);
2263	case MPOL_PREFERRED:
2264		/* a's ->flags is the same as b's */
2265		if (a->flags & MPOL_F_LOCAL)
2266			return true;
2267		return a->v.preferred_node == b->v.preferred_node;
2268	default:
2269		BUG();
2270		return false;
2271	}
2272}
2273
2274/*
2275 * Shared memory backing store policy support.
2276 *
2277 * Remember policies even when nobody has shared memory mapped.
2278 * The policies are kept in Red-Black tree linked from the inode.
2279 * They are protected by the sp->lock rwlock, which should be held
2280 * for any accesses to the tree.
2281 */
2282
2283/*
2284 * lookup first element intersecting start-end.  Caller holds sp->lock for
2285 * reading or for writing
2286 */
2287static struct sp_node *
2288sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2289{
2290	struct rb_node *n = sp->root.rb_node;
2291
2292	while (n) {
2293		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2294
2295		if (start >= p->end)
2296			n = n->rb_right;
2297		else if (end <= p->start)
2298			n = n->rb_left;
2299		else
2300			break;
2301	}
2302	if (!n)
2303		return NULL;
2304	for (;;) {
2305		struct sp_node *w = NULL;
2306		struct rb_node *prev = rb_prev(n);
2307		if (!prev)
2308			break;
2309		w = rb_entry(prev, struct sp_node, nd);
2310		if (w->end <= start)
2311			break;
2312		n = prev;
2313	}
2314	return rb_entry(n, struct sp_node, nd);
2315}
2316
2317/*
2318 * Insert a new shared policy into the list.  Caller holds sp->lock for
2319 * writing.
2320 */
2321static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2322{
2323	struct rb_node **p = &sp->root.rb_node;
2324	struct rb_node *parent = NULL;
2325	struct sp_node *nd;
2326
2327	while (*p) {
2328		parent = *p;
2329		nd = rb_entry(parent, struct sp_node, nd);
2330		if (new->start < nd->start)
2331			p = &(*p)->rb_left;
2332		else if (new->end > nd->end)
2333			p = &(*p)->rb_right;
2334		else
2335			BUG();
2336	}
2337	rb_link_node(&new->nd, parent, p);
2338	rb_insert_color(&new->nd, &sp->root);
2339	pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2340		 new->policy ? new->policy->mode : 0);
2341}
2342
2343/* Find shared policy intersecting idx */
2344struct mempolicy *
2345mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2346{
2347	struct mempolicy *pol = NULL;
2348	struct sp_node *sn;
2349
2350	if (!sp->root.rb_node)
2351		return NULL;
2352	read_lock(&sp->lock);
2353	sn = sp_lookup(sp, idx, idx+1);
2354	if (sn) {
2355		mpol_get(sn->policy);
2356		pol = sn->policy;
2357	}
2358	read_unlock(&sp->lock);
2359	return pol;
2360}
2361
2362static void sp_free(struct sp_node *n)
2363{
2364	mpol_put(n->policy);
2365	kmem_cache_free(sn_cache, n);
2366}
2367
2368/**
2369 * mpol_misplaced - check whether current page node is valid in policy
2370 *
2371 * @page: page to be checked
2372 * @vma: vm area where page mapped
2373 * @addr: virtual address where page mapped
2374 *
2375 * Lookup current policy node id for vma,addr and "compare to" page's
2376 * node id.
2377 *
2378 * Returns:
2379 *	-1	- not misplaced, page is in the right node
2380 *	node	- node id where the page should be
2381 *
2382 * Policy determination "mimics" alloc_page_vma().
2383 * Called from fault path where we know the vma and faulting address.
2384 */
2385int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2386{
2387	struct mempolicy *pol;
2388	struct zoneref *z;
2389	int curnid = page_to_nid(page);
2390	unsigned long pgoff;
2391	int thiscpu = raw_smp_processor_id();
2392	int thisnid = cpu_to_node(thiscpu);
2393	int polnid = NUMA_NO_NODE;
2394	int ret = -1;
2395
2396	pol = get_vma_policy(vma, addr);
2397	if (!(pol->flags & MPOL_F_MOF))
2398		goto out;
2399
2400	switch (pol->mode) {
2401	case MPOL_INTERLEAVE:
2402		pgoff = vma->vm_pgoff;
2403		pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2404		polnid = offset_il_node(pol, pgoff);
2405		break;
2406
2407	case MPOL_PREFERRED:
2408		if (pol->flags & MPOL_F_LOCAL)
2409			polnid = numa_node_id();
2410		else
2411			polnid = pol->v.preferred_node;
2412		break;
2413
2414	case MPOL_BIND:
2415
2416		/*
2417		 * allows binding to multiple nodes.
2418		 * use current page if in policy nodemask,
2419		 * else select nearest allowed node, if any.
2420		 * If no allowed nodes, use current [!misplaced].
2421		 */
2422		if (node_isset(curnid, pol->v.nodes))
2423			goto out;
2424		z = first_zones_zonelist(
2425				node_zonelist(numa_node_id(), GFP_HIGHUSER),
2426				gfp_zone(GFP_HIGHUSER),
2427				&pol->v.nodes);
2428		polnid = zone_to_nid(z->zone);
2429		break;
2430
2431	default:
2432		BUG();
2433	}
2434
2435	/* Migrate the page towards the node whose CPU is referencing it */
2436	if (pol->flags & MPOL_F_MORON) {
2437		polnid = thisnid;
2438
2439		if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2440			goto out;
2441	}
2442
2443	if (curnid != polnid)
2444		ret = polnid;
2445out:
2446	mpol_cond_put(pol);
2447
2448	return ret;
2449}
2450
2451/*
2452 * Drop the (possibly final) reference to task->mempolicy.  It needs to be
2453 * dropped after task->mempolicy is set to NULL so that any allocation done as
2454 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2455 * policy.
2456 */
2457void mpol_put_task_policy(struct task_struct *task)
2458{
2459	struct mempolicy *pol;
2460
2461	task_lock(task);
2462	pol = task->mempolicy;
2463	task->mempolicy = NULL;
2464	task_unlock(task);
2465	mpol_put(pol);
2466}
2467
2468static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2469{
2470	pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2471	rb_erase(&n->nd, &sp->root);
2472	sp_free(n);
2473}
2474
2475static void sp_node_init(struct sp_node *node, unsigned long start,
2476			unsigned long end, struct mempolicy *pol)
2477{
2478	node->start = start;
2479	node->end = end;
2480	node->policy = pol;
2481}
2482
2483static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2484				struct mempolicy *pol)
2485{
2486	struct sp_node *n;
2487	struct mempolicy *newpol;
2488
2489	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2490	if (!n)
2491		return NULL;
2492
2493	newpol = mpol_dup(pol);
2494	if (IS_ERR(newpol)) {
2495		kmem_cache_free(sn_cache, n);
2496		return NULL;
2497	}
2498	newpol->flags |= MPOL_F_SHARED;
2499	sp_node_init(n, start, end, newpol);
2500
2501	return n;
2502}
2503
2504/* Replace a policy range. */
2505static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2506				 unsigned long end, struct sp_node *new)
2507{
2508	struct sp_node *n;
2509	struct sp_node *n_new = NULL;
2510	struct mempolicy *mpol_new = NULL;
2511	int ret = 0;
2512
2513restart:
2514	write_lock(&sp->lock);
2515	n = sp_lookup(sp, start, end);
2516	/* Take care of old policies in the same range. */
2517	while (n && n->start < end) {
2518		struct rb_node *next = rb_next(&n->nd);
2519		if (n->start >= start) {
2520			if (n->end <= end)
2521				sp_delete(sp, n);
2522			else
2523				n->start = end;
2524		} else {
2525			/* Old policy spanning whole new range. */
2526			if (n->end > end) {
2527				if (!n_new)
2528					goto alloc_new;
2529
2530				*mpol_new = *n->policy;
2531				atomic_set(&mpol_new->refcnt, 1);
2532				sp_node_init(n_new, end, n->end, mpol_new);
2533				n->end = start;
2534				sp_insert(sp, n_new);
2535				n_new = NULL;
2536				mpol_new = NULL;
2537				break;
2538			} else
2539				n->end = start;
2540		}
2541		if (!next)
2542			break;
2543		n = rb_entry(next, struct sp_node, nd);
2544	}
2545	if (new)
2546		sp_insert(sp, new);
2547	write_unlock(&sp->lock);
2548	ret = 0;
2549
2550err_out:
2551	if (mpol_new)
2552		mpol_put(mpol_new);
2553	if (n_new)
2554		kmem_cache_free(sn_cache, n_new);
2555
2556	return ret;
2557
2558alloc_new:
2559	write_unlock(&sp->lock);
2560	ret = -ENOMEM;
2561	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2562	if (!n_new)
2563		goto err_out;
2564	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2565	if (!mpol_new)
2566		goto err_out;
2567	goto restart;
2568}
2569
2570/**
2571 * mpol_shared_policy_init - initialize shared policy for inode
2572 * @sp: pointer to inode shared policy
2573 * @mpol:  struct mempolicy to install
2574 *
2575 * Install non-NULL @mpol in inode's shared policy rb-tree.
2576 * On entry, the current task has a reference on a non-NULL @mpol.
2577 * This must be released on exit.
2578 * This is called at get_inode() calls and we can use GFP_KERNEL.
2579 */
2580void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2581{
2582	int ret;
2583
2584	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2585	rwlock_init(&sp->lock);
2586
2587	if (mpol) {
2588		struct vm_area_struct pvma;
2589		struct mempolicy *new;
2590		NODEMASK_SCRATCH(scratch);
2591
2592		if (!scratch)
2593			goto put_mpol;
2594		/* contextualize the tmpfs mount point mempolicy */
2595		new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2596		if (IS_ERR(new))
2597			goto free_scratch; /* no valid nodemask intersection */
2598
2599		task_lock(current);
2600		ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2601		task_unlock(current);
2602		if (ret)
2603			goto put_new;
2604
2605		/* Create pseudo-vma that contains just the policy */
2606		vma_init(&pvma, NULL);
2607		pvma.vm_end = TASK_SIZE;	/* policy covers entire file */
2608		mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2609
2610put_new:
2611		mpol_put(new);			/* drop initial ref */
2612free_scratch:
2613		NODEMASK_SCRATCH_FREE(scratch);
2614put_mpol:
2615		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
2616	}
2617}
2618
2619int mpol_set_shared_policy(struct shared_policy *info,
2620			struct vm_area_struct *vma, struct mempolicy *npol)
2621{
2622	int err;
2623	struct sp_node *new = NULL;
2624	unsigned long sz = vma_pages(vma);
2625
2626	pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2627		 vma->vm_pgoff,
2628		 sz, npol ? npol->mode : -1,
2629		 npol ? npol->flags : -1,
2630		 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2631
2632	if (npol) {
2633		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2634		if (!new)
2635			return -ENOMEM;
2636	}
2637	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2638	if (err && new)
2639		sp_free(new);
2640	return err;
2641}
2642
2643/* Free a backing policy store on inode delete. */
2644void mpol_free_shared_policy(struct shared_policy *p)
2645{
2646	struct sp_node *n;
2647	struct rb_node *next;
2648
2649	if (!p->root.rb_node)
2650		return;
2651	write_lock(&p->lock);
2652	next = rb_first(&p->root);
2653	while (next) {
2654		n = rb_entry(next, struct sp_node, nd);
2655		next = rb_next(&n->nd);
2656		sp_delete(p, n);
2657	}
2658	write_unlock(&p->lock);
2659}
2660
2661#ifdef CONFIG_NUMA_BALANCING
2662static int __initdata numabalancing_override;
2663
2664static void __init check_numabalancing_enable(void)
2665{
2666	bool numabalancing_default = false;
2667
2668	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2669		numabalancing_default = true;
2670
2671	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2672	if (numabalancing_override)
2673		set_numabalancing_state(numabalancing_override == 1);
2674
2675	if (num_online_nodes() > 1 && !numabalancing_override) {
2676		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2677			numabalancing_default ? "Enabling" : "Disabling");
2678		set_numabalancing_state(numabalancing_default);
2679	}
2680}
2681
2682static int __init setup_numabalancing(char *str)
2683{
2684	int ret = 0;
2685	if (!str)
2686		goto out;
2687
2688	if (!strcmp(str, "enable")) {
2689		numabalancing_override = 1;
2690		ret = 1;
2691	} else if (!strcmp(str, "disable")) {
2692		numabalancing_override = -1;
2693		ret = 1;
2694	}
2695out:
2696	if (!ret)
2697		pr_warn("Unable to parse numa_balancing=\n");
2698
2699	return ret;
2700}
2701__setup("numa_balancing=", setup_numabalancing);
2702#else
2703static inline void __init check_numabalancing_enable(void)
2704{
2705}
2706#endif /* CONFIG_NUMA_BALANCING */
2707
2708/* assumes fs == KERNEL_DS */
2709void __init numa_policy_init(void)
2710{
2711	nodemask_t interleave_nodes;
2712	unsigned long largest = 0;
2713	int nid, prefer = 0;
2714
2715	policy_cache = kmem_cache_create("numa_policy",
2716					 sizeof(struct mempolicy),
2717					 0, SLAB_PANIC, NULL);
2718
2719	sn_cache = kmem_cache_create("shared_policy_node",
2720				     sizeof(struct sp_node),
2721				     0, SLAB_PANIC, NULL);
2722
2723	for_each_node(nid) {
2724		preferred_node_policy[nid] = (struct mempolicy) {
2725			.refcnt = ATOMIC_INIT(1),
2726			.mode = MPOL_PREFERRED,
2727			.flags = MPOL_F_MOF | MPOL_F_MORON,
2728			.v = { .preferred_node = nid, },
2729		};
2730	}
2731
2732	/*
2733	 * Set interleaving policy for system init. Interleaving is only
2734	 * enabled across suitably sized nodes (default is >= 16MB), or
2735	 * fall back to the largest node if they're all smaller.
2736	 */
2737	nodes_clear(interleave_nodes);
2738	for_each_node_state(nid, N_MEMORY) {
2739		unsigned long total_pages = node_present_pages(nid);
2740
2741		/* Preserve the largest node */
2742		if (largest < total_pages) {
2743			largest = total_pages;
2744			prefer = nid;
2745		}
2746
2747		/* Interleave this node? */
2748		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2749			node_set(nid, interleave_nodes);
2750	}
2751
2752	/* All too small, use the largest */
2753	if (unlikely(nodes_empty(interleave_nodes)))
2754		node_set(prefer, interleave_nodes);
2755
2756	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2757		pr_err("%s: interleaving failed\n", __func__);
2758
2759	check_numabalancing_enable();
2760}
2761
2762/* Reset policy of current process to default */
2763void numa_default_policy(void)
2764{
2765	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2766}
2767
2768/*
2769 * Parse and format mempolicy from/to strings
2770 */
2771
2772/*
2773 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2774 */
2775static const char * const policy_modes[] =
2776{
2777	[MPOL_DEFAULT]    = "default",
2778	[MPOL_PREFERRED]  = "prefer",
2779	[MPOL_BIND]       = "bind",
2780	[MPOL_INTERLEAVE] = "interleave",
2781	[MPOL_LOCAL]      = "local",
2782};
2783
2784
2785#ifdef CONFIG_TMPFS
2786/**
2787 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2788 * @str:  string containing mempolicy to parse
2789 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2790 *
2791 * Format of input:
2792 *	<mode>[=<flags>][:<nodelist>]
2793 *
2794 * On success, returns 0, else 1
2795 */
2796int mpol_parse_str(char *str, struct mempolicy **mpol)
2797{
2798	struct mempolicy *new = NULL;
 
2799	unsigned short mode_flags;
2800	nodemask_t nodes;
2801	char *nodelist = strchr(str, ':');
2802	char *flags = strchr(str, '=');
2803	int err = 1, mode;
2804
2805	if (nodelist) {
2806		/* NUL-terminate mode or flags string */
2807		*nodelist++ = '\0';
2808		if (nodelist_parse(nodelist, nodes))
2809			goto out;
2810		if (!nodes_subset(nodes, node_states[N_MEMORY]))
2811			goto out;
2812	} else
2813		nodes_clear(nodes);
2814
2815	if (flags)
2816		*flags++ = '\0';	/* terminate mode string */
2817
2818	mode = match_string(policy_modes, MPOL_MAX, str);
2819	if (mode < 0)
 
 
 
 
2820		goto out;
2821
2822	switch (mode) {
2823	case MPOL_PREFERRED:
2824		/*
2825		 * Insist on a nodelist of one node only
2826		 */
2827		if (nodelist) {
2828			char *rest = nodelist;
2829			while (isdigit(*rest))
2830				rest++;
2831			if (*rest)
2832				goto out;
2833		}
2834		break;
2835	case MPOL_INTERLEAVE:
2836		/*
2837		 * Default to online nodes with memory if no nodelist
2838		 */
2839		if (!nodelist)
2840			nodes = node_states[N_MEMORY];
2841		break;
2842	case MPOL_LOCAL:
2843		/*
2844		 * Don't allow a nodelist;  mpol_new() checks flags
2845		 */
2846		if (nodelist)
2847			goto out;
2848		mode = MPOL_PREFERRED;
2849		break;
2850	case MPOL_DEFAULT:
2851		/*
2852		 * Insist on a empty nodelist
2853		 */
2854		if (!nodelist)
2855			err = 0;
2856		goto out;
2857	case MPOL_BIND:
2858		/*
2859		 * Insist on a nodelist
2860		 */
2861		if (!nodelist)
2862			goto out;
2863	}
2864
2865	mode_flags = 0;
2866	if (flags) {
2867		/*
2868		 * Currently, we only support two mutually exclusive
2869		 * mode flags.
2870		 */
2871		if (!strcmp(flags, "static"))
2872			mode_flags |= MPOL_F_STATIC_NODES;
2873		else if (!strcmp(flags, "relative"))
2874			mode_flags |= MPOL_F_RELATIVE_NODES;
2875		else
2876			goto out;
2877	}
2878
2879	new = mpol_new(mode, mode_flags, &nodes);
2880	if (IS_ERR(new))
2881		goto out;
2882
2883	/*
2884	 * Save nodes for mpol_to_str() to show the tmpfs mount options
2885	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2886	 */
2887	if (mode != MPOL_PREFERRED)
2888		new->v.nodes = nodes;
2889	else if (nodelist)
2890		new->v.preferred_node = first_node(nodes);
2891	else
2892		new->flags |= MPOL_F_LOCAL;
2893
2894	/*
2895	 * Save nodes for contextualization: this will be used to "clone"
2896	 * the mempolicy in a specific context [cpuset] at a later time.
2897	 */
2898	new->w.user_nodemask = nodes;
2899
2900	err = 0;
2901
2902out:
2903	/* Restore string for error message */
2904	if (nodelist)
2905		*--nodelist = ':';
2906	if (flags)
2907		*--flags = '=';
2908	if (!err)
2909		*mpol = new;
2910	return err;
2911}
2912#endif /* CONFIG_TMPFS */
2913
2914/**
2915 * mpol_to_str - format a mempolicy structure for printing
2916 * @buffer:  to contain formatted mempolicy string
2917 * @maxlen:  length of @buffer
2918 * @pol:  pointer to mempolicy to be formatted
2919 *
2920 * Convert @pol into a string.  If @buffer is too short, truncate the string.
2921 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2922 * longest flag, "relative", and to display at least a few node ids.
2923 */
2924void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2925{
2926	char *p = buffer;
2927	nodemask_t nodes = NODE_MASK_NONE;
2928	unsigned short mode = MPOL_DEFAULT;
2929	unsigned short flags = 0;
2930
2931	if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2932		mode = pol->mode;
2933		flags = pol->flags;
2934	}
2935
2936	switch (mode) {
2937	case MPOL_DEFAULT:
2938		break;
2939	case MPOL_PREFERRED:
2940		if (flags & MPOL_F_LOCAL)
2941			mode = MPOL_LOCAL;
2942		else
2943			node_set(pol->v.preferred_node, nodes);
2944		break;
2945	case MPOL_BIND:
2946	case MPOL_INTERLEAVE:
2947		nodes = pol->v.nodes;
2948		break;
2949	default:
2950		WARN_ON_ONCE(1);
2951		snprintf(p, maxlen, "unknown");
2952		return;
2953	}
2954
2955	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2956
2957	if (flags & MPOL_MODE_FLAGS) {
2958		p += snprintf(p, buffer + maxlen - p, "=");
2959
2960		/*
2961		 * Currently, the only defined flags are mutually exclusive
2962		 */
2963		if (flags & MPOL_F_STATIC_NODES)
2964			p += snprintf(p, buffer + maxlen - p, "static");
2965		else if (flags & MPOL_F_RELATIVE_NODES)
2966			p += snprintf(p, buffer + maxlen - p, "relative");
2967	}
2968
2969	if (!nodes_empty(nodes))
2970		p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2971			       nodemask_pr_args(&nodes));
2972}
v4.17
 
   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#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  69
  70#include <linux/mempolicy.h>
  71#include <linux/mm.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_PREFERRED,
 125	.flags = MPOL_F_LOCAL,
 126};
 127
 128static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 129
 130struct mempolicy *get_task_policy(struct task_struct *p)
 131{
 132	struct mempolicy *pol = p->mempolicy;
 133	int node;
 134
 135	if (pol)
 136		return pol;
 137
 138	node = numa_node_id();
 139	if (node != NUMA_NO_NODE) {
 140		pol = &preferred_node_policy[node];
 141		/* preferred_node_policy is not initialised early in boot */
 142		if (pol->mode)
 143			return pol;
 144	}
 145
 146	return &default_policy;
 147}
 148
 149static const struct mempolicy_operations {
 150	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 151	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
 152} mpol_ops[MPOL_MAX];
 153
 154static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 155{
 156	return pol->flags & MPOL_MODE_FLAGS;
 157}
 158
 159static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 160				   const nodemask_t *rel)
 161{
 162	nodemask_t tmp;
 163	nodes_fold(tmp, *orig, nodes_weight(*rel));
 164	nodes_onto(*ret, tmp, *rel);
 165}
 166
 167static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
 168{
 169	if (nodes_empty(*nodes))
 170		return -EINVAL;
 171	pol->v.nodes = *nodes;
 172	return 0;
 173}
 174
 175static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 176{
 177	if (!nodes)
 178		pol->flags |= MPOL_F_LOCAL;	/* local allocation */
 179	else if (nodes_empty(*nodes))
 180		return -EINVAL;			/*  no allowed nodes */
 181	else
 182		pol->v.preferred_node = first_node(*nodes);
 183	return 0;
 184}
 185
 186static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
 187{
 188	if (nodes_empty(*nodes))
 189		return -EINVAL;
 190	pol->v.nodes = *nodes;
 191	return 0;
 192}
 193
 194/*
 195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 196 * any, for the new policy.  mpol_new() has already validated the nodes
 197 * parameter with respect to the policy mode and flags.  But, we need to
 198 * handle an empty nodemask with MPOL_PREFERRED here.
 199 *
 200 * Must be called holding task's alloc_lock to protect task's mems_allowed
 201 * and mempolicy.  May also be called holding the mmap_semaphore for write.
 202 */
 203static int mpol_set_nodemask(struct mempolicy *pol,
 204		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 205{
 206	int ret;
 207
 208	/* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
 209	if (pol == NULL)
 210		return 0;
 211	/* Check N_MEMORY */
 212	nodes_and(nsc->mask1,
 213		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 214
 215	VM_BUG_ON(!nodes);
 216	if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
 217		nodes = NULL;	/* explicit local allocation */
 218	else {
 219		if (pol->flags & MPOL_F_RELATIVE_NODES)
 220			mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
 221		else
 222			nodes_and(nsc->mask2, *nodes, nsc->mask1);
 223
 224		if (mpol_store_user_nodemask(pol))
 225			pol->w.user_nodemask = *nodes;
 226		else
 227			pol->w.cpuset_mems_allowed =
 228						cpuset_current_mems_allowed;
 229	}
 230
 231	if (nodes)
 232		ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 233	else
 234		ret = mpol_ops[pol->mode].create(pol, NULL);
 235	return ret;
 236}
 237
 238/*
 239 * This function just creates a new policy, does some check and simple
 240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 241 */
 242static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 243				  nodemask_t *nodes)
 244{
 245	struct mempolicy *policy;
 246
 247	pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 248		 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
 249
 250	if (mode == MPOL_DEFAULT) {
 251		if (nodes && !nodes_empty(*nodes))
 252			return ERR_PTR(-EINVAL);
 253		return NULL;
 254	}
 255	VM_BUG_ON(!nodes);
 256
 257	/*
 258	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 259	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 260	 * All other modes require a valid pointer to a non-empty nodemask.
 261	 */
 262	if (mode == MPOL_PREFERRED) {
 263		if (nodes_empty(*nodes)) {
 264			if (((flags & MPOL_F_STATIC_NODES) ||
 265			     (flags & MPOL_F_RELATIVE_NODES)))
 266				return ERR_PTR(-EINVAL);
 267		}
 268	} else if (mode == MPOL_LOCAL) {
 269		if (!nodes_empty(*nodes) ||
 270		    (flags & MPOL_F_STATIC_NODES) ||
 271		    (flags & MPOL_F_RELATIVE_NODES))
 272			return ERR_PTR(-EINVAL);
 273		mode = MPOL_PREFERRED;
 274	} else if (nodes_empty(*nodes))
 275		return ERR_PTR(-EINVAL);
 276	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 277	if (!policy)
 278		return ERR_PTR(-ENOMEM);
 279	atomic_set(&policy->refcnt, 1);
 280	policy->mode = mode;
 281	policy->flags = flags;
 282
 283	return policy;
 284}
 285
 286/* Slow path of a mpol destructor. */
 287void __mpol_put(struct mempolicy *p)
 288{
 289	if (!atomic_dec_and_test(&p->refcnt))
 290		return;
 291	kmem_cache_free(policy_cache, p);
 292}
 293
 294static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
 295{
 296}
 297
 298static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 299{
 300	nodemask_t tmp;
 301
 302	if (pol->flags & MPOL_F_STATIC_NODES)
 303		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 304	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 305		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 306	else {
 307		nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
 308								*nodes);
 309		pol->w.cpuset_mems_allowed = tmp;
 310	}
 311
 312	if (nodes_empty(tmp))
 313		tmp = *nodes;
 314
 315	pol->v.nodes = tmp;
 316}
 317
 318static void mpol_rebind_preferred(struct mempolicy *pol,
 319						const nodemask_t *nodes)
 320{
 321	nodemask_t tmp;
 322
 323	if (pol->flags & MPOL_F_STATIC_NODES) {
 324		int node = first_node(pol->w.user_nodemask);
 325
 326		if (node_isset(node, *nodes)) {
 327			pol->v.preferred_node = node;
 328			pol->flags &= ~MPOL_F_LOCAL;
 329		} else
 330			pol->flags |= MPOL_F_LOCAL;
 331	} else if (pol->flags & MPOL_F_RELATIVE_NODES) {
 332		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 333		pol->v.preferred_node = first_node(tmp);
 334	} else if (!(pol->flags & MPOL_F_LOCAL)) {
 335		pol->v.preferred_node = node_remap(pol->v.preferred_node,
 336						   pol->w.cpuset_mems_allowed,
 337						   *nodes);
 338		pol->w.cpuset_mems_allowed = *nodes;
 339	}
 340}
 341
 342/*
 343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 344 *
 345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
 346 * policies are protected by task->mems_allowed_seq to prevent a premature
 347 * OOM/allocation failure due to parallel nodemask modification.
 348 */
 349static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
 350{
 351	if (!pol)
 352		return;
 353	if (!mpol_store_user_nodemask(pol) &&
 354	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 355		return;
 356
 357	mpol_ops[pol->mode].rebind(pol, newmask);
 358}
 359
 360/*
 361 * Wrapper for mpol_rebind_policy() that just requires task
 362 * pointer, and updates task mempolicy.
 363 *
 364 * Called with task's alloc_lock held.
 365 */
 366
 367void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
 368{
 369	mpol_rebind_policy(tsk->mempolicy, new);
 370}
 371
 372/*
 373 * Rebind each vma in mm to new nodemask.
 374 *
 375 * Call holding a reference to mm.  Takes mm->mmap_sem during call.
 376 */
 377
 378void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 379{
 380	struct vm_area_struct *vma;
 381
 382	down_write(&mm->mmap_sem);
 383	for (vma = mm->mmap; vma; vma = vma->vm_next)
 384		mpol_rebind_policy(vma->vm_policy, new);
 385	up_write(&mm->mmap_sem);
 386}
 387
 388static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 389	[MPOL_DEFAULT] = {
 390		.rebind = mpol_rebind_default,
 391	},
 392	[MPOL_INTERLEAVE] = {
 393		.create = mpol_new_interleave,
 394		.rebind = mpol_rebind_nodemask,
 395	},
 396	[MPOL_PREFERRED] = {
 397		.create = mpol_new_preferred,
 398		.rebind = mpol_rebind_preferred,
 399	},
 400	[MPOL_BIND] = {
 401		.create = mpol_new_bind,
 402		.rebind = mpol_rebind_nodemask,
 403	},
 404};
 405
 406static void migrate_page_add(struct page *page, struct list_head *pagelist,
 407				unsigned long flags);
 408
 409struct queue_pages {
 410	struct list_head *pagelist;
 411	unsigned long flags;
 412	nodemask_t *nmask;
 413	struct vm_area_struct *prev;
 414};
 415
 416/*
 417 * Check if the page's nid is in qp->nmask.
 418 *
 419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
 420 * in the invert of qp->nmask.
 421 */
 422static inline bool queue_pages_required(struct page *page,
 423					struct queue_pages *qp)
 424{
 425	int nid = page_to_nid(page);
 426	unsigned long flags = qp->flags;
 427
 428	return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
 429}
 430
 
 
 
 
 
 
 
 
 
 
 431static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
 432				unsigned long end, struct mm_walk *walk)
 433{
 434	int ret = 0;
 435	struct page *page;
 436	struct queue_pages *qp = walk->private;
 437	unsigned long flags;
 438
 439	if (unlikely(is_pmd_migration_entry(*pmd))) {
 440		ret = 1;
 441		goto unlock;
 442	}
 443	page = pmd_page(*pmd);
 444	if (is_huge_zero_page(page)) {
 445		spin_unlock(ptl);
 446		__split_huge_pmd(walk->vma, pmd, addr, false, NULL);
 
 447		goto out;
 448	}
 449	if (!queue_pages_required(page, qp)) {
 450		ret = 1;
 451		goto unlock;
 452	}
 453
 454	ret = 1;
 455	flags = qp->flags;
 456	/* go to thp migration */
 457	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
 458		migrate_page_add(page, qp->pagelist, flags);
 
 
 
 
 
 
 459unlock:
 460	spin_unlock(ptl);
 461out:
 462	return ret;
 463}
 464
 465/*
 466 * Scan through pages checking if pages follow certain conditions,
 467 * and move them to the pagelist if they do.
 
 
 
 
 
 
 
 468 */
 469static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
 470			unsigned long end, struct mm_walk *walk)
 471{
 472	struct vm_area_struct *vma = walk->vma;
 473	struct page *page;
 474	struct queue_pages *qp = walk->private;
 475	unsigned long flags = qp->flags;
 476	int ret;
 
 477	pte_t *pte;
 478	spinlock_t *ptl;
 479
 480	ptl = pmd_trans_huge_lock(pmd, vma);
 481	if (ptl) {
 482		ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
 483		if (ret)
 484			return 0;
 485	}
 
 486
 487	if (pmd_trans_unstable(pmd))
 488		return 0;
 489
 490	pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
 491	for (; addr != end; pte++, addr += PAGE_SIZE) {
 492		if (!pte_present(*pte))
 493			continue;
 494		page = vm_normal_page(vma, addr, *pte);
 495		if (!page)
 496			continue;
 497		/*
 498		 * vm_normal_page() filters out zero pages, but there might
 499		 * still be PageReserved pages to skip, perhaps in a VDSO.
 500		 */
 501		if (PageReserved(page))
 502			continue;
 503		if (!queue_pages_required(page, qp))
 504			continue;
 505		migrate_page_add(page, qp->pagelist, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 506	}
 507	pte_unmap_unlock(pte - 1, ptl);
 508	cond_resched();
 509	return 0;
 
 
 
 
 510}
 511
 512static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
 513			       unsigned long addr, unsigned long end,
 514			       struct mm_walk *walk)
 515{
 516#ifdef CONFIG_HUGETLB_PAGE
 517	struct queue_pages *qp = walk->private;
 518	unsigned long flags = qp->flags;
 519	struct page *page;
 520	spinlock_t *ptl;
 521	pte_t entry;
 522
 523	ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 524	entry = huge_ptep_get(pte);
 525	if (!pte_present(entry))
 526		goto unlock;
 527	page = pte_page(entry);
 528	if (!queue_pages_required(page, qp))
 529		goto unlock;
 530	/* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
 531	if (flags & (MPOL_MF_MOVE_ALL) ||
 532	    (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
 533		isolate_huge_page(page, qp->pagelist);
 534unlock:
 535	spin_unlock(ptl);
 536#else
 537	BUG();
 538#endif
 539	return 0;
 540}
 541
 542#ifdef CONFIG_NUMA_BALANCING
 543/*
 544 * This is used to mark a range of virtual addresses to be inaccessible.
 545 * These are later cleared by a NUMA hinting fault. Depending on these
 546 * faults, pages may be migrated for better NUMA placement.
 547 *
 548 * This is assuming that NUMA faults are handled using PROT_NONE. If
 549 * an architecture makes a different choice, it will need further
 550 * changes to the core.
 551 */
 552unsigned long change_prot_numa(struct vm_area_struct *vma,
 553			unsigned long addr, unsigned long end)
 554{
 555	int nr_updated;
 556
 557	nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
 558	if (nr_updated)
 559		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 560
 561	return nr_updated;
 562}
 563#else
 564static unsigned long change_prot_numa(struct vm_area_struct *vma,
 565			unsigned long addr, unsigned long end)
 566{
 567	return 0;
 568}
 569#endif /* CONFIG_NUMA_BALANCING */
 570
 571static int queue_pages_test_walk(unsigned long start, unsigned long end,
 572				struct mm_walk *walk)
 573{
 574	struct vm_area_struct *vma = walk->vma;
 575	struct queue_pages *qp = walk->private;
 576	unsigned long endvma = vma->vm_end;
 577	unsigned long flags = qp->flags;
 578
 579	if (!vma_migratable(vma))
 
 
 
 
 
 580		return 1;
 581
 582	if (endvma > end)
 583		endvma = end;
 584	if (vma->vm_start > start)
 585		start = vma->vm_start;
 586
 587	if (!(flags & MPOL_MF_DISCONTIG_OK)) {
 588		if (!vma->vm_next && vma->vm_end < end)
 589			return -EFAULT;
 590		if (qp->prev && qp->prev->vm_end < vma->vm_start)
 591			return -EFAULT;
 592	}
 593
 594	qp->prev = vma;
 595
 596	if (flags & MPOL_MF_LAZY) {
 597		/* Similar to task_numa_work, skip inaccessible VMAs */
 598		if (!is_vm_hugetlb_page(vma) &&
 599			(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
 600			!(vma->vm_flags & VM_MIXEDMAP))
 601			change_prot_numa(vma, start, endvma);
 602		return 1;
 603	}
 604
 605	/* queue pages from current vma */
 606	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
 607		return 0;
 608	return 1;
 609}
 610
 
 
 
 
 
 
 611/*
 612 * Walk through page tables and collect pages to be migrated.
 613 *
 614 * If pages found in a given range are on a set of nodes (determined by
 615 * @nodes and @flags,) it's isolated and queued to the pagelist which is
 616 * passed via @private.)
 
 
 
 
 
 
 
 
 617 */
 618static int
 619queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 620		nodemask_t *nodes, unsigned long flags,
 621		struct list_head *pagelist)
 622{
 623	struct queue_pages qp = {
 624		.pagelist = pagelist,
 625		.flags = flags,
 626		.nmask = nodes,
 627		.prev = NULL,
 628	};
 629	struct mm_walk queue_pages_walk = {
 630		.hugetlb_entry = queue_pages_hugetlb,
 631		.pmd_entry = queue_pages_pte_range,
 632		.test_walk = queue_pages_test_walk,
 633		.mm = mm,
 634		.private = &qp,
 635	};
 636
 637	return walk_page_range(start, end, &queue_pages_walk);
 638}
 639
 640/*
 641 * Apply policy to a single VMA
 642 * This must be called with the mmap_sem held for writing.
 643 */
 644static int vma_replace_policy(struct vm_area_struct *vma,
 645						struct mempolicy *pol)
 646{
 647	int err;
 648	struct mempolicy *old;
 649	struct mempolicy *new;
 650
 651	pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 652		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 653		 vma->vm_ops, vma->vm_file,
 654		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 655
 656	new = mpol_dup(pol);
 657	if (IS_ERR(new))
 658		return PTR_ERR(new);
 659
 660	if (vma->vm_ops && vma->vm_ops->set_policy) {
 661		err = vma->vm_ops->set_policy(vma, new);
 662		if (err)
 663			goto err_out;
 664	}
 665
 666	old = vma->vm_policy;
 667	vma->vm_policy = new; /* protected by mmap_sem */
 668	mpol_put(old);
 669
 670	return 0;
 671 err_out:
 672	mpol_put(new);
 673	return err;
 674}
 675
 676/* Step 2: apply policy to a range and do splits. */
 677static int mbind_range(struct mm_struct *mm, unsigned long start,
 678		       unsigned long end, struct mempolicy *new_pol)
 679{
 680	struct vm_area_struct *next;
 681	struct vm_area_struct *prev;
 682	struct vm_area_struct *vma;
 683	int err = 0;
 684	pgoff_t pgoff;
 685	unsigned long vmstart;
 686	unsigned long vmend;
 687
 688	vma = find_vma(mm, start);
 689	if (!vma || vma->vm_start > start)
 690		return -EFAULT;
 691
 692	prev = vma->vm_prev;
 693	if (start > vma->vm_start)
 694		prev = vma;
 695
 696	for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 697		next = vma->vm_next;
 698		vmstart = max(start, vma->vm_start);
 699		vmend   = min(end, vma->vm_end);
 700
 701		if (mpol_equal(vma_policy(vma), new_pol))
 702			continue;
 703
 704		pgoff = vma->vm_pgoff +
 705			((vmstart - vma->vm_start) >> PAGE_SHIFT);
 706		prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 707				 vma->anon_vma, vma->vm_file, pgoff,
 708				 new_pol, vma->vm_userfaultfd_ctx);
 709		if (prev) {
 710			vma = prev;
 711			next = vma->vm_next;
 712			if (mpol_equal(vma_policy(vma), new_pol))
 713				continue;
 714			/* vma_merge() joined vma && vma->next, case 8 */
 715			goto replace;
 716		}
 717		if (vma->vm_start != vmstart) {
 718			err = split_vma(vma->vm_mm, vma, vmstart, 1);
 719			if (err)
 720				goto out;
 721		}
 722		if (vma->vm_end != vmend) {
 723			err = split_vma(vma->vm_mm, vma, vmend, 0);
 724			if (err)
 725				goto out;
 726		}
 727 replace:
 728		err = vma_replace_policy(vma, new_pol);
 729		if (err)
 730			goto out;
 731	}
 732
 733 out:
 734	return err;
 735}
 736
 737/* Set the process memory policy */
 738static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 739			     nodemask_t *nodes)
 740{
 741	struct mempolicy *new, *old;
 742	NODEMASK_SCRATCH(scratch);
 743	int ret;
 744
 745	if (!scratch)
 746		return -ENOMEM;
 747
 748	new = mpol_new(mode, flags, nodes);
 749	if (IS_ERR(new)) {
 750		ret = PTR_ERR(new);
 751		goto out;
 752	}
 753
 754	task_lock(current);
 755	ret = mpol_set_nodemask(new, nodes, scratch);
 756	if (ret) {
 757		task_unlock(current);
 758		mpol_put(new);
 759		goto out;
 760	}
 761	old = current->mempolicy;
 762	current->mempolicy = new;
 763	if (new && new->mode == MPOL_INTERLEAVE)
 764		current->il_prev = MAX_NUMNODES-1;
 765	task_unlock(current);
 766	mpol_put(old);
 767	ret = 0;
 768out:
 769	NODEMASK_SCRATCH_FREE(scratch);
 770	return ret;
 771}
 772
 773/*
 774 * Return nodemask for policy for get_mempolicy() query
 775 *
 776 * Called with task's alloc_lock held
 777 */
 778static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 779{
 780	nodes_clear(*nodes);
 781	if (p == &default_policy)
 782		return;
 783
 784	switch (p->mode) {
 785	case MPOL_BIND:
 786		/* Fall through */
 787	case MPOL_INTERLEAVE:
 788		*nodes = p->v.nodes;
 789		break;
 790	case MPOL_PREFERRED:
 791		if (!(p->flags & MPOL_F_LOCAL))
 792			node_set(p->v.preferred_node, *nodes);
 793		/* else return empty node mask for local allocation */
 794		break;
 795	default:
 796		BUG();
 797	}
 798}
 799
 800static int lookup_node(unsigned long addr)
 801{
 802	struct page *p;
 803	int err;
 804
 805	err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
 
 806	if (err >= 0) {
 807		err = page_to_nid(p);
 808		put_page(p);
 809	}
 
 
 810	return err;
 811}
 812
 813/* Retrieve NUMA policy */
 814static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 815			     unsigned long addr, unsigned long flags)
 816{
 817	int err;
 818	struct mm_struct *mm = current->mm;
 819	struct vm_area_struct *vma = NULL;
 820	struct mempolicy *pol = current->mempolicy;
 821
 822	if (flags &
 823		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 824		return -EINVAL;
 825
 826	if (flags & MPOL_F_MEMS_ALLOWED) {
 827		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 828			return -EINVAL;
 829		*policy = 0;	/* just so it's initialized */
 830		task_lock(current);
 831		*nmask  = cpuset_current_mems_allowed;
 832		task_unlock(current);
 833		return 0;
 834	}
 835
 836	if (flags & MPOL_F_ADDR) {
 837		/*
 838		 * Do NOT fall back to task policy if the
 839		 * vma/shared policy at addr is NULL.  We
 840		 * want to return MPOL_DEFAULT in this case.
 841		 */
 842		down_read(&mm->mmap_sem);
 843		vma = find_vma_intersection(mm, addr, addr+1);
 844		if (!vma) {
 845			up_read(&mm->mmap_sem);
 846			return -EFAULT;
 847		}
 848		if (vma->vm_ops && vma->vm_ops->get_policy)
 849			pol = vma->vm_ops->get_policy(vma, addr);
 850		else
 851			pol = vma->vm_policy;
 852	} else if (addr)
 853		return -EINVAL;
 854
 855	if (!pol)
 856		pol = &default_policy;	/* indicates default behavior */
 857
 858	if (flags & MPOL_F_NODE) {
 859		if (flags & MPOL_F_ADDR) {
 860			err = lookup_node(addr);
 
 
 
 
 
 
 
 
 
 861			if (err < 0)
 862				goto out;
 863			*policy = err;
 864		} else if (pol == current->mempolicy &&
 865				pol->mode == MPOL_INTERLEAVE) {
 866			*policy = next_node_in(current->il_prev, pol->v.nodes);
 867		} else {
 868			err = -EINVAL;
 869			goto out;
 870		}
 871	} else {
 872		*policy = pol == &default_policy ? MPOL_DEFAULT :
 873						pol->mode;
 874		/*
 875		 * Internal mempolicy flags must be masked off before exposing
 876		 * the policy to userspace.
 877		 */
 878		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 879	}
 880
 881	err = 0;
 882	if (nmask) {
 883		if (mpol_store_user_nodemask(pol)) {
 884			*nmask = pol->w.user_nodemask;
 885		} else {
 886			task_lock(current);
 887			get_policy_nodemask(pol, nmask);
 888			task_unlock(current);
 889		}
 890	}
 891
 892 out:
 893	mpol_cond_put(pol);
 894	if (vma)
 895		up_read(&current->mm->mmap_sem);
 
 
 896	return err;
 897}
 898
 899#ifdef CONFIG_MIGRATION
 900/*
 901 * page migration, thp tail pages can be passed.
 902 */
 903static void migrate_page_add(struct page *page, struct list_head *pagelist,
 904				unsigned long flags)
 905{
 906	struct page *head = compound_head(page);
 907	/*
 908	 * Avoid migrating a page that is shared with others.
 909	 */
 910	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
 911		if (!isolate_lru_page(head)) {
 912			list_add_tail(&head->lru, pagelist);
 913			mod_node_page_state(page_pgdat(head),
 914				NR_ISOLATED_ANON + page_is_file_cache(head),
 915				hpage_nr_pages(head));
 
 
 
 
 
 
 
 
 
 916		}
 917	}
 
 
 918}
 919
 920/* page allocation callback for NUMA node migration */
 921struct page *alloc_new_node_page(struct page *page, unsigned long node)
 922{
 923	if (PageHuge(page))
 924		return alloc_huge_page_node(page_hstate(compound_head(page)),
 925					node);
 926	else if (PageTransHuge(page)) {
 927		struct page *thp;
 928
 929		thp = alloc_pages_node(node,
 930			(GFP_TRANSHUGE | __GFP_THISNODE),
 931			HPAGE_PMD_ORDER);
 932		if (!thp)
 933			return NULL;
 934		prep_transhuge_page(thp);
 935		return thp;
 936	} else
 937		return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
 938						    __GFP_THISNODE, 0);
 939}
 940
 941/*
 942 * Migrate pages from one node to a target node.
 943 * Returns error or the number of pages not migrated.
 944 */
 945static int migrate_to_node(struct mm_struct *mm, int source, int dest,
 946			   int flags)
 947{
 948	nodemask_t nmask;
 949	LIST_HEAD(pagelist);
 950	int err = 0;
 951
 952	nodes_clear(nmask);
 953	node_set(source, nmask);
 954
 955	/*
 956	 * This does not "check" the range but isolates all pages that
 957	 * need migration.  Between passing in the full user address
 958	 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
 959	 */
 960	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
 961	queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
 962			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
 963
 964	if (!list_empty(&pagelist)) {
 965		err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
 966					MIGRATE_SYNC, MR_SYSCALL);
 967		if (err)
 968			putback_movable_pages(&pagelist);
 969	}
 970
 971	return err;
 972}
 973
 974/*
 975 * Move pages between the two nodesets so as to preserve the physical
 976 * layout as much as possible.
 977 *
 978 * Returns the number of page that could not be moved.
 979 */
 980int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
 981		     const nodemask_t *to, int flags)
 982{
 983	int busy = 0;
 984	int err;
 985	nodemask_t tmp;
 986
 987	err = migrate_prep();
 988	if (err)
 989		return err;
 990
 991	down_read(&mm->mmap_sem);
 992
 993	/*
 994	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
 995	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
 996	 * bit in 'tmp', and return that <source, dest> pair for migration.
 997	 * The pair of nodemasks 'to' and 'from' define the map.
 998	 *
 999	 * If no pair of bits is found that way, fallback to picking some
1000	 * pair of 'source' and 'dest' bits that are not the same.  If the
1001	 * 'source' and 'dest' bits are the same, this represents a node
1002	 * that will be migrating to itself, so no pages need move.
1003	 *
1004	 * If no bits are left in 'tmp', or if all remaining bits left
1005	 * in 'tmp' correspond to the same bit in 'to', return false
1006	 * (nothing left to migrate).
1007	 *
1008	 * This lets us pick a pair of nodes to migrate between, such that
1009	 * if possible the dest node is not already occupied by some other
1010	 * source node, minimizing the risk of overloading the memory on a
1011	 * node that would happen if we migrated incoming memory to a node
1012	 * before migrating outgoing memory source that same node.
1013	 *
1014	 * A single scan of tmp is sufficient.  As we go, we remember the
1015	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1016	 * that not only moved, but what's better, moved to an empty slot
1017	 * (d is not set in tmp), then we break out then, with that pair.
1018	 * Otherwise when we finish scanning from_tmp, we at least have the
1019	 * most recent <s, d> pair that moved.  If we get all the way through
1020	 * the scan of tmp without finding any node that moved, much less
1021	 * moved to an empty node, then there is nothing left worth migrating.
1022	 */
1023
1024	tmp = *from;
1025	while (!nodes_empty(tmp)) {
1026		int s,d;
1027		int source = NUMA_NO_NODE;
1028		int dest = 0;
1029
1030		for_each_node_mask(s, tmp) {
1031
1032			/*
1033			 * do_migrate_pages() tries to maintain the relative
1034			 * node relationship of the pages established between
1035			 * threads and memory areas.
1036                         *
1037			 * However if the number of source nodes is not equal to
1038			 * the number of destination nodes we can not preserve
1039			 * this node relative relationship.  In that case, skip
1040			 * copying memory from a node that is in the destination
1041			 * mask.
1042			 *
1043			 * Example: [2,3,4] -> [3,4,5] moves everything.
1044			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1045			 */
1046
1047			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1048						(node_isset(s, *to)))
1049				continue;
1050
1051			d = node_remap(s, *from, *to);
1052			if (s == d)
1053				continue;
1054
1055			source = s;	/* Node moved. Memorize */
1056			dest = d;
1057
1058			/* dest not in remaining from nodes? */
1059			if (!node_isset(dest, tmp))
1060				break;
1061		}
1062		if (source == NUMA_NO_NODE)
1063			break;
1064
1065		node_clear(source, tmp);
1066		err = migrate_to_node(mm, source, dest, flags);
1067		if (err > 0)
1068			busy += err;
1069		if (err < 0)
1070			break;
1071	}
1072	up_read(&mm->mmap_sem);
1073	if (err < 0)
1074		return err;
1075	return busy;
1076
1077}
1078
1079/*
1080 * Allocate a new page for page migration based on vma policy.
1081 * Start by assuming the page is mapped by the same vma as contains @start.
1082 * Search forward from there, if not.  N.B., this assumes that the
1083 * list of pages handed to migrate_pages()--which is how we get here--
1084 * is in virtual address order.
1085 */
1086static struct page *new_page(struct page *page, unsigned long start)
1087{
1088	struct vm_area_struct *vma;
1089	unsigned long uninitialized_var(address);
1090
1091	vma = find_vma(current->mm, start);
1092	while (vma) {
1093		address = page_address_in_vma(page, vma);
1094		if (address != -EFAULT)
1095			break;
1096		vma = vma->vm_next;
1097	}
1098
1099	if (PageHuge(page)) {
1100		return alloc_huge_page_vma(page_hstate(compound_head(page)),
1101				vma, address);
1102	} else if (PageTransHuge(page)) {
1103		struct page *thp;
1104
1105		thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1106					 HPAGE_PMD_ORDER);
1107		if (!thp)
1108			return NULL;
1109		prep_transhuge_page(thp);
1110		return thp;
1111	}
1112	/*
1113	 * if !vma, alloc_page_vma() will use task or system default policy
1114	 */
1115	return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1116			vma, address);
1117}
1118#else
1119
1120static void migrate_page_add(struct page *page, struct list_head *pagelist,
1121				unsigned long flags)
1122{
 
1123}
1124
1125int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1126		     const nodemask_t *to, int flags)
1127{
1128	return -ENOSYS;
1129}
1130
1131static struct page *new_page(struct page *page, unsigned long start)
1132{
1133	return NULL;
1134}
1135#endif
1136
1137static long do_mbind(unsigned long start, unsigned long len,
1138		     unsigned short mode, unsigned short mode_flags,
1139		     nodemask_t *nmask, unsigned long flags)
1140{
1141	struct mm_struct *mm = current->mm;
1142	struct mempolicy *new;
1143	unsigned long end;
1144	int err;
 
1145	LIST_HEAD(pagelist);
1146
1147	if (flags & ~(unsigned long)MPOL_MF_VALID)
1148		return -EINVAL;
1149	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1150		return -EPERM;
1151
1152	if (start & ~PAGE_MASK)
1153		return -EINVAL;
1154
1155	if (mode == MPOL_DEFAULT)
1156		flags &= ~MPOL_MF_STRICT;
1157
1158	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1159	end = start + len;
1160
1161	if (end < start)
1162		return -EINVAL;
1163	if (end == start)
1164		return 0;
1165
1166	new = mpol_new(mode, mode_flags, nmask);
1167	if (IS_ERR(new))
1168		return PTR_ERR(new);
1169
1170	if (flags & MPOL_MF_LAZY)
1171		new->flags |= MPOL_F_MOF;
1172
1173	/*
1174	 * If we are using the default policy then operation
1175	 * on discontinuous address spaces is okay after all
1176	 */
1177	if (!new)
1178		flags |= MPOL_MF_DISCONTIG_OK;
1179
1180	pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1181		 start, start + len, mode, mode_flags,
1182		 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1183
1184	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1185
1186		err = migrate_prep();
1187		if (err)
1188			goto mpol_out;
1189	}
1190	{
1191		NODEMASK_SCRATCH(scratch);
1192		if (scratch) {
1193			down_write(&mm->mmap_sem);
1194			task_lock(current);
1195			err = mpol_set_nodemask(new, nmask, scratch);
1196			task_unlock(current);
1197			if (err)
1198				up_write(&mm->mmap_sem);
1199		} else
1200			err = -ENOMEM;
1201		NODEMASK_SCRATCH_FREE(scratch);
1202	}
1203	if (err)
1204		goto mpol_out;
1205
1206	err = queue_pages_range(mm, start, end, nmask,
1207			  flags | MPOL_MF_INVERT, &pagelist);
1208	if (!err)
1209		err = mbind_range(mm, start, end, new);
 
 
 
 
 
1210
1211	if (!err) {
1212		int nr_failed = 0;
1213
1214		if (!list_empty(&pagelist)) {
1215			WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1216			nr_failed = migrate_pages(&pagelist, new_page, NULL,
1217				start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1218			if (nr_failed)
1219				putback_movable_pages(&pagelist);
1220		}
1221
1222		if (nr_failed && (flags & MPOL_MF_STRICT))
1223			err = -EIO;
1224	} else
1225		putback_movable_pages(&pagelist);
 
 
 
1226
1227	up_write(&mm->mmap_sem);
1228 mpol_out:
1229	mpol_put(new);
1230	return err;
1231}
1232
1233/*
1234 * User space interface with variable sized bitmaps for nodelists.
1235 */
1236
1237/* Copy a node mask from user space. */
1238static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1239		     unsigned long maxnode)
1240{
1241	unsigned long k;
1242	unsigned long t;
1243	unsigned long nlongs;
1244	unsigned long endmask;
1245
1246	--maxnode;
1247	nodes_clear(*nodes);
1248	if (maxnode == 0 || !nmask)
1249		return 0;
1250	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1251		return -EINVAL;
1252
1253	nlongs = BITS_TO_LONGS(maxnode);
1254	if ((maxnode % BITS_PER_LONG) == 0)
1255		endmask = ~0UL;
1256	else
1257		endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1258
1259	/*
1260	 * When the user specified more nodes than supported just check
1261	 * if the non supported part is all zero.
1262	 *
1263	 * If maxnode have more longs than MAX_NUMNODES, check
1264	 * the bits in that area first. And then go through to
1265	 * check the rest bits which equal or bigger than MAX_NUMNODES.
1266	 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1267	 */
1268	if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1269		for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1270			if (get_user(t, nmask + k))
1271				return -EFAULT;
1272			if (k == nlongs - 1) {
1273				if (t & endmask)
1274					return -EINVAL;
1275			} else if (t)
1276				return -EINVAL;
1277		}
1278		nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1279		endmask = ~0UL;
1280	}
1281
1282	if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1283		unsigned long valid_mask = endmask;
1284
1285		valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1286		if (get_user(t, nmask + nlongs - 1))
1287			return -EFAULT;
1288		if (t & valid_mask)
1289			return -EINVAL;
1290	}
1291
1292	if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1293		return -EFAULT;
1294	nodes_addr(*nodes)[nlongs-1] &= endmask;
1295	return 0;
1296}
1297
1298/* Copy a kernel node mask to user space */
1299static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1300			      nodemask_t *nodes)
1301{
1302	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1303	const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1304
1305	if (copy > nbytes) {
1306		if (copy > PAGE_SIZE)
1307			return -EINVAL;
1308		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1309			return -EFAULT;
1310		copy = nbytes;
1311	}
1312	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1313}
1314
1315static long kernel_mbind(unsigned long start, unsigned long len,
1316			 unsigned long mode, const unsigned long __user *nmask,
1317			 unsigned long maxnode, unsigned int flags)
1318{
1319	nodemask_t nodes;
1320	int err;
1321	unsigned short mode_flags;
1322
 
1323	mode_flags = mode & MPOL_MODE_FLAGS;
1324	mode &= ~MPOL_MODE_FLAGS;
1325	if (mode >= MPOL_MAX)
1326		return -EINVAL;
1327	if ((mode_flags & MPOL_F_STATIC_NODES) &&
1328	    (mode_flags & MPOL_F_RELATIVE_NODES))
1329		return -EINVAL;
1330	err = get_nodes(&nodes, nmask, maxnode);
1331	if (err)
1332		return err;
1333	return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1334}
1335
1336SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1337		unsigned long, mode, const unsigned long __user *, nmask,
1338		unsigned long, maxnode, unsigned int, flags)
1339{
1340	return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1341}
1342
1343/* Set the process memory policy */
1344static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1345				 unsigned long maxnode)
1346{
1347	int err;
1348	nodemask_t nodes;
1349	unsigned short flags;
1350
1351	flags = mode & MPOL_MODE_FLAGS;
1352	mode &= ~MPOL_MODE_FLAGS;
1353	if ((unsigned int)mode >= MPOL_MAX)
1354		return -EINVAL;
1355	if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1356		return -EINVAL;
1357	err = get_nodes(&nodes, nmask, maxnode);
1358	if (err)
1359		return err;
1360	return do_set_mempolicy(mode, flags, &nodes);
1361}
1362
1363SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1364		unsigned long, maxnode)
1365{
1366	return kernel_set_mempolicy(mode, nmask, maxnode);
1367}
1368
1369static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1370				const unsigned long __user *old_nodes,
1371				const unsigned long __user *new_nodes)
1372{
1373	struct mm_struct *mm = NULL;
1374	struct task_struct *task;
1375	nodemask_t task_nodes;
1376	int err;
1377	nodemask_t *old;
1378	nodemask_t *new;
1379	NODEMASK_SCRATCH(scratch);
1380
1381	if (!scratch)
1382		return -ENOMEM;
1383
1384	old = &scratch->mask1;
1385	new = &scratch->mask2;
1386
1387	err = get_nodes(old, old_nodes, maxnode);
1388	if (err)
1389		goto out;
1390
1391	err = get_nodes(new, new_nodes, maxnode);
1392	if (err)
1393		goto out;
1394
1395	/* Find the mm_struct */
1396	rcu_read_lock();
1397	task = pid ? find_task_by_vpid(pid) : current;
1398	if (!task) {
1399		rcu_read_unlock();
1400		err = -ESRCH;
1401		goto out;
1402	}
1403	get_task_struct(task);
1404
1405	err = -EINVAL;
1406
1407	/*
1408	 * Check if this process has the right to modify the specified process.
1409	 * Use the regular "ptrace_may_access()" checks.
1410	 */
1411	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1412		rcu_read_unlock();
1413		err = -EPERM;
1414		goto out_put;
1415	}
1416	rcu_read_unlock();
1417
1418	task_nodes = cpuset_mems_allowed(task);
1419	/* Is the user allowed to access the target nodes? */
1420	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1421		err = -EPERM;
1422		goto out_put;
1423	}
1424
1425	task_nodes = cpuset_mems_allowed(current);
1426	nodes_and(*new, *new, task_nodes);
1427	if (nodes_empty(*new))
1428		goto out_put;
1429
1430	nodes_and(*new, *new, node_states[N_MEMORY]);
1431	if (nodes_empty(*new))
1432		goto out_put;
1433
1434	err = security_task_movememory(task);
1435	if (err)
1436		goto out_put;
1437
1438	mm = get_task_mm(task);
1439	put_task_struct(task);
1440
1441	if (!mm) {
1442		err = -EINVAL;
1443		goto out;
1444	}
1445
1446	err = do_migrate_pages(mm, old, new,
1447		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1448
1449	mmput(mm);
1450out:
1451	NODEMASK_SCRATCH_FREE(scratch);
1452
1453	return err;
1454
1455out_put:
1456	put_task_struct(task);
1457	goto out;
1458
1459}
1460
1461SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1462		const unsigned long __user *, old_nodes,
1463		const unsigned long __user *, new_nodes)
1464{
1465	return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1466}
1467
1468
1469/* Retrieve NUMA policy */
1470static int kernel_get_mempolicy(int __user *policy,
1471				unsigned long __user *nmask,
1472				unsigned long maxnode,
1473				unsigned long addr,
1474				unsigned long flags)
1475{
1476	int err;
1477	int uninitialized_var(pval);
1478	nodemask_t nodes;
1479
1480	if (nmask != NULL && maxnode < MAX_NUMNODES)
 
 
1481		return -EINVAL;
1482
1483	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1484
1485	if (err)
1486		return err;
1487
1488	if (policy && put_user(pval, policy))
1489		return -EFAULT;
1490
1491	if (nmask)
1492		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1493
1494	return err;
1495}
1496
1497SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1498		unsigned long __user *, nmask, unsigned long, maxnode,
1499		unsigned long, addr, unsigned long, flags)
1500{
1501	return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1502}
1503
1504#ifdef CONFIG_COMPAT
1505
1506COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1507		       compat_ulong_t __user *, nmask,
1508		       compat_ulong_t, maxnode,
1509		       compat_ulong_t, addr, compat_ulong_t, flags)
1510{
1511	long err;
1512	unsigned long __user *nm = NULL;
1513	unsigned long nr_bits, alloc_size;
1514	DECLARE_BITMAP(bm, MAX_NUMNODES);
1515
1516	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1517	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1518
1519	if (nmask)
1520		nm = compat_alloc_user_space(alloc_size);
1521
1522	err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1523
1524	if (!err && nmask) {
1525		unsigned long copy_size;
1526		copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1527		err = copy_from_user(bm, nm, copy_size);
1528		/* ensure entire bitmap is zeroed */
1529		err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1530		err |= compat_put_bitmap(nmask, bm, nr_bits);
1531	}
1532
1533	return err;
1534}
1535
1536COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1537		       compat_ulong_t, maxnode)
1538{
1539	unsigned long __user *nm = NULL;
1540	unsigned long nr_bits, alloc_size;
1541	DECLARE_BITMAP(bm, MAX_NUMNODES);
1542
1543	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1544	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1545
1546	if (nmask) {
1547		if (compat_get_bitmap(bm, nmask, nr_bits))
1548			return -EFAULT;
1549		nm = compat_alloc_user_space(alloc_size);
1550		if (copy_to_user(nm, bm, alloc_size))
1551			return -EFAULT;
1552	}
1553
1554	return kernel_set_mempolicy(mode, nm, nr_bits+1);
1555}
1556
1557COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1558		       compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1559		       compat_ulong_t, maxnode, compat_ulong_t, flags)
1560{
1561	unsigned long __user *nm = NULL;
1562	unsigned long nr_bits, alloc_size;
1563	nodemask_t bm;
1564
1565	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1566	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1567
1568	if (nmask) {
1569		if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1570			return -EFAULT;
1571		nm = compat_alloc_user_space(alloc_size);
1572		if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1573			return -EFAULT;
1574	}
1575
1576	return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1577}
1578
1579COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1580		       compat_ulong_t, maxnode,
1581		       const compat_ulong_t __user *, old_nodes,
1582		       const compat_ulong_t __user *, new_nodes)
1583{
1584	unsigned long __user *old = NULL;
1585	unsigned long __user *new = NULL;
1586	nodemask_t tmp_mask;
1587	unsigned long nr_bits;
1588	unsigned long size;
1589
1590	nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1591	size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1592	if (old_nodes) {
1593		if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1594			return -EFAULT;
1595		old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1596		if (new_nodes)
1597			new = old + size / sizeof(unsigned long);
1598		if (copy_to_user(old, nodes_addr(tmp_mask), size))
1599			return -EFAULT;
1600	}
1601	if (new_nodes) {
1602		if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1603			return -EFAULT;
1604		if (new == NULL)
1605			new = compat_alloc_user_space(size);
1606		if (copy_to_user(new, nodes_addr(tmp_mask), size))
1607			return -EFAULT;
1608	}
1609	return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1610}
1611
1612#endif /* CONFIG_COMPAT */
1613
1614struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1615						unsigned long addr)
1616{
1617	struct mempolicy *pol = NULL;
1618
1619	if (vma) {
1620		if (vma->vm_ops && vma->vm_ops->get_policy) {
1621			pol = vma->vm_ops->get_policy(vma, addr);
1622		} else if (vma->vm_policy) {
1623			pol = vma->vm_policy;
1624
1625			/*
1626			 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1627			 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1628			 * count on these policies which will be dropped by
1629			 * mpol_cond_put() later
1630			 */
1631			if (mpol_needs_cond_ref(pol))
1632				mpol_get(pol);
1633		}
1634	}
1635
1636	return pol;
1637}
1638
1639/*
1640 * get_vma_policy(@vma, @addr)
1641 * @vma: virtual memory area whose policy is sought
1642 * @addr: address in @vma for shared policy lookup
1643 *
1644 * Returns effective policy for a VMA at specified address.
1645 * Falls back to current->mempolicy or system default policy, as necessary.
1646 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1647 * count--added by the get_policy() vm_op, as appropriate--to protect against
1648 * freeing by another task.  It is the caller's responsibility to free the
1649 * extra reference for shared policies.
1650 */
1651static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1652						unsigned long addr)
1653{
1654	struct mempolicy *pol = __get_vma_policy(vma, addr);
1655
1656	if (!pol)
1657		pol = get_task_policy(current);
1658
1659	return pol;
1660}
1661
1662bool vma_policy_mof(struct vm_area_struct *vma)
1663{
1664	struct mempolicy *pol;
1665
1666	if (vma->vm_ops && vma->vm_ops->get_policy) {
1667		bool ret = false;
1668
1669		pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1670		if (pol && (pol->flags & MPOL_F_MOF))
1671			ret = true;
1672		mpol_cond_put(pol);
1673
1674		return ret;
1675	}
1676
1677	pol = vma->vm_policy;
1678	if (!pol)
1679		pol = get_task_policy(current);
1680
1681	return pol->flags & MPOL_F_MOF;
1682}
1683
1684static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1685{
1686	enum zone_type dynamic_policy_zone = policy_zone;
1687
1688	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1689
1690	/*
1691	 * if policy->v.nodes has movable memory only,
1692	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1693	 *
1694	 * policy->v.nodes is intersect with node_states[N_MEMORY].
1695	 * so if the following test faile, it implies
1696	 * policy->v.nodes has movable memory only.
1697	 */
1698	if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1699		dynamic_policy_zone = ZONE_MOVABLE;
1700
1701	return zone >= dynamic_policy_zone;
1702}
1703
1704/*
1705 * Return a nodemask representing a mempolicy for filtering nodes for
1706 * page allocation
1707 */
1708static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1709{
1710	/* Lower zones don't get a nodemask applied for MPOL_BIND */
1711	if (unlikely(policy->mode == MPOL_BIND) &&
1712			apply_policy_zone(policy, gfp_zone(gfp)) &&
1713			cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1714		return &policy->v.nodes;
1715
1716	return NULL;
1717}
1718
1719/* Return the node id preferred by the given mempolicy, or the given id */
1720static int policy_node(gfp_t gfp, struct mempolicy *policy,
1721								int nd)
1722{
1723	if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1724		nd = policy->v.preferred_node;
1725	else {
1726		/*
1727		 * __GFP_THISNODE shouldn't even be used with the bind policy
1728		 * because we might easily break the expectation to stay on the
1729		 * requested node and not break the policy.
1730		 */
1731		WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1732	}
1733
1734	return nd;
1735}
1736
1737/* Do dynamic interleaving for a process */
1738static unsigned interleave_nodes(struct mempolicy *policy)
1739{
1740	unsigned next;
1741	struct task_struct *me = current;
1742
1743	next = next_node_in(me->il_prev, policy->v.nodes);
1744	if (next < MAX_NUMNODES)
1745		me->il_prev = next;
1746	return next;
1747}
1748
1749/*
1750 * Depending on the memory policy provide a node from which to allocate the
1751 * next slab entry.
1752 */
1753unsigned int mempolicy_slab_node(void)
1754{
1755	struct mempolicy *policy;
1756	int node = numa_mem_id();
1757
1758	if (in_interrupt())
1759		return node;
1760
1761	policy = current->mempolicy;
1762	if (!policy || policy->flags & MPOL_F_LOCAL)
1763		return node;
1764
1765	switch (policy->mode) {
1766	case MPOL_PREFERRED:
1767		/*
1768		 * handled MPOL_F_LOCAL above
1769		 */
1770		return policy->v.preferred_node;
1771
1772	case MPOL_INTERLEAVE:
1773		return interleave_nodes(policy);
1774
1775	case MPOL_BIND: {
1776		struct zoneref *z;
1777
1778		/*
1779		 * Follow bind policy behavior and start allocation at the
1780		 * first node.
1781		 */
1782		struct zonelist *zonelist;
1783		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1784		zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1785		z = first_zones_zonelist(zonelist, highest_zoneidx,
1786							&policy->v.nodes);
1787		return z->zone ? z->zone->node : node;
1788	}
1789
1790	default:
1791		BUG();
1792	}
1793}
1794
1795/*
1796 * Do static interleaving for a VMA with known offset @n.  Returns the n'th
1797 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1798 * number of present nodes.
1799 */
1800static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1801{
1802	unsigned nnodes = nodes_weight(pol->v.nodes);
1803	unsigned target;
1804	int i;
1805	int nid;
1806
1807	if (!nnodes)
1808		return numa_node_id();
1809	target = (unsigned int)n % nnodes;
1810	nid = first_node(pol->v.nodes);
1811	for (i = 0; i < target; i++)
1812		nid = next_node(nid, pol->v.nodes);
1813	return nid;
1814}
1815
1816/* Determine a node number for interleave */
1817static inline unsigned interleave_nid(struct mempolicy *pol,
1818		 struct vm_area_struct *vma, unsigned long addr, int shift)
1819{
1820	if (vma) {
1821		unsigned long off;
1822
1823		/*
1824		 * for small pages, there is no difference between
1825		 * shift and PAGE_SHIFT, so the bit-shift is safe.
1826		 * for huge pages, since vm_pgoff is in units of small
1827		 * pages, we need to shift off the always 0 bits to get
1828		 * a useful offset.
1829		 */
1830		BUG_ON(shift < PAGE_SHIFT);
1831		off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1832		off += (addr - vma->vm_start) >> shift;
1833		return offset_il_node(pol, off);
1834	} else
1835		return interleave_nodes(pol);
1836}
1837
1838#ifdef CONFIG_HUGETLBFS
1839/*
1840 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1841 * @vma: virtual memory area whose policy is sought
1842 * @addr: address in @vma for shared policy lookup and interleave policy
1843 * @gfp_flags: for requested zone
1844 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1845 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1846 *
1847 * Returns a nid suitable for a huge page allocation and a pointer
1848 * to the struct mempolicy for conditional unref after allocation.
1849 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1850 * @nodemask for filtering the zonelist.
1851 *
1852 * Must be protected by read_mems_allowed_begin()
1853 */
1854int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1855				struct mempolicy **mpol, nodemask_t **nodemask)
1856{
1857	int nid;
1858
1859	*mpol = get_vma_policy(vma, addr);
1860	*nodemask = NULL;	/* assume !MPOL_BIND */
1861
1862	if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1863		nid = interleave_nid(*mpol, vma, addr,
1864					huge_page_shift(hstate_vma(vma)));
1865	} else {
1866		nid = policy_node(gfp_flags, *mpol, numa_node_id());
1867		if ((*mpol)->mode == MPOL_BIND)
1868			*nodemask = &(*mpol)->v.nodes;
1869	}
1870	return nid;
1871}
1872
1873/*
1874 * init_nodemask_of_mempolicy
1875 *
1876 * If the current task's mempolicy is "default" [NULL], return 'false'
1877 * to indicate default policy.  Otherwise, extract the policy nodemask
1878 * for 'bind' or 'interleave' policy into the argument nodemask, or
1879 * initialize the argument nodemask to contain the single node for
1880 * 'preferred' or 'local' policy and return 'true' to indicate presence
1881 * of non-default mempolicy.
1882 *
1883 * We don't bother with reference counting the mempolicy [mpol_get/put]
1884 * because the current task is examining it's own mempolicy and a task's
1885 * mempolicy is only ever changed by the task itself.
1886 *
1887 * N.B., it is the caller's responsibility to free a returned nodemask.
1888 */
1889bool init_nodemask_of_mempolicy(nodemask_t *mask)
1890{
1891	struct mempolicy *mempolicy;
1892	int nid;
1893
1894	if (!(mask && current->mempolicy))
1895		return false;
1896
1897	task_lock(current);
1898	mempolicy = current->mempolicy;
1899	switch (mempolicy->mode) {
1900	case MPOL_PREFERRED:
1901		if (mempolicy->flags & MPOL_F_LOCAL)
1902			nid = numa_node_id();
1903		else
1904			nid = mempolicy->v.preferred_node;
1905		init_nodemask_of_node(mask, nid);
1906		break;
1907
1908	case MPOL_BIND:
1909		/* Fall through */
1910	case MPOL_INTERLEAVE:
1911		*mask =  mempolicy->v.nodes;
1912		break;
1913
1914	default:
1915		BUG();
1916	}
1917	task_unlock(current);
1918
1919	return true;
1920}
1921#endif
1922
1923/*
1924 * mempolicy_nodemask_intersects
1925 *
1926 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1927 * policy.  Otherwise, check for intersection between mask and the policy
1928 * nodemask for 'bind' or 'interleave' policy.  For 'perferred' or 'local'
1929 * policy, always return true since it may allocate elsewhere on fallback.
1930 *
1931 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1932 */
1933bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1934					const nodemask_t *mask)
1935{
1936	struct mempolicy *mempolicy;
1937	bool ret = true;
1938
1939	if (!mask)
1940		return ret;
1941	task_lock(tsk);
1942	mempolicy = tsk->mempolicy;
1943	if (!mempolicy)
1944		goto out;
1945
1946	switch (mempolicy->mode) {
1947	case MPOL_PREFERRED:
1948		/*
1949		 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1950		 * allocate from, they may fallback to other nodes when oom.
1951		 * Thus, it's possible for tsk to have allocated memory from
1952		 * nodes in mask.
1953		 */
1954		break;
1955	case MPOL_BIND:
1956	case MPOL_INTERLEAVE:
1957		ret = nodes_intersects(mempolicy->v.nodes, *mask);
1958		break;
1959	default:
1960		BUG();
1961	}
1962out:
1963	task_unlock(tsk);
1964	return ret;
1965}
1966
1967/* Allocate a page in interleaved policy.
1968   Own path because it needs to do special accounting. */
1969static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1970					unsigned nid)
1971{
1972	struct page *page;
1973
1974	page = __alloc_pages(gfp, order, nid);
1975	/* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1976	if (!static_branch_likely(&vm_numa_stat_key))
1977		return page;
1978	if (page && page_to_nid(page) == nid) {
1979		preempt_disable();
1980		__inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1981		preempt_enable();
1982	}
1983	return page;
1984}
1985
1986/**
1987 * 	alloc_pages_vma	- Allocate a page for a VMA.
1988 *
1989 * 	@gfp:
1990 *      %GFP_USER    user allocation.
1991 *      %GFP_KERNEL  kernel allocations,
1992 *      %GFP_HIGHMEM highmem/user allocations,
1993 *      %GFP_FS      allocation should not call back into a file system.
1994 *      %GFP_ATOMIC  don't sleep.
1995 *
1996 *	@order:Order of the GFP allocation.
1997 * 	@vma:  Pointer to VMA or NULL if not available.
1998 *	@addr: Virtual Address of the allocation. Must be inside the VMA.
1999 *	@node: Which node to prefer for allocation (modulo policy).
2000 *	@hugepage: for hugepages try only the preferred node if possible
2001 *
2002 * 	This function allocates a page from the kernel page pool and applies
2003 *	a NUMA policy associated with the VMA or the current process.
2004 *	When VMA is not NULL caller must hold down_read on the mmap_sem of the
2005 *	mm_struct of the VMA to prevent it from going away. Should be used for
2006 *	all allocations for pages that will be mapped into user space. Returns
2007 *	NULL when no page can be allocated.
2008 */
2009struct page *
2010alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2011		unsigned long addr, int node, bool hugepage)
2012{
2013	struct mempolicy *pol;
2014	struct page *page;
2015	int preferred_nid;
2016	nodemask_t *nmask;
2017
2018	pol = get_vma_policy(vma, addr);
2019
2020	if (pol->mode == MPOL_INTERLEAVE) {
2021		unsigned nid;
2022
2023		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2024		mpol_cond_put(pol);
2025		page = alloc_page_interleave(gfp, order, nid);
2026		goto out;
2027	}
2028
2029	if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2030		int hpage_node = node;
2031
2032		/*
2033		 * For hugepage allocation and non-interleave policy which
2034		 * allows the current node (or other explicitly preferred
2035		 * node) we only try to allocate from the current/preferred
2036		 * node and don't fall back to other nodes, as the cost of
2037		 * remote accesses would likely offset THP benefits.
2038		 *
2039		 * If the policy is interleave, or does not allow the current
2040		 * node in its nodemask, we allocate the standard way.
2041		 */
2042		if (pol->mode == MPOL_PREFERRED &&
2043						!(pol->flags & MPOL_F_LOCAL))
2044			hpage_node = pol->v.preferred_node;
2045
2046		nmask = policy_nodemask(gfp, pol);
2047		if (!nmask || node_isset(hpage_node, *nmask)) {
2048			mpol_cond_put(pol);
2049			page = __alloc_pages_node(hpage_node,
2050						gfp | __GFP_THISNODE, order);
 
 
 
 
 
 
 
 
 
 
 
2051			goto out;
2052		}
2053	}
2054
2055	nmask = policy_nodemask(gfp, pol);
2056	preferred_nid = policy_node(gfp, pol, node);
2057	page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2058	mpol_cond_put(pol);
2059out:
2060	return page;
2061}
 
2062
2063/**
2064 * 	alloc_pages_current - Allocate pages.
2065 *
2066 *	@gfp:
2067 *		%GFP_USER   user allocation,
2068 *      	%GFP_KERNEL kernel allocation,
2069 *      	%GFP_HIGHMEM highmem allocation,
2070 *      	%GFP_FS     don't call back into a file system.
2071 *      	%GFP_ATOMIC don't sleep.
2072 *	@order: Power of two of allocation size in pages. 0 is a single page.
2073 *
2074 *	Allocate a page from the kernel page pool.  When not in
2075 *	interrupt context and apply the current process NUMA policy.
2076 *	Returns NULL when no page can be allocated.
2077 */
2078struct page *alloc_pages_current(gfp_t gfp, unsigned order)
2079{
2080	struct mempolicy *pol = &default_policy;
2081	struct page *page;
2082
2083	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2084		pol = get_task_policy(current);
2085
2086	/*
2087	 * No reference counting needed for current->mempolicy
2088	 * nor system default_policy
2089	 */
2090	if (pol->mode == MPOL_INTERLEAVE)
2091		page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2092	else
2093		page = __alloc_pages_nodemask(gfp, order,
2094				policy_node(gfp, pol, numa_node_id()),
2095				policy_nodemask(gfp, pol));
2096
2097	return page;
2098}
2099EXPORT_SYMBOL(alloc_pages_current);
2100
2101int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2102{
2103	struct mempolicy *pol = mpol_dup(vma_policy(src));
2104
2105	if (IS_ERR(pol))
2106		return PTR_ERR(pol);
2107	dst->vm_policy = pol;
2108	return 0;
2109}
2110
2111/*
2112 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2113 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2114 * with the mems_allowed returned by cpuset_mems_allowed().  This
2115 * keeps mempolicies cpuset relative after its cpuset moves.  See
2116 * further kernel/cpuset.c update_nodemask().
2117 *
2118 * current's mempolicy may be rebinded by the other task(the task that changes
2119 * cpuset's mems), so we needn't do rebind work for current task.
2120 */
2121
2122/* Slow path of a mempolicy duplicate */
2123struct mempolicy *__mpol_dup(struct mempolicy *old)
2124{
2125	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2126
2127	if (!new)
2128		return ERR_PTR(-ENOMEM);
2129
2130	/* task's mempolicy is protected by alloc_lock */
2131	if (old == current->mempolicy) {
2132		task_lock(current);
2133		*new = *old;
2134		task_unlock(current);
2135	} else
2136		*new = *old;
2137
2138	if (current_cpuset_is_being_rebound()) {
2139		nodemask_t mems = cpuset_mems_allowed(current);
2140		mpol_rebind_policy(new, &mems);
2141	}
2142	atomic_set(&new->refcnt, 1);
2143	return new;
2144}
2145
2146/* Slow path of a mempolicy comparison */
2147bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2148{
2149	if (!a || !b)
2150		return false;
2151	if (a->mode != b->mode)
2152		return false;
2153	if (a->flags != b->flags)
2154		return false;
2155	if (mpol_store_user_nodemask(a))
2156		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2157			return false;
2158
2159	switch (a->mode) {
2160	case MPOL_BIND:
2161		/* Fall through */
2162	case MPOL_INTERLEAVE:
2163		return !!nodes_equal(a->v.nodes, b->v.nodes);
2164	case MPOL_PREFERRED:
2165		/* a's ->flags is the same as b's */
2166		if (a->flags & MPOL_F_LOCAL)
2167			return true;
2168		return a->v.preferred_node == b->v.preferred_node;
2169	default:
2170		BUG();
2171		return false;
2172	}
2173}
2174
2175/*
2176 * Shared memory backing store policy support.
2177 *
2178 * Remember policies even when nobody has shared memory mapped.
2179 * The policies are kept in Red-Black tree linked from the inode.
2180 * They are protected by the sp->lock rwlock, which should be held
2181 * for any accesses to the tree.
2182 */
2183
2184/*
2185 * lookup first element intersecting start-end.  Caller holds sp->lock for
2186 * reading or for writing
2187 */
2188static struct sp_node *
2189sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2190{
2191	struct rb_node *n = sp->root.rb_node;
2192
2193	while (n) {
2194		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2195
2196		if (start >= p->end)
2197			n = n->rb_right;
2198		else if (end <= p->start)
2199			n = n->rb_left;
2200		else
2201			break;
2202	}
2203	if (!n)
2204		return NULL;
2205	for (;;) {
2206		struct sp_node *w = NULL;
2207		struct rb_node *prev = rb_prev(n);
2208		if (!prev)
2209			break;
2210		w = rb_entry(prev, struct sp_node, nd);
2211		if (w->end <= start)
2212			break;
2213		n = prev;
2214	}
2215	return rb_entry(n, struct sp_node, nd);
2216}
2217
2218/*
2219 * Insert a new shared policy into the list.  Caller holds sp->lock for
2220 * writing.
2221 */
2222static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2223{
2224	struct rb_node **p = &sp->root.rb_node;
2225	struct rb_node *parent = NULL;
2226	struct sp_node *nd;
2227
2228	while (*p) {
2229		parent = *p;
2230		nd = rb_entry(parent, struct sp_node, nd);
2231		if (new->start < nd->start)
2232			p = &(*p)->rb_left;
2233		else if (new->end > nd->end)
2234			p = &(*p)->rb_right;
2235		else
2236			BUG();
2237	}
2238	rb_link_node(&new->nd, parent, p);
2239	rb_insert_color(&new->nd, &sp->root);
2240	pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2241		 new->policy ? new->policy->mode : 0);
2242}
2243
2244/* Find shared policy intersecting idx */
2245struct mempolicy *
2246mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2247{
2248	struct mempolicy *pol = NULL;
2249	struct sp_node *sn;
2250
2251	if (!sp->root.rb_node)
2252		return NULL;
2253	read_lock(&sp->lock);
2254	sn = sp_lookup(sp, idx, idx+1);
2255	if (sn) {
2256		mpol_get(sn->policy);
2257		pol = sn->policy;
2258	}
2259	read_unlock(&sp->lock);
2260	return pol;
2261}
2262
2263static void sp_free(struct sp_node *n)
2264{
2265	mpol_put(n->policy);
2266	kmem_cache_free(sn_cache, n);
2267}
2268
2269/**
2270 * mpol_misplaced - check whether current page node is valid in policy
2271 *
2272 * @page: page to be checked
2273 * @vma: vm area where page mapped
2274 * @addr: virtual address where page mapped
2275 *
2276 * Lookup current policy node id for vma,addr and "compare to" page's
2277 * node id.
2278 *
2279 * Returns:
2280 *	-1	- not misplaced, page is in the right node
2281 *	node	- node id where the page should be
2282 *
2283 * Policy determination "mimics" alloc_page_vma().
2284 * Called from fault path where we know the vma and faulting address.
2285 */
2286int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2287{
2288	struct mempolicy *pol;
2289	struct zoneref *z;
2290	int curnid = page_to_nid(page);
2291	unsigned long pgoff;
2292	int thiscpu = raw_smp_processor_id();
2293	int thisnid = cpu_to_node(thiscpu);
2294	int polnid = -1;
2295	int ret = -1;
2296
2297	pol = get_vma_policy(vma, addr);
2298	if (!(pol->flags & MPOL_F_MOF))
2299		goto out;
2300
2301	switch (pol->mode) {
2302	case MPOL_INTERLEAVE:
2303		pgoff = vma->vm_pgoff;
2304		pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2305		polnid = offset_il_node(pol, pgoff);
2306		break;
2307
2308	case MPOL_PREFERRED:
2309		if (pol->flags & MPOL_F_LOCAL)
2310			polnid = numa_node_id();
2311		else
2312			polnid = pol->v.preferred_node;
2313		break;
2314
2315	case MPOL_BIND:
2316
2317		/*
2318		 * allows binding to multiple nodes.
2319		 * use current page if in policy nodemask,
2320		 * else select nearest allowed node, if any.
2321		 * If no allowed nodes, use current [!misplaced].
2322		 */
2323		if (node_isset(curnid, pol->v.nodes))
2324			goto out;
2325		z = first_zones_zonelist(
2326				node_zonelist(numa_node_id(), GFP_HIGHUSER),
2327				gfp_zone(GFP_HIGHUSER),
2328				&pol->v.nodes);
2329		polnid = z->zone->node;
2330		break;
2331
2332	default:
2333		BUG();
2334	}
2335
2336	/* Migrate the page towards the node whose CPU is referencing it */
2337	if (pol->flags & MPOL_F_MORON) {
2338		polnid = thisnid;
2339
2340		if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2341			goto out;
2342	}
2343
2344	if (curnid != polnid)
2345		ret = polnid;
2346out:
2347	mpol_cond_put(pol);
2348
2349	return ret;
2350}
2351
2352/*
2353 * Drop the (possibly final) reference to task->mempolicy.  It needs to be
2354 * dropped after task->mempolicy is set to NULL so that any allocation done as
2355 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2356 * policy.
2357 */
2358void mpol_put_task_policy(struct task_struct *task)
2359{
2360	struct mempolicy *pol;
2361
2362	task_lock(task);
2363	pol = task->mempolicy;
2364	task->mempolicy = NULL;
2365	task_unlock(task);
2366	mpol_put(pol);
2367}
2368
2369static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2370{
2371	pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2372	rb_erase(&n->nd, &sp->root);
2373	sp_free(n);
2374}
2375
2376static void sp_node_init(struct sp_node *node, unsigned long start,
2377			unsigned long end, struct mempolicy *pol)
2378{
2379	node->start = start;
2380	node->end = end;
2381	node->policy = pol;
2382}
2383
2384static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2385				struct mempolicy *pol)
2386{
2387	struct sp_node *n;
2388	struct mempolicy *newpol;
2389
2390	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2391	if (!n)
2392		return NULL;
2393
2394	newpol = mpol_dup(pol);
2395	if (IS_ERR(newpol)) {
2396		kmem_cache_free(sn_cache, n);
2397		return NULL;
2398	}
2399	newpol->flags |= MPOL_F_SHARED;
2400	sp_node_init(n, start, end, newpol);
2401
2402	return n;
2403}
2404
2405/* Replace a policy range. */
2406static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2407				 unsigned long end, struct sp_node *new)
2408{
2409	struct sp_node *n;
2410	struct sp_node *n_new = NULL;
2411	struct mempolicy *mpol_new = NULL;
2412	int ret = 0;
2413
2414restart:
2415	write_lock(&sp->lock);
2416	n = sp_lookup(sp, start, end);
2417	/* Take care of old policies in the same range. */
2418	while (n && n->start < end) {
2419		struct rb_node *next = rb_next(&n->nd);
2420		if (n->start >= start) {
2421			if (n->end <= end)
2422				sp_delete(sp, n);
2423			else
2424				n->start = end;
2425		} else {
2426			/* Old policy spanning whole new range. */
2427			if (n->end > end) {
2428				if (!n_new)
2429					goto alloc_new;
2430
2431				*mpol_new = *n->policy;
2432				atomic_set(&mpol_new->refcnt, 1);
2433				sp_node_init(n_new, end, n->end, mpol_new);
2434				n->end = start;
2435				sp_insert(sp, n_new);
2436				n_new = NULL;
2437				mpol_new = NULL;
2438				break;
2439			} else
2440				n->end = start;
2441		}
2442		if (!next)
2443			break;
2444		n = rb_entry(next, struct sp_node, nd);
2445	}
2446	if (new)
2447		sp_insert(sp, new);
2448	write_unlock(&sp->lock);
2449	ret = 0;
2450
2451err_out:
2452	if (mpol_new)
2453		mpol_put(mpol_new);
2454	if (n_new)
2455		kmem_cache_free(sn_cache, n_new);
2456
2457	return ret;
2458
2459alloc_new:
2460	write_unlock(&sp->lock);
2461	ret = -ENOMEM;
2462	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2463	if (!n_new)
2464		goto err_out;
2465	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2466	if (!mpol_new)
2467		goto err_out;
2468	goto restart;
2469}
2470
2471/**
2472 * mpol_shared_policy_init - initialize shared policy for inode
2473 * @sp: pointer to inode shared policy
2474 * @mpol:  struct mempolicy to install
2475 *
2476 * Install non-NULL @mpol in inode's shared policy rb-tree.
2477 * On entry, the current task has a reference on a non-NULL @mpol.
2478 * This must be released on exit.
2479 * This is called at get_inode() calls and we can use GFP_KERNEL.
2480 */
2481void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2482{
2483	int ret;
2484
2485	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2486	rwlock_init(&sp->lock);
2487
2488	if (mpol) {
2489		struct vm_area_struct pvma;
2490		struct mempolicy *new;
2491		NODEMASK_SCRATCH(scratch);
2492
2493		if (!scratch)
2494			goto put_mpol;
2495		/* contextualize the tmpfs mount point mempolicy */
2496		new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2497		if (IS_ERR(new))
2498			goto free_scratch; /* no valid nodemask intersection */
2499
2500		task_lock(current);
2501		ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2502		task_unlock(current);
2503		if (ret)
2504			goto put_new;
2505
2506		/* Create pseudo-vma that contains just the policy */
2507		memset(&pvma, 0, sizeof(struct vm_area_struct));
2508		pvma.vm_end = TASK_SIZE;	/* policy covers entire file */
2509		mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2510
2511put_new:
2512		mpol_put(new);			/* drop initial ref */
2513free_scratch:
2514		NODEMASK_SCRATCH_FREE(scratch);
2515put_mpol:
2516		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
2517	}
2518}
2519
2520int mpol_set_shared_policy(struct shared_policy *info,
2521			struct vm_area_struct *vma, struct mempolicy *npol)
2522{
2523	int err;
2524	struct sp_node *new = NULL;
2525	unsigned long sz = vma_pages(vma);
2526
2527	pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2528		 vma->vm_pgoff,
2529		 sz, npol ? npol->mode : -1,
2530		 npol ? npol->flags : -1,
2531		 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2532
2533	if (npol) {
2534		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2535		if (!new)
2536			return -ENOMEM;
2537	}
2538	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2539	if (err && new)
2540		sp_free(new);
2541	return err;
2542}
2543
2544/* Free a backing policy store on inode delete. */
2545void mpol_free_shared_policy(struct shared_policy *p)
2546{
2547	struct sp_node *n;
2548	struct rb_node *next;
2549
2550	if (!p->root.rb_node)
2551		return;
2552	write_lock(&p->lock);
2553	next = rb_first(&p->root);
2554	while (next) {
2555		n = rb_entry(next, struct sp_node, nd);
2556		next = rb_next(&n->nd);
2557		sp_delete(p, n);
2558	}
2559	write_unlock(&p->lock);
2560}
2561
2562#ifdef CONFIG_NUMA_BALANCING
2563static int __initdata numabalancing_override;
2564
2565static void __init check_numabalancing_enable(void)
2566{
2567	bool numabalancing_default = false;
2568
2569	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2570		numabalancing_default = true;
2571
2572	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2573	if (numabalancing_override)
2574		set_numabalancing_state(numabalancing_override == 1);
2575
2576	if (num_online_nodes() > 1 && !numabalancing_override) {
2577		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2578			numabalancing_default ? "Enabling" : "Disabling");
2579		set_numabalancing_state(numabalancing_default);
2580	}
2581}
2582
2583static int __init setup_numabalancing(char *str)
2584{
2585	int ret = 0;
2586	if (!str)
2587		goto out;
2588
2589	if (!strcmp(str, "enable")) {
2590		numabalancing_override = 1;
2591		ret = 1;
2592	} else if (!strcmp(str, "disable")) {
2593		numabalancing_override = -1;
2594		ret = 1;
2595	}
2596out:
2597	if (!ret)
2598		pr_warn("Unable to parse numa_balancing=\n");
2599
2600	return ret;
2601}
2602__setup("numa_balancing=", setup_numabalancing);
2603#else
2604static inline void __init check_numabalancing_enable(void)
2605{
2606}
2607#endif /* CONFIG_NUMA_BALANCING */
2608
2609/* assumes fs == KERNEL_DS */
2610void __init numa_policy_init(void)
2611{
2612	nodemask_t interleave_nodes;
2613	unsigned long largest = 0;
2614	int nid, prefer = 0;
2615
2616	policy_cache = kmem_cache_create("numa_policy",
2617					 sizeof(struct mempolicy),
2618					 0, SLAB_PANIC, NULL);
2619
2620	sn_cache = kmem_cache_create("shared_policy_node",
2621				     sizeof(struct sp_node),
2622				     0, SLAB_PANIC, NULL);
2623
2624	for_each_node(nid) {
2625		preferred_node_policy[nid] = (struct mempolicy) {
2626			.refcnt = ATOMIC_INIT(1),
2627			.mode = MPOL_PREFERRED,
2628			.flags = MPOL_F_MOF | MPOL_F_MORON,
2629			.v = { .preferred_node = nid, },
2630		};
2631	}
2632
2633	/*
2634	 * Set interleaving policy for system init. Interleaving is only
2635	 * enabled across suitably sized nodes (default is >= 16MB), or
2636	 * fall back to the largest node if they're all smaller.
2637	 */
2638	nodes_clear(interleave_nodes);
2639	for_each_node_state(nid, N_MEMORY) {
2640		unsigned long total_pages = node_present_pages(nid);
2641
2642		/* Preserve the largest node */
2643		if (largest < total_pages) {
2644			largest = total_pages;
2645			prefer = nid;
2646		}
2647
2648		/* Interleave this node? */
2649		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2650			node_set(nid, interleave_nodes);
2651	}
2652
2653	/* All too small, use the largest */
2654	if (unlikely(nodes_empty(interleave_nodes)))
2655		node_set(prefer, interleave_nodes);
2656
2657	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2658		pr_err("%s: interleaving failed\n", __func__);
2659
2660	check_numabalancing_enable();
2661}
2662
2663/* Reset policy of current process to default */
2664void numa_default_policy(void)
2665{
2666	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2667}
2668
2669/*
2670 * Parse and format mempolicy from/to strings
2671 */
2672
2673/*
2674 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2675 */
2676static const char * const policy_modes[] =
2677{
2678	[MPOL_DEFAULT]    = "default",
2679	[MPOL_PREFERRED]  = "prefer",
2680	[MPOL_BIND]       = "bind",
2681	[MPOL_INTERLEAVE] = "interleave",
2682	[MPOL_LOCAL]      = "local",
2683};
2684
2685
2686#ifdef CONFIG_TMPFS
2687/**
2688 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2689 * @str:  string containing mempolicy to parse
2690 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2691 *
2692 * Format of input:
2693 *	<mode>[=<flags>][:<nodelist>]
2694 *
2695 * On success, returns 0, else 1
2696 */
2697int mpol_parse_str(char *str, struct mempolicy **mpol)
2698{
2699	struct mempolicy *new = NULL;
2700	unsigned short mode;
2701	unsigned short mode_flags;
2702	nodemask_t nodes;
2703	char *nodelist = strchr(str, ':');
2704	char *flags = strchr(str, '=');
2705	int err = 1;
2706
2707	if (nodelist) {
2708		/* NUL-terminate mode or flags string */
2709		*nodelist++ = '\0';
2710		if (nodelist_parse(nodelist, nodes))
2711			goto out;
2712		if (!nodes_subset(nodes, node_states[N_MEMORY]))
2713			goto out;
2714	} else
2715		nodes_clear(nodes);
2716
2717	if (flags)
2718		*flags++ = '\0';	/* terminate mode string */
2719
2720	for (mode = 0; mode < MPOL_MAX; mode++) {
2721		if (!strcmp(str, policy_modes[mode])) {
2722			break;
2723		}
2724	}
2725	if (mode >= MPOL_MAX)
2726		goto out;
2727
2728	switch (mode) {
2729	case MPOL_PREFERRED:
2730		/*
2731		 * Insist on a nodelist of one node only
2732		 */
2733		if (nodelist) {
2734			char *rest = nodelist;
2735			while (isdigit(*rest))
2736				rest++;
2737			if (*rest)
2738				goto out;
2739		}
2740		break;
2741	case MPOL_INTERLEAVE:
2742		/*
2743		 * Default to online nodes with memory if no nodelist
2744		 */
2745		if (!nodelist)
2746			nodes = node_states[N_MEMORY];
2747		break;
2748	case MPOL_LOCAL:
2749		/*
2750		 * Don't allow a nodelist;  mpol_new() checks flags
2751		 */
2752		if (nodelist)
2753			goto out;
2754		mode = MPOL_PREFERRED;
2755		break;
2756	case MPOL_DEFAULT:
2757		/*
2758		 * Insist on a empty nodelist
2759		 */
2760		if (!nodelist)
2761			err = 0;
2762		goto out;
2763	case MPOL_BIND:
2764		/*
2765		 * Insist on a nodelist
2766		 */
2767		if (!nodelist)
2768			goto out;
2769	}
2770
2771	mode_flags = 0;
2772	if (flags) {
2773		/*
2774		 * Currently, we only support two mutually exclusive
2775		 * mode flags.
2776		 */
2777		if (!strcmp(flags, "static"))
2778			mode_flags |= MPOL_F_STATIC_NODES;
2779		else if (!strcmp(flags, "relative"))
2780			mode_flags |= MPOL_F_RELATIVE_NODES;
2781		else
2782			goto out;
2783	}
2784
2785	new = mpol_new(mode, mode_flags, &nodes);
2786	if (IS_ERR(new))
2787		goto out;
2788
2789	/*
2790	 * Save nodes for mpol_to_str() to show the tmpfs mount options
2791	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2792	 */
2793	if (mode != MPOL_PREFERRED)
2794		new->v.nodes = nodes;
2795	else if (nodelist)
2796		new->v.preferred_node = first_node(nodes);
2797	else
2798		new->flags |= MPOL_F_LOCAL;
2799
2800	/*
2801	 * Save nodes for contextualization: this will be used to "clone"
2802	 * the mempolicy in a specific context [cpuset] at a later time.
2803	 */
2804	new->w.user_nodemask = nodes;
2805
2806	err = 0;
2807
2808out:
2809	/* Restore string for error message */
2810	if (nodelist)
2811		*--nodelist = ':';
2812	if (flags)
2813		*--flags = '=';
2814	if (!err)
2815		*mpol = new;
2816	return err;
2817}
2818#endif /* CONFIG_TMPFS */
2819
2820/**
2821 * mpol_to_str - format a mempolicy structure for printing
2822 * @buffer:  to contain formatted mempolicy string
2823 * @maxlen:  length of @buffer
2824 * @pol:  pointer to mempolicy to be formatted
2825 *
2826 * Convert @pol into a string.  If @buffer is too short, truncate the string.
2827 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2828 * longest flag, "relative", and to display at least a few node ids.
2829 */
2830void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2831{
2832	char *p = buffer;
2833	nodemask_t nodes = NODE_MASK_NONE;
2834	unsigned short mode = MPOL_DEFAULT;
2835	unsigned short flags = 0;
2836
2837	if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2838		mode = pol->mode;
2839		flags = pol->flags;
2840	}
2841
2842	switch (mode) {
2843	case MPOL_DEFAULT:
2844		break;
2845	case MPOL_PREFERRED:
2846		if (flags & MPOL_F_LOCAL)
2847			mode = MPOL_LOCAL;
2848		else
2849			node_set(pol->v.preferred_node, nodes);
2850		break;
2851	case MPOL_BIND:
2852	case MPOL_INTERLEAVE:
2853		nodes = pol->v.nodes;
2854		break;
2855	default:
2856		WARN_ON_ONCE(1);
2857		snprintf(p, maxlen, "unknown");
2858		return;
2859	}
2860
2861	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2862
2863	if (flags & MPOL_MODE_FLAGS) {
2864		p += snprintf(p, buffer + maxlen - p, "=");
2865
2866		/*
2867		 * Currently, the only defined flags are mutually exclusive
2868		 */
2869		if (flags & MPOL_F_STATIC_NODES)
2870			p += snprintf(p, buffer + maxlen - p, "static");
2871		else if (flags & MPOL_F_RELATIVE_NODES)
2872			p += snprintf(p, buffer + maxlen - p, "relative");
2873	}
2874
2875	if (!nodes_empty(nodes))
2876		p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2877			       nodemask_pr_args(&nodes));
2878}