Linux Audio

Check our new training course

Loading...
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
v3.1
   1/*
   2 *  linux/mm/vmalloc.c
   3 *
   4 *  Copyright (C) 1993  Linus Torvalds
   5 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   6 *  SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
   7 *  Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
   8 *  Numa awareness, Christoph Lameter, SGI, June 2005
   9 */
  10
  11#include <linux/vmalloc.h>
  12#include <linux/mm.h>
  13#include <linux/module.h>
  14#include <linux/highmem.h>
  15#include <linux/sched.h>
  16#include <linux/slab.h>
  17#include <linux/spinlock.h>
  18#include <linux/interrupt.h>
  19#include <linux/proc_fs.h>
  20#include <linux/seq_file.h>
  21#include <linux/debugobjects.h>
  22#include <linux/kallsyms.h>
  23#include <linux/list.h>
  24#include <linux/rbtree.h>
  25#include <linux/radix-tree.h>
  26#include <linux/rcupdate.h>
  27#include <linux/pfn.h>
  28#include <linux/kmemleak.h>
  29#include <linux/atomic.h>
 
 
 
 
  30#include <asm/uaccess.h>
  31#include <asm/tlbflush.h>
  32#include <asm/shmparam.h>
  33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  34/*** Page table manipulation functions ***/
  35
  36static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
  37{
  38	pte_t *pte;
  39
  40	pte = pte_offset_kernel(pmd, addr);
  41	do {
  42		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
  43		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
  44	} while (pte++, addr += PAGE_SIZE, addr != end);
  45}
  46
  47static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
  48{
  49	pmd_t *pmd;
  50	unsigned long next;
  51
  52	pmd = pmd_offset(pud, addr);
  53	do {
  54		next = pmd_addr_end(addr, end);
 
 
  55		if (pmd_none_or_clear_bad(pmd))
  56			continue;
  57		vunmap_pte_range(pmd, addr, next);
  58	} while (pmd++, addr = next, addr != end);
  59}
  60
  61static void vunmap_pud_range(pgd_t *pgd, unsigned long addr, unsigned long end)
  62{
  63	pud_t *pud;
  64	unsigned long next;
  65
  66	pud = pud_offset(pgd, addr);
  67	do {
  68		next = pud_addr_end(addr, end);
 
 
  69		if (pud_none_or_clear_bad(pud))
  70			continue;
  71		vunmap_pmd_range(pud, addr, next);
  72	} while (pud++, addr = next, addr != end);
  73}
  74
  75static void vunmap_page_range(unsigned long addr, unsigned long end)
  76{
  77	pgd_t *pgd;
  78	unsigned long next;
  79
  80	BUG_ON(addr >= end);
  81	pgd = pgd_offset_k(addr);
  82	do {
  83		next = pgd_addr_end(addr, end);
  84		if (pgd_none_or_clear_bad(pgd))
  85			continue;
  86		vunmap_pud_range(pgd, addr, next);
  87	} while (pgd++, addr = next, addr != end);
  88}
  89
  90static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
  91		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
  92{
  93	pte_t *pte;
  94
  95	/*
  96	 * nr is a running index into the array which helps higher level
  97	 * callers keep track of where we're up to.
  98	 */
  99
 100	pte = pte_alloc_kernel(pmd, addr);
 101	if (!pte)
 102		return -ENOMEM;
 103	do {
 104		struct page *page = pages[*nr];
 105
 106		if (WARN_ON(!pte_none(*pte)))
 107			return -EBUSY;
 108		if (WARN_ON(!page))
 109			return -ENOMEM;
 110		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 111		(*nr)++;
 112	} while (pte++, addr += PAGE_SIZE, addr != end);
 113	return 0;
 114}
 115
 116static int vmap_pmd_range(pud_t *pud, unsigned long addr,
 117		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 118{
 119	pmd_t *pmd;
 120	unsigned long next;
 121
 122	pmd = pmd_alloc(&init_mm, pud, addr);
 123	if (!pmd)
 124		return -ENOMEM;
 125	do {
 126		next = pmd_addr_end(addr, end);
 127		if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
 128			return -ENOMEM;
 129	} while (pmd++, addr = next, addr != end);
 130	return 0;
 131}
 132
 133static int vmap_pud_range(pgd_t *pgd, unsigned long addr,
 134		unsigned long end, pgprot_t prot, struct page **pages, int *nr)
 135{
 136	pud_t *pud;
 137	unsigned long next;
 138
 139	pud = pud_alloc(&init_mm, pgd, addr);
 140	if (!pud)
 141		return -ENOMEM;
 142	do {
 143		next = pud_addr_end(addr, end);
 144		if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
 145			return -ENOMEM;
 146	} while (pud++, addr = next, addr != end);
 147	return 0;
 148}
 149
 150/*
 151 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
 152 * will have pfns corresponding to the "pages" array.
 153 *
 154 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
 155 */
 156static int vmap_page_range_noflush(unsigned long start, unsigned long end,
 157				   pgprot_t prot, struct page **pages)
 158{
 159	pgd_t *pgd;
 160	unsigned long next;
 161	unsigned long addr = start;
 162	int err = 0;
 163	int nr = 0;
 164
 165	BUG_ON(addr >= end);
 166	pgd = pgd_offset_k(addr);
 167	do {
 168		next = pgd_addr_end(addr, end);
 169		err = vmap_pud_range(pgd, addr, next, prot, pages, &nr);
 170		if (err)
 171			return err;
 172	} while (pgd++, addr = next, addr != end);
 173
 174	return nr;
 175}
 176
 177static int vmap_page_range(unsigned long start, unsigned long end,
 178			   pgprot_t prot, struct page **pages)
 179{
 180	int ret;
 181
 182	ret = vmap_page_range_noflush(start, end, prot, pages);
 183	flush_cache_vmap(start, end);
 184	return ret;
 185}
 186
 187int is_vmalloc_or_module_addr(const void *x)
 188{
 189	/*
 190	 * ARM, x86-64 and sparc64 put modules in a special place,
 191	 * and fall back on vmalloc() if that fails. Others
 192	 * just put it in the vmalloc space.
 193	 */
 194#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 195	unsigned long addr = (unsigned long)x;
 196	if (addr >= MODULES_VADDR && addr < MODULES_END)
 197		return 1;
 198#endif
 199	return is_vmalloc_addr(x);
 200}
 201
 202/*
 203 * Walk a vmap address to the struct page it maps.
 204 */
 205struct page *vmalloc_to_page(const void *vmalloc_addr)
 206{
 207	unsigned long addr = (unsigned long) vmalloc_addr;
 208	struct page *page = NULL;
 209	pgd_t *pgd = pgd_offset_k(addr);
 210
 211	/*
 212	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 213	 * architectures that do not vmalloc module space
 214	 */
 215	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 216
 217	if (!pgd_none(*pgd)) {
 218		pud_t *pud = pud_offset(pgd, addr);
 219		if (!pud_none(*pud)) {
 220			pmd_t *pmd = pmd_offset(pud, addr);
 221			if (!pmd_none(*pmd)) {
 222				pte_t *ptep, pte;
 223
 224				ptep = pte_offset_map(pmd, addr);
 225				pte = *ptep;
 226				if (pte_present(pte))
 227					page = pte_page(pte);
 228				pte_unmap(ptep);
 229			}
 230		}
 231	}
 232	return page;
 233}
 234EXPORT_SYMBOL(vmalloc_to_page);
 235
 236/*
 237 * Map a vmalloc()-space virtual address to the physical page frame number.
 238 */
 239unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 240{
 241	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 242}
 243EXPORT_SYMBOL(vmalloc_to_pfn);
 244
 245
 246/*** Global kva allocator ***/
 247
 248#define VM_LAZY_FREE	0x01
 249#define VM_LAZY_FREEING	0x02
 250#define VM_VM_AREA	0x04
 251
 252struct vmap_area {
 253	unsigned long va_start;
 254	unsigned long va_end;
 255	unsigned long flags;
 256	struct rb_node rb_node;		/* address sorted rbtree */
 257	struct list_head list;		/* address sorted list */
 258	struct list_head purge_list;	/* "lazy purge" list */
 259	void *private;
 260	struct rcu_head rcu_head;
 261};
 262
 263static DEFINE_SPINLOCK(vmap_area_lock);
 264static LIST_HEAD(vmap_area_list);
 
 265static struct rb_root vmap_area_root = RB_ROOT;
 266
 267/* The vmap cache globals are protected by vmap_area_lock */
 268static struct rb_node *free_vmap_cache;
 269static unsigned long cached_hole_size;
 270static unsigned long cached_vstart;
 271static unsigned long cached_align;
 272
 273static unsigned long vmap_area_pcpu_hole;
 274
 275static struct vmap_area *__find_vmap_area(unsigned long addr)
 276{
 277	struct rb_node *n = vmap_area_root.rb_node;
 278
 279	while (n) {
 280		struct vmap_area *va;
 281
 282		va = rb_entry(n, struct vmap_area, rb_node);
 283		if (addr < va->va_start)
 284			n = n->rb_left;
 285		else if (addr > va->va_start)
 286			n = n->rb_right;
 287		else
 288			return va;
 289	}
 290
 291	return NULL;
 292}
 293
 294static void __insert_vmap_area(struct vmap_area *va)
 295{
 296	struct rb_node **p = &vmap_area_root.rb_node;
 297	struct rb_node *parent = NULL;
 298	struct rb_node *tmp;
 299
 300	while (*p) {
 301		struct vmap_area *tmp_va;
 302
 303		parent = *p;
 304		tmp_va = rb_entry(parent, struct vmap_area, rb_node);
 305		if (va->va_start < tmp_va->va_end)
 306			p = &(*p)->rb_left;
 307		else if (va->va_end > tmp_va->va_start)
 308			p = &(*p)->rb_right;
 309		else
 310			BUG();
 311	}
 312
 313	rb_link_node(&va->rb_node, parent, p);
 314	rb_insert_color(&va->rb_node, &vmap_area_root);
 315
 316	/* address-sort this list so it is usable like the vmlist */
 317	tmp = rb_prev(&va->rb_node);
 318	if (tmp) {
 319		struct vmap_area *prev;
 320		prev = rb_entry(tmp, struct vmap_area, rb_node);
 321		list_add_rcu(&va->list, &prev->list);
 322	} else
 323		list_add_rcu(&va->list, &vmap_area_list);
 324}
 325
 326static void purge_vmap_area_lazy(void);
 327
 328/*
 329 * Allocate a region of KVA of the specified size and alignment, within the
 330 * vstart and vend.
 331 */
 332static struct vmap_area *alloc_vmap_area(unsigned long size,
 333				unsigned long align,
 334				unsigned long vstart, unsigned long vend,
 335				int node, gfp_t gfp_mask)
 336{
 337	struct vmap_area *va;
 338	struct rb_node *n;
 339	unsigned long addr;
 340	int purged = 0;
 341	struct vmap_area *first;
 342
 343	BUG_ON(!size);
 344	BUG_ON(size & ~PAGE_MASK);
 345	BUG_ON(!is_power_of_2(align));
 346
 347	va = kmalloc_node(sizeof(struct vmap_area),
 348			gfp_mask & GFP_RECLAIM_MASK, node);
 349	if (unlikely(!va))
 350		return ERR_PTR(-ENOMEM);
 351
 
 
 
 
 
 
 352retry:
 353	spin_lock(&vmap_area_lock);
 354	/*
 355	 * Invalidate cache if we have more permissive parameters.
 356	 * cached_hole_size notes the largest hole noticed _below_
 357	 * the vmap_area cached in free_vmap_cache: if size fits
 358	 * into that hole, we want to scan from vstart to reuse
 359	 * the hole instead of allocating above free_vmap_cache.
 360	 * Note that __free_vmap_area may update free_vmap_cache
 361	 * without updating cached_hole_size or cached_align.
 362	 */
 363	if (!free_vmap_cache ||
 364			size < cached_hole_size ||
 365			vstart < cached_vstart ||
 366			align < cached_align) {
 367nocache:
 368		cached_hole_size = 0;
 369		free_vmap_cache = NULL;
 370	}
 371	/* record if we encounter less permissive parameters */
 372	cached_vstart = vstart;
 373	cached_align = align;
 374
 375	/* find starting point for our search */
 376	if (free_vmap_cache) {
 377		first = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 378		addr = ALIGN(first->va_end, align);
 379		if (addr < vstart)
 380			goto nocache;
 381		if (addr + size - 1 < addr)
 382			goto overflow;
 383
 384	} else {
 385		addr = ALIGN(vstart, align);
 386		if (addr + size - 1 < addr)
 387			goto overflow;
 388
 389		n = vmap_area_root.rb_node;
 390		first = NULL;
 391
 392		while (n) {
 393			struct vmap_area *tmp;
 394			tmp = rb_entry(n, struct vmap_area, rb_node);
 395			if (tmp->va_end >= addr) {
 396				first = tmp;
 397				if (tmp->va_start <= addr)
 398					break;
 399				n = n->rb_left;
 400			} else
 401				n = n->rb_right;
 402		}
 403
 404		if (!first)
 405			goto found;
 406	}
 407
 408	/* from the starting point, walk areas until a suitable hole is found */
 409	while (addr + size > first->va_start && addr + size <= vend) {
 410		if (addr + cached_hole_size < first->va_start)
 411			cached_hole_size = first->va_start - addr;
 412		addr = ALIGN(first->va_end, align);
 413		if (addr + size - 1 < addr)
 414			goto overflow;
 415
 416		n = rb_next(&first->rb_node);
 417		if (n)
 418			first = rb_entry(n, struct vmap_area, rb_node);
 419		else
 420			goto found;
 
 
 421	}
 422
 423found:
 424	if (addr + size > vend)
 425		goto overflow;
 426
 427	va->va_start = addr;
 428	va->va_end = addr + size;
 429	va->flags = 0;
 430	__insert_vmap_area(va);
 431	free_vmap_cache = &va->rb_node;
 432	spin_unlock(&vmap_area_lock);
 433
 434	BUG_ON(va->va_start & (align-1));
 435	BUG_ON(va->va_start < vstart);
 436	BUG_ON(va->va_end > vend);
 437
 438	return va;
 439
 440overflow:
 441	spin_unlock(&vmap_area_lock);
 442	if (!purged) {
 443		purge_vmap_area_lazy();
 444		purged = 1;
 445		goto retry;
 446	}
 447	if (printk_ratelimit())
 448		printk(KERN_WARNING
 449			"vmap allocation for size %lu failed: "
 450			"use vmalloc=<size> to increase size.\n", size);
 451	kfree(va);
 452	return ERR_PTR(-EBUSY);
 453}
 454
 455static void __free_vmap_area(struct vmap_area *va)
 456{
 457	BUG_ON(RB_EMPTY_NODE(&va->rb_node));
 458
 459	if (free_vmap_cache) {
 460		if (va->va_end < cached_vstart) {
 461			free_vmap_cache = NULL;
 462		} else {
 463			struct vmap_area *cache;
 464			cache = rb_entry(free_vmap_cache, struct vmap_area, rb_node);
 465			if (va->va_start <= cache->va_start) {
 466				free_vmap_cache = rb_prev(&va->rb_node);
 467				/*
 468				 * We don't try to update cached_hole_size or
 469				 * cached_align, but it won't go very wrong.
 470				 */
 471			}
 472		}
 473	}
 474	rb_erase(&va->rb_node, &vmap_area_root);
 475	RB_CLEAR_NODE(&va->rb_node);
 476	list_del_rcu(&va->list);
 477
 478	/*
 479	 * Track the highest possible candidate for pcpu area
 480	 * allocation.  Areas outside of vmalloc area can be returned
 481	 * here too, consider only end addresses which fall inside
 482	 * vmalloc area proper.
 483	 */
 484	if (va->va_end > VMALLOC_START && va->va_end <= VMALLOC_END)
 485		vmap_area_pcpu_hole = max(vmap_area_pcpu_hole, va->va_end);
 486
 487	kfree_rcu(va, rcu_head);
 488}
 489
 490/*
 491 * Free a region of KVA allocated by alloc_vmap_area
 492 */
 493static void free_vmap_area(struct vmap_area *va)
 494{
 495	spin_lock(&vmap_area_lock);
 496	__free_vmap_area(va);
 497	spin_unlock(&vmap_area_lock);
 498}
 499
 500/*
 501 * Clear the pagetable entries of a given vmap_area
 502 */
 503static void unmap_vmap_area(struct vmap_area *va)
 504{
 505	vunmap_page_range(va->va_start, va->va_end);
 506}
 507
 508static void vmap_debug_free_range(unsigned long start, unsigned long end)
 509{
 510	/*
 511	 * Unmap page tables and force a TLB flush immediately if
 512	 * CONFIG_DEBUG_PAGEALLOC is set. This catches use after free
 513	 * bugs similarly to those in linear kernel virtual address
 514	 * space after a page has been freed.
 515	 *
 516	 * All the lazy freeing logic is still retained, in order to
 517	 * minimise intrusiveness of this debugging feature.
 518	 *
 519	 * This is going to be *slow* (linear kernel virtual address
 520	 * debugging doesn't do a broadcast TLB flush so it is a lot
 521	 * faster).
 522	 */
 523#ifdef CONFIG_DEBUG_PAGEALLOC
 524	vunmap_page_range(start, end);
 525	flush_tlb_kernel_range(start, end);
 526#endif
 527}
 528
 529/*
 530 * lazy_max_pages is the maximum amount of virtual address space we gather up
 531 * before attempting to purge with a TLB flush.
 532 *
 533 * There is a tradeoff here: a larger number will cover more kernel page tables
 534 * and take slightly longer to purge, but it will linearly reduce the number of
 535 * global TLB flushes that must be performed. It would seem natural to scale
 536 * this number up linearly with the number of CPUs (because vmapping activity
 537 * could also scale linearly with the number of CPUs), however it is likely
 538 * that in practice, workloads might be constrained in other ways that mean
 539 * vmap activity will not scale linearly with CPUs. Also, I want to be
 540 * conservative and not introduce a big latency on huge systems, so go with
 541 * a less aggressive log scale. It will still be an improvement over the old
 542 * code, and it will be simple to change the scale factor if we find that it
 543 * becomes a problem on bigger systems.
 544 */
 545static unsigned long lazy_max_pages(void)
 546{
 547	unsigned int log;
 548
 549	log = fls(num_online_cpus());
 550
 551	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
 552}
 553
 554static atomic_t vmap_lazy_nr = ATOMIC_INIT(0);
 555
 556/* for per-CPU blocks */
 557static void purge_fragmented_blocks_allcpus(void);
 558
 559/*
 560 * called before a call to iounmap() if the caller wants vm_area_struct's
 561 * immediately freed.
 562 */
 563void set_iounmap_nonlazy(void)
 564{
 565	atomic_set(&vmap_lazy_nr, lazy_max_pages()+1);
 566}
 567
 568/*
 569 * Purges all lazily-freed vmap areas.
 570 *
 571 * If sync is 0 then don't purge if there is already a purge in progress.
 572 * If force_flush is 1, then flush kernel TLBs between *start and *end even
 573 * if we found no lazy vmap areas to unmap (callers can use this to optimise
 574 * their own TLB flushing).
 575 * Returns with *start = min(*start, lowest purged address)
 576 *              *end = max(*end, highest purged address)
 577 */
 578static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
 579					int sync, int force_flush)
 580{
 581	static DEFINE_SPINLOCK(purge_lock);
 582	LIST_HEAD(valist);
 583	struct vmap_area *va;
 584	struct vmap_area *n_va;
 585	int nr = 0;
 586
 587	/*
 588	 * If sync is 0 but force_flush is 1, we'll go sync anyway but callers
 589	 * should not expect such behaviour. This just simplifies locking for
 590	 * the case that isn't actually used at the moment anyway.
 591	 */
 592	if (!sync && !force_flush) {
 593		if (!spin_trylock(&purge_lock))
 594			return;
 595	} else
 596		spin_lock(&purge_lock);
 597
 598	if (sync)
 599		purge_fragmented_blocks_allcpus();
 600
 601	rcu_read_lock();
 602	list_for_each_entry_rcu(va, &vmap_area_list, list) {
 603		if (va->flags & VM_LAZY_FREE) {
 604			if (va->va_start < *start)
 605				*start = va->va_start;
 606			if (va->va_end > *end)
 607				*end = va->va_end;
 608			nr += (va->va_end - va->va_start) >> PAGE_SHIFT;
 609			list_add_tail(&va->purge_list, &valist);
 610			va->flags |= VM_LAZY_FREEING;
 611			va->flags &= ~VM_LAZY_FREE;
 612		}
 613	}
 614	rcu_read_unlock();
 615
 616	if (nr)
 617		atomic_sub(nr, &vmap_lazy_nr);
 618
 619	if (nr || force_flush)
 620		flush_tlb_kernel_range(*start, *end);
 621
 622	if (nr) {
 623		spin_lock(&vmap_area_lock);
 624		list_for_each_entry_safe(va, n_va, &valist, purge_list)
 625			__free_vmap_area(va);
 626		spin_unlock(&vmap_area_lock);
 627	}
 628	spin_unlock(&purge_lock);
 629}
 630
 631/*
 632 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
 633 * is already purging.
 634 */
 635static void try_purge_vmap_area_lazy(void)
 636{
 637	unsigned long start = ULONG_MAX, end = 0;
 638
 639	__purge_vmap_area_lazy(&start, &end, 0, 0);
 640}
 641
 642/*
 643 * Kick off a purge of the outstanding lazy areas.
 644 */
 645static void purge_vmap_area_lazy(void)
 646{
 647	unsigned long start = ULONG_MAX, end = 0;
 648
 649	__purge_vmap_area_lazy(&start, &end, 1, 0);
 650}
 651
 652/*
 653 * Free a vmap area, caller ensuring that the area has been unmapped
 654 * and flush_cache_vunmap had been called for the correct range
 655 * previously.
 656 */
 657static void free_vmap_area_noflush(struct vmap_area *va)
 658{
 659	va->flags |= VM_LAZY_FREE;
 660	atomic_add((va->va_end - va->va_start) >> PAGE_SHIFT, &vmap_lazy_nr);
 661	if (unlikely(atomic_read(&vmap_lazy_nr) > lazy_max_pages()))
 662		try_purge_vmap_area_lazy();
 663}
 664
 665/*
 666 * Free and unmap a vmap area, caller ensuring flush_cache_vunmap had been
 667 * called for the correct range previously.
 668 */
 669static void free_unmap_vmap_area_noflush(struct vmap_area *va)
 670{
 671	unmap_vmap_area(va);
 672	free_vmap_area_noflush(va);
 673}
 674
 675/*
 676 * Free and unmap a vmap area
 677 */
 678static void free_unmap_vmap_area(struct vmap_area *va)
 679{
 680	flush_cache_vunmap(va->va_start, va->va_end);
 681	free_unmap_vmap_area_noflush(va);
 682}
 683
 684static struct vmap_area *find_vmap_area(unsigned long addr)
 685{
 686	struct vmap_area *va;
 687
 688	spin_lock(&vmap_area_lock);
 689	va = __find_vmap_area(addr);
 690	spin_unlock(&vmap_area_lock);
 691
 692	return va;
 693}
 694
 695static void free_unmap_vmap_area_addr(unsigned long addr)
 696{
 697	struct vmap_area *va;
 698
 699	va = find_vmap_area(addr);
 700	BUG_ON(!va);
 701	free_unmap_vmap_area(va);
 702}
 703
 704
 705/*** Per cpu kva allocator ***/
 706
 707/*
 708 * vmap space is limited especially on 32 bit architectures. Ensure there is
 709 * room for at least 16 percpu vmap blocks per CPU.
 710 */
 711/*
 712 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
 713 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
 714 * instead (we just need a rough idea)
 715 */
 716#if BITS_PER_LONG == 32
 717#define VMALLOC_SPACE		(128UL*1024*1024)
 718#else
 719#define VMALLOC_SPACE		(128UL*1024*1024*1024)
 720#endif
 721
 722#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
 723#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
 724#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
 725#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
 726#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
 727#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
 728#define VMAP_BBMAP_BITS		\
 729		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
 730		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
 731			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
 732
 733#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
 734
 735static bool vmap_initialized __read_mostly = false;
 736
 737struct vmap_block_queue {
 738	spinlock_t lock;
 739	struct list_head free;
 740};
 741
 742struct vmap_block {
 743	spinlock_t lock;
 744	struct vmap_area *va;
 745	struct vmap_block_queue *vbq;
 746	unsigned long free, dirty;
 747	DECLARE_BITMAP(alloc_map, VMAP_BBMAP_BITS);
 748	DECLARE_BITMAP(dirty_map, VMAP_BBMAP_BITS);
 749	struct list_head free_list;
 750	struct rcu_head rcu_head;
 751	struct list_head purge;
 752};
 753
 754/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
 755static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
 756
 757/*
 758 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
 759 * in the free path. Could get rid of this if we change the API to return a
 760 * "cookie" from alloc, to be passed to free. But no big deal yet.
 761 */
 762static DEFINE_SPINLOCK(vmap_block_tree_lock);
 763static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
 764
 765/*
 766 * We should probably have a fallback mechanism to allocate virtual memory
 767 * out of partially filled vmap blocks. However vmap block sizing should be
 768 * fairly reasonable according to the vmalloc size, so it shouldn't be a
 769 * big problem.
 770 */
 771
 772static unsigned long addr_to_vb_idx(unsigned long addr)
 773{
 774	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
 775	addr /= VMAP_BLOCK_SIZE;
 776	return addr;
 777}
 778
 779static struct vmap_block *new_vmap_block(gfp_t gfp_mask)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 780{
 781	struct vmap_block_queue *vbq;
 782	struct vmap_block *vb;
 783	struct vmap_area *va;
 784	unsigned long vb_idx;
 785	int node, err;
 
 786
 787	node = numa_node_id();
 788
 789	vb = kmalloc_node(sizeof(struct vmap_block),
 790			gfp_mask & GFP_RECLAIM_MASK, node);
 791	if (unlikely(!vb))
 792		return ERR_PTR(-ENOMEM);
 793
 794	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
 795					VMALLOC_START, VMALLOC_END,
 796					node, gfp_mask);
 797	if (IS_ERR(va)) {
 798		kfree(vb);
 799		return ERR_CAST(va);
 800	}
 801
 802	err = radix_tree_preload(gfp_mask);
 803	if (unlikely(err)) {
 804		kfree(vb);
 805		free_vmap_area(va);
 806		return ERR_PTR(err);
 807	}
 808
 
 809	spin_lock_init(&vb->lock);
 810	vb->va = va;
 811	vb->free = VMAP_BBMAP_BITS;
 
 
 812	vb->dirty = 0;
 813	bitmap_zero(vb->alloc_map, VMAP_BBMAP_BITS);
 814	bitmap_zero(vb->dirty_map, VMAP_BBMAP_BITS);
 815	INIT_LIST_HEAD(&vb->free_list);
 816
 817	vb_idx = addr_to_vb_idx(va->va_start);
 818	spin_lock(&vmap_block_tree_lock);
 819	err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
 820	spin_unlock(&vmap_block_tree_lock);
 821	BUG_ON(err);
 822	radix_tree_preload_end();
 823
 824	vbq = &get_cpu_var(vmap_block_queue);
 825	vb->vbq = vbq;
 826	spin_lock(&vbq->lock);
 827	list_add_rcu(&vb->free_list, &vbq->free);
 828	spin_unlock(&vbq->lock);
 829	put_cpu_var(vmap_block_queue);
 830
 831	return vb;
 832}
 833
 834static void free_vmap_block(struct vmap_block *vb)
 835{
 836	struct vmap_block *tmp;
 837	unsigned long vb_idx;
 838
 839	vb_idx = addr_to_vb_idx(vb->va->va_start);
 840	spin_lock(&vmap_block_tree_lock);
 841	tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
 842	spin_unlock(&vmap_block_tree_lock);
 843	BUG_ON(tmp != vb);
 844
 845	free_vmap_area_noflush(vb->va);
 846	kfree_rcu(vb, rcu_head);
 847}
 848
 849static void purge_fragmented_blocks(int cpu)
 850{
 851	LIST_HEAD(purge);
 852	struct vmap_block *vb;
 853	struct vmap_block *n_vb;
 854	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
 855
 856	rcu_read_lock();
 857	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 858
 859		if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
 860			continue;
 861
 862		spin_lock(&vb->lock);
 863		if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
 864			vb->free = 0; /* prevent further allocs after releasing lock */
 865			vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
 866			bitmap_fill(vb->alloc_map, VMAP_BBMAP_BITS);
 867			bitmap_fill(vb->dirty_map, VMAP_BBMAP_BITS);
 868			spin_lock(&vbq->lock);
 869			list_del_rcu(&vb->free_list);
 870			spin_unlock(&vbq->lock);
 871			spin_unlock(&vb->lock);
 872			list_add_tail(&vb->purge, &purge);
 873		} else
 874			spin_unlock(&vb->lock);
 875	}
 876	rcu_read_unlock();
 877
 878	list_for_each_entry_safe(vb, n_vb, &purge, purge) {
 879		list_del(&vb->purge);
 880		free_vmap_block(vb);
 881	}
 882}
 883
 884static void purge_fragmented_blocks_thiscpu(void)
 885{
 886	purge_fragmented_blocks(smp_processor_id());
 887}
 888
 889static void purge_fragmented_blocks_allcpus(void)
 890{
 891	int cpu;
 892
 893	for_each_possible_cpu(cpu)
 894		purge_fragmented_blocks(cpu);
 895}
 896
 897static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
 898{
 899	struct vmap_block_queue *vbq;
 900	struct vmap_block *vb;
 901	unsigned long addr = 0;
 902	unsigned int order;
 903	int purge = 0;
 904
 905	BUG_ON(size & ~PAGE_MASK);
 906	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 
 
 
 
 
 
 
 
 907	order = get_order(size);
 908
 909again:
 910	rcu_read_lock();
 911	vbq = &get_cpu_var(vmap_block_queue);
 912	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 913		int i;
 914
 915		spin_lock(&vb->lock);
 916		if (vb->free < 1UL << order)
 917			goto next;
 918
 919		i = bitmap_find_free_region(vb->alloc_map,
 920						VMAP_BBMAP_BITS, order);
 921
 922		if (i < 0) {
 923			if (vb->free + vb->dirty == VMAP_BBMAP_BITS) {
 924				/* fragmented and no outstanding allocations */
 925				BUG_ON(vb->dirty != VMAP_BBMAP_BITS);
 926				purge = 1;
 927			}
 928			goto next;
 929		}
 930		addr = vb->va->va_start + (i << PAGE_SHIFT);
 931		BUG_ON(addr_to_vb_idx(addr) !=
 932				addr_to_vb_idx(vb->va->va_start));
 933		vb->free -= 1UL << order;
 934		if (vb->free == 0) {
 935			spin_lock(&vbq->lock);
 936			list_del_rcu(&vb->free_list);
 937			spin_unlock(&vbq->lock);
 938		}
 
 939		spin_unlock(&vb->lock);
 940		break;
 941next:
 942		spin_unlock(&vb->lock);
 943	}
 944
 945	if (purge)
 946		purge_fragmented_blocks_thiscpu();
 947
 948	put_cpu_var(vmap_block_queue);
 949	rcu_read_unlock();
 950
 951	if (!addr) {
 952		vb = new_vmap_block(gfp_mask);
 953		if (IS_ERR(vb))
 954			return vb;
 955		goto again;
 956	}
 957
 958	return (void *)addr;
 959}
 960
 961static void vb_free(const void *addr, unsigned long size)
 962{
 963	unsigned long offset;
 964	unsigned long vb_idx;
 965	unsigned int order;
 966	struct vmap_block *vb;
 967
 968	BUG_ON(size & ~PAGE_MASK);
 969	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
 970
 971	flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
 972
 973	order = get_order(size);
 974
 975	offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
 
 976
 977	vb_idx = addr_to_vb_idx((unsigned long)addr);
 978	rcu_read_lock();
 979	vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
 980	rcu_read_unlock();
 981	BUG_ON(!vb);
 982
 983	vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
 984
 985	spin_lock(&vb->lock);
 986	BUG_ON(bitmap_allocate_region(vb->dirty_map, offset >> PAGE_SHIFT, order));
 
 
 
 987
 988	vb->dirty += 1UL << order;
 989	if (vb->dirty == VMAP_BBMAP_BITS) {
 990		BUG_ON(vb->free);
 991		spin_unlock(&vb->lock);
 992		free_vmap_block(vb);
 993	} else
 994		spin_unlock(&vb->lock);
 995}
 996
 997/**
 998 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
 999 *
1000 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1001 * to amortize TLB flushing overheads. What this means is that any page you
1002 * have now, may, in a former life, have been mapped into kernel virtual
1003 * address by the vmap layer and so there might be some CPUs with TLB entries
1004 * still referencing that page (additional to the regular 1:1 kernel mapping).
1005 *
1006 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1007 * be sure that none of the pages we have control over will have any aliases
1008 * from the vmap layer.
1009 */
1010void vm_unmap_aliases(void)
1011{
1012	unsigned long start = ULONG_MAX, end = 0;
1013	int cpu;
1014	int flush = 0;
1015
1016	if (unlikely(!vmap_initialized))
1017		return;
1018
1019	for_each_possible_cpu(cpu) {
1020		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1021		struct vmap_block *vb;
1022
1023		rcu_read_lock();
1024		list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1025			int i;
1026
1027			spin_lock(&vb->lock);
1028			i = find_first_bit(vb->dirty_map, VMAP_BBMAP_BITS);
1029			while (i < VMAP_BBMAP_BITS) {
1030				unsigned long s, e;
1031				int j;
1032				j = find_next_zero_bit(vb->dirty_map,
1033					VMAP_BBMAP_BITS, i);
1034
1035				s = vb->va->va_start + (i << PAGE_SHIFT);
1036				e = vb->va->va_start + (j << PAGE_SHIFT);
 
 
 
 
1037				flush = 1;
1038
1039				if (s < start)
1040					start = s;
1041				if (e > end)
1042					end = e;
1043
1044				i = j;
1045				i = find_next_bit(vb->dirty_map,
1046							VMAP_BBMAP_BITS, i);
1047			}
1048			spin_unlock(&vb->lock);
1049		}
1050		rcu_read_unlock();
1051	}
1052
1053	__purge_vmap_area_lazy(&start, &end, 1, flush);
1054}
1055EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1056
1057/**
1058 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1059 * @mem: the pointer returned by vm_map_ram
1060 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1061 */
1062void vm_unmap_ram(const void *mem, unsigned int count)
1063{
1064	unsigned long size = count << PAGE_SHIFT;
1065	unsigned long addr = (unsigned long)mem;
1066
1067	BUG_ON(!addr);
1068	BUG_ON(addr < VMALLOC_START);
1069	BUG_ON(addr > VMALLOC_END);
1070	BUG_ON(addr & (PAGE_SIZE-1));
1071
1072	debug_check_no_locks_freed(mem, size);
1073	vmap_debug_free_range(addr, addr+size);
1074
1075	if (likely(count <= VMAP_MAX_ALLOC))
1076		vb_free(mem, size);
1077	else
1078		free_unmap_vmap_area_addr(addr);
1079}
1080EXPORT_SYMBOL(vm_unmap_ram);
1081
1082/**
1083 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1084 * @pages: an array of pointers to the pages to be mapped
1085 * @count: number of pages
1086 * @node: prefer to allocate data structures on this node
1087 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1088 *
 
 
 
 
 
 
1089 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1090 */
1091void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1092{
1093	unsigned long size = count << PAGE_SHIFT;
1094	unsigned long addr;
1095	void *mem;
1096
1097	if (likely(count <= VMAP_MAX_ALLOC)) {
1098		mem = vb_alloc(size, GFP_KERNEL);
1099		if (IS_ERR(mem))
1100			return NULL;
1101		addr = (unsigned long)mem;
1102	} else {
1103		struct vmap_area *va;
1104		va = alloc_vmap_area(size, PAGE_SIZE,
1105				VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1106		if (IS_ERR(va))
1107			return NULL;
1108
1109		addr = va->va_start;
1110		mem = (void *)addr;
1111	}
1112	if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1113		vm_unmap_ram(mem, count);
1114		return NULL;
1115	}
1116	return mem;
1117}
1118EXPORT_SYMBOL(vm_map_ram);
1119
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1120/**
1121 * vm_area_register_early - register vmap area early during boot
1122 * @vm: vm_struct to register
1123 * @align: requested alignment
1124 *
1125 * This function is used to register kernel vm area before
1126 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1127 * proper values on entry and other fields should be zero.  On return,
1128 * vm->addr contains the allocated address.
1129 *
1130 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1131 */
1132void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1133{
1134	static size_t vm_init_off __initdata;
1135	unsigned long addr;
1136
1137	addr = ALIGN(VMALLOC_START + vm_init_off, align);
1138	vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1139
1140	vm->addr = (void *)addr;
1141
1142	vm->next = vmlist;
1143	vmlist = vm;
1144}
1145
1146void __init vmalloc_init(void)
1147{
1148	struct vmap_area *va;
1149	struct vm_struct *tmp;
1150	int i;
1151
1152	for_each_possible_cpu(i) {
1153		struct vmap_block_queue *vbq;
 
1154
1155		vbq = &per_cpu(vmap_block_queue, i);
1156		spin_lock_init(&vbq->lock);
1157		INIT_LIST_HEAD(&vbq->free);
 
 
 
1158	}
1159
1160	/* Import existing vmlist entries. */
1161	for (tmp = vmlist; tmp; tmp = tmp->next) {
1162		va = kzalloc(sizeof(struct vmap_area), GFP_NOWAIT);
1163		va->flags = tmp->flags | VM_VM_AREA;
1164		va->va_start = (unsigned long)tmp->addr;
1165		va->va_end = va->va_start + tmp->size;
 
1166		__insert_vmap_area(va);
1167	}
1168
1169	vmap_area_pcpu_hole = VMALLOC_END;
1170
1171	vmap_initialized = true;
1172}
1173
1174/**
1175 * map_kernel_range_noflush - map kernel VM area with the specified pages
1176 * @addr: start of the VM area to map
1177 * @size: size of the VM area to map
1178 * @prot: page protection flags to use
1179 * @pages: pages to map
1180 *
1181 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1182 * specify should have been allocated using get_vm_area() and its
1183 * friends.
1184 *
1185 * NOTE:
1186 * This function does NOT do any cache flushing.  The caller is
1187 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1188 * before calling this function.
1189 *
1190 * RETURNS:
1191 * The number of pages mapped on success, -errno on failure.
1192 */
1193int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1194			     pgprot_t prot, struct page **pages)
1195{
1196	return vmap_page_range_noflush(addr, addr + size, prot, pages);
1197}
1198
1199/**
1200 * unmap_kernel_range_noflush - unmap kernel VM area
1201 * @addr: start of the VM area to unmap
1202 * @size: size of the VM area to unmap
1203 *
1204 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size
1205 * specify should have been allocated using get_vm_area() and its
1206 * friends.
1207 *
1208 * NOTE:
1209 * This function does NOT do any cache flushing.  The caller is
1210 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1211 * before calling this function and flush_tlb_kernel_range() after.
1212 */
1213void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1214{
1215	vunmap_page_range(addr, addr + size);
1216}
1217EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1218
1219/**
1220 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1221 * @addr: start of the VM area to unmap
1222 * @size: size of the VM area to unmap
1223 *
1224 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1225 * the unmapping and tlb after.
1226 */
1227void unmap_kernel_range(unsigned long addr, unsigned long size)
1228{
1229	unsigned long end = addr + size;
1230
1231	flush_cache_vunmap(addr, end);
1232	vunmap_page_range(addr, end);
1233	flush_tlb_kernel_range(addr, end);
1234}
 
1235
1236int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
1237{
1238	unsigned long addr = (unsigned long)area->addr;
1239	unsigned long end = addr + area->size - PAGE_SIZE;
1240	int err;
1241
1242	err = vmap_page_range(addr, end, prot, *pages);
1243	if (err > 0) {
1244		*pages += err;
1245		err = 0;
1246	}
1247
1248	return err;
1249}
1250EXPORT_SYMBOL_GPL(map_vm_area);
1251
1252/*** Old vmalloc interfaces ***/
1253DEFINE_RWLOCK(vmlist_lock);
1254struct vm_struct *vmlist;
1255
1256static void insert_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
1257			      unsigned long flags, void *caller)
1258{
1259	struct vm_struct *tmp, **p;
1260
1261	vm->flags = flags;
1262	vm->addr = (void *)va->va_start;
1263	vm->size = va->va_end - va->va_start;
1264	vm->caller = caller;
1265	va->private = vm;
1266	va->flags |= VM_VM_AREA;
 
 
1267
1268	write_lock(&vmlist_lock);
1269	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1270		if (tmp->addr >= vm->addr)
1271			break;
1272	}
1273	vm->next = *p;
1274	*p = vm;
1275	write_unlock(&vmlist_lock);
 
1276}
1277
1278static struct vm_struct *__get_vm_area_node(unsigned long size,
1279		unsigned long align, unsigned long flags, unsigned long start,
1280		unsigned long end, int node, gfp_t gfp_mask, void *caller)
1281{
1282	static struct vmap_area *va;
1283	struct vm_struct *area;
1284
1285	BUG_ON(in_interrupt());
1286	if (flags & VM_IOREMAP) {
1287		int bit = fls(size);
1288
1289		if (bit > IOREMAP_MAX_ORDER)
1290			bit = IOREMAP_MAX_ORDER;
1291		else if (bit < PAGE_SHIFT)
1292			bit = PAGE_SHIFT;
1293
1294		align = 1ul << bit;
1295	}
1296
1297	size = PAGE_ALIGN(size);
1298	if (unlikely(!size))
1299		return NULL;
1300
1301	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
1302	if (unlikely(!area))
1303		return NULL;
1304
1305	/*
1306	 * We always allocate a guard page.
1307	 */
1308	size += PAGE_SIZE;
1309
1310	va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
1311	if (IS_ERR(va)) {
1312		kfree(area);
1313		return NULL;
1314	}
1315
1316	insert_vmalloc_vm(area, va, flags, caller);
 
1317	return area;
1318}
1319
1320struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
1321				unsigned long start, unsigned long end)
1322{
1323	return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1324						__builtin_return_address(0));
1325}
1326EXPORT_SYMBOL_GPL(__get_vm_area);
1327
1328struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
1329				       unsigned long start, unsigned long end,
1330				       void *caller)
1331{
1332	return __get_vm_area_node(size, 1, flags, start, end, -1, GFP_KERNEL,
1333				  caller);
1334}
1335
1336/**
1337 *	get_vm_area  -  reserve a contiguous kernel virtual area
1338 *	@size:		size of the area
1339 *	@flags:		%VM_IOREMAP for I/O mappings or VM_ALLOC
1340 *
1341 *	Search an area of @size in the kernel virtual mapping area,
1342 *	and reserved it for out purposes.  Returns the area descriptor
1343 *	on success or %NULL on failure.
1344 */
1345struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
1346{
1347	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1348				-1, GFP_KERNEL, __builtin_return_address(0));
 
1349}
1350
1351struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
1352				void *caller)
1353{
1354	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
1355						-1, GFP_KERNEL, caller);
1356}
1357
1358static struct vm_struct *find_vm_area(const void *addr)
 
 
 
 
 
 
 
 
1359{
1360	struct vmap_area *va;
1361
1362	va = find_vmap_area((unsigned long)addr);
1363	if (va && va->flags & VM_VM_AREA)
1364		return va->private;
1365
1366	return NULL;
1367}
1368
1369/**
1370 *	remove_vm_area  -  find and remove a continuous kernel virtual area
1371 *	@addr:		base address
1372 *
1373 *	Search for the kernel VM area starting at @addr, and remove it.
1374 *	This function returns the found VM area, but using it is NOT safe
1375 *	on SMP machines, except for its size or flags.
1376 */
1377struct vm_struct *remove_vm_area(const void *addr)
1378{
1379	struct vmap_area *va;
1380
1381	va = find_vmap_area((unsigned long)addr);
1382	if (va && va->flags & VM_VM_AREA) {
1383		struct vm_struct *vm = va->private;
1384		struct vm_struct *tmp, **p;
1385		/*
1386		 * remove from list and disallow access to this vm_struct
1387		 * before unmap. (address range confliction is maintained by
1388		 * vmap.)
1389		 */
1390		write_lock(&vmlist_lock);
1391		for (p = &vmlist; (tmp = *p) != vm; p = &tmp->next)
1392			;
1393		*p = tmp->next;
1394		write_unlock(&vmlist_lock);
1395
1396		vmap_debug_free_range(va->va_start, va->va_end);
 
1397		free_unmap_vmap_area(va);
1398		vm->size -= PAGE_SIZE;
1399
1400		return vm;
1401	}
1402	return NULL;
1403}
1404
1405static void __vunmap(const void *addr, int deallocate_pages)
1406{
1407	struct vm_struct *area;
1408
1409	if (!addr)
1410		return;
1411
1412	if ((PAGE_SIZE-1) & (unsigned long)addr) {
1413		WARN(1, KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
1414		return;
1415	}
1416
1417	area = remove_vm_area(addr);
1418	if (unlikely(!area)) {
1419		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
1420				addr);
1421		return;
1422	}
1423
1424	debug_check_no_locks_freed(addr, area->size);
1425	debug_check_no_obj_freed(addr, area->size);
1426
1427	if (deallocate_pages) {
1428		int i;
1429
1430		for (i = 0; i < area->nr_pages; i++) {
1431			struct page *page = area->pages[i];
1432
1433			BUG_ON(!page);
1434			__free_page(page);
1435		}
1436
1437		if (area->flags & VM_VPAGES)
1438			vfree(area->pages);
1439		else
1440			kfree(area->pages);
1441	}
1442
1443	kfree(area);
1444	return;
1445}
1446
1447/**
1448 *	vfree  -  release memory allocated by vmalloc()
1449 *	@addr:		memory base address
1450 *
1451 *	Free the virtually continuous memory area starting at @addr, as
1452 *	obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
1453 *	NULL, no operation is performed.
1454 *
1455 *	Must not be called in interrupt context.
 
 
 
 
1456 */
1457void vfree(const void *addr)
1458{
1459	BUG_ON(in_interrupt());
1460
1461	kmemleak_free(addr);
1462
1463	__vunmap(addr, 1);
 
 
 
 
 
 
 
1464}
1465EXPORT_SYMBOL(vfree);
1466
1467/**
1468 *	vunmap  -  release virtual mapping obtained by vmap()
1469 *	@addr:		memory base address
1470 *
1471 *	Free the virtually contiguous memory area starting at @addr,
1472 *	which was created from the page array passed to vmap().
1473 *
1474 *	Must not be called in interrupt context.
1475 */
1476void vunmap(const void *addr)
1477{
1478	BUG_ON(in_interrupt());
1479	might_sleep();
1480	__vunmap(addr, 0);
 
1481}
1482EXPORT_SYMBOL(vunmap);
1483
1484/**
1485 *	vmap  -  map an array of pages into virtually contiguous space
1486 *	@pages:		array of page pointers
1487 *	@count:		number of pages to map
1488 *	@flags:		vm_area->flags
1489 *	@prot:		page protection for the mapping
1490 *
1491 *	Maps @count pages from @pages into contiguous kernel virtual
1492 *	space.
1493 */
1494void *vmap(struct page **pages, unsigned int count,
1495		unsigned long flags, pgprot_t prot)
1496{
1497	struct vm_struct *area;
1498
1499	might_sleep();
1500
1501	if (count > totalram_pages)
1502		return NULL;
1503
1504	area = get_vm_area_caller((count << PAGE_SHIFT), flags,
1505					__builtin_return_address(0));
1506	if (!area)
1507		return NULL;
1508
1509	if (map_vm_area(area, prot, &pages)) {
1510		vunmap(area->addr);
1511		return NULL;
1512	}
1513
1514	return area->addr;
1515}
1516EXPORT_SYMBOL(vmap);
1517
1518static void *__vmalloc_node(unsigned long size, unsigned long align,
1519			    gfp_t gfp_mask, pgprot_t prot,
1520			    int node, void *caller);
1521static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
1522				 pgprot_t prot, int node, void *caller)
1523{
1524	const int order = 0;
1525	struct page **pages;
1526	unsigned int nr_pages, array_size, i;
1527	gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
 
1528
1529	nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
1530	array_size = (nr_pages * sizeof(struct page *));
1531
1532	area->nr_pages = nr_pages;
1533	/* Please note that the recursion is strictly bounded. */
1534	if (array_size > PAGE_SIZE) {
1535		pages = __vmalloc_node(array_size, 1, nested_gfp|__GFP_HIGHMEM,
1536				PAGE_KERNEL, node, caller);
1537		area->flags |= VM_VPAGES;
1538	} else {
1539		pages = kmalloc_node(array_size, nested_gfp, node);
1540	}
1541	area->pages = pages;
1542	area->caller = caller;
1543	if (!area->pages) {
1544		remove_vm_area(area->addr);
1545		kfree(area);
1546		return NULL;
1547	}
1548
1549	for (i = 0; i < area->nr_pages; i++) {
1550		struct page *page;
1551		gfp_t tmp_mask = gfp_mask | __GFP_NOWARN;
1552
1553		if (node < 0)
1554			page = alloc_page(tmp_mask);
1555		else
1556			page = alloc_pages_node(node, tmp_mask, order);
1557
1558		if (unlikely(!page)) {
1559			/* Successfully allocated i pages, free them in __vunmap() */
1560			area->nr_pages = i;
1561			goto fail;
1562		}
1563		area->pages[i] = page;
 
 
1564	}
1565
1566	if (map_vm_area(area, prot, &pages))
1567		goto fail;
1568	return area->addr;
1569
1570fail:
1571	warn_alloc_failed(gfp_mask, order, "vmalloc: allocation failure, "
1572			  "allocated %ld of %ld bytes\n",
1573			  (area->nr_pages*PAGE_SIZE), area->size);
1574	vfree(area->addr);
1575	return NULL;
1576}
1577
1578/**
1579 *	__vmalloc_node_range  -  allocate virtually contiguous memory
1580 *	@size:		allocation size
1581 *	@align:		desired alignment
1582 *	@start:		vm area range start
1583 *	@end:		vm area range end
1584 *	@gfp_mask:	flags for the page level allocator
1585 *	@prot:		protection mask for the allocated pages
1586 *	@node:		node to use for allocation or -1
 
1587 *	@caller:	caller's return address
1588 *
1589 *	Allocate enough pages to cover @size from the page level
1590 *	allocator with @gfp_mask flags.  Map them into contiguous
1591 *	kernel virtual space, using a pagetable protection of @prot.
1592 */
1593void *__vmalloc_node_range(unsigned long size, unsigned long align,
1594			unsigned long start, unsigned long end, gfp_t gfp_mask,
1595			pgprot_t prot, int node, void *caller)
 
1596{
1597	struct vm_struct *area;
1598	void *addr;
1599	unsigned long real_size = size;
1600
1601	size = PAGE_ALIGN(size);
1602	if (!size || (size >> PAGE_SHIFT) > totalram_pages)
1603		return NULL;
1604
1605	area = __get_vm_area_node(size, align, VM_ALLOC, start, end, node,
1606				  gfp_mask, caller);
 
 
1607
1608	if (!area)
 
1609		return NULL;
1610
1611	addr = __vmalloc_area_node(area, gfp_mask, prot, node, caller);
 
 
 
 
 
1612
1613	/*
1614	 * A ref_count = 3 is needed because the vm_struct and vmap_area
1615	 * structures allocated in the __get_vm_area_node() function contain
1616	 * references to the virtual address of the vmalloc'ed block.
1617	 */
1618	kmemleak_alloc(addr, real_size, 3, gfp_mask);
1619
1620	return addr;
 
 
 
 
 
 
1621}
1622
1623/**
1624 *	__vmalloc_node  -  allocate virtually contiguous memory
1625 *	@size:		allocation size
1626 *	@align:		desired alignment
1627 *	@gfp_mask:	flags for the page level allocator
1628 *	@prot:		protection mask for the allocated pages
1629 *	@node:		node to use for allocation or -1
1630 *	@caller:	caller's return address
1631 *
1632 *	Allocate enough pages to cover @size from the page level
1633 *	allocator with @gfp_mask flags.  Map them into contiguous
1634 *	kernel virtual space, using a pagetable protection of @prot.
1635 */
1636static void *__vmalloc_node(unsigned long size, unsigned long align,
1637			    gfp_t gfp_mask, pgprot_t prot,
1638			    int node, void *caller)
1639{
1640	return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
1641				gfp_mask, prot, node, caller);
1642}
1643
1644void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
1645{
1646	return __vmalloc_node(size, 1, gfp_mask, prot, -1,
1647				__builtin_return_address(0));
1648}
1649EXPORT_SYMBOL(__vmalloc);
1650
1651static inline void *__vmalloc_node_flags(unsigned long size,
1652					int node, gfp_t flags)
1653{
1654	return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
1655					node, __builtin_return_address(0));
1656}
1657
1658/**
1659 *	vmalloc  -  allocate virtually contiguous memory
1660 *	@size:		allocation size
1661 *	Allocate enough pages to cover @size from the page level
1662 *	allocator and map them into contiguous kernel virtual space.
1663 *
1664 *	For tight control over page level allocator and protection flags
1665 *	use __vmalloc() instead.
1666 */
1667void *vmalloc(unsigned long size)
1668{
1669	return __vmalloc_node_flags(size, -1, GFP_KERNEL | __GFP_HIGHMEM);
 
1670}
1671EXPORT_SYMBOL(vmalloc);
1672
1673/**
1674 *	vzalloc - allocate virtually contiguous memory with zero fill
1675 *	@size:	allocation size
1676 *	Allocate enough pages to cover @size from the page level
1677 *	allocator and map them into contiguous kernel virtual space.
1678 *	The memory allocated is set to zero.
1679 *
1680 *	For tight control over page level allocator and protection flags
1681 *	use __vmalloc() instead.
1682 */
1683void *vzalloc(unsigned long size)
1684{
1685	return __vmalloc_node_flags(size, -1,
1686				GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1687}
1688EXPORT_SYMBOL(vzalloc);
1689
1690/**
1691 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
1692 * @size: allocation size
1693 *
1694 * The resulting memory area is zeroed so it can be mapped to userspace
1695 * without leaking data.
1696 */
1697void *vmalloc_user(unsigned long size)
1698{
1699	struct vm_struct *area;
1700	void *ret;
1701
1702	ret = __vmalloc_node(size, SHMLBA,
1703			     GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
1704			     PAGE_KERNEL, -1, __builtin_return_address(0));
 
1705	if (ret) {
1706		area = find_vm_area(ret);
1707		area->flags |= VM_USERMAP;
1708	}
1709	return ret;
1710}
1711EXPORT_SYMBOL(vmalloc_user);
1712
1713/**
1714 *	vmalloc_node  -  allocate memory on a specific node
1715 *	@size:		allocation size
1716 *	@node:		numa node
1717 *
1718 *	Allocate enough pages to cover @size from the page level
1719 *	allocator and map them into contiguous kernel virtual space.
1720 *
1721 *	For tight control over page level allocator and protection flags
1722 *	use __vmalloc() instead.
1723 */
1724void *vmalloc_node(unsigned long size, int node)
1725{
1726	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL,
1727					node, __builtin_return_address(0));
1728}
1729EXPORT_SYMBOL(vmalloc_node);
1730
1731/**
1732 * vzalloc_node - allocate memory on a specific node with zero fill
1733 * @size:	allocation size
1734 * @node:	numa node
1735 *
1736 * Allocate enough pages to cover @size from the page level
1737 * allocator and map them into contiguous kernel virtual space.
1738 * The memory allocated is set to zero.
1739 *
1740 * For tight control over page level allocator and protection flags
1741 * use __vmalloc_node() instead.
1742 */
1743void *vzalloc_node(unsigned long size, int node)
1744{
1745	return __vmalloc_node_flags(size, node,
1746			 GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
1747}
1748EXPORT_SYMBOL(vzalloc_node);
1749
1750#ifndef PAGE_KERNEL_EXEC
1751# define PAGE_KERNEL_EXEC PAGE_KERNEL
1752#endif
1753
1754/**
1755 *	vmalloc_exec  -  allocate virtually contiguous, executable memory
1756 *	@size:		allocation size
1757 *
1758 *	Kernel-internal function to allocate enough pages to cover @size
1759 *	the page level allocator and map them into contiguous and
1760 *	executable kernel virtual space.
1761 *
1762 *	For tight control over page level allocator and protection flags
1763 *	use __vmalloc() instead.
1764 */
1765
1766void *vmalloc_exec(unsigned long size)
1767{
1768	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC,
1769			      -1, __builtin_return_address(0));
1770}
1771
1772#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
1773#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
1774#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
1775#define GFP_VMALLOC32 GFP_DMA | GFP_KERNEL
1776#else
1777#define GFP_VMALLOC32 GFP_KERNEL
1778#endif
1779
1780/**
1781 *	vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
1782 *	@size:		allocation size
1783 *
1784 *	Allocate enough 32bit PA addressable pages to cover @size from the
1785 *	page level allocator and map them into contiguous kernel virtual space.
1786 */
1787void *vmalloc_32(unsigned long size)
1788{
1789	return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
1790			      -1, __builtin_return_address(0));
1791}
1792EXPORT_SYMBOL(vmalloc_32);
1793
1794/**
1795 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
1796 *	@size:		allocation size
1797 *
1798 * The resulting memory area is 32bit addressable and zeroed so it can be
1799 * mapped to userspace without leaking data.
1800 */
1801void *vmalloc_32_user(unsigned long size)
1802{
1803	struct vm_struct *area;
1804	void *ret;
1805
1806	ret = __vmalloc_node(size, 1, GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
1807			     -1, __builtin_return_address(0));
1808	if (ret) {
1809		area = find_vm_area(ret);
1810		area->flags |= VM_USERMAP;
1811	}
1812	return ret;
1813}
1814EXPORT_SYMBOL(vmalloc_32_user);
1815
1816/*
1817 * small helper routine , copy contents to buf from addr.
1818 * If the page is not present, fill zero.
1819 */
1820
1821static int aligned_vread(char *buf, char *addr, unsigned long count)
1822{
1823	struct page *p;
1824	int copied = 0;
1825
1826	while (count) {
1827		unsigned long offset, length;
1828
1829		offset = (unsigned long)addr & ~PAGE_MASK;
1830		length = PAGE_SIZE - offset;
1831		if (length > count)
1832			length = count;
1833		p = vmalloc_to_page(addr);
1834		/*
1835		 * To do safe access to this _mapped_ area, we need
1836		 * lock. But adding lock here means that we need to add
1837		 * overhead of vmalloc()/vfree() calles for this _debug_
1838		 * interface, rarely used. Instead of that, we'll use
1839		 * kmap() and get small overhead in this access function.
1840		 */
1841		if (p) {
1842			/*
1843			 * we can expect USER0 is not used (see vread/vwrite's
1844			 * function description)
1845			 */
1846			void *map = kmap_atomic(p, KM_USER0);
1847			memcpy(buf, map + offset, length);
1848			kunmap_atomic(map, KM_USER0);
1849		} else
1850			memset(buf, 0, length);
1851
1852		addr += length;
1853		buf += length;
1854		copied += length;
1855		count -= length;
1856	}
1857	return copied;
1858}
1859
1860static int aligned_vwrite(char *buf, char *addr, unsigned long count)
1861{
1862	struct page *p;
1863	int copied = 0;
1864
1865	while (count) {
1866		unsigned long offset, length;
1867
1868		offset = (unsigned long)addr & ~PAGE_MASK;
1869		length = PAGE_SIZE - offset;
1870		if (length > count)
1871			length = count;
1872		p = vmalloc_to_page(addr);
1873		/*
1874		 * To do safe access to this _mapped_ area, we need
1875		 * lock. But adding lock here means that we need to add
1876		 * overhead of vmalloc()/vfree() calles for this _debug_
1877		 * interface, rarely used. Instead of that, we'll use
1878		 * kmap() and get small overhead in this access function.
1879		 */
1880		if (p) {
1881			/*
1882			 * we can expect USER0 is not used (see vread/vwrite's
1883			 * function description)
1884			 */
1885			void *map = kmap_atomic(p, KM_USER0);
1886			memcpy(map + offset, buf, length);
1887			kunmap_atomic(map, KM_USER0);
1888		}
1889		addr += length;
1890		buf += length;
1891		copied += length;
1892		count -= length;
1893	}
1894	return copied;
1895}
1896
1897/**
1898 *	vread() -  read vmalloc area in a safe way.
1899 *	@buf:		buffer for reading data
1900 *	@addr:		vm address.
1901 *	@count:		number of bytes to be read.
1902 *
1903 *	Returns # of bytes which addr and buf should be increased.
1904 *	(same number to @count). Returns 0 if [addr...addr+count) doesn't
1905 *	includes any intersect with alive vmalloc area.
1906 *
1907 *	This function checks that addr is a valid vmalloc'ed area, and
1908 *	copy data from that area to a given buffer. If the given memory range
1909 *	of [addr...addr+count) includes some valid address, data is copied to
1910 *	proper area of @buf. If there are memory holes, they'll be zero-filled.
1911 *	IOREMAP area is treated as memory hole and no copy is done.
1912 *
1913 *	If [addr...addr+count) doesn't includes any intersects with alive
1914 *	vm_struct area, returns 0.
1915 *	@buf should be kernel's buffer. Because	this function uses KM_USER0,
1916 *	the caller should guarantee KM_USER0 is not used.
1917 *
1918 *	Note: In usual ops, vread() is never necessary because the caller
1919 *	should know vmalloc() area is valid and can use memcpy().
1920 *	This is for routines which have to access vmalloc area without
1921 *	any informaion, as /dev/kmem.
1922 *
1923 */
1924
1925long vread(char *buf, char *addr, unsigned long count)
1926{
1927	struct vm_struct *tmp;
 
1928	char *vaddr, *buf_start = buf;
1929	unsigned long buflen = count;
1930	unsigned long n;
1931
1932	/* Don't allow overflow */
1933	if ((unsigned long) addr + count < count)
1934		count = -(unsigned long) addr;
1935
1936	read_lock(&vmlist_lock);
1937	for (tmp = vmlist; count && tmp; tmp = tmp->next) {
1938		vaddr = (char *) tmp->addr;
1939		if (addr >= vaddr + tmp->size - PAGE_SIZE)
 
 
 
 
 
 
 
1940			continue;
1941		while (addr < vaddr) {
1942			if (count == 0)
1943				goto finished;
1944			*buf = '\0';
1945			buf++;
1946			addr++;
1947			count--;
1948		}
1949		n = vaddr + tmp->size - PAGE_SIZE - addr;
1950		if (n > count)
1951			n = count;
1952		if (!(tmp->flags & VM_IOREMAP))
1953			aligned_vread(buf, addr, n);
1954		else /* IOREMAP area is treated as memory hole */
1955			memset(buf, 0, n);
1956		buf += n;
1957		addr += n;
1958		count -= n;
1959	}
1960finished:
1961	read_unlock(&vmlist_lock);
1962
1963	if (buf == buf_start)
1964		return 0;
1965	/* zero-fill memory holes */
1966	if (buf != buf_start + buflen)
1967		memset(buf, 0, buflen - (buf - buf_start));
1968
1969	return buflen;
1970}
1971
1972/**
1973 *	vwrite() -  write vmalloc area in a safe way.
1974 *	@buf:		buffer for source data
1975 *	@addr:		vm address.
1976 *	@count:		number of bytes to be read.
1977 *
1978 *	Returns # of bytes which addr and buf should be incresed.
1979 *	(same number to @count).
1980 *	If [addr...addr+count) doesn't includes any intersect with valid
1981 *	vmalloc area, returns 0.
1982 *
1983 *	This function checks that addr is a valid vmalloc'ed area, and
1984 *	copy data from a buffer to the given addr. If specified range of
1985 *	[addr...addr+count) includes some valid address, data is copied from
1986 *	proper area of @buf. If there are memory holes, no copy to hole.
1987 *	IOREMAP area is treated as memory hole and no copy is done.
1988 *
1989 *	If [addr...addr+count) doesn't includes any intersects with alive
1990 *	vm_struct area, returns 0.
1991 *	@buf should be kernel's buffer. Because	this function uses KM_USER0,
1992 *	the caller should guarantee KM_USER0 is not used.
1993 *
1994 *	Note: In usual ops, vwrite() is never necessary because the caller
1995 *	should know vmalloc() area is valid and can use memcpy().
1996 *	This is for routines which have to access vmalloc area without
1997 *	any informaion, as /dev/kmem.
1998 */
1999
2000long vwrite(char *buf, char *addr, unsigned long count)
2001{
2002	struct vm_struct *tmp;
 
2003	char *vaddr;
2004	unsigned long n, buflen;
2005	int copied = 0;
2006
2007	/* Don't allow overflow */
2008	if ((unsigned long) addr + count < count)
2009		count = -(unsigned long) addr;
2010	buflen = count;
2011
2012	read_lock(&vmlist_lock);
2013	for (tmp = vmlist; count && tmp; tmp = tmp->next) {
2014		vaddr = (char *) tmp->addr;
2015		if (addr >= vaddr + tmp->size - PAGE_SIZE)
 
 
 
 
 
 
 
2016			continue;
2017		while (addr < vaddr) {
2018			if (count == 0)
2019				goto finished;
2020			buf++;
2021			addr++;
2022			count--;
2023		}
2024		n = vaddr + tmp->size - PAGE_SIZE - addr;
2025		if (n > count)
2026			n = count;
2027		if (!(tmp->flags & VM_IOREMAP)) {
2028			aligned_vwrite(buf, addr, n);
2029			copied++;
2030		}
2031		buf += n;
2032		addr += n;
2033		count -= n;
2034	}
2035finished:
2036	read_unlock(&vmlist_lock);
2037	if (!copied)
2038		return 0;
2039	return buflen;
2040}
2041
2042/**
2043 *	remap_vmalloc_range  -  map vmalloc pages to userspace
2044 *	@vma:		vma to cover (map full range of vma)
2045 *	@addr:		vmalloc memory
2046 *	@pgoff:		number of pages into addr before first page to map
 
2047 *
2048 *	Returns:	0 for success, -Exxx on failure
2049 *
2050 *	This function checks that addr is a valid vmalloc'ed area, and
2051 *	that it is big enough to cover the vma. Will return failure if
2052 *	that criteria isn't met.
 
2053 *
2054 *	Similar to remap_pfn_range() (see mm/memory.c)
2055 */
2056int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
2057						unsigned long pgoff)
2058{
2059	struct vm_struct *area;
2060	unsigned long uaddr = vma->vm_start;
2061	unsigned long usize = vma->vm_end - vma->vm_start;
2062
2063	if ((PAGE_SIZE-1) & (unsigned long)addr)
 
 
2064		return -EINVAL;
2065
2066	area = find_vm_area(addr);
2067	if (!area)
2068		return -EINVAL;
2069
2070	if (!(area->flags & VM_USERMAP))
2071		return -EINVAL;
2072
2073	if (usize + (pgoff << PAGE_SHIFT) > area->size - PAGE_SIZE)
2074		return -EINVAL;
2075
2076	addr += pgoff << PAGE_SHIFT;
2077	do {
2078		struct page *page = vmalloc_to_page(addr);
2079		int ret;
2080
2081		ret = vm_insert_page(vma, uaddr, page);
2082		if (ret)
2083			return ret;
2084
2085		uaddr += PAGE_SIZE;
2086		addr += PAGE_SIZE;
2087		usize -= PAGE_SIZE;
2088	} while (usize > 0);
2089
2090	/* Prevent "things" like memory migration? VM_flags need a cleanup... */
2091	vma->vm_flags |= VM_RESERVED;
2092
2093	return 0;
2094}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2095EXPORT_SYMBOL(remap_vmalloc_range);
2096
2097/*
2098 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
2099 * have one.
2100 */
2101void  __attribute__((weak)) vmalloc_sync_all(void)
2102{
2103}
2104
2105
2106static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
2107{
2108	/* apply_to_page_range() does all the hard work. */
 
 
 
 
 
2109	return 0;
2110}
2111
2112/**
2113 *	alloc_vm_area - allocate a range of kernel address space
2114 *	@size:		size of the area
 
2115 *
2116 *	Returns:	NULL on failure, vm_struct on success
2117 *
2118 *	This function reserves a range of kernel address space, and
2119 *	allocates pagetables to map that range.  No actual mappings
2120 *	are created.  If the kernel address space is not shared
2121 *	between processes, it syncs the pagetable across all
2122 *	processes.
 
2123 */
2124struct vm_struct *alloc_vm_area(size_t size)
2125{
2126	struct vm_struct *area;
2127
2128	area = get_vm_area_caller(size, VM_IOREMAP,
2129				__builtin_return_address(0));
2130	if (area == NULL)
2131		return NULL;
2132
2133	/*
2134	 * This ensures that page tables are constructed for this region
2135	 * of kernel virtual address space and mapped into init_mm.
2136	 */
2137	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2138				area->size, f, NULL)) {
2139		free_vm_area(area);
2140		return NULL;
2141	}
2142
2143	/*
2144	 * If the allocated address space is passed to a hypercall
2145	 * before being used then we cannot rely on a page fault to
2146	 * trigger an update of the page tables.  So sync all the page
2147	 * tables here.
2148	 */
2149	vmalloc_sync_all();
2150
2151	return area;
2152}
2153EXPORT_SYMBOL_GPL(alloc_vm_area);
2154
2155void free_vm_area(struct vm_struct *area)
2156{
2157	struct vm_struct *ret;
2158	ret = remove_vm_area(area->addr);
2159	BUG_ON(ret != area);
2160	kfree(area);
2161}
2162EXPORT_SYMBOL_GPL(free_vm_area);
2163
2164#ifdef CONFIG_SMP
2165static struct vmap_area *node_to_va(struct rb_node *n)
2166{
2167	return n ? rb_entry(n, struct vmap_area, rb_node) : NULL;
2168}
2169
2170/**
2171 * pvm_find_next_prev - find the next and prev vmap_area surrounding @end
2172 * @end: target address
2173 * @pnext: out arg for the next vmap_area
2174 * @pprev: out arg for the previous vmap_area
2175 *
2176 * Returns: %true if either or both of next and prev are found,
2177 *	    %false if no vmap_area exists
2178 *
2179 * Find vmap_areas end addresses of which enclose @end.  ie. if not
2180 * NULL, *pnext->va_end > @end and *pprev->va_end <= @end.
2181 */
2182static bool pvm_find_next_prev(unsigned long end,
2183			       struct vmap_area **pnext,
2184			       struct vmap_area **pprev)
2185{
2186	struct rb_node *n = vmap_area_root.rb_node;
2187	struct vmap_area *va = NULL;
2188
2189	while (n) {
2190		va = rb_entry(n, struct vmap_area, rb_node);
2191		if (end < va->va_end)
2192			n = n->rb_left;
2193		else if (end > va->va_end)
2194			n = n->rb_right;
2195		else
2196			break;
2197	}
2198
2199	if (!va)
2200		return false;
2201
2202	if (va->va_end > end) {
2203		*pnext = va;
2204		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2205	} else {
2206		*pprev = va;
2207		*pnext = node_to_va(rb_next(&(*pprev)->rb_node));
2208	}
2209	return true;
2210}
2211
2212/**
2213 * pvm_determine_end - find the highest aligned address between two vmap_areas
2214 * @pnext: in/out arg for the next vmap_area
2215 * @pprev: in/out arg for the previous vmap_area
2216 * @align: alignment
2217 *
2218 * Returns: determined end address
2219 *
2220 * Find the highest aligned address between *@pnext and *@pprev below
2221 * VMALLOC_END.  *@pnext and *@pprev are adjusted so that the aligned
2222 * down address is between the end addresses of the two vmap_areas.
2223 *
2224 * Please note that the address returned by this function may fall
2225 * inside *@pnext vmap_area.  The caller is responsible for checking
2226 * that.
2227 */
2228static unsigned long pvm_determine_end(struct vmap_area **pnext,
2229				       struct vmap_area **pprev,
2230				       unsigned long align)
2231{
2232	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2233	unsigned long addr;
2234
2235	if (*pnext)
2236		addr = min((*pnext)->va_start & ~(align - 1), vmalloc_end);
2237	else
2238		addr = vmalloc_end;
2239
2240	while (*pprev && (*pprev)->va_end > addr) {
2241		*pnext = *pprev;
2242		*pprev = node_to_va(rb_prev(&(*pnext)->rb_node));
2243	}
2244
2245	return addr;
2246}
2247
2248/**
2249 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
2250 * @offsets: array containing offset of each area
2251 * @sizes: array containing size of each area
2252 * @nr_vms: the number of areas to allocate
2253 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
2254 *
2255 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
2256 *	    vm_structs on success, %NULL on failure
2257 *
2258 * Percpu allocator wants to use congruent vm areas so that it can
2259 * maintain the offsets among percpu areas.  This function allocates
2260 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
2261 * be scattered pretty far, distance between two areas easily going up
2262 * to gigabytes.  To avoid interacting with regular vmallocs, these
2263 * areas are allocated from top.
2264 *
2265 * Despite its complicated look, this allocator is rather simple.  It
2266 * does everything top-down and scans areas from the end looking for
2267 * matching slot.  While scanning, if any of the areas overlaps with
2268 * existing vmap_area, the base address is pulled down to fit the
2269 * area.  Scanning is repeated till all the areas fit and then all
2270 * necessary data structres are inserted and the result is returned.
2271 */
2272struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
2273				     const size_t *sizes, int nr_vms,
2274				     size_t align)
2275{
2276	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
2277	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
2278	struct vmap_area **vas, *prev, *next;
2279	struct vm_struct **vms;
2280	int area, area2, last_area, term_area;
2281	unsigned long base, start, end, last_end;
2282	bool purged = false;
2283
2284	/* verify parameters and allocate data structures */
2285	BUG_ON(align & ~PAGE_MASK || !is_power_of_2(align));
2286	for (last_area = 0, area = 0; area < nr_vms; area++) {
2287		start = offsets[area];
2288		end = start + sizes[area];
2289
2290		/* is everything aligned properly? */
2291		BUG_ON(!IS_ALIGNED(offsets[area], align));
2292		BUG_ON(!IS_ALIGNED(sizes[area], align));
2293
2294		/* detect the area with the highest address */
2295		if (start > offsets[last_area])
2296			last_area = area;
2297
2298		for (area2 = 0; area2 < nr_vms; area2++) {
2299			unsigned long start2 = offsets[area2];
2300			unsigned long end2 = start2 + sizes[area2];
2301
2302			if (area2 == area)
2303				continue;
2304
2305			BUG_ON(start2 >= start && start2 < end);
2306			BUG_ON(end2 <= end && end2 > start);
2307		}
2308	}
2309	last_end = offsets[last_area] + sizes[last_area];
2310
2311	if (vmalloc_end - vmalloc_start < last_end) {
2312		WARN_ON(true);
2313		return NULL;
2314	}
2315
2316	vms = kzalloc(sizeof(vms[0]) * nr_vms, GFP_KERNEL);
2317	vas = kzalloc(sizeof(vas[0]) * nr_vms, GFP_KERNEL);
2318	if (!vas || !vms)
2319		goto err_free;
2320
2321	for (area = 0; area < nr_vms; area++) {
2322		vas[area] = kzalloc(sizeof(struct vmap_area), GFP_KERNEL);
2323		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
2324		if (!vas[area] || !vms[area])
2325			goto err_free;
2326	}
2327retry:
2328	spin_lock(&vmap_area_lock);
2329
2330	/* start scanning - we scan from the top, begin with the last area */
2331	area = term_area = last_area;
2332	start = offsets[area];
2333	end = start + sizes[area];
2334
2335	if (!pvm_find_next_prev(vmap_area_pcpu_hole, &next, &prev)) {
2336		base = vmalloc_end - last_end;
2337		goto found;
2338	}
2339	base = pvm_determine_end(&next, &prev, align) - end;
2340
2341	while (true) {
2342		BUG_ON(next && next->va_end <= base + end);
2343		BUG_ON(prev && prev->va_end > base + end);
2344
2345		/*
2346		 * base might have underflowed, add last_end before
2347		 * comparing.
2348		 */
2349		if (base + last_end < vmalloc_start + last_end) {
2350			spin_unlock(&vmap_area_lock);
2351			if (!purged) {
2352				purge_vmap_area_lazy();
2353				purged = true;
2354				goto retry;
2355			}
2356			goto err_free;
2357		}
2358
2359		/*
2360		 * If next overlaps, move base downwards so that it's
2361		 * right below next and then recheck.
2362		 */
2363		if (next && next->va_start < base + end) {
2364			base = pvm_determine_end(&next, &prev, align) - end;
2365			term_area = area;
2366			continue;
2367		}
2368
2369		/*
2370		 * If prev overlaps, shift down next and prev and move
2371		 * base so that it's right below new next and then
2372		 * recheck.
2373		 */
2374		if (prev && prev->va_end > base + start)  {
2375			next = prev;
2376			prev = node_to_va(rb_prev(&next->rb_node));
2377			base = pvm_determine_end(&next, &prev, align) - end;
2378			term_area = area;
2379			continue;
2380		}
2381
2382		/*
2383		 * This area fits, move on to the previous one.  If
2384		 * the previous one is the terminal one, we're done.
2385		 */
2386		area = (area + nr_vms - 1) % nr_vms;
2387		if (area == term_area)
2388			break;
2389		start = offsets[area];
2390		end = start + sizes[area];
2391		pvm_find_next_prev(base + end, &next, &prev);
2392	}
2393found:
2394	/* we've found a fitting base, insert all va's */
2395	for (area = 0; area < nr_vms; area++) {
2396		struct vmap_area *va = vas[area];
2397
2398		va->va_start = base + offsets[area];
2399		va->va_end = va->va_start + sizes[area];
2400		__insert_vmap_area(va);
2401	}
2402
2403	vmap_area_pcpu_hole = base + offsets[last_area];
2404
2405	spin_unlock(&vmap_area_lock);
2406
2407	/* insert all vm's */
2408	for (area = 0; area < nr_vms; area++)
2409		insert_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
2410				  pcpu_get_vm_areas);
2411
2412	kfree(vas);
2413	return vms;
2414
2415err_free:
2416	for (area = 0; area < nr_vms; area++) {
2417		if (vas)
2418			kfree(vas[area]);
2419		if (vms)
2420			kfree(vms[area]);
2421	}
 
2422	kfree(vas);
2423	kfree(vms);
2424	return NULL;
2425}
2426
2427/**
2428 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
2429 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
2430 * @nr_vms: the number of allocated areas
2431 *
2432 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
2433 */
2434void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
2435{
2436	int i;
2437
2438	for (i = 0; i < nr_vms; i++)
2439		free_vm_area(vms[i]);
2440	kfree(vms);
2441}
2442#endif	/* CONFIG_SMP */
2443
2444#ifdef CONFIG_PROC_FS
2445static void *s_start(struct seq_file *m, loff_t *pos)
2446	__acquires(&vmlist_lock)
2447{
2448	loff_t n = *pos;
2449	struct vm_struct *v;
2450
2451	read_lock(&vmlist_lock);
2452	v = vmlist;
2453	while (n > 0 && v) {
2454		n--;
2455		v = v->next;
2456	}
2457	if (!n)
2458		return v;
2459
2460	return NULL;
2461
2462}
2463
2464static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2465{
2466	struct vm_struct *v = p;
2467
2468	++*pos;
2469	return v->next;
 
 
 
 
2470}
2471
2472static void s_stop(struct seq_file *m, void *p)
2473	__releases(&vmlist_lock)
2474{
2475	read_unlock(&vmlist_lock);
2476}
2477
2478static void show_numa_info(struct seq_file *m, struct vm_struct *v)
2479{
2480	if (NUMA_BUILD) {
2481		unsigned int nr, *counters = m->private;
2482
2483		if (!counters)
2484			return;
2485
 
 
 
 
 
2486		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
2487
2488		for (nr = 0; nr < v->nr_pages; nr++)
2489			counters[page_to_nid(v->pages[nr])]++;
2490
2491		for_each_node_state(nr, N_HIGH_MEMORY)
2492			if (counters[nr])
2493				seq_printf(m, " N%u=%u", nr, counters[nr]);
2494	}
2495}
2496
2497static int s_show(struct seq_file *m, void *p)
2498{
2499	struct vm_struct *v = p;
 
 
 
 
 
 
 
 
2500
2501	seq_printf(m, "0x%p-0x%p %7ld",
 
 
2502		v->addr, v->addr + v->size, v->size);
2503
2504	if (v->caller)
2505		seq_printf(m, " %pS", v->caller);
2506
2507	if (v->nr_pages)
2508		seq_printf(m, " pages=%d", v->nr_pages);
2509
2510	if (v->phys_addr)
2511		seq_printf(m, " phys=%llx", (unsigned long long)v->phys_addr);
2512
2513	if (v->flags & VM_IOREMAP)
2514		seq_printf(m, " ioremap");
2515
2516	if (v->flags & VM_ALLOC)
2517		seq_printf(m, " vmalloc");
2518
2519	if (v->flags & VM_MAP)
2520		seq_printf(m, " vmap");
2521
2522	if (v->flags & VM_USERMAP)
2523		seq_printf(m, " user");
2524
2525	if (v->flags & VM_VPAGES)
2526		seq_printf(m, " vpages");
2527
2528	show_numa_info(m, v);
2529	seq_putc(m, '\n');
2530	return 0;
2531}
2532
2533static const struct seq_operations vmalloc_op = {
2534	.start = s_start,
2535	.next = s_next,
2536	.stop = s_stop,
2537	.show = s_show,
2538};
2539
2540static int vmalloc_open(struct inode *inode, struct file *file)
2541{
2542	unsigned int *ptr = NULL;
2543	int ret;
2544
2545	if (NUMA_BUILD) {
2546		ptr = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
2547		if (ptr == NULL)
2548			return -ENOMEM;
2549	}
2550	ret = seq_open(file, &vmalloc_op);
2551	if (!ret) {
2552		struct seq_file *m = file->private_data;
2553		m->private = ptr;
2554	} else
2555		kfree(ptr);
2556	return ret;
2557}
2558
2559static const struct file_operations proc_vmalloc_operations = {
2560	.open		= vmalloc_open,
2561	.read		= seq_read,
2562	.llseek		= seq_lseek,
2563	.release	= seq_release_private,
2564};
2565
2566static int __init proc_vmalloc_init(void)
2567{
2568	proc_create("vmallocinfo", S_IRUSR, NULL, &proc_vmalloc_operations);
2569	return 0;
2570}
2571module_init(proc_vmalloc_init);
 
2572#endif
2573