linux/mm/mempolicy.c
<<
>>
Prefs
   1/*
   2 * Simple NUMA memory policy for the Linux kernel.
   3 *
   4 * Copyright 2003,2004 Andi Kleen, SuSE Labs.
   5 * (C) Copyright 2005 Christoph Lameter, Silicon Graphics, Inc.
   6 * Subject to the GNU Public License, version 2.
   7 *
   8 * NUMA policy allows the user to give hints in which node(s) memory should
   9 * be allocated.
  10 *
  11 * Support four policies per VMA and per process:
  12 *
  13 * The VMA policy has priority over the process policy for a page fault.
  14 *
  15 * interleave     Allocate memory interleaved over a set of nodes,
  16 *                with normal fallback if it fails.
  17 *                For VMA based allocations this interleaves based on the
  18 *                offset into the backing object or offset into the mapping
  19 *                for anonymous memory. For process policy an process counter
  20 *                is used.
  21 *
  22 * bind           Only allocate memory on a specific set of nodes,
  23 *                no fallback.
  24 *                FIXME: memory is allocated starting with the first node
  25 *                to the last. It would be better if bind would truly restrict
  26 *                the allocation to memory nodes instead
  27 *
  28 * preferred       Try a specific node first before normal fallback.
  29 *                As a special case node -1 here means do the allocation
  30 *                on the local CPU. This is normally identical to default,
  31 *                but useful to set in a VMA when you have a non default
  32 *                process policy.
  33 *
  34 * default        Allocate on the local node first, or when on a VMA
  35 *                use the process policy. This is what Linux always did
  36 *                in a NUMA aware kernel and still does by, ahem, default.
  37 *
  38 * The process policy is applied for most non interrupt memory allocations
  39 * in that process' context. Interrupts ignore the policies and always
  40 * try to allocate on the local CPU. The VMA policy is only applied for memory
  41 * allocations for a VMA in the VM.
  42 *
  43 * Currently there are a few corner cases in swapping where the policy
  44 * is not applied, but the majority should be handled. When process policy
  45 * is used it is not remembered over swap outs/swap ins.
  46 *
  47 * Only the highest zone in the zone hierarchy gets policied. Allocations
  48 * requesting a lower zone just use default policy. This implies that
  49 * on systems with highmem kernel lowmem allocation don't get policied.
  50 * Same with GFP_DMA allocations.
  51 *
  52 * For shmfs/tmpfs/hugetlbfs shared memory the policy is shared between
  53 * all users and remembered even when nobody has memory mapped.
  54 */
  55
  56/* Notebook:
  57   fix mmap readahead to honour policy and enable policy for any page cache
  58   object
  59   statistics for bigpages
  60   global policy for page cache? currently it uses process policy. Requires
  61   first item above.
  62   handle mremap for shared memory (currently ignored for the policy)
  63   grows down?
  64   make bind policy root only? It can trigger oom much faster and the
  65   kernel is not always grateful with that.
  66*/
  67
  68#include <linux/mempolicy.h>
  69#include <linux/mm.h>
  70#include <linux/highmem.h>
  71#include <linux/hugetlb.h>
  72#include <linux/kernel.h>
  73#include <linux/sched.h>
  74#include <linux/nodemask.h>
  75#include <linux/cpuset.h>
  76#include <linux/slab.h>
  77#include <linux/string.h>
  78#include <linux/export.h>
  79#include <linux/nsproxy.h>
  80#include <linux/interrupt.h>
  81#include <linux/init.h>
  82#include <linux/compat.h>
  83#include <linux/swap.h>
  84#include <linux/seq_file.h>
  85#include <linux/proc_fs.h>
  86#include <linux/migrate.h>
  87#include <linux/ksm.h>
  88#include <linux/rmap.h>
  89#include <linux/security.h>
  90#include <linux/syscalls.h>
  91#include <linux/ctype.h>
  92#include <linux/mm_inline.h>
  93
  94#include <asm/tlbflush.h>
  95#include <asm/uaccess.h>
  96#include <linux/random.h>
  97
  98#include "internal.h"
  99
 100/* Internal flags */
 101#define MPOL_MF_DISCONTIG_OK (MPOL_MF_INTERNAL << 0)    /* Skip checks for continuous vmas */
 102#define MPOL_MF_INVERT (MPOL_MF_INTERNAL << 1)          /* Invert check for nodemask */
 103
 104static struct kmem_cache *policy_cache;
 105static struct kmem_cache *sn_cache;
 106
 107/* Highest zone. An specific allocation for a zone below that is not
 108   policied. */
 109enum zone_type policy_zone = 0;
 110
 111/*
 112 * run-time system-wide default policy => local allocation
 113 */
 114static struct mempolicy default_policy = {
 115        .refcnt = ATOMIC_INIT(1), /* never free it */
 116        .mode = MPOL_PREFERRED,
 117        .flags = MPOL_F_LOCAL,
 118};
 119
 120static const struct mempolicy_operations {
 121        int (*create)(struct mempolicy *pol, const nodemask_t *nodes);
 122        /*
 123         * If read-side task has no lock to protect task->mempolicy, write-side
 124         * task will rebind the task->mempolicy by two step. The first step is
 125         * setting all the newly nodes, and the second step is cleaning all the
 126         * disallowed nodes. In this way, we can avoid finding no node to alloc
 127         * page.
 128         * If we have a lock to protect task->mempolicy in read-side, we do
 129         * rebind directly.
 130         *
 131         * step:
 132         *      MPOL_REBIND_ONCE - do rebind work at once
 133         *      MPOL_REBIND_STEP1 - set all the newly nodes
 134         *      MPOL_REBIND_STEP2 - clean all the disallowed nodes
 135         */
 136        void (*rebind)(struct mempolicy *pol, const nodemask_t *nodes,
 137                        enum mpol_rebind_step step);
 138} mpol_ops[MPOL_MAX];
 139
 140/* Check that the nodemask contains at least one populated zone */
 141static int is_valid_nodemask(const nodemask_t *nodemask)
 142{
 143        int nd, k;
 144
 145        for_each_node_mask(nd, *nodemask) {
 146                struct zone *z;
 147
 148                for (k = 0; k <= policy_zone; k++) {
 149                        z = &NODE_DATA(nd)->node_zones[k];
 150                        if (z->present_pages > 0)
 151                                return 1;
 152                }
 153        }
 154
 155        return 0;
 156}
 157
 158static inline int mpol_store_user_nodemask(const struct mempolicy *pol)
 159{
 160        return pol->flags & MPOL_MODE_FLAGS;
 161}
 162
 163static void mpol_relative_nodemask(nodemask_t *ret, const nodemask_t *orig,
 164                                   const nodemask_t *rel)
 165{
 166        nodemask_t tmp;
 167        nodes_fold(tmp, *orig, nodes_weight(*rel));
 168        nodes_onto(*ret, tmp, *rel);
 169}
 170
 171static int mpol_new_interleave(struct mempolicy *pol, const nodemask_t *nodes)
 172{
 173        if (nodes_empty(*nodes))
 174                return -EINVAL;
 175        pol->v.nodes = *nodes;
 176        return 0;
 177}
 178
 179static int mpol_new_preferred(struct mempolicy *pol, const nodemask_t *nodes)
 180{
 181        if (!nodes)
 182                pol->flags |= MPOL_F_LOCAL;     /* local allocation */
 183        else if (nodes_empty(*nodes))
 184                return -EINVAL;                 /*  no allowed nodes */
 185        else
 186                pol->v.preferred_node = first_node(*nodes);
 187        return 0;
 188}
 189
 190static int mpol_new_bind(struct mempolicy *pol, const nodemask_t *nodes)
 191{
 192        if (!is_valid_nodemask(nodes))
 193                return -EINVAL;
 194        pol->v.nodes = *nodes;
 195        return 0;
 196}
 197
 198/*
 199 * mpol_set_nodemask is called after mpol_new() to set up the nodemask, if
 200 * any, for the new policy.  mpol_new() has already validated the nodes
 201 * parameter with respect to the policy mode and flags.  But, we need to
 202 * handle an empty nodemask with MPOL_PREFERRED here.
 203 *
 204 * Must be called holding task's alloc_lock to protect task's mems_allowed
 205 * and mempolicy.  May also be called holding the mmap_semaphore for write.
 206 */
 207static int mpol_set_nodemask(struct mempolicy *pol,
 208                     const nodemask_t *nodes, struct nodemask_scratch *nsc)
 209{
 210        int ret;
 211
 212        /* if mode is MPOL_DEFAULT, pol is NULL. This is right. */
 213        if (pol == NULL)
 214                return 0;
 215        /* Check N_HIGH_MEMORY */
 216        nodes_and(nsc->mask1,
 217                  cpuset_current_mems_allowed, node_states[N_HIGH_MEMORY]);
 218
 219        VM_BUG_ON(!nodes);
 220        if (pol->mode == MPOL_PREFERRED && nodes_empty(*nodes))
 221                nodes = NULL;   /* explicit local allocation */
 222        else {
 223                if (pol->flags & MPOL_F_RELATIVE_NODES)
 224                        mpol_relative_nodemask(&nsc->mask2, nodes,&nsc->mask1);
 225                else
 226                        nodes_and(nsc->mask2, *nodes, nsc->mask1);
 227
 228                if (mpol_store_user_nodemask(pol))
 229                        pol->w.user_nodemask = *nodes;
 230                else
 231                        pol->w.cpuset_mems_allowed =
 232                                                cpuset_current_mems_allowed;
 233        }
 234
 235        if (nodes)
 236                ret = mpol_ops[pol->mode].create(pol, &nsc->mask2);
 237        else
 238                ret = mpol_ops[pol->mode].create(pol, NULL);
 239        return ret;
 240}
 241
 242/*
 243 * This function just creates a new policy, does some check and simple
 244 * initialization. You must invoke mpol_set_nodemask() to set nodes.
 245 */
 246static struct mempolicy *mpol_new(unsigned short mode, unsigned short flags,
 247                                  nodemask_t *nodes)
 248{
 249        struct mempolicy *policy;
 250
 251        pr_debug("setting mode %d flags %d nodes[0] %lx\n",
 252                 mode, flags, nodes ? nodes_addr(*nodes)[0] : -1);
 253
 254        if (mode == MPOL_DEFAULT) {
 255                if (nodes && !nodes_empty(*nodes))
 256                        return ERR_PTR(-EINVAL);
 257                return NULL;    /* simply delete any existing policy */
 258        }
 259        VM_BUG_ON(!nodes);
 260
 261        /*
 262         * MPOL_PREFERRED cannot be used with MPOL_F_STATIC_NODES or
 263         * MPOL_F_RELATIVE_NODES if the nodemask is empty (local allocation).
 264         * All other modes require a valid pointer to a non-empty nodemask.
 265         */
 266        if (mode == MPOL_PREFERRED) {
 267                if (nodes_empty(*nodes)) {
 268                        if (((flags & MPOL_F_STATIC_NODES) ||
 269                             (flags & MPOL_F_RELATIVE_NODES)))
 270                                return ERR_PTR(-EINVAL);
 271                }
 272        } else if (nodes_empty(*nodes))
 273                return ERR_PTR(-EINVAL);
 274        policy = kmem_cache_alloc(policy_cache, GFP_KERNEL);
 275        if (!policy)
 276                return ERR_PTR(-ENOMEM);
 277        atomic_set(&policy->refcnt, 1);
 278        policy->mode = mode;
 279        policy->flags = flags;
 280
 281        return policy;
 282}
 283
 284/* Slow path of a mpol destructor. */
 285void __mpol_put(struct mempolicy *p)
 286{
 287        if (!atomic_dec_and_test(&p->refcnt))
 288                return;
 289        kmem_cache_free(policy_cache, p);
 290}
 291
 292static void mpol_rebind_default(struct mempolicy *pol, const nodemask_t *nodes,
 293                                enum mpol_rebind_step step)
 294{
 295}
 296
 297/*
 298 * step:
 299 *      MPOL_REBIND_ONCE  - do rebind work at once
 300 *      MPOL_REBIND_STEP1 - set all the newly nodes
 301 *      MPOL_REBIND_STEP2 - clean all the disallowed nodes
 302 */
 303static void mpol_rebind_nodemask(struct mempolicy *pol, const nodemask_t *nodes,
 304                                 enum mpol_rebind_step step)
 305{
 306        nodemask_t tmp;
 307
 308        if (pol->flags & MPOL_F_STATIC_NODES)
 309                nodes_and(tmp, pol->w.user_nodemask, *nodes);
 310        else if (pol->flags & MPOL_F_RELATIVE_NODES)
 311                mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 312        else {
 313                /*
 314                 * if step == 1, we use ->w.cpuset_mems_allowed to cache the
 315                 * result
 316                 */
 317                if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP1) {
 318                        nodes_remap(tmp, pol->v.nodes,
 319                                        pol->w.cpuset_mems_allowed, *nodes);
 320                        pol->w.cpuset_mems_allowed = step ? tmp : *nodes;
 321                } else if (step == MPOL_REBIND_STEP2) {
 322                        tmp = pol->w.cpuset_mems_allowed;
 323                        pol->w.cpuset_mems_allowed = *nodes;
 324                } else
 325                        BUG();
 326        }
 327
 328        if (nodes_empty(tmp))
 329                tmp = *nodes;
 330
 331        if (step == MPOL_REBIND_STEP1)
 332                nodes_or(pol->v.nodes, pol->v.nodes, tmp);
 333        else if (step == MPOL_REBIND_ONCE || step == MPOL_REBIND_STEP2)
 334                pol->v.nodes = tmp;
 335        else
 336                BUG();
 337
 338        if (!node_isset(current->il_next, tmp)) {
 339                current->il_next = next_node(current->il_next, tmp);
 340                if (current->il_next >= MAX_NUMNODES)
 341                        current->il_next = first_node(tmp);
 342                if (current->il_next >= MAX_NUMNODES)
 343                        current->il_next = numa_node_id();
 344        }
 345}
 346
 347static void mpol_rebind_preferred(struct mempolicy *pol,
 348                                  const nodemask_t *nodes,
 349                                  enum mpol_rebind_step step)
 350{
 351        nodemask_t tmp;
 352
 353        if (pol->flags & MPOL_F_STATIC_NODES) {
 354                int node = first_node(pol->w.user_nodemask);
 355
 356                if (node_isset(node, *nodes)) {
 357                        pol->v.preferred_node = node;
 358                        pol->flags &= ~MPOL_F_LOCAL;
 359                } else
 360                        pol->flags |= MPOL_F_LOCAL;
 361        } else if (pol->flags & MPOL_F_RELATIVE_NODES) {
 362                mpol_relative_nodemask(&tmp, &pol->w.user_nodemask, nodes);
 363                pol->v.preferred_node = first_node(tmp);
 364        } else if (!(pol->flags & MPOL_F_LOCAL)) {
 365                pol->v.preferred_node = node_remap(pol->v.preferred_node,
 366                                                   pol->w.cpuset_mems_allowed,
 367                                                   *nodes);
 368                pol->w.cpuset_mems_allowed = *nodes;
 369        }
 370}
 371
 372/*
 373 * mpol_rebind_policy - Migrate a policy to a different set of nodes
 374 *
 375 * If read-side task has no lock to protect task->mempolicy, write-side
 376 * task will rebind the task->mempolicy by two step. The first step is
 377 * setting all the newly nodes, and the second step is cleaning all the
 378 * disallowed nodes. In this way, we can avoid finding no node to alloc
 379 * page.
 380 * If we have a lock to protect task->mempolicy in read-side, we do
 381 * rebind directly.
 382 *
 383 * step:
 384 *      MPOL_REBIND_ONCE  - do rebind work at once
 385 *      MPOL_REBIND_STEP1 - set all the newly nodes
 386 *      MPOL_REBIND_STEP2 - clean all the disallowed nodes
 387 */
 388static void mpol_rebind_policy(struct mempolicy *pol, const nodemask_t *newmask,
 389                                enum mpol_rebind_step step)
 390{
 391        if (!pol)
 392                return;
 393        if (!mpol_store_user_nodemask(pol) && step == MPOL_REBIND_ONCE &&
 394            nodes_equal(pol->w.cpuset_mems_allowed, *newmask))
 395                return;
 396
 397        if (step == MPOL_REBIND_STEP1 && (pol->flags & MPOL_F_REBINDING))
 398                return;
 399
 400        if (step == MPOL_REBIND_STEP2 && !(pol->flags & MPOL_F_REBINDING))
 401                BUG();
 402
 403        if (step == MPOL_REBIND_STEP1)
 404                pol->flags |= MPOL_F_REBINDING;
 405        else if (step == MPOL_REBIND_STEP2)
 406                pol->flags &= ~MPOL_F_REBINDING;
 407        else if (step >= MPOL_REBIND_NSTEP)
 408                BUG();
 409
 410        mpol_ops[pol->mode].rebind(pol, newmask, step);
 411}
 412
 413/*
 414 * Wrapper for mpol_rebind_policy() that just requires task
 415 * pointer, and updates task mempolicy.
 416 *
 417 * Called with task's alloc_lock held.
 418 */
 419
 420void mpol_rebind_task(struct task_struct *tsk, const nodemask_t *new,
 421                        enum mpol_rebind_step step)
 422{
 423        mpol_rebind_policy(tsk->mempolicy, new, step);
 424}
 425
 426/*
 427 * Rebind each vma in mm to new nodemask.
 428 *
 429 * Call holding a reference to mm.  Takes mm->mmap_sem during call.
 430 */
 431
 432void mpol_rebind_mm(struct mm_struct *mm, nodemask_t *new)
 433{
 434        struct vm_area_struct *vma;
 435
 436        down_write(&mm->mmap_sem);
 437        for (vma = mm->mmap; vma; vma = vma->vm_next)
 438                mpol_rebind_policy(vma->vm_policy, new, MPOL_REBIND_ONCE);
 439        up_write(&mm->mmap_sem);
 440}
 441
 442static const struct mempolicy_operations mpol_ops[MPOL_MAX] = {
 443        [MPOL_DEFAULT] = {
 444                .rebind = mpol_rebind_default,
 445        },
 446        [MPOL_INTERLEAVE] = {
 447                .create = mpol_new_interleave,
 448                .rebind = mpol_rebind_nodemask,
 449        },
 450        [MPOL_PREFERRED] = {
 451                .create = mpol_new_preferred,
 452                .rebind = mpol_rebind_preferred,
 453        },
 454        [MPOL_BIND] = {
 455                .create = mpol_new_bind,
 456                .rebind = mpol_rebind_nodemask,
 457        },
 458};
 459
 460static void migrate_page_add(struct page *page, struct list_head *pagelist,
 461                                unsigned long flags);
 462
 463/* Scan through pages checking if pages follow certain conditions. */
 464static int check_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 465                unsigned long addr, unsigned long end,
 466                const nodemask_t *nodes, unsigned long flags,
 467                void *private)
 468{
 469        pte_t *orig_pte;
 470        pte_t *pte;
 471        spinlock_t *ptl;
 472
 473        orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 474        do {
 475                struct page *page;
 476                int nid;
 477
 478                if (!pte_present(*pte))
 479                        continue;
 480                page = vm_normal_page(vma, addr, *pte);
 481                if (!page)
 482                        continue;
 483                /*
 484                 * vm_normal_page() filters out zero pages, but there might
 485                 * still be PageReserved pages to skip, perhaps in a VDSO.
 486                 * And we cannot move PageKsm pages sensibly or safely yet.
 487                 */
 488                if (PageReserved(page) || PageKsm(page))
 489                        continue;
 490                nid = page_to_nid(page);
 491                if (node_isset(nid, *nodes) == !!(flags & MPOL_MF_INVERT))
 492                        continue;
 493
 494                if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
 495                        migrate_page_add(page, private, flags);
 496                else
 497                        break;
 498        } while (pte++, addr += PAGE_SIZE, addr != end);
 499        pte_unmap_unlock(orig_pte, ptl);
 500        return addr != end;
 501}
 502
 503static inline int check_pmd_range(struct vm_area_struct *vma, pud_t *pud,
 504                unsigned long addr, unsigned long end,
 505                const nodemask_t *nodes, unsigned long flags,
 506                void *private)
 507{
 508        pmd_t *pmd;
 509        unsigned long next;
 510
 511        pmd = pmd_offset(pud, addr);
 512        do {
 513                next = pmd_addr_end(addr, end);
 514                split_huge_page_pmd(vma->vm_mm, pmd);
 515                if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 516                        continue;
 517                if (check_pte_range(vma, pmd, addr, next, nodes,
 518                                    flags, private))
 519                        return -EIO;
 520        } while (pmd++, addr = next, addr != end);
 521        return 0;
 522}
 523
 524static inline int check_pud_range(struct vm_area_struct *vma, pgd_t *pgd,
 525                unsigned long addr, unsigned long end,
 526                const nodemask_t *nodes, unsigned long flags,
 527                void *private)
 528{
 529        pud_t *pud;
 530        unsigned long next;
 531
 532        pud = pud_offset(pgd, addr);
 533        do {
 534                next = pud_addr_end(addr, end);
 535                if (pud_none_or_clear_bad(pud))
 536                        continue;
 537                if (check_pmd_range(vma, pud, addr, next, nodes,
 538                                    flags, private))
 539                        return -EIO;
 540        } while (pud++, addr = next, addr != end);
 541        return 0;
 542}
 543
 544static inline int check_pgd_range(struct vm_area_struct *vma,
 545                unsigned long addr, unsigned long end,
 546                const nodemask_t *nodes, unsigned long flags,
 547                void *private)
 548{
 549        pgd_t *pgd;
 550        unsigned long next;
 551
 552        pgd = pgd_offset(vma->vm_mm, addr);
 553        do {
 554                next = pgd_addr_end(addr, end);
 555                if (pgd_none_or_clear_bad(pgd))
 556                        continue;
 557                if (check_pud_range(vma, pgd, addr, next, nodes,
 558                                    flags, private))
 559                        return -EIO;
 560        } while (pgd++, addr = next, addr != end);
 561        return 0;
 562}
 563
 564/*
 565 * Check if all pages in a range are on a set of nodes.
 566 * If pagelist != NULL then isolate pages from the LRU and
 567 * put them on the pagelist.
 568 */
 569static struct vm_area_struct *
 570check_range(struct mm_struct *mm, unsigned long start, unsigned long end,
 571                const nodemask_t *nodes, unsigned long flags, void *private)
 572{
 573        int err;
 574        struct vm_area_struct *first, *vma, *prev;
 575
 576
 577        first = find_vma(mm, start);
 578        if (!first)
 579                return ERR_PTR(-EFAULT);
 580        prev = NULL;
 581        for (vma = first; vma && vma->vm_start < end; vma = vma->vm_next) {
 582                if (!(flags & MPOL_MF_DISCONTIG_OK)) {
 583                        if (!vma->vm_next && vma->vm_end < end)
 584                                return ERR_PTR(-EFAULT);
 585                        if (prev && prev->vm_end < vma->vm_start)
 586                                return ERR_PTR(-EFAULT);
 587                }
 588                if (!is_vm_hugetlb_page(vma) &&
 589                    ((flags & MPOL_MF_STRICT) ||
 590                     ((flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) &&
 591                                vma_migratable(vma)))) {
 592                        unsigned long endvma = vma->vm_end;
 593
 594                        if (endvma > end)
 595                                endvma = end;
 596                        if (vma->vm_start > start)
 597                                start = vma->vm_start;
 598                        err = check_pgd_range(vma, start, endvma, nodes,
 599                                                flags, private);
 600                        if (err) {
 601                                first = ERR_PTR(err);
 602                                break;
 603                        }
 604                }
 605                prev = vma;
 606        }
 607        return first;
 608}
 609
 610/* Step 2: apply policy to a range and do splits. */
 611static int mbind_range(struct mm_struct *mm, unsigned long start,
 612                       unsigned long end, struct mempolicy *new_pol)
 613{
 614        struct vm_area_struct *next;
 615        struct vm_area_struct *prev;
 616        struct vm_area_struct *vma;
 617        int err = 0;
 618        pgoff_t pgoff;
 619        unsigned long vmstart;
 620        unsigned long vmend;
 621
 622        vma = find_vma(mm, start);
 623        if (!vma || vma->vm_start > start)
 624                return -EFAULT;
 625
 626        prev = vma->vm_prev;
 627        if (start > vma->vm_start)
 628                prev = vma;
 629
 630        for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 631                next = vma->vm_next;
 632                vmstart = max(start, vma->vm_start);
 633                vmend   = min(end, vma->vm_end);
 634
 635                if (mpol_equal(vma_policy(vma), new_pol))
 636                        continue;
 637
 638                pgoff = vma->vm_pgoff +
 639                        ((vmstart - vma->vm_start) >> PAGE_SHIFT);
 640                prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 641                                  vma->anon_vma, vma->vm_file, pgoff,
 642                                  new_pol);
 643                if (prev) {
 644                        vma = prev;
 645                        next = vma->vm_next;
 646                        continue;
 647                }
 648                if (vma->vm_start != vmstart) {
 649                        err = split_vma(vma->vm_mm, vma, vmstart, 1);
 650                        if (err)
 651                                goto out;
 652                }
 653                if (vma->vm_end != vmend) {
 654                        err = split_vma(vma->vm_mm, vma, vmend, 0);
 655                        if (err)
 656                                goto out;
 657                }
 658
 659                /*
 660                 * Apply policy to a single VMA. The reference counting of
 661                 * policy for vma_policy linkages has already been handled by
 662                 * vma_merge and split_vma as necessary. If this is a shared
 663                 * policy then ->set_policy will increment the reference count
 664                 * for an sp node.
 665                 */
 666                pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 667                        vma->vm_start, vma->vm_end, vma->vm_pgoff,
 668                        vma->vm_ops, vma->vm_file,
 669                        vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 670                if (vma->vm_ops && vma->vm_ops->set_policy) {
 671                        err = vma->vm_ops->set_policy(vma, new_pol);
 672                        if (err)
 673                                goto out;
 674                }
 675        }
 676
 677 out:
 678        return err;
 679}
 680
 681/*
 682 * Update task->flags PF_MEMPOLICY bit: set iff non-default
 683 * mempolicy.  Allows more rapid checking of this (combined perhaps
 684 * with other PF_* flag bits) on memory allocation hot code paths.
 685 *
 686 * If called from outside this file, the task 'p' should -only- be
 687 * a newly forked child not yet visible on the task list, because
 688 * manipulating the task flags of a visible task is not safe.
 689 *
 690 * The above limitation is why this routine has the funny name
 691 * mpol_fix_fork_child_flag().
 692 *
 693 * It is also safe to call this with a task pointer of current,
 694 * which the static wrapper mpol_set_task_struct_flag() does,
 695 * for use within this file.
 696 */
 697
 698void mpol_fix_fork_child_flag(struct task_struct *p)
 699{
 700        if (p->mempolicy)
 701                p->flags |= PF_MEMPOLICY;
 702        else
 703                p->flags &= ~PF_MEMPOLICY;
 704}
 705
 706static void mpol_set_task_struct_flag(void)
 707{
 708        mpol_fix_fork_child_flag(current);
 709}
 710
 711/* Set the process memory policy */
 712static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 713                             nodemask_t *nodes)
 714{
 715        struct mempolicy *new, *old;
 716        struct mm_struct *mm = current->mm;
 717        NODEMASK_SCRATCH(scratch);
 718        int ret;
 719
 720        if (!scratch)
 721                return -ENOMEM;
 722
 723        new = mpol_new(mode, flags, nodes);
 724        if (IS_ERR(new)) {
 725                ret = PTR_ERR(new);
 726                goto out;
 727        }
 728        /*
 729         * prevent changing our mempolicy while show_numa_maps()
 730         * is using it.
 731         * Note:  do_set_mempolicy() can be called at init time
 732         * with no 'mm'.
 733         */
 734        if (mm)
 735                down_write(&mm->mmap_sem);
 736        task_lock(current);
 737        ret = mpol_set_nodemask(new, nodes, scratch);
 738        if (ret) {
 739                task_unlock(current);
 740                if (mm)
 741                        up_write(&mm->mmap_sem);
 742                mpol_put(new);
 743                goto out;
 744        }
 745        old = current->mempolicy;
 746        current->mempolicy = new;
 747        mpol_set_task_struct_flag();
 748        if (new && new->mode == MPOL_INTERLEAVE &&
 749            nodes_weight(new->v.nodes))
 750                current->il_next = first_node(new->v.nodes);
 751        task_unlock(current);
 752        if (mm)
 753                up_write(&mm->mmap_sem);
 754
 755        mpol_put(old);
 756        ret = 0;
 757out:
 758        NODEMASK_SCRATCH_FREE(scratch);
 759        return ret;
 760}
 761
 762/*
 763 * Return nodemask for policy for get_mempolicy() query
 764 *
 765 * Called with task's alloc_lock held
 766 */
 767static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 768{
 769        nodes_clear(*nodes);
 770        if (p == &default_policy)
 771                return;
 772
 773        switch (p->mode) {
 774        case MPOL_BIND:
 775                /* Fall through */
 776        case MPOL_INTERLEAVE:
 777                *nodes = p->v.nodes;
 778                break;
 779        case MPOL_PREFERRED:
 780                if (!(p->flags & MPOL_F_LOCAL))
 781                        node_set(p->v.preferred_node, *nodes);
 782                /* else return empty node mask for local allocation */
 783                break;
 784        default:
 785                BUG();
 786        }
 787}
 788
 789static int lookup_node(struct mm_struct *mm, unsigned long addr)
 790{
 791        struct page *p;
 792        int err;
 793
 794        err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
 795        if (err >= 0) {
 796                err = page_to_nid(p);
 797                put_page(p);
 798        }
 799        return err;
 800}
 801
 802/* Retrieve NUMA policy */
 803static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 804                             unsigned long addr, unsigned long flags)
 805{
 806        int err;
 807        struct mm_struct *mm = current->mm;
 808        struct vm_area_struct *vma = NULL;
 809        struct mempolicy *pol = current->mempolicy;
 810
 811        if (flags &
 812                ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 813                return -EINVAL;
 814
 815        if (flags & MPOL_F_MEMS_ALLOWED) {
 816                if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 817                        return -EINVAL;
 818                *policy = 0;    /* just so it's initialized */
 819                task_lock(current);
 820                *nmask  = cpuset_current_mems_allowed;
 821                task_unlock(current);
 822                return 0;
 823        }
 824
 825        if (flags & MPOL_F_ADDR) {
 826                /*
 827                 * Do NOT fall back to task policy if the
 828                 * vma/shared policy at addr is NULL.  We
 829                 * want to return MPOL_DEFAULT in this case.
 830                 */
 831                down_read(&mm->mmap_sem);
 832                vma = find_vma_intersection(mm, addr, addr+1);
 833                if (!vma) {
 834                        up_read(&mm->mmap_sem);
 835                        return -EFAULT;
 836                }
 837                if (vma->vm_ops && vma->vm_ops->get_policy)
 838                        pol = vma->vm_ops->get_policy(vma, addr);
 839                else
 840                        pol = vma->vm_policy;
 841        } else if (addr)
 842                return -EINVAL;
 843
 844        if (!pol)
 845                pol = &default_policy;  /* indicates default behavior */
 846
 847        if (flags & MPOL_F_NODE) {
 848                if (flags & MPOL_F_ADDR) {
 849                        err = lookup_node(mm, addr);
 850                        if (err < 0)
 851                                goto out;
 852                        *policy = err;
 853                } else if (pol == current->mempolicy &&
 854                                pol->mode == MPOL_INTERLEAVE) {
 855                        *policy = current->il_next;
 856                } else {
 857                        err = -EINVAL;
 858                        goto out;
 859                }
 860        } else {
 861                *policy = pol == &default_policy ? MPOL_DEFAULT :
 862                                                pol->mode;
 863                /*
 864                 * Internal mempolicy flags must be masked off before exposing
 865                 * the policy to userspace.
 866                 */
 867                *policy |= (pol->flags & MPOL_MODE_FLAGS);
 868        }
 869
 870        if (vma) {
 871                up_read(&current->mm->mmap_sem);
 872                vma = NULL;
 873        }
 874
 875        err = 0;
 876        if (nmask) {
 877                if (mpol_store_user_nodemask(pol)) {
 878                        *nmask = pol->w.user_nodemask;
 879                } else {
 880                        task_lock(current);
 881                        get_policy_nodemask(pol, nmask);
 882                        task_unlock(current);
 883                }
 884        }
 885
 886 out:
 887        mpol_cond_put(pol);
 888        if (vma)
 889                up_read(&current->mm->mmap_sem);
 890        return err;
 891}
 892
 893#ifdef CONFIG_MIGRATION
 894/*
 895 * page migration
 896 */
 897static void migrate_page_add(struct page *page, struct list_head *pagelist,
 898                                unsigned long flags)
 899{
 900        /*
 901         * Avoid migrating a page that is shared with others.
 902         */
 903        if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
 904                if (!isolate_lru_page(page)) {
 905                        list_add_tail(&page->lru, pagelist);
 906                        inc_zone_page_state(page, NR_ISOLATED_ANON +
 907                                            page_is_file_cache(page));
 908                }
 909        }
 910}
 911
 912static struct page *new_node_page(struct page *page, unsigned long node, int **x)
 913{
 914        return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
 915}
 916
 917/*
 918 * Migrate pages from one node to a target node.
 919 * Returns error or the number of pages not migrated.
 920 */
 921static int migrate_to_node(struct mm_struct *mm, int source, int dest,
 922                           int flags)
 923{
 924        nodemask_t nmask;
 925        LIST_HEAD(pagelist);
 926        int err = 0;
 927        struct vm_area_struct *vma;
 928
 929        nodes_clear(nmask);
 930        node_set(source, nmask);
 931
 932        vma = check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
 933                        flags | MPOL_MF_DISCONTIG_OK, &pagelist);
 934        if (IS_ERR(vma))
 935                return PTR_ERR(vma);
 936
 937        if (!list_empty(&pagelist)) {
 938                err = migrate_pages(&pagelist, new_node_page, dest,
 939                                                        false, MIGRATE_SYNC);
 940                if (err)
 941                        putback_lru_pages(&pagelist);
 942        }
 943
 944        return err;
 945}
 946
 947/*
 948 * Move pages between the two nodesets so as to preserve the physical
 949 * layout as much as possible.
 950 *
 951 * Returns the number of page that could not be moved.
 952 */
 953int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
 954                     const nodemask_t *to, int flags)
 955{
 956        int busy = 0;
 957        int err;
 958        nodemask_t tmp;
 959
 960        err = migrate_prep();
 961        if (err)
 962                return err;
 963
 964        down_read(&mm->mmap_sem);
 965
 966        err = migrate_vmas(mm, from, to, flags);
 967        if (err)
 968                goto out;
 969
 970        /*
 971         * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
 972         * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
 973         * bit in 'tmp', and return that <source, dest> pair for migration.
 974         * The pair of nodemasks 'to' and 'from' define the map.
 975         *
 976         * If no pair of bits is found that way, fallback to picking some
 977         * pair of 'source' and 'dest' bits that are not the same.  If the
 978         * 'source' and 'dest' bits are the same, this represents a node
 979         * that will be migrating to itself, so no pages need move.
 980         *
 981         * If no bits are left in 'tmp', or if all remaining bits left
 982         * in 'tmp' correspond to the same bit in 'to', return false
 983         * (nothing left to migrate).
 984         *
 985         * This lets us pick a pair of nodes to migrate between, such that
 986         * if possible the dest node is not already occupied by some other
 987         * source node, minimizing the risk of overloading the memory on a
 988         * node that would happen if we migrated incoming memory to a node
 989         * before migrating outgoing memory source that same node.
 990         *
 991         * A single scan of tmp is sufficient.  As we go, we remember the
 992         * most recent <s, d> pair that moved (s != d).  If we find a pair
 993         * that not only moved, but what's better, moved to an empty slot
 994         * (d is not set in tmp), then we break out then, with that pair.
 995         * Otherwise when we finish scanning from_tmp, we at least have the
 996         * most recent <s, d> pair that moved.  If we get all the way through
 997         * the scan of tmp without finding any node that moved, much less
 998         * moved to an empty node, then there is nothing left worth migrating.
 999         */
1000
1001        tmp = *from;
1002        while (!nodes_empty(tmp)) {
1003                int s,d;
1004                int source = -1;
1005                int dest = 0;
1006
1007                for_each_node_mask(s, tmp) {
1008
1009                        /*
1010                         * do_migrate_pages() tries to maintain the relative
1011                         * node relationship of the pages established between
1012                         * threads and memory areas.
1013                         *
1014                         * However if the number of source nodes is not equal to
1015                         * the number of destination nodes we can not preserve
1016                         * this node relative relationship.  In that case, skip
1017                         * copying memory from a node that is in the destination
1018                         * mask.
1019                         *
1020                         * Example: [2,3,4] -> [3,4,5] moves everything.
1021                         *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1022                         */
1023
1024                        if ((nodes_weight(*from) != nodes_weight(*to)) &&
1025                                                (node_isset(s, *to)))
1026                                continue;
1027
1028                        d = node_remap(s, *from, *to);
1029                        if (s == d)
1030                                continue;
1031
1032                        source = s;     /* Node moved. Memorize */
1033                        dest = d;
1034
1035                        /* dest not in remaining from nodes? */
1036                        if (!node_isset(dest, tmp))
1037                                break;
1038                }
1039                if (source == -1)
1040                        break;
1041
1042                node_clear(source, tmp);
1043                err = migrate_to_node(mm, source, dest, flags);
1044                if (err > 0)
1045                        busy += err;
1046                if (err < 0)
1047                        break;
1048        }
1049out:
1050        up_read(&mm->mmap_sem);
1051        if (err < 0)
1052                return err;
1053        return busy;
1054
1055}
1056
1057/*
1058 * Allocate a new page for page migration based on vma policy.
1059 * Start assuming that page is mapped by vma pointed to by @private.
1060 * Search forward from there, if not.  N.B., this assumes that the
1061 * list of pages handed to migrate_pages()--which is how we get here--
1062 * is in virtual address order.
1063 */
1064static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1065{
1066        struct vm_area_struct *vma = (struct vm_area_struct *)private;
1067        unsigned long uninitialized_var(address);
1068
1069        while (vma) {
1070                address = page_address_in_vma(page, vma);
1071                if (address != -EFAULT)
1072                        break;
1073                vma = vma->vm_next;
1074        }
1075
1076        /*
1077         * if !vma, alloc_page_vma() will use task or system default policy
1078         */
1079        return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1080}
1081#else
1082
1083static void migrate_page_add(struct page *page, struct list_head *pagelist,
1084                                unsigned long flags)
1085{
1086}
1087
1088int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1089                     const nodemask_t *to, int flags)
1090{
1091        return -ENOSYS;
1092}
1093
1094static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1095{
1096        return NULL;
1097}
1098#endif
1099
1100static long do_mbind(unsigned long start, unsigned long len,
1101                     unsigned short mode, unsigned short mode_flags,
1102                     nodemask_t *nmask, unsigned long flags)
1103{
1104        struct vm_area_struct *vma;
1105        struct mm_struct *mm = current->mm;
1106        struct mempolicy *new;
1107        unsigned long end;
1108        int err;
1109        LIST_HEAD(pagelist);
1110
1111        if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1112                                     MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1113                return -EINVAL;
1114        if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1115                return -EPERM;
1116
1117        if (start & ~PAGE_MASK)
1118                return -EINVAL;
1119
1120        if (mode == MPOL_DEFAULT)
1121                flags &= ~MPOL_MF_STRICT;
1122
1123        len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1124        end = start + len;
1125
1126        if (end < start)
1127                return -EINVAL;
1128        if (end == start)
1129                return 0;
1130
1131        new = mpol_new(mode, mode_flags, nmask);
1132        if (IS_ERR(new))
1133                return PTR_ERR(new);
1134
1135        /*
1136         * If we are using the default policy then operation
1137         * on discontinuous address spaces is okay after all
1138         */
1139        if (!new)
1140                flags |= MPOL_MF_DISCONTIG_OK;
1141
1142        pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1143                 start, start + len, mode, mode_flags,
1144                 nmask ? nodes_addr(*nmask)[0] : -1);
1145
1146        if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1147
1148                err = migrate_prep();
1149                if (err)
1150                        goto mpol_out;
1151        }
1152        {
1153                NODEMASK_SCRATCH(scratch);
1154                if (scratch) {
1155                        down_write(&mm->mmap_sem);
1156                        task_lock(current);
1157                        err = mpol_set_nodemask(new, nmask, scratch);
1158                        task_unlock(current);
1159                        if (err)
1160                                up_write(&mm->mmap_sem);
1161                } else
1162                        err = -ENOMEM;
1163                NODEMASK_SCRATCH_FREE(scratch);
1164        }
1165        if (err)
1166                goto mpol_out;
1167
1168        vma = check_range(mm, start, end, nmask,
1169                          flags | MPOL_MF_INVERT, &pagelist);
1170
1171        err = PTR_ERR(vma);
1172        if (!IS_ERR(vma)) {
1173                int nr_failed = 0;
1174
1175                err = mbind_range(mm, start, end, new);
1176
1177                if (!list_empty(&pagelist)) {
1178                        nr_failed = migrate_pages(&pagelist, new_vma_page,
1179                                                (unsigned long)vma,
1180                                                false, MIGRATE_SYNC);
1181                        if (nr_failed)
1182                                putback_lru_pages(&pagelist);
1183                }
1184
1185                if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1186                        err = -EIO;
1187        } else
1188                putback_lru_pages(&pagelist);
1189
1190        up_write(&mm->mmap_sem);
1191 mpol_out:
1192        mpol_put(new);
1193        return err;
1194}
1195
1196/*
1197 * User space interface with variable sized bitmaps for nodelists.
1198 */
1199
1200/* Copy a node mask from user space. */
1201static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1202                     unsigned long maxnode)
1203{
1204        unsigned long k;
1205        unsigned long nlongs;
1206        unsigned long endmask;
1207
1208        --maxnode;
1209        nodes_clear(*nodes);
1210        if (maxnode == 0 || !nmask)
1211                return 0;
1212        if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1213                return -EINVAL;
1214
1215        nlongs = BITS_TO_LONGS(maxnode);
1216        if ((maxnode % BITS_PER_LONG) == 0)
1217                endmask = ~0UL;
1218        else
1219                endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1220
1221        /* When the user specified more nodes than supported just check
1222           if the non supported part is all zero. */
1223        if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1224                if (nlongs > PAGE_SIZE/sizeof(long))
1225                        return -EINVAL;
1226                for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1227                        unsigned long t;
1228                        if (get_user(t, nmask + k))
1229                                return -EFAULT;
1230                        if (k == nlongs - 1) {
1231                                if (t & endmask)
1232                                        return -EINVAL;
1233                        } else if (t)
1234                                return -EINVAL;
1235                }
1236                nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1237                endmask = ~0UL;
1238        }
1239
1240        if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1241                return -EFAULT;
1242        nodes_addr(*nodes)[nlongs-1] &= endmask;
1243        return 0;
1244}
1245
1246/* Copy a kernel node mask to user space */
1247static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1248                              nodemask_t *nodes)
1249{
1250        unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1251        const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1252
1253        if (copy > nbytes) {
1254                if (copy > PAGE_SIZE)
1255                        return -EINVAL;
1256                if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1257                        return -EFAULT;
1258                copy = nbytes;
1259        }
1260        return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1261}
1262
1263SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1264                unsigned long, mode, unsigned long __user *, nmask,
1265                unsigned long, maxnode, unsigned, flags)
1266{
1267        nodemask_t nodes;
1268        int err;
1269        unsigned short mode_flags;
1270
1271        mode_flags = mode & MPOL_MODE_FLAGS;
1272        mode &= ~MPOL_MODE_FLAGS;
1273        if (mode >= MPOL_MAX)
1274                return -EINVAL;
1275        if ((mode_flags & MPOL_F_STATIC_NODES) &&
1276            (mode_flags & MPOL_F_RELATIVE_NODES))
1277                return -EINVAL;
1278        err = get_nodes(&nodes, nmask, maxnode);
1279        if (err)
1280                return err;
1281        return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1282}
1283
1284/* Set the process memory policy */
1285SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1286                unsigned long, maxnode)
1287{
1288        int err;
1289        nodemask_t nodes;
1290        unsigned short flags;
1291
1292        flags = mode & MPOL_MODE_FLAGS;
1293        mode &= ~MPOL_MODE_FLAGS;
1294        if ((unsigned int)mode >= MPOL_MAX)
1295                return -EINVAL;
1296        if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1297                return -EINVAL;
1298        err = get_nodes(&nodes, nmask, maxnode);
1299        if (err)
1300                return err;
1301        return do_set_mempolicy(mode, flags, &nodes);
1302}
1303
1304SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1305                const unsigned long __user *, old_nodes,
1306                const unsigned long __user *, new_nodes)
1307{
1308        const struct cred *cred = current_cred(), *tcred;
1309        struct mm_struct *mm = NULL;
1310        struct task_struct *task;
1311        nodemask_t task_nodes;
1312        int err;
1313        nodemask_t *old;
1314        nodemask_t *new;
1315        NODEMASK_SCRATCH(scratch);
1316
1317        if (!scratch)
1318                return -ENOMEM;
1319
1320        old = &scratch->mask1;
1321        new = &scratch->mask2;
1322
1323        err = get_nodes(old, old_nodes, maxnode);
1324        if (err)
1325                goto out;
1326
1327        err = get_nodes(new, new_nodes, maxnode);
1328        if (err)
1329                goto out;
1330
1331        /* Find the mm_struct */
1332        rcu_read_lock();
1333        task = pid ? find_task_by_vpid(pid) : current;
1334        if (!task) {
1335                rcu_read_unlock();
1336                err = -ESRCH;
1337                goto out;
1338        }
1339        get_task_struct(task);
1340
1341        err = -EINVAL;
1342
1343        /*
1344         * Check if this process has the right to modify the specified
1345         * process. The right exists if the process has administrative
1346         * capabilities, superuser privileges or the same
1347         * userid as the target process.
1348         */
1349        tcred = __task_cred(task);
1350        if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1351            !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1352            !capable(CAP_SYS_NICE)) {
1353                rcu_read_unlock();
1354                err = -EPERM;
1355                goto out_put;
1356        }
1357        rcu_read_unlock();
1358
1359        task_nodes = cpuset_mems_allowed(task);
1360        /* Is the user allowed to access the target nodes? */
1361        if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1362                err = -EPERM;
1363                goto out_put;
1364        }
1365
1366        if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1367                err = -EINVAL;
1368                goto out_put;
1369        }
1370
1371        err = security_task_movememory(task);
1372        if (err)
1373                goto out_put;
1374
1375        mm = get_task_mm(task);
1376        put_task_struct(task);
1377
1378        if (!mm) {
1379                err = -EINVAL;
1380                goto out;
1381        }
1382
1383        err = do_migrate_pages(mm, old, new,
1384                capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1385
1386        mmput(mm);
1387out:
1388        NODEMASK_SCRATCH_FREE(scratch);
1389
1390        return err;
1391
1392out_put:
1393        put_task_struct(task);
1394        goto out;
1395
1396}
1397
1398
1399/* Retrieve NUMA policy */
1400SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1401                unsigned long __user *, nmask, unsigned long, maxnode,
1402                unsigned long, addr, unsigned long, flags)
1403{
1404        int err;
1405        int uninitialized_var(pval);
1406        nodemask_t nodes;
1407
1408        if (nmask != NULL && maxnode < MAX_NUMNODES)
1409                return -EINVAL;
1410
1411        err = do_get_mempolicy(&pval, &nodes, addr, flags);
1412
1413        if (err)
1414                return err;
1415
1416        if (policy && put_user(pval, policy))
1417                return -EFAULT;
1418
1419        if (nmask)
1420                err = copy_nodes_to_user(nmask, maxnode, &nodes);
1421
1422        return err;
1423}
1424
1425#ifdef CONFIG_COMPAT
1426
1427asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1428                                     compat_ulong_t __user *nmask,
1429                                     compat_ulong_t maxnode,
1430                                     compat_ulong_t addr, compat_ulong_t flags)
1431{
1432        long err;
1433        unsigned long __user *nm = NULL;
1434        unsigned long nr_bits, alloc_size;
1435        DECLARE_BITMAP(bm, MAX_NUMNODES);
1436
1437        nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1438        alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1439
1440        if (nmask)
1441                nm = compat_alloc_user_space(alloc_size);
1442
1443        err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1444
1445        if (!err && nmask) {
1446                unsigned long copy_size;
1447                copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1448                err = copy_from_user(bm, nm, copy_size);
1449                /* ensure entire bitmap is zeroed */
1450                err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1451                err |= compat_put_bitmap(nmask, bm, nr_bits);
1452        }
1453
1454        return err;
1455}
1456
1457asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1458                                     compat_ulong_t maxnode)
1459{
1460        long err = 0;
1461        unsigned long __user *nm = NULL;
1462        unsigned long nr_bits, alloc_size;
1463        DECLARE_BITMAP(bm, MAX_NUMNODES);
1464
1465        nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1466        alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1467
1468        if (nmask) {
1469                err = compat_get_bitmap(bm, nmask, nr_bits);
1470                nm = compat_alloc_user_space(alloc_size);
1471                err |= copy_to_user(nm, bm, alloc_size);
1472        }
1473
1474        if (err)
1475                return -EFAULT;
1476
1477        return sys_set_mempolicy(mode, nm, nr_bits+1);
1478}
1479
1480asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1481                             compat_ulong_t mode, compat_ulong_t __user *nmask,
1482                             compat_ulong_t maxnode, compat_ulong_t flags)
1483{
1484        long err = 0;
1485        unsigned long __user *nm = NULL;
1486        unsigned long nr_bits, alloc_size;
1487        nodemask_t bm;
1488
1489        nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1490        alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1491
1492        if (nmask) {
1493                err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1494                nm = compat_alloc_user_space(alloc_size);
1495                err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1496        }
1497
1498        if (err)
1499                return -EFAULT;
1500
1501        return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1502}
1503
1504#endif
1505
1506/*
1507 * get_vma_policy(@task, @vma, @addr)
1508 * @task - task for fallback if vma policy == default
1509 * @vma   - virtual memory area whose policy is sought
1510 * @addr  - address in @vma for shared policy lookup
1511 *
1512 * Returns effective policy for a VMA at specified address.
1513 * Falls back to @task or system default policy, as necessary.
1514 * Current or other task's task mempolicy and non-shared vma policies
1515 * are protected by the task's mmap_sem, which must be held for read by
1516 * the caller.
1517 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1518 * count--added by the get_policy() vm_op, as appropriate--to protect against
1519 * freeing by another task.  It is the caller's responsibility to free the
1520 * extra reference for shared policies.
1521 */
1522struct mempolicy *get_vma_policy(struct task_struct *task,
1523                struct vm_area_struct *vma, unsigned long addr)
1524{
1525        struct mempolicy *pol = task->mempolicy;
1526
1527        if (vma) {
1528                if (vma->vm_ops && vma->vm_ops->get_policy) {
1529                        struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1530                                                                        addr);
1531                        if (vpol)
1532                                pol = vpol;
1533                } else if (vma->vm_policy)
1534                        pol = vma->vm_policy;
1535        }
1536        if (!pol)
1537                pol = &default_policy;
1538        return pol;
1539}
1540
1541/*
1542 * Return a nodemask representing a mempolicy for filtering nodes for
1543 * page allocation
1544 */
1545static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1546{
1547        /* Lower zones don't get a nodemask applied for MPOL_BIND */
1548        if (unlikely(policy->mode == MPOL_BIND) &&
1549                        gfp_zone(gfp) >= policy_zone &&
1550                        cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1551                return &policy->v.nodes;
1552
1553        return NULL;
1554}
1555
1556/* Return a zonelist indicated by gfp for node representing a mempolicy */
1557static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1558        int nd)
1559{
1560        switch (policy->mode) {
1561        case MPOL_PREFERRED:
1562                if (!(policy->flags & MPOL_F_LOCAL))
1563                        nd = policy->v.preferred_node;
1564                break;
1565        case MPOL_BIND:
1566                /*
1567                 * Normally, MPOL_BIND allocations are node-local within the
1568                 * allowed nodemask.  However, if __GFP_THISNODE is set and the
1569                 * current node isn't part of the mask, we use the zonelist for
1570                 * the first node in the mask instead.
1571                 */
1572                if (unlikely(gfp & __GFP_THISNODE) &&
1573                                unlikely(!node_isset(nd, policy->v.nodes)))
1574                        nd = first_node(policy->v.nodes);
1575                break;
1576        default:
1577                BUG();
1578        }
1579        return node_zonelist(nd, gfp);
1580}
1581
1582/* Do dynamic interleaving for a process */
1583static unsigned interleave_nodes(struct mempolicy *policy)
1584{
1585        unsigned nid, next;
1586        struct task_struct *me = current;
1587
1588        nid = me->il_next;
1589        next = next_node(nid, policy->v.nodes);
1590        if (next >= MAX_NUMNODES)
1591                next = first_node(policy->v.nodes);
1592        if (next < MAX_NUMNODES)
1593                me->il_next = next;
1594        return nid;
1595}
1596
1597/*
1598 * Depending on the memory policy provide a node from which to allocate the
1599 * next slab entry.
1600 * @policy must be protected by freeing by the caller.  If @policy is
1601 * the current task's mempolicy, this protection is implicit, as only the
1602 * task can change it's policy.  The system default policy requires no
1603 * such protection.
1604 */
1605unsigned slab_node(struct mempolicy *policy)
1606{
1607        if (!policy || policy->flags & MPOL_F_LOCAL)
1608                return numa_node_id();
1609
1610        switch (policy->mode) {
1611        case MPOL_PREFERRED:
1612                /*
1613                 * handled MPOL_F_LOCAL above
1614                 */
1615                return policy->v.preferred_node;
1616
1617        case MPOL_INTERLEAVE:
1618                return interleave_nodes(policy);
1619
1620        case MPOL_BIND: {
1621                /*
1622                 * Follow bind policy behavior and start allocation at the
1623                 * first node.
1624                 */
1625                struct zonelist *zonelist;
1626                struct zone *zone;
1627                enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1628                zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1629                (void)first_zones_zonelist(zonelist, highest_zoneidx,
1630                                                        &policy->v.nodes,
1631                                                        &zone);
1632                return zone ? zone->node : numa_node_id();
1633        }
1634
1635        default:
1636                BUG();
1637        }
1638}
1639
1640/* Do static interleaving for a VMA with known offset. */
1641static unsigned offset_il_node(struct mempolicy *pol,
1642                struct vm_area_struct *vma, unsigned long off)
1643{
1644        unsigned nnodes = nodes_weight(pol->v.nodes);
1645        unsigned target;
1646        int c;
1647        int nid = -1;
1648
1649        if (!nnodes)
1650                return numa_node_id();
1651        target = (unsigned int)off % nnodes;
1652        c = 0;
1653        do {
1654                nid = next_node(nid, pol->v.nodes);
1655                c++;
1656        } while (c <= target);
1657        return nid;
1658}
1659
1660/* Determine a node number for interleave */
1661static inline unsigned interleave_nid(struct mempolicy *pol,
1662                 struct vm_area_struct *vma, unsigned long addr, int shift)
1663{
1664        if (vma) {
1665                unsigned long off;
1666
1667                /*
1668                 * for small pages, there is no difference between
1669                 * shift and PAGE_SHIFT, so the bit-shift is safe.
1670                 * for huge pages, since vm_pgoff is in units of small
1671                 * pages, we need to shift off the always 0 bits to get
1672                 * a useful offset.
1673                 */
1674                BUG_ON(shift < PAGE_SHIFT);
1675                off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1676                off += (addr - vma->vm_start) >> shift;
1677                return offset_il_node(pol, vma, off);
1678        } else
1679                return interleave_nodes(pol);
1680}
1681
1682/*
1683 * Return the bit number of a random bit set in the nodemask.
1684 * (returns -1 if nodemask is empty)
1685 */
1686int node_random(const nodemask_t *maskp)
1687{
1688        int w, bit = -1;
1689
1690        w = nodes_weight(*maskp);
1691        if (w)
1692                bit = bitmap_ord_to_pos(maskp->bits,
1693                        get_random_int() % w, MAX_NUMNODES);
1694        return bit;
1695}
1696
1697#ifdef CONFIG_HUGETLBFS
1698/*
1699 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1700 * @vma = virtual memory area whose policy is sought
1701 * @addr = address in @vma for shared policy lookup and interleave policy
1702 * @gfp_flags = for requested zone
1703 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1704 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1705 *
1706 * Returns a zonelist suitable for a huge page allocation and a pointer
1707 * to the struct mempolicy for conditional unref after allocation.
1708 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1709 * @nodemask for filtering the zonelist.
1710 *
1711 * Must be protected by get_mems_allowed()
1712 */
1713struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1714                                gfp_t gfp_flags, struct mempolicy **mpol,
1715                                nodemask_t **nodemask)
1716{
1717        struct zonelist *zl;
1718
1719        *mpol = get_vma_policy(current, vma, addr);
1720        *nodemask = NULL;       /* assume !MPOL_BIND */
1721
1722        if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1723                zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1724                                huge_page_shift(hstate_vma(vma))), gfp_flags);
1725        } else {
1726                zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1727                if ((*mpol)->mode == MPOL_BIND)
1728                        *nodemask = &(*mpol)->v.nodes;
1729        }
1730        return zl;
1731}
1732
1733/*
1734 * init_nodemask_of_mempolicy
1735 *
1736 * If the current task's mempolicy is "default" [NULL], return 'false'
1737 * to indicate default policy.  Otherwise, extract the policy nodemask
1738 * for 'bind' or 'interleave' policy into the argument nodemask, or
1739 * initialize the argument nodemask to contain the single node for
1740 * 'preferred' or 'local' policy and return 'true' to indicate presence
1741 * of non-default mempolicy.
1742 *
1743 * We don't bother with reference counting the mempolicy [mpol_get/put]
1744 * because the current task is examining it's own mempolicy and a task's
1745 * mempolicy is only ever changed by the task itself.
1746 *
1747 * N.B., it is the caller's responsibility to free a returned nodemask.
1748 */
1749bool init_nodemask_of_mempolicy(nodemask_t *mask)
1750{
1751        struct mempolicy *mempolicy;
1752        int nid;
1753
1754        if (!(mask && current->mempolicy))
1755                return false;
1756
1757        task_lock(current);
1758        mempolicy = current->mempolicy;
1759        switch (mempolicy->mode) {
1760        case MPOL_PREFERRED:
1761                if (mempolicy->flags & MPOL_F_LOCAL)
1762                        nid = numa_node_id();
1763                else
1764                        nid = mempolicy->v.preferred_node;
1765                init_nodemask_of_node(mask, nid);
1766                break;
1767
1768        case MPOL_BIND:
1769                /* Fall through */
1770        case MPOL_INTERLEAVE:
1771                *mask =  mempolicy->v.nodes;
1772                break;
1773
1774        default:
1775                BUG();
1776        }
1777        task_unlock(current);
1778
1779        return true;
1780}
1781#endif
1782
1783/*
1784 * mempolicy_nodemask_intersects
1785 *
1786 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1787 * policy.  Otherwise, check for intersection between mask and the policy
1788 * nodemask for 'bind' or 'interleave' policy.  For 'perferred' or 'local'
1789 * policy, always return true since it may allocate elsewhere on fallback.
1790 *
1791 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1792 */
1793bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1794                                        const nodemask_t *mask)
1795{
1796        struct mempolicy *mempolicy;
1797        bool ret = true;
1798
1799        if (!mask)
1800                return ret;
1801        task_lock(tsk);
1802        mempolicy = tsk->mempolicy;
1803        if (!mempolicy)
1804                goto out;
1805
1806        switch (mempolicy->mode) {
1807        case MPOL_PREFERRED:
1808                /*
1809                 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1810                 * allocate from, they may fallback to other nodes when oom.
1811                 * Thus, it's possible for tsk to have allocated memory from
1812                 * nodes in mask.
1813                 */
1814                break;
1815        case MPOL_BIND:
1816        case MPOL_INTERLEAVE:
1817                ret = nodes_intersects(mempolicy->v.nodes, *mask);
1818                break;
1819        default:
1820                BUG();
1821        }
1822out:
1823        task_unlock(tsk);
1824        return ret;
1825}
1826
1827/* Allocate a page in interleaved policy.
1828   Own path because it needs to do special accounting. */
1829static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1830                                        unsigned nid)
1831{
1832        struct zonelist *zl;
1833        struct page *page;
1834
1835        zl = node_zonelist(nid, gfp);
1836        page = __alloc_pages(gfp, order, zl);
1837        if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1838                inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1839        return page;
1840}
1841
1842/**
1843 *      alloc_pages_vma - Allocate a page for a VMA.
1844 *
1845 *      @gfp:
1846 *      %GFP_USER    user allocation.
1847 *      %GFP_KERNEL  kernel allocations,
1848 *      %GFP_HIGHMEM highmem/user allocations,
1849 *      %GFP_FS      allocation should not call back into a file system.
1850 *      %GFP_ATOMIC  don't sleep.
1851 *
1852 *      @order:Order of the GFP allocation.
1853 *      @vma:  Pointer to VMA or NULL if not available.
1854 *      @addr: Virtual Address of the allocation. Must be inside the VMA.
1855 *
1856 *      This function allocates a page from the kernel page pool and applies
1857 *      a NUMA policy associated with the VMA or the current process.
1858 *      When VMA is not NULL caller must hold down_read on the mmap_sem of the
1859 *      mm_struct of the VMA to prevent it from going away. Should be used for
1860 *      all allocations for pages that will be mapped into
1861 *      user space. Returns NULL when no page can be allocated.
1862 *
1863 *      Should be called with the mm_sem of the vma hold.
1864 */
1865struct page *
1866alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1867                unsigned long addr, int node)
1868{
1869        struct mempolicy *pol;
1870        struct zonelist *zl;
1871        struct page *page;
1872        unsigned int cpuset_mems_cookie;
1873
1874retry_cpuset:
1875        pol = get_vma_policy(current, vma, addr);
1876        cpuset_mems_cookie = get_mems_allowed();
1877
1878        if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1879                unsigned nid;
1880
1881                nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1882                mpol_cond_put(pol);
1883                page = alloc_page_interleave(gfp, order, nid);
1884                if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1885                        goto retry_cpuset;
1886
1887                return page;
1888        }
1889        zl = policy_zonelist(gfp, pol, node);
1890        if (unlikely(mpol_needs_cond_ref(pol))) {
1891                /*
1892                 * slow path: ref counted shared policy
1893                 */
1894                struct page *page =  __alloc_pages_nodemask(gfp, order,
1895                                                zl, policy_nodemask(gfp, pol));
1896                __mpol_put(pol);
1897                if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1898                        goto retry_cpuset;
1899                return page;
1900        }
1901        /*
1902         * fast path:  default or task policy
1903         */
1904        page = __alloc_pages_nodemask(gfp, order, zl,
1905                                      policy_nodemask(gfp, pol));
1906        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1907                goto retry_cpuset;
1908        return page;
1909}
1910
1911/**
1912 *      alloc_pages_current - Allocate pages.
1913 *
1914 *      @gfp:
1915 *              %GFP_USER   user allocation,
1916 *              %GFP_KERNEL kernel allocation,
1917 *              %GFP_HIGHMEM highmem allocation,
1918 *              %GFP_FS     don't call back into a file system.
1919 *              %GFP_ATOMIC don't sleep.
1920 *      @order: Power of two of allocation size in pages. 0 is a single page.
1921 *
1922 *      Allocate a page from the kernel page pool.  When not in
1923 *      interrupt context and apply the current process NUMA policy.
1924 *      Returns NULL when no page can be allocated.
1925 *
1926 *      Don't call cpuset_update_task_memory_state() unless
1927 *      1) it's ok to take cpuset_sem (can WAIT), and
1928 *      2) allocating for current task (not interrupt).
1929 */
1930struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1931{
1932        struct mempolicy *pol = current->mempolicy;
1933        struct page *page;
1934        unsigned int cpuset_mems_cookie;
1935
1936        if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1937                pol = &default_policy;
1938
1939retry_cpuset:
1940        cpuset_mems_cookie = get_mems_allowed();
1941
1942        /*
1943         * No reference counting needed for current->mempolicy
1944         * nor system default_policy
1945         */
1946        if (pol->mode == MPOL_INTERLEAVE)
1947                page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1948        else
1949                page = __alloc_pages_nodemask(gfp, order,
1950                                policy_zonelist(gfp, pol, numa_node_id()),
1951                                policy_nodemask(gfp, pol));
1952
1953        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1954                goto retry_cpuset;
1955
1956        return page;
1957}
1958EXPORT_SYMBOL(alloc_pages_current);
1959
1960/*
1961 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
1962 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
1963 * with the mems_allowed returned by cpuset_mems_allowed().  This
1964 * keeps mempolicies cpuset relative after its cpuset moves.  See
1965 * further kernel/cpuset.c update_nodemask().
1966 *
1967 * current's mempolicy may be rebinded by the other task(the task that changes
1968 * cpuset's mems), so we needn't do rebind work for current task.
1969 */
1970
1971/* Slow path of a mempolicy duplicate */
1972struct mempolicy *__mpol_dup(struct mempolicy *old)
1973{
1974        struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
1975
1976        if (!new)
1977                return ERR_PTR(-ENOMEM);
1978
1979        /* task's mempolicy is protected by alloc_lock */
1980        if (old == current->mempolicy) {
1981                task_lock(current);
1982                *new = *old;
1983                task_unlock(current);
1984        } else
1985                *new = *old;
1986
1987        rcu_read_lock();
1988        if (current_cpuset_is_being_rebound()) {
1989                nodemask_t mems = cpuset_mems_allowed(current);
1990                if (new->flags & MPOL_F_REBINDING)
1991                        mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
1992                else
1993                        mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
1994        }
1995        rcu_read_unlock();
1996        atomic_set(&new->refcnt, 1);
1997        return new;
1998}
1999
2000/*
2001 * If *frompol needs [has] an extra ref, copy *frompol to *tompol ,
2002 * eliminate the * MPOL_F_* flags that require conditional ref and
2003 * [NOTE!!!] drop the extra ref.  Not safe to reference *frompol directly
2004 * after return.  Use the returned value.
2005 *
2006 * Allows use of a mempolicy for, e.g., multiple allocations with a single
2007 * policy lookup, even if the policy needs/has extra ref on lookup.
2008 * shmem_readahead needs this.
2009 */
2010struct mempolicy *__mpol_cond_copy(struct mempolicy *tompol,
2011                                                struct mempolicy *frompol)
2012{
2013        if (!mpol_needs_cond_ref(frompol))
2014                return frompol;
2015
2016        *tompol = *frompol;
2017        tompol->flags &= ~MPOL_F_SHARED;        /* copy doesn't need unref */
2018        __mpol_put(frompol);
2019        return tompol;
2020}
2021
2022/* Slow path of a mempolicy comparison */
2023bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2024{
2025        if (!a || !b)
2026                return false;
2027        if (a->mode != b->mode)
2028                return false;
2029        if (a->flags != b->flags)
2030                return false;
2031        if (mpol_store_user_nodemask(a))
2032                if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2033                        return false;
2034
2035        switch (a->mode) {
2036        case MPOL_BIND:
2037                /* Fall through */
2038        case MPOL_INTERLEAVE:
2039                return !!nodes_equal(a->v.nodes, b->v.nodes);
2040        case MPOL_PREFERRED:
2041                return a->v.preferred_node == b->v.preferred_node;
2042        default:
2043                BUG();
2044                return false;
2045        }
2046}
2047
2048/*
2049 * Shared memory backing store policy support.
2050 *
2051 * Remember policies even when nobody has shared memory mapped.
2052 * The policies are kept in Red-Black tree linked from the inode.
2053 * They are protected by the sp->lock spinlock, which should be held
2054 * for any accesses to the tree.
2055 */
2056
2057/* lookup first element intersecting start-end */
2058/* Caller holds sp->lock */
2059static struct sp_node *
2060sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2061{
2062        struct rb_node *n = sp->root.rb_node;
2063
2064        while (n) {
2065                struct sp_node *p = rb_entry(n, struct sp_node, nd);
2066
2067                if (start >= p->end)
2068                        n = n->rb_right;
2069                else if (end <= p->start)
2070                        n = n->rb_left;
2071                else
2072                        break;
2073        }
2074        if (!n)
2075                return NULL;
2076        for (;;) {
2077                struct sp_node *w = NULL;
2078                struct rb_node *prev = rb_prev(n);
2079                if (!prev)
2080                        break;
2081                w = rb_entry(prev, struct sp_node, nd);
2082                if (w->end <= start)
2083                        break;
2084                n = prev;
2085        }
2086        return rb_entry(n, struct sp_node, nd);
2087}
2088
2089/* Insert a new shared policy into the list. */
2090/* Caller holds sp->lock */
2091static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2092{
2093        struct rb_node **p = &sp->root.rb_node;
2094        struct rb_node *parent = NULL;
2095        struct sp_node *nd;
2096
2097        while (*p) {
2098                parent = *p;
2099                nd = rb_entry(parent, struct sp_node, nd);
2100                if (new->start < nd->start)
2101                        p = &(*p)->rb_left;
2102                else if (new->end > nd->end)
2103                        p = &(*p)->rb_right;
2104                else
2105                        BUG();
2106        }
2107        rb_link_node(&new->nd, parent, p);
2108        rb_insert_color(&new->nd, &sp->root);
2109        pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2110                 new->policy ? new->policy->mode : 0);
2111}
2112
2113/* Find shared policy intersecting idx */
2114struct mempolicy *
2115mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2116{
2117        struct mempolicy *pol = NULL;
2118        struct sp_node *sn;
2119
2120        if (!sp->root.rb_node)
2121                return NULL;
2122        spin_lock(&sp->lock);
2123        sn = sp_lookup(sp, idx, idx+1);
2124        if (sn) {
2125                mpol_get(sn->policy);
2126                pol = sn->policy;
2127        }
2128        spin_unlock(&sp->lock);
2129        return pol;
2130}
2131
2132static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2133{
2134        pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2135        rb_erase(&n->nd, &sp->root);
2136        mpol_put(n->policy);
2137        kmem_cache_free(sn_cache, n);
2138}
2139
2140static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2141                                struct mempolicy *pol)
2142{
2143        struct sp_node *n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2144
2145        if (!n)
2146                return NULL;
2147        n->start = start;
2148        n->end = end;
2149        mpol_get(pol);
2150        pol->flags |= MPOL_F_SHARED;    /* for unref */
2151        n->policy = pol;
2152        return n;
2153}
2154
2155/* Replace a policy range. */
2156static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2157                                 unsigned long end, struct sp_node *new)
2158{
2159        struct sp_node *n, *new2 = NULL;
2160
2161restart:
2162        spin_lock(&sp->lock);
2163        n = sp_lookup(sp, start, end);
2164        /* Take care of old policies in the same range. */
2165        while (n && n->start < end) {
2166                struct rb_node *next = rb_next(&n->nd);
2167                if (n->start >= start) {
2168                        if (n->end <= end)
2169                                sp_delete(sp, n);
2170                        else
2171                                n->start = end;
2172                } else {
2173                        /* Old policy spanning whole new range. */
2174                        if (n->end > end) {
2175                                if (!new2) {
2176                                        spin_unlock(&sp->lock);
2177                                        new2 = sp_alloc(end, n->end, n->policy);
2178                                        if (!new2)
2179                                                return -ENOMEM;
2180                                        goto restart;
2181                                }
2182                                n->end = start;
2183                                sp_insert(sp, new2);
2184                                new2 = NULL;
2185                                break;
2186                        } else
2187                                n->end = start;
2188                }
2189                if (!next)
2190                        break;
2191                n = rb_entry(next, struct sp_node, nd);
2192        }
2193        if (new)
2194                sp_insert(sp, new);
2195        spin_unlock(&sp->lock);
2196        if (new2) {
2197                mpol_put(new2->policy);
2198                kmem_cache_free(sn_cache, new2);
2199        }
2200        return 0;
2201}
2202
2203/**
2204 * mpol_shared_policy_init - initialize shared policy for inode
2205 * @sp: pointer to inode shared policy
2206 * @mpol:  struct mempolicy to install
2207 *
2208 * Install non-NULL @mpol in inode's shared policy rb-tree.
2209 * On entry, the current task has a reference on a non-NULL @mpol.
2210 * This must be released on exit.
2211 * This is called at get_inode() calls and we can use GFP_KERNEL.
2212 */
2213void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2214{
2215        int ret;
2216
2217        sp->root = RB_ROOT;             /* empty tree == default mempolicy */
2218        spin_lock_init(&sp->lock);
2219
2220        if (mpol) {
2221                struct vm_area_struct pvma;
2222                struct mempolicy *new;
2223                NODEMASK_SCRATCH(scratch);
2224
2225                if (!scratch)
2226                        goto put_mpol;
2227                /* contextualize the tmpfs mount point mempolicy */
2228                new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2229                if (IS_ERR(new))
2230                        goto free_scratch; /* no valid nodemask intersection */
2231
2232                task_lock(current);
2233                ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2234                task_unlock(current);
2235                if (ret)
2236                        goto put_new;
2237
2238                /* Create pseudo-vma that contains just the policy */
2239                memset(&pvma, 0, sizeof(struct vm_area_struct));
2240                pvma.vm_end = TASK_SIZE;        /* policy covers entire file */
2241                mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2242
2243put_new:
2244                mpol_put(new);                  /* drop initial ref */
2245free_scratch:
2246                NODEMASK_SCRATCH_FREE(scratch);
2247put_mpol:
2248                mpol_put(mpol); /* drop our incoming ref on sb mpol */
2249        }
2250}
2251
2252int mpol_set_shared_policy(struct shared_policy *info,
2253                        struct vm_area_struct *vma, struct mempolicy *npol)
2254{
2255        int err;
2256        struct sp_node *new = NULL;
2257        unsigned long sz = vma_pages(vma);
2258
2259        pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2260                 vma->vm_pgoff,
2261                 sz, npol ? npol->mode : -1,
2262                 npol ? npol->flags : -1,
2263                 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2264
2265        if (npol) {
2266                new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2267                if (!new)
2268                        return -ENOMEM;
2269        }
2270        err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2271        if (err && new)
2272                kmem_cache_free(sn_cache, new);
2273        return err;
2274}
2275
2276/* Free a backing policy store on inode delete. */
2277void mpol_free_shared_policy(struct shared_policy *p)
2278{
2279        struct sp_node *n;
2280        struct rb_node *next;
2281
2282        if (!p->root.rb_node)
2283                return;
2284        spin_lock(&p->lock);
2285        next = rb_first(&p->root);
2286        while (next) {
2287                n = rb_entry(next, struct sp_node, nd);
2288                next = rb_next(&n->nd);
2289                rb_erase(&n->nd, &p->root);
2290                mpol_put(n->policy);
2291                kmem_cache_free(sn_cache, n);
2292        }
2293        spin_unlock(&p->lock);
2294}
2295
2296/* assumes fs == KERNEL_DS */
2297void __init numa_policy_init(void)
2298{
2299        nodemask_t interleave_nodes;
2300        unsigned long largest = 0;
2301        int nid, prefer = 0;
2302
2303        policy_cache = kmem_cache_create("numa_policy",
2304                                         sizeof(struct mempolicy),
2305                                         0, SLAB_PANIC, NULL);
2306
2307        sn_cache = kmem_cache_create("shared_policy_node",
2308                                     sizeof(struct sp_node),
2309                                     0, SLAB_PANIC, NULL);
2310
2311        /*
2312         * Set interleaving policy for system init. Interleaving is only
2313         * enabled across suitably sized nodes (default is >= 16MB), or
2314         * fall back to the largest node if they're all smaller.
2315         */
2316        nodes_clear(interleave_nodes);
2317        for_each_node_state(nid, N_HIGH_MEMORY) {
2318                unsigned long total_pages = node_present_pages(nid);
2319
2320                /* Preserve the largest node */
2321                if (largest < total_pages) {
2322                        largest = total_pages;
2323                        prefer = nid;
2324                }
2325
2326                /* Interleave this node? */
2327                if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2328                        node_set(nid, interleave_nodes);
2329        }
2330
2331        /* All too small, use the largest */
2332        if (unlikely(nodes_empty(interleave_nodes)))
2333                node_set(prefer, interleave_nodes);
2334
2335        if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2336                printk("numa_policy_init: interleaving failed\n");
2337}
2338
2339/* Reset policy of current process to default */
2340void numa_default_policy(void)
2341{
2342        do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2343}
2344
2345/*
2346 * Parse and format mempolicy from/to strings
2347 */
2348
2349/*
2350 * "local" is pseudo-policy:  MPOL_PREFERRED with MPOL_F_LOCAL flag
2351 * Used only for mpol_parse_str() and mpol_to_str()
2352 */
2353#define MPOL_LOCAL MPOL_MAX
2354static const char * const policy_modes[] =
2355{
2356        [MPOL_DEFAULT]    = "default",
2357        [MPOL_PREFERRED]  = "prefer",
2358        [MPOL_BIND]       = "bind",
2359        [MPOL_INTERLEAVE] = "interleave",
2360        [MPOL_LOCAL]      = "local"
2361};
2362
2363
2364#ifdef CONFIG_TMPFS
2365/**
2366 * mpol_parse_str - parse string to mempolicy
2367 * @str:  string containing mempolicy to parse
2368 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2369 * @no_context:  flag whether to "contextualize" the mempolicy
2370 *
2371 * Format of input:
2372 *      <mode>[=<flags>][:<nodelist>]
2373 *
2374 * if @no_context is true, save the input nodemask in w.user_nodemask in
2375 * the returned mempolicy.  This will be used to "clone" the mempolicy in
2376 * a specific context [cpuset] at a later time.  Used to parse tmpfs mpol
2377 * mount option.  Note that if 'static' or 'relative' mode flags were
2378 * specified, the input nodemask will already have been saved.  Saving
2379 * it again is redundant, but safe.
2380 *
2381 * On success, returns 0, else 1
2382 */
2383int mpol_parse_str(char *str, struct mempolicy **mpol, int no_context)
2384{
2385        struct mempolicy *new = NULL;
2386        unsigned short mode;
2387        unsigned short uninitialized_var(mode_flags);
2388        nodemask_t nodes;
2389        char *nodelist = strchr(str, ':');
2390        char *flags = strchr(str, '=');
2391        int err = 1;
2392
2393        if (nodelist) {
2394                /* NUL-terminate mode or flags string */
2395                *nodelist++ = '\0';
2396                if (nodelist_parse(nodelist, nodes))
2397                        goto out;
2398                if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2399                        goto out;
2400        } else
2401                nodes_clear(nodes);
2402
2403        if (flags)
2404                *flags++ = '\0';        /* terminate mode string */
2405
2406        for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2407                if (!strcmp(str, policy_modes[mode])) {
2408                        break;
2409                }
2410        }
2411        if (mode > MPOL_LOCAL)
2412                goto out;
2413
2414        switch (mode) {
2415        case MPOL_PREFERRED:
2416                /*
2417                 * Insist on a nodelist of one node only
2418                 */
2419                if (nodelist) {
2420                        char *rest = nodelist;
2421                        while (isdigit(*rest))
2422                                rest++;
2423                        if (*rest)
2424                                goto out;
2425                }
2426                break;
2427        case MPOL_INTERLEAVE:
2428                /*
2429                 * Default to online nodes with memory if no nodelist
2430                 */
2431                if (!nodelist)
2432                        nodes = node_states[N_HIGH_MEMORY];
2433                break;
2434        case MPOL_LOCAL:
2435                /*
2436                 * Don't allow a nodelist;  mpol_new() checks flags
2437                 */
2438                if (nodelist)
2439                        goto out;
2440                mode = MPOL_PREFERRED;
2441                break;
2442        case MPOL_DEFAULT:
2443                /*
2444                 * Insist on a empty nodelist
2445                 */
2446                if (!nodelist)
2447                        err = 0;
2448                goto out;
2449        case MPOL_BIND:
2450                /*
2451                 * Insist on a nodelist
2452                 */
2453                if (!nodelist)
2454                        goto out;
2455        }
2456
2457        mode_flags = 0;
2458        if (flags) {
2459                /*
2460                 * Currently, we only support two mutually exclusive
2461                 * mode flags.
2462                 */
2463                if (!strcmp(flags, "static"))
2464                        mode_flags |= MPOL_F_STATIC_NODES;
2465                else if (!strcmp(flags, "relative"))
2466                        mode_flags |= MPOL_F_RELATIVE_NODES;
2467                else
2468                        goto out;
2469        }
2470
2471        new = mpol_new(mode, mode_flags, &nodes);
2472        if (IS_ERR(new))
2473                goto out;
2474
2475        if (no_context) {
2476                /* save for contextualization */
2477                new->w.user_nodemask = nodes;
2478        } else {
2479                int ret;
2480                NODEMASK_SCRATCH(scratch);
2481                if (scratch) {
2482                        task_lock(current);
2483                        ret = mpol_set_nodemask(new, &nodes, scratch);
2484                        task_unlock(current);
2485                } else
2486                        ret = -ENOMEM;
2487                NODEMASK_SCRATCH_FREE(scratch);
2488                if (ret) {
2489                        mpol_put(new);
2490                        goto out;
2491                }
2492        }
2493        err = 0;
2494
2495out:
2496        /* Restore string for error message */
2497        if (nodelist)
2498                *--nodelist = ':';
2499        if (flags)
2500                *--flags = '=';
2501        if (!err)
2502                *mpol = new;
2503        return err;
2504}
2505#endif /* CONFIG_TMPFS */
2506
2507/**
2508 * mpol_to_str - format a mempolicy structure for printing
2509 * @buffer:  to contain formatted mempolicy string
2510 * @maxlen:  length of @buffer
2511 * @pol:  pointer to mempolicy to be formatted
2512 * @no_context:  "context free" mempolicy - use nodemask in w.user_nodemask
2513 *
2514 * Convert a mempolicy into a string.
2515 * Returns the number of characters in buffer (if positive)
2516 * or an error (negative)
2517 */
2518int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int no_context)
2519{
2520        char *p = buffer;
2521        int l;
2522        nodemask_t nodes;
2523        unsigned short mode;
2524        unsigned short flags = pol ? pol->flags : 0;
2525
2526        /*
2527         * Sanity check:  room for longest mode, flag and some nodes
2528         */
2529        VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2530
2531        if (!pol || pol == &default_policy)
2532                mode = MPOL_DEFAULT;
2533        else
2534                mode = pol->mode;
2535
2536        switch (mode) {
2537        case MPOL_DEFAULT:
2538                nodes_clear(nodes);
2539                break;
2540
2541        case MPOL_PREFERRED:
2542                nodes_clear(nodes);
2543                if (flags & MPOL_F_LOCAL)
2544                        mode = MPOL_LOCAL;      /* pseudo-policy */
2545                else
2546                        node_set(pol->v.preferred_node, nodes);
2547                break;
2548
2549        case MPOL_BIND:
2550                /* Fall through */
2551        case MPOL_INTERLEAVE:
2552                if (no_context)
2553                        nodes = pol->w.user_nodemask;
2554                else
2555                        nodes = pol->v.nodes;
2556                break;
2557
2558        default:
2559                return -EINVAL;
2560        }
2561
2562        l = strlen(policy_modes[mode]);
2563        if (buffer + maxlen < p + l + 1)
2564                return -ENOSPC;
2565
2566        strcpy(p, policy_modes[mode]);
2567        p += l;
2568
2569        if (flags & MPOL_MODE_FLAGS) {
2570                if (buffer + maxlen < p + 2)
2571                        return -ENOSPC;
2572                *p++ = '=';
2573
2574                /*
2575                 * Currently, the only defined flags are mutually exclusive
2576                 */
2577                if (flags & MPOL_F_STATIC_NODES)
2578                        p += snprintf(p, buffer + maxlen - p, "static");
2579                else if (flags & MPOL_F_RELATIVE_NODES)
2580                        p += snprintf(p, buffer + maxlen - p, "relative");
2581        }
2582
2583        if (!nodes_empty(nodes)) {
2584                if (buffer + maxlen < p + 2)
2585                        return -ENOSPC;
2586                *p++ = ':';
2587                p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2588        }
2589        return p - buffer;
2590}
2591
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.