linux/mm/vmalloc.c
<<
>>
Prefs
   1/*
   2 *  linux/mm/vmalloc.c
   3 *
   4 *  Copyright (C) 1993  Linus Torvalds
   5 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   6 *  SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
   7 *  Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
   8 *  Numa awareness, Christoph Lameter, SGI, June 2005
   9 */
  10
  11#include <linux/vmalloc.h>
  12#include <linux/mm.h>
  13#include <linux/module.h>
  14#include <linux/highmem.h>
  15#include <linux/sched.h>
  16#include <linux/slab.h>
  17#include <linux/spinlock.h>
  18#include <linux/interrupt.h>
  19#include <linux/proc_fs.h>
  20#include <linux/seq_file.h>
  21#include <linux/debugobjects.h>
  22#include <linux/kallsyms.h>
  23#include <linux/list.h>
  24#include <linux/rbtree.h>
  25#include <linux/radix-tree.h>
  26#include <linux/rcupdate.h>
  27#include <linux/pfn.h>
  28#include <linux/kmemleak.h>
  29#include <linux/atomic.h>
  30#include <asm/uaccess.h>
  31#include <asm/tlbflush.h>
  32#include <asm/shmparam.h>
  33
  34/*** Page table manipulation functions ***/
  35
  36static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
  37{
  38        pte_t *pte;
  39
  40        pte = pte_offset_kernel(pmd, addr);
  41        do {
  42                pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
  43                WARN_ON(!pte_none(ptent) && !pte_present(ptent));
  44        } while (pte++, addr += PAGE_SIZE, addr != end);
  45}
  46
  47static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
  48{
  49        pmd_t *pmd;
  50        unsigned long next;
  51
  52        pmd = pmd_offset(pud, addr);
  53        do {
  54                next = pmd_addr_end(addr, end);
  55                if (pmd_none_or_clear_bad(pmd))
  56                        continue;
  57                vunmap_pte_range(pmd, addr, next);
  58        } while (pmd++, addr = next, addr != end);
  59}
  60
  61static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
  62{
  63        pud_t *pud;
  64        unsigned long next;
  65
  66        pud = pud_offset(pgd, addr);
  67        do {
  68                next = pud_addr_end(addr, end);
  69                if (pud_none_or_clear_bad(pud))
  70                        continue;
  71                vunmap_pmd_range(pud, addr, next);
  72        } while (pud++, addr = next, addr != end);
  73}
  74
  75static void vunmap_page_range(unsigned long addr, unsigned long end)
  76{
  77        pgd_t *pgd;
  78        unsigned long next;
  79
  80        BUG_ON(addr >= end);
  81        pgd = pgd_offset_k(addr);
  82        do {
  83                next = pgd_addr_end(addr, end);
  84                if (pgd_none_or_clear_bad(pgd))
  85                        continue;
  86                vunmap_pud_range(pgd, addr, next);
  87        } while (pgd++, addr = next, addr != end);
  88}
  89
  90static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
  91                unsigned long end, pgprot_t prot, struct page **pages, int *nr)
  92{
  93        pte_t *pte;
  94
  95        /*
  96         * nr is a running index into the array which helps higher level
  97         * callers keep track of where we're up to.
  98         */
  99
 100        pte = pte_alloc_kernel(pmd, addr);
 101        if (!pte)
 102                return -ENOMEM;
 103        do {
 104                struct page *page = pages[*nr];
 105
 106                if (WARN_ON(!pte_none(*pte)))
 107                        return -EBUSY;
 108                if (WARN_ON(!page))
 109                        return -ENOMEM;
 110                set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 111                (*nr)++;
 112        } while (pte++, addr += PAGE_SIZE, addr != end);
 113        return 0;
 114}
 115
 116static int vmap_pmd_range(pud_t *pud, unsigned long addr,
 117                unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 118{
 119        pmd_t *pmd;
 120        unsigned long next;
 121
 122        pmd = pmd_alloc(&init_mm, pud, addr);
 123        if (!pmd)
 124                return -ENOMEM;
 125        do {
 126                next = pmd_addr_end(addr, end);
 127                if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
 128                        return -ENOMEM;
 129        } while (pmd++, addr = next, addr != end);
 130        return 0;
 131}
 132
 133static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
 134                unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 135{
 136        pud_t *pud;
 137        unsigned long next;
 138
 139        pud = pud_alloc(&init_mm, pgd, addr);
 140        if (!pud)
 141                return -ENOMEM;
 142        do {
 143                next = pud_addr_end(addr, end);
 144                if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
 145                        return -ENOMEM;
 146        } while (pud++, addr = next, addr != end);
 147        return 0;
 148}
 149
 150/*
 151 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
 152 * will have pfns corresponding to the "pages" array.
 153 *
 154 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
 155 */
 156static int vmap_page_range_noflush(unsigned long start, unsigned long end,
 157                                   pgprot_t prot, struct page **pages)
 158{
 159        pgd_t *pgd;
 160        unsigned long next;
 161        unsigned long addr = start;
 162        int err = 0;
 163        int nr = 0;
 164
 165        BUG_ON(addr >= end);
 166        pgd = pgd_offset_k(addr);
 167        do {
 168                next = pgd_addr_end(addr, end);
 169                err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
 170                if (err)
 171                        return err;
 172        } while (pgd++, addr = next, addr != end);
 173
 174        return nr;
 175}
 176
 177static int vmap_page_range(unsigned long start, unsigned long end,
 178                           pgprot_t prot, struct page **pages)
 179{
 180        int ret;
 181
 182        ret = vmap_page_range_noflush(start, end, prot, pages);
 183        flush_cache_vmap(start, end);
 184        return ret;
 185}
 186
 187int is_vmalloc_or_module_addr(const void *x)
 188{
 189        /*
 190         * ARM, x86-64 and sparc64 put modules in a special place,
 191         * and fall back on vmalloc() if that fails. Others
 192         * just put it in the vmalloc space.
 193         */
 194#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 195        unsigned long addr = (unsigned long)x;
 196        if (addr >= MODULES_VADDR && addr < MODULES_END)
 197                return 1;
 198#endif
 199        return is_vmalloc_addr(x);
 200}
 201
 202/*
 203 * Walk a vmap address to the struct page it maps.
 204 */
 205struct page *vmalloc_to_page(const void *vmalloc_addr)
 206{
 207        unsigned long addr = (unsigned long) vmalloc_addr;
 208        struct page *page = NULL;
 209        pgd_t *pgd = pgd_offset_k(addr);
 210
 211        /*
 212         * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 213         * architectures that do not vmalloc module space
 214         */
 215        VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 216
 217        if (!pgd_none(*pgd)) {
 218                pud_t *pud = pud_offset(pgd, addr);
 219                if (!pud_none(*pud)) {
 220                        pmd_t *pmd = pmd_offset(pud, addr);
 221                        if (!pmd_none(*pmd)) {
 222                                pte_t *ptep, pte;
 223
 224                                ptep = pte_offset_map(pmd, addr);
 225                                pte = *ptep;
 226                                if (pte_present(pte))
 227                                        page = pte_page(pte);
 228                                pte_unmap(ptep);
 229                        }
 230                }
 231        }
 232        return page;
 233}
 234EXPORT_SYMBOL(vmalloc_to_page);
 235
 236/*
 237 * Map a vmalloc()-space virtual address to the physical page frame number.
 238 */
 239unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 240{
 241        return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 242}
 243EXPORT_SYMBOL(vmalloc_to_pfn);
 244
 245
 246/*** Global kva allocator ***/
 247
 248#define VM_LAZY_FREE    0x01
 249#define VM_LAZY_FREEING 0x02
 250#define VM_VM_AREA      0x04
 251
 252struct vmap_area {
 253        unsigned long va_start;
 254        unsigned long va_end;
 255        unsigned long flags;
 256        struct rb_node rb_node;         /* address sorted rbtree */
 257        struct list_head list;          /* address sorted list */
 258        struct list_head purge_list;    /* "lazy purge" list */
 259        void *private;
 260        struct rcu_head rcu_head;
 261};
 262
 263static DEFINE_SPINLOCK(vmap_area_lock);
 264static LIST_HEAD(vmap_area_list);
 265static struct rb_root vmap_area_root = RB_ROOT;
 266
 267/* The vmap cache globals are protected by vmap_area_lock */
 268static struct rb_node *free_vmap_cache;
 269static unsigned long cached_hole_size;
 270static unsigned long cached_vstart;
 271static unsigned long cached_align;
 272
 273static unsigned long vmap_area_pcpu_hole;
 274
 275static struct vmap_area *__find_vmap_area(unsigned long addr)
 276{
 277        struct rb_node *n = vmap_area_root.rb_node;
 278
 279        while (n) {
 280                struct vmap_area *va;
 281
 282                va = rb_entry(n, struct vmap_area, rb_node);
 283                if (addr < va->va_start)
 284                        n = n->rb_left;
 285                else if (addr > va->va_start)
 286                        n = n->rb_right;
 287                else
 288                        return va;
 289        }
 290
 291        return NULL;
 292}
 293
 294static void __insert_vmap_area(struct vmap_area *va)
 295{
 296        struct rb_node **p = &vmap_area_root.rb_node;
 297        struct rb_node *parent = NULL;
 298        struct rb_node *tmp;
 299
 300        while (*p) {
 301                struct vmap_area *tmp_va;
 302
 303                parent = *p;
 304                tmp_va = rb_entry(parent, struct vmap_area, rb_node);
 305                if (va->va_start < tmp_va->va_end)
 306                        p = &(*p)->rb_left;
 307                else if (va->va_end > tmp_va->va_start)
 308                        p = &(*p)->rb_right;
 309                else
 310                        BUG();
 311        }
 312
 313        rb_link_node(&va->rb_node, parent, p);
 314        rb_insert_color(&va->rb_node, &vmap_area_root);
 315
 316        /* address-sort this list so it is usable like the vmlist */
 317        tmp = rb_prev(&va->rb_node);
 318        if (tmp) {
 319                struct vmap_area *prev;
 320                prev = rb_entry(tmp, struct vmap_area, rb_node);
 321                list_add_rcu(&va->list, &prev->list);
 322        } else
 323                list_add_rcu(&va->list, &vmap_area_list);
 324}
 325
 326static void purge_vmap_area_lazy(void);
 327
 328/*
 329 * Allocate a region of KVA of the specified size and alignment, within the
 330 * vstart and vend.
 331 */
 332static struct vmap_area *alloc_vmap_area(unsigned long size,
 333                                unsigned long align,
 334                                unsigned long vstart, unsigned long vend,
 335                                int node, gfp_t gfp_mask)
 336{
 337        struct vmap_area *va;
 338        struct rb_node *n;
 339        unsigned long addr;
 340        int purged = 0;
 341        struct vmap_area *first;
 342
 343        BUG_ON(!size);
 344        BUG_ON(size & ~PAGE_MASK);
 345        BUG_ON(!is_power_of_2(align));
 346
 347        va = kmalloc_node(sizeof(struct vmap_area),
 348                        gfp_mask & GFP_RECLAIM_MASK, node);
 349        if (unlikely(!va))
 350                return ERR_PTR(-ENOMEM);
 351
 352retry:
 353        spin_lock(&vmap_area_lock);
 354        /*
 355         * Invalidate cache if we have more permissive parameters.
 356         * cached_hole_size notes the largest hole noticed _below_
 357         * the vmap_area cached in free_vmap_cache: if size fits
 358         * into that hole, we want to scan from vstart to reuse
 359         * the hole instead of allocating above free_vmap_cache.
 360         * Note that __free_vmap_area may update free_vmap_cache
 361         * without updating cached_hole_size or cached_align.
 362         */
 363        if (!free_vmap_cache ||
 364                        size < cached_hole_size ||
 365                        vstart < cached_vstart ||
 366                        align < cached_align) {
 367nocache:
 368                cached_hole_size = 0;
 369                free_vmap_cache = NULL;
 370        }
 371        /* record if we encounter less permissive parameters */
 372        cached_vstart = vstart;
 373        cached_align = align;
 374
 375        /* find starting point for our search */
 376        if (free_vmap_cache) {
 377                first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 378                addr = ALIGN(first->va_end, align);
 379                if (addr < vstart)
 380                        goto nocache;
 381                if (addr + size - 1 < addr)
 382                        goto overflow;
 383
 384        } else {
 385                addr = ALIGN(vstart, align);
 386                if (addr + size - 1 < addr)
 387                        goto overflow;
 388
 389                n = vmap_area_root.rb_node;
 390                first = NULL;
 391
 392                while (n) {
 393                        struct vmap_area *tmp;
 394                        tmp = rb_entry(n, struct vmap_area, rb_node);
 395                        if (tmp->va_end >= addr) {
 396                                first = tmp;
 397                                if (tmp->va_start <= addr)
 398                                        break;
 399                                n = n->rb_left;
 400                        } else
 401                                n = n->rb_right;
 402                }
 403
 404                if (!first)
 405                        goto found;
 406        }
 407
 408        /* from the starting point, walk areas until a suitable hole is found */
 409        while (addr + size > first->va_start && addr + size <= vend) {
 410                if (addr + cached_hole_size < first->va_start)
 411                        cached_hole_size = first->va_start - addr;
 412                addr = ALIGN(first->va_end, align);
 413                if (addr + size - 1 < addr)
 414                        goto overflow;
 415
 416                n = rb_next(&first->rb_node);
 417                if (n)
 418                        first = rb_entry(n, struct vmap_area, rb_node);
 419                else
 420                        goto found;
 421        }
 422
 423found:
 424        if (addr + size > vend)
 425                goto overflow;
 426
 427        va->va_start = addr;
 428        va->va_end = addr + size;
 429        va->flags = 0;
 430        __insert_vmap_area(va);
 431        free_vmap_cache = &va->rb_node;
 432        spin_unlock(&vmap_area_lock);
 433
 434        BUG_ON(va->va_start & (align-1));
 435        BUG_ON(va->va_start < vstart);
 436        BUG_ON(va->va_end > vend);
 437
 438        return va;
 439
 440overflow:
 441        spin_unlock(&vmap_area_lock);
 442        if (!purged) {
 443                purge_vmap_area_lazy();
 444                purged = 1;
 445                goto retry;
 446        }
 447        if (printk_ratelimit())
 448                printk(KERN_WARNING
 449                        "vmap allocation for size %lu failed: "
 450                        "use vmalloc=<size> to increase size.\n", size);
 451        kfree(va);
 452        return ERR_PTR(-EBUSY);
 453}
 454
 455static void __free_vmap_area(struct vmap_area *va)
 456{
 457        BUG_ON(RB_EMPTY_NODE(&va->rb_node));
 458
 459        if (free_vmap_cache) {
 460                if (va->va_end < cached_vstart) {
 461                        free_vmap_cache = NULL;
 462                } else {
 463                        struct vmap_area *cache;
 464                        cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 465                        if (va->va_start <= cache->va_start) {
 466                                free_vmap_cache = rb_prev(&va->rb_node);
 467                                /*
 468                                 * We don't try to update cached_hole_size or
 469                                 * cached_align, but it won't go very wrong.
 470                                 */
 471                        }
 472                }
 473        }
 474        rb_erase(&va->rb_node, &vmap_area_root);
 475        RB_CLEAR_NODE(&va->rb_node);
 476        list_del_rcu(&va->list);
 477
 478        /*
 479         * Track the highest possible candidate for pcpu area
 480         * allocation.  Areas outside of vmalloc area can be returned
 481         * here too, consider only end addresses which fall inside
 482         * vmalloc area proper.
 483         */
 484        if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
 485                vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
 486
 487        kfree_rcu(va, rcu_head);
 488}
 489
 490/*
 491 * Free a region of KVA allocated by alloc_vmap_area
 492 */
 493static void free_vmap_area(struct vmap_area *va)
 494{
 495        spin_lock(&vmap_area_lock);
 496        __free_vmap_area(va);
 497        spin_unlock(&vmap_area_lock);
 498}
 499
 500/*
 501 * Clear the pagetable entries of a given vmap_area
 502 */
 503static void unmap_vmap_area(struct vmap_area *va)
 504{
 505        vunmap_page_range(va->va_start, va->va_end);
 506}
 507
 508static void vmap_debug_free_range(unsigned long start, unsigned long end)
 509{
 510        /*
 511         * Unmap page tables and force a TLB flush immediately if
 512         * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
 513         * bugs similarly to those in linear kernel virtual address
 514         * space after a page has been freed.
 515         *
 516         * All the lazy freeing logic is still retained, in order to
 517         * minimise intrusiveness of this debugging feature.
 518         *
 519         * This is going to be *slow* (linear kernel virtual address
 520         * debugging doesn't do a broadcast TLB flush so it is a lot
 521         * faster).
 522         */
 523#ifdef CONFIG_DEBUG_PAGEALLOC
 524        vunmap_page_range(start, end);
 525        flush_tlb_kernel_range(start, end);
 526#endif
 527}
 528
 529/*
 530 * lazy_max_pages is the maximum amount of virtual address space we gather up
 531 * before attempting to purge with a TLB flush.
 532 *
 533 * There is a tradeoff here: a larger number will cover more kernel page tables
 534 * and take slightly longer to purge, but it will linearly reduce the number of
 535 * global TLB flushes that must be performed. It would seem natural to scale
 536 * this number up linearly with the number of CPUs (because vmapping activity
 537 * could also scale linearly with the number of CPUs), however it is likely
 538 * that in practice, workloads might be constrained in other ways that mean
 539 * vmap activity will not scale linearly with CPUs. Also, I want to be
 540 * conservative and not introduce a big latency on huge systems, so go with
 541 * a less aggressive log scale. It will still be an improvement over the old
 542 * code, and it will be simple to change the scale factor if we find that it
 543 * becomes a problem on bigger systems.
 544 */
 545static unsigned long lazy_max_pages(void)
 546{
 547        unsigned int log;
 548
 549        log = fls(num_online_cpus());
 550
 551        return log * (32UL * 1024 * 1024 / PAGE_SIZE);
 552}
 553
 554static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
 555
 556/* for per-CPU blocks */
 557static void purge_fragmented_blocks_allcpus(void);
 558
 559/*
 560 * called before a call to iounmap() if the caller wants vm_area_struct's
 561 * immediately freed.
 562 */
 563void set_iounmap_nonlazy(void)
 564{
 565        atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
 566}
 567
 568/*
 569 * Purges all lazily-freed vmap areas.
 570 *
 571 * If sync is 0 then don't purge if there is already a purge in progress.
 572 * If force_flush is 1, then flush kernel TLBs between *start and *end even
 573 * if we found no lazy vmap areas to unmap (callers can use this to optimise
 574 * their own TLB flushing).
 575 * Returns with *start = min(*start, lowest purged address)
 576 *              *end = max(*end, highest purged address)
 577 */
 578static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
 579                                        int sync, int force_flush)
 580{
 581        static DEFINE_SPINLOCK(purge_lock);
 582        LIST_HEAD(valist);
 583        struct vmap_area *va;
 584        struct vmap_area *n_va;
 585        int nr = 0;
 586
 587        /*
 588         * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
 589         * should not expect such behaviour. This just simplifies locking for
 590         * the case that isn't actually used at the moment anyway.
 591         */
 592        if (!sync && !force_flush) {
 593                if (!spin_trylock(&purge_lock))
 594                        return;
 595        } else
 596                spin_lock(&purge_lock);
 597
 598        if (sync)
 599                purge_fragmented_blocks_allcpus();
 600
 601        rcu_read_lock();
 602        list_for_each_entry_rcu(va, &vmap_area_list, list) {
 603                if (va->flags & VM_LAZY_FREE) {
 604                        if (va->va_start < *start)
 605                                *start = va->va_start;
 606                        if (va->va_end > *end)
 607                                *end = va->va_end;
 608                        nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
 609                        list_add_tail(&va->purge_list, &valist);
 610                        va->flags |= VM_LAZY_FREEING;
 611                        va->flags &= ~VM_LAZY_FREE;
 612                }
 613        }
 614        rcu_read_unlock();
 615
 616        if (nr)
 617                atomic_sub(nr, &vmap_lazy_nr);
 618
 619        if (nr || force_flush)
 620                flush_tlb_kernel_range(*start, *end);
 621
 622        if (nr) {
 623                spin_lock(&vmap_area_lock);
 624                list_for_each_entry_safe(va, n_va, &valist, purge_list)
 625                        __free_vmap_area(va);
 626                spin_unlock(&vmap_area_lock);
 627        }
 628        spin_unlock(&purge_lock);
 629}
 630
 631/*
 632 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
 633 * is already purging.
 634 */
 635static void try_purge_vmap_area_lazy(void)
 636{
 637        unsigned long start = ULONG_MAX, end = 0;
 638
 639        __purge_vmap_area_lazy(&start, &end, 0, 0);
 640}
 641
 642/*
 643 * Kick off a purge of the outstanding lazy areas.
 644 */
 645static void purge_vmap_area_lazy(void)
 646{
 647        unsigned long start = ULONG_MAX, end = 0;
 648
 649        __purge_vmap_area_lazy(&start, &end, 1, 0);
 650}
 651
 652/*
 653 * Free a vmap area, caller ensuring that the area has been unmapped
 654 * and flush_cache_vunmap had been called for the correct range
 655 * previously.
 656 */
 657static void free_vmap_area_noflush(struct vmap_area *va)
 658{
 659        va->flags |= VM_LAZY_FREE;
 660        atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
 661        if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
 662                try_purge_vmap_area_lazy();
 663}
 664
 665/*
 666 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
 667 * called for the correct range previously.
 668 */
 669static void free_unmap_vmap_area_noflush(struct vmap_area *va)
 670{
 671        unmap_vmap_area(va);
 672        free_vmap_area_noflush(va);
 673}
 674
 675/*
 676 * Free and unmap a vmap area
 677 */
 678static void free_unmap_vmap_area(struct vmap_area *va)
 679{
 680        flush_cache_vunmap(va->va_start, va->va_end);
 681        free_unmap_vmap_area_noflush(va);
 682}
 683
 684static struct vmap_area *find_vmap_area(unsigned long addr)
 685{
 686        struct vmap_area *va;
 687
 688        spin_lock(&vmap_area_lock);
 689        va = __find_vmap_area(addr);
 690        spin_unlock(&vmap_area_lock);
 691
 692        return va;
 693}
 694
 695static void free_unmap_vmap_area_addr(unsigned long addr)
 696{
 697        struct vmap_area *va;
 698
 699        va = find_vmap_area(addr);
 700        BUG_ON(!va);
 701        free_unmap_vmap_area(va);
 702}
 703
 704
 705/*** Per cpu kva allocator ***/
 706
 707/*
 708 * vmap space is limited especially on 32 bit architectures. Ensure there is
 709 * room for at least 16 percpu vmap blocks per CPU.
 710 */
 711/*
 712 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
 713 * to #define VMALLOC_SPACE             (VMALLOC_END-VMALLOC_START). Guess
 714 * instead (we just need a rough idea)
 715 */
 716#if BITS_PER_LONG == 32
 717#define VMALLOC_SPACE           (128UL*1024*1024)
 718#else
 719#define VMALLOC_SPACE           (128UL*1024*1024*1024)
 720#endif
 721
 722#define VMALLOC_PAGES           (VMALLOC_SPACE / PAGE_SIZE)
 723#define VMAP_MAX_ALLOC          BITS_PER_LONG   /* 256K with 4K pages */
 724#define VMAP_BBMAP_BITS_MAX     1024    /* 4MB with 4K pages */
 725#define VMAP_BBMAP_BITS_MIN     (VMAP_MAX_ALLOC*2)
 726#define VMAP_MIN(x, y)          ((x) < (y) ? (x) : (y)) /* can't use min() */
 727#define VMAP_MAX(x, y)          ((x) > (y) ? (x) : (y)) /* can't use max() */
 728#define VMAP_BBMAP_BITS         \
 729                VMAP_MIN(VMAP_BBMAP_BITS_MAX,   \
 730                VMAP_MAX(VMAP_BBMAP_BITS_MIN,   \
 731                        VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
 732
 733#define VMAP_BLOCK_SIZE         (VMAP_BBMAP_BITS * PAGE_SIZE)
 734
 735static bool vmap_initialized __read_mostly = false;
 736
 737struct vmap_block_queue {
 738        spinlock_t lock;
 739        struct list_head free;
 740};
 741
 742struct vmap_block {
 743        spinlock_t lock;
 744        struct vmap_area *va;
 745        struct vmap_block_queue *vbq;
 746        unsigned long free, dirty;
 747        DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
 748        DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
 749        struct list_head free_list;
 750        struct rcu_head rcu_head;
 751        struct list_head purge;
 752};
 753
 754/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
 755static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
 756
 757/*
 758 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
 759 * in the free path. Could get rid of this if we change the API to return a
 760 * "cookie" from alloc, to be passed to free. But no big deal yet.
 761 */
 762static DEFINE_SPINLOCK(vmap_block_tree_lock);
 763static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
 764
 765/*
 766 * We should probably have a fallback mechanism to allocate virtual memory
 767 * out of partially filled vmap blocks. However vmap block sizing should be
 768 * fairly reasonable according to the vmalloc size, so it shouldn't be a
 769 * big problem.
 770 */
 771
 772static unsigned long addr_to_vb_idx(unsigned long addr)
 773{
 774        addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
 775        addr /= VMAP_BLOCK_SIZE;
 776        return addr;
 777}
 778
 779static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
 780{
 781        struct vmap_block_queue *vbq;
 782        struct vmap_block *vb;
 783        struct vmap_area *va;
 784        unsigned long vb_idx;
 785        int node, err;
 786
 787        node = numa_node_id();
 788
 789        vb = kmalloc_node(sizeof(struct vmap_block),
 790                        gfp_mask & GFP_RECLAIM_MASK, node);
 791        if (unlikely(!vb))
 792                return ERR_PTR(-ENOMEM);
 793
 794        va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
 795                                        VMALLOC_START, VMALLOC_END,
 796                                        node, gfp_mask);
 797        if (IS_ERR(va)) {
 798                kfree(vb);
 799                return ERR_CAST(va);
 800        }
 801
 802        err = radix_tree_preload(gfp_mask);
 803        if (unlikely(err)) {
 804                kfree(vb);
 805                free_vmap_area(va);
 806                return ERR_PTR(err);
 807        }
 808
 809        spin_lock_init(&vb->lock);
 810        vb->va = va;
 811        vb->free = VMAP_BBMAP_BITS;
 812        vb->dirty = 0;
 813        bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
 814        bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
 815        INIT_LIST_HEAD(&vb->free_list);
 816
 817        vb_idx = addr_to_vb_idx(va->va_start);
 818        spin_lock(&vmap_block_tree_lock);
 819        err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
 820        spin_unlock(&vmap_block_tree_lock);
 821        BUG_ON(err);
 822        radix_tree_preload_end();
 823
 824        vbq = &get_cpu_var(vmap_block_queue);
 825        vb->vbq = vbq;
 826        spin_lock(&vbq->lock);
 827        list_add_rcu(&vb->free_list, &vbq->free);
 828        spin_unlock(&vbq->lock);
 829        put_cpu_var(vmap_block_queue);
 830
 831        return vb;
 832}
 833
 834static void free_vmap_block(struct vmap_block *vb)
 835{
 836        struct vmap_block *tmp;
 837        unsigned long vb_idx;
 838
 839        vb_idx = addr_to_vb_idx(vb->va->va_start);
 840        spin_lock(&vmap_block_tree_lock);
 841        tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
 842        spin_unlock(&vmap_block_tree_lock);
 843        BUG_ON(tmp != vb);
 844
 845        free_vmap_area_noflush(vb->va);
 846        kfree_rcu(vb, rcu_head);
 847}
 848
 849static void purge_fragmented_blocks(int cpu)
 850{
 851        LIST_HEAD(purge);
 852        struct vmap_block *vb;
 853        struct vmap_block *n_vb;
 854        struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
 855
 856        rcu_read_lock();
 857        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 858
 859                if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
 860                        continue;
 861
 862                spin_lock(&vb->lock);
 863                if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
 864                        vb->free = 0; /* prevent further allocs after releasing lock */
 865                        vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
 866                        bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
 867                        bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
 868                        spin_lock(&vbq->lock);
 869                        list_del_rcu(&vb->free_list);
 870                        spin_unlock(&vbq->lock);
 871                        spin_unlock(&vb->lock);
 872                        list_add_tail(&vb->purge, &purge);
 873                } else
 874                        spin_unlock(&vb->lock);
 875        }
 876        rcu_read_unlock();
 877
 878        list_for_each_entry_safe(vb, n_vb, &purge, purge) {
 879                list_del(&vb->purge);
 880                free_vmap_block(vb);
 881        }
 882}
 883
 884static void purge_fragmented_blocks_thiscpu(void)
 885{
 886        purge_fragmented_blocks(smp_processor_id());
 887}
 888
 889static void purge_fragmented_blocks_allcpus(void)
 890{
 891        int cpu;
 892
 893        for_each_possible_cpu(cpu)
 894                purge_fragmented_blocks(cpu);
 895}
 896
 897static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
 898{
 899        struct vmap_block_queue *vbq;
 900        struct vmap_block *vb;
 901        unsigned long addr = 0;
 902        unsigned int order;
 903        int purge = 0;
 904
 905        BUG_ON(size & ~PAGE_MASK);
 906        BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 907        order = get_order(size);
 908
 909again:
 910        rcu_read_lock();
 911        vbq = &get_cpu_var(vmap_block_queue);
 912        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 913                int i;
 914
 915                spin_lock(&vb->lock);
 916                if (vb->free < 1UL << order)
 917                        goto next;
 918
 919                i = bitmap_find_free_region(vb->alloc_map,
 920                                                VMAP_BBMAP_BITS, order);
 921
 922                if (i < 0) {
 923                        if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
 924                                /* fragmented and no outstanding allocations */
 925                                BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
 926                                purge = 1;
 927                        }
 928                        goto next;
 929                }
 930                addr = vb->va->va_start + (i << PAGE_SHIFT);
 931                BUG_ON(addr_to_vb_idx(addr) !=
 932                                addr_to_vb_idx(vb->va->va_start));
 933                vb->free -= 1UL << order;
 934                if (vb->free == 0) {
 935                        spin_lock(&vbq->lock);
 936                        list_del_rcu(&vb->free_list);
 937                        spin_unlock(&vbq->lock);
 938                }
 939                spin_unlock(&vb->lock);
 940                break;
 941next:
 942                spin_unlock(&vb->lock);
 943        }
 944
 945        if (purge)
 946                purge_fragmented_blocks_thiscpu();
 947
 948        put_cpu_var(vmap_block_queue);
 949        rcu_read_unlock();
 950
 951        if (!addr) {
 952                vb = new_vmap_block(gfp_mask);
 953                if (IS_ERR(vb))
 954                        return vb;
 955                goto again;
 956        }
 957
 958        return (void *)addr;
 959}
 960
 961static void vb_free(const void *addr, unsigned long size)
 962{
 963        unsigned long offset;
 964        unsigned long vb_idx;
 965        unsigned int order;
 966        struct vmap_block *vb;
 967
 968        BUG_ON(size & ~PAGE_MASK);
 969        BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 970
 971        flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
 972
 973        order = get_order(size);
 974
 975        offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
 976
 977        vb_idx = addr_to_vb_idx((unsigned long)addr);
 978        rcu_read_lock();
 979        vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
 980        rcu_read_unlock();
 981        BUG_ON(!vb);
 982
 983        vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
 984
 985        spin_lock(&vb->lock);
 986        BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
 987
 988        vb->dirty += 1UL << order;
 989        if (vb->dirty == VMAP_BBMAP_BITS) {
 990                BUG_ON(vb->free);
 991                spin_unlock(&vb->lock);
 992                free_vmap_block(vb);
 993        } else
 994                spin_unlock(&vb->lock);
 995}
 996
 997/**
 998 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
 999 *
1000 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1001 * to amortize TLB flushing overheads. What this means is that any page you
1002 * have now, may, in a former life, have been mapped into kernel virtual
1003 * address by the vmap layer and so there might be some CPUs with TLB entries
1004 * still referencing that page (additional to the regular 1:1 kernel mapping).
1005 *
1006 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1007 * be sure that none of the pages we have control over will have any aliases
1008 * from the vmap layer.
1009 */
1010void vm_unmap_aliases(void)
1011{
1012        unsigned long start = ULONG_MAX, end = 0;
1013        int cpu;
1014        int flush = 0;
1015
1016        if (unlikely(!vmap_initialized))
1017                return;
1018
1019        for_each_possible_cpu(cpu) {
1020                struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1021                struct vmap_block *vb;
1022
1023                rcu_read_lock();
1024                list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1025                        int i;
1026
1027                        spin_lock(&vb->lock);
1028                        i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1029                        while (i < VMAP_BBMAP_BITS) {
1030                                unsigned long s, e;
1031                                int j;
1032                                j = find_next_zero_bit(vb->dirty_map,
1033                                        VMAP_BBMAP_BITS, i);
1034
1035                                s = vb->va->va_start + (i << PAGE_SHIFT);
1036                                e = vb->va->va_start + (j << PAGE_SHIFT);
1037                                flush = 1;
1038
1039                                if (s < start)
1040                                        start = s;
1041                                if (e > end)
1042                                        end = e;
1043
1044                                i = j;
1045                                i = find_next_bit(vb->dirty_map,
1046                                                        VMAP_BBMAP_BITS, i);
1047                        }
1048                        spin_unlock(&vb->lock);
1049                }
1050                rcu_read_unlock();
1051        }
1052
1053        __purge_vmap_area_lazy(&start, &end, 1, flush);
1054}
1055EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1056
1057/**
1058 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1059 * @mem: the pointer returned by vm_map_ram
1060 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1061 */
1062void vm_unmap_ram(const void *mem, unsigned int count)
1063{
1064        unsigned long size = count << PAGE_SHIFT;
1065        unsigned long addr = (unsigned long)mem;
1066
1067        BUG_ON(!addr);
1068        BUG_ON(addr < VMALLOC_START);
1069        BUG_ON(addr > VMALLOC_END);
1070        BUG_ON(addr & (PAGE_SIZE-1));
1071
1072        debug_check_no_locks_freed(mem, size);
1073        vmap_debug_free_range(addr, addr+size);
1074
1075        if (likely(count <= VMAP_MAX_ALLOC))
1076                vb_free(mem, size);
1077        else
1078                free_unmap_vmap_area_addr(addr);
1079}
1080EXPORT_SYMBOL(vm_unmap_ram);
1081
1082/**
1083 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1084 * @pages: an array of pointers to the pages to be mapped
1085 * @count: number of pages
1086 * @node: prefer to allocate data structures on this node
1087 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1088 *
1089 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1090 */
1091void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1092{
1093        unsigned long size = count << PAGE_SHIFT;
1094        unsigned long addr;
1095        void *mem;
1096
1097        if (likely(count <= VMAP_MAX_ALLOC)) {
1098                mem = vb_alloc(size, GFP_KERNEL);
1099                if (IS_ERR(mem))
1100                        return NULL;
1101                addr = (unsigned long)mem;
1102        } else {
1103                struct vmap_area *va;
1104                va = alloc_vmap_area(size, PAGE_SIZE,
1105                                VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1106                if (IS_ERR(va))
1107                        return NULL;
1108
1109                addr = va->va_start;
1110                mem = (void *)addr;
1111        }
1112        if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1113                vm_unmap_ram(mem, count);
1114                return NULL;
1115        }
1116        return mem;
1117}
1118EXPORT_SYMBOL(vm_map_ram);
1119
1120/**
1121 * vm_area_register_early - register vmap area early during boot
1122 * @vm: vm_struct to register
1123 * @align: requested alignment
1124 *
1125 * This function is used to register kernel vm area before
1126 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1127 * proper values on entry and other fields should be zero.  On return,
1128 * vm->addr contains the allocated address.
1129 *
1130 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1131 */
1132void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1133{
1134        static size_t vm_init_off __initdata;
1135        unsigned long addr;
1136
1137        addr = ALIGN(VMALLOC_START + vm_init_off, align);
1138        vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1139
1140        vm->addr = (void *)addr;
1141
1142        vm->next = vmlist;
1143        vmlist = vm;
1144}
1145
1146void __init vmalloc_init(void)
1147{
1148        struct vmap_area *va;
1149        struct vm_struct *tmp;
1150        int i;
1151
1152        for_each_possible_cpu(i) {
1153                struct vmap_block_queue *vbq;
1154
1155                vbq = &per_cpu(vmap_block_queue, i);
1156                spin_lock_init(&vbq->lock);
1157                INIT_LIST_HEAD(&vbq->free);
1158        }
1159
1160        /* Import existing vmlist entries. */
1161        for (tmp = vmlist; tmp; tmp = tmp->next) {
1162                va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1163                va->flags = tmp->flags | VM_VM_AREA;
1164                va->va_start = (unsigned long)tmp->addr;
1165                va->va_end = va->va_start + tmp->size;
1166                __insert_vmap_area(va);
1167        }
1168
1169        vmap_area_pcpu_hole = VMALLOC_END;
1170
1171        vmap_initialized = true;
1172}
1173
1174/**
1175 * map_kernel_range_noflush - map kernel VM area with the specified pages
1176 * @addr: start of the VM area to map
1177 * @size: size of the VM area to map
1178 * @prot: page protection flags to use
1179 * @pages: pages to map
1180 *
1181 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1182 * specify should have been allocated using get_vm_area() and its
1183 * friends.
1184 *
1185 * NOTE:
1186 * This function does NOT do any cache flushing.  The caller is
1187 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1188 * before calling this function.
1189 *
1190 * RETURNS:
1191 * The number of pages mapped on success, -errno on failure.
1192 */
1193int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1194                             pgprot_t prot, struct page **pages)
1195{
1196        return vmap_page_range_noflush(addr, addr + size, prot, pages);
1197}
1198
1199/**
1200 * unmap_kernel_range_noflush - unmap kernel VM area
1201 * @addr: start of the VM area to unmap
1202 * @size: size of the VM area to unmap
1203 *
1204 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1205 * specify should have been allocated using get_vm_area() and its
1206 * friends.
1207 *
1208 * NOTE:
1209 * This function does NOT do any cache flushing.  The caller is
1210 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1211 * before calling this function and flush_tlb_kernel_range() after.
1212 */
1213void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1214{
1215        vunmap_page_range(addr, addr + size);
1216}
1217EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1218
1219/**
1220 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1221 * @addr: start of the VM area to unmap
1222 * @size: size of the VM area to unmap
1223 *
1224 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1225 * the unmapping and tlb after.
1226 */
1227void unmap_kernel_range(unsigned long addr, unsigned long size)
1228{
1229        unsigned long end = addr + size;
1230
1231        flush_cache_vunmap(addr, end);
1232        vunmap_page_range(addr, end);
1233        flush_tlb_kernel_range(addr, end);
1234}
1235
1236int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1237{
1238        unsigned long addr = (unsigned long)area->addr;
1239        unsigned long end = addr + area->size - PAGE_SIZE;
1240        int err;
1241
1242        err = vmap_page_range(addr, end, prot, *pages);
1243        if (err > 0) {
1244                *pages += err;
1245                err = 0;
1246        }
1247
1248        return err;
1249}
1250EXPORT_SYMBOL_GPL(map_vm_area);
1251
1252/*** Old vmalloc interfaces ***/
1253DEFINE_RWLOCK(vmlist_lock);
1254struct vm_struct *vmlist;
1255
1256static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1257                              unsigned long flags, void *caller)
1258{
1259        vm->flags = flags;
1260        vm->addr = (void *)va->va_start;
1261        vm->size = va->va_end - va->va_start;
1262        vm->caller = caller;
1263        va->private = vm;
1264        va->flags |= VM_VM_AREA;
1265}
1266
1267static void insert_vmalloc_vmlist(struct vm_struct *vm)
1268{
1269        struct vm_struct *tmp, **p;
1270
1271        vm->flags &= ~VM_UNLIST;
1272        write_lock(&vmlist_lock);
1273        for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1274                if (tmp->addr >= vm->addr)
1275                        break;
1276        }
1277        vm->next = *p;
1278        *p = vm;
1279        write_unlock(&vmlist_lock);
1280}
1281
1282static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1283                              unsigned long flags, void *caller)
1284{
1285        setup_vmalloc_vm(vm, va, flags, caller);
1286        insert_vmalloc_vmlist(vm);
1287}
1288
1289static struct vm_struct *__get_vm_area_node(unsigned long size,
1290                unsigned long align, unsigned long flags, unsigned long start,
1291                unsigned long end, int node, gfp_t gfp_mask, void *caller)
1292{
1293        static struct vmap_area *va;
1294        struct vm_struct *area;
1295
1296        BUG_ON(in_interrupt());
1297        if (flags & VM_IOREMAP) {
1298                int bit = fls(size);
1299
1300                if (bit > IOREMAP_MAX_ORDER)
1301                        bit = IOREMAP_MAX_ORDER;
1302                else if (bit < PAGE_SHIFT)
1303                        bit = PAGE_SHIFT;
1304
1305                align = 1ul << bit;
1306        }
1307
1308        size = PAGE_ALIGN(size);
1309        if (unlikely(!size))
1310                return NULL;
1311
1312        area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1313        if (unlikely(!area))
1314                return NULL;
1315
1316        /*
1317         * We always allocate a guard page.
1318         */
1319        size += PAGE_SIZE;
1320
1321        va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1322        if (IS_ERR(va)) {
1323                kfree(area);
1324                return NULL;
1325        }
1326
1327        /*
1328         * When this function is called from __vmalloc_node_range,
1329         * we do not add vm_struct to vmlist here to avoid
1330         * accessing uninitialized members of vm_struct such as
1331         * pages and nr_pages fields. They will be set later.
1332         * To distinguish it from others, we use a VM_UNLIST flag.
1333         */
1334        if (flags & VM_UNLIST)
1335                setup_vmalloc_vm(area, va, flags, caller);
1336        else
1337                insert_vmalloc_vm(area, va, flags, caller);
1338
1339        return area;
1340}
1341
1342struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1343                                unsigned long start, unsigned long end)
1344{
1345        return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1346                                                __builtin_return_address(0));
1347}
1348EXPORT_SYMBOL_GPL(__get_vm_area);
1349
1350struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1351                                       unsigned long start, unsigned long end,
1352                                       void *caller)
1353{
1354        return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1355                                  caller);
1356}
1357
1358/**
1359 *      get_vm_area  -  reserve a contiguous kernel virtual area
1360 *      @size:          size of the area
1361 *      @flags:         %VM_IOREMAP for I/O mappings or VM_ALLOC
1362 *
1363 *      Search an area of @size in the kernel virtual mapping area,
1364 *      and reserved it for out purposes.  Returns the area descriptor
1365 *      on success or %NULL on failure.
1366 */
1367struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1368{
1369        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1370                                -1, GFP_KERNEL, __builtin_return_address(0));
1371}
1372
1373struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1374                                void *caller)
1375{
1376        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1377                                                -1, GFP_KERNEL, caller);
1378}
1379
1380static struct vm_struct *find_vm_area(const void *addr)
1381{
1382        struct vmap_area *va;
1383
1384        va = find_vmap_area((unsigned long)addr);
1385        if (va && va->flags & VM_VM_AREA)
1386                return va->private;
1387
1388        return NULL;
1389}
1390
1391/**
1392 *      remove_vm_area  -  find and remove a continuous kernel virtual area
1393 *      @addr:          base address
1394 *
1395 *      Search for the kernel VM area starting at @addr, and remove it.
1396 *      This function returns the found VM area, but using it is NOT safe
1397 *      on SMP machines, except for its size or flags.
1398 */
1399struct vm_struct *remove_vm_area(const void *addr)
1400{
1401        struct vmap_area *va;
1402
1403        va = find_vmap_area((unsigned long)addr);
1404        if (va && va->flags & VM_VM_AREA) {
1405                struct vm_struct *vm = va->private;
1406
1407                if (!(vm->flags & VM_UNLIST)) {
1408                        struct vm_struct *tmp, **p;
1409                        /*
1410                         * remove from list and disallow access to
1411                         * this vm_struct before unmap. (address range
1412                         * confliction is maintained by vmap.)
1413                         */
1414                        write_lock(&vmlist_lock);
1415                        for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1416                                ;
1417                        *p = tmp->next;
1418                        write_unlock(&vmlist_lock);
1419                }
1420
1421                vmap_debug_free_range(va->va_start, va->va_end);
1422                free_unmap_vmap_area(va);
1423                vm->size -= PAGE_SIZE;
1424
1425                return vm;
1426        }
1427        return NULL;
1428}
1429
1430static void __vunmap(const void *addr, int deallocate_pages)
1431{
1432        struct vm_struct *area;
1433
1434        if (!addr)
1435                return;
1436
1437        if ((PAGE_SIZE-1) & (unsigned long)addr) {
1438                WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1439                return;
1440        }
1441
1442        area = remove_vm_area(addr);
1443        if (unlikely(!area)) {
1444                WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1445                                addr);
1446                return;
1447        }
1448
1449        debug_check_no_locks_freed(addr, area->size);
1450        debug_check_no_obj_freed(addr, area->size);
1451
1452        if (deallocate_pages) {
1453                int i;
1454
1455                for (i = 0; i < area->nr_pages; i++) {
1456                        struct page *page = area->pages[i];
1457
1458                        BUG_ON(!page);
1459                        __free_page(page);
1460                }
1461
1462                if (area->flags & VM_VPAGES)
1463                        vfree(area->pages);
1464                else
1465                        kfree(area->pages);
1466        }
1467
1468        kfree(area);
1469        return;
1470}
1471
1472/**
1473 *      vfree  -  release memory allocated by vmalloc()
1474 *      @addr:          memory base address
1475 *
1476 *      Free the virtually continuous memory area starting at @addr, as
1477 *      obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1478 *      NULL, no operation is performed.
1479 *
1480 *      Must not be called in interrupt context.
1481 */
1482void vfree(const void *addr)
1483{
1484        BUG_ON(in_interrupt());
1485
1486        kmemleak_free(addr);
1487
1488        __vunmap(addr, 1);
1489}
1490EXPORT_SYMBOL(vfree);
1491
1492/**
1493 *      vunmap  -  release virtual mapping obtained by vmap()
1494 *      @addr:          memory base address
1495 *
1496 *      Free the virtually contiguous memory area starting at @addr,
1497 *      which was created from the page array passed to vmap().
1498 *
1499 *      Must not be called in interrupt context.
1500 */
1501void vunmap(const void *addr)
1502{
1503        BUG_ON(in_interrupt());
1504        might_sleep();
1505        __vunmap(addr, 0);
1506}
1507EXPORT_SYMBOL(vunmap);
1508
1509/**
1510 *      vmap  -  map an array of pages into virtually contiguous space
1511 *      @pages:         array of page pointers
1512 *      @count:         number of pages to map
1513 *      @flags:         vm_area->flags
1514 *      @prot:          page protection for the mapping
1515 *
1516 *      Maps @count pages from @pages into contiguous kernel virtual
1517 *      space.
1518 */
1519void *vmap(struct page **pages, unsigned int count,
1520                unsigned long flags, pgprot_t prot)
1521{
1522        struct vm_struct *area;
1523
1524        might_sleep();
1525
1526        if (count > totalram_pages)
1527                return NULL;
1528
1529        area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1530                                        __builtin_return_address(0));
1531        if (!area)
1532                return NULL;
1533
1534        if (map_vm_area(area, prot, &pages)) {
1535                vunmap(area->addr);
1536                return NULL;
1537        }
1538
1539        return area->addr;
1540}
1541EXPORT_SYMBOL(vmap);
1542
1543static void *__vmalloc_node(unsigned long size, unsigned long align,
1544                            gfp_t gfp_mask, pgprot_t prot,
1545                            int node, void *caller);
1546static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1547                                 pgprot_t prot, int node, void *caller)
1548{
1549        const int order = 0;
1550        struct page **pages;
1551        unsigned int nr_pages, array_size, i;
1552        gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1553
1554        nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1555        array_size = (nr_pages * sizeof(struct page *));
1556
1557        area->nr_pages = nr_pages;
1558        /* Please note that the recursion is strictly bounded. */
1559        if (array_size > PAGE_SIZE) {
1560                pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1561                                PAGE_KERNEL, node, caller);
1562                area->flags |= VM_VPAGES;
1563        } else {
1564                pages = kmalloc_node(array_size, nested_gfp, node);
1565        }
1566        area->pages = pages;
1567        area->caller = caller;
1568        if (!area->pages) {
1569                remove_vm_area(area->addr);
1570                kfree(area);
1571                return NULL;
1572        }
1573
1574        for (i = 0; i < area->nr_pages; i++) {
1575                struct page *page;
1576                gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1577
1578                if (node < 0)
1579                        page = alloc_page(tmp_mask);
1580                else
1581                        page = alloc_pages_node(node, tmp_mask, order);
1582
1583                if (unlikely(!page)) {
1584                        /* Successfully allocated i pages, free them in __vunmap() */
1585                        area->nr_pages = i;
1586                        goto fail;
1587                }
1588                area->pages[i] = page;
1589        }
1590
1591        if (map_vm_area(area, prot, &pages))
1592                goto fail;
1593        return area->addr;
1594
1595fail:
1596        warn_alloc_failed(gfp_mask, order, "vmalloc: allocation failure, "
1597                          "allocated %ld of %ld bytes\n",
1598                          (area->nr_pages*PAGE_SIZE), area->size);
1599        vfree(area->addr);
1600        return NULL;
1601}
1602
1603/**
1604 *      __vmalloc_node_range  -  allocate virtually contiguous memory
1605 *      @size:          allocation size
1606 *      @align:         desired alignment
1607 *      @start:         vm area range start
1608 *      @end:           vm area range end
1609 *      @gfp_mask:      flags for the page level allocator
1610 *      @prot:          protection mask for the allocated pages
1611 *      @node:          node to use for allocation or -1
1612 *      @caller:        caller's return address
1613 *
1614 *      Allocate enough pages to cover @size from the page level
1615 *      allocator with @gfp_mask flags.  Map them into contiguous
1616 *      kernel virtual space, using a pagetable protection of @prot.
1617 */
1618void *__vmalloc_node_range(unsigned long size, unsigned long align,
1619                        unsigned long start, unsigned long end, gfp_t gfp_mask,
1620                        pgprot_t prot, int node, void *caller)
1621{
1622        struct vm_struct *area;
1623        void *addr;
1624        unsigned long real_size = size;
1625
1626        size = PAGE_ALIGN(size);
1627        if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1628                return NULL;
1629
1630        area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNLIST,
1631                                  start, end, node, gfp_mask, caller);
1632
1633        if (!area)
1634                return NULL;
1635
1636        addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1637        if (!addr)
1638                return NULL;
1639
1640        /*
1641         * In this function, newly allocated vm_struct is not added
1642         * to vmlist at __get_vm_area_node(). so, it is added here.
1643         */
1644        insert_vmalloc_vmlist(area);
1645
1646        /*
1647         * A ref_count = 3 is needed because the vm_struct and vmap_area
1648         * structures allocated in the __get_vm_area_node() function contain
1649         * references to the virtual address of the vmalloc'ed block.
1650         */
1651        kmemleak_alloc(addr, real_size, 3, gfp_mask);
1652
1653        return addr;
1654}
1655
1656/**
1657 *      __vmalloc_node  -  allocate virtually contiguous memory
1658 *      @size:          allocation size
1659 *      @align:         desired alignment
1660 *      @gfp_mask:      flags for the page level allocator
1661 *      @prot:          protection mask for the allocated pages
1662 *      @node:          node to use for allocation or -1
1663 *      @caller:        caller's return address
1664 *
1665 *      Allocate enough pages to cover @size from the page level
1666 *      allocator with @gfp_mask flags.  Map them into contiguous
1667 *      kernel virtual space, using a pagetable protection of @prot.
1668 */
1669static void *__vmalloc_node(unsigned long size, unsigned long align,
1670                            gfp_t gfp_mask, pgprot_t prot,
1671                            int node, void *caller)
1672{
1673        return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1674                                gfp_mask, prot, node, caller);
1675}
1676
1677void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1678{
1679        return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1680                                __builtin_return_address(0));
1681}
1682EXPORT_SYMBOL(__vmalloc);
1683
1684static inline void *__vmalloc_node_flags(unsigned long size,
1685                                        int node, gfp_t flags)
1686{
1687        return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1688                                        node, __builtin_return_address(0));
1689}
1690
1691/**
1692 *      vmalloc  -  allocate virtually contiguous memory
1693 *      @size:          allocation size
1694 *      Allocate enough pages to cover @size from the page level
1695 *      allocator and map them into contiguous kernel virtual space.
1696 *
1697 *      For tight control over page level allocator and protection flags
1698 *      use __vmalloc() instead.
1699 */
1700void *vmalloc(unsigned long size)
1701{
1702        return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
1703}
1704EXPORT_SYMBOL(vmalloc);
1705
1706/**
1707 *      vzalloc - allocate virtually contiguous memory with zero fill
1708 *      @size:  allocation size
1709 *      Allocate enough pages to cover @size from the page level
1710 *      allocator and map them into contiguous kernel virtual space.
1711 *      The memory allocated is set to zero.
1712 *
1713 *      For tight control over page level allocator and protection flags
1714 *      use __vmalloc() instead.
1715 */
1716void *vzalloc(unsigned long size)
1717{
1718        return __vmalloc_node_flags(size, -1,
1719                                GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1720}
1721EXPORT_SYMBOL(vzalloc);
1722
1723/**
1724 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1725 * @size: allocation size
1726 *
1727 * The resulting memory area is zeroed so it can be mapped to userspace
1728 * without leaking data.
1729 */
1730void *vmalloc_user(unsigned long size)
1731{
1732        struct vm_struct *area;
1733        void *ret;
1734
1735        ret = __vmalloc_node(size, SHMLBA,
1736                             GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1737                             PAGE_KERNEL, -1, __builtin_return_address(0));
1738        if (ret) {
1739                area = find_vm_area(ret);
1740                area->flags |= VM_USERMAP;
1741        }
1742        return ret;
1743}
1744EXPORT_SYMBOL(vmalloc_user);
1745
1746/**
1747 *      vmalloc_node  -  allocate memory on a specific node
1748 *      @size:          allocation size
1749 *      @node:          numa node
1750 *
1751 *      Allocate enough pages to cover @size from the page level
1752 *      allocator and map them into contiguous kernel virtual space.
1753 *
1754 *      For tight control over page level allocator and protection flags
1755 *      use __vmalloc() instead.
1756 */
1757void *vmalloc_node(unsigned long size, int node)
1758{
1759        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1760                                        node, __builtin_return_address(0));
1761}
1762EXPORT_SYMBOL(vmalloc_node);
1763
1764/**
1765 * vzalloc_node - allocate memory on a specific node with zero fill
1766 * @size:       allocation size
1767 * @node:       numa node
1768 *
1769 * Allocate enough pages to cover @size from the page level
1770 * allocator and map them into contiguous kernel virtual space.
1771 * The memory allocated is set to zero.
1772 *
1773 * For tight control over page level allocator and protection flags
1774 * use __vmalloc_node() instead.
1775 */
1776void *vzalloc_node(unsigned long size, int node)
1777{
1778        return __vmalloc_node_flags(size, node,
1779                         GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1780}
1781EXPORT_SYMBOL(vzalloc_node);
1782
1783#ifndef PAGE_KERNEL_EXEC
1784# define PAGE_KERNEL_EXEC PAGE_KERNEL
1785#endif
1786
1787/**
1788 *      vmalloc_exec  -  allocate virtually contiguous, executable memory
1789 *      @size:          allocation size
1790 *
1791 *      Kernel-internal function to allocate enough pages to cover @size
1792 *      the page level allocator and map them into contiguous and
1793 *      executable kernel virtual space.
1794 *
1795 *      For tight control over page level allocator and protection flags
1796 *      use __vmalloc() instead.
1797 */
1798
1799void *vmalloc_exec(unsigned long size)
1800{
1801        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1802                              -1, __builtin_return_address(0));
1803}
1804
1805#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1806#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1807#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1808#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1809#else
1810#define GFP_VMALLOC32 GFP_KERNEL
1811#endif
1812
1813/**
1814 *      vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1815 *      @size:          allocation size
1816 *
1817 *      Allocate enough 32bit PA addressable pages to cover @size from the
1818 *      page level allocator and map them into contiguous kernel virtual space.
1819 */
1820void *vmalloc_32(unsigned long size)
1821{
1822        return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1823                              -1, __builtin_return_address(0));
1824}
1825EXPORT_SYMBOL(vmalloc_32);
1826
1827/**
1828 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1829 *      @size:          allocation size
1830 *
1831 * The resulting memory area is 32bit addressable and zeroed so it can be
1832 * mapped to userspace without leaking data.
1833 */
1834void *vmalloc_32_user(unsigned long size)
1835{
1836        struct vm_struct *area;
1837        void *ret;
1838
1839        ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1840                             -1, __builtin_return_address(0));
1841        if (ret) {
1842                area = find_vm_area(ret);
1843                area->flags |= VM_USERMAP;
1844        }
1845        return ret;
1846}
1847EXPORT_SYMBOL(vmalloc_32_user);
1848
1849/*
1850 * small helper routine , copy contents to buf from addr.
1851 * If the page is not present, fill zero.
1852 */
1853
1854static int aligned_vread(char *buf, char *addr, unsigned long count)
1855{
1856        struct page *p;
1857        int copied = 0;
1858
1859        while (count) {
1860                unsigned long offset, length;
1861
1862                offset = (unsigned long)addr & ~PAGE_MASK;
1863                length = PAGE_SIZE - offset;
1864                if (length > count)
1865                        length = count;
1866                p = vmalloc_to_page(addr);
1867                /*
1868                 * To do safe access to this _mapped_ area, we need
1869                 * lock. But adding lock here means that we need to add
1870                 * overhead of vmalloc()/vfree() calles for this _debug_
1871                 * interface, rarely used. Instead of that, we'll use
1872                 * kmap() and get small overhead in this access function.
1873                 */
1874                if (p) {
1875                        /*
1876                         * we can expect USER0 is not used (see vread/vwrite's
1877                         * function description)
1878                         */
1879                        void *map = kmap_atomic(p, KM_USER0);
1880                        memcpy(buf, map + offset, length);
1881                        kunmap_atomic(map, KM_USER0);
1882                } else
1883                        memset(buf, 0, length);
1884
1885                addr += length;
1886                buf += length;
1887                copied += length;
1888                count -= length;
1889        }
1890        return copied;
1891}
1892
1893static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1894{
1895        struct page *p;
1896        int copied = 0;
1897
1898        while (count) {
1899                unsigned long offset, length;
1900
1901                offset = (unsigned long)addr & ~PAGE_MASK;
1902                length = PAGE_SIZE - offset;
1903                if (length > count)
1904                        length = count;
1905                p = vmalloc_to_page(addr);
1906                /*
1907                 * To do safe access to this _mapped_ area, we need
1908                 * lock. But adding lock here means that we need to add
1909                 * overhead of vmalloc()/vfree() calles for this _debug_
1910                 * interface, rarely used. Instead of that, we'll use
1911                 * kmap() and get small overhead in this access function.
1912                 */
1913                if (p) {
1914                        /*
1915                         * we can expect USER0 is not used (see vread/vwrite's
1916                         * function description)
1917                         */
1918                        void *map = kmap_atomic(p, KM_USER0);
1919                        memcpy(map + offset, buf, length);
1920                        kunmap_atomic(map, KM_USER0);
1921                }
1922                addr += length;
1923                buf += length;
1924                copied += length;
1925                count -= length;
1926        }
1927        return copied;
1928}
1929
1930/**
1931 *      vread() -  read vmalloc area in a safe way.
1932 *      @buf:           buffer for reading data
1933 *      @addr:          vm address.
1934 *      @count:         number of bytes to be read.
1935 *
1936 *      Returns # of bytes which addr and buf should be increased.
1937 *      (same number to @count). Returns 0 if [addr...addr+count) doesn't
1938 *      includes any intersect with alive vmalloc area.
1939 *
1940 *      This function checks that addr is a valid vmalloc'ed area, and
1941 *      copy data from that area to a given buffer. If the given memory range
1942 *      of [addr...addr+count) includes some valid address, data is copied to
1943 *      proper area of @buf. If there are memory holes, they'll be zero-filled.
1944 *      IOREMAP area is treated as memory hole and no copy is done.
1945 *
1946 *      If [addr...addr+count) doesn't includes any intersects with alive
1947 *      vm_struct area, returns 0.
1948 *      @buf should be kernel's buffer. Because this function uses KM_USER0,
1949 *      the caller should guarantee KM_USER0 is not used.
1950 *
1951 *      Note: In usual ops, vread() is never necessary because the caller
1952 *      should know vmalloc() area is valid and can use memcpy().
1953 *      This is for routines which have to access vmalloc area without
1954 *      any informaion, as /dev/kmem.
1955 *
1956 */
1957
1958long vread(char *buf, char *addr, unsigned long count)
1959{
1960        struct vm_struct *tmp;
1961        char *vaddr, *buf_start = buf;
1962        unsigned long buflen = count;
1963        unsigned long n;
1964
1965        /* Don't allow overflow */
1966        if ((unsigned long) addr + count < count)
1967                count = -(unsigned long) addr;
1968
1969        read_lock(&vmlist_lock);
1970        for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1971                vaddr = (char *) tmp->addr;
1972                if (addr >= vaddr + tmp->size - PAGE_SIZE)
1973                        continue;
1974                while (addr < vaddr) {
1975                        if (count == 0)
1976                                goto finished;
1977                        *buf = '\0';
1978                        buf++;
1979                        addr++;
1980                        count--;
1981                }
1982                n = vaddr + tmp->size - PAGE_SIZE - addr;
1983                if (n > count)
1984                        n = count;
1985                if (!(tmp->flags & VM_IOREMAP))
1986                        aligned_vread(buf, addr, n);
1987                else /* IOREMAP area is treated as memory hole */
1988                        memset(buf, 0, n);
1989                buf += n;
1990                addr += n;
1991                count -= n;
1992        }
1993finished:
1994        read_unlock(&vmlist_lock);
1995
1996        if (buf == buf_start)
1997                return 0;
1998        /* zero-fill memory holes */
1999        if (buf != buf_start + buflen)
2000                memset(buf, 0, buflen - (buf - buf_start));
2001
2002        return buflen;
2003}
2004
2005/**
2006 *      vwrite() -  write vmalloc area in a safe way.
2007 *      @buf:           buffer for source data
2008 *      @addr:          vm address.
2009 *      @count:         number of bytes to be read.
2010 *
2011 *      Returns # of bytes which addr and buf should be incresed.
2012 *      (same number to @count).
2013 *      If [addr...addr+count) doesn't includes any intersect with valid
2014 *      vmalloc area, returns 0.
2015 *
2016 *      This function checks that addr is a valid vmalloc'ed area, and
2017 *      copy data from a buffer to the given addr. If specified range of
2018 *      [addr...addr+count) includes some valid address, data is copied from
2019 *      proper area of @buf. If there are memory holes, no copy to hole.
2020 *      IOREMAP area is treated as memory hole and no copy is done.
2021 *
2022 *      If [addr...addr+count) doesn't includes any intersects with alive
2023 *      vm_struct area, returns 0.
2024 *      @buf should be kernel's buffer. Because this function uses KM_USER0,
2025 *      the caller should guarantee KM_USER0 is not used.
2026 *
2027 *      Note: In usual ops, vwrite() is never necessary because the caller
2028 *      should know vmalloc() area is valid and can use memcpy().
2029 *      This is for routines which have to access vmalloc area without
2030 *      any informaion, as /dev/kmem.
2031 */
2032
2033long vwrite(char *buf, char *addr, unsigned long count)
2034{
2035        struct vm_struct *tmp;
2036        char *vaddr;
2037        unsigned long n, buflen;
2038        int copied = 0;
2039
2040        /* Don't allow overflow */
2041        if ((unsigned long) addr + count < count)
2042                count = -(unsigned long) addr;
2043        buflen = count;
2044
2045        read_lock(&vmlist_lock);
2046        for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2047                vaddr = (char *) tmp->addr;
2048                if (addr >= vaddr + tmp->size - PAGE_SIZE)
2049                        continue;
2050                while (addr < vaddr) {
2051                        if (count == 0)
2052                                goto finished;
2053                        buf++;
2054                        addr++;
2055                        count--;
2056                }
2057                n = vaddr + tmp->size - PAGE_SIZE - addr;
2058                if (n > count)
2059                        n = count;
2060                if (!(tmp->flags & VM_IOREMAP)) {
2061                        aligned_vwrite(buf, addr, n);
2062                        copied++;
2063                }
2064                buf += n;
2065                addr += n;
2066                count -= n;
2067        }
2068finished:
2069        read_unlock(&vmlist_lock);
2070        if (!copied)
2071                return 0;
2072        return buflen;
2073}
2074
2075/**
2076 *      remap_vmalloc_range  -  map vmalloc pages to userspace
2077 *      @vma:           vma to cover (map full range of vma)
2078 *      @addr:          vmalloc memory
2079 *      @pgoff:         number of pages into addr before first page to map
2080 *
2081 *      Returns:        0 for success, -Exxx on failure
2082 *
2083 *      This function checks that addr is a valid vmalloc'ed area, and
2084 *      that it is big enough to cover the vma. Will return failure if
2085 *      that criteria isn't met.
2086 *
2087 *      Similar to remap_pfn_range() (see mm/memory.c)
2088 */
2089int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2090                                                unsigned long pgoff)
2091{
2092        struct vm_struct *area;
2093        unsigned long uaddr = vma->vm_start;
2094        unsigned long usize = vma->vm_end - vma->vm_start;
2095
2096        if ((PAGE_SIZE-1) & (unsigned long)addr)
2097                return -EINVAL;
2098
2099        area = find_vm_area(addr);
2100        if (!area)
2101                return -EINVAL;
2102
2103        if (!(area->flags & VM_USERMAP))
2104                return -EINVAL;
2105
2106        if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2107                return -EINVAL;
2108
2109        addr += pgoff << PAGE_SHIFT;
2110        do {
2111                struct page *page = vmalloc_to_page(addr);
2112                int ret;
2113
2114                ret = vm_insert_page(vma, uaddr, page);
2115                if (ret)
2116                        return ret;
2117
2118                uaddr += PAGE_SIZE;
2119                addr += PAGE_SIZE;
2120                usize -= PAGE_SIZE;
2121        } while (usize > 0);
2122
2123        /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2124        vma->vm_flags |= VM_RESERVED;
2125
2126        return 0;
2127}
2128EXPORT_SYMBOL(remap_vmalloc_range);
2129
2130/*
2131 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2132 * have one.
2133 */
2134void  __attribute__((weak)) vmalloc_sync_all(void)
2135{
2136}
2137
2138
2139static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2140{
2141        /* apply_to_page_range() does all the hard work. */
2142        return 0;
2143}
2144
2145/**
2146 *      alloc_vm_area - allocate a range of kernel address space
2147 *      @size:          size of the area
2148 *
2149 *      Returns:        NULL on failure, vm_struct on success
2150 *
2151 *      This function reserves a range of kernel address space, and
2152 *      allocates pagetables to map that range.  No actual mappings
2153 *      are created.  If the kernel address space is not shared
2154 *      between processes, it syncs the pagetable across all
2155 *      processes.
2156 */
2157struct vm_struct *alloc_vm_area(size_t size)
2158{
2159        struct vm_struct *area;
2160
2161        area = get_vm_area_caller(size, VM_IOREMAP,
2162                                __builtin_return_address(0));
2163        if (area == NULL)
2164                return NULL;
2165
2166        /*
2167         * This ensures that page tables are constructed for this region
2168         * of kernel virtual address space and mapped into init_mm.
2169         */
2170        if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2171                                area->size, f, NULL)) {
2172                free_vm_area(area);
2173                return NULL;
2174        }
2175
2176        /*
2177         * If the allocated address space is passed to a hypercall
2178         * before being used then we cannot rely on a page fault to
2179         * trigger an update of the page tables.  So sync all the page
2180         * tables here.
2181         */
2182        vmalloc_sync_all();
2183
2184        return area;
2185}
2186EXPORT_SYMBOL_GPL(alloc_vm_area);
2187
2188void free_vm_area(struct vm_struct *area)
2189{
2190        struct vm_struct *ret;
2191        ret = remove_vm_area(area->addr);
2192        BUG_ON(ret != area);
2193        kfree(area);
2194}
2195EXPORT_SYMBOL_GPL(free_vm_area);
2196
2197#ifdef CONFIG_SMP
2198static struct vmap_area *node_to_va(struct rb_node *n)
2199{
2200        return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2201}
2202
2203/**
2204 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2205 * @end: target address
2206 * @pnext: out arg for the next vmap_area
2207 * @pprev: out arg for the previous vmap_area
2208 *
2209 * Returns: %true if either or both of next and prev are found,
2210 *          %false if no vmap_area exists
2211 *
2212 * Find vmap_areas end addresses of which enclose @end.  ie. if not
2213 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2214 */
2215static bool pvm_find_next_prev(unsigned long end,
2216                               struct vmap_area **pnext,
2217                               struct vmap_area **pprev)
2218{
2219        struct rb_node *n = vmap_area_root.rb_node;
2220        struct vmap_area *va = NULL;
2221
2222        while (n) {
2223                va = rb_entry(n, struct vmap_area, rb_node);
2224                if (end < va->va_end)
2225                        n = n->rb_left;
2226                else if (end > va->va_end)
2227                        n = n->rb_right;
2228                else
2229                        break;
2230        }
2231
2232        if (!va)
2233                return false;
2234
2235        if (va->va_end > end) {
2236                *pnext = va;
2237                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2238        } else {
2239                *pprev = va;
2240                *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2241        }
2242        return true;
2243}
2244
2245/**
2246 * pvm_determine_end - find the highest aligned address between two vmap_areas
2247 * @pnext: in/out arg for the next vmap_area
2248 * @pprev: in/out arg for the previous vmap_area
2249 * @align: alignment
2250 *
2251 * Returns: determined end address
2252 *
2253 * Find the highest aligned address between *@pnext and *@pprev below
2254 * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2255 * down address is between the end addresses of the two vmap_areas.
2256 *
2257 * Please note that the address returned by this function may fall
2258 * inside *@pnext vmap_area.  The caller is responsible for checking
2259 * that.
2260 */
2261static unsigned long pvm_determine_end(struct vmap_area **pnext,
2262                                       struct vmap_area **pprev,
2263                                       unsigned long align)
2264{
2265        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2266        unsigned long addr;
2267
2268        if (*pnext)
2269                addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2270        else
2271                addr = vmalloc_end;
2272
2273        while (*pprev && (*pprev)->va_end > addr) {
2274                *pnext = *pprev;
2275                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2276        }
2277
2278        return addr;
2279}
2280
2281/**
2282 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2283 * @offsets: array containing offset of each area
2284 * @sizes: array containing size of each area
2285 * @nr_vms: the number of areas to allocate
2286 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2287 *
2288 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2289 *          vm_structs on success, %NULL on failure
2290 *
2291 * Percpu allocator wants to use congruent vm areas so that it can
2292 * maintain the offsets among percpu areas.  This function allocates
2293 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
2294 * be scattered pretty far, distance between two areas easily going up
2295 * to gigabytes.  To avoid interacting with regular vmallocs, these
2296 * areas are allocated from top.
2297 *
2298 * Despite its complicated look, this allocator is rather simple.  It
2299 * does everything top-down and scans areas from the end looking for
2300 * matching slot.  While scanning, if any of the areas overlaps with
2301 * existing vmap_area, the base address is pulled down to fit the
2302 * area.  Scanning is repeated till all the areas fit and then all
2303 * necessary data structres are inserted and the result is returned.
2304 */
2305struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2306                                     const size_t *sizes, int nr_vms,
2307                                     size_t align)
2308{
2309        const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2310        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2311        struct vmap_area **vas, *prev, *next;
2312        struct vm_struct **vms;
2313        int area, area2, last_area, term_area;
2314        unsigned long base, start, end, last_end;
2315        bool purged = false;
2316
2317        /* verify parameters and allocate data structures */
2318        BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2319        for (last_area = 0, area = 0; area < nr_vms; area++) {
2320                start = offsets[area];
2321                end = start + sizes[area];
2322
2323                /* is everything aligned properly? */
2324                BUG_ON(!IS_ALIGNED(offsets[area], align));
2325                BUG_ON(!IS_ALIGNED(sizes[area], align));
2326
2327                /* detect the area with the highest address */
2328                if (start > offsets[last_area])
2329                        last_area = area;
2330
2331                for (area2 = 0; area2 < nr_vms; area2++) {
2332                        unsigned long start2 = offsets[area2];
2333                        unsigned long end2 = start2 + sizes[area2];
2334
2335                        if (area2 == area)
2336                                continue;
2337
2338                        BUG_ON(start2 >= start && start2 < end);
2339                        BUG_ON(end2 <= end && end2 > start);
2340                }
2341        }
2342        last_end = offsets[last_area] + sizes[last_area];
2343
2344        if (vmalloc_end - vmalloc_start < last_end) {
2345                WARN_ON(true);
2346                return NULL;
2347        }
2348
2349        vms = kzalloc(sizeof(vms[0]) * nr_vms, GFP_KERNEL);
2350        vas = kzalloc(sizeof(vas[0]) * nr_vms, GFP_KERNEL);
2351        if (!vas || !vms)
2352                goto err_free;
2353
2354        for (area = 0; area < nr_vms; area++) {
2355                vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2356                vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2357                if (!vas[area] || !vms[area])
2358                        goto err_free;
2359        }
2360retry:
2361        spin_lock(&vmap_area_lock);
2362
2363        /* start scanning - we scan from the top, begin with the last area */
2364        area = term_area = last_area;
2365        start = offsets[area];
2366        end = start + sizes[area];
2367
2368        if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2369                base = vmalloc_end - last_end;
2370                goto found;
2371        }
2372        base = pvm_determine_end(&next, &prev, align) - end;
2373
2374        while (true) {
2375                BUG_ON(next && next->va_end <= base + end);
2376                BUG_ON(prev && prev->va_end > base + end);
2377
2378                /*
2379                 * base might have underflowed, add last_end before
2380                 * comparing.
2381                 */
2382                if (base + last_end < vmalloc_start + last_end) {
2383                        spin_unlock(&vmap_area_lock);
2384                        if (!purged) {
2385                                purge_vmap_area_lazy();
2386                                purged = true;
2387                                goto retry;
2388                        }
2389                        goto err_free;
2390                }
2391
2392                /*
2393                 * If next overlaps, move base downwards so that it's
2394                 * right below next and then recheck.
2395                 */
2396                if (next && next->va_start < base + end) {
2397                        base = pvm_determine_end(&next, &prev, align) - end;
2398                        term_area = area;
2399                        continue;
2400                }
2401
2402                /*
2403                 * If prev overlaps, shift down next and prev and move
2404                 * base so that it's right below new next and then
2405                 * recheck.
2406                 */
2407                if (prev && prev->va_end > base + start)  {
2408                        next = prev;
2409                        prev = node_to_va(rb_prev(&next->rb_node));
2410                        base = pvm_determine_end(&next, &prev, align) - end;
2411                        term_area = area;
2412                        continue;
2413                }
2414
2415                /*
2416                 * This area fits, move on to the previous one.  If
2417                 * the previous one is the terminal one, we're done.
2418                 */
2419                area = (area + nr_vms - 1) % nr_vms;
2420                if (area == term_area)
2421                        break;
2422                start = offsets[area];
2423                end = start + sizes[area];
2424                pvm_find_next_prev(base + end, &next, &prev);
2425        }
2426found:
2427        /* we've found a fitting base, insert all va's */
2428        for (area = 0; area < nr_vms; area++) {
2429                struct vmap_area *va = vas[area];
2430
2431                va->va_start = base + offsets[area];
2432                va->va_end = va->va_start + sizes[area];
2433                __insert_vmap_area(va);
2434        }
2435
2436        vmap_area_pcpu_hole = base + offsets[last_area];
2437
2438        spin_unlock(&vmap_area_lock);
2439
2440        /* insert all vm's */
2441        for (area = 0; area < nr_vms; area++)
2442                insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2443                                  pcpu_get_vm_areas);
2444
2445        kfree(vas);
2446        return vms;
2447
2448err_free:
2449        for (area = 0; area < nr_vms; area++) {
2450                if (vas)
2451                        kfree(vas[area]);
2452                if (vms)
2453                        kfree(vms[area]);
2454        }
2455        kfree(vas);
2456        kfree(vms);
2457        return NULL;
2458}
2459
2460/**
2461 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2462 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2463 * @nr_vms: the number of allocated areas
2464 *
2465 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2466 */
2467void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2468{
2469        int i;
2470
2471        for (i = 0; i < nr_vms; i++)
2472                free_vm_area(vms[i]);
2473        kfree(vms);
2474}
2475#endif  /* CONFIG_SMP */
2476
2477#ifdef CONFIG_PROC_FS
2478static void *s_start(struct seq_file *m, loff_t *pos)
2479        __acquires(&vmlist_lock)
2480{
2481        loff_t n = *pos;
2482        struct vm_struct *v;
2483
2484        read_lock(&vmlist_lock);
2485        v = vmlist;
2486        while (n > 0 && v) {
2487                n--;
2488                v = v->next;
2489        }
2490        if (!n)
2491                return v;
2492
2493        return NULL;
2494
2495}
2496
2497static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2498{
2499        struct vm_struct *v = p;
2500
2501        ++*pos;
2502        return v->next;
2503}
2504
2505static void s_stop(struct seq_file *m, void *p)
2506        __releases(&vmlist_lock)
2507{
2508        read_unlock(&vmlist_lock);
2509}
2510
2511static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2512{
2513        if (NUMA_BUILD) {
2514                unsigned int nr, *counters = m->private;
2515
2516                if (!counters)
2517                        return;
2518
2519                memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2520
2521                for (nr = 0; nr < v->nr_pages; nr++)
2522                        counters[page_to_nid(v->pages[nr])]++;
2523
2524                for_each_node_state(nr, N_HIGH_MEMORY)
2525                        if (counters[nr])
2526                                seq_printf(m, " N%u=%u", nr, counters[nr]);
2527        }
2528}
2529
2530static int s_show(struct seq_file *m, void *p)
2531{
2532        struct vm_struct *v = p;
2533
2534        seq_printf(m, "0x%p-0x%p %7ld",
2535                v->addr, v->addr + v->size, v->size);
2536
2537        if (v->caller)
2538                seq_printf(m, " %pS", v->caller);
2539
2540        if (v->nr_pages)
2541                seq_printf(m, " pages=%d", v->nr_pages);
2542
2543        if (v->phys_addr)
2544                seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2545
2546        if (v->flags & VM_IOREMAP)
2547                seq_printf(m, " ioremap");
2548
2549        if (v->flags & VM_ALLOC)
2550                seq_printf(m, " vmalloc");
2551
2552        if (v->flags & VM_MAP)
2553                seq_printf(m, " vmap");
2554
2555        if (v->flags & VM_USERMAP)
2556                seq_printf(m, " user");
2557
2558        if (v->flags & VM_VPAGES)
2559                seq_printf(m, " vpages");
2560
2561        show_numa_info(m, v);
2562        seq_putc(m, '\n');
2563        return 0;
2564}
2565
2566static const struct seq_operations vmalloc_op = {
2567        .start = s_start,
2568        .next = s_next,
2569        .stop = s_stop,
2570        .show = s_show,
2571};
2572
2573static int vmalloc_open(struct inode *inode, struct file *file)
2574{
2575        unsigned int *ptr = NULL;
2576        int ret;
2577
2578        if (NUMA_BUILD) {
2579                ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2580                if (ptr == NULL)
2581                        return -ENOMEM;
2582        }
2583        ret = seq_open(file, &vmalloc_op);
2584        if (!ret) {
2585                struct seq_file *m = file->private_data;
2586                m->private = ptr;
2587        } else
2588                kfree(ptr);
2589        return ret;
2590}
2591
2592static const struct file_operations proc_vmalloc_operations = {
2593        .open           = vmalloc_open,
2594        .read           = seq_read,
2595        .llseek         = seq_lseek,
2596        .release        = seq_release_private,
2597};
2598
2599static int __init proc_vmalloc_init(void)
2600{
2601        proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2602        return 0;
2603}
2604module_init(proc_vmalloc_init);
2605#endif
2606
2607
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.