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 <linux/llist.h>
  31#include <asm/uaccess.h>
  32#include <asm/tlbflush.h>
  33#include <asm/shmparam.h>
  34
  35struct vfree_deferred {
  36        struct llist_head list;
  37        struct work_struct wq;
  38};
  39static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
  40
  41static void __vunmap(const void *, int);
  42
  43static void free_work(struct work_struct *w)
  44{
  45        struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
  46        struct llist_node *llnode = llist_del_all(&p->list);
  47        while (llnode) {
  48                void *p = llnode;
  49                llnode = llist_next(llnode);
  50                __vunmap(p, 1);
  51        }
  52}
  53
  54/*** Page table manipulation functions ***/
  55
  56static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
  57{
  58        pte_t *pte;
  59
  60        pte = pte_offset_kernel(pmd, addr);
  61        do {
  62                pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
  63                WARN_ON(!pte_none(ptent) && !pte_present(ptent));
  64        } while (pte++, addr += PAGE_SIZE, addr != end);
  65}
  66
  67static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
  68{
  69        pmd_t *pmd;
  70        unsigned long next;
  71
  72        pmd = pmd_offset(pud, addr);
  73        do {
  74                next = pmd_addr_end(addr, end);
  75                if (pmd_none_or_clear_bad(pmd))
  76                        continue;
  77                vunmap_pte_range(pmd, addr, next);
  78        } while (pmd++, addr = next, addr != end);
  79}
  80
  81static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
  82{
  83        pud_t *pud;
  84        unsigned long next;
  85
  86        pud = pud_offset(pgd, addr);
  87        do {
  88                next = pud_addr_end(addr, end);
  89                if (pud_none_or_clear_bad(pud))
  90                        continue;
  91                vunmap_pmd_range(pud, addr, next);
  92        } while (pud++, addr = next, addr != end);
  93}
  94
  95static void vunmap_page_range(unsigned long addr, unsigned long end)
  96{
  97        pgd_t *pgd;
  98        unsigned long next;
  99
 100        BUG_ON(addr >= end);
 101        pgd = pgd_offset_k(addr);
 102        do {
 103                next = pgd_addr_end(addr, end);
 104                if (pgd_none_or_clear_bad(pgd))
 105                        continue;
 106                vunmap_pud_range(pgd, addr, next);
 107        } while (pgd++, addr = next, addr != end);
 108}
 109
 110static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
 111                unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 112{
 113        pte_t *pte;
 114
 115        /*
 116         * nr is a running index into the array which helps higher level
 117         * callers keep track of where we're up to.
 118         */
 119
 120        pte = pte_alloc_kernel(pmd, addr);
 121        if (!pte)
 122                return -ENOMEM;
 123        do {
 124                struct page *page = pages[*nr];
 125
 126                if (WARN_ON(!pte_none(*pte)))
 127                        return -EBUSY;
 128                if (WARN_ON(!page))
 129                        return -ENOMEM;
 130                set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 131                (*nr)++;
 132        } while (pte++, addr += PAGE_SIZE, addr != end);
 133        return 0;
 134}
 135
 136static int vmap_pmd_range(pud_t *pud, unsigned long addr,
 137                unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 138{
 139        pmd_t *pmd;
 140        unsigned long next;
 141
 142        pmd = pmd_alloc(&init_mm, pud, addr);
 143        if (!pmd)
 144                return -ENOMEM;
 145        do {
 146                next = pmd_addr_end(addr, end);
 147                if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
 148                        return -ENOMEM;
 149        } while (pmd++, addr = next, addr != end);
 150        return 0;
 151}
 152
 153static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
 154                unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 155{
 156        pud_t *pud;
 157        unsigned long next;
 158
 159        pud = pud_alloc(&init_mm, pgd, addr);
 160        if (!pud)
 161                return -ENOMEM;
 162        do {
 163                next = pud_addr_end(addr, end);
 164                if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
 165                        return -ENOMEM;
 166        } while (pud++, addr = next, addr != end);
 167        return 0;
 168}
 169
 170/*
 171 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
 172 * will have pfns corresponding to the "pages" array.
 173 *
 174 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
 175 */
 176static int vmap_page_range_noflush(unsigned long start, unsigned long end,
 177                                   pgprot_t prot, struct page **pages)
 178{
 179        pgd_t *pgd;
 180        unsigned long next;
 181        unsigned long addr = start;
 182        int err = 0;
 183        int nr = 0;
 184
 185        BUG_ON(addr >= end);
 186        pgd = pgd_offset_k(addr);
 187        do {
 188                next = pgd_addr_end(addr, end);
 189                err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
 190                if (err)
 191                        return err;
 192        } while (pgd++, addr = next, addr != end);
 193
 194        return nr;
 195}
 196
 197static int vmap_page_range(unsigned long start, unsigned long end,
 198                           pgprot_t prot, struct page **pages)
 199{
 200        int ret;
 201
 202        ret = vmap_page_range_noflush(start, end, prot, pages);
 203        flush_cache_vmap(start, end);
 204        return ret;
 205}
 206
 207int is_vmalloc_or_module_addr(const void *x)
 208{
 209        /*
 210         * ARM, x86-64 and sparc64 put modules in a special place,
 211         * and fall back on vmalloc() if that fails. Others
 212         * just put it in the vmalloc space.
 213         */
 214#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 215        unsigned long addr = (unsigned long)x;
 216        if (addr >= MODULES_VADDR && addr < MODULES_END)
 217                return 1;
 218#endif
 219        return is_vmalloc_addr(x);
 220}
 221
 222/*
 223 * Walk a vmap address to the struct page it maps.
 224 */
 225struct page *vmalloc_to_page(const void *vmalloc_addr)
 226{
 227        unsigned long addr = (unsigned long) vmalloc_addr;
 228        struct page *page = NULL;
 229        pgd_t *pgd = pgd_offset_k(addr);
 230
 231        /*
 232         * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 233         * architectures that do not vmalloc module space
 234         */
 235        VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 236
 237        if (!pgd_none(*pgd)) {
 238                pud_t *pud = pud_offset(pgd, addr);
 239                if (!pud_none(*pud)) {
 240                        pmd_t *pmd = pmd_offset(pud, addr);
 241                        if (!pmd_none(*pmd)) {
 242                                pte_t *ptep, pte;
 243
 244                                ptep = pte_offset_map(pmd, addr);
 245                                pte = *ptep;
 246                                if (pte_present(pte))
 247                                        page = pte_page(pte);
 248                                pte_unmap(ptep);
 249                        }
 250                }
 251        }
 252        return page;
 253}
 254EXPORT_SYMBOL(vmalloc_to_page);
 255
 256/*
 257 * Map a vmalloc()-space virtual address to the physical page frame number.
 258 */
 259unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 260{
 261        return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 262}
 263EXPORT_SYMBOL(vmalloc_to_pfn);
 264
 265
 266/*** Global kva allocator ***/
 267
 268#define VM_LAZY_FREE    0x01
 269#define VM_LAZY_FREEING 0x02
 270#define VM_VM_AREA      0x04
 271
 272static DEFINE_SPINLOCK(vmap_area_lock);
 273/* Export for kexec only */
 274LIST_HEAD(vmap_area_list);
 275static struct rb_root vmap_area_root = RB_ROOT;
 276
 277/* The vmap cache globals are protected by vmap_area_lock */
 278static struct rb_node *free_vmap_cache;
 279static unsigned long cached_hole_size;
 280static unsigned long cached_vstart;
 281static unsigned long cached_align;
 282
 283static unsigned long vmap_area_pcpu_hole;
 284
 285static struct vmap_area *__find_vmap_area(unsigned long addr)
 286{
 287        struct rb_node *n = vmap_area_root.rb_node;
 288
 289        while (n) {
 290                struct vmap_area *va;
 291
 292                va = rb_entry(n, struct vmap_area, rb_node);
 293                if (addr < va->va_start)
 294                        n = n->rb_left;
 295                else if (addr >= va->va_end)
 296                        n = n->rb_right;
 297                else
 298                        return va;
 299        }
 300
 301        return NULL;
 302}
 303
 304static void __insert_vmap_area(struct vmap_area *va)
 305{
 306        struct rb_node **p = &vmap_area_root.rb_node;
 307        struct rb_node *parent = NULL;
 308        struct rb_node *tmp;
 309
 310        while (*p) {
 311                struct vmap_area *tmp_va;
 312
 313                parent = *p;
 314                tmp_va = rb_entry(parent, struct vmap_area, rb_node);
 315                if (va->va_start < tmp_va->va_end)
 316                        p = &(*p)->rb_left;
 317                else if (va->va_end > tmp_va->va_start)
 318                        p = &(*p)->rb_right;
 319                else
 320                        BUG();
 321        }
 322
 323        rb_link_node(&va->rb_node, parent, p);
 324        rb_insert_color(&va->rb_node, &vmap_area_root);
 325
 326        /* address-sort this list */
 327        tmp = rb_prev(&va->rb_node);
 328        if (tmp) {
 329                struct vmap_area *prev;
 330                prev = rb_entry(tmp, struct vmap_area, rb_node);
 331                list_add_rcu(&va->list, &prev->list);
 332        } else
 333                list_add_rcu(&va->list, &vmap_area_list);
 334}
 335
 336static void purge_vmap_area_lazy(void);
 337
 338/*
 339 * Allocate a region of KVA of the specified size and alignment, within the
 340 * vstart and vend.
 341 */
 342static struct vmap_area *alloc_vmap_area(unsigned long size,
 343                                unsigned long align,
 344                                unsigned long vstart, unsigned long vend,
 345                                int node, gfp_t gfp_mask)
 346{
 347        struct vmap_area *va;
 348        struct rb_node *n;
 349        unsigned long addr;
 350        int purged = 0;
 351        struct vmap_area *first;
 352
 353        BUG_ON(!size);
 354        BUG_ON(size & ~PAGE_MASK);
 355        BUG_ON(!is_power_of_2(align));
 356
 357        va = kmalloc_node(sizeof(struct vmap_area),
 358                        gfp_mask & GFP_RECLAIM_MASK, node);
 359        if (unlikely(!va))
 360                return ERR_PTR(-ENOMEM);
 361
 362retry:
 363        spin_lock(&vmap_area_lock);
 364        /*
 365         * Invalidate cache if we have more permissive parameters.
 366         * cached_hole_size notes the largest hole noticed _below_
 367         * the vmap_area cached in free_vmap_cache: if size fits
 368         * into that hole, we want to scan from vstart to reuse
 369         * the hole instead of allocating above free_vmap_cache.
 370         * Note that __free_vmap_area may update free_vmap_cache
 371         * without updating cached_hole_size or cached_align.
 372         */
 373        if (!free_vmap_cache ||
 374                        size < cached_hole_size ||
 375                        vstart < cached_vstart ||
 376                        align < cached_align) {
 377nocache:
 378                cached_hole_size = 0;
 379                free_vmap_cache = NULL;
 380        }
 381        /* record if we encounter less permissive parameters */
 382        cached_vstart = vstart;
 383        cached_align = align;
 384
 385        /* find starting point for our search */
 386        if (free_vmap_cache) {
 387                first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 388                addr = ALIGN(first->va_end, align);
 389                if (addr < vstart)
 390                        goto nocache;
 391                if (addr + size < addr)
 392                        goto overflow;
 393
 394        } else {
 395                addr = ALIGN(vstart, align);
 396                if (addr + size < addr)
 397                        goto overflow;
 398
 399                n = vmap_area_root.rb_node;
 400                first = NULL;
 401
 402                while (n) {
 403                        struct vmap_area *tmp;
 404                        tmp = rb_entry(n, struct vmap_area, rb_node);
 405                        if (tmp->va_end >= addr) {
 406                                first = tmp;
 407                                if (tmp->va_start <= addr)
 408                                        break;
 409                                n = n->rb_left;
 410                        } else
 411                                n = n->rb_right;
 412                }
 413
 414                if (!first)
 415                        goto found;
 416        }
 417
 418        /* from the starting point, walk areas until a suitable hole is found */
 419        while (addr + size > first->va_start && addr + size <= vend) {
 420                if (addr + cached_hole_size < first->va_start)
 421                        cached_hole_size = first->va_start - addr;
 422                addr = ALIGN(first->va_end, align);
 423                if (addr + size < addr)
 424                        goto overflow;
 425
 426                if (list_is_last(&first->list, &vmap_area_list))
 427                        goto found;
 428
 429                first = list_entry(first->list.next,
 430                                struct vmap_area, list);
 431        }
 432
 433found:
 434        if (addr + size > vend)
 435                goto overflow;
 436
 437        va->va_start = addr;
 438        va->va_end = addr + size;
 439        va->flags = 0;
 440        __insert_vmap_area(va);
 441        free_vmap_cache = &va->rb_node;
 442        spin_unlock(&vmap_area_lock);
 443
 444        BUG_ON(va->va_start & (align-1));
 445        BUG_ON(va->va_start < vstart);
 446        BUG_ON(va->va_end > vend);
 447
 448        return va;
 449
 450overflow:
 451        spin_unlock(&vmap_area_lock);
 452        if (!purged) {
 453                purge_vmap_area_lazy();
 454                purged = 1;
 455                goto retry;
 456        }
 457        if (printk_ratelimit())
 458                printk(KERN_WARNING
 459                        "vmap allocation for size %lu failed: "
 460                        "use vmalloc=<size> to increase size.\n", size);
 461        kfree(va);
 462        return ERR_PTR(-EBUSY);
 463}
 464
 465static void __free_vmap_area(struct vmap_area *va)
 466{
 467        BUG_ON(RB_EMPTY_NODE(&va->rb_node));
 468
 469        if (free_vmap_cache) {
 470                if (va->va_end < cached_vstart) {
 471                        free_vmap_cache = NULL;
 472                } else {
 473                        struct vmap_area *cache;
 474                        cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 475                        if (va->va_start <= cache->va_start) {
 476                                free_vmap_cache = rb_prev(&va->rb_node);
 477                                /*
 478                                 * We don't try to update cached_hole_size or
 479                                 * cached_align, but it won't go very wrong.
 480                                 */
 481                        }
 482                }
 483        }
 484        rb_erase(&va->rb_node, &vmap_area_root);
 485        RB_CLEAR_NODE(&va->rb_node);
 486        list_del_rcu(&va->list);
 487
 488        /*
 489         * Track the highest possible candidate for pcpu area
 490         * allocation.  Areas outside of vmalloc area can be returned
 491         * here too, consider only end addresses which fall inside
 492         * vmalloc area proper.
 493         */
 494        if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
 495                vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
 496
 497        kfree_rcu(va, rcu_head);
 498}
 499
 500/*
 501 * Free a region of KVA allocated by alloc_vmap_area
 502 */
 503static void free_vmap_area(struct vmap_area *va)
 504{
 505        spin_lock(&vmap_area_lock);
 506        __free_vmap_area(va);
 507        spin_unlock(&vmap_area_lock);
 508}
 509
 510/*
 511 * Clear the pagetable entries of a given vmap_area
 512 */
 513static void unmap_vmap_area(struct vmap_area *va)
 514{
 515        vunmap_page_range(va->va_start, va->va_end);
 516}
 517
 518static void vmap_debug_free_range(unsigned long start, unsigned long end)
 519{
 520        /*
 521         * Unmap page tables and force a TLB flush immediately if
 522         * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
 523         * bugs similarly to those in linear kernel virtual address
 524         * space after a page has been freed.
 525         *
 526         * All the lazy freeing logic is still retained, in order to
 527         * minimise intrusiveness of this debugging feature.
 528         *
 529         * This is going to be *slow* (linear kernel virtual address
 530         * debugging doesn't do a broadcast TLB flush so it is a lot
 531         * faster).
 532         */
 533#ifdef CONFIG_DEBUG_PAGEALLOC
 534        vunmap_page_range(start, end);
 535        flush_tlb_kernel_range(start, end);
 536#endif
 537}
 538
 539/*
 540 * lazy_max_pages is the maximum amount of virtual address space we gather up
 541 * before attempting to purge with a TLB flush.
 542 *
 543 * There is a tradeoff here: a larger number will cover more kernel page tables
 544 * and take slightly longer to purge, but it will linearly reduce the number of
 545 * global TLB flushes that must be performed. It would seem natural to scale
 546 * this number up linearly with the number of CPUs (because vmapping activity
 547 * could also scale linearly with the number of CPUs), however it is likely
 548 * that in practice, workloads might be constrained in other ways that mean
 549 * vmap activity will not scale linearly with CPUs. Also, I want to be
 550 * conservative and not introduce a big latency on huge systems, so go with
 551 * a less aggressive log scale. It will still be an improvement over the old
 552 * code, and it will be simple to change the scale factor if we find that it
 553 * becomes a problem on bigger systems.
 554 */
 555static unsigned long lazy_max_pages(void)
 556{
 557        unsigned int log;
 558
 559        log = fls(num_online_cpus());
 560
 561        return log * (32UL * 1024 * 1024 / PAGE_SIZE);
 562}
 563
 564static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
 565
 566/* for per-CPU blocks */
 567static void purge_fragmented_blocks_allcpus(void);
 568
 569/*
 570 * called before a call to iounmap() if the caller wants vm_area_struct's
 571 * immediately freed.
 572 */
 573void set_iounmap_nonlazy(void)
 574{
 575        atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
 576}
 577
 578/*
 579 * Purges all lazily-freed vmap areas.
 580 *
 581 * If sync is 0 then don't purge if there is already a purge in progress.
 582 * If force_flush is 1, then flush kernel TLBs between *start and *end even
 583 * if we found no lazy vmap areas to unmap (callers can use this to optimise
 584 * their own TLB flushing).
 585 * Returns with *start = min(*start, lowest purged address)
 586 *              *end = max(*end, highest purged address)
 587 */
 588static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
 589                                        int sync, int force_flush)
 590{
 591        static DEFINE_SPINLOCK(purge_lock);
 592        LIST_HEAD(valist);
 593        struct vmap_area *va;
 594        struct vmap_area *n_va;
 595        int nr = 0;
 596
 597        /*
 598         * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
 599         * should not expect such behaviour. This just simplifies locking for
 600         * the case that isn't actually used at the moment anyway.
 601         */
 602        if (!sync && !force_flush) {
 603                if (!spin_trylock(&purge_lock))
 604                        return;
 605        } else
 606                spin_lock(&purge_lock);
 607
 608        if (sync)
 609                purge_fragmented_blocks_allcpus();
 610
 611        rcu_read_lock();
 612        list_for_each_entry_rcu(va, &vmap_area_list, list) {
 613                if (va->flags & VM_LAZY_FREE) {
 614                        if (va->va_start < *start)
 615                                *start = va->va_start;
 616                        if (va->va_end > *end)
 617                                *end = va->va_end;
 618                        nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
 619                        list_add_tail(&va->purge_list, &valist);
 620                        va->flags |= VM_LAZY_FREEING;
 621                        va->flags &= ~VM_LAZY_FREE;
 622                }
 623        }
 624        rcu_read_unlock();
 625
 626        if (nr)
 627                atomic_sub(nr, &vmap_lazy_nr);
 628
 629        if (nr || force_flush)
 630                flush_tlb_kernel_range(*start, *end);
 631
 632        if (nr) {
 633                spin_lock(&vmap_area_lock);
 634                list_for_each_entry_safe(va, n_va, &valist, purge_list)
 635                        __free_vmap_area(va);
 636                spin_unlock(&vmap_area_lock);
 637        }
 638        spin_unlock(&purge_lock);
 639}
 640
 641/*
 642 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
 643 * is already purging.
 644 */
 645static void try_purge_vmap_area_lazy(void)
 646{
 647        unsigned long start = ULONG_MAX, end = 0;
 648
 649        __purge_vmap_area_lazy(&start, &end, 0, 0);
 650}
 651
 652/*
 653 * Kick off a purge of the outstanding lazy areas.
 654 */
 655static void purge_vmap_area_lazy(void)
 656{
 657        unsigned long start = ULONG_MAX, end = 0;
 658
 659        __purge_vmap_area_lazy(&start, &end, 1, 0);
 660}
 661
 662/*
 663 * Free a vmap area, caller ensuring that the area has been unmapped
 664 * and flush_cache_vunmap had been called for the correct range
 665 * previously.
 666 */
 667static void free_vmap_area_noflush(struct vmap_area *va)
 668{
 669        va->flags |= VM_LAZY_FREE;
 670        atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
 671        if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
 672                try_purge_vmap_area_lazy();
 673}
 674
 675/*
 676 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
 677 * called for the correct range previously.
 678 */
 679static void free_unmap_vmap_area_noflush(struct vmap_area *va)
 680{
 681        unmap_vmap_area(va);
 682        free_vmap_area_noflush(va);
 683}
 684
 685/*
 686 * Free and unmap a vmap area
 687 */
 688static void free_unmap_vmap_area(struct vmap_area *va)
 689{
 690        flush_cache_vunmap(va->va_start, va->va_end);
 691        free_unmap_vmap_area_noflush(va);
 692}
 693
 694static struct vmap_area *find_vmap_area(unsigned long addr)
 695{
 696        struct vmap_area *va;
 697
 698        spin_lock(&vmap_area_lock);
 699        va = __find_vmap_area(addr);
 700        spin_unlock(&vmap_area_lock);
 701
 702        return va;
 703}
 704
 705static void free_unmap_vmap_area_addr(unsigned long addr)
 706{
 707        struct vmap_area *va;
 708
 709        va = find_vmap_area(addr);
 710        BUG_ON(!va);
 711        free_unmap_vmap_area(va);
 712}
 713
 714
 715/*** Per cpu kva allocator ***/
 716
 717/*
 718 * vmap space is limited especially on 32 bit architectures. Ensure there is
 719 * room for at least 16 percpu vmap blocks per CPU.
 720 */
 721/*
 722 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
 723 * to #define VMALLOC_SPACE             (VMALLOC_END-VMALLOC_START). Guess
 724 * instead (we just need a rough idea)
 725 */
 726#if BITS_PER_LONG == 32
 727#define VMALLOC_SPACE           (128UL*1024*1024)
 728#else
 729#define VMALLOC_SPACE           (128UL*1024*1024*1024)
 730#endif
 731
 732#define VMALLOC_PAGES           (VMALLOC_SPACE / PAGE_SIZE)
 733#define VMAP_MAX_ALLOC          BITS_PER_LONG   /* 256K with 4K pages */
 734#define VMAP_BBMAP_BITS_MAX     1024    /* 4MB with 4K pages */
 735#define VMAP_BBMAP_BITS_MIN     (VMAP_MAX_ALLOC*2)
 736#define VMAP_MIN(x, y)          ((x) < (y) ? (x) : (y)) /* can't use min() */
 737#define VMAP_MAX(x, y)          ((x) > (y) ? (x) : (y)) /* can't use max() */
 738#define VMAP_BBMAP_BITS         \
 739                VMAP_MIN(VMAP_BBMAP_BITS_MAX,   \
 740                VMAP_MAX(VMAP_BBMAP_BITS_MIN,   \
 741                        VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
 742
 743#define VMAP_BLOCK_SIZE         (VMAP_BBMAP_BITS * PAGE_SIZE)
 744
 745static bool vmap_initialized __read_mostly = false;
 746
 747struct vmap_block_queue {
 748        spinlock_t lock;
 749        struct list_head free;
 750};
 751
 752struct vmap_block {
 753        spinlock_t lock;
 754        struct vmap_area *va;
 755        struct vmap_block_queue *vbq;
 756        unsigned long free, dirty;
 757        DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
 758        struct list_head free_list;
 759        struct rcu_head rcu_head;
 760        struct list_head purge;
 761};
 762
 763/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
 764static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
 765
 766/*
 767 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
 768 * in the free path. Could get rid of this if we change the API to return a
 769 * "cookie" from alloc, to be passed to free. But no big deal yet.
 770 */
 771static DEFINE_SPINLOCK(vmap_block_tree_lock);
 772static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
 773
 774/*
 775 * We should probably have a fallback mechanism to allocate virtual memory
 776 * out of partially filled vmap blocks. However vmap block sizing should be
 777 * fairly reasonable according to the vmalloc size, so it shouldn't be a
 778 * big problem.
 779 */
 780
 781static unsigned long addr_to_vb_idx(unsigned long addr)
 782{
 783        addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
 784        addr /= VMAP_BLOCK_SIZE;
 785        return addr;
 786}
 787
 788static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
 789{
 790        struct vmap_block_queue *vbq;
 791        struct vmap_block *vb;
 792        struct vmap_area *va;
 793        unsigned long vb_idx;
 794        int node, err;
 795
 796        node = numa_node_id();
 797
 798        vb = kmalloc_node(sizeof(struct vmap_block),
 799                        gfp_mask & GFP_RECLAIM_MASK, node);
 800        if (unlikely(!vb))
 801                return ERR_PTR(-ENOMEM);
 802
 803        va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
 804                                        VMALLOC_START, VMALLOC_END,
 805                                        node, gfp_mask);
 806        if (IS_ERR(va)) {
 807                kfree(vb);
 808                return ERR_CAST(va);
 809        }
 810
 811        err = radix_tree_preload(gfp_mask);
 812        if (unlikely(err)) {
 813                kfree(vb);
 814                free_vmap_area(va);
 815                return ERR_PTR(err);
 816        }
 817
 818        spin_lock_init(&vb->lock);
 819        vb->va = va;
 820        vb->free = VMAP_BBMAP_BITS;
 821        vb->dirty = 0;
 822        bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
 823        INIT_LIST_HEAD(&vb->free_list);
 824
 825        vb_idx = addr_to_vb_idx(va->va_start);
 826        spin_lock(&vmap_block_tree_lock);
 827        err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
 828        spin_unlock(&vmap_block_tree_lock);
 829        BUG_ON(err);
 830        radix_tree_preload_end();
 831
 832        vbq = &get_cpu_var(vmap_block_queue);
 833        vb->vbq = vbq;
 834        spin_lock(&vbq->lock);
 835        list_add_rcu(&vb->free_list, &vbq->free);
 836        spin_unlock(&vbq->lock);
 837        put_cpu_var(vmap_block_queue);
 838
 839        return vb;
 840}
 841
 842static void free_vmap_block(struct vmap_block *vb)
 843{
 844        struct vmap_block *tmp;
 845        unsigned long vb_idx;
 846
 847        vb_idx = addr_to_vb_idx(vb->va->va_start);
 848        spin_lock(&vmap_block_tree_lock);
 849        tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
 850        spin_unlock(&vmap_block_tree_lock);
 851        BUG_ON(tmp != vb);
 852
 853        free_vmap_area_noflush(vb->va);
 854        kfree_rcu(vb, rcu_head);
 855}
 856
 857static void purge_fragmented_blocks(int cpu)
 858{
 859        LIST_HEAD(purge);
 860        struct vmap_block *vb;
 861        struct vmap_block *n_vb;
 862        struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
 863
 864        rcu_read_lock();
 865        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 866
 867                if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
 868                        continue;
 869
 870                spin_lock(&vb->lock);
 871                if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
 872                        vb->free = 0; /* prevent further allocs after releasing lock */
 873                        vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
 874                        bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
 875                        spin_lock(&vbq->lock);
 876                        list_del_rcu(&vb->free_list);
 877                        spin_unlock(&vbq->lock);
 878                        spin_unlock(&vb->lock);
 879                        list_add_tail(&vb->purge, &purge);
 880                } else
 881                        spin_unlock(&vb->lock);
 882        }
 883        rcu_read_unlock();
 884
 885        list_for_each_entry_safe(vb, n_vb, &purge, purge) {
 886                list_del(&vb->purge);
 887                free_vmap_block(vb);
 888        }
 889}
 890
 891static void purge_fragmented_blocks_allcpus(void)
 892{
 893        int cpu;
 894
 895        for_each_possible_cpu(cpu)
 896                purge_fragmented_blocks(cpu);
 897}
 898
 899static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
 900{
 901        struct vmap_block_queue *vbq;
 902        struct vmap_block *vb;
 903        unsigned long addr = 0;
 904        unsigned int order;
 905
 906        BUG_ON(size & ~PAGE_MASK);
 907        BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 908        if (WARN_ON(size == 0)) {
 909                /*
 910                 * Allocating 0 bytes isn't what caller wants since
 911                 * get_order(0) returns funny result. Just warn and terminate
 912                 * early.
 913                 */
 914                return NULL;
 915        }
 916        order = get_order(size);
 917
 918again:
 919        rcu_read_lock();
 920        vbq = &get_cpu_var(vmap_block_queue);
 921        list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 922                int i;
 923
 924                spin_lock(&vb->lock);
 925                if (vb->free < 1UL << order)
 926                        goto next;
 927
 928                i = VMAP_BBMAP_BITS - vb->free;
 929                addr = vb->va->va_start + (i << PAGE_SHIFT);
 930                BUG_ON(addr_to_vb_idx(addr) !=
 931                                addr_to_vb_idx(vb->va->va_start));
 932                vb->free -= 1UL << order;
 933                if (vb->free == 0) {
 934                        spin_lock(&vbq->lock);
 935                        list_del_rcu(&vb->free_list);
 936                        spin_unlock(&vbq->lock);
 937                }
 938                spin_unlock(&vb->lock);
 939                break;
 940next:
 941                spin_unlock(&vb->lock);
 942        }
 943
 944        put_cpu_var(vmap_block_queue);
 945        rcu_read_unlock();
 946
 947        if (!addr) {
 948                vb = new_vmap_block(gfp_mask);
 949                if (IS_ERR(vb))
 950                        return vb;
 951                goto again;
 952        }
 953
 954        return (void *)addr;
 955}
 956
 957static void vb_free(const void *addr, unsigned long size)
 958{
 959        unsigned long offset;
 960        unsigned long vb_idx;
 961        unsigned int order;
 962        struct vmap_block *vb;
 963
 964        BUG_ON(size & ~PAGE_MASK);
 965        BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 966
 967        flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
 968
 969        order = get_order(size);
 970
 971        offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
 972
 973        vb_idx = addr_to_vb_idx((unsigned long)addr);
 974        rcu_read_lock();
 975        vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
 976        rcu_read_unlock();
 977        BUG_ON(!vb);
 978
 979        vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
 980
 981        spin_lock(&vb->lock);
 982        BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
 983
 984        vb->dirty += 1UL << order;
 985        if (vb->dirty == VMAP_BBMAP_BITS) {
 986                BUG_ON(vb->free);
 987                spin_unlock(&vb->lock);
 988                free_vmap_block(vb);
 989        } else
 990                spin_unlock(&vb->lock);
 991}
 992
 993/**
 994 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
 995 *
 996 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
 997 * to amortize TLB flushing overheads. What this means is that any page you
 998 * have now, may, in a former life, have been mapped into kernel virtual
 999 * address by the vmap layer and so there might be some CPUs with TLB entries
1000 * still referencing that page (additional to the regular 1:1 kernel mapping).
1001 *
1002 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1003 * be sure that none of the pages we have control over will have any aliases
1004 * from the vmap layer.
1005 */
1006void vm_unmap_aliases(void)
1007{
1008        unsigned long start = ULONG_MAX, end = 0;
1009        int cpu;
1010        int flush = 0;
1011
1012        if (unlikely(!vmap_initialized))
1013                return;
1014
1015        for_each_possible_cpu(cpu) {
1016                struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1017                struct vmap_block *vb;
1018
1019                rcu_read_lock();
1020                list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1021                        int i;
1022
1023                        spin_lock(&vb->lock);
1024                        i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1025                        while (i < VMAP_BBMAP_BITS) {
1026                                unsigned long s, e;
1027                                int j;
1028                                j = find_next_zero_bit(vb->dirty_map,
1029                                        VMAP_BBMAP_BITS, i);
1030
1031                                s = vb->va->va_start + (i << PAGE_SHIFT);
1032                                e = vb->va->va_start + (j << PAGE_SHIFT);
1033                                flush = 1;
1034
1035                                if (s < start)
1036                                        start = s;
1037                                if (e > end)
1038                                        end = e;
1039
1040                                i = j;
1041                                i = find_next_bit(vb->dirty_map,
1042                                                        VMAP_BBMAP_BITS, i);
1043                        }
1044                        spin_unlock(&vb->lock);
1045                }
1046                rcu_read_unlock();
1047        }
1048
1049        __purge_vmap_area_lazy(&start, &end, 1, flush);
1050}
1051EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1052
1053/**
1054 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1055 * @mem: the pointer returned by vm_map_ram
1056 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1057 */
1058void vm_unmap_ram(const void *mem, unsigned int count)
1059{
1060        unsigned long size = count << PAGE_SHIFT;
1061        unsigned long addr = (unsigned long)mem;
1062
1063        BUG_ON(!addr);
1064        BUG_ON(addr < VMALLOC_START);
1065        BUG_ON(addr > VMALLOC_END);
1066        BUG_ON(addr & (PAGE_SIZE-1));
1067
1068        debug_check_no_locks_freed(mem, size);
1069        vmap_debug_free_range(addr, addr+size);
1070
1071        if (likely(count <= VMAP_MAX_ALLOC))
1072                vb_free(mem, size);
1073        else
1074                free_unmap_vmap_area_addr(addr);
1075}
1076EXPORT_SYMBOL(vm_unmap_ram);
1077
1078/**
1079 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1080 * @pages: an array of pointers to the pages to be mapped
1081 * @count: number of pages
1082 * @node: prefer to allocate data structures on this node
1083 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1084 *
1085 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1086 */
1087void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1088{
1089        unsigned long size = count << PAGE_SHIFT;
1090        unsigned long addr;
1091        void *mem;
1092
1093        if (likely(count <= VMAP_MAX_ALLOC)) {
1094                mem = vb_alloc(size, GFP_KERNEL);
1095                if (IS_ERR(mem))
1096                        return NULL;
1097                addr = (unsigned long)mem;
1098        } else {
1099                struct vmap_area *va;
1100                va = alloc_vmap_area(size, PAGE_SIZE,
1101                                VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1102                if (IS_ERR(va))
1103                        return NULL;
1104
1105                addr = va->va_start;
1106                mem = (void *)addr;
1107        }
1108        if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1109                vm_unmap_ram(mem, count);
1110                return NULL;
1111        }
1112        return mem;
1113}
1114EXPORT_SYMBOL(vm_map_ram);
1115
1116static struct vm_struct *vmlist __initdata;
1117/**
1118 * vm_area_add_early - add vmap area early during boot
1119 * @vm: vm_struct to add
1120 *
1121 * This function is used to add fixed kernel vm area to vmlist before
1122 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
1123 * should contain proper values and the other fields should be zero.
1124 *
1125 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1126 */
1127void __init vm_area_add_early(struct vm_struct *vm)
1128{
1129        struct vm_struct *tmp, **p;
1130
1131        BUG_ON(vmap_initialized);
1132        for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1133                if (tmp->addr >= vm->addr) {
1134                        BUG_ON(tmp->addr < vm->addr + vm->size);
1135                        break;
1136                } else
1137                        BUG_ON(tmp->addr + tmp->size > vm->addr);
1138        }
1139        vm->next = *p;
1140        *p = vm;
1141}
1142
1143/**
1144 * vm_area_register_early - register vmap area early during boot
1145 * @vm: vm_struct to register
1146 * @align: requested alignment
1147 *
1148 * This function is used to register kernel vm area before
1149 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1150 * proper values on entry and other fields should be zero.  On return,
1151 * vm->addr contains the allocated address.
1152 *
1153 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1154 */
1155void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1156{
1157        static size_t vm_init_off __initdata;
1158        unsigned long addr;
1159
1160        addr = ALIGN(VMALLOC_START + vm_init_off, align);
1161        vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1162
1163        vm->addr = (void *)addr;
1164
1165        vm_area_add_early(vm);
1166}
1167
1168void __init vmalloc_init(void)
1169{
1170        struct vmap_area *va;
1171        struct vm_struct *tmp;
1172        int i;
1173
1174        for_each_possible_cpu(i) {
1175                struct vmap_block_queue *vbq;
1176                struct vfree_deferred *p;
1177
1178                vbq = &per_cpu(vmap_block_queue, i);
1179                spin_lock_init(&vbq->lock);
1180                INIT_LIST_HEAD(&vbq->free);
1181                p = &per_cpu(vfree_deferred, i);
1182                init_llist_head(&p->list);
1183                INIT_WORK(&p->wq, free_work);
1184        }
1185
1186        /* Import existing vmlist entries. */
1187        for (tmp = vmlist; tmp; tmp = tmp->next) {
1188                va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1189                va->flags = VM_VM_AREA;
1190                va->va_start = (unsigned long)tmp->addr;
1191                va->va_end = va->va_start + tmp->size;
1192                va->vm = tmp;
1193                __insert_vmap_area(va);
1194        }
1195
1196        vmap_area_pcpu_hole = VMALLOC_END;
1197
1198        vmap_initialized = true;
1199}
1200
1201/**
1202 * map_kernel_range_noflush - map kernel VM area with the specified pages
1203 * @addr: start of the VM area to map
1204 * @size: size of the VM area to map
1205 * @prot: page protection flags to use
1206 * @pages: pages to map
1207 *
1208 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1209 * specify should have been allocated using get_vm_area() and its
1210 * friends.
1211 *
1212 * NOTE:
1213 * This function does NOT do any cache flushing.  The caller is
1214 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1215 * before calling this function.
1216 *
1217 * RETURNS:
1218 * The number of pages mapped on success, -errno on failure.
1219 */
1220int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1221                             pgprot_t prot, struct page **pages)
1222{
1223        return vmap_page_range_noflush(addr, addr + size, prot, pages);
1224}
1225
1226/**
1227 * unmap_kernel_range_noflush - unmap kernel VM area
1228 * @addr: start of the VM area to unmap
1229 * @size: size of the VM area to unmap
1230 *
1231 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1232 * specify should have been allocated using get_vm_area() and its
1233 * friends.
1234 *
1235 * NOTE:
1236 * This function does NOT do any cache flushing.  The caller is
1237 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1238 * before calling this function and flush_tlb_kernel_range() after.
1239 */
1240void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1241{
1242        vunmap_page_range(addr, addr + size);
1243}
1244EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1245
1246/**
1247 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1248 * @addr: start of the VM area to unmap
1249 * @size: size of the VM area to unmap
1250 *
1251 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1252 * the unmapping and tlb after.
1253 */
1254void unmap_kernel_range(unsigned long addr, unsigned long size)
1255{
1256        unsigned long end = addr + size;
1257
1258        flush_cache_vunmap(addr, end);
1259        vunmap_page_range(addr, end);
1260        flush_tlb_kernel_range(addr, end);
1261}
1262
1263int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1264{
1265        unsigned long addr = (unsigned long)area->addr;
1266        unsigned long end = addr + area->size - PAGE_SIZE;
1267        int err;
1268
1269        err = vmap_page_range(addr, end, prot, *pages);
1270        if (err > 0) {
1271                *pages += err;
1272                err = 0;
1273        }
1274
1275        return err;
1276}
1277EXPORT_SYMBOL_GPL(map_vm_area);
1278
1279static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1280                              unsigned long flags, const void *caller)
1281{
1282        spin_lock(&vmap_area_lock);
1283        vm->flags = flags;
1284        vm->addr = (void *)va->va_start;
1285        vm->size = va->va_end - va->va_start;
1286        vm->caller = caller;
1287        va->vm = vm;
1288        va->flags |= VM_VM_AREA;
1289        spin_unlock(&vmap_area_lock);
1290}
1291
1292static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1293{
1294        /*
1295         * Before removing VM_UNINITIALIZED,
1296         * we should make sure that vm has proper values.
1297         * Pair with smp_rmb() in show_numa_info().
1298         */
1299        smp_wmb();
1300        vm->flags &= ~VM_UNINITIALIZED;
1301}
1302
1303static struct vm_struct *__get_vm_area_node(unsigned long size,
1304                unsigned long align, unsigned long flags, unsigned long start,
1305                unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1306{
1307        struct vmap_area *va;
1308        struct vm_struct *area;
1309
1310        BUG_ON(in_interrupt());
1311        if (flags & VM_IOREMAP)
1312                align = 1ul << clamp(fls(size), PAGE_SHIFT, IOREMAP_MAX_ORDER);
1313
1314        size = PAGE_ALIGN(size);
1315        if (unlikely(!size))
1316                return NULL;
1317
1318        area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1319        if (unlikely(!area))
1320                return NULL;
1321
1322        /*
1323         * We always allocate a guard page.
1324         */
1325        size += PAGE_SIZE;
1326
1327        va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1328        if (IS_ERR(va)) {
1329                kfree(area);
1330                return NULL;
1331        }
1332
1333        setup_vmalloc_vm(area, va, flags, caller);
1334
1335        return area;
1336}
1337
1338struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1339                                unsigned long start, unsigned long end)
1340{
1341        return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1342                                  GFP_KERNEL, __builtin_return_address(0));
1343}
1344EXPORT_SYMBOL_GPL(__get_vm_area);
1345
1346struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1347                                       unsigned long start, unsigned long end,
1348                                       const void *caller)
1349{
1350        return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1351                                  GFP_KERNEL, caller);
1352}
1353
1354/**
1355 *      get_vm_area  -  reserve a contiguous kernel virtual area
1356 *      @size:          size of the area
1357 *      @flags:         %VM_IOREMAP for I/O mappings or VM_ALLOC
1358 *
1359 *      Search an area of @size in the kernel virtual mapping area,
1360 *      and reserved it for out purposes.  Returns the area descriptor
1361 *      on success or %NULL on failure.
1362 */
1363struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1364{
1365        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1366                                  NUMA_NO_NODE, GFP_KERNEL,
1367                                  __builtin_return_address(0));
1368}
1369
1370struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1371                                const void *caller)
1372{
1373        return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1374                                  NUMA_NO_NODE, GFP_KERNEL, caller);
1375}
1376
1377/**
1378 *      find_vm_area  -  find a continuous kernel virtual area
1379 *      @addr:          base address
1380 *
1381 *      Search for the kernel VM area starting at @addr, and return it.
1382 *      It is up to the caller to do all required locking to keep the returned
1383 *      pointer valid.
1384 */
1385struct vm_struct *find_vm_area(const void *addr)
1386{
1387        struct vmap_area *va;
1388
1389        va = find_vmap_area((unsigned long)addr);
1390        if (va && va->flags & VM_VM_AREA)
1391                return va->vm;
1392
1393        return NULL;
1394}
1395
1396/**
1397 *      remove_vm_area  -  find and remove a continuous kernel virtual area
1398 *      @addr:          base address
1399 *
1400 *      Search for the kernel VM area starting at @addr, and remove it.
1401 *      This function returns the found VM area, but using it is NOT safe
1402 *      on SMP machines, except for its size or flags.
1403 */
1404struct vm_struct *remove_vm_area(const void *addr)
1405{
1406        struct vmap_area *va;
1407
1408        va = find_vmap_area((unsigned long)addr);
1409        if (va && va->flags & VM_VM_AREA) {
1410                struct vm_struct *vm = va->vm;
1411
1412                spin_lock(&vmap_area_lock);
1413                va->vm = NULL;
1414                va->flags &= ~VM_VM_AREA;
1415                spin_unlock(&vmap_area_lock);
1416
1417                vmap_debug_free_range(va->va_start, va->va_end);
1418                free_unmap_vmap_area(va);
1419                vm->size -= PAGE_SIZE;
1420
1421                return vm;
1422        }
1423        return NULL;
1424}
1425
1426static void __vunmap(const void *addr, int deallocate_pages)
1427{
1428        struct vm_struct *area;
1429
1430        if (!addr)
1431                return;
1432
1433        if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1434                        addr))
1435                return;
1436
1437        area = remove_vm_area(addr);
1438        if (unlikely(!area)) {
1439                WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1440                                addr);
1441                return;
1442        }
1443
1444        debug_check_no_locks_freed(addr, area->size);
1445        debug_check_no_obj_freed(addr, area->size);
1446
1447        if (deallocate_pages) {
1448                int i;
1449
1450                for (i = 0; i < area->nr_pages; i++) {
1451                        struct page *page = area->pages[i];
1452
1453                        BUG_ON(!page);
1454                        __free_page(page);
1455                }
1456
1457                if (area->flags & VM_VPAGES)
1458                        vfree(area->pages);
1459                else
1460                        kfree(area->pages);
1461        }
1462
1463        kfree(area);
1464        return;
1465}
1466 
1467/**
1468 *      vfree  -  release memory allocated by vmalloc()
1469 *      @addr:          memory base address
1470 *
1471 *      Free the virtually continuous memory area starting at @addr, as
1472 *      obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1473 *      NULL, no operation is performed.
1474 *
1475 *      Must not be called in NMI context (strictly speaking, only if we don't
1476 *      have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1477 *      conventions for vfree() arch-depenedent would be a really bad idea)
1478 *
1479 *      NOTE: assumes that the object at *addr has a size >= sizeof(llist_node)
1480 */
1481void vfree(const void *addr)
1482{
1483        BUG_ON(in_nmi());
1484
1485        kmemleak_free(addr);
1486
1487        if (!addr)
1488                return;
1489        if (unlikely(in_interrupt())) {
1490                struct vfree_deferred *p = &__get_cpu_var(vfree_deferred);
1491                if (llist_add((struct llist_node *)addr, &p->list))
1492                        schedule_work(&p->wq);
1493        } else
1494                __vunmap(addr, 1);
1495}
1496EXPORT_SYMBOL(vfree);
1497
1498/**
1499 *      vunmap  -  release virtual mapping obtained by vmap()
1500 *      @addr:          memory base address
1501 *
1502 *      Free the virtually contiguous memory area starting at @addr,
1503 *      which was created from the page array passed to vmap().
1504 *
1505 *      Must not be called in interrupt context.
1506 */
1507void vunmap(const void *addr)
1508{
1509        BUG_ON(in_interrupt());
1510        might_sleep();
1511        if (addr)
1512                __vunmap(addr, 0);
1513}
1514EXPORT_SYMBOL(vunmap);
1515
1516/**
1517 *      vmap  -  map an array of pages into virtually contiguous space
1518 *      @pages:         array of page pointers
1519 *      @count:         number of pages to map
1520 *      @flags:         vm_area->flags
1521 *      @prot:          page protection for the mapping
1522 *
1523 *      Maps @count pages from @pages into contiguous kernel virtual
1524 *      space.
1525 */
1526void *vmap(struct page **pages, unsigned int count,
1527                unsigned long flags, pgprot_t prot)
1528{
1529        struct vm_struct *area;
1530
1531        might_sleep();
1532
1533        if (count > totalram_pages)
1534                return NULL;
1535
1536        area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1537                                        __builtin_return_address(0));
1538        if (!area)
1539                return NULL;
1540
1541        if (map_vm_area(area, prot, &pages)) {
1542                vunmap(area->addr);
1543                return NULL;
1544        }
1545
1546        return area->addr;
1547}
1548EXPORT_SYMBOL(vmap);
1549
1550static void *__vmalloc_node(unsigned long size, unsigned long align,
1551                            gfp_t gfp_mask, pgprot_t prot,
1552                            int node, const void *caller);
1553static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1554                                 pgprot_t prot, int node, const void *caller)
1555{
1556        const int order = 0;
1557        struct page **pages;
1558        unsigned int nr_pages, array_size, i;
1559        gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1560
1561        nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1562        array_size = (nr_pages * sizeof(struct page *));
1563
1564        area->nr_pages = nr_pages;
1565        /* Please note that the recursion is strictly bounded. */
1566        if (array_size > PAGE_SIZE) {
1567                pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1568                                PAGE_KERNEL, node, caller);
1569                area->flags |= VM_VPAGES;
1570        } else {
1571                pages = kmalloc_node(array_size, nested_gfp, node);
1572        }
1573        area->pages = pages;
1574        area->caller = caller;
1575        if (!area->pages) {
1576                remove_vm_area(area->addr);
1577                kfree(area);
1578                return NULL;
1579        }
1580
1581        for (i = 0; i < area->nr_pages; i++) {
1582                struct page *page;
1583                gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1584
1585                if (node < 0)
1586                        page = alloc_page(tmp_mask);
1587                else
1588                        page = alloc_pages_node(node, tmp_mask, order);
1589
1590                if (unlikely(!page)) {
1591                        /* Successfully allocated i pages, free them in __vunmap() */
1592                        area->nr_pages = i;
1593                        goto fail;
1594                }
1595                area->pages[i] = page;
1596        }
1597
1598        if (map_vm_area(area, prot, &pages))
1599                goto fail;
1600        return area->addr;
1601
1602fail:
1603        warn_alloc_failed(gfp_mask, order,
1604                          "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1605                          (area->nr_pages*PAGE_SIZE), area->size);
1606        vfree(area->addr);
1607        return NULL;
1608}
1609
1610/**
1611 *      __vmalloc_node_range  -  allocate virtually contiguous memory
1612 *      @size:          allocation size
1613 *      @align:         desired alignment
1614 *      @start:         vm area range start
1615 *      @end:           vm area range end
1616 *      @gfp_mask:      flags for the page level allocator
1617 *      @prot:          protection mask for the allocated pages
1618 *      @node:          node to use for allocation or NUMA_NO_NODE
1619 *      @caller:        caller's return address
1620 *
1621 *      Allocate enough pages to cover @size from the page level
1622 *      allocator with @gfp_mask flags.  Map them into contiguous
1623 *      kernel virtual space, using a pagetable protection of @prot.
1624 */
1625void *__vmalloc_node_range(unsigned long size, unsigned long align,
1626                        unsigned long start, unsigned long end, gfp_t gfp_mask,
1627                        pgprot_t prot, int node, const void *caller)
1628{
1629        struct vm_struct *area;
1630        void *addr;
1631        unsigned long real_size = size;
1632
1633        size = PAGE_ALIGN(size);
1634        if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1635                goto fail;
1636
1637        area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED,
1638                                  start, end, node, gfp_mask, caller);
1639        if (!area)
1640                goto fail;
1641
1642        addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
1643        if (!addr)
1644                goto fail;
1645
1646        /*
1647         * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1648         * flag. It means that vm_struct is not fully initialized.
1649         * Now, it is fully initialized, so remove this flag here.
1650         */
1651        clear_vm_uninitialized_flag(area);
1652
1653        /*
1654         * A ref_count = 3 is needed because the vm_struct and vmap_area
1655         * structures allocated in the __get_vm_area_node() function contain
1656         * references to the virtual address of the vmalloc'ed block.
1657         */
1658        kmemleak_alloc(addr, real_size, 3, gfp_mask);
1659
1660        return addr;
1661
1662fail:
1663        warn_alloc_failed(gfp_mask, 0,
1664                          "vmalloc: allocation failure: %lu bytes\n",
1665                          real_size);
1666        return NULL;
1667}
1668
1669/**
1670 *      __vmalloc_node  -  allocate virtually contiguous memory
1671 *      @size:          allocation size
1672 *      @align:         desired alignment
1673 *      @gfp_mask:      flags for the page level allocator
1674 *      @prot:          protection mask for the allocated pages
1675 *      @node:          node to use for allocation or NUMA_NO_NODE
1676 *      @caller:        caller's return address
1677 *
1678 *      Allocate enough pages to cover @size from the page level
1679 *      allocator with @gfp_mask flags.  Map them into contiguous
1680 *      kernel virtual space, using a pagetable protection of @prot.
1681 */
1682static void *__vmalloc_node(unsigned long size, unsigned long align,
1683                            gfp_t gfp_mask, pgprot_t prot,
1684                            int node, const void *caller)
1685{
1686        return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1687                                gfp_mask, prot, node, caller);
1688}
1689
1690void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1691{
1692        return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1693                                __builtin_return_address(0));
1694}
1695EXPORT_SYMBOL(__vmalloc);
1696
1697static inline void *__vmalloc_node_flags(unsigned long size,
1698                                        int node, gfp_t flags)
1699{
1700        return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1701                                        node, __builtin_return_address(0));
1702}
1703
1704/**
1705 *      vmalloc  -  allocate virtually contiguous memory
1706 *      @size:          allocation size
1707 *      Allocate enough pages to cover @size from the page level
1708 *      allocator and map them into contiguous kernel virtual space.
1709 *
1710 *      For tight control over page level allocator and protection flags
1711 *      use __vmalloc() instead.
1712 */
1713void *vmalloc(unsigned long size)
1714{
1715        return __vmalloc_node_flags(size, NUMA_NO_NODE,
1716                                    GFP_KERNEL | __GFP_HIGHMEM);
1717}
1718EXPORT_SYMBOL(vmalloc);
1719
1720/**
1721 *      vzalloc - allocate virtually contiguous memory with zero fill
1722 *      @size:  allocation size
1723 *      Allocate enough pages to cover @size from the page level
1724 *      allocator and map them into contiguous kernel virtual space.
1725 *      The memory allocated is set to zero.
1726 *
1727 *      For tight control over page level allocator and protection flags
1728 *      use __vmalloc() instead.
1729 */
1730void *vzalloc(unsigned long size)
1731{
1732        return __vmalloc_node_flags(size, NUMA_NO_NODE,
1733                                GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1734}
1735EXPORT_SYMBOL(vzalloc);
1736
1737/**
1738 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1739 * @size: allocation size
1740 *
1741 * The resulting memory area is zeroed so it can be mapped to userspace
1742 * without leaking data.
1743 */
1744void *vmalloc_user(unsigned long size)
1745{
1746        struct vm_struct *area;
1747        void *ret;
1748
1749        ret = __vmalloc_node(size, SHMLBA,
1750                             GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1751                             PAGE_KERNEL, NUMA_NO_NODE,
1752                             __builtin_return_address(0));
1753        if (ret) {
1754                area = find_vm_area(ret);
1755                area->flags |= VM_USERMAP;
1756        }
1757        return ret;
1758}
1759EXPORT_SYMBOL(vmalloc_user);
1760
1761/**
1762 *      vmalloc_node  -  allocate memory on a specific node
1763 *      @size:          allocation size
1764 *      @node:          numa node
1765 *
1766 *      Allocate enough pages to cover @size from the page level
1767 *      allocator and map them into contiguous kernel virtual space.
1768 *
1769 *      For tight control over page level allocator and protection flags
1770 *      use __vmalloc() instead.
1771 */
1772void *vmalloc_node(unsigned long size, int node)
1773{
1774        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1775                                        node, __builtin_return_address(0));
1776}
1777EXPORT_SYMBOL(vmalloc_node);
1778
1779/**
1780 * vzalloc_node - allocate memory on a specific node with zero fill
1781 * @size:       allocation size
1782 * @node:       numa node
1783 *
1784 * Allocate enough pages to cover @size from the page level
1785 * allocator and map them into contiguous kernel virtual space.
1786 * The memory allocated is set to zero.
1787 *
1788 * For tight control over page level allocator and protection flags
1789 * use __vmalloc_node() instead.
1790 */
1791void *vzalloc_node(unsigned long size, int node)
1792{
1793        return __vmalloc_node_flags(size, node,
1794                         GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1795}
1796EXPORT_SYMBOL(vzalloc_node);
1797
1798#ifndef PAGE_KERNEL_EXEC
1799# define PAGE_KERNEL_EXEC PAGE_KERNEL
1800#endif
1801
1802/**
1803 *      vmalloc_exec  -  allocate virtually contiguous, executable memory
1804 *      @size:          allocation size
1805 *
1806 *      Kernel-internal function to allocate enough pages to cover @size
1807 *      the page level allocator and map them into contiguous and
1808 *      executable kernel virtual space.
1809 *
1810 *      For tight control over page level allocator and protection flags
1811 *      use __vmalloc() instead.
1812 */
1813
1814void *vmalloc_exec(unsigned long size)
1815{
1816        return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1817                              NUMA_NO_NODE, __builtin_return_address(0));
1818}
1819
1820#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1821#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1822#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1823#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1824#else
1825#define GFP_VMALLOC32 GFP_KERNEL
1826#endif
1827
1828/**
1829 *      vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1830 *      @size:          allocation size
1831 *
1832 *      Allocate enough 32bit PA addressable pages to cover @size from the
1833 *      page level allocator and map them into contiguous kernel virtual space.
1834 */
1835void *vmalloc_32(unsigned long size)
1836{
1837        return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1838                              NUMA_NO_NODE, __builtin_return_address(0));
1839}
1840EXPORT_SYMBOL(vmalloc_32);
1841
1842/**
1843 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1844 *      @size:          allocation size
1845 *
1846 * The resulting memory area is 32bit addressable and zeroed so it can be
1847 * mapped to userspace without leaking data.
1848 */
1849void *vmalloc_32_user(unsigned long size)
1850{
1851        struct vm_struct *area;
1852        void *ret;
1853
1854        ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1855                             NUMA_NO_NODE, __builtin_return_address(0));
1856        if (ret) {
1857                area = find_vm_area(ret);
1858                area->flags |= VM_USERMAP;
1859        }
1860        return ret;
1861}
1862EXPORT_SYMBOL(vmalloc_32_user);
1863
1864/*
1865 * small helper routine , copy contents to buf from addr.
1866 * If the page is not present, fill zero.
1867 */
1868
1869static int aligned_vread(char *buf, char *addr, unsigned long count)
1870{
1871        struct page *p;
1872        int copied = 0;
1873
1874        while (count) {
1875                unsigned long offset, length;
1876
1877                offset = (unsigned long)addr & ~PAGE_MASK;
1878                length = PAGE_SIZE - offset;
1879                if (length > count)
1880                        length = count;
1881                p = vmalloc_to_page(addr);
1882                /*
1883                 * To do safe access to this _mapped_ area, we need
1884                 * lock. But adding lock here means that we need to add
1885                 * overhead of vmalloc()/vfree() calles for this _debug_
1886                 * interface, rarely used. Instead of that, we'll use
1887                 * kmap() and get small overhead in this access function.
1888                 */
1889                if (p) {
1890                        /*
1891                         * we can expect USER0 is not used (see vread/vwrite's
1892                         * function description)
1893                         */
1894                        void *map = kmap_atomic(p);
1895                        memcpy(buf, map + offset, length);
1896                        kunmap_atomic(map);
1897                } else
1898                        memset(buf, 0, length);
1899
1900                addr += length;
1901                buf += length;
1902                copied += length;
1903                count -= length;
1904        }
1905        return copied;
1906}
1907
1908static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1909{
1910        struct page *p;
1911        int copied = 0;
1912
1913        while (count) {
1914                unsigned long offset, length;
1915
1916                offset = (unsigned long)addr & ~PAGE_MASK;
1917                length = PAGE_SIZE - offset;
1918                if (length > count)
1919                        length = count;
1920                p = vmalloc_to_page(addr);
1921                /*
1922                 * To do safe access to this _mapped_ area, we need
1923                 * lock. But adding lock here means that we need to add
1924                 * overhead of vmalloc()/vfree() calles for this _debug_
1925                 * interface, rarely used. Instead of that, we'll use
1926                 * kmap() and get small overhead in this access function.
1927                 */
1928                if (p) {
1929                        /*
1930                         * we can expect USER0 is not used (see vread/vwrite's
1931                         * function description)
1932                         */
1933                        void *map = kmap_atomic(p);
1934                        memcpy(map + offset, buf, length);
1935                        kunmap_atomic(map);
1936                }
1937                addr += length;
1938                buf += length;
1939                copied += length;
1940                count -= length;
1941        }
1942        return copied;
1943}
1944
1945/**
1946 *      vread() -  read vmalloc area in a safe way.
1947 *      @buf:           buffer for reading data
1948 *      @addr:          vm address.
1949 *      @count:         number of bytes to be read.
1950 *
1951 *      Returns # of bytes which addr and buf should be increased.
1952 *      (same number to @count). Returns 0 if [addr...addr+count) doesn't
1953 *      includes any intersect with alive vmalloc area.
1954 *
1955 *      This function checks that addr is a valid vmalloc'ed area, and
1956 *      copy data from that area to a given buffer. If the given memory range
1957 *      of [addr...addr+count) includes some valid address, data is copied to
1958 *      proper area of @buf. If there are memory holes, they'll be zero-filled.
1959 *      IOREMAP area is treated as memory hole and no copy is done.
1960 *
1961 *      If [addr...addr+count) doesn't includes any intersects with alive
1962 *      vm_struct area, returns 0. @buf should be kernel's buffer.
1963 *
1964 *      Note: In usual ops, vread() is never necessary because the caller
1965 *      should know vmalloc() area is valid and can use memcpy().
1966 *      This is for routines which have to access vmalloc area without
1967 *      any informaion, as /dev/kmem.
1968 *
1969 */
1970
1971long vread(char *buf, char *addr, unsigned long count)
1972{
1973        struct vmap_area *va;
1974        struct vm_struct *vm;
1975        char *vaddr, *buf_start = buf;
1976        unsigned long buflen = count;
1977        unsigned long n;
1978
1979        /* Don't allow overflow */
1980        if ((unsigned long) addr + count < count)
1981                count = -(unsigned long) addr;
1982
1983        spin_lock(&vmap_area_lock);
1984        list_for_each_entry(va, &vmap_area_list, list) {
1985                if (!count)
1986                        break;
1987
1988                if (!(va->flags & VM_VM_AREA))
1989                        continue;
1990
1991                vm = va->vm;
1992                vaddr = (char *) vm->addr;
1993                if (addr >= vaddr + vm->size - PAGE_SIZE)
1994                        continue;
1995                while (addr < vaddr) {
1996                        if (count == 0)
1997                                goto finished;
1998                        *buf = '\0';
1999                        buf++;
2000                        addr++;
2001                        count--;
2002                }
2003                n = vaddr + vm->size - PAGE_SIZE - addr;
2004                if (n > count)
2005                        n = count;
2006                if (!(vm->flags & VM_IOREMAP))
2007                        aligned_vread(buf, addr, n);
2008                else /* IOREMAP area is treated as memory hole */
2009                        memset(buf, 0, n);
2010                buf += n;
2011                addr += n;
2012                count -= n;
2013        }
2014finished:
2015        spin_unlock(&vmap_area_lock);
2016
2017        if (buf == buf_start)
2018                return 0;
2019        /* zero-fill memory holes */
2020        if (buf != buf_start + buflen)
2021                memset(buf, 0, buflen - (buf - buf_start));
2022
2023        return buflen;
2024}
2025
2026/**
2027 *      vwrite() -  write vmalloc area in a safe way.
2028 *      @buf:           buffer for source data
2029 *      @addr:          vm address.
2030 *      @count:         number of bytes to be read.
2031 *
2032 *      Returns # of bytes which addr and buf should be incresed.
2033 *      (same number to @count).
2034 *      If [addr...addr+count) doesn't includes any intersect with valid
2035 *      vmalloc area, returns 0.
2036 *
2037 *      This function checks that addr is a valid vmalloc'ed area, and
2038 *      copy data from a buffer to the given addr. If specified range of
2039 *      [addr...addr+count) includes some valid address, data is copied from
2040 *      proper area of @buf. If there are memory holes, no copy to hole.
2041 *      IOREMAP area is treated as memory hole and no copy is done.
2042 *
2043 *      If [addr...addr+count) doesn't includes any intersects with alive
2044 *      vm_struct area, returns 0. @buf should be kernel's buffer.
2045 *
2046 *      Note: In usual ops, vwrite() is never necessary because the caller
2047 *      should know vmalloc() area is valid and can use memcpy().
2048 *      This is for routines which have to access vmalloc area without
2049 *      any informaion, as /dev/kmem.
2050 */
2051
2052long vwrite(char *buf, char *addr, unsigned long count)
2053{
2054        struct vmap_area *va;
2055        struct vm_struct *vm;
2056        char *vaddr;
2057        unsigned long n, buflen;
2058        int copied = 0;
2059
2060        /* Don't allow overflow */
2061        if ((unsigned long) addr + count < count)
2062                count = -(unsigned long) addr;
2063        buflen = count;
2064
2065        spin_lock(&vmap_area_lock);
2066        list_for_each_entry(va, &vmap_area_list, list) {
2067                if (!count)
2068                        break;
2069
2070                if (!(va->flags & VM_VM_AREA))
2071                        continue;
2072
2073                vm = va->vm;
2074                vaddr = (char *) vm->addr;
2075                if (addr >= vaddr + vm->size - PAGE_SIZE)
2076                        continue;
2077                while (addr < vaddr) {
2078                        if (count == 0)
2079                                goto finished;
2080                        buf++;
2081                        addr++;
2082                        count--;
2083                }
2084                n = vaddr + vm->size - PAGE_SIZE - addr;
2085                if (n > count)
2086                        n = count;
2087                if (!(vm->flags & VM_IOREMAP)) {
2088                        aligned_vwrite(buf, addr, n);
2089                        copied++;
2090                }
2091                buf += n;
2092                addr += n;
2093                count -= n;
2094        }
2095finished:
2096        spin_unlock(&vmap_area_lock);
2097        if (!copied)
2098                return 0;
2099        return buflen;
2100}
2101
2102/**
2103 *      remap_vmalloc_range_partial  -  map vmalloc pages to userspace
2104 *      @vma:           vma to cover
2105 *      @uaddr:         target user address to start at
2106 *      @kaddr:         virtual address of vmalloc kernel memory
2107 *      @size:          size of map area
2108 *
2109 *      Returns:        0 for success, -Exxx on failure
2110 *
2111 *      This function checks that @kaddr is a valid vmalloc'ed area,
2112 *      and that it is big enough to cover the range starting at
2113 *      @uaddr in @vma. Will return failure if that criteria isn't
2114 *      met.
2115 *
2116 *      Similar to remap_pfn_range() (see mm/memory.c)
2117 */
2118int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2119                                void *kaddr, unsigned long size)
2120{
2121        struct vm_struct *area;
2122
2123        size = PAGE_ALIGN(size);
2124
2125        if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2126                return -EINVAL;
2127
2128        area = find_vm_area(kaddr);
2129        if (!area)
2130                return -EINVAL;
2131
2132        if (!(area->flags & VM_USERMAP))
2133                return -EINVAL;
2134
2135        if (kaddr + size > area->addr + area->size)
2136                return -EINVAL;
2137
2138        do {
2139                struct page *page = vmalloc_to_page(kaddr);
2140                int ret;
2141
2142                ret = vm_insert_page(vma, uaddr, page);
2143                if (ret)
2144                        return ret;
2145
2146                uaddr += PAGE_SIZE;
2147                kaddr += PAGE_SIZE;
2148                size -= PAGE_SIZE;
2149        } while (size > 0);
2150
2151        vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2152
2153        return 0;
2154}
2155EXPORT_SYMBOL(remap_vmalloc_range_partial);
2156
2157/**
2158 *      remap_vmalloc_range  -  map vmalloc pages to userspace
2159 *      @vma:           vma to cover (map full range of vma)
2160 *      @addr:          vmalloc memory
2161 *      @pgoff:         number of pages into addr before first page to map
2162 *
2163 *      Returns:        0 for success, -Exxx on failure
2164 *
2165 *      This function checks that addr is a valid vmalloc'ed area, and
2166 *      that it is big enough to cover the vma. Will return failure if
2167 *      that criteria isn't met.
2168 *
2169 *      Similar to remap_pfn_range() (see mm/memory.c)
2170 */
2171int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2172                                                unsigned long pgoff)
2173{
2174        return remap_vmalloc_range_partial(vma, vma->vm_start,
2175                                           addr + (pgoff << PAGE_SHIFT),
2176                                           vma->vm_end - vma->vm_start);
2177}
2178EXPORT_SYMBOL(remap_vmalloc_range);
2179
2180/*
2181 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2182 * have one.
2183 */
2184void  __attribute__((weak)) vmalloc_sync_all(void)
2185{
2186}
2187
2188
2189static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2190{
2191        pte_t ***p = data;
2192
2193        if (p) {
2194                *(*p) = pte;
2195                (*p)++;
2196        }
2197        return 0;
2198}
2199
2200/**
2201 *      alloc_vm_area - allocate a range of kernel address space
2202 *      @size:          size of the area
2203 *      @ptes:          returns the PTEs for the address space
2204 *
2205 *      Returns:        NULL on failure, vm_struct on success
2206 *
2207 *      This function reserves a range of kernel address space, and
2208 *      allocates pagetables to map that range.  No actual mappings
2209 *      are created.
2210 *
2211 *      If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2212 *      allocated for the VM area are returned.
2213 */
2214struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2215{
2216        struct vm_struct *area;
2217
2218        area = get_vm_area_caller(size, VM_IOREMAP,
2219                                __builtin_return_address(0));
2220        if (area == NULL)
2221                return NULL;
2222
2223        /*
2224         * This ensures that page tables are constructed for this region
2225         * of kernel virtual address space and mapped into init_mm.
2226         */
2227        if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2228                                size, f, ptes ? &ptes : NULL)) {
2229                free_vm_area(area);
2230                return NULL;
2231        }
2232
2233        return area;
2234}
2235EXPORT_SYMBOL_GPL(alloc_vm_area);
2236
2237void free_vm_area(struct vm_struct *area)
2238{
2239        struct vm_struct *ret;
2240        ret = remove_vm_area(area->addr);
2241        BUG_ON(ret != area);
2242        kfree(area);
2243}
2244EXPORT_SYMBOL_GPL(free_vm_area);
2245
2246#ifdef CONFIG_SMP
2247static struct vmap_area *node_to_va(struct rb_node *n)
2248{
2249        return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2250}
2251
2252/**
2253 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2254 * @end: target address
2255 * @pnext: out arg for the next vmap_area
2256 * @pprev: out arg for the previous vmap_area
2257 *
2258 * Returns: %true if either or both of next and prev are found,
2259 *          %false if no vmap_area exists
2260 *
2261 * Find vmap_areas end addresses of which enclose @end.  ie. if not
2262 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2263 */
2264static bool pvm_find_next_prev(unsigned long end,
2265                               struct vmap_area **pnext,
2266                               struct vmap_area **pprev)
2267{
2268        struct rb_node *n = vmap_area_root.rb_node;
2269        struct vmap_area *va = NULL;
2270
2271        while (n) {
2272                va = rb_entry(n, struct vmap_area, rb_node);
2273                if (end < va->va_end)
2274                        n = n->rb_left;
2275                else if (end > va->va_end)
2276                        n = n->rb_right;
2277                else
2278                        break;
2279        }
2280
2281        if (!va)
2282                return false;
2283
2284        if (va->va_end > end) {
2285                *pnext = va;
2286                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2287        } else {
2288                *pprev = va;
2289                *pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2290        }
2291        return true;
2292}
2293
2294/**
2295 * pvm_determine_end - find the highest aligned address between two vmap_areas
2296 * @pnext: in/out arg for the next vmap_area
2297 * @pprev: in/out arg for the previous vmap_area
2298 * @align: alignment
2299 *
2300 * Returns: determined end address
2301 *
2302 * Find the highest aligned address between *@pnext and *@pprev below
2303 * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2304 * down address is between the end addresses of the two vmap_areas.
2305 *
2306 * Please note that the address returned by this function may fall
2307 * inside *@pnext vmap_area.  The caller is responsible for checking
2308 * that.
2309 */
2310static unsigned long pvm_determine_end(struct vmap_area **pnext,
2311                                       struct vmap_area **pprev,
2312                                       unsigned long align)
2313{
2314        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2315        unsigned long addr;
2316
2317        if (*pnext)
2318                addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2319        else
2320                addr = vmalloc_end;
2321
2322        while (*pprev && (*pprev)->va_end > addr) {
2323                *pnext = *pprev;
2324                *pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2325        }
2326
2327        return addr;
2328}
2329
2330/**
2331 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2332 * @offsets: array containing offset of each area
2333 * @sizes: array containing size of each area
2334 * @nr_vms: the number of areas to allocate
2335 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2336 *
2337 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2338 *          vm_structs on success, %NULL on failure
2339 *
2340 * Percpu allocator wants to use congruent vm areas so that it can
2341 * maintain the offsets among percpu areas.  This function allocates
2342 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
2343 * be scattered pretty far, distance between two areas easily going up
2344 * to gigabytes.  To avoid interacting with regular vmallocs, these
2345 * areas are allocated from top.
2346 *
2347 * Despite its complicated look, this allocator is rather simple.  It
2348 * does everything top-down and scans areas from the end looking for
2349 * matching slot.  While scanning, if any of the areas overlaps with
2350 * existing vmap_area, the base address is pulled down to fit the
2351 * area.  Scanning is repeated till all the areas fit and then all
2352 * necessary data structres are inserted and the result is returned.
2353 */
2354struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2355                                     const size_t *sizes, int nr_vms,
2356                                     size_t align)
2357{
2358        const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2359        const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2360        struct vmap_area **vas, *prev, *next;
2361        struct vm_struct **vms;
2362        int area, area2, last_area, term_area;
2363        unsigned long base, start, end, last_end;
2364        bool purged = false;
2365
2366        /* verify parameters and allocate data structures */
2367        BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2368        for (last_area = 0, area = 0; area < nr_vms; area++) {
2369                start = offsets[area];
2370                end = start + sizes[area];
2371
2372                /* is everything aligned properly? */
2373                BUG_ON(!IS_ALIGNED(offsets[area], align));
2374                BUG_ON(!IS_ALIGNED(sizes[area], align));
2375
2376                /* detect the area with the highest address */
2377                if (start > offsets[last_area])
2378                        last_area = area;
2379
2380                for (area2 = 0; area2 < nr_vms; area2++) {
2381                        unsigned long start2 = offsets[area2];
2382                        unsigned long end2 = start2 + sizes[area2];
2383
2384                        if (area2 == area)
2385                                continue;
2386
2387                        BUG_ON(start2 >= start && start2 < end);
2388                        BUG_ON(end2 <= end && end2 > start);
2389                }
2390        }
2391        last_end = offsets[last_area] + sizes[last_area];
2392
2393        if (vmalloc_end - vmalloc_start < last_end) {
2394                WARN_ON(true);
2395                return NULL;
2396        }
2397
2398        vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2399        vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2400        if (!vas || !vms)
2401                goto err_free2;
2402
2403        for (area = 0; area < nr_vms; area++) {
2404                vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2405                vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2406                if (!vas[area] || !vms[area])
2407                        goto err_free;
2408        }
2409retry:
2410        spin_lock(&vmap_area_lock);
2411
2412        /* start scanning - we scan from the top, begin with the last area */
2413        area = term_area = last_area;
2414        start = offsets[area];
2415        end = start + sizes[area];
2416
2417        if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2418                base = vmalloc_end - last_end;
2419                goto found;
2420        }
2421        base = pvm_determine_end(&next, &prev, align) - end;
2422
2423        while (true) {
2424                BUG_ON(next && next->va_end <= base + end);
2425                BUG_ON(prev && prev->va_end > base + end);
2426
2427                /*
2428                 * base might have underflowed, add last_end before
2429                 * comparing.
2430                 */
2431                if (base + last_end < vmalloc_start + last_end) {
2432                        spin_unlock(&vmap_area_lock);
2433                        if (!purged) {
2434                                purge_vmap_area_lazy();
2435                                purged = true;
2436                                goto retry;
2437                        }
2438                        goto err_free;
2439                }
2440
2441                /*
2442                 * If next overlaps, move base downwards so that it's
2443                 * right below next and then recheck.
2444                 */
2445                if (next && next->va_start < base + end) {
2446                        base = pvm_determine_end(&next, &prev, align) - end;
2447                        term_area = area;
2448                        continue;
2449                }
2450
2451                /*
2452                 * If prev overlaps, shift down next and prev and move
2453                 * base so that it's right below new next and then
2454                 * recheck.
2455                 */
2456                if (prev && prev->va_end > base + start)  {
2457                        next = prev;
2458                        prev = node_to_va(rb_prev(&next->rb_node));
2459                        base = pvm_determine_end(&next, &prev, align) - end;
2460                        term_area = area;
2461                        continue;
2462                }
2463
2464                /*
2465                 * This area fits, move on to the previous one.  If
2466                 * the previous one is the terminal one, we're done.
2467                 */
2468                area = (area + nr_vms - 1) % nr_vms;
2469                if (area == term_area)
2470                        break;
2471                start = offsets[area];
2472                end = start + sizes[area];
2473                pvm_find_next_prev(base + end, &next, &prev);
2474        }
2475found:
2476        /* we've found a fitting base, insert all va's */
2477        for (area = 0; area < nr_vms; area++) {
2478                struct vmap_area *va = vas[area];
2479
2480                va->va_start = base + offsets[area];
2481                va->va_end = va->va_start + sizes[area];
2482                __insert_vmap_area(va);
2483        }
2484
2485        vmap_area_pcpu_hole = base + offsets[last_area];
2486
2487        spin_unlock(&vmap_area_lock);
2488
2489        /* insert all vm's */
2490        for (area = 0; area < nr_vms; area++)
2491                setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2492                                 pcpu_get_vm_areas);
2493
2494        kfree(vas);
2495        return vms;
2496
2497err_free:
2498        for (area = 0; area < nr_vms; area++) {
2499                kfree(vas[area]);
2500                kfree(vms[area]);
2501        }
2502err_free2:
2503        kfree(vas);
2504        kfree(vms);
2505        return NULL;
2506}
2507
2508/**
2509 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2510 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2511 * @nr_vms: the number of allocated areas
2512 *
2513 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2514 */
2515void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2516{
2517        int i;
2518
2519        for (i = 0; i < nr_vms; i++)
2520                free_vm_area(vms[i]);
2521        kfree(vms);
2522}
2523#endif  /* CONFIG_SMP */
2524
2525#ifdef CONFIG_PROC_FS
2526static void *s_start(struct seq_file *m, loff_t *pos)
2527        __acquires(&vmap_area_lock)
2528{
2529        loff_t n = *pos;
2530        struct vmap_area *va;
2531
2532        spin_lock(&vmap_area_lock);
2533        va = list_entry((&vmap_area_list)->next, typeof(*va), list);
2534        while (n > 0 && &va->list != &vmap_area_list) {
2535                n--;
2536                va = list_entry(va->list.next, typeof(*va), list);
2537        }
2538        if (!n && &va->list != &vmap_area_list)
2539                return va;
2540
2541        return NULL;
2542
2543}
2544
2545static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2546{
2547        struct vmap_area *va = p, *next;
2548
2549        ++*pos;
2550        next = list_entry(va->list.next, typeof(*va), list);
2551        if (&next->list != &vmap_area_list)
2552                return next;
2553
2554        return NULL;
2555}
2556
2557static void s_stop(struct seq_file *m, void *p)
2558        __releases(&vmap_area_lock)
2559{
2560        spin_unlock(&vmap_area_lock);
2561}
2562
2563static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2564{
2565        if (IS_ENABLED(CONFIG_NUMA)) {
2566                unsigned int nr, *counters = m->private;
2567
2568                if (!counters)
2569                        return;
2570
2571                memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2572
2573                for (nr = 0; nr < v->nr_pages; nr++)
2574                        counters[page_to_nid(v->pages[nr])]++;
2575
2576                for_each_node_state(nr, N_HIGH_MEMORY)
2577                        if (counters[nr])
2578                                seq_printf(m, " N%u=%u", nr, counters[nr]);
2579        }
2580}
2581
2582static int s_show(struct seq_file *m, void *p)
2583{
2584        struct vmap_area *va = p;
2585        struct vm_struct *v;
2586
2587        if (va->flags & (VM_LAZY_FREE | VM_LAZY_FREEING))
2588                return 0;
2589
2590        if (!(va->flags & VM_VM_AREA)) {
2591                seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
2592                        (void *)va->va_start, (void *)va->va_end,
2593                                        va->va_end - va->va_start);
2594                return 0;
2595        }
2596
2597        v = va->vm;
2598
2599        /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2600        smp_rmb();
2601        if (v->flags & VM_UNINITIALIZED)
2602                return 0;
2603
2604        seq_printf(m, "0x%pK-0x%pK %7ld",
2605                v->addr, v->addr + v->size, v->size);
2606
2607        if (v->caller)
2608                seq_printf(m, " %pS", v->caller);
2609
2610        if (v->nr_pages)
2611                seq_printf(m, " pages=%d", v->nr_pages);
2612
2613        if (v->phys_addr)
2614                seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2615
2616        if (v->flags & VM_IOREMAP)
2617                seq_printf(m, " ioremap");
2618
2619        if (v->flags & VM_ALLOC)
2620                seq_printf(m, " vmalloc");
2621
2622        if (v->flags & VM_MAP)
2623                seq_printf(m, " vmap");
2624
2625        if (v->flags & VM_USERMAP)
2626                seq_printf(m, " user");
2627
2628        if (v->flags & VM_VPAGES)
2629                seq_printf(m, " vpages");
2630
2631        show_numa_info(m, v);
2632        seq_putc(m, '\n');
2633        return 0;
2634}
2635
2636static const struct seq_operations vmalloc_op = {
2637        .start = s_start,
2638        .next = s_next,
2639        .stop = s_stop,
2640        .show = s_show,
2641};
2642
2643static int vmalloc_open(struct inode *inode, struct file *file)
2644{
2645        unsigned int *ptr = NULL;
2646        int ret;
2647
2648        if (IS_ENABLED(CONFIG_NUMA)) {
2649                ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2650                if (ptr == NULL)
2651                        return -ENOMEM;
2652        }
2653        ret = seq_open(file, &vmalloc_op);
2654        if (!ret) {
2655                struct seq_file *m = file->private_data;
2656                m->private = ptr;
2657        } else
2658                kfree(ptr);
2659        return ret;
2660}
2661
2662static const struct file_operations proc_vmalloc_operations = {
2663        .open           = vmalloc_open,
2664        .read           = seq_read,
2665        .llseek         = seq_lseek,
2666        .release        = seq_release_private,
2667};
2668
2669static int __init proc_vmalloc_init(void)
2670{
2671        proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2672        return 0;
2673}
2674module_init(proc_vmalloc_init);
2675
2676void get_vmalloc_info(struct vmalloc_info *vmi)
2677{
2678        struct vmap_area *va;
2679        unsigned long free_area_size;
2680        unsigned long prev_end;
2681
2682        vmi->used = 0;
2683        vmi->largest_chunk = 0;
2684
2685        prev_end = VMALLOC_START;
2686
2687        spin_lock(&vmap_area_lock);
2688
2689        if (list_empty(&vmap_area_list)) {
2690                vmi->largest_chunk = VMALLOC_TOTAL;
2691                goto out;
2692        }
2693
2694        list_for_each_entry(va, &vmap_area_list, list) {
2695                unsigned long addr = va->va_start;
2696
2697                /*
2698                 * Some archs keep another range for modules in vmalloc space
2699                 */
2700                if (addr < VMALLOC_START)
2701                        continue;
2702                if (addr >= VMALLOC_END)
2703                        break;
2704
2705                if (va->flags & (VM_LAZY_FREE | VM_LAZY_FREEING))
2706                        continue;
2707
2708                vmi->used += (va->va_end - va->va_start);
2709
2710                free_area_size = addr - prev_end;
2711                if (vmi->largest_chunk < free_area_size)
2712                        vmi->largest_chunk = free_area_size;
2713
2714                prev_end = va->va_end;
2715        }
2716
2717        if (VMALLOC_END - prev_end > vmi->largest_chunk)
2718                vmi->largest_chunk = VMALLOC_END - prev_end;
2719
2720out:
2721        spin_unlock(&vmap_area_lock);
2722}
2723#endif
2724
2725
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.