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/*
 611 * Apply policy to a single VMA
 612 * This must be called with the mmap_sem held for writing.
 613 */
 614static int vma_replace_policy(struct vm_area_struct *vma,
 615                                                struct mempolicy *pol)
 616{
 617        int err;
 618        struct mempolicy *old;
 619        struct mempolicy *new;
 620
 621        pr_debug("vma %lx-%lx/%lx vm_ops %p vm_file %p set_policy %p\n",
 622                 vma->vm_start, vma->vm_end, vma->vm_pgoff,
 623                 vma->vm_ops, vma->vm_file,
 624                 vma->vm_ops ? vma->vm_ops->set_policy : NULL);
 625
 626        new = mpol_dup(pol);
 627        if (IS_ERR(new))
 628                return PTR_ERR(new);
 629
 630        if (vma->vm_ops && vma->vm_ops->set_policy) {
 631                err = vma->vm_ops->set_policy(vma, new);
 632                if (err)
 633                        goto err_out;
 634        }
 635
 636        old = vma->vm_policy;
 637        vma->vm_policy = new; /* protected by mmap_sem */
 638        mpol_put(old);
 639
 640        return 0;
 641 err_out:
 642        mpol_put(new);
 643        return err;
 644}
 645
 646/* Step 2: apply policy to a range and do splits. */
 647static int mbind_range(struct mm_struct *mm, unsigned long start,
 648                       unsigned long end, struct mempolicy *new_pol)
 649{
 650        struct vm_area_struct *next;
 651        struct vm_area_struct *prev;
 652        struct vm_area_struct *vma;
 653        int err = 0;
 654        pgoff_t pgoff;
 655        unsigned long vmstart;
 656        unsigned long vmend;
 657
 658        vma = find_vma(mm, start);
 659        if (!vma || vma->vm_start > start)
 660                return -EFAULT;
 661
 662        prev = vma->vm_prev;
 663        if (start > vma->vm_start)
 664                prev = vma;
 665
 666        for (; vma && vma->vm_start < end; prev = vma, vma = next) {
 667                next = vma->vm_next;
 668                vmstart = max(start, vma->vm_start);
 669                vmend   = min(end, vma->vm_end);
 670
 671                if (mpol_equal(vma_policy(vma), new_pol))
 672                        continue;
 673
 674                pgoff = vma->vm_pgoff +
 675                        ((vmstart - vma->vm_start) >> PAGE_SHIFT);
 676                prev = vma_merge(mm, prev, vmstart, vmend, vma->vm_flags,
 677                                  vma->anon_vma, vma->vm_file, pgoff,
 678                                  new_pol);
 679                if (prev) {
 680                        vma = prev;
 681                        next = vma->vm_next;
 682                        continue;
 683                }
 684                if (vma->vm_start != vmstart) {
 685                        err = split_vma(vma->vm_mm, vma, vmstart, 1);
 686                        if (err)
 687                                goto out;
 688                }
 689                if (vma->vm_end != vmend) {
 690                        err = split_vma(vma->vm_mm, vma, vmend, 0);
 691                        if (err)
 692                                goto out;
 693                }
 694                err = vma_replace_policy(vma, new_pol);
 695                if (err)
 696                        goto out;
 697        }
 698
 699 out:
 700        return err;
 701}
 702
 703/*
 704 * Update task->flags PF_MEMPOLICY bit: set iff non-default
 705 * mempolicy.  Allows more rapid checking of this (combined perhaps
 706 * with other PF_* flag bits) on memory allocation hot code paths.
 707 *
 708 * If called from outside this file, the task 'p' should -only- be
 709 * a newly forked child not yet visible on the task list, because
 710 * manipulating the task flags of a visible task is not safe.
 711 *
 712 * The above limitation is why this routine has the funny name
 713 * mpol_fix_fork_child_flag().
 714 *
 715 * It is also safe to call this with a task pointer of current,
 716 * which the static wrapper mpol_set_task_struct_flag() does,
 717 * for use within this file.
 718 */
 719
 720void mpol_fix_fork_child_flag(struct task_struct *p)
 721{
 722        if (p->mempolicy)
 723                p->flags |= PF_MEMPOLICY;
 724        else
 725                p->flags &= ~PF_MEMPOLICY;
 726}
 727
 728static void mpol_set_task_struct_flag(void)
 729{
 730        mpol_fix_fork_child_flag(current);
 731}
 732
 733/* Set the process memory policy */
 734static long do_set_mempolicy(unsigned short mode, unsigned short flags,
 735                             nodemask_t *nodes)
 736{
 737        struct mempolicy *new, *old;
 738        struct mm_struct *mm = current->mm;
 739        NODEMASK_SCRATCH(scratch);
 740        int ret;
 741
 742        if (!scratch)
 743                return -ENOMEM;
 744
 745        new = mpol_new(mode, flags, nodes);
 746        if (IS_ERR(new)) {
 747                ret = PTR_ERR(new);
 748                goto out;
 749        }
 750        /*
 751         * prevent changing our mempolicy while show_numa_maps()
 752         * is using it.
 753         * Note:  do_set_mempolicy() can be called at init time
 754         * with no 'mm'.
 755         */
 756        if (mm)
 757                down_write(&mm->mmap_sem);
 758        task_lock(current);
 759        ret = mpol_set_nodemask(new, nodes, scratch);
 760        if (ret) {
 761                task_unlock(current);
 762                if (mm)
 763                        up_write(&mm->mmap_sem);
 764                mpol_put(new);
 765                goto out;
 766        }
 767        old = current->mempolicy;
 768        current->mempolicy = new;
 769        mpol_set_task_struct_flag();
 770        if (new && new->mode == MPOL_INTERLEAVE &&
 771            nodes_weight(new->v.nodes))
 772                current->il_next = first_node(new->v.nodes);
 773        task_unlock(current);
 774        if (mm)
 775                up_write(&mm->mmap_sem);
 776
 777        mpol_put(old);
 778        ret = 0;
 779out:
 780        NODEMASK_SCRATCH_FREE(scratch);
 781        return ret;
 782}
 783
 784/*
 785 * Return nodemask for policy for get_mempolicy() query
 786 *
 787 * Called with task's alloc_lock held
 788 */
 789static void get_policy_nodemask(struct mempolicy *p, nodemask_t *nodes)
 790{
 791        nodes_clear(*nodes);
 792        if (p == &default_policy)
 793                return;
 794
 795        switch (p->mode) {
 796        case MPOL_BIND:
 797                /* Fall through */
 798        case MPOL_INTERLEAVE:
 799                *nodes = p->v.nodes;
 800                break;
 801        case MPOL_PREFERRED:
 802                if (!(p->flags & MPOL_F_LOCAL))
 803                        node_set(p->v.preferred_node, *nodes);
 804                /* else return empty node mask for local allocation */
 805                break;
 806        default:
 807                BUG();
 808        }
 809}
 810
 811static int lookup_node(struct mm_struct *mm, unsigned long addr)
 812{
 813        struct page *p;
 814        int err;
 815
 816        err = get_user_pages(current, mm, addr & PAGE_MASK, 1, 0, 0, &p, NULL);
 817        if (err >= 0) {
 818                err = page_to_nid(p);
 819                put_page(p);
 820        }
 821        return err;
 822}
 823
 824/* Retrieve NUMA policy */
 825static long do_get_mempolicy(int *policy, nodemask_t *nmask,
 826                             unsigned long addr, unsigned long flags)
 827{
 828        int err;
 829        struct mm_struct *mm = current->mm;
 830        struct vm_area_struct *vma = NULL;
 831        struct mempolicy *pol = current->mempolicy;
 832
 833        if (flags &
 834                ~(unsigned long)(MPOL_F_NODE|MPOL_F_ADDR|MPOL_F_MEMS_ALLOWED))
 835                return -EINVAL;
 836
 837        if (flags & MPOL_F_MEMS_ALLOWED) {
 838                if (flags & (MPOL_F_NODE|MPOL_F_ADDR))
 839                        return -EINVAL;
 840                *policy = 0;    /* just so it's initialized */
 841                task_lock(current);
 842                *nmask  = cpuset_current_mems_allowed;
 843                task_unlock(current);
 844                return 0;
 845        }
 846
 847        if (flags & MPOL_F_ADDR) {
 848                /*
 849                 * Do NOT fall back to task policy if the
 850                 * vma/shared policy at addr is NULL.  We
 851                 * want to return MPOL_DEFAULT in this case.
 852                 */
 853                down_read(&mm->mmap_sem);
 854                vma = find_vma_intersection(mm, addr, addr+1);
 855                if (!vma) {
 856                        up_read(&mm->mmap_sem);
 857                        return -EFAULT;
 858                }
 859                if (vma->vm_ops && vma->vm_ops->get_policy)
 860                        pol = vma->vm_ops->get_policy(vma, addr);
 861                else
 862                        pol = vma->vm_policy;
 863        } else if (addr)
 864                return -EINVAL;
 865
 866        if (!pol)
 867                pol = &default_policy;  /* indicates default behavior */
 868
 869        if (flags & MPOL_F_NODE) {
 870                if (flags & MPOL_F_ADDR) {
 871                        err = lookup_node(mm, addr);
 872                        if (err < 0)
 873                                goto out;
 874                        *policy = err;
 875                } else if (pol == current->mempolicy &&
 876                                pol->mode == MPOL_INTERLEAVE) {
 877                        *policy = current->il_next;
 878                } else {
 879                        err = -EINVAL;
 880                        goto out;
 881                }
 882        } else {
 883                *policy = pol == &default_policy ? MPOL_DEFAULT :
 884                                                pol->mode;
 885                /*
 886                 * Internal mempolicy flags must be masked off before exposing
 887                 * the policy to userspace.
 888                 */
 889                *policy |= (pol->flags & MPOL_MODE_FLAGS);
 890        }
 891
 892        if (vma) {
 893                up_read(&current->mm->mmap_sem);
 894                vma = NULL;
 895        }
 896
 897        err = 0;
 898        if (nmask) {
 899                if (mpol_store_user_nodemask(pol)) {
 900                        *nmask = pol->w.user_nodemask;
 901                } else {
 902                        task_lock(current);
 903                        get_policy_nodemask(pol, nmask);
 904                        task_unlock(current);
 905                }
 906        }
 907
 908 out:
 909        mpol_cond_put(pol);
 910        if (vma)
 911                up_read(&current->mm->mmap_sem);
 912        return err;
 913}
 914
 915#ifdef CONFIG_MIGRATION
 916/*
 917 * page migration
 918 */
 919static void migrate_page_add(struct page *page, struct list_head *pagelist,
 920                                unsigned long flags)
 921{
 922        /*
 923         * Avoid migrating a page that is shared with others.
 924         */
 925        if ((flags & MPOL_MF_MOVE_ALL) || page_mapcount(page) == 1) {
 926                if (!isolate_lru_page(page)) {
 927                        list_add_tail(&page->lru, pagelist);
 928                        inc_zone_page_state(page, NR_ISOLATED_ANON +
 929                                            page_is_file_cache(page));
 930                }
 931        }
 932}
 933
 934static struct page *new_node_page(struct page *page, unsigned long node, int **x)
 935{
 936        return alloc_pages_exact_node(node, GFP_HIGHUSER_MOVABLE, 0);
 937}
 938
 939/*
 940 * Migrate pages from one node to a target node.
 941 * Returns error or the number of pages not migrated.
 942 */
 943static int migrate_to_node(struct mm_struct *mm, int source, int dest,
 944                           int flags)
 945{
 946        nodemask_t nmask;
 947        LIST_HEAD(pagelist);
 948        int err = 0;
 949
 950        nodes_clear(nmask);
 951        node_set(source, nmask);
 952
 953        /*
 954         * This does not "check" the range but isolates all pages that
 955         * need migration.  Between passing in the full user address
 956         * space range and MPOL_MF_DISCONTIG_OK, this call can not fail.
 957         */
 958        VM_BUG_ON(!(flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)));
 959        check_range(mm, mm->mmap->vm_start, mm->task_size, &nmask,
 960                        flags | MPOL_MF_DISCONTIG_OK, &pagelist);
 961
 962        if (!list_empty(&pagelist)) {
 963                err = migrate_pages(&pagelist, new_node_page, dest,
 964                                                        false, MIGRATE_SYNC);
 965                if (err)
 966                        putback_lru_pages(&pagelist);
 967        }
 968
 969        return err;
 970}
 971
 972/*
 973 * Move pages between the two nodesets so as to preserve the physical
 974 * layout as much as possible.
 975 *
 976 * Returns the number of page that could not be moved.
 977 */
 978int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
 979                     const nodemask_t *to, int flags)
 980{
 981        int busy = 0;
 982        int err;
 983        nodemask_t tmp;
 984
 985        err = migrate_prep();
 986        if (err)
 987                return err;
 988
 989        down_read(&mm->mmap_sem);
 990
 991        err = migrate_vmas(mm, from, to, flags);
 992        if (err)
 993                goto out;
 994
 995        /*
 996         * Find a 'source' bit set in 'tmp' whose corresponding 'dest'
 997         * bit in 'to' is not also set in 'tmp'.  Clear the found 'source'
 998         * bit in 'tmp', and return that <source, dest> pair for migration.
 999         * The pair of nodemasks 'to' and 'from' define the map.
1000         *
1001         * If no pair of bits is found that way, fallback to picking some
1002         * pair of 'source' and 'dest' bits that are not the same.  If the
1003         * 'source' and 'dest' bits are the same, this represents a node
1004         * that will be migrating to itself, so no pages need move.
1005         *
1006         * If no bits are left in 'tmp', or if all remaining bits left
1007         * in 'tmp' correspond to the same bit in 'to', return false
1008         * (nothing left to migrate).
1009         *
1010         * This lets us pick a pair of nodes to migrate between, such that
1011         * if possible the dest node is not already occupied by some other
1012         * source node, minimizing the risk of overloading the memory on a
1013         * node that would happen if we migrated incoming memory to a node
1014         * before migrating outgoing memory source that same node.
1015         *
1016         * A single scan of tmp is sufficient.  As we go, we remember the
1017         * most recent <s, d> pair that moved (s != d).  If we find a pair
1018         * that not only moved, but what's better, moved to an empty slot
1019         * (d is not set in tmp), then we break out then, with that pair.
1020         * Otherwise when we finish scanning from_tmp, we at least have the
1021         * most recent <s, d> pair that moved.  If we get all the way through
1022         * the scan of tmp without finding any node that moved, much less
1023         * moved to an empty node, then there is nothing left worth migrating.
1024         */
1025
1026        tmp = *from;
1027        while (!nodes_empty(tmp)) {
1028                int s,d;
1029                int source = -1;
1030                int dest = 0;
1031
1032                for_each_node_mask(s, tmp) {
1033
1034                        /*
1035                         * do_migrate_pages() tries to maintain the relative
1036                         * node relationship of the pages established between
1037                         * threads and memory areas.
1038                         *
1039                         * However if the number of source nodes is not equal to
1040                         * the number of destination nodes we can not preserve
1041                         * this node relative relationship.  In that case, skip
1042                         * copying memory from a node that is in the destination
1043                         * mask.
1044                         *
1045                         * Example: [2,3,4] -> [3,4,5] moves everything.
1046                         *          [0-7] - > [3,4,5] moves only 0,1,2,6,7.
1047                         */
1048
1049                        if ((nodes_weight(*from) != nodes_weight(*to)) &&
1050                                                (node_isset(s, *to)))
1051                                continue;
1052
1053                        d = node_remap(s, *from, *to);
1054                        if (s == d)
1055                                continue;
1056
1057                        source = s;     /* Node moved. Memorize */
1058                        dest = d;
1059
1060                        /* dest not in remaining from nodes? */
1061                        if (!node_isset(dest, tmp))
1062                                break;
1063                }
1064                if (source == -1)
1065                        break;
1066
1067                node_clear(source, tmp);
1068                err = migrate_to_node(mm, source, dest, flags);
1069                if (err > 0)
1070                        busy += err;
1071                if (err < 0)
1072                        break;
1073        }
1074out:
1075        up_read(&mm->mmap_sem);
1076        if (err < 0)
1077                return err;
1078        return busy;
1079
1080}
1081
1082/*
1083 * Allocate a new page for page migration based on vma policy.
1084 * Start assuming that page is mapped by vma pointed to by @private.
1085 * Search forward from there, if not.  N.B., this assumes that the
1086 * list of pages handed to migrate_pages()--which is how we get here--
1087 * is in virtual address order.
1088 */
1089static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1090{
1091        struct vm_area_struct *vma = (struct vm_area_struct *)private;
1092        unsigned long uninitialized_var(address);
1093
1094        while (vma) {
1095                address = page_address_in_vma(page, vma);
1096                if (address != -EFAULT)
1097                        break;
1098                vma = vma->vm_next;
1099        }
1100
1101        /*
1102         * if !vma, alloc_page_vma() will use task or system default policy
1103         */
1104        return alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
1105}
1106#else
1107
1108static void migrate_page_add(struct page *page, struct list_head *pagelist,
1109                                unsigned long flags)
1110{
1111}
1112
1113int do_migrate_pages(struct mm_struct *mm, const nodemask_t *from,
1114                     const nodemask_t *to, int flags)
1115{
1116        return -ENOSYS;
1117}
1118
1119static struct page *new_vma_page(struct page *page, unsigned long private, int **x)
1120{
1121        return NULL;
1122}
1123#endif
1124
1125static long do_mbind(unsigned long start, unsigned long len,
1126                     unsigned short mode, unsigned short mode_flags,
1127                     nodemask_t *nmask, unsigned long flags)
1128{
1129        struct vm_area_struct *vma;
1130        struct mm_struct *mm = current->mm;
1131        struct mempolicy *new;
1132        unsigned long end;
1133        int err;
1134        LIST_HEAD(pagelist);
1135
1136        if (flags & ~(unsigned long)(MPOL_MF_STRICT |
1137                                     MPOL_MF_MOVE | MPOL_MF_MOVE_ALL))
1138                return -EINVAL;
1139        if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE))
1140                return -EPERM;
1141
1142        if (start & ~PAGE_MASK)
1143                return -EINVAL;
1144
1145        if (mode == MPOL_DEFAULT)
1146                flags &= ~MPOL_MF_STRICT;
1147
1148        len = (len + PAGE_SIZE - 1) & PAGE_MASK;
1149        end = start + len;
1150
1151        if (end < start)
1152                return -EINVAL;
1153        if (end == start)
1154                return 0;
1155
1156        new = mpol_new(mode, mode_flags, nmask);
1157        if (IS_ERR(new))
1158                return PTR_ERR(new);
1159
1160        /*
1161         * If we are using the default policy then operation
1162         * on discontinuous address spaces is okay after all
1163         */
1164        if (!new)
1165                flags |= MPOL_MF_DISCONTIG_OK;
1166
1167        pr_debug("mbind %lx-%lx mode:%d flags:%d nodes:%lx\n",
1168                 start, start + len, mode, mode_flags,
1169                 nmask ? nodes_addr(*nmask)[0] : -1);
1170
1171        if (flags & (MPOL_MF_MOVE | MPOL_MF_MOVE_ALL)) {
1172
1173                err = migrate_prep();
1174                if (err)
1175                        goto mpol_out;
1176        }
1177        {
1178                NODEMASK_SCRATCH(scratch);
1179                if (scratch) {
1180                        down_write(&mm->mmap_sem);
1181                        task_lock(current);
1182                        err = mpol_set_nodemask(new, nmask, scratch);
1183                        task_unlock(current);
1184                        if (err)
1185                                up_write(&mm->mmap_sem);
1186                } else
1187                        err = -ENOMEM;
1188                NODEMASK_SCRATCH_FREE(scratch);
1189        }
1190        if (err)
1191                goto mpol_out;
1192
1193        vma = check_range(mm, start, end, nmask,
1194                          flags | MPOL_MF_INVERT, &pagelist);
1195
1196        err = PTR_ERR(vma);
1197        if (!IS_ERR(vma)) {
1198                int nr_failed = 0;
1199
1200                err = mbind_range(mm, start, end, new);
1201
1202                if (!list_empty(&pagelist)) {
1203                        nr_failed = migrate_pages(&pagelist, new_vma_page,
1204                                                (unsigned long)vma,
1205                                                false, MIGRATE_SYNC);
1206                        if (nr_failed)
1207                                putback_lru_pages(&pagelist);
1208                }
1209
1210                if (!err && nr_failed && (flags & MPOL_MF_STRICT))
1211                        err = -EIO;
1212        } else
1213                putback_lru_pages(&pagelist);
1214
1215        up_write(&mm->mmap_sem);
1216 mpol_out:
1217        mpol_put(new);
1218        return err;
1219}
1220
1221/*
1222 * User space interface with variable sized bitmaps for nodelists.
1223 */
1224
1225/* Copy a node mask from user space. */
1226static int get_nodes(nodemask_t *nodes, const unsigned long __user *nmask,
1227                     unsigned long maxnode)
1228{
1229        unsigned long k;
1230        unsigned long nlongs;
1231        unsigned long endmask;
1232
1233        --maxnode;
1234        nodes_clear(*nodes);
1235        if (maxnode == 0 || !nmask)
1236                return 0;
1237        if (maxnode > PAGE_SIZE*BITS_PER_BYTE)
1238                return -EINVAL;
1239
1240        nlongs = BITS_TO_LONGS(maxnode);
1241        if ((maxnode % BITS_PER_LONG) == 0)
1242                endmask = ~0UL;
1243        else
1244                endmask = (1UL << (maxnode % BITS_PER_LONG)) - 1;
1245
1246        /* When the user specified more nodes than supported just check
1247           if the non supported part is all zero. */
1248        if (nlongs > BITS_TO_LONGS(MAX_NUMNODES)) {
1249                if (nlongs > PAGE_SIZE/sizeof(long))
1250                        return -EINVAL;
1251                for (k = BITS_TO_LONGS(MAX_NUMNODES); k < nlongs; k++) {
1252                        unsigned long t;
1253                        if (get_user(t, nmask + k))
1254                                return -EFAULT;
1255                        if (k == nlongs - 1) {
1256                                if (t & endmask)
1257                                        return -EINVAL;
1258                        } else if (t)
1259                                return -EINVAL;
1260                }
1261                nlongs = BITS_TO_LONGS(MAX_NUMNODES);
1262                endmask = ~0UL;
1263        }
1264
1265        if (copy_from_user(nodes_addr(*nodes), nmask, nlongs*sizeof(unsigned long)))
1266                return -EFAULT;
1267        nodes_addr(*nodes)[nlongs-1] &= endmask;
1268        return 0;
1269}
1270
1271/* Copy a kernel node mask to user space */
1272static int copy_nodes_to_user(unsigned long __user *mask, unsigned long maxnode,
1273                              nodemask_t *nodes)
1274{
1275        unsigned long copy = ALIGN(maxnode-1, 64) / 8;
1276        const int nbytes = BITS_TO_LONGS(MAX_NUMNODES) * sizeof(long);
1277
1278        if (copy > nbytes) {
1279                if (copy > PAGE_SIZE)
1280                        return -EINVAL;
1281                if (clear_user((char __user *)mask + nbytes, copy - nbytes))
1282                        return -EFAULT;
1283                copy = nbytes;
1284        }
1285        return copy_to_user(mask, nodes_addr(*nodes), copy) ? -EFAULT : 0;
1286}
1287
1288SYSCALL_DEFINE6(mbind, unsigned long, start, unsigned long, len,
1289                unsigned long, mode, unsigned long __user *, nmask,
1290                unsigned long, maxnode, unsigned, flags)
1291{
1292        nodemask_t nodes;
1293        int err;
1294        unsigned short mode_flags;
1295
1296        mode_flags = mode & MPOL_MODE_FLAGS;
1297        mode &= ~MPOL_MODE_FLAGS;
1298        if (mode >= MPOL_MAX)
1299                return -EINVAL;
1300        if ((mode_flags & MPOL_F_STATIC_NODES) &&
1301            (mode_flags & MPOL_F_RELATIVE_NODES))
1302                return -EINVAL;
1303        err = get_nodes(&nodes, nmask, maxnode);
1304        if (err)
1305                return err;
1306        return do_mbind(start, len, mode, mode_flags, &nodes, flags);
1307}
1308
1309/* Set the process memory policy */
1310SYSCALL_DEFINE3(set_mempolicy, int, mode, unsigned long __user *, nmask,
1311                unsigned long, maxnode)
1312{
1313        int err;
1314        nodemask_t nodes;
1315        unsigned short flags;
1316
1317        flags = mode & MPOL_MODE_FLAGS;
1318        mode &= ~MPOL_MODE_FLAGS;
1319        if ((unsigned int)mode >= MPOL_MAX)
1320                return -EINVAL;
1321        if ((flags & MPOL_F_STATIC_NODES) && (flags & MPOL_F_RELATIVE_NODES))
1322                return -EINVAL;
1323        err = get_nodes(&nodes, nmask, maxnode);
1324        if (err)
1325                return err;
1326        return do_set_mempolicy(mode, flags, &nodes);
1327}
1328
1329SYSCALL_DEFINE4(migrate_pages, pid_t, pid, unsigned long, maxnode,
1330                const unsigned long __user *, old_nodes,
1331                const unsigned long __user *, new_nodes)
1332{
1333        const struct cred *cred = current_cred(), *tcred;
1334        struct mm_struct *mm = NULL;
1335        struct task_struct *task;
1336        nodemask_t task_nodes;
1337        int err;
1338        nodemask_t *old;
1339        nodemask_t *new;
1340        NODEMASK_SCRATCH(scratch);
1341
1342        if (!scratch)
1343                return -ENOMEM;
1344
1345        old = &scratch->mask1;
1346        new = &scratch->mask2;
1347
1348        err = get_nodes(old, old_nodes, maxnode);
1349        if (err)
1350                goto out;
1351
1352        err = get_nodes(new, new_nodes, maxnode);
1353        if (err)
1354                goto out;
1355
1356        /* Find the mm_struct */
1357        rcu_read_lock();
1358        task = pid ? find_task_by_vpid(pid) : current;
1359        if (!task) {
1360                rcu_read_unlock();
1361                err = -ESRCH;
1362                goto out;
1363        }
1364        get_task_struct(task);
1365
1366        err = -EINVAL;
1367
1368        /*
1369         * Check if this process has the right to modify the specified
1370         * process. The right exists if the process has administrative
1371         * capabilities, superuser privileges or the same
1372         * userid as the target process.
1373         */
1374        tcred = __task_cred(task);
1375        if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) &&
1376            !uid_eq(cred->uid,  tcred->suid) && !uid_eq(cred->uid,  tcred->uid) &&
1377            !capable(CAP_SYS_NICE)) {
1378                rcu_read_unlock();
1379                err = -EPERM;
1380                goto out_put;
1381        }
1382        rcu_read_unlock();
1383
1384        task_nodes = cpuset_mems_allowed(task);
1385        /* Is the user allowed to access the target nodes? */
1386        if (!nodes_subset(*new, task_nodes) && !capable(CAP_SYS_NICE)) {
1387                err = -EPERM;
1388                goto out_put;
1389        }
1390
1391        if (!nodes_subset(*new, node_states[N_HIGH_MEMORY])) {
1392                err = -EINVAL;
1393                goto out_put;
1394        }
1395
1396        err = security_task_movememory(task);
1397        if (err)
1398                goto out_put;
1399
1400        mm = get_task_mm(task);
1401        put_task_struct(task);
1402
1403        if (!mm) {
1404                err = -EINVAL;
1405                goto out;
1406        }
1407
1408        err = do_migrate_pages(mm, old, new,
1409                capable(CAP_SYS_NICE) ? MPOL_MF_MOVE_ALL : MPOL_MF_MOVE);
1410
1411        mmput(mm);
1412out:
1413        NODEMASK_SCRATCH_FREE(scratch);
1414
1415        return err;
1416
1417out_put:
1418        put_task_struct(task);
1419        goto out;
1420
1421}
1422
1423
1424/* Retrieve NUMA policy */
1425SYSCALL_DEFINE5(get_mempolicy, int __user *, policy,
1426                unsigned long __user *, nmask, unsigned long, maxnode,
1427                unsigned long, addr, unsigned long, flags)
1428{
1429        int err;
1430        int uninitialized_var(pval);
1431        nodemask_t nodes;
1432
1433        if (nmask != NULL && maxnode < MAX_NUMNODES)
1434                return -EINVAL;
1435
1436        err = do_get_mempolicy(&pval, &nodes, addr, flags);
1437
1438        if (err)
1439                return err;
1440
1441        if (policy && put_user(pval, policy))
1442                return -EFAULT;
1443
1444        if (nmask)
1445                err = copy_nodes_to_user(nmask, maxnode, &nodes);
1446
1447        return err;
1448}
1449
1450#ifdef CONFIG_COMPAT
1451
1452asmlinkage long compat_sys_get_mempolicy(int __user *policy,
1453                                     compat_ulong_t __user *nmask,
1454                                     compat_ulong_t maxnode,
1455                                     compat_ulong_t addr, compat_ulong_t flags)
1456{
1457        long err;
1458        unsigned long __user *nm = NULL;
1459        unsigned long nr_bits, alloc_size;
1460        DECLARE_BITMAP(bm, MAX_NUMNODES);
1461
1462        nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1463        alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1464
1465        if (nmask)
1466                nm = compat_alloc_user_space(alloc_size);
1467
1468        err = sys_get_mempolicy(policy, nm, nr_bits+1, addr, flags);
1469
1470        if (!err && nmask) {
1471                unsigned long copy_size;
1472                copy_size = min_t(unsigned long, sizeof(bm), alloc_size);
1473                err = copy_from_user(bm, nm, copy_size);
1474                /* ensure entire bitmap is zeroed */
1475                err |= clear_user(nmask, ALIGN(maxnode-1, 8) / 8);
1476                err |= compat_put_bitmap(nmask, bm, nr_bits);
1477        }
1478
1479        return err;
1480}
1481
1482asmlinkage long compat_sys_set_mempolicy(int mode, compat_ulong_t __user *nmask,
1483                                     compat_ulong_t maxnode)
1484{
1485        long err = 0;
1486        unsigned long __user *nm = NULL;
1487        unsigned long nr_bits, alloc_size;
1488        DECLARE_BITMAP(bm, MAX_NUMNODES);
1489
1490        nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1491        alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1492
1493        if (nmask) {
1494                err = compat_get_bitmap(bm, nmask, nr_bits);
1495                nm = compat_alloc_user_space(alloc_size);
1496                err |= copy_to_user(nm, bm, alloc_size);
1497        }
1498
1499        if (err)
1500                return -EFAULT;
1501
1502        return sys_set_mempolicy(mode, nm, nr_bits+1);
1503}
1504
1505asmlinkage long compat_sys_mbind(compat_ulong_t start, compat_ulong_t len,
1506                             compat_ulong_t mode, compat_ulong_t __user *nmask,
1507                             compat_ulong_t maxnode, compat_ulong_t flags)
1508{
1509        long err = 0;
1510        unsigned long __user *nm = NULL;
1511        unsigned long nr_bits, alloc_size;
1512        nodemask_t bm;
1513
1514        nr_bits = min_t(unsigned long, maxnode-1, MAX_NUMNODES);
1515        alloc_size = ALIGN(nr_bits, BITS_PER_LONG) / 8;
1516
1517        if (nmask) {
1518                err = compat_get_bitmap(nodes_addr(bm), nmask, nr_bits);
1519                nm = compat_alloc_user_space(alloc_size);
1520                err |= copy_to_user(nm, nodes_addr(bm), alloc_size);
1521        }
1522
1523        if (err)
1524                return -EFAULT;
1525
1526        return sys_mbind(start, len, mode, nm, nr_bits+1, flags);
1527}
1528
1529#endif
1530
1531/*
1532 * get_vma_policy(@task, @vma, @addr)
1533 * @task - task for fallback if vma policy == default
1534 * @vma   - virtual memory area whose policy is sought
1535 * @addr  - address in @vma for shared policy lookup
1536 *
1537 * Returns effective policy for a VMA at specified address.
1538 * Falls back to @task or system default policy, as necessary.
1539 * Current or other task's task mempolicy and non-shared vma policies must be
1540 * protected by task_lock(task) by the caller.
1541 * Shared policies [those marked as MPOL_F_SHARED] require an extra reference
1542 * count--added by the get_policy() vm_op, as appropriate--to protect against
1543 * freeing by another task.  It is the caller's responsibility to free the
1544 * extra reference for shared policies.
1545 */
1546struct mempolicy *get_vma_policy(struct task_struct *task,
1547                struct vm_area_struct *vma, unsigned long addr)
1548{
1549        struct mempolicy *pol = task->mempolicy;
1550
1551        if (vma) {
1552                if (vma->vm_ops && vma->vm_ops->get_policy) {
1553                        struct mempolicy *vpol = vma->vm_ops->get_policy(vma,
1554                                                                        addr);
1555                        if (vpol)
1556                                pol = vpol;
1557                } else if (vma->vm_policy) {
1558                        pol = vma->vm_policy;
1559
1560                        /*
1561                         * shmem_alloc_page() passes MPOL_F_SHARED policy with
1562                         * a pseudo vma whose vma->vm_ops=NULL. Take a reference
1563                         * count on these policies which will be dropped by
1564                         * mpol_cond_put() later
1565                         */
1566                        if (mpol_needs_cond_ref(pol))
1567                                mpol_get(pol);
1568                }
1569        }
1570        if (!pol)
1571                pol = &default_policy;
1572        return pol;
1573}
1574
1575/*
1576 * Return a nodemask representing a mempolicy for filtering nodes for
1577 * page allocation
1578 */
1579static nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
1580{
1581        /* Lower zones don't get a nodemask applied for MPOL_BIND */
1582        if (unlikely(policy->mode == MPOL_BIND) &&
1583                        gfp_zone(gfp) >= policy_zone &&
1584                        cpuset_nodemask_valid_mems_allowed(&policy->v.nodes))
1585                return &policy->v.nodes;
1586
1587        return NULL;
1588}
1589
1590/* Return a zonelist indicated by gfp for node representing a mempolicy */
1591static struct zonelist *policy_zonelist(gfp_t gfp, struct mempolicy *policy,
1592        int nd)
1593{
1594        switch (policy->mode) {
1595        case MPOL_PREFERRED:
1596                if (!(policy->flags & MPOL_F_LOCAL))
1597                        nd = policy->v.preferred_node;
1598                break;
1599        case MPOL_BIND:
1600                /*
1601                 * Normally, MPOL_BIND allocations are node-local within the
1602                 * allowed nodemask.  However, if __GFP_THISNODE is set and the
1603                 * current node isn't part of the mask, we use the zonelist for
1604                 * the first node in the mask instead.
1605                 */
1606                if (unlikely(gfp & __GFP_THISNODE) &&
1607                                unlikely(!node_isset(nd, policy->v.nodes)))
1608                        nd = first_node(policy->v.nodes);
1609                break;
1610        default:
1611                BUG();
1612        }
1613        return node_zonelist(nd, gfp);
1614}
1615
1616/* Do dynamic interleaving for a process */
1617static unsigned interleave_nodes(struct mempolicy *policy)
1618{
1619        unsigned nid, next;
1620        struct task_struct *me = current;
1621
1622        nid = me->il_next;
1623        next = next_node(nid, policy->v.nodes);
1624        if (next >= MAX_NUMNODES)
1625                next = first_node(policy->v.nodes);
1626        if (next < MAX_NUMNODES)
1627                me->il_next = next;
1628        return nid;
1629}
1630
1631/*
1632 * Depending on the memory policy provide a node from which to allocate the
1633 * next slab entry.
1634 * @policy must be protected by freeing by the caller.  If @policy is
1635 * the current task's mempolicy, this protection is implicit, as only the
1636 * task can change it's policy.  The system default policy requires no
1637 * such protection.
1638 */
1639unsigned slab_node(void)
1640{
1641        struct mempolicy *policy;
1642
1643        if (in_interrupt())
1644                return numa_node_id();
1645
1646        policy = current->mempolicy;
1647        if (!policy || policy->flags & MPOL_F_LOCAL)
1648                return numa_node_id();
1649
1650        switch (policy->mode) {
1651        case MPOL_PREFERRED:
1652                /*
1653                 * handled MPOL_F_LOCAL above
1654                 */
1655                return policy->v.preferred_node;
1656
1657        case MPOL_INTERLEAVE:
1658                return interleave_nodes(policy);
1659
1660        case MPOL_BIND: {
1661                /*
1662                 * Follow bind policy behavior and start allocation at the
1663                 * first node.
1664                 */
1665                struct zonelist *zonelist;
1666                struct zone *zone;
1667                enum zone_type highest_zoneidx = gfp_zone(GFP_KERNEL);
1668                zonelist = &NODE_DATA(numa_node_id())->node_zonelists[0];
1669                (void)first_zones_zonelist(zonelist, highest_zoneidx,
1670                                                        &policy->v.nodes,
1671                                                        &zone);
1672                return zone ? zone->node : numa_node_id();
1673        }
1674
1675        default:
1676                BUG();
1677        }
1678}
1679
1680/* Do static interleaving for a VMA with known offset. */
1681static unsigned offset_il_node(struct mempolicy *pol,
1682                struct vm_area_struct *vma, unsigned long off)
1683{
1684        unsigned nnodes = nodes_weight(pol->v.nodes);
1685        unsigned target;
1686        int c;
1687        int nid = -1;
1688
1689        if (!nnodes)
1690                return numa_node_id();
1691        target = (unsigned int)off % nnodes;
1692        c = 0;
1693        do {
1694                nid = next_node(nid, pol->v.nodes);
1695                c++;
1696        } while (c <= target);
1697        return nid;
1698}
1699
1700/* Determine a node number for interleave */
1701static inline unsigned interleave_nid(struct mempolicy *pol,
1702                 struct vm_area_struct *vma, unsigned long addr, int shift)
1703{
1704        if (vma) {
1705                unsigned long off;
1706
1707                /*
1708                 * for small pages, there is no difference between
1709                 * shift and PAGE_SHIFT, so the bit-shift is safe.
1710                 * for huge pages, since vm_pgoff is in units of small
1711                 * pages, we need to shift off the always 0 bits to get
1712                 * a useful offset.
1713                 */
1714                BUG_ON(shift < PAGE_SHIFT);
1715                off = vma->vm_pgoff >> (shift - PAGE_SHIFT);
1716                off += (addr - vma->vm_start) >> shift;
1717                return offset_il_node(pol, vma, off);
1718        } else
1719                return interleave_nodes(pol);
1720}
1721
1722/*
1723 * Return the bit number of a random bit set in the nodemask.
1724 * (returns -1 if nodemask is empty)
1725 */
1726int node_random(const nodemask_t *maskp)
1727{
1728        int w, bit = -1;
1729
1730        w = nodes_weight(*maskp);
1731        if (w)
1732                bit = bitmap_ord_to_pos(maskp->bits,
1733                        get_random_int() % w, MAX_NUMNODES);
1734        return bit;
1735}
1736
1737#ifdef CONFIG_HUGETLBFS
1738/*
1739 * huge_zonelist(@vma, @addr, @gfp_flags, @mpol)
1740 * @vma = virtual memory area whose policy is sought
1741 * @addr = address in @vma for shared policy lookup and interleave policy
1742 * @gfp_flags = for requested zone
1743 * @mpol = pointer to mempolicy pointer for reference counted mempolicy
1744 * @nodemask = pointer to nodemask pointer for MPOL_BIND nodemask
1745 *
1746 * Returns a zonelist suitable for a huge page allocation and a pointer
1747 * to the struct mempolicy for conditional unref after allocation.
1748 * If the effective policy is 'BIND, returns a pointer to the mempolicy's
1749 * @nodemask for filtering the zonelist.
1750 *
1751 * Must be protected by get_mems_allowed()
1752 */
1753struct zonelist *huge_zonelist(struct vm_area_struct *vma, unsigned long addr,
1754                                gfp_t gfp_flags, struct mempolicy **mpol,
1755                                nodemask_t **nodemask)
1756{
1757        struct zonelist *zl;
1758
1759        *mpol = get_vma_policy(current, vma, addr);
1760        *nodemask = NULL;       /* assume !MPOL_BIND */
1761
1762        if (unlikely((*mpol)->mode == MPOL_INTERLEAVE)) {
1763                zl = node_zonelist(interleave_nid(*mpol, vma, addr,
1764                                huge_page_shift(hstate_vma(vma))), gfp_flags);
1765        } else {
1766                zl = policy_zonelist(gfp_flags, *mpol, numa_node_id());
1767                if ((*mpol)->mode == MPOL_BIND)
1768                        *nodemask = &(*mpol)->v.nodes;
1769        }
1770        return zl;
1771}
1772
1773/*
1774 * init_nodemask_of_mempolicy
1775 *
1776 * If the current task's mempolicy is "default" [NULL], return 'false'
1777 * to indicate default policy.  Otherwise, extract the policy nodemask
1778 * for 'bind' or 'interleave' policy into the argument nodemask, or
1779 * initialize the argument nodemask to contain the single node for
1780 * 'preferred' or 'local' policy and return 'true' to indicate presence
1781 * of non-default mempolicy.
1782 *
1783 * We don't bother with reference counting the mempolicy [mpol_get/put]
1784 * because the current task is examining it's own mempolicy and a task's
1785 * mempolicy is only ever changed by the task itself.
1786 *
1787 * N.B., it is the caller's responsibility to free a returned nodemask.
1788 */
1789bool init_nodemask_of_mempolicy(nodemask_t *mask)
1790{
1791        struct mempolicy *mempolicy;
1792        int nid;
1793
1794        if (!(mask && current->mempolicy))
1795                return false;
1796
1797        task_lock(current);
1798        mempolicy = current->mempolicy;
1799        switch (mempolicy->mode) {
1800        case MPOL_PREFERRED:
1801                if (mempolicy->flags & MPOL_F_LOCAL)
1802                        nid = numa_node_id();
1803                else
1804                        nid = mempolicy->v.preferred_node;
1805                init_nodemask_of_node(mask, nid);
1806                break;
1807
1808        case MPOL_BIND:
1809                /* Fall through */
1810        case MPOL_INTERLEAVE:
1811                *mask =  mempolicy->v.nodes;
1812                break;
1813
1814        default:
1815                BUG();
1816        }
1817        task_unlock(current);
1818
1819        return true;
1820}
1821#endif
1822
1823/*
1824 * mempolicy_nodemask_intersects
1825 *
1826 * If tsk's mempolicy is "default" [NULL], return 'true' to indicate default
1827 * policy.  Otherwise, check for intersection between mask and the policy
1828 * nodemask for 'bind' or 'interleave' policy.  For 'perferred' or 'local'
1829 * policy, always return true since it may allocate elsewhere on fallback.
1830 *
1831 * Takes task_lock(tsk) to prevent freeing of its mempolicy.
1832 */
1833bool mempolicy_nodemask_intersects(struct task_struct *tsk,
1834                                        const nodemask_t *mask)
1835{
1836        struct mempolicy *mempolicy;
1837        bool ret = true;
1838
1839        if (!mask)
1840                return ret;
1841        task_lock(tsk);
1842        mempolicy = tsk->mempolicy;
1843        if (!mempolicy)
1844                goto out;
1845
1846        switch (mempolicy->mode) {
1847        case MPOL_PREFERRED:
1848                /*
1849                 * MPOL_PREFERRED and MPOL_F_LOCAL are only preferred nodes to
1850                 * allocate from, they may fallback to other nodes when oom.
1851                 * Thus, it's possible for tsk to have allocated memory from
1852                 * nodes in mask.
1853                 */
1854                break;
1855        case MPOL_BIND:
1856        case MPOL_INTERLEAVE:
1857                ret = nodes_intersects(mempolicy->v.nodes, *mask);
1858                break;
1859        default:
1860                BUG();
1861        }
1862out:
1863        task_unlock(tsk);
1864        return ret;
1865}
1866
1867/* Allocate a page in interleaved policy.
1868   Own path because it needs to do special accounting. */
1869static struct page *alloc_page_interleave(gfp_t gfp, unsigned order,
1870                                        unsigned nid)
1871{
1872        struct zonelist *zl;
1873        struct page *page;
1874
1875        zl = node_zonelist(nid, gfp);
1876        page = __alloc_pages(gfp, order, zl);
1877        if (page && page_zone(page) == zonelist_zone(&zl->_zonerefs[0]))
1878                inc_zone_page_state(page, NUMA_INTERLEAVE_HIT);
1879        return page;
1880}
1881
1882/**
1883 *      alloc_pages_vma - Allocate a page for a VMA.
1884 *
1885 *      @gfp:
1886 *      %GFP_USER    user allocation.
1887 *      %GFP_KERNEL  kernel allocations,
1888 *      %GFP_HIGHMEM highmem/user allocations,
1889 *      %GFP_FS      allocation should not call back into a file system.
1890 *      %GFP_ATOMIC  don't sleep.
1891 *
1892 *      @order:Order of the GFP allocation.
1893 *      @vma:  Pointer to VMA or NULL if not available.
1894 *      @addr: Virtual Address of the allocation. Must be inside the VMA.
1895 *
1896 *      This function allocates a page from the kernel page pool and applies
1897 *      a NUMA policy associated with the VMA or the current process.
1898 *      When VMA is not NULL caller must hold down_read on the mmap_sem of the
1899 *      mm_struct of the VMA to prevent it from going away. Should be used for
1900 *      all allocations for pages that will be mapped into
1901 *      user space. Returns NULL when no page can be allocated.
1902 *
1903 *      Should be called with the mm_sem of the vma hold.
1904 */
1905struct page *
1906alloc_pages_vma(gfp_t gfp, int order, struct vm_area_struct *vma,
1907                unsigned long addr, int node)
1908{
1909        struct mempolicy *pol;
1910        struct zonelist *zl;
1911        struct page *page;
1912        unsigned int cpuset_mems_cookie;
1913
1914retry_cpuset:
1915        pol = get_vma_policy(current, vma, addr);
1916        cpuset_mems_cookie = get_mems_allowed();
1917
1918        if (unlikely(pol->mode == MPOL_INTERLEAVE)) {
1919                unsigned nid;
1920
1921                nid = interleave_nid(pol, vma, addr, PAGE_SHIFT + order);
1922                mpol_cond_put(pol);
1923                page = alloc_page_interleave(gfp, order, nid);
1924                if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1925                        goto retry_cpuset;
1926
1927                return page;
1928        }
1929        zl = policy_zonelist(gfp, pol, node);
1930        if (unlikely(mpol_needs_cond_ref(pol))) {
1931                /*
1932                 * slow path: ref counted shared policy
1933                 */
1934                struct page *page =  __alloc_pages_nodemask(gfp, order,
1935                                                zl, policy_nodemask(gfp, pol));
1936                __mpol_put(pol);
1937                if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1938                        goto retry_cpuset;
1939                return page;
1940        }
1941        /*
1942         * fast path:  default or task policy
1943         */
1944        page = __alloc_pages_nodemask(gfp, order, zl,
1945                                      policy_nodemask(gfp, pol));
1946        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1947                goto retry_cpuset;
1948        return page;
1949}
1950
1951/**
1952 *      alloc_pages_current - Allocate pages.
1953 *
1954 *      @gfp:
1955 *              %GFP_USER   user allocation,
1956 *              %GFP_KERNEL kernel allocation,
1957 *              %GFP_HIGHMEM highmem allocation,
1958 *              %GFP_FS     don't call back into a file system.
1959 *              %GFP_ATOMIC don't sleep.
1960 *      @order: Power of two of allocation size in pages. 0 is a single page.
1961 *
1962 *      Allocate a page from the kernel page pool.  When not in
1963 *      interrupt context and apply the current process NUMA policy.
1964 *      Returns NULL when no page can be allocated.
1965 *
1966 *      Don't call cpuset_update_task_memory_state() unless
1967 *      1) it's ok to take cpuset_sem (can WAIT), and
1968 *      2) allocating for current task (not interrupt).
1969 */
1970struct page *alloc_pages_current(gfp_t gfp, unsigned order)
1971{
1972        struct mempolicy *pol = current->mempolicy;
1973        struct page *page;
1974        unsigned int cpuset_mems_cookie;
1975
1976        if (!pol || in_interrupt() || (gfp & __GFP_THISNODE))
1977                pol = &default_policy;
1978
1979retry_cpuset:
1980        cpuset_mems_cookie = get_mems_allowed();
1981
1982        /*
1983         * No reference counting needed for current->mempolicy
1984         * nor system default_policy
1985         */
1986        if (pol->mode == MPOL_INTERLEAVE)
1987                page = alloc_page_interleave(gfp, order, interleave_nodes(pol));
1988        else
1989                page = __alloc_pages_nodemask(gfp, order,
1990                                policy_zonelist(gfp, pol, numa_node_id()),
1991                                policy_nodemask(gfp, pol));
1992
1993        if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !page))
1994                goto retry_cpuset;
1995
1996        return page;
1997}
1998EXPORT_SYMBOL(alloc_pages_current);
1999
2000/*
2001 * If mpol_dup() sees current->cpuset == cpuset_being_rebound, then it
2002 * rebinds the mempolicy its copying by calling mpol_rebind_policy()
2003 * with the mems_allowed returned by cpuset_mems_allowed().  This
2004 * keeps mempolicies cpuset relative after its cpuset moves.  See
2005 * further kernel/cpuset.c update_nodemask().
2006 *
2007 * current's mempolicy may be rebinded by the other task(the task that changes
2008 * cpuset's mems), so we needn't do rebind work for current task.
2009 */
2010
2011/* Slow path of a mempolicy duplicate */
2012struct mempolicy *__mpol_dup(struct mempolicy *old)
2013{
2014        struct mempolicy *new = kmem_cache_alloc(policy_cache, GFP_KERNEL);
2015
2016        if (!new)
2017                return ERR_PTR(-ENOMEM);
2018
2019        /* task's mempolicy is protected by alloc_lock */
2020        if (old == current->mempolicy) {
2021                task_lock(current);
2022                *new = *old;
2023                task_unlock(current);
2024        } else
2025                *new = *old;
2026
2027        rcu_read_lock();
2028        if (current_cpuset_is_being_rebound()) {
2029                nodemask_t mems = cpuset_mems_allowed(current);
2030                if (new->flags & MPOL_F_REBINDING)
2031                        mpol_rebind_policy(new, &mems, MPOL_REBIND_STEP2);
2032                else
2033                        mpol_rebind_policy(new, &mems, MPOL_REBIND_ONCE);
2034        }
2035        rcu_read_unlock();
2036        atomic_set(&new->refcnt, 1);
2037        return new;
2038}
2039
2040/* Slow path of a mempolicy comparison */
2041bool __mpol_equal(struct mempolicy *a, struct mempolicy *b)
2042{
2043        if (!a || !b)
2044                return false;
2045        if (a->mode != b->mode)
2046                return false;
2047        if (a->flags != b->flags)
2048                return false;
2049        if (mpol_store_user_nodemask(a))
2050                if (!nodes_equal(a->w.user_nodemask, b->w.user_nodemask))
2051                        return false;
2052
2053        switch (a->mode) {
2054        case MPOL_BIND:
2055                /* Fall through */
2056        case MPOL_INTERLEAVE:
2057                return !!nodes_equal(a->v.nodes, b->v.nodes);
2058        case MPOL_PREFERRED:
2059                return a->v.preferred_node == b->v.preferred_node;
2060        default:
2061                BUG();
2062                return false;
2063        }
2064}
2065
2066/*
2067 * Shared memory backing store policy support.
2068 *
2069 * Remember policies even when nobody has shared memory mapped.
2070 * The policies are kept in Red-Black tree linked from the inode.
2071 * They are protected by the sp->lock spinlock, which should be held
2072 * for any accesses to the tree.
2073 */
2074
2075/* lookup first element intersecting start-end */
2076/* Caller holds sp->mutex */
2077static struct sp_node *
2078sp_lookup(struct shared_policy *sp, unsigned long start, unsigned long end)
2079{
2080        struct rb_node *n = sp->root.rb_node;
2081
2082        while (n) {
2083                struct sp_node *p = rb_entry(n, struct sp_node, nd);
2084
2085                if (start >= p->end)
2086                        n = n->rb_right;
2087                else if (end <= p->start)
2088                        n = n->rb_left;
2089                else
2090                        break;
2091        }
2092        if (!n)
2093                return NULL;
2094        for (;;) {
2095                struct sp_node *w = NULL;
2096                struct rb_node *prev = rb_prev(n);
2097                if (!prev)
2098                        break;
2099                w = rb_entry(prev, struct sp_node, nd);
2100                if (w->end <= start)
2101                        break;
2102                n = prev;
2103        }
2104        return rb_entry(n, struct sp_node, nd);
2105}
2106
2107/* Insert a new shared policy into the list. */
2108/* Caller holds sp->lock */
2109static void sp_insert(struct shared_policy *sp, struct sp_node *new)
2110{
2111        struct rb_node **p = &sp->root.rb_node;
2112        struct rb_node *parent = NULL;
2113        struct sp_node *nd;
2114
2115        while (*p) {
2116                parent = *p;
2117                nd = rb_entry(parent, struct sp_node, nd);
2118                if (new->start < nd->start)
2119                        p = &(*p)->rb_left;
2120                else if (new->end > nd->end)
2121                        p = &(*p)->rb_right;
2122                else
2123                        BUG();
2124        }
2125        rb_link_node(&new->nd, parent, p);
2126        rb_insert_color(&new->nd, &sp->root);
2127        pr_debug("inserting %lx-%lx: %d\n", new->start, new->end,
2128                 new->policy ? new->policy->mode : 0);
2129}
2130
2131/* Find shared policy intersecting idx */
2132struct mempolicy *
2133mpol_shared_policy_lookup(struct shared_policy *sp, unsigned long idx)
2134{
2135        struct mempolicy *pol = NULL;
2136        struct sp_node *sn;
2137
2138        if (!sp->root.rb_node)
2139                return NULL;
2140        mutex_lock(&sp->mutex);
2141        sn = sp_lookup(sp, idx, idx+1);
2142        if (sn) {
2143                mpol_get(sn->policy);
2144                pol = sn->policy;
2145        }
2146        mutex_unlock(&sp->mutex);
2147        return pol;
2148}
2149
2150static void sp_free(struct sp_node *n)
2151{
2152        mpol_put(n->policy);
2153        kmem_cache_free(sn_cache, n);
2154}
2155
2156static void sp_delete(struct shared_policy *sp, struct sp_node *n)
2157{
2158        pr_debug("deleting %lx-l%lx\n", n->start, n->end);
2159        rb_erase(&n->nd, &sp->root);
2160        sp_free(n);
2161}
2162
2163static struct sp_node *sp_alloc(unsigned long start, unsigned long end,
2164                                struct mempolicy *pol)
2165{
2166        struct sp_node *n;
2167        struct mempolicy *newpol;
2168
2169        n = kmem_cache_alloc(sn_cache, GFP_KERNEL);
2170        if (!n)
2171                return NULL;
2172
2173        newpol = mpol_dup(pol);
2174        if (IS_ERR(newpol)) {
2175                kmem_cache_free(sn_cache, n);
2176                return NULL;
2177        }
2178        newpol->flags |= MPOL_F_SHARED;
2179
2180        n->start = start;
2181        n->end = end;
2182        n->policy = newpol;
2183
2184        return n;
2185}
2186
2187/* Replace a policy range. */
2188static int shared_policy_replace(struct shared_policy *sp, unsigned long start,
2189                                 unsigned long end, struct sp_node *new)
2190{
2191        struct sp_node *n;
2192        int ret = 0;
2193
2194        mutex_lock(&sp->mutex);
2195        n = sp_lookup(sp, start, end);
2196        /* Take care of old policies in the same range. */
2197        while (n && n->start < end) {
2198                struct rb_node *next = rb_next(&n->nd);
2199                if (n->start >= start) {
2200                        if (n->end <= end)
2201                                sp_delete(sp, n);
2202                        else
2203                                n->start = end;
2204                } else {
2205                        /* Old policy spanning whole new range. */
2206                        if (n->end > end) {
2207                                struct sp_node *new2;
2208                                new2 = sp_alloc(end, n->end, n->policy);
2209                                if (!new2) {
2210                                        ret = -ENOMEM;
2211                                        goto out;
2212                                }
2213                                n->end = start;
2214                                sp_insert(sp, new2);
2215                                break;
2216                        } else
2217                                n->end = start;
2218                }
2219                if (!next)
2220                        break;
2221                n = rb_entry(next, struct sp_node, nd);
2222        }
2223        if (new)
2224                sp_insert(sp, new);
2225out:
2226        mutex_unlock(&sp->mutex);
2227        return ret;
2228}
2229
2230/**
2231 * mpol_shared_policy_init - initialize shared policy for inode
2232 * @sp: pointer to inode shared policy
2233 * @mpol:  struct mempolicy to install
2234 *
2235 * Install non-NULL @mpol in inode's shared policy rb-tree.
2236 * On entry, the current task has a reference on a non-NULL @mpol.
2237 * This must be released on exit.
2238 * This is called at get_inode() calls and we can use GFP_KERNEL.
2239 */
2240void mpol_shared_policy_init(struct shared_policy *sp, struct mempolicy *mpol)
2241{
2242        int ret;
2243
2244        sp->root = RB_ROOT;             /* empty tree == default mempolicy */
2245        mutex_init(&sp->mutex);
2246
2247        if (mpol) {
2248                struct vm_area_struct pvma;
2249                struct mempolicy *new;
2250                NODEMASK_SCRATCH(scratch);
2251
2252                if (!scratch)
2253                        goto put_mpol;
2254                /* contextualize the tmpfs mount point mempolicy */
2255                new = mpol_new(mpol->mode, mpol->flags, &mpol->w.user_nodemask);
2256                if (IS_ERR(new))
2257                        goto free_scratch; /* no valid nodemask intersection */
2258
2259                task_lock(current);
2260                ret = mpol_set_nodemask(new, &mpol->w.user_nodemask, scratch);
2261                task_unlock(current);
2262                if (ret)
2263                        goto put_new;
2264
2265                /* Create pseudo-vma that contains just the policy */
2266                memset(&pvma, 0, sizeof(struct vm_area_struct));
2267                pvma.vm_end = TASK_SIZE;        /* policy covers entire file */
2268                mpol_set_shared_policy(sp, &pvma, new); /* adds ref */
2269
2270put_new:
2271                mpol_put(new);                  /* drop initial ref */
2272free_scratch:
2273                NODEMASK_SCRATCH_FREE(scratch);
2274put_mpol:
2275                mpol_put(mpol); /* drop our incoming ref on sb mpol */
2276        }
2277}
2278
2279int mpol_set_shared_policy(struct shared_policy *info,
2280                        struct vm_area_struct *vma, struct mempolicy *npol)
2281{
2282        int err;
2283        struct sp_node *new = NULL;
2284        unsigned long sz = vma_pages(vma);
2285
2286        pr_debug("set_shared_policy %lx sz %lu %d %d %lx\n",
2287                 vma->vm_pgoff,
2288                 sz, npol ? npol->mode : -1,
2289                 npol ? npol->flags : -1,
2290                 npol ? nodes_addr(npol->v.nodes)[0] : -1);
2291
2292        if (npol) {
2293                new = sp_alloc(vma->vm_pgoff, vma->vm_pgoff + sz, npol);
2294                if (!new)
2295                        return -ENOMEM;
2296        }
2297        err = shared_policy_replace(info, vma->vm_pgoff, vma->vm_pgoff+sz, new);
2298        if (err && new)
2299                sp_free(new);
2300        return err;
2301}
2302
2303/* Free a backing policy store on inode delete. */
2304void mpol_free_shared_policy(struct shared_policy *p)
2305{
2306        struct sp_node *n;
2307        struct rb_node *next;
2308
2309        if (!p->root.rb_node)
2310                return;
2311        mutex_lock(&p->mutex);
2312        next = rb_first(&p->root);
2313        while (next) {
2314                n = rb_entry(next, struct sp_node, nd);
2315                next = rb_next(&n->nd);
2316                sp_delete(p, n);
2317        }
2318        mutex_unlock(&p->mutex);
2319}
2320
2321/* assumes fs == KERNEL_DS */
2322void __init numa_policy_init(void)
2323{
2324        nodemask_t interleave_nodes;
2325        unsigned long largest = 0;
2326        int nid, prefer = 0;
2327
2328        policy_cache = kmem_cache_create("numa_policy",
2329                                         sizeof(struct mempolicy),
2330                                         0, SLAB_PANIC, NULL);
2331
2332        sn_cache = kmem_cache_create("shared_policy_node",
2333                                     sizeof(struct sp_node),
2334                                     0, SLAB_PANIC, NULL);
2335
2336        /*
2337         * Set interleaving policy for system init. Interleaving is only
2338         * enabled across suitably sized nodes (default is >= 16MB), or
2339         * fall back to the largest node if they're all smaller.
2340         */
2341        nodes_clear(interleave_nodes);
2342        for_each_node_state(nid, N_HIGH_MEMORY) {
2343                unsigned long total_pages = node_present_pages(nid);
2344
2345                /* Preserve the largest node */
2346                if (largest < total_pages) {
2347                        largest = total_pages;
2348                        prefer = nid;
2349                }
2350
2351                /* Interleave this node? */
2352                if ((total_pages << PAGE_SHIFT) >= (16 << 20))
2353                        node_set(nid, interleave_nodes);
2354        }
2355
2356        /* All too small, use the largest */
2357        if (unlikely(nodes_empty(interleave_nodes)))
2358                node_set(prefer, interleave_nodes);
2359
2360        if (do_set_mempolicy(MPOL_INTERLEAVE, 0, &interleave_nodes))
2361                printk("numa_policy_init: interleaving failed\n");
2362}
2363
2364/* Reset policy of current process to default */
2365void numa_default_policy(void)
2366{
2367        do_set_mempolicy(MPOL_DEFAULT, 0, NULL);
2368}
2369
2370/*
2371 * Parse and format mempolicy from/to strings
2372 */
2373
2374/*
2375 * "local" is implemented internally by MPOL_PREFERRED with MPOL_F_LOCAL flag.
2376 */
2377#define MPOL_LOCAL MPOL_MAX
2378static const char * const policy_modes[] =
2379{
2380        [MPOL_DEFAULT]    = "default",
2381        [MPOL_PREFERRED]  = "prefer",
2382        [MPOL_BIND]       = "bind",
2383        [MPOL_INTERLEAVE] = "interleave",
2384        [MPOL_LOCAL]      = "local"
2385};
2386
2387
2388#ifdef CONFIG_TMPFS
2389/**
2390 * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option.
2391 * @str:  string containing mempolicy to parse
2392 * @mpol:  pointer to struct mempolicy pointer, returned on success.
2393 * @unused:  redundant argument, to be removed later.
2394 *
2395 * Format of input:
2396 *      <mode>[=<flags>][:<nodelist>]
2397 *
2398 * On success, returns 0, else 1
2399 */
2400int mpol_parse_str(char *str, struct mempolicy **mpol, int unused)
2401{
2402        struct mempolicy *new = NULL;
2403        unsigned short mode;
2404        unsigned short mode_flags;
2405        nodemask_t nodes;
2406        char *nodelist = strchr(str, ':');
2407        char *flags = strchr(str, '=');
2408        int err = 1;
2409
2410        if (nodelist) {
2411                /* NUL-terminate mode or flags string */
2412                *nodelist++ = '\0';
2413                if (nodelist_parse(nodelist, nodes))
2414                        goto out;
2415                if (!nodes_subset(nodes, node_states[N_HIGH_MEMORY]))
2416                        goto out;
2417        } else
2418                nodes_clear(nodes);
2419
2420        if (flags)
2421                *flags++ = '\0';        /* terminate mode string */
2422
2423        for (mode = 0; mode <= MPOL_LOCAL; mode++) {
2424                if (!strcmp(str, policy_modes[mode])) {
2425                        break;
2426                }
2427        }
2428        if (mode > MPOL_LOCAL)
2429                goto out;
2430
2431        switch (mode) {
2432        case MPOL_PREFERRED:
2433                /*
2434                 * Insist on a nodelist of one node only
2435                 */
2436                if (nodelist) {
2437                        char *rest = nodelist;
2438                        while (isdigit(*rest))
2439                                rest++;
2440                        if (*rest)
2441                                goto out;
2442                }
2443                break;
2444        case MPOL_INTERLEAVE:
2445                /*
2446                 * Default to online nodes with memory if no nodelist
2447                 */
2448                if (!nodelist)
2449                        nodes = node_states[N_HIGH_MEMORY];
2450                break;
2451        case MPOL_LOCAL:
2452                /*
2453                 * Don't allow a nodelist;  mpol_new() checks flags
2454                 */
2455                if (nodelist)
2456                        goto out;
2457                mode = MPOL_PREFERRED;
2458                break;
2459        case MPOL_DEFAULT:
2460                /*
2461                 * Insist on a empty nodelist
2462                 */
2463                if (!nodelist)
2464                        err = 0;
2465                goto out;
2466        case MPOL_BIND:
2467                /*
2468                 * Insist on a nodelist
2469                 */
2470                if (!nodelist)
2471                        goto out;
2472        }
2473
2474        mode_flags = 0;
2475        if (flags) {
2476                /*
2477                 * Currently, we only support two mutually exclusive
2478                 * mode flags.
2479                 */
2480                if (!strcmp(flags, "static"))
2481                        mode_flags |= MPOL_F_STATIC_NODES;
2482                else if (!strcmp(flags, "relative"))
2483                        mode_flags |= MPOL_F_RELATIVE_NODES;
2484                else
2485                        goto out;
2486        }
2487
2488        new = mpol_new(mode, mode_flags, &nodes);
2489        if (IS_ERR(new))
2490                goto out;
2491
2492        /*
2493         * Save nodes for mpol_to_str() to show the tmpfs mount options
2494         * for /proc/mounts, /proc/pid/mounts and /proc/pid/mountinfo.
2495         */
2496        if (mode != MPOL_PREFERRED)
2497                new->v.nodes = nodes;
2498        else if (nodelist)
2499                new->v.preferred_node = first_node(nodes);
2500        else
2501                new->flags |= MPOL_F_LOCAL;
2502
2503        /*
2504         * Save nodes for contextualization: this will be used to "clone"
2505         * the mempolicy in a specific context [cpuset] at a later time.
2506         */
2507        new->w.user_nodemask = nodes;
2508
2509        err = 0;
2510
2511out:
2512        /* Restore string for error message */
2513        if (nodelist)
2514                *--nodelist = ':';
2515        if (flags)
2516                *--flags = '=';
2517        if (!err)
2518                *mpol = new;
2519        return err;
2520}
2521#endif /* CONFIG_TMPFS */
2522
2523/**
2524 * mpol_to_str - format a mempolicy structure for printing
2525 * @buffer:  to contain formatted mempolicy string
2526 * @maxlen:  length of @buffer
2527 * @pol:  pointer to mempolicy to be formatted
2528 * @unused:  redundant argument, to be removed later.
2529 *
2530 * Convert a mempolicy into a string.
2531 * Returns the number of characters in buffer (if positive)
2532 * or an error (negative)
2533 */
2534int mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol, int unused)
2535{
2536        char *p = buffer;
2537        int l;
2538        nodemask_t nodes;
2539        unsigned short mode;
2540        unsigned short flags = pol ? pol->flags : 0;
2541
2542        /*
2543         * Sanity check:  room for longest mode, flag and some nodes
2544         */
2545        VM_BUG_ON(maxlen < strlen("interleave") + strlen("relative") + 16);
2546
2547        if (!pol || pol == &default_policy)
2548                mode = MPOL_DEFAULT;
2549        else
2550                mode = pol->mode;
2551
2552        switch (mode) {
2553        case MPOL_DEFAULT:
2554                nodes_clear(nodes);
2555                break;
2556
2557        case MPOL_PREFERRED:
2558                nodes_clear(nodes);
2559                if (flags & MPOL_F_LOCAL)
2560                        mode = MPOL_LOCAL;
2561                else
2562                        node_set(pol->v.preferred_node, nodes);
2563                break;
2564
2565        case MPOL_BIND:
2566                /* Fall through */
2567        case MPOL_INTERLEAVE:
2568                nodes = pol->v.nodes;
2569                break;
2570
2571        default:
2572                return -EINVAL;
2573        }
2574
2575        l = strlen(policy_modes[mode]);
2576        if (buffer + maxlen < p + l + 1)
2577                return -ENOSPC;
2578
2579        strcpy(p, policy_modes[mode]);
2580        p += l;
2581
2582        if (flags & MPOL_MODE_FLAGS) {
2583                if (buffer + maxlen < p + 2)
2584                        return -ENOSPC;
2585                *p++ = '=';
2586
2587                /*
2588                 * Currently, the only defined flags are mutually exclusive
2589                 */
2590                if (flags & MPOL_F_STATIC_NODES)
2591                        p += snprintf(p, buffer + maxlen - p, "static");
2592                else if (flags & MPOL_F_RELATIVE_NODES)
2593                        p += snprintf(p, buffer + maxlen - p, "relative");
2594        }
2595
2596        if (!nodes_empty(nodes)) {
2597                if (buffer + maxlen < p + 2)
2598                        return -ENOSPC;
2599                *p++ = ':';
2600                p += nodelist_scnprintf(p, buffer + maxlen - p, nodes);
2601        }
2602        return p - buffer;
2603}
2604
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.