Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
 
 
   3 *  Copyright (C) 1993  Linus Torvalds
   4 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   5 *  SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
   6 *  Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
   7 *  Numa awareness, Christoph Lameter, SGI, June 2005
   8 *  Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
   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/signal.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/set_memory.h>
  22#include <linux/debugobjects.h>
  23#include <linux/kallsyms.h>
  24#include <linux/list.h>
  25#include <linux/notifier.h>
  26#include <linux/rbtree.h>
  27#include <linux/xarray.h>
  28#include <linux/io.h>
  29#include <linux/rcupdate.h>
  30#include <linux/pfn.h>
  31#include <linux/kmemleak.h>
  32#include <linux/atomic.h>
  33#include <linux/compiler.h>
  34#include <linux/memcontrol.h>
  35#include <linux/llist.h>
  36#include <linux/bitops.h>
  37#include <linux/rbtree_augmented.h>
  38#include <linux/overflow.h>
  39#include <linux/pgtable.h>
  40#include <linux/uaccess.h>
  41#include <linux/hugetlb.h>
  42#include <linux/sched/mm.h>
  43#include <asm/tlbflush.h>
  44#include <asm/shmparam.h>
  45
  46#define CREATE_TRACE_POINTS
  47#include <trace/events/vmalloc.h>
  48
  49#include "internal.h"
  50#include "pgalloc-track.h"
  51
  52#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
  53static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
  54
  55static int __init set_nohugeiomap(char *str)
  56{
  57	ioremap_max_page_shift = PAGE_SHIFT;
  58	return 0;
  59}
  60early_param("nohugeiomap", set_nohugeiomap);
  61#else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
  62static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
  63#endif	/* CONFIG_HAVE_ARCH_HUGE_VMAP */
  64
  65#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
  66static bool __ro_after_init vmap_allow_huge = true;
  67
  68static int __init set_nohugevmalloc(char *str)
  69{
  70	vmap_allow_huge = false;
  71	return 0;
  72}
  73early_param("nohugevmalloc", set_nohugevmalloc);
  74#else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
  75static const bool vmap_allow_huge = false;
  76#endif	/* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
  77
  78bool is_vmalloc_addr(const void *x)
  79{
  80	unsigned long addr = (unsigned long)kasan_reset_tag(x);
  81
  82	return addr >= VMALLOC_START && addr < VMALLOC_END;
  83}
  84EXPORT_SYMBOL(is_vmalloc_addr);
  85
  86struct vfree_deferred {
  87	struct llist_head list;
  88	struct work_struct wq;
  89};
  90static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
  91
  92static void __vunmap(const void *, int);
  93
  94static void free_work(struct work_struct *w)
  95{
  96	struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
  97	struct llist_node *t, *llnode;
  98
  99	llist_for_each_safe(llnode, t, llist_del_all(&p->list))
 100		__vunmap((void *)llnode, 1);
 101}
 102
 103/*** Page table manipulation functions ***/
 104static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 105			phys_addr_t phys_addr, pgprot_t prot,
 106			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 107{
 108	pte_t *pte;
 109	u64 pfn;
 110	unsigned long size = PAGE_SIZE;
 111
 112	pfn = phys_addr >> PAGE_SHIFT;
 113	pte = pte_alloc_kernel_track(pmd, addr, mask);
 114	if (!pte)
 115		return -ENOMEM;
 116	do {
 117		BUG_ON(!pte_none(*pte));
 118
 119#ifdef CONFIG_HUGETLB_PAGE
 120		size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
 121		if (size != PAGE_SIZE) {
 122			pte_t entry = pfn_pte(pfn, prot);
 123
 124			entry = arch_make_huge_pte(entry, ilog2(size), 0);
 125			set_huge_pte_at(&init_mm, addr, pte, entry);
 126			pfn += PFN_DOWN(size);
 127			continue;
 128		}
 129#endif
 130		set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
 131		pfn++;
 132	} while (pte += PFN_DOWN(size), addr += size, addr != end);
 133	*mask |= PGTBL_PTE_MODIFIED;
 134	return 0;
 135}
 136
 137static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
 138			phys_addr_t phys_addr, pgprot_t prot,
 139			unsigned int max_page_shift)
 140{
 141	if (max_page_shift < PMD_SHIFT)
 142		return 0;
 143
 144	if (!arch_vmap_pmd_supported(prot))
 145		return 0;
 146
 147	if ((end - addr) != PMD_SIZE)
 148		return 0;
 149
 150	if (!IS_ALIGNED(addr, PMD_SIZE))
 151		return 0;
 152
 153	if (!IS_ALIGNED(phys_addr, PMD_SIZE))
 154		return 0;
 155
 156	if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
 157		return 0;
 158
 159	return pmd_set_huge(pmd, phys_addr, prot);
 160}
 161
 162static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 163			phys_addr_t phys_addr, pgprot_t prot,
 164			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 165{
 166	pmd_t *pmd;
 167	unsigned long next;
 168
 169	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 170	if (!pmd)
 171		return -ENOMEM;
 172	do {
 173		next = pmd_addr_end(addr, end);
 174
 175		if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
 176					max_page_shift)) {
 177			*mask |= PGTBL_PMD_MODIFIED;
 178			continue;
 179		}
 180
 181		if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
 182			return -ENOMEM;
 183	} while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
 184	return 0;
 185}
 186
 187static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
 188			phys_addr_t phys_addr, pgprot_t prot,
 189			unsigned int max_page_shift)
 190{
 191	if (max_page_shift < PUD_SHIFT)
 192		return 0;
 193
 194	if (!arch_vmap_pud_supported(prot))
 195		return 0;
 196
 197	if ((end - addr) != PUD_SIZE)
 198		return 0;
 199
 200	if (!IS_ALIGNED(addr, PUD_SIZE))
 201		return 0;
 202
 203	if (!IS_ALIGNED(phys_addr, PUD_SIZE))
 204		return 0;
 205
 206	if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
 207		return 0;
 208
 209	return pud_set_huge(pud, phys_addr, prot);
 210}
 211
 212static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 213			phys_addr_t phys_addr, pgprot_t prot,
 214			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 215{
 216	pud_t *pud;
 217	unsigned long next;
 218
 219	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 220	if (!pud)
 221		return -ENOMEM;
 222	do {
 223		next = pud_addr_end(addr, end);
 224
 225		if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
 226					max_page_shift)) {
 227			*mask |= PGTBL_PUD_MODIFIED;
 228			continue;
 229		}
 230
 231		if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
 232					max_page_shift, mask))
 233			return -ENOMEM;
 234	} while (pud++, phys_addr += (next - addr), addr = next, addr != end);
 235	return 0;
 236}
 237
 238static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
 239			phys_addr_t phys_addr, pgprot_t prot,
 240			unsigned int max_page_shift)
 241{
 242	if (max_page_shift < P4D_SHIFT)
 243		return 0;
 244
 245	if (!arch_vmap_p4d_supported(prot))
 246		return 0;
 247
 248	if ((end - addr) != P4D_SIZE)
 249		return 0;
 250
 251	if (!IS_ALIGNED(addr, P4D_SIZE))
 252		return 0;
 253
 254	if (!IS_ALIGNED(phys_addr, P4D_SIZE))
 255		return 0;
 256
 257	if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
 258		return 0;
 259
 260	return p4d_set_huge(p4d, phys_addr, prot);
 261}
 262
 263static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 264			phys_addr_t phys_addr, pgprot_t prot,
 265			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 266{
 267	p4d_t *p4d;
 268	unsigned long next;
 269
 270	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 271	if (!p4d)
 272		return -ENOMEM;
 273	do {
 274		next = p4d_addr_end(addr, end);
 275
 276		if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
 277					max_page_shift)) {
 278			*mask |= PGTBL_P4D_MODIFIED;
 279			continue;
 280		}
 281
 282		if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
 283					max_page_shift, mask))
 284			return -ENOMEM;
 285	} while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
 286	return 0;
 287}
 288
 289static int vmap_range_noflush(unsigned long addr, unsigned long end,
 290			phys_addr_t phys_addr, pgprot_t prot,
 291			unsigned int max_page_shift)
 292{
 293	pgd_t *pgd;
 294	unsigned long start;
 295	unsigned long next;
 296	int err;
 297	pgtbl_mod_mask mask = 0;
 298
 299	might_sleep();
 300	BUG_ON(addr >= end);
 301
 302	start = addr;
 303	pgd = pgd_offset_k(addr);
 304	do {
 305		next = pgd_addr_end(addr, end);
 306		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
 307					max_page_shift, &mask);
 308		if (err)
 309			break;
 310	} while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
 311
 312	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 313		arch_sync_kernel_mappings(start, end);
 314
 315	return err;
 316}
 317
 318int ioremap_page_range(unsigned long addr, unsigned long end,
 319		phys_addr_t phys_addr, pgprot_t prot)
 320{
 321	int err;
 322
 323	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
 324				 ioremap_max_page_shift);
 325	flush_cache_vmap(addr, end);
 326	if (!err)
 327		kmsan_ioremap_page_range(addr, end, phys_addr, prot,
 328					 ioremap_max_page_shift);
 329	return err;
 330}
 331
 332static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 333			     pgtbl_mod_mask *mask)
 334{
 335	pte_t *pte;
 336
 337	pte = pte_offset_kernel(pmd, addr);
 338	do {
 339		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
 340		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
 341	} while (pte++, addr += PAGE_SIZE, addr != end);
 342	*mask |= PGTBL_PTE_MODIFIED;
 343}
 344
 345static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 346			     pgtbl_mod_mask *mask)
 347{
 348	pmd_t *pmd;
 349	unsigned long next;
 350	int cleared;
 351
 352	pmd = pmd_offset(pud, addr);
 353	do {
 354		next = pmd_addr_end(addr, end);
 355
 356		cleared = pmd_clear_huge(pmd);
 357		if (cleared || pmd_bad(*pmd))
 358			*mask |= PGTBL_PMD_MODIFIED;
 359
 360		if (cleared)
 361			continue;
 362		if (pmd_none_or_clear_bad(pmd))
 363			continue;
 364		vunmap_pte_range(pmd, addr, next, mask);
 365
 366		cond_resched();
 367	} while (pmd++, addr = next, addr != end);
 368}
 369
 370static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 371			     pgtbl_mod_mask *mask)
 372{
 373	pud_t *pud;
 374	unsigned long next;
 375	int cleared;
 376
 377	pud = pud_offset(p4d, addr);
 378	do {
 379		next = pud_addr_end(addr, end);
 380
 381		cleared = pud_clear_huge(pud);
 382		if (cleared || pud_bad(*pud))
 383			*mask |= PGTBL_PUD_MODIFIED;
 384
 385		if (cleared)
 386			continue;
 387		if (pud_none_or_clear_bad(pud))
 388			continue;
 389		vunmap_pmd_range(pud, addr, next, mask);
 390	} while (pud++, addr = next, addr != end);
 391}
 392
 393static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 394			     pgtbl_mod_mask *mask)
 395{
 396	p4d_t *p4d;
 397	unsigned long next;
 
 398
 399	p4d = p4d_offset(pgd, addr);
 400	do {
 401		next = p4d_addr_end(addr, end);
 402
 403		p4d_clear_huge(p4d);
 404		if (p4d_bad(*p4d))
 405			*mask |= PGTBL_P4D_MODIFIED;
 406
 
 
 407		if (p4d_none_or_clear_bad(p4d))
 408			continue;
 409		vunmap_pud_range(p4d, addr, next, mask);
 410	} while (p4d++, addr = next, addr != end);
 411}
 412
 413/*
 414 * vunmap_range_noflush is similar to vunmap_range, but does not
 415 * flush caches or TLBs.
 
 416 *
 417 * The caller is responsible for calling flush_cache_vmap() before calling
 418 * this function, and flush_tlb_kernel_range after it has returned
 419 * successfully (and before the addresses are expected to cause a page fault
 420 * or be re-mapped for something else, if TLB flushes are being delayed or
 421 * coalesced).
 422 *
 423 * This is an internal function only. Do not use outside mm/.
 
 
 
 424 */
 425void __vunmap_range_noflush(unsigned long start, unsigned long end)
 426{
 
 427	unsigned long next;
 428	pgd_t *pgd;
 429	unsigned long addr = start;
 430	pgtbl_mod_mask mask = 0;
 431
 432	BUG_ON(addr >= end);
 433	pgd = pgd_offset_k(addr);
 434	do {
 435		next = pgd_addr_end(addr, end);
 436		if (pgd_bad(*pgd))
 437			mask |= PGTBL_PGD_MODIFIED;
 438		if (pgd_none_or_clear_bad(pgd))
 439			continue;
 440		vunmap_p4d_range(pgd, addr, next, &mask);
 441	} while (pgd++, addr = next, addr != end);
 442
 443	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 444		arch_sync_kernel_mappings(start, end);
 445}
 446
 447void vunmap_range_noflush(unsigned long start, unsigned long end)
 448{
 449	kmsan_vunmap_range_noflush(start, end);
 450	__vunmap_range_noflush(start, end);
 451}
 452
 453/**
 454 * vunmap_range - unmap kernel virtual addresses
 455 * @addr: start of the VM area to unmap
 456 * @end: end of the VM area to unmap (non-inclusive)
 457 *
 458 * Clears any present PTEs in the virtual address range, flushes TLBs and
 459 * caches. Any subsequent access to the address before it has been re-mapped
 460 * is a kernel bug.
 461 */
 462void vunmap_range(unsigned long addr, unsigned long end)
 463{
 464	flush_cache_vunmap(addr, end);
 465	vunmap_range_noflush(addr, end);
 466	flush_tlb_kernel_range(addr, end);
 467}
 468
 469static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
 470		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 471		pgtbl_mod_mask *mask)
 472{
 473	pte_t *pte;
 474
 475	/*
 476	 * nr is a running index into the array which helps higher level
 477	 * callers keep track of where we're up to.
 478	 */
 479
 480	pte = pte_alloc_kernel_track(pmd, addr, mask);
 481	if (!pte)
 482		return -ENOMEM;
 483	do {
 484		struct page *page = pages[*nr];
 485
 486		if (WARN_ON(!pte_none(*pte)))
 487			return -EBUSY;
 488		if (WARN_ON(!page))
 489			return -ENOMEM;
 490		if (WARN_ON(!pfn_valid(page_to_pfn(page))))
 491			return -EINVAL;
 492
 493		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 494		(*nr)++;
 495	} while (pte++, addr += PAGE_SIZE, addr != end);
 496	*mask |= PGTBL_PTE_MODIFIED;
 497	return 0;
 498}
 499
 500static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
 501		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 502		pgtbl_mod_mask *mask)
 503{
 504	pmd_t *pmd;
 505	unsigned long next;
 506
 507	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 508	if (!pmd)
 509		return -ENOMEM;
 510	do {
 511		next = pmd_addr_end(addr, end);
 512		if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
 513			return -ENOMEM;
 514	} while (pmd++, addr = next, addr != end);
 515	return 0;
 516}
 517
 518static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
 519		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 520		pgtbl_mod_mask *mask)
 521{
 522	pud_t *pud;
 523	unsigned long next;
 524
 525	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 526	if (!pud)
 527		return -ENOMEM;
 528	do {
 529		next = pud_addr_end(addr, end);
 530		if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
 531			return -ENOMEM;
 532	} while (pud++, addr = next, addr != end);
 533	return 0;
 534}
 535
 536static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
 537		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 538		pgtbl_mod_mask *mask)
 539{
 540	p4d_t *p4d;
 541	unsigned long next;
 542
 543	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 544	if (!p4d)
 545		return -ENOMEM;
 546	do {
 547		next = p4d_addr_end(addr, end);
 548		if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
 549			return -ENOMEM;
 550	} while (p4d++, addr = next, addr != end);
 551	return 0;
 552}
 553
 554static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
 555		pgprot_t prot, struct page **pages)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 556{
 557	unsigned long start = addr;
 558	pgd_t *pgd;
 559	unsigned long next;
 
 560	int err = 0;
 561	int nr = 0;
 562	pgtbl_mod_mask mask = 0;
 563
 564	BUG_ON(addr >= end);
 565	pgd = pgd_offset_k(addr);
 566	do {
 567		next = pgd_addr_end(addr, end);
 568		if (pgd_bad(*pgd))
 569			mask |= PGTBL_PGD_MODIFIED;
 570		err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
 571		if (err)
 572			return err;
 573	} while (pgd++, addr = next, addr != end);
 574
 575	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 576		arch_sync_kernel_mappings(start, end);
 577
 578	return 0;
 579}
 580
 581/*
 582 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
 583 * flush caches.
 584 *
 585 * The caller is responsible for calling flush_cache_vmap() after this
 586 * function returns successfully and before the addresses are accessed.
 587 *
 588 * This is an internal function only. Do not use outside mm/.
 589 */
 590int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 591		pgprot_t prot, struct page **pages, unsigned int page_shift)
 592{
 593	unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
 594
 595	WARN_ON(page_shift < PAGE_SHIFT);
 596
 597	if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
 598			page_shift == PAGE_SHIFT)
 599		return vmap_small_pages_range_noflush(addr, end, prot, pages);
 600
 601	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
 602		int err;
 603
 604		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
 605					page_to_phys(pages[i]), prot,
 606					page_shift);
 607		if (err)
 608			return err;
 609
 610		addr += 1UL << page_shift;
 611	}
 612
 613	return 0;
 614}
 615
 616int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 617		pgprot_t prot, struct page **pages, unsigned int page_shift)
 618{
 619	kmsan_vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 620	return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 621}
 622
 623/**
 624 * vmap_pages_range - map pages to a kernel virtual address
 625 * @addr: start of the VM area to map
 626 * @end: end of the VM area to map (non-inclusive)
 627 * @prot: page protection flags to use
 628 * @pages: pages to map (always PAGE_SIZE pages)
 629 * @page_shift: maximum shift that the pages may be mapped with, @pages must
 630 * be aligned and contiguous up to at least this shift.
 631 *
 632 * RETURNS:
 633 * 0 on success, -errno on failure.
 634 */
 635static int vmap_pages_range(unsigned long addr, unsigned long end,
 636		pgprot_t prot, struct page **pages, unsigned int page_shift)
 637{
 638	int err;
 639
 640	err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 641	flush_cache_vmap(addr, end);
 642	return err;
 643}
 644
 645int is_vmalloc_or_module_addr(const void *x)
 646{
 647	/*
 648	 * ARM, x86-64 and sparc64 put modules in a special place,
 649	 * and fall back on vmalloc() if that fails. Others
 650	 * just put it in the vmalloc space.
 651	 */
 652#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 653	unsigned long addr = (unsigned long)kasan_reset_tag(x);
 654	if (addr >= MODULES_VADDR && addr < MODULES_END)
 655		return 1;
 656#endif
 657	return is_vmalloc_addr(x);
 658}
 659
 660/*
 661 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
 662 * return the tail page that corresponds to the base page address, which
 663 * matches small vmap mappings.
 664 */
 665struct page *vmalloc_to_page(const void *vmalloc_addr)
 666{
 667	unsigned long addr = (unsigned long) vmalloc_addr;
 668	struct page *page = NULL;
 669	pgd_t *pgd = pgd_offset_k(addr);
 670	p4d_t *p4d;
 671	pud_t *pud;
 672	pmd_t *pmd;
 673	pte_t *ptep, pte;
 674
 675	/*
 676	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 677	 * architectures that do not vmalloc module space
 678	 */
 679	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 680
 681	if (pgd_none(*pgd))
 682		return NULL;
 683	if (WARN_ON_ONCE(pgd_leaf(*pgd)))
 684		return NULL; /* XXX: no allowance for huge pgd */
 685	if (WARN_ON_ONCE(pgd_bad(*pgd)))
 686		return NULL;
 687
 688	p4d = p4d_offset(pgd, addr);
 689	if (p4d_none(*p4d))
 690		return NULL;
 691	if (p4d_leaf(*p4d))
 692		return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
 693	if (WARN_ON_ONCE(p4d_bad(*p4d)))
 694		return NULL;
 695
 696	pud = pud_offset(p4d, addr);
 697	if (pud_none(*pud))
 698		return NULL;
 699	if (pud_leaf(*pud))
 700		return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 701	if (WARN_ON_ONCE(pud_bad(*pud)))
 702		return NULL;
 703
 704	pmd = pmd_offset(pud, addr);
 705	if (pmd_none(*pmd))
 
 
 
 
 
 
 
 
 706		return NULL;
 707	if (pmd_leaf(*pmd))
 708		return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 709	if (WARN_ON_ONCE(pmd_bad(*pmd)))
 710		return NULL;
 711
 712	ptep = pte_offset_map(pmd, addr);
 713	pte = *ptep;
 714	if (pte_present(pte))
 715		page = pte_page(pte);
 716	pte_unmap(ptep);
 717
 718	return page;
 719}
 720EXPORT_SYMBOL(vmalloc_to_page);
 721
 722/*
 723 * Map a vmalloc()-space virtual address to the physical page frame number.
 724 */
 725unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 726{
 727	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 728}
 729EXPORT_SYMBOL(vmalloc_to_pfn);
 730
 731
 732/*** Global kva allocator ***/
 733
 734#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
 735#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
 736
 737
 738static DEFINE_SPINLOCK(vmap_area_lock);
 739static DEFINE_SPINLOCK(free_vmap_area_lock);
 740/* Export for kexec only */
 741LIST_HEAD(vmap_area_list);
 
 742static struct rb_root vmap_area_root = RB_ROOT;
 743static bool vmap_initialized __read_mostly;
 744
 745static struct rb_root purge_vmap_area_root = RB_ROOT;
 746static LIST_HEAD(purge_vmap_area_list);
 747static DEFINE_SPINLOCK(purge_vmap_area_lock);
 748
 749/*
 750 * This kmem_cache is used for vmap_area objects. Instead of
 751 * allocating from slab we reuse an object from this cache to
 752 * make things faster. Especially in "no edge" splitting of
 753 * free block.
 754 */
 755static struct kmem_cache *vmap_area_cachep;
 756
 757/*
 758 * This linked list is used in pair with free_vmap_area_root.
 759 * It gives O(1) access to prev/next to perform fast coalescing.
 760 */
 761static LIST_HEAD(free_vmap_area_list);
 762
 763/*
 764 * This augment red-black tree represents the free vmap space.
 765 * All vmap_area objects in this tree are sorted by va->va_start
 766 * address. It is used for allocation and merging when a vmap
 767 * object is released.
 768 *
 769 * Each vmap_area node contains a maximum available free block
 770 * of its sub-tree, right or left. Therefore it is possible to
 771 * find a lowest match of free area.
 772 */
 773static struct rb_root free_vmap_area_root = RB_ROOT;
 774
 775/*
 776 * Preload a CPU with one object for "no edge" split case. The
 777 * aim is to get rid of allocations from the atomic context, thus
 778 * to use more permissive allocation masks.
 779 */
 780static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
 781
 782static __always_inline unsigned long
 783va_size(struct vmap_area *va)
 784{
 785	return (va->va_end - va->va_start);
 786}
 787
 788static __always_inline unsigned long
 789get_subtree_max_size(struct rb_node *node)
 790{
 791	struct vmap_area *va;
 792
 793	va = rb_entry_safe(node, struct vmap_area, rb_node);
 794	return va ? va->subtree_max_size : 0;
 795}
 796
 
 
 
 
 
 
 
 
 
 
 
 797RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
 798	struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
 799
 800static void purge_vmap_area_lazy(void);
 801static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
 802static void drain_vmap_area_work(struct work_struct *work);
 803static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
 804
 805static atomic_long_t nr_vmalloc_pages;
 806
 807unsigned long vmalloc_nr_pages(void)
 808{
 809	return atomic_long_read(&nr_vmalloc_pages);
 810}
 811
 812/* Look up the first VA which satisfies addr < va_end, NULL if none. */
 813static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
 814{
 815	struct vmap_area *va = NULL;
 816	struct rb_node *n = vmap_area_root.rb_node;
 817
 818	addr = (unsigned long)kasan_reset_tag((void *)addr);
 819
 820	while (n) {
 821		struct vmap_area *tmp;
 822
 823		tmp = rb_entry(n, struct vmap_area, rb_node);
 824		if (tmp->va_end > addr) {
 825			va = tmp;
 826			if (tmp->va_start <= addr)
 827				break;
 828
 829			n = n->rb_left;
 830		} else
 831			n = n->rb_right;
 832	}
 833
 834	return va;
 835}
 836
 837static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
 838{
 839	struct rb_node *n = root->rb_node;
 840
 841	addr = (unsigned long)kasan_reset_tag((void *)addr);
 842
 843	while (n) {
 844		struct vmap_area *va;
 845
 846		va = rb_entry(n, struct vmap_area, rb_node);
 847		if (addr < va->va_start)
 848			n = n->rb_left;
 849		else if (addr >= va->va_end)
 850			n = n->rb_right;
 851		else
 852			return va;
 853	}
 854
 855	return NULL;
 856}
 857
 858/*
 859 * This function returns back addresses of parent node
 860 * and its left or right link for further processing.
 861 *
 862 * Otherwise NULL is returned. In that case all further
 863 * steps regarding inserting of conflicting overlap range
 864 * have to be declined and actually considered as a bug.
 865 */
 866static __always_inline struct rb_node **
 867find_va_links(struct vmap_area *va,
 868	struct rb_root *root, struct rb_node *from,
 869	struct rb_node **parent)
 870{
 871	struct vmap_area *tmp_va;
 872	struct rb_node **link;
 873
 874	if (root) {
 875		link = &root->rb_node;
 876		if (unlikely(!*link)) {
 877			*parent = NULL;
 878			return link;
 879		}
 880	} else {
 881		link = &from;
 882	}
 883
 884	/*
 885	 * Go to the bottom of the tree. When we hit the last point
 886	 * we end up with parent rb_node and correct direction, i name
 887	 * it link, where the new va->rb_node will be attached to.
 888	 */
 889	do {
 890		tmp_va = rb_entry(*link, struct vmap_area, rb_node);
 891
 892		/*
 893		 * During the traversal we also do some sanity check.
 894		 * Trigger the BUG() if there are sides(left/right)
 895		 * or full overlaps.
 896		 */
 897		if (va->va_end <= tmp_va->va_start)
 
 898			link = &(*link)->rb_left;
 899		else if (va->va_start >= tmp_va->va_end)
 
 900			link = &(*link)->rb_right;
 901		else {
 902			WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
 903				va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
 904
 905			return NULL;
 906		}
 907	} while (*link);
 908
 909	*parent = &tmp_va->rb_node;
 910	return link;
 911}
 912
 913static __always_inline struct list_head *
 914get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
 915{
 916	struct list_head *list;
 917
 918	if (unlikely(!parent))
 919		/*
 920		 * The red-black tree where we try to find VA neighbors
 921		 * before merging or inserting is empty, i.e. it means
 922		 * there is no free vmap space. Normally it does not
 923		 * happen but we handle this case anyway.
 924		 */
 925		return NULL;
 926
 927	list = &rb_entry(parent, struct vmap_area, rb_node)->list;
 928	return (&parent->rb_right == link ? list->next : list);
 929}
 930
 931static __always_inline void
 932__link_va(struct vmap_area *va, struct rb_root *root,
 933	struct rb_node *parent, struct rb_node **link,
 934	struct list_head *head, bool augment)
 935{
 936	/*
 937	 * VA is still not in the list, but we can
 938	 * identify its future previous list_head node.
 939	 */
 940	if (likely(parent)) {
 941		head = &rb_entry(parent, struct vmap_area, rb_node)->list;
 942		if (&parent->rb_right != link)
 943			head = head->prev;
 944	}
 945
 946	/* Insert to the rb-tree */
 947	rb_link_node(&va->rb_node, parent, link);
 948	if (augment) {
 949		/*
 950		 * Some explanation here. Just perform simple insertion
 951		 * to the tree. We do not set va->subtree_max_size to
 952		 * its current size before calling rb_insert_augmented().
 953		 * It is because we populate the tree from the bottom
 954		 * to parent levels when the node _is_ in the tree.
 955		 *
 956		 * Therefore we set subtree_max_size to zero after insertion,
 957		 * to let __augment_tree_propagate_from() puts everything to
 958		 * the correct order later on.
 959		 */
 960		rb_insert_augmented(&va->rb_node,
 961			root, &free_vmap_area_rb_augment_cb);
 962		va->subtree_max_size = 0;
 963	} else {
 964		rb_insert_color(&va->rb_node, root);
 965	}
 966
 967	/* Address-sort this list */
 968	list_add(&va->list, head);
 969}
 970
 971static __always_inline void
 972link_va(struct vmap_area *va, struct rb_root *root,
 973	struct rb_node *parent, struct rb_node **link,
 974	struct list_head *head)
 975{
 976	__link_va(va, root, parent, link, head, false);
 977}
 978
 979static __always_inline void
 980link_va_augment(struct vmap_area *va, struct rb_root *root,
 981	struct rb_node *parent, struct rb_node **link,
 982	struct list_head *head)
 983{
 984	__link_va(va, root, parent, link, head, true);
 985}
 986
 987static __always_inline void
 988__unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
 989{
 990	if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
 991		return;
 992
 993	if (augment)
 994		rb_erase_augmented(&va->rb_node,
 995			root, &free_vmap_area_rb_augment_cb);
 996	else
 997		rb_erase(&va->rb_node, root);
 998
 999	list_del_init(&va->list);
1000	RB_CLEAR_NODE(&va->rb_node);
1001}
1002
1003static __always_inline void
1004unlink_va(struct vmap_area *va, struct rb_root *root)
1005{
1006	__unlink_va(va, root, false);
1007}
1008
1009static __always_inline void
1010unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1011{
1012	__unlink_va(va, root, true);
1013}
1014
1015#if DEBUG_AUGMENT_PROPAGATE_CHECK
1016/*
1017 * Gets called when remove the node and rotate.
1018 */
1019static __always_inline unsigned long
1020compute_subtree_max_size(struct vmap_area *va)
1021{
1022	return max3(va_size(va),
1023		get_subtree_max_size(va->rb_node.rb_left),
1024		get_subtree_max_size(va->rb_node.rb_right));
1025}
1026
1027static void
1028augment_tree_propagate_check(void)
1029{
1030	struct vmap_area *va;
1031	unsigned long computed_size;
1032
1033	list_for_each_entry(va, &free_vmap_area_list, list) {
1034		computed_size = compute_subtree_max_size(va);
1035		if (computed_size != va->subtree_max_size)
1036			pr_emerg("tree is corrupted: %lu, %lu\n",
1037				va_size(va), va->subtree_max_size);
1038	}
1039}
1040#endif
1041
1042/*
1043 * This function populates subtree_max_size from bottom to upper
1044 * levels starting from VA point. The propagation must be done
1045 * when VA size is modified by changing its va_start/va_end. Or
1046 * in case of newly inserting of VA to the tree.
1047 *
1048 * It means that __augment_tree_propagate_from() must be called:
1049 * - After VA has been inserted to the tree(free path);
1050 * - After VA has been shrunk(allocation path);
1051 * - After VA has been increased(merging path).
1052 *
1053 * Please note that, it does not mean that upper parent nodes
1054 * and their subtree_max_size are recalculated all the time up
1055 * to the root node.
1056 *
1057 *       4--8
1058 *        /\
1059 *       /  \
1060 *      /    \
1061 *    2--2  8--8
1062 *
1063 * For example if we modify the node 4, shrinking it to 2, then
1064 * no any modification is required. If we shrink the node 2 to 1
1065 * its subtree_max_size is updated only, and set to 1. If we shrink
1066 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1067 * node becomes 4--6.
1068 */
1069static __always_inline void
1070augment_tree_propagate_from(struct vmap_area *va)
1071{
1072	/*
1073	 * Populate the tree from bottom towards the root until
1074	 * the calculated maximum available size of checked node
1075	 * is equal to its current one.
1076	 */
1077	free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1078
1079#if DEBUG_AUGMENT_PROPAGATE_CHECK
1080	augment_tree_propagate_check();
1081#endif
1082}
1083
1084static void
1085insert_vmap_area(struct vmap_area *va,
1086	struct rb_root *root, struct list_head *head)
1087{
1088	struct rb_node **link;
1089	struct rb_node *parent;
1090
1091	link = find_va_links(va, root, NULL, &parent);
1092	if (link)
1093		link_va(va, root, parent, link, head);
1094}
1095
1096static void
1097insert_vmap_area_augment(struct vmap_area *va,
1098	struct rb_node *from, struct rb_root *root,
1099	struct list_head *head)
1100{
1101	struct rb_node **link;
1102	struct rb_node *parent;
1103
1104	if (from)
1105		link = find_va_links(va, NULL, from, &parent);
1106	else
1107		link = find_va_links(va, root, NULL, &parent);
1108
1109	if (link) {
1110		link_va_augment(va, root, parent, link, head);
1111		augment_tree_propagate_from(va);
1112	}
1113}
1114
1115/*
1116 * Merge de-allocated chunk of VA memory with previous
1117 * and next free blocks. If coalesce is not done a new
1118 * free area is inserted. If VA has been merged, it is
1119 * freed.
1120 *
1121 * Please note, it can return NULL in case of overlap
1122 * ranges, followed by WARN() report. Despite it is a
1123 * buggy behaviour, a system can be alive and keep
1124 * ongoing.
1125 */
1126static __always_inline struct vmap_area *
1127__merge_or_add_vmap_area(struct vmap_area *va,
1128	struct rb_root *root, struct list_head *head, bool augment)
1129{
1130	struct vmap_area *sibling;
1131	struct list_head *next;
1132	struct rb_node **link;
1133	struct rb_node *parent;
1134	bool merged = false;
1135
1136	/*
1137	 * Find a place in the tree where VA potentially will be
1138	 * inserted, unless it is merged with its sibling/siblings.
1139	 */
1140	link = find_va_links(va, root, NULL, &parent);
1141	if (!link)
1142		return NULL;
1143
1144	/*
1145	 * Get next node of VA to check if merging can be done.
1146	 */
1147	next = get_va_next_sibling(parent, link);
1148	if (unlikely(next == NULL))
1149		goto insert;
1150
1151	/*
1152	 * start            end
1153	 * |                |
1154	 * |<------VA------>|<-----Next----->|
1155	 *                  |                |
1156	 *                  start            end
1157	 */
1158	if (next != head) {
1159		sibling = list_entry(next, struct vmap_area, list);
1160		if (sibling->va_start == va->va_end) {
1161			sibling->va_start = va->va_start;
1162
1163			/* Free vmap_area object. */
1164			kmem_cache_free(vmap_area_cachep, va);
1165
1166			/* Point to the new merged area. */
1167			va = sibling;
1168			merged = true;
1169		}
1170	}
1171
1172	/*
1173	 * start            end
1174	 * |                |
1175	 * |<-----Prev----->|<------VA------>|
1176	 *                  |                |
1177	 *                  start            end
1178	 */
1179	if (next->prev != head) {
1180		sibling = list_entry(next->prev, struct vmap_area, list);
1181		if (sibling->va_end == va->va_start) {
1182			/*
1183			 * If both neighbors are coalesced, it is important
1184			 * to unlink the "next" node first, followed by merging
1185			 * with "previous" one. Otherwise the tree might not be
1186			 * fully populated if a sibling's augmented value is
1187			 * "normalized" because of rotation operations.
1188			 */
1189			if (merged)
1190				__unlink_va(va, root, augment);
1191
1192			sibling->va_end = va->va_end;
1193
1194			/* Free vmap_area object. */
1195			kmem_cache_free(vmap_area_cachep, va);
1196
1197			/* Point to the new merged area. */
1198			va = sibling;
1199			merged = true;
1200		}
1201	}
1202
1203insert:
1204	if (!merged)
1205		__link_va(va, root, parent, link, head, augment);
1206
1207	return va;
1208}
1209
1210static __always_inline struct vmap_area *
1211merge_or_add_vmap_area(struct vmap_area *va,
1212	struct rb_root *root, struct list_head *head)
1213{
1214	return __merge_or_add_vmap_area(va, root, head, false);
1215}
1216
1217static __always_inline struct vmap_area *
1218merge_or_add_vmap_area_augment(struct vmap_area *va,
1219	struct rb_root *root, struct list_head *head)
1220{
1221	va = __merge_or_add_vmap_area(va, root, head, true);
1222	if (va)
1223		augment_tree_propagate_from(va);
1224
 
 
 
 
1225	return va;
1226}
1227
1228static __always_inline bool
1229is_within_this_va(struct vmap_area *va, unsigned long size,
1230	unsigned long align, unsigned long vstart)
1231{
1232	unsigned long nva_start_addr;
1233
1234	if (va->va_start > vstart)
1235		nva_start_addr = ALIGN(va->va_start, align);
1236	else
1237		nva_start_addr = ALIGN(vstart, align);
1238
1239	/* Can be overflowed due to big size or alignment. */
1240	if (nva_start_addr + size < nva_start_addr ||
1241			nva_start_addr < vstart)
1242		return false;
1243
1244	return (nva_start_addr + size <= va->va_end);
1245}
1246
1247/*
1248 * Find the first free block(lowest start address) in the tree,
1249 * that will accomplish the request corresponding to passing
1250 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1251 * a search length is adjusted to account for worst case alignment
1252 * overhead.
1253 */
1254static __always_inline struct vmap_area *
1255find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1256	unsigned long align, unsigned long vstart, bool adjust_search_size)
1257{
1258	struct vmap_area *va;
1259	struct rb_node *node;
1260	unsigned long length;
1261
1262	/* Start from the root. */
1263	node = root->rb_node;
1264
1265	/* Adjust the search size for alignment overhead. */
1266	length = adjust_search_size ? size + align - 1 : size;
1267
1268	while (node) {
1269		va = rb_entry(node, struct vmap_area, rb_node);
1270
1271		if (get_subtree_max_size(node->rb_left) >= length &&
1272				vstart < va->va_start) {
1273			node = node->rb_left;
1274		} else {
1275			if (is_within_this_va(va, size, align, vstart))
1276				return va;
1277
1278			/*
1279			 * Does not make sense to go deeper towards the right
1280			 * sub-tree if it does not have a free block that is
1281			 * equal or bigger to the requested search length.
1282			 */
1283			if (get_subtree_max_size(node->rb_right) >= length) {
1284				node = node->rb_right;
1285				continue;
1286			}
1287
1288			/*
1289			 * OK. We roll back and find the first right sub-tree,
1290			 * that will satisfy the search criteria. It can happen
1291			 * due to "vstart" restriction or an alignment overhead
1292			 * that is bigger then PAGE_SIZE.
1293			 */
1294			while ((node = rb_parent(node))) {
1295				va = rb_entry(node, struct vmap_area, rb_node);
1296				if (is_within_this_va(va, size, align, vstart))
1297					return va;
1298
1299				if (get_subtree_max_size(node->rb_right) >= length &&
1300						vstart <= va->va_start) {
1301					/*
1302					 * Shift the vstart forward. Please note, we update it with
1303					 * parent's start address adding "1" because we do not want
1304					 * to enter same sub-tree after it has already been checked
1305					 * and no suitable free block found there.
1306					 */
1307					vstart = va->va_start + 1;
1308					node = node->rb_right;
1309					break;
1310				}
1311			}
1312		}
1313	}
1314
1315	return NULL;
1316}
1317
1318#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1319#include <linux/random.h>
1320
1321static struct vmap_area *
1322find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1323	unsigned long align, unsigned long vstart)
1324{
1325	struct vmap_area *va;
1326
1327	list_for_each_entry(va, head, list) {
1328		if (!is_within_this_va(va, size, align, vstart))
1329			continue;
1330
1331		return va;
1332	}
1333
1334	return NULL;
1335}
1336
1337static void
1338find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1339			     unsigned long size, unsigned long align)
1340{
1341	struct vmap_area *va_1, *va_2;
1342	unsigned long vstart;
1343	unsigned int rnd;
1344
1345	get_random_bytes(&rnd, sizeof(rnd));
1346	vstart = VMALLOC_START + rnd;
1347
1348	va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1349	va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1350
1351	if (va_1 != va_2)
1352		pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1353			va_1, va_2, vstart);
1354}
1355#endif
1356
1357enum fit_type {
1358	NOTHING_FIT = 0,
1359	FL_FIT_TYPE = 1,	/* full fit */
1360	LE_FIT_TYPE = 2,	/* left edge fit */
1361	RE_FIT_TYPE = 3,	/* right edge fit */
1362	NE_FIT_TYPE = 4		/* no edge fit */
1363};
1364
1365static __always_inline enum fit_type
1366classify_va_fit_type(struct vmap_area *va,
1367	unsigned long nva_start_addr, unsigned long size)
1368{
1369	enum fit_type type;
1370
1371	/* Check if it is within VA. */
1372	if (nva_start_addr < va->va_start ||
1373			nva_start_addr + size > va->va_end)
1374		return NOTHING_FIT;
1375
1376	/* Now classify. */
1377	if (va->va_start == nva_start_addr) {
1378		if (va->va_end == nva_start_addr + size)
1379			type = FL_FIT_TYPE;
1380		else
1381			type = LE_FIT_TYPE;
1382	} else if (va->va_end == nva_start_addr + size) {
1383		type = RE_FIT_TYPE;
1384	} else {
1385		type = NE_FIT_TYPE;
1386	}
1387
1388	return type;
1389}
1390
1391static __always_inline int
1392adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1393		      struct vmap_area *va, unsigned long nva_start_addr,
1394		      unsigned long size)
1395{
1396	struct vmap_area *lva = NULL;
1397	enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1398
1399	if (type == FL_FIT_TYPE) {
1400		/*
1401		 * No need to split VA, it fully fits.
1402		 *
1403		 * |               |
1404		 * V      NVA      V
1405		 * |---------------|
1406		 */
1407		unlink_va_augment(va, root);
1408		kmem_cache_free(vmap_area_cachep, va);
1409	} else if (type == LE_FIT_TYPE) {
1410		/*
1411		 * Split left edge of fit VA.
1412		 *
1413		 * |       |
1414		 * V  NVA  V   R
1415		 * |-------|-------|
1416		 */
1417		va->va_start += size;
1418	} else if (type == RE_FIT_TYPE) {
1419		/*
1420		 * Split right edge of fit VA.
1421		 *
1422		 *         |       |
1423		 *     L   V  NVA  V
1424		 * |-------|-------|
1425		 */
1426		va->va_end = nva_start_addr;
1427	} else if (type == NE_FIT_TYPE) {
1428		/*
1429		 * Split no edge of fit VA.
1430		 *
1431		 *     |       |
1432		 *   L V  NVA  V R
1433		 * |---|-------|---|
1434		 */
1435		lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1436		if (unlikely(!lva)) {
1437			/*
1438			 * For percpu allocator we do not do any pre-allocation
1439			 * and leave it as it is. The reason is it most likely
1440			 * never ends up with NE_FIT_TYPE splitting. In case of
1441			 * percpu allocations offsets and sizes are aligned to
1442			 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1443			 * are its main fitting cases.
1444			 *
1445			 * There are a few exceptions though, as an example it is
1446			 * a first allocation (early boot up) when we have "one"
1447			 * big free space that has to be split.
1448			 *
1449			 * Also we can hit this path in case of regular "vmap"
1450			 * allocations, if "this" current CPU was not preloaded.
1451			 * See the comment in alloc_vmap_area() why. If so, then
1452			 * GFP_NOWAIT is used instead to get an extra object for
1453			 * split purpose. That is rare and most time does not
1454			 * occur.
1455			 *
1456			 * What happens if an allocation gets failed. Basically,
1457			 * an "overflow" path is triggered to purge lazily freed
1458			 * areas to free some memory, then, the "retry" path is
1459			 * triggered to repeat one more time. See more details
1460			 * in alloc_vmap_area() function.
1461			 */
1462			lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1463			if (!lva)
1464				return -1;
1465		}
1466
1467		/*
1468		 * Build the remainder.
1469		 */
1470		lva->va_start = va->va_start;
1471		lva->va_end = nva_start_addr;
1472
1473		/*
1474		 * Shrink this VA to remaining size.
1475		 */
1476		va->va_start = nva_start_addr + size;
1477	} else {
1478		return -1;
1479	}
1480
1481	if (type != FL_FIT_TYPE) {
1482		augment_tree_propagate_from(va);
1483
1484		if (lva)	/* type == NE_FIT_TYPE */
1485			insert_vmap_area_augment(lva, &va->rb_node, root, head);
 
1486	}
1487
1488	return 0;
1489}
1490
1491/*
1492 * Returns a start address of the newly allocated area, if success.
1493 * Otherwise a vend is returned that indicates failure.
1494 */
1495static __always_inline unsigned long
1496__alloc_vmap_area(struct rb_root *root, struct list_head *head,
1497	unsigned long size, unsigned long align,
1498	unsigned long vstart, unsigned long vend)
1499{
1500	bool adjust_search_size = true;
1501	unsigned long nva_start_addr;
1502	struct vmap_area *va;
 
1503	int ret;
1504
1505	/*
1506	 * Do not adjust when:
1507	 *   a) align <= PAGE_SIZE, because it does not make any sense.
1508	 *      All blocks(their start addresses) are at least PAGE_SIZE
1509	 *      aligned anyway;
1510	 *   b) a short range where a requested size corresponds to exactly
1511	 *      specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1512	 *      With adjusted search length an allocation would not succeed.
1513	 */
1514	if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1515		adjust_search_size = false;
1516
1517	va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1518	if (unlikely(!va))
1519		return vend;
1520
1521	if (va->va_start > vstart)
1522		nva_start_addr = ALIGN(va->va_start, align);
1523	else
1524		nva_start_addr = ALIGN(vstart, align);
1525
1526	/* Check the "vend" restriction. */
1527	if (nva_start_addr + size > vend)
1528		return vend;
1529
 
 
 
 
 
1530	/* Update the free vmap_area. */
1531	ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1532	if (WARN_ON_ONCE(ret))
1533		return vend;
1534
1535#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1536	find_vmap_lowest_match_check(root, head, size, align);
1537#endif
1538
1539	return nva_start_addr;
1540}
1541
1542/*
1543 * Free a region of KVA allocated by alloc_vmap_area
1544 */
1545static void free_vmap_area(struct vmap_area *va)
1546{
1547	/*
1548	 * Remove from the busy tree/list.
1549	 */
1550	spin_lock(&vmap_area_lock);
1551	unlink_va(va, &vmap_area_root);
1552	spin_unlock(&vmap_area_lock);
1553
1554	/*
1555	 * Insert/Merge it back to the free tree/list.
1556	 */
1557	spin_lock(&free_vmap_area_lock);
1558	merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1559	spin_unlock(&free_vmap_area_lock);
1560}
1561
1562static inline void
1563preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1564{
1565	struct vmap_area *va = NULL;
1566
1567	/*
1568	 * Preload this CPU with one extra vmap_area object. It is used
1569	 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1570	 * a CPU that does an allocation is preloaded.
1571	 *
1572	 * We do it in non-atomic context, thus it allows us to use more
1573	 * permissive allocation masks to be more stable under low memory
1574	 * condition and high memory pressure.
1575	 */
1576	if (!this_cpu_read(ne_fit_preload_node))
1577		va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1578
1579	spin_lock(lock);
1580
1581	if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1582		kmem_cache_free(vmap_area_cachep, va);
1583}
1584
1585/*
1586 * Allocate a region of KVA of the specified size and alignment, within the
1587 * vstart and vend.
1588 */
1589static struct vmap_area *alloc_vmap_area(unsigned long size,
1590				unsigned long align,
1591				unsigned long vstart, unsigned long vend,
1592				int node, gfp_t gfp_mask)
1593{
1594	struct vmap_area *va;
1595	unsigned long freed;
1596	unsigned long addr;
1597	int purged = 0;
1598	int ret;
1599
1600	BUG_ON(!size);
1601	BUG_ON(offset_in_page(size));
1602	BUG_ON(!is_power_of_2(align));
1603
1604	if (unlikely(!vmap_initialized))
1605		return ERR_PTR(-EBUSY);
1606
1607	might_sleep();
1608	gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1609
1610	va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1611	if (unlikely(!va))
1612		return ERR_PTR(-ENOMEM);
1613
1614	/*
1615	 * Only scan the relevant parts containing pointers to other objects
1616	 * to avoid false negatives.
1617	 */
1618	kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1619
1620retry:
1621	preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1622	addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1623		size, align, vstart, vend);
1624	spin_unlock(&free_vmap_area_lock);
 
 
 
 
 
 
 
 
 
 
 
 
1625
1626	trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
 
 
 
 
 
 
 
 
 
 
 
1627
1628	/*
1629	 * If an allocation fails, the "vend" address is
1630	 * returned. Therefore trigger the overflow path.
1631	 */
 
 
 
1632	if (unlikely(addr == vend))
1633		goto overflow;
1634
1635	va->va_start = addr;
1636	va->va_end = addr + size;
1637	va->vm = NULL;
1638
 
1639	spin_lock(&vmap_area_lock);
1640	insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1641	spin_unlock(&vmap_area_lock);
1642
1643	BUG_ON(!IS_ALIGNED(va->va_start, align));
1644	BUG_ON(va->va_start < vstart);
1645	BUG_ON(va->va_end > vend);
1646
1647	ret = kasan_populate_vmalloc(addr, size);
1648	if (ret) {
1649		free_vmap_area(va);
1650		return ERR_PTR(ret);
1651	}
1652
1653	return va;
1654
1655overflow:
1656	if (!purged) {
1657		purge_vmap_area_lazy();
1658		purged = 1;
1659		goto retry;
1660	}
1661
1662	freed = 0;
1663	blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1664
1665	if (freed > 0) {
1666		purged = 0;
1667		goto retry;
 
1668	}
1669
1670	if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1671		pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1672			size);
1673
1674	kmem_cache_free(vmap_area_cachep, va);
1675	return ERR_PTR(-EBUSY);
1676}
1677
1678int register_vmap_purge_notifier(struct notifier_block *nb)
1679{
1680	return blocking_notifier_chain_register(&vmap_notify_list, nb);
1681}
1682EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1683
1684int unregister_vmap_purge_notifier(struct notifier_block *nb)
1685{
1686	return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1687}
1688EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1689
1690/*
1691 * lazy_max_pages is the maximum amount of virtual address space we gather up
1692 * before attempting to purge with a TLB flush.
1693 *
1694 * There is a tradeoff here: a larger number will cover more kernel page tables
1695 * and take slightly longer to purge, but it will linearly reduce the number of
1696 * global TLB flushes that must be performed. It would seem natural to scale
1697 * this number up linearly with the number of CPUs (because vmapping activity
1698 * could also scale linearly with the number of CPUs), however it is likely
1699 * that in practice, workloads might be constrained in other ways that mean
1700 * vmap activity will not scale linearly with CPUs. Also, I want to be
1701 * conservative and not introduce a big latency on huge systems, so go with
1702 * a less aggressive log scale. It will still be an improvement over the old
1703 * code, and it will be simple to change the scale factor if we find that it
1704 * becomes a problem on bigger systems.
1705 */
1706static unsigned long lazy_max_pages(void)
1707{
1708	unsigned int log;
1709
1710	log = fls(num_online_cpus());
1711
1712	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1713}
1714
1715static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1716
1717/*
1718 * Serialize vmap purging.  There is no actual critical section protected
1719 * by this lock, but we want to avoid concurrent calls for performance
1720 * reasons and to make the pcpu_get_vm_areas more deterministic.
1721 */
1722static DEFINE_MUTEX(vmap_purge_lock);
1723
1724/* for per-CPU blocks */
1725static void purge_fragmented_blocks_allcpus(void);
1726
1727/*
 
 
 
 
 
 
 
 
 
1728 * Purges all lazily-freed vmap areas.
1729 */
1730static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1731{
1732	unsigned long resched_threshold;
1733	unsigned int num_purged_areas = 0;
1734	struct list_head local_purge_list;
1735	struct vmap_area *va, *n_va;
1736
1737	lockdep_assert_held(&vmap_purge_lock);
1738
1739	spin_lock(&purge_vmap_area_lock);
1740	purge_vmap_area_root = RB_ROOT;
1741	list_replace_init(&purge_vmap_area_list, &local_purge_list);
1742	spin_unlock(&purge_vmap_area_lock);
1743
1744	if (unlikely(list_empty(&local_purge_list)))
1745		goto out;
1746
1747	start = min(start,
1748		list_first_entry(&local_purge_list,
1749			struct vmap_area, list)->va_start);
1750
1751	end = max(end,
1752		list_last_entry(&local_purge_list,
1753			struct vmap_area, list)->va_end);
1754
1755	flush_tlb_kernel_range(start, end);
1756	resched_threshold = lazy_max_pages() << 1;
1757
1758	spin_lock(&free_vmap_area_lock);
1759	list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1760		unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1761		unsigned long orig_start = va->va_start;
1762		unsigned long orig_end = va->va_end;
1763
1764		/*
1765		 * Finally insert or merge lazily-freed area. It is
1766		 * detached and there is no need to "unlink" it from
1767		 * anything.
1768		 */
1769		va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1770				&free_vmap_area_list);
1771
1772		if (!va)
1773			continue;
1774
1775		if (is_vmalloc_or_module_addr((void *)orig_start))
1776			kasan_release_vmalloc(orig_start, orig_end,
1777					      va->va_start, va->va_end);
1778
1779		atomic_long_sub(nr, &vmap_lazy_nr);
1780		num_purged_areas++;
1781
1782		if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1783			cond_resched_lock(&free_vmap_area_lock);
1784	}
1785	spin_unlock(&free_vmap_area_lock);
 
 
1786
1787out:
1788	trace_purge_vmap_area_lazy(start, end, num_purged_areas);
1789	return num_purged_areas > 0;
 
 
 
 
 
 
 
1790}
1791
1792/*
1793 * Kick off a purge of the outstanding lazy areas.
1794 */
1795static void purge_vmap_area_lazy(void)
1796{
1797	mutex_lock(&vmap_purge_lock);
1798	purge_fragmented_blocks_allcpus();
1799	__purge_vmap_area_lazy(ULONG_MAX, 0);
1800	mutex_unlock(&vmap_purge_lock);
1801}
1802
1803static void drain_vmap_area_work(struct work_struct *work)
1804{
1805	unsigned long nr_lazy;
1806
1807	do {
1808		mutex_lock(&vmap_purge_lock);
1809		__purge_vmap_area_lazy(ULONG_MAX, 0);
1810		mutex_unlock(&vmap_purge_lock);
1811
1812		/* Recheck if further work is required. */
1813		nr_lazy = atomic_long_read(&vmap_lazy_nr);
1814	} while (nr_lazy > lazy_max_pages());
1815}
1816
1817/*
1818 * Free a vmap area, caller ensuring that the area has been unmapped
1819 * and flush_cache_vunmap had been called for the correct range
1820 * previously.
1821 */
1822static void free_vmap_area_noflush(struct vmap_area *va)
1823{
1824	unsigned long nr_lazy_max = lazy_max_pages();
1825	unsigned long va_start = va->va_start;
1826	unsigned long nr_lazy;
1827
1828	spin_lock(&vmap_area_lock);
1829	unlink_va(va, &vmap_area_root);
1830	spin_unlock(&vmap_area_lock);
1831
1832	nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1833				PAGE_SHIFT, &vmap_lazy_nr);
1834
1835	/*
1836	 * Merge or place it to the purge tree/list.
1837	 */
1838	spin_lock(&purge_vmap_area_lock);
1839	merge_or_add_vmap_area(va,
1840		&purge_vmap_area_root, &purge_vmap_area_list);
1841	spin_unlock(&purge_vmap_area_lock);
1842
1843	trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
1844
1845	/* After this point, we may free va at any time */
1846	if (unlikely(nr_lazy > nr_lazy_max))
1847		schedule_work(&drain_vmap_work);
 
 
1848}
1849
1850/*
1851 * Free and unmap a vmap area
1852 */
1853static void free_unmap_vmap_area(struct vmap_area *va)
1854{
1855	flush_cache_vunmap(va->va_start, va->va_end);
1856	vunmap_range_noflush(va->va_start, va->va_end);
1857	if (debug_pagealloc_enabled_static())
1858		flush_tlb_kernel_range(va->va_start, va->va_end);
1859
1860	free_vmap_area_noflush(va);
1861}
1862
1863struct vmap_area *find_vmap_area(unsigned long addr)
1864{
1865	struct vmap_area *va;
1866
1867	spin_lock(&vmap_area_lock);
1868	va = __find_vmap_area(addr, &vmap_area_root);
1869	spin_unlock(&vmap_area_lock);
1870
1871	return va;
1872}
1873
1874/*** Per cpu kva allocator ***/
1875
1876/*
1877 * vmap space is limited especially on 32 bit architectures. Ensure there is
1878 * room for at least 16 percpu vmap blocks per CPU.
1879 */
1880/*
1881 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1882 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
1883 * instead (we just need a rough idea)
1884 */
1885#if BITS_PER_LONG == 32
1886#define VMALLOC_SPACE		(128UL*1024*1024)
1887#else
1888#define VMALLOC_SPACE		(128UL*1024*1024*1024)
1889#endif
1890
1891#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
1892#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
1893#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
1894#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
1895#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
1896#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
1897#define VMAP_BBMAP_BITS		\
1898		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
1899		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
1900			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1901
1902#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
1903
1904struct vmap_block_queue {
1905	spinlock_t lock;
1906	struct list_head free;
1907};
1908
1909struct vmap_block {
1910	spinlock_t lock;
1911	struct vmap_area *va;
1912	unsigned long free, dirty;
1913	unsigned long dirty_min, dirty_max; /*< dirty range */
1914	struct list_head free_list;
1915	struct rcu_head rcu_head;
1916	struct list_head purge;
1917};
1918
1919/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1920static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1921
1922/*
1923 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1924 * in the free path. Could get rid of this if we change the API to return a
1925 * "cookie" from alloc, to be passed to free. But no big deal yet.
1926 */
1927static DEFINE_XARRAY(vmap_blocks);
1928
1929/*
1930 * We should probably have a fallback mechanism to allocate virtual memory
1931 * out of partially filled vmap blocks. However vmap block sizing should be
1932 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1933 * big problem.
1934 */
1935
1936static unsigned long addr_to_vb_idx(unsigned long addr)
1937{
1938	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1939	addr /= VMAP_BLOCK_SIZE;
1940	return addr;
1941}
1942
1943static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1944{
1945	unsigned long addr;
1946
1947	addr = va_start + (pages_off << PAGE_SHIFT);
1948	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1949	return (void *)addr;
1950}
1951
1952/**
1953 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1954 *                  block. Of course pages number can't exceed VMAP_BBMAP_BITS
1955 * @order:    how many 2^order pages should be occupied in newly allocated block
1956 * @gfp_mask: flags for the page level allocator
1957 *
1958 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1959 */
1960static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1961{
1962	struct vmap_block_queue *vbq;
1963	struct vmap_block *vb;
1964	struct vmap_area *va;
1965	unsigned long vb_idx;
1966	int node, err;
1967	void *vaddr;
1968
1969	node = numa_node_id();
1970
1971	vb = kmalloc_node(sizeof(struct vmap_block),
1972			gfp_mask & GFP_RECLAIM_MASK, node);
1973	if (unlikely(!vb))
1974		return ERR_PTR(-ENOMEM);
1975
1976	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1977					VMALLOC_START, VMALLOC_END,
1978					node, gfp_mask);
1979	if (IS_ERR(va)) {
1980		kfree(vb);
1981		return ERR_CAST(va);
1982	}
1983
1984	vaddr = vmap_block_vaddr(va->va_start, 0);
1985	spin_lock_init(&vb->lock);
1986	vb->va = va;
1987	/* At least something should be left free */
1988	BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1989	vb->free = VMAP_BBMAP_BITS - (1UL << order);
1990	vb->dirty = 0;
1991	vb->dirty_min = VMAP_BBMAP_BITS;
1992	vb->dirty_max = 0;
1993	INIT_LIST_HEAD(&vb->free_list);
1994
1995	vb_idx = addr_to_vb_idx(va->va_start);
1996	err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1997	if (err) {
1998		kfree(vb);
1999		free_vmap_area(va);
2000		return ERR_PTR(err);
2001	}
2002
2003	vbq = raw_cpu_ptr(&vmap_block_queue);
2004	spin_lock(&vbq->lock);
2005	list_add_tail_rcu(&vb->free_list, &vbq->free);
2006	spin_unlock(&vbq->lock);
 
2007
2008	return vaddr;
2009}
2010
2011static void free_vmap_block(struct vmap_block *vb)
2012{
2013	struct vmap_block *tmp;
2014
2015	tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
2016	BUG_ON(tmp != vb);
2017
2018	free_vmap_area_noflush(vb->va);
2019	kfree_rcu(vb, rcu_head);
2020}
2021
2022static void purge_fragmented_blocks(int cpu)
2023{
2024	LIST_HEAD(purge);
2025	struct vmap_block *vb;
2026	struct vmap_block *n_vb;
2027	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2028
2029	rcu_read_lock();
2030	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2031
2032		if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
2033			continue;
2034
2035		spin_lock(&vb->lock);
2036		if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
2037			vb->free = 0; /* prevent further allocs after releasing lock */
2038			vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
2039			vb->dirty_min = 0;
2040			vb->dirty_max = VMAP_BBMAP_BITS;
2041			spin_lock(&vbq->lock);
2042			list_del_rcu(&vb->free_list);
2043			spin_unlock(&vbq->lock);
2044			spin_unlock(&vb->lock);
2045			list_add_tail(&vb->purge, &purge);
2046		} else
2047			spin_unlock(&vb->lock);
2048	}
2049	rcu_read_unlock();
2050
2051	list_for_each_entry_safe(vb, n_vb, &purge, purge) {
2052		list_del(&vb->purge);
2053		free_vmap_block(vb);
2054	}
2055}
2056
2057static void purge_fragmented_blocks_allcpus(void)
2058{
2059	int cpu;
2060
2061	for_each_possible_cpu(cpu)
2062		purge_fragmented_blocks(cpu);
2063}
2064
2065static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2066{
2067	struct vmap_block_queue *vbq;
2068	struct vmap_block *vb;
2069	void *vaddr = NULL;
2070	unsigned int order;
2071
2072	BUG_ON(offset_in_page(size));
2073	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2074	if (WARN_ON(size == 0)) {
2075		/*
2076		 * Allocating 0 bytes isn't what caller wants since
2077		 * get_order(0) returns funny result. Just warn and terminate
2078		 * early.
2079		 */
2080		return NULL;
2081	}
2082	order = get_order(size);
2083
2084	rcu_read_lock();
2085	vbq = raw_cpu_ptr(&vmap_block_queue);
2086	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2087		unsigned long pages_off;
2088
2089		spin_lock(&vb->lock);
2090		if (vb->free < (1UL << order)) {
2091			spin_unlock(&vb->lock);
2092			continue;
2093		}
2094
2095		pages_off = VMAP_BBMAP_BITS - vb->free;
2096		vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2097		vb->free -= 1UL << order;
2098		if (vb->free == 0) {
2099			spin_lock(&vbq->lock);
2100			list_del_rcu(&vb->free_list);
2101			spin_unlock(&vbq->lock);
2102		}
2103
2104		spin_unlock(&vb->lock);
2105		break;
2106	}
2107
 
2108	rcu_read_unlock();
2109
2110	/* Allocate new block if nothing was found */
2111	if (!vaddr)
2112		vaddr = new_vmap_block(order, gfp_mask);
2113
2114	return vaddr;
2115}
2116
2117static void vb_free(unsigned long addr, unsigned long size)
2118{
2119	unsigned long offset;
2120	unsigned int order;
2121	struct vmap_block *vb;
2122
2123	BUG_ON(offset_in_page(size));
2124	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2125
2126	flush_cache_vunmap(addr, addr + size);
2127
2128	order = get_order(size);
2129	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2130	vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
2131
2132	vunmap_range_noflush(addr, addr + size);
2133
2134	if (debug_pagealloc_enabled_static())
2135		flush_tlb_kernel_range(addr, addr + size);
2136
2137	spin_lock(&vb->lock);
2138
2139	/* Expand dirty range */
2140	vb->dirty_min = min(vb->dirty_min, offset);
2141	vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2142
2143	vb->dirty += 1UL << order;
2144	if (vb->dirty == VMAP_BBMAP_BITS) {
2145		BUG_ON(vb->free);
2146		spin_unlock(&vb->lock);
2147		free_vmap_block(vb);
2148	} else
2149		spin_unlock(&vb->lock);
2150}
2151
2152static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2153{
2154	int cpu;
2155
2156	if (unlikely(!vmap_initialized))
2157		return;
2158
2159	might_sleep();
2160
2161	for_each_possible_cpu(cpu) {
2162		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2163		struct vmap_block *vb;
2164
2165		rcu_read_lock();
2166		list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2167			spin_lock(&vb->lock);
2168			if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
2169				unsigned long va_start = vb->va->va_start;
2170				unsigned long s, e;
2171
2172				s = va_start + (vb->dirty_min << PAGE_SHIFT);
2173				e = va_start + (vb->dirty_max << PAGE_SHIFT);
2174
2175				start = min(s, start);
2176				end   = max(e, end);
2177
2178				flush = 1;
2179			}
2180			spin_unlock(&vb->lock);
2181		}
2182		rcu_read_unlock();
2183	}
2184
2185	mutex_lock(&vmap_purge_lock);
2186	purge_fragmented_blocks_allcpus();
2187	if (!__purge_vmap_area_lazy(start, end) && flush)
2188		flush_tlb_kernel_range(start, end);
2189	mutex_unlock(&vmap_purge_lock);
2190}
2191
2192/**
2193 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2194 *
2195 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2196 * to amortize TLB flushing overheads. What this means is that any page you
2197 * have now, may, in a former life, have been mapped into kernel virtual
2198 * address by the vmap layer and so there might be some CPUs with TLB entries
2199 * still referencing that page (additional to the regular 1:1 kernel mapping).
2200 *
2201 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2202 * be sure that none of the pages we have control over will have any aliases
2203 * from the vmap layer.
2204 */
2205void vm_unmap_aliases(void)
2206{
2207	unsigned long start = ULONG_MAX, end = 0;
2208	int flush = 0;
2209
2210	_vm_unmap_aliases(start, end, flush);
2211}
2212EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2213
2214/**
2215 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2216 * @mem: the pointer returned by vm_map_ram
2217 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2218 */
2219void vm_unmap_ram(const void *mem, unsigned int count)
2220{
2221	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2222	unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2223	struct vmap_area *va;
2224
2225	might_sleep();
2226	BUG_ON(!addr);
2227	BUG_ON(addr < VMALLOC_START);
2228	BUG_ON(addr > VMALLOC_END);
2229	BUG_ON(!PAGE_ALIGNED(addr));
2230
2231	kasan_poison_vmalloc(mem, size);
2232
2233	if (likely(count <= VMAP_MAX_ALLOC)) {
2234		debug_check_no_locks_freed(mem, size);
2235		vb_free(addr, size);
2236		return;
2237	}
2238
2239	va = find_vmap_area(addr);
2240	BUG_ON(!va);
2241	debug_check_no_locks_freed((void *)va->va_start,
2242				    (va->va_end - va->va_start));
2243	free_unmap_vmap_area(va);
2244}
2245EXPORT_SYMBOL(vm_unmap_ram);
2246
2247/**
2248 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2249 * @pages: an array of pointers to the pages to be mapped
2250 * @count: number of pages
2251 * @node: prefer to allocate data structures on this node
2252 *
2253 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2254 * faster than vmap so it's good.  But if you mix long-life and short-life
2255 * objects with vm_map_ram(), it could consume lots of address space through
2256 * fragmentation (especially on a 32bit machine).  You could see failures in
2257 * the end.  Please use this function for short-lived objects.
2258 *
2259 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2260 */
2261void *vm_map_ram(struct page **pages, unsigned int count, int node)
2262{
2263	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2264	unsigned long addr;
2265	void *mem;
2266
2267	if (likely(count <= VMAP_MAX_ALLOC)) {
2268		mem = vb_alloc(size, GFP_KERNEL);
2269		if (IS_ERR(mem))
2270			return NULL;
2271		addr = (unsigned long)mem;
2272	} else {
2273		struct vmap_area *va;
2274		va = alloc_vmap_area(size, PAGE_SIZE,
2275				VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
2276		if (IS_ERR(va))
2277			return NULL;
2278
2279		addr = va->va_start;
2280		mem = (void *)addr;
2281	}
2282
2283	if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2284				pages, PAGE_SHIFT) < 0) {
 
2285		vm_unmap_ram(mem, count);
2286		return NULL;
2287	}
2288
2289	/*
2290	 * Mark the pages as accessible, now that they are mapped.
2291	 * With hardware tag-based KASAN, marking is skipped for
2292	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2293	 */
2294	mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2295
2296	return mem;
2297}
2298EXPORT_SYMBOL(vm_map_ram);
2299
2300static struct vm_struct *vmlist __initdata;
2301
2302static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2303{
2304#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2305	return vm->page_order;
2306#else
2307	return 0;
2308#endif
2309}
2310
2311static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2312{
2313#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2314	vm->page_order = order;
2315#else
2316	BUG_ON(order != 0);
2317#endif
2318}
2319
2320/**
2321 * vm_area_add_early - add vmap area early during boot
2322 * @vm: vm_struct to add
2323 *
2324 * This function is used to add fixed kernel vm area to vmlist before
2325 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
2326 * should contain proper values and the other fields should be zero.
2327 *
2328 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2329 */
2330void __init vm_area_add_early(struct vm_struct *vm)
2331{
2332	struct vm_struct *tmp, **p;
2333
2334	BUG_ON(vmap_initialized);
2335	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2336		if (tmp->addr >= vm->addr) {
2337			BUG_ON(tmp->addr < vm->addr + vm->size);
2338			break;
2339		} else
2340			BUG_ON(tmp->addr + tmp->size > vm->addr);
2341	}
2342	vm->next = *p;
2343	*p = vm;
2344}
2345
2346/**
2347 * vm_area_register_early - register vmap area early during boot
2348 * @vm: vm_struct to register
2349 * @align: requested alignment
2350 *
2351 * This function is used to register kernel vm area before
2352 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
2353 * proper values on entry and other fields should be zero.  On return,
2354 * vm->addr contains the allocated address.
2355 *
2356 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2357 */
2358void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2359{
2360	unsigned long addr = ALIGN(VMALLOC_START, align);
2361	struct vm_struct *cur, **p;
2362
2363	BUG_ON(vmap_initialized);
2364
2365	for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2366		if ((unsigned long)cur->addr - addr >= vm->size)
2367			break;
2368		addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2369	}
2370
2371	BUG_ON(addr > VMALLOC_END - vm->size);
2372	vm->addr = (void *)addr;
2373	vm->next = *p;
2374	*p = vm;
2375	kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2376}
2377
2378static void vmap_init_free_space(void)
2379{
2380	unsigned long vmap_start = 1;
2381	const unsigned long vmap_end = ULONG_MAX;
2382	struct vmap_area *busy, *free;
2383
2384	/*
2385	 *     B     F     B     B     B     F
2386	 * -|-----|.....|-----|-----|-----|.....|-
2387	 *  |           The KVA space           |
2388	 *  |<--------------------------------->|
2389	 */
2390	list_for_each_entry(busy, &vmap_area_list, list) {
2391		if (busy->va_start - vmap_start > 0) {
2392			free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2393			if (!WARN_ON_ONCE(!free)) {
2394				free->va_start = vmap_start;
2395				free->va_end = busy->va_start;
2396
2397				insert_vmap_area_augment(free, NULL,
2398					&free_vmap_area_root,
2399						&free_vmap_area_list);
2400			}
2401		}
2402
2403		vmap_start = busy->va_end;
2404	}
2405
2406	if (vmap_end - vmap_start > 0) {
2407		free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2408		if (!WARN_ON_ONCE(!free)) {
2409			free->va_start = vmap_start;
2410			free->va_end = vmap_end;
2411
2412			insert_vmap_area_augment(free, NULL,
2413				&free_vmap_area_root,
2414					&free_vmap_area_list);
2415		}
2416	}
2417}
2418
2419void __init vmalloc_init(void)
2420{
2421	struct vmap_area *va;
2422	struct vm_struct *tmp;
2423	int i;
2424
2425	/*
2426	 * Create the cache for vmap_area objects.
2427	 */
2428	vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2429
2430	for_each_possible_cpu(i) {
2431		struct vmap_block_queue *vbq;
2432		struct vfree_deferred *p;
2433
2434		vbq = &per_cpu(vmap_block_queue, i);
2435		spin_lock_init(&vbq->lock);
2436		INIT_LIST_HEAD(&vbq->free);
2437		p = &per_cpu(vfree_deferred, i);
2438		init_llist_head(&p->list);
2439		INIT_WORK(&p->wq, free_work);
2440	}
2441
2442	/* Import existing vmlist entries. */
2443	for (tmp = vmlist; tmp; tmp = tmp->next) {
2444		va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2445		if (WARN_ON_ONCE(!va))
2446			continue;
2447
2448		va->va_start = (unsigned long)tmp->addr;
2449		va->va_end = va->va_start + tmp->size;
2450		va->vm = tmp;
2451		insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2452	}
2453
2454	/*
2455	 * Now we can initialize a free vmap space.
2456	 */
2457	vmap_init_free_space();
2458	vmap_initialized = true;
2459}
2460
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2461static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2462	struct vmap_area *va, unsigned long flags, const void *caller)
2463{
2464	vm->flags = flags;
2465	vm->addr = (void *)va->va_start;
2466	vm->size = va->va_end - va->va_start;
2467	vm->caller = caller;
2468	va->vm = vm;
2469}
2470
2471static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2472			      unsigned long flags, const void *caller)
2473{
2474	spin_lock(&vmap_area_lock);
2475	setup_vmalloc_vm_locked(vm, va, flags, caller);
2476	spin_unlock(&vmap_area_lock);
2477}
2478
2479static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2480{
2481	/*
2482	 * Before removing VM_UNINITIALIZED,
2483	 * we should make sure that vm has proper values.
2484	 * Pair with smp_rmb() in show_numa_info().
2485	 */
2486	smp_wmb();
2487	vm->flags &= ~VM_UNINITIALIZED;
2488}
2489
2490static struct vm_struct *__get_vm_area_node(unsigned long size,
2491		unsigned long align, unsigned long shift, unsigned long flags,
2492		unsigned long start, unsigned long end, int node,
2493		gfp_t gfp_mask, const void *caller)
2494{
2495	struct vmap_area *va;
2496	struct vm_struct *area;
2497	unsigned long requested_size = size;
2498
2499	BUG_ON(in_interrupt());
2500	size = ALIGN(size, 1ul << shift);
2501	if (unlikely(!size))
2502		return NULL;
2503
2504	if (flags & VM_IOREMAP)
2505		align = 1ul << clamp_t(int, get_count_order_long(size),
2506				       PAGE_SHIFT, IOREMAP_MAX_ORDER);
2507
2508	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2509	if (unlikely(!area))
2510		return NULL;
2511
2512	if (!(flags & VM_NO_GUARD))
2513		size += PAGE_SIZE;
2514
2515	va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2516	if (IS_ERR(va)) {
2517		kfree(area);
2518		return NULL;
2519	}
2520
2521	setup_vmalloc_vm(area, va, flags, caller);
2522
2523	/*
2524	 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2525	 * best-effort approach, as they can be mapped outside of vmalloc code.
2526	 * For VM_ALLOC mappings, the pages are marked as accessible after
2527	 * getting mapped in __vmalloc_node_range().
2528	 * With hardware tag-based KASAN, marking is skipped for
2529	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2530	 */
2531	if (!(flags & VM_ALLOC))
2532		area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2533						    KASAN_VMALLOC_PROT_NORMAL);
2534
2535	return area;
2536}
2537
2538struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2539				       unsigned long start, unsigned long end,
2540				       const void *caller)
2541{
2542	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2543				  NUMA_NO_NODE, GFP_KERNEL, caller);
2544}
2545
2546/**
2547 * get_vm_area - reserve a contiguous kernel virtual area
2548 * @size:	 size of the area
2549 * @flags:	 %VM_IOREMAP for I/O mappings or VM_ALLOC
2550 *
2551 * Search an area of @size in the kernel virtual mapping area,
2552 * and reserved it for out purposes.  Returns the area descriptor
2553 * on success or %NULL on failure.
2554 *
2555 * Return: the area descriptor on success or %NULL on failure.
2556 */
2557struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2558{
2559	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2560				  VMALLOC_START, VMALLOC_END,
2561				  NUMA_NO_NODE, GFP_KERNEL,
2562				  __builtin_return_address(0));
2563}
2564
2565struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2566				const void *caller)
2567{
2568	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2569				  VMALLOC_START, VMALLOC_END,
2570				  NUMA_NO_NODE, GFP_KERNEL, caller);
2571}
2572
2573/**
2574 * find_vm_area - find a continuous kernel virtual area
2575 * @addr:	  base address
2576 *
2577 * Search for the kernel VM area starting at @addr, and return it.
2578 * It is up to the caller to do all required locking to keep the returned
2579 * pointer valid.
2580 *
2581 * Return: the area descriptor on success or %NULL on failure.
2582 */
2583struct vm_struct *find_vm_area(const void *addr)
2584{
2585	struct vmap_area *va;
2586
2587	va = find_vmap_area((unsigned long)addr);
2588	if (!va)
2589		return NULL;
2590
2591	return va->vm;
2592}
2593
2594/**
2595 * remove_vm_area - find and remove a continuous kernel virtual area
2596 * @addr:	    base address
2597 *
2598 * Search for the kernel VM area starting at @addr, and remove it.
2599 * This function returns the found VM area, but using it is NOT safe
2600 * on SMP machines, except for its size or flags.
2601 *
2602 * Return: the area descriptor on success or %NULL on failure.
2603 */
2604struct vm_struct *remove_vm_area(const void *addr)
2605{
2606	struct vmap_area *va;
2607
2608	might_sleep();
2609
2610	spin_lock(&vmap_area_lock);
2611	va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
2612	if (va && va->vm) {
2613		struct vm_struct *vm = va->vm;
2614
2615		va->vm = NULL;
2616		spin_unlock(&vmap_area_lock);
2617
2618		kasan_free_module_shadow(vm);
2619		free_unmap_vmap_area(va);
2620
2621		return vm;
2622	}
2623
2624	spin_unlock(&vmap_area_lock);
2625	return NULL;
2626}
2627
2628static inline void set_area_direct_map(const struct vm_struct *area,
2629				       int (*set_direct_map)(struct page *page))
2630{
2631	int i;
2632
2633	/* HUGE_VMALLOC passes small pages to set_direct_map */
2634	for (i = 0; i < area->nr_pages; i++)
2635		if (page_address(area->pages[i]))
2636			set_direct_map(area->pages[i]);
2637}
2638
2639/* Handle removing and resetting vm mappings related to the vm_struct. */
2640static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2641{
2642	unsigned long start = ULONG_MAX, end = 0;
2643	unsigned int page_order = vm_area_page_order(area);
2644	int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2645	int flush_dmap = 0;
2646	int i;
2647
2648	remove_vm_area(area->addr);
2649
2650	/* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2651	if (!flush_reset)
2652		return;
2653
2654	/*
2655	 * If not deallocating pages, just do the flush of the VM area and
2656	 * return.
2657	 */
2658	if (!deallocate_pages) {
2659		vm_unmap_aliases();
2660		return;
2661	}
2662
2663	/*
2664	 * If execution gets here, flush the vm mapping and reset the direct
2665	 * map. Find the start and end range of the direct mappings to make sure
2666	 * the vm_unmap_aliases() flush includes the direct map.
2667	 */
2668	for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2669		unsigned long addr = (unsigned long)page_address(area->pages[i]);
2670		if (addr) {
2671			unsigned long page_size;
2672
2673			page_size = PAGE_SIZE << page_order;
2674			start = min(addr, start);
2675			end = max(addr + page_size, end);
2676			flush_dmap = 1;
2677		}
2678	}
2679
2680	/*
2681	 * Set direct map to something invalid so that it won't be cached if
2682	 * there are any accesses after the TLB flush, then flush the TLB and
2683	 * reset the direct map permissions to the default.
2684	 */
2685	set_area_direct_map(area, set_direct_map_invalid_noflush);
2686	_vm_unmap_aliases(start, end, flush_dmap);
2687	set_area_direct_map(area, set_direct_map_default_noflush);
2688}
2689
2690static void __vunmap(const void *addr, int deallocate_pages)
2691{
2692	struct vm_struct *area;
2693
2694	if (!addr)
2695		return;
2696
2697	if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2698			addr))
2699		return;
2700
2701	area = find_vm_area(addr);
2702	if (unlikely(!area)) {
2703		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2704				addr);
2705		return;
2706	}
2707
2708	debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2709	debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2710
2711	kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2712
2713	vm_remove_mappings(area, deallocate_pages);
2714
2715	if (deallocate_pages) {
2716		int i;
2717
2718		for (i = 0; i < area->nr_pages; i++) {
2719			struct page *page = area->pages[i];
2720
2721			BUG_ON(!page);
2722			mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2723			/*
2724			 * High-order allocs for huge vmallocs are split, so
2725			 * can be freed as an array of order-0 allocations
2726			 */
2727			__free_pages(page, 0);
2728			cond_resched();
2729		}
2730		atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2731
2732		kvfree(area->pages);
2733	}
2734
2735	kfree(area);
 
2736}
2737
2738static inline void __vfree_deferred(const void *addr)
2739{
2740	/*
2741	 * Use raw_cpu_ptr() because this can be called from preemptible
2742	 * context. Preemption is absolutely fine here, because the llist_add()
2743	 * implementation is lockless, so it works even if we are adding to
2744	 * another cpu's list. schedule_work() should be fine with this too.
2745	 */
2746	struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2747
2748	if (llist_add((struct llist_node *)addr, &p->list))
2749		schedule_work(&p->wq);
2750}
2751
2752/**
2753 * vfree_atomic - release memory allocated by vmalloc()
2754 * @addr:	  memory base address
2755 *
2756 * This one is just like vfree() but can be called in any atomic context
2757 * except NMIs.
2758 */
2759void vfree_atomic(const void *addr)
2760{
2761	BUG_ON(in_nmi());
2762
2763	kmemleak_free(addr);
2764
2765	if (!addr)
2766		return;
2767	__vfree_deferred(addr);
2768}
2769
2770static void __vfree(const void *addr)
2771{
2772	if (unlikely(in_interrupt()))
2773		__vfree_deferred(addr);
2774	else
2775		__vunmap(addr, 1);
2776}
2777
2778/**
2779 * vfree - Release memory allocated by vmalloc()
2780 * @addr:  Memory base address
2781 *
2782 * Free the virtually continuous memory area starting at @addr, as obtained
2783 * from one of the vmalloc() family of APIs.  This will usually also free the
2784 * physical memory underlying the virtual allocation, but that memory is
2785 * reference counted, so it will not be freed until the last user goes away.
2786 *
2787 * If @addr is NULL, no operation is performed.
 
 
 
 
 
 
2788 *
2789 * Context:
2790 * May sleep if called *not* from interrupt context.
2791 * Must not be called in NMI context (strictly speaking, it could be
2792 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2793 * conventions for vfree() arch-dependent would be a really bad idea).
2794 */
2795void vfree(const void *addr)
2796{
2797	BUG_ON(in_nmi());
2798
2799	kmemleak_free(addr);
2800
2801	might_sleep_if(!in_interrupt());
2802
2803	if (!addr)
2804		return;
2805
2806	__vfree(addr);
2807}
2808EXPORT_SYMBOL(vfree);
2809
2810/**
2811 * vunmap - release virtual mapping obtained by vmap()
2812 * @addr:   memory base address
2813 *
2814 * Free the virtually contiguous memory area starting at @addr,
2815 * which was created from the page array passed to vmap().
2816 *
2817 * Must not be called in interrupt context.
2818 */
2819void vunmap(const void *addr)
2820{
2821	BUG_ON(in_interrupt());
2822	might_sleep();
2823	if (addr)
2824		__vunmap(addr, 0);
2825}
2826EXPORT_SYMBOL(vunmap);
2827
2828/**
2829 * vmap - map an array of pages into virtually contiguous space
2830 * @pages: array of page pointers
2831 * @count: number of pages to map
2832 * @flags: vm_area->flags
2833 * @prot: page protection for the mapping
2834 *
2835 * Maps @count pages from @pages into contiguous kernel virtual space.
2836 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2837 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2838 * are transferred from the caller to vmap(), and will be freed / dropped when
2839 * vfree() is called on the return value.
2840 *
2841 * Return: the address of the area or %NULL on failure
2842 */
2843void *vmap(struct page **pages, unsigned int count,
2844	   unsigned long flags, pgprot_t prot)
2845{
2846	struct vm_struct *area;
2847	unsigned long addr;
2848	unsigned long size;		/* In bytes */
2849
2850	might_sleep();
2851
2852	/*
2853	 * Your top guard is someone else's bottom guard. Not having a top
2854	 * guard compromises someone else's mappings too.
2855	 */
2856	if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2857		flags &= ~VM_NO_GUARD;
2858
2859	if (count > totalram_pages())
2860		return NULL;
2861
2862	size = (unsigned long)count << PAGE_SHIFT;
2863	area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2864	if (!area)
2865		return NULL;
2866
2867	addr = (unsigned long)area->addr;
2868	if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2869				pages, PAGE_SHIFT) < 0) {
2870		vunmap(area->addr);
2871		return NULL;
2872	}
2873
2874	if (flags & VM_MAP_PUT_PAGES) {
2875		area->pages = pages;
2876		area->nr_pages = count;
2877	}
2878	return area->addr;
2879}
2880EXPORT_SYMBOL(vmap);
2881
2882#ifdef CONFIG_VMAP_PFN
2883struct vmap_pfn_data {
2884	unsigned long	*pfns;
2885	pgprot_t	prot;
2886	unsigned int	idx;
2887};
2888
2889static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2890{
2891	struct vmap_pfn_data *data = private;
2892
2893	if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2894		return -EINVAL;
2895	*pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
2896	return 0;
2897}
2898
2899/**
2900 * vmap_pfn - map an array of PFNs into virtually contiguous space
2901 * @pfns: array of PFNs
2902 * @count: number of pages to map
2903 * @prot: page protection for the mapping
2904 *
2905 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2906 * the start address of the mapping.
2907 */
2908void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2909{
2910	struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2911	struct vm_struct *area;
2912
2913	area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2914			__builtin_return_address(0));
2915	if (!area)
2916		return NULL;
2917	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2918			count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2919		free_vm_area(area);
2920		return NULL;
2921	}
2922	return area->addr;
2923}
2924EXPORT_SYMBOL_GPL(vmap_pfn);
2925#endif /* CONFIG_VMAP_PFN */
2926
2927static inline unsigned int
2928vm_area_alloc_pages(gfp_t gfp, int nid,
2929		unsigned int order, unsigned int nr_pages, struct page **pages)
2930{
2931	unsigned int nr_allocated = 0;
2932	struct page *page;
2933	int i;
2934
2935	/*
2936	 * For order-0 pages we make use of bulk allocator, if
2937	 * the page array is partly or not at all populated due
2938	 * to fails, fallback to a single page allocator that is
2939	 * more permissive.
2940	 */
2941	if (!order) {
2942		gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2943
2944		while (nr_allocated < nr_pages) {
2945			unsigned int nr, nr_pages_request;
2946
2947			/*
2948			 * A maximum allowed request is hard-coded and is 100
2949			 * pages per call. That is done in order to prevent a
2950			 * long preemption off scenario in the bulk-allocator
2951			 * so the range is [1:100].
2952			 */
2953			nr_pages_request = min(100U, nr_pages - nr_allocated);
2954
2955			/* memory allocation should consider mempolicy, we can't
2956			 * wrongly use nearest node when nid == NUMA_NO_NODE,
2957			 * otherwise memory may be allocated in only one node,
2958			 * but mempolicy wants to alloc memory by interleaving.
2959			 */
2960			if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2961				nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2962							nr_pages_request,
2963							pages + nr_allocated);
2964
2965			else
2966				nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2967							nr_pages_request,
2968							pages + nr_allocated);
2969
2970			nr_allocated += nr;
2971			cond_resched();
2972
2973			/*
2974			 * If zero or pages were obtained partly,
2975			 * fallback to a single page allocator.
2976			 */
2977			if (nr != nr_pages_request)
2978				break;
2979		}
2980	}
2981
2982	/* High-order pages or fallback path if "bulk" fails. */
2983
2984	while (nr_allocated < nr_pages) {
2985		if (fatal_signal_pending(current))
2986			break;
2987
2988		if (nid == NUMA_NO_NODE)
2989			page = alloc_pages(gfp, order);
2990		else
2991			page = alloc_pages_node(nid, gfp, order);
2992		if (unlikely(!page))
2993			break;
2994		/*
2995		 * Higher order allocations must be able to be treated as
2996		 * indepdenent small pages by callers (as they can with
2997		 * small-page vmallocs). Some drivers do their own refcounting
2998		 * on vmalloc_to_page() pages, some use page->mapping,
2999		 * page->lru, etc.
3000		 */
3001		if (order)
3002			split_page(page, order);
3003
3004		/*
3005		 * Careful, we allocate and map page-order pages, but
3006		 * tracking is done per PAGE_SIZE page so as to keep the
3007		 * vm_struct APIs independent of the physical/mapped size.
3008		 */
3009		for (i = 0; i < (1U << order); i++)
3010			pages[nr_allocated + i] = page + i;
3011
3012		cond_resched();
3013		nr_allocated += 1U << order;
3014	}
3015
3016	return nr_allocated;
3017}
3018
3019static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3020				 pgprot_t prot, unsigned int page_shift,
3021				 int node)
3022{
 
 
3023	const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3024	bool nofail = gfp_mask & __GFP_NOFAIL;
3025	unsigned long addr = (unsigned long)area->addr;
3026	unsigned long size = get_vm_area_size(area);
3027	unsigned long array_size;
3028	unsigned int nr_small_pages = size >> PAGE_SHIFT;
3029	unsigned int page_order;
3030	unsigned int flags;
3031	int ret;
3032
3033	array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3034	gfp_mask |= __GFP_NOWARN;
3035	if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3036		gfp_mask |= __GFP_HIGHMEM;
3037
3038	/* Please note that the recursion is strictly bounded. */
3039	if (array_size > PAGE_SIZE) {
3040		area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3041					area->caller);
3042	} else {
3043		area->pages = kmalloc_node(array_size, nested_gfp, node);
3044	}
3045
3046	if (!area->pages) {
3047		warn_alloc(gfp_mask, NULL,
3048			"vmalloc error: size %lu, failed to allocated page array size %lu",
3049			nr_small_pages * PAGE_SIZE, array_size);
3050		free_vm_area(area);
3051		return NULL;
3052	}
3053
3054	set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3055	page_order = vm_area_page_order(area);
3056
3057	area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3058		node, page_order, nr_small_pages, area->pages);
3059
3060	atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3061	if (gfp_mask & __GFP_ACCOUNT) {
3062		int i;
3063
3064		for (i = 0; i < area->nr_pages; i++)
3065			mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3066	}
 
3067
3068	/*
3069	 * If not enough pages were obtained to accomplish an
3070	 * allocation request, free them via __vfree() if any.
3071	 */
3072	if (area->nr_pages != nr_small_pages) {
3073		warn_alloc(gfp_mask, NULL,
3074			"vmalloc error: size %lu, page order %u, failed to allocate pages",
3075			area->nr_pages * PAGE_SIZE, page_order);
3076		goto fail;
3077	}
 
3078
3079	/*
3080	 * page tables allocations ignore external gfp mask, enforce it
3081	 * by the scope API
3082	 */
3083	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3084		flags = memalloc_nofs_save();
3085	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3086		flags = memalloc_noio_save();
3087
3088	do {
3089		ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3090			page_shift);
3091		if (nofail && (ret < 0))
3092			schedule_timeout_uninterruptible(1);
3093	} while (nofail && (ret < 0));
3094
3095	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3096		memalloc_nofs_restore(flags);
3097	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3098		memalloc_noio_restore(flags);
3099
3100	if (ret < 0) {
3101		warn_alloc(gfp_mask, NULL,
3102			"vmalloc error: size %lu, failed to map pages",
3103			area->nr_pages * PAGE_SIZE);
3104		goto fail;
3105	}
3106
3107	return area->addr;
3108
3109fail:
 
 
 
3110	__vfree(area->addr);
3111	return NULL;
3112}
3113
3114/**
3115 * __vmalloc_node_range - allocate virtually contiguous memory
3116 * @size:		  allocation size
3117 * @align:		  desired alignment
3118 * @start:		  vm area range start
3119 * @end:		  vm area range end
3120 * @gfp_mask:		  flags for the page level allocator
3121 * @prot:		  protection mask for the allocated pages
3122 * @vm_flags:		  additional vm area flags (e.g. %VM_NO_GUARD)
3123 * @node:		  node to use for allocation or NUMA_NO_NODE
3124 * @caller:		  caller's return address
3125 *
3126 * Allocate enough pages to cover @size from the page level
3127 * allocator with @gfp_mask flags. Please note that the full set of gfp
3128 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3129 * supported.
3130 * Zone modifiers are not supported. From the reclaim modifiers
3131 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3132 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3133 * __GFP_RETRY_MAYFAIL are not supported).
3134 *
3135 * __GFP_NOWARN can be used to suppress failures messages.
3136 *
3137 * Map them into contiguous kernel virtual space, using a pagetable
3138 * protection of @prot.
3139 *
3140 * Return: the address of the area or %NULL on failure
3141 */
3142void *__vmalloc_node_range(unsigned long size, unsigned long align,
3143			unsigned long start, unsigned long end, gfp_t gfp_mask,
3144			pgprot_t prot, unsigned long vm_flags, int node,
3145			const void *caller)
3146{
3147	struct vm_struct *area;
3148	void *ret;
3149	kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3150	unsigned long real_size = size;
3151	unsigned long real_align = align;
3152	unsigned int shift = PAGE_SHIFT;
3153
3154	if (WARN_ON_ONCE(!size))
3155		return NULL;
3156
3157	if ((size >> PAGE_SHIFT) > totalram_pages()) {
3158		warn_alloc(gfp_mask, NULL,
3159			"vmalloc error: size %lu, exceeds total pages",
3160			real_size);
3161		return NULL;
3162	}
3163
3164	if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3165		unsigned long size_per_node;
3166
3167		/*
3168		 * Try huge pages. Only try for PAGE_KERNEL allocations,
3169		 * others like modules don't yet expect huge pages in
3170		 * their allocations due to apply_to_page_range not
3171		 * supporting them.
3172		 */
3173
3174		size_per_node = size;
3175		if (node == NUMA_NO_NODE)
3176			size_per_node /= num_online_nodes();
3177		if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3178			shift = PMD_SHIFT;
3179		else
3180			shift = arch_vmap_pte_supported_shift(size_per_node);
3181
3182		align = max(real_align, 1UL << shift);
3183		size = ALIGN(real_size, 1UL << shift);
3184	}
3185
3186again:
3187	area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3188				  VM_UNINITIALIZED | vm_flags, start, end, node,
3189				  gfp_mask, caller);
3190	if (!area) {
3191		bool nofail = gfp_mask & __GFP_NOFAIL;
3192		warn_alloc(gfp_mask, NULL,
3193			"vmalloc error: size %lu, vm_struct allocation failed%s",
3194			real_size, (nofail) ? ". Retrying." : "");
3195		if (nofail) {
3196			schedule_timeout_uninterruptible(1);
3197			goto again;
3198		}
3199		goto fail;
3200	}
3201
3202	/*
3203	 * Prepare arguments for __vmalloc_area_node() and
3204	 * kasan_unpoison_vmalloc().
3205	 */
3206	if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3207		if (kasan_hw_tags_enabled()) {
3208			/*
3209			 * Modify protection bits to allow tagging.
3210			 * This must be done before mapping.
3211			 */
3212			prot = arch_vmap_pgprot_tagged(prot);
3213
3214			/*
3215			 * Skip page_alloc poisoning and zeroing for physical
3216			 * pages backing VM_ALLOC mapping. Memory is instead
3217			 * poisoned and zeroed by kasan_unpoison_vmalloc().
3218			 */
3219			gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3220		}
3221
3222		/* Take note that the mapping is PAGE_KERNEL. */
3223		kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3224	}
3225
3226	/* Allocate physical pages and map them into vmalloc space. */
3227	ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3228	if (!ret)
3229		goto fail;
3230
3231	/*
3232	 * Mark the pages as accessible, now that they are mapped.
3233	 * The condition for setting KASAN_VMALLOC_INIT should complement the
3234	 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3235	 * to make sure that memory is initialized under the same conditions.
3236	 * Tag-based KASAN modes only assign tags to normal non-executable
3237	 * allocations, see __kasan_unpoison_vmalloc().
3238	 */
3239	kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3240	if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3241	    (gfp_mask & __GFP_SKIP_ZERO))
3242		kasan_flags |= KASAN_VMALLOC_INIT;
3243	/* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3244	area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3245
3246	/*
3247	 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3248	 * flag. It means that vm_struct is not fully initialized.
3249	 * Now, it is fully initialized, so remove this flag here.
3250	 */
3251	clear_vm_uninitialized_flag(area);
3252
3253	size = PAGE_ALIGN(size);
3254	if (!(vm_flags & VM_DEFER_KMEMLEAK))
3255		kmemleak_vmalloc(area, size, gfp_mask);
3256
3257	return area->addr;
3258
3259fail:
3260	if (shift > PAGE_SHIFT) {
3261		shift = PAGE_SHIFT;
3262		align = real_align;
3263		size = real_size;
3264		goto again;
3265	}
3266
3267	return NULL;
3268}
3269
3270/**
3271 * __vmalloc_node - allocate virtually contiguous memory
3272 * @size:	    allocation size
3273 * @align:	    desired alignment
3274 * @gfp_mask:	    flags for the page level allocator
3275 * @node:	    node to use for allocation or NUMA_NO_NODE
3276 * @caller:	    caller's return address
3277 *
3278 * Allocate enough pages to cover @size from the page level allocator with
3279 * @gfp_mask flags.  Map them into contiguous kernel virtual space.
3280 *
3281 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3282 * and __GFP_NOFAIL are not supported
3283 *
3284 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3285 * with mm people.
3286 *
3287 * Return: pointer to the allocated memory or %NULL on error
3288 */
3289void *__vmalloc_node(unsigned long size, unsigned long align,
3290			    gfp_t gfp_mask, int node, const void *caller)
3291{
3292	return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3293				gfp_mask, PAGE_KERNEL, 0, node, caller);
3294}
3295/*
3296 * This is only for performance analysis of vmalloc and stress purpose.
3297 * It is required by vmalloc test module, therefore do not use it other
3298 * than that.
3299 */
3300#ifdef CONFIG_TEST_VMALLOC_MODULE
3301EXPORT_SYMBOL_GPL(__vmalloc_node);
3302#endif
3303
3304void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3305{
3306	return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3307				__builtin_return_address(0));
3308}
3309EXPORT_SYMBOL(__vmalloc);
3310
3311/**
3312 * vmalloc - allocate virtually contiguous memory
3313 * @size:    allocation size
3314 *
3315 * Allocate enough pages to cover @size from the page level
3316 * allocator and map them into contiguous kernel virtual space.
3317 *
3318 * For tight control over page level allocator and protection flags
3319 * use __vmalloc() instead.
3320 *
3321 * Return: pointer to the allocated memory or %NULL on error
3322 */
3323void *vmalloc(unsigned long size)
3324{
3325	return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3326				__builtin_return_address(0));
3327}
3328EXPORT_SYMBOL(vmalloc);
3329
3330/**
3331 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3332 * @size:      allocation size
3333 * @gfp_mask:  flags for the page level allocator
3334 *
3335 * Allocate enough pages to cover @size from the page level
3336 * allocator and map them into contiguous kernel virtual space.
3337 * If @size is greater than or equal to PMD_SIZE, allow using
3338 * huge pages for the memory
3339 *
3340 * Return: pointer to the allocated memory or %NULL on error
3341 */
3342void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3343{
3344	return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3345				    gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3346				    NUMA_NO_NODE, __builtin_return_address(0));
3347}
3348EXPORT_SYMBOL_GPL(vmalloc_huge);
3349
3350/**
3351 * vzalloc - allocate virtually contiguous memory with zero fill
3352 * @size:    allocation size
3353 *
3354 * Allocate enough pages to cover @size from the page level
3355 * allocator and map them into contiguous kernel virtual space.
3356 * The memory allocated is set to zero.
3357 *
3358 * For tight control over page level allocator and protection flags
3359 * use __vmalloc() instead.
3360 *
3361 * Return: pointer to the allocated memory or %NULL on error
3362 */
3363void *vzalloc(unsigned long size)
3364{
3365	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3366				__builtin_return_address(0));
3367}
3368EXPORT_SYMBOL(vzalloc);
3369
3370/**
3371 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3372 * @size: allocation size
3373 *
3374 * The resulting memory area is zeroed so it can be mapped to userspace
3375 * without leaking data.
3376 *
3377 * Return: pointer to the allocated memory or %NULL on error
3378 */
3379void *vmalloc_user(unsigned long size)
3380{
3381	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
3382				    GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3383				    VM_USERMAP, NUMA_NO_NODE,
3384				    __builtin_return_address(0));
3385}
3386EXPORT_SYMBOL(vmalloc_user);
3387
3388/**
3389 * vmalloc_node - allocate memory on a specific node
3390 * @size:	  allocation size
3391 * @node:	  numa node
3392 *
3393 * Allocate enough pages to cover @size from the page level
3394 * allocator and map them into contiguous kernel virtual space.
3395 *
3396 * For tight control over page level allocator and protection flags
3397 * use __vmalloc() instead.
3398 *
3399 * Return: pointer to the allocated memory or %NULL on error
3400 */
3401void *vmalloc_node(unsigned long size, int node)
3402{
3403	return __vmalloc_node(size, 1, GFP_KERNEL, node,
3404			__builtin_return_address(0));
3405}
3406EXPORT_SYMBOL(vmalloc_node);
3407
3408/**
3409 * vzalloc_node - allocate memory on a specific node with zero fill
3410 * @size:	allocation size
3411 * @node:	numa node
3412 *
3413 * Allocate enough pages to cover @size from the page level
3414 * allocator and map them into contiguous kernel virtual space.
3415 * The memory allocated is set to zero.
3416 *
3417 * Return: pointer to the allocated memory or %NULL on error
3418 */
3419void *vzalloc_node(unsigned long size, int node)
3420{
3421	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3422				__builtin_return_address(0));
3423}
3424EXPORT_SYMBOL(vzalloc_node);
3425
3426#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3427#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3428#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3429#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3430#else
3431/*
3432 * 64b systems should always have either DMA or DMA32 zones. For others
3433 * GFP_DMA32 should do the right thing and use the normal zone.
3434 */
3435#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3436#endif
3437
3438/**
3439 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3440 * @size:	allocation size
3441 *
3442 * Allocate enough 32bit PA addressable pages to cover @size from the
3443 * page level allocator and map them into contiguous kernel virtual space.
3444 *
3445 * Return: pointer to the allocated memory or %NULL on error
3446 */
3447void *vmalloc_32(unsigned long size)
3448{
3449	return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3450			__builtin_return_address(0));
3451}
3452EXPORT_SYMBOL(vmalloc_32);
3453
3454/**
3455 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3456 * @size:	     allocation size
3457 *
3458 * The resulting memory area is 32bit addressable and zeroed so it can be
3459 * mapped to userspace without leaking data.
3460 *
3461 * Return: pointer to the allocated memory or %NULL on error
3462 */
3463void *vmalloc_32_user(unsigned long size)
3464{
3465	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
3466				    GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3467				    VM_USERMAP, NUMA_NO_NODE,
3468				    __builtin_return_address(0));
3469}
3470EXPORT_SYMBOL(vmalloc_32_user);
3471
3472/*
3473 * small helper routine , copy contents to buf from addr.
3474 * If the page is not present, fill zero.
3475 */
3476
3477static int aligned_vread(char *buf, char *addr, unsigned long count)
3478{
3479	struct page *p;
3480	int copied = 0;
3481
3482	while (count) {
3483		unsigned long offset, length;
3484
3485		offset = offset_in_page(addr);
3486		length = PAGE_SIZE - offset;
3487		if (length > count)
3488			length = count;
3489		p = vmalloc_to_page(addr);
3490		/*
3491		 * To do safe access to this _mapped_ area, we need
3492		 * lock. But adding lock here means that we need to add
3493		 * overhead of vmalloc()/vfree() calls for this _debug_
3494		 * interface, rarely used. Instead of that, we'll use
3495		 * kmap() and get small overhead in this access function.
3496		 */
3497		if (p) {
3498			/* We can expect USER0 is not used -- see vread() */
 
 
 
3499			void *map = kmap_atomic(p);
3500			memcpy(buf, map + offset, length);
3501			kunmap_atomic(map);
3502		} else
3503			memset(buf, 0, length);
3504
3505		addr += length;
3506		buf += length;
3507		copied += length;
3508		count -= length;
3509	}
3510	return copied;
3511}
3512
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3513/**
3514 * vread() - read vmalloc area in a safe way.
3515 * @buf:     buffer for reading data
3516 * @addr:    vm address.
3517 * @count:   number of bytes to be read.
3518 *
3519 * This function checks that addr is a valid vmalloc'ed area, and
3520 * copy data from that area to a given buffer. If the given memory range
3521 * of [addr...addr+count) includes some valid address, data is copied to
3522 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3523 * IOREMAP area is treated as memory hole and no copy is done.
3524 *
3525 * If [addr...addr+count) doesn't includes any intersects with alive
3526 * vm_struct area, returns 0. @buf should be kernel's buffer.
3527 *
3528 * Note: In usual ops, vread() is never necessary because the caller
3529 * should know vmalloc() area is valid and can use memcpy().
3530 * This is for routines which have to access vmalloc area without
3531 * any information, as /proc/kcore.
3532 *
3533 * Return: number of bytes for which addr and buf should be increased
3534 * (same number as @count) or %0 if [addr...addr+count) doesn't
3535 * include any intersection with valid vmalloc area
3536 */
3537long vread(char *buf, char *addr, unsigned long count)
3538{
3539	struct vmap_area *va;
3540	struct vm_struct *vm;
3541	char *vaddr, *buf_start = buf;
3542	unsigned long buflen = count;
3543	unsigned long n;
3544
3545	addr = kasan_reset_tag(addr);
3546
3547	/* Don't allow overflow */
3548	if ((unsigned long) addr + count < count)
3549		count = -(unsigned long) addr;
3550
3551	spin_lock(&vmap_area_lock);
3552	va = find_vmap_area_exceed_addr((unsigned long)addr);
3553	if (!va)
3554		goto finished;
3555
3556	/* no intersects with alive vmap_area */
3557	if ((unsigned long)addr + count <= va->va_start)
3558		goto finished;
3559
3560	list_for_each_entry_from(va, &vmap_area_list, list) {
3561		if (!count)
3562			break;
3563
3564		if (!va->vm)
3565			continue;
3566
3567		vm = va->vm;
3568		vaddr = (char *) vm->addr;
3569		if (addr >= vaddr + get_vm_area_size(vm))
3570			continue;
3571		while (addr < vaddr) {
3572			if (count == 0)
3573				goto finished;
3574			*buf = '\0';
3575			buf++;
3576			addr++;
3577			count--;
3578		}
3579		n = vaddr + get_vm_area_size(vm) - addr;
3580		if (n > count)
3581			n = count;
3582		if (!(vm->flags & VM_IOREMAP))
3583			aligned_vread(buf, addr, n);
3584		else /* IOREMAP area is treated as memory hole */
3585			memset(buf, 0, n);
3586		buf += n;
3587		addr += n;
3588		count -= n;
3589	}
3590finished:
3591	spin_unlock(&vmap_area_lock);
3592
3593	if (buf == buf_start)
3594		return 0;
3595	/* zero-fill memory holes */
3596	if (buf != buf_start + buflen)
3597		memset(buf, 0, buflen - (buf - buf_start));
3598
3599	return buflen;
3600}
3601
3602/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3603 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3604 * @vma:		vma to cover
3605 * @uaddr:		target user address to start at
3606 * @kaddr:		virtual address of vmalloc kernel memory
3607 * @pgoff:		offset from @kaddr to start at
3608 * @size:		size of map area
3609 *
3610 * Returns:	0 for success, -Exxx on failure
3611 *
3612 * This function checks that @kaddr is a valid vmalloc'ed area,
3613 * and that it is big enough to cover the range starting at
3614 * @uaddr in @vma. Will return failure if that criteria isn't
3615 * met.
3616 *
3617 * Similar to remap_pfn_range() (see mm/memory.c)
3618 */
3619int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3620				void *kaddr, unsigned long pgoff,
3621				unsigned long size)
3622{
3623	struct vm_struct *area;
3624	unsigned long off;
3625	unsigned long end_index;
3626
3627	if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3628		return -EINVAL;
3629
3630	size = PAGE_ALIGN(size);
3631
3632	if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3633		return -EINVAL;
3634
3635	area = find_vm_area(kaddr);
3636	if (!area)
3637		return -EINVAL;
3638
3639	if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3640		return -EINVAL;
3641
3642	if (check_add_overflow(size, off, &end_index) ||
3643	    end_index > get_vm_area_size(area))
3644		return -EINVAL;
3645	kaddr += off;
3646
3647	do {
3648		struct page *page = vmalloc_to_page(kaddr);
3649		int ret;
3650
3651		ret = vm_insert_page(vma, uaddr, page);
3652		if (ret)
3653			return ret;
3654
3655		uaddr += PAGE_SIZE;
3656		kaddr += PAGE_SIZE;
3657		size -= PAGE_SIZE;
3658	} while (size > 0);
3659
3660	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3661
3662	return 0;
3663}
 
3664
3665/**
3666 * remap_vmalloc_range - map vmalloc pages to userspace
3667 * @vma:		vma to cover (map full range of vma)
3668 * @addr:		vmalloc memory
3669 * @pgoff:		number of pages into addr before first page to map
3670 *
3671 * Returns:	0 for success, -Exxx on failure
3672 *
3673 * This function checks that addr is a valid vmalloc'ed area, and
3674 * that it is big enough to cover the vma. Will return failure if
3675 * that criteria isn't met.
3676 *
3677 * Similar to remap_pfn_range() (see mm/memory.c)
3678 */
3679int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3680						unsigned long pgoff)
3681{
3682	return remap_vmalloc_range_partial(vma, vma->vm_start,
3683					   addr, pgoff,
3684					   vma->vm_end - vma->vm_start);
3685}
3686EXPORT_SYMBOL(remap_vmalloc_range);
3687
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3688void free_vm_area(struct vm_struct *area)
3689{
3690	struct vm_struct *ret;
3691	ret = remove_vm_area(area->addr);
3692	BUG_ON(ret != area);
3693	kfree(area);
3694}
3695EXPORT_SYMBOL_GPL(free_vm_area);
3696
3697#ifdef CONFIG_SMP
3698static struct vmap_area *node_to_va(struct rb_node *n)
3699{
3700	return rb_entry_safe(n, struct vmap_area, rb_node);
3701}
3702
3703/**
3704 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3705 * @addr: target address
3706 *
3707 * Returns: vmap_area if it is found. If there is no such area
3708 *   the first highest(reverse order) vmap_area is returned
3709 *   i.e. va->va_start < addr && va->va_end < addr or NULL
3710 *   if there are no any areas before @addr.
3711 */
3712static struct vmap_area *
3713pvm_find_va_enclose_addr(unsigned long addr)
3714{
3715	struct vmap_area *va, *tmp;
3716	struct rb_node *n;
3717
3718	n = free_vmap_area_root.rb_node;
3719	va = NULL;
3720
3721	while (n) {
3722		tmp = rb_entry(n, struct vmap_area, rb_node);
3723		if (tmp->va_start <= addr) {
3724			va = tmp;
3725			if (tmp->va_end >= addr)
3726				break;
3727
3728			n = n->rb_right;
3729		} else {
3730			n = n->rb_left;
3731		}
3732	}
3733
3734	return va;
3735}
3736
3737/**
3738 * pvm_determine_end_from_reverse - find the highest aligned address
3739 * of free block below VMALLOC_END
3740 * @va:
3741 *   in - the VA we start the search(reverse order);
3742 *   out - the VA with the highest aligned end address.
3743 * @align: alignment for required highest address
3744 *
3745 * Returns: determined end address within vmap_area
3746 */
3747static unsigned long
3748pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3749{
3750	unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3751	unsigned long addr;
3752
3753	if (likely(*va)) {
3754		list_for_each_entry_from_reverse((*va),
3755				&free_vmap_area_list, list) {
3756			addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3757			if ((*va)->va_start < addr)
3758				return addr;
3759		}
3760	}
3761
3762	return 0;
3763}
3764
3765/**
3766 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3767 * @offsets: array containing offset of each area
3768 * @sizes: array containing size of each area
3769 * @nr_vms: the number of areas to allocate
3770 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3771 *
3772 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3773 *	    vm_structs on success, %NULL on failure
3774 *
3775 * Percpu allocator wants to use congruent vm areas so that it can
3776 * maintain the offsets among percpu areas.  This function allocates
3777 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
3778 * be scattered pretty far, distance between two areas easily going up
3779 * to gigabytes.  To avoid interacting with regular vmallocs, these
3780 * areas are allocated from top.
3781 *
3782 * Despite its complicated look, this allocator is rather simple. It
3783 * does everything top-down and scans free blocks from the end looking
3784 * for matching base. While scanning, if any of the areas do not fit the
3785 * base address is pulled down to fit the area. Scanning is repeated till
3786 * all the areas fit and then all necessary data structures are inserted
3787 * and the result is returned.
3788 */
3789struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3790				     const size_t *sizes, int nr_vms,
3791				     size_t align)
3792{
3793	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3794	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3795	struct vmap_area **vas, *va;
3796	struct vm_struct **vms;
3797	int area, area2, last_area, term_area;
3798	unsigned long base, start, size, end, last_end, orig_start, orig_end;
3799	bool purged = false;
 
3800
3801	/* verify parameters and allocate data structures */
3802	BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3803	for (last_area = 0, area = 0; area < nr_vms; area++) {
3804		start = offsets[area];
3805		end = start + sizes[area];
3806
3807		/* is everything aligned properly? */
3808		BUG_ON(!IS_ALIGNED(offsets[area], align));
3809		BUG_ON(!IS_ALIGNED(sizes[area], align));
3810
3811		/* detect the area with the highest address */
3812		if (start > offsets[last_area])
3813			last_area = area;
3814
3815		for (area2 = area + 1; area2 < nr_vms; area2++) {
3816			unsigned long start2 = offsets[area2];
3817			unsigned long end2 = start2 + sizes[area2];
3818
3819			BUG_ON(start2 < end && start < end2);
3820		}
3821	}
3822	last_end = offsets[last_area] + sizes[last_area];
3823
3824	if (vmalloc_end - vmalloc_start < last_end) {
3825		WARN_ON(true);
3826		return NULL;
3827	}
3828
3829	vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3830	vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3831	if (!vas || !vms)
3832		goto err_free2;
3833
3834	for (area = 0; area < nr_vms; area++) {
3835		vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3836		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3837		if (!vas[area] || !vms[area])
3838			goto err_free;
3839	}
3840retry:
3841	spin_lock(&free_vmap_area_lock);
3842
3843	/* start scanning - we scan from the top, begin with the last area */
3844	area = term_area = last_area;
3845	start = offsets[area];
3846	end = start + sizes[area];
3847
3848	va = pvm_find_va_enclose_addr(vmalloc_end);
3849	base = pvm_determine_end_from_reverse(&va, align) - end;
3850
3851	while (true) {
3852		/*
3853		 * base might have underflowed, add last_end before
3854		 * comparing.
3855		 */
3856		if (base + last_end < vmalloc_start + last_end)
3857			goto overflow;
3858
3859		/*
3860		 * Fitting base has not been found.
3861		 */
3862		if (va == NULL)
3863			goto overflow;
3864
3865		/*
3866		 * If required width exceeds current VA block, move
3867		 * base downwards and then recheck.
3868		 */
3869		if (base + end > va->va_end) {
3870			base = pvm_determine_end_from_reverse(&va, align) - end;
3871			term_area = area;
3872			continue;
3873		}
3874
3875		/*
3876		 * If this VA does not fit, move base downwards and recheck.
3877		 */
3878		if (base + start < va->va_start) {
3879			va = node_to_va(rb_prev(&va->rb_node));
3880			base = pvm_determine_end_from_reverse(&va, align) - end;
3881			term_area = area;
3882			continue;
3883		}
3884
3885		/*
3886		 * This area fits, move on to the previous one.  If
3887		 * the previous one is the terminal one, we're done.
3888		 */
3889		area = (area + nr_vms - 1) % nr_vms;
3890		if (area == term_area)
3891			break;
3892
3893		start = offsets[area];
3894		end = start + sizes[area];
3895		va = pvm_find_va_enclose_addr(base + end);
3896	}
3897
3898	/* we've found a fitting base, insert all va's */
3899	for (area = 0; area < nr_vms; area++) {
3900		int ret;
3901
3902		start = base + offsets[area];
3903		size = sizes[area];
3904
3905		va = pvm_find_va_enclose_addr(start);
3906		if (WARN_ON_ONCE(va == NULL))
3907			/* It is a BUG(), but trigger recovery instead. */
3908			goto recovery;
3909
3910		ret = adjust_va_to_fit_type(&free_vmap_area_root,
3911					    &free_vmap_area_list,
3912					    va, start, size);
3913		if (WARN_ON_ONCE(unlikely(ret)))
3914			/* It is a BUG(), but trigger recovery instead. */
3915			goto recovery;
3916
 
 
 
 
3917		/* Allocated area. */
3918		va = vas[area];
3919		va->va_start = start;
3920		va->va_end = start + size;
3921	}
3922
3923	spin_unlock(&free_vmap_area_lock);
3924
3925	/* populate the kasan shadow space */
3926	for (area = 0; area < nr_vms; area++) {
3927		if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3928			goto err_free_shadow;
 
 
 
3929	}
3930
3931	/* insert all vm's */
3932	spin_lock(&vmap_area_lock);
3933	for (area = 0; area < nr_vms; area++) {
3934		insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3935
3936		setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3937				 pcpu_get_vm_areas);
3938	}
3939	spin_unlock(&vmap_area_lock);
3940
3941	/*
3942	 * Mark allocated areas as accessible. Do it now as a best-effort
3943	 * approach, as they can be mapped outside of vmalloc code.
3944	 * With hardware tag-based KASAN, marking is skipped for
3945	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3946	 */
3947	for (area = 0; area < nr_vms; area++)
3948		vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3949				vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3950
3951	kfree(vas);
3952	return vms;
3953
3954recovery:
3955	/*
3956	 * Remove previously allocated areas. There is no
3957	 * need in removing these areas from the busy tree,
3958	 * because they are inserted only on the final step
3959	 * and when pcpu_get_vm_areas() is success.
3960	 */
3961	while (area--) {
3962		orig_start = vas[area]->va_start;
3963		orig_end = vas[area]->va_end;
3964		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3965				&free_vmap_area_list);
3966		if (va)
3967			kasan_release_vmalloc(orig_start, orig_end,
3968				va->va_start, va->va_end);
3969		vas[area] = NULL;
3970	}
3971
3972overflow:
3973	spin_unlock(&free_vmap_area_lock);
3974	if (!purged) {
3975		purge_vmap_area_lazy();
3976		purged = true;
3977
3978		/* Before "retry", check if we recover. */
3979		for (area = 0; area < nr_vms; area++) {
3980			if (vas[area])
3981				continue;
3982
3983			vas[area] = kmem_cache_zalloc(
3984				vmap_area_cachep, GFP_KERNEL);
3985			if (!vas[area])
3986				goto err_free;
3987		}
3988
3989		goto retry;
3990	}
3991
3992err_free:
3993	for (area = 0; area < nr_vms; area++) {
3994		if (vas[area])
3995			kmem_cache_free(vmap_area_cachep, vas[area]);
3996
3997		kfree(vms[area]);
3998	}
3999err_free2:
4000	kfree(vas);
4001	kfree(vms);
4002	return NULL;
4003
4004err_free_shadow:
4005	spin_lock(&free_vmap_area_lock);
4006	/*
4007	 * We release all the vmalloc shadows, even the ones for regions that
4008	 * hadn't been successfully added. This relies on kasan_release_vmalloc
4009	 * being able to tolerate this case.
4010	 */
4011	for (area = 0; area < nr_vms; area++) {
4012		orig_start = vas[area]->va_start;
4013		orig_end = vas[area]->va_end;
4014		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4015				&free_vmap_area_list);
4016		if (va)
4017			kasan_release_vmalloc(orig_start, orig_end,
4018				va->va_start, va->va_end);
4019		vas[area] = NULL;
4020		kfree(vms[area]);
4021	}
4022	spin_unlock(&free_vmap_area_lock);
4023	kfree(vas);
4024	kfree(vms);
4025	return NULL;
4026}
4027
4028/**
4029 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4030 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4031 * @nr_vms: the number of allocated areas
4032 *
4033 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4034 */
4035void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4036{
4037	int i;
4038
4039	for (i = 0; i < nr_vms; i++)
4040		free_vm_area(vms[i]);
4041	kfree(vms);
4042}
4043#endif	/* CONFIG_SMP */
4044
4045#ifdef CONFIG_PRINTK
4046bool vmalloc_dump_obj(void *object)
4047{
4048	struct vm_struct *vm;
4049	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4050
4051	vm = find_vm_area(objp);
4052	if (!vm)
4053		return false;
4054	pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4055		vm->nr_pages, (unsigned long)vm->addr, vm->caller);
4056	return true;
4057}
4058#endif
4059
4060#ifdef CONFIG_PROC_FS
4061static void *s_start(struct seq_file *m, loff_t *pos)
4062	__acquires(&vmap_purge_lock)
4063	__acquires(&vmap_area_lock)
4064{
4065	mutex_lock(&vmap_purge_lock);
4066	spin_lock(&vmap_area_lock);
4067
4068	return seq_list_start(&vmap_area_list, *pos);
4069}
4070
4071static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4072{
4073	return seq_list_next(p, &vmap_area_list, pos);
4074}
4075
4076static void s_stop(struct seq_file *m, void *p)
4077	__releases(&vmap_area_lock)
4078	__releases(&vmap_purge_lock)
 
4079{
4080	spin_unlock(&vmap_area_lock);
4081	mutex_unlock(&vmap_purge_lock);
 
4082}
4083
4084static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4085{
4086	if (IS_ENABLED(CONFIG_NUMA)) {
4087		unsigned int nr, *counters = m->private;
4088		unsigned int step = 1U << vm_area_page_order(v);
4089
4090		if (!counters)
4091			return;
4092
4093		if (v->flags & VM_UNINITIALIZED)
4094			return;
4095		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4096		smp_rmb();
4097
4098		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4099
4100		for (nr = 0; nr < v->nr_pages; nr += step)
4101			counters[page_to_nid(v->pages[nr])] += step;
 
4102		for_each_node_state(nr, N_HIGH_MEMORY)
4103			if (counters[nr])
4104				seq_printf(m, " N%u=%u", nr, counters[nr]);
4105	}
4106}
4107
4108static void show_purge_info(struct seq_file *m)
4109{
 
4110	struct vmap_area *va;
4111
4112	spin_lock(&purge_vmap_area_lock);
4113	list_for_each_entry(va, &purge_vmap_area_list, list) {
 
 
 
4114		seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4115			(void *)va->va_start, (void *)va->va_end,
4116			va->va_end - va->va_start);
4117	}
4118	spin_unlock(&purge_vmap_area_lock);
4119}
4120
4121static int s_show(struct seq_file *m, void *p)
4122{
4123	struct vmap_area *va;
4124	struct vm_struct *v;
4125
4126	va = list_entry(p, struct vmap_area, list);
4127
4128	/*
4129	 * s_show can encounter race with remove_vm_area, !vm on behalf
4130	 * of vmap area is being tear down or vm_map_ram allocation.
4131	 */
4132	if (!va->vm) {
4133		seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4134			(void *)va->va_start, (void *)va->va_end,
4135			va->va_end - va->va_start);
4136
4137		goto final;
4138	}
4139
4140	v = va->vm;
4141
4142	seq_printf(m, "0x%pK-0x%pK %7ld",
4143		v->addr, v->addr + v->size, v->size);
4144
4145	if (v->caller)
4146		seq_printf(m, " %pS", v->caller);
4147
4148	if (v->nr_pages)
4149		seq_printf(m, " pages=%d", v->nr_pages);
4150
4151	if (v->phys_addr)
4152		seq_printf(m, " phys=%pa", &v->phys_addr);
4153
4154	if (v->flags & VM_IOREMAP)
4155		seq_puts(m, " ioremap");
4156
4157	if (v->flags & VM_ALLOC)
4158		seq_puts(m, " vmalloc");
4159
4160	if (v->flags & VM_MAP)
4161		seq_puts(m, " vmap");
4162
4163	if (v->flags & VM_USERMAP)
4164		seq_puts(m, " user");
4165
4166	if (v->flags & VM_DMA_COHERENT)
4167		seq_puts(m, " dma-coherent");
4168
4169	if (is_vmalloc_addr(v->pages))
4170		seq_puts(m, " vpages");
4171
4172	show_numa_info(m, v);
4173	seq_putc(m, '\n');
4174
4175	/*
4176	 * As a final step, dump "unpurged" areas.
 
 
 
4177	 */
4178final:
4179	if (list_is_last(&va->list, &vmap_area_list))
4180		show_purge_info(m);
4181
4182	return 0;
4183}
4184
4185static const struct seq_operations vmalloc_op = {
4186	.start = s_start,
4187	.next = s_next,
4188	.stop = s_stop,
4189	.show = s_show,
4190};
4191
4192static int __init proc_vmalloc_init(void)
4193{
4194	if (IS_ENABLED(CONFIG_NUMA))
4195		proc_create_seq_private("vmallocinfo", 0400, NULL,
4196				&vmalloc_op,
4197				nr_node_ids * sizeof(unsigned int), NULL);
4198	else
4199		proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4200	return 0;
4201}
4202module_init(proc_vmalloc_init);
4203
4204#endif
v5.9
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  linux/mm/vmalloc.c
   4 *
   5 *  Copyright (C) 1993  Linus Torvalds
   6 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   7 *  SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
   8 *  Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
   9 *  Numa awareness, Christoph Lameter, SGI, June 2005
  10 *  Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
  11 */
  12
  13#include <linux/vmalloc.h>
  14#include <linux/mm.h>
  15#include <linux/module.h>
  16#include <linux/highmem.h>
  17#include <linux/sched/signal.h>
  18#include <linux/slab.h>
  19#include <linux/spinlock.h>
  20#include <linux/interrupt.h>
  21#include <linux/proc_fs.h>
  22#include <linux/seq_file.h>
  23#include <linux/set_memory.h>
  24#include <linux/debugobjects.h>
  25#include <linux/kallsyms.h>
  26#include <linux/list.h>
  27#include <linux/notifier.h>
  28#include <linux/rbtree.h>
  29#include <linux/xarray.h>
 
  30#include <linux/rcupdate.h>
  31#include <linux/pfn.h>
  32#include <linux/kmemleak.h>
  33#include <linux/atomic.h>
  34#include <linux/compiler.h>
 
  35#include <linux/llist.h>
  36#include <linux/bitops.h>
  37#include <linux/rbtree_augmented.h>
  38#include <linux/overflow.h>
  39
  40#include <linux/uaccess.h>
 
 
  41#include <asm/tlbflush.h>
  42#include <asm/shmparam.h>
  43
 
 
 
  44#include "internal.h"
  45#include "pgalloc-track.h"
  46
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  47bool is_vmalloc_addr(const void *x)
  48{
  49	unsigned long addr = (unsigned long)x;
  50
  51	return addr >= VMALLOC_START && addr < VMALLOC_END;
  52}
  53EXPORT_SYMBOL(is_vmalloc_addr);
  54
  55struct vfree_deferred {
  56	struct llist_head list;
  57	struct work_struct wq;
  58};
  59static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
  60
  61static void __vunmap(const void *, int);
  62
  63static void free_work(struct work_struct *w)
  64{
  65	struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
  66	struct llist_node *t, *llnode;
  67
  68	llist_for_each_safe(llnode, t, llist_del_all(&p->list))
  69		__vunmap((void *)llnode, 1);
  70}
  71
  72/*** Page table manipulation functions ***/
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  73
  74static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
  75			     pgtbl_mod_mask *mask)
  76{
  77	pte_t *pte;
  78
  79	pte = pte_offset_kernel(pmd, addr);
  80	do {
  81		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
  82		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
  83	} while (pte++, addr += PAGE_SIZE, addr != end);
  84	*mask |= PGTBL_PTE_MODIFIED;
  85}
  86
  87static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
  88			     pgtbl_mod_mask *mask)
  89{
  90	pmd_t *pmd;
  91	unsigned long next;
  92	int cleared;
  93
  94	pmd = pmd_offset(pud, addr);
  95	do {
  96		next = pmd_addr_end(addr, end);
  97
  98		cleared = pmd_clear_huge(pmd);
  99		if (cleared || pmd_bad(*pmd))
 100			*mask |= PGTBL_PMD_MODIFIED;
 101
 102		if (cleared)
 103			continue;
 104		if (pmd_none_or_clear_bad(pmd))
 105			continue;
 106		vunmap_pte_range(pmd, addr, next, mask);
 107
 108		cond_resched();
 109	} while (pmd++, addr = next, addr != end);
 110}
 111
 112static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 113			     pgtbl_mod_mask *mask)
 114{
 115	pud_t *pud;
 116	unsigned long next;
 117	int cleared;
 118
 119	pud = pud_offset(p4d, addr);
 120	do {
 121		next = pud_addr_end(addr, end);
 122
 123		cleared = pud_clear_huge(pud);
 124		if (cleared || pud_bad(*pud))
 125			*mask |= PGTBL_PUD_MODIFIED;
 126
 127		if (cleared)
 128			continue;
 129		if (pud_none_or_clear_bad(pud))
 130			continue;
 131		vunmap_pmd_range(pud, addr, next, mask);
 132	} while (pud++, addr = next, addr != end);
 133}
 134
 135static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 136			     pgtbl_mod_mask *mask)
 137{
 138	p4d_t *p4d;
 139	unsigned long next;
 140	int cleared;
 141
 142	p4d = p4d_offset(pgd, addr);
 143	do {
 144		next = p4d_addr_end(addr, end);
 145
 146		cleared = p4d_clear_huge(p4d);
 147		if (cleared || p4d_bad(*p4d))
 148			*mask |= PGTBL_P4D_MODIFIED;
 149
 150		if (cleared)
 151			continue;
 152		if (p4d_none_or_clear_bad(p4d))
 153			continue;
 154		vunmap_pud_range(p4d, addr, next, mask);
 155	} while (p4d++, addr = next, addr != end);
 156}
 157
 158/**
 159 * unmap_kernel_range_noflush - unmap kernel VM area
 160 * @start: start of the VM area to unmap
 161 * @size: size of the VM area to unmap
 162 *
 163 * Unmap PFN_UP(@size) pages at @addr.  The VM area @addr and @size specify
 164 * should have been allocated using get_vm_area() and its friends.
 
 
 
 165 *
 166 * NOTE:
 167 * This function does NOT do any cache flushing.  The caller is responsible
 168 * for calling flush_cache_vunmap() on to-be-mapped areas before calling this
 169 * function and flush_tlb_kernel_range() after.
 170 */
 171void unmap_kernel_range_noflush(unsigned long start, unsigned long size)
 172{
 173	unsigned long end = start + size;
 174	unsigned long next;
 175	pgd_t *pgd;
 176	unsigned long addr = start;
 177	pgtbl_mod_mask mask = 0;
 178
 179	BUG_ON(addr >= end);
 180	pgd = pgd_offset_k(addr);
 181	do {
 182		next = pgd_addr_end(addr, end);
 183		if (pgd_bad(*pgd))
 184			mask |= PGTBL_PGD_MODIFIED;
 185		if (pgd_none_or_clear_bad(pgd))
 186			continue;
 187		vunmap_p4d_range(pgd, addr, next, &mask);
 188	} while (pgd++, addr = next, addr != end);
 189
 190	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 191		arch_sync_kernel_mappings(start, end);
 192}
 193
 194static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 195		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 196		pgtbl_mod_mask *mask)
 197{
 198	pte_t *pte;
 199
 200	/*
 201	 * nr is a running index into the array which helps higher level
 202	 * callers keep track of where we're up to.
 203	 */
 204
 205	pte = pte_alloc_kernel_track(pmd, addr, mask);
 206	if (!pte)
 207		return -ENOMEM;
 208	do {
 209		struct page *page = pages[*nr];
 210
 211		if (WARN_ON(!pte_none(*pte)))
 212			return -EBUSY;
 213		if (WARN_ON(!page))
 214			return -ENOMEM;
 
 
 
 215		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 216		(*nr)++;
 217	} while (pte++, addr += PAGE_SIZE, addr != end);
 218	*mask |= PGTBL_PTE_MODIFIED;
 219	return 0;
 220}
 221
 222static int vmap_pmd_range(pud_t *pud, unsigned long addr,
 223		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 224		pgtbl_mod_mask *mask)
 225{
 226	pmd_t *pmd;
 227	unsigned long next;
 228
 229	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 230	if (!pmd)
 231		return -ENOMEM;
 232	do {
 233		next = pmd_addr_end(addr, end);
 234		if (vmap_pte_range(pmd, addr, next, prot, pages, nr, mask))
 235			return -ENOMEM;
 236	} while (pmd++, addr = next, addr != end);
 237	return 0;
 238}
 239
 240static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
 241		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 242		pgtbl_mod_mask *mask)
 243{
 244	pud_t *pud;
 245	unsigned long next;
 246
 247	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 248	if (!pud)
 249		return -ENOMEM;
 250	do {
 251		next = pud_addr_end(addr, end);
 252		if (vmap_pmd_range(pud, addr, next, prot, pages, nr, mask))
 253			return -ENOMEM;
 254	} while (pud++, addr = next, addr != end);
 255	return 0;
 256}
 257
 258static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
 259		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 260		pgtbl_mod_mask *mask)
 261{
 262	p4d_t *p4d;
 263	unsigned long next;
 264
 265	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 266	if (!p4d)
 267		return -ENOMEM;
 268	do {
 269		next = p4d_addr_end(addr, end);
 270		if (vmap_pud_range(p4d, addr, next, prot, pages, nr, mask))
 271			return -ENOMEM;
 272	} while (p4d++, addr = next, addr != end);
 273	return 0;
 274}
 275
 276/**
 277 * map_kernel_range_noflush - map kernel VM area with the specified pages
 278 * @addr: start of the VM area to map
 279 * @size: size of the VM area to map
 280 * @prot: page protection flags to use
 281 * @pages: pages to map
 282 *
 283 * Map PFN_UP(@size) pages at @addr.  The VM area @addr and @size specify should
 284 * have been allocated using get_vm_area() and its friends.
 285 *
 286 * NOTE:
 287 * This function does NOT do any cache flushing.  The caller is responsible for
 288 * calling flush_cache_vmap() on to-be-mapped areas before calling this
 289 * function.
 290 *
 291 * RETURNS:
 292 * 0 on success, -errno on failure.
 293 */
 294int map_kernel_range_noflush(unsigned long addr, unsigned long size,
 295			     pgprot_t prot, struct page **pages)
 296{
 297	unsigned long start = addr;
 298	unsigned long end = addr + size;
 299	unsigned long next;
 300	pgd_t *pgd;
 301	int err = 0;
 302	int nr = 0;
 303	pgtbl_mod_mask mask = 0;
 304
 305	BUG_ON(addr >= end);
 306	pgd = pgd_offset_k(addr);
 307	do {
 308		next = pgd_addr_end(addr, end);
 309		if (pgd_bad(*pgd))
 310			mask |= PGTBL_PGD_MODIFIED;
 311		err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
 312		if (err)
 313			return err;
 314	} while (pgd++, addr = next, addr != end);
 315
 316	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 317		arch_sync_kernel_mappings(start, end);
 318
 319	return 0;
 320}
 321
 322int map_kernel_range(unsigned long start, unsigned long size, pgprot_t prot,
 323		struct page **pages)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 324{
 325	int ret;
 326
 327	ret = map_kernel_range_noflush(start, size, prot, pages);
 328	flush_cache_vmap(start, start + size);
 329	return ret;
 330}
 331
 332int is_vmalloc_or_module_addr(const void *x)
 333{
 334	/*
 335	 * ARM, x86-64 and sparc64 put modules in a special place,
 336	 * and fall back on vmalloc() if that fails. Others
 337	 * just put it in the vmalloc space.
 338	 */
 339#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 340	unsigned long addr = (unsigned long)x;
 341	if (addr >= MODULES_VADDR && addr < MODULES_END)
 342		return 1;
 343#endif
 344	return is_vmalloc_addr(x);
 345}
 346
 347/*
 348 * Walk a vmap address to the struct page it maps.
 
 
 349 */
 350struct page *vmalloc_to_page(const void *vmalloc_addr)
 351{
 352	unsigned long addr = (unsigned long) vmalloc_addr;
 353	struct page *page = NULL;
 354	pgd_t *pgd = pgd_offset_k(addr);
 355	p4d_t *p4d;
 356	pud_t *pud;
 357	pmd_t *pmd;
 358	pte_t *ptep, pte;
 359
 360	/*
 361	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 362	 * architectures that do not vmalloc module space
 363	 */
 364	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 365
 366	if (pgd_none(*pgd))
 367		return NULL;
 
 
 
 
 
 368	p4d = p4d_offset(pgd, addr);
 369	if (p4d_none(*p4d))
 370		return NULL;
 
 
 
 
 
 371	pud = pud_offset(p4d, addr);
 
 
 
 
 
 
 372
 373	/*
 374	 * Don't dereference bad PUD or PMD (below) entries. This will also
 375	 * identify huge mappings, which we may encounter on architectures
 376	 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
 377	 * identified as vmalloc addresses by is_vmalloc_addr(), but are
 378	 * not [unambiguously] associated with a struct page, so there is
 379	 * no correct value to return for them.
 380	 */
 381	WARN_ON_ONCE(pud_bad(*pud));
 382	if (pud_none(*pud) || pud_bad(*pud))
 383		return NULL;
 384	pmd = pmd_offset(pud, addr);
 385	WARN_ON_ONCE(pmd_bad(*pmd));
 386	if (pmd_none(*pmd) || pmd_bad(*pmd))
 387		return NULL;
 388
 389	ptep = pte_offset_map(pmd, addr);
 390	pte = *ptep;
 391	if (pte_present(pte))
 392		page = pte_page(pte);
 393	pte_unmap(ptep);
 
 394	return page;
 395}
 396EXPORT_SYMBOL(vmalloc_to_page);
 397
 398/*
 399 * Map a vmalloc()-space virtual address to the physical page frame number.
 400 */
 401unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 402{
 403	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 404}
 405EXPORT_SYMBOL(vmalloc_to_pfn);
 406
 407
 408/*** Global kva allocator ***/
 409
 410#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
 411#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
 412
 413
 414static DEFINE_SPINLOCK(vmap_area_lock);
 415static DEFINE_SPINLOCK(free_vmap_area_lock);
 416/* Export for kexec only */
 417LIST_HEAD(vmap_area_list);
 418static LLIST_HEAD(vmap_purge_list);
 419static struct rb_root vmap_area_root = RB_ROOT;
 420static bool vmap_initialized __read_mostly;
 421
 
 
 
 
 422/*
 423 * This kmem_cache is used for vmap_area objects. Instead of
 424 * allocating from slab we reuse an object from this cache to
 425 * make things faster. Especially in "no edge" splitting of
 426 * free block.
 427 */
 428static struct kmem_cache *vmap_area_cachep;
 429
 430/*
 431 * This linked list is used in pair with free_vmap_area_root.
 432 * It gives O(1) access to prev/next to perform fast coalescing.
 433 */
 434static LIST_HEAD(free_vmap_area_list);
 435
 436/*
 437 * This augment red-black tree represents the free vmap space.
 438 * All vmap_area objects in this tree are sorted by va->va_start
 439 * address. It is used for allocation and merging when a vmap
 440 * object is released.
 441 *
 442 * Each vmap_area node contains a maximum available free block
 443 * of its sub-tree, right or left. Therefore it is possible to
 444 * find a lowest match of free area.
 445 */
 446static struct rb_root free_vmap_area_root = RB_ROOT;
 447
 448/*
 449 * Preload a CPU with one object for "no edge" split case. The
 450 * aim is to get rid of allocations from the atomic context, thus
 451 * to use more permissive allocation masks.
 452 */
 453static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
 454
 455static __always_inline unsigned long
 456va_size(struct vmap_area *va)
 457{
 458	return (va->va_end - va->va_start);
 459}
 460
 461static __always_inline unsigned long
 462get_subtree_max_size(struct rb_node *node)
 463{
 464	struct vmap_area *va;
 465
 466	va = rb_entry_safe(node, struct vmap_area, rb_node);
 467	return va ? va->subtree_max_size : 0;
 468}
 469
 470/*
 471 * Gets called when remove the node and rotate.
 472 */
 473static __always_inline unsigned long
 474compute_subtree_max_size(struct vmap_area *va)
 475{
 476	return max3(va_size(va),
 477		get_subtree_max_size(va->rb_node.rb_left),
 478		get_subtree_max_size(va->rb_node.rb_right));
 479}
 480
 481RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
 482	struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
 483
 484static void purge_vmap_area_lazy(void);
 485static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
 486static unsigned long lazy_max_pages(void);
 
 487
 488static atomic_long_t nr_vmalloc_pages;
 489
 490unsigned long vmalloc_nr_pages(void)
 491{
 492	return atomic_long_read(&nr_vmalloc_pages);
 493}
 494
 495static struct vmap_area *__find_vmap_area(unsigned long addr)
 
 496{
 
 497	struct rb_node *n = vmap_area_root.rb_node;
 498
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 499	while (n) {
 500		struct vmap_area *va;
 501
 502		va = rb_entry(n, struct vmap_area, rb_node);
 503		if (addr < va->va_start)
 504			n = n->rb_left;
 505		else if (addr >= va->va_end)
 506			n = n->rb_right;
 507		else
 508			return va;
 509	}
 510
 511	return NULL;
 512}
 513
 514/*
 515 * This function returns back addresses of parent node
 516 * and its left or right link for further processing.
 517 *
 518 * Otherwise NULL is returned. In that case all further
 519 * steps regarding inserting of conflicting overlap range
 520 * have to be declined and actually considered as a bug.
 521 */
 522static __always_inline struct rb_node **
 523find_va_links(struct vmap_area *va,
 524	struct rb_root *root, struct rb_node *from,
 525	struct rb_node **parent)
 526{
 527	struct vmap_area *tmp_va;
 528	struct rb_node **link;
 529
 530	if (root) {
 531		link = &root->rb_node;
 532		if (unlikely(!*link)) {
 533			*parent = NULL;
 534			return link;
 535		}
 536	} else {
 537		link = &from;
 538	}
 539
 540	/*
 541	 * Go to the bottom of the tree. When we hit the last point
 542	 * we end up with parent rb_node and correct direction, i name
 543	 * it link, where the new va->rb_node will be attached to.
 544	 */
 545	do {
 546		tmp_va = rb_entry(*link, struct vmap_area, rb_node);
 547
 548		/*
 549		 * During the traversal we also do some sanity check.
 550		 * Trigger the BUG() if there are sides(left/right)
 551		 * or full overlaps.
 552		 */
 553		if (va->va_start < tmp_va->va_end &&
 554				va->va_end <= tmp_va->va_start)
 555			link = &(*link)->rb_left;
 556		else if (va->va_end > tmp_va->va_start &&
 557				va->va_start >= tmp_va->va_end)
 558			link = &(*link)->rb_right;
 559		else {
 560			WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
 561				va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
 562
 563			return NULL;
 564		}
 565	} while (*link);
 566
 567	*parent = &tmp_va->rb_node;
 568	return link;
 569}
 570
 571static __always_inline struct list_head *
 572get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
 573{
 574	struct list_head *list;
 575
 576	if (unlikely(!parent))
 577		/*
 578		 * The red-black tree where we try to find VA neighbors
 579		 * before merging or inserting is empty, i.e. it means
 580		 * there is no free vmap space. Normally it does not
 581		 * happen but we handle this case anyway.
 582		 */
 583		return NULL;
 584
 585	list = &rb_entry(parent, struct vmap_area, rb_node)->list;
 586	return (&parent->rb_right == link ? list->next : list);
 587}
 588
 589static __always_inline void
 590link_va(struct vmap_area *va, struct rb_root *root,
 591	struct rb_node *parent, struct rb_node **link, struct list_head *head)
 
 592{
 593	/*
 594	 * VA is still not in the list, but we can
 595	 * identify its future previous list_head node.
 596	 */
 597	if (likely(parent)) {
 598		head = &rb_entry(parent, struct vmap_area, rb_node)->list;
 599		if (&parent->rb_right != link)
 600			head = head->prev;
 601	}
 602
 603	/* Insert to the rb-tree */
 604	rb_link_node(&va->rb_node, parent, link);
 605	if (root == &free_vmap_area_root) {
 606		/*
 607		 * Some explanation here. Just perform simple insertion
 608		 * to the tree. We do not set va->subtree_max_size to
 609		 * its current size before calling rb_insert_augmented().
 610		 * It is because of we populate the tree from the bottom
 611		 * to parent levels when the node _is_ in the tree.
 612		 *
 613		 * Therefore we set subtree_max_size to zero after insertion,
 614		 * to let __augment_tree_propagate_from() puts everything to
 615		 * the correct order later on.
 616		 */
 617		rb_insert_augmented(&va->rb_node,
 618			root, &free_vmap_area_rb_augment_cb);
 619		va->subtree_max_size = 0;
 620	} else {
 621		rb_insert_color(&va->rb_node, root);
 622	}
 623
 624	/* Address-sort this list */
 625	list_add(&va->list, head);
 626}
 627
 628static __always_inline void
 629unlink_va(struct vmap_area *va, struct rb_root *root)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 630{
 631	if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
 632		return;
 633
 634	if (root == &free_vmap_area_root)
 635		rb_erase_augmented(&va->rb_node,
 636			root, &free_vmap_area_rb_augment_cb);
 637	else
 638		rb_erase(&va->rb_node, root);
 639
 640	list_del(&va->list);
 641	RB_CLEAR_NODE(&va->rb_node);
 642}
 643
 
 
 
 
 
 
 
 
 
 
 
 
 644#if DEBUG_AUGMENT_PROPAGATE_CHECK
 
 
 
 
 
 
 
 
 
 
 
 645static void
 646augment_tree_propagate_check(void)
 647{
 648	struct vmap_area *va;
 649	unsigned long computed_size;
 650
 651	list_for_each_entry(va, &free_vmap_area_list, list) {
 652		computed_size = compute_subtree_max_size(va);
 653		if (computed_size != va->subtree_max_size)
 654			pr_emerg("tree is corrupted: %lu, %lu\n",
 655				va_size(va), va->subtree_max_size);
 656	}
 657}
 658#endif
 659
 660/*
 661 * This function populates subtree_max_size from bottom to upper
 662 * levels starting from VA point. The propagation must be done
 663 * when VA size is modified by changing its va_start/va_end. Or
 664 * in case of newly inserting of VA to the tree.
 665 *
 666 * It means that __augment_tree_propagate_from() must be called:
 667 * - After VA has been inserted to the tree(free path);
 668 * - After VA has been shrunk(allocation path);
 669 * - After VA has been increased(merging path).
 670 *
 671 * Please note that, it does not mean that upper parent nodes
 672 * and their subtree_max_size are recalculated all the time up
 673 * to the root node.
 674 *
 675 *       4--8
 676 *        /\
 677 *       /  \
 678 *      /    \
 679 *    2--2  8--8
 680 *
 681 * For example if we modify the node 4, shrinking it to 2, then
 682 * no any modification is required. If we shrink the node 2 to 1
 683 * its subtree_max_size is updated only, and set to 1. If we shrink
 684 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
 685 * node becomes 4--6.
 686 */
 687static __always_inline void
 688augment_tree_propagate_from(struct vmap_area *va)
 689{
 690	/*
 691	 * Populate the tree from bottom towards the root until
 692	 * the calculated maximum available size of checked node
 693	 * is equal to its current one.
 694	 */
 695	free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
 696
 697#if DEBUG_AUGMENT_PROPAGATE_CHECK
 698	augment_tree_propagate_check();
 699#endif
 700}
 701
 702static void
 703insert_vmap_area(struct vmap_area *va,
 704	struct rb_root *root, struct list_head *head)
 705{
 706	struct rb_node **link;
 707	struct rb_node *parent;
 708
 709	link = find_va_links(va, root, NULL, &parent);
 710	if (link)
 711		link_va(va, root, parent, link, head);
 712}
 713
 714static void
 715insert_vmap_area_augment(struct vmap_area *va,
 716	struct rb_node *from, struct rb_root *root,
 717	struct list_head *head)
 718{
 719	struct rb_node **link;
 720	struct rb_node *parent;
 721
 722	if (from)
 723		link = find_va_links(va, NULL, from, &parent);
 724	else
 725		link = find_va_links(va, root, NULL, &parent);
 726
 727	if (link) {
 728		link_va(va, root, parent, link, head);
 729		augment_tree_propagate_from(va);
 730	}
 731}
 732
 733/*
 734 * Merge de-allocated chunk of VA memory with previous
 735 * and next free blocks. If coalesce is not done a new
 736 * free area is inserted. If VA has been merged, it is
 737 * freed.
 738 *
 739 * Please note, it can return NULL in case of overlap
 740 * ranges, followed by WARN() report. Despite it is a
 741 * buggy behaviour, a system can be alive and keep
 742 * ongoing.
 743 */
 744static __always_inline struct vmap_area *
 745merge_or_add_vmap_area(struct vmap_area *va,
 746	struct rb_root *root, struct list_head *head)
 747{
 748	struct vmap_area *sibling;
 749	struct list_head *next;
 750	struct rb_node **link;
 751	struct rb_node *parent;
 752	bool merged = false;
 753
 754	/*
 755	 * Find a place in the tree where VA potentially will be
 756	 * inserted, unless it is merged with its sibling/siblings.
 757	 */
 758	link = find_va_links(va, root, NULL, &parent);
 759	if (!link)
 760		return NULL;
 761
 762	/*
 763	 * Get next node of VA to check if merging can be done.
 764	 */
 765	next = get_va_next_sibling(parent, link);
 766	if (unlikely(next == NULL))
 767		goto insert;
 768
 769	/*
 770	 * start            end
 771	 * |                |
 772	 * |<------VA------>|<-----Next----->|
 773	 *                  |                |
 774	 *                  start            end
 775	 */
 776	if (next != head) {
 777		sibling = list_entry(next, struct vmap_area, list);
 778		if (sibling->va_start == va->va_end) {
 779			sibling->va_start = va->va_start;
 780
 781			/* Free vmap_area object. */
 782			kmem_cache_free(vmap_area_cachep, va);
 783
 784			/* Point to the new merged area. */
 785			va = sibling;
 786			merged = true;
 787		}
 788	}
 789
 790	/*
 791	 * start            end
 792	 * |                |
 793	 * |<-----Prev----->|<------VA------>|
 794	 *                  |                |
 795	 *                  start            end
 796	 */
 797	if (next->prev != head) {
 798		sibling = list_entry(next->prev, struct vmap_area, list);
 799		if (sibling->va_end == va->va_start) {
 800			/*
 801			 * If both neighbors are coalesced, it is important
 802			 * to unlink the "next" node first, followed by merging
 803			 * with "previous" one. Otherwise the tree might not be
 804			 * fully populated if a sibling's augmented value is
 805			 * "normalized" because of rotation operations.
 806			 */
 807			if (merged)
 808				unlink_va(va, root);
 809
 810			sibling->va_end = va->va_end;
 811
 812			/* Free vmap_area object. */
 813			kmem_cache_free(vmap_area_cachep, va);
 814
 815			/* Point to the new merged area. */
 816			va = sibling;
 817			merged = true;
 818		}
 819	}
 820
 821insert:
 822	if (!merged)
 823		link_va(va, root, parent, link, head);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 824
 825	/*
 826	 * Last step is to check and update the tree.
 827	 */
 828	augment_tree_propagate_from(va);
 829	return va;
 830}
 831
 832static __always_inline bool
 833is_within_this_va(struct vmap_area *va, unsigned long size,
 834	unsigned long align, unsigned long vstart)
 835{
 836	unsigned long nva_start_addr;
 837
 838	if (va->va_start > vstart)
 839		nva_start_addr = ALIGN(va->va_start, align);
 840	else
 841		nva_start_addr = ALIGN(vstart, align);
 842
 843	/* Can be overflowed due to big size or alignment. */
 844	if (nva_start_addr + size < nva_start_addr ||
 845			nva_start_addr < vstart)
 846		return false;
 847
 848	return (nva_start_addr + size <= va->va_end);
 849}
 850
 851/*
 852 * Find the first free block(lowest start address) in the tree,
 853 * that will accomplish the request corresponding to passing
 854 * parameters.
 
 
 855 */
 856static __always_inline struct vmap_area *
 857find_vmap_lowest_match(unsigned long size,
 858	unsigned long align, unsigned long vstart)
 859{
 860	struct vmap_area *va;
 861	struct rb_node *node;
 862	unsigned long length;
 863
 864	/* Start from the root. */
 865	node = free_vmap_area_root.rb_node;
 866
 867	/* Adjust the search size for alignment overhead. */
 868	length = size + align - 1;
 869
 870	while (node) {
 871		va = rb_entry(node, struct vmap_area, rb_node);
 872
 873		if (get_subtree_max_size(node->rb_left) >= length &&
 874				vstart < va->va_start) {
 875			node = node->rb_left;
 876		} else {
 877			if (is_within_this_va(va, size, align, vstart))
 878				return va;
 879
 880			/*
 881			 * Does not make sense to go deeper towards the right
 882			 * sub-tree if it does not have a free block that is
 883			 * equal or bigger to the requested search length.
 884			 */
 885			if (get_subtree_max_size(node->rb_right) >= length) {
 886				node = node->rb_right;
 887				continue;
 888			}
 889
 890			/*
 891			 * OK. We roll back and find the first right sub-tree,
 892			 * that will satisfy the search criteria. It can happen
 893			 * only once due to "vstart" restriction.
 
 894			 */
 895			while ((node = rb_parent(node))) {
 896				va = rb_entry(node, struct vmap_area, rb_node);
 897				if (is_within_this_va(va, size, align, vstart))
 898					return va;
 899
 900				if (get_subtree_max_size(node->rb_right) >= length &&
 901						vstart <= va->va_start) {
 
 
 
 
 
 
 
 902					node = node->rb_right;
 903					break;
 904				}
 905			}
 906		}
 907	}
 908
 909	return NULL;
 910}
 911
 912#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
 913#include <linux/random.h>
 914
 915static struct vmap_area *
 916find_vmap_lowest_linear_match(unsigned long size,
 917	unsigned long align, unsigned long vstart)
 918{
 919	struct vmap_area *va;
 920
 921	list_for_each_entry(va, &free_vmap_area_list, list) {
 922		if (!is_within_this_va(va, size, align, vstart))
 923			continue;
 924
 925		return va;
 926	}
 927
 928	return NULL;
 929}
 930
 931static void
 932find_vmap_lowest_match_check(unsigned long size)
 
 933{
 934	struct vmap_area *va_1, *va_2;
 935	unsigned long vstart;
 936	unsigned int rnd;
 937
 938	get_random_bytes(&rnd, sizeof(rnd));
 939	vstart = VMALLOC_START + rnd;
 940
 941	va_1 = find_vmap_lowest_match(size, 1, vstart);
 942	va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
 943
 944	if (va_1 != va_2)
 945		pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
 946			va_1, va_2, vstart);
 947}
 948#endif
 949
 950enum fit_type {
 951	NOTHING_FIT = 0,
 952	FL_FIT_TYPE = 1,	/* full fit */
 953	LE_FIT_TYPE = 2,	/* left edge fit */
 954	RE_FIT_TYPE = 3,	/* right edge fit */
 955	NE_FIT_TYPE = 4		/* no edge fit */
 956};
 957
 958static __always_inline enum fit_type
 959classify_va_fit_type(struct vmap_area *va,
 960	unsigned long nva_start_addr, unsigned long size)
 961{
 962	enum fit_type type;
 963
 964	/* Check if it is within VA. */
 965	if (nva_start_addr < va->va_start ||
 966			nva_start_addr + size > va->va_end)
 967		return NOTHING_FIT;
 968
 969	/* Now classify. */
 970	if (va->va_start == nva_start_addr) {
 971		if (va->va_end == nva_start_addr + size)
 972			type = FL_FIT_TYPE;
 973		else
 974			type = LE_FIT_TYPE;
 975	} else if (va->va_end == nva_start_addr + size) {
 976		type = RE_FIT_TYPE;
 977	} else {
 978		type = NE_FIT_TYPE;
 979	}
 980
 981	return type;
 982}
 983
 984static __always_inline int
 985adjust_va_to_fit_type(struct vmap_area *va,
 986	unsigned long nva_start_addr, unsigned long size,
 987	enum fit_type type)
 988{
 989	struct vmap_area *lva = NULL;
 
 990
 991	if (type == FL_FIT_TYPE) {
 992		/*
 993		 * No need to split VA, it fully fits.
 994		 *
 995		 * |               |
 996		 * V      NVA      V
 997		 * |---------------|
 998		 */
 999		unlink_va(va, &free_vmap_area_root);
1000		kmem_cache_free(vmap_area_cachep, va);
1001	} else if (type == LE_FIT_TYPE) {
1002		/*
1003		 * Split left edge of fit VA.
1004		 *
1005		 * |       |
1006		 * V  NVA  V   R
1007		 * |-------|-------|
1008		 */
1009		va->va_start += size;
1010	} else if (type == RE_FIT_TYPE) {
1011		/*
1012		 * Split right edge of fit VA.
1013		 *
1014		 *         |       |
1015		 *     L   V  NVA  V
1016		 * |-------|-------|
1017		 */
1018		va->va_end = nva_start_addr;
1019	} else if (type == NE_FIT_TYPE) {
1020		/*
1021		 * Split no edge of fit VA.
1022		 *
1023		 *     |       |
1024		 *   L V  NVA  V R
1025		 * |---|-------|---|
1026		 */
1027		lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1028		if (unlikely(!lva)) {
1029			/*
1030			 * For percpu allocator we do not do any pre-allocation
1031			 * and leave it as it is. The reason is it most likely
1032			 * never ends up with NE_FIT_TYPE splitting. In case of
1033			 * percpu allocations offsets and sizes are aligned to
1034			 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1035			 * are its main fitting cases.
1036			 *
1037			 * There are a few exceptions though, as an example it is
1038			 * a first allocation (early boot up) when we have "one"
1039			 * big free space that has to be split.
1040			 *
1041			 * Also we can hit this path in case of regular "vmap"
1042			 * allocations, if "this" current CPU was not preloaded.
1043			 * See the comment in alloc_vmap_area() why. If so, then
1044			 * GFP_NOWAIT is used instead to get an extra object for
1045			 * split purpose. That is rare and most time does not
1046			 * occur.
1047			 *
1048			 * What happens if an allocation gets failed. Basically,
1049			 * an "overflow" path is triggered to purge lazily freed
1050			 * areas to free some memory, then, the "retry" path is
1051			 * triggered to repeat one more time. See more details
1052			 * in alloc_vmap_area() function.
1053			 */
1054			lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1055			if (!lva)
1056				return -1;
1057		}
1058
1059		/*
1060		 * Build the remainder.
1061		 */
1062		lva->va_start = va->va_start;
1063		lva->va_end = nva_start_addr;
1064
1065		/*
1066		 * Shrink this VA to remaining size.
1067		 */
1068		va->va_start = nva_start_addr + size;
1069	} else {
1070		return -1;
1071	}
1072
1073	if (type != FL_FIT_TYPE) {
1074		augment_tree_propagate_from(va);
1075
1076		if (lva)	/* type == NE_FIT_TYPE */
1077			insert_vmap_area_augment(lva, &va->rb_node,
1078				&free_vmap_area_root, &free_vmap_area_list);
1079	}
1080
1081	return 0;
1082}
1083
1084/*
1085 * Returns a start address of the newly allocated area, if success.
1086 * Otherwise a vend is returned that indicates failure.
1087 */
1088static __always_inline unsigned long
1089__alloc_vmap_area(unsigned long size, unsigned long align,
 
1090	unsigned long vstart, unsigned long vend)
1091{
 
1092	unsigned long nva_start_addr;
1093	struct vmap_area *va;
1094	enum fit_type type;
1095	int ret;
1096
1097	va = find_vmap_lowest_match(size, align, vstart);
 
 
 
 
 
 
 
 
 
 
 
 
1098	if (unlikely(!va))
1099		return vend;
1100
1101	if (va->va_start > vstart)
1102		nva_start_addr = ALIGN(va->va_start, align);
1103	else
1104		nva_start_addr = ALIGN(vstart, align);
1105
1106	/* Check the "vend" restriction. */
1107	if (nva_start_addr + size > vend)
1108		return vend;
1109
1110	/* Classify what we have found. */
1111	type = classify_va_fit_type(va, nva_start_addr, size);
1112	if (WARN_ON_ONCE(type == NOTHING_FIT))
1113		return vend;
1114
1115	/* Update the free vmap_area. */
1116	ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1117	if (ret)
1118		return vend;
1119
1120#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1121	find_vmap_lowest_match_check(size);
1122#endif
1123
1124	return nva_start_addr;
1125}
1126
1127/*
1128 * Free a region of KVA allocated by alloc_vmap_area
1129 */
1130static void free_vmap_area(struct vmap_area *va)
1131{
1132	/*
1133	 * Remove from the busy tree/list.
1134	 */
1135	spin_lock(&vmap_area_lock);
1136	unlink_va(va, &vmap_area_root);
1137	spin_unlock(&vmap_area_lock);
1138
1139	/*
1140	 * Insert/Merge it back to the free tree/list.
1141	 */
1142	spin_lock(&free_vmap_area_lock);
1143	merge_or_add_vmap_area(va, &free_vmap_area_root, &free_vmap_area_list);
1144	spin_unlock(&free_vmap_area_lock);
1145}
1146
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1147/*
1148 * Allocate a region of KVA of the specified size and alignment, within the
1149 * vstart and vend.
1150 */
1151static struct vmap_area *alloc_vmap_area(unsigned long size,
1152				unsigned long align,
1153				unsigned long vstart, unsigned long vend,
1154				int node, gfp_t gfp_mask)
1155{
1156	struct vmap_area *va, *pva;
 
1157	unsigned long addr;
1158	int purged = 0;
1159	int ret;
1160
1161	BUG_ON(!size);
1162	BUG_ON(offset_in_page(size));
1163	BUG_ON(!is_power_of_2(align));
1164
1165	if (unlikely(!vmap_initialized))
1166		return ERR_PTR(-EBUSY);
1167
1168	might_sleep();
1169	gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1170
1171	va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1172	if (unlikely(!va))
1173		return ERR_PTR(-ENOMEM);
1174
1175	/*
1176	 * Only scan the relevant parts containing pointers to other objects
1177	 * to avoid false negatives.
1178	 */
1179	kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1180
1181retry:
1182	/*
1183	 * Preload this CPU with one extra vmap_area object. It is used
1184	 * when fit type of free area is NE_FIT_TYPE. Please note, it
1185	 * does not guarantee that an allocation occurs on a CPU that
1186	 * is preloaded, instead we minimize the case when it is not.
1187	 * It can happen because of cpu migration, because there is a
1188	 * race until the below spinlock is taken.
1189	 *
1190	 * The preload is done in non-atomic context, thus it allows us
1191	 * to use more permissive allocation masks to be more stable under
1192	 * low memory condition and high memory pressure. In rare case,
1193	 * if not preloaded, GFP_NOWAIT is used.
1194	 *
1195	 * Set "pva" to NULL here, because of "retry" path.
1196	 */
1197	pva = NULL;
1198
1199	if (!this_cpu_read(ne_fit_preload_node))
1200		/*
1201		 * Even if it fails we do not really care about that.
1202		 * Just proceed as it is. If needed "overflow" path
1203		 * will refill the cache we allocate from.
1204		 */
1205		pva = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1206
1207	spin_lock(&free_vmap_area_lock);
1208
1209	if (pva && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva))
1210		kmem_cache_free(vmap_area_cachep, pva);
1211
1212	/*
1213	 * If an allocation fails, the "vend" address is
1214	 * returned. Therefore trigger the overflow path.
1215	 */
1216	addr = __alloc_vmap_area(size, align, vstart, vend);
1217	spin_unlock(&free_vmap_area_lock);
1218
1219	if (unlikely(addr == vend))
1220		goto overflow;
1221
1222	va->va_start = addr;
1223	va->va_end = addr + size;
1224	va->vm = NULL;
1225
1226
1227	spin_lock(&vmap_area_lock);
1228	insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1229	spin_unlock(&vmap_area_lock);
1230
1231	BUG_ON(!IS_ALIGNED(va->va_start, align));
1232	BUG_ON(va->va_start < vstart);
1233	BUG_ON(va->va_end > vend);
1234
1235	ret = kasan_populate_vmalloc(addr, size);
1236	if (ret) {
1237		free_vmap_area(va);
1238		return ERR_PTR(ret);
1239	}
1240
1241	return va;
1242
1243overflow:
1244	if (!purged) {
1245		purge_vmap_area_lazy();
1246		purged = 1;
1247		goto retry;
1248	}
1249
1250	if (gfpflags_allow_blocking(gfp_mask)) {
1251		unsigned long freed = 0;
1252		blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1253		if (freed > 0) {
1254			purged = 0;
1255			goto retry;
1256		}
1257	}
1258
1259	if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1260		pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1261			size);
1262
1263	kmem_cache_free(vmap_area_cachep, va);
1264	return ERR_PTR(-EBUSY);
1265}
1266
1267int register_vmap_purge_notifier(struct notifier_block *nb)
1268{
1269	return blocking_notifier_chain_register(&vmap_notify_list, nb);
1270}
1271EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1272
1273int unregister_vmap_purge_notifier(struct notifier_block *nb)
1274{
1275	return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1276}
1277EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1278
1279/*
1280 * lazy_max_pages is the maximum amount of virtual address space we gather up
1281 * before attempting to purge with a TLB flush.
1282 *
1283 * There is a tradeoff here: a larger number will cover more kernel page tables
1284 * and take slightly longer to purge, but it will linearly reduce the number of
1285 * global TLB flushes that must be performed. It would seem natural to scale
1286 * this number up linearly with the number of CPUs (because vmapping activity
1287 * could also scale linearly with the number of CPUs), however it is likely
1288 * that in practice, workloads might be constrained in other ways that mean
1289 * vmap activity will not scale linearly with CPUs. Also, I want to be
1290 * conservative and not introduce a big latency on huge systems, so go with
1291 * a less aggressive log scale. It will still be an improvement over the old
1292 * code, and it will be simple to change the scale factor if we find that it
1293 * becomes a problem on bigger systems.
1294 */
1295static unsigned long lazy_max_pages(void)
1296{
1297	unsigned int log;
1298
1299	log = fls(num_online_cpus());
1300
1301	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1302}
1303
1304static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1305
1306/*
1307 * Serialize vmap purging.  There is no actual criticial section protected
1308 * by this look, but we want to avoid concurrent calls for performance
1309 * reasons and to make the pcpu_get_vm_areas more deterministic.
1310 */
1311static DEFINE_MUTEX(vmap_purge_lock);
1312
1313/* for per-CPU blocks */
1314static void purge_fragmented_blocks_allcpus(void);
1315
1316/*
1317 * called before a call to iounmap() if the caller wants vm_area_struct's
1318 * immediately freed.
1319 */
1320void set_iounmap_nonlazy(void)
1321{
1322	atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1323}
1324
1325/*
1326 * Purges all lazily-freed vmap areas.
1327 */
1328static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1329{
1330	unsigned long resched_threshold;
1331	struct llist_node *valist;
1332	struct vmap_area *va;
1333	struct vmap_area *n_va;
1334
1335	lockdep_assert_held(&vmap_purge_lock);
1336
1337	valist = llist_del_all(&vmap_purge_list);
1338	if (unlikely(valist == NULL))
1339		return false;
1340
1341	/*
1342	 * TODO: to calculate a flush range without looping.
1343	 * The list can be up to lazy_max_pages() elements.
1344	 */
1345	llist_for_each_entry(va, valist, purge_list) {
1346		if (va->va_start < start)
1347			start = va->va_start;
1348		if (va->va_end > end)
1349			end = va->va_end;
1350	}
 
1351
1352	flush_tlb_kernel_range(start, end);
1353	resched_threshold = lazy_max_pages() << 1;
1354
1355	spin_lock(&free_vmap_area_lock);
1356	llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1357		unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1358		unsigned long orig_start = va->va_start;
1359		unsigned long orig_end = va->va_end;
1360
1361		/*
1362		 * Finally insert or merge lazily-freed area. It is
1363		 * detached and there is no need to "unlink" it from
1364		 * anything.
1365		 */
1366		va = merge_or_add_vmap_area(va, &free_vmap_area_root,
1367					    &free_vmap_area_list);
1368
1369		if (!va)
1370			continue;
1371
1372		if (is_vmalloc_or_module_addr((void *)orig_start))
1373			kasan_release_vmalloc(orig_start, orig_end,
1374					      va->va_start, va->va_end);
1375
1376		atomic_long_sub(nr, &vmap_lazy_nr);
 
1377
1378		if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1379			cond_resched_lock(&free_vmap_area_lock);
1380	}
1381	spin_unlock(&free_vmap_area_lock);
1382	return true;
1383}
1384
1385/*
1386 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1387 * is already purging.
1388 */
1389static void try_purge_vmap_area_lazy(void)
1390{
1391	if (mutex_trylock(&vmap_purge_lock)) {
1392		__purge_vmap_area_lazy(ULONG_MAX, 0);
1393		mutex_unlock(&vmap_purge_lock);
1394	}
1395}
1396
1397/*
1398 * Kick off a purge of the outstanding lazy areas.
1399 */
1400static void purge_vmap_area_lazy(void)
1401{
1402	mutex_lock(&vmap_purge_lock);
1403	purge_fragmented_blocks_allcpus();
1404	__purge_vmap_area_lazy(ULONG_MAX, 0);
1405	mutex_unlock(&vmap_purge_lock);
1406}
1407
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1408/*
1409 * Free a vmap area, caller ensuring that the area has been unmapped
1410 * and flush_cache_vunmap had been called for the correct range
1411 * previously.
1412 */
1413static void free_vmap_area_noflush(struct vmap_area *va)
1414{
 
 
1415	unsigned long nr_lazy;
1416
1417	spin_lock(&vmap_area_lock);
1418	unlink_va(va, &vmap_area_root);
1419	spin_unlock(&vmap_area_lock);
1420
1421	nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1422				PAGE_SHIFT, &vmap_lazy_nr);
1423
 
 
 
 
 
 
 
 
 
 
1424	/* After this point, we may free va at any time */
1425	llist_add(&va->purge_list, &vmap_purge_list);
1426
1427	if (unlikely(nr_lazy > lazy_max_pages()))
1428		try_purge_vmap_area_lazy();
1429}
1430
1431/*
1432 * Free and unmap a vmap area
1433 */
1434static void free_unmap_vmap_area(struct vmap_area *va)
1435{
1436	flush_cache_vunmap(va->va_start, va->va_end);
1437	unmap_kernel_range_noflush(va->va_start, va->va_end - va->va_start);
1438	if (debug_pagealloc_enabled_static())
1439		flush_tlb_kernel_range(va->va_start, va->va_end);
1440
1441	free_vmap_area_noflush(va);
1442}
1443
1444static struct vmap_area *find_vmap_area(unsigned long addr)
1445{
1446	struct vmap_area *va;
1447
1448	spin_lock(&vmap_area_lock);
1449	va = __find_vmap_area(addr);
1450	spin_unlock(&vmap_area_lock);
1451
1452	return va;
1453}
1454
1455/*** Per cpu kva allocator ***/
1456
1457/*
1458 * vmap space is limited especially on 32 bit architectures. Ensure there is
1459 * room for at least 16 percpu vmap blocks per CPU.
1460 */
1461/*
1462 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1463 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
1464 * instead (we just need a rough idea)
1465 */
1466#if BITS_PER_LONG == 32
1467#define VMALLOC_SPACE		(128UL*1024*1024)
1468#else
1469#define VMALLOC_SPACE		(128UL*1024*1024*1024)
1470#endif
1471
1472#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
1473#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
1474#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
1475#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
1476#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
1477#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
1478#define VMAP_BBMAP_BITS		\
1479		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
1480		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
1481			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1482
1483#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
1484
1485struct vmap_block_queue {
1486	spinlock_t lock;
1487	struct list_head free;
1488};
1489
1490struct vmap_block {
1491	spinlock_t lock;
1492	struct vmap_area *va;
1493	unsigned long free, dirty;
1494	unsigned long dirty_min, dirty_max; /*< dirty range */
1495	struct list_head free_list;
1496	struct rcu_head rcu_head;
1497	struct list_head purge;
1498};
1499
1500/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1501static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1502
1503/*
1504 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1505 * in the free path. Could get rid of this if we change the API to return a
1506 * "cookie" from alloc, to be passed to free. But no big deal yet.
1507 */
1508static DEFINE_XARRAY(vmap_blocks);
1509
1510/*
1511 * We should probably have a fallback mechanism to allocate virtual memory
1512 * out of partially filled vmap blocks. However vmap block sizing should be
1513 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1514 * big problem.
1515 */
1516
1517static unsigned long addr_to_vb_idx(unsigned long addr)
1518{
1519	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1520	addr /= VMAP_BLOCK_SIZE;
1521	return addr;
1522}
1523
1524static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1525{
1526	unsigned long addr;
1527
1528	addr = va_start + (pages_off << PAGE_SHIFT);
1529	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1530	return (void *)addr;
1531}
1532
1533/**
1534 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1535 *                  block. Of course pages number can't exceed VMAP_BBMAP_BITS
1536 * @order:    how many 2^order pages should be occupied in newly allocated block
1537 * @gfp_mask: flags for the page level allocator
1538 *
1539 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1540 */
1541static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1542{
1543	struct vmap_block_queue *vbq;
1544	struct vmap_block *vb;
1545	struct vmap_area *va;
1546	unsigned long vb_idx;
1547	int node, err;
1548	void *vaddr;
1549
1550	node = numa_node_id();
1551
1552	vb = kmalloc_node(sizeof(struct vmap_block),
1553			gfp_mask & GFP_RECLAIM_MASK, node);
1554	if (unlikely(!vb))
1555		return ERR_PTR(-ENOMEM);
1556
1557	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1558					VMALLOC_START, VMALLOC_END,
1559					node, gfp_mask);
1560	if (IS_ERR(va)) {
1561		kfree(vb);
1562		return ERR_CAST(va);
1563	}
1564
1565	vaddr = vmap_block_vaddr(va->va_start, 0);
1566	spin_lock_init(&vb->lock);
1567	vb->va = va;
1568	/* At least something should be left free */
1569	BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1570	vb->free = VMAP_BBMAP_BITS - (1UL << order);
1571	vb->dirty = 0;
1572	vb->dirty_min = VMAP_BBMAP_BITS;
1573	vb->dirty_max = 0;
1574	INIT_LIST_HEAD(&vb->free_list);
1575
1576	vb_idx = addr_to_vb_idx(va->va_start);
1577	err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1578	if (err) {
1579		kfree(vb);
1580		free_vmap_area(va);
1581		return ERR_PTR(err);
1582	}
1583
1584	vbq = &get_cpu_var(vmap_block_queue);
1585	spin_lock(&vbq->lock);
1586	list_add_tail_rcu(&vb->free_list, &vbq->free);
1587	spin_unlock(&vbq->lock);
1588	put_cpu_var(vmap_block_queue);
1589
1590	return vaddr;
1591}
1592
1593static void free_vmap_block(struct vmap_block *vb)
1594{
1595	struct vmap_block *tmp;
1596
1597	tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1598	BUG_ON(tmp != vb);
1599
1600	free_vmap_area_noflush(vb->va);
1601	kfree_rcu(vb, rcu_head);
1602}
1603
1604static void purge_fragmented_blocks(int cpu)
1605{
1606	LIST_HEAD(purge);
1607	struct vmap_block *vb;
1608	struct vmap_block *n_vb;
1609	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1610
1611	rcu_read_lock();
1612	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1613
1614		if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1615			continue;
1616
1617		spin_lock(&vb->lock);
1618		if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1619			vb->free = 0; /* prevent further allocs after releasing lock */
1620			vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1621			vb->dirty_min = 0;
1622			vb->dirty_max = VMAP_BBMAP_BITS;
1623			spin_lock(&vbq->lock);
1624			list_del_rcu(&vb->free_list);
1625			spin_unlock(&vbq->lock);
1626			spin_unlock(&vb->lock);
1627			list_add_tail(&vb->purge, &purge);
1628		} else
1629			spin_unlock(&vb->lock);
1630	}
1631	rcu_read_unlock();
1632
1633	list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1634		list_del(&vb->purge);
1635		free_vmap_block(vb);
1636	}
1637}
1638
1639static void purge_fragmented_blocks_allcpus(void)
1640{
1641	int cpu;
1642
1643	for_each_possible_cpu(cpu)
1644		purge_fragmented_blocks(cpu);
1645}
1646
1647static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1648{
1649	struct vmap_block_queue *vbq;
1650	struct vmap_block *vb;
1651	void *vaddr = NULL;
1652	unsigned int order;
1653
1654	BUG_ON(offset_in_page(size));
1655	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1656	if (WARN_ON(size == 0)) {
1657		/*
1658		 * Allocating 0 bytes isn't what caller wants since
1659		 * get_order(0) returns funny result. Just warn and terminate
1660		 * early.
1661		 */
1662		return NULL;
1663	}
1664	order = get_order(size);
1665
1666	rcu_read_lock();
1667	vbq = &get_cpu_var(vmap_block_queue);
1668	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1669		unsigned long pages_off;
1670
1671		spin_lock(&vb->lock);
1672		if (vb->free < (1UL << order)) {
1673			spin_unlock(&vb->lock);
1674			continue;
1675		}
1676
1677		pages_off = VMAP_BBMAP_BITS - vb->free;
1678		vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1679		vb->free -= 1UL << order;
1680		if (vb->free == 0) {
1681			spin_lock(&vbq->lock);
1682			list_del_rcu(&vb->free_list);
1683			spin_unlock(&vbq->lock);
1684		}
1685
1686		spin_unlock(&vb->lock);
1687		break;
1688	}
1689
1690	put_cpu_var(vmap_block_queue);
1691	rcu_read_unlock();
1692
1693	/* Allocate new block if nothing was found */
1694	if (!vaddr)
1695		vaddr = new_vmap_block(order, gfp_mask);
1696
1697	return vaddr;
1698}
1699
1700static void vb_free(unsigned long addr, unsigned long size)
1701{
1702	unsigned long offset;
1703	unsigned int order;
1704	struct vmap_block *vb;
1705
1706	BUG_ON(offset_in_page(size));
1707	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1708
1709	flush_cache_vunmap(addr, addr + size);
1710
1711	order = get_order(size);
1712	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
1713	vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
1714
1715	unmap_kernel_range_noflush(addr, size);
1716
1717	if (debug_pagealloc_enabled_static())
1718		flush_tlb_kernel_range(addr, addr + size);
1719
1720	spin_lock(&vb->lock);
1721
1722	/* Expand dirty range */
1723	vb->dirty_min = min(vb->dirty_min, offset);
1724	vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1725
1726	vb->dirty += 1UL << order;
1727	if (vb->dirty == VMAP_BBMAP_BITS) {
1728		BUG_ON(vb->free);
1729		spin_unlock(&vb->lock);
1730		free_vmap_block(vb);
1731	} else
1732		spin_unlock(&vb->lock);
1733}
1734
1735static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1736{
1737	int cpu;
1738
1739	if (unlikely(!vmap_initialized))
1740		return;
1741
1742	might_sleep();
1743
1744	for_each_possible_cpu(cpu) {
1745		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1746		struct vmap_block *vb;
1747
1748		rcu_read_lock();
1749		list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1750			spin_lock(&vb->lock);
1751			if (vb->dirty) {
1752				unsigned long va_start = vb->va->va_start;
1753				unsigned long s, e;
1754
1755				s = va_start + (vb->dirty_min << PAGE_SHIFT);
1756				e = va_start + (vb->dirty_max << PAGE_SHIFT);
1757
1758				start = min(s, start);
1759				end   = max(e, end);
1760
1761				flush = 1;
1762			}
1763			spin_unlock(&vb->lock);
1764		}
1765		rcu_read_unlock();
1766	}
1767
1768	mutex_lock(&vmap_purge_lock);
1769	purge_fragmented_blocks_allcpus();
1770	if (!__purge_vmap_area_lazy(start, end) && flush)
1771		flush_tlb_kernel_range(start, end);
1772	mutex_unlock(&vmap_purge_lock);
1773}
1774
1775/**
1776 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1777 *
1778 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1779 * to amortize TLB flushing overheads. What this means is that any page you
1780 * have now, may, in a former life, have been mapped into kernel virtual
1781 * address by the vmap layer and so there might be some CPUs with TLB entries
1782 * still referencing that page (additional to the regular 1:1 kernel mapping).
1783 *
1784 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1785 * be sure that none of the pages we have control over will have any aliases
1786 * from the vmap layer.
1787 */
1788void vm_unmap_aliases(void)
1789{
1790	unsigned long start = ULONG_MAX, end = 0;
1791	int flush = 0;
1792
1793	_vm_unmap_aliases(start, end, flush);
1794}
1795EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1796
1797/**
1798 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1799 * @mem: the pointer returned by vm_map_ram
1800 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1801 */
1802void vm_unmap_ram(const void *mem, unsigned int count)
1803{
1804	unsigned long size = (unsigned long)count << PAGE_SHIFT;
1805	unsigned long addr = (unsigned long)mem;
1806	struct vmap_area *va;
1807
1808	might_sleep();
1809	BUG_ON(!addr);
1810	BUG_ON(addr < VMALLOC_START);
1811	BUG_ON(addr > VMALLOC_END);
1812	BUG_ON(!PAGE_ALIGNED(addr));
1813
1814	kasan_poison_vmalloc(mem, size);
1815
1816	if (likely(count <= VMAP_MAX_ALLOC)) {
1817		debug_check_no_locks_freed(mem, size);
1818		vb_free(addr, size);
1819		return;
1820	}
1821
1822	va = find_vmap_area(addr);
1823	BUG_ON(!va);
1824	debug_check_no_locks_freed((void *)va->va_start,
1825				    (va->va_end - va->va_start));
1826	free_unmap_vmap_area(va);
1827}
1828EXPORT_SYMBOL(vm_unmap_ram);
1829
1830/**
1831 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1832 * @pages: an array of pointers to the pages to be mapped
1833 * @count: number of pages
1834 * @node: prefer to allocate data structures on this node
1835 *
1836 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1837 * faster than vmap so it's good.  But if you mix long-life and short-life
1838 * objects with vm_map_ram(), it could consume lots of address space through
1839 * fragmentation (especially on a 32bit machine).  You could see failures in
1840 * the end.  Please use this function for short-lived objects.
1841 *
1842 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1843 */
1844void *vm_map_ram(struct page **pages, unsigned int count, int node)
1845{
1846	unsigned long size = (unsigned long)count << PAGE_SHIFT;
1847	unsigned long addr;
1848	void *mem;
1849
1850	if (likely(count <= VMAP_MAX_ALLOC)) {
1851		mem = vb_alloc(size, GFP_KERNEL);
1852		if (IS_ERR(mem))
1853			return NULL;
1854		addr = (unsigned long)mem;
1855	} else {
1856		struct vmap_area *va;
1857		va = alloc_vmap_area(size, PAGE_SIZE,
1858				VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1859		if (IS_ERR(va))
1860			return NULL;
1861
1862		addr = va->va_start;
1863		mem = (void *)addr;
1864	}
1865
1866	kasan_unpoison_vmalloc(mem, size);
1867
1868	if (map_kernel_range(addr, size, PAGE_KERNEL, pages) < 0) {
1869		vm_unmap_ram(mem, count);
1870		return NULL;
1871	}
 
 
 
 
 
 
 
 
1872	return mem;
1873}
1874EXPORT_SYMBOL(vm_map_ram);
1875
1876static struct vm_struct *vmlist __initdata;
1877
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1878/**
1879 * vm_area_add_early - add vmap area early during boot
1880 * @vm: vm_struct to add
1881 *
1882 * This function is used to add fixed kernel vm area to vmlist before
1883 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
1884 * should contain proper values and the other fields should be zero.
1885 *
1886 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1887 */
1888void __init vm_area_add_early(struct vm_struct *vm)
1889{
1890	struct vm_struct *tmp, **p;
1891
1892	BUG_ON(vmap_initialized);
1893	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1894		if (tmp->addr >= vm->addr) {
1895			BUG_ON(tmp->addr < vm->addr + vm->size);
1896			break;
1897		} else
1898			BUG_ON(tmp->addr + tmp->size > vm->addr);
1899	}
1900	vm->next = *p;
1901	*p = vm;
1902}
1903
1904/**
1905 * vm_area_register_early - register vmap area early during boot
1906 * @vm: vm_struct to register
1907 * @align: requested alignment
1908 *
1909 * This function is used to register kernel vm area before
1910 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
1911 * proper values on entry and other fields should be zero.  On return,
1912 * vm->addr contains the allocated address.
1913 *
1914 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1915 */
1916void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1917{
1918	static size_t vm_init_off __initdata;
1919	unsigned long addr;
 
 
1920
1921	addr = ALIGN(VMALLOC_START + vm_init_off, align);
1922	vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
 
 
 
1923
 
1924	vm->addr = (void *)addr;
1925
1926	vm_area_add_early(vm);
 
1927}
1928
1929static void vmap_init_free_space(void)
1930{
1931	unsigned long vmap_start = 1;
1932	const unsigned long vmap_end = ULONG_MAX;
1933	struct vmap_area *busy, *free;
1934
1935	/*
1936	 *     B     F     B     B     B     F
1937	 * -|-----|.....|-----|-----|-----|.....|-
1938	 *  |           The KVA space           |
1939	 *  |<--------------------------------->|
1940	 */
1941	list_for_each_entry(busy, &vmap_area_list, list) {
1942		if (busy->va_start - vmap_start > 0) {
1943			free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1944			if (!WARN_ON_ONCE(!free)) {
1945				free->va_start = vmap_start;
1946				free->va_end = busy->va_start;
1947
1948				insert_vmap_area_augment(free, NULL,
1949					&free_vmap_area_root,
1950						&free_vmap_area_list);
1951			}
1952		}
1953
1954		vmap_start = busy->va_end;
1955	}
1956
1957	if (vmap_end - vmap_start > 0) {
1958		free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1959		if (!WARN_ON_ONCE(!free)) {
1960			free->va_start = vmap_start;
1961			free->va_end = vmap_end;
1962
1963			insert_vmap_area_augment(free, NULL,
1964				&free_vmap_area_root,
1965					&free_vmap_area_list);
1966		}
1967	}
1968}
1969
1970void __init vmalloc_init(void)
1971{
1972	struct vmap_area *va;
1973	struct vm_struct *tmp;
1974	int i;
1975
1976	/*
1977	 * Create the cache for vmap_area objects.
1978	 */
1979	vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1980
1981	for_each_possible_cpu(i) {
1982		struct vmap_block_queue *vbq;
1983		struct vfree_deferred *p;
1984
1985		vbq = &per_cpu(vmap_block_queue, i);
1986		spin_lock_init(&vbq->lock);
1987		INIT_LIST_HEAD(&vbq->free);
1988		p = &per_cpu(vfree_deferred, i);
1989		init_llist_head(&p->list);
1990		INIT_WORK(&p->wq, free_work);
1991	}
1992
1993	/* Import existing vmlist entries. */
1994	for (tmp = vmlist; tmp; tmp = tmp->next) {
1995		va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1996		if (WARN_ON_ONCE(!va))
1997			continue;
1998
1999		va->va_start = (unsigned long)tmp->addr;
2000		va->va_end = va->va_start + tmp->size;
2001		va->vm = tmp;
2002		insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2003	}
2004
2005	/*
2006	 * Now we can initialize a free vmap space.
2007	 */
2008	vmap_init_free_space();
2009	vmap_initialized = true;
2010}
2011
2012/**
2013 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2014 * @addr: start of the VM area to unmap
2015 * @size: size of the VM area to unmap
2016 *
2017 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2018 * the unmapping and tlb after.
2019 */
2020void unmap_kernel_range(unsigned long addr, unsigned long size)
2021{
2022	unsigned long end = addr + size;
2023
2024	flush_cache_vunmap(addr, end);
2025	unmap_kernel_range_noflush(addr, size);
2026	flush_tlb_kernel_range(addr, end);
2027}
2028
2029static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2030	struct vmap_area *va, unsigned long flags, const void *caller)
2031{
2032	vm->flags = flags;
2033	vm->addr = (void *)va->va_start;
2034	vm->size = va->va_end - va->va_start;
2035	vm->caller = caller;
2036	va->vm = vm;
2037}
2038
2039static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2040			      unsigned long flags, const void *caller)
2041{
2042	spin_lock(&vmap_area_lock);
2043	setup_vmalloc_vm_locked(vm, va, flags, caller);
2044	spin_unlock(&vmap_area_lock);
2045}
2046
2047static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2048{
2049	/*
2050	 * Before removing VM_UNINITIALIZED,
2051	 * we should make sure that vm has proper values.
2052	 * Pair with smp_rmb() in show_numa_info().
2053	 */
2054	smp_wmb();
2055	vm->flags &= ~VM_UNINITIALIZED;
2056}
2057
2058static struct vm_struct *__get_vm_area_node(unsigned long size,
2059		unsigned long align, unsigned long flags, unsigned long start,
2060		unsigned long end, int node, gfp_t gfp_mask, const void *caller)
 
2061{
2062	struct vmap_area *va;
2063	struct vm_struct *area;
2064	unsigned long requested_size = size;
2065
2066	BUG_ON(in_interrupt());
2067	size = PAGE_ALIGN(size);
2068	if (unlikely(!size))
2069		return NULL;
2070
2071	if (flags & VM_IOREMAP)
2072		align = 1ul << clamp_t(int, get_count_order_long(size),
2073				       PAGE_SHIFT, IOREMAP_MAX_ORDER);
2074
2075	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2076	if (unlikely(!area))
2077		return NULL;
2078
2079	if (!(flags & VM_NO_GUARD))
2080		size += PAGE_SIZE;
2081
2082	va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2083	if (IS_ERR(va)) {
2084		kfree(area);
2085		return NULL;
2086	}
2087
2088	kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2089
2090	setup_vmalloc_vm(area, va, flags, caller);
 
 
 
 
 
 
 
 
 
 
2091
2092	return area;
2093}
2094
2095struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2096				       unsigned long start, unsigned long end,
2097				       const void *caller)
2098{
2099	return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2100				  GFP_KERNEL, caller);
2101}
2102
2103/**
2104 * get_vm_area - reserve a contiguous kernel virtual area
2105 * @size:	 size of the area
2106 * @flags:	 %VM_IOREMAP for I/O mappings or VM_ALLOC
2107 *
2108 * Search an area of @size in the kernel virtual mapping area,
2109 * and reserved it for out purposes.  Returns the area descriptor
2110 * on success or %NULL on failure.
2111 *
2112 * Return: the area descriptor on success or %NULL on failure.
2113 */
2114struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2115{
2116	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
 
2117				  NUMA_NO_NODE, GFP_KERNEL,
2118				  __builtin_return_address(0));
2119}
2120
2121struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2122				const void *caller)
2123{
2124	return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
 
2125				  NUMA_NO_NODE, GFP_KERNEL, caller);
2126}
2127
2128/**
2129 * find_vm_area - find a continuous kernel virtual area
2130 * @addr:	  base address
2131 *
2132 * Search for the kernel VM area starting at @addr, and return it.
2133 * It is up to the caller to do all required locking to keep the returned
2134 * pointer valid.
2135 *
2136 * Return: pointer to the found area or %NULL on faulure
2137 */
2138struct vm_struct *find_vm_area(const void *addr)
2139{
2140	struct vmap_area *va;
2141
2142	va = find_vmap_area((unsigned long)addr);
2143	if (!va)
2144		return NULL;
2145
2146	return va->vm;
2147}
2148
2149/**
2150 * remove_vm_area - find and remove a continuous kernel virtual area
2151 * @addr:	    base address
2152 *
2153 * Search for the kernel VM area starting at @addr, and remove it.
2154 * This function returns the found VM area, but using it is NOT safe
2155 * on SMP machines, except for its size or flags.
2156 *
2157 * Return: pointer to the found area or %NULL on faulure
2158 */
2159struct vm_struct *remove_vm_area(const void *addr)
2160{
2161	struct vmap_area *va;
2162
2163	might_sleep();
2164
2165	spin_lock(&vmap_area_lock);
2166	va = __find_vmap_area((unsigned long)addr);
2167	if (va && va->vm) {
2168		struct vm_struct *vm = va->vm;
2169
2170		va->vm = NULL;
2171		spin_unlock(&vmap_area_lock);
2172
2173		kasan_free_shadow(vm);
2174		free_unmap_vmap_area(va);
2175
2176		return vm;
2177	}
2178
2179	spin_unlock(&vmap_area_lock);
2180	return NULL;
2181}
2182
2183static inline void set_area_direct_map(const struct vm_struct *area,
2184				       int (*set_direct_map)(struct page *page))
2185{
2186	int i;
2187
 
2188	for (i = 0; i < area->nr_pages; i++)
2189		if (page_address(area->pages[i]))
2190			set_direct_map(area->pages[i]);
2191}
2192
2193/* Handle removing and resetting vm mappings related to the vm_struct. */
2194static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2195{
2196	unsigned long start = ULONG_MAX, end = 0;
 
2197	int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2198	int flush_dmap = 0;
2199	int i;
2200
2201	remove_vm_area(area->addr);
2202
2203	/* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2204	if (!flush_reset)
2205		return;
2206
2207	/*
2208	 * If not deallocating pages, just do the flush of the VM area and
2209	 * return.
2210	 */
2211	if (!deallocate_pages) {
2212		vm_unmap_aliases();
2213		return;
2214	}
2215
2216	/*
2217	 * If execution gets here, flush the vm mapping and reset the direct
2218	 * map. Find the start and end range of the direct mappings to make sure
2219	 * the vm_unmap_aliases() flush includes the direct map.
2220	 */
2221	for (i = 0; i < area->nr_pages; i++) {
2222		unsigned long addr = (unsigned long)page_address(area->pages[i]);
2223		if (addr) {
 
 
 
2224			start = min(addr, start);
2225			end = max(addr + PAGE_SIZE, end);
2226			flush_dmap = 1;
2227		}
2228	}
2229
2230	/*
2231	 * Set direct map to something invalid so that it won't be cached if
2232	 * there are any accesses after the TLB flush, then flush the TLB and
2233	 * reset the direct map permissions to the default.
2234	 */
2235	set_area_direct_map(area, set_direct_map_invalid_noflush);
2236	_vm_unmap_aliases(start, end, flush_dmap);
2237	set_area_direct_map(area, set_direct_map_default_noflush);
2238}
2239
2240static void __vunmap(const void *addr, int deallocate_pages)
2241{
2242	struct vm_struct *area;
2243
2244	if (!addr)
2245		return;
2246
2247	if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2248			addr))
2249		return;
2250
2251	area = find_vm_area(addr);
2252	if (unlikely(!area)) {
2253		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2254				addr);
2255		return;
2256	}
2257
2258	debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2259	debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2260
2261	kasan_poison_vmalloc(area->addr, area->size);
2262
2263	vm_remove_mappings(area, deallocate_pages);
2264
2265	if (deallocate_pages) {
2266		int i;
2267
2268		for (i = 0; i < area->nr_pages; i++) {
2269			struct page *page = area->pages[i];
2270
2271			BUG_ON(!page);
 
 
 
 
 
2272			__free_pages(page, 0);
 
2273		}
2274		atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2275
2276		kvfree(area->pages);
2277	}
2278
2279	kfree(area);
2280	return;
2281}
2282
2283static inline void __vfree_deferred(const void *addr)
2284{
2285	/*
2286	 * Use raw_cpu_ptr() because this can be called from preemptible
2287	 * context. Preemption is absolutely fine here, because the llist_add()
2288	 * implementation is lockless, so it works even if we are adding to
2289	 * another cpu's list. schedule_work() should be fine with this too.
2290	 */
2291	struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2292
2293	if (llist_add((struct llist_node *)addr, &p->list))
2294		schedule_work(&p->wq);
2295}
2296
2297/**
2298 * vfree_atomic - release memory allocated by vmalloc()
2299 * @addr:	  memory base address
2300 *
2301 * This one is just like vfree() but can be called in any atomic context
2302 * except NMIs.
2303 */
2304void vfree_atomic(const void *addr)
2305{
2306	BUG_ON(in_nmi());
2307
2308	kmemleak_free(addr);
2309
2310	if (!addr)
2311		return;
2312	__vfree_deferred(addr);
2313}
2314
2315static void __vfree(const void *addr)
2316{
2317	if (unlikely(in_interrupt()))
2318		__vfree_deferred(addr);
2319	else
2320		__vunmap(addr, 1);
2321}
2322
2323/**
2324 * vfree - release memory allocated by vmalloc()
2325 * @addr:  memory base address
 
 
 
 
 
2326 *
2327 * Free the virtually continuous memory area starting at @addr, as
2328 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2329 * NULL, no operation is performed.
2330 *
2331 * Must not be called in NMI context (strictly speaking, only if we don't
2332 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2333 * conventions for vfree() arch-depenedent would be a really bad idea)
2334 *
 
2335 * May sleep if called *not* from interrupt context.
2336 *
2337 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
 
2338 */
2339void vfree(const void *addr)
2340{
2341	BUG_ON(in_nmi());
2342
2343	kmemleak_free(addr);
2344
2345	might_sleep_if(!in_interrupt());
2346
2347	if (!addr)
2348		return;
2349
2350	__vfree(addr);
2351}
2352EXPORT_SYMBOL(vfree);
2353
2354/**
2355 * vunmap - release virtual mapping obtained by vmap()
2356 * @addr:   memory base address
2357 *
2358 * Free the virtually contiguous memory area starting at @addr,
2359 * which was created from the page array passed to vmap().
2360 *
2361 * Must not be called in interrupt context.
2362 */
2363void vunmap(const void *addr)
2364{
2365	BUG_ON(in_interrupt());
2366	might_sleep();
2367	if (addr)
2368		__vunmap(addr, 0);
2369}
2370EXPORT_SYMBOL(vunmap);
2371
2372/**
2373 * vmap - map an array of pages into virtually contiguous space
2374 * @pages: array of page pointers
2375 * @count: number of pages to map
2376 * @flags: vm_area->flags
2377 * @prot: page protection for the mapping
2378 *
2379 * Maps @count pages from @pages into contiguous kernel virtual
2380 * space.
 
 
 
2381 *
2382 * Return: the address of the area or %NULL on failure
2383 */
2384void *vmap(struct page **pages, unsigned int count,
2385	   unsigned long flags, pgprot_t prot)
2386{
2387	struct vm_struct *area;
 
2388	unsigned long size;		/* In bytes */
2389
2390	might_sleep();
2391
 
 
 
 
 
 
 
2392	if (count > totalram_pages())
2393		return NULL;
2394
2395	size = (unsigned long)count << PAGE_SHIFT;
2396	area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2397	if (!area)
2398		return NULL;
2399
2400	if (map_kernel_range((unsigned long)area->addr, size, pgprot_nx(prot),
2401			pages) < 0) {
 
2402		vunmap(area->addr);
2403		return NULL;
2404	}
2405
 
 
 
 
2406	return area->addr;
2407}
2408EXPORT_SYMBOL(vmap);
2409
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2410static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2411				 pgprot_t prot, int node)
 
2412{
2413	struct page **pages;
2414	unsigned int nr_pages, array_size, i;
2415	const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2416	const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2417	const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2418					0 :
2419					__GFP_HIGHMEM;
 
 
 
 
2420
2421	nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2422	array_size = (nr_pages * sizeof(struct page *));
 
 
2423
2424	/* Please note that the recursion is strictly bounded. */
2425	if (array_size > PAGE_SIZE) {
2426		pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2427				node, area->caller);
2428	} else {
2429		pages = kmalloc_node(array_size, nested_gfp, node);
2430	}
2431
2432	if (!pages) {
2433		remove_vm_area(area->addr);
2434		kfree(area);
 
 
2435		return NULL;
2436	}
2437
2438	area->pages = pages;
2439	area->nr_pages = nr_pages;
 
 
 
2440
2441	for (i = 0; i < area->nr_pages; i++) {
2442		struct page *page;
 
2443
2444		if (node == NUMA_NO_NODE)
2445			page = alloc_page(alloc_mask|highmem_mask);
2446		else
2447			page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2448
2449		if (unlikely(!page)) {
2450			/* Successfully allocated i pages, free them in __vunmap() */
2451			area->nr_pages = i;
2452			atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2453			goto fail;
2454		}
2455		area->pages[i] = page;
2456		if (gfpflags_allow_blocking(gfp_mask))
2457			cond_resched();
2458	}
2459	atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2460
2461	if (map_kernel_range((unsigned long)area->addr, get_vm_area_size(area),
2462			prot, pages) < 0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2463		goto fail;
 
2464
2465	return area->addr;
2466
2467fail:
2468	warn_alloc(gfp_mask, NULL,
2469			  "vmalloc: allocation failure, allocated %ld of %ld bytes",
2470			  (area->nr_pages*PAGE_SIZE), area->size);
2471	__vfree(area->addr);
2472	return NULL;
2473}
2474
2475/**
2476 * __vmalloc_node_range - allocate virtually contiguous memory
2477 * @size:		  allocation size
2478 * @align:		  desired alignment
2479 * @start:		  vm area range start
2480 * @end:		  vm area range end
2481 * @gfp_mask:		  flags for the page level allocator
2482 * @prot:		  protection mask for the allocated pages
2483 * @vm_flags:		  additional vm area flags (e.g. %VM_NO_GUARD)
2484 * @node:		  node to use for allocation or NUMA_NO_NODE
2485 * @caller:		  caller's return address
2486 *
2487 * Allocate enough pages to cover @size from the page level
2488 * allocator with @gfp_mask flags.  Map them into contiguous
2489 * kernel virtual space, using a pagetable protection of @prot.
 
 
 
 
 
 
 
 
 
 
2490 *
2491 * Return: the address of the area or %NULL on failure
2492 */
2493void *__vmalloc_node_range(unsigned long size, unsigned long align,
2494			unsigned long start, unsigned long end, gfp_t gfp_mask,
2495			pgprot_t prot, unsigned long vm_flags, int node,
2496			const void *caller)
2497{
2498	struct vm_struct *area;
2499	void *addr;
 
2500	unsigned long real_size = size;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2501
2502	size = PAGE_ALIGN(size);
2503	if (!size || (size >> PAGE_SHIFT) > totalram_pages())
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2504		goto fail;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2505
2506	area = __get_vm_area_node(real_size, align, VM_ALLOC | VM_UNINITIALIZED |
2507				vm_flags, start, end, node, gfp_mask, caller);
2508	if (!area)
2509		goto fail;
2510
2511	addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2512	if (!addr)
2513		return NULL;
 
 
 
 
 
 
 
 
 
 
 
2514
2515	/*
2516	 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2517	 * flag. It means that vm_struct is not fully initialized.
2518	 * Now, it is fully initialized, so remove this flag here.
2519	 */
2520	clear_vm_uninitialized_flag(area);
2521
2522	kmemleak_vmalloc(area, size, gfp_mask);
 
 
2523
2524	return addr;
2525
2526fail:
2527	warn_alloc(gfp_mask, NULL,
2528			  "vmalloc: allocation failure: %lu bytes", real_size);
 
 
 
 
 
2529	return NULL;
2530}
2531
2532/**
2533 * __vmalloc_node - allocate virtually contiguous memory
2534 * @size:	    allocation size
2535 * @align:	    desired alignment
2536 * @gfp_mask:	    flags for the page level allocator
2537 * @node:	    node to use for allocation or NUMA_NO_NODE
2538 * @caller:	    caller's return address
2539 *
2540 * Allocate enough pages to cover @size from the page level allocator with
2541 * @gfp_mask flags.  Map them into contiguous kernel virtual space.
2542 *
2543 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2544 * and __GFP_NOFAIL are not supported
2545 *
2546 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2547 * with mm people.
2548 *
2549 * Return: pointer to the allocated memory or %NULL on error
2550 */
2551void *__vmalloc_node(unsigned long size, unsigned long align,
2552			    gfp_t gfp_mask, int node, const void *caller)
2553{
2554	return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2555				gfp_mask, PAGE_KERNEL, 0, node, caller);
2556}
2557/*
2558 * This is only for performance analysis of vmalloc and stress purpose.
2559 * It is required by vmalloc test module, therefore do not use it other
2560 * than that.
2561 */
2562#ifdef CONFIG_TEST_VMALLOC_MODULE
2563EXPORT_SYMBOL_GPL(__vmalloc_node);
2564#endif
2565
2566void *__vmalloc(unsigned long size, gfp_t gfp_mask)
2567{
2568	return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
2569				__builtin_return_address(0));
2570}
2571EXPORT_SYMBOL(__vmalloc);
2572
2573/**
2574 * vmalloc - allocate virtually contiguous memory
2575 * @size:    allocation size
2576 *
2577 * Allocate enough pages to cover @size from the page level
2578 * allocator and map them into contiguous kernel virtual space.
2579 *
2580 * For tight control over page level allocator and protection flags
2581 * use __vmalloc() instead.
2582 *
2583 * Return: pointer to the allocated memory or %NULL on error
2584 */
2585void *vmalloc(unsigned long size)
2586{
2587	return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
2588				__builtin_return_address(0));
2589}
2590EXPORT_SYMBOL(vmalloc);
2591
2592/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2593 * vzalloc - allocate virtually contiguous memory with zero fill
2594 * @size:    allocation size
2595 *
2596 * Allocate enough pages to cover @size from the page level
2597 * allocator and map them into contiguous kernel virtual space.
2598 * The memory allocated is set to zero.
2599 *
2600 * For tight control over page level allocator and protection flags
2601 * use __vmalloc() instead.
2602 *
2603 * Return: pointer to the allocated memory or %NULL on error
2604 */
2605void *vzalloc(unsigned long size)
2606{
2607	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
2608				__builtin_return_address(0));
2609}
2610EXPORT_SYMBOL(vzalloc);
2611
2612/**
2613 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2614 * @size: allocation size
2615 *
2616 * The resulting memory area is zeroed so it can be mapped to userspace
2617 * without leaking data.
2618 *
2619 * Return: pointer to the allocated memory or %NULL on error
2620 */
2621void *vmalloc_user(unsigned long size)
2622{
2623	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
2624				    GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2625				    VM_USERMAP, NUMA_NO_NODE,
2626				    __builtin_return_address(0));
2627}
2628EXPORT_SYMBOL(vmalloc_user);
2629
2630/**
2631 * vmalloc_node - allocate memory on a specific node
2632 * @size:	  allocation size
2633 * @node:	  numa node
2634 *
2635 * Allocate enough pages to cover @size from the page level
2636 * allocator and map them into contiguous kernel virtual space.
2637 *
2638 * For tight control over page level allocator and protection flags
2639 * use __vmalloc() instead.
2640 *
2641 * Return: pointer to the allocated memory or %NULL on error
2642 */
2643void *vmalloc_node(unsigned long size, int node)
2644{
2645	return __vmalloc_node(size, 1, GFP_KERNEL, node,
2646			__builtin_return_address(0));
2647}
2648EXPORT_SYMBOL(vmalloc_node);
2649
2650/**
2651 * vzalloc_node - allocate memory on a specific node with zero fill
2652 * @size:	allocation size
2653 * @node:	numa node
2654 *
2655 * Allocate enough pages to cover @size from the page level
2656 * allocator and map them into contiguous kernel virtual space.
2657 * The memory allocated is set to zero.
2658 *
2659 * Return: pointer to the allocated memory or %NULL on error
2660 */
2661void *vzalloc_node(unsigned long size, int node)
2662{
2663	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
2664				__builtin_return_address(0));
2665}
2666EXPORT_SYMBOL(vzalloc_node);
2667
2668#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2669#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2670#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2671#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2672#else
2673/*
2674 * 64b systems should always have either DMA or DMA32 zones. For others
2675 * GFP_DMA32 should do the right thing and use the normal zone.
2676 */
2677#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2678#endif
2679
2680/**
2681 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2682 * @size:	allocation size
2683 *
2684 * Allocate enough 32bit PA addressable pages to cover @size from the
2685 * page level allocator and map them into contiguous kernel virtual space.
2686 *
2687 * Return: pointer to the allocated memory or %NULL on error
2688 */
2689void *vmalloc_32(unsigned long size)
2690{
2691	return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
2692			__builtin_return_address(0));
2693}
2694EXPORT_SYMBOL(vmalloc_32);
2695
2696/**
2697 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2698 * @size:	     allocation size
2699 *
2700 * The resulting memory area is 32bit addressable and zeroed so it can be
2701 * mapped to userspace without leaking data.
2702 *
2703 * Return: pointer to the allocated memory or %NULL on error
2704 */
2705void *vmalloc_32_user(unsigned long size)
2706{
2707	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
2708				    GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2709				    VM_USERMAP, NUMA_NO_NODE,
2710				    __builtin_return_address(0));
2711}
2712EXPORT_SYMBOL(vmalloc_32_user);
2713
2714/*
2715 * small helper routine , copy contents to buf from addr.
2716 * If the page is not present, fill zero.
2717 */
2718
2719static int aligned_vread(char *buf, char *addr, unsigned long count)
2720{
2721	struct page *p;
2722	int copied = 0;
2723
2724	while (count) {
2725		unsigned long offset, length;
2726
2727		offset = offset_in_page(addr);
2728		length = PAGE_SIZE - offset;
2729		if (length > count)
2730			length = count;
2731		p = vmalloc_to_page(addr);
2732		/*
2733		 * To do safe access to this _mapped_ area, we need
2734		 * lock. But adding lock here means that we need to add
2735		 * overhead of vmalloc()/vfree() calles for this _debug_
2736		 * interface, rarely used. Instead of that, we'll use
2737		 * kmap() and get small overhead in this access function.
2738		 */
2739		if (p) {
2740			/*
2741			 * we can expect USER0 is not used (see vread/vwrite's
2742			 * function description)
2743			 */
2744			void *map = kmap_atomic(p);
2745			memcpy(buf, map + offset, length);
2746			kunmap_atomic(map);
2747		} else
2748			memset(buf, 0, length);
2749
2750		addr += length;
2751		buf += length;
2752		copied += length;
2753		count -= length;
2754	}
2755	return copied;
2756}
2757
2758static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2759{
2760	struct page *p;
2761	int copied = 0;
2762
2763	while (count) {
2764		unsigned long offset, length;
2765
2766		offset = offset_in_page(addr);
2767		length = PAGE_SIZE - offset;
2768		if (length > count)
2769			length = count;
2770		p = vmalloc_to_page(addr);
2771		/*
2772		 * To do safe access to this _mapped_ area, we need
2773		 * lock. But adding lock here means that we need to add
2774		 * overhead of vmalloc()/vfree() calles for this _debug_
2775		 * interface, rarely used. Instead of that, we'll use
2776		 * kmap() and get small overhead in this access function.
2777		 */
2778		if (p) {
2779			/*
2780			 * we can expect USER0 is not used (see vread/vwrite's
2781			 * function description)
2782			 */
2783			void *map = kmap_atomic(p);
2784			memcpy(map + offset, buf, length);
2785			kunmap_atomic(map);
2786		}
2787		addr += length;
2788		buf += length;
2789		copied += length;
2790		count -= length;
2791	}
2792	return copied;
2793}
2794
2795/**
2796 * vread() - read vmalloc area in a safe way.
2797 * @buf:     buffer for reading data
2798 * @addr:    vm address.
2799 * @count:   number of bytes to be read.
2800 *
2801 * This function checks that addr is a valid vmalloc'ed area, and
2802 * copy data from that area to a given buffer. If the given memory range
2803 * of [addr...addr+count) includes some valid address, data is copied to
2804 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2805 * IOREMAP area is treated as memory hole and no copy is done.
2806 *
2807 * If [addr...addr+count) doesn't includes any intersects with alive
2808 * vm_struct area, returns 0. @buf should be kernel's buffer.
2809 *
2810 * Note: In usual ops, vread() is never necessary because the caller
2811 * should know vmalloc() area is valid and can use memcpy().
2812 * This is for routines which have to access vmalloc area without
2813 * any information, as /dev/kmem.
2814 *
2815 * Return: number of bytes for which addr and buf should be increased
2816 * (same number as @count) or %0 if [addr...addr+count) doesn't
2817 * include any intersection with valid vmalloc area
2818 */
2819long vread(char *buf, char *addr, unsigned long count)
2820{
2821	struct vmap_area *va;
2822	struct vm_struct *vm;
2823	char *vaddr, *buf_start = buf;
2824	unsigned long buflen = count;
2825	unsigned long n;
2826
 
 
2827	/* Don't allow overflow */
2828	if ((unsigned long) addr + count < count)
2829		count = -(unsigned long) addr;
2830
2831	spin_lock(&vmap_area_lock);
2832	list_for_each_entry(va, &vmap_area_list, list) {
 
 
 
 
 
 
 
 
2833		if (!count)
2834			break;
2835
2836		if (!va->vm)
2837			continue;
2838
2839		vm = va->vm;
2840		vaddr = (char *) vm->addr;
2841		if (addr >= vaddr + get_vm_area_size(vm))
2842			continue;
2843		while (addr < vaddr) {
2844			if (count == 0)
2845				goto finished;
2846			*buf = '\0';
2847			buf++;
2848			addr++;
2849			count--;
2850		}
2851		n = vaddr + get_vm_area_size(vm) - addr;
2852		if (n > count)
2853			n = count;
2854		if (!(vm->flags & VM_IOREMAP))
2855			aligned_vread(buf, addr, n);
2856		else /* IOREMAP area is treated as memory hole */
2857			memset(buf, 0, n);
2858		buf += n;
2859		addr += n;
2860		count -= n;
2861	}
2862finished:
2863	spin_unlock(&vmap_area_lock);
2864
2865	if (buf == buf_start)
2866		return 0;
2867	/* zero-fill memory holes */
2868	if (buf != buf_start + buflen)
2869		memset(buf, 0, buflen - (buf - buf_start));
2870
2871	return buflen;
2872}
2873
2874/**
2875 * vwrite() - write vmalloc area in a safe way.
2876 * @buf:      buffer for source data
2877 * @addr:     vm address.
2878 * @count:    number of bytes to be read.
2879 *
2880 * This function checks that addr is a valid vmalloc'ed area, and
2881 * copy data from a buffer to the given addr. If specified range of
2882 * [addr...addr+count) includes some valid address, data is copied from
2883 * proper area of @buf. If there are memory holes, no copy to hole.
2884 * IOREMAP area is treated as memory hole and no copy is done.
2885 *
2886 * If [addr...addr+count) doesn't includes any intersects with alive
2887 * vm_struct area, returns 0. @buf should be kernel's buffer.
2888 *
2889 * Note: In usual ops, vwrite() is never necessary because the caller
2890 * should know vmalloc() area is valid and can use memcpy().
2891 * This is for routines which have to access vmalloc area without
2892 * any information, as /dev/kmem.
2893 *
2894 * Return: number of bytes for which addr and buf should be
2895 * increased (same number as @count) or %0 if [addr...addr+count)
2896 * doesn't include any intersection with valid vmalloc area
2897 */
2898long vwrite(char *buf, char *addr, unsigned long count)
2899{
2900	struct vmap_area *va;
2901	struct vm_struct *vm;
2902	char *vaddr;
2903	unsigned long n, buflen;
2904	int copied = 0;
2905
2906	/* Don't allow overflow */
2907	if ((unsigned long) addr + count < count)
2908		count = -(unsigned long) addr;
2909	buflen = count;
2910
2911	spin_lock(&vmap_area_lock);
2912	list_for_each_entry(va, &vmap_area_list, list) {
2913		if (!count)
2914			break;
2915
2916		if (!va->vm)
2917			continue;
2918
2919		vm = va->vm;
2920		vaddr = (char *) vm->addr;
2921		if (addr >= vaddr + get_vm_area_size(vm))
2922			continue;
2923		while (addr < vaddr) {
2924			if (count == 0)
2925				goto finished;
2926			buf++;
2927			addr++;
2928			count--;
2929		}
2930		n = vaddr + get_vm_area_size(vm) - addr;
2931		if (n > count)
2932			n = count;
2933		if (!(vm->flags & VM_IOREMAP)) {
2934			aligned_vwrite(buf, addr, n);
2935			copied++;
2936		}
2937		buf += n;
2938		addr += n;
2939		count -= n;
2940	}
2941finished:
2942	spin_unlock(&vmap_area_lock);
2943	if (!copied)
2944		return 0;
2945	return buflen;
2946}
2947
2948/**
2949 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2950 * @vma:		vma to cover
2951 * @uaddr:		target user address to start at
2952 * @kaddr:		virtual address of vmalloc kernel memory
2953 * @pgoff:		offset from @kaddr to start at
2954 * @size:		size of map area
2955 *
2956 * Returns:	0 for success, -Exxx on failure
2957 *
2958 * This function checks that @kaddr is a valid vmalloc'ed area,
2959 * and that it is big enough to cover the range starting at
2960 * @uaddr in @vma. Will return failure if that criteria isn't
2961 * met.
2962 *
2963 * Similar to remap_pfn_range() (see mm/memory.c)
2964 */
2965int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2966				void *kaddr, unsigned long pgoff,
2967				unsigned long size)
2968{
2969	struct vm_struct *area;
2970	unsigned long off;
2971	unsigned long end_index;
2972
2973	if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
2974		return -EINVAL;
2975
2976	size = PAGE_ALIGN(size);
2977
2978	if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2979		return -EINVAL;
2980
2981	area = find_vm_area(kaddr);
2982	if (!area)
2983		return -EINVAL;
2984
2985	if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
2986		return -EINVAL;
2987
2988	if (check_add_overflow(size, off, &end_index) ||
2989	    end_index > get_vm_area_size(area))
2990		return -EINVAL;
2991	kaddr += off;
2992
2993	do {
2994		struct page *page = vmalloc_to_page(kaddr);
2995		int ret;
2996
2997		ret = vm_insert_page(vma, uaddr, page);
2998		if (ret)
2999			return ret;
3000
3001		uaddr += PAGE_SIZE;
3002		kaddr += PAGE_SIZE;
3003		size -= PAGE_SIZE;
3004	} while (size > 0);
3005
3006	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3007
3008	return 0;
3009}
3010EXPORT_SYMBOL(remap_vmalloc_range_partial);
3011
3012/**
3013 * remap_vmalloc_range - map vmalloc pages to userspace
3014 * @vma:		vma to cover (map full range of vma)
3015 * @addr:		vmalloc memory
3016 * @pgoff:		number of pages into addr before first page to map
3017 *
3018 * Returns:	0 for success, -Exxx on failure
3019 *
3020 * This function checks that addr is a valid vmalloc'ed area, and
3021 * that it is big enough to cover the vma. Will return failure if
3022 * that criteria isn't met.
3023 *
3024 * Similar to remap_pfn_range() (see mm/memory.c)
3025 */
3026int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3027						unsigned long pgoff)
3028{
3029	return remap_vmalloc_range_partial(vma, vma->vm_start,
3030					   addr, pgoff,
3031					   vma->vm_end - vma->vm_start);
3032}
3033EXPORT_SYMBOL(remap_vmalloc_range);
3034
3035static int f(pte_t *pte, unsigned long addr, void *data)
3036{
3037	pte_t ***p = data;
3038
3039	if (p) {
3040		*(*p) = pte;
3041		(*p)++;
3042	}
3043	return 0;
3044}
3045
3046/**
3047 * alloc_vm_area - allocate a range of kernel address space
3048 * @size:	   size of the area
3049 * @ptes:	   returns the PTEs for the address space
3050 *
3051 * Returns:	NULL on failure, vm_struct on success
3052 *
3053 * This function reserves a range of kernel address space, and
3054 * allocates pagetables to map that range.  No actual mappings
3055 * are created.
3056 *
3057 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3058 * allocated for the VM area are returned.
3059 */
3060struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3061{
3062	struct vm_struct *area;
3063
3064	area = get_vm_area_caller(size, VM_IOREMAP,
3065				__builtin_return_address(0));
3066	if (area == NULL)
3067		return NULL;
3068
3069	/*
3070	 * This ensures that page tables are constructed for this region
3071	 * of kernel virtual address space and mapped into init_mm.
3072	 */
3073	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3074				size, f, ptes ? &ptes : NULL)) {
3075		free_vm_area(area);
3076		return NULL;
3077	}
3078
3079	return area;
3080}
3081EXPORT_SYMBOL_GPL(alloc_vm_area);
3082
3083void free_vm_area(struct vm_struct *area)
3084{
3085	struct vm_struct *ret;
3086	ret = remove_vm_area(area->addr);
3087	BUG_ON(ret != area);
3088	kfree(area);
3089}
3090EXPORT_SYMBOL_GPL(free_vm_area);
3091
3092#ifdef CONFIG_SMP
3093static struct vmap_area *node_to_va(struct rb_node *n)
3094{
3095	return rb_entry_safe(n, struct vmap_area, rb_node);
3096}
3097
3098/**
3099 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3100 * @addr: target address
3101 *
3102 * Returns: vmap_area if it is found. If there is no such area
3103 *   the first highest(reverse order) vmap_area is returned
3104 *   i.e. va->va_start < addr && va->va_end < addr or NULL
3105 *   if there are no any areas before @addr.
3106 */
3107static struct vmap_area *
3108pvm_find_va_enclose_addr(unsigned long addr)
3109{
3110	struct vmap_area *va, *tmp;
3111	struct rb_node *n;
3112
3113	n = free_vmap_area_root.rb_node;
3114	va = NULL;
3115
3116	while (n) {
3117		tmp = rb_entry(n, struct vmap_area, rb_node);
3118		if (tmp->va_start <= addr) {
3119			va = tmp;
3120			if (tmp->va_end >= addr)
3121				break;
3122
3123			n = n->rb_right;
3124		} else {
3125			n = n->rb_left;
3126		}
3127	}
3128
3129	return va;
3130}
3131
3132/**
3133 * pvm_determine_end_from_reverse - find the highest aligned address
3134 * of free block below VMALLOC_END
3135 * @va:
3136 *   in - the VA we start the search(reverse order);
3137 *   out - the VA with the highest aligned end address.
 
3138 *
3139 * Returns: determined end address within vmap_area
3140 */
3141static unsigned long
3142pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3143{
3144	unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3145	unsigned long addr;
3146
3147	if (likely(*va)) {
3148		list_for_each_entry_from_reverse((*va),
3149				&free_vmap_area_list, list) {
3150			addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3151			if ((*va)->va_start < addr)
3152				return addr;
3153		}
3154	}
3155
3156	return 0;
3157}
3158
3159/**
3160 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3161 * @offsets: array containing offset of each area
3162 * @sizes: array containing size of each area
3163 * @nr_vms: the number of areas to allocate
3164 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3165 *
3166 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3167 *	    vm_structs on success, %NULL on failure
3168 *
3169 * Percpu allocator wants to use congruent vm areas so that it can
3170 * maintain the offsets among percpu areas.  This function allocates
3171 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
3172 * be scattered pretty far, distance between two areas easily going up
3173 * to gigabytes.  To avoid interacting with regular vmallocs, these
3174 * areas are allocated from top.
3175 *
3176 * Despite its complicated look, this allocator is rather simple. It
3177 * does everything top-down and scans free blocks from the end looking
3178 * for matching base. While scanning, if any of the areas do not fit the
3179 * base address is pulled down to fit the area. Scanning is repeated till
3180 * all the areas fit and then all necessary data structures are inserted
3181 * and the result is returned.
3182 */
3183struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3184				     const size_t *sizes, int nr_vms,
3185				     size_t align)
3186{
3187	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3188	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3189	struct vmap_area **vas, *va;
3190	struct vm_struct **vms;
3191	int area, area2, last_area, term_area;
3192	unsigned long base, start, size, end, last_end, orig_start, orig_end;
3193	bool purged = false;
3194	enum fit_type type;
3195
3196	/* verify parameters and allocate data structures */
3197	BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3198	for (last_area = 0, area = 0; area < nr_vms; area++) {
3199		start = offsets[area];
3200		end = start + sizes[area];
3201
3202		/* is everything aligned properly? */
3203		BUG_ON(!IS_ALIGNED(offsets[area], align));
3204		BUG_ON(!IS_ALIGNED(sizes[area], align));
3205
3206		/* detect the area with the highest address */
3207		if (start > offsets[last_area])
3208			last_area = area;
3209
3210		for (area2 = area + 1; area2 < nr_vms; area2++) {
3211			unsigned long start2 = offsets[area2];
3212			unsigned long end2 = start2 + sizes[area2];
3213
3214			BUG_ON(start2 < end && start < end2);
3215		}
3216	}
3217	last_end = offsets[last_area] + sizes[last_area];
3218
3219	if (vmalloc_end - vmalloc_start < last_end) {
3220		WARN_ON(true);
3221		return NULL;
3222	}
3223
3224	vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3225	vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3226	if (!vas || !vms)
3227		goto err_free2;
3228
3229	for (area = 0; area < nr_vms; area++) {
3230		vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3231		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3232		if (!vas[area] || !vms[area])
3233			goto err_free;
3234	}
3235retry:
3236	spin_lock(&free_vmap_area_lock);
3237
3238	/* start scanning - we scan from the top, begin with the last area */
3239	area = term_area = last_area;
3240	start = offsets[area];
3241	end = start + sizes[area];
3242
3243	va = pvm_find_va_enclose_addr(vmalloc_end);
3244	base = pvm_determine_end_from_reverse(&va, align) - end;
3245
3246	while (true) {
3247		/*
3248		 * base might have underflowed, add last_end before
3249		 * comparing.
3250		 */
3251		if (base + last_end < vmalloc_start + last_end)
3252			goto overflow;
3253
3254		/*
3255		 * Fitting base has not been found.
3256		 */
3257		if (va == NULL)
3258			goto overflow;
3259
3260		/*
3261		 * If required width exceeds current VA block, move
3262		 * base downwards and then recheck.
3263		 */
3264		if (base + end > va->va_end) {
3265			base = pvm_determine_end_from_reverse(&va, align) - end;
3266			term_area = area;
3267			continue;
3268		}
3269
3270		/*
3271		 * If this VA does not fit, move base downwards and recheck.
3272		 */
3273		if (base + start < va->va_start) {
3274			va = node_to_va(rb_prev(&va->rb_node));
3275			base = pvm_determine_end_from_reverse(&va, align) - end;
3276			term_area = area;
3277			continue;
3278		}
3279
3280		/*
3281		 * This area fits, move on to the previous one.  If
3282		 * the previous one is the terminal one, we're done.
3283		 */
3284		area = (area + nr_vms - 1) % nr_vms;
3285		if (area == term_area)
3286			break;
3287
3288		start = offsets[area];
3289		end = start + sizes[area];
3290		va = pvm_find_va_enclose_addr(base + end);
3291	}
3292
3293	/* we've found a fitting base, insert all va's */
3294	for (area = 0; area < nr_vms; area++) {
3295		int ret;
3296
3297		start = base + offsets[area];
3298		size = sizes[area];
3299
3300		va = pvm_find_va_enclose_addr(start);
3301		if (WARN_ON_ONCE(va == NULL))
3302			/* It is a BUG(), but trigger recovery instead. */
3303			goto recovery;
3304
3305		type = classify_va_fit_type(va, start, size);
3306		if (WARN_ON_ONCE(type == NOTHING_FIT))
 
 
3307			/* It is a BUG(), but trigger recovery instead. */
3308			goto recovery;
3309
3310		ret = adjust_va_to_fit_type(va, start, size, type);
3311		if (unlikely(ret))
3312			goto recovery;
3313
3314		/* Allocated area. */
3315		va = vas[area];
3316		va->va_start = start;
3317		va->va_end = start + size;
3318	}
3319
3320	spin_unlock(&free_vmap_area_lock);
3321
3322	/* populate the kasan shadow space */
3323	for (area = 0; area < nr_vms; area++) {
3324		if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3325			goto err_free_shadow;
3326
3327		kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3328				       sizes[area]);
3329	}
3330
3331	/* insert all vm's */
3332	spin_lock(&vmap_area_lock);
3333	for (area = 0; area < nr_vms; area++) {
3334		insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3335
3336		setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3337				 pcpu_get_vm_areas);
3338	}
3339	spin_unlock(&vmap_area_lock);
3340
 
 
 
 
 
 
 
 
 
 
3341	kfree(vas);
3342	return vms;
3343
3344recovery:
3345	/*
3346	 * Remove previously allocated areas. There is no
3347	 * need in removing these areas from the busy tree,
3348	 * because they are inserted only on the final step
3349	 * and when pcpu_get_vm_areas() is success.
3350	 */
3351	while (area--) {
3352		orig_start = vas[area]->va_start;
3353		orig_end = vas[area]->va_end;
3354		va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3355					    &free_vmap_area_list);
3356		if (va)
3357			kasan_release_vmalloc(orig_start, orig_end,
3358				va->va_start, va->va_end);
3359		vas[area] = NULL;
3360	}
3361
3362overflow:
3363	spin_unlock(&free_vmap_area_lock);
3364	if (!purged) {
3365		purge_vmap_area_lazy();
3366		purged = true;
3367
3368		/* Before "retry", check if we recover. */
3369		for (area = 0; area < nr_vms; area++) {
3370			if (vas[area])
3371				continue;
3372
3373			vas[area] = kmem_cache_zalloc(
3374				vmap_area_cachep, GFP_KERNEL);
3375			if (!vas[area])
3376				goto err_free;
3377		}
3378
3379		goto retry;
3380	}
3381
3382err_free:
3383	for (area = 0; area < nr_vms; area++) {
3384		if (vas[area])
3385			kmem_cache_free(vmap_area_cachep, vas[area]);
3386
3387		kfree(vms[area]);
3388	}
3389err_free2:
3390	kfree(vas);
3391	kfree(vms);
3392	return NULL;
3393
3394err_free_shadow:
3395	spin_lock(&free_vmap_area_lock);
3396	/*
3397	 * We release all the vmalloc shadows, even the ones for regions that
3398	 * hadn't been successfully added. This relies on kasan_release_vmalloc
3399	 * being able to tolerate this case.
3400	 */
3401	for (area = 0; area < nr_vms; area++) {
3402		orig_start = vas[area]->va_start;
3403		orig_end = vas[area]->va_end;
3404		va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3405					    &free_vmap_area_list);
3406		if (va)
3407			kasan_release_vmalloc(orig_start, orig_end,
3408				va->va_start, va->va_end);
3409		vas[area] = NULL;
3410		kfree(vms[area]);
3411	}
3412	spin_unlock(&free_vmap_area_lock);
3413	kfree(vas);
3414	kfree(vms);
3415	return NULL;
3416}
3417
3418/**
3419 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3420 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3421 * @nr_vms: the number of allocated areas
3422 *
3423 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3424 */
3425void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3426{
3427	int i;
3428
3429	for (i = 0; i < nr_vms; i++)
3430		free_vm_area(vms[i]);
3431	kfree(vms);
3432}
3433#endif	/* CONFIG_SMP */
3434
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3435#ifdef CONFIG_PROC_FS
3436static void *s_start(struct seq_file *m, loff_t *pos)
3437	__acquires(&vmap_purge_lock)
3438	__acquires(&vmap_area_lock)
3439{
3440	mutex_lock(&vmap_purge_lock);
3441	spin_lock(&vmap_area_lock);
3442
3443	return seq_list_start(&vmap_area_list, *pos);
3444}
3445
3446static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3447{
3448	return seq_list_next(p, &vmap_area_list, pos);
3449}
3450
3451static void s_stop(struct seq_file *m, void *p)
 
3452	__releases(&vmap_purge_lock)
3453	__releases(&vmap_area_lock)
3454{
 
3455	mutex_unlock(&vmap_purge_lock);
3456	spin_unlock(&vmap_area_lock);
3457}
3458
3459static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3460{
3461	if (IS_ENABLED(CONFIG_NUMA)) {
3462		unsigned int nr, *counters = m->private;
 
3463
3464		if (!counters)
3465			return;
3466
3467		if (v->flags & VM_UNINITIALIZED)
3468			return;
3469		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3470		smp_rmb();
3471
3472		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3473
3474		for (nr = 0; nr < v->nr_pages; nr++)
3475			counters[page_to_nid(v->pages[nr])]++;
3476
3477		for_each_node_state(nr, N_HIGH_MEMORY)
3478			if (counters[nr])
3479				seq_printf(m, " N%u=%u", nr, counters[nr]);
3480	}
3481}
3482
3483static void show_purge_info(struct seq_file *m)
3484{
3485	struct llist_node *head;
3486	struct vmap_area *va;
3487
3488	head = READ_ONCE(vmap_purge_list.first);
3489	if (head == NULL)
3490		return;
3491
3492	llist_for_each_entry(va, head, purge_list) {
3493		seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3494			(void *)va->va_start, (void *)va->va_end,
3495			va->va_end - va->va_start);
3496	}
 
3497}
3498
3499static int s_show(struct seq_file *m, void *p)
3500{
3501	struct vmap_area *va;
3502	struct vm_struct *v;
3503
3504	va = list_entry(p, struct vmap_area, list);
3505
3506	/*
3507	 * s_show can encounter race with remove_vm_area, !vm on behalf
3508	 * of vmap area is being tear down or vm_map_ram allocation.
3509	 */
3510	if (!va->vm) {
3511		seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3512			(void *)va->va_start, (void *)va->va_end,
3513			va->va_end - va->va_start);
3514
3515		return 0;
3516	}
3517
3518	v = va->vm;
3519
3520	seq_printf(m, "0x%pK-0x%pK %7ld",
3521		v->addr, v->addr + v->size, v->size);
3522
3523	if (v->caller)
3524		seq_printf(m, " %pS", v->caller);
3525
3526	if (v->nr_pages)
3527		seq_printf(m, " pages=%d", v->nr_pages);
3528
3529	if (v->phys_addr)
3530		seq_printf(m, " phys=%pa", &v->phys_addr);
3531
3532	if (v->flags & VM_IOREMAP)
3533		seq_puts(m, " ioremap");
3534
3535	if (v->flags & VM_ALLOC)
3536		seq_puts(m, " vmalloc");
3537
3538	if (v->flags & VM_MAP)
3539		seq_puts(m, " vmap");
3540
3541	if (v->flags & VM_USERMAP)
3542		seq_puts(m, " user");
3543
3544	if (v->flags & VM_DMA_COHERENT)
3545		seq_puts(m, " dma-coherent");
3546
3547	if (is_vmalloc_addr(v->pages))
3548		seq_puts(m, " vpages");
3549
3550	show_numa_info(m, v);
3551	seq_putc(m, '\n');
3552
3553	/*
3554	 * As a final step, dump "unpurged" areas. Note,
3555	 * that entire "/proc/vmallocinfo" output will not
3556	 * be address sorted, because the purge list is not
3557	 * sorted.
3558	 */
 
3559	if (list_is_last(&va->list, &vmap_area_list))
3560		show_purge_info(m);
3561
3562	return 0;
3563}
3564
3565static const struct seq_operations vmalloc_op = {
3566	.start = s_start,
3567	.next = s_next,
3568	.stop = s_stop,
3569	.show = s_show,
3570};
3571
3572static int __init proc_vmalloc_init(void)
3573{
3574	if (IS_ENABLED(CONFIG_NUMA))
3575		proc_create_seq_private("vmallocinfo", 0400, NULL,
3576				&vmalloc_op,
3577				nr_node_ids * sizeof(unsigned int), NULL);
3578	else
3579		proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3580	return 0;
3581}
3582module_init(proc_vmalloc_init);
3583
3584#endif