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        struct vm_struct *vm;
 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_add_early - add vmap area early during boot
1122 * @vm: vm_struct to add
1123 *
1124 * This function is used to add fixed kernel vm area to vmlist before
1125 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
1126 * should contain proper values and the other fields should be zero.
1127 *
1128 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1129 */
1130void __init vm_area_add_early(struct vm_struct *vm)
1131{
1132        struct vm_struct *tmp, **p;
1133
1134        BUG_ON(vmap_initialized);
1135        for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1136                if (tmp->addr >= vm->addr) {
1137                        BUG_ON(tmp->addr < vm->addr + vm->size);
1138                        break;
1139                } else
1140                        BUG_ON(tmp->addr + tmp->size > vm->addr);
1141        }
1142        vm->next = *p;
1143        *p = vm;
1144}
1145
1146/**
1147 * vm_area_register_early - register vmap area early during boot
1148 * @vm: vm_struct to register
1149 * @align: requested alignment
1150 *
1151 * This function is used to register kernel vm area before
1152 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1153 * proper values on entry and other fields should be zero.  On return,
1154 * vm->addr contains the allocated address.
1155 *
1156 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1157 */
1158void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1159{
1160        static size_t vm_init_off __initdata;
1161        unsigned long addr;
1162
1163        addr = ALIGN(VMALLOC_START + vm_init_off, align);
1164        vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1165
1166        vm->addr = (void *)addr;
1167
1168        vm_area_add_early(vm);
1169}
1170
1171void __init vmalloc_init(void)
1172{
1173        struct vmap_area *va;
1174        struct vm_struct *tmp;
1175        int i;
1176
1177        for_each_possible_cpu(i) {
1178                struct vmap_block_queue *vbq;
1179
1180                vbq = &per_cpu(vmap_block_queue, i);
1181                spin_lock_init(&vbq->lock);
1182                INIT_LIST_HEAD(&vbq->free);
1183        }
1184
1185        /* Import existing vmlist entries. */
1186        for (tmp = vmlist; tmp; tmp = tmp->next) {
1187                va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1188                va->flags = VM_VM_AREA;
1189                va->va_start = (unsigned long)tmp->addr;
1190                va->va_end = va->va_start + tmp->size;
1191                va->vm = tmp;
1192                __insert_vmap_area(va);
1193        }
1194
1195        vmap_area_pcpu_hole = VMALLOC_END;
1196
1197        vmap_initialized = true;
1198}
1199
1200/**
1201 * map_kernel_range_noflush - map kernel VM area with the specified pages
1202 * @addr: start of the VM area to map
1203 * @size: size of the VM area to map
1204 * @prot: page protection flags to use
1205 * @pages: pages to map
1206 *
1207 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1208 * specify should have been allocated using get_vm_area() and its
1209 * friends.
1210 *
1211 * NOTE:
1212 * This function does NOT do any cache flushing.  The caller is
1213 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1214 * before calling this function.
1215 *
1216 * RETURNS:
1217 * The number of pages mapped on success, -errno on failure.
1218 */
1219int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1220                             pgprot_t prot, struct page **pages)
1221{
1222        return vmap_page_range_noflush(addr, addr + size, prot, pages);
1223}
1224
1225/**
1226 * unmap_kernel_range_noflush - unmap kernel VM area
1227 * @addr: start of the VM area to unmap
1228 * @size: size of the VM area to unmap
1229 *
1230 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1231 * specify should have been allocated using get_vm_area() and its
1232 * friends.
1233 *
1234 * NOTE:
1235 * This function does NOT do any cache flushing.  The caller is
1236 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1237 * before calling this function and flush_tlb_kernel_range() after.
1238 */
1239void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1240{
1241        vunmap_page_range(addr, addr + size);
1242}
1243EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1244
1245/**
1246 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1247 * @addr: start of the VM area to unmap
1248 * @size: size of the VM area to unmap
1249 *
1250 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1251 * the unmapping and tlb after.
1252 */
1253void unmap_kernel_range(unsigned long addr, unsigned long size)
1254{
1255        unsigned long end = addr + size;
1256
1257        flush_cache_vunmap(addr, end);
1258        vunmap_page_range(addr, end);
1259        flush_tlb_kernel_range(addr, end);
1260}
1261
1262int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1263{
1264        unsigned long addr = (unsigned long)area->addr;
1265        unsigned long end = addr + area->size - PAGE_SIZE;
1266        int err;
1267
1268        err = vmap_page_range(addr, end, prot, *pages);
1269        if (err > 0) {
1270                *pages += err;
1271                err = 0;
1272        }
1273
1274        return err;
1275}
1276EXPORT_SYMBOL_GPL(map_vm_area);
1277
1278/*** Old vmalloc interfaces ***/
1279DEFINE_RWLOCK(vmlist_lock);
1280struct vm_struct *vmlist;
1281
1282static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1283                              unsigned long flags, void *caller)
1284{
1285        vm->flags = flags;
1286        vm->addr = (void *)va->va_start;
1287        vm->size = va->va_end - va->va_start;
1288        vm->caller = caller;
1289        va->vm = vm;
1290        va->flags |= VM_VM_AREA;
1291}
1292
1293static void insert_vmalloc_vmlist(struct vm_struct *vm)
1294{
1295        struct vm_struct *tmp, **p;
1296
1297        vm->flags &= ~VM_UNLIST;
1298        write_lock(&vmlist_lock);
1299        for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1300                if (tmp->addr >= vm->addr)
1301                        break;
1302        }
1303        vm->next = *p;
1304        *p = vm;
1305        write_unlock(&vmlist_lock);
1306}
1307
1308static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1309                              unsigned long flags, void *caller)
1310{
1311        setup_vmalloc_vm(vm, va, flags, caller);
1312        insert_vmalloc_vmlist(vm);
1313}
1314
1315static struct vm_struct *__get_vm_area_node(unsigned long size,
1316                unsigned long align, unsigned long flags, unsigned long start,
1317                unsigned long end, int node, gfp_t gfp_mask, void *caller)
1318{
1319        struct vmap_area *va;
1320        struct vm_struct *area;
1321
1322        BUG_ON(in_interrupt());
1323        if (flags & VM_IOREMAP) {
1324                int bit = fls(size);
1325
1326                if (bit > IOREMAP_MAX_ORDER)
1327                        bit = IOREMAP_MAX_ORDER;
1328                else if (bit < PAGE_SHIFT)
1329                        bit = PAGE_SHIFT;
1330
1331                align = 1ul << bit;
1332        }
1333
1334        size = PAGE_ALIGN(size);
1335        if (unlikely(!size))
1336                return NULL;
1337
1338        area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1339        if (unlikely(!area))
1340                return NULL;
1341
1342        /*
1343         * We always allocate a guard page.
1344         */
1345        size += PAGE_SIZE;
1346
1347        va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1348        if (IS_ERR(va)) {
1349                kfree(area);
1350                return NULL;
1351        }
1352
1353        /*
1354         * When this function is called from __vmalloc_node_range,
1355         * we do not add vm_struct to vmlist here to avoid
1356         * accessing uninitialized members of vm_struct such as
1357         * pages and nr_pages fields. They will be set later.
1358         * To distinguish it from others, we use a VM_UNLIST flag.
1359         */
1360        if (flags & VM_UNLIST)
1361                setup_vmalloc_vm(area, va, flags, caller);
1362        else
1363                insert_vmalloc_vm(area, va, flags, caller);
1364
1365        return area;
1366}
1367
1368struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1369                                unsigned long start, unsigned long end)
1370{
1371        return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1372                                                __builtin_return_address(0));
1373}
1374EXPORT_SYMBOL_GPL(__get_vm_area);
1375
1376struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1377                                       unsigned long start, unsigned long end,
1378                                       void *caller)
1379{
1380        return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1381                                  caller);
1382}
1383
1384/**
1385 *      get_vm_area  -  reserve a contiguous kernel virtual area
1386 *      @size:          size of the area
1387 *      @flags:         %VM_IOREMAP for I/O mappings or VM_ALLOC
1388 *
1389 *      Search an area of @size in the kernel virtual mapping area,
1390 *      and reserved it for out purposes.  Returns the area descriptor
1391 *      on success or %NULL on failure.
1392 */
1393struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1394{
1395        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1396                                -1, GFP_KERNEL, __builtin_return_address(0));
1397}
1398
1399struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1400                                void *caller)
1401{
1402        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1403                                                -1, GFP_KERNEL, caller);
1404}
1405
1406static struct vm_struct *find_vm_area(const void *addr)
1407{
1408        struct vmap_area *va;
1409
1410        va = find_vmap_area((unsigned long)addr);
1411        if (va && va->flags & VM_VM_AREA)
1412                return va->vm;
1413
1414        return NULL;
1415}
1416
1417/**
1418 *      remove_vm_area  -  find and remove a continuous kernel virtual area
1419 *      @addr:          base address
1420 *
1421 *      Search for the kernel VM area starting at @addr, and remove it.
1422 *      This function returns the found VM area, but using it is NOT safe
1423 *      on SMP machines, except for its size or flags.
1424 */
1425struct vm_struct *remove_vm_area(const void *addr)
1426{
1427        struct vmap_area *va;
1428
1429        va = find_vmap_area((unsigned long)addr);
1430        if (va && va->flags & VM_VM_AREA) {
1431                struct vm_struct *vm = va->vm;
1432
1433                if (!(vm->flags & VM_UNLIST)) {
1434                        struct vm_struct *tmp, **p;
1435                        /*
1436                         * remove from list and disallow access to
1437                         * this vm_struct before unmap. (address range
1438                         * confliction is maintained by vmap.)
1439                         */
1440                        write_lock(&vmlist_lock);
1441                        for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1442                                ;
1443                        *p = tmp->next;
1444                        write_unlock(&vmlist_lock);
1445                }
1446
1447                vmap_debug_free_range(va->va_start, va->va_end);
1448                free_unmap_vmap_area(va);
1449                vm->size -= PAGE_SIZE;
1450
1451                return vm;
1452        }
1453        return NULL;
1454}
1455
1456static void __vunmap(const void *addr, int deallocate_pages)
1457{
1458        struct vm_struct *area;
1459
1460        if (!addr)
1461                return;
1462
1463        if ((PAGE_SIZE-1) & (unsigned long)addr) {
1464                WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1465                return;
1466        }
1467
1468        area = remove_vm_area(addr);
1469        if (unlikely(!area)) {
1470                WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1471                                addr);
1472                return;
1473        }
1474
1475        debug_check_no_locks_freed(addr, area->size);
1476        debug_check_no_obj_freed(addr, area->size);
1477
1478        if (deallocate_pages) {
1479                int i;
1480
1481                for (i = 0; i < area->nr_pages; i++) {
1482                        struct page *page = area->pages[i];
1483
1484                        BUG_ON(!page);
1485                        __free_page(page);
1486                }
1487
1488                if (area->flags & VM_VPAGES)
1489                        vfree(area->pages);
1490                else
1491                        kfree(area->pages);
1492        }
1493
1494        kfree(area);
1495        return;
1496}
1497
1498/**
1499 *      vfree  -  release memory allocated by vmalloc()
1500 *      @addr:          memory base address
1501 *
1502 *      Free the virtually continuous memory area starting at @addr, as
1503 *      obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1504 *      NULL, no operation is performed.
1505 *
1506 *      Must not be called in interrupt context.
1507 */
1508void vfree(const void *addr)
1509{
1510        BUG_ON(in_interrupt());
1511
1512        kmemleak_free(addr);
1513
1514        __vunmap(addr, 1);
1515}
1516EXPORT_SYMBOL(vfree);
1517
1518/**
1519 *      vunmap  -  release virtual mapping obtained by vmap()
1520 *      @addr:          memory base address
1521 *
1522 *      Free the virtually contiguous memory area starting at @addr,
1523 *      which was created from the page array passed to vmap().
1524 *
1525 *      Must not be called in interrupt context.
1526 */
1527void vunmap(const void *addr)
1528{
1529        BUG_ON(in_interrupt());
1530        might_sleep();
1531        __vunmap(addr, 0);
1532}
1533EXPORT_SYMBOL(vunmap);
1534
1535/**
1536 *      vmap  -  map an array of pages into virtually contiguous space
1537 *      @pages:         array of page pointers
1538 *      @count:         number of pages to map
1539 *      @flags:         vm_area->flags
1540 *      @prot:          page protection for the mapping
1541 *
1542 *      Maps @count pages from @pages into contiguous kernel virtual
1543 *      space.
1544 */
1545void *vmap(struct page **pages, unsigned int count,
1546                unsigned long flags, pgprot_t prot)
1547{
1548        struct vm_struct *area;
1549
1550        might_sleep();
1551
1552        if (count > totalram_pages)
1553                return NULL;
1554
1555        area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1556                                        __builtin_return_address(0));
1557        if (!area)
1558                return NULL;
1559
1560        if (map_vm_area(area, prot, &pages)) {
1561                vunmap(area->addr);
1562                return NULL;
1563        }
1564
1565        return area->addr;
1566}
1567EXPORT_SYMBOL(vmap);
1568
1569static void *__vmalloc_node(unsigned long size, unsigned long align,
1570                            gfp_t gfp_mask, pgprot_t prot,
1571                            int node, void *caller);
1572static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1573                                 pgprot_t prot, int node, void *caller)
1574{
1575        const int order = 0;
1576        struct page **pages;
1577        unsigned int nr_pages, array_size, i;
1578        gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1579
1580        nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1581        array_size = (nr_pages * sizeof(struct page *));
1582
1583        area->nr_pages = nr_pages;
1584        /* Please note that the recursion is strictly bounded. */
1585        if (array_size > PAGE_SIZE) {
1586                pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1587                                PAGE_KERNEL, node, caller);
1588                area->flags |= VM_VPAGES;
1589        } else {
1590                pages = kmalloc_node(array_size, nested_gfp, node);
1591        }
1592        area->pages = pages;
1593        area->caller = caller;
1594        if (!area->pages) {
1595                remove_vm_area(area->addr);
1596                kfree(area);
1597                return NULL;
1598        }
1599
1600        for (i = 0; i < area->nr_pages; i++) {
1601                struct page *page;
1602                gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1603
1604                if (node < 0)
1605                        page = alloc_page(tmp_mask);
1606                else
1607                        page = alloc_pages_node(node, tmp_mask, order);
1608
1609                if (unlikely(!page)) {
1610                        /* Successfully allocated i pages, free them in __vunmap() */
1611                        area->nr_pages = i;
1612                        goto fail;
1613                }
1614                area->pages[i] = page;
1615        }
1616
1617        if (map_vm_area(area, prot, &pages))
1618                goto fail;
1619        return area->addr;
1620
1621fail:
1622        warn_alloc_failed(gfp_mask, order,
1623                          "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1624                          (area->nr_pages*PAGE_SIZE), area->size);
1625        vfree(area->addr);
1626        return NULL;
1627}
1628
1629/**
1630 *      __vmalloc_node_range  -  allocate virtually contiguous memory
1631 *      @size:          allocation size
1632 *      @align:         desired alignment
1633 *      @start:         vm area range start
1634 *      @end:           vm area range end
1635 *      @gfp_mask:      flags for the page level allocator
1636 *      @prot:          protection mask for the allocated pages
1637 *      @node:          node to use for allocation or -1
1638 *      @caller:        caller's return address
1639 *
1640 *      Allocate enough pages to cover @size from the page level
1641 *      allocator with @gfp_mask flags.  Map them into contiguous
1642 *      kernel virtual space, using a pagetable protection of @prot.
1643 */
1644void *__vmalloc_node_range(unsigned long size, unsigned long align,
1645                        unsigned long start, unsigned long end, gfp_t gfp_mask,
1646                        pgprot_t prot, int node, void *caller)
1647{
1648        struct vm_struct *area;
1649        void *addr;
1650        unsigned long real_size = size;
1651
1652        size = PAGE_ALIGN(size);
1653        if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1654                goto fail;
1655
1656        area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNLIST,
1657                                  start, end, node, gfp_mask, caller);
1658        if (!area)
1659                goto fail;
1660
1661        addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1662        if (!addr)
1663                return NULL;
1664
1665        /*
1666         * In this function, newly allocated vm_struct is not added
1667         * to vmlist at __get_vm_area_node(). so, it is added here.
1668         */
1669        insert_vmalloc_vmlist(area);
1670
1671        /*
1672         * A ref_count = 3 is needed because the vm_struct and vmap_area
1673         * structures allocated in the __get_vm_area_node() function contain
1674         * references to the virtual address of the vmalloc'ed block.
1675         */
1676        kmemleak_alloc(addr, real_size, 3, gfp_mask);
1677
1678        return addr;
1679
1680fail:
1681        warn_alloc_failed(gfp_mask, 0,
1682                          "vmalloc: allocation failure: %lu bytes\n",
1683                          real_size);
1684        return NULL;
1685}
1686
1687/**
1688 *      __vmalloc_node  -  allocate virtually contiguous memory
1689 *      @size:          allocation size
1690 *      @align:         desired alignment
1691 *      @gfp_mask:      flags for the page level allocator
1692 *      @prot:          protection mask for the allocated pages
1693 *      @node:          node to use for allocation or -1
1694 *      @caller:        caller's return address
1695 *
1696 *      Allocate enough pages to cover @size from the page level
1697 *      allocator with @gfp_mask flags.  Map them into contiguous
1698 *      kernel virtual space, using a pagetable protection of @prot.
1699 */
1700static void *__vmalloc_node(unsigned long size, unsigned long align,
1701                            gfp_t gfp_mask, pgprot_t prot,
1702                            int node, void *caller)
1703{
1704        return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1705                                gfp_mask, prot, node, caller);
1706}
1707
1708void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1709{
1710        return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1711                                __builtin_return_address(0));
1712}
1713EXPORT_SYMBOL(__vmalloc);
1714
1715static inline void *__vmalloc_node_flags(unsigned long size,
1716                                        int node, gfp_t flags)
1717{
1718        return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1719                                        node, __builtin_return_address(0));
1720}
1721
1722/**
1723 *      vmalloc  -  allocate virtually contiguous memory
1724 *      @size:          allocation size
1725 *      Allocate enough pages to cover @size from the page level
1726 *      allocator and map them into contiguous kernel virtual space.
1727 *
1728 *      For tight control over page level allocator and protection flags
1729 *      use __vmalloc() instead.
1730 */
1731void *vmalloc(unsigned long size)
1732{
1733        return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
1734}
1735EXPORT_SYMBOL(vmalloc);
1736
1737/**
1738 *      vzalloc - allocate virtually contiguous memory with zero fill
1739 *      @size:  allocation size
1740 *      Allocate enough pages to cover @size from the page level
1741 *      allocator and map them into contiguous kernel virtual space.
1742 *      The memory allocated is set to zero.
1743 *
1744 *      For tight control over page level allocator and protection flags
1745 *      use __vmalloc() instead.
1746 */
1747void *vzalloc(unsigned long size)
1748{
1749        return __vmalloc_node_flags(size, -1,
1750                                GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1751}
1752EXPORT_SYMBOL(vzalloc);
1753
1754/**
1755 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1756 * @size: allocation size
1757 *
1758 * The resulting memory area is zeroed so it can be mapped to userspace
1759 * without leaking data.
1760 */
1761void *vmalloc_user(unsigned long size)
1762{
1763        struct vm_struct *area;
1764        void *ret;
1765
1766        ret = __vmalloc_node(size, SHMLBA,
1767                             GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1768                             PAGE_KERNEL, -1, __builtin_return_address(0));
1769        if (ret) {
1770                area = find_vm_area(ret);
1771                area->flags |= VM_USERMAP;
1772        }
1773        return ret;
1774}
1775EXPORT_SYMBOL(vmalloc_user);
1776
1777/**
1778 *      vmalloc_node  -  allocate memory on a specific node
1779 *      @size:          allocation size
1780 *      @node:          numa node
1781 *
1782 *      Allocate enough pages to cover @size from the page level
1783 *      allocator and map them into contiguous kernel virtual space.
1784 *
1785 *      For tight control over page level allocator and protection flags
1786 *      use __vmalloc() instead.
1787 */
1788void *vmalloc_node(unsigned long size, int node)
1789{
1790        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1791                                        node, __builtin_return_address(0));
1792}
1793EXPORT_SYMBOL(vmalloc_node);
1794
1795/**
1796 * vzalloc_node - allocate memory on a specific node with zero fill
1797 * @size:       allocation size
1798 * @node:       numa node
1799 *
1800 * Allocate enough pages to cover @size from the page level
1801 * allocator and map them into contiguous kernel virtual space.
1802 * The memory allocated is set to zero.
1803 *
1804 * For tight control over page level allocator and protection flags
1805 * use __vmalloc_node() instead.
1806 */
1807void *vzalloc_node(unsigned long size, int node)
1808{
1809        return __vmalloc_node_flags(size, node,
1810                         GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1811}
1812EXPORT_SYMBOL(vzalloc_node);
1813
1814#ifndef PAGE_KERNEL_EXEC
1815# define PAGE_KERNEL_EXEC PAGE_KERNEL
1816#endif
1817
1818/**
1819 *      vmalloc_exec  -  allocate virtually contiguous, executable memory
1820 *      @size:          allocation size
1821 *
1822 *      Kernel-internal function to allocate enough pages to cover @size
1823 *      the page level allocator and map them into contiguous and
1824 *      executable kernel virtual space.
1825 *
1826 *      For tight control over page level allocator and protection flags
1827 *      use __vmalloc() instead.
1828 */
1829
1830void *vmalloc_exec(unsigned long size)
1831{
1832        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1833                              -1, __builtin_return_address(0));
1834}
1835
1836#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1837#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1838#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1839#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1840#else
1841#define GFP_VMALLOC32 GFP_KERNEL
1842#endif
1843
1844/**
1845 *      vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1846 *      @size:          allocation size
1847 *
1848 *      Allocate enough 32bit PA addressable pages to cover @size from the
1849 *      page level allocator and map them into contiguous kernel virtual space.
1850 */
1851void *vmalloc_32(unsigned long size)
1852{
1853        return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1854                              -1, __builtin_return_address(0));
1855}
1856EXPORT_SYMBOL(vmalloc_32);
1857
1858/**
1859 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1860 *      @size:          allocation size
1861 *
1862 * The resulting memory area is 32bit addressable and zeroed so it can be
1863 * mapped to userspace without leaking data.
1864 */
1865void *vmalloc_32_user(unsigned long size)
1866{
1867        struct vm_struct *area;
1868        void *ret;
1869
1870        ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1871                             -1, __builtin_return_address(0));
1872        if (ret) {
1873                area = find_vm_area(ret);
1874                area->flags |= VM_USERMAP;
1875        }
1876        return ret;
1877}
1878EXPORT_SYMBOL(vmalloc_32_user);
1879
1880/*
1881 * small helper routine , copy contents to buf from addr.
1882 * If the page is not present, fill zero.
1883 */
1884
1885static int aligned_vread(char *buf, char *addr, unsigned long count)
1886{
1887        struct page *p;
1888        int copied = 0;
1889
1890        while (count) {
1891                unsigned long offset, length;
1892
1893                offset = (unsigned long)addr & ~PAGE_MASK;
1894                length = PAGE_SIZE - offset;
1895                if (length > count)
1896                        length = count;
1897                p = vmalloc_to_page(addr);
1898                /*
1899                 * To do safe access to this _mapped_ area, we need
1900                 * lock. But adding lock here means that we need to add
1901                 * overhead of vmalloc()/vfree() calles for this _debug_
1902                 * interface, rarely used. Instead of that, we'll use
1903                 * kmap() and get small overhead in this access function.
1904                 */
1905                if (p) {
1906                        /*
1907                         * we can expect USER0 is not used (see vread/vwrite's
1908                         * function description)
1909                         */
1910                        void *map = kmap_atomic(p);
1911                        memcpy(buf, map + offset, length);
1912                        kunmap_atomic(map);
1913                } else
1914                        memset(buf, 0, length);
1915
1916                addr += length;
1917                buf += length;
1918                copied += length;
1919                count -= length;
1920        }
1921        return copied;
1922}
1923
1924static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1925{
1926        struct page *p;
1927        int copied = 0;
1928
1929        while (count) {
1930                unsigned long offset, length;
1931
1932                offset = (unsigned long)addr & ~PAGE_MASK;
1933                length = PAGE_SIZE - offset;
1934                if (length > count)
1935                        length = count;
1936                p = vmalloc_to_page(addr);
1937                /*
1938                 * To do safe access to this _mapped_ area, we need
1939                 * lock. But adding lock here means that we need to add
1940                 * overhead of vmalloc()/vfree() calles for this _debug_
1941                 * interface, rarely used. Instead of that, we'll use
1942                 * kmap() and get small overhead in this access function.
1943                 */
1944                if (p) {
1945                        /*
1946                         * we can expect USER0 is not used (see vread/vwrite's
1947                         * function description)
1948                         */
1949                        void *map = kmap_atomic(p);
1950                        memcpy(map + offset, buf, length);
1951                        kunmap_atomic(map);
1952                }
1953                addr += length;
1954                buf += length;
1955                copied += length;
1956                count -= length;
1957        }
1958        return copied;
1959}
1960
1961/**
1962 *      vread() -  read vmalloc area in a safe way.
1963 *      @buf:           buffer for reading data
1964 *      @addr:          vm address.
1965 *      @count:         number of bytes to be read.
1966 *
1967 *      Returns # of bytes which addr and buf should be increased.
1968 *      (same number to @count). Returns 0 if [addr...addr+count) doesn't
1969 *      includes any intersect with alive vmalloc area.
1970 *
1971 *      This function checks that addr is a valid vmalloc'ed area, and
1972 *      copy data from that area to a given buffer. If the given memory range
1973 *      of [addr...addr+count) includes some valid address, data is copied to
1974 *      proper area of @buf. If there are memory holes, they'll be zero-filled.
1975 *      IOREMAP area is treated as memory hole and no copy is done.
1976 *
1977 *      If [addr...addr+count) doesn't includes any intersects with alive
1978 *      vm_struct area, returns 0.
1979 *      @buf should be kernel's buffer. Because this function uses KM_USER0,
1980 *      the caller should guarantee KM_USER0 is not used.
1981 *
1982 *      Note: In usual ops, vread() is never necessary because the caller
1983 *      should know vmalloc() area is valid and can use memcpy().
1984 *      This is for routines which have to access vmalloc area without
1985 *      any informaion, as /dev/kmem.
1986 *
1987 */
1988
1989long vread(char *buf, char *addr, unsigned long count)
1990{
1991        struct vm_struct *tmp;
1992        char *vaddr, *buf_start = buf;
1993        unsigned long buflen = count;
1994        unsigned long n;
1995
1996        /* Don't allow overflow */
1997        if ((unsigned long) addr + count < count)
1998                count = -(unsigned long) addr;
1999
2000        read_lock(&vmlist_lock);
2001        for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2002                vaddr = (char *) tmp->addr;
2003                if (addr >= vaddr + tmp->size - PAGE_SIZE)
2004                        continue;
2005                while (addr < vaddr) {
2006                        if (count == 0)
2007                                goto finished;
2008                        *buf = '\0';
2009                        buf++;
2010                        addr++;
2011                        count--;
2012                }
2013                n = vaddr + tmp->size - PAGE_SIZE - addr;
2014                if (n > count)
2015                        n = count;
2016                if (!(tmp->flags & VM_IOREMAP))
2017                        aligned_vread(buf, addr, n);
2018                else /* IOREMAP area is treated as memory hole */
2019                        memset(buf, 0, n);
2020                buf += n;
2021                addr += n;
2022                count -= n;
2023        }
2024finished:
2025        read_unlock(&vmlist_lock);
2026
2027        if (buf == buf_start)
2028                return 0;
2029        /* zero-fill memory holes */
2030        if (buf != buf_start + buflen)
2031                memset(buf, 0, buflen - (buf - buf_start));
2032
2033        return buflen;
2034}
2035
2036/**
2037 *      vwrite() -  write vmalloc area in a safe way.
2038 *      @buf:           buffer for source data
2039 *      @addr:          vm address.
2040 *      @count:         number of bytes to be read.
2041 *
2042 *      Returns # of bytes which addr and buf should be incresed.
2043 *      (same number to @count).
2044 *      If [addr...addr+count) doesn't includes any intersect with valid
2045 *      vmalloc area, returns 0.
2046 *
2047 *      This function checks that addr is a valid vmalloc'ed area, and
2048 *      copy data from a buffer to the given addr. If specified range of
2049 *      [addr...addr+count) includes some valid address, data is copied from
2050 *      proper area of @buf. If there are memory holes, no copy to hole.
2051 *      IOREMAP area is treated as memory hole and no copy is done.
2052 *
2053 *      If [addr...addr+count) doesn't includes any intersects with alive
2054 *      vm_struct area, returns 0.
2055 *      @buf should be kernel's buffer. Because this function uses KM_USER0,
2056 *      the caller should guarantee KM_USER0 is not used.
2057 *
2058 *      Note: In usual ops, vwrite() is never necessary because the caller
2059 *      should know vmalloc() area is valid and can use memcpy().
2060 *      This is for routines which have to access vmalloc area without
2061 *      any informaion, as /dev/kmem.
2062 */
2063
2064long vwrite(char *buf, char *addr, unsigned long count)
2065{
2066        struct vm_struct *tmp;
2067        char *vaddr;
2068        unsigned long n, buflen;
2069        int copied = 0;
2070
2071        /* Don't allow overflow */
2072        if ((unsigned long) addr + count < count)
2073                count = -(unsigned long) addr;
2074        buflen = count;
2075
2076        read_lock(&vmlist_lock);
2077        for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2078                vaddr = (char *) tmp->addr;
2079                if (addr >= vaddr + tmp->size - PAGE_SIZE)
2080                        continue;
2081                while (addr < vaddr) {
2082                        if (count == 0)
2083                                goto finished;
2084                        buf++;
2085                        addr++;
2086                        count--;
2087                }
2088                n = vaddr + tmp->size - PAGE_SIZE - addr;
2089                if (n > count)
2090                        n = count;
2091                if (!(tmp->flags & VM_IOREMAP)) {
2092                        aligned_vwrite(buf, addr, n);
2093                        copied++;
2094                }
2095                buf += n;
2096                addr += n;
2097                count -= n;
2098        }
2099finished:
2100        read_unlock(&vmlist_lock);
2101        if (!copied)
2102                return 0;
2103        return buflen;
2104}
2105
2106/**
2107 *      remap_vmalloc_range  -  map vmalloc pages to userspace
2108 *      @vma:           vma to cover (map full range of vma)
2109 *      @addr:          vmalloc memory
2110 *      @pgoff:         number of pages into addr before first page to map
2111 *
2112 *      Returns:        0 for success, -Exxx on failure
2113 *
2114 *      This function checks that addr is a valid vmalloc'ed area, and
2115 *      that it is big enough to cover the vma. Will return failure if
2116 *      that criteria isn't met.
2117 *
2118 *      Similar to remap_pfn_range() (see mm/memory.c)
2119 */
2120int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2121                                                unsigned long pgoff)
2122{
2123        struct vm_struct *area;
2124        unsigned long uaddr = vma->vm_start;
2125        unsigned long usize = vma->vm_end - vma->vm_start;
2126
2127        if ((PAGE_SIZE-1) & (unsigned long)addr)
2128                return -EINVAL;
2129
2130        area = find_vm_area(addr);
2131        if (!area)
2132                return -EINVAL;
2133
2134        if (!(area->flags & VM_USERMAP))
2135                return -EINVAL;
2136
2137        if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2138                return -EINVAL;
2139
2140        addr += pgoff << PAGE_SHIFT;
2141        do {
2142                struct page *page = vmalloc_to_page(addr);
2143                int ret;
2144
2145                ret = vm_insert_page(vma, uaddr, page);
2146                if (ret)
2147                        return ret;
2148
2149                uaddr += PAGE_SIZE;
2150                addr += PAGE_SIZE;
2151                usize -= PAGE_SIZE;
2152        } while (usize > 0);
2153
2154        /* Prevent "things" like memory migration? VM_flags need a cleanup... */
2155        vma->vm_flags |= VM_RESERVED;
2156
2157        return 0;
2158}
2159EXPORT_SYMBOL(remap_vmalloc_range);
2160
2161/*
2162 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2163 * have one.
2164 */
2165void  __attribute__((weak)) vmalloc_sync_all(void)
2166{
2167}
2168
2169
2170static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2171{
2172        pte_t ***p = data;
2173
2174        if (p) {
2175                *(*p) = pte;
2176                (*p)++;
2177        }
2178        return 0;
2179}
2180
2181/**
2182 *      alloc_vm_area - allocate a range of kernel address space
2183 *      @size:          size of the area
2184 *      @ptes:          returns the PTEs for the address space
2185 *
2186 *      Returns:        NULL on failure, vm_struct on success
2187 *
2188 *      This function reserves a range of kernel address space, and
2189 *      allocates pagetables to map that range.  No actual mappings
2190 *      are created.
2191 *
2192 *      If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2193 *      allocated for the VM area are returned.
2194 */
2195struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2196{
2197        struct vm_struct *area;
2198
2199        area = get_vm_area_caller(size, VM_IOREMAP,
2200                                __builtin_return_address(0));
2201        if (area == NULL)
2202                return NULL;
2203
2204        /*
2205         * This ensures that page tables are constructed for this region
2206         * of kernel virtual address space and mapped into init_mm.
2207         */
2208        if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2209                                size, f, ptes ? &ptes : NULL)) {
2210                free_vm_area(area);
2211                return NULL;
2212        }
2213
2214        return area;
2215}
2216EXPORT_SYMBOL_GPL(alloc_vm_area);
2217
2218void free_vm_area(struct vm_struct *area)
2219{
2220        struct vm_struct *ret;
2221        ret = remove_vm_area(area->addr);
2222        BUG_ON(ret != area);
2223        kfree(area);
2224}
2225EXPORT_SYMBOL_GPL(free_vm_area);
2226
2227#ifdef CONFIG_SMP
2228static struct vmap_area *node_to_va(struct rb_node *n)
2229{
2230        return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2231}
2232
2233/**
2234 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2235 * @end: target address
2236 * @pnext: out arg for the next vmap_area
2237 * @pprev: out arg for the previous vmap_area
2238 *
2239 * Returns: %true if either or both of next and prev are found,
2240 *          %false if no vmap_area exists
2241 *
2242 * Find vmap_areas end addresses of which enclose @end.  ie. if not
2243 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2244 */
2245static bool pvm_find_next_prev(unsigned long end,
2246                               struct vmap_area **pnext,
2247                               struct vmap_area **pprev)
2248{
2249        struct rb_node *n = vmap_area_root.rb_node;
2250        struct vmap_area *va = NULL;
2251
2252        while (n) {
2253                va = rb_entry(n, struct vmap_area, rb_node);
2254                if (end < va->va_end)
2255                        n = n->rb_left;
2256                else if (end > va->va_end)
2257                        n = n->rb_right;
2258                else
2259                        break;
2260        }
2261
2262        if (!va)
2263                return false;
2264
2265        if (va->va_end > end) {
2266                *pnext = va;
2267                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2268        } else {
2269                *pprev = va;
2270                *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2271        }
2272        return true;
2273}
2274
2275/**
2276 * pvm_determine_end - find the highest aligned address between two vmap_areas
2277 * @pnext: in/out arg for the next vmap_area
2278 * @pprev: in/out arg for the previous vmap_area
2279 * @align: alignment
2280 *
2281 * Returns: determined end address
2282 *
2283 * Find the highest aligned address between *@pnext and *@pprev below
2284 * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2285 * down address is between the end addresses of the two vmap_areas.
2286 *
2287 * Please note that the address returned by this function may fall
2288 * inside *@pnext vmap_area.  The caller is responsible for checking
2289 * that.
2290 */
2291static unsigned long pvm_determine_end(struct vmap_area **pnext,
2292                                       struct vmap_area **pprev,
2293                                       unsigned long align)
2294{
2295        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2296        unsigned long addr;
2297
2298        if (*pnext)
2299                addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2300        else
2301                addr = vmalloc_end;
2302
2303        while (*pprev && (*pprev)->va_end > addr) {
2304                *pnext = *pprev;
2305                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2306        }
2307
2308        return addr;
2309}
2310
2311/**
2312 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2313 * @offsets: array containing offset of each area
2314 * @sizes: array containing size of each area
2315 * @nr_vms: the number of areas to allocate
2316 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2317 *
2318 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2319 *          vm_structs on success, %NULL on failure
2320 *
2321 * Percpu allocator wants to use congruent vm areas so that it can
2322 * maintain the offsets among percpu areas.  This function allocates
2323 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
2324 * be scattered pretty far, distance between two areas easily going up
2325 * to gigabytes.  To avoid interacting with regular vmallocs, these
2326 * areas are allocated from top.
2327 *
2328 * Despite its complicated look, this allocator is rather simple.  It
2329 * does everything top-down and scans areas from the end looking for
2330 * matching slot.  While scanning, if any of the areas overlaps with
2331 * existing vmap_area, the base address is pulled down to fit the
2332 * area.  Scanning is repeated till all the areas fit and then all
2333 * necessary data structres are inserted and the result is returned.
2334 */
2335struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2336                                     const size_t *sizes, int nr_vms,
2337                                     size_t align)
2338{
2339        const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2340        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2341        struct vmap_area **vas, *prev, *next;
2342        struct vm_struct **vms;
2343        int area, area2, last_area, term_area;
2344        unsigned long base, start, end, last_end;
2345        bool purged = false;
2346
2347        /* verify parameters and allocate data structures */
2348        BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2349        for (last_area = 0, area = 0; area < nr_vms; area++) {
2350                start = offsets[area];
2351                end = start + sizes[area];
2352
2353                /* is everything aligned properly? */
2354                BUG_ON(!IS_ALIGNED(offsets[area], align));
2355                BUG_ON(!IS_ALIGNED(sizes[area], align));
2356
2357                /* detect the area with the highest address */
2358                if (start > offsets[last_area])
2359                        last_area = area;
2360
2361                for (area2 = 0; area2 < nr_vms; area2++) {
2362                        unsigned long start2 = offsets[area2];
2363                        unsigned long end2 = start2 + sizes[area2];
2364
2365                        if (area2 == area)
2366                                continue;
2367
2368                        BUG_ON(start2 >= start && start2 < end);
2369                        BUG_ON(end2 <= end && end2 > start);
2370                }
2371        }
2372        last_end = offsets[last_area] + sizes[last_area];
2373
2374        if (vmalloc_end - vmalloc_start < last_end) {
2375                WARN_ON(true);
2376                return NULL;
2377        }
2378
2379        vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2380        vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2381        if (!vas || !vms)
2382                goto err_free2;
2383
2384        for (area = 0; area < nr_vms; area++) {
2385                vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2386                vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2387                if (!vas[area] || !vms[area])
2388                        goto err_free;
2389        }
2390retry:
2391        spin_lock(&vmap_area_lock);
2392
2393        /* start scanning - we scan from the top, begin with the last area */
2394        area = term_area = last_area;
2395        start = offsets[area];
2396        end = start + sizes[area];
2397
2398        if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2399                base = vmalloc_end - last_end;
2400                goto found;
2401        }
2402        base = pvm_determine_end(&next, &prev, align) - end;
2403
2404        while (true) {
2405                BUG_ON(next && next->va_end <= base + end);
2406                BUG_ON(prev && prev->va_end > base + end);
2407
2408                /*
2409                 * base might have underflowed, add last_end before
2410                 * comparing.
2411                 */
2412                if (base + last_end < vmalloc_start + last_end) {
2413                        spin_unlock(&vmap_area_lock);
2414                        if (!purged) {
2415                                purge_vmap_area_lazy();
2416                                purged = true;
2417                                goto retry;
2418                        }
2419                        goto err_free;
2420                }
2421
2422                /*
2423                 * If next overlaps, move base downwards so that it's
2424                 * right below next and then recheck.
2425                 */
2426                if (next && next->va_start < base + end) {
2427                        base = pvm_determine_end(&next, &prev, align) - end;
2428                        term_area = area;
2429                        continue;
2430                }
2431
2432                /*
2433                 * If prev overlaps, shift down next and prev and move
2434                 * base so that it's right below new next and then
2435                 * recheck.
2436                 */
2437                if (prev && prev->va_end > base + start)  {
2438                        next = prev;
2439                        prev = node_to_va(rb_prev(&next->rb_node));
2440                        base = pvm_determine_end(&next, &prev, align) - end;
2441                        term_area = area;
2442                        continue;
2443                }
2444
2445                /*
2446                 * This area fits, move on to the previous one.  If
2447                 * the previous one is the terminal one, we're done.
2448                 */
2449                area = (area + nr_vms - 1) % nr_vms;
2450                if (area == term_area)
2451                        break;
2452                start = offsets[area];
2453                end = start + sizes[area];
2454                pvm_find_next_prev(base + end, &next, &prev);
2455        }
2456found:
2457        /* we've found a fitting base, insert all va's */
2458        for (area = 0; area < nr_vms; area++) {
2459                struct vmap_area *va = vas[area];
2460
2461                va->va_start = base + offsets[area];
2462                va->va_end = va->va_start + sizes[area];
2463                __insert_vmap_area(va);
2464        }
2465
2466        vmap_area_pcpu_hole = base + offsets[last_area];
2467
2468        spin_unlock(&vmap_area_lock);
2469
2470        /* insert all vm's */
2471        for (area = 0; area < nr_vms; area++)
2472                insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2473                                  pcpu_get_vm_areas);
2474
2475        kfree(vas);
2476        return vms;
2477
2478err_free:
2479        for (area = 0; area < nr_vms; area++) {
2480                kfree(vas[area]);
2481                kfree(vms[area]);
2482        }
2483err_free2:
2484        kfree(vas);
2485        kfree(vms);
2486        return NULL;
2487}
2488
2489/**
2490 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2491 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2492 * @nr_vms: the number of allocated areas
2493 *
2494 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2495 */
2496void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2497{
2498        int i;
2499
2500        for (i = 0; i < nr_vms; i++)
2501                free_vm_area(vms[i]);
2502        kfree(vms);
2503}
2504#endif  /* CONFIG_SMP */
2505
2506#ifdef CONFIG_PROC_FS
2507static void *s_start(struct seq_file *m, loff_t *pos)
2508        __acquires(&vmlist_lock)
2509{
2510        loff_t n = *pos;
2511        struct vm_struct *v;
2512
2513        read_lock(&vmlist_lock);
2514        v = vmlist;
2515        while (n > 0 && v) {
2516                n--;
2517                v = v->next;
2518        }
2519        if (!n)
2520                return v;
2521
2522        return NULL;
2523
2524}
2525
2526static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2527{
2528        struct vm_struct *v = p;
2529
2530        ++*pos;
2531        return v->next;
2532}
2533
2534static void s_stop(struct seq_file *m, void *p)
2535        __releases(&vmlist_lock)
2536{
2537        read_unlock(&vmlist_lock);
2538}
2539
2540static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2541{
2542        if (NUMA_BUILD) {
2543                unsigned int nr, *counters = m->private;
2544
2545                if (!counters)
2546                        return;
2547
2548                memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2549
2550                for (nr = 0; nr < v->nr_pages; nr++)
2551                        counters[page_to_nid(v->pages[nr])]++;
2552
2553                for_each_node_state(nr, N_HIGH_MEMORY)
2554                        if (counters[nr])
2555                                seq_printf(m, " N%u=%u", nr, counters[nr]);
2556        }
2557}
2558
2559static int s_show(struct seq_file *m, void *p)
2560{
2561        struct vm_struct *v = p;
2562
2563        seq_printf(m, "0x%p-0x%p %7ld",
2564                v->addr, v->addr + v->size, v->size);
2565
2566        if (v->caller)
2567                seq_printf(m, " %pS", v->caller);
2568
2569        if (v->nr_pages)
2570                seq_printf(m, " pages=%d", v->nr_pages);
2571
2572        if (v->phys_addr)
2573                seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2574
2575        if (v->flags & VM_IOREMAP)
2576                seq_printf(m, " ioremap");
2577
2578        if (v->flags & VM_ALLOC)
2579                seq_printf(m, " vmalloc");
2580
2581        if (v->flags & VM_MAP)
2582                seq_printf(m, " vmap");
2583
2584        if (v->flags & VM_USERMAP)
2585                seq_printf(m, " user");
2586
2587        if (v->flags & VM_VPAGES)
2588                seq_printf(m, " vpages");
2589
2590        show_numa_info(m, v);
2591        seq_putc(m, '\n');
2592        return 0;
2593}
2594
2595static const struct seq_operations vmalloc_op = {
2596        .start = s_start,
2597        .next = s_next,
2598        .stop = s_stop,
2599        .show = s_show,
2600};
2601
2602static int vmalloc_open(struct inode *inode, struct file *file)
2603{
2604        unsigned int *ptr = NULL;
2605        int ret;
2606
2607        if (NUMA_BUILD) {
2608                ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2609                if (ptr == NULL)
2610                        return -ENOMEM;
2611        }
2612        ret = seq_open(file, &vmalloc_op);
2613        if (!ret) {
2614                struct seq_file *m = file->private_data;
2615                m->private = ptr;
2616        } else
2617                kfree(ptr);
2618        return ret;
2619}
2620
2621static const struct file_operations proc_vmalloc_operations = {
2622        .open           = vmalloc_open,
2623        .read           = seq_read,
2624        .llseek         = seq_lseek,
2625        .release        = seq_release_private,
2626};
2627
2628static int __init proc_vmalloc_init(void)
2629{
2630        proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2631        return 0;
2632}
2633module_init(proc_vmalloc_init);
2634#endif
2635
2636
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.