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                if (list_is_last(&first->list, &vmap_area_list))
 417                        goto found;
 418
 419                first = list_entry(first->list.next,
 420                                struct vmap_area, list);
 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        if (WARN_ON(size == 0)) {
 908                /*
 909                 * Allocating 0 bytes isn't what caller wants since
 910                 * get_order(0) returns funny result. Just warn and terminate
 911                 * early.
 912                 */
 913                return NULL;
 914        }
 915        order = get_order(size);
 916
 917again:
 918        rcu_read_lock();
 919        vbq = &get_cpu_var(vmap_block_queue);
 920        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 921                int i;
 922
 923                spin_lock(&vb->lock);
 924                if (vb->free < 1UL << order)
 925                        goto next;
 926
 927                i = bitmap_find_free_region(vb->alloc_map,
 928                                                VMAP_BBMAP_BITS, order);
 929
 930                if (i < 0) {
 931                        if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
 932                                /* fragmented and no outstanding allocations */
 933                                BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
 934                                purge = 1;
 935                        }
 936                        goto next;
 937                }
 938                addr = vb->va->va_start + (i << PAGE_SHIFT);
 939                BUG_ON(addr_to_vb_idx(addr) !=
 940                                addr_to_vb_idx(vb->va->va_start));
 941                vb->free -= 1UL << order;
 942                if (vb->free == 0) {
 943                        spin_lock(&vbq->lock);
 944                        list_del_rcu(&vb->free_list);
 945                        spin_unlock(&vbq->lock);
 946                }
 947                spin_unlock(&vb->lock);
 948                break;
 949next:
 950                spin_unlock(&vb->lock);
 951        }
 952
 953        if (purge)
 954                purge_fragmented_blocks_thiscpu();
 955
 956        put_cpu_var(vmap_block_queue);
 957        rcu_read_unlock();
 958
 959        if (!addr) {
 960                vb = new_vmap_block(gfp_mask);
 961                if (IS_ERR(vb))
 962                        return vb;
 963                goto again;
 964        }
 965
 966        return (void *)addr;
 967}
 968
 969static void vb_free(const void *addr, unsigned long size)
 970{
 971        unsigned long offset;
 972        unsigned long vb_idx;
 973        unsigned int order;
 974        struct vmap_block *vb;
 975
 976        BUG_ON(size & ~PAGE_MASK);
 977        BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 978
 979        flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
 980
 981        order = get_order(size);
 982
 983        offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
 984
 985        vb_idx = addr_to_vb_idx((unsigned long)addr);
 986        rcu_read_lock();
 987        vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
 988        rcu_read_unlock();
 989        BUG_ON(!vb);
 990
 991        vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
 992
 993        spin_lock(&vb->lock);
 994        BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
 995
 996        vb->dirty += 1UL << order;
 997        if (vb->dirty == VMAP_BBMAP_BITS) {
 998                BUG_ON(vb->free);
 999                spin_unlock(&vb->lock);
1000                free_vmap_block(vb);
1001        } else
1002                spin_unlock(&vb->lock);
1003}
1004
1005/**
1006 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1007 *
1008 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1009 * to amortize TLB flushing overheads. What this means is that any page you
1010 * have now, may, in a former life, have been mapped into kernel virtual
1011 * address by the vmap layer and so there might be some CPUs with TLB entries
1012 * still referencing that page (additional to the regular 1:1 kernel mapping).
1013 *
1014 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1015 * be sure that none of the pages we have control over will have any aliases
1016 * from the vmap layer.
1017 */
1018void vm_unmap_aliases(void)
1019{
1020        unsigned long start = ULONG_MAX, end = 0;
1021        int cpu;
1022        int flush = 0;
1023
1024        if (unlikely(!vmap_initialized))
1025                return;
1026
1027        for_each_possible_cpu(cpu) {
1028                struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1029                struct vmap_block *vb;
1030
1031                rcu_read_lock();
1032                list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1033                        int i;
1034
1035                        spin_lock(&vb->lock);
1036                        i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1037                        while (i < VMAP_BBMAP_BITS) {
1038                                unsigned long s, e;
1039                                int j;
1040                                j = find_next_zero_bit(vb->dirty_map,
1041                                        VMAP_BBMAP_BITS, i);
1042
1043                                s = vb->va->va_start + (i << PAGE_SHIFT);
1044                                e = vb->va->va_start + (j << PAGE_SHIFT);
1045                                flush = 1;
1046
1047                                if (s < start)
1048                                        start = s;
1049                                if (e > end)
1050                                        end = e;
1051
1052                                i = j;
1053                                i = find_next_bit(vb->dirty_map,
1054                                                        VMAP_BBMAP_BITS, i);
1055                        }
1056                        spin_unlock(&vb->lock);
1057                }
1058                rcu_read_unlock();
1059        }
1060
1061        __purge_vmap_area_lazy(&start, &end, 1, flush);
1062}
1063EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1064
1065/**
1066 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1067 * @mem: the pointer returned by vm_map_ram
1068 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1069 */
1070void vm_unmap_ram(const void *mem, unsigned int count)
1071{
1072        unsigned long size = count << PAGE_SHIFT;
1073        unsigned long addr = (unsigned long)mem;
1074
1075        BUG_ON(!addr);
1076        BUG_ON(addr < VMALLOC_START);
1077        BUG_ON(addr > VMALLOC_END);
1078        BUG_ON(addr & (PAGE_SIZE-1));
1079
1080        debug_check_no_locks_freed(mem, size);
1081        vmap_debug_free_range(addr, addr+size);
1082
1083        if (likely(count <= VMAP_MAX_ALLOC))
1084                vb_free(mem, size);
1085        else
1086                free_unmap_vmap_area_addr(addr);
1087}
1088EXPORT_SYMBOL(vm_unmap_ram);
1089
1090/**
1091 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1092 * @pages: an array of pointers to the pages to be mapped
1093 * @count: number of pages
1094 * @node: prefer to allocate data structures on this node
1095 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1096 *
1097 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1098 */
1099void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1100{
1101        unsigned long size = count << PAGE_SHIFT;
1102        unsigned long addr;
1103        void *mem;
1104
1105        if (likely(count <= VMAP_MAX_ALLOC)) {
1106                mem = vb_alloc(size, GFP_KERNEL);
1107                if (IS_ERR(mem))
1108                        return NULL;
1109                addr = (unsigned long)mem;
1110        } else {
1111                struct vmap_area *va;
1112                va = alloc_vmap_area(size, PAGE_SIZE,
1113                                VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1114                if (IS_ERR(va))
1115                        return NULL;
1116
1117                addr = va->va_start;
1118                mem = (void *)addr;
1119        }
1120        if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1121                vm_unmap_ram(mem, count);
1122                return NULL;
1123        }
1124        return mem;
1125}
1126EXPORT_SYMBOL(vm_map_ram);
1127
1128/**
1129 * vm_area_add_early - add vmap area early during boot
1130 * @vm: vm_struct to add
1131 *
1132 * This function is used to add fixed kernel vm area to vmlist before
1133 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
1134 * should contain proper values and the other fields should be zero.
1135 *
1136 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1137 */
1138void __init vm_area_add_early(struct vm_struct *vm)
1139{
1140        struct vm_struct *tmp, **p;
1141
1142        BUG_ON(vmap_initialized);
1143        for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1144                if (tmp->addr >= vm->addr) {
1145                        BUG_ON(tmp->addr < vm->addr + vm->size);
1146                        break;
1147                } else
1148                        BUG_ON(tmp->addr + tmp->size > vm->addr);
1149        }
1150        vm->next = *p;
1151        *p = vm;
1152}
1153
1154/**
1155 * vm_area_register_early - register vmap area early during boot
1156 * @vm: vm_struct to register
1157 * @align: requested alignment
1158 *
1159 * This function is used to register kernel vm area before
1160 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1161 * proper values on entry and other fields should be zero.  On return,
1162 * vm->addr contains the allocated address.
1163 *
1164 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1165 */
1166void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1167{
1168        static size_t vm_init_off __initdata;
1169        unsigned long addr;
1170
1171        addr = ALIGN(VMALLOC_START + vm_init_off, align);
1172        vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1173
1174        vm->addr = (void *)addr;
1175
1176        vm_area_add_early(vm);
1177}
1178
1179void __init vmalloc_init(void)
1180{
1181        struct vmap_area *va;
1182        struct vm_struct *tmp;
1183        int i;
1184
1185        for_each_possible_cpu(i) {
1186                struct vmap_block_queue *vbq;
1187
1188                vbq = &per_cpu(vmap_block_queue, i);
1189                spin_lock_init(&vbq->lock);
1190                INIT_LIST_HEAD(&vbq->free);
1191        }
1192
1193        /* Import existing vmlist entries. */
1194        for (tmp = vmlist; tmp; tmp = tmp->next) {
1195                va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1196                va->flags = VM_VM_AREA;
1197                va->va_start = (unsigned long)tmp->addr;
1198                va->va_end = va->va_start + tmp->size;
1199                va->vm = tmp;
1200                __insert_vmap_area(va);
1201        }
1202
1203        vmap_area_pcpu_hole = VMALLOC_END;
1204
1205        vmap_initialized = true;
1206}
1207
1208/**
1209 * map_kernel_range_noflush - map kernel VM area with the specified pages
1210 * @addr: start of the VM area to map
1211 * @size: size of the VM area to map
1212 * @prot: page protection flags to use
1213 * @pages: pages to map
1214 *
1215 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1216 * specify should have been allocated using get_vm_area() and its
1217 * friends.
1218 *
1219 * NOTE:
1220 * This function does NOT do any cache flushing.  The caller is
1221 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1222 * before calling this function.
1223 *
1224 * RETURNS:
1225 * The number of pages mapped on success, -errno on failure.
1226 */
1227int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1228                             pgprot_t prot, struct page **pages)
1229{
1230        return vmap_page_range_noflush(addr, addr + size, prot, pages);
1231}
1232
1233/**
1234 * unmap_kernel_range_noflush - unmap kernel VM area
1235 * @addr: start of the VM area to unmap
1236 * @size: size of the VM area to unmap
1237 *
1238 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1239 * specify should have been allocated using get_vm_area() and its
1240 * friends.
1241 *
1242 * NOTE:
1243 * This function does NOT do any cache flushing.  The caller is
1244 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1245 * before calling this function and flush_tlb_kernel_range() after.
1246 */
1247void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1248{
1249        vunmap_page_range(addr, addr + size);
1250}
1251EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1252
1253/**
1254 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1255 * @addr: start of the VM area to unmap
1256 * @size: size of the VM area to unmap
1257 *
1258 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1259 * the unmapping and tlb after.
1260 */
1261void unmap_kernel_range(unsigned long addr, unsigned long size)
1262{
1263        unsigned long end = addr + size;
1264
1265        flush_cache_vunmap(addr, end);
1266        vunmap_page_range(addr, end);
1267        flush_tlb_kernel_range(addr, end);
1268}
1269
1270int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1271{
1272        unsigned long addr = (unsigned long)area->addr;
1273        unsigned long end = addr + area->size - PAGE_SIZE;
1274        int err;
1275
1276        err = vmap_page_range(addr, end, prot, *pages);
1277        if (err > 0) {
1278                *pages += err;
1279                err = 0;
1280        }
1281
1282        return err;
1283}
1284EXPORT_SYMBOL_GPL(map_vm_area);
1285
1286/*** Old vmalloc interfaces ***/
1287DEFINE_RWLOCK(vmlist_lock);
1288struct vm_struct *vmlist;
1289
1290static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1291                              unsigned long flags, const void *caller)
1292{
1293        vm->flags = flags;
1294        vm->addr = (void *)va->va_start;
1295        vm->size = va->va_end - va->va_start;
1296        vm->caller = caller;
1297        va->vm = vm;
1298        va->flags |= VM_VM_AREA;
1299}
1300
1301static void insert_vmalloc_vmlist(struct vm_struct *vm)
1302{
1303        struct vm_struct *tmp, **p;
1304
1305        vm->flags &= ~VM_UNLIST;
1306        write_lock(&vmlist_lock);
1307        for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1308                if (tmp->addr >= vm->addr)
1309                        break;
1310        }
1311        vm->next = *p;
1312        *p = vm;
1313        write_unlock(&vmlist_lock);
1314}
1315
1316static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1317                              unsigned long flags, const void *caller)
1318{
1319        setup_vmalloc_vm(vm, va, flags, caller);
1320        insert_vmalloc_vmlist(vm);
1321}
1322
1323static struct vm_struct *__get_vm_area_node(unsigned long size,
1324                unsigned long align, unsigned long flags, unsigned long start,
1325                unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1326{
1327        struct vmap_area *va;
1328        struct vm_struct *area;
1329
1330        BUG_ON(in_interrupt());
1331        if (flags & VM_IOREMAP) {
1332                int bit = fls(size);
1333
1334                if (bit > IOREMAP_MAX_ORDER)
1335                        bit = IOREMAP_MAX_ORDER;
1336                else if (bit < PAGE_SHIFT)
1337                        bit = PAGE_SHIFT;
1338
1339                align = 1ul << bit;
1340        }
1341
1342        size = PAGE_ALIGN(size);
1343        if (unlikely(!size))
1344                return NULL;
1345
1346        area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1347        if (unlikely(!area))
1348                return NULL;
1349
1350        /*
1351         * We always allocate a guard page.
1352         */
1353        size += PAGE_SIZE;
1354
1355        va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1356        if (IS_ERR(va)) {
1357                kfree(area);
1358                return NULL;
1359        }
1360
1361        /*
1362         * When this function is called from __vmalloc_node_range,
1363         * we do not add vm_struct to vmlist here to avoid
1364         * accessing uninitialized members of vm_struct such as
1365         * pages and nr_pages fields. They will be set later.
1366         * To distinguish it from others, we use a VM_UNLIST flag.
1367         */
1368        if (flags & VM_UNLIST)
1369                setup_vmalloc_vm(area, va, flags, caller);
1370        else
1371                insert_vmalloc_vm(area, va, flags, caller);
1372
1373        return area;
1374}
1375
1376struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1377                                unsigned long start, unsigned long end)
1378{
1379        return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1380                                                __builtin_return_address(0));
1381}
1382EXPORT_SYMBOL_GPL(__get_vm_area);
1383
1384struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1385                                       unsigned long start, unsigned long end,
1386                                       const void *caller)
1387{
1388        return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1389                                  caller);
1390}
1391
1392/**
1393 *      get_vm_area  -  reserve a contiguous kernel virtual area
1394 *      @size:          size of the area
1395 *      @flags:         %VM_IOREMAP for I/O mappings or VM_ALLOC
1396 *
1397 *      Search an area of @size in the kernel virtual mapping area,
1398 *      and reserved it for out purposes.  Returns the area descriptor
1399 *      on success or %NULL on failure.
1400 */
1401struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1402{
1403        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1404                                -1, GFP_KERNEL, __builtin_return_address(0));
1405}
1406
1407struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1408                                const void *caller)
1409{
1410        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1411                                                -1, GFP_KERNEL, caller);
1412}
1413
1414/**
1415 *      find_vm_area  -  find a continuous kernel virtual area
1416 *      @addr:          base address
1417 *
1418 *      Search for the kernel VM area starting at @addr, and return it.
1419 *      It is up to the caller to do all required locking to keep the returned
1420 *      pointer valid.
1421 */
1422struct vm_struct *find_vm_area(const void *addr)
1423{
1424        struct vmap_area *va;
1425
1426        va = find_vmap_area((unsigned long)addr);
1427        if (va && va->flags & VM_VM_AREA)
1428                return va->vm;
1429
1430        return NULL;
1431}
1432
1433/**
1434 *      remove_vm_area  -  find and remove a continuous kernel virtual area
1435 *      @addr:          base address
1436 *
1437 *      Search for the kernel VM area starting at @addr, and remove it.
1438 *      This function returns the found VM area, but using it is NOT safe
1439 *      on SMP machines, except for its size or flags.
1440 */
1441struct vm_struct *remove_vm_area(const void *addr)
1442{
1443        struct vmap_area *va;
1444
1445        va = find_vmap_area((unsigned long)addr);
1446        if (va && va->flags & VM_VM_AREA) {
1447                struct vm_struct *vm = va->vm;
1448
1449                if (!(vm->flags & VM_UNLIST)) {
1450                        struct vm_struct *tmp, **p;
1451                        /*
1452                         * remove from list and disallow access to
1453                         * this vm_struct before unmap. (address range
1454                         * confliction is maintained by vmap.)
1455                         */
1456                        write_lock(&vmlist_lock);
1457                        for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1458                                ;
1459                        *p = tmp->next;
1460                        write_unlock(&vmlist_lock);
1461                }
1462
1463                vmap_debug_free_range(va->va_start, va->va_end);
1464                free_unmap_vmap_area(va);
1465                vm->size -= PAGE_SIZE;
1466
1467                return vm;
1468        }
1469        return NULL;
1470}
1471
1472static void __vunmap(const void *addr, int deallocate_pages)
1473{
1474        struct vm_struct *area;
1475
1476        if (!addr)
1477                return;
1478
1479        if ((PAGE_SIZE-1) & (unsigned long)addr) {
1480                WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1481                return;
1482        }
1483
1484        area = remove_vm_area(addr);
1485        if (unlikely(!area)) {
1486                WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1487                                addr);
1488                return;
1489        }
1490
1491        debug_check_no_locks_freed(addr, area->size);
1492        debug_check_no_obj_freed(addr, area->size);
1493
1494        if (deallocate_pages) {
1495                int i;
1496
1497                for (i = 0; i < area->nr_pages; i++) {
1498                        struct page *page = area->pages[i];
1499
1500                        BUG_ON(!page);
1501                        __free_page(page);
1502                }
1503
1504                if (area->flags & VM_VPAGES)
1505                        vfree(area->pages);
1506                else
1507                        kfree(area->pages);
1508        }
1509
1510        kfree(area);
1511        return;
1512}
1513
1514/**
1515 *      vfree  -  release memory allocated by vmalloc()
1516 *      @addr:          memory base address
1517 *
1518 *      Free the virtually continuous memory area starting at @addr, as
1519 *      obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1520 *      NULL, no operation is performed.
1521 *
1522 *      Must not be called in interrupt context.
1523 */
1524void vfree(const void *addr)
1525{
1526        BUG_ON(in_interrupt());
1527
1528        kmemleak_free(addr);
1529
1530        __vunmap(addr, 1);
1531}
1532EXPORT_SYMBOL(vfree);
1533
1534/**
1535 *      vunmap  -  release virtual mapping obtained by vmap()
1536 *      @addr:          memory base address
1537 *
1538 *      Free the virtually contiguous memory area starting at @addr,
1539 *      which was created from the page array passed to vmap().
1540 *
1541 *      Must not be called in interrupt context.
1542 */
1543void vunmap(const void *addr)
1544{
1545        BUG_ON(in_interrupt());
1546        might_sleep();
1547        __vunmap(addr, 0);
1548}
1549EXPORT_SYMBOL(vunmap);
1550
1551/**
1552 *      vmap  -  map an array of pages into virtually contiguous space
1553 *      @pages:         array of page pointers
1554 *      @count:         number of pages to map
1555 *      @flags:         vm_area->flags
1556 *      @prot:          page protection for the mapping
1557 *
1558 *      Maps @count pages from @pages into contiguous kernel virtual
1559 *      space.
1560 */
1561void *vmap(struct page **pages, unsigned int count,
1562                unsigned long flags, pgprot_t prot)
1563{
1564        struct vm_struct *area;
1565
1566        might_sleep();
1567
1568        if (count > totalram_pages)
1569                return NULL;
1570
1571        area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1572                                        __builtin_return_address(0));
1573        if (!area)
1574                return NULL;
1575
1576        if (map_vm_area(area, prot, &pages)) {
1577                vunmap(area->addr);
1578                return NULL;
1579        }
1580
1581        return area->addr;
1582}
1583EXPORT_SYMBOL(vmap);
1584
1585static void *__vmalloc_node(unsigned long size, unsigned long align,
1586                            gfp_t gfp_mask, pgprot_t prot,
1587                            int node, const void *caller);
1588static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1589                                 pgprot_t prot, int node, const void *caller)
1590{
1591        const int order = 0;
1592        struct page **pages;
1593        unsigned int nr_pages, array_size, i;
1594        gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1595
1596        nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1597        array_size = (nr_pages * sizeof(struct page *));
1598
1599        area->nr_pages = nr_pages;
1600        /* Please note that the recursion is strictly bounded. */
1601        if (array_size > PAGE_SIZE) {
1602                pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1603                                PAGE_KERNEL, node, caller);
1604                area->flags |= VM_VPAGES;
1605        } else {
1606                pages = kmalloc_node(array_size, nested_gfp, node);
1607        }
1608        area->pages = pages;
1609        area->caller = caller;
1610        if (!area->pages) {
1611                remove_vm_area(area->addr);
1612                kfree(area);
1613                return NULL;
1614        }
1615
1616        for (i = 0; i < area->nr_pages; i++) {
1617                struct page *page;
1618                gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1619
1620                if (node < 0)
1621                        page = alloc_page(tmp_mask);
1622                else
1623                        page = alloc_pages_node(node, tmp_mask, order);
1624
1625                if (unlikely(!page)) {
1626                        /* Successfully allocated i pages, free them in __vunmap() */
1627                        area->nr_pages = i;
1628                        goto fail;
1629                }
1630                area->pages[i] = page;
1631        }
1632
1633        if (map_vm_area(area, prot, &pages))
1634                goto fail;
1635        return area->addr;
1636
1637fail:
1638        warn_alloc_failed(gfp_mask, order,
1639                          "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1640                          (area->nr_pages*PAGE_SIZE), area->size);
1641        vfree(area->addr);
1642        return NULL;
1643}
1644
1645/**
1646 *      __vmalloc_node_range  -  allocate virtually contiguous memory
1647 *      @size:          allocation size
1648 *      @align:         desired alignment
1649 *      @start:         vm area range start
1650 *      @end:           vm area range end
1651 *      @gfp_mask:      flags for the page level allocator
1652 *      @prot:          protection mask for the allocated pages
1653 *      @node:          node to use for allocation or -1
1654 *      @caller:        caller's return address
1655 *
1656 *      Allocate enough pages to cover @size from the page level
1657 *      allocator with @gfp_mask flags.  Map them into contiguous
1658 *      kernel virtual space, using a pagetable protection of @prot.
1659 */
1660void *__vmalloc_node_range(unsigned long size, unsigned long align,
1661                        unsigned long start, unsigned long end, gfp_t gfp_mask,
1662                        pgprot_t prot, int node, const void *caller)
1663{
1664        struct vm_struct *area;
1665        void *addr;
1666        unsigned long real_size = size;
1667
1668        size = PAGE_ALIGN(size);
1669        if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1670                goto fail;
1671
1672        area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNLIST,
1673                                  start, end, node, gfp_mask, caller);
1674        if (!area)
1675                goto fail;
1676
1677        addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1678        if (!addr)
1679                return NULL;
1680
1681        /*
1682         * In this function, newly allocated vm_struct is not added
1683         * to vmlist at __get_vm_area_node(). so, it is added here.
1684         */
1685        insert_vmalloc_vmlist(area);
1686
1687        /*
1688         * A ref_count = 3 is needed because the vm_struct and vmap_area
1689         * structures allocated in the __get_vm_area_node() function contain
1690         * references to the virtual address of the vmalloc'ed block.
1691         */
1692        kmemleak_alloc(addr, real_size, 3, gfp_mask);
1693
1694        return addr;
1695
1696fail:
1697        warn_alloc_failed(gfp_mask, 0,
1698                          "vmalloc: allocation failure: %lu bytes\n",
1699                          real_size);
1700        return NULL;
1701}
1702
1703/**
1704 *      __vmalloc_node  -  allocate virtually contiguous memory
1705 *      @size:          allocation size
1706 *      @align:         desired alignment
1707 *      @gfp_mask:      flags for the page level allocator
1708 *      @prot:          protection mask for the allocated pages
1709 *      @node:          node to use for allocation or -1
1710 *      @caller:        caller's return address
1711 *
1712 *      Allocate enough pages to cover @size from the page level
1713 *      allocator with @gfp_mask flags.  Map them into contiguous
1714 *      kernel virtual space, using a pagetable protection of @prot.
1715 */
1716static void *__vmalloc_node(unsigned long size, unsigned long align,
1717                            gfp_t gfp_mask, pgprot_t prot,
1718                            int node, const void *caller)
1719{
1720        return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1721                                gfp_mask, prot, node, caller);
1722}
1723
1724void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1725{
1726        return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1727                                __builtin_return_address(0));
1728}
1729EXPORT_SYMBOL(__vmalloc);
1730
1731static inline void *__vmalloc_node_flags(unsigned long size,
1732                                        int node, gfp_t flags)
1733{
1734        return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1735                                        node, __builtin_return_address(0));
1736}
1737
1738/**
1739 *      vmalloc  -  allocate virtually contiguous memory
1740 *      @size:          allocation size
1741 *      Allocate enough pages to cover @size from the page level
1742 *      allocator and map them into contiguous kernel virtual space.
1743 *
1744 *      For tight control over page level allocator and protection flags
1745 *      use __vmalloc() instead.
1746 */
1747void *vmalloc(unsigned long size)
1748{
1749        return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
1750}
1751EXPORT_SYMBOL(vmalloc);
1752
1753/**
1754 *      vzalloc - allocate virtually contiguous memory with zero fill
1755 *      @size:  allocation size
1756 *      Allocate enough pages to cover @size from the page level
1757 *      allocator and map them into contiguous kernel virtual space.
1758 *      The memory allocated is set to zero.
1759 *
1760 *      For tight control over page level allocator and protection flags
1761 *      use __vmalloc() instead.
1762 */
1763void *vzalloc(unsigned long size)
1764{
1765        return __vmalloc_node_flags(size, -1,
1766                                GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1767}
1768EXPORT_SYMBOL(vzalloc);
1769
1770/**
1771 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1772 * @size: allocation size
1773 *
1774 * The resulting memory area is zeroed so it can be mapped to userspace
1775 * without leaking data.
1776 */
1777void *vmalloc_user(unsigned long size)
1778{
1779        struct vm_struct *area;
1780        void *ret;
1781
1782        ret = __vmalloc_node(size, SHMLBA,
1783                             GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1784                             PAGE_KERNEL, -1, __builtin_return_address(0));
1785        if (ret) {
1786                area = find_vm_area(ret);
1787                area->flags |= VM_USERMAP;
1788        }
1789        return ret;
1790}
1791EXPORT_SYMBOL(vmalloc_user);
1792
1793/**
1794 *      vmalloc_node  -  allocate memory on a specific node
1795 *      @size:          allocation size
1796 *      @node:          numa node
1797 *
1798 *      Allocate enough pages to cover @size from the page level
1799 *      allocator and map them into contiguous kernel virtual space.
1800 *
1801 *      For tight control over page level allocator and protection flags
1802 *      use __vmalloc() instead.
1803 */
1804void *vmalloc_node(unsigned long size, int node)
1805{
1806        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1807                                        node, __builtin_return_address(0));
1808}
1809EXPORT_SYMBOL(vmalloc_node);
1810
1811/**
1812 * vzalloc_node - allocate memory on a specific node with zero fill
1813 * @size:       allocation size
1814 * @node:       numa node
1815 *
1816 * Allocate enough pages to cover @size from the page level
1817 * allocator and map them into contiguous kernel virtual space.
1818 * The memory allocated is set to zero.
1819 *
1820 * For tight control over page level allocator and protection flags
1821 * use __vmalloc_node() instead.
1822 */
1823void *vzalloc_node(unsigned long size, int node)
1824{
1825        return __vmalloc_node_flags(size, node,
1826                         GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1827}
1828EXPORT_SYMBOL(vzalloc_node);
1829
1830#ifndef PAGE_KERNEL_EXEC
1831# define PAGE_KERNEL_EXEC PAGE_KERNEL
1832#endif
1833
1834/**
1835 *      vmalloc_exec  -  allocate virtually contiguous, executable memory
1836 *      @size:          allocation size
1837 *
1838 *      Kernel-internal function to allocate enough pages to cover @size
1839 *      the page level allocator and map them into contiguous and
1840 *      executable kernel virtual space.
1841 *
1842 *      For tight control over page level allocator and protection flags
1843 *      use __vmalloc() instead.
1844 */
1845
1846void *vmalloc_exec(unsigned long size)
1847{
1848        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1849                              -1, __builtin_return_address(0));
1850}
1851
1852#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1853#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1854#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1855#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1856#else
1857#define GFP_VMALLOC32 GFP_KERNEL
1858#endif
1859
1860/**
1861 *      vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1862 *      @size:          allocation size
1863 *
1864 *      Allocate enough 32bit PA addressable pages to cover @size from the
1865 *      page level allocator and map them into contiguous kernel virtual space.
1866 */
1867void *vmalloc_32(unsigned long size)
1868{
1869        return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1870                              -1, __builtin_return_address(0));
1871}
1872EXPORT_SYMBOL(vmalloc_32);
1873
1874/**
1875 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1876 *      @size:          allocation size
1877 *
1878 * The resulting memory area is 32bit addressable and zeroed so it can be
1879 * mapped to userspace without leaking data.
1880 */
1881void *vmalloc_32_user(unsigned long size)
1882{
1883        struct vm_struct *area;
1884        void *ret;
1885
1886        ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1887                             -1, __builtin_return_address(0));
1888        if (ret) {
1889                area = find_vm_area(ret);
1890                area->flags |= VM_USERMAP;
1891        }
1892        return ret;
1893}
1894EXPORT_SYMBOL(vmalloc_32_user);
1895
1896/*
1897 * small helper routine , copy contents to buf from addr.
1898 * If the page is not present, fill zero.
1899 */
1900
1901static int aligned_vread(char *buf, char *addr, unsigned long count)
1902{
1903        struct page *p;
1904        int copied = 0;
1905
1906        while (count) {
1907                unsigned long offset, length;
1908
1909                offset = (unsigned long)addr & ~PAGE_MASK;
1910                length = PAGE_SIZE - offset;
1911                if (length > count)
1912                        length = count;
1913                p = vmalloc_to_page(addr);
1914                /*
1915                 * To do safe access to this _mapped_ area, we need
1916                 * lock. But adding lock here means that we need to add
1917                 * overhead of vmalloc()/vfree() calles for this _debug_
1918                 * interface, rarely used. Instead of that, we'll use
1919                 * kmap() and get small overhead in this access function.
1920                 */
1921                if (p) {
1922                        /*
1923                         * we can expect USER0 is not used (see vread/vwrite's
1924                         * function description)
1925                         */
1926                        void *map = kmap_atomic(p);
1927                        memcpy(buf, map + offset, length);
1928                        kunmap_atomic(map);
1929                } else
1930                        memset(buf, 0, length);
1931
1932                addr += length;
1933                buf += length;
1934                copied += length;
1935                count -= length;
1936        }
1937        return copied;
1938}
1939
1940static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1941{
1942        struct page *p;
1943        int copied = 0;
1944
1945        while (count) {
1946                unsigned long offset, length;
1947
1948                offset = (unsigned long)addr & ~PAGE_MASK;
1949                length = PAGE_SIZE - offset;
1950                if (length > count)
1951                        length = count;
1952                p = vmalloc_to_page(addr);
1953                /*
1954                 * To do safe access to this _mapped_ area, we need
1955                 * lock. But adding lock here means that we need to add
1956                 * overhead of vmalloc()/vfree() calles for this _debug_
1957                 * interface, rarely used. Instead of that, we'll use
1958                 * kmap() and get small overhead in this access function.
1959                 */
1960                if (p) {
1961                        /*
1962                         * we can expect USER0 is not used (see vread/vwrite's
1963                         * function description)
1964                         */
1965                        void *map = kmap_atomic(p);
1966                        memcpy(map + offset, buf, length);
1967                        kunmap_atomic(map);
1968                }
1969                addr += length;
1970                buf += length;
1971                copied += length;
1972                count -= length;
1973        }
1974        return copied;
1975}
1976
1977/**
1978 *      vread() -  read vmalloc area in a safe way.
1979 *      @buf:           buffer for reading data
1980 *      @addr:          vm address.
1981 *      @count:         number of bytes to be read.
1982 *
1983 *      Returns # of bytes which addr and buf should be increased.
1984 *      (same number to @count). Returns 0 if [addr...addr+count) doesn't
1985 *      includes any intersect with alive vmalloc area.
1986 *
1987 *      This function checks that addr is a valid vmalloc'ed area, and
1988 *      copy data from that area to a given buffer. If the given memory range
1989 *      of [addr...addr+count) includes some valid address, data is copied to
1990 *      proper area of @buf. If there are memory holes, they'll be zero-filled.
1991 *      IOREMAP area is treated as memory hole and no copy is done.
1992 *
1993 *      If [addr...addr+count) doesn't includes any intersects with alive
1994 *      vm_struct area, returns 0. @buf should be kernel's buffer.
1995 *
1996 *      Note: In usual ops, vread() is never necessary because the caller
1997 *      should know vmalloc() area is valid and can use memcpy().
1998 *      This is for routines which have to access vmalloc area without
1999 *      any informaion, as /dev/kmem.
2000 *
2001 */
2002
2003long vread(char *buf, char *addr, unsigned long count)
2004{
2005        struct vm_struct *tmp;
2006        char *vaddr, *buf_start = buf;
2007        unsigned long buflen = count;
2008        unsigned long n;
2009
2010        /* Don't allow overflow */
2011        if ((unsigned long) addr + count < count)
2012                count = -(unsigned long) addr;
2013
2014        read_lock(&vmlist_lock);
2015        for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2016                vaddr = (char *) tmp->addr;
2017                if (addr >= vaddr + tmp->size - PAGE_SIZE)
2018                        continue;
2019                while (addr < vaddr) {
2020                        if (count == 0)
2021                                goto finished;
2022                        *buf = '\0';
2023                        buf++;
2024                        addr++;
2025                        count--;
2026                }
2027                n = vaddr + tmp->size - PAGE_SIZE - addr;
2028                if (n > count)
2029                        n = count;
2030                if (!(tmp->flags & VM_IOREMAP))
2031                        aligned_vread(buf, addr, n);
2032                else /* IOREMAP area is treated as memory hole */
2033                        memset(buf, 0, n);
2034                buf += n;
2035                addr += n;
2036                count -= n;
2037        }
2038finished:
2039        read_unlock(&vmlist_lock);
2040
2041        if (buf == buf_start)
2042                return 0;
2043        /* zero-fill memory holes */
2044        if (buf != buf_start + buflen)
2045                memset(buf, 0, buflen - (buf - buf_start));
2046
2047        return buflen;
2048}
2049
2050/**
2051 *      vwrite() -  write vmalloc area in a safe way.
2052 *      @buf:           buffer for source data
2053 *      @addr:          vm address.
2054 *      @count:         number of bytes to be read.
2055 *
2056 *      Returns # of bytes which addr and buf should be incresed.
2057 *      (same number to @count).
2058 *      If [addr...addr+count) doesn't includes any intersect with valid
2059 *      vmalloc area, returns 0.
2060 *
2061 *      This function checks that addr is a valid vmalloc'ed area, and
2062 *      copy data from a buffer to the given addr. If specified range of
2063 *      [addr...addr+count) includes some valid address, data is copied from
2064 *      proper area of @buf. If there are memory holes, no copy to hole.
2065 *      IOREMAP area is treated as memory hole and no copy is done.
2066 *
2067 *      If [addr...addr+count) doesn't includes any intersects with alive
2068 *      vm_struct area, returns 0. @buf should be kernel's buffer.
2069 *
2070 *      Note: In usual ops, vwrite() is never necessary because the caller
2071 *      should know vmalloc() area is valid and can use memcpy().
2072 *      This is for routines which have to access vmalloc area without
2073 *      any informaion, as /dev/kmem.
2074 */
2075
2076long vwrite(char *buf, char *addr, unsigned long count)
2077{
2078        struct vm_struct *tmp;
2079        char *vaddr;
2080        unsigned long n, buflen;
2081        int copied = 0;
2082
2083        /* Don't allow overflow */
2084        if ((unsigned long) addr + count < count)
2085                count = -(unsigned long) addr;
2086        buflen = count;
2087
2088        read_lock(&vmlist_lock);
2089        for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2090                vaddr = (char *) tmp->addr;
2091                if (addr >= vaddr + tmp->size - PAGE_SIZE)
2092                        continue;
2093                while (addr < vaddr) {
2094                        if (count == 0)
2095                                goto finished;
2096                        buf++;
2097                        addr++;
2098                        count--;
2099                }
2100                n = vaddr + tmp->size - PAGE_SIZE - addr;
2101                if (n > count)
2102                        n = count;
2103                if (!(tmp->flags & VM_IOREMAP)) {
2104                        aligned_vwrite(buf, addr, n);
2105                        copied++;
2106                }
2107                buf += n;
2108                addr += n;
2109                count -= n;
2110        }
2111finished:
2112        read_unlock(&vmlist_lock);
2113        if (!copied)
2114                return 0;
2115        return buflen;
2116}
2117
2118/**
2119 *      remap_vmalloc_range  -  map vmalloc pages to userspace
2120 *      @vma:           vma to cover (map full range of vma)
2121 *      @addr:          vmalloc memory
2122 *      @pgoff:         number of pages into addr before first page to map
2123 *
2124 *      Returns:        0 for success, -Exxx on failure
2125 *
2126 *      This function checks that addr is a valid vmalloc'ed area, and
2127 *      that it is big enough to cover the vma. Will return failure if
2128 *      that criteria isn't met.
2129 *
2130 *      Similar to remap_pfn_range() (see mm/memory.c)
2131 */
2132int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2133                                                unsigned long pgoff)
2134{
2135        struct vm_struct *area;
2136        unsigned long uaddr = vma->vm_start;
2137        unsigned long usize = vma->vm_end - vma->vm_start;
2138
2139        if ((PAGE_SIZE-1) & (unsigned long)addr)
2140                return -EINVAL;
2141
2142        area = find_vm_area(addr);
2143        if (!area)
2144                return -EINVAL;
2145
2146        if (!(area->flags & VM_USERMAP))
2147                return -EINVAL;
2148
2149        if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2150                return -EINVAL;
2151
2152        addr += pgoff << PAGE_SHIFT;
2153        do {
2154                struct page *page = vmalloc_to_page(addr);
2155                int ret;
2156
2157                ret = vm_insert_page(vma, uaddr, page);
2158                if (ret)
2159                        return ret;
2160
2161                uaddr += PAGE_SIZE;
2162                addr += PAGE_SIZE;
2163                usize -= PAGE_SIZE;
2164        } while (usize > 0);
2165
2166        vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2167
2168        return 0;
2169}
2170EXPORT_SYMBOL(remap_vmalloc_range);
2171
2172/*
2173 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2174 * have one.
2175 */
2176void  __attribute__((weak)) vmalloc_sync_all(void)
2177{
2178}
2179
2180
2181static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2182{
2183        pte_t ***p = data;
2184
2185        if (p) {
2186                *(*p) = pte;
2187                (*p)++;
2188        }
2189        return 0;
2190}
2191
2192/**
2193 *      alloc_vm_area - allocate a range of kernel address space
2194 *      @size:          size of the area
2195 *      @ptes:          returns the PTEs for the address space
2196 *
2197 *      Returns:        NULL on failure, vm_struct on success
2198 *
2199 *      This function reserves a range of kernel address space, and
2200 *      allocates pagetables to map that range.  No actual mappings
2201 *      are created.
2202 *
2203 *      If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2204 *      allocated for the VM area are returned.
2205 */
2206struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2207{
2208        struct vm_struct *area;
2209
2210        area = get_vm_area_caller(size, VM_IOREMAP,
2211                                __builtin_return_address(0));
2212        if (area == NULL)
2213                return NULL;
2214
2215        /*
2216         * This ensures that page tables are constructed for this region
2217         * of kernel virtual address space and mapped into init_mm.
2218         */
2219        if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2220                                size, f, ptes ? &ptes : NULL)) {
2221                free_vm_area(area);
2222                return NULL;
2223        }
2224
2225        return area;
2226}
2227EXPORT_SYMBOL_GPL(alloc_vm_area);
2228
2229void free_vm_area(struct vm_struct *area)
2230{
2231        struct vm_struct *ret;
2232        ret = remove_vm_area(area->addr);
2233        BUG_ON(ret != area);
2234        kfree(area);
2235}
2236EXPORT_SYMBOL_GPL(free_vm_area);
2237
2238#ifdef CONFIG_SMP
2239static struct vmap_area *node_to_va(struct rb_node *n)
2240{
2241        return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2242}
2243
2244/**
2245 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2246 * @end: target address
2247 * @pnext: out arg for the next vmap_area
2248 * @pprev: out arg for the previous vmap_area
2249 *
2250 * Returns: %true if either or both of next and prev are found,
2251 *          %false if no vmap_area exists
2252 *
2253 * Find vmap_areas end addresses of which enclose @end.  ie. if not
2254 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2255 */
2256static bool pvm_find_next_prev(unsigned long end,
2257                               struct vmap_area **pnext,
2258                               struct vmap_area **pprev)
2259{
2260        struct rb_node *n = vmap_area_root.rb_node;
2261        struct vmap_area *va = NULL;
2262
2263        while (n) {
2264                va = rb_entry(n, struct vmap_area, rb_node);
2265                if (end < va->va_end)
2266                        n = n->rb_left;
2267                else if (end > va->va_end)
2268                        n = n->rb_right;
2269                else
2270                        break;
2271        }
2272
2273        if (!va)
2274                return false;
2275
2276        if (va->va_end > end) {
2277                *pnext = va;
2278                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2279        } else {
2280                *pprev = va;
2281                *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2282        }
2283        return true;
2284}
2285
2286/**
2287 * pvm_determine_end - find the highest aligned address between two vmap_areas
2288 * @pnext: in/out arg for the next vmap_area
2289 * @pprev: in/out arg for the previous vmap_area
2290 * @align: alignment
2291 *
2292 * Returns: determined end address
2293 *
2294 * Find the highest aligned address between *@pnext and *@pprev below
2295 * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2296 * down address is between the end addresses of the two vmap_areas.
2297 *
2298 * Please note that the address returned by this function may fall
2299 * inside *@pnext vmap_area.  The caller is responsible for checking
2300 * that.
2301 */
2302static unsigned long pvm_determine_end(struct vmap_area **pnext,
2303                                       struct vmap_area **pprev,
2304                                       unsigned long align)
2305{
2306        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2307        unsigned long addr;
2308
2309        if (*pnext)
2310                addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2311        else
2312                addr = vmalloc_end;
2313
2314        while (*pprev && (*pprev)->va_end > addr) {
2315                *pnext = *pprev;
2316                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2317        }
2318
2319        return addr;
2320}
2321
2322/**
2323 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2324 * @offsets: array containing offset of each area
2325 * @sizes: array containing size of each area
2326 * @nr_vms: the number of areas to allocate
2327 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2328 *
2329 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2330 *          vm_structs on success, %NULL on failure
2331 *
2332 * Percpu allocator wants to use congruent vm areas so that it can
2333 * maintain the offsets among percpu areas.  This function allocates
2334 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
2335 * be scattered pretty far, distance between two areas easily going up
2336 * to gigabytes.  To avoid interacting with regular vmallocs, these
2337 * areas are allocated from top.
2338 *
2339 * Despite its complicated look, this allocator is rather simple.  It
2340 * does everything top-down and scans areas from the end looking for
2341 * matching slot.  While scanning, if any of the areas overlaps with
2342 * existing vmap_area, the base address is pulled down to fit the
2343 * area.  Scanning is repeated till all the areas fit and then all
2344 * necessary data structres are inserted and the result is returned.
2345 */
2346struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2347                                     const size_t *sizes, int nr_vms,
2348                                     size_t align)
2349{
2350        const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2351        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2352        struct vmap_area **vas, *prev, *next;
2353        struct vm_struct **vms;
2354        int area, area2, last_area, term_area;
2355        unsigned long base, start, end, last_end;
2356        bool purged = false;
2357
2358        /* verify parameters and allocate data structures */
2359        BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2360        for (last_area = 0, area = 0; area < nr_vms; area++) {
2361                start = offsets[area];
2362                end = start + sizes[area];
2363
2364                /* is everything aligned properly? */
2365                BUG_ON(!IS_ALIGNED(offsets[area], align));
2366                BUG_ON(!IS_ALIGNED(sizes[area], align));
2367
2368                /* detect the area with the highest address */
2369                if (start > offsets[last_area])
2370                        last_area = area;
2371
2372                for (area2 = 0; area2 < nr_vms; area2++) {
2373                        unsigned long start2 = offsets[area2];
2374                        unsigned long end2 = start2 + sizes[area2];
2375
2376                        if (area2 == area)
2377                                continue;
2378
2379                        BUG_ON(start2 >= start && start2 < end);
2380                        BUG_ON(end2 <= end && end2 > start);
2381                }
2382        }
2383        last_end = offsets[last_area] + sizes[last_area];
2384
2385        if (vmalloc_end - vmalloc_start < last_end) {
2386                WARN_ON(true);
2387                return NULL;
2388        }
2389
2390        vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2391        vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2392        if (!vas || !vms)
2393                goto err_free2;
2394
2395        for (area = 0; area < nr_vms; area++) {
2396                vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2397                vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2398                if (!vas[area] || !vms[area])
2399                        goto err_free;
2400        }
2401retry:
2402        spin_lock(&vmap_area_lock);
2403
2404        /* start scanning - we scan from the top, begin with the last area */
2405        area = term_area = last_area;
2406        start = offsets[area];
2407        end = start + sizes[area];
2408
2409        if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2410                base = vmalloc_end - last_end;
2411                goto found;
2412        }
2413        base = pvm_determine_end(&next, &prev, align) - end;
2414
2415        while (true) {
2416                BUG_ON(next && next->va_end <= base + end);
2417                BUG_ON(prev && prev->va_end > base + end);
2418
2419                /*
2420                 * base might have underflowed, add last_end before
2421                 * comparing.
2422                 */
2423                if (base + last_end < vmalloc_start + last_end) {
2424                        spin_unlock(&vmap_area_lock);
2425                        if (!purged) {
2426                                purge_vmap_area_lazy();
2427                                purged = true;
2428                                goto retry;
2429                        }
2430                        goto err_free;
2431                }
2432
2433                /*
2434                 * If next overlaps, move base downwards so that it's
2435                 * right below next and then recheck.
2436                 */
2437                if (next && next->va_start < base + end) {
2438                        base = pvm_determine_end(&next, &prev, align) - end;
2439                        term_area = area;
2440                        continue;
2441                }
2442
2443                /*
2444                 * If prev overlaps, shift down next and prev and move
2445                 * base so that it's right below new next and then
2446                 * recheck.
2447                 */
2448                if (prev && prev->va_end > base + start)  {
2449                        next = prev;
2450                        prev = node_to_va(rb_prev(&next->rb_node));
2451                        base = pvm_determine_end(&next, &prev, align) - end;
2452                        term_area = area;
2453                        continue;
2454                }
2455
2456                /*
2457                 * This area fits, move on to the previous one.  If
2458                 * the previous one is the terminal one, we're done.
2459                 */
2460                area = (area + nr_vms - 1) % nr_vms;
2461                if (area == term_area)
2462                        break;
2463                start = offsets[area];
2464                end = start + sizes[area];
2465                pvm_find_next_prev(base + end, &next, &prev);
2466        }
2467found:
2468        /* we've found a fitting base, insert all va's */
2469        for (area = 0; area < nr_vms; area++) {
2470                struct vmap_area *va = vas[area];
2471
2472                va->va_start = base + offsets[area];
2473                va->va_end = va->va_start + sizes[area];
2474                __insert_vmap_area(va);
2475        }
2476
2477        vmap_area_pcpu_hole = base + offsets[last_area];
2478
2479        spin_unlock(&vmap_area_lock);
2480
2481        /* insert all vm's */
2482        for (area = 0; area < nr_vms; area++)
2483                insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2484                                  pcpu_get_vm_areas);
2485
2486        kfree(vas);
2487        return vms;
2488
2489err_free:
2490        for (area = 0; area < nr_vms; area++) {
2491                kfree(vas[area]);
2492                kfree(vms[area]);
2493        }
2494err_free2:
2495        kfree(vas);
2496        kfree(vms);
2497        return NULL;
2498}
2499
2500/**
2501 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2502 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2503 * @nr_vms: the number of allocated areas
2504 *
2505 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2506 */
2507void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2508{
2509        int i;
2510
2511        for (i = 0; i < nr_vms; i++)
2512                free_vm_area(vms[i]);
2513        kfree(vms);
2514}
2515#endif  /* CONFIG_SMP */
2516
2517#ifdef CONFIG_PROC_FS
2518static void *s_start(struct seq_file *m, loff_t *pos)
2519        __acquires(&vmlist_lock)
2520{
2521        loff_t n = *pos;
2522        struct vm_struct *v;
2523
2524        read_lock(&vmlist_lock);
2525        v = vmlist;
2526        while (n > 0 && v) {
2527                n--;
2528                v = v->next;
2529        }
2530        if (!n)
2531                return v;
2532
2533        return NULL;
2534
2535}
2536
2537static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2538{
2539        struct vm_struct *v = p;
2540
2541        ++*pos;
2542        return v->next;
2543}
2544
2545static void s_stop(struct seq_file *m, void *p)
2546        __releases(&vmlist_lock)
2547{
2548        read_unlock(&vmlist_lock);
2549}
2550
2551static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2552{
2553        if (IS_ENABLED(CONFIG_NUMA)) {
2554                unsigned int nr, *counters = m->private;
2555
2556                if (!counters)
2557                        return;
2558
2559                memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2560
2561                for (nr = 0; nr < v->nr_pages; nr++)
2562                        counters[page_to_nid(v->pages[nr])]++;
2563
2564                for_each_node_state(nr, N_HIGH_MEMORY)
2565                        if (counters[nr])
2566                                seq_printf(m, " N%u=%u", nr, counters[nr]);
2567        }
2568}
2569
2570static int s_show(struct seq_file *m, void *p)
2571{
2572        struct vm_struct *v = p;
2573
2574        seq_printf(m, "0x%pK-0x%pK %7ld",
2575                v->addr, v->addr + v->size, v->size);
2576
2577        if (v->caller)
2578                seq_printf(m, " %pS", v->caller);
2579
2580        if (v->nr_pages)
2581                seq_printf(m, " pages=%d", v->nr_pages);
2582
2583        if (v->phys_addr)
2584                seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2585
2586        if (v->flags & VM_IOREMAP)
2587                seq_printf(m, " ioremap");
2588
2589        if (v->flags & VM_ALLOC)
2590                seq_printf(m, " vmalloc");
2591
2592        if (v->flags & VM_MAP)
2593                seq_printf(m, " vmap");
2594
2595        if (v->flags & VM_USERMAP)
2596                seq_printf(m, " user");
2597
2598        if (v->flags & VM_VPAGES)
2599                seq_printf(m, " vpages");
2600
2601        show_numa_info(m, v);
2602        seq_putc(m, '\n');
2603        return 0;
2604}
2605
2606static const struct seq_operations vmalloc_op = {
2607        .start = s_start,
2608        .next = s_next,
2609        .stop = s_stop,
2610        .show = s_show,
2611};
2612
2613static int vmalloc_open(struct inode *inode, struct file *file)
2614{
2615        unsigned int *ptr = NULL;
2616        int ret;
2617
2618        if (IS_ENABLED(CONFIG_NUMA)) {
2619                ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2620                if (ptr == NULL)
2621                        return -ENOMEM;
2622        }
2623        ret = seq_open(file, &vmalloc_op);
2624        if (!ret) {
2625                struct seq_file *m = file->private_data;
2626                m->private = ptr;
2627        } else
2628                kfree(ptr);
2629        return ret;
2630}
2631
2632static const struct file_operations proc_vmalloc_operations = {
2633        .open           = vmalloc_open,
2634        .read           = seq_read,
2635        .llseek         = seq_lseek,
2636        .release        = seq_release_private,
2637};
2638
2639static int __init proc_vmalloc_init(void)
2640{
2641        proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2642        return 0;
2643}
2644module_init(proc_vmalloc_init);
2645#endif
2646
2647
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.