Linux Audio

Check our new training course

Loading...
v4.17
 
   1/*
   2 * Simple NUMA memory policy for the Linux kernel.
   3 *
   4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
   5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
   6 * Subject to the GNU Public License, version 2.
   7 *
   8 * NUMA policy allows the user to give hints in which node(s) memory should
   9 * be allocated.
  10 *
  11 * Support four policies per VMA and per process:
  12 *
  13 * The VMA policy has priority over the process policy for a page fault.
  14 *
  15 * interleave     Allocate memory interleaved over a set of nodes,
  16 *                with normal fallback if it fails.
  17 *                For VMA based allocations this interleaves based on the
  18 *                offset into the backing object or offset into the mapping
  19 *                for anonymous memory. For process policy an process counter
  20 *                is used.
  21 *
 
 
 
 
 
 
 
  22 * bind           Only allocate memory on a specific set of nodes,
  23 *                no fallback.
  24 *                FIXME: memory is allocated starting with the first node
  25 *                to the last. It would be better if bind would truly restrict
  26 *                the allocation to memory nodes instead
  27 *
  28 * preferred       Try a specific node first before normal fallback.
  29 *                As a special case NUMA_NO_NODE here means do the allocation
  30 *                on the local CPU. This is normally identical to default,
  31 *                but useful to set in a VMA when you have a non default
  32 *                process policy.
  33 *
 
 
 
  34 * default        Allocate on the local node first, or when on a VMA
  35 *                use the process policy. This is what Linux always did
  36 *		  in a NUMA aware kernel and still does by, ahem, default.
  37 *
  38 * The process policy is applied for most non interrupt memory allocations
  39 * in that process' context. Interrupts ignore the policies and always
  40 * try to allocate on the local CPU. The VMA policy is only applied for memory
  41 * allocations for a VMA in the VM.
  42 *
  43 * Currently there are a few corner cases in swapping where the policy
  44 * is not applied, but the majority should be handled. When process policy
  45 * is used it is not remembered over swap outs/swap ins.
  46 *
  47 * Only the highest zone in the zone hierarchy gets policied. Allocations
  48 * requesting a lower zone just use default policy. This implies that
  49 * on systems with highmem kernel lowmem allocation don't get policied.
  50 * Same with GFP_DMA allocations.
  51 *
  52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
  53 * all users and remembered even when nobody has memory mapped.
  54 */
  55
  56/* Notebook:
  57   fix mmap readahead to honour policy and enable policy for any page cache
  58   object
  59   statistics for bigpages
  60   global policy for page cache? currently it uses process policy. Requires
  61   first item above.
  62   handle mremap for shared memory (currently ignored for the policy)
  63   grows down?
  64   make bind policy root only? It can trigger oom much faster and the
  65   kernel is not always grateful with that.
  66*/
  67
  68#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  69
  70#include <linux/mempolicy.h>
  71#include <linux/mm.h>
  72#include <linux/highmem.h>
  73#include <linux/hugetlb.h>
  74#include <linux/kernel.h>
  75#include <linux/sched.h>
  76#include <linux/sched/mm.h>
  77#include <linux/sched/numa_balancing.h>
  78#include <linux/sched/task.h>
  79#include <linux/nodemask.h>
  80#include <linux/cpuset.h>
  81#include <linux/slab.h>
  82#include <linux/string.h>
  83#include <linux/export.h>
  84#include <linux/nsproxy.h>
  85#include <linux/interrupt.h>
  86#include <linux/init.h>
  87#include <linux/compat.h>
  88#include <linux/ptrace.h>
  89#include <linux/swap.h>
  90#include <linux/seq_file.h>
  91#include <linux/proc_fs.h>
  92#include <linux/migrate.h>
  93#include <linux/ksm.h>
  94#include <linux/rmap.h>
  95#include <linux/security.h>
  96#include <linux/syscalls.h>
  97#include <linux/ctype.h>
  98#include <linux/mm_inline.h>
  99#include <linux/mmu_notifier.h>
 100#include <linux/printk.h>
 101#include <linux/swapops.h>
 102
 103#include <asm/tlbflush.h>
 
 104#include <linux/uaccess.h>
 105
 106#include "internal.h"
 107
 108/* Internal flags */
 109#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)	/* Skip checks for continuous vmas */
 110#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)		/* Invert check for nodemask */
 
 111
 112static struct kmem_cache *policy_cache;
 113static struct kmem_cache *sn_cache;
 114
 115/* Highest zone. An specific allocation for a zone below that is not
 116   policied. */
 117enum zone_type policy_zone = 0;
 118
 119/*
 120 * run-time system-wide default policy => local allocation
 121 */
 122static struct mempolicy default_policy = {
 123	.refcnt = ATOMIC_INIT(1), /* never free it */
 124	.mode = MPOL_PREFERRED,
 125	.flags = MPOL_F_LOCAL,
 126};
 127
 128static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 129
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 130struct mempolicy *get_task_policy(struct task_struct *p)
 131{
 132	struct mempolicy *pol = p->mempolicy;
 133	int node;
 134
 135	if (pol)
 136		return pol;
 137
 138	node = numa_node_id();
 139	if (node != NUMA_NO_NODE) {
 140		pol = &preferred_node_policy[node];
 141		/* preferred_node_policy is not initialised early in boot */
 142		if (pol->mode)
 143			return pol;
 144	}
 145
 146	return &default_policy;
 147}
 148
 149static const struct mempolicy_operations {
 150	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 151	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
 152} mpol_ops[MPOL_MAX];
 153
 154static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 155{
 156	return pol->flags & MPOL_MODE_FLAGS;
 157}
 158
 159static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 160				   const nodemask_t *rel)
 161{
 162	nodemask_t tmp;
 163	nodes_fold(tmp, *orig, nodes_weight(*rel));
 164	nodes_onto(*ret, tmp, *rel);
 165}
 166
 167static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
 168{
 169	if (nodes_empty(*nodes))
 170		return -EINVAL;
 171	pol->v.nodes = *nodes;
 172	return 0;
 173}
 174
 175static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 176{
 177	if (!nodes)
 178		pol->flags |= MPOL_F_LOCAL;	/* local allocation */
 179	else if (nodes_empty(*nodes))
 180		return -EINVAL;			/*  no allowed nodes */
 181	else
 182		pol->v.preferred_node = first_node(*nodes);
 183	return 0;
 184}
 185
 186static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
 187{
 188	if (nodes_empty(*nodes))
 189		return -EINVAL;
 190	pol->v.nodes = *nodes;
 
 
 191	return 0;
 192}
 193
 194/*
 195 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 196 * any, for the new policy.  mpol_new() has already validated the nodes
 197 * parameter with respect to the policy mode and flags.  But, we need to
 198 * handle an empty nodemask with MPOL_PREFERRED here.
 199 *
 200 * Must be called holding task's alloc_lock to protect task's mems_allowed
 201 * and mempolicy.  May also be called holding the mmap_semaphore for write.
 202 */
 203static int mpol_set_nodemask(struct mempolicy *pol,
 204		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 205{
 206	int ret;
 207
 208	/* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
 209	if (pol == NULL)
 
 
 
 
 210		return 0;
 
 211	/* Check N_MEMORY */
 212	nodes_and(nsc->mask1,
 213		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 214
 215	VM_BUG_ON(!nodes);
 216	if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
 217		nodes = NULL;	/* explicit local allocation */
 218	else {
 219		if (pol->flags & MPOL_F_RELATIVE_NODES)
 220			mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
 221		else
 222			nodes_and(nsc->mask2, *nodes, nsc->mask1);
 223
 224		if (mpol_store_user_nodemask(pol))
 225			pol->w.user_nodemask = *nodes;
 226		else
 227			pol->w.cpuset_mems_allowed =
 228						cpuset_current_mems_allowed;
 229	}
 230
 231	if (nodes)
 232		ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 233	else
 234		ret = mpol_ops[pol->mode].create(pol, NULL);
 
 
 235	return ret;
 236}
 237
 238/*
 239 * This function just creates a new policy, does some check and simple
 240 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 241 */
 242static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 243				  nodemask_t *nodes)
 244{
 245	struct mempolicy *policy;
 246
 247	pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 248		 mode, flags, nodes ? nodes_addr(*nodes)[0] : NUMA_NO_NODE);
 249
 250	if (mode == MPOL_DEFAULT) {
 251		if (nodes && !nodes_empty(*nodes))
 252			return ERR_PTR(-EINVAL);
 253		return NULL;
 254	}
 255	VM_BUG_ON(!nodes);
 256
 257	/*
 258	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 259	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 260	 * All other modes require a valid pointer to a non-empty nodemask.
 261	 */
 262	if (mode == MPOL_PREFERRED) {
 263		if (nodes_empty(*nodes)) {
 264			if (((flags & MPOL_F_STATIC_NODES) ||
 265			     (flags & MPOL_F_RELATIVE_NODES)))
 266				return ERR_PTR(-EINVAL);
 
 
 267		}
 268	} else if (mode == MPOL_LOCAL) {
 269		if (!nodes_empty(*nodes) ||
 270		    (flags & MPOL_F_STATIC_NODES) ||
 271		    (flags & MPOL_F_RELATIVE_NODES))
 272			return ERR_PTR(-EINVAL);
 273		mode = MPOL_PREFERRED;
 274	} else if (nodes_empty(*nodes))
 275		return ERR_PTR(-EINVAL);
 
 276	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 277	if (!policy)
 278		return ERR_PTR(-ENOMEM);
 279	atomic_set(&policy->refcnt, 1);
 280	policy->mode = mode;
 281	policy->flags = flags;
 
 282
 283	return policy;
 284}
 285
 286/* Slow path of a mpol destructor. */
 287void __mpol_put(struct mempolicy *p)
 288{
 289	if (!atomic_dec_and_test(&p->refcnt))
 290		return;
 291	kmem_cache_free(policy_cache, p);
 292}
 293
 294static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
 295{
 296}
 297
 298static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 299{
 300	nodemask_t tmp;
 301
 302	if (pol->flags & MPOL_F_STATIC_NODES)
 303		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 304	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 305		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 306	else {
 307		nodes_remap(tmp, pol->v.nodes,pol->w.cpuset_mems_allowed,
 308								*nodes);
 309		pol->w.cpuset_mems_allowed = tmp;
 310	}
 311
 312	if (nodes_empty(tmp))
 313		tmp = *nodes;
 314
 315	pol->v.nodes = tmp;
 316}
 317
 318static void mpol_rebind_preferred(struct mempolicy *pol,
 319						const nodemask_t *nodes)
 320{
 321	nodemask_t tmp;
 322
 323	if (pol->flags & MPOL_F_STATIC_NODES) {
 324		int node = first_node(pol->w.user_nodemask);
 325
 326		if (node_isset(node, *nodes)) {
 327			pol->v.preferred_node = node;
 328			pol->flags &= ~MPOL_F_LOCAL;
 329		} else
 330			pol->flags |= MPOL_F_LOCAL;
 331	} else if (pol->flags & MPOL_F_RELATIVE_NODES) {
 332		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 333		pol->v.preferred_node = first_node(tmp);
 334	} else if (!(pol->flags & MPOL_F_LOCAL)) {
 335		pol->v.preferred_node = node_remap(pol->v.preferred_node,
 336						   pol->w.cpuset_mems_allowed,
 337						   *nodes);
 338		pol->w.cpuset_mems_allowed = *nodes;
 339	}
 340}
 341
 342/*
 343 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 344 *
 345 * Per-vma policies are protected by mmap_sem. Allocations using per-task
 346 * policies are protected by task->mems_allowed_seq to prevent a premature
 347 * OOM/allocation failure due to parallel nodemask modification.
 348 */
 349static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
 350{
 351	if (!pol)
 352		return;
 353	if (!mpol_store_user_nodemask(pol) &&
 354	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 355		return;
 356
 357	mpol_ops[pol->mode].rebind(pol, newmask);
 358}
 359
 360/*
 361 * Wrapper for mpol_rebind_policy() that just requires task
 362 * pointer, and updates task mempolicy.
 363 *
 364 * Called with task's alloc_lock held.
 365 */
 366
 367void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
 368{
 369	mpol_rebind_policy(tsk->mempolicy, new);
 370}
 371
 372/*
 373 * Rebind each vma in mm to new nodemask.
 374 *
 375 * Call holding a reference to mm.  Takes mm->mmap_sem during call.
 376 */
 377
 378void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 379{
 380	struct vm_area_struct *vma;
 
 381
 382	down_write(&mm->mmap_sem);
 383	for (vma = mm->mmap; vma; vma = vma->vm_next)
 
 384		mpol_rebind_policy(vma->vm_policy, new);
 385	up_write(&mm->mmap_sem);
 
 386}
 387
 388static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 389	[MPOL_DEFAULT] = {
 390		.rebind = mpol_rebind_default,
 391	},
 392	[MPOL_INTERLEAVE] = {
 393		.create = mpol_new_interleave,
 394		.rebind = mpol_rebind_nodemask,
 395	},
 396	[MPOL_PREFERRED] = {
 397		.create = mpol_new_preferred,
 398		.rebind = mpol_rebind_preferred,
 399	},
 400	[MPOL_BIND] = {
 401		.create = mpol_new_bind,
 
 
 
 
 
 
 
 
 
 
 
 402		.rebind = mpol_rebind_nodemask,
 403	},
 404};
 405
 406static void migrate_page_add(struct page *page, struct list_head *pagelist,
 407				unsigned long flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 408
 409struct queue_pages {
 410	struct list_head *pagelist;
 411	unsigned long flags;
 412	nodemask_t *nmask;
 413	struct vm_area_struct *prev;
 
 
 
 
 414};
 415
 416/*
 417 * Check if the page's nid is in qp->nmask.
 418 *
 419 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
 420 * in the invert of qp->nmask.
 421 */
 422static inline bool queue_pages_required(struct page *page,
 423					struct queue_pages *qp)
 424{
 425	int nid = page_to_nid(page);
 426	unsigned long flags = qp->flags;
 427
 428	return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
 429}
 430
 431static int queue_pages_pmd(pmd_t *pmd, spinlock_t *ptl, unsigned long addr,
 432				unsigned long end, struct mm_walk *walk)
 433{
 434	int ret = 0;
 435	struct page *page;
 436	struct queue_pages *qp = walk->private;
 437	unsigned long flags;
 438
 439	if (unlikely(is_pmd_migration_entry(*pmd))) {
 440		ret = 1;
 441		goto unlock;
 442	}
 443	page = pmd_page(*pmd);
 444	if (is_huge_zero_page(page)) {
 445		spin_unlock(ptl);
 446		__split_huge_pmd(walk->vma, pmd, addr, false, NULL);
 447		goto out;
 448	}
 449	if (!queue_pages_required(page, qp)) {
 450		ret = 1;
 451		goto unlock;
 
 452	}
 453
 454	ret = 1;
 455	flags = qp->flags;
 456	/* go to thp migration */
 457	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
 458		migrate_page_add(page, qp->pagelist, flags);
 459unlock:
 460	spin_unlock(ptl);
 461out:
 462	return ret;
 463}
 464
 465/*
 466 * Scan through pages checking if pages follow certain conditions,
 467 * and move them to the pagelist if they do.
 
 
 
 
 
 
 468 */
 469static int queue_pages_pte_range(pmd_t *pmd, unsigned long addr,
 470			unsigned long end, struct mm_walk *walk)
 471{
 472	struct vm_area_struct *vma = walk->vma;
 473	struct page *page;
 474	struct queue_pages *qp = walk->private;
 475	unsigned long flags = qp->flags;
 476	int ret;
 477	pte_t *pte;
 478	spinlock_t *ptl;
 479
 480	ptl = pmd_trans_huge_lock(pmd, vma);
 481	if (ptl) {
 482		ret = queue_pages_pmd(pmd, ptl, addr, end, walk);
 483		if (ret)
 484			return 0;
 485	}
 486
 487	if (pmd_trans_unstable(pmd))
 
 
 488		return 0;
 489
 490	pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
 491	for (; addr != end; pte++, addr += PAGE_SIZE) {
 492		if (!pte_present(*pte))
 
 
 
 
 
 493			continue;
 494		page = vm_normal_page(vma, addr, *pte);
 495		if (!page)
 
 496			continue;
 497		/*
 498		 * vm_normal_page() filters out zero pages, but there might
 499		 * still be PageReserved pages to skip, perhaps in a VDSO.
 500		 */
 501		if (PageReserved(page))
 502			continue;
 503		if (!queue_pages_required(page, qp))
 504			continue;
 505		migrate_page_add(page, qp->pagelist, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 506	}
 507	pte_unmap_unlock(pte - 1, ptl);
 508	cond_resched();
 
 
 
 509	return 0;
 510}
 511
 512static int queue_pages_hugetlb(pte_t *pte, unsigned long hmask,
 513			       unsigned long addr, unsigned long end,
 514			       struct mm_walk *walk)
 515{
 516#ifdef CONFIG_HUGETLB_PAGE
 517	struct queue_pages *qp = walk->private;
 518	unsigned long flags = qp->flags;
 519	struct page *page;
 520	spinlock_t *ptl;
 521	pte_t entry;
 522
 523	ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 524	entry = huge_ptep_get(pte);
 525	if (!pte_present(entry))
 
 
 
 
 
 
 526		goto unlock;
 527	page = pte_page(entry);
 528	if (!queue_pages_required(page, qp))
 
 529		goto unlock;
 530	/* With MPOL_MF_MOVE, we migrate only unshared hugepage. */
 531	if (flags & (MPOL_MF_MOVE_ALL) ||
 532	    (flags & MPOL_MF_MOVE && page_mapcount(page) == 1))
 533		isolate_huge_page(page, qp->pagelist);
 
 
 
 
 
 
 
 
 534unlock:
 535	spin_unlock(ptl);
 536#else
 537	BUG();
 538#endif
 539	return 0;
 540}
 541
 542#ifdef CONFIG_NUMA_BALANCING
 543/*
 544 * This is used to mark a range of virtual addresses to be inaccessible.
 545 * These are later cleared by a NUMA hinting fault. Depending on these
 546 * faults, pages may be migrated for better NUMA placement.
 547 *
 548 * This is assuming that NUMA faults are handled using PROT_NONE. If
 549 * an architecture makes a different choice, it will need further
 550 * changes to the core.
 551 */
 552unsigned long change_prot_numa(struct vm_area_struct *vma,
 553			unsigned long addr, unsigned long end)
 554{
 555	int nr_updated;
 
 
 
 556
 557	nr_updated = change_protection(vma, addr, end, PAGE_NONE, 0, 1);
 558	if (nr_updated)
 559		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 
 
 
 
 560
 561	return nr_updated;
 562}
 563#else
 564static unsigned long change_prot_numa(struct vm_area_struct *vma,
 565			unsigned long addr, unsigned long end)
 566{
 567	return 0;
 568}
 569#endif /* CONFIG_NUMA_BALANCING */
 570
 571static int queue_pages_test_walk(unsigned long start, unsigned long end,
 572				struct mm_walk *walk)
 573{
 574	struct vm_area_struct *vma = walk->vma;
 575	struct queue_pages *qp = walk->private;
 576	unsigned long endvma = vma->vm_end;
 577	unsigned long flags = qp->flags;
 578
 579	if (!vma_migratable(vma))
 580		return 1;
 581
 582	if (endvma > end)
 583		endvma = end;
 584	if (vma->vm_start > start)
 585		start = vma->vm_start;
 586
 587	if (!(flags & MPOL_MF_DISCONTIG_OK)) {
 588		if (!vma->vm_next && vma->vm_end < end)
 589			return -EFAULT;
 590		if (qp->prev && qp->prev->vm_end < vma->vm_start)
 
 591			return -EFAULT;
 592	}
 
 
 
 
 
 
 593
 594	qp->prev = vma;
 595
 596	if (flags & MPOL_MF_LAZY) {
 597		/* Similar to task_numa_work, skip inaccessible VMAs */
 598		if (!is_vm_hugetlb_page(vma) &&
 599			(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)) &&
 600			!(vma->vm_flags & VM_MIXEDMAP))
 601			change_prot_numa(vma, start, endvma);
 602		return 1;
 603	}
 604
 605	/* queue pages from current vma */
 606	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
 
 
 
 607		return 0;
 608	return 1;
 609}
 610
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 611/*
 612 * Walk through page tables and collect pages to be migrated.
 613 *
 614 * If pages found in a given range are on a set of nodes (determined by
 615 * @nodes and @flags,) it's isolated and queued to the pagelist which is
 616 * passed via @private.)
 
 
 
 
 
 
 
 617 */
 618static int
 619queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 620		nodemask_t *nodes, unsigned long flags,
 621		struct list_head *pagelist)
 622{
 
 623	struct queue_pages qp = {
 624		.pagelist = pagelist,
 625		.flags = flags,
 626		.nmask = nodes,
 627		.prev = NULL,
 628	};
 629	struct mm_walk queue_pages_walk = {
 630		.hugetlb_entry = queue_pages_hugetlb,
 631		.pmd_entry = queue_pages_pte_range,
 632		.test_walk = queue_pages_test_walk,
 633		.mm = mm,
 634		.private = &qp,
 635	};
 
 
 
 
 
 
 
 
 636
 637	return walk_page_range(start, end, &queue_pages_walk);
 638}
 639
 640/*
 641 * Apply policy to a single VMA
 642 * This must be called with the mmap_sem held for writing.
 643 */
 644static int vma_replace_policy(struct vm_area_struct *vma,
 645						struct mempolicy *pol)
 646{
 647	int err;
 648	struct mempolicy *old;
 649	struct mempolicy *new;
 650
 651	pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 652		 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 653		 vma->vm_ops, vma->vm_file,
 654		 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 655
 656	new = mpol_dup(pol);
 657	if (IS_ERR(new))
 658		return PTR_ERR(new);
 659
 660	if (vma->vm_ops && vma->vm_ops->set_policy) {
 661		err = vma->vm_ops->set_policy(vma, new);
 662		if (err)
 663			goto err_out;
 664	}
 665
 666	old = vma->vm_policy;
 667	vma->vm_policy = new; /* protected by mmap_sem */
 668	mpol_put(old);
 669
 670	return 0;
 671 err_out:
 672	mpol_put(new);
 673	return err;
 674}
 675
 676/* Step 2: apply policy to a range and do splits. */
 677static int mbind_range(struct mm_struct *mm, unsigned long start,
 678		       unsigned long end, struct mempolicy *new_pol)
 679{
 680	struct vm_area_struct *next;
 681	struct vm_area_struct *prev;
 682	struct vm_area_struct *vma;
 683	int err = 0;
 684	pgoff_t pgoff;
 685	unsigned long vmstart;
 686	unsigned long vmend;
 687
 688	vma = find_vma(mm, start);
 689	if (!vma || vma->vm_start > start)
 690		return -EFAULT;
 691
 692	prev = vma->vm_prev;
 693	if (start > vma->vm_start)
 694		prev = vma;
 695
 696	for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 697		next = vma->vm_next;
 698		vmstart = max(start, vma->vm_start);
 699		vmend   = min(end, vma->vm_end);
 700
 701		if (mpol_equal(vma_policy(vma), new_pol))
 702			continue;
 703
 704		pgoff = vma->vm_pgoff +
 705			((vmstart - vma->vm_start) >> PAGE_SHIFT);
 706		prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 707				 vma->anon_vma, vma->vm_file, pgoff,
 708				 new_pol, vma->vm_userfaultfd_ctx);
 709		if (prev) {
 710			vma = prev;
 711			next = vma->vm_next;
 712			if (mpol_equal(vma_policy(vma), new_pol))
 713				continue;
 714			/* vma_merge() joined vma && vma->next, case 8 */
 715			goto replace;
 716		}
 717		if (vma->vm_start != vmstart) {
 718			err = split_vma(vma->vm_mm, vma, vmstart, 1);
 719			if (err)
 720				goto out;
 721		}
 722		if (vma->vm_end != vmend) {
 723			err = split_vma(vma->vm_mm, vma, vmend, 0);
 724			if (err)
 725				goto out;
 726		}
 727 replace:
 728		err = vma_replace_policy(vma, new_pol);
 729		if (err)
 730			goto out;
 731	}
 732
 733 out:
 734	return err;
 
 
 
 
 735}
 736
 737/* Set the process memory policy */
 738static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 739			     nodemask_t *nodes)
 740{
 741	struct mempolicy *new, *old;
 742	NODEMASK_SCRATCH(scratch);
 743	int ret;
 744
 745	if (!scratch)
 746		return -ENOMEM;
 747
 748	new = mpol_new(mode, flags, nodes);
 749	if (IS_ERR(new)) {
 750		ret = PTR_ERR(new);
 751		goto out;
 752	}
 753
 754	task_lock(current);
 755	ret = mpol_set_nodemask(new, nodes, scratch);
 756	if (ret) {
 757		task_unlock(current);
 758		mpol_put(new);
 759		goto out;
 760	}
 
 761	old = current->mempolicy;
 762	current->mempolicy = new;
 763	if (new && new->mode == MPOL_INTERLEAVE)
 
 764		current->il_prev = MAX_NUMNODES-1;
 
 
 765	task_unlock(current);
 766	mpol_put(old);
 767	ret = 0;
 768out:
 769	NODEMASK_SCRATCH_FREE(scratch);
 770	return ret;
 771}
 772
 773/*
 774 * Return nodemask for policy for get_mempolicy() query
 775 *
 776 * Called with task's alloc_lock held
 777 */
 778static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 779{
 780	nodes_clear(*nodes);
 781	if (p == &default_policy)
 782		return;
 783
 784	switch (p->mode) {
 785	case MPOL_BIND:
 786		/* Fall through */
 787	case MPOL_INTERLEAVE:
 788		*nodes = p->v.nodes;
 789		break;
 790	case MPOL_PREFERRED:
 791		if (!(p->flags & MPOL_F_LOCAL))
 792			node_set(p->v.preferred_node, *nodes);
 793		/* else return empty node mask for local allocation */
 
 
 
 794		break;
 795	default:
 796		BUG();
 797	}
 798}
 799
 800static int lookup_node(unsigned long addr)
 801{
 802	struct page *p;
 803	int err;
 804
 805	err = get_user_pages(addr & PAGE_MASK, 1, 0, &p, NULL);
 806	if (err >= 0) {
 807		err = page_to_nid(p);
 808		put_page(p);
 809	}
 810	return err;
 811}
 812
 813/* Retrieve NUMA policy */
 814static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 815			     unsigned long addr, unsigned long flags)
 816{
 817	int err;
 818	struct mm_struct *mm = current->mm;
 819	struct vm_area_struct *vma = NULL;
 820	struct mempolicy *pol = current->mempolicy;
 821
 822	if (flags &
 823		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 824		return -EINVAL;
 825
 826	if (flags & MPOL_F_MEMS_ALLOWED) {
 827		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 828			return -EINVAL;
 829		*policy = 0;	/* just so it's initialized */
 830		task_lock(current);
 831		*nmask  = cpuset_current_mems_allowed;
 832		task_unlock(current);
 833		return 0;
 834	}
 835
 836	if (flags & MPOL_F_ADDR) {
 
 837		/*
 838		 * Do NOT fall back to task policy if the
 839		 * vma/shared policy at addr is NULL.  We
 840		 * want to return MPOL_DEFAULT in this case.
 841		 */
 842		down_read(&mm->mmap_sem);
 843		vma = find_vma_intersection(mm, addr, addr+1);
 844		if (!vma) {
 845			up_read(&mm->mmap_sem);
 846			return -EFAULT;
 847		}
 848		if (vma->vm_ops && vma->vm_ops->get_policy)
 849			pol = vma->vm_ops->get_policy(vma, addr);
 850		else
 851			pol = vma->vm_policy;
 852	} else if (addr)
 853		return -EINVAL;
 854
 855	if (!pol)
 856		pol = &default_policy;	/* indicates default behavior */
 857
 858	if (flags & MPOL_F_NODE) {
 859		if (flags & MPOL_F_ADDR) {
 860			err = lookup_node(addr);
 
 
 
 
 
 
 
 
 
 861			if (err < 0)
 862				goto out;
 863			*policy = err;
 864		} else if (pol == current->mempolicy &&
 865				pol->mode == MPOL_INTERLEAVE) {
 866			*policy = next_node_in(current->il_prev, pol->v.nodes);
 
 
 
 
 
 
 
 867		} else {
 868			err = -EINVAL;
 869			goto out;
 870		}
 871	} else {
 872		*policy = pol == &default_policy ? MPOL_DEFAULT :
 873						pol->mode;
 874		/*
 875		 * Internal mempolicy flags must be masked off before exposing
 876		 * the policy to userspace.
 877		 */
 878		*policy |= (pol->flags & MPOL_MODE_FLAGS);
 879	}
 880
 881	err = 0;
 882	if (nmask) {
 883		if (mpol_store_user_nodemask(pol)) {
 884			*nmask = pol->w.user_nodemask;
 885		} else {
 886			task_lock(current);
 887			get_policy_nodemask(pol, nmask);
 888			task_unlock(current);
 889		}
 890	}
 891
 892 out:
 893	mpol_cond_put(pol);
 894	if (vma)
 895		up_read(&current->mm->mmap_sem);
 
 
 896	return err;
 897}
 898
 899#ifdef CONFIG_MIGRATION
 900/*
 901 * page migration, thp tail pages can be passed.
 902 */
 903static void migrate_page_add(struct page *page, struct list_head *pagelist,
 904				unsigned long flags)
 905{
 906	struct page *head = compound_head(page);
 907	/*
 908	 * Avoid migrating a page that is shared with others.
 
 
 
 
 909	 */
 910	if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(head) == 1) {
 911		if (!isolate_lru_page(head)) {
 912			list_add_tail(&head->lru, pagelist);
 913			mod_node_page_state(page_pgdat(head),
 914				NR_ISOLATED_ANON + page_is_file_cache(head),
 915				hpage_nr_pages(head));
 
 
 
 
 
 
 
 
 916		}
 917	}
 918}
 919
 920/* page allocation callback for NUMA node migration */
 921struct page *alloc_new_node_page(struct page *page, unsigned long node)
 922{
 923	if (PageHuge(page))
 924		return alloc_huge_page_node(page_hstate(compound_head(page)),
 925					node);
 926	else if (PageTransHuge(page)) {
 927		struct page *thp;
 928
 929		thp = alloc_pages_node(node,
 930			(GFP_TRANSHUGE | __GFP_THISNODE),
 931			HPAGE_PMD_ORDER);
 932		if (!thp)
 933			return NULL;
 934		prep_transhuge_page(thp);
 935		return thp;
 936	} else
 937		return __alloc_pages_node(node, GFP_HIGHUSER_MOVABLE |
 938						    __GFP_THISNODE, 0);
 939}
 940
 941/*
 942 * Migrate pages from one node to a target node.
 943 * Returns error or the number of pages not migrated.
 944 */
 945static int migrate_to_node(struct mm_struct *mm, int source, int dest,
 946			   int flags)
 947{
 948	nodemask_t nmask;
 
 949	LIST_HEAD(pagelist);
 950	int err = 0;
 
 
 
 
 
 
 951
 952	nodes_clear(nmask);
 953	node_set(source, nmask);
 954
 
 
 
 
 
 
 
 
 
 955	/*
 956	 * This does not "check" the range but isolates all pages that
 957	 * need migration.  Between passing in the full user address
 958	 * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
 
 959	 */
 960	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
 961	queue_pages_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
 962			flags | MPOL_MF_DISCONTIG_OK, &pagelist);
 963
 964	if (!list_empty(&pagelist)) {
 965		err = migrate_pages(&pagelist, alloc_new_node_page, NULL, dest,
 966					MIGRATE_SYNC, MR_SYSCALL);
 967		if (err)
 968			putback_movable_pages(&pagelist);
 969	}
 970
 
 
 971	return err;
 972}
 973
 974/*
 975 * Move pages between the two nodesets so as to preserve the physical
 976 * layout as much as possible.
 977 *
 978 * Returns the number of page that could not be moved.
 979 */
 980int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
 981		     const nodemask_t *to, int flags)
 982{
 983	int busy = 0;
 984	int err;
 985	nodemask_t tmp;
 986
 987	err = migrate_prep();
 988	if (err)
 989		return err;
 990
 991	down_read(&mm->mmap_sem);
 992
 993	/*
 994	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
 995	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
 996	 * bit in 'tmp', and return that <source, dest> pair for migration.
 997	 * The pair of nodemasks 'to' and 'from' define the map.
 998	 *
 999	 * If no pair of bits is found that way, fallback to picking some
1000	 * pair of 'source' and 'dest' bits that are not the same.  If the
1001	 * 'source' and 'dest' bits are the same, this represents a node
1002	 * that will be migrating to itself, so no pages need move.
1003	 *
1004	 * If no bits are left in 'tmp', or if all remaining bits left
1005	 * in 'tmp' correspond to the same bit in 'to', return false
1006	 * (nothing left to migrate).
1007	 *
1008	 * This lets us pick a pair of nodes to migrate between, such that
1009	 * if possible the dest node is not already occupied by some other
1010	 * source node, minimizing the risk of overloading the memory on a
1011	 * node that would happen if we migrated incoming memory to a node
1012	 * before migrating outgoing memory source that same node.
1013	 *
1014	 * A single scan of tmp is sufficient.  As we go, we remember the
1015	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1016	 * that not only moved, but what's better, moved to an empty slot
1017	 * (d is not set in tmp), then we break out then, with that pair.
1018	 * Otherwise when we finish scanning from_tmp, we at least have the
1019	 * most recent <s, d> pair that moved.  If we get all the way through
1020	 * the scan of tmp without finding any node that moved, much less
1021	 * moved to an empty node, then there is nothing left worth migrating.
1022	 */
1023
1024	tmp = *from;
1025	while (!nodes_empty(tmp)) {
1026		int s,d;
1027		int source = NUMA_NO_NODE;
1028		int dest = 0;
1029
1030		for_each_node_mask(s, tmp) {
1031
1032			/*
1033			 * do_migrate_pages() tries to maintain the relative
1034			 * node relationship of the pages established between
1035			 * threads and memory areas.
1036                         *
1037			 * However if the number of source nodes is not equal to
1038			 * the number of destination nodes we can not preserve
1039			 * this node relative relationship.  In that case, skip
1040			 * copying memory from a node that is in the destination
1041			 * mask.
1042			 *
1043			 * Example: [2,3,4] -> [3,4,5] moves everything.
1044			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1045			 */
1046
1047			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1048						(node_isset(s, *to)))
1049				continue;
1050
1051			d = node_remap(s, *from, *to);
1052			if (s == d)
1053				continue;
1054
1055			source = s;	/* Node moved. Memorize */
1056			dest = d;
1057
1058			/* dest not in remaining from nodes? */
1059			if (!node_isset(dest, tmp))
1060				break;
1061		}
1062		if (source == NUMA_NO_NODE)
1063			break;
1064
1065		node_clear(source, tmp);
1066		err = migrate_to_node(mm, source, dest, flags);
1067		if (err > 0)
1068			busy += err;
1069		if (err < 0)
1070			break;
1071	}
1072	up_read(&mm->mmap_sem);
 
1073	if (err < 0)
1074		return err;
1075	return busy;
1076
1077}
1078
1079/*
1080 * Allocate a new page for page migration based on vma policy.
1081 * Start by assuming the page is mapped by the same vma as contains @start.
1082 * Search forward from there, if not.  N.B., this assumes that the
1083 * list of pages handed to migrate_pages()--which is how we get here--
1084 * is in virtual address order.
1085 */
1086static struct page *new_page(struct page *page, unsigned long start)
 
1087{
1088	struct vm_area_struct *vma;
1089	unsigned long uninitialized_var(address);
 
 
 
 
1090
1091	vma = find_vma(current->mm, start);
1092	while (vma) {
1093		address = page_address_in_vma(page, vma);
1094		if (address != -EFAULT)
1095			break;
1096		vma = vma->vm_next;
1097	}
1098
1099	if (PageHuge(page)) {
1100		return alloc_huge_page_vma(page_hstate(compound_head(page)),
1101				vma, address);
1102	} else if (PageTransHuge(page)) {
1103		struct page *thp;
1104
1105		thp = alloc_hugepage_vma(GFP_TRANSHUGE, vma, address,
1106					 HPAGE_PMD_ORDER);
1107		if (!thp)
1108			return NULL;
1109		prep_transhuge_page(thp);
1110		return thp;
1111	}
1112	/*
1113	 * if !vma, alloc_page_vma() will use task or system default policy
1114	 */
1115	return alloc_page_vma(GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL,
1116			vma, address);
 
 
1117}
1118#else
1119
1120static void migrate_page_add(struct page *page, struct list_head *pagelist,
1121				unsigned long flags)
1122{
 
1123}
1124
1125int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1126		     const nodemask_t *to, int flags)
1127{
1128	return -ENOSYS;
1129}
1130
1131static struct page *new_page(struct page *page, unsigned long start)
 
1132{
1133	return NULL;
1134}
1135#endif
1136
1137static long do_mbind(unsigned long start, unsigned long len,
1138		     unsigned short mode, unsigned short mode_flags,
1139		     nodemask_t *nmask, unsigned long flags)
1140{
1141	struct mm_struct *mm = current->mm;
 
 
 
1142	struct mempolicy *new;
1143	unsigned long end;
1144	int err;
 
1145	LIST_HEAD(pagelist);
1146
1147	if (flags & ~(unsigned long)MPOL_MF_VALID)
1148		return -EINVAL;
1149	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1150		return -EPERM;
1151
1152	if (start & ~PAGE_MASK)
1153		return -EINVAL;
1154
1155	if (mode == MPOL_DEFAULT)
1156		flags &= ~MPOL_MF_STRICT;
1157
1158	len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1159	end = start + len;
1160
1161	if (end < start)
1162		return -EINVAL;
1163	if (end == start)
1164		return 0;
1165
1166	new = mpol_new(mode, mode_flags, nmask);
1167	if (IS_ERR(new))
1168		return PTR_ERR(new);
1169
1170	if (flags & MPOL_MF_LAZY)
1171		new->flags |= MPOL_F_MOF;
1172
1173	/*
1174	 * If we are using the default policy then operation
1175	 * on discontinuous address spaces is okay after all
1176	 */
1177	if (!new)
1178		flags |= MPOL_MF_DISCONTIG_OK;
1179
1180	pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1181		 start, start + len, mode, mode_flags,
1182		 nmask ? nodes_addr(*nmask)[0] : NUMA_NO_NODE);
1183
1184	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1185
1186		err = migrate_prep();
1187		if (err)
1188			goto mpol_out;
1189	}
1190	{
1191		NODEMASK_SCRATCH(scratch);
1192		if (scratch) {
1193			down_write(&mm->mmap_sem);
1194			task_lock(current);
1195			err = mpol_set_nodemask(new, nmask, scratch);
1196			task_unlock(current);
1197			if (err)
1198				up_write(&mm->mmap_sem);
1199		} else
1200			err = -ENOMEM;
1201		NODEMASK_SCRATCH_FREE(scratch);
1202	}
1203	if (err)
1204		goto mpol_out;
1205
1206	err = queue_pages_range(mm, start, end, nmask,
1207			  flags | MPOL_MF_INVERT, &pagelist);
1208	if (!err)
1209		err = mbind_range(mm, start, end, new);
 
 
1210
1211	if (!err) {
1212		int nr_failed = 0;
 
 
 
 
 
 
 
 
 
 
1213
1214		if (!list_empty(&pagelist)) {
1215			WARN_ON_ONCE(flags & MPOL_MF_LAZY);
1216			nr_failed = migrate_pages(&pagelist, new_page, NULL,
1217				start, MIGRATE_SYNC, MR_MEMPOLICY_MBIND);
1218			if (nr_failed)
1219				putback_movable_pages(&pagelist);
1220		}
 
 
1221
1222		if (nr_failed && (flags & MPOL_MF_STRICT))
1223			err = -EIO;
1224	} else
1225		putback_movable_pages(&pagelist);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1226
1227	up_write(&mm->mmap_sem);
1228 mpol_out:
 
 
 
1229	mpol_put(new);
 
 
1230	return err;
1231}
1232
1233/*
1234 * User space interface with variable sized bitmaps for nodelists.
1235 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1236
1237/* Copy a node mask from user space. */
1238static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1239		     unsigned long maxnode)
1240{
1241	unsigned long k;
1242	unsigned long t;
1243	unsigned long nlongs;
1244	unsigned long endmask;
1245
1246	--maxnode;
1247	nodes_clear(*nodes);
1248	if (maxnode == 0 || !nmask)
1249		return 0;
1250	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1251		return -EINVAL;
1252
1253	nlongs = BITS_TO_LONGS(maxnode);
1254	if ((maxnode % BITS_PER_LONG) == 0)
1255		endmask = ~0UL;
1256	else
1257		endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1258
1259	/*
1260	 * When the user specified more nodes than supported just check
1261	 * if the non supported part is all zero.
1262	 *
1263	 * If maxnode have more longs than MAX_NUMNODES, check
1264	 * the bits in that area first. And then go through to
1265	 * check the rest bits which equal or bigger than MAX_NUMNODES.
1266	 * Otherwise, just check bits [MAX_NUMNODES, maxnode).
1267	 */
1268	if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1269		for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1270			if (get_user(t, nmask + k))
1271				return -EFAULT;
1272			if (k == nlongs - 1) {
1273				if (t & endmask)
1274					return -EINVAL;
1275			} else if (t)
1276				return -EINVAL;
1277		}
1278		nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1279		endmask = ~0UL;
1280	}
1281
1282	if (maxnode > MAX_NUMNODES && MAX_NUMNODES % BITS_PER_LONG != 0) {
1283		unsigned long valid_mask = endmask;
1284
1285		valid_mask &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1286		if (get_user(t, nmask + nlongs - 1))
1287			return -EFAULT;
1288		if (t & valid_mask)
 
 
 
 
 
 
 
1289			return -EINVAL;
1290	}
1291
1292	if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1293		return -EFAULT;
1294	nodes_addr(*nodes)[nlongs-1] &= endmask;
1295	return 0;
1296}
1297
1298/* Copy a kernel node mask to user space */
1299static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1300			      nodemask_t *nodes)
1301{
1302	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1303	const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
 
 
 
 
1304
1305	if (copy > nbytes) {
1306		if (copy > PAGE_SIZE)
1307			return -EINVAL;
1308		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1309			return -EFAULT;
1310		copy = nbytes;
 
1311	}
 
 
 
 
 
1312	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1313}
1314
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1315static long kernel_mbind(unsigned long start, unsigned long len,
1316			 unsigned long mode, const unsigned long __user *nmask,
1317			 unsigned long maxnode, unsigned int flags)
1318{
 
1319	nodemask_t nodes;
 
1320	int err;
1321	unsigned short mode_flags;
1322
1323	mode_flags = mode & MPOL_MODE_FLAGS;
1324	mode &= ~MPOL_MODE_FLAGS;
1325	if (mode >= MPOL_MAX)
1326		return -EINVAL;
1327	if ((mode_flags & MPOL_F_STATIC_NODES) &&
1328	    (mode_flags & MPOL_F_RELATIVE_NODES))
1329		return -EINVAL;
1330	err = get_nodes(&nodes, nmask, maxnode);
1331	if (err)
1332		return err;
1333	return do_mbind(start, len, mode, mode_flags, &nodes, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1334}
1335
1336SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1337		unsigned long, mode, const unsigned long __user *, nmask,
1338		unsigned long, maxnode, unsigned int, flags)
1339{
1340	return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1341}
1342
1343/* Set the process memory policy */
1344static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1345				 unsigned long maxnode)
1346{
1347	int err;
1348	nodemask_t nodes;
1349	unsigned short flags;
 
 
 
 
 
1350
1351	flags = mode & MPOL_MODE_FLAGS;
1352	mode &= ~MPOL_MODE_FLAGS;
1353	if ((unsigned int)mode >= MPOL_MAX)
1354		return -EINVAL;
1355	if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1356		return -EINVAL;
1357	err = get_nodes(&nodes, nmask, maxnode);
1358	if (err)
1359		return err;
1360	return do_set_mempolicy(mode, flags, &nodes);
 
1361}
1362
1363SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1364		unsigned long, maxnode)
1365{
1366	return kernel_set_mempolicy(mode, nmask, maxnode);
1367}
1368
1369static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1370				const unsigned long __user *old_nodes,
1371				const unsigned long __user *new_nodes)
1372{
1373	struct mm_struct *mm = NULL;
1374	struct task_struct *task;
1375	nodemask_t task_nodes;
1376	int err;
1377	nodemask_t *old;
1378	nodemask_t *new;
1379	NODEMASK_SCRATCH(scratch);
1380
1381	if (!scratch)
1382		return -ENOMEM;
1383
1384	old = &scratch->mask1;
1385	new = &scratch->mask2;
1386
1387	err = get_nodes(old, old_nodes, maxnode);
1388	if (err)
1389		goto out;
1390
1391	err = get_nodes(new, new_nodes, maxnode);
1392	if (err)
1393		goto out;
1394
1395	/* Find the mm_struct */
1396	rcu_read_lock();
1397	task = pid ? find_task_by_vpid(pid) : current;
1398	if (!task) {
1399		rcu_read_unlock();
1400		err = -ESRCH;
1401		goto out;
1402	}
1403	get_task_struct(task);
1404
1405	err = -EINVAL;
1406
1407	/*
1408	 * Check if this process has the right to modify the specified process.
1409	 * Use the regular "ptrace_may_access()" checks.
1410	 */
1411	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1412		rcu_read_unlock();
1413		err = -EPERM;
1414		goto out_put;
1415	}
1416	rcu_read_unlock();
1417
1418	task_nodes = cpuset_mems_allowed(task);
1419	/* Is the user allowed to access the target nodes? */
1420	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1421		err = -EPERM;
1422		goto out_put;
1423	}
1424
1425	task_nodes = cpuset_mems_allowed(current);
1426	nodes_and(*new, *new, task_nodes);
1427	if (nodes_empty(*new))
1428		goto out_put;
1429
1430	nodes_and(*new, *new, node_states[N_MEMORY]);
1431	if (nodes_empty(*new))
1432		goto out_put;
1433
1434	err = security_task_movememory(task);
1435	if (err)
1436		goto out_put;
1437
1438	mm = get_task_mm(task);
1439	put_task_struct(task);
1440
1441	if (!mm) {
1442		err = -EINVAL;
1443		goto out;
1444	}
1445
1446	err = do_migrate_pages(mm, old, new,
1447		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1448
1449	mmput(mm);
1450out:
1451	NODEMASK_SCRATCH_FREE(scratch);
1452
1453	return err;
1454
1455out_put:
1456	put_task_struct(task);
1457	goto out;
1458
1459}
1460
1461SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1462		const unsigned long __user *, old_nodes,
1463		const unsigned long __user *, new_nodes)
1464{
1465	return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1466}
1467
1468
1469/* Retrieve NUMA policy */
1470static int kernel_get_mempolicy(int __user *policy,
1471				unsigned long __user *nmask,
1472				unsigned long maxnode,
1473				unsigned long addr,
1474				unsigned long flags)
1475{
1476	int err;
1477	int uninitialized_var(pval);
1478	nodemask_t nodes;
1479
1480	if (nmask != NULL && maxnode < MAX_NUMNODES)
1481		return -EINVAL;
1482
 
 
1483	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1484
1485	if (err)
1486		return err;
1487
1488	if (policy && put_user(pval, policy))
1489		return -EFAULT;
1490
1491	if (nmask)
1492		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1493
1494	return err;
1495}
1496
1497SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1498		unsigned long __user *, nmask, unsigned long, maxnode,
1499		unsigned long, addr, unsigned long, flags)
1500{
1501	return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1502}
1503
1504#ifdef CONFIG_COMPAT
1505
1506COMPAT_SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1507		       compat_ulong_t __user *, nmask,
1508		       compat_ulong_t, maxnode,
1509		       compat_ulong_t, addr, compat_ulong_t, flags)
1510{
1511	long err;
1512	unsigned long __user *nm = NULL;
1513	unsigned long nr_bits, alloc_size;
1514	DECLARE_BITMAP(bm, MAX_NUMNODES);
1515
1516	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1517	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1518
1519	if (nmask)
1520		nm = compat_alloc_user_space(alloc_size);
1521
1522	err = kernel_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1523
1524	if (!err && nmask) {
1525		unsigned long copy_size;
1526		copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1527		err = copy_from_user(bm, nm, copy_size);
1528		/* ensure entire bitmap is zeroed */
1529		err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1530		err |= compat_put_bitmap(nmask, bm, nr_bits);
1531	}
1532
1533	return err;
1534}
1535
1536COMPAT_SYSCALL_DEFINE3(set_mempolicy, int, mode, compat_ulong_t __user *, nmask,
1537		       compat_ulong_t, maxnode)
1538{
1539	unsigned long __user *nm = NULL;
1540	unsigned long nr_bits, alloc_size;
1541	DECLARE_BITMAP(bm, MAX_NUMNODES);
1542
1543	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1544	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1545
1546	if (nmask) {
1547		if (compat_get_bitmap(bm, nmask, nr_bits))
1548			return -EFAULT;
1549		nm = compat_alloc_user_space(alloc_size);
1550		if (copy_to_user(nm, bm, alloc_size))
1551			return -EFAULT;
1552	}
1553
1554	return kernel_set_mempolicy(mode, nm, nr_bits+1);
1555}
1556
1557COMPAT_SYSCALL_DEFINE6(mbind, compat_ulong_t, start, compat_ulong_t, len,
1558		       compat_ulong_t, mode, compat_ulong_t __user *, nmask,
1559		       compat_ulong_t, maxnode, compat_ulong_t, flags)
1560{
1561	unsigned long __user *nm = NULL;
1562	unsigned long nr_bits, alloc_size;
1563	nodemask_t bm;
1564
1565	nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1566	alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1567
1568	if (nmask) {
1569		if (compat_get_bitmap(nodes_addr(bm), nmask, nr_bits))
1570			return -EFAULT;
1571		nm = compat_alloc_user_space(alloc_size);
1572		if (copy_to_user(nm, nodes_addr(bm), alloc_size))
1573			return -EFAULT;
1574	}
1575
1576	return kernel_mbind(start, len, mode, nm, nr_bits+1, flags);
1577}
 
1578
1579COMPAT_SYSCALL_DEFINE4(migrate_pages, compat_pid_t, pid,
1580		       compat_ulong_t, maxnode,
1581		       const compat_ulong_t __user *, old_nodes,
1582		       const compat_ulong_t __user *, new_nodes)
1583{
1584	unsigned long __user *old = NULL;
1585	unsigned long __user *new = NULL;
1586	nodemask_t tmp_mask;
1587	unsigned long nr_bits;
1588	unsigned long size;
1589
1590	nr_bits = min_t(unsigned long, maxnode - 1, MAX_NUMNODES);
1591	size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1592	if (old_nodes) {
1593		if (compat_get_bitmap(nodes_addr(tmp_mask), old_nodes, nr_bits))
1594			return -EFAULT;
1595		old = compat_alloc_user_space(new_nodes ? size * 2 : size);
1596		if (new_nodes)
1597			new = old + size / sizeof(unsigned long);
1598		if (copy_to_user(old, nodes_addr(tmp_mask), size))
1599			return -EFAULT;
1600	}
1601	if (new_nodes) {
1602		if (compat_get_bitmap(nodes_addr(tmp_mask), new_nodes, nr_bits))
1603			return -EFAULT;
1604		if (new == NULL)
1605			new = compat_alloc_user_space(size);
1606		if (copy_to_user(new, nodes_addr(tmp_mask), size))
1607			return -EFAULT;
1608	}
1609	return kernel_migrate_pages(pid, nr_bits + 1, old, new);
1610}
1611
1612#endif /* CONFIG_COMPAT */
1613
1614struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1615						unsigned long addr)
1616{
1617	struct mempolicy *pol = NULL;
1618
1619	if (vma) {
1620		if (vma->vm_ops && vma->vm_ops->get_policy) {
1621			pol = vma->vm_ops->get_policy(vma, addr);
1622		} else if (vma->vm_policy) {
1623			pol = vma->vm_policy;
1624
1625			/*
1626			 * shmem_alloc_page() passes MPOL_F_SHARED policy with
1627			 * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1628			 * count on these policies which will be dropped by
1629			 * mpol_cond_put() later
1630			 */
1631			if (mpol_needs_cond_ref(pol))
1632				mpol_get(pol);
1633		}
1634	}
1635
1636	return pol;
1637}
1638
1639/*
1640 * get_vma_policy(@vma, @addr)
1641 * @vma: virtual memory area whose policy is sought
1642 * @addr: address in @vma for shared policy lookup
 
 
 
1643 *
1644 * Returns effective policy for a VMA at specified address.
1645 * Falls back to current->mempolicy or system default policy, as necessary.
1646 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1647 * count--added by the get_policy() vm_op, as appropriate--to protect against
1648 * freeing by another task.  It is the caller's responsibility to free the
1649 * extra reference for shared policies.
1650 */
1651static struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1652						unsigned long addr)
1653{
1654	struct mempolicy *pol = __get_vma_policy(vma, addr);
1655
 
1656	if (!pol)
1657		pol = get_task_policy(current);
1658
 
 
 
 
1659	return pol;
1660}
1661
1662bool vma_policy_mof(struct vm_area_struct *vma)
1663{
1664	struct mempolicy *pol;
1665
1666	if (vma->vm_ops && vma->vm_ops->get_policy) {
1667		bool ret = false;
 
1668
1669		pol = vma->vm_ops->get_policy(vma, vma->vm_start);
1670		if (pol && (pol->flags & MPOL_F_MOF))
1671			ret = true;
1672		mpol_cond_put(pol);
1673
1674		return ret;
1675	}
1676
1677	pol = vma->vm_policy;
1678	if (!pol)
1679		pol = get_task_policy(current);
1680
1681	return pol->flags & MPOL_F_MOF;
1682}
1683
1684static int apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1685{
1686	enum zone_type dynamic_policy_zone = policy_zone;
1687
1688	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1689
1690	/*
1691	 * if policy->v.nodes has movable memory only,
1692	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1693	 *
1694	 * policy->v.nodes is intersect with node_states[N_MEMORY].
1695	 * so if the following test faile, it implies
1696	 * policy->v.nodes has movable memory only.
1697	 */
1698	if (!nodes_intersects(policy->v.nodes, node_states[N_HIGH_MEMORY]))
1699		dynamic_policy_zone = ZONE_MOVABLE;
1700
1701	return zone >= dynamic_policy_zone;
1702}
1703
1704/*
1705 * Return a nodemask representing a mempolicy for filtering nodes for
1706 * page allocation
1707 */
1708static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1709{
1710	/* Lower zones don't get a nodemask applied for MPOL_BIND */
1711	if (unlikely(policy->mode == MPOL_BIND) &&
1712			apply_policy_zone(policy, gfp_zone(gfp)) &&
1713			cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1714		return &policy->v.nodes;
1715
1716	return NULL;
1717}
1718
1719/* Return the node id preferred by the given mempolicy, or the given id */
1720static int policy_node(gfp_t gfp, struct mempolicy *policy,
1721								int nd)
1722{
1723	if (policy->mode == MPOL_PREFERRED && !(policy->flags & MPOL_F_LOCAL))
1724		nd = policy->v.preferred_node;
1725	else {
1726		/*
1727		 * __GFP_THISNODE shouldn't even be used with the bind policy
1728		 * because we might easily break the expectation to stay on the
1729		 * requested node and not break the policy.
1730		 */
1731		WARN_ON_ONCE(policy->mode == MPOL_BIND && (gfp & __GFP_THISNODE));
1732	}
1733
1734	return nd;
1735}
1736
1737/* Do dynamic interleaving for a process */
1738static unsigned interleave_nodes(struct mempolicy *policy)
1739{
1740	unsigned next;
1741	struct task_struct *me = current;
 
 
 
 
 
 
1742
1743	next = next_node_in(me->il_prev, policy->v.nodes);
1744	if (next < MAX_NUMNODES)
1745		me->il_prev = next;
1746	return next;
1747}
1748
1749/*
1750 * Depending on the memory policy provide a node from which to allocate the
1751 * next slab entry.
1752 */
1753unsigned int mempolicy_slab_node(void)
1754{
1755	struct mempolicy *policy;
1756	int node = numa_mem_id();
1757
1758	if (in_interrupt())
1759		return node;
1760
1761	policy = current->mempolicy;
1762	if (!policy || policy->flags & MPOL_F_LOCAL)
1763		return node;
1764
1765	switch (policy->mode) {
1766	case MPOL_PREFERRED:
1767		/*
1768		 * handled MPOL_F_LOCAL above
1769		 */
1770		return policy->v.preferred_node;
1771
1772	case MPOL_INTERLEAVE:
1773		return interleave_nodes(policy);
1774
1775	case MPOL_BIND: {
 
 
 
 
 
1776		struct zoneref *z;
1777
1778		/*
1779		 * Follow bind policy behavior and start allocation at the
1780		 * first node.
1781		 */
1782		struct zonelist *zonelist;
1783		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1784		zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1785		z = first_zones_zonelist(zonelist, highest_zoneidx,
1786							&policy->v.nodes);
1787		return z->zone ? z->zone->node : node;
1788	}
 
 
1789
1790	default:
1791		BUG();
1792	}
1793}
1794
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1795/*
1796 * Do static interleaving for a VMA with known offset @n.  Returns the n'th
1797 * node in pol->v.nodes (starting from n=0), wrapping around if n exceeds the
1798 * number of present nodes.
1799 */
1800static unsigned offset_il_node(struct mempolicy *pol, unsigned long n)
1801{
1802	unsigned nnodes = nodes_weight(pol->v.nodes);
1803	unsigned target;
1804	int i;
1805	int nid;
1806
 
1807	if (!nnodes)
1808		return numa_node_id();
1809	target = (unsigned int)n % nnodes;
1810	nid = first_node(pol->v.nodes);
1811	for (i = 0; i < target; i++)
1812		nid = next_node(nid, pol->v.nodes);
1813	return nid;
1814}
1815
1816/* Determine a node number for interleave */
1817static inline unsigned interleave_nid(struct mempolicy *pol,
1818		 struct vm_area_struct *vma, unsigned long addr, int shift)
 
 
 
1819{
1820	if (vma) {
1821		unsigned long off;
1822
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1823		/*
1824		 * for small pages, there is no difference between
1825		 * shift and PAGE_SHIFT, so the bit-shift is safe.
1826		 * for huge pages, since vm_pgoff is in units of small
1827		 * pages, we need to shift off the always 0 bits to get
1828		 * a useful offset.
1829		 */
1830		BUG_ON(shift < PAGE_SHIFT);
1831		off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1832		off += (addr - vma->vm_start) >> shift;
1833		return offset_il_node(pol, off);
1834	} else
1835		return interleave_nodes(pol);
 
 
 
 
 
 
 
 
 
1836}
1837
1838#ifdef CONFIG_HUGETLBFS
1839/*
1840 * huge_node(@vma, @addr, @gfp_flags, @mpol)
1841 * @vma: virtual memory area whose policy is sought
1842 * @addr: address in @vma for shared policy lookup and interleave policy
1843 * @gfp_flags: for requested zone
1844 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
1845 * @nodemask: pointer to nodemask pointer for MPOL_BIND nodemask
1846 *
1847 * Returns a nid suitable for a huge page allocation and a pointer
1848 * to the struct mempolicy for conditional unref after allocation.
1849 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1850 * @nodemask for filtering the zonelist.
1851 *
1852 * Must be protected by read_mems_allowed_begin()
1853 */
1854int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
1855				struct mempolicy **mpol, nodemask_t **nodemask)
1856{
 
1857	int nid;
1858
1859	*mpol = get_vma_policy(vma, addr);
1860	*nodemask = NULL;	/* assume !MPOL_BIND */
1861
1862	if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1863		nid = interleave_nid(*mpol, vma, addr,
1864					huge_page_shift(hstate_vma(vma)));
1865	} else {
1866		nid = policy_node(gfp_flags, *mpol, numa_node_id());
1867		if ((*mpol)->mode == MPOL_BIND)
1868			*nodemask = &(*mpol)->v.nodes;
1869	}
1870	return nid;
1871}
1872
1873/*
1874 * init_nodemask_of_mempolicy
1875 *
1876 * If the current task's mempolicy is "default" [NULL], return 'false'
1877 * to indicate default policy.  Otherwise, extract the policy nodemask
1878 * for 'bind' or 'interleave' policy into the argument nodemask, or
1879 * initialize the argument nodemask to contain the single node for
1880 * 'preferred' or 'local' policy and return 'true' to indicate presence
1881 * of non-default mempolicy.
1882 *
1883 * We don't bother with reference counting the mempolicy [mpol_get/put]
1884 * because the current task is examining it's own mempolicy and a task's
1885 * mempolicy is only ever changed by the task itself.
1886 *
1887 * N.B., it is the caller's responsibility to free a returned nodemask.
1888 */
1889bool init_nodemask_of_mempolicy(nodemask_t *mask)
1890{
1891	struct mempolicy *mempolicy;
1892	int nid;
1893
1894	if (!(mask && current->mempolicy))
1895		return false;
1896
1897	task_lock(current);
1898	mempolicy = current->mempolicy;
1899	switch (mempolicy->mode) {
1900	case MPOL_PREFERRED:
1901		if (mempolicy->flags & MPOL_F_LOCAL)
1902			nid = numa_node_id();
1903		else
1904			nid = mempolicy->v.preferred_node;
1905		init_nodemask_of_node(mask, nid);
1906		break;
1907
1908	case MPOL_BIND:
1909		/* Fall through */
1910	case MPOL_INTERLEAVE:
1911		*mask =  mempolicy->v.nodes;
 
 
 
 
 
1912		break;
1913
1914	default:
1915		BUG();
1916	}
1917	task_unlock(current);
1918
1919	return true;
1920}
1921#endif
1922
1923/*
1924 * mempolicy_nodemask_intersects
1925 *
1926 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1927 * policy.  Otherwise, check for intersection between mask and the policy
1928 * nodemask for 'bind' or 'interleave' policy.  For 'perferred' or 'local'
1929 * policy, always return true since it may allocate elsewhere on fallback.
1930 *
1931 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1932 */
1933bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1934					const nodemask_t *mask)
1935{
1936	struct mempolicy *mempolicy;
1937	bool ret = true;
1938
1939	if (!mask)
1940		return ret;
 
1941	task_lock(tsk);
1942	mempolicy = tsk->mempolicy;
1943	if (!mempolicy)
1944		goto out;
1945
1946	switch (mempolicy->mode) {
1947	case MPOL_PREFERRED:
1948		/*
1949		 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1950		 * allocate from, they may fallback to other nodes when oom.
1951		 * Thus, it's possible for tsk to have allocated memory from
1952		 * nodes in mask.
1953		 */
1954		break;
1955	case MPOL_BIND:
1956	case MPOL_INTERLEAVE:
1957		ret = nodes_intersects(mempolicy->v.nodes, *mask);
1958		break;
1959	default:
1960		BUG();
1961	}
1962out:
1963	task_unlock(tsk);
 
1964	return ret;
1965}
1966
1967/* Allocate a page in interleaved policy.
1968   Own path because it needs to do special accounting. */
1969static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1970					unsigned nid)
1971{
1972	struct page *page;
 
 
 
 
 
 
 
 
 
 
 
 
 
1973
1974	page = __alloc_pages(gfp, order, nid);
1975	/* skip NUMA_INTERLEAVE_HIT counter update if numa stats is disabled */
1976	if (!static_branch_likely(&vm_numa_stat_key))
1977		return page;
1978	if (page && page_to_nid(page) == nid) {
1979		preempt_disable();
1980		__inc_numa_state(page_zone(page), NUMA_INTERLEAVE_HIT);
1981		preempt_enable();
1982	}
1983	return page;
1984}
1985
1986/**
1987 * 	alloc_pages_vma	- Allocate a page for a VMA.
 
 
 
 
 
1988 *
1989 * 	@gfp:
1990 *      %GFP_USER    user allocation.
1991 *      %GFP_KERNEL  kernel allocations,
1992 *      %GFP_HIGHMEM highmem/user allocations,
1993 *      %GFP_FS      allocation should not call back into a file system.
1994 *      %GFP_ATOMIC  don't sleep.
1995 *
1996 *	@order:Order of the GFP allocation.
1997 * 	@vma:  Pointer to VMA or NULL if not available.
1998 *	@addr: Virtual Address of the allocation. Must be inside the VMA.
1999 *	@node: Which node to prefer for allocation (modulo policy).
2000 *	@hugepage: for hugepages try only the preferred node if possible
2001 *
2002 * 	This function allocates a page from the kernel page pool and applies
2003 *	a NUMA policy associated with the VMA or the current process.
2004 *	When VMA is not NULL caller must hold down_read on the mmap_sem of the
2005 *	mm_struct of the VMA to prevent it from going away. Should be used for
2006 *	all allocations for pages that will be mapped into user space. Returns
2007 *	NULL when no page can be allocated.
2008 */
2009struct page *
2010alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
2011		unsigned long addr, int node, bool hugepage)
2012{
2013	struct mempolicy *pol;
2014	struct page *page;
2015	int preferred_nid;
2016	nodemask_t *nmask;
2017
2018	pol = get_vma_policy(vma, addr);
2019
2020	if (pol->mode == MPOL_INTERLEAVE) {
2021		unsigned nid;
2022
2023		nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
2024		mpol_cond_put(pol);
2025		page = alloc_page_interleave(gfp, order, nid);
2026		goto out;
2027	}
2028
2029	if (unlikely(IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && hugepage)) {
2030		int hpage_node = node;
2031
 
 
 
2032		/*
2033		 * For hugepage allocation and non-interleave policy which
2034		 * allows the current node (or other explicitly preferred
2035		 * node) we only try to allocate from the current/preferred
2036		 * node and don't fall back to other nodes, as the cost of
2037		 * remote accesses would likely offset THP benefits.
2038		 *
2039		 * If the policy is interleave, or does not allow the current
2040		 * node in its nodemask, we allocate the standard way.
2041		 */
2042		if (pol->mode == MPOL_PREFERRED &&
2043						!(pol->flags & MPOL_F_LOCAL))
2044			hpage_node = pol->v.preferred_node;
2045
2046		nmask = policy_nodemask(gfp, pol);
2047		if (!nmask || node_isset(hpage_node, *nmask)) {
2048			mpol_cond_put(pol);
2049			page = __alloc_pages_node(hpage_node,
2050						gfp | __GFP_THISNODE, order);
2051			goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2052		}
2053	}
2054
2055	nmask = policy_nodemask(gfp, pol);
2056	preferred_nid = policy_node(gfp, pol, node);
2057	page = __alloc_pages_nodemask(gfp, order, preferred_nid, nmask);
2058	mpol_cond_put(pol);
2059out:
2060	return page;
2061}
2062
 
 
 
 
 
 
 
2063/**
2064 * 	alloc_pages_current - Allocate pages.
 
 
 
 
 
 
 
 
 
 
2065 *
2066 *	@gfp:
2067 *		%GFP_USER   user allocation,
2068 *      	%GFP_KERNEL kernel allocation,
2069 *      	%GFP_HIGHMEM highmem allocation,
2070 *      	%GFP_FS     don't call back into a file system.
2071 *      	%GFP_ATOMIC don't sleep.
2072 *	@order: Power of two of allocation size in pages. 0 is a single page.
2073 *
2074 *	Allocate a page from the kernel page pool.  When not in
2075 *	interrupt context and apply the current process NUMA policy.
2076 *	Returns NULL when no page can be allocated.
2077 */
2078struct page *alloc_pages_current(gfp_t gfp, unsigned order)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2079{
2080	struct mempolicy *pol = &default_policy;
2081	struct page *page;
2082
 
 
 
 
2083	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2084		pol = get_task_policy(current);
2085
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2086	/*
2087	 * No reference counting needed for current->mempolicy
2088	 * nor system default_policy
 
 
 
2089	 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2090	if (pol->mode == MPOL_INTERLEAVE)
2091		page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
2092	else
2093		page = __alloc_pages_nodemask(gfp, order,
2094				policy_node(gfp, pol, numa_node_id()),
2095				policy_nodemask(gfp, pol));
2096
2097	return page;
 
 
 
 
 
 
 
 
 
 
 
2098}
2099EXPORT_SYMBOL(alloc_pages_current);
2100
2101int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2102{
2103	struct mempolicy *pol = mpol_dup(vma_policy(src));
2104
2105	if (IS_ERR(pol))
2106		return PTR_ERR(pol);
2107	dst->vm_policy = pol;
2108	return 0;
2109}
2110
2111/*
2112 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2113 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2114 * with the mems_allowed returned by cpuset_mems_allowed().  This
2115 * keeps mempolicies cpuset relative after its cpuset moves.  See
2116 * further kernel/cpuset.c update_nodemask().
2117 *
2118 * current's mempolicy may be rebinded by the other task(the task that changes
2119 * cpuset's mems), so we needn't do rebind work for current task.
2120 */
2121
2122/* Slow path of a mempolicy duplicate */
2123struct mempolicy *__mpol_dup(struct mempolicy *old)
2124{
2125	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2126
2127	if (!new)
2128		return ERR_PTR(-ENOMEM);
2129
2130	/* task's mempolicy is protected by alloc_lock */
2131	if (old == current->mempolicy) {
2132		task_lock(current);
2133		*new = *old;
2134		task_unlock(current);
2135	} else
2136		*new = *old;
2137
2138	if (current_cpuset_is_being_rebound()) {
2139		nodemask_t mems = cpuset_mems_allowed(current);
2140		mpol_rebind_policy(new, &mems);
2141	}
2142	atomic_set(&new->refcnt, 1);
2143	return new;
2144}
2145
2146/* Slow path of a mempolicy comparison */
2147bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2148{
2149	if (!a || !b)
2150		return false;
2151	if (a->mode != b->mode)
2152		return false;
2153	if (a->flags != b->flags)
2154		return false;
 
 
2155	if (mpol_store_user_nodemask(a))
2156		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2157			return false;
2158
2159	switch (a->mode) {
2160	case MPOL_BIND:
2161		/* Fall through */
2162	case MPOL_INTERLEAVE:
2163		return !!nodes_equal(a->v.nodes, b->v.nodes);
2164	case MPOL_PREFERRED:
2165		/* a's ->flags is the same as b's */
2166		if (a->flags & MPOL_F_LOCAL)
2167			return true;
2168		return a->v.preferred_node == b->v.preferred_node;
 
2169	default:
2170		BUG();
2171		return false;
2172	}
2173}
2174
2175/*
2176 * Shared memory backing store policy support.
2177 *
2178 * Remember policies even when nobody has shared memory mapped.
2179 * The policies are kept in Red-Black tree linked from the inode.
2180 * They are protected by the sp->lock rwlock, which should be held
2181 * for any accesses to the tree.
2182 */
2183
2184/*
2185 * lookup first element intersecting start-end.  Caller holds sp->lock for
2186 * reading or for writing
2187 */
2188static struct sp_node *
2189sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2190{
2191	struct rb_node *n = sp->root.rb_node;
2192
2193	while (n) {
2194		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2195
2196		if (start >= p->end)
2197			n = n->rb_right;
2198		else if (end <= p->start)
2199			n = n->rb_left;
2200		else
2201			break;
2202	}
2203	if (!n)
2204		return NULL;
2205	for (;;) {
2206		struct sp_node *w = NULL;
2207		struct rb_node *prev = rb_prev(n);
2208		if (!prev)
2209			break;
2210		w = rb_entry(prev, struct sp_node, nd);
2211		if (w->end <= start)
2212			break;
2213		n = prev;
2214	}
2215	return rb_entry(n, struct sp_node, nd);
2216}
2217
2218/*
2219 * Insert a new shared policy into the list.  Caller holds sp->lock for
2220 * writing.
2221 */
2222static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2223{
2224	struct rb_node **p = &sp->root.rb_node;
2225	struct rb_node *parent = NULL;
2226	struct sp_node *nd;
2227
2228	while (*p) {
2229		parent = *p;
2230		nd = rb_entry(parent, struct sp_node, nd);
2231		if (new->start < nd->start)
2232			p = &(*p)->rb_left;
2233		else if (new->end > nd->end)
2234			p = &(*p)->rb_right;
2235		else
2236			BUG();
2237	}
2238	rb_link_node(&new->nd, parent, p);
2239	rb_insert_color(&new->nd, &sp->root);
2240	pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2241		 new->policy ? new->policy->mode : 0);
2242}
2243
2244/* Find shared policy intersecting idx */
2245struct mempolicy *
2246mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2247{
2248	struct mempolicy *pol = NULL;
2249	struct sp_node *sn;
2250
2251	if (!sp->root.rb_node)
2252		return NULL;
2253	read_lock(&sp->lock);
2254	sn = sp_lookup(sp, idx, idx+1);
2255	if (sn) {
2256		mpol_get(sn->policy);
2257		pol = sn->policy;
2258	}
2259	read_unlock(&sp->lock);
2260	return pol;
2261}
2262
2263static void sp_free(struct sp_node *n)
2264{
2265	mpol_put(n->policy);
2266	kmem_cache_free(sn_cache, n);
2267}
2268
2269/**
2270 * mpol_misplaced - check whether current page node is valid in policy
2271 *
2272 * @page: page to be checked
2273 * @vma: vm area where page mapped
2274 * @addr: virtual address where page mapped
2275 *
2276 * Lookup current policy node id for vma,addr and "compare to" page's
2277 * node id.
2278 *
2279 * Returns:
2280 *	-1	- not misplaced, page is in the right node
2281 *	node	- node id where the page should be
2282 *
2283 * Policy determination "mimics" alloc_page_vma().
2284 * Called from fault path where we know the vma and faulting address.
 
 
 
2285 */
2286int mpol_misplaced(struct page *page, struct vm_area_struct *vma, unsigned long addr)
 
2287{
2288	struct mempolicy *pol;
 
2289	struct zoneref *z;
2290	int curnid = page_to_nid(page);
2291	unsigned long pgoff;
2292	int thiscpu = raw_smp_processor_id();
2293	int thisnid = cpu_to_node(thiscpu);
2294	int polnid = -1;
2295	int ret = -1;
2296
2297	pol = get_vma_policy(vma, addr);
 
 
 
 
 
2298	if (!(pol->flags & MPOL_F_MOF))
2299		goto out;
2300
2301	switch (pol->mode) {
2302	case MPOL_INTERLEAVE:
2303		pgoff = vma->vm_pgoff;
2304		pgoff += (addr - vma->vm_start) >> PAGE_SHIFT;
2305		polnid = offset_il_node(pol, pgoff);
 
 
2306		break;
2307
2308	case MPOL_PREFERRED:
2309		if (pol->flags & MPOL_F_LOCAL)
2310			polnid = numa_node_id();
2311		else
2312			polnid = pol->v.preferred_node;
 
 
 
2313		break;
2314
2315	case MPOL_BIND:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2316
2317		/*
2318		 * allows binding to multiple nodes.
2319		 * use current page if in policy nodemask,
2320		 * else select nearest allowed node, if any.
2321		 * If no allowed nodes, use current [!misplaced].
2322		 */
2323		if (node_isset(curnid, pol->v.nodes))
2324			goto out;
2325		z = first_zones_zonelist(
2326				node_zonelist(numa_node_id(), GFP_HIGHUSER),
2327				gfp_zone(GFP_HIGHUSER),
2328				&pol->v.nodes);
2329		polnid = z->zone->node;
2330		break;
2331
2332	default:
2333		BUG();
2334	}
2335
2336	/* Migrate the page towards the node whose CPU is referencing it */
2337	if (pol->flags & MPOL_F_MORON) {
2338		polnid = thisnid;
2339
2340		if (!should_numa_migrate_memory(current, page, curnid, thiscpu))
 
2341			goto out;
2342	}
2343
2344	if (curnid != polnid)
2345		ret = polnid;
2346out:
2347	mpol_cond_put(pol);
2348
2349	return ret;
2350}
2351
2352/*
2353 * Drop the (possibly final) reference to task->mempolicy.  It needs to be
2354 * dropped after task->mempolicy is set to NULL so that any allocation done as
2355 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2356 * policy.
2357 */
2358void mpol_put_task_policy(struct task_struct *task)
2359{
2360	struct mempolicy *pol;
2361
2362	task_lock(task);
2363	pol = task->mempolicy;
2364	task->mempolicy = NULL;
2365	task_unlock(task);
2366	mpol_put(pol);
2367}
2368
2369static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2370{
2371	pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2372	rb_erase(&n->nd, &sp->root);
2373	sp_free(n);
2374}
2375
2376static void sp_node_init(struct sp_node *node, unsigned long start,
2377			unsigned long end, struct mempolicy *pol)
2378{
2379	node->start = start;
2380	node->end = end;
2381	node->policy = pol;
2382}
2383
2384static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2385				struct mempolicy *pol)
2386{
2387	struct sp_node *n;
2388	struct mempolicy *newpol;
2389
2390	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2391	if (!n)
2392		return NULL;
2393
2394	newpol = mpol_dup(pol);
2395	if (IS_ERR(newpol)) {
2396		kmem_cache_free(sn_cache, n);
2397		return NULL;
2398	}
2399	newpol->flags |= MPOL_F_SHARED;
2400	sp_node_init(n, start, end, newpol);
2401
2402	return n;
2403}
2404
2405/* Replace a policy range. */
2406static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2407				 unsigned long end, struct sp_node *new)
2408{
2409	struct sp_node *n;
2410	struct sp_node *n_new = NULL;
2411	struct mempolicy *mpol_new = NULL;
2412	int ret = 0;
2413
2414restart:
2415	write_lock(&sp->lock);
2416	n = sp_lookup(sp, start, end);
2417	/* Take care of old policies in the same range. */
2418	while (n && n->start < end) {
2419		struct rb_node *next = rb_next(&n->nd);
2420		if (n->start >= start) {
2421			if (n->end <= end)
2422				sp_delete(sp, n);
2423			else
2424				n->start = end;
2425		} else {
2426			/* Old policy spanning whole new range. */
2427			if (n->end > end) {
2428				if (!n_new)
2429					goto alloc_new;
2430
2431				*mpol_new = *n->policy;
2432				atomic_set(&mpol_new->refcnt, 1);
2433				sp_node_init(n_new, end, n->end, mpol_new);
2434				n->end = start;
2435				sp_insert(sp, n_new);
2436				n_new = NULL;
2437				mpol_new = NULL;
2438				break;
2439			} else
2440				n->end = start;
2441		}
2442		if (!next)
2443			break;
2444		n = rb_entry(next, struct sp_node, nd);
2445	}
2446	if (new)
2447		sp_insert(sp, new);
2448	write_unlock(&sp->lock);
2449	ret = 0;
2450
2451err_out:
2452	if (mpol_new)
2453		mpol_put(mpol_new);
2454	if (n_new)
2455		kmem_cache_free(sn_cache, n_new);
2456
2457	return ret;
2458
2459alloc_new:
2460	write_unlock(&sp->lock);
2461	ret = -ENOMEM;
2462	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2463	if (!n_new)
2464		goto err_out;
2465	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2466	if (!mpol_new)
2467		goto err_out;
 
2468	goto restart;
2469}
2470
2471/**
2472 * mpol_shared_policy_init - initialize shared policy for inode
2473 * @sp: pointer to inode shared policy
2474 * @mpol:  struct mempolicy to install
2475 *
2476 * Install non-NULL @mpol in inode's shared policy rb-tree.
2477 * On entry, the current task has a reference on a non-NULL @mpol.
2478 * This must be released on exit.
2479 * This is called at get_inode() calls and we can use GFP_KERNEL.
2480 */
2481void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2482{
2483	int ret;
2484
2485	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2486	rwlock_init(&sp->lock);
2487
2488	if (mpol) {
2489		struct vm_area_struct pvma;
2490		struct mempolicy *new;
2491		NODEMASK_SCRATCH(scratch);
2492
2493		if (!scratch)
2494			goto put_mpol;
2495		/* contextualize the tmpfs mount point mempolicy */
2496		new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2497		if (IS_ERR(new))
 
2498			goto free_scratch; /* no valid nodemask intersection */
2499
2500		task_lock(current);
2501		ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2502		task_unlock(current);
2503		if (ret)
2504			goto put_new;
2505
2506		/* Create pseudo-vma that contains just the policy */
2507		memset(&pvma, 0, sizeof(struct vm_area_struct));
2508		pvma.vm_end = TASK_SIZE;	/* policy covers entire file */
2509		mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2510
2511put_new:
2512		mpol_put(new);			/* drop initial ref */
2513free_scratch:
2514		NODEMASK_SCRATCH_FREE(scratch);
2515put_mpol:
2516		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
2517	}
2518}
2519
2520int mpol_set_shared_policy(struct shared_policy *info,
2521			struct vm_area_struct *vma, struct mempolicy *npol)
2522{
2523	int err;
2524	struct sp_node *new = NULL;
2525	unsigned long sz = vma_pages(vma);
2526
2527	pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2528		 vma->vm_pgoff,
2529		 sz, npol ? npol->mode : -1,
2530		 npol ? npol->flags : -1,
2531		 npol ? nodes_addr(npol->v.nodes)[0] : NUMA_NO_NODE);
2532
2533	if (npol) {
2534		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2535		if (!new)
2536			return -ENOMEM;
2537	}
2538	err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2539	if (err && new)
2540		sp_free(new);
2541	return err;
2542}
2543
2544/* Free a backing policy store on inode delete. */
2545void mpol_free_shared_policy(struct shared_policy *p)
2546{
2547	struct sp_node *n;
2548	struct rb_node *next;
2549
2550	if (!p->root.rb_node)
2551		return;
2552	write_lock(&p->lock);
2553	next = rb_first(&p->root);
2554	while (next) {
2555		n = rb_entry(next, struct sp_node, nd);
2556		next = rb_next(&n->nd);
2557		sp_delete(p, n);
2558	}
2559	write_unlock(&p->lock);
2560}
2561
2562#ifdef CONFIG_NUMA_BALANCING
2563static int __initdata numabalancing_override;
2564
2565static void __init check_numabalancing_enable(void)
2566{
2567	bool numabalancing_default = false;
2568
2569	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
2570		numabalancing_default = true;
2571
2572	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
2573	if (numabalancing_override)
2574		set_numabalancing_state(numabalancing_override == 1);
2575
2576	if (num_online_nodes() > 1 && !numabalancing_override) {
2577		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
2578			numabalancing_default ? "Enabling" : "Disabling");
2579		set_numabalancing_state(numabalancing_default);
2580	}
2581}
2582
2583static int __init setup_numabalancing(char *str)
2584{
2585	int ret = 0;
2586	if (!str)
2587		goto out;
2588
2589	if (!strcmp(str, "enable")) {
2590		numabalancing_override = 1;
2591		ret = 1;
2592	} else if (!strcmp(str, "disable")) {
2593		numabalancing_override = -1;
2594		ret = 1;
2595	}
2596out:
2597	if (!ret)
2598		pr_warn("Unable to parse numa_balancing=\n");
2599
2600	return ret;
2601}
2602__setup("numa_balancing=", setup_numabalancing);
2603#else
2604static inline void __init check_numabalancing_enable(void)
2605{
2606}
2607#endif /* CONFIG_NUMA_BALANCING */
2608
2609/* assumes fs == KERNEL_DS */
2610void __init numa_policy_init(void)
2611{
2612	nodemask_t interleave_nodes;
2613	unsigned long largest = 0;
2614	int nid, prefer = 0;
2615
2616	policy_cache = kmem_cache_create("numa_policy",
2617					 sizeof(struct mempolicy),
2618					 0, SLAB_PANIC, NULL);
2619
2620	sn_cache = kmem_cache_create("shared_policy_node",
2621				     sizeof(struct sp_node),
2622				     0, SLAB_PANIC, NULL);
2623
2624	for_each_node(nid) {
2625		preferred_node_policy[nid] = (struct mempolicy) {
2626			.refcnt = ATOMIC_INIT(1),
2627			.mode = MPOL_PREFERRED,
2628			.flags = MPOL_F_MOF | MPOL_F_MORON,
2629			.v = { .preferred_node = nid, },
2630		};
2631	}
2632
2633	/*
2634	 * Set interleaving policy for system init. Interleaving is only
2635	 * enabled across suitably sized nodes (default is >= 16MB), or
2636	 * fall back to the largest node if they're all smaller.
2637	 */
2638	nodes_clear(interleave_nodes);
2639	for_each_node_state(nid, N_MEMORY) {
2640		unsigned long total_pages = node_present_pages(nid);
2641
2642		/* Preserve the largest node */
2643		if (largest < total_pages) {
2644			largest = total_pages;
2645			prefer = nid;
2646		}
2647
2648		/* Interleave this node? */
2649		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2650			node_set(nid, interleave_nodes);
2651	}
2652
2653	/* All too small, use the largest */
2654	if (unlikely(nodes_empty(interleave_nodes)))
2655		node_set(prefer, interleave_nodes);
2656
2657	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2658		pr_err("%s: interleaving failed\n", __func__);
2659
2660	check_numabalancing_enable();
2661}
2662
2663/* Reset policy of current process to default */
2664void numa_default_policy(void)
2665{
2666	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2667}
2668
2669/*
2670 * Parse and format mempolicy from/to strings
2671 */
2672
2673/*
2674 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2675 */
2676static const char * const policy_modes[] =
2677{
2678	[MPOL_DEFAULT]    = "default",
2679	[MPOL_PREFERRED]  = "prefer",
2680	[MPOL_BIND]       = "bind",
2681	[MPOL_INTERLEAVE] = "interleave",
 
2682	[MPOL_LOCAL]      = "local",
 
2683};
2684
2685
2686#ifdef CONFIG_TMPFS
2687/**
2688 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2689 * @str:  string containing mempolicy to parse
2690 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2691 *
2692 * Format of input:
2693 *	<mode>[=<flags>][:<nodelist>]
2694 *
2695 * On success, returns 0, else 1
2696 */
2697int mpol_parse_str(char *str, struct mempolicy **mpol)
2698{
2699	struct mempolicy *new = NULL;
2700	unsigned short mode;
2701	unsigned short mode_flags;
2702	nodemask_t nodes;
2703	char *nodelist = strchr(str, ':');
2704	char *flags = strchr(str, '=');
2705	int err = 1;
 
 
 
2706
2707	if (nodelist) {
2708		/* NUL-terminate mode or flags string */
2709		*nodelist++ = '\0';
2710		if (nodelist_parse(nodelist, nodes))
2711			goto out;
2712		if (!nodes_subset(nodes, node_states[N_MEMORY]))
2713			goto out;
2714	} else
2715		nodes_clear(nodes);
2716
2717	if (flags)
2718		*flags++ = '\0';	/* terminate mode string */
2719
2720	for (mode = 0; mode < MPOL_MAX; mode++) {
2721		if (!strcmp(str, policy_modes[mode])) {
2722			break;
2723		}
2724	}
2725	if (mode >= MPOL_MAX)
2726		goto out;
2727
2728	switch (mode) {
2729	case MPOL_PREFERRED:
2730		/*
2731		 * Insist on a nodelist of one node only
 
 
2732		 */
2733		if (nodelist) {
2734			char *rest = nodelist;
2735			while (isdigit(*rest))
2736				rest++;
2737			if (*rest)
2738				goto out;
 
 
2739		}
2740		break;
2741	case MPOL_INTERLEAVE:
 
2742		/*
2743		 * Default to online nodes with memory if no nodelist
2744		 */
2745		if (!nodelist)
2746			nodes = node_states[N_MEMORY];
2747		break;
2748	case MPOL_LOCAL:
2749		/*
2750		 * Don't allow a nodelist;  mpol_new() checks flags
2751		 */
2752		if (nodelist)
2753			goto out;
2754		mode = MPOL_PREFERRED;
2755		break;
2756	case MPOL_DEFAULT:
2757		/*
2758		 * Insist on a empty nodelist
2759		 */
2760		if (!nodelist)
2761			err = 0;
2762		goto out;
 
2763	case MPOL_BIND:
2764		/*
2765		 * Insist on a nodelist
2766		 */
2767		if (!nodelist)
2768			goto out;
2769	}
2770
2771	mode_flags = 0;
2772	if (flags) {
2773		/*
2774		 * Currently, we only support two mutually exclusive
2775		 * mode flags.
2776		 */
2777		if (!strcmp(flags, "static"))
2778			mode_flags |= MPOL_F_STATIC_NODES;
2779		else if (!strcmp(flags, "relative"))
2780			mode_flags |= MPOL_F_RELATIVE_NODES;
2781		else
2782			goto out;
2783	}
2784
2785	new = mpol_new(mode, mode_flags, &nodes);
2786	if (IS_ERR(new))
2787		goto out;
2788
2789	/*
2790	 * Save nodes for mpol_to_str() to show the tmpfs mount options
2791	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2792	 */
2793	if (mode != MPOL_PREFERRED)
2794		new->v.nodes = nodes;
2795	else if (nodelist)
2796		new->v.preferred_node = first_node(nodes);
2797	else
2798		new->flags |= MPOL_F_LOCAL;
 
 
2799
2800	/*
2801	 * Save nodes for contextualization: this will be used to "clone"
2802	 * the mempolicy in a specific context [cpuset] at a later time.
2803	 */
2804	new->w.user_nodemask = nodes;
2805
2806	err = 0;
2807
2808out:
2809	/* Restore string for error message */
2810	if (nodelist)
2811		*--nodelist = ':';
2812	if (flags)
2813		*--flags = '=';
2814	if (!err)
2815		*mpol = new;
2816	return err;
2817}
2818#endif /* CONFIG_TMPFS */
2819
2820/**
2821 * mpol_to_str - format a mempolicy structure for printing
2822 * @buffer:  to contain formatted mempolicy string
2823 * @maxlen:  length of @buffer
2824 * @pol:  pointer to mempolicy to be formatted
2825 *
2826 * Convert @pol into a string.  If @buffer is too short, truncate the string.
2827 * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the
2828 * longest flag, "relative", and to display at least a few node ids.
 
2829 */
2830void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
2831{
2832	char *p = buffer;
2833	nodemask_t nodes = NODE_MASK_NONE;
2834	unsigned short mode = MPOL_DEFAULT;
2835	unsigned short flags = 0;
2836
2837	if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {
 
 
 
2838		mode = pol->mode;
2839		flags = pol->flags;
2840	}
2841
2842	switch (mode) {
2843	case MPOL_DEFAULT:
 
2844		break;
2845	case MPOL_PREFERRED:
2846		if (flags & MPOL_F_LOCAL)
2847			mode = MPOL_LOCAL;
2848		else
2849			node_set(pol->v.preferred_node, nodes);
2850		break;
2851	case MPOL_BIND:
2852	case MPOL_INTERLEAVE:
2853		nodes = pol->v.nodes;
 
2854		break;
2855	default:
2856		WARN_ON_ONCE(1);
2857		snprintf(p, maxlen, "unknown");
2858		return;
2859	}
2860
2861	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
2862
2863	if (flags & MPOL_MODE_FLAGS) {
2864		p += snprintf(p, buffer + maxlen - p, "=");
2865
2866		/*
2867		 * Currently, the only defined flags are mutually exclusive
2868		 */
2869		if (flags & MPOL_F_STATIC_NODES)
2870			p += snprintf(p, buffer + maxlen - p, "static");
2871		else if (flags & MPOL_F_RELATIVE_NODES)
2872			p += snprintf(p, buffer + maxlen - p, "relative");
 
 
 
 
 
 
2873	}
2874
2875	if (!nodes_empty(nodes))
2876		p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
2877			       nodemask_pr_args(&nodes));
2878}
v6.13.7
   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 six 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 * weighted interleave
  23 *                Allocate memory interleaved over a set of nodes based on
  24 *                a set of weights (per-node), with normal fallback if it
  25 *                fails.  Otherwise operates the same as interleave.
  26 *                Example: nodeset(0,1) & weights (2,1) - 2 pages allocated
  27 *                on node 0 for every 1 page allocated on node 1.
  28 *
  29 * bind           Only allocate memory on a specific set of nodes,
  30 *                no fallback.
  31 *                FIXME: memory is allocated starting with the first node
  32 *                to the last. It would be better if bind would truly restrict
  33 *                the allocation to memory nodes instead
  34 *
  35 * preferred      Try a specific node first before normal fallback.
  36 *                As a special case NUMA_NO_NODE here means do the allocation
  37 *                on the local CPU. This is normally identical to default,
  38 *                but useful to set in a VMA when you have a non default
  39 *                process policy.
  40 *
  41 * preferred many Try a set of nodes first before normal fallback. This is
  42 *                similar to preferred without the special case.
  43 *
  44 * default        Allocate on the local node first, or when on a VMA
  45 *                use the process policy. This is what Linux always did
  46 *		  in a NUMA aware kernel and still does by, ahem, default.
  47 *
  48 * The process policy is applied for most non interrupt memory allocations
  49 * in that process' context. Interrupts ignore the policies and always
  50 * try to allocate on the local CPU. The VMA policy is only applied for memory
  51 * allocations for a VMA in the VM.
  52 *
  53 * Currently there are a few corner cases in swapping where the policy
  54 * is not applied, but the majority should be handled. When process policy
  55 * is used it is not remembered over swap outs/swap ins.
  56 *
  57 * Only the highest zone in the zone hierarchy gets policied. Allocations
  58 * requesting a lower zone just use default policy. This implies that
  59 * on systems with highmem kernel lowmem allocation don't get policied.
  60 * Same with GFP_DMA allocations.
  61 *
  62 * For shmem/tmpfs shared memory the policy is shared between
  63 * all users and remembered even when nobody has memory mapped.
  64 */
  65
  66/* Notebook:
  67   fix mmap readahead to honour policy and enable policy for any page cache
  68   object
  69   statistics for bigpages
  70   global policy for page cache? currently it uses process policy. Requires
  71   first item above.
  72   handle mremap for shared memory (currently ignored for the policy)
  73   grows down?
  74   make bind policy root only? It can trigger oom much faster and the
  75   kernel is not always grateful with that.
  76*/
  77
  78#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  79
  80#include <linux/mempolicy.h>
  81#include <linux/pagewalk.h>
  82#include <linux/highmem.h>
  83#include <linux/hugetlb.h>
  84#include <linux/kernel.h>
  85#include <linux/sched.h>
  86#include <linux/sched/mm.h>
  87#include <linux/sched/numa_balancing.h>
  88#include <linux/sched/task.h>
  89#include <linux/nodemask.h>
  90#include <linux/cpuset.h>
  91#include <linux/slab.h>
  92#include <linux/string.h>
  93#include <linux/export.h>
  94#include <linux/nsproxy.h>
  95#include <linux/interrupt.h>
  96#include <linux/init.h>
  97#include <linux/compat.h>
  98#include <linux/ptrace.h>
  99#include <linux/swap.h>
 100#include <linux/seq_file.h>
 101#include <linux/proc_fs.h>
 102#include <linux/migrate.h>
 103#include <linux/ksm.h>
 104#include <linux/rmap.h>
 105#include <linux/security.h>
 106#include <linux/syscalls.h>
 107#include <linux/ctype.h>
 108#include <linux/mm_inline.h>
 109#include <linux/mmu_notifier.h>
 110#include <linux/printk.h>
 111#include <linux/swapops.h>
 112
 113#include <asm/tlbflush.h>
 114#include <asm/tlb.h>
 115#include <linux/uaccess.h>
 116
 117#include "internal.h"
 118
 119/* Internal flags */
 120#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)	/* Skip checks for continuous vmas */
 121#define MPOL_MF_INVERT       (MPOL_MF_INTERNAL << 1)	/* Invert check for nodemask */
 122#define MPOL_MF_WRLOCK       (MPOL_MF_INTERNAL << 2)	/* Write-lock walked vmas */
 123
 124static struct kmem_cache *policy_cache;
 125static struct kmem_cache *sn_cache;
 126
 127/* Highest zone. An specific allocation for a zone below that is not
 128   policied. */
 129enum zone_type policy_zone = 0;
 130
 131/*
 132 * run-time system-wide default policy => local allocation
 133 */
 134static struct mempolicy default_policy = {
 135	.refcnt = ATOMIC_INIT(1), /* never free it */
 136	.mode = MPOL_LOCAL,
 
 137};
 138
 139static struct mempolicy preferred_node_policy[MAX_NUMNODES];
 140
 141/*
 142 * iw_table is the sysfs-set interleave weight table, a value of 0 denotes
 143 * system-default value should be used. A NULL iw_table also denotes that
 144 * system-default values should be used. Until the system-default table
 145 * is implemented, the system-default is always 1.
 146 *
 147 * iw_table is RCU protected
 148 */
 149static u8 __rcu *iw_table;
 150static DEFINE_MUTEX(iw_table_lock);
 151
 152static u8 get_il_weight(int node)
 153{
 154	u8 *table;
 155	u8 weight;
 156
 157	rcu_read_lock();
 158	table = rcu_dereference(iw_table);
 159	/* if no iw_table, use system default */
 160	weight = table ? table[node] : 1;
 161	/* if value in iw_table is 0, use system default */
 162	weight = weight ? weight : 1;
 163	rcu_read_unlock();
 164	return weight;
 165}
 166
 167/**
 168 * numa_nearest_node - Find nearest node by state
 169 * @node: Node id to start the search
 170 * @state: State to filter the search
 171 *
 172 * Lookup the closest node by distance if @nid is not in state.
 173 *
 174 * Return: this @node if it is in state, otherwise the closest node by distance
 175 */
 176int numa_nearest_node(int node, unsigned int state)
 177{
 178	int min_dist = INT_MAX, dist, n, min_node;
 179
 180	if (state >= NR_NODE_STATES)
 181		return -EINVAL;
 182
 183	if (node == NUMA_NO_NODE || node_state(node, state))
 184		return node;
 185
 186	min_node = node;
 187	for_each_node_state(n, state) {
 188		dist = node_distance(node, n);
 189		if (dist < min_dist) {
 190			min_dist = dist;
 191			min_node = n;
 192		}
 193	}
 194
 195	return min_node;
 196}
 197EXPORT_SYMBOL_GPL(numa_nearest_node);
 198
 199struct mempolicy *get_task_policy(struct task_struct *p)
 200{
 201	struct mempolicy *pol = p->mempolicy;
 202	int node;
 203
 204	if (pol)
 205		return pol;
 206
 207	node = numa_node_id();
 208	if (node != NUMA_NO_NODE) {
 209		pol = &preferred_node_policy[node];
 210		/* preferred_node_policy is not initialised early in boot */
 211		if (pol->mode)
 212			return pol;
 213	}
 214
 215	return &default_policy;
 216}
 217
 218static const struct mempolicy_operations {
 219	int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 220	void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes);
 221} mpol_ops[MPOL_MAX];
 222
 223static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 224{
 225	return pol->flags & MPOL_MODE_FLAGS;
 226}
 227
 228static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 229				   const nodemask_t *rel)
 230{
 231	nodemask_t tmp;
 232	nodes_fold(tmp, *orig, nodes_weight(*rel));
 233	nodes_onto(*ret, tmp, *rel);
 234}
 235
 236static int mpol_new_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 237{
 238	if (nodes_empty(*nodes))
 239		return -EINVAL;
 240	pol->nodes = *nodes;
 241	return 0;
 242}
 243
 244static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 245{
 
 
 
 
 
 
 
 
 
 
 
 246	if (nodes_empty(*nodes))
 247		return -EINVAL;
 248
 249	nodes_clear(pol->nodes);
 250	node_set(first_node(*nodes), pol->nodes);
 251	return 0;
 252}
 253
 254/*
 255 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 256 * any, for the new policy.  mpol_new() has already validated the nodes
 257 * parameter with respect to the policy mode and flags.
 
 258 *
 259 * Must be called holding task's alloc_lock to protect task's mems_allowed
 260 * and mempolicy.  May also be called holding the mmap_lock for write.
 261 */
 262static int mpol_set_nodemask(struct mempolicy *pol,
 263		     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 264{
 265	int ret;
 266
 267	/*
 268	 * Default (pol==NULL) resp. local memory policies are not a
 269	 * subject of any remapping. They also do not need any special
 270	 * constructor.
 271	 */
 272	if (!pol || pol->mode == MPOL_LOCAL)
 273		return 0;
 274
 275	/* Check N_MEMORY */
 276	nodes_and(nsc->mask1,
 277		  cpuset_current_mems_allowed, node_states[N_MEMORY]);
 278
 279	VM_BUG_ON(!nodes);
 
 
 
 
 
 
 
 280
 281	if (pol->flags & MPOL_F_RELATIVE_NODES)
 282		mpol_relative_nodemask(&nsc->mask2, nodes, &nsc->mask1);
 283	else
 284		nodes_and(nsc->mask2, *nodes, nsc->mask1);
 
 
 285
 286	if (mpol_store_user_nodemask(pol))
 287		pol->w.user_nodemask = *nodes;
 288	else
 289		pol->w.cpuset_mems_allowed = cpuset_current_mems_allowed;
 290
 291	ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 292	return ret;
 293}
 294
 295/*
 296 * This function just creates a new policy, does some check and simple
 297 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 298 */
 299static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 300				  nodemask_t *nodes)
 301{
 302	struct mempolicy *policy;
 303
 
 
 
 304	if (mode == MPOL_DEFAULT) {
 305		if (nodes && !nodes_empty(*nodes))
 306			return ERR_PTR(-EINVAL);
 307		return NULL;
 308	}
 309	VM_BUG_ON(!nodes);
 310
 311	/*
 312	 * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 313	 * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 314	 * All other modes require a valid pointer to a non-empty nodemask.
 315	 */
 316	if (mode == MPOL_PREFERRED) {
 317		if (nodes_empty(*nodes)) {
 318			if (((flags & MPOL_F_STATIC_NODES) ||
 319			     (flags & MPOL_F_RELATIVE_NODES)))
 320				return ERR_PTR(-EINVAL);
 321
 322			mode = MPOL_LOCAL;
 323		}
 324	} else if (mode == MPOL_LOCAL) {
 325		if (!nodes_empty(*nodes) ||
 326		    (flags & MPOL_F_STATIC_NODES) ||
 327		    (flags & MPOL_F_RELATIVE_NODES))
 328			return ERR_PTR(-EINVAL);
 
 329	} else if (nodes_empty(*nodes))
 330		return ERR_PTR(-EINVAL);
 331
 332	policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 333	if (!policy)
 334		return ERR_PTR(-ENOMEM);
 335	atomic_set(&policy->refcnt, 1);
 336	policy->mode = mode;
 337	policy->flags = flags;
 338	policy->home_node = NUMA_NO_NODE;
 339
 340	return policy;
 341}
 342
 343/* Slow path of a mpol destructor. */
 344void __mpol_put(struct mempolicy *pol)
 345{
 346	if (!atomic_dec_and_test(&pol->refcnt))
 347		return;
 348	kmem_cache_free(policy_cache, pol);
 349}
 350
 351static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes)
 352{
 353}
 354
 355static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes)
 356{
 357	nodemask_t tmp;
 358
 359	if (pol->flags & MPOL_F_STATIC_NODES)
 360		nodes_and(tmp, pol->w.user_nodemask, *nodes);
 361	else if (pol->flags & MPOL_F_RELATIVE_NODES)
 362		mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 363	else {
 364		nodes_remap(tmp, pol->nodes, pol->w.cpuset_mems_allowed,
 365								*nodes);
 366		pol->w.cpuset_mems_allowed = *nodes;
 367	}
 368
 369	if (nodes_empty(tmp))
 370		tmp = *nodes;
 371
 372	pol->nodes = tmp;
 373}
 374
 375static void mpol_rebind_preferred(struct mempolicy *pol,
 376						const nodemask_t *nodes)
 377{
 378	pol->w.cpuset_mems_allowed = *nodes;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 379}
 380
 381/*
 382 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 383 *
 384 * Per-vma policies are protected by mmap_lock. Allocations using per-task
 385 * policies are protected by task->mems_allowed_seq to prevent a premature
 386 * OOM/allocation failure due to parallel nodemask modification.
 387 */
 388static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask)
 389{
 390	if (!pol || pol->mode == MPOL_LOCAL)
 391		return;
 392	if (!mpol_store_user_nodemask(pol) &&
 393	    nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 394		return;
 395
 396	mpol_ops[pol->mode].rebind(pol, newmask);
 397}
 398
 399/*
 400 * Wrapper for mpol_rebind_policy() that just requires task
 401 * pointer, and updates task mempolicy.
 402 *
 403 * Called with task's alloc_lock held.
 404 */
 
 405void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new)
 406{
 407	mpol_rebind_policy(tsk->mempolicy, new);
 408}
 409
 410/*
 411 * Rebind each vma in mm to new nodemask.
 412 *
 413 * Call holding a reference to mm.  Takes mm->mmap_lock during call.
 414 */
 
 415void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 416{
 417	struct vm_area_struct *vma;
 418	VMA_ITERATOR(vmi, mm, 0);
 419
 420	mmap_write_lock(mm);
 421	for_each_vma(vmi, vma) {
 422		vma_start_write(vma);
 423		mpol_rebind_policy(vma->vm_policy, new);
 424	}
 425	mmap_write_unlock(mm);
 426}
 427
 428static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 429	[MPOL_DEFAULT] = {
 430		.rebind = mpol_rebind_default,
 431	},
 432	[MPOL_INTERLEAVE] = {
 433		.create = mpol_new_nodemask,
 434		.rebind = mpol_rebind_nodemask,
 435	},
 436	[MPOL_PREFERRED] = {
 437		.create = mpol_new_preferred,
 438		.rebind = mpol_rebind_preferred,
 439	},
 440	[MPOL_BIND] = {
 441		.create = mpol_new_nodemask,
 442		.rebind = mpol_rebind_nodemask,
 443	},
 444	[MPOL_LOCAL] = {
 445		.rebind = mpol_rebind_default,
 446	},
 447	[MPOL_PREFERRED_MANY] = {
 448		.create = mpol_new_nodemask,
 449		.rebind = mpol_rebind_preferred,
 450	},
 451	[MPOL_WEIGHTED_INTERLEAVE] = {
 452		.create = mpol_new_nodemask,
 453		.rebind = mpol_rebind_nodemask,
 454	},
 455};
 456
 457static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
 458				unsigned long flags);
 459static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
 460				pgoff_t ilx, int *nid);
 461
 462static bool strictly_unmovable(unsigned long flags)
 463{
 464	/*
 465	 * STRICT without MOVE flags lets do_mbind() fail immediately with -EIO
 466	 * if any misplaced page is found.
 467	 */
 468	return (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ==
 469			 MPOL_MF_STRICT;
 470}
 471
 472struct migration_mpol {		/* for alloc_migration_target_by_mpol() */
 473	struct mempolicy *pol;
 474	pgoff_t ilx;
 475};
 476
 477struct queue_pages {
 478	struct list_head *pagelist;
 479	unsigned long flags;
 480	nodemask_t *nmask;
 481	unsigned long start;
 482	unsigned long end;
 483	struct vm_area_struct *first;
 484	struct folio *large;		/* note last large folio encountered */
 485	long nr_failed;			/* could not be isolated at this time */
 486};
 487
 488/*
 489 * Check if the folio's nid is in qp->nmask.
 490 *
 491 * If MPOL_MF_INVERT is set in qp->flags, check if the nid is
 492 * in the invert of qp->nmask.
 493 */
 494static inline bool queue_folio_required(struct folio *folio,
 495					struct queue_pages *qp)
 496{
 497	int nid = folio_nid(folio);
 498	unsigned long flags = qp->flags;
 499
 500	return node_isset(nid, *qp->nmask) == !(flags & MPOL_MF_INVERT);
 501}
 502
 503static void queue_folios_pmd(pmd_t *pmd, struct mm_walk *walk)
 
 504{
 505	struct folio *folio;
 
 506	struct queue_pages *qp = walk->private;
 
 507
 508	if (unlikely(is_pmd_migration_entry(*pmd))) {
 509		qp->nr_failed++;
 510		return;
 
 
 
 
 
 
 511	}
 512	folio = pmd_folio(*pmd);
 513	if (is_huge_zero_folio(folio)) {
 514		walk->action = ACTION_CONTINUE;
 515		return;
 516	}
 517	if (!queue_folio_required(folio, qp))
 518		return;
 519	if (!(qp->flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
 520	    !vma_migratable(walk->vma) ||
 521	    !migrate_folio_add(folio, qp->pagelist, qp->flags))
 522		qp->nr_failed++;
 
 
 
 
 523}
 524
 525/*
 526 * Scan through folios, checking if they satisfy the required conditions,
 527 * moving them from LRU to local pagelist for migration if they do (or not).
 528 *
 529 * queue_folios_pte_range() has two possible return values:
 530 * 0 - continue walking to scan for more, even if an existing folio on the
 531 *     wrong node could not be isolated and queued for migration.
 532 * -EIO - only MPOL_MF_STRICT was specified, without MPOL_MF_MOVE or ..._ALL,
 533 *        and an existing folio was on a node that does not follow the policy.
 534 */
 535static int queue_folios_pte_range(pmd_t *pmd, unsigned long addr,
 536			unsigned long end, struct mm_walk *walk)
 537{
 538	struct vm_area_struct *vma = walk->vma;
 539	struct folio *folio;
 540	struct queue_pages *qp = walk->private;
 541	unsigned long flags = qp->flags;
 542	pte_t *pte, *mapped_pte;
 543	pte_t ptent;
 544	spinlock_t *ptl;
 545
 546	ptl = pmd_trans_huge_lock(pmd, vma);
 547	if (ptl) {
 548		queue_folios_pmd(pmd, walk);
 549		spin_unlock(ptl);
 550		goto out;
 551	}
 552
 553	mapped_pte = pte = pte_offset_map_lock(walk->mm, pmd, addr, &ptl);
 554	if (!pte) {
 555		walk->action = ACTION_AGAIN;
 556		return 0;
 557	}
 
 558	for (; addr != end; pte++, addr += PAGE_SIZE) {
 559		ptent = ptep_get(pte);
 560		if (pte_none(ptent))
 561			continue;
 562		if (!pte_present(ptent)) {
 563			if (is_migration_entry(pte_to_swp_entry(ptent)))
 564				qp->nr_failed++;
 565			continue;
 566		}
 567		folio = vm_normal_folio(vma, addr, ptent);
 568		if (!folio || folio_is_zone_device(folio))
 569			continue;
 570		/*
 571		 * vm_normal_folio() filters out zero pages, but there might
 572		 * still be reserved folios to skip, perhaps in a VDSO.
 573		 */
 574		if (folio_test_reserved(folio))
 575			continue;
 576		if (!queue_folio_required(folio, qp))
 577			continue;
 578		if (folio_test_large(folio)) {
 579			/*
 580			 * A large folio can only be isolated from LRU once,
 581			 * but may be mapped by many PTEs (and Copy-On-Write may
 582			 * intersperse PTEs of other, order 0, folios).  This is
 583			 * a common case, so don't mistake it for failure (but
 584			 * there can be other cases of multi-mapped pages which
 585			 * this quick check does not help to filter out - and a
 586			 * search of the pagelist might grow to be prohibitive).
 587			 *
 588			 * migrate_pages(&pagelist) returns nr_failed folios, so
 589			 * check "large" now so that queue_pages_range() returns
 590			 * a comparable nr_failed folios.  This does imply that
 591			 * if folio could not be isolated for some racy reason
 592			 * at its first PTE, later PTEs will not give it another
 593			 * chance of isolation; but keeps the accounting simple.
 594			 */
 595			if (folio == qp->large)
 596				continue;
 597			qp->large = folio;
 598		}
 599		if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
 600		    !vma_migratable(vma) ||
 601		    !migrate_folio_add(folio, qp->pagelist, flags)) {
 602			qp->nr_failed++;
 603			if (strictly_unmovable(flags))
 604				break;
 605		}
 606	}
 607	pte_unmap_unlock(mapped_pte, ptl);
 608	cond_resched();
 609out:
 610	if (qp->nr_failed && strictly_unmovable(flags))
 611		return -EIO;
 612	return 0;
 613}
 614
 615static int queue_folios_hugetlb(pte_t *pte, unsigned long hmask,
 616			       unsigned long addr, unsigned long end,
 617			       struct mm_walk *walk)
 618{
 619#ifdef CONFIG_HUGETLB_PAGE
 620	struct queue_pages *qp = walk->private;
 621	unsigned long flags = qp->flags;
 622	struct folio *folio;
 623	spinlock_t *ptl;
 624	pte_t entry;
 625
 626	ptl = huge_pte_lock(hstate_vma(walk->vma), walk->mm, pte);
 627	entry = huge_ptep_get(walk->mm, addr, pte);
 628	if (!pte_present(entry)) {
 629		if (unlikely(is_hugetlb_entry_migration(entry)))
 630			qp->nr_failed++;
 631		goto unlock;
 632	}
 633	folio = pfn_folio(pte_pfn(entry));
 634	if (!queue_folio_required(folio, qp))
 635		goto unlock;
 636	if (!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) ||
 637	    !vma_migratable(walk->vma)) {
 638		qp->nr_failed++;
 639		goto unlock;
 640	}
 641	/*
 642	 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
 643	 * Choosing not to migrate a shared folio is not counted as a failure.
 644	 *
 645	 * See folio_likely_mapped_shared() on possible imprecision when we
 646	 * cannot easily detect if a folio is shared.
 647	 */
 648	if ((flags & MPOL_MF_MOVE_ALL) ||
 649	    (!folio_likely_mapped_shared(folio) && !hugetlb_pmd_shared(pte)))
 650		if (!isolate_hugetlb(folio, qp->pagelist))
 651			qp->nr_failed++;
 652unlock:
 653	spin_unlock(ptl);
 654	if (qp->nr_failed && strictly_unmovable(flags))
 655		return -EIO;
 656#endif
 657	return 0;
 658}
 659
 660#ifdef CONFIG_NUMA_BALANCING
 661/*
 662 * This is used to mark a range of virtual addresses to be inaccessible.
 663 * These are later cleared by a NUMA hinting fault. Depending on these
 664 * faults, pages may be migrated for better NUMA placement.
 665 *
 666 * This is assuming that NUMA faults are handled using PROT_NONE. If
 667 * an architecture makes a different choice, it will need further
 668 * changes to the core.
 669 */
 670unsigned long change_prot_numa(struct vm_area_struct *vma,
 671			unsigned long addr, unsigned long end)
 672{
 673	struct mmu_gather tlb;
 674	long nr_updated;
 675
 676	tlb_gather_mmu(&tlb, vma->vm_mm);
 677
 678	nr_updated = change_protection(&tlb, vma, addr, end, MM_CP_PROT_NUMA);
 679	if (nr_updated > 0) {
 680		count_vm_numa_events(NUMA_PTE_UPDATES, nr_updated);
 681		count_memcg_events_mm(vma->vm_mm, NUMA_PTE_UPDATES, nr_updated);
 682	}
 683
 684	tlb_finish_mmu(&tlb);
 685
 686	return nr_updated;
 687}
 
 
 
 
 
 
 688#endif /* CONFIG_NUMA_BALANCING */
 689
 690static int queue_pages_test_walk(unsigned long start, unsigned long end,
 691				struct mm_walk *walk)
 692{
 693	struct vm_area_struct *next, *vma = walk->vma;
 694	struct queue_pages *qp = walk->private;
 
 695	unsigned long flags = qp->flags;
 696
 697	/* range check first */
 698	VM_BUG_ON_VMA(!range_in_vma(vma, start, end), vma);
 
 
 
 
 
 699
 700	if (!qp->first) {
 701		qp->first = vma;
 702		if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 703			(qp->start < vma->vm_start))
 704			/* hole at head side of range */
 705			return -EFAULT;
 706	}
 707	next = find_vma(vma->vm_mm, vma->vm_end);
 708	if (!(flags & MPOL_MF_DISCONTIG_OK) &&
 709		((vma->vm_end < qp->end) &&
 710		(!next || vma->vm_end < next->vm_start)))
 711		/* hole at middle or tail of range */
 712		return -EFAULT;
 713
 714	/*
 715	 * Need check MPOL_MF_STRICT to return -EIO if possible
 716	 * regardless of vma_migratable
 717	 */
 718	if (!vma_migratable(vma) &&
 719	    !(flags & MPOL_MF_STRICT))
 
 
 720		return 1;
 
 721
 722	/*
 723	 * Check page nodes, and queue pages to move, in the current vma.
 724	 * But if no moving, and no strict checking, the scan can be skipped.
 725	 */
 726	if (flags & (MPOL_MF_STRICT | MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
 727		return 0;
 728	return 1;
 729}
 730
 731static const struct mm_walk_ops queue_pages_walk_ops = {
 732	.hugetlb_entry		= queue_folios_hugetlb,
 733	.pmd_entry		= queue_folios_pte_range,
 734	.test_walk		= queue_pages_test_walk,
 735	.walk_lock		= PGWALK_RDLOCK,
 736};
 737
 738static const struct mm_walk_ops queue_pages_lock_vma_walk_ops = {
 739	.hugetlb_entry		= queue_folios_hugetlb,
 740	.pmd_entry		= queue_folios_pte_range,
 741	.test_walk		= queue_pages_test_walk,
 742	.walk_lock		= PGWALK_WRLOCK,
 743};
 744
 745/*
 746 * Walk through page tables and collect pages to be migrated.
 747 *
 748 * If pages found in a given range are not on the required set of @nodes,
 749 * and migration is allowed, they are isolated and queued to @pagelist.
 750 *
 751 * queue_pages_range() may return:
 752 * 0 - all pages already on the right node, or successfully queued for moving
 753 *     (or neither strict checking nor moving requested: only range checking).
 754 * >0 - this number of misplaced folios could not be queued for moving
 755 *      (a hugetlbfs page or a transparent huge page being counted as 1).
 756 * -EIO - a misplaced page found, when MPOL_MF_STRICT specified without MOVEs.
 757 * -EFAULT - a hole in the memory range, when MPOL_MF_DISCONTIG_OK unspecified.
 758 */
 759static long
 760queue_pages_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 761		nodemask_t *nodes, unsigned long flags,
 762		struct list_head *pagelist)
 763{
 764	int err;
 765	struct queue_pages qp = {
 766		.pagelist = pagelist,
 767		.flags = flags,
 768		.nmask = nodes,
 769		.start = start,
 770		.end = end,
 771		.first = NULL,
 
 
 
 
 
 772	};
 773	const struct mm_walk_ops *ops = (flags & MPOL_MF_WRLOCK) ?
 774			&queue_pages_lock_vma_walk_ops : &queue_pages_walk_ops;
 775
 776	err = walk_page_range(mm, start, end, ops, &qp);
 777
 778	if (!qp.first)
 779		/* whole range in hole */
 780		err = -EFAULT;
 781
 782	return err ? : qp.nr_failed;
 783}
 784
 785/*
 786 * Apply policy to a single VMA
 787 * This must be called with the mmap_lock held for writing.
 788 */
 789static int vma_replace_policy(struct vm_area_struct *vma,
 790				struct mempolicy *pol)
 791{
 792	int err;
 793	struct mempolicy *old;
 794	struct mempolicy *new;
 795
 796	vma_assert_write_locked(vma);
 
 
 
 797
 798	new = mpol_dup(pol);
 799	if (IS_ERR(new))
 800		return PTR_ERR(new);
 801
 802	if (vma->vm_ops && vma->vm_ops->set_policy) {
 803		err = vma->vm_ops->set_policy(vma, new);
 804		if (err)
 805			goto err_out;
 806	}
 807
 808	old = vma->vm_policy;
 809	vma->vm_policy = new; /* protected by mmap_lock */
 810	mpol_put(old);
 811
 812	return 0;
 813 err_out:
 814	mpol_put(new);
 815	return err;
 816}
 817
 818/* Split or merge the VMA (if required) and apply the new policy */
 819static int mbind_range(struct vma_iterator *vmi, struct vm_area_struct *vma,
 820		struct vm_area_struct **prev, unsigned long start,
 821		unsigned long end, struct mempolicy *new_pol)
 822{
 823	unsigned long vmstart, vmend;
 824
 825	vmend = min(end, vma->vm_end);
 826	if (start > vma->vm_start) {
 827		*prev = vma;
 828		vmstart = start;
 829	} else {
 830		vmstart = vma->vm_start;
 831	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 832
 833	if (mpol_equal(vma->vm_policy, new_pol)) {
 834		*prev = vma;
 835		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 836	}
 837
 838	vma =  vma_modify_policy(vmi, *prev, vma, vmstart, vmend, new_pol);
 839	if (IS_ERR(vma))
 840		return PTR_ERR(vma);
 841
 842	*prev = vma;
 843	return vma_replace_policy(vma, new_pol);
 844}
 845
 846/* Set the process memory policy */
 847static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 848			     nodemask_t *nodes)
 849{
 850	struct mempolicy *new, *old;
 851	NODEMASK_SCRATCH(scratch);
 852	int ret;
 853
 854	if (!scratch)
 855		return -ENOMEM;
 856
 857	new = mpol_new(mode, flags, nodes);
 858	if (IS_ERR(new)) {
 859		ret = PTR_ERR(new);
 860		goto out;
 861	}
 862
 863	task_lock(current);
 864	ret = mpol_set_nodemask(new, nodes, scratch);
 865	if (ret) {
 866		task_unlock(current);
 867		mpol_put(new);
 868		goto out;
 869	}
 870
 871	old = current->mempolicy;
 872	current->mempolicy = new;
 873	if (new && (new->mode == MPOL_INTERLEAVE ||
 874		    new->mode == MPOL_WEIGHTED_INTERLEAVE)) {
 875		current->il_prev = MAX_NUMNODES-1;
 876		current->il_weight = 0;
 877	}
 878	task_unlock(current);
 879	mpol_put(old);
 880	ret = 0;
 881out:
 882	NODEMASK_SCRATCH_FREE(scratch);
 883	return ret;
 884}
 885
 886/*
 887 * Return nodemask for policy for get_mempolicy() query
 888 *
 889 * Called with task's alloc_lock held
 890 */
 891static void get_policy_nodemask(struct mempolicy *pol, nodemask_t *nodes)
 892{
 893	nodes_clear(*nodes);
 894	if (pol == &default_policy)
 895		return;
 896
 897	switch (pol->mode) {
 898	case MPOL_BIND:
 
 899	case MPOL_INTERLEAVE:
 
 
 900	case MPOL_PREFERRED:
 901	case MPOL_PREFERRED_MANY:
 902	case MPOL_WEIGHTED_INTERLEAVE:
 903		*nodes = pol->nodes;
 904		break;
 905	case MPOL_LOCAL:
 906		/* return empty node mask for local allocation */
 907		break;
 908	default:
 909		BUG();
 910	}
 911}
 912
 913static int lookup_node(struct mm_struct *mm, unsigned long addr)
 914{
 915	struct page *p = NULL;
 916	int ret;
 917
 918	ret = get_user_pages_fast(addr & PAGE_MASK, 1, 0, &p);
 919	if (ret > 0) {
 920		ret = page_to_nid(p);
 921		put_page(p);
 922	}
 923	return ret;
 924}
 925
 926/* Retrieve NUMA policy */
 927static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 928			     unsigned long addr, unsigned long flags)
 929{
 930	int err;
 931	struct mm_struct *mm = current->mm;
 932	struct vm_area_struct *vma = NULL;
 933	struct mempolicy *pol = current->mempolicy, *pol_refcount = NULL;
 934
 935	if (flags &
 936		~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 937		return -EINVAL;
 938
 939	if (flags & MPOL_F_MEMS_ALLOWED) {
 940		if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 941			return -EINVAL;
 942		*policy = 0;	/* just so it's initialized */
 943		task_lock(current);
 944		*nmask  = cpuset_current_mems_allowed;
 945		task_unlock(current);
 946		return 0;
 947	}
 948
 949	if (flags & MPOL_F_ADDR) {
 950		pgoff_t ilx;		/* ignored here */
 951		/*
 952		 * Do NOT fall back to task policy if the
 953		 * vma/shared policy at addr is NULL.  We
 954		 * want to return MPOL_DEFAULT in this case.
 955		 */
 956		mmap_read_lock(mm);
 957		vma = vma_lookup(mm, addr);
 958		if (!vma) {
 959			mmap_read_unlock(mm);
 960			return -EFAULT;
 961		}
 962		pol = __get_vma_policy(vma, addr, &ilx);
 
 
 
 963	} else if (addr)
 964		return -EINVAL;
 965
 966	if (!pol)
 967		pol = &default_policy;	/* indicates default behavior */
 968
 969	if (flags & MPOL_F_NODE) {
 970		if (flags & MPOL_F_ADDR) {
 971			/*
 972			 * Take a refcount on the mpol, because we are about to
 973			 * drop the mmap_lock, after which only "pol" remains
 974			 * valid, "vma" is stale.
 975			 */
 976			pol_refcount = pol;
 977			vma = NULL;
 978			mpol_get(pol);
 979			mmap_read_unlock(mm);
 980			err = lookup_node(mm, addr);
 981			if (err < 0)
 982				goto out;
 983			*policy = err;
 984		} else if (pol == current->mempolicy &&
 985				pol->mode == MPOL_INTERLEAVE) {
 986			*policy = next_node_in(current->il_prev, pol->nodes);
 987		} else if (pol == current->mempolicy &&
 988				pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
 989			if (current->il_weight)
 990				*policy = current->il_prev;
 991			else
 992				*policy = next_node_in(current->il_prev,
 993						       pol->nodes);
 994		} else {
 995			err = -EINVAL;
 996			goto out;
 997		}
 998	} else {
 999		*policy = pol == &default_policy ? MPOL_DEFAULT :
1000						pol->mode;
1001		/*
1002		 * Internal mempolicy flags must be masked off before exposing
1003		 * the policy to userspace.
1004		 */
1005		*policy |= (pol->flags & MPOL_MODE_FLAGS);
1006	}
1007
1008	err = 0;
1009	if (nmask) {
1010		if (mpol_store_user_nodemask(pol)) {
1011			*nmask = pol->w.user_nodemask;
1012		} else {
1013			task_lock(current);
1014			get_policy_nodemask(pol, nmask);
1015			task_unlock(current);
1016		}
1017	}
1018
1019 out:
1020	mpol_cond_put(pol);
1021	if (vma)
1022		mmap_read_unlock(mm);
1023	if (pol_refcount)
1024		mpol_put(pol_refcount);
1025	return err;
1026}
1027
1028#ifdef CONFIG_MIGRATION
1029static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
 
 
 
1030				unsigned long flags)
1031{
 
1032	/*
1033	 * Unless MPOL_MF_MOVE_ALL, we try to avoid migrating a shared folio.
1034	 * Choosing not to migrate a shared folio is not counted as a failure.
1035	 *
1036	 * See folio_likely_mapped_shared() on possible imprecision when we
1037	 * cannot easily detect if a folio is shared.
1038	 */
1039	if ((flags & MPOL_MF_MOVE_ALL) || !folio_likely_mapped_shared(folio)) {
1040		if (folio_isolate_lru(folio)) {
1041			list_add_tail(&folio->lru, foliolist);
1042			node_stat_mod_folio(folio,
1043				NR_ISOLATED_ANON + folio_is_file_lru(folio),
1044				folio_nr_pages(folio));
1045		} else {
1046			/*
1047			 * Non-movable folio may reach here.  And, there may be
1048			 * temporary off LRU folios or non-LRU movable folios.
1049			 * Treat them as unmovable folios since they can't be
1050			 * isolated, so they can't be moved at the moment.
1051			 */
1052			return false;
1053		}
1054	}
1055	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 long 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	long nr_failed;
1069	long err = 0;
1070	struct migration_target_control mtc = {
1071		.nid = dest,
1072		.gfp_mask = GFP_HIGHUSER_MOVABLE | __GFP_THISNODE,
1073		.reason = MR_SYSCALL,
1074	};
1075
1076	nodes_clear(nmask);
1077	node_set(source, nmask);
1078
1079	VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
1080
1081	mmap_read_lock(mm);
1082	vma = find_vma(mm, 0);
1083	if (unlikely(!vma)) {
1084		mmap_read_unlock(mm);
1085		return 0;
1086	}
1087
1088	/*
1089	 * This does not migrate the range, but isolates all pages that
1090	 * need migration.  Between passing in the full user address
1091	 * space range and MPOL_MF_DISCONTIG_OK, this call cannot fail,
1092	 * but passes back the count of pages which could not be isolated.
1093	 */
1094	nr_failed = queue_pages_range(mm, vma->vm_start, mm->task_size, &nmask,
1095				      flags | MPOL_MF_DISCONTIG_OK, &pagelist);
1096	mmap_read_unlock(mm);
1097
1098	if (!list_empty(&pagelist)) {
1099		err = migrate_pages(&pagelist, alloc_migration_target, NULL,
1100			(unsigned long)&mtc, MIGRATE_SYNC, MR_SYSCALL, NULL);
1101		if (err)
1102			putback_movable_pages(&pagelist);
1103	}
1104
1105	if (err >= 0)
1106		err += nr_failed;
1107	return err;
1108}
1109
1110/*
1111 * Move pages between the two nodesets so as to preserve the physical
1112 * layout as much as possible.
1113 *
1114 * Returns the number of page that could not be moved.
1115 */
1116int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1117		     const nodemask_t *to, int flags)
1118{
1119	long nr_failed = 0;
1120	long err = 0;
1121	nodemask_t tmp;
1122
1123	lru_cache_disable();
 
 
 
 
1124
1125	/*
1126	 * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
1127	 * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
1128	 * bit in 'tmp', and return that <source, dest> pair for migration.
1129	 * The pair of nodemasks 'to' and 'from' define the map.
1130	 *
1131	 * If no pair of bits is found that way, fallback to picking some
1132	 * pair of 'source' and 'dest' bits that are not the same.  If the
1133	 * 'source' and 'dest' bits are the same, this represents a node
1134	 * that will be migrating to itself, so no pages need move.
1135	 *
1136	 * If no bits are left in 'tmp', or if all remaining bits left
1137	 * in 'tmp' correspond to the same bit in 'to', return false
1138	 * (nothing left to migrate).
1139	 *
1140	 * This lets us pick a pair of nodes to migrate between, such that
1141	 * if possible the dest node is not already occupied by some other
1142	 * source node, minimizing the risk of overloading the memory on a
1143	 * node that would happen if we migrated incoming memory to a node
1144	 * before migrating outgoing memory source that same node.
1145	 *
1146	 * A single scan of tmp is sufficient.  As we go, we remember the
1147	 * most recent <s, d> pair that moved (s != d).  If we find a pair
1148	 * that not only moved, but what's better, moved to an empty slot
1149	 * (d is not set in tmp), then we break out then, with that pair.
1150	 * Otherwise when we finish scanning from_tmp, we at least have the
1151	 * most recent <s, d> pair that moved.  If we get all the way through
1152	 * the scan of tmp without finding any node that moved, much less
1153	 * moved to an empty node, then there is nothing left worth migrating.
1154	 */
1155
1156	tmp = *from;
1157	while (!nodes_empty(tmp)) {
1158		int s, d;
1159		int source = NUMA_NO_NODE;
1160		int dest = 0;
1161
1162		for_each_node_mask(s, tmp) {
1163
1164			/*
1165			 * do_migrate_pages() tries to maintain the relative
1166			 * node relationship of the pages established between
1167			 * threads and memory areas.
1168                         *
1169			 * However if the number of source nodes is not equal to
1170			 * the number of destination nodes we can not preserve
1171			 * this node relative relationship.  In that case, skip
1172			 * copying memory from a node that is in the destination
1173			 * mask.
1174			 *
1175			 * Example: [2,3,4] -> [3,4,5] moves everything.
1176			 *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1177			 */
1178
1179			if ((nodes_weight(*from) != nodes_weight(*to)) &&
1180						(node_isset(s, *to)))
1181				continue;
1182
1183			d = node_remap(s, *from, *to);
1184			if (s == d)
1185				continue;
1186
1187			source = s;	/* Node moved. Memorize */
1188			dest = d;
1189
1190			/* dest not in remaining from nodes? */
1191			if (!node_isset(dest, tmp))
1192				break;
1193		}
1194		if (source == NUMA_NO_NODE)
1195			break;
1196
1197		node_clear(source, tmp);
1198		err = migrate_to_node(mm, source, dest, flags);
1199		if (err > 0)
1200			nr_failed += err;
1201		if (err < 0)
1202			break;
1203	}
1204
1205	lru_cache_enable();
1206	if (err < 0)
1207		return err;
1208	return (nr_failed < INT_MAX) ? nr_failed : INT_MAX;
 
1209}
1210
1211/*
1212 * Allocate a new folio for page migration, according to NUMA mempolicy.
 
 
 
 
1213 */
1214static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1215						    unsigned long private)
1216{
1217	struct migration_mpol *mmpol = (struct migration_mpol *)private;
1218	struct mempolicy *pol = mmpol->pol;
1219	pgoff_t ilx = mmpol->ilx;
1220	unsigned int order;
1221	int nid = numa_node_id();
1222	gfp_t gfp;
1223
1224	order = folio_order(src);
1225	ilx += src->index >> order;
 
 
 
 
 
1226
1227	if (folio_test_hugetlb(src)) {
1228		nodemask_t *nodemask;
1229		struct hstate *h;
 
 
1230
1231		h = folio_hstate(src);
1232		gfp = htlb_alloc_mask(h);
1233		nodemask = policy_nodemask(gfp, pol, ilx, &nid);
1234		return alloc_hugetlb_folio_nodemask(h, nid, nodemask, gfp,
1235				htlb_allow_alloc_fallback(MR_MEMPOLICY_MBIND));
 
1236	}
1237
1238	if (folio_test_large(src))
1239		gfp = GFP_TRANSHUGE;
1240	else
1241		gfp = GFP_HIGHUSER_MOVABLE | __GFP_RETRY_MAYFAIL | __GFP_COMP;
1242
1243	return folio_alloc_mpol(gfp, order, pol, ilx, nid);
1244}
1245#else
1246
1247static bool migrate_folio_add(struct folio *folio, struct list_head *foliolist,
1248				unsigned long flags)
1249{
1250	return false;
1251}
1252
1253int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1254		     const nodemask_t *to, int flags)
1255{
1256	return -ENOSYS;
1257}
1258
1259static struct folio *alloc_migration_target_by_mpol(struct folio *src,
1260						    unsigned long private)
1261{
1262	return NULL;
1263}
1264#endif
1265
1266static long do_mbind(unsigned long start, unsigned long len,
1267		     unsigned short mode, unsigned short mode_flags,
1268		     nodemask_t *nmask, unsigned long flags)
1269{
1270	struct mm_struct *mm = current->mm;
1271	struct vm_area_struct *vma, *prev;
1272	struct vma_iterator vmi;
1273	struct migration_mpol mmpol;
1274	struct mempolicy *new;
1275	unsigned long end;
1276	long err;
1277	long nr_failed;
1278	LIST_HEAD(pagelist);
1279
1280	if (flags & ~(unsigned long)MPOL_MF_VALID)
1281		return -EINVAL;
1282	if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1283		return -EPERM;
1284
1285	if (start & ~PAGE_MASK)
1286		return -EINVAL;
1287
1288	if (mode == MPOL_DEFAULT)
1289		flags &= ~MPOL_MF_STRICT;
1290
1291	len = PAGE_ALIGN(len);
1292	end = start + len;
1293
1294	if (end < start)
1295		return -EINVAL;
1296	if (end == start)
1297		return 0;
1298
1299	new = mpol_new(mode, mode_flags, nmask);
1300	if (IS_ERR(new))
1301		return PTR_ERR(new);
1302
 
 
 
1303	/*
1304	 * If we are using the default policy then operation
1305	 * on discontinuous address spaces is okay after all
1306	 */
1307	if (!new)
1308		flags |= MPOL_MF_DISCONTIG_OK;
1309
1310	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1311		lru_cache_disable();
 
 
 
 
 
 
 
 
1312	{
1313		NODEMASK_SCRATCH(scratch);
1314		if (scratch) {
1315			mmap_write_lock(mm);
 
1316			err = mpol_set_nodemask(new, nmask, scratch);
 
1317			if (err)
1318				mmap_write_unlock(mm);
1319		} else
1320			err = -ENOMEM;
1321		NODEMASK_SCRATCH_FREE(scratch);
1322	}
1323	if (err)
1324		goto mpol_out;
1325
1326	/*
1327	 * Lock the VMAs before scanning for pages to migrate,
1328	 * to ensure we don't miss a concurrently inserted page.
1329	 */
1330	nr_failed = queue_pages_range(mm, start, end, nmask,
1331			flags | MPOL_MF_INVERT | MPOL_MF_WRLOCK, &pagelist);
1332
1333	if (nr_failed < 0) {
1334		err = nr_failed;
1335		nr_failed = 0;
1336	} else {
1337		vma_iter_init(&vmi, mm, start);
1338		prev = vma_prev(&vmi);
1339		for_each_vma_range(vmi, vma, end) {
1340			err = mbind_range(&vmi, vma, &prev, start, end, new);
1341			if (err)
1342				break;
1343		}
1344	}
1345
1346	if (!err && !list_empty(&pagelist)) {
1347		/* Convert MPOL_DEFAULT's NULL to task or default policy */
1348		if (!new) {
1349			new = get_task_policy(current);
1350			mpol_get(new);
 
1351		}
1352		mmpol.pol = new;
1353		mmpol.ilx = 0;
1354
1355		/*
1356		 * In the interleaved case, attempt to allocate on exactly the
1357		 * targeted nodes, for the first VMA to be migrated; for later
1358		 * VMAs, the nodes will still be interleaved from the targeted
1359		 * nodemask, but one by one may be selected differently.
1360		 */
1361		if (new->mode == MPOL_INTERLEAVE ||
1362		    new->mode == MPOL_WEIGHTED_INTERLEAVE) {
1363			struct folio *folio;
1364			unsigned int order;
1365			unsigned long addr = -EFAULT;
1366
1367			list_for_each_entry(folio, &pagelist, lru) {
1368				if (!folio_test_ksm(folio))
1369					break;
1370			}
1371			if (!list_entry_is_head(folio, &pagelist, lru)) {
1372				vma_iter_init(&vmi, mm, start);
1373				for_each_vma_range(vmi, vma, end) {
1374					addr = page_address_in_vma(folio,
1375						folio_page(folio, 0), vma);
1376					if (addr != -EFAULT)
1377						break;
1378				}
1379			}
1380			if (addr != -EFAULT) {
1381				order = folio_order(folio);
1382				/* We already know the pol, but not the ilx */
1383				mpol_cond_put(get_vma_policy(vma, addr, order,
1384							     &mmpol.ilx));
1385				/* Set base from which to increment by index */
1386				mmpol.ilx -= folio->index >> order;
1387			}
1388		}
1389	}
1390
1391	mmap_write_unlock(mm);
1392
1393	if (!err && !list_empty(&pagelist)) {
1394		nr_failed |= migrate_pages(&pagelist,
1395				alloc_migration_target_by_mpol, NULL,
1396				(unsigned long)&mmpol, MIGRATE_SYNC,
1397				MR_MEMPOLICY_MBIND, NULL);
1398	}
1399
1400	if (nr_failed && (flags & MPOL_MF_STRICT))
1401		err = -EIO;
1402	if (!list_empty(&pagelist))
1403		putback_movable_pages(&pagelist);
1404mpol_out:
1405	mpol_put(new);
1406	if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1407		lru_cache_enable();
1408	return err;
1409}
1410
1411/*
1412 * User space interface with variable sized bitmaps for nodelists.
1413 */
1414static int get_bitmap(unsigned long *mask, const unsigned long __user *nmask,
1415		      unsigned long maxnode)
1416{
1417	unsigned long nlongs = BITS_TO_LONGS(maxnode);
1418	int ret;
1419
1420	if (in_compat_syscall())
1421		ret = compat_get_bitmap(mask,
1422					(const compat_ulong_t __user *)nmask,
1423					maxnode);
1424	else
1425		ret = copy_from_user(mask, nmask,
1426				     nlongs * sizeof(unsigned long));
1427
1428	if (ret)
1429		return -EFAULT;
1430
1431	if (maxnode % BITS_PER_LONG)
1432		mask[nlongs - 1] &= (1UL << (maxnode % BITS_PER_LONG)) - 1;
1433
1434	return 0;
1435}
1436
1437/* Copy a node mask from user space. */
1438static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1439		     unsigned long maxnode)
1440{
 
 
 
 
 
1441	--maxnode;
1442	nodes_clear(*nodes);
1443	if (maxnode == 0 || !nmask)
1444		return 0;
1445	if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1446		return -EINVAL;
1447
 
 
 
 
 
 
1448	/*
1449	 * When the user specified more nodes than supported just check
1450	 * if the non supported part is all zero, one word at a time,
1451	 * starting at the end.
1452	 */
1453	while (maxnode > MAX_NUMNODES) {
1454		unsigned long bits = min_t(unsigned long, maxnode, BITS_PER_LONG);
1455		unsigned long t;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1456
1457		if (get_bitmap(&t, &nmask[(maxnode - 1) / BITS_PER_LONG], bits))
 
1458			return -EFAULT;
1459
1460		if (maxnode - bits >= MAX_NUMNODES) {
1461			maxnode -= bits;
1462		} else {
1463			maxnode = MAX_NUMNODES;
1464			t &= ~((1UL << (MAX_NUMNODES % BITS_PER_LONG)) - 1);
1465		}
1466		if (t)
1467			return -EINVAL;
1468	}
1469
1470	return get_bitmap(nodes_addr(*nodes), nmask, maxnode);
 
 
 
1471}
1472
1473/* Copy a kernel node mask to user space */
1474static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1475			      nodemask_t *nodes)
1476{
1477	unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1478	unsigned int nbytes = BITS_TO_LONGS(nr_node_ids) * sizeof(long);
1479	bool compat = in_compat_syscall();
1480
1481	if (compat)
1482		nbytes = BITS_TO_COMPAT_LONGS(nr_node_ids) * sizeof(compat_long_t);
1483
1484	if (copy > nbytes) {
1485		if (copy > PAGE_SIZE)
1486			return -EINVAL;
1487		if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1488			return -EFAULT;
1489		copy = nbytes;
1490		maxnode = nr_node_ids;
1491	}
1492
1493	if (compat)
1494		return compat_put_bitmap((compat_ulong_t __user *)mask,
1495					 nodes_addr(*nodes), maxnode);
1496
1497	return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1498}
1499
1500/* Basic parameter sanity check used by both mbind() and set_mempolicy() */
1501static inline int sanitize_mpol_flags(int *mode, unsigned short *flags)
1502{
1503	*flags = *mode & MPOL_MODE_FLAGS;
1504	*mode &= ~MPOL_MODE_FLAGS;
1505
1506	if ((unsigned int)(*mode) >=  MPOL_MAX)
1507		return -EINVAL;
1508	if ((*flags & MPOL_F_STATIC_NODES) && (*flags & MPOL_F_RELATIVE_NODES))
1509		return -EINVAL;
1510	if (*flags & MPOL_F_NUMA_BALANCING) {
1511		if (*mode == MPOL_BIND || *mode == MPOL_PREFERRED_MANY)
1512			*flags |= (MPOL_F_MOF | MPOL_F_MORON);
1513		else
1514			return -EINVAL;
1515	}
1516	return 0;
1517}
1518
1519static long kernel_mbind(unsigned long start, unsigned long len,
1520			 unsigned long mode, const unsigned long __user *nmask,
1521			 unsigned long maxnode, unsigned int flags)
1522{
1523	unsigned short mode_flags;
1524	nodemask_t nodes;
1525	int lmode = mode;
1526	int err;
 
1527
1528	start = untagged_addr(start);
1529	err = sanitize_mpol_flags(&lmode, &mode_flags);
1530	if (err)
1531		return err;
1532
 
 
1533	err = get_nodes(&nodes, nmask, maxnode);
1534	if (err)
1535		return err;
1536
1537	return do_mbind(start, len, lmode, mode_flags, &nodes, flags);
1538}
1539
1540SYSCALL_DEFINE4(set_mempolicy_home_node, unsigned long, start, unsigned long, len,
1541		unsigned long, home_node, unsigned long, flags)
1542{
1543	struct mm_struct *mm = current->mm;
1544	struct vm_area_struct *vma, *prev;
1545	struct mempolicy *new, *old;
1546	unsigned long end;
1547	int err = -ENOENT;
1548	VMA_ITERATOR(vmi, mm, start);
1549
1550	start = untagged_addr(start);
1551	if (start & ~PAGE_MASK)
1552		return -EINVAL;
1553	/*
1554	 * flags is used for future extension if any.
1555	 */
1556	if (flags != 0)
1557		return -EINVAL;
1558
1559	/*
1560	 * Check home_node is online to avoid accessing uninitialized
1561	 * NODE_DATA.
1562	 */
1563	if (home_node >= MAX_NUMNODES || !node_online(home_node))
1564		return -EINVAL;
1565
1566	len = PAGE_ALIGN(len);
1567	end = start + len;
1568
1569	if (end < start)
1570		return -EINVAL;
1571	if (end == start)
1572		return 0;
1573	mmap_write_lock(mm);
1574	prev = vma_prev(&vmi);
1575	for_each_vma_range(vmi, vma, end) {
1576		/*
1577		 * If any vma in the range got policy other than MPOL_BIND
1578		 * or MPOL_PREFERRED_MANY we return error. We don't reset
1579		 * the home node for vmas we already updated before.
1580		 */
1581		old = vma_policy(vma);
1582		if (!old) {
1583			prev = vma;
1584			continue;
1585		}
1586		if (old->mode != MPOL_BIND && old->mode != MPOL_PREFERRED_MANY) {
1587			err = -EOPNOTSUPP;
1588			break;
1589		}
1590		new = mpol_dup(old);
1591		if (IS_ERR(new)) {
1592			err = PTR_ERR(new);
1593			break;
1594		}
1595
1596		vma_start_write(vma);
1597		new->home_node = home_node;
1598		err = mbind_range(&vmi, vma, &prev, start, end, new);
1599		mpol_put(new);
1600		if (err)
1601			break;
1602	}
1603	mmap_write_unlock(mm);
1604	return err;
1605}
1606
1607SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1608		unsigned long, mode, const unsigned long __user *, nmask,
1609		unsigned long, maxnode, unsigned int, flags)
1610{
1611	return kernel_mbind(start, len, mode, nmask, maxnode, flags);
1612}
1613
1614/* Set the process memory policy */
1615static long kernel_set_mempolicy(int mode, const unsigned long __user *nmask,
1616				 unsigned long maxnode)
1617{
1618	unsigned short mode_flags;
1619	nodemask_t nodes;
1620	int lmode = mode;
1621	int err;
1622
1623	err = sanitize_mpol_flags(&lmode, &mode_flags);
1624	if (err)
1625		return err;
1626
 
 
 
 
 
 
1627	err = get_nodes(&nodes, nmask, maxnode);
1628	if (err)
1629		return err;
1630
1631	return do_set_mempolicy(lmode, mode_flags, &nodes);
1632}
1633
1634SYSCALL_DEFINE3(set_mempolicy, int, mode, const unsigned long __user *, nmask,
1635		unsigned long, maxnode)
1636{
1637	return kernel_set_mempolicy(mode, nmask, maxnode);
1638}
1639
1640static int kernel_migrate_pages(pid_t pid, unsigned long maxnode,
1641				const unsigned long __user *old_nodes,
1642				const unsigned long __user *new_nodes)
1643{
1644	struct mm_struct *mm = NULL;
1645	struct task_struct *task;
1646	nodemask_t task_nodes;
1647	int err;
1648	nodemask_t *old;
1649	nodemask_t *new;
1650	NODEMASK_SCRATCH(scratch);
1651
1652	if (!scratch)
1653		return -ENOMEM;
1654
1655	old = &scratch->mask1;
1656	new = &scratch->mask2;
1657
1658	err = get_nodes(old, old_nodes, maxnode);
1659	if (err)
1660		goto out;
1661
1662	err = get_nodes(new, new_nodes, maxnode);
1663	if (err)
1664		goto out;
1665
1666	/* Find the mm_struct */
1667	rcu_read_lock();
1668	task = pid ? find_task_by_vpid(pid) : current;
1669	if (!task) {
1670		rcu_read_unlock();
1671		err = -ESRCH;
1672		goto out;
1673	}
1674	get_task_struct(task);
1675
1676	err = -EINVAL;
1677
1678	/*
1679	 * Check if this process has the right to modify the specified process.
1680	 * Use the regular "ptrace_may_access()" checks.
1681	 */
1682	if (!ptrace_may_access(task, PTRACE_MODE_READ_REALCREDS)) {
1683		rcu_read_unlock();
1684		err = -EPERM;
1685		goto out_put;
1686	}
1687	rcu_read_unlock();
1688
1689	task_nodes = cpuset_mems_allowed(task);
1690	/* Is the user allowed to access the target nodes? */
1691	if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1692		err = -EPERM;
1693		goto out_put;
1694	}
1695
1696	task_nodes = cpuset_mems_allowed(current);
1697	nodes_and(*new, *new, task_nodes);
1698	if (nodes_empty(*new))
1699		goto out_put;
1700
 
 
 
 
1701	err = security_task_movememory(task);
1702	if (err)
1703		goto out_put;
1704
1705	mm = get_task_mm(task);
1706	put_task_struct(task);
1707
1708	if (!mm) {
1709		err = -EINVAL;
1710		goto out;
1711	}
1712
1713	err = do_migrate_pages(mm, old, new,
1714		capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1715
1716	mmput(mm);
1717out:
1718	NODEMASK_SCRATCH_FREE(scratch);
1719
1720	return err;
1721
1722out_put:
1723	put_task_struct(task);
1724	goto out;
 
1725}
1726
1727SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1728		const unsigned long __user *, old_nodes,
1729		const unsigned long __user *, new_nodes)
1730{
1731	return kernel_migrate_pages(pid, maxnode, old_nodes, new_nodes);
1732}
1733
 
1734/* Retrieve NUMA policy */
1735static int kernel_get_mempolicy(int __user *policy,
1736				unsigned long __user *nmask,
1737				unsigned long maxnode,
1738				unsigned long addr,
1739				unsigned long flags)
1740{
1741	int err;
1742	int pval;
1743	nodemask_t nodes;
1744
1745	if (nmask != NULL && maxnode < nr_node_ids)
1746		return -EINVAL;
1747
1748	addr = untagged_addr(addr);
1749
1750	err = do_get_mempolicy(&pval, &nodes, addr, flags);
1751
1752	if (err)
1753		return err;
1754
1755	if (policy && put_user(pval, policy))
1756		return -EFAULT;
1757
1758	if (nmask)
1759		err = copy_nodes_to_user(nmask, maxnode, &nodes);
1760
1761	return err;
1762}
1763
1764SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1765		unsigned long __user *, nmask, unsigned long, maxnode,
1766		unsigned long, addr, unsigned long, flags)
1767{
1768	return kernel_get_mempolicy(policy, nmask, maxnode, addr, flags);
1769}
1770
1771bool vma_migratable(struct vm_area_struct *vma)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1772{
1773	if (vma->vm_flags & (VM_IO | VM_PFNMAP))
1774		return false;
 
 
 
 
1775
1776	/*
1777	 * DAX device mappings require predictable access latency, so avoid
1778	 * incurring periodic faults.
1779	 */
1780	if (vma_is_dax(vma))
1781		return false;
 
1782
1783	if (is_vm_hugetlb_page(vma) &&
1784		!hugepage_migration_supported(hstate_vma(vma)))
1785		return false;
1786
1787	/*
1788	 * Migration allocates pages in the highest zone. If we cannot
1789	 * do so then migration (at least from node to node) is not
1790	 * possible.
1791	 */
1792	if (vma->vm_file &&
1793		gfp_zone(mapping_gfp_mask(vma->vm_file->f_mapping))
1794			< policy_zone)
1795		return false;
1796	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1797}
1798
 
 
1799struct mempolicy *__get_vma_policy(struct vm_area_struct *vma,
1800				   unsigned long addr, pgoff_t *ilx)
1801{
1802	*ilx = 0;
1803	return (vma->vm_ops && vma->vm_ops->get_policy) ?
1804		vma->vm_ops->get_policy(vma, addr, ilx) : vma->vm_policy;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1805}
1806
1807/*
1808 * get_vma_policy(@vma, @addr, @order, @ilx)
1809 * @vma: virtual memory area whose policy is sought
1810 * @addr: address in @vma for shared policy lookup
1811 * @order: 0, or appropriate huge_page_order for interleaving
1812 * @ilx: interleave index (output), for use only when MPOL_INTERLEAVE or
1813 *       MPOL_WEIGHTED_INTERLEAVE
1814 *
1815 * Returns effective policy for a VMA at specified address.
1816 * Falls back to current->mempolicy or system default policy, as necessary.
1817 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1818 * count--added by the get_policy() vm_op, as appropriate--to protect against
1819 * freeing by another task.  It is the caller's responsibility to free the
1820 * extra reference for shared policies.
1821 */
1822struct mempolicy *get_vma_policy(struct vm_area_struct *vma,
1823				 unsigned long addr, int order, pgoff_t *ilx)
1824{
1825	struct mempolicy *pol;
1826
1827	pol = __get_vma_policy(vma, addr, ilx);
1828	if (!pol)
1829		pol = get_task_policy(current);
1830	if (pol->mode == MPOL_INTERLEAVE ||
1831	    pol->mode == MPOL_WEIGHTED_INTERLEAVE) {
1832		*ilx += vma->vm_pgoff >> order;
1833		*ilx += (addr - vma->vm_start) >> (PAGE_SHIFT + order);
1834	}
1835	return pol;
1836}
1837
1838bool vma_policy_mof(struct vm_area_struct *vma)
1839{
1840	struct mempolicy *pol;
1841
1842	if (vma->vm_ops && vma->vm_ops->get_policy) {
1843		bool ret = false;
1844		pgoff_t ilx;		/* ignored here */
1845
1846		pol = vma->vm_ops->get_policy(vma, vma->vm_start, &ilx);
1847		if (pol && (pol->flags & MPOL_F_MOF))
1848			ret = true;
1849		mpol_cond_put(pol);
1850
1851		return ret;
1852	}
1853
1854	pol = vma->vm_policy;
1855	if (!pol)
1856		pol = get_task_policy(current);
1857
1858	return pol->flags & MPOL_F_MOF;
1859}
1860
1861bool apply_policy_zone(struct mempolicy *policy, enum zone_type zone)
1862{
1863	enum zone_type dynamic_policy_zone = policy_zone;
1864
1865	BUG_ON(dynamic_policy_zone == ZONE_MOVABLE);
1866
1867	/*
1868	 * if policy->nodes has movable memory only,
1869	 * we apply policy when gfp_zone(gfp) = ZONE_MOVABLE only.
1870	 *
1871	 * policy->nodes is intersect with node_states[N_MEMORY].
1872	 * so if the following test fails, it implies
1873	 * policy->nodes has movable memory only.
1874	 */
1875	if (!nodes_intersects(policy->nodes, node_states[N_HIGH_MEMORY]))
1876		dynamic_policy_zone = ZONE_MOVABLE;
1877
1878	return zone >= dynamic_policy_zone;
1879}
1880
1881static unsigned int weighted_interleave_nodes(struct mempolicy *policy)
 
 
 
 
1882{
1883	unsigned int node;
1884	unsigned int cpuset_mems_cookie;
 
 
 
 
 
 
1885
1886retry:
1887	/* to prevent miscount use tsk->mems_allowed_seq to detect rebind */
1888	cpuset_mems_cookie = read_mems_allowed_begin();
1889	node = current->il_prev;
1890	if (!current->il_weight || !node_isset(node, policy->nodes)) {
1891		node = next_node_in(node, policy->nodes);
1892		if (read_mems_allowed_retry(cpuset_mems_cookie))
1893			goto retry;
1894		if (node == MAX_NUMNODES)
1895			return node;
1896		current->il_prev = node;
1897		current->il_weight = get_il_weight(node);
 
1898	}
1899	current->il_weight--;
1900	return node;
1901}
1902
1903/* Do dynamic interleaving for a process */
1904static unsigned int interleave_nodes(struct mempolicy *policy)
1905{
1906	unsigned int nid;
1907	unsigned int cpuset_mems_cookie;
1908
1909	/* to prevent miscount, use tsk->mems_allowed_seq to detect rebind */
1910	do {
1911		cpuset_mems_cookie = read_mems_allowed_begin();
1912		nid = next_node_in(current->il_prev, policy->nodes);
1913	} while (read_mems_allowed_retry(cpuset_mems_cookie));
1914
1915	if (nid < MAX_NUMNODES)
1916		current->il_prev = nid;
1917	return nid;
 
1918}
1919
1920/*
1921 * Depending on the memory policy provide a node from which to allocate the
1922 * next slab entry.
1923 */
1924unsigned int mempolicy_slab_node(void)
1925{
1926	struct mempolicy *policy;
1927	int node = numa_mem_id();
1928
1929	if (!in_task())
1930		return node;
1931
1932	policy = current->mempolicy;
1933	if (!policy)
1934		return node;
1935
1936	switch (policy->mode) {
1937	case MPOL_PREFERRED:
1938		return first_node(policy->nodes);
 
 
 
1939
1940	case MPOL_INTERLEAVE:
1941		return interleave_nodes(policy);
1942
1943	case MPOL_WEIGHTED_INTERLEAVE:
1944		return weighted_interleave_nodes(policy);
1945
1946	case MPOL_BIND:
1947	case MPOL_PREFERRED_MANY:
1948	{
1949		struct zoneref *z;
1950
1951		/*
1952		 * Follow bind policy behavior and start allocation at the
1953		 * first node.
1954		 */
1955		struct zonelist *zonelist;
1956		enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1957		zonelist = &NODE_DATA(node)->node_zonelists[ZONELIST_FALLBACK];
1958		z = first_zones_zonelist(zonelist, highest_zoneidx,
1959							&policy->nodes);
1960		return zonelist_zone(z) ? zonelist_node_idx(z) : node;
1961	}
1962	case MPOL_LOCAL:
1963		return node;
1964
1965	default:
1966		BUG();
1967	}
1968}
1969
1970static unsigned int read_once_policy_nodemask(struct mempolicy *pol,
1971					      nodemask_t *mask)
1972{
1973	/*
1974	 * barrier stabilizes the nodemask locally so that it can be iterated
1975	 * over safely without concern for changes. Allocators validate node
1976	 * selection does not violate mems_allowed, so this is safe.
1977	 */
1978	barrier();
1979	memcpy(mask, &pol->nodes, sizeof(nodemask_t));
1980	barrier();
1981	return nodes_weight(*mask);
1982}
1983
1984static unsigned int weighted_interleave_nid(struct mempolicy *pol, pgoff_t ilx)
1985{
1986	nodemask_t nodemask;
1987	unsigned int target, nr_nodes;
1988	u8 *table;
1989	unsigned int weight_total = 0;
1990	u8 weight;
1991	int nid;
1992
1993	nr_nodes = read_once_policy_nodemask(pol, &nodemask);
1994	if (!nr_nodes)
1995		return numa_node_id();
1996
1997	rcu_read_lock();
1998	table = rcu_dereference(iw_table);
1999	/* calculate the total weight */
2000	for_each_node_mask(nid, nodemask) {
2001		/* detect system default usage */
2002		weight = table ? table[nid] : 1;
2003		weight = weight ? weight : 1;
2004		weight_total += weight;
2005	}
2006
2007	/* Calculate the node offset based on totals */
2008	target = ilx % weight_total;
2009	nid = first_node(nodemask);
2010	while (target) {
2011		/* detect system default usage */
2012		weight = table ? table[nid] : 1;
2013		weight = weight ? weight : 1;
2014		if (target < weight)
2015			break;
2016		target -= weight;
2017		nid = next_node_in(nid, nodemask);
2018	}
2019	rcu_read_unlock();
2020	return nid;
2021}
2022
2023/*
2024 * Do static interleaving for interleave index @ilx.  Returns the ilx'th
2025 * node in pol->nodes (starting from ilx=0), wrapping around if ilx
2026 * exceeds the number of present nodes.
2027 */
2028static unsigned int interleave_nid(struct mempolicy *pol, pgoff_t ilx)
2029{
2030	nodemask_t nodemask;
2031	unsigned int target, nnodes;
2032	int i;
2033	int nid;
2034
2035	nnodes = read_once_policy_nodemask(pol, &nodemask);
2036	if (!nnodes)
2037		return numa_node_id();
2038	target = ilx % nnodes;
2039	nid = first_node(nodemask);
2040	for (i = 0; i < target; i++)
2041		nid = next_node(nid, nodemask);
2042	return nid;
2043}
2044
2045/*
2046 * Return a nodemask representing a mempolicy for filtering nodes for
2047 * page allocation, together with preferred node id (or the input node id).
2048 */
2049static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *pol,
2050				   pgoff_t ilx, int *nid)
2051{
2052	nodemask_t *nodemask = NULL;
 
2053
2054	switch (pol->mode) {
2055	case MPOL_PREFERRED:
2056		/* Override input node id */
2057		*nid = first_node(pol->nodes);
2058		break;
2059	case MPOL_PREFERRED_MANY:
2060		nodemask = &pol->nodes;
2061		if (pol->home_node != NUMA_NO_NODE)
2062			*nid = pol->home_node;
2063		break;
2064	case MPOL_BIND:
2065		/* Restrict to nodemask (but not on lower zones) */
2066		if (apply_policy_zone(pol, gfp_zone(gfp)) &&
2067		    cpuset_nodemask_valid_mems_allowed(&pol->nodes))
2068			nodemask = &pol->nodes;
2069		if (pol->home_node != NUMA_NO_NODE)
2070			*nid = pol->home_node;
2071		/*
2072		 * __GFP_THISNODE shouldn't even be used with the bind policy
2073		 * because we might easily break the expectation to stay on the
2074		 * requested node and not break the policy.
 
 
2075		 */
2076		WARN_ON_ONCE(gfp & __GFP_THISNODE);
2077		break;
2078	case MPOL_INTERLEAVE:
2079		/* Override input node id */
2080		*nid = (ilx == NO_INTERLEAVE_INDEX) ?
2081			interleave_nodes(pol) : interleave_nid(pol, ilx);
2082		break;
2083	case MPOL_WEIGHTED_INTERLEAVE:
2084		*nid = (ilx == NO_INTERLEAVE_INDEX) ?
2085			weighted_interleave_nodes(pol) :
2086			weighted_interleave_nid(pol, ilx);
2087		break;
2088	}
2089
2090	return nodemask;
2091}
2092
2093#ifdef CONFIG_HUGETLBFS
2094/*
2095 * huge_node(@vma, @addr, @gfp_flags, @mpol)
2096 * @vma: virtual memory area whose policy is sought
2097 * @addr: address in @vma for shared policy lookup and interleave policy
2098 * @gfp_flags: for requested zone
2099 * @mpol: pointer to mempolicy pointer for reference counted mempolicy
2100 * @nodemask: pointer to nodemask pointer for 'bind' and 'prefer-many' policy
2101 *
2102 * Returns a nid suitable for a huge page allocation and a pointer
2103 * to the struct mempolicy for conditional unref after allocation.
2104 * If the effective policy is 'bind' or 'prefer-many', returns a pointer
2105 * to the mempolicy's @nodemask for filtering the zonelist.
 
 
2106 */
2107int huge_node(struct vm_area_struct *vma, unsigned long addr, gfp_t gfp_flags,
2108		struct mempolicy **mpol, nodemask_t **nodemask)
2109{
2110	pgoff_t ilx;
2111	int nid;
2112
2113	nid = numa_node_id();
2114	*mpol = get_vma_policy(vma, addr, hstate_vma(vma)->order, &ilx);
2115	*nodemask = policy_nodemask(gfp_flags, *mpol, ilx, &nid);
 
 
 
 
 
 
 
 
2116	return nid;
2117}
2118
2119/*
2120 * init_nodemask_of_mempolicy
2121 *
2122 * If the current task's mempolicy is "default" [NULL], return 'false'
2123 * to indicate default policy.  Otherwise, extract the policy nodemask
2124 * for 'bind' or 'interleave' policy into the argument nodemask, or
2125 * initialize the argument nodemask to contain the single node for
2126 * 'preferred' or 'local' policy and return 'true' to indicate presence
2127 * of non-default mempolicy.
2128 *
2129 * We don't bother with reference counting the mempolicy [mpol_get/put]
2130 * because the current task is examining it's own mempolicy and a task's
2131 * mempolicy is only ever changed by the task itself.
2132 *
2133 * N.B., it is the caller's responsibility to free a returned nodemask.
2134 */
2135bool init_nodemask_of_mempolicy(nodemask_t *mask)
2136{
2137	struct mempolicy *mempolicy;
 
2138
2139	if (!(mask && current->mempolicy))
2140		return false;
2141
2142	task_lock(current);
2143	mempolicy = current->mempolicy;
2144	switch (mempolicy->mode) {
2145	case MPOL_PREFERRED:
2146	case MPOL_PREFERRED_MANY:
 
 
 
 
 
 
2147	case MPOL_BIND:
 
2148	case MPOL_INTERLEAVE:
2149	case MPOL_WEIGHTED_INTERLEAVE:
2150		*mask = mempolicy->nodes;
2151		break;
2152
2153	case MPOL_LOCAL:
2154		init_nodemask_of_node(mask, numa_node_id());
2155		break;
2156
2157	default:
2158		BUG();
2159	}
2160	task_unlock(current);
2161
2162	return true;
2163}
2164#endif
2165
2166/*
2167 * mempolicy_in_oom_domain
2168 *
2169 * If tsk's mempolicy is "bind", check for intersection between mask and
2170 * the policy nodemask. Otherwise, return true for all other policies
2171 * including "interleave", as a tsk with "interleave" policy may have
2172 * memory allocated from all nodes in system.
2173 *
2174 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
2175 */
2176bool mempolicy_in_oom_domain(struct task_struct *tsk,
2177					const nodemask_t *mask)
2178{
2179	struct mempolicy *mempolicy;
2180	bool ret = true;
2181
2182	if (!mask)
2183		return ret;
2184
2185	task_lock(tsk);
2186	mempolicy = tsk->mempolicy;
2187	if (mempolicy && mempolicy->mode == MPOL_BIND)
2188		ret = nodes_intersects(mempolicy->nodes, *mask);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2189	task_unlock(tsk);
2190
2191	return ret;
2192}
2193
2194static struct page *alloc_pages_preferred_many(gfp_t gfp, unsigned int order,
2195						int nid, nodemask_t *nodemask)
 
 
2196{
2197	struct page *page;
2198	gfp_t preferred_gfp;
2199
2200	/*
2201	 * This is a two pass approach. The first pass will only try the
2202	 * preferred nodes but skip the direct reclaim and allow the
2203	 * allocation to fail, while the second pass will try all the
2204	 * nodes in system.
2205	 */
2206	preferred_gfp = gfp | __GFP_NOWARN;
2207	preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2208	page = __alloc_pages_noprof(preferred_gfp, order, nid, nodemask);
2209	if (!page)
2210		page = __alloc_pages_noprof(gfp, order, nid, NULL);
2211
 
 
 
 
 
 
 
 
 
2212	return page;
2213}
2214
2215/**
2216 * alloc_pages_mpol - Allocate pages according to NUMA mempolicy.
2217 * @gfp: GFP flags.
2218 * @order: Order of the page allocation.
2219 * @pol: Pointer to the NUMA mempolicy.
2220 * @ilx: Index for interleave mempolicy (also distinguishes alloc_pages()).
2221 * @nid: Preferred node (usually numa_node_id() but @mpol may override it).
2222 *
2223 * Return: The page on success or NULL if allocation fails.
2224 */
2225struct page *alloc_pages_mpol_noprof(gfp_t gfp, unsigned int order,
2226		struct mempolicy *pol, pgoff_t ilx, int nid)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2227{
2228	nodemask_t *nodemask;
2229	struct page *page;
 
 
 
 
 
 
 
2230
2231	nodemask = policy_nodemask(gfp, pol, ilx, &nid);
 
 
 
 
2232
2233	if (pol->mode == MPOL_PREFERRED_MANY)
2234		return alloc_pages_preferred_many(gfp, order, nid, nodemask);
2235
2236	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
2237	    /* filter "hugepage" allocation, unless from alloc_pages() */
2238	    order == HPAGE_PMD_ORDER && ilx != NO_INTERLEAVE_INDEX) {
2239		/*
2240		 * For hugepage allocation and non-interleave policy which
2241		 * allows the current node (or other explicitly preferred
2242		 * node) we only try to allocate from the current/preferred
2243		 * node and don't fall back to other nodes, as the cost of
2244		 * remote accesses would likely offset THP benefits.
2245		 *
2246		 * If the policy is interleave or does not allow the current
2247		 * node in its nodemask, we allocate the standard way.
2248		 */
2249		if (pol->mode != MPOL_INTERLEAVE &&
2250		    pol->mode != MPOL_WEIGHTED_INTERLEAVE &&
2251		    (!nodemask || node_isset(nid, *nodemask))) {
2252			/*
2253			 * First, try to allocate THP only on local node, but
2254			 * don't reclaim unnecessarily, just compact.
2255			 */
2256			page = __alloc_pages_node_noprof(nid,
2257				gfp | __GFP_THISNODE | __GFP_NORETRY, order);
2258			if (page || !(gfp & __GFP_DIRECT_RECLAIM))
2259				return page;
2260			/*
2261			 * If hugepage allocations are configured to always
2262			 * synchronous compact or the vma has been madvised
2263			 * to prefer hugepage backing, retry allowing remote
2264			 * memory with both reclaim and compact as well.
2265			 */
2266		}
2267	}
2268
2269	page = __alloc_pages_noprof(gfp, order, nid, nodemask);
2270
2271	if (unlikely(pol->mode == MPOL_INTERLEAVE ||
2272		     pol->mode == MPOL_WEIGHTED_INTERLEAVE) && page) {
2273		/* skip NUMA_INTERLEAVE_HIT update if numa stats is disabled */
2274		if (static_branch_likely(&vm_numa_stat_key) &&
2275		    page_to_nid(page) == nid) {
2276			preempt_disable();
2277			__count_numa_event(page_zone(page), NUMA_INTERLEAVE_HIT);
2278			preempt_enable();
2279		}
2280	}
2281
 
 
 
 
 
2282	return page;
2283}
2284
2285struct folio *folio_alloc_mpol_noprof(gfp_t gfp, unsigned int order,
2286		struct mempolicy *pol, pgoff_t ilx, int nid)
2287{
2288	return page_rmappable_folio(alloc_pages_mpol_noprof(gfp | __GFP_COMP,
2289							order, pol, ilx, nid));
2290}
2291
2292/**
2293 * vma_alloc_folio - Allocate a folio for a VMA.
2294 * @gfp: GFP flags.
2295 * @order: Order of the folio.
2296 * @vma: Pointer to VMA.
2297 * @addr: Virtual address of the allocation.  Must be inside @vma.
2298 *
2299 * Allocate a folio for a specific address in @vma, using the appropriate
2300 * NUMA policy.  The caller must hold the mmap_lock of the mm_struct of the
2301 * VMA to prevent it from going away.  Should be used for all allocations
2302 * for folios that will be mapped into user space, excepting hugetlbfs, and
2303 * excepting where direct use of alloc_pages_mpol() is more appropriate.
2304 *
2305 * Return: The folio on success or NULL if allocation fails.
 
 
 
 
 
 
 
 
 
 
2306 */
2307struct folio *vma_alloc_folio_noprof(gfp_t gfp, int order, struct vm_area_struct *vma,
2308		unsigned long addr)
2309{
2310	struct mempolicy *pol;
2311	pgoff_t ilx;
2312	struct folio *folio;
2313
2314	if (vma->vm_flags & VM_DROPPABLE)
2315		gfp |= __GFP_NOWARN;
2316
2317	pol = get_vma_policy(vma, addr, order, &ilx);
2318	folio = folio_alloc_mpol_noprof(gfp, order, pol, ilx, numa_node_id());
2319	mpol_cond_put(pol);
2320	return folio;
2321}
2322EXPORT_SYMBOL(vma_alloc_folio_noprof);
2323
2324/**
2325 * alloc_pages - Allocate pages.
2326 * @gfp: GFP flags.
2327 * @order: Power of two of number of pages to allocate.
2328 *
2329 * Allocate 1 << @order contiguous pages.  The physical address of the
2330 * first page is naturally aligned (eg an order-3 allocation will be aligned
2331 * to a multiple of 8 * PAGE_SIZE bytes).  The NUMA policy of the current
2332 * process is honoured when in process context.
2333 *
2334 * Context: Can be called from any context, providing the appropriate GFP
2335 * flags are used.
2336 * Return: The page on success or NULL if allocation fails.
2337 */
2338struct page *alloc_pages_noprof(gfp_t gfp, unsigned int order)
2339{
2340	struct mempolicy *pol = &default_policy;
 
2341
2342	/*
2343	 * No reference counting needed for current->mempolicy
2344	 * nor system default_policy
2345	 */
2346	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2347		pol = get_task_policy(current);
2348
2349	return alloc_pages_mpol_noprof(gfp, order, pol, NO_INTERLEAVE_INDEX,
2350				       numa_node_id());
2351}
2352EXPORT_SYMBOL(alloc_pages_noprof);
2353
2354struct folio *folio_alloc_noprof(gfp_t gfp, unsigned int order)
2355{
2356	return page_rmappable_folio(alloc_pages_noprof(gfp | __GFP_COMP, order));
2357}
2358EXPORT_SYMBOL(folio_alloc_noprof);
2359
2360static unsigned long alloc_pages_bulk_array_interleave(gfp_t gfp,
2361		struct mempolicy *pol, unsigned long nr_pages,
2362		struct page **page_array)
2363{
2364	int nodes;
2365	unsigned long nr_pages_per_node;
2366	int delta;
2367	int i;
2368	unsigned long nr_allocated;
2369	unsigned long total_allocated = 0;
2370
2371	nodes = nodes_weight(pol->nodes);
2372	nr_pages_per_node = nr_pages / nodes;
2373	delta = nr_pages - nodes * nr_pages_per_node;
2374
2375	for (i = 0; i < nodes; i++) {
2376		if (delta) {
2377			nr_allocated = alloc_pages_bulk_noprof(gfp,
2378					interleave_nodes(pol), NULL,
2379					nr_pages_per_node + 1, NULL,
2380					page_array);
2381			delta--;
2382		} else {
2383			nr_allocated = alloc_pages_bulk_noprof(gfp,
2384					interleave_nodes(pol), NULL,
2385					nr_pages_per_node, NULL, page_array);
2386		}
2387
2388		page_array += nr_allocated;
2389		total_allocated += nr_allocated;
2390	}
2391
2392	return total_allocated;
2393}
2394
2395static unsigned long alloc_pages_bulk_array_weighted_interleave(gfp_t gfp,
2396		struct mempolicy *pol, unsigned long nr_pages,
2397		struct page **page_array)
2398{
2399	struct task_struct *me = current;
2400	unsigned int cpuset_mems_cookie;
2401	unsigned long total_allocated = 0;
2402	unsigned long nr_allocated = 0;
2403	unsigned long rounds;
2404	unsigned long node_pages, delta;
2405	u8 *table, *weights, weight;
2406	unsigned int weight_total = 0;
2407	unsigned long rem_pages = nr_pages;
2408	nodemask_t nodes;
2409	int nnodes, node;
2410	int resume_node = MAX_NUMNODES - 1;
2411	u8 resume_weight = 0;
2412	int prev_node;
2413	int i;
2414
2415	if (!nr_pages)
2416		return 0;
2417
2418	/* read the nodes onto the stack, retry if done during rebind */
2419	do {
2420		cpuset_mems_cookie = read_mems_allowed_begin();
2421		nnodes = read_once_policy_nodemask(pol, &nodes);
2422	} while (read_mems_allowed_retry(cpuset_mems_cookie));
2423
2424	/* if the nodemask has become invalid, we cannot do anything */
2425	if (!nnodes)
2426		return 0;
2427
2428	/* Continue allocating from most recent node and adjust the nr_pages */
2429	node = me->il_prev;
2430	weight = me->il_weight;
2431	if (weight && node_isset(node, nodes)) {
2432		node_pages = min(rem_pages, weight);
2433		nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
2434						  NULL, page_array);
2435		page_array += nr_allocated;
2436		total_allocated += nr_allocated;
2437		/* if that's all the pages, no need to interleave */
2438		if (rem_pages <= weight) {
2439			me->il_weight -= rem_pages;
2440			return total_allocated;
2441		}
2442		/* Otherwise we adjust remaining pages, continue from there */
2443		rem_pages -= weight;
2444	}
2445	/* clear active weight in case of an allocation failure */
2446	me->il_weight = 0;
2447	prev_node = node;
2448
2449	/* create a local copy of node weights to operate on outside rcu */
2450	weights = kzalloc(nr_node_ids, GFP_KERNEL);
2451	if (!weights)
2452		return total_allocated;
2453
2454	rcu_read_lock();
2455	table = rcu_dereference(iw_table);
2456	if (table)
2457		memcpy(weights, table, nr_node_ids);
2458	rcu_read_unlock();
2459
2460	/* calculate total, detect system default usage */
2461	for_each_node_mask(node, nodes) {
2462		if (!weights[node])
2463			weights[node] = 1;
2464		weight_total += weights[node];
2465	}
2466
2467	/*
2468	 * Calculate rounds/partial rounds to minimize __alloc_pages_bulk calls.
2469	 * Track which node weighted interleave should resume from.
2470	 *
2471	 * if (rounds > 0) and (delta == 0), resume_node will always be
2472	 * the node following prev_node and its weight.
2473	 */
2474	rounds = rem_pages / weight_total;
2475	delta = rem_pages % weight_total;
2476	resume_node = next_node_in(prev_node, nodes);
2477	resume_weight = weights[resume_node];
2478	for (i = 0; i < nnodes; i++) {
2479		node = next_node_in(prev_node, nodes);
2480		weight = weights[node];
2481		node_pages = weight * rounds;
2482		/* If a delta exists, add this node's portion of the delta */
2483		if (delta > weight) {
2484			node_pages += weight;
2485			delta -= weight;
2486		} else if (delta) {
2487			/* when delta is depleted, resume from that node */
2488			node_pages += delta;
2489			resume_node = node;
2490			resume_weight = weight - delta;
2491			delta = 0;
2492		}
2493		/* node_pages can be 0 if an allocation fails and rounds == 0 */
2494		if (!node_pages)
2495			break;
2496		nr_allocated = __alloc_pages_bulk(gfp, node, NULL, node_pages,
2497						  NULL, page_array);
2498		page_array += nr_allocated;
2499		total_allocated += nr_allocated;
2500		if (total_allocated == nr_pages)
2501			break;
2502		prev_node = node;
2503	}
2504	me->il_prev = resume_node;
2505	me->il_weight = resume_weight;
2506	kfree(weights);
2507	return total_allocated;
2508}
2509
2510static unsigned long alloc_pages_bulk_array_preferred_many(gfp_t gfp, int nid,
2511		struct mempolicy *pol, unsigned long nr_pages,
2512		struct page **page_array)
2513{
2514	gfp_t preferred_gfp;
2515	unsigned long nr_allocated = 0;
2516
2517	preferred_gfp = gfp | __GFP_NOWARN;
2518	preferred_gfp &= ~(__GFP_DIRECT_RECLAIM | __GFP_NOFAIL);
2519
2520	nr_allocated  = alloc_pages_bulk_noprof(preferred_gfp, nid, &pol->nodes,
2521					   nr_pages, NULL, page_array);
2522
2523	if (nr_allocated < nr_pages)
2524		nr_allocated += alloc_pages_bulk_noprof(gfp, numa_node_id(), NULL,
2525				nr_pages - nr_allocated, NULL,
2526				page_array + nr_allocated);
2527	return nr_allocated;
2528}
2529
2530/* alloc pages bulk and mempolicy should be considered at the
2531 * same time in some situation such as vmalloc.
2532 *
2533 * It can accelerate memory allocation especially interleaving
2534 * allocate memory.
2535 */
2536unsigned long alloc_pages_bulk_array_mempolicy_noprof(gfp_t gfp,
2537		unsigned long nr_pages, struct page **page_array)
2538{
2539	struct mempolicy *pol = &default_policy;
2540	nodemask_t *nodemask;
2541	int nid;
2542
2543	if (!in_interrupt() && !(gfp & __GFP_THISNODE))
2544		pol = get_task_policy(current);
2545
2546	if (pol->mode == MPOL_INTERLEAVE)
2547		return alloc_pages_bulk_array_interleave(gfp, pol,
2548							 nr_pages, page_array);
 
 
 
2549
2550	if (pol->mode == MPOL_WEIGHTED_INTERLEAVE)
2551		return alloc_pages_bulk_array_weighted_interleave(
2552				  gfp, pol, nr_pages, page_array);
2553
2554	if (pol->mode == MPOL_PREFERRED_MANY)
2555		return alloc_pages_bulk_array_preferred_many(gfp,
2556				numa_node_id(), pol, nr_pages, page_array);
2557
2558	nid = numa_node_id();
2559	nodemask = policy_nodemask(gfp, pol, NO_INTERLEAVE_INDEX, &nid);
2560	return alloc_pages_bulk_noprof(gfp, nid, nodemask,
2561				       nr_pages, NULL, page_array);
2562}
 
2563
2564int vma_dup_policy(struct vm_area_struct *src, struct vm_area_struct *dst)
2565{
2566	struct mempolicy *pol = mpol_dup(src->vm_policy);
2567
2568	if (IS_ERR(pol))
2569		return PTR_ERR(pol);
2570	dst->vm_policy = pol;
2571	return 0;
2572}
2573
2574/*
2575 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2576 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2577 * with the mems_allowed returned by cpuset_mems_allowed().  This
2578 * keeps mempolicies cpuset relative after its cpuset moves.  See
2579 * further kernel/cpuset.c update_nodemask().
2580 *
2581 * current's mempolicy may be rebinded by the other task(the task that changes
2582 * cpuset's mems), so we needn't do rebind work for current task.
2583 */
2584
2585/* Slow path of a mempolicy duplicate */
2586struct mempolicy *__mpol_dup(struct mempolicy *old)
2587{
2588	struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2589
2590	if (!new)
2591		return ERR_PTR(-ENOMEM);
2592
2593	/* task's mempolicy is protected by alloc_lock */
2594	if (old == current->mempolicy) {
2595		task_lock(current);
2596		*new = *old;
2597		task_unlock(current);
2598	} else
2599		*new = *old;
2600
2601	if (current_cpuset_is_being_rebound()) {
2602		nodemask_t mems = cpuset_mems_allowed(current);
2603		mpol_rebind_policy(new, &mems);
2604	}
2605	atomic_set(&new->refcnt, 1);
2606	return new;
2607}
2608
2609/* Slow path of a mempolicy comparison */
2610bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2611{
2612	if (!a || !b)
2613		return false;
2614	if (a->mode != b->mode)
2615		return false;
2616	if (a->flags != b->flags)
2617		return false;
2618	if (a->home_node != b->home_node)
2619		return false;
2620	if (mpol_store_user_nodemask(a))
2621		if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2622			return false;
2623
2624	switch (a->mode) {
2625	case MPOL_BIND:
 
2626	case MPOL_INTERLEAVE:
 
2627	case MPOL_PREFERRED:
2628	case MPOL_PREFERRED_MANY:
2629	case MPOL_WEIGHTED_INTERLEAVE:
2630		return !!nodes_equal(a->nodes, b->nodes);
2631	case MPOL_LOCAL:
2632		return true;
2633	default:
2634		BUG();
2635		return false;
2636	}
2637}
2638
2639/*
2640 * Shared memory backing store policy support.
2641 *
2642 * Remember policies even when nobody has shared memory mapped.
2643 * The policies are kept in Red-Black tree linked from the inode.
2644 * They are protected by the sp->lock rwlock, which should be held
2645 * for any accesses to the tree.
2646 */
2647
2648/*
2649 * lookup first element intersecting start-end.  Caller holds sp->lock for
2650 * reading or for writing
2651 */
2652static struct sp_node *sp_lookup(struct shared_policy *sp,
2653					pgoff_t start, pgoff_t end)
2654{
2655	struct rb_node *n = sp->root.rb_node;
2656
2657	while (n) {
2658		struct sp_node *p = rb_entry(n, struct sp_node, nd);
2659
2660		if (start >= p->end)
2661			n = n->rb_right;
2662		else if (end <= p->start)
2663			n = n->rb_left;
2664		else
2665			break;
2666	}
2667	if (!n)
2668		return NULL;
2669	for (;;) {
2670		struct sp_node *w = NULL;
2671		struct rb_node *prev = rb_prev(n);
2672		if (!prev)
2673			break;
2674		w = rb_entry(prev, struct sp_node, nd);
2675		if (w->end <= start)
2676			break;
2677		n = prev;
2678	}
2679	return rb_entry(n, struct sp_node, nd);
2680}
2681
2682/*
2683 * Insert a new shared policy into the list.  Caller holds sp->lock for
2684 * writing.
2685 */
2686static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2687{
2688	struct rb_node **p = &sp->root.rb_node;
2689	struct rb_node *parent = NULL;
2690	struct sp_node *nd;
2691
2692	while (*p) {
2693		parent = *p;
2694		nd = rb_entry(parent, struct sp_node, nd);
2695		if (new->start < nd->start)
2696			p = &(*p)->rb_left;
2697		else if (new->end > nd->end)
2698			p = &(*p)->rb_right;
2699		else
2700			BUG();
2701	}
2702	rb_link_node(&new->nd, parent, p);
2703	rb_insert_color(&new->nd, &sp->root);
 
 
2704}
2705
2706/* Find shared policy intersecting idx */
2707struct mempolicy *mpol_shared_policy_lookup(struct shared_policy *sp,
2708						pgoff_t idx)
2709{
2710	struct mempolicy *pol = NULL;
2711	struct sp_node *sn;
2712
2713	if (!sp->root.rb_node)
2714		return NULL;
2715	read_lock(&sp->lock);
2716	sn = sp_lookup(sp, idx, idx+1);
2717	if (sn) {
2718		mpol_get(sn->policy);
2719		pol = sn->policy;
2720	}
2721	read_unlock(&sp->lock);
2722	return pol;
2723}
2724
2725static void sp_free(struct sp_node *n)
2726{
2727	mpol_put(n->policy);
2728	kmem_cache_free(sn_cache, n);
2729}
2730
2731/**
2732 * mpol_misplaced - check whether current folio node is valid in policy
2733 *
2734 * @folio: folio to be checked
2735 * @vmf: structure describing the fault
2736 * @addr: virtual address in @vma for shared policy lookup and interleave policy
2737 *
2738 * Lookup current policy node id for vma,addr and "compare to" folio's
2739 * node id.  Policy determination "mimics" alloc_page_vma().
 
 
 
 
 
 
2740 * Called from fault path where we know the vma and faulting address.
2741 *
2742 * Return: NUMA_NO_NODE if the page is in a node that is valid for this
2743 * policy, or a suitable node ID to allocate a replacement folio from.
2744 */
2745int mpol_misplaced(struct folio *folio, struct vm_fault *vmf,
2746		   unsigned long addr)
2747{
2748	struct mempolicy *pol;
2749	pgoff_t ilx;
2750	struct zoneref *z;
2751	int curnid = folio_nid(folio);
2752	struct vm_area_struct *vma = vmf->vma;
2753	int thiscpu = raw_smp_processor_id();
2754	int thisnid = numa_node_id();
2755	int polnid = NUMA_NO_NODE;
2756	int ret = NUMA_NO_NODE;
2757
2758	/*
2759	 * Make sure ptl is held so that we don't preempt and we
2760	 * have a stable smp processor id
2761	 */
2762	lockdep_assert_held(vmf->ptl);
2763	pol = get_vma_policy(vma, addr, folio_order(folio), &ilx);
2764	if (!(pol->flags & MPOL_F_MOF))
2765		goto out;
2766
2767	switch (pol->mode) {
2768	case MPOL_INTERLEAVE:
2769		polnid = interleave_nid(pol, ilx);
2770		break;
2771
2772	case MPOL_WEIGHTED_INTERLEAVE:
2773		polnid = weighted_interleave_nid(pol, ilx);
2774		break;
2775
2776	case MPOL_PREFERRED:
2777		if (node_isset(curnid, pol->nodes))
2778			goto out;
2779		polnid = first_node(pol->nodes);
2780		break;
2781
2782	case MPOL_LOCAL:
2783		polnid = numa_node_id();
2784		break;
2785
2786	case MPOL_BIND:
2787	case MPOL_PREFERRED_MANY:
2788		/*
2789		 * Even though MPOL_PREFERRED_MANY can allocate pages outside
2790		 * policy nodemask we don't allow numa migration to nodes
2791		 * outside policy nodemask for now. This is done so that if we
2792		 * want demotion to slow memory to happen, before allocating
2793		 * from some DRAM node say 'x', we will end up using a
2794		 * MPOL_PREFERRED_MANY mask excluding node 'x'. In such scenario
2795		 * we should not promote to node 'x' from slow memory node.
2796		 */
2797		if (pol->flags & MPOL_F_MORON) {
2798			/*
2799			 * Optimize placement among multiple nodes
2800			 * via NUMA balancing
2801			 */
2802			if (node_isset(thisnid, pol->nodes))
2803				break;
2804			goto out;
2805		}
2806
2807		/*
 
2808		 * use current page if in policy nodemask,
2809		 * else select nearest allowed node, if any.
2810		 * If no allowed nodes, use current [!misplaced].
2811		 */
2812		if (node_isset(curnid, pol->nodes))
2813			goto out;
2814		z = first_zones_zonelist(
2815				node_zonelist(thisnid, GFP_HIGHUSER),
2816				gfp_zone(GFP_HIGHUSER),
2817				&pol->nodes);
2818		polnid = zonelist_node_idx(z);
2819		break;
2820
2821	default:
2822		BUG();
2823	}
2824
2825	/* Migrate the folio towards the node whose CPU is referencing it */
2826	if (pol->flags & MPOL_F_MORON) {
2827		polnid = thisnid;
2828
2829		if (!should_numa_migrate_memory(current, folio, curnid,
2830						thiscpu))
2831			goto out;
2832	}
2833
2834	if (curnid != polnid)
2835		ret = polnid;
2836out:
2837	mpol_cond_put(pol);
2838
2839	return ret;
2840}
2841
2842/*
2843 * Drop the (possibly final) reference to task->mempolicy.  It needs to be
2844 * dropped after task->mempolicy is set to NULL so that any allocation done as
2845 * part of its kmem_cache_free(), such as by KASAN, doesn't reference a freed
2846 * policy.
2847 */
2848void mpol_put_task_policy(struct task_struct *task)
2849{
2850	struct mempolicy *pol;
2851
2852	task_lock(task);
2853	pol = task->mempolicy;
2854	task->mempolicy = NULL;
2855	task_unlock(task);
2856	mpol_put(pol);
2857}
2858
2859static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2860{
 
2861	rb_erase(&n->nd, &sp->root);
2862	sp_free(n);
2863}
2864
2865static void sp_node_init(struct sp_node *node, unsigned long start,
2866			unsigned long end, struct mempolicy *pol)
2867{
2868	node->start = start;
2869	node->end = end;
2870	node->policy = pol;
2871}
2872
2873static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2874				struct mempolicy *pol)
2875{
2876	struct sp_node *n;
2877	struct mempolicy *newpol;
2878
2879	n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2880	if (!n)
2881		return NULL;
2882
2883	newpol = mpol_dup(pol);
2884	if (IS_ERR(newpol)) {
2885		kmem_cache_free(sn_cache, n);
2886		return NULL;
2887	}
2888	newpol->flags |= MPOL_F_SHARED;
2889	sp_node_init(n, start, end, newpol);
2890
2891	return n;
2892}
2893
2894/* Replace a policy range. */
2895static int shared_policy_replace(struct shared_policy *sp, pgoff_t start,
2896				 pgoff_t end, struct sp_node *new)
2897{
2898	struct sp_node *n;
2899	struct sp_node *n_new = NULL;
2900	struct mempolicy *mpol_new = NULL;
2901	int ret = 0;
2902
2903restart:
2904	write_lock(&sp->lock);
2905	n = sp_lookup(sp, start, end);
2906	/* Take care of old policies in the same range. */
2907	while (n && n->start < end) {
2908		struct rb_node *next = rb_next(&n->nd);
2909		if (n->start >= start) {
2910			if (n->end <= end)
2911				sp_delete(sp, n);
2912			else
2913				n->start = end;
2914		} else {
2915			/* Old policy spanning whole new range. */
2916			if (n->end > end) {
2917				if (!n_new)
2918					goto alloc_new;
2919
2920				*mpol_new = *n->policy;
2921				atomic_set(&mpol_new->refcnt, 1);
2922				sp_node_init(n_new, end, n->end, mpol_new);
2923				n->end = start;
2924				sp_insert(sp, n_new);
2925				n_new = NULL;
2926				mpol_new = NULL;
2927				break;
2928			} else
2929				n->end = start;
2930		}
2931		if (!next)
2932			break;
2933		n = rb_entry(next, struct sp_node, nd);
2934	}
2935	if (new)
2936		sp_insert(sp, new);
2937	write_unlock(&sp->lock);
2938	ret = 0;
2939
2940err_out:
2941	if (mpol_new)
2942		mpol_put(mpol_new);
2943	if (n_new)
2944		kmem_cache_free(sn_cache, n_new);
2945
2946	return ret;
2947
2948alloc_new:
2949	write_unlock(&sp->lock);
2950	ret = -ENOMEM;
2951	n_new = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2952	if (!n_new)
2953		goto err_out;
2954	mpol_new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2955	if (!mpol_new)
2956		goto err_out;
2957	atomic_set(&mpol_new->refcnt, 1);
2958	goto restart;
2959}
2960
2961/**
2962 * mpol_shared_policy_init - initialize shared policy for inode
2963 * @sp: pointer to inode shared policy
2964 * @mpol:  struct mempolicy to install
2965 *
2966 * Install non-NULL @mpol in inode's shared policy rb-tree.
2967 * On entry, the current task has a reference on a non-NULL @mpol.
2968 * This must be released on exit.
2969 * This is called at get_inode() calls and we can use GFP_KERNEL.
2970 */
2971void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2972{
2973	int ret;
2974
2975	sp->root = RB_ROOT;		/* empty tree == default mempolicy */
2976	rwlock_init(&sp->lock);
2977
2978	if (mpol) {
2979		struct sp_node *sn;
2980		struct mempolicy *npol;
2981		NODEMASK_SCRATCH(scratch);
2982
2983		if (!scratch)
2984			goto put_mpol;
2985
2986		/* contextualize the tmpfs mount point mempolicy to this file */
2987		npol = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2988		if (IS_ERR(npol))
2989			goto free_scratch; /* no valid nodemask intersection */
2990
2991		task_lock(current);
2992		ret = mpol_set_nodemask(npol, &mpol->w.user_nodemask, scratch);
2993		task_unlock(current);
2994		if (ret)
2995			goto put_npol;
2996
2997		/* alloc node covering entire file; adds ref to file's npol */
2998		sn = sp_alloc(0, MAX_LFS_FILESIZE >> PAGE_SHIFT, npol);
2999		if (sn)
3000			sp_insert(sp, sn);
3001put_npol:
3002		mpol_put(npol);	/* drop initial ref on file's npol */
 
3003free_scratch:
3004		NODEMASK_SCRATCH_FREE(scratch);
3005put_mpol:
3006		mpol_put(mpol);	/* drop our incoming ref on sb mpol */
3007	}
3008}
3009
3010int mpol_set_shared_policy(struct shared_policy *sp,
3011			struct vm_area_struct *vma, struct mempolicy *pol)
3012{
3013	int err;
3014	struct sp_node *new = NULL;
3015	unsigned long sz = vma_pages(vma);
3016
3017	if (pol) {
3018		new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, pol);
 
 
 
 
 
 
3019		if (!new)
3020			return -ENOMEM;
3021	}
3022	err = shared_policy_replace(sp, vma->vm_pgoff, vma->vm_pgoff + sz, new);
3023	if (err && new)
3024		sp_free(new);
3025	return err;
3026}
3027
3028/* Free a backing policy store on inode delete. */
3029void mpol_free_shared_policy(struct shared_policy *sp)
3030{
3031	struct sp_node *n;
3032	struct rb_node *next;
3033
3034	if (!sp->root.rb_node)
3035		return;
3036	write_lock(&sp->lock);
3037	next = rb_first(&sp->root);
3038	while (next) {
3039		n = rb_entry(next, struct sp_node, nd);
3040		next = rb_next(&n->nd);
3041		sp_delete(sp, n);
3042	}
3043	write_unlock(&sp->lock);
3044}
3045
3046#ifdef CONFIG_NUMA_BALANCING
3047static int __initdata numabalancing_override;
3048
3049static void __init check_numabalancing_enable(void)
3050{
3051	bool numabalancing_default = false;
3052
3053	if (IS_ENABLED(CONFIG_NUMA_BALANCING_DEFAULT_ENABLED))
3054		numabalancing_default = true;
3055
3056	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
3057	if (numabalancing_override)
3058		set_numabalancing_state(numabalancing_override == 1);
3059
3060	if (num_online_nodes() > 1 && !numabalancing_override) {
3061		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
3062			numabalancing_default ? "Enabling" : "Disabling");
3063		set_numabalancing_state(numabalancing_default);
3064	}
3065}
3066
3067static int __init setup_numabalancing(char *str)
3068{
3069	int ret = 0;
3070	if (!str)
3071		goto out;
3072
3073	if (!strcmp(str, "enable")) {
3074		numabalancing_override = 1;
3075		ret = 1;
3076	} else if (!strcmp(str, "disable")) {
3077		numabalancing_override = -1;
3078		ret = 1;
3079	}
3080out:
3081	if (!ret)
3082		pr_warn("Unable to parse numa_balancing=\n");
3083
3084	return ret;
3085}
3086__setup("numa_balancing=", setup_numabalancing);
3087#else
3088static inline void __init check_numabalancing_enable(void)
3089{
3090}
3091#endif /* CONFIG_NUMA_BALANCING */
3092
 
3093void __init numa_policy_init(void)
3094{
3095	nodemask_t interleave_nodes;
3096	unsigned long largest = 0;
3097	int nid, prefer = 0;
3098
3099	policy_cache = kmem_cache_create("numa_policy",
3100					 sizeof(struct mempolicy),
3101					 0, SLAB_PANIC, NULL);
3102
3103	sn_cache = kmem_cache_create("shared_policy_node",
3104				     sizeof(struct sp_node),
3105				     0, SLAB_PANIC, NULL);
3106
3107	for_each_node(nid) {
3108		preferred_node_policy[nid] = (struct mempolicy) {
3109			.refcnt = ATOMIC_INIT(1),
3110			.mode = MPOL_PREFERRED,
3111			.flags = MPOL_F_MOF | MPOL_F_MORON,
3112			.nodes = nodemask_of_node(nid),
3113		};
3114	}
3115
3116	/*
3117	 * Set interleaving policy for system init. Interleaving is only
3118	 * enabled across suitably sized nodes (default is >= 16MB), or
3119	 * fall back to the largest node if they're all smaller.
3120	 */
3121	nodes_clear(interleave_nodes);
3122	for_each_node_state(nid, N_MEMORY) {
3123		unsigned long total_pages = node_present_pages(nid);
3124
3125		/* Preserve the largest node */
3126		if (largest < total_pages) {
3127			largest = total_pages;
3128			prefer = nid;
3129		}
3130
3131		/* Interleave this node? */
3132		if ((total_pages << PAGE_SHIFT) >= (16 << 20))
3133			node_set(nid, interleave_nodes);
3134	}
3135
3136	/* All too small, use the largest */
3137	if (unlikely(nodes_empty(interleave_nodes)))
3138		node_set(prefer, interleave_nodes);
3139
3140	if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
3141		pr_err("%s: interleaving failed\n", __func__);
3142
3143	check_numabalancing_enable();
3144}
3145
3146/* Reset policy of current process to default */
3147void numa_default_policy(void)
3148{
3149	do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
3150}
3151
3152/*
3153 * Parse and format mempolicy from/to strings
3154 */
 
 
 
 
3155static const char * const policy_modes[] =
3156{
3157	[MPOL_DEFAULT]    = "default",
3158	[MPOL_PREFERRED]  = "prefer",
3159	[MPOL_BIND]       = "bind",
3160	[MPOL_INTERLEAVE] = "interleave",
3161	[MPOL_WEIGHTED_INTERLEAVE] = "weighted interleave",
3162	[MPOL_LOCAL]      = "local",
3163	[MPOL_PREFERRED_MANY]  = "prefer (many)",
3164};
3165
 
3166#ifdef CONFIG_TMPFS
3167/**
3168 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
3169 * @str:  string containing mempolicy to parse
3170 * @mpol:  pointer to struct mempolicy pointer, returned on success.
3171 *
3172 * Format of input:
3173 *	<mode>[=<flags>][:<nodelist>]
3174 *
3175 * Return: %0 on success, else %1
3176 */
3177int mpol_parse_str(char *str, struct mempolicy **mpol)
3178{
3179	struct mempolicy *new = NULL;
 
3180	unsigned short mode_flags;
3181	nodemask_t nodes;
3182	char *nodelist = strchr(str, ':');
3183	char *flags = strchr(str, '=');
3184	int err = 1, mode;
3185
3186	if (flags)
3187		*flags++ = '\0';	/* terminate mode string */
3188
3189	if (nodelist) {
3190		/* NUL-terminate mode or flags string */
3191		*nodelist++ = '\0';
3192		if (nodelist_parse(nodelist, nodes))
3193			goto out;
3194		if (!nodes_subset(nodes, node_states[N_MEMORY]))
3195			goto out;
3196	} else
3197		nodes_clear(nodes);
3198
3199	mode = match_string(policy_modes, MPOL_MAX, str);
3200	if (mode < 0)
 
 
 
 
 
 
 
3201		goto out;
3202
3203	switch (mode) {
3204	case MPOL_PREFERRED:
3205		/*
3206		 * Insist on a nodelist of one node only, although later
3207		 * we use first_node(nodes) to grab a single node, so here
3208		 * nodelist (or nodes) cannot be empty.
3209		 */
3210		if (nodelist) {
3211			char *rest = nodelist;
3212			while (isdigit(*rest))
3213				rest++;
3214			if (*rest)
3215				goto out;
3216			if (nodes_empty(nodes))
3217				goto out;
3218		}
3219		break;
3220	case MPOL_INTERLEAVE:
3221	case MPOL_WEIGHTED_INTERLEAVE:
3222		/*
3223		 * Default to online nodes with memory if no nodelist
3224		 */
3225		if (!nodelist)
3226			nodes = node_states[N_MEMORY];
3227		break;
3228	case MPOL_LOCAL:
3229		/*
3230		 * Don't allow a nodelist;  mpol_new() checks flags
3231		 */
3232		if (nodelist)
3233			goto out;
 
3234		break;
3235	case MPOL_DEFAULT:
3236		/*
3237		 * Insist on a empty nodelist
3238		 */
3239		if (!nodelist)
3240			err = 0;
3241		goto out;
3242	case MPOL_PREFERRED_MANY:
3243	case MPOL_BIND:
3244		/*
3245		 * Insist on a nodelist
3246		 */
3247		if (!nodelist)
3248			goto out;
3249	}
3250
3251	mode_flags = 0;
3252	if (flags) {
3253		/*
3254		 * Currently, we only support two mutually exclusive
3255		 * mode flags.
3256		 */
3257		if (!strcmp(flags, "static"))
3258			mode_flags |= MPOL_F_STATIC_NODES;
3259		else if (!strcmp(flags, "relative"))
3260			mode_flags |= MPOL_F_RELATIVE_NODES;
3261		else
3262			goto out;
3263	}
3264
3265	new = mpol_new(mode, mode_flags, &nodes);
3266	if (IS_ERR(new))
3267		goto out;
3268
3269	/*
3270	 * Save nodes for mpol_to_str() to show the tmpfs mount options
3271	 * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
3272	 */
3273	if (mode != MPOL_PREFERRED) {
3274		new->nodes = nodes;
3275	} else if (nodelist) {
3276		nodes_clear(new->nodes);
3277		node_set(first_node(nodes), new->nodes);
3278	} else {
3279		new->mode = MPOL_LOCAL;
3280	}
3281
3282	/*
3283	 * Save nodes for contextualization: this will be used to "clone"
3284	 * the mempolicy in a specific context [cpuset] at a later time.
3285	 */
3286	new->w.user_nodemask = nodes;
3287
3288	err = 0;
3289
3290out:
3291	/* Restore string for error message */
3292	if (nodelist)
3293		*--nodelist = ':';
3294	if (flags)
3295		*--flags = '=';
3296	if (!err)
3297		*mpol = new;
3298	return err;
3299}
3300#endif /* CONFIG_TMPFS */
3301
3302/**
3303 * mpol_to_str - format a mempolicy structure for printing
3304 * @buffer:  to contain formatted mempolicy string
3305 * @maxlen:  length of @buffer
3306 * @pol:  pointer to mempolicy to be formatted
3307 *
3308 * Convert @pol into a string.  If @buffer is too short, truncate the string.
3309 * Recommend a @maxlen of at least 51 for the longest mode, "weighted
3310 * interleave", plus the longest flag flags, "relative|balancing", and to
3311 * display at least a few node ids.
3312 */
3313void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol)
3314{
3315	char *p = buffer;
3316	nodemask_t nodes = NODE_MASK_NONE;
3317	unsigned short mode = MPOL_DEFAULT;
3318	unsigned short flags = 0;
3319
3320	if (pol &&
3321	    pol != &default_policy &&
3322	    !(pol >= &preferred_node_policy[0] &&
3323	      pol <= &preferred_node_policy[ARRAY_SIZE(preferred_node_policy) - 1])) {
3324		mode = pol->mode;
3325		flags = pol->flags;
3326	}
3327
3328	switch (mode) {
3329	case MPOL_DEFAULT:
3330	case MPOL_LOCAL:
3331		break;
3332	case MPOL_PREFERRED:
3333	case MPOL_PREFERRED_MANY:
 
 
 
 
3334	case MPOL_BIND:
3335	case MPOL_INTERLEAVE:
3336	case MPOL_WEIGHTED_INTERLEAVE:
3337		nodes = pol->nodes;
3338		break;
3339	default:
3340		WARN_ON_ONCE(1);
3341		snprintf(p, maxlen, "unknown");
3342		return;
3343	}
3344
3345	p += snprintf(p, maxlen, "%s", policy_modes[mode]);
3346
3347	if (flags & MPOL_MODE_FLAGS) {
3348		p += snprintf(p, buffer + maxlen - p, "=");
3349
3350		/*
3351		 * Static and relative are mutually exclusive.
3352		 */
3353		if (flags & MPOL_F_STATIC_NODES)
3354			p += snprintf(p, buffer + maxlen - p, "static");
3355		else if (flags & MPOL_F_RELATIVE_NODES)
3356			p += snprintf(p, buffer + maxlen - p, "relative");
3357
3358		if (flags & MPOL_F_NUMA_BALANCING) {
3359			if (!is_power_of_2(flags & MPOL_MODE_FLAGS))
3360				p += snprintf(p, buffer + maxlen - p, "|");
3361			p += snprintf(p, buffer + maxlen - p, "balancing");
3362		}
3363	}
3364
3365	if (!nodes_empty(nodes))
3366		p += scnprintf(p, buffer + maxlen - p, ":%*pbl",
3367			       nodemask_pr_args(&nodes));
3368}
3369
3370#ifdef CONFIG_SYSFS
3371struct iw_node_attr {
3372	struct kobj_attribute kobj_attr;
3373	int nid;
3374};
3375
3376static ssize_t node_show(struct kobject *kobj, struct kobj_attribute *attr,
3377			 char *buf)
3378{
3379	struct iw_node_attr *node_attr;
3380	u8 weight;
3381
3382	node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
3383	weight = get_il_weight(node_attr->nid);
3384	return sysfs_emit(buf, "%d\n", weight);
3385}
3386
3387static ssize_t node_store(struct kobject *kobj, struct kobj_attribute *attr,
3388			  const char *buf, size_t count)
3389{
3390	struct iw_node_attr *node_attr;
3391	u8 *new;
3392	u8 *old;
3393	u8 weight = 0;
3394
3395	node_attr = container_of(attr, struct iw_node_attr, kobj_attr);
3396	if (count == 0 || sysfs_streq(buf, ""))
3397		weight = 0;
3398	else if (kstrtou8(buf, 0, &weight))
3399		return -EINVAL;
3400
3401	new = kzalloc(nr_node_ids, GFP_KERNEL);
3402	if (!new)
3403		return -ENOMEM;
3404
3405	mutex_lock(&iw_table_lock);
3406	old = rcu_dereference_protected(iw_table,
3407					lockdep_is_held(&iw_table_lock));
3408	if (old)
3409		memcpy(new, old, nr_node_ids);
3410	new[node_attr->nid] = weight;
3411	rcu_assign_pointer(iw_table, new);
3412	mutex_unlock(&iw_table_lock);
3413	synchronize_rcu();
3414	kfree(old);
3415	return count;
3416}
3417
3418static struct iw_node_attr **node_attrs;
3419
3420static void sysfs_wi_node_release(struct iw_node_attr *node_attr,
3421				  struct kobject *parent)
3422{
3423	if (!node_attr)
3424		return;
3425	sysfs_remove_file(parent, &node_attr->kobj_attr.attr);
3426	kfree(node_attr->kobj_attr.attr.name);
3427	kfree(node_attr);
3428}
3429
3430static void sysfs_wi_release(struct kobject *wi_kobj)
3431{
3432	int i;
3433
3434	for (i = 0; i < nr_node_ids; i++)
3435		sysfs_wi_node_release(node_attrs[i], wi_kobj);
3436	kobject_put(wi_kobj);
3437}
3438
3439static const struct kobj_type wi_ktype = {
3440	.sysfs_ops = &kobj_sysfs_ops,
3441	.release = sysfs_wi_release,
3442};
3443
3444static int add_weight_node(int nid, struct kobject *wi_kobj)
3445{
3446	struct iw_node_attr *node_attr;
3447	char *name;
3448
3449	node_attr = kzalloc(sizeof(*node_attr), GFP_KERNEL);
3450	if (!node_attr)
3451		return -ENOMEM;
3452
3453	name = kasprintf(GFP_KERNEL, "node%d", nid);
3454	if (!name) {
3455		kfree(node_attr);
3456		return -ENOMEM;
3457	}
3458
3459	sysfs_attr_init(&node_attr->kobj_attr.attr);
3460	node_attr->kobj_attr.attr.name = name;
3461	node_attr->kobj_attr.attr.mode = 0644;
3462	node_attr->kobj_attr.show = node_show;
3463	node_attr->kobj_attr.store = node_store;
3464	node_attr->nid = nid;
3465
3466	if (sysfs_create_file(wi_kobj, &node_attr->kobj_attr.attr)) {
3467		kfree(node_attr->kobj_attr.attr.name);
3468		kfree(node_attr);
3469		pr_err("failed to add attribute to weighted_interleave\n");
3470		return -ENOMEM;
3471	}
3472
3473	node_attrs[nid] = node_attr;
3474	return 0;
3475}
3476
3477static int add_weighted_interleave_group(struct kobject *root_kobj)
3478{
3479	struct kobject *wi_kobj;
3480	int nid, err;
3481
3482	wi_kobj = kzalloc(sizeof(struct kobject), GFP_KERNEL);
3483	if (!wi_kobj)
3484		return -ENOMEM;
3485
3486	err = kobject_init_and_add(wi_kobj, &wi_ktype, root_kobj,
3487				   "weighted_interleave");
3488	if (err) {
3489		kfree(wi_kobj);
3490		return err;
3491	}
3492
3493	for_each_node_state(nid, N_POSSIBLE) {
3494		err = add_weight_node(nid, wi_kobj);
3495		if (err) {
3496			pr_err("failed to add sysfs [node%d]\n", nid);
3497			break;
3498		}
3499	}
3500	if (err)
3501		kobject_put(wi_kobj);
3502	return 0;
3503}
3504
3505static void mempolicy_kobj_release(struct kobject *kobj)
3506{
3507	u8 *old;
3508
3509	mutex_lock(&iw_table_lock);
3510	old = rcu_dereference_protected(iw_table,
3511					lockdep_is_held(&iw_table_lock));
3512	rcu_assign_pointer(iw_table, NULL);
3513	mutex_unlock(&iw_table_lock);
3514	synchronize_rcu();
3515	kfree(old);
3516	kfree(node_attrs);
3517	kfree(kobj);
3518}
3519
3520static const struct kobj_type mempolicy_ktype = {
3521	.release = mempolicy_kobj_release
3522};
3523
3524static int __init mempolicy_sysfs_init(void)
3525{
3526	int err;
3527	static struct kobject *mempolicy_kobj;
3528
3529	mempolicy_kobj = kzalloc(sizeof(*mempolicy_kobj), GFP_KERNEL);
3530	if (!mempolicy_kobj) {
3531		err = -ENOMEM;
3532		goto err_out;
3533	}
3534
3535	node_attrs = kcalloc(nr_node_ids, sizeof(struct iw_node_attr *),
3536			     GFP_KERNEL);
3537	if (!node_attrs) {
3538		err = -ENOMEM;
3539		goto mempol_out;
3540	}
3541
3542	err = kobject_init_and_add(mempolicy_kobj, &mempolicy_ktype, mm_kobj,
3543				   "mempolicy");
3544	if (err)
3545		goto node_out;
3546
3547	err = add_weighted_interleave_group(mempolicy_kobj);
3548	if (err) {
3549		pr_err("mempolicy sysfs structure failed to initialize\n");
3550		kobject_put(mempolicy_kobj);
3551		return err;
3552	}
3553
3554	return err;
3555node_out:
3556	kfree(node_attrs);
3557mempol_out:
3558	kfree(mempolicy_kobj);
3559err_out:
3560	pr_err("failed to add mempolicy kobject to the system\n");
3561	return err;
3562}
3563
3564late_initcall(mempolicy_sysfs_init);
3565#endif /* CONFIG_SYSFS */