Linux Audio

Check our new training course

Loading...
v5.14.15
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Simple NUMA memory policy for the Linux kernel.
   4 *
   5 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
   6 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
   7 *
   8 * NUMA policy allows the user to give hints in which node(s) memory should
   9 * be allocated.
  10 *
  11 * Support four policies per VMA and per process:
  12 *
  13 * The VMA policy has priority over the process policy for a page fault.
  14 *
  15 * interleave     Allocate memory interleaved over a set of nodes,
  16 *                with normal fallback if it fails.
  17 *                For VMA based allocations this interleaves based on the
  18 *                offset into the backing object or offset into the mapping
  19 *                for anonymous memory. For process policy an process counter
  20 *                is used.
  21 *
  22 * bind           Only allocate memory on a specific set of nodes,
  23 *                no fallback.
  24 *                FIXME: memory is allocated starting with the first node
  25 *                to the last. It would be better if bind would truly restrict
  26 *                the allocation to memory nodes instead
  27 *
  28 * preferred       Try a specific node first before normal fallback.
  29 *                As a special case NUMA_NO_NODE here means do the allocation
  30 *                on the local CPU. This is normally identical to default,
  31 *                but useful to set in a VMA when you have a non default
  32 *                process policy.
  33 *
 
 
 
  34 * default        Allocate on the local node first, or when on a VMA
  35 *                use the process policy. This is what Linux always did
  36 *		  in a NUMA aware kernel and still does by, ahem, default.
  37 *
  38 * The process policy is applied for most non interrupt memory allocations
  39 * in that process' context. Interrupts ignore the policies and always
  40 * try to allocate on the local CPU. The VMA policy is only applied for memory
  41 * allocations for a VMA in the VM.
  42 *
  43 * Currently there are a few corner cases in swapping where the policy
  44 * is not applied, but the majority should be handled. When process policy
  45 * is used it is not remembered over swap outs/swap ins.
  46 *
  47 * Only the highest zone in the zone hierarchy gets policied. Allocations
  48 * requesting a lower zone just use default policy. This implies that
  49 * on systems with highmem kernel lowmem allocation don't get policied.
  50 * Same with GFP_DMA allocations.
  51 *
  52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
  53 * all users and remembered even when nobody has memory mapped.
  54 */
  55
  56/* Notebook:
  57   fix mmap readahead to honour policy and enable policy for any page cache
  58   object
  59   statistics for bigpages
  60   global policy for page cache? currently it uses process policy. Requires
  61   first item above.
  62   handle mremap for shared memory (currently ignored for the policy)
  63   grows down?
  64   make bind policy root only? It can trigger oom much faster and the
  65   kernel is not always grateful with that.
  66*/
  67
  68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  69
  70#include <linux/mempolicy.h>
  71#include <linux/pagewalk.h>
  72#include <linux/highmem.h>
  73#include <linux/hugetlb.h>
  74#include <linux/kernel.h>
  75#include <linux/sched.h>
  76#include <linux/sched/mm.h>
  77#include <linux/sched/numa_balancing.h>
  78#include <linux/sched/task.h>
  79#include <linux/nodemask.h>
  80#include <linux/cpuset.h>
  81#include <linux/slab.h>
  82#include <linux/string.h>
  83#include <linux/export.h>
  84#include <linux/nsproxy.h>
  85#include <linux/interrupt.h>
  86#include <linux/init.h>
  87#include <linux/compat.h>
  88#include <linux/ptrace.h>
  89#include <linux/swap.h>
  90#include <linux/seq_file.h>
  91#include <linux/proc_fs.h>
  92#include <linux/migrate.h>
  93#include <linux/ksm.h>
  94#include <linux/rmap.h>
  95#include <linux/security.h>
  96#include <linux/syscalls.h>
  97#include <linux/ctype.h>
  98#include <linux/mm_inline.h>
  99#include <linux/mmu_notifier.h>
 100#include <linux/printk.h>
 101#include <linux/swapops.h>
 102
 103#include <asm/tlbflush.h>
 
 104#include <linux/uaccess.h>
 105
 106#include "internal.h"
 107
 108/* Internal flags */
 109#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)	/* Skip checks for continuous vmas */
 110#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)		/* Invert check for nodemask */
 111
 112static struct kmem_cache *policy_cache;
 113static struct kmem_cache *sn_cache;
 114
 115/* Highest zone. An specific allocation for a zone below that is not
 116   policied. */
 117enum zone_type policy_zone = 0;
 118
 119/*
 120 * run-time system-wide default policy => local allocation
 121 */
 122static struct mempolicy default_policy = {
 123	.refcnt = ATOMIC_INIT(1), /* never free it */
 124	.mode = MPOL_LOCAL,
 125};
 126
 127static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 128
 129/**
 130 * numa_map_to_online_node - Find closest online node
 131 * @node: Node id to start the search
 132 *
 133 * Lookup the next closest node by distance if @nid is not online.
 
 
 134 */
 135int numa_map_to_online_node(int node)
 136{
 137	int min_dist = INT_MAX, dist, n, min_node;
 138
 139	if (node == NUMA_NO_NODE || node_online(node))
 140		return node;
 141
 142	min_node = node;
 143	for_each_online_node(n) {
 144		dist = node_distance(node, n);
 145		if (dist < min_dist) {
 146			min_dist = dist;
 147			min_node = n;
 148		}
 149	}
 150
 151	return min_node;
 152}
 153EXPORT_SYMBOL_GPL(numa_map_to_online_node);
 154
 155struct mempolicy *get_task_policy(struct task_struct *p)
 156{
 157	struct mempolicy *pol = p->mempolicy;
 158	int node;
 159
 160	if (pol)
 161		return pol;
 162
 163	node = numa_node_id();
 164	if (node != NUMA_NO_NODE) {
 165		pol = &preferred_node_policy[node];
 166		/* preferred_node_policy is not initialised early in boot */
 167		if (pol->mode)
 168			return pol;
 169	}
 170
 171	return &default_policy;
 172}
 173
 174static const struct mempolicy_operations {
 175	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 176	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
 177} mpol_ops[MPOL_MAX];
 178
 179static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 180{
 181	return pol->flags & MPOL_MODE_FLAGS;
 182}
 183
 184static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 185				   const nodemask_t *rel)
 186{
 187	nodemask_t tmp;
 188	nodes_fold(tmp, *orig, nodes_weight(*rel));
 189	nodes_onto(*ret, tmp, *rel);
 190}
 191
 192static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
 193{
 194	if (nodes_empty(*nodes))
 195		return -EINVAL;
 196	pol->nodes = *nodes;
 197	return 0;
 198}
 199
 200static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 201{
 202	if (nodes_empty(*nodes))
 203		return -EINVAL;
 204
 205	nodes_clear(pol->nodes);
 206	node_set(first_node(*nodes), pol->nodes);
 207	return 0;
 208}
 209
 210static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
 211{
 212	if (nodes_empty(*nodes))
 213		return -EINVAL;
 214	pol->nodes = *nodes;
 215	return 0;
 216}
 217
 218/*
 219 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 220 * any, for the new policy.  mpol_new() has already validated the nodes
 221 * parameter with respect to the policy mode and flags.
 222 *
 223 * Must be called holding task's alloc_lock to protect task's mems_allowed
 224 * and mempolicy.  May also be called holding the mmap_lock for write.
 225 */
 226static int mpol_set_nodemask(struct mempolicy *pol,
 227		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 228{
 229	int ret;
 230
 231	/*
 232	 * Default (pol==NULL) resp. local memory policies are not a
 233	 * subject of any remapping. They also do not need any special
 234	 * constructor.
 235	 */
 236	if (!pol || pol->mode == MPOL_LOCAL)
 237		return 0;
 238
 239	/* Check N_MEMORY */
 240	nodes_and(nsc->mask1,
 241		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 242
 243	VM_BUG_ON(!nodes);
 244
 245	if (pol->flags & MPOL_F_RELATIVE_NODES)
 246		mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
 247	else
 248		nodes_and(nsc->mask2, *nodes, nsc->mask1);
 249
 250	if (mpol_store_user_nodemask(pol))
 251		pol->w.user_nodemask = *nodes;
 252	else
 253		pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
 254
 255	ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 256	return ret;
 257}
 258
 259/*
 260 * This function just creates a new policy, does some check and simple
 261 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 262 */
 263static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 264				  nodemask_t *nodes)
 265{
 266	struct mempolicy *policy;
 267
 268	pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 269		 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
 270
 271	if (mode == MPOL_DEFAULT) {
 272		if (nodes && !nodes_empty(*nodes))
 273			return ERR_PTR(-EINVAL);
 274		return NULL;
 275	}
 276	VM_BUG_ON(!nodes);
 277
 278	/*
 279	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 280	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 281	 * All other modes require a valid pointer to a non-empty nodemask.
 282	 */
 283	if (mode == MPOL_PREFERRED) {
 284		if (nodes_empty(*nodes)) {
 285			if (((flags & MPOL_F_STATIC_NODES) ||
 286			     (flags & MPOL_F_RELATIVE_NODES)))
 287				return ERR_PTR(-EINVAL);
 288
 289			mode = MPOL_LOCAL;
 290		}
 291	} else if (mode == MPOL_LOCAL) {
 292		if (!nodes_empty(*nodes) ||
 293		    (flags & MPOL_F_STATIC_NODES) ||
 294		    (flags & MPOL_F_RELATIVE_NODES))
 295			return ERR_PTR(-EINVAL);
 296	} else if (nodes_empty(*nodes))
 297		return ERR_PTR(-EINVAL);
 298	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 299	if (!policy)
 300		return ERR_PTR(-ENOMEM);
 301	atomic_set(&policy->refcnt, 1);
 302	policy->mode = mode;
 303	policy->flags = flags;
 
 304
 305	return policy;
 306}
 307
 308/* Slow path of a mpol destructor. */
 309void __mpol_put(struct mempolicy *p)
 310{
 311	if (!atomic_dec_and_test(&p->refcnt))
 312		return;
 313	kmem_cache_free(policy_cache, p);
 314}
 315
 316static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
 317{
 318}
 319
 320static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 321{
 322	nodemask_t tmp;
 323
 324	if (pol->flags & MPOL_F_STATIC_NODES)
 325		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 326	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 327		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 328	else {
 329		nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
 330								*nodes);
 331		pol->w.cpuset_mems_allowed = *nodes;
 332	}
 333
 334	if (nodes_empty(tmp))
 335		tmp = *nodes;
 336
 337	pol->nodes = tmp;
 338}
 339
 340static void mpol_rebind_preferred(struct mempolicy *pol,
 341						const nodemask_t *nodes)
 342{
 343	pol->w.cpuset_mems_allowed = *nodes;
 344}
 345
 346/*
 347 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 348 *
 349 * Per-vma policies are protected by mmap_lock. Allocations using per-task
 350 * policies are protected by task->mems_allowed_seq to prevent a premature
 351 * OOM/allocation failure due to parallel nodemask modification.
 352 */
 353static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
 354{
 355	if (!pol)
 356		return;
 357	if (!mpol_store_user_nodemask(pol) &&
 358	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 359		return;
 360
 361	mpol_ops[pol->mode].rebind(pol, newmask);
 362}
 363
 364/*
 365 * Wrapper for mpol_rebind_policy() that just requires task
 366 * pointer, and updates task mempolicy.
 367 *
 368 * Called with task's alloc_lock held.
 369 */
 370
 371void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
 372{
 373	mpol_rebind_policy(tsk->mempolicy, new);
 374}
 375
 376/*
 377 * Rebind each vma in mm to new nodemask.
 378 *
 379 * Call holding a reference to mm.  Takes mm->mmap_lock during call.
 380 */
 381
 382void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 383{
 384	struct vm_area_struct *vma;
 
 385
 386	mmap_write_lock(mm);
 387	for (vma = mm->mmap; vma; vma = vma->vm_next)
 388		mpol_rebind_policy(vma->vm_policy, new);
 389	mmap_write_unlock(mm);
 390}
 391
 392static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 393	[MPOL_DEFAULT] = {
 394		.rebind = mpol_rebind_default,
 395	},
 396	[MPOL_INTERLEAVE] = {
 397		.create = mpol_new_interleave,
 398		.rebind = mpol_rebind_nodemask,
 399	},
 400	[MPOL_PREFERRED] = {
 401		.create = mpol_new_preferred,
 402		.rebind = mpol_rebind_preferred,
 403	},
 404	[MPOL_BIND] = {
 405		.create = mpol_new_bind,
 406		.rebind = mpol_rebind_nodemask,
 407	},
 408	[MPOL_LOCAL] = {
 409		.rebind = mpol_rebind_default,
 410	},
 
 
 
 
 411};
 412
 413static int migrate_page_add(struct page *page, struct list_head *pagelist,
 414				unsigned long flags);
 415
 416struct queue_pages {
 417	struct list_head *pagelist;
 418	unsigned long flags;
 419	nodemask_t *nmask;
 420	unsigned long start;
 421	unsigned long end;
 422	struct vm_area_struct *first;
 423};
 424
 425/*
 426 * Check if the page's nid is in qp->nmask.
 427 *
 428 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
 429 * in the invert of qp->nmask.
 430 */
 431static inline bool queue_pages_required(struct page *page,
 432					struct queue_pages *qp)
 433{
 434	int nid = page_to_nid(page);
 435	unsigned long flags = qp->flags;
 436
 437	return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
 438}
 439
 440/*
 441 * queue_pages_pmd() has four possible return values:
 442 * 0 - pages are placed on the right node or queued successfully, or
 443 *     special page is met, i.e. huge zero page.
 444 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 445 *     specified.
 446 * 2 - THP was split.
 447 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
 448 *        existing page was already on a node that does not follow the
 449 *        policy.
 450 */
 451static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
 452				unsigned long end, struct mm_walk *walk)
 453	__releases(ptl)
 454{
 455	int ret = 0;
 456	struct page *page;
 457	struct queue_pages *qp = walk->private;
 458	unsigned long flags;
 459
 460	if (unlikely(is_pmd_migration_entry(*pmd))) {
 461		ret = -EIO;
 462		goto unlock;
 463	}
 464	page = pmd_page(*pmd);
 465	if (is_huge_zero_page(page)) {
 466		spin_unlock(ptl);
 467		walk->action = ACTION_CONTINUE;
 468		goto out;
 469	}
 470	if (!queue_pages_required(page, qp))
 471		goto unlock;
 472
 473	flags = qp->flags;
 474	/* go to thp migration */
 475	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 476		if (!vma_migratable(walk->vma) ||
 477		    migrate_page_add(page, qp->pagelist, flags)) {
 478			ret = 1;
 479			goto unlock;
 480		}
 481	} else
 482		ret = -EIO;
 483unlock:
 484	spin_unlock(ptl);
 485out:
 486	return ret;
 487}
 488
 489/*
 490 * Scan through pages checking if pages follow certain conditions,
 491 * and move them to the pagelist if they do.
 492 *
 493 * queue_pages_pte_range() has three possible return values:
 494 * 0 - pages are placed on the right node or queued successfully, or
 495 *     special page is met, i.e. zero page.
 496 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 497 *     specified.
 498 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
 499 *        on a node that does not follow the policy.
 500 */
 501static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
 502			unsigned long end, struct mm_walk *walk)
 503{
 504	struct vm_area_struct *vma = walk->vma;
 505	struct page *page;
 506	struct queue_pages *qp = walk->private;
 507	unsigned long flags = qp->flags;
 508	int ret;
 509	bool has_unmovable = false;
 510	pte_t *pte, *mapped_pte;
 511	spinlock_t *ptl;
 512
 513	ptl = pmd_trans_huge_lock(pmd, vma);
 514	if (ptl) {
 515		ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
 516		if (ret != 2)
 517			return ret;
 518	}
 519	/* THP was split, fall through to pte walk */
 520
 521	if (pmd_trans_unstable(pmd))
 522		return 0;
 523
 524	mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
 525	for (; addr != end; pte++, addr += PAGE_SIZE) {
 526		if (!pte_present(*pte))
 527			continue;
 528		page = vm_normal_page(vma, addr, *pte);
 529		if (!page)
 530			continue;
 531		/*
 532		 * vm_normal_page() filters out zero pages, but there might
 533		 * still be PageReserved pages to skip, perhaps in a VDSO.
 534		 */
 535		if (PageReserved(page))
 536			continue;
 537		if (!queue_pages_required(page, qp))
 538			continue;
 539		if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 540			/* MPOL_MF_STRICT must be specified if we get here */
 541			if (!vma_migratable(vma)) {
 542				has_unmovable = true;
 543				break;
 544			}
 545
 546			/*
 547			 * Do not abort immediately since there may be
 548			 * temporary off LRU pages in the range.  Still
 549			 * need migrate other LRU pages.
 550			 */
 551			if (migrate_page_add(page, qp->pagelist, flags))
 552				has_unmovable = true;
 553		} else
 554			break;
 555	}
 556	pte_unmap_unlock(mapped_pte, ptl);
 557	cond_resched();
 558
 559	if (has_unmovable)
 560		return 1;
 561
 562	return addr != end ? -EIO : 0;
 563}
 564
 565static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
 566			       unsigned long addr, unsigned long end,
 567			       struct mm_walk *walk)
 568{
 569	int ret = 0;
 570#ifdef CONFIG_HUGETLB_PAGE
 571	struct queue_pages *qp = walk->private;
 572	unsigned long flags = (qp->flags & MPOL_MF_VALID);
 573	struct page *page;
 574	spinlock_t *ptl;
 575	pte_t entry;
 576
 577	ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 578	entry = huge_ptep_get(pte);
 579	if (!pte_present(entry))
 580		goto unlock;
 581	page = pte_page(entry);
 582	if (!queue_pages_required(page, qp))
 583		goto unlock;
 584
 585	if (flags == MPOL_MF_STRICT) {
 586		/*
 587		 * STRICT alone means only detecting misplaced page and no
 588		 * need to further check other vma.
 589		 */
 590		ret = -EIO;
 591		goto unlock;
 592	}
 593
 594	if (!vma_migratable(walk->vma)) {
 595		/*
 596		 * Must be STRICT with MOVE*, otherwise .test_walk() have
 597		 * stopped walking current vma.
 598		 * Detecting misplaced page but allow migrating pages which
 599		 * have been queued.
 600		 */
 601		ret = 1;
 602		goto unlock;
 603	}
 604
 605	/* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
 606	if (flags & (MPOL_MF_MOVE_ALL) ||
 607	    (flags & MPOL_MF_MOVE && page_mapcount(page) == 1)) {
 608		if (!isolate_huge_page(page, qp->pagelist) &&
 
 609			(flags & MPOL_MF_STRICT))
 610			/*
 611			 * Failed to isolate page but allow migrating pages
 612			 * which have been queued.
 613			 */
 614			ret = 1;
 615	}
 616unlock:
 617	spin_unlock(ptl);
 618#else
 619	BUG();
 620#endif
 621	return ret;
 622}
 623
 624#ifdef CONFIG_NUMA_BALANCING
 625/*
 626 * This is used to mark a range of virtual addresses to be inaccessible.
 627 * These are later cleared by a NUMA hinting fault. Depending on these
 628 * faults, pages may be migrated for better NUMA placement.
 629 *
 630 * This is assuming that NUMA faults are handled using PROT_NONE. If
 631 * an architecture makes a different choice, it will need further
 632 * changes to the core.
 633 */
 634unsigned long change_prot_numa(struct vm_area_struct *vma,
 635			unsigned long addr, unsigned long end)
 636{
 
 637	int nr_updated;
 638
 639	nr_updated = change_protection(vma, addr, end, PAGE_NONE, MM_CP_PROT_NUMA);
 
 
 
 640	if (nr_updated)
 641		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 642
 
 
 643	return nr_updated;
 644}
 645#else
 646static unsigned long change_prot_numa(struct vm_area_struct *vma,
 647			unsigned long addr, unsigned long end)
 648{
 649	return 0;
 650}
 651#endif /* CONFIG_NUMA_BALANCING */
 652
 653static int queue_pages_test_walk(unsigned long start, unsigned long end,
 654				struct mm_walk *walk)
 655{
 656	struct vm_area_struct *vma = walk->vma;
 657	struct queue_pages *qp = walk->private;
 658	unsigned long endvma = vma->vm_end;
 659	unsigned long flags = qp->flags;
 660
 661	/* range check first */
 662	VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
 663
 664	if (!qp->first) {
 665		qp->first = vma;
 666		if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 667			(qp->start < vma->vm_start))
 668			/* hole at head side of range */
 669			return -EFAULT;
 670	}
 
 671	if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 672		((vma->vm_end < qp->end) &&
 673		(!vma->vm_next || vma->vm_end < vma->vm_next->vm_start)))
 674		/* hole at middle or tail of range */
 675		return -EFAULT;
 676
 677	/*
 678	 * Need check MPOL_MF_STRICT to return -EIO if possible
 679	 * regardless of vma_migratable
 680	 */
 681	if (!vma_migratable(vma) &&
 682	    !(flags & MPOL_MF_STRICT))
 683		return 1;
 684
 685	if (endvma > end)
 686		endvma = end;
 687
 688	if (flags & MPOL_MF_LAZY) {
 689		/* Similar to task_numa_work, skip inaccessible VMAs */
 690		if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
 691			!(vma->vm_flags & VM_MIXEDMAP))
 692			change_prot_numa(vma, start, endvma);
 693		return 1;
 694	}
 695
 696	/* queue pages from current vma */
 697	if (flags & MPOL_MF_VALID)
 698		return 0;
 699	return 1;
 700}
 701
 702static const struct mm_walk_ops queue_pages_walk_ops = {
 703	.hugetlb_entry		= queue_pages_hugetlb,
 704	.pmd_entry		= queue_pages_pte_range,
 705	.test_walk		= queue_pages_test_walk,
 706};
 707
 708/*
 709 * Walk through page tables and collect pages to be migrated.
 710 *
 711 * If pages found in a given range are on a set of nodes (determined by
 712 * @nodes and @flags,) it's isolated and queued to the pagelist which is
 713 * passed via @private.
 714 *
 715 * queue_pages_range() has three possible return values:
 716 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
 717 *     specified.
 718 * 0 - queue pages successfully or no misplaced page.
 719 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
 720 *         memory range specified by nodemask and maxnode points outside
 721 *         your accessible address space (-EFAULT)
 722 */
 723static int
 724queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 725		nodemask_t *nodes, unsigned long flags,
 726		struct list_head *pagelist)
 727{
 728	int err;
 729	struct queue_pages qp = {
 730		.pagelist = pagelist,
 731		.flags = flags,
 732		.nmask = nodes,
 733		.start = start,
 734		.end = end,
 735		.first = NULL,
 736	};
 737
 738	err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
 739
 740	if (!qp.first)
 741		/* whole range in hole */
 742		err = -EFAULT;
 743
 744	return err;
 745}
 746
 747/*
 748 * Apply policy to a single VMA
 749 * This must be called with the mmap_lock held for writing.
 750 */
 751static int vma_replace_policy(struct vm_area_struct *vma,
 752						struct mempolicy *pol)
 753{
 754	int err;
 755	struct mempolicy *old;
 756	struct mempolicy *new;
 757
 758	pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 759		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 760		 vma->vm_ops, vma->vm_file,
 761		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 762
 763	new = mpol_dup(pol);
 764	if (IS_ERR(new))
 765		return PTR_ERR(new);
 766
 767	if (vma->vm_ops && vma->vm_ops->set_policy) {
 768		err = vma->vm_ops->set_policy(vma, new);
 769		if (err)
 770			goto err_out;
 771	}
 772
 773	old = vma->vm_policy;
 774	vma->vm_policy = new; /* protected by mmap_lock */
 775	mpol_put(old);
 776
 777	return 0;
 778 err_out:
 779	mpol_put(new);
 780	return err;
 781}
 782
 783/* Step 2: apply policy to a range and do splits. */
 784static int mbind_range(struct mm_struct *mm, unsigned long start,
 785		       unsigned long end, struct mempolicy *new_pol)
 786{
 787	struct vm_area_struct *next;
 788	struct vm_area_struct *prev;
 789	struct vm_area_struct *vma;
 790	int err = 0;
 791	pgoff_t pgoff;
 792	unsigned long vmstart;
 793	unsigned long vmend;
 794
 795	vma = find_vma(mm, start);
 796	VM_BUG_ON(!vma);
 
 
 
 
 
 797
 798	prev = vma->vm_prev;
 799	if (start > vma->vm_start)
 800		prev = vma;
 801
 802	for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 803		next = vma->vm_next;
 804		vmstart = max(start, vma->vm_start);
 805		vmend   = min(end, vma->vm_end);
 806
 807		if (mpol_equal(vma_policy(vma), new_pol))
 808			continue;
 809
 810		pgoff = vma->vm_pgoff +
 811			((vmstart - vma->vm_start) >> PAGE_SHIFT);
 812		prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 813				 vma->anon_vma, vma->vm_file, pgoff,
 814				 new_pol, vma->vm_userfaultfd_ctx);
 
 815		if (prev) {
 
 
 816			vma = prev;
 817			next = vma->vm_next;
 818			if (mpol_equal(vma_policy(vma), new_pol))
 819				continue;
 820			/* vma_merge() joined vma && vma->next, case 8 */
 821			goto replace;
 822		}
 823		if (vma->vm_start != vmstart) {
 824			err = split_vma(vma->vm_mm, vma, vmstart, 1);
 825			if (err)
 826				goto out;
 
 
 827		}
 828		if (vma->vm_end != vmend) {
 829			err = split_vma(vma->vm_mm, vma, vmend, 0);
 830			if (err)
 831				goto out;
 
 
 832		}
 833 replace:
 834		err = vma_replace_policy(vma, new_pol);
 835		if (err)
 836			goto out;
 
 
 837	}
 838
 839 out:
 840	return err;
 841}
 842
 843/* Set the process memory policy */
 844static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 845			     nodemask_t *nodes)
 846{
 847	struct mempolicy *new, *old;
 848	NODEMASK_SCRATCH(scratch);
 849	int ret;
 850
 851	if (!scratch)
 852		return -ENOMEM;
 853
 854	new = mpol_new(mode, flags, nodes);
 855	if (IS_ERR(new)) {
 856		ret = PTR_ERR(new);
 857		goto out;
 858	}
 859
 
 860	ret = mpol_set_nodemask(new, nodes, scratch);
 861	if (ret) {
 
 862		mpol_put(new);
 863		goto out;
 864	}
 865	task_lock(current);
 866	old = current->mempolicy;
 867	current->mempolicy = new;
 868	if (new && new->mode == MPOL_INTERLEAVE)
 869		current->il_prev = MAX_NUMNODES-1;
 870	task_unlock(current);
 871	mpol_put(old);
 872	ret = 0;
 873out:
 874	NODEMASK_SCRATCH_FREE(scratch);
 875	return ret;
 876}
 877
 878/*
 879 * Return nodemask for policy for get_mempolicy() query
 880 *
 881 * Called with task's alloc_lock held
 882 */
 883static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 884{
 885	nodes_clear(*nodes);
 886	if (p == &default_policy)
 887		return;
 888
 889	switch (p->mode) {
 890	case MPOL_BIND:
 891	case MPOL_INTERLEAVE:
 892	case MPOL_PREFERRED:
 
 893		*nodes = p->nodes;
 894		break;
 895	case MPOL_LOCAL:
 896		/* return empty node mask for local allocation */
 897		break;
 898	default:
 899		BUG();
 900	}
 901}
 902
 903static int lookup_node(struct mm_struct *mm, unsigned long addr)
 904{
 905	struct page *p = NULL;
 906	int err;
 907
 908	int locked = 1;
 909	err = get_user_pages_locked(addr & PAGE_MASK, 1, 0, &p, &locked);
 910	if (err > 0) {
 911		err = page_to_nid(p);
 912		put_page(p);
 913	}
 914	if (locked)
 915		mmap_read_unlock(mm);
 916	return err;
 917}
 918
 919/* Retrieve NUMA policy */
 920static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 921			     unsigned long addr, unsigned long flags)
 922{
 923	int err;
 924	struct mm_struct *mm = current->mm;
 925	struct vm_area_struct *vma = NULL;
 926	struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
 927
 928	if (flags &
 929		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 930		return -EINVAL;
 931
 932	if (flags & MPOL_F_MEMS_ALLOWED) {
 933		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 934			return -EINVAL;
 935		*policy = 0;	/* just so it's initialized */
 936		task_lock(current);
 937		*nmask  = cpuset_current_mems_allowed;
 938		task_unlock(current);
 939		return 0;
 940	}
 941
 942	if (flags & MPOL_F_ADDR) {
 943		/*
 944		 * Do NOT fall back to task policy if the
 945		 * vma/shared policy at addr is NULL.  We
 946		 * want to return MPOL_DEFAULT in this case.
 947		 */
 948		mmap_read_lock(mm);
 949		vma = vma_lookup(mm, addr);
 950		if (!vma) {
 951			mmap_read_unlock(mm);
 952			return -EFAULT;
 953		}
 954		if (vma->vm_ops && vma->vm_ops->get_policy)
 955			pol = vma->vm_ops->get_policy(vma, addr);
 956		else
 957			pol = vma->vm_policy;
 958	} else if (addr)
 959		return -EINVAL;
 960
 961	if (!pol)
 962		pol = &default_policy;	/* indicates default behavior */
 963
 964	if (flags & MPOL_F_NODE) {
 965		if (flags & MPOL_F_ADDR) {
 966			/*
 967			 * Take a refcount on the mpol, lookup_node()
 968			 * will drop the mmap_lock, so after calling
 969			 * lookup_node() only "pol" remains valid, "vma"
 970			 * is stale.
 971			 */
 972			pol_refcount = pol;
 973			vma = NULL;
 974			mpol_get(pol);
 
 975			err = lookup_node(mm, addr);
 976			if (err < 0)
 977				goto out;
 978			*policy = err;
 979		} else if (pol == current->mempolicy &&
 980				pol->mode == MPOL_INTERLEAVE) {
 981			*policy = next_node_in(current->il_prev, pol->nodes);
 982		} else {
 983			err = -EINVAL;
 984			goto out;
 985		}
 986	} else {
 987		*policy = pol == &default_policy ? MPOL_DEFAULT :
 988						pol->mode;
 989		/*
 990		 * Internal mempolicy flags must be masked off before exposing
 991		 * the policy to userspace.
 992		 */
 993		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 994	}
 995
 996	err = 0;
 997	if (nmask) {
 998		if (mpol_store_user_nodemask(pol)) {
 999			*nmask = pol->w.user_nodemask;
1000		} else {
1001			task_lock(current);
1002			get_policy_nodemask(pol, nmask);
1003			task_unlock(current);
1004		}
1005	}
1006
1007 out:
1008	mpol_cond_put(pol);
1009	if (vma)
1010		mmap_read_unlock(mm);
1011	if (pol_refcount)
1012		mpol_put(pol_refcount);
1013	return err;
1014}
1015
1016#ifdef CONFIG_MIGRATION
1017/*
1018 * page migration, thp tail pages can be passed.
1019 */
1020static int migrate_page_add(struct page *page, struct list_head *pagelist,
1021				unsigned long flags)
1022{
1023	struct page *head = compound_head(page);
1024	/*
1025	 * Avoid migrating a page that is shared with others.
1026	 */
1027	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1028		if (!isolate_lru_page(head)) {
1029			list_add_tail(&head->lru, pagelist);
1030			mod_node_page_state(page_pgdat(head),
1031				NR_ISOLATED_ANON + page_is_file_lru(head),
1032				thp_nr_pages(head));
1033		} else if (flags & MPOL_MF_STRICT) {
1034			/*
1035			 * Non-movable page may reach here.  And, there may be
1036			 * temporary off LRU pages or non-LRU movable pages.
1037			 * Treat them as unmovable pages since they can't be
1038			 * isolated, so they can't be moved at the moment.  It
1039			 * should return -EIO for this case too.
1040			 */
1041			return -EIO;
1042		}
1043	}
1044
1045	return 0;
1046}
1047
1048/*
1049 * Migrate pages from one node to a target node.
1050 * Returns error or the number of pages not migrated.
1051 */
1052static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1053			   int flags)
1054{
1055	nodemask_t nmask;
 
1056	LIST_HEAD(pagelist);
1057	int err = 0;
1058	struct migration_target_control mtc = {
1059		.nid = dest,
1060		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1061	};
1062
1063	nodes_clear(nmask);
1064	node_set(source, nmask);
1065
1066	/*
1067	 * This does not "check" the range but isolates all pages that
1068	 * need migration.  Between passing in the full user address
1069	 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1070	 */
 
1071	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1072	queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
1073			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1074
1075	if (!list_empty(&pagelist)) {
1076		err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1077				(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL);
1078		if (err)
1079			putback_movable_pages(&pagelist);
1080	}
1081
1082	return err;
1083}
1084
1085/*
1086 * Move pages between the two nodesets so as to preserve the physical
1087 * layout as much as possible.
1088 *
1089 * Returns the number of page that could not be moved.
1090 */
1091int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1092		     const nodemask_t *to, int flags)
1093{
1094	int busy = 0;
1095	int err = 0;
1096	nodemask_t tmp;
1097
1098	lru_cache_disable();
1099
1100	mmap_read_lock(mm);
1101
1102	/*
1103	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1104	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
1105	 * bit in 'tmp', and return that <source, dest> pair for migration.
1106	 * The pair of nodemasks 'to' and 'from' define the map.
1107	 *
1108	 * If no pair of bits is found that way, fallback to picking some
1109	 * pair of 'source' and 'dest' bits that are not the same.  If the
1110	 * 'source' and 'dest' bits are the same, this represents a node
1111	 * that will be migrating to itself, so no pages need move.
1112	 *
1113	 * If no bits are left in 'tmp', or if all remaining bits left
1114	 * in 'tmp' correspond to the same bit in 'to', return false
1115	 * (nothing left to migrate).
1116	 *
1117	 * This lets us pick a pair of nodes to migrate between, such that
1118	 * if possible the dest node is not already occupied by some other
1119	 * source node, minimizing the risk of overloading the memory on a
1120	 * node that would happen if we migrated incoming memory to a node
1121	 * before migrating outgoing memory source that same node.
1122	 *
1123	 * A single scan of tmp is sufficient.  As we go, we remember the
1124	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1125	 * that not only moved, but what's better, moved to an empty slot
1126	 * (d is not set in tmp), then we break out then, with that pair.
1127	 * Otherwise when we finish scanning from_tmp, we at least have the
1128	 * most recent <s, d> pair that moved.  If we get all the way through
1129	 * the scan of tmp without finding any node that moved, much less
1130	 * moved to an empty node, then there is nothing left worth migrating.
1131	 */
1132
1133	tmp = *from;
1134	while (!nodes_empty(tmp)) {
1135		int s, d;
1136		int source = NUMA_NO_NODE;
1137		int dest = 0;
1138
1139		for_each_node_mask(s, tmp) {
1140
1141			/*
1142			 * do_migrate_pages() tries to maintain the relative
1143			 * node relationship of the pages established between
1144			 * threads and memory areas.
1145                         *
1146			 * However if the number of source nodes is not equal to
1147			 * the number of destination nodes we can not preserve
1148			 * this node relative relationship.  In that case, skip
1149			 * copying memory from a node that is in the destination
1150			 * mask.
1151			 *
1152			 * Example: [2,3,4] -> [3,4,5] moves everything.
1153			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1154			 */
1155
1156			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1157						(node_isset(s, *to)))
1158				continue;
1159
1160			d = node_remap(s, *from, *to);
1161			if (s == d)
1162				continue;
1163
1164			source = s;	/* Node moved. Memorize */
1165			dest = d;
1166
1167			/* dest not in remaining from nodes? */
1168			if (!node_isset(dest, tmp))
1169				break;
1170		}
1171		if (source == NUMA_NO_NODE)
1172			break;
1173
1174		node_clear(source, tmp);
1175		err = migrate_to_node(mm, source, dest, flags);
1176		if (err > 0)
1177			busy += err;
1178		if (err < 0)
1179			break;
1180	}
1181	mmap_read_unlock(mm);
1182
1183	lru_cache_enable();
1184	if (err < 0)
1185		return err;
1186	return busy;
1187
1188}
1189
1190/*
1191 * Allocate a new page for page migration based on vma policy.
1192 * Start by assuming the page is mapped by the same vma as contains @start.
1193 * Search forward from there, if not.  N.B., this assumes that the
1194 * list of pages handed to migrate_pages()--which is how we get here--
1195 * is in virtual address order.
1196 */
1197static struct page *new_page(struct page *page, unsigned long start)
1198{
 
1199	struct vm_area_struct *vma;
1200	unsigned long address;
 
 
1201
1202	vma = find_vma(current->mm, start);
1203	while (vma) {
1204		address = page_address_in_vma(page, vma);
1205		if (address != -EFAULT)
1206			break;
1207		vma = vma->vm_next;
1208	}
1209
1210	if (PageHuge(page)) {
1211		return alloc_huge_page_vma(page_hstate(compound_head(page)),
1212				vma, address);
1213	} else if (PageTransHuge(page)) {
1214		struct page *thp;
1215
1216		thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1217					 HPAGE_PMD_ORDER);
1218		if (!thp)
1219			return NULL;
1220		prep_transhuge_page(thp);
1221		return thp;
1222	}
1223	/*
1224	 * if !vma, alloc_page_vma() will use task or system default policy
1225	 */
1226	return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1227			vma, address);
 
1228}
1229#else
1230
1231static int migrate_page_add(struct page *page, struct list_head *pagelist,
1232				unsigned long flags)
1233{
1234	return -EIO;
1235}
1236
1237int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1238		     const nodemask_t *to, int flags)
1239{
1240	return -ENOSYS;
1241}
1242
1243static struct page *new_page(struct page *page, unsigned long start)
1244{
1245	return NULL;
1246}
1247#endif
1248
1249static long do_mbind(unsigned long start, unsigned long len,
1250		     unsigned short mode, unsigned short mode_flags,
1251		     nodemask_t *nmask, unsigned long flags)
1252{
1253	struct mm_struct *mm = current->mm;
1254	struct mempolicy *new;
1255	unsigned long end;
1256	int err;
1257	int ret;
1258	LIST_HEAD(pagelist);
1259
1260	if (flags & ~(unsigned long)MPOL_MF_VALID)
1261		return -EINVAL;
1262	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1263		return -EPERM;
1264
1265	if (start & ~PAGE_MASK)
1266		return -EINVAL;
1267
1268	if (mode == MPOL_DEFAULT)
1269		flags &= ~MPOL_MF_STRICT;
1270
1271	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1272	end = start + len;
1273
1274	if (end < start)
1275		return -EINVAL;
1276	if (end == start)
1277		return 0;
1278
1279	new = mpol_new(mode, mode_flags, nmask);
1280	if (IS_ERR(new))
1281		return PTR_ERR(new);
1282
1283	if (flags & MPOL_MF_LAZY)
1284		new->flags |= MPOL_F_MOF;
1285
1286	/*
1287	 * If we are using the default policy then operation
1288	 * on discontinuous address spaces is okay after all
1289	 */
1290	if (!new)
1291		flags |= MPOL_MF_DISCONTIG_OK;
1292
1293	pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1294		 start, start + len, mode, mode_flags,
1295		 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1296
1297	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1298
1299		lru_cache_disable();
1300	}
1301	{
1302		NODEMASK_SCRATCH(scratch);
1303		if (scratch) {
1304			mmap_write_lock(mm);
1305			err = mpol_set_nodemask(new, nmask, scratch);
1306			if (err)
1307				mmap_write_unlock(mm);
1308		} else
1309			err = -ENOMEM;
1310		NODEMASK_SCRATCH_FREE(scratch);
1311	}
1312	if (err)
1313		goto mpol_out;
1314
1315	ret = queue_pages_range(mm, start, end, nmask,
1316			  flags | MPOL_MF_INVERT, &pagelist);
1317
1318	if (ret < 0) {
1319		err = ret;
1320		goto up_out;
1321	}
1322
1323	err = mbind_range(mm, start, end, new);
1324
1325	if (!err) {
1326		int nr_failed = 0;
1327
1328		if (!list_empty(&pagelist)) {
1329			WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1330			nr_failed = migrate_pages(&pagelist, new_page, NULL,
1331				start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1332			if (nr_failed)
1333				putback_movable_pages(&pagelist);
1334		}
1335
1336		if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1337			err = -EIO;
1338	} else {
1339up_out:
1340		if (!list_empty(&pagelist))
1341			putback_movable_pages(&pagelist);
1342	}
1343
1344	mmap_write_unlock(mm);
1345mpol_out:
1346	mpol_put(new);
1347	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1348		lru_cache_enable();
1349	return err;
1350}
1351
1352/*
1353 * User space interface with variable sized bitmaps for nodelists.
1354 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1355
1356/* Copy a node mask from user space. */
1357static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1358		     unsigned long maxnode)
1359{
1360	unsigned long k;
1361	unsigned long t;
1362	unsigned long nlongs;
1363	unsigned long endmask;
1364
1365	--maxnode;
1366	nodes_clear(*nodes);
1367	if (maxnode == 0 || !nmask)
1368		return 0;
1369	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1370		return -EINVAL;
1371
1372	nlongs = BITS_TO_LONGS(maxnode);
1373	if ((maxnode % BITS_PER_LONG) == 0)
1374		endmask = ~0UL;
1375	else
1376		endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1377
1378	/*
1379	 * When the user specified more nodes than supported just check
1380	 * if the non supported part is all zero.
1381	 *
1382	 * If maxnode have more longs than MAX_NUMNODES, check
1383	 * the bits in that area first. And then go through to
1384	 * check the rest bits which equal or bigger than MAX_NUMNODES.
1385	 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1386	 */
1387	if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1388		for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1389			if (get_user(t, nmask + k))
1390				return -EFAULT;
1391			if (k == nlongs - 1) {
1392				if (t & endmask)
1393					return -EINVAL;
1394			} else if (t)
1395				return -EINVAL;
1396		}
1397		nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1398		endmask = ~0UL;
1399	}
1400
1401	if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1402		unsigned long valid_mask = endmask;
1403
1404		valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1405		if (get_user(t, nmask + nlongs - 1))
1406			return -EFAULT;
1407		if (t & valid_mask)
 
 
 
 
 
 
 
1408			return -EINVAL;
1409	}
1410
1411	if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1412		return -EFAULT;
1413	nodes_addr(*nodes)[nlongs-1] &= endmask;
1414	return 0;
1415}
1416
1417/* Copy a kernel node mask to user space */
1418static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1419			      nodemask_t *nodes)
1420{
1421	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1422	unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
 
 
 
 
1423
1424	if (copy > nbytes) {
1425		if (copy > PAGE_SIZE)
1426			return -EINVAL;
1427		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1428			return -EFAULT;
1429		copy = nbytes;
 
1430	}
 
 
 
 
 
1431	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1432}
1433
1434/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1435static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1436{
1437	*flags = *mode & MPOL_MODE_FLAGS;
1438	*mode &= ~MPOL_MODE_FLAGS;
1439	if ((unsigned int)(*mode) >= MPOL_MAX)
 
1440		return -EINVAL;
1441	if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1442		return -EINVAL;
1443	if (*flags & MPOL_F_NUMA_BALANCING) {
1444		if (*mode != MPOL_BIND)
1445			return -EINVAL;
1446		*flags |= (MPOL_F_MOF | MPOL_F_MORON);
1447	}
1448	return 0;
1449}
1450
1451static long kernel_mbind(unsigned long start, unsigned long len,
1452			 unsigned long mode, const unsigned long __user *nmask,
1453			 unsigned long maxnode, unsigned int flags)
1454{
1455	unsigned short mode_flags;
1456	nodemask_t nodes;
1457	int lmode = mode;
1458	int err;
1459
1460	start = untagged_addr(start);
1461	err = sanitize_mpol_flags(&lmode, &mode_flags);
1462	if (err)
1463		return err;
1464
1465	err = get_nodes(&nodes, nmask, maxnode);
1466	if (err)
1467		return err;
1468
1469	return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1470}
1471
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1472SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1473		unsigned long, mode, const unsigned long __user *, nmask,
1474		unsigned long, maxnode, unsigned int, flags)
1475{
1476	return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1477}
1478
1479/* Set the process memory policy */
1480static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1481				 unsigned long maxnode)
1482{
1483	unsigned short mode_flags;
1484	nodemask_t nodes;
1485	int lmode = mode;
1486	int err;
1487
1488	err = sanitize_mpol_flags(&lmode, &mode_flags);
1489	if (err)
1490		return err;
1491
1492	err = get_nodes(&nodes, nmask, maxnode);
1493	if (err)
1494		return err;
1495
1496	return do_set_mempolicy(lmode, mode_flags, &nodes);
1497}
1498
1499SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1500		unsigned long, maxnode)
1501{
1502	return kernel_set_mempolicy(mode, nmask, maxnode);
1503}
1504
1505static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1506				const unsigned long __user *old_nodes,
1507				const unsigned long __user *new_nodes)
1508{
1509	struct mm_struct *mm = NULL;
1510	struct task_struct *task;
1511	nodemask_t task_nodes;
1512	int err;
1513	nodemask_t *old;
1514	nodemask_t *new;
1515	NODEMASK_SCRATCH(scratch);
1516
1517	if (!scratch)
1518		return -ENOMEM;
1519
1520	old = &scratch->mask1;
1521	new = &scratch->mask2;
1522
1523	err = get_nodes(old, old_nodes, maxnode);
1524	if (err)
1525		goto out;
1526
1527	err = get_nodes(new, new_nodes, maxnode);
1528	if (err)
1529		goto out;
1530
1531	/* Find the mm_struct */
1532	rcu_read_lock();
1533	task = pid ? find_task_by_vpid(pid) : current;
1534	if (!task) {
1535		rcu_read_unlock();
1536		err = -ESRCH;
1537		goto out;
1538	}
1539	get_task_struct(task);
1540
1541	err = -EINVAL;
1542
1543	/*
1544	 * Check if this process has the right to modify the specified process.
1545	 * Use the regular "ptrace_may_access()" checks.
1546	 */
1547	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1548		rcu_read_unlock();
1549		err = -EPERM;
1550		goto out_put;
1551	}
1552	rcu_read_unlock();
1553
1554	task_nodes = cpuset_mems_allowed(task);
1555	/* Is the user allowed to access the target nodes? */
1556	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1557		err = -EPERM;
1558		goto out_put;
1559	}
1560
1561	task_nodes = cpuset_mems_allowed(current);
1562	nodes_and(*new, *new, task_nodes);
1563	if (nodes_empty(*new))
1564		goto out_put;
1565
1566	err = security_task_movememory(task);
1567	if (err)
1568		goto out_put;
1569
1570	mm = get_task_mm(task);
1571	put_task_struct(task);
1572
1573	if (!mm) {
1574		err = -EINVAL;
1575		goto out;
1576	}
1577
1578	err = do_migrate_pages(mm, old, new,
1579		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1580
1581	mmput(mm);
1582out:
1583	NODEMASK_SCRATCH_FREE(scratch);
1584
1585	return err;
1586
1587out_put:
1588	put_task_struct(task);
1589	goto out;
1590
1591}
1592
1593SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1594		const unsigned long __user *, old_nodes,
1595		const unsigned long __user *, new_nodes)
1596{
1597	return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1598}
1599
1600
1601/* Retrieve NUMA policy */
1602static int kernel_get_mempolicy(int __user *policy,
1603				unsigned long __user *nmask,
1604				unsigned long maxnode,
1605				unsigned long addr,
1606				unsigned long flags)
1607{
1608	int err;
1609	int pval;
1610	nodemask_t nodes;
1611
1612	if (nmask != NULL && maxnode < nr_node_ids)
1613		return -EINVAL;
1614
1615	addr = untagged_addr(addr);
1616
1617	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1618
1619	if (err)
1620		return err;
1621
1622	if (policy && put_user(pval, policy))
1623		return -EFAULT;
1624
1625	if (nmask)
1626		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1627
1628	return err;
1629}
1630
1631SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1632		unsigned long __user *, nmask, unsigned long, maxnode,
1633		unsigned long, addr, unsigned long, flags)
1634{
1635	return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1636}
1637
1638#ifdef CONFIG_COMPAT
1639
1640COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1641		       compat_ulong_t __user *, nmask,
1642		       compat_ulong_t, maxnode,
1643		       compat_ulong_t, addr, compat_ulong_t, flags)
1644{
1645	long err;
1646	unsigned long __user *nm = NULL;
1647	unsigned long nr_bits, alloc_size;
1648	DECLARE_BITMAP(bm, MAX_NUMNODES);
1649
1650	nr_bits = min_t(unsigned long, maxnode-1, nr_node_ids);
1651	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1652
1653	if (nmask)
1654		nm = compat_alloc_user_space(alloc_size);
1655
1656	err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1657
1658	if (!err && nmask) {
1659		unsigned long copy_size;
1660		copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1661		err = copy_from_user(bm, nm, copy_size);
1662		/* ensure entire bitmap is zeroed */
1663		err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1664		err |= compat_put_bitmap(nmask, bm, nr_bits);
1665	}
1666
1667	return err;
1668}
1669
1670COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1671		       compat_ulong_t, maxnode)
1672{
1673	unsigned long __user *nm = NULL;
1674	unsigned long nr_bits, alloc_size;
1675	DECLARE_BITMAP(bm, MAX_NUMNODES);
1676
1677	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1678	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1679
1680	if (nmask) {
1681		if (compat_get_bitmap(bm, nmask, nr_bits))
1682			return -EFAULT;
1683		nm = compat_alloc_user_space(alloc_size);
1684		if (copy_to_user(nm, bm, alloc_size))
1685			return -EFAULT;
1686	}
1687
1688	return kernel_set_mempolicy(mode, nm, nr_bits+1);
1689}
1690
1691COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1692		       compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1693		       compat_ulong_t, maxnode, compat_ulong_t, flags)
1694{
1695	unsigned long __user *nm = NULL;
1696	unsigned long nr_bits, alloc_size;
1697	nodemask_t bm;
1698
1699	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1700	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1701
1702	if (nmask) {
1703		if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1704			return -EFAULT;
1705		nm = compat_alloc_user_space(alloc_size);
1706		if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1707			return -EFAULT;
1708	}
1709
1710	return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1711}
1712
1713COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1714		       compat_ulong_t, maxnode,
1715		       const compat_ulong_t __user *, old_nodes,
1716		       const compat_ulong_t __user *, new_nodes)
1717{
1718	unsigned long __user *old = NULL;
1719	unsigned long __user *new = NULL;
1720	nodemask_t tmp_mask;
1721	unsigned long nr_bits;
1722	unsigned long size;
1723
1724	nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1725	size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1726	if (old_nodes) {
1727		if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1728			return -EFAULT;
1729		old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1730		if (new_nodes)
1731			new = old + size / sizeof(unsigned long);
1732		if (copy_to_user(old, nodes_addr(tmp_mask), size))
1733			return -EFAULT;
1734	}
1735	if (new_nodes) {
1736		if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1737			return -EFAULT;
1738		if (new == NULL)
1739			new = compat_alloc_user_space(size);
1740		if (copy_to_user(new, nodes_addr(tmp_mask), size))
1741			return -EFAULT;
1742	}
1743	return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1744}
1745
1746#endif /* CONFIG_COMPAT */
1747
1748bool vma_migratable(struct vm_area_struct *vma)
1749{
1750	if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1751		return false;
1752
1753	/*
1754	 * DAX device mappings require predictable access latency, so avoid
1755	 * incurring periodic faults.
1756	 */
1757	if (vma_is_dax(vma))
1758		return false;
1759
1760	if (is_vm_hugetlb_page(vma) &&
1761		!hugepage_migration_supported(hstate_vma(vma)))
1762		return false;
1763
1764	/*
1765	 * Migration allocates pages in the highest zone. If we cannot
1766	 * do so then migration (at least from node to node) is not
1767	 * possible.
1768	 */
1769	if (vma->vm_file &&
1770		gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1771			< policy_zone)
1772		return false;
1773	return true;
1774}
1775
1776struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1777						unsigned long addr)
1778{
1779	struct mempolicy *pol = NULL;
1780
1781	if (vma) {
1782		if (vma->vm_ops && vma->vm_ops->get_policy) {
1783			pol = vma->vm_ops->get_policy(vma, addr);
1784		} else if (vma->vm_policy) {
1785			pol = vma->vm_policy;
1786
1787			/*
1788			 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1789			 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1790			 * count on these policies which will be dropped by
1791			 * mpol_cond_put() later
1792			 */
1793			if (mpol_needs_cond_ref(pol))
1794				mpol_get(pol);
1795		}
1796	}
1797
1798	return pol;
1799}
1800
1801/*
1802 * get_vma_policy(@vma, @addr)
1803 * @vma: virtual memory area whose policy is sought
1804 * @addr: address in @vma for shared policy lookup
1805 *
1806 * Returns effective policy for a VMA at specified address.
1807 * Falls back to current->mempolicy or system default policy, as necessary.
1808 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1809 * count--added by the get_policy() vm_op, as appropriate--to protect against
1810 * freeing by another task.  It is the caller's responsibility to free the
1811 * extra reference for shared policies.
1812 */
1813static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1814						unsigned long addr)
1815{
1816	struct mempolicy *pol = __get_vma_policy(vma, addr);
1817
1818	if (!pol)
1819		pol = get_task_policy(current);
1820
1821	return pol;
1822}
1823
1824bool vma_policy_mof(struct vm_area_struct *vma)
1825{
1826	struct mempolicy *pol;
1827
1828	if (vma->vm_ops && vma->vm_ops->get_policy) {
1829		bool ret = false;
1830
1831		pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1832		if (pol && (pol->flags & MPOL_F_MOF))
1833			ret = true;
1834		mpol_cond_put(pol);
1835
1836		return ret;
1837	}
1838
1839	pol = vma->vm_policy;
1840	if (!pol)
1841		pol = get_task_policy(current);
1842
1843	return pol->flags & MPOL_F_MOF;
1844}
1845
1846static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1847{
1848	enum zone_type dynamic_policy_zone = policy_zone;
1849
1850	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1851
1852	/*
1853	 * if policy->nodes has movable memory only,
1854	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1855	 *
1856	 * policy->nodes is intersect with node_states[N_MEMORY].
1857	 * so if the following test fails, it implies
1858	 * policy->nodes has movable memory only.
1859	 */
1860	if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1861		dynamic_policy_zone = ZONE_MOVABLE;
1862
1863	return zone >= dynamic_policy_zone;
1864}
1865
1866/*
1867 * Return a nodemask representing a mempolicy for filtering nodes for
1868 * page allocation
1869 */
1870nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1871{
 
 
1872	/* Lower zones don't get a nodemask applied for MPOL_BIND */
1873	if (unlikely(policy->mode == MPOL_BIND) &&
1874			apply_policy_zone(policy, gfp_zone(gfp)) &&
1875			cpuset_nodemask_valid_mems_allowed(&policy->nodes))
 
 
 
1876		return &policy->nodes;
1877
1878	return NULL;
1879}
1880
1881/* Return the node id preferred by the given mempolicy, or the given id */
 
 
 
 
 
 
1882static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1883{
1884	if (policy->mode == MPOL_PREFERRED) {
1885		nd = first_node(policy->nodes);
1886	} else {
1887		/*
1888		 * __GFP_THISNODE shouldn't even be used with the bind policy
1889		 * because we might easily break the expectation to stay on the
1890		 * requested node and not break the policy.
1891		 */
1892		WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1893	}
1894
 
 
 
 
 
1895	return nd;
1896}
1897
1898/* Do dynamic interleaving for a process */
1899static unsigned interleave_nodes(struct mempolicy *policy)
1900{
1901	unsigned next;
1902	struct task_struct *me = current;
1903
1904	next = next_node_in(me->il_prev, policy->nodes);
1905	if (next < MAX_NUMNODES)
1906		me->il_prev = next;
1907	return next;
1908}
1909
1910/*
1911 * Depending on the memory policy provide a node from which to allocate the
1912 * next slab entry.
1913 */
1914unsigned int mempolicy_slab_node(void)
1915{
1916	struct mempolicy *policy;
1917	int node = numa_mem_id();
1918
1919	if (in_interrupt())
1920		return node;
1921
1922	policy = current->mempolicy;
1923	if (!policy)
1924		return node;
1925
1926	switch (policy->mode) {
1927	case MPOL_PREFERRED:
1928		return first_node(policy->nodes);
1929
1930	case MPOL_INTERLEAVE:
1931		return interleave_nodes(policy);
1932
1933	case MPOL_BIND: {
 
 
1934		struct zoneref *z;
1935
1936		/*
1937		 * Follow bind policy behavior and start allocation at the
1938		 * first node.
1939		 */
1940		struct zonelist *zonelist;
1941		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1942		zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1943		z = first_zones_zonelist(zonelist, highest_zoneidx,
1944							&policy->nodes);
1945		return z->zone ? zone_to_nid(z->zone) : node;
1946	}
1947	case MPOL_LOCAL:
1948		return node;
1949
1950	default:
1951		BUG();
1952	}
1953}
1954
1955/*
1956 * Do static interleaving for a VMA with known offset @n.  Returns the n'th
1957 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1958 * number of present nodes.
1959 */
1960static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1961{
1962	nodemask_t nodemask = pol->nodes;
1963	unsigned int target, nnodes;
1964	int i;
1965	int nid;
1966	/*
1967	 * The barrier will stabilize the nodemask in a register or on
1968	 * the stack so that it will stop changing under the code.
1969	 *
1970	 * Between first_node() and next_node(), pol->nodes could be changed
1971	 * by other threads. So we put pol->nodes in a local stack.
1972	 */
1973	barrier();
1974
1975	nnodes = nodes_weight(nodemask);
1976	if (!nnodes)
1977		return numa_node_id();
1978	target = (unsigned int)n % nnodes;
1979	nid = first_node(nodemask);
1980	for (i = 0; i < target; i++)
1981		nid = next_node(nid, nodemask);
1982	return nid;
1983}
1984
1985/* Determine a node number for interleave */
1986static inline unsigned interleave_nid(struct mempolicy *pol,
1987		 struct vm_area_struct *vma, unsigned long addr, int shift)
1988{
1989	if (vma) {
1990		unsigned long off;
1991
1992		/*
1993		 * for small pages, there is no difference between
1994		 * shift and PAGE_SHIFT, so the bit-shift is safe.
1995		 * for huge pages, since vm_pgoff is in units of small
1996		 * pages, we need to shift off the always 0 bits to get
1997		 * a useful offset.
1998		 */
1999		BUG_ON(shift < PAGE_SHIFT);
2000		off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
2001		off += (addr - vma->vm_start) >> shift;
2002		return offset_il_node(pol, off);
2003	} else
2004		return interleave_nodes(pol);
2005}
2006
2007#ifdef CONFIG_HUGETLBFS
2008/*
2009 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2010 * @vma: virtual memory area whose policy is sought
2011 * @addr: address in @vma for shared policy lookup and interleave policy
2012 * @gfp_flags: for requested zone
2013 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2014 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
2015 *
2016 * Returns a nid suitable for a huge page allocation and a pointer
2017 * to the struct mempolicy for conditional unref after allocation.
2018 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
2019 * @nodemask for filtering the zonelist.
2020 *
2021 * Must be protected by read_mems_allowed_begin()
2022 */
2023int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2024				struct mempolicy **mpol, nodemask_t **nodemask)
2025{
2026	int nid;
 
2027
2028	*mpol = get_vma_policy(vma, addr);
2029	*nodemask = NULL;	/* assume !MPOL_BIND */
 
2030
2031	if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
2032		nid = interleave_nid(*mpol, vma, addr,
2033					huge_page_shift(hstate_vma(vma)));
2034	} else {
2035		nid = policy_node(gfp_flags, *mpol, numa_node_id());
2036		if ((*mpol)->mode == MPOL_BIND)
2037			*nodemask = &(*mpol)->nodes;
2038	}
2039	return nid;
2040}
2041
2042/*
2043 * init_nodemask_of_mempolicy
2044 *
2045 * If the current task's mempolicy is "default" [NULL], return 'false'
2046 * to indicate default policy.  Otherwise, extract the policy nodemask
2047 * for 'bind' or 'interleave' policy into the argument nodemask, or
2048 * initialize the argument nodemask to contain the single node for
2049 * 'preferred' or 'local' policy and return 'true' to indicate presence
2050 * of non-default mempolicy.
2051 *
2052 * We don't bother with reference counting the mempolicy [mpol_get/put]
2053 * because the current task is examining it's own mempolicy and a task's
2054 * mempolicy is only ever changed by the task itself.
2055 *
2056 * N.B., it is the caller's responsibility to free a returned nodemask.
2057 */
2058bool init_nodemask_of_mempolicy(nodemask_t *mask)
2059{
2060	struct mempolicy *mempolicy;
2061
2062	if (!(mask && current->mempolicy))
2063		return false;
2064
2065	task_lock(current);
2066	mempolicy = current->mempolicy;
2067	switch (mempolicy->mode) {
2068	case MPOL_PREFERRED:
 
2069	case MPOL_BIND:
2070	case MPOL_INTERLEAVE:
2071		*mask = mempolicy->nodes;
2072		break;
2073
2074	case MPOL_LOCAL:
2075		init_nodemask_of_node(mask, numa_node_id());
2076		break;
2077
2078	default:
2079		BUG();
2080	}
2081	task_unlock(current);
2082
2083	return true;
2084}
2085#endif
2086
2087/*
2088 * mempolicy_in_oom_domain
2089 *
2090 * If tsk's mempolicy is "bind", check for intersection between mask and
2091 * the policy nodemask. Otherwise, return true for all other policies
2092 * including "interleave", as a tsk with "interleave" policy may have
2093 * memory allocated from all nodes in system.
2094 *
2095 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2096 */
2097bool mempolicy_in_oom_domain(struct task_struct *tsk,
2098					const nodemask_t *mask)
2099{
2100	struct mempolicy *mempolicy;
2101	bool ret = true;
2102
2103	if (!mask)
2104		return ret;
2105
2106	task_lock(tsk);
2107	mempolicy = tsk->mempolicy;
2108	if (mempolicy && mempolicy->mode == MPOL_BIND)
2109		ret = nodes_intersects(mempolicy->nodes, *mask);
2110	task_unlock(tsk);
2111
2112	return ret;
2113}
2114
2115/* Allocate a page in interleaved policy.
2116   Own path because it needs to do special accounting. */
2117static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2118					unsigned nid)
2119{
2120	struct page *page;
2121
2122	page = __alloc_pages(gfp, order, nid, NULL);
2123	/* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2124	if (!static_branch_likely(&vm_numa_stat_key))
2125		return page;
2126	if (page && page_to_nid(page) == nid) {
2127		preempt_disable();
2128		__count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2129		preempt_enable();
2130	}
2131	return page;
2132}
2133
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2134/**
2135 * alloc_pages_vma - Allocate a page for a VMA.
2136 * @gfp: GFP flags.
2137 * @order: Order of the GFP allocation.
2138 * @vma: Pointer to VMA or NULL if not available.
2139 * @addr: Virtual address of the allocation.  Must be inside @vma.
2140 * @node: Which node to prefer for allocation (modulo policy).
2141 * @hugepage: For hugepages try only the preferred node if possible.
2142 *
2143 * Allocate a page for a specific address in @vma, using the appropriate
2144 * NUMA policy.  When @vma is not NULL the caller must hold the mmap_lock
2145 * of the mm_struct of the VMA to prevent it from going away.  Should be
2146 * used for all allocations for pages that will be mapped into user space.
2147 *
2148 * Return: The page on success or NULL if allocation fails.
2149 */
2150struct page *alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2151		unsigned long addr, int node, bool hugepage)
2152{
2153	struct mempolicy *pol;
2154	struct page *page;
 
2155	int preferred_nid;
2156	nodemask_t *nmask;
2157
2158	pol = get_vma_policy(vma, addr);
2159
2160	if (pol->mode == MPOL_INTERLEAVE) {
 
2161		unsigned nid;
2162
2163		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2164		mpol_cond_put(pol);
 
2165		page = alloc_page_interleave(gfp, order, nid);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2166		goto out;
2167	}
2168
2169	if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2170		int hpage_node = node;
2171
2172		/*
2173		 * For hugepage allocation and non-interleave policy which
2174		 * allows the current node (or other explicitly preferred
2175		 * node) we only try to allocate from the current/preferred
2176		 * node and don't fall back to other nodes, as the cost of
2177		 * remote accesses would likely offset THP benefits.
2178		 *
2179		 * If the policy is interleave, or does not allow the current
2180		 * node in its nodemask, we allocate the standard way.
2181		 */
2182		if (pol->mode == MPOL_PREFERRED)
2183			hpage_node = first_node(pol->nodes);
2184
2185		nmask = policy_nodemask(gfp, pol);
2186		if (!nmask || node_isset(hpage_node, *nmask)) {
2187			mpol_cond_put(pol);
2188			/*
2189			 * First, try to allocate THP only on local node, but
2190			 * don't reclaim unnecessarily, just compact.
2191			 */
2192			page = __alloc_pages_node(hpage_node,
2193				gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2194
2195			/*
2196			 * If hugepage allocations are configured to always
2197			 * synchronous compact or the vma has been madvised
2198			 * to prefer hugepage backing, retry allowing remote
2199			 * memory with both reclaim and compact as well.
2200			 */
2201			if (!page && (gfp & __GFP_DIRECT_RECLAIM))
2202				page = __alloc_pages_node(hpage_node,
2203								gfp, order);
2204
2205			goto out;
2206		}
2207	}
2208
2209	nmask = policy_nodemask(gfp, pol);
2210	preferred_nid = policy_node(gfp, pol, node);
2211	page = __alloc_pages(gfp, order, preferred_nid, nmask);
2212	mpol_cond_put(pol);
2213out:
2214	return page;
2215}
2216EXPORT_SYMBOL(alloc_pages_vma);
2217
2218/**
2219 * alloc_pages - Allocate pages.
2220 * @gfp: GFP flags.
2221 * @order: Power of two of number of pages to allocate.
2222 *
2223 * Allocate 1 << @order contiguous pages.  The physical address of the
2224 * first page is naturally aligned (eg an order-3 allocation will be aligned
2225 * to a multiple of 8 * PAGE_SIZE bytes).  The NUMA policy of the current
2226 * process is honoured when in process context.
2227 *
2228 * Context: Can be called from any context, providing the appropriate GFP
2229 * flags are used.
2230 * Return: The page on success or NULL if allocation fails.
2231 */
2232struct page *alloc_pages(gfp_t gfp, unsigned order)
2233{
2234	struct mempolicy *pol = &default_policy;
2235	struct page *page;
2236
2237	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2238		pol = get_task_policy(current);
2239
2240	/*
2241	 * No reference counting needed for current->mempolicy
2242	 * nor system default_policy
2243	 */
2244	if (pol->mode == MPOL_INTERLEAVE)
2245		page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
 
 
 
2246	else
2247		page = __alloc_pages(gfp, order,
2248				policy_node(gfp, pol, numa_node_id()),
2249				policy_nodemask(gfp, pol));
2250
2251	return page;
2252}
2253EXPORT_SYMBOL(alloc_pages);
2254
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2255int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2256{
2257	struct mempolicy *pol = mpol_dup(vma_policy(src));
2258
2259	if (IS_ERR(pol))
2260		return PTR_ERR(pol);
2261	dst->vm_policy = pol;
2262	return 0;
2263}
2264
2265/*
2266 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2267 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2268 * with the mems_allowed returned by cpuset_mems_allowed().  This
2269 * keeps mempolicies cpuset relative after its cpuset moves.  See
2270 * further kernel/cpuset.c update_nodemask().
2271 *
2272 * current's mempolicy may be rebinded by the other task(the task that changes
2273 * cpuset's mems), so we needn't do rebind work for current task.
2274 */
2275
2276/* Slow path of a mempolicy duplicate */
2277struct mempolicy *__mpol_dup(struct mempolicy *old)
2278{
2279	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2280
2281	if (!new)
2282		return ERR_PTR(-ENOMEM);
2283
2284	/* task's mempolicy is protected by alloc_lock */
2285	if (old == current->mempolicy) {
2286		task_lock(current);
2287		*new = *old;
2288		task_unlock(current);
2289	} else
2290		*new = *old;
2291
2292	if (current_cpuset_is_being_rebound()) {
2293		nodemask_t mems = cpuset_mems_allowed(current);
2294		mpol_rebind_policy(new, &mems);
2295	}
2296	atomic_set(&new->refcnt, 1);
2297	return new;
2298}
2299
2300/* Slow path of a mempolicy comparison */
2301bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2302{
2303	if (!a || !b)
2304		return false;
2305	if (a->mode != b->mode)
2306		return false;
2307	if (a->flags != b->flags)
2308		return false;
 
 
2309	if (mpol_store_user_nodemask(a))
2310		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2311			return false;
2312
2313	switch (a->mode) {
2314	case MPOL_BIND:
2315	case MPOL_INTERLEAVE:
2316	case MPOL_PREFERRED:
 
2317		return !!nodes_equal(a->nodes, b->nodes);
2318	case MPOL_LOCAL:
2319		return true;
2320	default:
2321		BUG();
2322		return false;
2323	}
2324}
2325
2326/*
2327 * Shared memory backing store policy support.
2328 *
2329 * Remember policies even when nobody has shared memory mapped.
2330 * The policies are kept in Red-Black tree linked from the inode.
2331 * They are protected by the sp->lock rwlock, which should be held
2332 * for any accesses to the tree.
2333 */
2334
2335/*
2336 * lookup first element intersecting start-end.  Caller holds sp->lock for
2337 * reading or for writing
2338 */
2339static struct sp_node *
2340sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2341{
2342	struct rb_node *n = sp->root.rb_node;
2343
2344	while (n) {
2345		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2346
2347		if (start >= p->end)
2348			n = n->rb_right;
2349		else if (end <= p->start)
2350			n = n->rb_left;
2351		else
2352			break;
2353	}
2354	if (!n)
2355		return NULL;
2356	for (;;) {
2357		struct sp_node *w = NULL;
2358		struct rb_node *prev = rb_prev(n);
2359		if (!prev)
2360			break;
2361		w = rb_entry(prev, struct sp_node, nd);
2362		if (w->end <= start)
2363			break;
2364		n = prev;
2365	}
2366	return rb_entry(n, struct sp_node, nd);
2367}
2368
2369/*
2370 * Insert a new shared policy into the list.  Caller holds sp->lock for
2371 * writing.
2372 */
2373static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2374{
2375	struct rb_node **p = &sp->root.rb_node;
2376	struct rb_node *parent = NULL;
2377	struct sp_node *nd;
2378
2379	while (*p) {
2380		parent = *p;
2381		nd = rb_entry(parent, struct sp_node, nd);
2382		if (new->start < nd->start)
2383			p = &(*p)->rb_left;
2384		else if (new->end > nd->end)
2385			p = &(*p)->rb_right;
2386		else
2387			BUG();
2388	}
2389	rb_link_node(&new->nd, parent, p);
2390	rb_insert_color(&new->nd, &sp->root);
2391	pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2392		 new->policy ? new->policy->mode : 0);
2393}
2394
2395/* Find shared policy intersecting idx */
2396struct mempolicy *
2397mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2398{
2399	struct mempolicy *pol = NULL;
2400	struct sp_node *sn;
2401
2402	if (!sp->root.rb_node)
2403		return NULL;
2404	read_lock(&sp->lock);
2405	sn = sp_lookup(sp, idx, idx+1);
2406	if (sn) {
2407		mpol_get(sn->policy);
2408		pol = sn->policy;
2409	}
2410	read_unlock(&sp->lock);
2411	return pol;
2412}
2413
2414static void sp_free(struct sp_node *n)
2415{
2416	mpol_put(n->policy);
2417	kmem_cache_free(sn_cache, n);
2418}
2419
2420/**
2421 * mpol_misplaced - check whether current page node is valid in policy
2422 *
2423 * @page: page to be checked
2424 * @vma: vm area where page mapped
2425 * @addr: virtual address where page mapped
2426 *
2427 * Lookup current policy node id for vma,addr and "compare to" page's
2428 * node id.  Policy determination "mimics" alloc_page_vma().
2429 * Called from fault path where we know the vma and faulting address.
2430 *
2431 * Return: -1 if the page is in a node that is valid for this policy, or a
2432 * suitable node ID to allocate a replacement page from.
2433 */
2434int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2435{
2436	struct mempolicy *pol;
2437	struct zoneref *z;
2438	int curnid = page_to_nid(page);
2439	unsigned long pgoff;
2440	int thiscpu = raw_smp_processor_id();
2441	int thisnid = cpu_to_node(thiscpu);
2442	int polnid = NUMA_NO_NODE;
2443	int ret = -1;
2444
2445	pol = get_vma_policy(vma, addr);
2446	if (!(pol->flags & MPOL_F_MOF))
2447		goto out;
2448
2449	switch (pol->mode) {
2450	case MPOL_INTERLEAVE:
2451		pgoff = vma->vm_pgoff;
2452		pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2453		polnid = offset_il_node(pol, pgoff);
2454		break;
2455
2456	case MPOL_PREFERRED:
 
 
2457		polnid = first_node(pol->nodes);
2458		break;
2459
2460	case MPOL_LOCAL:
2461		polnid = numa_node_id();
2462		break;
2463
2464	case MPOL_BIND:
2465		/* Optimize placement among multiple nodes via NUMA balancing */
2466		if (pol->flags & MPOL_F_MORON) {
2467			if (node_isset(thisnid, pol->nodes))
2468				break;
2469			goto out;
2470		}
 
2471
 
2472		/*
2473		 * allows binding to multiple nodes.
2474		 * use current page if in policy nodemask,
2475		 * else select nearest allowed node, if any.
2476		 * If no allowed nodes, use current [!misplaced].
2477		 */
2478		if (node_isset(curnid, pol->nodes))
2479			goto out;
2480		z = first_zones_zonelist(
2481				node_zonelist(numa_node_id(), GFP_HIGHUSER),
2482				gfp_zone(GFP_HIGHUSER),
2483				&pol->nodes);
2484		polnid = zone_to_nid(z->zone);
2485		break;
2486
2487	default:
2488		BUG();
2489	}
2490
2491	/* Migrate the page towards the node whose CPU is referencing it */
2492	if (pol->flags & MPOL_F_MORON) {
2493		polnid = thisnid;
2494
2495		if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2496			goto out;
2497	}
2498
2499	if (curnid != polnid)
2500		ret = polnid;
2501out:
2502	mpol_cond_put(pol);
2503
2504	return ret;
2505}
2506
2507/*
2508 * Drop the (possibly final) reference to task->mempolicy.  It needs to be
2509 * dropped after task->mempolicy is set to NULL so that any allocation done as
2510 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2511 * policy.
2512 */
2513void mpol_put_task_policy(struct task_struct *task)
2514{
2515	struct mempolicy *pol;
2516
2517	task_lock(task);
2518	pol = task->mempolicy;
2519	task->mempolicy = NULL;
2520	task_unlock(task);
2521	mpol_put(pol);
2522}
2523
2524static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2525{
2526	pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2527	rb_erase(&n->nd, &sp->root);
2528	sp_free(n);
2529}
2530
2531static void sp_node_init(struct sp_node *node, unsigned long start,
2532			unsigned long end, struct mempolicy *pol)
2533{
2534	node->start = start;
2535	node->end = end;
2536	node->policy = pol;
2537}
2538
2539static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2540				struct mempolicy *pol)
2541{
2542	struct sp_node *n;
2543	struct mempolicy *newpol;
2544
2545	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2546	if (!n)
2547		return NULL;
2548
2549	newpol = mpol_dup(pol);
2550	if (IS_ERR(newpol)) {
2551		kmem_cache_free(sn_cache, n);
2552		return NULL;
2553	}
2554	newpol->flags |= MPOL_F_SHARED;
2555	sp_node_init(n, start, end, newpol);
2556
2557	return n;
2558}
2559
2560/* Replace a policy range. */
2561static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2562				 unsigned long end, struct sp_node *new)
2563{
2564	struct sp_node *n;
2565	struct sp_node *n_new = NULL;
2566	struct mempolicy *mpol_new = NULL;
2567	int ret = 0;
2568
2569restart:
2570	write_lock(&sp->lock);
2571	n = sp_lookup(sp, start, end);
2572	/* Take care of old policies in the same range. */
2573	while (n && n->start < end) {
2574		struct rb_node *next = rb_next(&n->nd);
2575		if (n->start >= start) {
2576			if (n->end <= end)
2577				sp_delete(sp, n);
2578			else
2579				n->start = end;
2580		} else {
2581			/* Old policy spanning whole new range. */
2582			if (n->end > end) {
2583				if (!n_new)
2584					goto alloc_new;
2585
2586				*mpol_new = *n->policy;
2587				atomic_set(&mpol_new->refcnt, 1);
2588				sp_node_init(n_new, end, n->end, mpol_new);
2589				n->end = start;
2590				sp_insert(sp, n_new);
2591				n_new = NULL;
2592				mpol_new = NULL;
2593				break;
2594			} else
2595				n->end = start;
2596		}
2597		if (!next)
2598			break;
2599		n = rb_entry(next, struct sp_node, nd);
2600	}
2601	if (new)
2602		sp_insert(sp, new);
2603	write_unlock(&sp->lock);
2604	ret = 0;
2605
2606err_out:
2607	if (mpol_new)
2608		mpol_put(mpol_new);
2609	if (n_new)
2610		kmem_cache_free(sn_cache, n_new);
2611
2612	return ret;
2613
2614alloc_new:
2615	write_unlock(&sp->lock);
2616	ret = -ENOMEM;
2617	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2618	if (!n_new)
2619		goto err_out;
2620	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2621	if (!mpol_new)
2622		goto err_out;
 
2623	goto restart;
2624}
2625
2626/**
2627 * mpol_shared_policy_init - initialize shared policy for inode
2628 * @sp: pointer to inode shared policy
2629 * @mpol:  struct mempolicy to install
2630 *
2631 * Install non-NULL @mpol in inode's shared policy rb-tree.
2632 * On entry, the current task has a reference on a non-NULL @mpol.
2633 * This must be released on exit.
2634 * This is called at get_inode() calls and we can use GFP_KERNEL.
2635 */
2636void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2637{
2638	int ret;
2639
2640	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2641	rwlock_init(&sp->lock);
2642
2643	if (mpol) {
2644		struct vm_area_struct pvma;
2645		struct mempolicy *new;
2646		NODEMASK_SCRATCH(scratch);
2647
2648		if (!scratch)
2649			goto put_mpol;
2650		/* contextualize the tmpfs mount point mempolicy */
2651		new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2652		if (IS_ERR(new))
2653			goto free_scratch; /* no valid nodemask intersection */
2654
2655		task_lock(current);
2656		ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2657		task_unlock(current);
2658		if (ret)
2659			goto put_new;
2660
2661		/* Create pseudo-vma that contains just the policy */
2662		vma_init(&pvma, NULL);
2663		pvma.vm_end = TASK_SIZE;	/* policy covers entire file */
2664		mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2665
2666put_new:
2667		mpol_put(new);			/* drop initial ref */
2668free_scratch:
2669		NODEMASK_SCRATCH_FREE(scratch);
2670put_mpol:
2671		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
2672	}
2673}
2674
2675int mpol_set_shared_policy(struct shared_policy *info,
2676			struct vm_area_struct *vma, struct mempolicy *npol)
2677{
2678	int err;
2679	struct sp_node *new = NULL;
2680	unsigned long sz = vma_pages(vma);
2681
2682	pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2683		 vma->vm_pgoff,
2684		 sz, npol ? npol->mode : -1,
2685		 npol ? npol->flags : -1,
2686		 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2687
2688	if (npol) {
2689		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2690		if (!new)
2691			return -ENOMEM;
2692	}
2693	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2694	if (err && new)
2695		sp_free(new);
2696	return err;
2697}
2698
2699/* Free a backing policy store on inode delete. */
2700void mpol_free_shared_policy(struct shared_policy *p)
2701{
2702	struct sp_node *n;
2703	struct rb_node *next;
2704
2705	if (!p->root.rb_node)
2706		return;
2707	write_lock(&p->lock);
2708	next = rb_first(&p->root);
2709	while (next) {
2710		n = rb_entry(next, struct sp_node, nd);
2711		next = rb_next(&n->nd);
2712		sp_delete(p, n);
2713	}
2714	write_unlock(&p->lock);
2715}
2716
2717#ifdef CONFIG_NUMA_BALANCING
2718static int __initdata numabalancing_override;
2719
2720static void __init check_numabalancing_enable(void)
2721{
2722	bool numabalancing_default = false;
2723
2724	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2725		numabalancing_default = true;
2726
2727	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2728	if (numabalancing_override)
2729		set_numabalancing_state(numabalancing_override == 1);
2730
2731	if (num_online_nodes() > 1 && !numabalancing_override) {
2732		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2733			numabalancing_default ? "Enabling" : "Disabling");
2734		set_numabalancing_state(numabalancing_default);
2735	}
2736}
2737
2738static int __init setup_numabalancing(char *str)
2739{
2740	int ret = 0;
2741	if (!str)
2742		goto out;
2743
2744	if (!strcmp(str, "enable")) {
2745		numabalancing_override = 1;
2746		ret = 1;
2747	} else if (!strcmp(str, "disable")) {
2748		numabalancing_override = -1;
2749		ret = 1;
2750	}
2751out:
2752	if (!ret)
2753		pr_warn("Unable to parse numa_balancing=\n");
2754
2755	return ret;
2756}
2757__setup("numa_balancing=", setup_numabalancing);
2758#else
2759static inline void __init check_numabalancing_enable(void)
2760{
2761}
2762#endif /* CONFIG_NUMA_BALANCING */
2763
2764/* assumes fs == KERNEL_DS */
2765void __init numa_policy_init(void)
2766{
2767	nodemask_t interleave_nodes;
2768	unsigned long largest = 0;
2769	int nid, prefer = 0;
2770
2771	policy_cache = kmem_cache_create("numa_policy",
2772					 sizeof(struct mempolicy),
2773					 0, SLAB_PANIC, NULL);
2774
2775	sn_cache = kmem_cache_create("shared_policy_node",
2776				     sizeof(struct sp_node),
2777				     0, SLAB_PANIC, NULL);
2778
2779	for_each_node(nid) {
2780		preferred_node_policy[nid] = (struct mempolicy) {
2781			.refcnt = ATOMIC_INIT(1),
2782			.mode = MPOL_PREFERRED,
2783			.flags = MPOL_F_MOF | MPOL_F_MORON,
2784			.nodes = nodemask_of_node(nid),
2785		};
2786	}
2787
2788	/*
2789	 * Set interleaving policy for system init. Interleaving is only
2790	 * enabled across suitably sized nodes (default is >= 16MB), or
2791	 * fall back to the largest node if they're all smaller.
2792	 */
2793	nodes_clear(interleave_nodes);
2794	for_each_node_state(nid, N_MEMORY) {
2795		unsigned long total_pages = node_present_pages(nid);
2796
2797		/* Preserve the largest node */
2798		if (largest < total_pages) {
2799			largest = total_pages;
2800			prefer = nid;
2801		}
2802
2803		/* Interleave this node? */
2804		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2805			node_set(nid, interleave_nodes);
2806	}
2807
2808	/* All too small, use the largest */
2809	if (unlikely(nodes_empty(interleave_nodes)))
2810		node_set(prefer, interleave_nodes);
2811
2812	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2813		pr_err("%s: interleaving failed\n", __func__);
2814
2815	check_numabalancing_enable();
2816}
2817
2818/* Reset policy of current process to default */
2819void numa_default_policy(void)
2820{
2821	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2822}
2823
2824/*
2825 * Parse and format mempolicy from/to strings
2826 */
2827
2828static const char * const policy_modes[] =
2829{
2830	[MPOL_DEFAULT]    = "default",
2831	[MPOL_PREFERRED]  = "prefer",
2832	[MPOL_BIND]       = "bind",
2833	[MPOL_INTERLEAVE] = "interleave",
2834	[MPOL_LOCAL]      = "local",
 
2835};
2836
2837
2838#ifdef CONFIG_TMPFS
2839/**
2840 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2841 * @str:  string containing mempolicy to parse
2842 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2843 *
2844 * Format of input:
2845 *	<mode>[=<flags>][:<nodelist>]
2846 *
2847 * On success, returns 0, else 1
2848 */
2849int mpol_parse_str(char *str, struct mempolicy **mpol)
2850{
2851	struct mempolicy *new = NULL;
2852	unsigned short mode_flags;
2853	nodemask_t nodes;
2854	char *nodelist = strchr(str, ':');
2855	char *flags = strchr(str, '=');
2856	int err = 1, mode;
2857
2858	if (flags)
2859		*flags++ = '\0';	/* terminate mode string */
2860
2861	if (nodelist) {
2862		/* NUL-terminate mode or flags string */
2863		*nodelist++ = '\0';
2864		if (nodelist_parse(nodelist, nodes))
2865			goto out;
2866		if (!nodes_subset(nodes, node_states[N_MEMORY]))
2867			goto out;
2868	} else
2869		nodes_clear(nodes);
2870
2871	mode = match_string(policy_modes, MPOL_MAX, str);
2872	if (mode < 0)
2873		goto out;
2874
2875	switch (mode) {
2876	case MPOL_PREFERRED:
2877		/*
2878		 * Insist on a nodelist of one node only, although later
2879		 * we use first_node(nodes) to grab a single node, so here
2880		 * nodelist (or nodes) cannot be empty.
2881		 */
2882		if (nodelist) {
2883			char *rest = nodelist;
2884			while (isdigit(*rest))
2885				rest++;
2886			if (*rest)
2887				goto out;
2888			if (nodes_empty(nodes))
2889				goto out;
2890		}
2891		break;
2892	case MPOL_INTERLEAVE:
2893		/*
2894		 * Default to online nodes with memory if no nodelist
2895		 */
2896		if (!nodelist)
2897			nodes = node_states[N_MEMORY];
2898		break;
2899	case MPOL_LOCAL:
2900		/*
2901		 * Don't allow a nodelist;  mpol_new() checks flags
2902		 */
2903		if (nodelist)
2904			goto out;
2905		break;
2906	case MPOL_DEFAULT:
2907		/*
2908		 * Insist on a empty nodelist
2909		 */
2910		if (!nodelist)
2911			err = 0;
2912		goto out;
 
2913	case MPOL_BIND:
2914		/*
2915		 * Insist on a nodelist
2916		 */
2917		if (!nodelist)
2918			goto out;
2919	}
2920
2921	mode_flags = 0;
2922	if (flags) {
2923		/*
2924		 * Currently, we only support two mutually exclusive
2925		 * mode flags.
2926		 */
2927		if (!strcmp(flags, "static"))
2928			mode_flags |= MPOL_F_STATIC_NODES;
2929		else if (!strcmp(flags, "relative"))
2930			mode_flags |= MPOL_F_RELATIVE_NODES;
2931		else
2932			goto out;
2933	}
2934
2935	new = mpol_new(mode, mode_flags, &nodes);
2936	if (IS_ERR(new))
2937		goto out;
2938
2939	/*
2940	 * Save nodes for mpol_to_str() to show the tmpfs mount options
2941	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2942	 */
2943	if (mode != MPOL_PREFERRED) {
2944		new->nodes = nodes;
2945	} else if (nodelist) {
2946		nodes_clear(new->nodes);
2947		node_set(first_node(nodes), new->nodes);
2948	} else {
2949		new->mode = MPOL_LOCAL;
2950	}
2951
2952	/*
2953	 * Save nodes for contextualization: this will be used to "clone"
2954	 * the mempolicy in a specific context [cpuset] at a later time.
2955	 */
2956	new->w.user_nodemask = nodes;
2957
2958	err = 0;
2959
2960out:
2961	/* Restore string for error message */
2962	if (nodelist)
2963		*--nodelist = ':';
2964	if (flags)
2965		*--flags = '=';
2966	if (!err)
2967		*mpol = new;
2968	return err;
2969}
2970#endif /* CONFIG_TMPFS */
2971
2972/**
2973 * mpol_to_str - format a mempolicy structure for printing
2974 * @buffer:  to contain formatted mempolicy string
2975 * @maxlen:  length of @buffer
2976 * @pol:  pointer to mempolicy to be formatted
2977 *
2978 * Convert @pol into a string.  If @buffer is too short, truncate the string.
2979 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2980 * longest flag, "relative", and to display at least a few node ids.
2981 */
2982void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2983{
2984	char *p = buffer;
2985	nodemask_t nodes = NODE_MASK_NONE;
2986	unsigned short mode = MPOL_DEFAULT;
2987	unsigned short flags = 0;
2988
2989	if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
2990		mode = pol->mode;
2991		flags = pol->flags;
2992	}
2993
2994	switch (mode) {
2995	case MPOL_DEFAULT:
2996	case MPOL_LOCAL:
2997		break;
2998	case MPOL_PREFERRED:
 
2999	case MPOL_BIND:
3000	case MPOL_INTERLEAVE:
3001		nodes = pol->nodes;
3002		break;
3003	default:
3004		WARN_ON_ONCE(1);
3005		snprintf(p, maxlen, "unknown");
3006		return;
3007	}
3008
3009	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3010
3011	if (flags & MPOL_MODE_FLAGS) {
3012		p += snprintf(p, buffer + maxlen - p, "=");
3013
3014		/*
3015		 * Currently, the only defined flags are mutually exclusive
3016		 */
3017		if (flags & MPOL_F_STATIC_NODES)
3018			p += snprintf(p, buffer + maxlen - p, "static");
3019		else if (flags & MPOL_F_RELATIVE_NODES)
3020			p += snprintf(p, buffer + maxlen - p, "relative");
3021	}
3022
3023	if (!nodes_empty(nodes))
3024		p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3025			       nodemask_pr_args(&nodes));
3026}
v6.2
   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 * preferred many Try a set of nodes first before normal fallback. This is
  35 *                similar to preferred without the special case.
  36 *
  37 * default        Allocate on the local node first, or when on a VMA
  38 *                use the process policy. This is what Linux always did
  39 *		  in a NUMA aware kernel and still does by, ahem, default.
  40 *
  41 * The process policy is applied for most non interrupt memory allocations
  42 * in that process' context. Interrupts ignore the policies and always
  43 * try to allocate on the local CPU. The VMA policy is only applied for memory
  44 * allocations for a VMA in the VM.
  45 *
  46 * Currently there are a few corner cases in swapping where the policy
  47 * is not applied, but the majority should be handled. When process policy
  48 * is used it is not remembered over swap outs/swap ins.
  49 *
  50 * Only the highest zone in the zone hierarchy gets policied. Allocations
  51 * requesting a lower zone just use default policy. This implies that
  52 * on systems with highmem kernel lowmem allocation don't get policied.
  53 * Same with GFP_DMA allocations.
  54 *
  55 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
  56 * all users and remembered even when nobody has memory mapped.
  57 */
  58
  59/* Notebook:
  60   fix mmap readahead to honour policy and enable policy for any page cache
  61   object
  62   statistics for bigpages
  63   global policy for page cache? currently it uses process policy. Requires
  64   first item above.
  65   handle mremap for shared memory (currently ignored for the policy)
  66   grows down?
  67   make bind policy root only? It can trigger oom much faster and the
  68   kernel is not always grateful with that.
  69*/
  70
  71#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  72
  73#include <linux/mempolicy.h>
  74#include <linux/pagewalk.h>
  75#include <linux/highmem.h>
  76#include <linux/hugetlb.h>
  77#include <linux/kernel.h>
  78#include <linux/sched.h>
  79#include <linux/sched/mm.h>
  80#include <linux/sched/numa_balancing.h>
  81#include <linux/sched/task.h>
  82#include <linux/nodemask.h>
  83#include <linux/cpuset.h>
  84#include <linux/slab.h>
  85#include <linux/string.h>
  86#include <linux/export.h>
  87#include <linux/nsproxy.h>
  88#include <linux/interrupt.h>
  89#include <linux/init.h>
  90#include <linux/compat.h>
  91#include <linux/ptrace.h>
  92#include <linux/swap.h>
  93#include <linux/seq_file.h>
  94#include <linux/proc_fs.h>
  95#include <linux/migrate.h>
  96#include <linux/ksm.h>
  97#include <linux/rmap.h>
  98#include <linux/security.h>
  99#include <linux/syscalls.h>
 100#include <linux/ctype.h>
 101#include <linux/mm_inline.h>
 102#include <linux/mmu_notifier.h>
 103#include <linux/printk.h>
 104#include <linux/swapops.h>
 105
 106#include <asm/tlbflush.h>
 107#include <asm/tlb.h>
 108#include <linux/uaccess.h>
 109
 110#include "internal.h"
 111
 112/* Internal flags */
 113#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)	/* Skip checks for continuous vmas */
 114#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)		/* Invert check for nodemask */
 115
 116static struct kmem_cache *policy_cache;
 117static struct kmem_cache *sn_cache;
 118
 119/* Highest zone. An specific allocation for a zone below that is not
 120   policied. */
 121enum zone_type policy_zone = 0;
 122
 123/*
 124 * run-time system-wide default policy => local allocation
 125 */
 126static struct mempolicy default_policy = {
 127	.refcnt = ATOMIC_INIT(1), /* never free it */
 128	.mode = MPOL_LOCAL,
 129};
 130
 131static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 132
 133/**
 134 * numa_map_to_online_node - Find closest online node
 135 * @node: Node id to start the search
 136 *
 137 * Lookup the next closest node by distance if @nid is not online.
 138 *
 139 * Return: this @node if it is online, otherwise the closest node by distance
 140 */
 141int numa_map_to_online_node(int node)
 142{
 143	int min_dist = INT_MAX, dist, n, min_node;
 144
 145	if (node == NUMA_NO_NODE || node_online(node))
 146		return node;
 147
 148	min_node = node;
 149	for_each_online_node(n) {
 150		dist = node_distance(node, n);
 151		if (dist < min_dist) {
 152			min_dist = dist;
 153			min_node = n;
 154		}
 155	}
 156
 157	return min_node;
 158}
 159EXPORT_SYMBOL_GPL(numa_map_to_online_node);
 160
 161struct mempolicy *get_task_policy(struct task_struct *p)
 162{
 163	struct mempolicy *pol = p->mempolicy;
 164	int node;
 165
 166	if (pol)
 167		return pol;
 168
 169	node = numa_node_id();
 170	if (node != NUMA_NO_NODE) {
 171		pol = &preferred_node_policy[node];
 172		/* preferred_node_policy is not initialised early in boot */
 173		if (pol->mode)
 174			return pol;
 175	}
 176
 177	return &default_policy;
 178}
 179
 180static const struct mempolicy_operations {
 181	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 182	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
 183} mpol_ops[MPOL_MAX];
 184
 185static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 186{
 187	return pol->flags & MPOL_MODE_FLAGS;
 188}
 189
 190static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 191				   const nodemask_t *rel)
 192{
 193	nodemask_t tmp;
 194	nodes_fold(tmp, *orig, nodes_weight(*rel));
 195	nodes_onto(*ret, tmp, *rel);
 196}
 197
 198static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 199{
 200	if (nodes_empty(*nodes))
 201		return -EINVAL;
 202	pol->nodes = *nodes;
 203	return 0;
 204}
 205
 206static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 207{
 208	if (nodes_empty(*nodes))
 209		return -EINVAL;
 210
 211	nodes_clear(pol->nodes);
 212	node_set(first_node(*nodes), pol->nodes);
 213	return 0;
 214}
 215
 
 
 
 
 
 
 
 
 216/*
 217 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 218 * any, for the new policy.  mpol_new() has already validated the nodes
 219 * parameter with respect to the policy mode and flags.
 220 *
 221 * Must be called holding task's alloc_lock to protect task's mems_allowed
 222 * and mempolicy.  May also be called holding the mmap_lock for write.
 223 */
 224static int mpol_set_nodemask(struct mempolicy *pol,
 225		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 226{
 227	int ret;
 228
 229	/*
 230	 * Default (pol==NULL) resp. local memory policies are not a
 231	 * subject of any remapping. They also do not need any special
 232	 * constructor.
 233	 */
 234	if (!pol || pol->mode == MPOL_LOCAL)
 235		return 0;
 236
 237	/* Check N_MEMORY */
 238	nodes_and(nsc->mask1,
 239		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 240
 241	VM_BUG_ON(!nodes);
 242
 243	if (pol->flags & MPOL_F_RELATIVE_NODES)
 244		mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
 245	else
 246		nodes_and(nsc->mask2, *nodes, nsc->mask1);
 247
 248	if (mpol_store_user_nodemask(pol))
 249		pol->w.user_nodemask = *nodes;
 250	else
 251		pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
 252
 253	ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 254	return ret;
 255}
 256
 257/*
 258 * This function just creates a new policy, does some check and simple
 259 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 260 */
 261static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 262				  nodemask_t *nodes)
 263{
 264	struct mempolicy *policy;
 265
 266	pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 267		 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
 268
 269	if (mode == MPOL_DEFAULT) {
 270		if (nodes && !nodes_empty(*nodes))
 271			return ERR_PTR(-EINVAL);
 272		return NULL;
 273	}
 274	VM_BUG_ON(!nodes);
 275
 276	/*
 277	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 278	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 279	 * All other modes require a valid pointer to a non-empty nodemask.
 280	 */
 281	if (mode == MPOL_PREFERRED) {
 282		if (nodes_empty(*nodes)) {
 283			if (((flags & MPOL_F_STATIC_NODES) ||
 284			     (flags & MPOL_F_RELATIVE_NODES)))
 285				return ERR_PTR(-EINVAL);
 286
 287			mode = MPOL_LOCAL;
 288		}
 289	} else if (mode == MPOL_LOCAL) {
 290		if (!nodes_empty(*nodes) ||
 291		    (flags & MPOL_F_STATIC_NODES) ||
 292		    (flags & MPOL_F_RELATIVE_NODES))
 293			return ERR_PTR(-EINVAL);
 294	} else if (nodes_empty(*nodes))
 295		return ERR_PTR(-EINVAL);
 296	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 297	if (!policy)
 298		return ERR_PTR(-ENOMEM);
 299	atomic_set(&policy->refcnt, 1);
 300	policy->mode = mode;
 301	policy->flags = flags;
 302	policy->home_node = NUMA_NO_NODE;
 303
 304	return policy;
 305}
 306
 307/* Slow path of a mpol destructor. */
 308void __mpol_put(struct mempolicy *p)
 309{
 310	if (!atomic_dec_and_test(&p->refcnt))
 311		return;
 312	kmem_cache_free(policy_cache, p);
 313}
 314
 315static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
 316{
 317}
 318
 319static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 320{
 321	nodemask_t tmp;
 322
 323	if (pol->flags & MPOL_F_STATIC_NODES)
 324		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 325	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 326		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 327	else {
 328		nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
 329								*nodes);
 330		pol->w.cpuset_mems_allowed = *nodes;
 331	}
 332
 333	if (nodes_empty(tmp))
 334		tmp = *nodes;
 335
 336	pol->nodes = tmp;
 337}
 338
 339static void mpol_rebind_preferred(struct mempolicy *pol,
 340						const nodemask_t *nodes)
 341{
 342	pol->w.cpuset_mems_allowed = *nodes;
 343}
 344
 345/*
 346 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 347 *
 348 * Per-vma policies are protected by mmap_lock. Allocations using per-task
 349 * policies are protected by task->mems_allowed_seq to prevent a premature
 350 * OOM/allocation failure due to parallel nodemask modification.
 351 */
 352static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
 353{
 354	if (!pol || pol->mode == MPOL_LOCAL)
 355		return;
 356	if (!mpol_store_user_nodemask(pol) &&
 357	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 358		return;
 359
 360	mpol_ops[pol->mode].rebind(pol, newmask);
 361}
 362
 363/*
 364 * Wrapper for mpol_rebind_policy() that just requires task
 365 * pointer, and updates task mempolicy.
 366 *
 367 * Called with task's alloc_lock held.
 368 */
 369
 370void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
 371{
 372	mpol_rebind_policy(tsk->mempolicy, new);
 373}
 374
 375/*
 376 * Rebind each vma in mm to new nodemask.
 377 *
 378 * Call holding a reference to mm.  Takes mm->mmap_lock during call.
 379 */
 380
 381void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 382{
 383	struct vm_area_struct *vma;
 384	VMA_ITERATOR(vmi, mm, 0);
 385
 386	mmap_write_lock(mm);
 387	for_each_vma(vmi, vma)
 388		mpol_rebind_policy(vma->vm_policy, new);
 389	mmap_write_unlock(mm);
 390}
 391
 392static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 393	[MPOL_DEFAULT] = {
 394		.rebind = mpol_rebind_default,
 395	},
 396	[MPOL_INTERLEAVE] = {
 397		.create = mpol_new_nodemask,
 398		.rebind = mpol_rebind_nodemask,
 399	},
 400	[MPOL_PREFERRED] = {
 401		.create = mpol_new_preferred,
 402		.rebind = mpol_rebind_preferred,
 403	},
 404	[MPOL_BIND] = {
 405		.create = mpol_new_nodemask,
 406		.rebind = mpol_rebind_nodemask,
 407	},
 408	[MPOL_LOCAL] = {
 409		.rebind = mpol_rebind_default,
 410	},
 411	[MPOL_PREFERRED_MANY] = {
 412		.create = mpol_new_nodemask,
 413		.rebind = mpol_rebind_preferred,
 414	},
 415};
 416
 417static int migrate_page_add(struct page *page, struct list_head *pagelist,
 418				unsigned long flags);
 419
 420struct queue_pages {
 421	struct list_head *pagelist;
 422	unsigned long flags;
 423	nodemask_t *nmask;
 424	unsigned long start;
 425	unsigned long end;
 426	struct vm_area_struct *first;
 427};
 428
 429/*
 430 * Check if the page's nid is in qp->nmask.
 431 *
 432 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
 433 * in the invert of qp->nmask.
 434 */
 435static inline bool queue_pages_required(struct page *page,
 436					struct queue_pages *qp)
 437{
 438	int nid = page_to_nid(page);
 439	unsigned long flags = qp->flags;
 440
 441	return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
 442}
 443
 444/*
 445 * queue_pages_pmd() has three possible return values:
 446 * 0 - pages are placed on the right node or queued successfully, or
 447 *     special page is met, i.e. huge zero page.
 448 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 449 *     specified.
 
 450 * -EIO - is migration entry or only MPOL_MF_STRICT was specified and an
 451 *        existing page was already on a node that does not follow the
 452 *        policy.
 453 */
 454static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
 455				unsigned long end, struct mm_walk *walk)
 456	__releases(ptl)
 457{
 458	int ret = 0;
 459	struct page *page;
 460	struct queue_pages *qp = walk->private;
 461	unsigned long flags;
 462
 463	if (unlikely(is_pmd_migration_entry(*pmd))) {
 464		ret = -EIO;
 465		goto unlock;
 466	}
 467	page = pmd_page(*pmd);
 468	if (is_huge_zero_page(page)) {
 
 469		walk->action = ACTION_CONTINUE;
 470		goto unlock;
 471	}
 472	if (!queue_pages_required(page, qp))
 473		goto unlock;
 474
 475	flags = qp->flags;
 476	/* go to thp migration */
 477	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 478		if (!vma_migratable(walk->vma) ||
 479		    migrate_page_add(page, qp->pagelist, flags)) {
 480			ret = 1;
 481			goto unlock;
 482		}
 483	} else
 484		ret = -EIO;
 485unlock:
 486	spin_unlock(ptl);
 
 487	return ret;
 488}
 489
 490/*
 491 * Scan through pages checking if pages follow certain conditions,
 492 * and move them to the pagelist if they do.
 493 *
 494 * queue_pages_pte_range() has three possible return values:
 495 * 0 - pages are placed on the right node or queued successfully, or
 496 *     special page is met, i.e. zero page.
 497 * 1 - there is unmovable page, and MPOL_MF_MOVE* & MPOL_MF_STRICT were
 498 *     specified.
 499 * -EIO - only MPOL_MF_STRICT was specified and an existing page was already
 500 *        on a node that does not follow the policy.
 501 */
 502static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
 503			unsigned long end, struct mm_walk *walk)
 504{
 505	struct vm_area_struct *vma = walk->vma;
 506	struct page *page;
 507	struct queue_pages *qp = walk->private;
 508	unsigned long flags = qp->flags;
 
 509	bool has_unmovable = false;
 510	pte_t *pte, *mapped_pte;
 511	spinlock_t *ptl;
 512
 513	ptl = pmd_trans_huge_lock(pmd, vma);
 514	if (ptl)
 515		return queue_pages_pmd(pmd, ptl, addr, end, walk);
 
 
 
 
 516
 517	if (pmd_trans_unstable(pmd))
 518		return 0;
 519
 520	mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
 521	for (; addr != end; pte++, addr += PAGE_SIZE) {
 522		if (!pte_present(*pte))
 523			continue;
 524		page = vm_normal_page(vma, addr, *pte);
 525		if (!page || is_zone_device_page(page))
 526			continue;
 527		/*
 528		 * vm_normal_page() filters out zero pages, but there might
 529		 * still be PageReserved pages to skip, perhaps in a VDSO.
 530		 */
 531		if (PageReserved(page))
 532			continue;
 533		if (!queue_pages_required(page, qp))
 534			continue;
 535		if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
 536			/* MPOL_MF_STRICT must be specified if we get here */
 537			if (!vma_migratable(vma)) {
 538				has_unmovable = true;
 539				break;
 540			}
 541
 542			/*
 543			 * Do not abort immediately since there may be
 544			 * temporary off LRU pages in the range.  Still
 545			 * need migrate other LRU pages.
 546			 */
 547			if (migrate_page_add(page, qp->pagelist, flags))
 548				has_unmovable = true;
 549		} else
 550			break;
 551	}
 552	pte_unmap_unlock(mapped_pte, ptl);
 553	cond_resched();
 554
 555	if (has_unmovable)
 556		return 1;
 557
 558	return addr != end ? -EIO : 0;
 559}
 560
 561static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
 562			       unsigned long addr, unsigned long end,
 563			       struct mm_walk *walk)
 564{
 565	int ret = 0;
 566#ifdef CONFIG_HUGETLB_PAGE
 567	struct queue_pages *qp = walk->private;
 568	unsigned long flags = (qp->flags & MPOL_MF_VALID);
 569	struct page *page;
 570	spinlock_t *ptl;
 571	pte_t entry;
 572
 573	ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 574	entry = huge_ptep_get(pte);
 575	if (!pte_present(entry))
 576		goto unlock;
 577	page = pte_page(entry);
 578	if (!queue_pages_required(page, qp))
 579		goto unlock;
 580
 581	if (flags == MPOL_MF_STRICT) {
 582		/*
 583		 * STRICT alone means only detecting misplaced page and no
 584		 * need to further check other vma.
 585		 */
 586		ret = -EIO;
 587		goto unlock;
 588	}
 589
 590	if (!vma_migratable(walk->vma)) {
 591		/*
 592		 * Must be STRICT with MOVE*, otherwise .test_walk() have
 593		 * stopped walking current vma.
 594		 * Detecting misplaced page but allow migrating pages which
 595		 * have been queued.
 596		 */
 597		ret = 1;
 598		goto unlock;
 599	}
 600
 601	/* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
 602	if (flags & (MPOL_MF_MOVE_ALL) ||
 603	    (flags & MPOL_MF_MOVE && page_mapcount(page) == 1 &&
 604	     !hugetlb_pmd_shared(pte))) {
 605		if (isolate_hugetlb(page, qp->pagelist) &&
 606			(flags & MPOL_MF_STRICT))
 607			/*
 608			 * Failed to isolate page but allow migrating pages
 609			 * which have been queued.
 610			 */
 611			ret = 1;
 612	}
 613unlock:
 614	spin_unlock(ptl);
 615#else
 616	BUG();
 617#endif
 618	return ret;
 619}
 620
 621#ifdef CONFIG_NUMA_BALANCING
 622/*
 623 * This is used to mark a range of virtual addresses to be inaccessible.
 624 * These are later cleared by a NUMA hinting fault. Depending on these
 625 * faults, pages may be migrated for better NUMA placement.
 626 *
 627 * This is assuming that NUMA faults are handled using PROT_NONE. If
 628 * an architecture makes a different choice, it will need further
 629 * changes to the core.
 630 */
 631unsigned long change_prot_numa(struct vm_area_struct *vma,
 632			unsigned long addr, unsigned long end)
 633{
 634	struct mmu_gather tlb;
 635	int nr_updated;
 636
 637	tlb_gather_mmu(&tlb, vma->vm_mm);
 638
 639	nr_updated = change_protection(&tlb, vma, addr, end, PAGE_NONE,
 640				       MM_CP_PROT_NUMA);
 641	if (nr_updated)
 642		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 643
 644	tlb_finish_mmu(&tlb);
 645
 646	return nr_updated;
 647}
 648#else
 649static unsigned long change_prot_numa(struct vm_area_struct *vma,
 650			unsigned long addr, unsigned long end)
 651{
 652	return 0;
 653}
 654#endif /* CONFIG_NUMA_BALANCING */
 655
 656static int queue_pages_test_walk(unsigned long start, unsigned long end,
 657				struct mm_walk *walk)
 658{
 659	struct vm_area_struct *next, *vma = walk->vma;
 660	struct queue_pages *qp = walk->private;
 661	unsigned long endvma = vma->vm_end;
 662	unsigned long flags = qp->flags;
 663
 664	/* range check first */
 665	VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
 666
 667	if (!qp->first) {
 668		qp->first = vma;
 669		if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 670			(qp->start < vma->vm_start))
 671			/* hole at head side of range */
 672			return -EFAULT;
 673	}
 674	next = find_vma(vma->vm_mm, vma->vm_end);
 675	if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 676		((vma->vm_end < qp->end) &&
 677		(!next || vma->vm_end < next->vm_start)))
 678		/* hole at middle or tail of range */
 679		return -EFAULT;
 680
 681	/*
 682	 * Need check MPOL_MF_STRICT to return -EIO if possible
 683	 * regardless of vma_migratable
 684	 */
 685	if (!vma_migratable(vma) &&
 686	    !(flags & MPOL_MF_STRICT))
 687		return 1;
 688
 689	if (endvma > end)
 690		endvma = end;
 691
 692	if (flags & MPOL_MF_LAZY) {
 693		/* Similar to task_numa_work, skip inaccessible VMAs */
 694		if (!is_vm_hugetlb_page(vma) && vma_is_accessible(vma) &&
 695			!(vma->vm_flags & VM_MIXEDMAP))
 696			change_prot_numa(vma, start, endvma);
 697		return 1;
 698	}
 699
 700	/* queue pages from current vma */
 701	if (flags & MPOL_MF_VALID)
 702		return 0;
 703	return 1;
 704}
 705
 706static const struct mm_walk_ops queue_pages_walk_ops = {
 707	.hugetlb_entry		= queue_pages_hugetlb,
 708	.pmd_entry		= queue_pages_pte_range,
 709	.test_walk		= queue_pages_test_walk,
 710};
 711
 712/*
 713 * Walk through page tables and collect pages to be migrated.
 714 *
 715 * If pages found in a given range are on a set of nodes (determined by
 716 * @nodes and @flags,) it's isolated and queued to the pagelist which is
 717 * passed via @private.
 718 *
 719 * queue_pages_range() has three possible return values:
 720 * 1 - there is unmovable page, but MPOL_MF_MOVE* & MPOL_MF_STRICT were
 721 *     specified.
 722 * 0 - queue pages successfully or no misplaced page.
 723 * errno - i.e. misplaced pages with MPOL_MF_STRICT specified (-EIO) or
 724 *         memory range specified by nodemask and maxnode points outside
 725 *         your accessible address space (-EFAULT)
 726 */
 727static int
 728queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 729		nodemask_t *nodes, unsigned long flags,
 730		struct list_head *pagelist)
 731{
 732	int err;
 733	struct queue_pages qp = {
 734		.pagelist = pagelist,
 735		.flags = flags,
 736		.nmask = nodes,
 737		.start = start,
 738		.end = end,
 739		.first = NULL,
 740	};
 741
 742	err = walk_page_range(mm, start, end, &queue_pages_walk_ops, &qp);
 743
 744	if (!qp.first)
 745		/* whole range in hole */
 746		err = -EFAULT;
 747
 748	return err;
 749}
 750
 751/*
 752 * Apply policy to a single VMA
 753 * This must be called with the mmap_lock held for writing.
 754 */
 755static int vma_replace_policy(struct vm_area_struct *vma,
 756						struct mempolicy *pol)
 757{
 758	int err;
 759	struct mempolicy *old;
 760	struct mempolicy *new;
 761
 762	pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 763		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 764		 vma->vm_ops, vma->vm_file,
 765		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 766
 767	new = mpol_dup(pol);
 768	if (IS_ERR(new))
 769		return PTR_ERR(new);
 770
 771	if (vma->vm_ops && vma->vm_ops->set_policy) {
 772		err = vma->vm_ops->set_policy(vma, new);
 773		if (err)
 774			goto err_out;
 775	}
 776
 777	old = vma->vm_policy;
 778	vma->vm_policy = new; /* protected by mmap_lock */
 779	mpol_put(old);
 780
 781	return 0;
 782 err_out:
 783	mpol_put(new);
 784	return err;
 785}
 786
 787/* Step 2: apply policy to a range and do splits. */
 788static int mbind_range(struct mm_struct *mm, unsigned long start,
 789		       unsigned long end, struct mempolicy *new_pol)
 790{
 791	MA_STATE(mas, &mm->mm_mt, start, start);
 792	struct vm_area_struct *prev;
 793	struct vm_area_struct *vma;
 794	int err = 0;
 795	pgoff_t pgoff;
 
 
 796
 797	prev = mas_prev(&mas, 0);
 798	if (unlikely(!prev))
 799		mas_set(&mas, start);
 800
 801	vma = mas_find(&mas, end - 1);
 802	if (WARN_ON(!vma))
 803		return 0;
 804
 
 805	if (start > vma->vm_start)
 806		prev = vma;
 807
 808	for (; vma; vma = mas_next(&mas, end - 1)) {
 809		unsigned long vmstart = max(start, vma->vm_start);
 810		unsigned long vmend = min(end, vma->vm_end);
 
 811
 812		if (mpol_equal(vma_policy(vma), new_pol))
 813			goto next;
 814
 815		pgoff = vma->vm_pgoff +
 816			((vmstart - vma->vm_start) >> PAGE_SHIFT);
 817		prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 818				 vma->anon_vma, vma->vm_file, pgoff,
 819				 new_pol, vma->vm_userfaultfd_ctx,
 820				 anon_vma_name(vma));
 821		if (prev) {
 822			/* vma_merge() invalidated the mas */
 823			mas_pause(&mas);
 824			vma = prev;
 
 
 
 
 825			goto replace;
 826		}
 827		if (vma->vm_start != vmstart) {
 828			err = split_vma(vma->vm_mm, vma, vmstart, 1);
 829			if (err)
 830				goto out;
 831			/* split_vma() invalidated the mas */
 832			mas_pause(&mas);
 833		}
 834		if (vma->vm_end != vmend) {
 835			err = split_vma(vma->vm_mm, vma, vmend, 0);
 836			if (err)
 837				goto out;
 838			/* split_vma() invalidated the mas */
 839			mas_pause(&mas);
 840		}
 841replace:
 842		err = vma_replace_policy(vma, new_pol);
 843		if (err)
 844			goto out;
 845next:
 846		prev = vma;
 847	}
 848
 849out:
 850	return err;
 851}
 852
 853/* Set the process memory policy */
 854static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 855			     nodemask_t *nodes)
 856{
 857	struct mempolicy *new, *old;
 858	NODEMASK_SCRATCH(scratch);
 859	int ret;
 860
 861	if (!scratch)
 862		return -ENOMEM;
 863
 864	new = mpol_new(mode, flags, nodes);
 865	if (IS_ERR(new)) {
 866		ret = PTR_ERR(new);
 867		goto out;
 868	}
 869
 870	task_lock(current);
 871	ret = mpol_set_nodemask(new, nodes, scratch);
 872	if (ret) {
 873		task_unlock(current);
 874		mpol_put(new);
 875		goto out;
 876	}
 877
 878	old = current->mempolicy;
 879	current->mempolicy = new;
 880	if (new && new->mode == MPOL_INTERLEAVE)
 881		current->il_prev = MAX_NUMNODES-1;
 882	task_unlock(current);
 883	mpol_put(old);
 884	ret = 0;
 885out:
 886	NODEMASK_SCRATCH_FREE(scratch);
 887	return ret;
 888}
 889
 890/*
 891 * Return nodemask for policy for get_mempolicy() query
 892 *
 893 * Called with task's alloc_lock held
 894 */
 895static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 896{
 897	nodes_clear(*nodes);
 898	if (p == &default_policy)
 899		return;
 900
 901	switch (p->mode) {
 902	case MPOL_BIND:
 903	case MPOL_INTERLEAVE:
 904	case MPOL_PREFERRED:
 905	case MPOL_PREFERRED_MANY:
 906		*nodes = p->nodes;
 907		break;
 908	case MPOL_LOCAL:
 909		/* return empty node mask for local allocation */
 910		break;
 911	default:
 912		BUG();
 913	}
 914}
 915
 916static int lookup_node(struct mm_struct *mm, unsigned long addr)
 917{
 918	struct page *p = NULL;
 919	int ret;
 920
 921	ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
 922	if (ret > 0) {
 923		ret = page_to_nid(p);
 
 924		put_page(p);
 925	}
 926	return ret;
 
 
 927}
 928
 929/* Retrieve NUMA policy */
 930static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 931			     unsigned long addr, unsigned long flags)
 932{
 933	int err;
 934	struct mm_struct *mm = current->mm;
 935	struct vm_area_struct *vma = NULL;
 936	struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
 937
 938	if (flags &
 939		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 940		return -EINVAL;
 941
 942	if (flags & MPOL_F_MEMS_ALLOWED) {
 943		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 944			return -EINVAL;
 945		*policy = 0;	/* just so it's initialized */
 946		task_lock(current);
 947		*nmask  = cpuset_current_mems_allowed;
 948		task_unlock(current);
 949		return 0;
 950	}
 951
 952	if (flags & MPOL_F_ADDR) {
 953		/*
 954		 * Do NOT fall back to task policy if the
 955		 * vma/shared policy at addr is NULL.  We
 956		 * want to return MPOL_DEFAULT in this case.
 957		 */
 958		mmap_read_lock(mm);
 959		vma = vma_lookup(mm, addr);
 960		if (!vma) {
 961			mmap_read_unlock(mm);
 962			return -EFAULT;
 963		}
 964		if (vma->vm_ops && vma->vm_ops->get_policy)
 965			pol = vma->vm_ops->get_policy(vma, addr);
 966		else
 967			pol = vma->vm_policy;
 968	} else if (addr)
 969		return -EINVAL;
 970
 971	if (!pol)
 972		pol = &default_policy;	/* indicates default behavior */
 973
 974	if (flags & MPOL_F_NODE) {
 975		if (flags & MPOL_F_ADDR) {
 976			/*
 977			 * Take a refcount on the mpol, because we are about to
 978			 * drop the mmap_lock, after which only "pol" remains
 979			 * valid, "vma" is stale.
 
 980			 */
 981			pol_refcount = pol;
 982			vma = NULL;
 983			mpol_get(pol);
 984			mmap_read_unlock(mm);
 985			err = lookup_node(mm, addr);
 986			if (err < 0)
 987				goto out;
 988			*policy = err;
 989		} else if (pol == current->mempolicy &&
 990				pol->mode == MPOL_INTERLEAVE) {
 991			*policy = next_node_in(current->il_prev, pol->nodes);
 992		} else {
 993			err = -EINVAL;
 994			goto out;
 995		}
 996	} else {
 997		*policy = pol == &default_policy ? MPOL_DEFAULT :
 998						pol->mode;
 999		/*
1000		 * Internal mempolicy flags must be masked off before exposing
1001		 * the policy to userspace.
1002		 */
1003		*policy |= (pol->flags & MPOL_MODE_FLAGS);
1004	}
1005
1006	err = 0;
1007	if (nmask) {
1008		if (mpol_store_user_nodemask(pol)) {
1009			*nmask = pol->w.user_nodemask;
1010		} else {
1011			task_lock(current);
1012			get_policy_nodemask(pol, nmask);
1013			task_unlock(current);
1014		}
1015	}
1016
1017 out:
1018	mpol_cond_put(pol);
1019	if (vma)
1020		mmap_read_unlock(mm);
1021	if (pol_refcount)
1022		mpol_put(pol_refcount);
1023	return err;
1024}
1025
1026#ifdef CONFIG_MIGRATION
1027/*
1028 * page migration, thp tail pages can be passed.
1029 */
1030static int migrate_page_add(struct page *page, struct list_head *pagelist,
1031				unsigned long flags)
1032{
1033	struct page *head = compound_head(page);
1034	/*
1035	 * Avoid migrating a page that is shared with others.
1036	 */
1037	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
1038		if (!isolate_lru_page(head)) {
1039			list_add_tail(&head->lru, pagelist);
1040			mod_node_page_state(page_pgdat(head),
1041				NR_ISOLATED_ANON + page_is_file_lru(head),
1042				thp_nr_pages(head));
1043		} else if (flags & MPOL_MF_STRICT) {
1044			/*
1045			 * Non-movable page may reach here.  And, there may be
1046			 * temporary off LRU pages or non-LRU movable pages.
1047			 * Treat them as unmovable pages since they can't be
1048			 * isolated, so they can't be moved at the moment.  It
1049			 * should return -EIO for this case too.
1050			 */
1051			return -EIO;
1052		}
1053	}
1054
1055	return 0;
1056}
1057
1058/*
1059 * Migrate pages from one node to a target node.
1060 * Returns error or the number of pages not migrated.
1061 */
1062static int migrate_to_node(struct mm_struct *mm, int source, int dest,
1063			   int flags)
1064{
1065	nodemask_t nmask;
1066	struct vm_area_struct *vma;
1067	LIST_HEAD(pagelist);
1068	int err = 0;
1069	struct migration_target_control mtc = {
1070		.nid = dest,
1071		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1072	};
1073
1074	nodes_clear(nmask);
1075	node_set(source, nmask);
1076
1077	/*
1078	 * This does not "check" the range but isolates all pages that
1079	 * need migration.  Between passing in the full user address
1080	 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
1081	 */
1082	vma = find_vma(mm, 0);
1083	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1084	queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1085			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1086
1087	if (!list_empty(&pagelist)) {
1088		err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1089				(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1090		if (err)
1091			putback_movable_pages(&pagelist);
1092	}
1093
1094	return err;
1095}
1096
1097/*
1098 * Move pages between the two nodesets so as to preserve the physical
1099 * layout as much as possible.
1100 *
1101 * Returns the number of page that could not be moved.
1102 */
1103int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1104		     const nodemask_t *to, int flags)
1105{
1106	int busy = 0;
1107	int err = 0;
1108	nodemask_t tmp;
1109
1110	lru_cache_disable();
1111
1112	mmap_read_lock(mm);
1113
1114	/*
1115	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1116	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
1117	 * bit in 'tmp', and return that <source, dest> pair for migration.
1118	 * The pair of nodemasks 'to' and 'from' define the map.
1119	 *
1120	 * If no pair of bits is found that way, fallback to picking some
1121	 * pair of 'source' and 'dest' bits that are not the same.  If the
1122	 * 'source' and 'dest' bits are the same, this represents a node
1123	 * that will be migrating to itself, so no pages need move.
1124	 *
1125	 * If no bits are left in 'tmp', or if all remaining bits left
1126	 * in 'tmp' correspond to the same bit in 'to', return false
1127	 * (nothing left to migrate).
1128	 *
1129	 * This lets us pick a pair of nodes to migrate between, such that
1130	 * if possible the dest node is not already occupied by some other
1131	 * source node, minimizing the risk of overloading the memory on a
1132	 * node that would happen if we migrated incoming memory to a node
1133	 * before migrating outgoing memory source that same node.
1134	 *
1135	 * A single scan of tmp is sufficient.  As we go, we remember the
1136	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1137	 * that not only moved, but what's better, moved to an empty slot
1138	 * (d is not set in tmp), then we break out then, with that pair.
1139	 * Otherwise when we finish scanning from_tmp, we at least have the
1140	 * most recent <s, d> pair that moved.  If we get all the way through
1141	 * the scan of tmp without finding any node that moved, much less
1142	 * moved to an empty node, then there is nothing left worth migrating.
1143	 */
1144
1145	tmp = *from;
1146	while (!nodes_empty(tmp)) {
1147		int s, d;
1148		int source = NUMA_NO_NODE;
1149		int dest = 0;
1150
1151		for_each_node_mask(s, tmp) {
1152
1153			/*
1154			 * do_migrate_pages() tries to maintain the relative
1155			 * node relationship of the pages established between
1156			 * threads and memory areas.
1157                         *
1158			 * However if the number of source nodes is not equal to
1159			 * the number of destination nodes we can not preserve
1160			 * this node relative relationship.  In that case, skip
1161			 * copying memory from a node that is in the destination
1162			 * mask.
1163			 *
1164			 * Example: [2,3,4] -> [3,4,5] moves everything.
1165			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1166			 */
1167
1168			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1169						(node_isset(s, *to)))
1170				continue;
1171
1172			d = node_remap(s, *from, *to);
1173			if (s == d)
1174				continue;
1175
1176			source = s;	/* Node moved. Memorize */
1177			dest = d;
1178
1179			/* dest not in remaining from nodes? */
1180			if (!node_isset(dest, tmp))
1181				break;
1182		}
1183		if (source == NUMA_NO_NODE)
1184			break;
1185
1186		node_clear(source, tmp);
1187		err = migrate_to_node(mm, source, dest, flags);
1188		if (err > 0)
1189			busy += err;
1190		if (err < 0)
1191			break;
1192	}
1193	mmap_read_unlock(mm);
1194
1195	lru_cache_enable();
1196	if (err < 0)
1197		return err;
1198	return busy;
1199
1200}
1201
1202/*
1203 * Allocate a new page for page migration based on vma policy.
1204 * Start by assuming the page is mapped by the same vma as contains @start.
1205 * Search forward from there, if not.  N.B., this assumes that the
1206 * list of pages handed to migrate_pages()--which is how we get here--
1207 * is in virtual address order.
1208 */
1209static struct page *new_page(struct page *page, unsigned long start)
1210{
1211	struct folio *dst, *src = page_folio(page);
1212	struct vm_area_struct *vma;
1213	unsigned long address;
1214	VMA_ITERATOR(vmi, current->mm, start);
1215	gfp_t gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL;
1216
1217	for_each_vma(vmi, vma) {
 
1218		address = page_address_in_vma(page, vma);
1219		if (address != -EFAULT)
1220			break;
 
1221	}
1222
1223	if (folio_test_hugetlb(src))
1224		return alloc_huge_page_vma(page_hstate(&src->page),
1225				vma, address);
 
 
1226
1227	if (folio_test_large(src))
1228		gfp = GFP_TRANSHUGE;
1229
 
 
 
 
1230	/*
1231	 * if !vma, vma_alloc_folio() will use task or system default policy
1232	 */
1233	dst = vma_alloc_folio(gfp, folio_order(src), vma, address,
1234			folio_test_large(src));
1235	return &dst->page;
1236}
1237#else
1238
1239static int migrate_page_add(struct page *page, struct list_head *pagelist,
1240				unsigned long flags)
1241{
1242	return -EIO;
1243}
1244
1245int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1246		     const nodemask_t *to, int flags)
1247{
1248	return -ENOSYS;
1249}
1250
1251static struct page *new_page(struct page *page, unsigned long start)
1252{
1253	return NULL;
1254}
1255#endif
1256
1257static long do_mbind(unsigned long start, unsigned long len,
1258		     unsigned short mode, unsigned short mode_flags,
1259		     nodemask_t *nmask, unsigned long flags)
1260{
1261	struct mm_struct *mm = current->mm;
1262	struct mempolicy *new;
1263	unsigned long end;
1264	int err;
1265	int ret;
1266	LIST_HEAD(pagelist);
1267
1268	if (flags & ~(unsigned long)MPOL_MF_VALID)
1269		return -EINVAL;
1270	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1271		return -EPERM;
1272
1273	if (start & ~PAGE_MASK)
1274		return -EINVAL;
1275
1276	if (mode == MPOL_DEFAULT)
1277		flags &= ~MPOL_MF_STRICT;
1278
1279	len = PAGE_ALIGN(len);
1280	end = start + len;
1281
1282	if (end < start)
1283		return -EINVAL;
1284	if (end == start)
1285		return 0;
1286
1287	new = mpol_new(mode, mode_flags, nmask);
1288	if (IS_ERR(new))
1289		return PTR_ERR(new);
1290
1291	if (flags & MPOL_MF_LAZY)
1292		new->flags |= MPOL_F_MOF;
1293
1294	/*
1295	 * If we are using the default policy then operation
1296	 * on discontinuous address spaces is okay after all
1297	 */
1298	if (!new)
1299		flags |= MPOL_MF_DISCONTIG_OK;
1300
1301	pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1302		 start, start + len, mode, mode_flags,
1303		 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1304
1305	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1306
1307		lru_cache_disable();
1308	}
1309	{
1310		NODEMASK_SCRATCH(scratch);
1311		if (scratch) {
1312			mmap_write_lock(mm);
1313			err = mpol_set_nodemask(new, nmask, scratch);
1314			if (err)
1315				mmap_write_unlock(mm);
1316		} else
1317			err = -ENOMEM;
1318		NODEMASK_SCRATCH_FREE(scratch);
1319	}
1320	if (err)
1321		goto mpol_out;
1322
1323	ret = queue_pages_range(mm, start, end, nmask,
1324			  flags | MPOL_MF_INVERT, &pagelist);
1325
1326	if (ret < 0) {
1327		err = ret;
1328		goto up_out;
1329	}
1330
1331	err = mbind_range(mm, start, end, new);
1332
1333	if (!err) {
1334		int nr_failed = 0;
1335
1336		if (!list_empty(&pagelist)) {
1337			WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1338			nr_failed = migrate_pages(&pagelist, new_page, NULL,
1339				start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND, NULL);
1340			if (nr_failed)
1341				putback_movable_pages(&pagelist);
1342		}
1343
1344		if ((ret > 0) || (nr_failed && (flags & MPOL_MF_STRICT)))
1345			err = -EIO;
1346	} else {
1347up_out:
1348		if (!list_empty(&pagelist))
1349			putback_movable_pages(&pagelist);
1350	}
1351
1352	mmap_write_unlock(mm);
1353mpol_out:
1354	mpol_put(new);
1355	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1356		lru_cache_enable();
1357	return err;
1358}
1359
1360/*
1361 * User space interface with variable sized bitmaps for nodelists.
1362 */
1363static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1364		      unsigned long maxnode)
1365{
1366	unsigned long nlongs = BITS_TO_LONGS(maxnode);
1367	int ret;
1368
1369	if (in_compat_syscall())
1370		ret = compat_get_bitmap(mask,
1371					(const compat_ulong_t __user *)nmask,
1372					maxnode);
1373	else
1374		ret = copy_from_user(mask, nmask,
1375				     nlongs * sizeof(unsigned long));
1376
1377	if (ret)
1378		return -EFAULT;
1379
1380	if (maxnode % BITS_PER_LONG)
1381		mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1382
1383	return 0;
1384}
1385
1386/* Copy a node mask from user space. */
1387static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1388		     unsigned long maxnode)
1389{
 
 
 
 
 
1390	--maxnode;
1391	nodes_clear(*nodes);
1392	if (maxnode == 0 || !nmask)
1393		return 0;
1394	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1395		return -EINVAL;
1396
 
 
 
 
 
 
1397	/*
1398	 * When the user specified more nodes than supported just check
1399	 * if the non supported part is all zero, one word at a time,
1400	 * starting at the end.
1401	 */
1402	while (maxnode > MAX_NUMNODES) {
1403		unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1404		unsigned long t;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1405
1406		if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
 
1407			return -EFAULT;
1408
1409		if (maxnode - bits >= MAX_NUMNODES) {
1410			maxnode -= bits;
1411		} else {
1412			maxnode = MAX_NUMNODES;
1413			t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1414		}
1415		if (t)
1416			return -EINVAL;
1417	}
1418
1419	return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
 
 
 
1420}
1421
1422/* Copy a kernel node mask to user space */
1423static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1424			      nodemask_t *nodes)
1425{
1426	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1427	unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1428	bool compat = in_compat_syscall();
1429
1430	if (compat)
1431		nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1432
1433	if (copy > nbytes) {
1434		if (copy > PAGE_SIZE)
1435			return -EINVAL;
1436		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1437			return -EFAULT;
1438		copy = nbytes;
1439		maxnode = nr_node_ids;
1440	}
1441
1442	if (compat)
1443		return compat_put_bitmap((compat_ulong_t __user *)mask,
1444					 nodes_addr(*nodes), maxnode);
1445
1446	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1447}
1448
1449/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1450static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1451{
1452	*flags = *mode & MPOL_MODE_FLAGS;
1453	*mode &= ~MPOL_MODE_FLAGS;
1454
1455	if ((unsigned int)(*mode) >=  MPOL_MAX)
1456		return -EINVAL;
1457	if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1458		return -EINVAL;
1459	if (*flags & MPOL_F_NUMA_BALANCING) {
1460		if (*mode != MPOL_BIND)
1461			return -EINVAL;
1462		*flags |= (MPOL_F_MOF | MPOL_F_MORON);
1463	}
1464	return 0;
1465}
1466
1467static long kernel_mbind(unsigned long start, unsigned long len,
1468			 unsigned long mode, const unsigned long __user *nmask,
1469			 unsigned long maxnode, unsigned int flags)
1470{
1471	unsigned short mode_flags;
1472	nodemask_t nodes;
1473	int lmode = mode;
1474	int err;
1475
1476	start = untagged_addr(start);
1477	err = sanitize_mpol_flags(&lmode, &mode_flags);
1478	if (err)
1479		return err;
1480
1481	err = get_nodes(&nodes, nmask, maxnode);
1482	if (err)
1483		return err;
1484
1485	return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1486}
1487
1488SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1489		unsigned long, home_node, unsigned long, flags)
1490{
1491	struct mm_struct *mm = current->mm;
1492	struct vm_area_struct *vma;
1493	struct mempolicy *new;
1494	unsigned long vmstart;
1495	unsigned long vmend;
1496	unsigned long end;
1497	int err = -ENOENT;
1498	VMA_ITERATOR(vmi, mm, start);
1499
1500	start = untagged_addr(start);
1501	if (start & ~PAGE_MASK)
1502		return -EINVAL;
1503	/*
1504	 * flags is used for future extension if any.
1505	 */
1506	if (flags != 0)
1507		return -EINVAL;
1508
1509	/*
1510	 * Check home_node is online to avoid accessing uninitialized
1511	 * NODE_DATA.
1512	 */
1513	if (home_node >= MAX_NUMNODES || !node_online(home_node))
1514		return -EINVAL;
1515
1516	len = PAGE_ALIGN(len);
1517	end = start + len;
1518
1519	if (end < start)
1520		return -EINVAL;
1521	if (end == start)
1522		return 0;
1523	mmap_write_lock(mm);
1524	for_each_vma_range(vmi, vma, end) {
1525		vmstart = max(start, vma->vm_start);
1526		vmend   = min(end, vma->vm_end);
1527		new = mpol_dup(vma_policy(vma));
1528		if (IS_ERR(new)) {
1529			err = PTR_ERR(new);
1530			break;
1531		}
1532		/*
1533		 * Only update home node if there is an existing vma policy
1534		 */
1535		if (!new)
1536			continue;
1537
1538		/*
1539		 * If any vma in the range got policy other than MPOL_BIND
1540		 * or MPOL_PREFERRED_MANY we return error. We don't reset
1541		 * the home node for vmas we already updated before.
1542		 */
1543		if (new->mode != MPOL_BIND && new->mode != MPOL_PREFERRED_MANY) {
1544			mpol_put(new);
1545			err = -EOPNOTSUPP;
1546			break;
1547		}
1548
1549		new->home_node = home_node;
1550		err = mbind_range(mm, vmstart, vmend, new);
1551		mpol_put(new);
1552		if (err)
1553			break;
1554	}
1555	mmap_write_unlock(mm);
1556	return err;
1557}
1558
1559SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1560		unsigned long, mode, const unsigned long __user *, nmask,
1561		unsigned long, maxnode, unsigned int, flags)
1562{
1563	return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1564}
1565
1566/* Set the process memory policy */
1567static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1568				 unsigned long maxnode)
1569{
1570	unsigned short mode_flags;
1571	nodemask_t nodes;
1572	int lmode = mode;
1573	int err;
1574
1575	err = sanitize_mpol_flags(&lmode, &mode_flags);
1576	if (err)
1577		return err;
1578
1579	err = get_nodes(&nodes, nmask, maxnode);
1580	if (err)
1581		return err;
1582
1583	return do_set_mempolicy(lmode, mode_flags, &nodes);
1584}
1585
1586SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1587		unsigned long, maxnode)
1588{
1589	return kernel_set_mempolicy(mode, nmask, maxnode);
1590}
1591
1592static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1593				const unsigned long __user *old_nodes,
1594				const unsigned long __user *new_nodes)
1595{
1596	struct mm_struct *mm = NULL;
1597	struct task_struct *task;
1598	nodemask_t task_nodes;
1599	int err;
1600	nodemask_t *old;
1601	nodemask_t *new;
1602	NODEMASK_SCRATCH(scratch);
1603
1604	if (!scratch)
1605		return -ENOMEM;
1606
1607	old = &scratch->mask1;
1608	new = &scratch->mask2;
1609
1610	err = get_nodes(old, old_nodes, maxnode);
1611	if (err)
1612		goto out;
1613
1614	err = get_nodes(new, new_nodes, maxnode);
1615	if (err)
1616		goto out;
1617
1618	/* Find the mm_struct */
1619	rcu_read_lock();
1620	task = pid ? find_task_by_vpid(pid) : current;
1621	if (!task) {
1622		rcu_read_unlock();
1623		err = -ESRCH;
1624		goto out;
1625	}
1626	get_task_struct(task);
1627
1628	err = -EINVAL;
1629
1630	/*
1631	 * Check if this process has the right to modify the specified process.
1632	 * Use the regular "ptrace_may_access()" checks.
1633	 */
1634	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1635		rcu_read_unlock();
1636		err = -EPERM;
1637		goto out_put;
1638	}
1639	rcu_read_unlock();
1640
1641	task_nodes = cpuset_mems_allowed(task);
1642	/* Is the user allowed to access the target nodes? */
1643	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1644		err = -EPERM;
1645		goto out_put;
1646	}
1647
1648	task_nodes = cpuset_mems_allowed(current);
1649	nodes_and(*new, *new, task_nodes);
1650	if (nodes_empty(*new))
1651		goto out_put;
1652
1653	err = security_task_movememory(task);
1654	if (err)
1655		goto out_put;
1656
1657	mm = get_task_mm(task);
1658	put_task_struct(task);
1659
1660	if (!mm) {
1661		err = -EINVAL;
1662		goto out;
1663	}
1664
1665	err = do_migrate_pages(mm, old, new,
1666		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1667
1668	mmput(mm);
1669out:
1670	NODEMASK_SCRATCH_FREE(scratch);
1671
1672	return err;
1673
1674out_put:
1675	put_task_struct(task);
1676	goto out;
1677
1678}
1679
1680SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1681		const unsigned long __user *, old_nodes,
1682		const unsigned long __user *, new_nodes)
1683{
1684	return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1685}
1686
1687
1688/* Retrieve NUMA policy */
1689static int kernel_get_mempolicy(int __user *policy,
1690				unsigned long __user *nmask,
1691				unsigned long maxnode,
1692				unsigned long addr,
1693				unsigned long flags)
1694{
1695	int err;
1696	int pval;
1697	nodemask_t nodes;
1698
1699	if (nmask != NULL && maxnode < nr_node_ids)
1700		return -EINVAL;
1701
1702	addr = untagged_addr(addr);
1703
1704	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1705
1706	if (err)
1707		return err;
1708
1709	if (policy && put_user(pval, policy))
1710		return -EFAULT;
1711
1712	if (nmask)
1713		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1714
1715	return err;
1716}
1717
1718SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1719		unsigned long __user *, nmask, unsigned long, maxnode,
1720		unsigned long, addr, unsigned long, flags)
1721{
1722	return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1723}
1724
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1725bool vma_migratable(struct vm_area_struct *vma)
1726{
1727	if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1728		return false;
1729
1730	/*
1731	 * DAX device mappings require predictable access latency, so avoid
1732	 * incurring periodic faults.
1733	 */
1734	if (vma_is_dax(vma))
1735		return false;
1736
1737	if (is_vm_hugetlb_page(vma) &&
1738		!hugepage_migration_supported(hstate_vma(vma)))
1739		return false;
1740
1741	/*
1742	 * Migration allocates pages in the highest zone. If we cannot
1743	 * do so then migration (at least from node to node) is not
1744	 * possible.
1745	 */
1746	if (vma->vm_file &&
1747		gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1748			< policy_zone)
1749		return false;
1750	return true;
1751}
1752
1753struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1754						unsigned long addr)
1755{
1756	struct mempolicy *pol = NULL;
1757
1758	if (vma) {
1759		if (vma->vm_ops && vma->vm_ops->get_policy) {
1760			pol = vma->vm_ops->get_policy(vma, addr);
1761		} else if (vma->vm_policy) {
1762			pol = vma->vm_policy;
1763
1764			/*
1765			 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1766			 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1767			 * count on these policies which will be dropped by
1768			 * mpol_cond_put() later
1769			 */
1770			if (mpol_needs_cond_ref(pol))
1771				mpol_get(pol);
1772		}
1773	}
1774
1775	return pol;
1776}
1777
1778/*
1779 * get_vma_policy(@vma, @addr)
1780 * @vma: virtual memory area whose policy is sought
1781 * @addr: address in @vma for shared policy lookup
1782 *
1783 * Returns effective policy for a VMA at specified address.
1784 * Falls back to current->mempolicy or system default policy, as necessary.
1785 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1786 * count--added by the get_policy() vm_op, as appropriate--to protect against
1787 * freeing by another task.  It is the caller's responsibility to free the
1788 * extra reference for shared policies.
1789 */
1790static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1791						unsigned long addr)
1792{
1793	struct mempolicy *pol = __get_vma_policy(vma, addr);
1794
1795	if (!pol)
1796		pol = get_task_policy(current);
1797
1798	return pol;
1799}
1800
1801bool vma_policy_mof(struct vm_area_struct *vma)
1802{
1803	struct mempolicy *pol;
1804
1805	if (vma->vm_ops && vma->vm_ops->get_policy) {
1806		bool ret = false;
1807
1808		pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1809		if (pol && (pol->flags & MPOL_F_MOF))
1810			ret = true;
1811		mpol_cond_put(pol);
1812
1813		return ret;
1814	}
1815
1816	pol = vma->vm_policy;
1817	if (!pol)
1818		pol = get_task_policy(current);
1819
1820	return pol->flags & MPOL_F_MOF;
1821}
1822
1823bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1824{
1825	enum zone_type dynamic_policy_zone = policy_zone;
1826
1827	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1828
1829	/*
1830	 * if policy->nodes has movable memory only,
1831	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1832	 *
1833	 * policy->nodes is intersect with node_states[N_MEMORY].
1834	 * so if the following test fails, it implies
1835	 * policy->nodes has movable memory only.
1836	 */
1837	if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1838		dynamic_policy_zone = ZONE_MOVABLE;
1839
1840	return zone >= dynamic_policy_zone;
1841}
1842
1843/*
1844 * Return a nodemask representing a mempolicy for filtering nodes for
1845 * page allocation
1846 */
1847nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1848{
1849	int mode = policy->mode;
1850
1851	/* Lower zones don't get a nodemask applied for MPOL_BIND */
1852	if (unlikely(mode == MPOL_BIND) &&
1853		apply_policy_zone(policy, gfp_zone(gfp)) &&
1854		cpuset_nodemask_valid_mems_allowed(&policy->nodes))
1855		return &policy->nodes;
1856
1857	if (mode == MPOL_PREFERRED_MANY)
1858		return &policy->nodes;
1859
1860	return NULL;
1861}
1862
1863/*
1864 * Return the  preferred node id for 'prefer' mempolicy, and return
1865 * the given id for all other policies.
1866 *
1867 * policy_node() is always coupled with policy_nodemask(), which
1868 * secures the nodemask limit for 'bind' and 'prefer-many' policy.
1869 */
1870static int policy_node(gfp_t gfp, struct mempolicy *policy, int nd)
1871{
1872	if (policy->mode == MPOL_PREFERRED) {
1873		nd = first_node(policy->nodes);
1874	} else {
1875		/*
1876		 * __GFP_THISNODE shouldn't even be used with the bind policy
1877		 * because we might easily break the expectation to stay on the
1878		 * requested node and not break the policy.
1879		 */
1880		WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1881	}
1882
1883	if ((policy->mode == MPOL_BIND ||
1884	     policy->mode == MPOL_PREFERRED_MANY) &&
1885	    policy->home_node != NUMA_NO_NODE)
1886		return policy->home_node;
1887
1888	return nd;
1889}
1890
1891/* Do dynamic interleaving for a process */
1892static unsigned interleave_nodes(struct mempolicy *policy)
1893{
1894	unsigned next;
1895	struct task_struct *me = current;
1896
1897	next = next_node_in(me->il_prev, policy->nodes);
1898	if (next < MAX_NUMNODES)
1899		me->il_prev = next;
1900	return next;
1901}
1902
1903/*
1904 * Depending on the memory policy provide a node from which to allocate the
1905 * next slab entry.
1906 */
1907unsigned int mempolicy_slab_node(void)
1908{
1909	struct mempolicy *policy;
1910	int node = numa_mem_id();
1911
1912	if (!in_task())
1913		return node;
1914
1915	policy = current->mempolicy;
1916	if (!policy)
1917		return node;
1918
1919	switch (policy->mode) {
1920	case MPOL_PREFERRED:
1921		return first_node(policy->nodes);
1922
1923	case MPOL_INTERLEAVE:
1924		return interleave_nodes(policy);
1925
1926	case MPOL_BIND:
1927	case MPOL_PREFERRED_MANY:
1928	{
1929		struct zoneref *z;
1930
1931		/*
1932		 * Follow bind policy behavior and start allocation at the
1933		 * first node.
1934		 */
1935		struct zonelist *zonelist;
1936		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1937		zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1938		z = first_zones_zonelist(zonelist, highest_zoneidx,
1939							&policy->nodes);
1940		return z->zone ? zone_to_nid(z->zone) : node;
1941	}
1942	case MPOL_LOCAL:
1943		return node;
1944
1945	default:
1946		BUG();
1947	}
1948}
1949
1950/*
1951 * Do static interleaving for a VMA with known offset @n.  Returns the n'th
1952 * node in pol->nodes (starting from n=0), wrapping around if n exceeds the
1953 * number of present nodes.
1954 */
1955static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1956{
1957	nodemask_t nodemask = pol->nodes;
1958	unsigned int target, nnodes;
1959	int i;
1960	int nid;
1961	/*
1962	 * The barrier will stabilize the nodemask in a register or on
1963	 * the stack so that it will stop changing under the code.
1964	 *
1965	 * Between first_node() and next_node(), pol->nodes could be changed
1966	 * by other threads. So we put pol->nodes in a local stack.
1967	 */
1968	barrier();
1969
1970	nnodes = nodes_weight(nodemask);
1971	if (!nnodes)
1972		return numa_node_id();
1973	target = (unsigned int)n % nnodes;
1974	nid = first_node(nodemask);
1975	for (i = 0; i < target; i++)
1976		nid = next_node(nid, nodemask);
1977	return nid;
1978}
1979
1980/* Determine a node number for interleave */
1981static inline unsigned interleave_nid(struct mempolicy *pol,
1982		 struct vm_area_struct *vma, unsigned long addr, int shift)
1983{
1984	if (vma) {
1985		unsigned long off;
1986
1987		/*
1988		 * for small pages, there is no difference between
1989		 * shift and PAGE_SHIFT, so the bit-shift is safe.
1990		 * for huge pages, since vm_pgoff is in units of small
1991		 * pages, we need to shift off the always 0 bits to get
1992		 * a useful offset.
1993		 */
1994		BUG_ON(shift < PAGE_SHIFT);
1995		off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1996		off += (addr - vma->vm_start) >> shift;
1997		return offset_il_node(pol, off);
1998	} else
1999		return interleave_nodes(pol);
2000}
2001
2002#ifdef CONFIG_HUGETLBFS
2003/*
2004 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2005 * @vma: virtual memory area whose policy is sought
2006 * @addr: address in @vma for shared policy lookup and interleave policy
2007 * @gfp_flags: for requested zone
2008 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2009 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2010 *
2011 * Returns a nid suitable for a huge page allocation and a pointer
2012 * to the struct mempolicy for conditional unref after allocation.
2013 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2014 * to the mempolicy's @nodemask for filtering the zonelist.
2015 *
2016 * Must be protected by read_mems_allowed_begin()
2017 */
2018int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2019				struct mempolicy **mpol, nodemask_t **nodemask)
2020{
2021	int nid;
2022	int mode;
2023
2024	*mpol = get_vma_policy(vma, addr);
2025	*nodemask = NULL;
2026	mode = (*mpol)->mode;
2027
2028	if (unlikely(mode == MPOL_INTERLEAVE)) {
2029		nid = interleave_nid(*mpol, vma, addr,
2030					huge_page_shift(hstate_vma(vma)));
2031	} else {
2032		nid = policy_node(gfp_flags, *mpol, numa_node_id());
2033		if (mode == MPOL_BIND || mode == MPOL_PREFERRED_MANY)
2034			*nodemask = &(*mpol)->nodes;
2035	}
2036	return nid;
2037}
2038
2039/*
2040 * init_nodemask_of_mempolicy
2041 *
2042 * If the current task's mempolicy is "default" [NULL], return 'false'
2043 * to indicate default policy.  Otherwise, extract the policy nodemask
2044 * for 'bind' or 'interleave' policy into the argument nodemask, or
2045 * initialize the argument nodemask to contain the single node for
2046 * 'preferred' or 'local' policy and return 'true' to indicate presence
2047 * of non-default mempolicy.
2048 *
2049 * We don't bother with reference counting the mempolicy [mpol_get/put]
2050 * because the current task is examining it's own mempolicy and a task's
2051 * mempolicy is only ever changed by the task itself.
2052 *
2053 * N.B., it is the caller's responsibility to free a returned nodemask.
2054 */
2055bool init_nodemask_of_mempolicy(nodemask_t *mask)
2056{
2057	struct mempolicy *mempolicy;
2058
2059	if (!(mask && current->mempolicy))
2060		return false;
2061
2062	task_lock(current);
2063	mempolicy = current->mempolicy;
2064	switch (mempolicy->mode) {
2065	case MPOL_PREFERRED:
2066	case MPOL_PREFERRED_MANY:
2067	case MPOL_BIND:
2068	case MPOL_INTERLEAVE:
2069		*mask = mempolicy->nodes;
2070		break;
2071
2072	case MPOL_LOCAL:
2073		init_nodemask_of_node(mask, numa_node_id());
2074		break;
2075
2076	default:
2077		BUG();
2078	}
2079	task_unlock(current);
2080
2081	return true;
2082}
2083#endif
2084
2085/*
2086 * mempolicy_in_oom_domain
2087 *
2088 * If tsk's mempolicy is "bind", check for intersection between mask and
2089 * the policy nodemask. Otherwise, return true for all other policies
2090 * including "interleave", as a tsk with "interleave" policy may have
2091 * memory allocated from all nodes in system.
2092 *
2093 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2094 */
2095bool mempolicy_in_oom_domain(struct task_struct *tsk,
2096					const nodemask_t *mask)
2097{
2098	struct mempolicy *mempolicy;
2099	bool ret = true;
2100
2101	if (!mask)
2102		return ret;
2103
2104	task_lock(tsk);
2105	mempolicy = tsk->mempolicy;
2106	if (mempolicy && mempolicy->mode == MPOL_BIND)
2107		ret = nodes_intersects(mempolicy->nodes, *mask);
2108	task_unlock(tsk);
2109
2110	return ret;
2111}
2112
2113/* Allocate a page in interleaved policy.
2114   Own path because it needs to do special accounting. */
2115static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
2116					unsigned nid)
2117{
2118	struct page *page;
2119
2120	page = __alloc_pages(gfp, order, nid, NULL);
2121	/* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
2122	if (!static_branch_likely(&vm_numa_stat_key))
2123		return page;
2124	if (page && page_to_nid(page) == nid) {
2125		preempt_disable();
2126		__count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2127		preempt_enable();
2128	}
2129	return page;
2130}
2131
2132static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2133						int nid, struct mempolicy *pol)
2134{
2135	struct page *page;
2136	gfp_t preferred_gfp;
2137
2138	/*
2139	 * This is a two pass approach. The first pass will only try the
2140	 * preferred nodes but skip the direct reclaim and allow the
2141	 * allocation to fail, while the second pass will try all the
2142	 * nodes in system.
2143	 */
2144	preferred_gfp = gfp | __GFP_NOWARN;
2145	preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2146	page = __alloc_pages(preferred_gfp, order, nid, &pol->nodes);
2147	if (!page)
2148		page = __alloc_pages(gfp, order, nid, NULL);
2149
2150	return page;
2151}
2152
2153/**
2154 * vma_alloc_folio - Allocate a folio for a VMA.
2155 * @gfp: GFP flags.
2156 * @order: Order of the folio.
2157 * @vma: Pointer to VMA or NULL if not available.
2158 * @addr: Virtual address of the allocation.  Must be inside @vma.
 
2159 * @hugepage: For hugepages try only the preferred node if possible.
2160 *
2161 * Allocate a folio for a specific address in @vma, using the appropriate
2162 * NUMA policy.  When @vma is not NULL the caller must hold the mmap_lock
2163 * of the mm_struct of the VMA to prevent it from going away.  Should be
2164 * used for all allocations for folios that will be mapped into user space.
2165 *
2166 * Return: The folio on success or NULL if allocation fails.
2167 */
2168struct folio *vma_alloc_folio(gfp_t gfp, int order, struct vm_area_struct *vma,
2169		unsigned long addr, bool hugepage)
2170{
2171	struct mempolicy *pol;
2172	int node = numa_node_id();
2173	struct folio *folio;
2174	int preferred_nid;
2175	nodemask_t *nmask;
2176
2177	pol = get_vma_policy(vma, addr);
2178
2179	if (pol->mode == MPOL_INTERLEAVE) {
2180		struct page *page;
2181		unsigned nid;
2182
2183		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2184		mpol_cond_put(pol);
2185		gfp |= __GFP_COMP;
2186		page = alloc_page_interleave(gfp, order, nid);
2187		if (page && order > 1)
2188			prep_transhuge_page(page);
2189		folio = (struct folio *)page;
2190		goto out;
2191	}
2192
2193	if (pol->mode == MPOL_PREFERRED_MANY) {
2194		struct page *page;
2195
2196		node = policy_node(gfp, pol, node);
2197		gfp |= __GFP_COMP;
2198		page = alloc_pages_preferred_many(gfp, order, node, pol);
2199		mpol_cond_put(pol);
2200		if (page && order > 1)
2201			prep_transhuge_page(page);
2202		folio = (struct folio *)page;
2203		goto out;
2204	}
2205
2206	if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2207		int hpage_node = node;
2208
2209		/*
2210		 * For hugepage allocation and non-interleave policy which
2211		 * allows the current node (or other explicitly preferred
2212		 * node) we only try to allocate from the current/preferred
2213		 * node and don't fall back to other nodes, as the cost of
2214		 * remote accesses would likely offset THP benefits.
2215		 *
2216		 * If the policy is interleave or does not allow the current
2217		 * node in its nodemask, we allocate the standard way.
2218		 */
2219		if (pol->mode == MPOL_PREFERRED)
2220			hpage_node = first_node(pol->nodes);
2221
2222		nmask = policy_nodemask(gfp, pol);
2223		if (!nmask || node_isset(hpage_node, *nmask)) {
2224			mpol_cond_put(pol);
2225			/*
2226			 * First, try to allocate THP only on local node, but
2227			 * don't reclaim unnecessarily, just compact.
2228			 */
2229			folio = __folio_alloc_node(gfp | __GFP_THISNODE |
2230					__GFP_NORETRY, order, hpage_node);
2231
2232			/*
2233			 * If hugepage allocations are configured to always
2234			 * synchronous compact or the vma has been madvised
2235			 * to prefer hugepage backing, retry allowing remote
2236			 * memory with both reclaim and compact as well.
2237			 */
2238			if (!folio && (gfp & __GFP_DIRECT_RECLAIM))
2239				folio = __folio_alloc(gfp, order, hpage_node,
2240						      nmask);
2241
2242			goto out;
2243		}
2244	}
2245
2246	nmask = policy_nodemask(gfp, pol);
2247	preferred_nid = policy_node(gfp, pol, node);
2248	folio = __folio_alloc(gfp, order, preferred_nid, nmask);
2249	mpol_cond_put(pol);
2250out:
2251	return folio;
2252}
2253EXPORT_SYMBOL(vma_alloc_folio);
2254
2255/**
2256 * alloc_pages - Allocate pages.
2257 * @gfp: GFP flags.
2258 * @order: Power of two of number of pages to allocate.
2259 *
2260 * Allocate 1 << @order contiguous pages.  The physical address of the
2261 * first page is naturally aligned (eg an order-3 allocation will be aligned
2262 * to a multiple of 8 * PAGE_SIZE bytes).  The NUMA policy of the current
2263 * process is honoured when in process context.
2264 *
2265 * Context: Can be called from any context, providing the appropriate GFP
2266 * flags are used.
2267 * Return: The page on success or NULL if allocation fails.
2268 */
2269struct page *alloc_pages(gfp_t gfp, unsigned order)
2270{
2271	struct mempolicy *pol = &default_policy;
2272	struct page *page;
2273
2274	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2275		pol = get_task_policy(current);
2276
2277	/*
2278	 * No reference counting needed for current->mempolicy
2279	 * nor system default_policy
2280	 */
2281	if (pol->mode == MPOL_INTERLEAVE)
2282		page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2283	else if (pol->mode == MPOL_PREFERRED_MANY)
2284		page = alloc_pages_preferred_many(gfp, order,
2285				  policy_node(gfp, pol, numa_node_id()), pol);
2286	else
2287		page = __alloc_pages(gfp, order,
2288				policy_node(gfp, pol, numa_node_id()),
2289				policy_nodemask(gfp, pol));
2290
2291	return page;
2292}
2293EXPORT_SYMBOL(alloc_pages);
2294
2295struct folio *folio_alloc(gfp_t gfp, unsigned order)
2296{
2297	struct page *page = alloc_pages(gfp | __GFP_COMP, order);
2298
2299	if (page && order > 1)
2300		prep_transhuge_page(page);
2301	return (struct folio *)page;
2302}
2303EXPORT_SYMBOL(folio_alloc);
2304
2305static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2306		struct mempolicy *pol, unsigned long nr_pages,
2307		struct page **page_array)
2308{
2309	int nodes;
2310	unsigned long nr_pages_per_node;
2311	int delta;
2312	int i;
2313	unsigned long nr_allocated;
2314	unsigned long total_allocated = 0;
2315
2316	nodes = nodes_weight(pol->nodes);
2317	nr_pages_per_node = nr_pages / nodes;
2318	delta = nr_pages - nodes * nr_pages_per_node;
2319
2320	for (i = 0; i < nodes; i++) {
2321		if (delta) {
2322			nr_allocated = __alloc_pages_bulk(gfp,
2323					interleave_nodes(pol), NULL,
2324					nr_pages_per_node + 1, NULL,
2325					page_array);
2326			delta--;
2327		} else {
2328			nr_allocated = __alloc_pages_bulk(gfp,
2329					interleave_nodes(pol), NULL,
2330					nr_pages_per_node, NULL, page_array);
2331		}
2332
2333		page_array += nr_allocated;
2334		total_allocated += nr_allocated;
2335	}
2336
2337	return total_allocated;
2338}
2339
2340static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2341		struct mempolicy *pol, unsigned long nr_pages,
2342		struct page **page_array)
2343{
2344	gfp_t preferred_gfp;
2345	unsigned long nr_allocated = 0;
2346
2347	preferred_gfp = gfp | __GFP_NOWARN;
2348	preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2349
2350	nr_allocated  = __alloc_pages_bulk(preferred_gfp, nid, &pol->nodes,
2351					   nr_pages, NULL, page_array);
2352
2353	if (nr_allocated < nr_pages)
2354		nr_allocated += __alloc_pages_bulk(gfp, numa_node_id(), NULL,
2355				nr_pages - nr_allocated, NULL,
2356				page_array + nr_allocated);
2357	return nr_allocated;
2358}
2359
2360/* alloc pages bulk and mempolicy should be considered at the
2361 * same time in some situation such as vmalloc.
2362 *
2363 * It can accelerate memory allocation especially interleaving
2364 * allocate memory.
2365 */
2366unsigned long alloc_pages_bulk_array_mempolicy(gfp_t gfp,
2367		unsigned long nr_pages, struct page **page_array)
2368{
2369	struct mempolicy *pol = &default_policy;
2370
2371	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2372		pol = get_task_policy(current);
2373
2374	if (pol->mode == MPOL_INTERLEAVE)
2375		return alloc_pages_bulk_array_interleave(gfp, pol,
2376							 nr_pages, page_array);
2377
2378	if (pol->mode == MPOL_PREFERRED_MANY)
2379		return alloc_pages_bulk_array_preferred_many(gfp,
2380				numa_node_id(), pol, nr_pages, page_array);
2381
2382	return __alloc_pages_bulk(gfp, policy_node(gfp, pol, numa_node_id()),
2383				  policy_nodemask(gfp, pol), nr_pages, NULL,
2384				  page_array);
2385}
2386
2387int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2388{
2389	struct mempolicy *pol = mpol_dup(vma_policy(src));
2390
2391	if (IS_ERR(pol))
2392		return PTR_ERR(pol);
2393	dst->vm_policy = pol;
2394	return 0;
2395}
2396
2397/*
2398 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2399 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2400 * with the mems_allowed returned by cpuset_mems_allowed().  This
2401 * keeps mempolicies cpuset relative after its cpuset moves.  See
2402 * further kernel/cpuset.c update_nodemask().
2403 *
2404 * current's mempolicy may be rebinded by the other task(the task that changes
2405 * cpuset's mems), so we needn't do rebind work for current task.
2406 */
2407
2408/* Slow path of a mempolicy duplicate */
2409struct mempolicy *__mpol_dup(struct mempolicy *old)
2410{
2411	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2412
2413	if (!new)
2414		return ERR_PTR(-ENOMEM);
2415
2416	/* task's mempolicy is protected by alloc_lock */
2417	if (old == current->mempolicy) {
2418		task_lock(current);
2419		*new = *old;
2420		task_unlock(current);
2421	} else
2422		*new = *old;
2423
2424	if (current_cpuset_is_being_rebound()) {
2425		nodemask_t mems = cpuset_mems_allowed(current);
2426		mpol_rebind_policy(new, &mems);
2427	}
2428	atomic_set(&new->refcnt, 1);
2429	return new;
2430}
2431
2432/* Slow path of a mempolicy comparison */
2433bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2434{
2435	if (!a || !b)
2436		return false;
2437	if (a->mode != b->mode)
2438		return false;
2439	if (a->flags != b->flags)
2440		return false;
2441	if (a->home_node != b->home_node)
2442		return false;
2443	if (mpol_store_user_nodemask(a))
2444		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2445			return false;
2446
2447	switch (a->mode) {
2448	case MPOL_BIND:
2449	case MPOL_INTERLEAVE:
2450	case MPOL_PREFERRED:
2451	case MPOL_PREFERRED_MANY:
2452		return !!nodes_equal(a->nodes, b->nodes);
2453	case MPOL_LOCAL:
2454		return true;
2455	default:
2456		BUG();
2457		return false;
2458	}
2459}
2460
2461/*
2462 * Shared memory backing store policy support.
2463 *
2464 * Remember policies even when nobody has shared memory mapped.
2465 * The policies are kept in Red-Black tree linked from the inode.
2466 * They are protected by the sp->lock rwlock, which should be held
2467 * for any accesses to the tree.
2468 */
2469
2470/*
2471 * lookup first element intersecting start-end.  Caller holds sp->lock for
2472 * reading or for writing
2473 */
2474static struct sp_node *
2475sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2476{
2477	struct rb_node *n = sp->root.rb_node;
2478
2479	while (n) {
2480		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2481
2482		if (start >= p->end)
2483			n = n->rb_right;
2484		else if (end <= p->start)
2485			n = n->rb_left;
2486		else
2487			break;
2488	}
2489	if (!n)
2490		return NULL;
2491	for (;;) {
2492		struct sp_node *w = NULL;
2493		struct rb_node *prev = rb_prev(n);
2494		if (!prev)
2495			break;
2496		w = rb_entry(prev, struct sp_node, nd);
2497		if (w->end <= start)
2498			break;
2499		n = prev;
2500	}
2501	return rb_entry(n, struct sp_node, nd);
2502}
2503
2504/*
2505 * Insert a new shared policy into the list.  Caller holds sp->lock for
2506 * writing.
2507 */
2508static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2509{
2510	struct rb_node **p = &sp->root.rb_node;
2511	struct rb_node *parent = NULL;
2512	struct sp_node *nd;
2513
2514	while (*p) {
2515		parent = *p;
2516		nd = rb_entry(parent, struct sp_node, nd);
2517		if (new->start < nd->start)
2518			p = &(*p)->rb_left;
2519		else if (new->end > nd->end)
2520			p = &(*p)->rb_right;
2521		else
2522			BUG();
2523	}
2524	rb_link_node(&new->nd, parent, p);
2525	rb_insert_color(&new->nd, &sp->root);
2526	pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2527		 new->policy ? new->policy->mode : 0);
2528}
2529
2530/* Find shared policy intersecting idx */
2531struct mempolicy *
2532mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2533{
2534	struct mempolicy *pol = NULL;
2535	struct sp_node *sn;
2536
2537	if (!sp->root.rb_node)
2538		return NULL;
2539	read_lock(&sp->lock);
2540	sn = sp_lookup(sp, idx, idx+1);
2541	if (sn) {
2542		mpol_get(sn->policy);
2543		pol = sn->policy;
2544	}
2545	read_unlock(&sp->lock);
2546	return pol;
2547}
2548
2549static void sp_free(struct sp_node *n)
2550{
2551	mpol_put(n->policy);
2552	kmem_cache_free(sn_cache, n);
2553}
2554
2555/**
2556 * mpol_misplaced - check whether current page node is valid in policy
2557 *
2558 * @page: page to be checked
2559 * @vma: vm area where page mapped
2560 * @addr: virtual address where page mapped
2561 *
2562 * Lookup current policy node id for vma,addr and "compare to" page's
2563 * node id.  Policy determination "mimics" alloc_page_vma().
2564 * Called from fault path where we know the vma and faulting address.
2565 *
2566 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2567 * policy, or a suitable node ID to allocate a replacement page from.
2568 */
2569int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
2570{
2571	struct mempolicy *pol;
2572	struct zoneref *z;
2573	int curnid = page_to_nid(page);
2574	unsigned long pgoff;
2575	int thiscpu = raw_smp_processor_id();
2576	int thisnid = cpu_to_node(thiscpu);
2577	int polnid = NUMA_NO_NODE;
2578	int ret = NUMA_NO_NODE;
2579
2580	pol = get_vma_policy(vma, addr);
2581	if (!(pol->flags & MPOL_F_MOF))
2582		goto out;
2583
2584	switch (pol->mode) {
2585	case MPOL_INTERLEAVE:
2586		pgoff = vma->vm_pgoff;
2587		pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2588		polnid = offset_il_node(pol, pgoff);
2589		break;
2590
2591	case MPOL_PREFERRED:
2592		if (node_isset(curnid, pol->nodes))
2593			goto out;
2594		polnid = first_node(pol->nodes);
2595		break;
2596
2597	case MPOL_LOCAL:
2598		polnid = numa_node_id();
2599		break;
2600
2601	case MPOL_BIND:
2602		/* Optimize placement among multiple nodes via NUMA balancing */
2603		if (pol->flags & MPOL_F_MORON) {
2604			if (node_isset(thisnid, pol->nodes))
2605				break;
2606			goto out;
2607		}
2608		fallthrough;
2609
2610	case MPOL_PREFERRED_MANY:
2611		/*
 
2612		 * use current page if in policy nodemask,
2613		 * else select nearest allowed node, if any.
2614		 * If no allowed nodes, use current [!misplaced].
2615		 */
2616		if (node_isset(curnid, pol->nodes))
2617			goto out;
2618		z = first_zones_zonelist(
2619				node_zonelist(numa_node_id(), GFP_HIGHUSER),
2620				gfp_zone(GFP_HIGHUSER),
2621				&pol->nodes);
2622		polnid = zone_to_nid(z->zone);
2623		break;
2624
2625	default:
2626		BUG();
2627	}
2628
2629	/* Migrate the page towards the node whose CPU is referencing it */
2630	if (pol->flags & MPOL_F_MORON) {
2631		polnid = thisnid;
2632
2633		if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
2634			goto out;
2635	}
2636
2637	if (curnid != polnid)
2638		ret = polnid;
2639out:
2640	mpol_cond_put(pol);
2641
2642	return ret;
2643}
2644
2645/*
2646 * Drop the (possibly final) reference to task->mempolicy.  It needs to be
2647 * dropped after task->mempolicy is set to NULL so that any allocation done as
2648 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2649 * policy.
2650 */
2651void mpol_put_task_policy(struct task_struct *task)
2652{
2653	struct mempolicy *pol;
2654
2655	task_lock(task);
2656	pol = task->mempolicy;
2657	task->mempolicy = NULL;
2658	task_unlock(task);
2659	mpol_put(pol);
2660}
2661
2662static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2663{
2664	pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2665	rb_erase(&n->nd, &sp->root);
2666	sp_free(n);
2667}
2668
2669static void sp_node_init(struct sp_node *node, unsigned long start,
2670			unsigned long end, struct mempolicy *pol)
2671{
2672	node->start = start;
2673	node->end = end;
2674	node->policy = pol;
2675}
2676
2677static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2678				struct mempolicy *pol)
2679{
2680	struct sp_node *n;
2681	struct mempolicy *newpol;
2682
2683	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2684	if (!n)
2685		return NULL;
2686
2687	newpol = mpol_dup(pol);
2688	if (IS_ERR(newpol)) {
2689		kmem_cache_free(sn_cache, n);
2690		return NULL;
2691	}
2692	newpol->flags |= MPOL_F_SHARED;
2693	sp_node_init(n, start, end, newpol);
2694
2695	return n;
2696}
2697
2698/* Replace a policy range. */
2699static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2700				 unsigned long end, struct sp_node *new)
2701{
2702	struct sp_node *n;
2703	struct sp_node *n_new = NULL;
2704	struct mempolicy *mpol_new = NULL;
2705	int ret = 0;
2706
2707restart:
2708	write_lock(&sp->lock);
2709	n = sp_lookup(sp, start, end);
2710	/* Take care of old policies in the same range. */
2711	while (n && n->start < end) {
2712		struct rb_node *next = rb_next(&n->nd);
2713		if (n->start >= start) {
2714			if (n->end <= end)
2715				sp_delete(sp, n);
2716			else
2717				n->start = end;
2718		} else {
2719			/* Old policy spanning whole new range. */
2720			if (n->end > end) {
2721				if (!n_new)
2722					goto alloc_new;
2723
2724				*mpol_new = *n->policy;
2725				atomic_set(&mpol_new->refcnt, 1);
2726				sp_node_init(n_new, end, n->end, mpol_new);
2727				n->end = start;
2728				sp_insert(sp, n_new);
2729				n_new = NULL;
2730				mpol_new = NULL;
2731				break;
2732			} else
2733				n->end = start;
2734		}
2735		if (!next)
2736			break;
2737		n = rb_entry(next, struct sp_node, nd);
2738	}
2739	if (new)
2740		sp_insert(sp, new);
2741	write_unlock(&sp->lock);
2742	ret = 0;
2743
2744err_out:
2745	if (mpol_new)
2746		mpol_put(mpol_new);
2747	if (n_new)
2748		kmem_cache_free(sn_cache, n_new);
2749
2750	return ret;
2751
2752alloc_new:
2753	write_unlock(&sp->lock);
2754	ret = -ENOMEM;
2755	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2756	if (!n_new)
2757		goto err_out;
2758	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2759	if (!mpol_new)
2760		goto err_out;
2761	atomic_set(&mpol_new->refcnt, 1);
2762	goto restart;
2763}
2764
2765/**
2766 * mpol_shared_policy_init - initialize shared policy for inode
2767 * @sp: pointer to inode shared policy
2768 * @mpol:  struct mempolicy to install
2769 *
2770 * Install non-NULL @mpol in inode's shared policy rb-tree.
2771 * On entry, the current task has a reference on a non-NULL @mpol.
2772 * This must be released on exit.
2773 * This is called at get_inode() calls and we can use GFP_KERNEL.
2774 */
2775void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2776{
2777	int ret;
2778
2779	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2780	rwlock_init(&sp->lock);
2781
2782	if (mpol) {
2783		struct vm_area_struct pvma;
2784		struct mempolicy *new;
2785		NODEMASK_SCRATCH(scratch);
2786
2787		if (!scratch)
2788			goto put_mpol;
2789		/* contextualize the tmpfs mount point mempolicy */
2790		new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2791		if (IS_ERR(new))
2792			goto free_scratch; /* no valid nodemask intersection */
2793
2794		task_lock(current);
2795		ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2796		task_unlock(current);
2797		if (ret)
2798			goto put_new;
2799
2800		/* Create pseudo-vma that contains just the policy */
2801		vma_init(&pvma, NULL);
2802		pvma.vm_end = TASK_SIZE;	/* policy covers entire file */
2803		mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2804
2805put_new:
2806		mpol_put(new);			/* drop initial ref */
2807free_scratch:
2808		NODEMASK_SCRATCH_FREE(scratch);
2809put_mpol:
2810		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
2811	}
2812}
2813
2814int mpol_set_shared_policy(struct shared_policy *info,
2815			struct vm_area_struct *vma, struct mempolicy *npol)
2816{
2817	int err;
2818	struct sp_node *new = NULL;
2819	unsigned long sz = vma_pages(vma);
2820
2821	pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2822		 vma->vm_pgoff,
2823		 sz, npol ? npol->mode : -1,
2824		 npol ? npol->flags : -1,
2825		 npol ? nodes_addr(npol->nodes)[0] : NUMA_NO_NODE);
2826
2827	if (npol) {
2828		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2829		if (!new)
2830			return -ENOMEM;
2831	}
2832	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2833	if (err && new)
2834		sp_free(new);
2835	return err;
2836}
2837
2838/* Free a backing policy store on inode delete. */
2839void mpol_free_shared_policy(struct shared_policy *p)
2840{
2841	struct sp_node *n;
2842	struct rb_node *next;
2843
2844	if (!p->root.rb_node)
2845		return;
2846	write_lock(&p->lock);
2847	next = rb_first(&p->root);
2848	while (next) {
2849		n = rb_entry(next, struct sp_node, nd);
2850		next = rb_next(&n->nd);
2851		sp_delete(p, n);
2852	}
2853	write_unlock(&p->lock);
2854}
2855
2856#ifdef CONFIG_NUMA_BALANCING
2857static int __initdata numabalancing_override;
2858
2859static void __init check_numabalancing_enable(void)
2860{
2861	bool numabalancing_default = false;
2862
2863	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2864		numabalancing_default = true;
2865
2866	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2867	if (numabalancing_override)
2868		set_numabalancing_state(numabalancing_override == 1);
2869
2870	if (num_online_nodes() > 1 && !numabalancing_override) {
2871		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2872			numabalancing_default ? "Enabling" : "Disabling");
2873		set_numabalancing_state(numabalancing_default);
2874	}
2875}
2876
2877static int __init setup_numabalancing(char *str)
2878{
2879	int ret = 0;
2880	if (!str)
2881		goto out;
2882
2883	if (!strcmp(str, "enable")) {
2884		numabalancing_override = 1;
2885		ret = 1;
2886	} else if (!strcmp(str, "disable")) {
2887		numabalancing_override = -1;
2888		ret = 1;
2889	}
2890out:
2891	if (!ret)
2892		pr_warn("Unable to parse numa_balancing=\n");
2893
2894	return ret;
2895}
2896__setup("numa_balancing=", setup_numabalancing);
2897#else
2898static inline void __init check_numabalancing_enable(void)
2899{
2900}
2901#endif /* CONFIG_NUMA_BALANCING */
2902
2903/* assumes fs == KERNEL_DS */
2904void __init numa_policy_init(void)
2905{
2906	nodemask_t interleave_nodes;
2907	unsigned long largest = 0;
2908	int nid, prefer = 0;
2909
2910	policy_cache = kmem_cache_create("numa_policy",
2911					 sizeof(struct mempolicy),
2912					 0, SLAB_PANIC, NULL);
2913
2914	sn_cache = kmem_cache_create("shared_policy_node",
2915				     sizeof(struct sp_node),
2916				     0, SLAB_PANIC, NULL);
2917
2918	for_each_node(nid) {
2919		preferred_node_policy[nid] = (struct mempolicy) {
2920			.refcnt = ATOMIC_INIT(1),
2921			.mode = MPOL_PREFERRED,
2922			.flags = MPOL_F_MOF | MPOL_F_MORON,
2923			.nodes = nodemask_of_node(nid),
2924		};
2925	}
2926
2927	/*
2928	 * Set interleaving policy for system init. Interleaving is only
2929	 * enabled across suitably sized nodes (default is >= 16MB), or
2930	 * fall back to the largest node if they're all smaller.
2931	 */
2932	nodes_clear(interleave_nodes);
2933	for_each_node_state(nid, N_MEMORY) {
2934		unsigned long total_pages = node_present_pages(nid);
2935
2936		/* Preserve the largest node */
2937		if (largest < total_pages) {
2938			largest = total_pages;
2939			prefer = nid;
2940		}
2941
2942		/* Interleave this node? */
2943		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2944			node_set(nid, interleave_nodes);
2945	}
2946
2947	/* All too small, use the largest */
2948	if (unlikely(nodes_empty(interleave_nodes)))
2949		node_set(prefer, interleave_nodes);
2950
2951	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2952		pr_err("%s: interleaving failed\n", __func__);
2953
2954	check_numabalancing_enable();
2955}
2956
2957/* Reset policy of current process to default */
2958void numa_default_policy(void)
2959{
2960	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2961}
2962
2963/*
2964 * Parse and format mempolicy from/to strings
2965 */
2966
2967static const char * const policy_modes[] =
2968{
2969	[MPOL_DEFAULT]    = "default",
2970	[MPOL_PREFERRED]  = "prefer",
2971	[MPOL_BIND]       = "bind",
2972	[MPOL_INTERLEAVE] = "interleave",
2973	[MPOL_LOCAL]      = "local",
2974	[MPOL_PREFERRED_MANY]  = "prefer (many)",
2975};
2976
2977
2978#ifdef CONFIG_TMPFS
2979/**
2980 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2981 * @str:  string containing mempolicy to parse
2982 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2983 *
2984 * Format of input:
2985 *	<mode>[=<flags>][:<nodelist>]
2986 *
2987 * Return: %0 on success, else %1
2988 */
2989int mpol_parse_str(char *str, struct mempolicy **mpol)
2990{
2991	struct mempolicy *new = NULL;
2992	unsigned short mode_flags;
2993	nodemask_t nodes;
2994	char *nodelist = strchr(str, ':');
2995	char *flags = strchr(str, '=');
2996	int err = 1, mode;
2997
2998	if (flags)
2999		*flags++ = '\0';	/* terminate mode string */
3000
3001	if (nodelist) {
3002		/* NUL-terminate mode or flags string */
3003		*nodelist++ = '\0';
3004		if (nodelist_parse(nodelist, nodes))
3005			goto out;
3006		if (!nodes_subset(nodes, node_states[N_MEMORY]))
3007			goto out;
3008	} else
3009		nodes_clear(nodes);
3010
3011	mode = match_string(policy_modes, MPOL_MAX, str);
3012	if (mode < 0)
3013		goto out;
3014
3015	switch (mode) {
3016	case MPOL_PREFERRED:
3017		/*
3018		 * Insist on a nodelist of one node only, although later
3019		 * we use first_node(nodes) to grab a single node, so here
3020		 * nodelist (or nodes) cannot be empty.
3021		 */
3022		if (nodelist) {
3023			char *rest = nodelist;
3024			while (isdigit(*rest))
3025				rest++;
3026			if (*rest)
3027				goto out;
3028			if (nodes_empty(nodes))
3029				goto out;
3030		}
3031		break;
3032	case MPOL_INTERLEAVE:
3033		/*
3034		 * Default to online nodes with memory if no nodelist
3035		 */
3036		if (!nodelist)
3037			nodes = node_states[N_MEMORY];
3038		break;
3039	case MPOL_LOCAL:
3040		/*
3041		 * Don't allow a nodelist;  mpol_new() checks flags
3042		 */
3043		if (nodelist)
3044			goto out;
3045		break;
3046	case MPOL_DEFAULT:
3047		/*
3048		 * Insist on a empty nodelist
3049		 */
3050		if (!nodelist)
3051			err = 0;
3052		goto out;
3053	case MPOL_PREFERRED_MANY:
3054	case MPOL_BIND:
3055		/*
3056		 * Insist on a nodelist
3057		 */
3058		if (!nodelist)
3059			goto out;
3060	}
3061
3062	mode_flags = 0;
3063	if (flags) {
3064		/*
3065		 * Currently, we only support two mutually exclusive
3066		 * mode flags.
3067		 */
3068		if (!strcmp(flags, "static"))
3069			mode_flags |= MPOL_F_STATIC_NODES;
3070		else if (!strcmp(flags, "relative"))
3071			mode_flags |= MPOL_F_RELATIVE_NODES;
3072		else
3073			goto out;
3074	}
3075
3076	new = mpol_new(mode, mode_flags, &nodes);
3077	if (IS_ERR(new))
3078		goto out;
3079
3080	/*
3081	 * Save nodes for mpol_to_str() to show the tmpfs mount options
3082	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3083	 */
3084	if (mode != MPOL_PREFERRED) {
3085		new->nodes = nodes;
3086	} else if (nodelist) {
3087		nodes_clear(new->nodes);
3088		node_set(first_node(nodes), new->nodes);
3089	} else {
3090		new->mode = MPOL_LOCAL;
3091	}
3092
3093	/*
3094	 * Save nodes for contextualization: this will be used to "clone"
3095	 * the mempolicy in a specific context [cpuset] at a later time.
3096	 */
3097	new->w.user_nodemask = nodes;
3098
3099	err = 0;
3100
3101out:
3102	/* Restore string for error message */
3103	if (nodelist)
3104		*--nodelist = ':';
3105	if (flags)
3106		*--flags = '=';
3107	if (!err)
3108		*mpol = new;
3109	return err;
3110}
3111#endif /* CONFIG_TMPFS */
3112
3113/**
3114 * mpol_to_str - format a mempolicy structure for printing
3115 * @buffer:  to contain formatted mempolicy string
3116 * @maxlen:  length of @buffer
3117 * @pol:  pointer to mempolicy to be formatted
3118 *
3119 * Convert @pol into a string.  If @buffer is too short, truncate the string.
3120 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
3121 * longest flag, "relative", and to display at least a few node ids.
3122 */
3123void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3124{
3125	char *p = buffer;
3126	nodemask_t nodes = NODE_MASK_NONE;
3127	unsigned short mode = MPOL_DEFAULT;
3128	unsigned short flags = 0;
3129
3130	if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
3131		mode = pol->mode;
3132		flags = pol->flags;
3133	}
3134
3135	switch (mode) {
3136	case MPOL_DEFAULT:
3137	case MPOL_LOCAL:
3138		break;
3139	case MPOL_PREFERRED:
3140	case MPOL_PREFERRED_MANY:
3141	case MPOL_BIND:
3142	case MPOL_INTERLEAVE:
3143		nodes = pol->nodes;
3144		break;
3145	default:
3146		WARN_ON_ONCE(1);
3147		snprintf(p, maxlen, "unknown");
3148		return;
3149	}
3150
3151	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3152
3153	if (flags & MPOL_MODE_FLAGS) {
3154		p += snprintf(p, buffer + maxlen - p, "=");
3155
3156		/*
3157		 * Currently, the only defined flags are mutually exclusive
3158		 */
3159		if (flags & MPOL_F_STATIC_NODES)
3160			p += snprintf(p, buffer + maxlen - p, "static");
3161		else if (flags & MPOL_F_RELATIVE_NODES)
3162			p += snprintf(p, buffer + maxlen - p, "relative");
3163	}
3164
3165	if (!nodes_empty(nodes))
3166		p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3167			       nodemask_pr_args(&nodes));
3168}