Linux Audio

Check our new training course

Loading...
v3.15
   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/compiler.h>
  31#include <linux/llist.h>
 
  32
  33#include <asm/uaccess.h>
  34#include <asm/tlbflush.h>
  35#include <asm/shmparam.h>
  36
 
 
  37struct vfree_deferred {
  38	struct llist_head list;
  39	struct work_struct wq;
  40};
  41static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
  42
  43static void __vunmap(const void *, int);
  44
  45static void free_work(struct work_struct *w)
  46{
  47	struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
  48	struct llist_node *llnode = llist_del_all(&p->list);
  49	while (llnode) {
  50		void *p = llnode;
  51		llnode = llist_next(llnode);
  52		__vunmap(p, 1);
  53	}
  54}
  55
  56/*** Page table manipulation functions ***/
  57
  58static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
  59{
  60	pte_t *pte;
  61
  62	pte = pte_offset_kernel(pmd, addr);
  63	do {
  64		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
  65		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
  66	} while (pte++, addr += PAGE_SIZE, addr != end);
  67}
  68
  69static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
  70{
  71	pmd_t *pmd;
  72	unsigned long next;
  73
  74	pmd = pmd_offset(pud, addr);
  75	do {
  76		next = pmd_addr_end(addr, end);
 
 
  77		if (pmd_none_or_clear_bad(pmd))
  78			continue;
  79		vunmap_pte_range(pmd, addr, next);
  80	} while (pmd++, addr = next, addr != end);
  81}
  82
  83static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
  84{
  85	pud_t *pud;
  86	unsigned long next;
  87
  88	pud = pud_offset(pgd, addr);
  89	do {
  90		next = pud_addr_end(addr, end);
 
 
  91		if (pud_none_or_clear_bad(pud))
  92			continue;
  93		vunmap_pmd_range(pud, addr, next);
  94	} while (pud++, addr = next, addr != end);
  95}
  96
  97static void vunmap_page_range(unsigned long addr, unsigned long end)
  98{
  99	pgd_t *pgd;
 100	unsigned long next;
 101
 102	BUG_ON(addr >= end);
 103	pgd = pgd_offset_k(addr);
 104	do {
 105		next = pgd_addr_end(addr, end);
 106		if (pgd_none_or_clear_bad(pgd))
 107			continue;
 108		vunmap_pud_range(pgd, addr, next);
 109	} while (pgd++, addr = next, addr != end);
 110}
 111
 112static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
 113		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 114{
 115	pte_t *pte;
 116
 117	/*
 118	 * nr is a running index into the array which helps higher level
 119	 * callers keep track of where we're up to.
 120	 */
 121
 122	pte = pte_alloc_kernel(pmd, addr);
 123	if (!pte)
 124		return -ENOMEM;
 125	do {
 126		struct page *page = pages[*nr];
 127
 128		if (WARN_ON(!pte_none(*pte)))
 129			return -EBUSY;
 130		if (WARN_ON(!page))
 131			return -ENOMEM;
 132		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 133		(*nr)++;
 134	} while (pte++, addr += PAGE_SIZE, addr != end);
 135	return 0;
 136}
 137
 138static int vmap_pmd_range(pud_t *pud, unsigned long addr,
 139		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 140{
 141	pmd_t *pmd;
 142	unsigned long next;
 143
 144	pmd = pmd_alloc(&init_mm, pud, addr);
 145	if (!pmd)
 146		return -ENOMEM;
 147	do {
 148		next = pmd_addr_end(addr, end);
 149		if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
 150			return -ENOMEM;
 151	} while (pmd++, addr = next, addr != end);
 152	return 0;
 153}
 154
 155static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
 156		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 157{
 158	pud_t *pud;
 159	unsigned long next;
 160
 161	pud = pud_alloc(&init_mm, pgd, addr);
 162	if (!pud)
 163		return -ENOMEM;
 164	do {
 165		next = pud_addr_end(addr, end);
 166		if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
 167			return -ENOMEM;
 168	} while (pud++, addr = next, addr != end);
 169	return 0;
 170}
 171
 172/*
 173 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
 174 * will have pfns corresponding to the "pages" array.
 175 *
 176 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
 177 */
 178static int vmap_page_range_noflush(unsigned long start, unsigned long end,
 179				   pgprot_t prot, struct page **pages)
 180{
 181	pgd_t *pgd;
 182	unsigned long next;
 183	unsigned long addr = start;
 184	int err = 0;
 185	int nr = 0;
 186
 187	BUG_ON(addr >= end);
 188	pgd = pgd_offset_k(addr);
 189	do {
 190		next = pgd_addr_end(addr, end);
 191		err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
 192		if (err)
 193			return err;
 194	} while (pgd++, addr = next, addr != end);
 195
 196	return nr;
 197}
 198
 199static int vmap_page_range(unsigned long start, unsigned long end,
 200			   pgprot_t prot, struct page **pages)
 201{
 202	int ret;
 203
 204	ret = vmap_page_range_noflush(start, end, prot, pages);
 205	flush_cache_vmap(start, end);
 206	return ret;
 207}
 208
 209int is_vmalloc_or_module_addr(const void *x)
 210{
 211	/*
 212	 * ARM, x86-64 and sparc64 put modules in a special place,
 213	 * and fall back on vmalloc() if that fails. Others
 214	 * just put it in the vmalloc space.
 215	 */
 216#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 217	unsigned long addr = (unsigned long)x;
 218	if (addr >= MODULES_VADDR && addr < MODULES_END)
 219		return 1;
 220#endif
 221	return is_vmalloc_addr(x);
 222}
 223
 224/*
 225 * Walk a vmap address to the struct page it maps.
 226 */
 227struct page *vmalloc_to_page(const void *vmalloc_addr)
 228{
 229	unsigned long addr = (unsigned long) vmalloc_addr;
 230	struct page *page = NULL;
 231	pgd_t *pgd = pgd_offset_k(addr);
 232
 233	/*
 234	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 235	 * architectures that do not vmalloc module space
 236	 */
 237	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 238
 239	if (!pgd_none(*pgd)) {
 240		pud_t *pud = pud_offset(pgd, addr);
 241		if (!pud_none(*pud)) {
 242			pmd_t *pmd = pmd_offset(pud, addr);
 243			if (!pmd_none(*pmd)) {
 244				pte_t *ptep, pte;
 245
 246				ptep = pte_offset_map(pmd, addr);
 247				pte = *ptep;
 248				if (pte_present(pte))
 249					page = pte_page(pte);
 250				pte_unmap(ptep);
 251			}
 252		}
 253	}
 254	return page;
 255}
 256EXPORT_SYMBOL(vmalloc_to_page);
 257
 258/*
 259 * Map a vmalloc()-space virtual address to the physical page frame number.
 260 */
 261unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 262{
 263	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 264}
 265EXPORT_SYMBOL(vmalloc_to_pfn);
 266
 267
 268/*** Global kva allocator ***/
 269
 270#define VM_LAZY_FREE	0x01
 271#define VM_LAZY_FREEING	0x02
 272#define VM_VM_AREA	0x04
 273
 274static DEFINE_SPINLOCK(vmap_area_lock);
 275/* Export for kexec only */
 276LIST_HEAD(vmap_area_list);
 277static struct rb_root vmap_area_root = RB_ROOT;
 278
 279/* The vmap cache globals are protected by vmap_area_lock */
 280static struct rb_node *free_vmap_cache;
 281static unsigned long cached_hole_size;
 282static unsigned long cached_vstart;
 283static unsigned long cached_align;
 284
 285static unsigned long vmap_area_pcpu_hole;
 286
 287static struct vmap_area *__find_vmap_area(unsigned long addr)
 288{
 289	struct rb_node *n = vmap_area_root.rb_node;
 290
 291	while (n) {
 292		struct vmap_area *va;
 293
 294		va = rb_entry(n, struct vmap_area, rb_node);
 295		if (addr < va->va_start)
 296			n = n->rb_left;
 297		else if (addr >= va->va_end)
 298			n = n->rb_right;
 299		else
 300			return va;
 301	}
 302
 303	return NULL;
 304}
 305
 306static void __insert_vmap_area(struct vmap_area *va)
 307{
 308	struct rb_node **p = &vmap_area_root.rb_node;
 309	struct rb_node *parent = NULL;
 310	struct rb_node *tmp;
 311
 312	while (*p) {
 313		struct vmap_area *tmp_va;
 314
 315		parent = *p;
 316		tmp_va = rb_entry(parent, struct vmap_area, rb_node);
 317		if (va->va_start < tmp_va->va_end)
 318			p = &(*p)->rb_left;
 319		else if (va->va_end > tmp_va->va_start)
 320			p = &(*p)->rb_right;
 321		else
 322			BUG();
 323	}
 324
 325	rb_link_node(&va->rb_node, parent, p);
 326	rb_insert_color(&va->rb_node, &vmap_area_root);
 327
 328	/* address-sort this list */
 329	tmp = rb_prev(&va->rb_node);
 330	if (tmp) {
 331		struct vmap_area *prev;
 332		prev = rb_entry(tmp, struct vmap_area, rb_node);
 333		list_add_rcu(&va->list, &prev->list);
 334	} else
 335		list_add_rcu(&va->list, &vmap_area_list);
 336}
 337
 338static void purge_vmap_area_lazy(void);
 339
 340/*
 341 * Allocate a region of KVA of the specified size and alignment, within the
 342 * vstart and vend.
 343 */
 344static struct vmap_area *alloc_vmap_area(unsigned long size,
 345				unsigned long align,
 346				unsigned long vstart, unsigned long vend,
 347				int node, gfp_t gfp_mask)
 348{
 349	struct vmap_area *va;
 350	struct rb_node *n;
 351	unsigned long addr;
 352	int purged = 0;
 353	struct vmap_area *first;
 354
 355	BUG_ON(!size);
 356	BUG_ON(size & ~PAGE_MASK);
 357	BUG_ON(!is_power_of_2(align));
 358
 359	va = kmalloc_node(sizeof(struct vmap_area),
 360			gfp_mask & GFP_RECLAIM_MASK, node);
 361	if (unlikely(!va))
 362		return ERR_PTR(-ENOMEM);
 363
 364	/*
 365	 * Only scan the relevant parts containing pointers to other objects
 366	 * to avoid false negatives.
 367	 */
 368	kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
 369
 370retry:
 371	spin_lock(&vmap_area_lock);
 372	/*
 373	 * Invalidate cache if we have more permissive parameters.
 374	 * cached_hole_size notes the largest hole noticed _below_
 375	 * the vmap_area cached in free_vmap_cache: if size fits
 376	 * into that hole, we want to scan from vstart to reuse
 377	 * the hole instead of allocating above free_vmap_cache.
 378	 * Note that __free_vmap_area may update free_vmap_cache
 379	 * without updating cached_hole_size or cached_align.
 380	 */
 381	if (!free_vmap_cache ||
 382			size < cached_hole_size ||
 383			vstart < cached_vstart ||
 384			align < cached_align) {
 385nocache:
 386		cached_hole_size = 0;
 387		free_vmap_cache = NULL;
 388	}
 389	/* record if we encounter less permissive parameters */
 390	cached_vstart = vstart;
 391	cached_align = align;
 392
 393	/* find starting point for our search */
 394	if (free_vmap_cache) {
 395		first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 396		addr = ALIGN(first->va_end, align);
 397		if (addr < vstart)
 398			goto nocache;
 399		if (addr + size < addr)
 400			goto overflow;
 401
 402	} else {
 403		addr = ALIGN(vstart, align);
 404		if (addr + size < addr)
 405			goto overflow;
 406
 407		n = vmap_area_root.rb_node;
 408		first = NULL;
 409
 410		while (n) {
 411			struct vmap_area *tmp;
 412			tmp = rb_entry(n, struct vmap_area, rb_node);
 413			if (tmp->va_end >= addr) {
 414				first = tmp;
 415				if (tmp->va_start <= addr)
 416					break;
 417				n = n->rb_left;
 418			} else
 419				n = n->rb_right;
 420		}
 421
 422		if (!first)
 423			goto found;
 424	}
 425
 426	/* from the starting point, walk areas until a suitable hole is found */
 427	while (addr + size > first->va_start && addr + size <= vend) {
 428		if (addr + cached_hole_size < first->va_start)
 429			cached_hole_size = first->va_start - addr;
 430		addr = ALIGN(first->va_end, align);
 431		if (addr + size < addr)
 432			goto overflow;
 433
 434		if (list_is_last(&first->list, &vmap_area_list))
 435			goto found;
 436
 437		first = list_entry(first->list.next,
 438				struct vmap_area, list);
 439	}
 440
 441found:
 442	if (addr + size > vend)
 443		goto overflow;
 444
 445	va->va_start = addr;
 446	va->va_end = addr + size;
 447	va->flags = 0;
 448	__insert_vmap_area(va);
 449	free_vmap_cache = &va->rb_node;
 450	spin_unlock(&vmap_area_lock);
 451
 452	BUG_ON(va->va_start & (align-1));
 453	BUG_ON(va->va_start < vstart);
 454	BUG_ON(va->va_end > vend);
 455
 456	return va;
 457
 458overflow:
 459	spin_unlock(&vmap_area_lock);
 460	if (!purged) {
 461		purge_vmap_area_lazy();
 462		purged = 1;
 463		goto retry;
 464	}
 465	if (printk_ratelimit())
 466		printk(KERN_WARNING
 467			"vmap allocation for size %lu failed: "
 468			"use vmalloc=<size> to increase size.\n", size);
 469	kfree(va);
 470	return ERR_PTR(-EBUSY);
 471}
 472
 473static void __free_vmap_area(struct vmap_area *va)
 474{
 475	BUG_ON(RB_EMPTY_NODE(&va->rb_node));
 476
 477	if (free_vmap_cache) {
 478		if (va->va_end < cached_vstart) {
 479			free_vmap_cache = NULL;
 480		} else {
 481			struct vmap_area *cache;
 482			cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 483			if (va->va_start <= cache->va_start) {
 484				free_vmap_cache = rb_prev(&va->rb_node);
 485				/*
 486				 * We don't try to update cached_hole_size or
 487				 * cached_align, but it won't go very wrong.
 488				 */
 489			}
 490		}
 491	}
 492	rb_erase(&va->rb_node, &vmap_area_root);
 493	RB_CLEAR_NODE(&va->rb_node);
 494	list_del_rcu(&va->list);
 495
 496	/*
 497	 * Track the highest possible candidate for pcpu area
 498	 * allocation.  Areas outside of vmalloc area can be returned
 499	 * here too, consider only end addresses which fall inside
 500	 * vmalloc area proper.
 501	 */
 502	if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
 503		vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
 504
 505	kfree_rcu(va, rcu_head);
 506}
 507
 508/*
 509 * Free a region of KVA allocated by alloc_vmap_area
 510 */
 511static void free_vmap_area(struct vmap_area *va)
 512{
 513	spin_lock(&vmap_area_lock);
 514	__free_vmap_area(va);
 515	spin_unlock(&vmap_area_lock);
 516}
 517
 518/*
 519 * Clear the pagetable entries of a given vmap_area
 520 */
 521static void unmap_vmap_area(struct vmap_area *va)
 522{
 523	vunmap_page_range(va->va_start, va->va_end);
 524}
 525
 526static void vmap_debug_free_range(unsigned long start, unsigned long end)
 527{
 528	/*
 529	 * Unmap page tables and force a TLB flush immediately if
 530	 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
 531	 * bugs similarly to those in linear kernel virtual address
 532	 * space after a page has been freed.
 533	 *
 534	 * All the lazy freeing logic is still retained, in order to
 535	 * minimise intrusiveness of this debugging feature.
 536	 *
 537	 * This is going to be *slow* (linear kernel virtual address
 538	 * debugging doesn't do a broadcast TLB flush so it is a lot
 539	 * faster).
 540	 */
 541#ifdef CONFIG_DEBUG_PAGEALLOC
 542	vunmap_page_range(start, end);
 543	flush_tlb_kernel_range(start, end);
 544#endif
 545}
 546
 547/*
 548 * lazy_max_pages is the maximum amount of virtual address space we gather up
 549 * before attempting to purge with a TLB flush.
 550 *
 551 * There is a tradeoff here: a larger number will cover more kernel page tables
 552 * and take slightly longer to purge, but it will linearly reduce the number of
 553 * global TLB flushes that must be performed. It would seem natural to scale
 554 * this number up linearly with the number of CPUs (because vmapping activity
 555 * could also scale linearly with the number of CPUs), however it is likely
 556 * that in practice, workloads might be constrained in other ways that mean
 557 * vmap activity will not scale linearly with CPUs. Also, I want to be
 558 * conservative and not introduce a big latency on huge systems, so go with
 559 * a less aggressive log scale. It will still be an improvement over the old
 560 * code, and it will be simple to change the scale factor if we find that it
 561 * becomes a problem on bigger systems.
 562 */
 563static unsigned long lazy_max_pages(void)
 564{
 565	unsigned int log;
 566
 567	log = fls(num_online_cpus());
 568
 569	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
 570}
 571
 572static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
 573
 574/* for per-CPU blocks */
 575static void purge_fragmented_blocks_allcpus(void);
 576
 577/*
 578 * called before a call to iounmap() if the caller wants vm_area_struct's
 579 * immediately freed.
 580 */
 581void set_iounmap_nonlazy(void)
 582{
 583	atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
 584}
 585
 586/*
 587 * Purges all lazily-freed vmap areas.
 588 *
 589 * If sync is 0 then don't purge if there is already a purge in progress.
 590 * If force_flush is 1, then flush kernel TLBs between *start and *end even
 591 * if we found no lazy vmap areas to unmap (callers can use this to optimise
 592 * their own TLB flushing).
 593 * Returns with *start = min(*start, lowest purged address)
 594 *              *end = max(*end, highest purged address)
 595 */
 596static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
 597					int sync, int force_flush)
 598{
 599	static DEFINE_SPINLOCK(purge_lock);
 600	LIST_HEAD(valist);
 601	struct vmap_area *va;
 602	struct vmap_area *n_va;
 603	int nr = 0;
 604
 605	/*
 606	 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
 607	 * should not expect such behaviour. This just simplifies locking for
 608	 * the case that isn't actually used at the moment anyway.
 609	 */
 610	if (!sync && !force_flush) {
 611		if (!spin_trylock(&purge_lock))
 612			return;
 613	} else
 614		spin_lock(&purge_lock);
 615
 616	if (sync)
 617		purge_fragmented_blocks_allcpus();
 618
 619	rcu_read_lock();
 620	list_for_each_entry_rcu(va, &vmap_area_list, list) {
 621		if (va->flags & VM_LAZY_FREE) {
 622			if (va->va_start < *start)
 623				*start = va->va_start;
 624			if (va->va_end > *end)
 625				*end = va->va_end;
 626			nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
 627			list_add_tail(&va->purge_list, &valist);
 628			va->flags |= VM_LAZY_FREEING;
 629			va->flags &= ~VM_LAZY_FREE;
 630		}
 631	}
 632	rcu_read_unlock();
 633
 634	if (nr)
 635		atomic_sub(nr, &vmap_lazy_nr);
 636
 637	if (nr || force_flush)
 638		flush_tlb_kernel_range(*start, *end);
 639
 640	if (nr) {
 641		spin_lock(&vmap_area_lock);
 642		list_for_each_entry_safe(va, n_va, &valist, purge_list)
 643			__free_vmap_area(va);
 644		spin_unlock(&vmap_area_lock);
 645	}
 646	spin_unlock(&purge_lock);
 647}
 648
 649/*
 650 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
 651 * is already purging.
 652 */
 653static void try_purge_vmap_area_lazy(void)
 654{
 655	unsigned long start = ULONG_MAX, end = 0;
 656
 657	__purge_vmap_area_lazy(&start, &end, 0, 0);
 658}
 659
 660/*
 661 * Kick off a purge of the outstanding lazy areas.
 662 */
 663static void purge_vmap_area_lazy(void)
 664{
 665	unsigned long start = ULONG_MAX, end = 0;
 666
 667	__purge_vmap_area_lazy(&start, &end, 1, 0);
 668}
 669
 670/*
 671 * Free a vmap area, caller ensuring that the area has been unmapped
 672 * and flush_cache_vunmap had been called for the correct range
 673 * previously.
 674 */
 675static void free_vmap_area_noflush(struct vmap_area *va)
 676{
 677	va->flags |= VM_LAZY_FREE;
 678	atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
 679	if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
 680		try_purge_vmap_area_lazy();
 681}
 682
 683/*
 684 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
 685 * called for the correct range previously.
 686 */
 687static void free_unmap_vmap_area_noflush(struct vmap_area *va)
 688{
 689	unmap_vmap_area(va);
 690	free_vmap_area_noflush(va);
 691}
 692
 693/*
 694 * Free and unmap a vmap area
 695 */
 696static void free_unmap_vmap_area(struct vmap_area *va)
 697{
 698	flush_cache_vunmap(va->va_start, va->va_end);
 699	free_unmap_vmap_area_noflush(va);
 700}
 701
 702static struct vmap_area *find_vmap_area(unsigned long addr)
 703{
 704	struct vmap_area *va;
 705
 706	spin_lock(&vmap_area_lock);
 707	va = __find_vmap_area(addr);
 708	spin_unlock(&vmap_area_lock);
 709
 710	return va;
 711}
 712
 713static void free_unmap_vmap_area_addr(unsigned long addr)
 714{
 715	struct vmap_area *va;
 716
 717	va = find_vmap_area(addr);
 718	BUG_ON(!va);
 719	free_unmap_vmap_area(va);
 720}
 721
 722
 723/*** Per cpu kva allocator ***/
 724
 725/*
 726 * vmap space is limited especially on 32 bit architectures. Ensure there is
 727 * room for at least 16 percpu vmap blocks per CPU.
 728 */
 729/*
 730 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
 731 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
 732 * instead (we just need a rough idea)
 733 */
 734#if BITS_PER_LONG == 32
 735#define VMALLOC_SPACE		(128UL*1024*1024)
 736#else
 737#define VMALLOC_SPACE		(128UL*1024*1024*1024)
 738#endif
 739
 740#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
 741#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
 742#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
 743#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
 744#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
 745#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
 746#define VMAP_BBMAP_BITS		\
 747		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
 748		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
 749			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
 750
 751#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
 752
 753static bool vmap_initialized __read_mostly = false;
 754
 755struct vmap_block_queue {
 756	spinlock_t lock;
 757	struct list_head free;
 758};
 759
 760struct vmap_block {
 761	spinlock_t lock;
 762	struct vmap_area *va;
 763	unsigned long free, dirty;
 764	DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
 765	struct list_head free_list;
 766	struct rcu_head rcu_head;
 767	struct list_head purge;
 768};
 769
 770/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
 771static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
 772
 773/*
 774 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
 775 * in the free path. Could get rid of this if we change the API to return a
 776 * "cookie" from alloc, to be passed to free. But no big deal yet.
 777 */
 778static DEFINE_SPINLOCK(vmap_block_tree_lock);
 779static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
 780
 781/*
 782 * We should probably have a fallback mechanism to allocate virtual memory
 783 * out of partially filled vmap blocks. However vmap block sizing should be
 784 * fairly reasonable according to the vmalloc size, so it shouldn't be a
 785 * big problem.
 786 */
 787
 788static unsigned long addr_to_vb_idx(unsigned long addr)
 789{
 790	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
 791	addr /= VMAP_BLOCK_SIZE;
 792	return addr;
 793}
 794
 795static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 796{
 797	struct vmap_block_queue *vbq;
 798	struct vmap_block *vb;
 799	struct vmap_area *va;
 800	unsigned long vb_idx;
 801	int node, err;
 
 802
 803	node = numa_node_id();
 804
 805	vb = kmalloc_node(sizeof(struct vmap_block),
 806			gfp_mask & GFP_RECLAIM_MASK, node);
 807	if (unlikely(!vb))
 808		return ERR_PTR(-ENOMEM);
 809
 810	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
 811					VMALLOC_START, VMALLOC_END,
 812					node, gfp_mask);
 813	if (IS_ERR(va)) {
 814		kfree(vb);
 815		return ERR_CAST(va);
 816	}
 817
 818	err = radix_tree_preload(gfp_mask);
 819	if (unlikely(err)) {
 820		kfree(vb);
 821		free_vmap_area(va);
 822		return ERR_PTR(err);
 823	}
 824
 
 825	spin_lock_init(&vb->lock);
 826	vb->va = va;
 827	vb->free = VMAP_BBMAP_BITS;
 
 
 828	vb->dirty = 0;
 829	bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
 
 830	INIT_LIST_HEAD(&vb->free_list);
 831
 832	vb_idx = addr_to_vb_idx(va->va_start);
 833	spin_lock(&vmap_block_tree_lock);
 834	err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
 835	spin_unlock(&vmap_block_tree_lock);
 836	BUG_ON(err);
 837	radix_tree_preload_end();
 838
 839	vbq = &get_cpu_var(vmap_block_queue);
 840	spin_lock(&vbq->lock);
 841	list_add_rcu(&vb->free_list, &vbq->free);
 842	spin_unlock(&vbq->lock);
 843	put_cpu_var(vmap_block_queue);
 844
 845	return vb;
 846}
 847
 848static void free_vmap_block(struct vmap_block *vb)
 849{
 850	struct vmap_block *tmp;
 851	unsigned long vb_idx;
 852
 853	vb_idx = addr_to_vb_idx(vb->va->va_start);
 854	spin_lock(&vmap_block_tree_lock);
 855	tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
 856	spin_unlock(&vmap_block_tree_lock);
 857	BUG_ON(tmp != vb);
 858
 859	free_vmap_area_noflush(vb->va);
 860	kfree_rcu(vb, rcu_head);
 861}
 862
 863static void purge_fragmented_blocks(int cpu)
 864{
 865	LIST_HEAD(purge);
 866	struct vmap_block *vb;
 867	struct vmap_block *n_vb;
 868	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
 869
 870	rcu_read_lock();
 871	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 872
 873		if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
 874			continue;
 875
 876		spin_lock(&vb->lock);
 877		if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
 878			vb->free = 0; /* prevent further allocs after releasing lock */
 879			vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
 880			bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
 
 881			spin_lock(&vbq->lock);
 882			list_del_rcu(&vb->free_list);
 883			spin_unlock(&vbq->lock);
 884			spin_unlock(&vb->lock);
 885			list_add_tail(&vb->purge, &purge);
 886		} else
 887			spin_unlock(&vb->lock);
 888	}
 889	rcu_read_unlock();
 890
 891	list_for_each_entry_safe(vb, n_vb, &purge, purge) {
 892		list_del(&vb->purge);
 893		free_vmap_block(vb);
 894	}
 895}
 896
 897static void purge_fragmented_blocks_allcpus(void)
 898{
 899	int cpu;
 900
 901	for_each_possible_cpu(cpu)
 902		purge_fragmented_blocks(cpu);
 903}
 904
 905static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
 906{
 907	struct vmap_block_queue *vbq;
 908	struct vmap_block *vb;
 909	unsigned long addr = 0;
 910	unsigned int order;
 911
 912	BUG_ON(size & ~PAGE_MASK);
 913	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 914	if (WARN_ON(size == 0)) {
 915		/*
 916		 * Allocating 0 bytes isn't what caller wants since
 917		 * get_order(0) returns funny result. Just warn and terminate
 918		 * early.
 919		 */
 920		return NULL;
 921	}
 922	order = get_order(size);
 923
 924again:
 925	rcu_read_lock();
 926	vbq = &get_cpu_var(vmap_block_queue);
 927	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 928		int i;
 929
 930		spin_lock(&vb->lock);
 931		if (vb->free < 1UL << order)
 932			goto next;
 
 
 933
 934		i = VMAP_BBMAP_BITS - vb->free;
 935		addr = vb->va->va_start + (i << PAGE_SHIFT);
 936		BUG_ON(addr_to_vb_idx(addr) !=
 937				addr_to_vb_idx(vb->va->va_start));
 938		vb->free -= 1UL << order;
 939		if (vb->free == 0) {
 940			spin_lock(&vbq->lock);
 941			list_del_rcu(&vb->free_list);
 942			spin_unlock(&vbq->lock);
 943		}
 
 944		spin_unlock(&vb->lock);
 945		break;
 946next:
 947		spin_unlock(&vb->lock);
 948	}
 949
 950	put_cpu_var(vmap_block_queue);
 951	rcu_read_unlock();
 952
 953	if (!addr) {
 954		vb = new_vmap_block(gfp_mask);
 955		if (IS_ERR(vb))
 956			return vb;
 957		goto again;
 958	}
 959
 960	return (void *)addr;
 961}
 962
 963static void vb_free(const void *addr, unsigned long size)
 964{
 965	unsigned long offset;
 966	unsigned long vb_idx;
 967	unsigned int order;
 968	struct vmap_block *vb;
 969
 970	BUG_ON(size & ~PAGE_MASK);
 971	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 972
 973	flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
 974
 975	order = get_order(size);
 976
 977	offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
 
 978
 979	vb_idx = addr_to_vb_idx((unsigned long)addr);
 980	rcu_read_lock();
 981	vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
 982	rcu_read_unlock();
 983	BUG_ON(!vb);
 984
 985	vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
 986
 987	spin_lock(&vb->lock);
 988	BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
 
 
 
 989
 990	vb->dirty += 1UL << order;
 991	if (vb->dirty == VMAP_BBMAP_BITS) {
 992		BUG_ON(vb->free);
 993		spin_unlock(&vb->lock);
 994		free_vmap_block(vb);
 995	} else
 996		spin_unlock(&vb->lock);
 997}
 998
 999/**
1000 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1001 *
1002 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1003 * to amortize TLB flushing overheads. What this means is that any page you
1004 * have now, may, in a former life, have been mapped into kernel virtual
1005 * address by the vmap layer and so there might be some CPUs with TLB entries
1006 * still referencing that page (additional to the regular 1:1 kernel mapping).
1007 *
1008 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1009 * be sure that none of the pages we have control over will have any aliases
1010 * from the vmap layer.
1011 */
1012void vm_unmap_aliases(void)
1013{
1014	unsigned long start = ULONG_MAX, end = 0;
1015	int cpu;
1016	int flush = 0;
1017
1018	if (unlikely(!vmap_initialized))
1019		return;
1020
1021	for_each_possible_cpu(cpu) {
1022		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1023		struct vmap_block *vb;
1024
1025		rcu_read_lock();
1026		list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1027			int i, j;
1028
1029			spin_lock(&vb->lock);
1030			i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1031			if (i < VMAP_BBMAP_BITS) {
1032				unsigned long s, e;
1033
1034				j = find_last_bit(vb->dirty_map,
1035							VMAP_BBMAP_BITS);
1036				j = j + 1; /* need exclusive index */
1037
1038				s = vb->va->va_start + (i << PAGE_SHIFT);
1039				e = vb->va->va_start + (j << PAGE_SHIFT);
1040				flush = 1;
1041
1042				if (s < start)
1043					start = s;
1044				if (e > end)
1045					end = e;
1046			}
1047			spin_unlock(&vb->lock);
1048		}
1049		rcu_read_unlock();
1050	}
1051
1052	__purge_vmap_area_lazy(&start, &end, 1, flush);
1053}
1054EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1055
1056/**
1057 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1058 * @mem: the pointer returned by vm_map_ram
1059 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1060 */
1061void vm_unmap_ram(const void *mem, unsigned int count)
1062{
1063	unsigned long size = count << PAGE_SHIFT;
1064	unsigned long addr = (unsigned long)mem;
1065
1066	BUG_ON(!addr);
1067	BUG_ON(addr < VMALLOC_START);
1068	BUG_ON(addr > VMALLOC_END);
1069	BUG_ON(addr & (PAGE_SIZE-1));
1070
1071	debug_check_no_locks_freed(mem, size);
1072	vmap_debug_free_range(addr, addr+size);
1073
1074	if (likely(count <= VMAP_MAX_ALLOC))
1075		vb_free(mem, size);
1076	else
1077		free_unmap_vmap_area_addr(addr);
1078}
1079EXPORT_SYMBOL(vm_unmap_ram);
1080
1081/**
1082 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1083 * @pages: an array of pointers to the pages to be mapped
1084 * @count: number of pages
1085 * @node: prefer to allocate data structures on this node
1086 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1087 *
1088 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1089 * faster than vmap so it's good.  But if you mix long-life and short-life
1090 * objects with vm_map_ram(), it could consume lots of address space through
1091 * fragmentation (especially on a 32bit machine).  You could see failures in
1092 * the end.  Please use this function for short-lived objects.
1093 *
1094 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1095 */
1096void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1097{
1098	unsigned long size = count << PAGE_SHIFT;
1099	unsigned long addr;
1100	void *mem;
1101
1102	if (likely(count <= VMAP_MAX_ALLOC)) {
1103		mem = vb_alloc(size, GFP_KERNEL);
1104		if (IS_ERR(mem))
1105			return NULL;
1106		addr = (unsigned long)mem;
1107	} else {
1108		struct vmap_area *va;
1109		va = alloc_vmap_area(size, PAGE_SIZE,
1110				VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1111		if (IS_ERR(va))
1112			return NULL;
1113
1114		addr = va->va_start;
1115		mem = (void *)addr;
1116	}
1117	if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1118		vm_unmap_ram(mem, count);
1119		return NULL;
1120	}
1121	return mem;
1122}
1123EXPORT_SYMBOL(vm_map_ram);
1124
1125static struct vm_struct *vmlist __initdata;
1126/**
1127 * vm_area_add_early - add vmap area early during boot
1128 * @vm: vm_struct to add
1129 *
1130 * This function is used to add fixed kernel vm area to vmlist before
1131 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
1132 * should contain proper values and the other fields should be zero.
1133 *
1134 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1135 */
1136void __init vm_area_add_early(struct vm_struct *vm)
1137{
1138	struct vm_struct *tmp, **p;
1139
1140	BUG_ON(vmap_initialized);
1141	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1142		if (tmp->addr >= vm->addr) {
1143			BUG_ON(tmp->addr < vm->addr + vm->size);
1144			break;
1145		} else
1146			BUG_ON(tmp->addr + tmp->size > vm->addr);
1147	}
1148	vm->next = *p;
1149	*p = vm;
1150}
1151
1152/**
1153 * vm_area_register_early - register vmap area early during boot
1154 * @vm: vm_struct to register
1155 * @align: requested alignment
1156 *
1157 * This function is used to register kernel vm area before
1158 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1159 * proper values on entry and other fields should be zero.  On return,
1160 * vm->addr contains the allocated address.
1161 *
1162 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1163 */
1164void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1165{
1166	static size_t vm_init_off __initdata;
1167	unsigned long addr;
1168
1169	addr = ALIGN(VMALLOC_START + vm_init_off, align);
1170	vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1171
1172	vm->addr = (void *)addr;
1173
1174	vm_area_add_early(vm);
1175}
1176
1177void __init vmalloc_init(void)
1178{
1179	struct vmap_area *va;
1180	struct vm_struct *tmp;
1181	int i;
1182
1183	for_each_possible_cpu(i) {
1184		struct vmap_block_queue *vbq;
1185		struct vfree_deferred *p;
1186
1187		vbq = &per_cpu(vmap_block_queue, i);
1188		spin_lock_init(&vbq->lock);
1189		INIT_LIST_HEAD(&vbq->free);
1190		p = &per_cpu(vfree_deferred, i);
1191		init_llist_head(&p->list);
1192		INIT_WORK(&p->wq, free_work);
1193	}
1194
1195	/* Import existing vmlist entries. */
1196	for (tmp = vmlist; tmp; tmp = tmp->next) {
1197		va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1198		va->flags = VM_VM_AREA;
1199		va->va_start = (unsigned long)tmp->addr;
1200		va->va_end = va->va_start + tmp->size;
1201		va->vm = tmp;
1202		__insert_vmap_area(va);
1203	}
1204
1205	vmap_area_pcpu_hole = VMALLOC_END;
1206
1207	vmap_initialized = true;
1208}
1209
1210/**
1211 * map_kernel_range_noflush - map kernel VM area with the specified pages
1212 * @addr: start of the VM area to map
1213 * @size: size of the VM area to map
1214 * @prot: page protection flags to use
1215 * @pages: pages to map
1216 *
1217 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1218 * specify should have been allocated using get_vm_area() and its
1219 * friends.
1220 *
1221 * NOTE:
1222 * This function does NOT do any cache flushing.  The caller is
1223 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1224 * before calling this function.
1225 *
1226 * RETURNS:
1227 * The number of pages mapped on success, -errno on failure.
1228 */
1229int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1230			     pgprot_t prot, struct page **pages)
1231{
1232	return vmap_page_range_noflush(addr, addr + size, prot, pages);
1233}
1234
1235/**
1236 * unmap_kernel_range_noflush - unmap kernel VM area
1237 * @addr: start of the VM area to unmap
1238 * @size: size of the VM area to unmap
1239 *
1240 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1241 * specify should have been allocated using get_vm_area() and its
1242 * friends.
1243 *
1244 * NOTE:
1245 * This function does NOT do any cache flushing.  The caller is
1246 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1247 * before calling this function and flush_tlb_kernel_range() after.
1248 */
1249void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1250{
1251	vunmap_page_range(addr, addr + size);
1252}
1253EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1254
1255/**
1256 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1257 * @addr: start of the VM area to unmap
1258 * @size: size of the VM area to unmap
1259 *
1260 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1261 * the unmapping and tlb after.
1262 */
1263void unmap_kernel_range(unsigned long addr, unsigned long size)
1264{
1265	unsigned long end = addr + size;
1266
1267	flush_cache_vunmap(addr, end);
1268	vunmap_page_range(addr, end);
1269	flush_tlb_kernel_range(addr, end);
1270}
 
1271
1272int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1273{
1274	unsigned long addr = (unsigned long)area->addr;
1275	unsigned long end = addr + get_vm_area_size(area);
1276	int err;
1277
1278	err = vmap_page_range(addr, end, prot, *pages);
1279	if (err > 0) {
1280		*pages += err;
1281		err = 0;
1282	}
1283
1284	return err;
1285}
1286EXPORT_SYMBOL_GPL(map_vm_area);
1287
1288static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1289			      unsigned long flags, const void *caller)
1290{
1291	spin_lock(&vmap_area_lock);
1292	vm->flags = flags;
1293	vm->addr = (void *)va->va_start;
1294	vm->size = va->va_end - va->va_start;
1295	vm->caller = caller;
1296	va->vm = vm;
1297	va->flags |= VM_VM_AREA;
1298	spin_unlock(&vmap_area_lock);
1299}
1300
1301static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1302{
1303	/*
1304	 * Before removing VM_UNINITIALIZED,
1305	 * we should make sure that vm has proper values.
1306	 * Pair with smp_rmb() in show_numa_info().
1307	 */
1308	smp_wmb();
1309	vm->flags &= ~VM_UNINITIALIZED;
1310}
1311
1312static struct vm_struct *__get_vm_area_node(unsigned long size,
1313		unsigned long align, unsigned long flags, unsigned long start,
1314		unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1315{
1316	struct vmap_area *va;
1317	struct vm_struct *area;
1318
1319	BUG_ON(in_interrupt());
1320	if (flags & VM_IOREMAP)
1321		align = 1ul << clamp(fls(size), PAGE_SHIFT, IOREMAP_MAX_ORDER);
 
1322
1323	size = PAGE_ALIGN(size);
1324	if (unlikely(!size))
1325		return NULL;
1326
1327	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1328	if (unlikely(!area))
1329		return NULL;
1330
1331	/*
1332	 * We always allocate a guard page.
1333	 */
1334	size += PAGE_SIZE;
1335
1336	va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1337	if (IS_ERR(va)) {
1338		kfree(area);
1339		return NULL;
1340	}
1341
1342	setup_vmalloc_vm(area, va, flags, caller);
1343
1344	return area;
1345}
1346
1347struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1348				unsigned long start, unsigned long end)
1349{
1350	return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1351				  GFP_KERNEL, __builtin_return_address(0));
1352}
1353EXPORT_SYMBOL_GPL(__get_vm_area);
1354
1355struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1356				       unsigned long start, unsigned long end,
1357				       const void *caller)
1358{
1359	return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1360				  GFP_KERNEL, caller);
1361}
1362
1363/**
1364 *	get_vm_area  -  reserve a contiguous kernel virtual area
1365 *	@size:		size of the area
1366 *	@flags:		%VM_IOREMAP for I/O mappings or VM_ALLOC
1367 *
1368 *	Search an area of @size in the kernel virtual mapping area,
1369 *	and reserved it for out purposes.  Returns the area descriptor
1370 *	on success or %NULL on failure.
1371 */
1372struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1373{
1374	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1375				  NUMA_NO_NODE, GFP_KERNEL,
1376				  __builtin_return_address(0));
1377}
1378
1379struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1380				const void *caller)
1381{
1382	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1383				  NUMA_NO_NODE, GFP_KERNEL, caller);
1384}
1385
1386/**
1387 *	find_vm_area  -  find a continuous kernel virtual area
1388 *	@addr:		base address
1389 *
1390 *	Search for the kernel VM area starting at @addr, and return it.
1391 *	It is up to the caller to do all required locking to keep the returned
1392 *	pointer valid.
1393 */
1394struct vm_struct *find_vm_area(const void *addr)
1395{
1396	struct vmap_area *va;
1397
1398	va = find_vmap_area((unsigned long)addr);
1399	if (va && va->flags & VM_VM_AREA)
1400		return va->vm;
1401
1402	return NULL;
1403}
1404
1405/**
1406 *	remove_vm_area  -  find and remove a continuous kernel virtual area
1407 *	@addr:		base address
1408 *
1409 *	Search for the kernel VM area starting at @addr, and remove it.
1410 *	This function returns the found VM area, but using it is NOT safe
1411 *	on SMP machines, except for its size or flags.
1412 */
1413struct vm_struct *remove_vm_area(const void *addr)
1414{
1415	struct vmap_area *va;
1416
1417	va = find_vmap_area((unsigned long)addr);
1418	if (va && va->flags & VM_VM_AREA) {
1419		struct vm_struct *vm = va->vm;
1420
1421		spin_lock(&vmap_area_lock);
1422		va->vm = NULL;
1423		va->flags &= ~VM_VM_AREA;
1424		spin_unlock(&vmap_area_lock);
1425
1426		vmap_debug_free_range(va->va_start, va->va_end);
 
1427		free_unmap_vmap_area(va);
1428		vm->size -= PAGE_SIZE;
1429
1430		return vm;
1431	}
1432	return NULL;
1433}
1434
1435static void __vunmap(const void *addr, int deallocate_pages)
1436{
1437	struct vm_struct *area;
1438
1439	if (!addr)
1440		return;
1441
1442	if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1443			addr))
1444		return;
1445
1446	area = remove_vm_area(addr);
1447	if (unlikely(!area)) {
1448		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1449				addr);
1450		return;
1451	}
1452
1453	debug_check_no_locks_freed(addr, area->size);
1454	debug_check_no_obj_freed(addr, area->size);
1455
1456	if (deallocate_pages) {
1457		int i;
1458
1459		for (i = 0; i < area->nr_pages; i++) {
1460			struct page *page = area->pages[i];
1461
1462			BUG_ON(!page);
1463			__free_page(page);
1464		}
1465
1466		if (area->flags & VM_VPAGES)
1467			vfree(area->pages);
1468		else
1469			kfree(area->pages);
1470	}
1471
1472	kfree(area);
1473	return;
1474}
1475 
1476/**
1477 *	vfree  -  release memory allocated by vmalloc()
1478 *	@addr:		memory base address
1479 *
1480 *	Free the virtually continuous memory area starting at @addr, as
1481 *	obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1482 *	NULL, no operation is performed.
1483 *
1484 *	Must not be called in NMI context (strictly speaking, only if we don't
1485 *	have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1486 *	conventions for vfree() arch-depenedent would be a really bad idea)
1487 *
1488 *	NOTE: assumes that the object at *addr has a size >= sizeof(llist_node)
1489 */
1490void vfree(const void *addr)
1491{
1492	BUG_ON(in_nmi());
1493
1494	kmemleak_free(addr);
1495
1496	if (!addr)
1497		return;
1498	if (unlikely(in_interrupt())) {
1499		struct vfree_deferred *p = &__get_cpu_var(vfree_deferred);
1500		if (llist_add((struct llist_node *)addr, &p->list))
1501			schedule_work(&p->wq);
1502	} else
1503		__vunmap(addr, 1);
1504}
1505EXPORT_SYMBOL(vfree);
1506
1507/**
1508 *	vunmap  -  release virtual mapping obtained by vmap()
1509 *	@addr:		memory base address
1510 *
1511 *	Free the virtually contiguous memory area starting at @addr,
1512 *	which was created from the page array passed to vmap().
1513 *
1514 *	Must not be called in interrupt context.
1515 */
1516void vunmap(const void *addr)
1517{
1518	BUG_ON(in_interrupt());
1519	might_sleep();
1520	if (addr)
1521		__vunmap(addr, 0);
1522}
1523EXPORT_SYMBOL(vunmap);
1524
1525/**
1526 *	vmap  -  map an array of pages into virtually contiguous space
1527 *	@pages:		array of page pointers
1528 *	@count:		number of pages to map
1529 *	@flags:		vm_area->flags
1530 *	@prot:		page protection for the mapping
1531 *
1532 *	Maps @count pages from @pages into contiguous kernel virtual
1533 *	space.
1534 */
1535void *vmap(struct page **pages, unsigned int count,
1536		unsigned long flags, pgprot_t prot)
1537{
1538	struct vm_struct *area;
1539
1540	might_sleep();
1541
1542	if (count > totalram_pages)
1543		return NULL;
1544
1545	area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1546					__builtin_return_address(0));
1547	if (!area)
1548		return NULL;
1549
1550	if (map_vm_area(area, prot, &pages)) {
1551		vunmap(area->addr);
1552		return NULL;
1553	}
1554
1555	return area->addr;
1556}
1557EXPORT_SYMBOL(vmap);
1558
1559static void *__vmalloc_node(unsigned long size, unsigned long align,
1560			    gfp_t gfp_mask, pgprot_t prot,
1561			    int node, const void *caller);
1562static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1563				 pgprot_t prot, int node)
1564{
1565	const int order = 0;
1566	struct page **pages;
1567	unsigned int nr_pages, array_size, i;
1568	gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
 
1569
1570	nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1571	array_size = (nr_pages * sizeof(struct page *));
1572
1573	area->nr_pages = nr_pages;
1574	/* Please note that the recursion is strictly bounded. */
1575	if (array_size > PAGE_SIZE) {
1576		pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1577				PAGE_KERNEL, node, area->caller);
1578		area->flags |= VM_VPAGES;
1579	} else {
1580		pages = kmalloc_node(array_size, nested_gfp, node);
1581	}
1582	area->pages = pages;
1583	if (!area->pages) {
1584		remove_vm_area(area->addr);
1585		kfree(area);
1586		return NULL;
1587	}
1588
1589	for (i = 0; i < area->nr_pages; i++) {
1590		struct page *page;
1591		gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1592
1593		if (node == NUMA_NO_NODE)
1594			page = alloc_page(tmp_mask);
1595		else
1596			page = alloc_pages_node(node, tmp_mask, order);
1597
1598		if (unlikely(!page)) {
1599			/* Successfully allocated i pages, free them in __vunmap() */
1600			area->nr_pages = i;
1601			goto fail;
1602		}
1603		area->pages[i] = page;
 
 
1604	}
1605
1606	if (map_vm_area(area, prot, &pages))
1607		goto fail;
1608	return area->addr;
1609
1610fail:
1611	warn_alloc_failed(gfp_mask, order,
1612			  "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1613			  (area->nr_pages*PAGE_SIZE), area->size);
1614	vfree(area->addr);
1615	return NULL;
1616}
1617
1618/**
1619 *	__vmalloc_node_range  -  allocate virtually contiguous memory
1620 *	@size:		allocation size
1621 *	@align:		desired alignment
1622 *	@start:		vm area range start
1623 *	@end:		vm area range end
1624 *	@gfp_mask:	flags for the page level allocator
1625 *	@prot:		protection mask for the allocated pages
 
1626 *	@node:		node to use for allocation or NUMA_NO_NODE
1627 *	@caller:	caller's return address
1628 *
1629 *	Allocate enough pages to cover @size from the page level
1630 *	allocator with @gfp_mask flags.  Map them into contiguous
1631 *	kernel virtual space, using a pagetable protection of @prot.
1632 */
1633void *__vmalloc_node_range(unsigned long size, unsigned long align,
1634			unsigned long start, unsigned long end, gfp_t gfp_mask,
1635			pgprot_t prot, int node, const void *caller)
 
1636{
1637	struct vm_struct *area;
1638	void *addr;
1639	unsigned long real_size = size;
1640
1641	size = PAGE_ALIGN(size);
1642	if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1643		goto fail;
1644
1645	area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED,
1646				  start, end, node, gfp_mask, caller);
1647	if (!area)
1648		goto fail;
1649
1650	addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1651	if (!addr)
1652		return NULL;
1653
1654	/*
1655	 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1656	 * flag. It means that vm_struct is not fully initialized.
1657	 * Now, it is fully initialized, so remove this flag here.
1658	 */
1659	clear_vm_uninitialized_flag(area);
1660
1661	/*
1662	 * A ref_count = 2 is needed because vm_struct allocated in
1663	 * __get_vm_area_node() contains a reference to the virtual address of
1664	 * the vmalloc'ed block.
1665	 */
1666	kmemleak_alloc(addr, real_size, 2, gfp_mask);
1667
1668	return addr;
1669
1670fail:
1671	warn_alloc_failed(gfp_mask, 0,
1672			  "vmalloc: allocation failure: %lu bytes\n",
1673			  real_size);
1674	return NULL;
1675}
1676
1677/**
1678 *	__vmalloc_node  -  allocate virtually contiguous memory
1679 *	@size:		allocation size
1680 *	@align:		desired alignment
1681 *	@gfp_mask:	flags for the page level allocator
1682 *	@prot:		protection mask for the allocated pages
1683 *	@node:		node to use for allocation or NUMA_NO_NODE
1684 *	@caller:	caller's return address
1685 *
1686 *	Allocate enough pages to cover @size from the page level
1687 *	allocator with @gfp_mask flags.  Map them into contiguous
1688 *	kernel virtual space, using a pagetable protection of @prot.
1689 */
1690static void *__vmalloc_node(unsigned long size, unsigned long align,
1691			    gfp_t gfp_mask, pgprot_t prot,
1692			    int node, const void *caller)
1693{
1694	return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1695				gfp_mask, prot, node, caller);
1696}
1697
1698void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1699{
1700	return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1701				__builtin_return_address(0));
1702}
1703EXPORT_SYMBOL(__vmalloc);
1704
1705static inline void *__vmalloc_node_flags(unsigned long size,
1706					int node, gfp_t flags)
1707{
1708	return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1709					node, __builtin_return_address(0));
1710}
1711
1712/**
1713 *	vmalloc  -  allocate virtually contiguous memory
1714 *	@size:		allocation size
1715 *	Allocate enough pages to cover @size from the page level
1716 *	allocator and map them into contiguous kernel virtual space.
1717 *
1718 *	For tight control over page level allocator and protection flags
1719 *	use __vmalloc() instead.
1720 */
1721void *vmalloc(unsigned long size)
1722{
1723	return __vmalloc_node_flags(size, NUMA_NO_NODE,
1724				    GFP_KERNEL | __GFP_HIGHMEM);
1725}
1726EXPORT_SYMBOL(vmalloc);
1727
1728/**
1729 *	vzalloc - allocate virtually contiguous memory with zero fill
1730 *	@size:	allocation size
1731 *	Allocate enough pages to cover @size from the page level
1732 *	allocator and map them into contiguous kernel virtual space.
1733 *	The memory allocated is set to zero.
1734 *
1735 *	For tight control over page level allocator and protection flags
1736 *	use __vmalloc() instead.
1737 */
1738void *vzalloc(unsigned long size)
1739{
1740	return __vmalloc_node_flags(size, NUMA_NO_NODE,
1741				GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1742}
1743EXPORT_SYMBOL(vzalloc);
1744
1745/**
1746 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1747 * @size: allocation size
1748 *
1749 * The resulting memory area is zeroed so it can be mapped to userspace
1750 * without leaking data.
1751 */
1752void *vmalloc_user(unsigned long size)
1753{
1754	struct vm_struct *area;
1755	void *ret;
1756
1757	ret = __vmalloc_node(size, SHMLBA,
1758			     GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1759			     PAGE_KERNEL, NUMA_NO_NODE,
1760			     __builtin_return_address(0));
1761	if (ret) {
1762		area = find_vm_area(ret);
1763		area->flags |= VM_USERMAP;
1764	}
1765	return ret;
1766}
1767EXPORT_SYMBOL(vmalloc_user);
1768
1769/**
1770 *	vmalloc_node  -  allocate memory on a specific node
1771 *	@size:		allocation size
1772 *	@node:		numa node
1773 *
1774 *	Allocate enough pages to cover @size from the page level
1775 *	allocator and map them into contiguous kernel virtual space.
1776 *
1777 *	For tight control over page level allocator and protection flags
1778 *	use __vmalloc() instead.
1779 */
1780void *vmalloc_node(unsigned long size, int node)
1781{
1782	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1783					node, __builtin_return_address(0));
1784}
1785EXPORT_SYMBOL(vmalloc_node);
1786
1787/**
1788 * vzalloc_node - allocate memory on a specific node with zero fill
1789 * @size:	allocation size
1790 * @node:	numa node
1791 *
1792 * Allocate enough pages to cover @size from the page level
1793 * allocator and map them into contiguous kernel virtual space.
1794 * The memory allocated is set to zero.
1795 *
1796 * For tight control over page level allocator and protection flags
1797 * use __vmalloc_node() instead.
1798 */
1799void *vzalloc_node(unsigned long size, int node)
1800{
1801	return __vmalloc_node_flags(size, node,
1802			 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1803}
1804EXPORT_SYMBOL(vzalloc_node);
1805
1806#ifndef PAGE_KERNEL_EXEC
1807# define PAGE_KERNEL_EXEC PAGE_KERNEL
1808#endif
1809
1810/**
1811 *	vmalloc_exec  -  allocate virtually contiguous, executable memory
1812 *	@size:		allocation size
1813 *
1814 *	Kernel-internal function to allocate enough pages to cover @size
1815 *	the page level allocator and map them into contiguous and
1816 *	executable kernel virtual space.
1817 *
1818 *	For tight control over page level allocator and protection flags
1819 *	use __vmalloc() instead.
1820 */
1821
1822void *vmalloc_exec(unsigned long size)
1823{
1824	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1825			      NUMA_NO_NODE, __builtin_return_address(0));
1826}
1827
1828#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1829#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1830#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1831#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1832#else
1833#define GFP_VMALLOC32 GFP_KERNEL
1834#endif
1835
1836/**
1837 *	vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1838 *	@size:		allocation size
1839 *
1840 *	Allocate enough 32bit PA addressable pages to cover @size from the
1841 *	page level allocator and map them into contiguous kernel virtual space.
1842 */
1843void *vmalloc_32(unsigned long size)
1844{
1845	return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1846			      NUMA_NO_NODE, __builtin_return_address(0));
1847}
1848EXPORT_SYMBOL(vmalloc_32);
1849
1850/**
1851 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1852 *	@size:		allocation size
1853 *
1854 * The resulting memory area is 32bit addressable and zeroed so it can be
1855 * mapped to userspace without leaking data.
1856 */
1857void *vmalloc_32_user(unsigned long size)
1858{
1859	struct vm_struct *area;
1860	void *ret;
1861
1862	ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1863			     NUMA_NO_NODE, __builtin_return_address(0));
1864	if (ret) {
1865		area = find_vm_area(ret);
1866		area->flags |= VM_USERMAP;
1867	}
1868	return ret;
1869}
1870EXPORT_SYMBOL(vmalloc_32_user);
1871
1872/*
1873 * small helper routine , copy contents to buf from addr.
1874 * If the page is not present, fill zero.
1875 */
1876
1877static int aligned_vread(char *buf, char *addr, unsigned long count)
1878{
1879	struct page *p;
1880	int copied = 0;
1881
1882	while (count) {
1883		unsigned long offset, length;
1884
1885		offset = (unsigned long)addr & ~PAGE_MASK;
1886		length = PAGE_SIZE - offset;
1887		if (length > count)
1888			length = count;
1889		p = vmalloc_to_page(addr);
1890		/*
1891		 * To do safe access to this _mapped_ area, we need
1892		 * lock. But adding lock here means that we need to add
1893		 * overhead of vmalloc()/vfree() calles for this _debug_
1894		 * interface, rarely used. Instead of that, we'll use
1895		 * kmap() and get small overhead in this access function.
1896		 */
1897		if (p) {
1898			/*
1899			 * we can expect USER0 is not used (see vread/vwrite's
1900			 * function description)
1901			 */
1902			void *map = kmap_atomic(p);
1903			memcpy(buf, map + offset, length);
1904			kunmap_atomic(map);
1905		} else
1906			memset(buf, 0, length);
1907
1908		addr += length;
1909		buf += length;
1910		copied += length;
1911		count -= length;
1912	}
1913	return copied;
1914}
1915
1916static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1917{
1918	struct page *p;
1919	int copied = 0;
1920
1921	while (count) {
1922		unsigned long offset, length;
1923
1924		offset = (unsigned long)addr & ~PAGE_MASK;
1925		length = PAGE_SIZE - offset;
1926		if (length > count)
1927			length = count;
1928		p = vmalloc_to_page(addr);
1929		/*
1930		 * To do safe access to this _mapped_ area, we need
1931		 * lock. But adding lock here means that we need to add
1932		 * overhead of vmalloc()/vfree() calles for this _debug_
1933		 * interface, rarely used. Instead of that, we'll use
1934		 * kmap() and get small overhead in this access function.
1935		 */
1936		if (p) {
1937			/*
1938			 * we can expect USER0 is not used (see vread/vwrite's
1939			 * function description)
1940			 */
1941			void *map = kmap_atomic(p);
1942			memcpy(map + offset, buf, length);
1943			kunmap_atomic(map);
1944		}
1945		addr += length;
1946		buf += length;
1947		copied += length;
1948		count -= length;
1949	}
1950	return copied;
1951}
1952
1953/**
1954 *	vread() -  read vmalloc area in a safe way.
1955 *	@buf:		buffer for reading data
1956 *	@addr:		vm address.
1957 *	@count:		number of bytes to be read.
1958 *
1959 *	Returns # of bytes which addr and buf should be increased.
1960 *	(same number to @count). Returns 0 if [addr...addr+count) doesn't
1961 *	includes any intersect with alive vmalloc area.
1962 *
1963 *	This function checks that addr is a valid vmalloc'ed area, and
1964 *	copy data from that area to a given buffer. If the given memory range
1965 *	of [addr...addr+count) includes some valid address, data is copied to
1966 *	proper area of @buf. If there are memory holes, they'll be zero-filled.
1967 *	IOREMAP area is treated as memory hole and no copy is done.
1968 *
1969 *	If [addr...addr+count) doesn't includes any intersects with alive
1970 *	vm_struct area, returns 0. @buf should be kernel's buffer.
1971 *
1972 *	Note: In usual ops, vread() is never necessary because the caller
1973 *	should know vmalloc() area is valid and can use memcpy().
1974 *	This is for routines which have to access vmalloc area without
1975 *	any informaion, as /dev/kmem.
1976 *
1977 */
1978
1979long vread(char *buf, char *addr, unsigned long count)
1980{
1981	struct vmap_area *va;
1982	struct vm_struct *vm;
1983	char *vaddr, *buf_start = buf;
1984	unsigned long buflen = count;
1985	unsigned long n;
1986
1987	/* Don't allow overflow */
1988	if ((unsigned long) addr + count < count)
1989		count = -(unsigned long) addr;
1990
1991	spin_lock(&vmap_area_lock);
1992	list_for_each_entry(va, &vmap_area_list, list) {
1993		if (!count)
1994			break;
1995
1996		if (!(va->flags & VM_VM_AREA))
1997			continue;
1998
1999		vm = va->vm;
2000		vaddr = (char *) vm->addr;
2001		if (addr >= vaddr + get_vm_area_size(vm))
2002			continue;
2003		while (addr < vaddr) {
2004			if (count == 0)
2005				goto finished;
2006			*buf = '\0';
2007			buf++;
2008			addr++;
2009			count--;
2010		}
2011		n = vaddr + get_vm_area_size(vm) - addr;
2012		if (n > count)
2013			n = count;
2014		if (!(vm->flags & VM_IOREMAP))
2015			aligned_vread(buf, addr, n);
2016		else /* IOREMAP area is treated as memory hole */
2017			memset(buf, 0, n);
2018		buf += n;
2019		addr += n;
2020		count -= n;
2021	}
2022finished:
2023	spin_unlock(&vmap_area_lock);
2024
2025	if (buf == buf_start)
2026		return 0;
2027	/* zero-fill memory holes */
2028	if (buf != buf_start + buflen)
2029		memset(buf, 0, buflen - (buf - buf_start));
2030
2031	return buflen;
2032}
2033
2034/**
2035 *	vwrite() -  write vmalloc area in a safe way.
2036 *	@buf:		buffer for source data
2037 *	@addr:		vm address.
2038 *	@count:		number of bytes to be read.
2039 *
2040 *	Returns # of bytes which addr and buf should be incresed.
2041 *	(same number to @count).
2042 *	If [addr...addr+count) doesn't includes any intersect with valid
2043 *	vmalloc area, returns 0.
2044 *
2045 *	This function checks that addr is a valid vmalloc'ed area, and
2046 *	copy data from a buffer to the given addr. If specified range of
2047 *	[addr...addr+count) includes some valid address, data is copied from
2048 *	proper area of @buf. If there are memory holes, no copy to hole.
2049 *	IOREMAP area is treated as memory hole and no copy is done.
2050 *
2051 *	If [addr...addr+count) doesn't includes any intersects with alive
2052 *	vm_struct area, returns 0. @buf should be kernel's buffer.
2053 *
2054 *	Note: In usual ops, vwrite() is never necessary because the caller
2055 *	should know vmalloc() area is valid and can use memcpy().
2056 *	This is for routines which have to access vmalloc area without
2057 *	any informaion, as /dev/kmem.
2058 */
2059
2060long vwrite(char *buf, char *addr, unsigned long count)
2061{
2062	struct vmap_area *va;
2063	struct vm_struct *vm;
2064	char *vaddr;
2065	unsigned long n, buflen;
2066	int copied = 0;
2067
2068	/* Don't allow overflow */
2069	if ((unsigned long) addr + count < count)
2070		count = -(unsigned long) addr;
2071	buflen = count;
2072
2073	spin_lock(&vmap_area_lock);
2074	list_for_each_entry(va, &vmap_area_list, list) {
2075		if (!count)
2076			break;
2077
2078		if (!(va->flags & VM_VM_AREA))
2079			continue;
2080
2081		vm = va->vm;
2082		vaddr = (char *) vm->addr;
2083		if (addr >= vaddr + get_vm_area_size(vm))
2084			continue;
2085		while (addr < vaddr) {
2086			if (count == 0)
2087				goto finished;
2088			buf++;
2089			addr++;
2090			count--;
2091		}
2092		n = vaddr + get_vm_area_size(vm) - addr;
2093		if (n > count)
2094			n = count;
2095		if (!(vm->flags & VM_IOREMAP)) {
2096			aligned_vwrite(buf, addr, n);
2097			copied++;
2098		}
2099		buf += n;
2100		addr += n;
2101		count -= n;
2102	}
2103finished:
2104	spin_unlock(&vmap_area_lock);
2105	if (!copied)
2106		return 0;
2107	return buflen;
2108}
2109
2110/**
2111 *	remap_vmalloc_range_partial  -  map vmalloc pages to userspace
2112 *	@vma:		vma to cover
2113 *	@uaddr:		target user address to start at
2114 *	@kaddr:		virtual address of vmalloc kernel memory
2115 *	@size:		size of map area
2116 *
2117 *	Returns:	0 for success, -Exxx on failure
2118 *
2119 *	This function checks that @kaddr is a valid vmalloc'ed area,
2120 *	and that it is big enough to cover the range starting at
2121 *	@uaddr in @vma. Will return failure if that criteria isn't
2122 *	met.
2123 *
2124 *	Similar to remap_pfn_range() (see mm/memory.c)
2125 */
2126int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2127				void *kaddr, unsigned long size)
2128{
2129	struct vm_struct *area;
2130
2131	size = PAGE_ALIGN(size);
2132
2133	if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2134		return -EINVAL;
2135
2136	area = find_vm_area(kaddr);
2137	if (!area)
2138		return -EINVAL;
2139
2140	if (!(area->flags & VM_USERMAP))
2141		return -EINVAL;
2142
2143	if (kaddr + size > area->addr + area->size)
2144		return -EINVAL;
2145
2146	do {
2147		struct page *page = vmalloc_to_page(kaddr);
2148		int ret;
2149
2150		ret = vm_insert_page(vma, uaddr, page);
2151		if (ret)
2152			return ret;
2153
2154		uaddr += PAGE_SIZE;
2155		kaddr += PAGE_SIZE;
2156		size -= PAGE_SIZE;
2157	} while (size > 0);
2158
2159	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2160
2161	return 0;
2162}
2163EXPORT_SYMBOL(remap_vmalloc_range_partial);
2164
2165/**
2166 *	remap_vmalloc_range  -  map vmalloc pages to userspace
2167 *	@vma:		vma to cover (map full range of vma)
2168 *	@addr:		vmalloc memory
2169 *	@pgoff:		number of pages into addr before first page to map
2170 *
2171 *	Returns:	0 for success, -Exxx on failure
2172 *
2173 *	This function checks that addr is a valid vmalloc'ed area, and
2174 *	that it is big enough to cover the vma. Will return failure if
2175 *	that criteria isn't met.
2176 *
2177 *	Similar to remap_pfn_range() (see mm/memory.c)
2178 */
2179int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2180						unsigned long pgoff)
2181{
2182	return remap_vmalloc_range_partial(vma, vma->vm_start,
2183					   addr + (pgoff << PAGE_SHIFT),
2184					   vma->vm_end - vma->vm_start);
2185}
2186EXPORT_SYMBOL(remap_vmalloc_range);
2187
2188/*
2189 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2190 * have one.
2191 */
2192void __weak vmalloc_sync_all(void)
2193{
2194}
2195
2196
2197static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2198{
2199	pte_t ***p = data;
2200
2201	if (p) {
2202		*(*p) = pte;
2203		(*p)++;
2204	}
2205	return 0;
2206}
2207
2208/**
2209 *	alloc_vm_area - allocate a range of kernel address space
2210 *	@size:		size of the area
2211 *	@ptes:		returns the PTEs for the address space
2212 *
2213 *	Returns:	NULL on failure, vm_struct on success
2214 *
2215 *	This function reserves a range of kernel address space, and
2216 *	allocates pagetables to map that range.  No actual mappings
2217 *	are created.
2218 *
2219 *	If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2220 *	allocated for the VM area are returned.
2221 */
2222struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2223{
2224	struct vm_struct *area;
2225
2226	area = get_vm_area_caller(size, VM_IOREMAP,
2227				__builtin_return_address(0));
2228	if (area == NULL)
2229		return NULL;
2230
2231	/*
2232	 * This ensures that page tables are constructed for this region
2233	 * of kernel virtual address space and mapped into init_mm.
2234	 */
2235	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2236				size, f, ptes ? &ptes : NULL)) {
2237		free_vm_area(area);
2238		return NULL;
2239	}
2240
2241	return area;
2242}
2243EXPORT_SYMBOL_GPL(alloc_vm_area);
2244
2245void free_vm_area(struct vm_struct *area)
2246{
2247	struct vm_struct *ret;
2248	ret = remove_vm_area(area->addr);
2249	BUG_ON(ret != area);
2250	kfree(area);
2251}
2252EXPORT_SYMBOL_GPL(free_vm_area);
2253
2254#ifdef CONFIG_SMP
2255static struct vmap_area *node_to_va(struct rb_node *n)
2256{
2257	return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2258}
2259
2260/**
2261 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2262 * @end: target address
2263 * @pnext: out arg for the next vmap_area
2264 * @pprev: out arg for the previous vmap_area
2265 *
2266 * Returns: %true if either or both of next and prev are found,
2267 *	    %false if no vmap_area exists
2268 *
2269 * Find vmap_areas end addresses of which enclose @end.  ie. if not
2270 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2271 */
2272static bool pvm_find_next_prev(unsigned long end,
2273			       struct vmap_area **pnext,
2274			       struct vmap_area **pprev)
2275{
2276	struct rb_node *n = vmap_area_root.rb_node;
2277	struct vmap_area *va = NULL;
2278
2279	while (n) {
2280		va = rb_entry(n, struct vmap_area, rb_node);
2281		if (end < va->va_end)
2282			n = n->rb_left;
2283		else if (end > va->va_end)
2284			n = n->rb_right;
2285		else
2286			break;
2287	}
2288
2289	if (!va)
2290		return false;
2291
2292	if (va->va_end > end) {
2293		*pnext = va;
2294		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2295	} else {
2296		*pprev = va;
2297		*pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2298	}
2299	return true;
2300}
2301
2302/**
2303 * pvm_determine_end - find the highest aligned address between two vmap_areas
2304 * @pnext: in/out arg for the next vmap_area
2305 * @pprev: in/out arg for the previous vmap_area
2306 * @align: alignment
2307 *
2308 * Returns: determined end address
2309 *
2310 * Find the highest aligned address between *@pnext and *@pprev below
2311 * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2312 * down address is between the end addresses of the two vmap_areas.
2313 *
2314 * Please note that the address returned by this function may fall
2315 * inside *@pnext vmap_area.  The caller is responsible for checking
2316 * that.
2317 */
2318static unsigned long pvm_determine_end(struct vmap_area **pnext,
2319				       struct vmap_area **pprev,
2320				       unsigned long align)
2321{
2322	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2323	unsigned long addr;
2324
2325	if (*pnext)
2326		addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2327	else
2328		addr = vmalloc_end;
2329
2330	while (*pprev && (*pprev)->va_end > addr) {
2331		*pnext = *pprev;
2332		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2333	}
2334
2335	return addr;
2336}
2337
2338/**
2339 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2340 * @offsets: array containing offset of each area
2341 * @sizes: array containing size of each area
2342 * @nr_vms: the number of areas to allocate
2343 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2344 *
2345 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2346 *	    vm_structs on success, %NULL on failure
2347 *
2348 * Percpu allocator wants to use congruent vm areas so that it can
2349 * maintain the offsets among percpu areas.  This function allocates
2350 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
2351 * be scattered pretty far, distance between two areas easily going up
2352 * to gigabytes.  To avoid interacting with regular vmallocs, these
2353 * areas are allocated from top.
2354 *
2355 * Despite its complicated look, this allocator is rather simple.  It
2356 * does everything top-down and scans areas from the end looking for
2357 * matching slot.  While scanning, if any of the areas overlaps with
2358 * existing vmap_area, the base address is pulled down to fit the
2359 * area.  Scanning is repeated till all the areas fit and then all
2360 * necessary data structres are inserted and the result is returned.
2361 */
2362struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2363				     const size_t *sizes, int nr_vms,
2364				     size_t align)
2365{
2366	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2367	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2368	struct vmap_area **vas, *prev, *next;
2369	struct vm_struct **vms;
2370	int area, area2, last_area, term_area;
2371	unsigned long base, start, end, last_end;
2372	bool purged = false;
2373
2374	/* verify parameters and allocate data structures */
2375	BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2376	for (last_area = 0, area = 0; area < nr_vms; area++) {
2377		start = offsets[area];
2378		end = start + sizes[area];
2379
2380		/* is everything aligned properly? */
2381		BUG_ON(!IS_ALIGNED(offsets[area], align));
2382		BUG_ON(!IS_ALIGNED(sizes[area], align));
2383
2384		/* detect the area with the highest address */
2385		if (start > offsets[last_area])
2386			last_area = area;
2387
2388		for (area2 = 0; area2 < nr_vms; area2++) {
2389			unsigned long start2 = offsets[area2];
2390			unsigned long end2 = start2 + sizes[area2];
2391
2392			if (area2 == area)
2393				continue;
2394
2395			BUG_ON(start2 >= start && start2 < end);
2396			BUG_ON(end2 <= end && end2 > start);
2397		}
2398	}
2399	last_end = offsets[last_area] + sizes[last_area];
2400
2401	if (vmalloc_end - vmalloc_start < last_end) {
2402		WARN_ON(true);
2403		return NULL;
2404	}
2405
2406	vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2407	vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2408	if (!vas || !vms)
2409		goto err_free2;
2410
2411	for (area = 0; area < nr_vms; area++) {
2412		vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2413		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2414		if (!vas[area] || !vms[area])
2415			goto err_free;
2416	}
2417retry:
2418	spin_lock(&vmap_area_lock);
2419
2420	/* start scanning - we scan from the top, begin with the last area */
2421	area = term_area = last_area;
2422	start = offsets[area];
2423	end = start + sizes[area];
2424
2425	if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2426		base = vmalloc_end - last_end;
2427		goto found;
2428	}
2429	base = pvm_determine_end(&next, &prev, align) - end;
2430
2431	while (true) {
2432		BUG_ON(next && next->va_end <= base + end);
2433		BUG_ON(prev && prev->va_end > base + end);
2434
2435		/*
2436		 * base might have underflowed, add last_end before
2437		 * comparing.
2438		 */
2439		if (base + last_end < vmalloc_start + last_end) {
2440			spin_unlock(&vmap_area_lock);
2441			if (!purged) {
2442				purge_vmap_area_lazy();
2443				purged = true;
2444				goto retry;
2445			}
2446			goto err_free;
2447		}
2448
2449		/*
2450		 * If next overlaps, move base downwards so that it's
2451		 * right below next and then recheck.
2452		 */
2453		if (next && next->va_start < base + end) {
2454			base = pvm_determine_end(&next, &prev, align) - end;
2455			term_area = area;
2456			continue;
2457		}
2458
2459		/*
2460		 * If prev overlaps, shift down next and prev and move
2461		 * base so that it's right below new next and then
2462		 * recheck.
2463		 */
2464		if (prev && prev->va_end > base + start)  {
2465			next = prev;
2466			prev = node_to_va(rb_prev(&next->rb_node));
2467			base = pvm_determine_end(&next, &prev, align) - end;
2468			term_area = area;
2469			continue;
2470		}
2471
2472		/*
2473		 * This area fits, move on to the previous one.  If
2474		 * the previous one is the terminal one, we're done.
2475		 */
2476		area = (area + nr_vms - 1) % nr_vms;
2477		if (area == term_area)
2478			break;
2479		start = offsets[area];
2480		end = start + sizes[area];
2481		pvm_find_next_prev(base + end, &next, &prev);
2482	}
2483found:
2484	/* we've found a fitting base, insert all va's */
2485	for (area = 0; area < nr_vms; area++) {
2486		struct vmap_area *va = vas[area];
2487
2488		va->va_start = base + offsets[area];
2489		va->va_end = va->va_start + sizes[area];
2490		__insert_vmap_area(va);
2491	}
2492
2493	vmap_area_pcpu_hole = base + offsets[last_area];
2494
2495	spin_unlock(&vmap_area_lock);
2496
2497	/* insert all vm's */
2498	for (area = 0; area < nr_vms; area++)
2499		setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2500				 pcpu_get_vm_areas);
2501
2502	kfree(vas);
2503	return vms;
2504
2505err_free:
2506	for (area = 0; area < nr_vms; area++) {
2507		kfree(vas[area]);
2508		kfree(vms[area]);
2509	}
2510err_free2:
2511	kfree(vas);
2512	kfree(vms);
2513	return NULL;
2514}
2515
2516/**
2517 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2518 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2519 * @nr_vms: the number of allocated areas
2520 *
2521 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2522 */
2523void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2524{
2525	int i;
2526
2527	for (i = 0; i < nr_vms; i++)
2528		free_vm_area(vms[i]);
2529	kfree(vms);
2530}
2531#endif	/* CONFIG_SMP */
2532
2533#ifdef CONFIG_PROC_FS
2534static void *s_start(struct seq_file *m, loff_t *pos)
2535	__acquires(&vmap_area_lock)
2536{
2537	loff_t n = *pos;
2538	struct vmap_area *va;
2539
2540	spin_lock(&vmap_area_lock);
2541	va = list_entry((&vmap_area_list)->next, typeof(*va), list);
2542	while (n > 0 && &va->list != &vmap_area_list) {
2543		n--;
2544		va = list_entry(va->list.next, typeof(*va), list);
2545	}
2546	if (!n && &va->list != &vmap_area_list)
2547		return va;
2548
2549	return NULL;
2550
2551}
2552
2553static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2554{
2555	struct vmap_area *va = p, *next;
2556
2557	++*pos;
2558	next = list_entry(va->list.next, typeof(*va), list);
2559	if (&next->list != &vmap_area_list)
2560		return next;
2561
2562	return NULL;
2563}
2564
2565static void s_stop(struct seq_file *m, void *p)
2566	__releases(&vmap_area_lock)
2567{
2568	spin_unlock(&vmap_area_lock);
2569}
2570
2571static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2572{
2573	if (IS_ENABLED(CONFIG_NUMA)) {
2574		unsigned int nr, *counters = m->private;
2575
2576		if (!counters)
2577			return;
2578
2579		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2580		smp_rmb();
2581		if (v->flags & VM_UNINITIALIZED)
2582			return;
 
 
2583
2584		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2585
2586		for (nr = 0; nr < v->nr_pages; nr++)
2587			counters[page_to_nid(v->pages[nr])]++;
2588
2589		for_each_node_state(nr, N_HIGH_MEMORY)
2590			if (counters[nr])
2591				seq_printf(m, " N%u=%u", nr, counters[nr]);
2592	}
2593}
2594
2595static int s_show(struct seq_file *m, void *p)
2596{
2597	struct vmap_area *va = p;
2598	struct vm_struct *v;
2599
2600	/*
2601	 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2602	 * behalf of vmap area is being tear down or vm_map_ram allocation.
2603	 */
2604	if (!(va->flags & VM_VM_AREA))
2605		return 0;
2606
2607	v = va->vm;
2608
2609	seq_printf(m, "0x%pK-0x%pK %7ld",
2610		v->addr, v->addr + v->size, v->size);
2611
2612	if (v->caller)
2613		seq_printf(m, " %pS", v->caller);
2614
2615	if (v->nr_pages)
2616		seq_printf(m, " pages=%d", v->nr_pages);
2617
2618	if (v->phys_addr)
2619		seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2620
2621	if (v->flags & VM_IOREMAP)
2622		seq_printf(m, " ioremap");
2623
2624	if (v->flags & VM_ALLOC)
2625		seq_printf(m, " vmalloc");
2626
2627	if (v->flags & VM_MAP)
2628		seq_printf(m, " vmap");
2629
2630	if (v->flags & VM_USERMAP)
2631		seq_printf(m, " user");
2632
2633	if (v->flags & VM_VPAGES)
2634		seq_printf(m, " vpages");
2635
2636	show_numa_info(m, v);
2637	seq_putc(m, '\n');
2638	return 0;
2639}
2640
2641static const struct seq_operations vmalloc_op = {
2642	.start = s_start,
2643	.next = s_next,
2644	.stop = s_stop,
2645	.show = s_show,
2646};
2647
2648static int vmalloc_open(struct inode *inode, struct file *file)
2649{
2650	unsigned int *ptr = NULL;
2651	int ret;
2652
2653	if (IS_ENABLED(CONFIG_NUMA)) {
2654		ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2655		if (ptr == NULL)
2656			return -ENOMEM;
2657	}
2658	ret = seq_open(file, &vmalloc_op);
2659	if (!ret) {
2660		struct seq_file *m = file->private_data;
2661		m->private = ptr;
2662	} else
2663		kfree(ptr);
2664	return ret;
2665}
2666
2667static const struct file_operations proc_vmalloc_operations = {
2668	.open		= vmalloc_open,
2669	.read		= seq_read,
2670	.llseek		= seq_lseek,
2671	.release	= seq_release_private,
2672};
2673
2674static int __init proc_vmalloc_init(void)
2675{
2676	proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2677	return 0;
2678}
2679module_init(proc_vmalloc_init);
2680
2681void get_vmalloc_info(struct vmalloc_info *vmi)
2682{
2683	struct vmap_area *va;
2684	unsigned long free_area_size;
2685	unsigned long prev_end;
2686
2687	vmi->used = 0;
2688	vmi->largest_chunk = 0;
2689
2690	prev_end = VMALLOC_START;
2691
2692	spin_lock(&vmap_area_lock);
2693
2694	if (list_empty(&vmap_area_list)) {
2695		vmi->largest_chunk = VMALLOC_TOTAL;
2696		goto out;
2697	}
2698
2699	list_for_each_entry(va, &vmap_area_list, list) {
2700		unsigned long addr = va->va_start;
2701
2702		/*
2703		 * Some archs keep another range for modules in vmalloc space
2704		 */
2705		if (addr < VMALLOC_START)
2706			continue;
2707		if (addr >= VMALLOC_END)
2708			break;
2709
2710		if (va->flags & (VM_LAZY_FREE | VM_LAZY_FREEING))
2711			continue;
2712
2713		vmi->used += (va->va_end - va->va_start);
2714
2715		free_area_size = addr - prev_end;
2716		if (vmi->largest_chunk < free_area_size)
2717			vmi->largest_chunk = free_area_size;
2718
2719		prev_end = va->va_end;
2720	}
2721
2722	if (VMALLOC_END - prev_end > vmi->largest_chunk)
2723		vmi->largest_chunk = VMALLOC_END - prev_end;
2724
2725out:
2726	spin_unlock(&vmap_area_lock);
2727}
2728#endif
2729
v4.6
   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/compiler.h>
  31#include <linux/llist.h>
  32#include <linux/bitops.h>
  33
  34#include <asm/uaccess.h>
  35#include <asm/tlbflush.h>
  36#include <asm/shmparam.h>
  37
  38#include "internal.h"
  39
  40struct vfree_deferred {
  41	struct llist_head list;
  42	struct work_struct wq;
  43};
  44static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
  45
  46static void __vunmap(const void *, int);
  47
  48static void free_work(struct work_struct *w)
  49{
  50	struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
  51	struct llist_node *llnode = llist_del_all(&p->list);
  52	while (llnode) {
  53		void *p = llnode;
  54		llnode = llist_next(llnode);
  55		__vunmap(p, 1);
  56	}
  57}
  58
  59/*** Page table manipulation functions ***/
  60
  61static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
  62{
  63	pte_t *pte;
  64
  65	pte = pte_offset_kernel(pmd, addr);
  66	do {
  67		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
  68		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
  69	} while (pte++, addr += PAGE_SIZE, addr != end);
  70}
  71
  72static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
  73{
  74	pmd_t *pmd;
  75	unsigned long next;
  76
  77	pmd = pmd_offset(pud, addr);
  78	do {
  79		next = pmd_addr_end(addr, end);
  80		if (pmd_clear_huge(pmd))
  81			continue;
  82		if (pmd_none_or_clear_bad(pmd))
  83			continue;
  84		vunmap_pte_range(pmd, addr, next);
  85	} while (pmd++, addr = next, addr != end);
  86}
  87
  88static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
  89{
  90	pud_t *pud;
  91	unsigned long next;
  92
  93	pud = pud_offset(pgd, addr);
  94	do {
  95		next = pud_addr_end(addr, end);
  96		if (pud_clear_huge(pud))
  97			continue;
  98		if (pud_none_or_clear_bad(pud))
  99			continue;
 100		vunmap_pmd_range(pud, addr, next);
 101	} while (pud++, addr = next, addr != end);
 102}
 103
 104static void vunmap_page_range(unsigned long addr, unsigned long end)
 105{
 106	pgd_t *pgd;
 107	unsigned long next;
 108
 109	BUG_ON(addr >= end);
 110	pgd = pgd_offset_k(addr);
 111	do {
 112		next = pgd_addr_end(addr, end);
 113		if (pgd_none_or_clear_bad(pgd))
 114			continue;
 115		vunmap_pud_range(pgd, addr, next);
 116	} while (pgd++, addr = next, addr != end);
 117}
 118
 119static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
 120		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 121{
 122	pte_t *pte;
 123
 124	/*
 125	 * nr is a running index into the array which helps higher level
 126	 * callers keep track of where we're up to.
 127	 */
 128
 129	pte = pte_alloc_kernel(pmd, addr);
 130	if (!pte)
 131		return -ENOMEM;
 132	do {
 133		struct page *page = pages[*nr];
 134
 135		if (WARN_ON(!pte_none(*pte)))
 136			return -EBUSY;
 137		if (WARN_ON(!page))
 138			return -ENOMEM;
 139		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 140		(*nr)++;
 141	} while (pte++, addr += PAGE_SIZE, addr != end);
 142	return 0;
 143}
 144
 145static int vmap_pmd_range(pud_t *pud, unsigned long addr,
 146		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 147{
 148	pmd_t *pmd;
 149	unsigned long next;
 150
 151	pmd = pmd_alloc(&init_mm, pud, addr);
 152	if (!pmd)
 153		return -ENOMEM;
 154	do {
 155		next = pmd_addr_end(addr, end);
 156		if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
 157			return -ENOMEM;
 158	} while (pmd++, addr = next, addr != end);
 159	return 0;
 160}
 161
 162static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
 163		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 164{
 165	pud_t *pud;
 166	unsigned long next;
 167
 168	pud = pud_alloc(&init_mm, pgd, addr);
 169	if (!pud)
 170		return -ENOMEM;
 171	do {
 172		next = pud_addr_end(addr, end);
 173		if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
 174			return -ENOMEM;
 175	} while (pud++, addr = next, addr != end);
 176	return 0;
 177}
 178
 179/*
 180 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
 181 * will have pfns corresponding to the "pages" array.
 182 *
 183 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
 184 */
 185static int vmap_page_range_noflush(unsigned long start, unsigned long end,
 186				   pgprot_t prot, struct page **pages)
 187{
 188	pgd_t *pgd;
 189	unsigned long next;
 190	unsigned long addr = start;
 191	int err = 0;
 192	int nr = 0;
 193
 194	BUG_ON(addr >= end);
 195	pgd = pgd_offset_k(addr);
 196	do {
 197		next = pgd_addr_end(addr, end);
 198		err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
 199		if (err)
 200			return err;
 201	} while (pgd++, addr = next, addr != end);
 202
 203	return nr;
 204}
 205
 206static int vmap_page_range(unsigned long start, unsigned long end,
 207			   pgprot_t prot, struct page **pages)
 208{
 209	int ret;
 210
 211	ret = vmap_page_range_noflush(start, end, prot, pages);
 212	flush_cache_vmap(start, end);
 213	return ret;
 214}
 215
 216int is_vmalloc_or_module_addr(const void *x)
 217{
 218	/*
 219	 * ARM, x86-64 and sparc64 put modules in a special place,
 220	 * and fall back on vmalloc() if that fails. Others
 221	 * just put it in the vmalloc space.
 222	 */
 223#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 224	unsigned long addr = (unsigned long)x;
 225	if (addr >= MODULES_VADDR && addr < MODULES_END)
 226		return 1;
 227#endif
 228	return is_vmalloc_addr(x);
 229}
 230
 231/*
 232 * Walk a vmap address to the struct page it maps.
 233 */
 234struct page *vmalloc_to_page(const void *vmalloc_addr)
 235{
 236	unsigned long addr = (unsigned long) vmalloc_addr;
 237	struct page *page = NULL;
 238	pgd_t *pgd = pgd_offset_k(addr);
 239
 240	/*
 241	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 242	 * architectures that do not vmalloc module space
 243	 */
 244	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 245
 246	if (!pgd_none(*pgd)) {
 247		pud_t *pud = pud_offset(pgd, addr);
 248		if (!pud_none(*pud)) {
 249			pmd_t *pmd = pmd_offset(pud, addr);
 250			if (!pmd_none(*pmd)) {
 251				pte_t *ptep, pte;
 252
 253				ptep = pte_offset_map(pmd, addr);
 254				pte = *ptep;
 255				if (pte_present(pte))
 256					page = pte_page(pte);
 257				pte_unmap(ptep);
 258			}
 259		}
 260	}
 261	return page;
 262}
 263EXPORT_SYMBOL(vmalloc_to_page);
 264
 265/*
 266 * Map a vmalloc()-space virtual address to the physical page frame number.
 267 */
 268unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 269{
 270	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 271}
 272EXPORT_SYMBOL(vmalloc_to_pfn);
 273
 274
 275/*** Global kva allocator ***/
 276
 277#define VM_LAZY_FREE	0x01
 278#define VM_LAZY_FREEING	0x02
 279#define VM_VM_AREA	0x04
 280
 281static DEFINE_SPINLOCK(vmap_area_lock);
 282/* Export for kexec only */
 283LIST_HEAD(vmap_area_list);
 284static struct rb_root vmap_area_root = RB_ROOT;
 285
 286/* The vmap cache globals are protected by vmap_area_lock */
 287static struct rb_node *free_vmap_cache;
 288static unsigned long cached_hole_size;
 289static unsigned long cached_vstart;
 290static unsigned long cached_align;
 291
 292static unsigned long vmap_area_pcpu_hole;
 293
 294static struct vmap_area *__find_vmap_area(unsigned long addr)
 295{
 296	struct rb_node *n = vmap_area_root.rb_node;
 297
 298	while (n) {
 299		struct vmap_area *va;
 300
 301		va = rb_entry(n, struct vmap_area, rb_node);
 302		if (addr < va->va_start)
 303			n = n->rb_left;
 304		else if (addr >= va->va_end)
 305			n = n->rb_right;
 306		else
 307			return va;
 308	}
 309
 310	return NULL;
 311}
 312
 313static void __insert_vmap_area(struct vmap_area *va)
 314{
 315	struct rb_node **p = &vmap_area_root.rb_node;
 316	struct rb_node *parent = NULL;
 317	struct rb_node *tmp;
 318
 319	while (*p) {
 320		struct vmap_area *tmp_va;
 321
 322		parent = *p;
 323		tmp_va = rb_entry(parent, struct vmap_area, rb_node);
 324		if (va->va_start < tmp_va->va_end)
 325			p = &(*p)->rb_left;
 326		else if (va->va_end > tmp_va->va_start)
 327			p = &(*p)->rb_right;
 328		else
 329			BUG();
 330	}
 331
 332	rb_link_node(&va->rb_node, parent, p);
 333	rb_insert_color(&va->rb_node, &vmap_area_root);
 334
 335	/* address-sort this list */
 336	tmp = rb_prev(&va->rb_node);
 337	if (tmp) {
 338		struct vmap_area *prev;
 339		prev = rb_entry(tmp, struct vmap_area, rb_node);
 340		list_add_rcu(&va->list, &prev->list);
 341	} else
 342		list_add_rcu(&va->list, &vmap_area_list);
 343}
 344
 345static void purge_vmap_area_lazy(void);
 346
 347/*
 348 * Allocate a region of KVA of the specified size and alignment, within the
 349 * vstart and vend.
 350 */
 351static struct vmap_area *alloc_vmap_area(unsigned long size,
 352				unsigned long align,
 353				unsigned long vstart, unsigned long vend,
 354				int node, gfp_t gfp_mask)
 355{
 356	struct vmap_area *va;
 357	struct rb_node *n;
 358	unsigned long addr;
 359	int purged = 0;
 360	struct vmap_area *first;
 361
 362	BUG_ON(!size);
 363	BUG_ON(offset_in_page(size));
 364	BUG_ON(!is_power_of_2(align));
 365
 366	va = kmalloc_node(sizeof(struct vmap_area),
 367			gfp_mask & GFP_RECLAIM_MASK, node);
 368	if (unlikely(!va))
 369		return ERR_PTR(-ENOMEM);
 370
 371	/*
 372	 * Only scan the relevant parts containing pointers to other objects
 373	 * to avoid false negatives.
 374	 */
 375	kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
 376
 377retry:
 378	spin_lock(&vmap_area_lock);
 379	/*
 380	 * Invalidate cache if we have more permissive parameters.
 381	 * cached_hole_size notes the largest hole noticed _below_
 382	 * the vmap_area cached in free_vmap_cache: if size fits
 383	 * into that hole, we want to scan from vstart to reuse
 384	 * the hole instead of allocating above free_vmap_cache.
 385	 * Note that __free_vmap_area may update free_vmap_cache
 386	 * without updating cached_hole_size or cached_align.
 387	 */
 388	if (!free_vmap_cache ||
 389			size < cached_hole_size ||
 390			vstart < cached_vstart ||
 391			align < cached_align) {
 392nocache:
 393		cached_hole_size = 0;
 394		free_vmap_cache = NULL;
 395	}
 396	/* record if we encounter less permissive parameters */
 397	cached_vstart = vstart;
 398	cached_align = align;
 399
 400	/* find starting point for our search */
 401	if (free_vmap_cache) {
 402		first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 403		addr = ALIGN(first->va_end, align);
 404		if (addr < vstart)
 405			goto nocache;
 406		if (addr + size < addr)
 407			goto overflow;
 408
 409	} else {
 410		addr = ALIGN(vstart, align);
 411		if (addr + size < addr)
 412			goto overflow;
 413
 414		n = vmap_area_root.rb_node;
 415		first = NULL;
 416
 417		while (n) {
 418			struct vmap_area *tmp;
 419			tmp = rb_entry(n, struct vmap_area, rb_node);
 420			if (tmp->va_end >= addr) {
 421				first = tmp;
 422				if (tmp->va_start <= addr)
 423					break;
 424				n = n->rb_left;
 425			} else
 426				n = n->rb_right;
 427		}
 428
 429		if (!first)
 430			goto found;
 431	}
 432
 433	/* from the starting point, walk areas until a suitable hole is found */
 434	while (addr + size > first->va_start && addr + size <= vend) {
 435		if (addr + cached_hole_size < first->va_start)
 436			cached_hole_size = first->va_start - addr;
 437		addr = ALIGN(first->va_end, align);
 438		if (addr + size < addr)
 439			goto overflow;
 440
 441		if (list_is_last(&first->list, &vmap_area_list))
 442			goto found;
 443
 444		first = list_next_entry(first, list);
 
 445	}
 446
 447found:
 448	if (addr + size > vend)
 449		goto overflow;
 450
 451	va->va_start = addr;
 452	va->va_end = addr + size;
 453	va->flags = 0;
 454	__insert_vmap_area(va);
 455	free_vmap_cache = &va->rb_node;
 456	spin_unlock(&vmap_area_lock);
 457
 458	BUG_ON(!IS_ALIGNED(va->va_start, align));
 459	BUG_ON(va->va_start < vstart);
 460	BUG_ON(va->va_end > vend);
 461
 462	return va;
 463
 464overflow:
 465	spin_unlock(&vmap_area_lock);
 466	if (!purged) {
 467		purge_vmap_area_lazy();
 468		purged = 1;
 469		goto retry;
 470	}
 471	if (printk_ratelimit())
 472		pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
 473			size);
 
 474	kfree(va);
 475	return ERR_PTR(-EBUSY);
 476}
 477
 478static void __free_vmap_area(struct vmap_area *va)
 479{
 480	BUG_ON(RB_EMPTY_NODE(&va->rb_node));
 481
 482	if (free_vmap_cache) {
 483		if (va->va_end < cached_vstart) {
 484			free_vmap_cache = NULL;
 485		} else {
 486			struct vmap_area *cache;
 487			cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 488			if (va->va_start <= cache->va_start) {
 489				free_vmap_cache = rb_prev(&va->rb_node);
 490				/*
 491				 * We don't try to update cached_hole_size or
 492				 * cached_align, but it won't go very wrong.
 493				 */
 494			}
 495		}
 496	}
 497	rb_erase(&va->rb_node, &vmap_area_root);
 498	RB_CLEAR_NODE(&va->rb_node);
 499	list_del_rcu(&va->list);
 500
 501	/*
 502	 * Track the highest possible candidate for pcpu area
 503	 * allocation.  Areas outside of vmalloc area can be returned
 504	 * here too, consider only end addresses which fall inside
 505	 * vmalloc area proper.
 506	 */
 507	if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
 508		vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
 509
 510	kfree_rcu(va, rcu_head);
 511}
 512
 513/*
 514 * Free a region of KVA allocated by alloc_vmap_area
 515 */
 516static void free_vmap_area(struct vmap_area *va)
 517{
 518	spin_lock(&vmap_area_lock);
 519	__free_vmap_area(va);
 520	spin_unlock(&vmap_area_lock);
 521}
 522
 523/*
 524 * Clear the pagetable entries of a given vmap_area
 525 */
 526static void unmap_vmap_area(struct vmap_area *va)
 527{
 528	vunmap_page_range(va->va_start, va->va_end);
 529}
 530
 531static void vmap_debug_free_range(unsigned long start, unsigned long end)
 532{
 533	/*
 534	 * Unmap page tables and force a TLB flush immediately if pagealloc
 535	 * debugging is enabled.  This catches use after free bugs similarly to
 536	 * those in linear kernel virtual address space after a page has been
 537	 * freed.
 538	 *
 539	 * All the lazy freeing logic is still retained, in order to minimise
 540	 * intrusiveness of this debugging feature.
 541	 *
 542	 * This is going to be *slow* (linear kernel virtual address debugging
 543	 * doesn't do a broadcast TLB flush so it is a lot faster).
 
 544	 */
 545	if (debug_pagealloc_enabled()) {
 546		vunmap_page_range(start, end);
 547		flush_tlb_kernel_range(start, end);
 548	}
 549}
 550
 551/*
 552 * lazy_max_pages is the maximum amount of virtual address space we gather up
 553 * before attempting to purge with a TLB flush.
 554 *
 555 * There is a tradeoff here: a larger number will cover more kernel page tables
 556 * and take slightly longer to purge, but it will linearly reduce the number of
 557 * global TLB flushes that must be performed. It would seem natural to scale
 558 * this number up linearly with the number of CPUs (because vmapping activity
 559 * could also scale linearly with the number of CPUs), however it is likely
 560 * that in practice, workloads might be constrained in other ways that mean
 561 * vmap activity will not scale linearly with CPUs. Also, I want to be
 562 * conservative and not introduce a big latency on huge systems, so go with
 563 * a less aggressive log scale. It will still be an improvement over the old
 564 * code, and it will be simple to change the scale factor if we find that it
 565 * becomes a problem on bigger systems.
 566 */
 567static unsigned long lazy_max_pages(void)
 568{
 569	unsigned int log;
 570
 571	log = fls(num_online_cpus());
 572
 573	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
 574}
 575
 576static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
 577
 578/* for per-CPU blocks */
 579static void purge_fragmented_blocks_allcpus(void);
 580
 581/*
 582 * called before a call to iounmap() if the caller wants vm_area_struct's
 583 * immediately freed.
 584 */
 585void set_iounmap_nonlazy(void)
 586{
 587	atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
 588}
 589
 590/*
 591 * Purges all lazily-freed vmap areas.
 592 *
 593 * If sync is 0 then don't purge if there is already a purge in progress.
 594 * If force_flush is 1, then flush kernel TLBs between *start and *end even
 595 * if we found no lazy vmap areas to unmap (callers can use this to optimise
 596 * their own TLB flushing).
 597 * Returns with *start = min(*start, lowest purged address)
 598 *              *end = max(*end, highest purged address)
 599 */
 600static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
 601					int sync, int force_flush)
 602{
 603	static DEFINE_SPINLOCK(purge_lock);
 604	LIST_HEAD(valist);
 605	struct vmap_area *va;
 606	struct vmap_area *n_va;
 607	int nr = 0;
 608
 609	/*
 610	 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
 611	 * should not expect such behaviour. This just simplifies locking for
 612	 * the case that isn't actually used at the moment anyway.
 613	 */
 614	if (!sync && !force_flush) {
 615		if (!spin_trylock(&purge_lock))
 616			return;
 617	} else
 618		spin_lock(&purge_lock);
 619
 620	if (sync)
 621		purge_fragmented_blocks_allcpus();
 622
 623	rcu_read_lock();
 624	list_for_each_entry_rcu(va, &vmap_area_list, list) {
 625		if (va->flags & VM_LAZY_FREE) {
 626			if (va->va_start < *start)
 627				*start = va->va_start;
 628			if (va->va_end > *end)
 629				*end = va->va_end;
 630			nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
 631			list_add_tail(&va->purge_list, &valist);
 632			va->flags |= VM_LAZY_FREEING;
 633			va->flags &= ~VM_LAZY_FREE;
 634		}
 635	}
 636	rcu_read_unlock();
 637
 638	if (nr)
 639		atomic_sub(nr, &vmap_lazy_nr);
 640
 641	if (nr || force_flush)
 642		flush_tlb_kernel_range(*start, *end);
 643
 644	if (nr) {
 645		spin_lock(&vmap_area_lock);
 646		list_for_each_entry_safe(va, n_va, &valist, purge_list)
 647			__free_vmap_area(va);
 648		spin_unlock(&vmap_area_lock);
 649	}
 650	spin_unlock(&purge_lock);
 651}
 652
 653/*
 654 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
 655 * is already purging.
 656 */
 657static void try_purge_vmap_area_lazy(void)
 658{
 659	unsigned long start = ULONG_MAX, end = 0;
 660
 661	__purge_vmap_area_lazy(&start, &end, 0, 0);
 662}
 663
 664/*
 665 * Kick off a purge of the outstanding lazy areas.
 666 */
 667static void purge_vmap_area_lazy(void)
 668{
 669	unsigned long start = ULONG_MAX, end = 0;
 670
 671	__purge_vmap_area_lazy(&start, &end, 1, 0);
 672}
 673
 674/*
 675 * Free a vmap area, caller ensuring that the area has been unmapped
 676 * and flush_cache_vunmap had been called for the correct range
 677 * previously.
 678 */
 679static void free_vmap_area_noflush(struct vmap_area *va)
 680{
 681	va->flags |= VM_LAZY_FREE;
 682	atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
 683	if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
 684		try_purge_vmap_area_lazy();
 685}
 686
 687/*
 688 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
 689 * called for the correct range previously.
 690 */
 691static void free_unmap_vmap_area_noflush(struct vmap_area *va)
 692{
 693	unmap_vmap_area(va);
 694	free_vmap_area_noflush(va);
 695}
 696
 697/*
 698 * Free and unmap a vmap area
 699 */
 700static void free_unmap_vmap_area(struct vmap_area *va)
 701{
 702	flush_cache_vunmap(va->va_start, va->va_end);
 703	free_unmap_vmap_area_noflush(va);
 704}
 705
 706static struct vmap_area *find_vmap_area(unsigned long addr)
 707{
 708	struct vmap_area *va;
 709
 710	spin_lock(&vmap_area_lock);
 711	va = __find_vmap_area(addr);
 712	spin_unlock(&vmap_area_lock);
 713
 714	return va;
 715}
 716
 717static void free_unmap_vmap_area_addr(unsigned long addr)
 718{
 719	struct vmap_area *va;
 720
 721	va = find_vmap_area(addr);
 722	BUG_ON(!va);
 723	free_unmap_vmap_area(va);
 724}
 725
 726
 727/*** Per cpu kva allocator ***/
 728
 729/*
 730 * vmap space is limited especially on 32 bit architectures. Ensure there is
 731 * room for at least 16 percpu vmap blocks per CPU.
 732 */
 733/*
 734 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
 735 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
 736 * instead (we just need a rough idea)
 737 */
 738#if BITS_PER_LONG == 32
 739#define VMALLOC_SPACE		(128UL*1024*1024)
 740#else
 741#define VMALLOC_SPACE		(128UL*1024*1024*1024)
 742#endif
 743
 744#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
 745#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
 746#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
 747#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
 748#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
 749#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
 750#define VMAP_BBMAP_BITS		\
 751		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
 752		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
 753			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
 754
 755#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
 756
 757static bool vmap_initialized __read_mostly = false;
 758
 759struct vmap_block_queue {
 760	spinlock_t lock;
 761	struct list_head free;
 762};
 763
 764struct vmap_block {
 765	spinlock_t lock;
 766	struct vmap_area *va;
 767	unsigned long free, dirty;
 768	unsigned long dirty_min, dirty_max; /*< dirty range */
 769	struct list_head free_list;
 770	struct rcu_head rcu_head;
 771	struct list_head purge;
 772};
 773
 774/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
 775static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
 776
 777/*
 778 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
 779 * in the free path. Could get rid of this if we change the API to return a
 780 * "cookie" from alloc, to be passed to free. But no big deal yet.
 781 */
 782static DEFINE_SPINLOCK(vmap_block_tree_lock);
 783static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
 784
 785/*
 786 * We should probably have a fallback mechanism to allocate virtual memory
 787 * out of partially filled vmap blocks. However vmap block sizing should be
 788 * fairly reasonable according to the vmalloc size, so it shouldn't be a
 789 * big problem.
 790 */
 791
 792static unsigned long addr_to_vb_idx(unsigned long addr)
 793{
 794	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
 795	addr /= VMAP_BLOCK_SIZE;
 796	return addr;
 797}
 798
 799static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
 800{
 801	unsigned long addr;
 802
 803	addr = va_start + (pages_off << PAGE_SHIFT);
 804	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
 805	return (void *)addr;
 806}
 807
 808/**
 809 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
 810 *                  block. Of course pages number can't exceed VMAP_BBMAP_BITS
 811 * @order:    how many 2^order pages should be occupied in newly allocated block
 812 * @gfp_mask: flags for the page level allocator
 813 *
 814 * Returns: virtual address in a newly allocated block or ERR_PTR(-errno)
 815 */
 816static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
 817{
 818	struct vmap_block_queue *vbq;
 819	struct vmap_block *vb;
 820	struct vmap_area *va;
 821	unsigned long vb_idx;
 822	int node, err;
 823	void *vaddr;
 824
 825	node = numa_node_id();
 826
 827	vb = kmalloc_node(sizeof(struct vmap_block),
 828			gfp_mask & GFP_RECLAIM_MASK, node);
 829	if (unlikely(!vb))
 830		return ERR_PTR(-ENOMEM);
 831
 832	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
 833					VMALLOC_START, VMALLOC_END,
 834					node, gfp_mask);
 835	if (IS_ERR(va)) {
 836		kfree(vb);
 837		return ERR_CAST(va);
 838	}
 839
 840	err = radix_tree_preload(gfp_mask);
 841	if (unlikely(err)) {
 842		kfree(vb);
 843		free_vmap_area(va);
 844		return ERR_PTR(err);
 845	}
 846
 847	vaddr = vmap_block_vaddr(va->va_start, 0);
 848	spin_lock_init(&vb->lock);
 849	vb->va = va;
 850	/* At least something should be left free */
 851	BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
 852	vb->free = VMAP_BBMAP_BITS - (1UL << order);
 853	vb->dirty = 0;
 854	vb->dirty_min = VMAP_BBMAP_BITS;
 855	vb->dirty_max = 0;
 856	INIT_LIST_HEAD(&vb->free_list);
 857
 858	vb_idx = addr_to_vb_idx(va->va_start);
 859	spin_lock(&vmap_block_tree_lock);
 860	err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
 861	spin_unlock(&vmap_block_tree_lock);
 862	BUG_ON(err);
 863	radix_tree_preload_end();
 864
 865	vbq = &get_cpu_var(vmap_block_queue);
 866	spin_lock(&vbq->lock);
 867	list_add_tail_rcu(&vb->free_list, &vbq->free);
 868	spin_unlock(&vbq->lock);
 869	put_cpu_var(vmap_block_queue);
 870
 871	return vaddr;
 872}
 873
 874static void free_vmap_block(struct vmap_block *vb)
 875{
 876	struct vmap_block *tmp;
 877	unsigned long vb_idx;
 878
 879	vb_idx = addr_to_vb_idx(vb->va->va_start);
 880	spin_lock(&vmap_block_tree_lock);
 881	tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
 882	spin_unlock(&vmap_block_tree_lock);
 883	BUG_ON(tmp != vb);
 884
 885	free_vmap_area_noflush(vb->va);
 886	kfree_rcu(vb, rcu_head);
 887}
 888
 889static void purge_fragmented_blocks(int cpu)
 890{
 891	LIST_HEAD(purge);
 892	struct vmap_block *vb;
 893	struct vmap_block *n_vb;
 894	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
 895
 896	rcu_read_lock();
 897	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 898
 899		if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
 900			continue;
 901
 902		spin_lock(&vb->lock);
 903		if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
 904			vb->free = 0; /* prevent further allocs after releasing lock */
 905			vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
 906			vb->dirty_min = 0;
 907			vb->dirty_max = VMAP_BBMAP_BITS;
 908			spin_lock(&vbq->lock);
 909			list_del_rcu(&vb->free_list);
 910			spin_unlock(&vbq->lock);
 911			spin_unlock(&vb->lock);
 912			list_add_tail(&vb->purge, &purge);
 913		} else
 914			spin_unlock(&vb->lock);
 915	}
 916	rcu_read_unlock();
 917
 918	list_for_each_entry_safe(vb, n_vb, &purge, purge) {
 919		list_del(&vb->purge);
 920		free_vmap_block(vb);
 921	}
 922}
 923
 924static void purge_fragmented_blocks_allcpus(void)
 925{
 926	int cpu;
 927
 928	for_each_possible_cpu(cpu)
 929		purge_fragmented_blocks(cpu);
 930}
 931
 932static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
 933{
 934	struct vmap_block_queue *vbq;
 935	struct vmap_block *vb;
 936	void *vaddr = NULL;
 937	unsigned int order;
 938
 939	BUG_ON(offset_in_page(size));
 940	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 941	if (WARN_ON(size == 0)) {
 942		/*
 943		 * Allocating 0 bytes isn't what caller wants since
 944		 * get_order(0) returns funny result. Just warn and terminate
 945		 * early.
 946		 */
 947		return NULL;
 948	}
 949	order = get_order(size);
 950
 
 951	rcu_read_lock();
 952	vbq = &get_cpu_var(vmap_block_queue);
 953	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 954		unsigned long pages_off;
 955
 956		spin_lock(&vb->lock);
 957		if (vb->free < (1UL << order)) {
 958			spin_unlock(&vb->lock);
 959			continue;
 960		}
 961
 962		pages_off = VMAP_BBMAP_BITS - vb->free;
 963		vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
 
 
 964		vb->free -= 1UL << order;
 965		if (vb->free == 0) {
 966			spin_lock(&vbq->lock);
 967			list_del_rcu(&vb->free_list);
 968			spin_unlock(&vbq->lock);
 969		}
 970
 971		spin_unlock(&vb->lock);
 972		break;
 
 
 973	}
 974
 975	put_cpu_var(vmap_block_queue);
 976	rcu_read_unlock();
 977
 978	/* Allocate new block if nothing was found */
 979	if (!vaddr)
 980		vaddr = new_vmap_block(order, gfp_mask);
 
 
 
 981
 982	return vaddr;
 983}
 984
 985static void vb_free(const void *addr, unsigned long size)
 986{
 987	unsigned long offset;
 988	unsigned long vb_idx;
 989	unsigned int order;
 990	struct vmap_block *vb;
 991
 992	BUG_ON(offset_in_page(size));
 993	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 994
 995	flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
 996
 997	order = get_order(size);
 998
 999	offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1000	offset >>= PAGE_SHIFT;
1001
1002	vb_idx = addr_to_vb_idx((unsigned long)addr);
1003	rcu_read_lock();
1004	vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1005	rcu_read_unlock();
1006	BUG_ON(!vb);
1007
1008	vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1009
1010	spin_lock(&vb->lock);
1011
1012	/* Expand dirty range */
1013	vb->dirty_min = min(vb->dirty_min, offset);
1014	vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1015
1016	vb->dirty += 1UL << order;
1017	if (vb->dirty == VMAP_BBMAP_BITS) {
1018		BUG_ON(vb->free);
1019		spin_unlock(&vb->lock);
1020		free_vmap_block(vb);
1021	} else
1022		spin_unlock(&vb->lock);
1023}
1024
1025/**
1026 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1027 *
1028 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1029 * to amortize TLB flushing overheads. What this means is that any page you
1030 * have now, may, in a former life, have been mapped into kernel virtual
1031 * address by the vmap layer and so there might be some CPUs with TLB entries
1032 * still referencing that page (additional to the regular 1:1 kernel mapping).
1033 *
1034 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1035 * be sure that none of the pages we have control over will have any aliases
1036 * from the vmap layer.
1037 */
1038void vm_unmap_aliases(void)
1039{
1040	unsigned long start = ULONG_MAX, end = 0;
1041	int cpu;
1042	int flush = 0;
1043
1044	if (unlikely(!vmap_initialized))
1045		return;
1046
1047	for_each_possible_cpu(cpu) {
1048		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1049		struct vmap_block *vb;
1050
1051		rcu_read_lock();
1052		list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 
 
1053			spin_lock(&vb->lock);
1054			if (vb->dirty) {
1055				unsigned long va_start = vb->va->va_start;
1056				unsigned long s, e;
1057
1058				s = va_start + (vb->dirty_min << PAGE_SHIFT);
1059				e = va_start + (vb->dirty_max << PAGE_SHIFT);
 
1060
1061				start = min(s, start);
1062				end   = max(e, end);
 
1063
1064				flush = 1;
 
 
 
1065			}
1066			spin_unlock(&vb->lock);
1067		}
1068		rcu_read_unlock();
1069	}
1070
1071	__purge_vmap_area_lazy(&start, &end, 1, flush);
1072}
1073EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1074
1075/**
1076 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1077 * @mem: the pointer returned by vm_map_ram
1078 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1079 */
1080void vm_unmap_ram(const void *mem, unsigned int count)
1081{
1082	unsigned long size = count << PAGE_SHIFT;
1083	unsigned long addr = (unsigned long)mem;
1084
1085	BUG_ON(!addr);
1086	BUG_ON(addr < VMALLOC_START);
1087	BUG_ON(addr > VMALLOC_END);
1088	BUG_ON(!PAGE_ALIGNED(addr));
1089
1090	debug_check_no_locks_freed(mem, size);
1091	vmap_debug_free_range(addr, addr+size);
1092
1093	if (likely(count <= VMAP_MAX_ALLOC))
1094		vb_free(mem, size);
1095	else
1096		free_unmap_vmap_area_addr(addr);
1097}
1098EXPORT_SYMBOL(vm_unmap_ram);
1099
1100/**
1101 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1102 * @pages: an array of pointers to the pages to be mapped
1103 * @count: number of pages
1104 * @node: prefer to allocate data structures on this node
1105 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1106 *
1107 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1108 * faster than vmap so it's good.  But if you mix long-life and short-life
1109 * objects with vm_map_ram(), it could consume lots of address space through
1110 * fragmentation (especially on a 32bit machine).  You could see failures in
1111 * the end.  Please use this function for short-lived objects.
1112 *
1113 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1114 */
1115void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1116{
1117	unsigned long size = count << PAGE_SHIFT;
1118	unsigned long addr;
1119	void *mem;
1120
1121	if (likely(count <= VMAP_MAX_ALLOC)) {
1122		mem = vb_alloc(size, GFP_KERNEL);
1123		if (IS_ERR(mem))
1124			return NULL;
1125		addr = (unsigned long)mem;
1126	} else {
1127		struct vmap_area *va;
1128		va = alloc_vmap_area(size, PAGE_SIZE,
1129				VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1130		if (IS_ERR(va))
1131			return NULL;
1132
1133		addr = va->va_start;
1134		mem = (void *)addr;
1135	}
1136	if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1137		vm_unmap_ram(mem, count);
1138		return NULL;
1139	}
1140	return mem;
1141}
1142EXPORT_SYMBOL(vm_map_ram);
1143
1144static struct vm_struct *vmlist __initdata;
1145/**
1146 * vm_area_add_early - add vmap area early during boot
1147 * @vm: vm_struct to add
1148 *
1149 * This function is used to add fixed kernel vm area to vmlist before
1150 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
1151 * should contain proper values and the other fields should be zero.
1152 *
1153 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1154 */
1155void __init vm_area_add_early(struct vm_struct *vm)
1156{
1157	struct vm_struct *tmp, **p;
1158
1159	BUG_ON(vmap_initialized);
1160	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1161		if (tmp->addr >= vm->addr) {
1162			BUG_ON(tmp->addr < vm->addr + vm->size);
1163			break;
1164		} else
1165			BUG_ON(tmp->addr + tmp->size > vm->addr);
1166	}
1167	vm->next = *p;
1168	*p = vm;
1169}
1170
1171/**
1172 * vm_area_register_early - register vmap area early during boot
1173 * @vm: vm_struct to register
1174 * @align: requested alignment
1175 *
1176 * This function is used to register kernel vm area before
1177 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1178 * proper values on entry and other fields should be zero.  On return,
1179 * vm->addr contains the allocated address.
1180 *
1181 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1182 */
1183void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1184{
1185	static size_t vm_init_off __initdata;
1186	unsigned long addr;
1187
1188	addr = ALIGN(VMALLOC_START + vm_init_off, align);
1189	vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1190
1191	vm->addr = (void *)addr;
1192
1193	vm_area_add_early(vm);
1194}
1195
1196void __init vmalloc_init(void)
1197{
1198	struct vmap_area *va;
1199	struct vm_struct *tmp;
1200	int i;
1201
1202	for_each_possible_cpu(i) {
1203		struct vmap_block_queue *vbq;
1204		struct vfree_deferred *p;
1205
1206		vbq = &per_cpu(vmap_block_queue, i);
1207		spin_lock_init(&vbq->lock);
1208		INIT_LIST_HEAD(&vbq->free);
1209		p = &per_cpu(vfree_deferred, i);
1210		init_llist_head(&p->list);
1211		INIT_WORK(&p->wq, free_work);
1212	}
1213
1214	/* Import existing vmlist entries. */
1215	for (tmp = vmlist; tmp; tmp = tmp->next) {
1216		va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1217		va->flags = VM_VM_AREA;
1218		va->va_start = (unsigned long)tmp->addr;
1219		va->va_end = va->va_start + tmp->size;
1220		va->vm = tmp;
1221		__insert_vmap_area(va);
1222	}
1223
1224	vmap_area_pcpu_hole = VMALLOC_END;
1225
1226	vmap_initialized = true;
1227}
1228
1229/**
1230 * map_kernel_range_noflush - map kernel VM area with the specified pages
1231 * @addr: start of the VM area to map
1232 * @size: size of the VM area to map
1233 * @prot: page protection flags to use
1234 * @pages: pages to map
1235 *
1236 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1237 * specify should have been allocated using get_vm_area() and its
1238 * friends.
1239 *
1240 * NOTE:
1241 * This function does NOT do any cache flushing.  The caller is
1242 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1243 * before calling this function.
1244 *
1245 * RETURNS:
1246 * The number of pages mapped on success, -errno on failure.
1247 */
1248int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1249			     pgprot_t prot, struct page **pages)
1250{
1251	return vmap_page_range_noflush(addr, addr + size, prot, pages);
1252}
1253
1254/**
1255 * unmap_kernel_range_noflush - unmap kernel VM area
1256 * @addr: start of the VM area to unmap
1257 * @size: size of the VM area to unmap
1258 *
1259 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1260 * specify should have been allocated using get_vm_area() and its
1261 * friends.
1262 *
1263 * NOTE:
1264 * This function does NOT do any cache flushing.  The caller is
1265 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1266 * before calling this function and flush_tlb_kernel_range() after.
1267 */
1268void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1269{
1270	vunmap_page_range(addr, addr + size);
1271}
1272EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1273
1274/**
1275 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1276 * @addr: start of the VM area to unmap
1277 * @size: size of the VM area to unmap
1278 *
1279 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1280 * the unmapping and tlb after.
1281 */
1282void unmap_kernel_range(unsigned long addr, unsigned long size)
1283{
1284	unsigned long end = addr + size;
1285
1286	flush_cache_vunmap(addr, end);
1287	vunmap_page_range(addr, end);
1288	flush_tlb_kernel_range(addr, end);
1289}
1290EXPORT_SYMBOL_GPL(unmap_kernel_range);
1291
1292int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
1293{
1294	unsigned long addr = (unsigned long)area->addr;
1295	unsigned long end = addr + get_vm_area_size(area);
1296	int err;
1297
1298	err = vmap_page_range(addr, end, prot, pages);
 
 
 
 
1299
1300	return err > 0 ? 0 : err;
1301}
1302EXPORT_SYMBOL_GPL(map_vm_area);
1303
1304static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1305			      unsigned long flags, const void *caller)
1306{
1307	spin_lock(&vmap_area_lock);
1308	vm->flags = flags;
1309	vm->addr = (void *)va->va_start;
1310	vm->size = va->va_end - va->va_start;
1311	vm->caller = caller;
1312	va->vm = vm;
1313	va->flags |= VM_VM_AREA;
1314	spin_unlock(&vmap_area_lock);
1315}
1316
1317static void clear_vm_uninitialized_flag(struct vm_struct *vm)
1318{
1319	/*
1320	 * Before removing VM_UNINITIALIZED,
1321	 * we should make sure that vm has proper values.
1322	 * Pair with smp_rmb() in show_numa_info().
1323	 */
1324	smp_wmb();
1325	vm->flags &= ~VM_UNINITIALIZED;
1326}
1327
1328static struct vm_struct *__get_vm_area_node(unsigned long size,
1329		unsigned long align, unsigned long flags, unsigned long start,
1330		unsigned long end, int node, gfp_t gfp_mask, const void *caller)
1331{
1332	struct vmap_area *va;
1333	struct vm_struct *area;
1334
1335	BUG_ON(in_interrupt());
1336	if (flags & VM_IOREMAP)
1337		align = 1ul << clamp_t(int, fls_long(size),
1338				       PAGE_SHIFT, IOREMAP_MAX_ORDER);
1339
1340	size = PAGE_ALIGN(size);
1341	if (unlikely(!size))
1342		return NULL;
1343
1344	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1345	if (unlikely(!area))
1346		return NULL;
1347
1348	if (!(flags & VM_NO_GUARD))
1349		size += PAGE_SIZE;
 
 
1350
1351	va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1352	if (IS_ERR(va)) {
1353		kfree(area);
1354		return NULL;
1355	}
1356
1357	setup_vmalloc_vm(area, va, flags, caller);
1358
1359	return area;
1360}
1361
1362struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1363				unsigned long start, unsigned long end)
1364{
1365	return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1366				  GFP_KERNEL, __builtin_return_address(0));
1367}
1368EXPORT_SYMBOL_GPL(__get_vm_area);
1369
1370struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1371				       unsigned long start, unsigned long end,
1372				       const void *caller)
1373{
1374	return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
1375				  GFP_KERNEL, caller);
1376}
1377
1378/**
1379 *	get_vm_area  -  reserve a contiguous kernel virtual area
1380 *	@size:		size of the area
1381 *	@flags:		%VM_IOREMAP for I/O mappings or VM_ALLOC
1382 *
1383 *	Search an area of @size in the kernel virtual mapping area,
1384 *	and reserved it for out purposes.  Returns the area descriptor
1385 *	on success or %NULL on failure.
1386 */
1387struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1388{
1389	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1390				  NUMA_NO_NODE, GFP_KERNEL,
1391				  __builtin_return_address(0));
1392}
1393
1394struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1395				const void *caller)
1396{
1397	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1398				  NUMA_NO_NODE, GFP_KERNEL, caller);
1399}
1400
1401/**
1402 *	find_vm_area  -  find a continuous kernel virtual area
1403 *	@addr:		base address
1404 *
1405 *	Search for the kernel VM area starting at @addr, and return it.
1406 *	It is up to the caller to do all required locking to keep the returned
1407 *	pointer valid.
1408 */
1409struct vm_struct *find_vm_area(const void *addr)
1410{
1411	struct vmap_area *va;
1412
1413	va = find_vmap_area((unsigned long)addr);
1414	if (va && va->flags & VM_VM_AREA)
1415		return va->vm;
1416
1417	return NULL;
1418}
1419
1420/**
1421 *	remove_vm_area  -  find and remove a continuous kernel virtual area
1422 *	@addr:		base address
1423 *
1424 *	Search for the kernel VM area starting at @addr, and remove it.
1425 *	This function returns the found VM area, but using it is NOT safe
1426 *	on SMP machines, except for its size or flags.
1427 */
1428struct vm_struct *remove_vm_area(const void *addr)
1429{
1430	struct vmap_area *va;
1431
1432	va = find_vmap_area((unsigned long)addr);
1433	if (va && va->flags & VM_VM_AREA) {
1434		struct vm_struct *vm = va->vm;
1435
1436		spin_lock(&vmap_area_lock);
1437		va->vm = NULL;
1438		va->flags &= ~VM_VM_AREA;
1439		spin_unlock(&vmap_area_lock);
1440
1441		vmap_debug_free_range(va->va_start, va->va_end);
1442		kasan_free_shadow(vm);
1443		free_unmap_vmap_area(va);
 
1444
1445		return vm;
1446	}
1447	return NULL;
1448}
1449
1450static void __vunmap(const void *addr, int deallocate_pages)
1451{
1452	struct vm_struct *area;
1453
1454	if (!addr)
1455		return;
1456
1457	if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
1458			addr))
1459		return;
1460
1461	area = remove_vm_area(addr);
1462	if (unlikely(!area)) {
1463		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1464				addr);
1465		return;
1466	}
1467
1468	debug_check_no_locks_freed(addr, get_vm_area_size(area));
1469	debug_check_no_obj_freed(addr, get_vm_area_size(area));
1470
1471	if (deallocate_pages) {
1472		int i;
1473
1474		for (i = 0; i < area->nr_pages; i++) {
1475			struct page *page = area->pages[i];
1476
1477			BUG_ON(!page);
1478			__free_kmem_pages(page, 0);
1479		}
1480
1481		kvfree(area->pages);
 
 
 
1482	}
1483
1484	kfree(area);
1485	return;
1486}
1487 
1488/**
1489 *	vfree  -  release memory allocated by vmalloc()
1490 *	@addr:		memory base address
1491 *
1492 *	Free the virtually continuous memory area starting at @addr, as
1493 *	obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1494 *	NULL, no operation is performed.
1495 *
1496 *	Must not be called in NMI context (strictly speaking, only if we don't
1497 *	have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
1498 *	conventions for vfree() arch-depenedent would be a really bad idea)
1499 *
1500 *	NOTE: assumes that the object at *addr has a size >= sizeof(llist_node)
1501 */
1502void vfree(const void *addr)
1503{
1504	BUG_ON(in_nmi());
1505
1506	kmemleak_free(addr);
1507
1508	if (!addr)
1509		return;
1510	if (unlikely(in_interrupt())) {
1511		struct vfree_deferred *p = this_cpu_ptr(&vfree_deferred);
1512		if (llist_add((struct llist_node *)addr, &p->list))
1513			schedule_work(&p->wq);
1514	} else
1515		__vunmap(addr, 1);
1516}
1517EXPORT_SYMBOL(vfree);
1518
1519/**
1520 *	vunmap  -  release virtual mapping obtained by vmap()
1521 *	@addr:		memory base address
1522 *
1523 *	Free the virtually contiguous memory area starting at @addr,
1524 *	which was created from the page array passed to vmap().
1525 *
1526 *	Must not be called in interrupt context.
1527 */
1528void vunmap(const void *addr)
1529{
1530	BUG_ON(in_interrupt());
1531	might_sleep();
1532	if (addr)
1533		__vunmap(addr, 0);
1534}
1535EXPORT_SYMBOL(vunmap);
1536
1537/**
1538 *	vmap  -  map an array of pages into virtually contiguous space
1539 *	@pages:		array of page pointers
1540 *	@count:		number of pages to map
1541 *	@flags:		vm_area->flags
1542 *	@prot:		page protection for the mapping
1543 *
1544 *	Maps @count pages from @pages into contiguous kernel virtual
1545 *	space.
1546 */
1547void *vmap(struct page **pages, unsigned int count,
1548		unsigned long flags, pgprot_t prot)
1549{
1550	struct vm_struct *area;
1551
1552	might_sleep();
1553
1554	if (count > totalram_pages)
1555		return NULL;
1556
1557	area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1558					__builtin_return_address(0));
1559	if (!area)
1560		return NULL;
1561
1562	if (map_vm_area(area, prot, pages)) {
1563		vunmap(area->addr);
1564		return NULL;
1565	}
1566
1567	return area->addr;
1568}
1569EXPORT_SYMBOL(vmap);
1570
1571static void *__vmalloc_node(unsigned long size, unsigned long align,
1572			    gfp_t gfp_mask, pgprot_t prot,
1573			    int node, const void *caller);
1574static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1575				 pgprot_t prot, int node)
1576{
1577	const int order = 0;
1578	struct page **pages;
1579	unsigned int nr_pages, array_size, i;
1580	const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
1581	const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
1582
1583	nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
1584	array_size = (nr_pages * sizeof(struct page *));
1585
1586	area->nr_pages = nr_pages;
1587	/* Please note that the recursion is strictly bounded. */
1588	if (array_size > PAGE_SIZE) {
1589		pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1590				PAGE_KERNEL, node, area->caller);
 
1591	} else {
1592		pages = kmalloc_node(array_size, nested_gfp, node);
1593	}
1594	area->pages = pages;
1595	if (!area->pages) {
1596		remove_vm_area(area->addr);
1597		kfree(area);
1598		return NULL;
1599	}
1600
1601	for (i = 0; i < area->nr_pages; i++) {
1602		struct page *page;
 
1603
1604		if (node == NUMA_NO_NODE)
1605			page = alloc_kmem_pages(alloc_mask, order);
1606		else
1607			page = alloc_kmem_pages_node(node, alloc_mask, order);
1608
1609		if (unlikely(!page)) {
1610			/* Successfully allocated i pages, free them in __vunmap() */
1611			area->nr_pages = i;
1612			goto fail;
1613		}
1614		area->pages[i] = page;
1615		if (gfpflags_allow_blocking(gfp_mask))
1616			cond_resched();
1617	}
1618
1619	if (map_vm_area(area, prot, pages))
1620		goto fail;
1621	return area->addr;
1622
1623fail:
1624	warn_alloc_failed(gfp_mask, order,
1625			  "vmalloc: allocation failure, allocated %ld of %ld bytes\n",
1626			  (area->nr_pages*PAGE_SIZE), area->size);
1627	vfree(area->addr);
1628	return NULL;
1629}
1630
1631/**
1632 *	__vmalloc_node_range  -  allocate virtually contiguous memory
1633 *	@size:		allocation size
1634 *	@align:		desired alignment
1635 *	@start:		vm area range start
1636 *	@end:		vm area range end
1637 *	@gfp_mask:	flags for the page level allocator
1638 *	@prot:		protection mask for the allocated pages
1639 *	@vm_flags:	additional vm area flags (e.g. %VM_NO_GUARD)
1640 *	@node:		node to use for allocation or NUMA_NO_NODE
1641 *	@caller:	caller's return address
1642 *
1643 *	Allocate enough pages to cover @size from the page level
1644 *	allocator with @gfp_mask flags.  Map them into contiguous
1645 *	kernel virtual space, using a pagetable protection of @prot.
1646 */
1647void *__vmalloc_node_range(unsigned long size, unsigned long align,
1648			unsigned long start, unsigned long end, gfp_t gfp_mask,
1649			pgprot_t prot, unsigned long vm_flags, int node,
1650			const void *caller)
1651{
1652	struct vm_struct *area;
1653	void *addr;
1654	unsigned long real_size = size;
1655
1656	size = PAGE_ALIGN(size);
1657	if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1658		goto fail;
1659
1660	area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
1661				vm_flags, start, end, node, gfp_mask, caller);
1662	if (!area)
1663		goto fail;
1664
1665	addr = __vmalloc_area_node(area, gfp_mask, prot, node);
1666	if (!addr)
1667		return NULL;
1668
1669	/*
1670	 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
1671	 * flag. It means that vm_struct is not fully initialized.
1672	 * Now, it is fully initialized, so remove this flag here.
1673	 */
1674	clear_vm_uninitialized_flag(area);
1675
1676	/*
1677	 * A ref_count = 2 is needed because vm_struct allocated in
1678	 * __get_vm_area_node() contains a reference to the virtual address of
1679	 * the vmalloc'ed block.
1680	 */
1681	kmemleak_alloc(addr, real_size, 2, gfp_mask);
1682
1683	return addr;
1684
1685fail:
1686	warn_alloc_failed(gfp_mask, 0,
1687			  "vmalloc: allocation failure: %lu bytes\n",
1688			  real_size);
1689	return NULL;
1690}
1691
1692/**
1693 *	__vmalloc_node  -  allocate virtually contiguous memory
1694 *	@size:		allocation size
1695 *	@align:		desired alignment
1696 *	@gfp_mask:	flags for the page level allocator
1697 *	@prot:		protection mask for the allocated pages
1698 *	@node:		node to use for allocation or NUMA_NO_NODE
1699 *	@caller:	caller's return address
1700 *
1701 *	Allocate enough pages to cover @size from the page level
1702 *	allocator with @gfp_mask flags.  Map them into contiguous
1703 *	kernel virtual space, using a pagetable protection of @prot.
1704 */
1705static void *__vmalloc_node(unsigned long size, unsigned long align,
1706			    gfp_t gfp_mask, pgprot_t prot,
1707			    int node, const void *caller)
1708{
1709	return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1710				gfp_mask, prot, 0, node, caller);
1711}
1712
1713void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1714{
1715	return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
1716				__builtin_return_address(0));
1717}
1718EXPORT_SYMBOL(__vmalloc);
1719
1720static inline void *__vmalloc_node_flags(unsigned long size,
1721					int node, gfp_t flags)
1722{
1723	return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1724					node, __builtin_return_address(0));
1725}
1726
1727/**
1728 *	vmalloc  -  allocate virtually contiguous memory
1729 *	@size:		allocation size
1730 *	Allocate enough pages to cover @size from the page level
1731 *	allocator and map them into contiguous kernel virtual space.
1732 *
1733 *	For tight control over page level allocator and protection flags
1734 *	use __vmalloc() instead.
1735 */
1736void *vmalloc(unsigned long size)
1737{
1738	return __vmalloc_node_flags(size, NUMA_NO_NODE,
1739				    GFP_KERNEL | __GFP_HIGHMEM);
1740}
1741EXPORT_SYMBOL(vmalloc);
1742
1743/**
1744 *	vzalloc - allocate virtually contiguous memory with zero fill
1745 *	@size:	allocation size
1746 *	Allocate enough pages to cover @size from the page level
1747 *	allocator and map them into contiguous kernel virtual space.
1748 *	The memory allocated is set to zero.
1749 *
1750 *	For tight control over page level allocator and protection flags
1751 *	use __vmalloc() instead.
1752 */
1753void *vzalloc(unsigned long size)
1754{
1755	return __vmalloc_node_flags(size, NUMA_NO_NODE,
1756				GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1757}
1758EXPORT_SYMBOL(vzalloc);
1759
1760/**
1761 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1762 * @size: allocation size
1763 *
1764 * The resulting memory area is zeroed so it can be mapped to userspace
1765 * without leaking data.
1766 */
1767void *vmalloc_user(unsigned long size)
1768{
1769	struct vm_struct *area;
1770	void *ret;
1771
1772	ret = __vmalloc_node(size, SHMLBA,
1773			     GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1774			     PAGE_KERNEL, NUMA_NO_NODE,
1775			     __builtin_return_address(0));
1776	if (ret) {
1777		area = find_vm_area(ret);
1778		area->flags |= VM_USERMAP;
1779	}
1780	return ret;
1781}
1782EXPORT_SYMBOL(vmalloc_user);
1783
1784/**
1785 *	vmalloc_node  -  allocate memory on a specific node
1786 *	@size:		allocation size
1787 *	@node:		numa node
1788 *
1789 *	Allocate enough pages to cover @size from the page level
1790 *	allocator and map them into contiguous kernel virtual space.
1791 *
1792 *	For tight control over page level allocator and protection flags
1793 *	use __vmalloc() instead.
1794 */
1795void *vmalloc_node(unsigned long size, int node)
1796{
1797	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1798					node, __builtin_return_address(0));
1799}
1800EXPORT_SYMBOL(vmalloc_node);
1801
1802/**
1803 * vzalloc_node - allocate memory on a specific node with zero fill
1804 * @size:	allocation size
1805 * @node:	numa node
1806 *
1807 * Allocate enough pages to cover @size from the page level
1808 * allocator and map them into contiguous kernel virtual space.
1809 * The memory allocated is set to zero.
1810 *
1811 * For tight control over page level allocator and protection flags
1812 * use __vmalloc_node() instead.
1813 */
1814void *vzalloc_node(unsigned long size, int node)
1815{
1816	return __vmalloc_node_flags(size, node,
1817			 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1818}
1819EXPORT_SYMBOL(vzalloc_node);
1820
1821#ifndef PAGE_KERNEL_EXEC
1822# define PAGE_KERNEL_EXEC PAGE_KERNEL
1823#endif
1824
1825/**
1826 *	vmalloc_exec  -  allocate virtually contiguous, executable memory
1827 *	@size:		allocation size
1828 *
1829 *	Kernel-internal function to allocate enough pages to cover @size
1830 *	the page level allocator and map them into contiguous and
1831 *	executable kernel virtual space.
1832 *
1833 *	For tight control over page level allocator and protection flags
1834 *	use __vmalloc() instead.
1835 */
1836
1837void *vmalloc_exec(unsigned long size)
1838{
1839	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1840			      NUMA_NO_NODE, __builtin_return_address(0));
1841}
1842
1843#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1844#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1845#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1846#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1847#else
1848#define GFP_VMALLOC32 GFP_KERNEL
1849#endif
1850
1851/**
1852 *	vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1853 *	@size:		allocation size
1854 *
1855 *	Allocate enough 32bit PA addressable pages to cover @size from the
1856 *	page level allocator and map them into contiguous kernel virtual space.
1857 */
1858void *vmalloc_32(unsigned long size)
1859{
1860	return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1861			      NUMA_NO_NODE, __builtin_return_address(0));
1862}
1863EXPORT_SYMBOL(vmalloc_32);
1864
1865/**
1866 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1867 *	@size:		allocation size
1868 *
1869 * The resulting memory area is 32bit addressable and zeroed so it can be
1870 * mapped to userspace without leaking data.
1871 */
1872void *vmalloc_32_user(unsigned long size)
1873{
1874	struct vm_struct *area;
1875	void *ret;
1876
1877	ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1878			     NUMA_NO_NODE, __builtin_return_address(0));
1879	if (ret) {
1880		area = find_vm_area(ret);
1881		area->flags |= VM_USERMAP;
1882	}
1883	return ret;
1884}
1885EXPORT_SYMBOL(vmalloc_32_user);
1886
1887/*
1888 * small helper routine , copy contents to buf from addr.
1889 * If the page is not present, fill zero.
1890 */
1891
1892static int aligned_vread(char *buf, char *addr, unsigned long count)
1893{
1894	struct page *p;
1895	int copied = 0;
1896
1897	while (count) {
1898		unsigned long offset, length;
1899
1900		offset = offset_in_page(addr);
1901		length = PAGE_SIZE - offset;
1902		if (length > count)
1903			length = count;
1904		p = vmalloc_to_page(addr);
1905		/*
1906		 * To do safe access to this _mapped_ area, we need
1907		 * lock. But adding lock here means that we need to add
1908		 * overhead of vmalloc()/vfree() calles for this _debug_
1909		 * interface, rarely used. Instead of that, we'll use
1910		 * kmap() and get small overhead in this access function.
1911		 */
1912		if (p) {
1913			/*
1914			 * we can expect USER0 is not used (see vread/vwrite's
1915			 * function description)
1916			 */
1917			void *map = kmap_atomic(p);
1918			memcpy(buf, map + offset, length);
1919			kunmap_atomic(map);
1920		} else
1921			memset(buf, 0, length);
1922
1923		addr += length;
1924		buf += length;
1925		copied += length;
1926		count -= length;
1927	}
1928	return copied;
1929}
1930
1931static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1932{
1933	struct page *p;
1934	int copied = 0;
1935
1936	while (count) {
1937		unsigned long offset, length;
1938
1939		offset = offset_in_page(addr);
1940		length = PAGE_SIZE - offset;
1941		if (length > count)
1942			length = count;
1943		p = vmalloc_to_page(addr);
1944		/*
1945		 * To do safe access to this _mapped_ area, we need
1946		 * lock. But adding lock here means that we need to add
1947		 * overhead of vmalloc()/vfree() calles for this _debug_
1948		 * interface, rarely used. Instead of that, we'll use
1949		 * kmap() and get small overhead in this access function.
1950		 */
1951		if (p) {
1952			/*
1953			 * we can expect USER0 is not used (see vread/vwrite's
1954			 * function description)
1955			 */
1956			void *map = kmap_atomic(p);
1957			memcpy(map + offset, buf, length);
1958			kunmap_atomic(map);
1959		}
1960		addr += length;
1961		buf += length;
1962		copied += length;
1963		count -= length;
1964	}
1965	return copied;
1966}
1967
1968/**
1969 *	vread() -  read vmalloc area in a safe way.
1970 *	@buf:		buffer for reading data
1971 *	@addr:		vm address.
1972 *	@count:		number of bytes to be read.
1973 *
1974 *	Returns # of bytes which addr and buf should be increased.
1975 *	(same number to @count). Returns 0 if [addr...addr+count) doesn't
1976 *	includes any intersect with alive vmalloc area.
1977 *
1978 *	This function checks that addr is a valid vmalloc'ed area, and
1979 *	copy data from that area to a given buffer. If the given memory range
1980 *	of [addr...addr+count) includes some valid address, data is copied to
1981 *	proper area of @buf. If there are memory holes, they'll be zero-filled.
1982 *	IOREMAP area is treated as memory hole and no copy is done.
1983 *
1984 *	If [addr...addr+count) doesn't includes any intersects with alive
1985 *	vm_struct area, returns 0. @buf should be kernel's buffer.
1986 *
1987 *	Note: In usual ops, vread() is never necessary because the caller
1988 *	should know vmalloc() area is valid and can use memcpy().
1989 *	This is for routines which have to access vmalloc area without
1990 *	any informaion, as /dev/kmem.
1991 *
1992 */
1993
1994long vread(char *buf, char *addr, unsigned long count)
1995{
1996	struct vmap_area *va;
1997	struct vm_struct *vm;
1998	char *vaddr, *buf_start = buf;
1999	unsigned long buflen = count;
2000	unsigned long n;
2001
2002	/* Don't allow overflow */
2003	if ((unsigned long) addr + count < count)
2004		count = -(unsigned long) addr;
2005
2006	spin_lock(&vmap_area_lock);
2007	list_for_each_entry(va, &vmap_area_list, list) {
2008		if (!count)
2009			break;
2010
2011		if (!(va->flags & VM_VM_AREA))
2012			continue;
2013
2014		vm = va->vm;
2015		vaddr = (char *) vm->addr;
2016		if (addr >= vaddr + get_vm_area_size(vm))
2017			continue;
2018		while (addr < vaddr) {
2019			if (count == 0)
2020				goto finished;
2021			*buf = '\0';
2022			buf++;
2023			addr++;
2024			count--;
2025		}
2026		n = vaddr + get_vm_area_size(vm) - addr;
2027		if (n > count)
2028			n = count;
2029		if (!(vm->flags & VM_IOREMAP))
2030			aligned_vread(buf, addr, n);
2031		else /* IOREMAP area is treated as memory hole */
2032			memset(buf, 0, n);
2033		buf += n;
2034		addr += n;
2035		count -= n;
2036	}
2037finished:
2038	spin_unlock(&vmap_area_lock);
2039
2040	if (buf == buf_start)
2041		return 0;
2042	/* zero-fill memory holes */
2043	if (buf != buf_start + buflen)
2044		memset(buf, 0, buflen - (buf - buf_start));
2045
2046	return buflen;
2047}
2048
2049/**
2050 *	vwrite() -  write vmalloc area in a safe way.
2051 *	@buf:		buffer for source data
2052 *	@addr:		vm address.
2053 *	@count:		number of bytes to be read.
2054 *
2055 *	Returns # of bytes which addr and buf should be incresed.
2056 *	(same number to @count).
2057 *	If [addr...addr+count) doesn't includes any intersect with valid
2058 *	vmalloc area, returns 0.
2059 *
2060 *	This function checks that addr is a valid vmalloc'ed area, and
2061 *	copy data from a buffer to the given addr. If specified range of
2062 *	[addr...addr+count) includes some valid address, data is copied from
2063 *	proper area of @buf. If there are memory holes, no copy to hole.
2064 *	IOREMAP area is treated as memory hole and no copy is done.
2065 *
2066 *	If [addr...addr+count) doesn't includes any intersects with alive
2067 *	vm_struct area, returns 0. @buf should be kernel's buffer.
2068 *
2069 *	Note: In usual ops, vwrite() is never necessary because the caller
2070 *	should know vmalloc() area is valid and can use memcpy().
2071 *	This is for routines which have to access vmalloc area without
2072 *	any informaion, as /dev/kmem.
2073 */
2074
2075long vwrite(char *buf, char *addr, unsigned long count)
2076{
2077	struct vmap_area *va;
2078	struct vm_struct *vm;
2079	char *vaddr;
2080	unsigned long n, buflen;
2081	int copied = 0;
2082
2083	/* Don't allow overflow */
2084	if ((unsigned long) addr + count < count)
2085		count = -(unsigned long) addr;
2086	buflen = count;
2087
2088	spin_lock(&vmap_area_lock);
2089	list_for_each_entry(va, &vmap_area_list, list) {
2090		if (!count)
2091			break;
2092
2093		if (!(va->flags & VM_VM_AREA))
2094			continue;
2095
2096		vm = va->vm;
2097		vaddr = (char *) vm->addr;
2098		if (addr >= vaddr + get_vm_area_size(vm))
2099			continue;
2100		while (addr < vaddr) {
2101			if (count == 0)
2102				goto finished;
2103			buf++;
2104			addr++;
2105			count--;
2106		}
2107		n = vaddr + get_vm_area_size(vm) - addr;
2108		if (n > count)
2109			n = count;
2110		if (!(vm->flags & VM_IOREMAP)) {
2111			aligned_vwrite(buf, addr, n);
2112			copied++;
2113		}
2114		buf += n;
2115		addr += n;
2116		count -= n;
2117	}
2118finished:
2119	spin_unlock(&vmap_area_lock);
2120	if (!copied)
2121		return 0;
2122	return buflen;
2123}
2124
2125/**
2126 *	remap_vmalloc_range_partial  -  map vmalloc pages to userspace
2127 *	@vma:		vma to cover
2128 *	@uaddr:		target user address to start at
2129 *	@kaddr:		virtual address of vmalloc kernel memory
2130 *	@size:		size of map area
2131 *
2132 *	Returns:	0 for success, -Exxx on failure
2133 *
2134 *	This function checks that @kaddr is a valid vmalloc'ed area,
2135 *	and that it is big enough to cover the range starting at
2136 *	@uaddr in @vma. Will return failure if that criteria isn't
2137 *	met.
2138 *
2139 *	Similar to remap_pfn_range() (see mm/memory.c)
2140 */
2141int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2142				void *kaddr, unsigned long size)
2143{
2144	struct vm_struct *area;
2145
2146	size = PAGE_ALIGN(size);
2147
2148	if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2149		return -EINVAL;
2150
2151	area = find_vm_area(kaddr);
2152	if (!area)
2153		return -EINVAL;
2154
2155	if (!(area->flags & VM_USERMAP))
2156		return -EINVAL;
2157
2158	if (kaddr + size > area->addr + area->size)
2159		return -EINVAL;
2160
2161	do {
2162		struct page *page = vmalloc_to_page(kaddr);
2163		int ret;
2164
2165		ret = vm_insert_page(vma, uaddr, page);
2166		if (ret)
2167			return ret;
2168
2169		uaddr += PAGE_SIZE;
2170		kaddr += PAGE_SIZE;
2171		size -= PAGE_SIZE;
2172	} while (size > 0);
2173
2174	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
2175
2176	return 0;
2177}
2178EXPORT_SYMBOL(remap_vmalloc_range_partial);
2179
2180/**
2181 *	remap_vmalloc_range  -  map vmalloc pages to userspace
2182 *	@vma:		vma to cover (map full range of vma)
2183 *	@addr:		vmalloc memory
2184 *	@pgoff:		number of pages into addr before first page to map
2185 *
2186 *	Returns:	0 for success, -Exxx on failure
2187 *
2188 *	This function checks that addr is a valid vmalloc'ed area, and
2189 *	that it is big enough to cover the vma. Will return failure if
2190 *	that criteria isn't met.
2191 *
2192 *	Similar to remap_pfn_range() (see mm/memory.c)
2193 */
2194int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2195						unsigned long pgoff)
2196{
2197	return remap_vmalloc_range_partial(vma, vma->vm_start,
2198					   addr + (pgoff << PAGE_SHIFT),
2199					   vma->vm_end - vma->vm_start);
2200}
2201EXPORT_SYMBOL(remap_vmalloc_range);
2202
2203/*
2204 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2205 * have one.
2206 */
2207void __weak vmalloc_sync_all(void)
2208{
2209}
2210
2211
2212static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2213{
2214	pte_t ***p = data;
2215
2216	if (p) {
2217		*(*p) = pte;
2218		(*p)++;
2219	}
2220	return 0;
2221}
2222
2223/**
2224 *	alloc_vm_area - allocate a range of kernel address space
2225 *	@size:		size of the area
2226 *	@ptes:		returns the PTEs for the address space
2227 *
2228 *	Returns:	NULL on failure, vm_struct on success
2229 *
2230 *	This function reserves a range of kernel address space, and
2231 *	allocates pagetables to map that range.  No actual mappings
2232 *	are created.
2233 *
2234 *	If @ptes is non-NULL, pointers to the PTEs (in init_mm)
2235 *	allocated for the VM area are returned.
2236 */
2237struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
2238{
2239	struct vm_struct *area;
2240
2241	area = get_vm_area_caller(size, VM_IOREMAP,
2242				__builtin_return_address(0));
2243	if (area == NULL)
2244		return NULL;
2245
2246	/*
2247	 * This ensures that page tables are constructed for this region
2248	 * of kernel virtual address space and mapped into init_mm.
2249	 */
2250	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2251				size, f, ptes ? &ptes : NULL)) {
2252		free_vm_area(area);
2253		return NULL;
2254	}
2255
2256	return area;
2257}
2258EXPORT_SYMBOL_GPL(alloc_vm_area);
2259
2260void free_vm_area(struct vm_struct *area)
2261{
2262	struct vm_struct *ret;
2263	ret = remove_vm_area(area->addr);
2264	BUG_ON(ret != area);
2265	kfree(area);
2266}
2267EXPORT_SYMBOL_GPL(free_vm_area);
2268
2269#ifdef CONFIG_SMP
2270static struct vmap_area *node_to_va(struct rb_node *n)
2271{
2272	return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2273}
2274
2275/**
2276 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2277 * @end: target address
2278 * @pnext: out arg for the next vmap_area
2279 * @pprev: out arg for the previous vmap_area
2280 *
2281 * Returns: %true if either or both of next and prev are found,
2282 *	    %false if no vmap_area exists
2283 *
2284 * Find vmap_areas end addresses of which enclose @end.  ie. if not
2285 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2286 */
2287static bool pvm_find_next_prev(unsigned long end,
2288			       struct vmap_area **pnext,
2289			       struct vmap_area **pprev)
2290{
2291	struct rb_node *n = vmap_area_root.rb_node;
2292	struct vmap_area *va = NULL;
2293
2294	while (n) {
2295		va = rb_entry(n, struct vmap_area, rb_node);
2296		if (end < va->va_end)
2297			n = n->rb_left;
2298		else if (end > va->va_end)
2299			n = n->rb_right;
2300		else
2301			break;
2302	}
2303
2304	if (!va)
2305		return false;
2306
2307	if (va->va_end > end) {
2308		*pnext = va;
2309		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2310	} else {
2311		*pprev = va;
2312		*pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2313	}
2314	return true;
2315}
2316
2317/**
2318 * pvm_determine_end - find the highest aligned address between two vmap_areas
2319 * @pnext: in/out arg for the next vmap_area
2320 * @pprev: in/out arg for the previous vmap_area
2321 * @align: alignment
2322 *
2323 * Returns: determined end address
2324 *
2325 * Find the highest aligned address between *@pnext and *@pprev below
2326 * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2327 * down address is between the end addresses of the two vmap_areas.
2328 *
2329 * Please note that the address returned by this function may fall
2330 * inside *@pnext vmap_area.  The caller is responsible for checking
2331 * that.
2332 */
2333static unsigned long pvm_determine_end(struct vmap_area **pnext,
2334				       struct vmap_area **pprev,
2335				       unsigned long align)
2336{
2337	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2338	unsigned long addr;
2339
2340	if (*pnext)
2341		addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2342	else
2343		addr = vmalloc_end;
2344
2345	while (*pprev && (*pprev)->va_end > addr) {
2346		*pnext = *pprev;
2347		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2348	}
2349
2350	return addr;
2351}
2352
2353/**
2354 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2355 * @offsets: array containing offset of each area
2356 * @sizes: array containing size of each area
2357 * @nr_vms: the number of areas to allocate
2358 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2359 *
2360 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2361 *	    vm_structs on success, %NULL on failure
2362 *
2363 * Percpu allocator wants to use congruent vm areas so that it can
2364 * maintain the offsets among percpu areas.  This function allocates
2365 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
2366 * be scattered pretty far, distance between two areas easily going up
2367 * to gigabytes.  To avoid interacting with regular vmallocs, these
2368 * areas are allocated from top.
2369 *
2370 * Despite its complicated look, this allocator is rather simple.  It
2371 * does everything top-down and scans areas from the end looking for
2372 * matching slot.  While scanning, if any of the areas overlaps with
2373 * existing vmap_area, the base address is pulled down to fit the
2374 * area.  Scanning is repeated till all the areas fit and then all
2375 * necessary data structres are inserted and the result is returned.
2376 */
2377struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2378				     const size_t *sizes, int nr_vms,
2379				     size_t align)
2380{
2381	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2382	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2383	struct vmap_area **vas, *prev, *next;
2384	struct vm_struct **vms;
2385	int area, area2, last_area, term_area;
2386	unsigned long base, start, end, last_end;
2387	bool purged = false;
2388
2389	/* verify parameters and allocate data structures */
2390	BUG_ON(offset_in_page(align) || !is_power_of_2(align));
2391	for (last_area = 0, area = 0; area < nr_vms; area++) {
2392		start = offsets[area];
2393		end = start + sizes[area];
2394
2395		/* is everything aligned properly? */
2396		BUG_ON(!IS_ALIGNED(offsets[area], align));
2397		BUG_ON(!IS_ALIGNED(sizes[area], align));
2398
2399		/* detect the area with the highest address */
2400		if (start > offsets[last_area])
2401			last_area = area;
2402
2403		for (area2 = 0; area2 < nr_vms; area2++) {
2404			unsigned long start2 = offsets[area2];
2405			unsigned long end2 = start2 + sizes[area2];
2406
2407			if (area2 == area)
2408				continue;
2409
2410			BUG_ON(start2 >= start && start2 < end);
2411			BUG_ON(end2 <= end && end2 > start);
2412		}
2413	}
2414	last_end = offsets[last_area] + sizes[last_area];
2415
2416	if (vmalloc_end - vmalloc_start < last_end) {
2417		WARN_ON(true);
2418		return NULL;
2419	}
2420
2421	vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
2422	vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
2423	if (!vas || !vms)
2424		goto err_free2;
2425
2426	for (area = 0; area < nr_vms; area++) {
2427		vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2428		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2429		if (!vas[area] || !vms[area])
2430			goto err_free;
2431	}
2432retry:
2433	spin_lock(&vmap_area_lock);
2434
2435	/* start scanning - we scan from the top, begin with the last area */
2436	area = term_area = last_area;
2437	start = offsets[area];
2438	end = start + sizes[area];
2439
2440	if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2441		base = vmalloc_end - last_end;
2442		goto found;
2443	}
2444	base = pvm_determine_end(&next, &prev, align) - end;
2445
2446	while (true) {
2447		BUG_ON(next && next->va_end <= base + end);
2448		BUG_ON(prev && prev->va_end > base + end);
2449
2450		/*
2451		 * base might have underflowed, add last_end before
2452		 * comparing.
2453		 */
2454		if (base + last_end < vmalloc_start + last_end) {
2455			spin_unlock(&vmap_area_lock);
2456			if (!purged) {
2457				purge_vmap_area_lazy();
2458				purged = true;
2459				goto retry;
2460			}
2461			goto err_free;
2462		}
2463
2464		/*
2465		 * If next overlaps, move base downwards so that it's
2466		 * right below next and then recheck.
2467		 */
2468		if (next && next->va_start < base + end) {
2469			base = pvm_determine_end(&next, &prev, align) - end;
2470			term_area = area;
2471			continue;
2472		}
2473
2474		/*
2475		 * If prev overlaps, shift down next and prev and move
2476		 * base so that it's right below new next and then
2477		 * recheck.
2478		 */
2479		if (prev && prev->va_end > base + start)  {
2480			next = prev;
2481			prev = node_to_va(rb_prev(&next->rb_node));
2482			base = pvm_determine_end(&next, &prev, align) - end;
2483			term_area = area;
2484			continue;
2485		}
2486
2487		/*
2488		 * This area fits, move on to the previous one.  If
2489		 * the previous one is the terminal one, we're done.
2490		 */
2491		area = (area + nr_vms - 1) % nr_vms;
2492		if (area == term_area)
2493			break;
2494		start = offsets[area];
2495		end = start + sizes[area];
2496		pvm_find_next_prev(base + end, &next, &prev);
2497	}
2498found:
2499	/* we've found a fitting base, insert all va's */
2500	for (area = 0; area < nr_vms; area++) {
2501		struct vmap_area *va = vas[area];
2502
2503		va->va_start = base + offsets[area];
2504		va->va_end = va->va_start + sizes[area];
2505		__insert_vmap_area(va);
2506	}
2507
2508	vmap_area_pcpu_hole = base + offsets[last_area];
2509
2510	spin_unlock(&vmap_area_lock);
2511
2512	/* insert all vm's */
2513	for (area = 0; area < nr_vms; area++)
2514		setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2515				 pcpu_get_vm_areas);
2516
2517	kfree(vas);
2518	return vms;
2519
2520err_free:
2521	for (area = 0; area < nr_vms; area++) {
2522		kfree(vas[area]);
2523		kfree(vms[area]);
2524	}
2525err_free2:
2526	kfree(vas);
2527	kfree(vms);
2528	return NULL;
2529}
2530
2531/**
2532 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2533 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2534 * @nr_vms: the number of allocated areas
2535 *
2536 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2537 */
2538void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2539{
2540	int i;
2541
2542	for (i = 0; i < nr_vms; i++)
2543		free_vm_area(vms[i]);
2544	kfree(vms);
2545}
2546#endif	/* CONFIG_SMP */
2547
2548#ifdef CONFIG_PROC_FS
2549static void *s_start(struct seq_file *m, loff_t *pos)
2550	__acquires(&vmap_area_lock)
2551{
2552	loff_t n = *pos;
2553	struct vmap_area *va;
2554
2555	spin_lock(&vmap_area_lock);
2556	va = list_first_entry(&vmap_area_list, typeof(*va), list);
2557	while (n > 0 && &va->list != &vmap_area_list) {
2558		n--;
2559		va = list_next_entry(va, list);
2560	}
2561	if (!n && &va->list != &vmap_area_list)
2562		return va;
2563
2564	return NULL;
2565
2566}
2567
2568static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2569{
2570	struct vmap_area *va = p, *next;
2571
2572	++*pos;
2573	next = list_next_entry(va, list);
2574	if (&next->list != &vmap_area_list)
2575		return next;
2576
2577	return NULL;
2578}
2579
2580static void s_stop(struct seq_file *m, void *p)
2581	__releases(&vmap_area_lock)
2582{
2583	spin_unlock(&vmap_area_lock);
2584}
2585
2586static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2587{
2588	if (IS_ENABLED(CONFIG_NUMA)) {
2589		unsigned int nr, *counters = m->private;
2590
2591		if (!counters)
2592			return;
2593
 
 
2594		if (v->flags & VM_UNINITIALIZED)
2595			return;
2596		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
2597		smp_rmb();
2598
2599		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2600
2601		for (nr = 0; nr < v->nr_pages; nr++)
2602			counters[page_to_nid(v->pages[nr])]++;
2603
2604		for_each_node_state(nr, N_HIGH_MEMORY)
2605			if (counters[nr])
2606				seq_printf(m, " N%u=%u", nr, counters[nr]);
2607	}
2608}
2609
2610static int s_show(struct seq_file *m, void *p)
2611{
2612	struct vmap_area *va = p;
2613	struct vm_struct *v;
2614
2615	/*
2616	 * s_show can encounter race with remove_vm_area, !VM_VM_AREA on
2617	 * behalf of vmap area is being tear down or vm_map_ram allocation.
2618	 */
2619	if (!(va->flags & VM_VM_AREA))
2620		return 0;
2621
2622	v = va->vm;
2623
2624	seq_printf(m, "0x%pK-0x%pK %7ld",
2625		v->addr, v->addr + v->size, v->size);
2626
2627	if (v->caller)
2628		seq_printf(m, " %pS", v->caller);
2629
2630	if (v->nr_pages)
2631		seq_printf(m, " pages=%d", v->nr_pages);
2632
2633	if (v->phys_addr)
2634		seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2635
2636	if (v->flags & VM_IOREMAP)
2637		seq_puts(m, " ioremap");
2638
2639	if (v->flags & VM_ALLOC)
2640		seq_puts(m, " vmalloc");
2641
2642	if (v->flags & VM_MAP)
2643		seq_puts(m, " vmap");
2644
2645	if (v->flags & VM_USERMAP)
2646		seq_puts(m, " user");
2647
2648	if (is_vmalloc_addr(v->pages))
2649		seq_puts(m, " vpages");
2650
2651	show_numa_info(m, v);
2652	seq_putc(m, '\n');
2653	return 0;
2654}
2655
2656static const struct seq_operations vmalloc_op = {
2657	.start = s_start,
2658	.next = s_next,
2659	.stop = s_stop,
2660	.show = s_show,
2661};
2662
2663static int vmalloc_open(struct inode *inode, struct file *file)
2664{
2665	if (IS_ENABLED(CONFIG_NUMA))
2666		return seq_open_private(file, &vmalloc_op,
2667					nr_node_ids * sizeof(unsigned int));
2668	else
2669		return seq_open(file, &vmalloc_op);
 
 
 
 
 
 
 
 
 
 
2670}
2671
2672static const struct file_operations proc_vmalloc_operations = {
2673	.open		= vmalloc_open,
2674	.read		= seq_read,
2675	.llseek		= seq_lseek,
2676	.release	= seq_release_private,
2677};
2678
2679static int __init proc_vmalloc_init(void)
2680{
2681	proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2682	return 0;
2683}
2684module_init(proc_vmalloc_init);
2685
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2686#endif
2687