Linux Audio

Check our new training course

Loading...
v6.8
   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/uio.h>
  37#include <linux/bitops.h>
  38#include <linux/rbtree_augmented.h>
  39#include <linux/overflow.h>
  40#include <linux/pgtable.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
 
 
 
 
 
 
 
 
 
 
 
  92/*** Page table manipulation functions ***/
  93static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
  94			phys_addr_t phys_addr, pgprot_t prot,
  95			unsigned int max_page_shift, pgtbl_mod_mask *mask)
  96{
  97	pte_t *pte;
  98	u64 pfn;
  99	unsigned long size = PAGE_SIZE;
 100
 101	pfn = phys_addr >> PAGE_SHIFT;
 102	pte = pte_alloc_kernel_track(pmd, addr, mask);
 103	if (!pte)
 104		return -ENOMEM;
 105	do {
 106		BUG_ON(!pte_none(ptep_get(pte)));
 107
 108#ifdef CONFIG_HUGETLB_PAGE
 109		size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
 110		if (size != PAGE_SIZE) {
 111			pte_t entry = pfn_pte(pfn, prot);
 112
 113			entry = arch_make_huge_pte(entry, ilog2(size), 0);
 114			set_huge_pte_at(&init_mm, addr, pte, entry, size);
 115			pfn += PFN_DOWN(size);
 116			continue;
 117		}
 118#endif
 119		set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
 120		pfn++;
 121	} while (pte += PFN_DOWN(size), addr += size, addr != end);
 122	*mask |= PGTBL_PTE_MODIFIED;
 123	return 0;
 124}
 125
 126static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
 127			phys_addr_t phys_addr, pgprot_t prot,
 128			unsigned int max_page_shift)
 129{
 130	if (max_page_shift < PMD_SHIFT)
 131		return 0;
 132
 133	if (!arch_vmap_pmd_supported(prot))
 134		return 0;
 135
 136	if ((end - addr) != PMD_SIZE)
 137		return 0;
 138
 139	if (!IS_ALIGNED(addr, PMD_SIZE))
 140		return 0;
 141
 142	if (!IS_ALIGNED(phys_addr, PMD_SIZE))
 143		return 0;
 144
 145	if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
 146		return 0;
 147
 148	return pmd_set_huge(pmd, phys_addr, prot);
 149}
 150
 151static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 152			phys_addr_t phys_addr, pgprot_t prot,
 153			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 154{
 155	pmd_t *pmd;
 156	unsigned long next;
 157
 158	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 159	if (!pmd)
 160		return -ENOMEM;
 161	do {
 162		next = pmd_addr_end(addr, end);
 163
 164		if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
 165					max_page_shift)) {
 166			*mask |= PGTBL_PMD_MODIFIED;
 167			continue;
 168		}
 169
 170		if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
 171			return -ENOMEM;
 172	} while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
 173	return 0;
 174}
 175
 176static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
 177			phys_addr_t phys_addr, pgprot_t prot,
 178			unsigned int max_page_shift)
 179{
 180	if (max_page_shift < PUD_SHIFT)
 181		return 0;
 182
 183	if (!arch_vmap_pud_supported(prot))
 184		return 0;
 185
 186	if ((end - addr) != PUD_SIZE)
 187		return 0;
 188
 189	if (!IS_ALIGNED(addr, PUD_SIZE))
 190		return 0;
 191
 192	if (!IS_ALIGNED(phys_addr, PUD_SIZE))
 193		return 0;
 194
 195	if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
 196		return 0;
 197
 198	return pud_set_huge(pud, phys_addr, prot);
 199}
 200
 201static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 202			phys_addr_t phys_addr, pgprot_t prot,
 203			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 204{
 205	pud_t *pud;
 206	unsigned long next;
 207
 208	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 209	if (!pud)
 210		return -ENOMEM;
 211	do {
 212		next = pud_addr_end(addr, end);
 213
 214		if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
 215					max_page_shift)) {
 216			*mask |= PGTBL_PUD_MODIFIED;
 217			continue;
 218		}
 219
 220		if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
 221					max_page_shift, mask))
 222			return -ENOMEM;
 223	} while (pud++, phys_addr += (next - addr), addr = next, addr != end);
 224	return 0;
 225}
 226
 227static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
 228			phys_addr_t phys_addr, pgprot_t prot,
 229			unsigned int max_page_shift)
 230{
 231	if (max_page_shift < P4D_SHIFT)
 232		return 0;
 233
 234	if (!arch_vmap_p4d_supported(prot))
 235		return 0;
 236
 237	if ((end - addr) != P4D_SIZE)
 238		return 0;
 239
 240	if (!IS_ALIGNED(addr, P4D_SIZE))
 241		return 0;
 242
 243	if (!IS_ALIGNED(phys_addr, P4D_SIZE))
 244		return 0;
 245
 246	if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
 247		return 0;
 248
 249	return p4d_set_huge(p4d, phys_addr, prot);
 250}
 251
 252static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 253			phys_addr_t phys_addr, pgprot_t prot,
 254			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 255{
 256	p4d_t *p4d;
 257	unsigned long next;
 258
 259	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 260	if (!p4d)
 261		return -ENOMEM;
 262	do {
 263		next = p4d_addr_end(addr, end);
 264
 265		if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
 266					max_page_shift)) {
 267			*mask |= PGTBL_P4D_MODIFIED;
 268			continue;
 269		}
 270
 271		if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
 272					max_page_shift, mask))
 273			return -ENOMEM;
 274	} while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
 275	return 0;
 276}
 277
 278static int vmap_range_noflush(unsigned long addr, unsigned long end,
 279			phys_addr_t phys_addr, pgprot_t prot,
 280			unsigned int max_page_shift)
 281{
 282	pgd_t *pgd;
 283	unsigned long start;
 284	unsigned long next;
 285	int err;
 286	pgtbl_mod_mask mask = 0;
 287
 288	might_sleep();
 289	BUG_ON(addr >= end);
 290
 291	start = addr;
 292	pgd = pgd_offset_k(addr);
 293	do {
 294		next = pgd_addr_end(addr, end);
 295		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
 296					max_page_shift, &mask);
 297		if (err)
 298			break;
 299	} while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
 300
 301	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 302		arch_sync_kernel_mappings(start, end);
 303
 304	return err;
 305}
 306
 307int ioremap_page_range(unsigned long addr, unsigned long end,
 308		phys_addr_t phys_addr, pgprot_t prot)
 309{
 310	int err;
 311
 312	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
 313				 ioremap_max_page_shift);
 314	flush_cache_vmap(addr, end);
 315	if (!err)
 316		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
 317					       ioremap_max_page_shift);
 318	return err;
 319}
 320
 321static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 322			     pgtbl_mod_mask *mask)
 323{
 324	pte_t *pte;
 325
 326	pte = pte_offset_kernel(pmd, addr);
 327	do {
 328		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
 329		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
 330	} while (pte++, addr += PAGE_SIZE, addr != end);
 331	*mask |= PGTBL_PTE_MODIFIED;
 332}
 333
 334static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 335			     pgtbl_mod_mask *mask)
 336{
 337	pmd_t *pmd;
 338	unsigned long next;
 339	int cleared;
 340
 341	pmd = pmd_offset(pud, addr);
 342	do {
 343		next = pmd_addr_end(addr, end);
 344
 345		cleared = pmd_clear_huge(pmd);
 346		if (cleared || pmd_bad(*pmd))
 347			*mask |= PGTBL_PMD_MODIFIED;
 348
 349		if (cleared)
 350			continue;
 351		if (pmd_none_or_clear_bad(pmd))
 352			continue;
 353		vunmap_pte_range(pmd, addr, next, mask);
 354
 355		cond_resched();
 356	} while (pmd++, addr = next, addr != end);
 357}
 358
 359static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 360			     pgtbl_mod_mask *mask)
 361{
 362	pud_t *pud;
 363	unsigned long next;
 364	int cleared;
 365
 366	pud = pud_offset(p4d, addr);
 367	do {
 368		next = pud_addr_end(addr, end);
 369
 370		cleared = pud_clear_huge(pud);
 371		if (cleared || pud_bad(*pud))
 372			*mask |= PGTBL_PUD_MODIFIED;
 373
 374		if (cleared)
 375			continue;
 376		if (pud_none_or_clear_bad(pud))
 377			continue;
 378		vunmap_pmd_range(pud, addr, next, mask);
 379	} while (pud++, addr = next, addr != end);
 380}
 381
 382static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 383			     pgtbl_mod_mask *mask)
 384{
 385	p4d_t *p4d;
 386	unsigned long next;
 387
 388	p4d = p4d_offset(pgd, addr);
 389	do {
 390		next = p4d_addr_end(addr, end);
 391
 392		p4d_clear_huge(p4d);
 393		if (p4d_bad(*p4d))
 394			*mask |= PGTBL_P4D_MODIFIED;
 395
 396		if (p4d_none_or_clear_bad(p4d))
 397			continue;
 398		vunmap_pud_range(p4d, addr, next, mask);
 399	} while (p4d++, addr = next, addr != end);
 400}
 401
 402/*
 403 * vunmap_range_noflush is similar to vunmap_range, but does not
 404 * flush caches or TLBs.
 405 *
 406 * The caller is responsible for calling flush_cache_vmap() before calling
 407 * this function, and flush_tlb_kernel_range after it has returned
 408 * successfully (and before the addresses are expected to cause a page fault
 409 * or be re-mapped for something else, if TLB flushes are being delayed or
 410 * coalesced).
 411 *
 412 * This is an internal function only. Do not use outside mm/.
 413 */
 414void __vunmap_range_noflush(unsigned long start, unsigned long end)
 415{
 416	unsigned long next;
 417	pgd_t *pgd;
 418	unsigned long addr = start;
 419	pgtbl_mod_mask mask = 0;
 420
 421	BUG_ON(addr >= end);
 422	pgd = pgd_offset_k(addr);
 423	do {
 424		next = pgd_addr_end(addr, end);
 425		if (pgd_bad(*pgd))
 426			mask |= PGTBL_PGD_MODIFIED;
 427		if (pgd_none_or_clear_bad(pgd))
 428			continue;
 429		vunmap_p4d_range(pgd, addr, next, &mask);
 430	} while (pgd++, addr = next, addr != end);
 431
 432	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 433		arch_sync_kernel_mappings(start, end);
 434}
 435
 436void vunmap_range_noflush(unsigned long start, unsigned long end)
 437{
 438	kmsan_vunmap_range_noflush(start, end);
 439	__vunmap_range_noflush(start, end);
 440}
 441
 442/**
 443 * vunmap_range - unmap kernel virtual addresses
 444 * @addr: start of the VM area to unmap
 445 * @end: end of the VM area to unmap (non-inclusive)
 446 *
 447 * Clears any present PTEs in the virtual address range, flushes TLBs and
 448 * caches. Any subsequent access to the address before it has been re-mapped
 449 * is a kernel bug.
 450 */
 451void vunmap_range(unsigned long addr, unsigned long end)
 452{
 453	flush_cache_vunmap(addr, end);
 454	vunmap_range_noflush(addr, end);
 455	flush_tlb_kernel_range(addr, end);
 456}
 457
 458static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
 459		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 460		pgtbl_mod_mask *mask)
 461{
 462	pte_t *pte;
 463
 464	/*
 465	 * nr is a running index into the array which helps higher level
 466	 * callers keep track of where we're up to.
 467	 */
 468
 469	pte = pte_alloc_kernel_track(pmd, addr, mask);
 470	if (!pte)
 471		return -ENOMEM;
 472	do {
 473		struct page *page = pages[*nr];
 474
 475		if (WARN_ON(!pte_none(ptep_get(pte))))
 476			return -EBUSY;
 477		if (WARN_ON(!page))
 478			return -ENOMEM;
 479		if (WARN_ON(!pfn_valid(page_to_pfn(page))))
 480			return -EINVAL;
 481
 482		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 483		(*nr)++;
 484	} while (pte++, addr += PAGE_SIZE, addr != end);
 485	*mask |= PGTBL_PTE_MODIFIED;
 486	return 0;
 487}
 488
 489static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
 490		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 491		pgtbl_mod_mask *mask)
 492{
 493	pmd_t *pmd;
 494	unsigned long next;
 495
 496	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 497	if (!pmd)
 498		return -ENOMEM;
 499	do {
 500		next = pmd_addr_end(addr, end);
 501		if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
 502			return -ENOMEM;
 503	} while (pmd++, addr = next, addr != end);
 504	return 0;
 505}
 506
 507static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
 508		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 509		pgtbl_mod_mask *mask)
 510{
 511	pud_t *pud;
 512	unsigned long next;
 513
 514	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 515	if (!pud)
 516		return -ENOMEM;
 517	do {
 518		next = pud_addr_end(addr, end);
 519		if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
 520			return -ENOMEM;
 521	} while (pud++, addr = next, addr != end);
 522	return 0;
 523}
 524
 525static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
 526		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 527		pgtbl_mod_mask *mask)
 528{
 529	p4d_t *p4d;
 530	unsigned long next;
 531
 532	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 533	if (!p4d)
 534		return -ENOMEM;
 535	do {
 536		next = p4d_addr_end(addr, end);
 537		if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
 538			return -ENOMEM;
 539	} while (p4d++, addr = next, addr != end);
 540	return 0;
 541}
 542
 543static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
 544		pgprot_t prot, struct page **pages)
 545{
 546	unsigned long start = addr;
 547	pgd_t *pgd;
 548	unsigned long next;
 549	int err = 0;
 550	int nr = 0;
 551	pgtbl_mod_mask mask = 0;
 552
 553	BUG_ON(addr >= end);
 554	pgd = pgd_offset_k(addr);
 555	do {
 556		next = pgd_addr_end(addr, end);
 557		if (pgd_bad(*pgd))
 558			mask |= PGTBL_PGD_MODIFIED;
 559		err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
 560		if (err)
 561			return err;
 562	} while (pgd++, addr = next, addr != end);
 563
 564	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 565		arch_sync_kernel_mappings(start, end);
 566
 567	return 0;
 568}
 569
 570/*
 571 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
 572 * flush caches.
 573 *
 574 * The caller is responsible for calling flush_cache_vmap() after this
 575 * function returns successfully and before the addresses are accessed.
 576 *
 577 * This is an internal function only. Do not use outside mm/.
 578 */
 579int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 580		pgprot_t prot, struct page **pages, unsigned int page_shift)
 581{
 582	unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
 583
 584	WARN_ON(page_shift < PAGE_SHIFT);
 585
 586	if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
 587			page_shift == PAGE_SHIFT)
 588		return vmap_small_pages_range_noflush(addr, end, prot, pages);
 589
 590	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
 591		int err;
 592
 593		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
 594					page_to_phys(pages[i]), prot,
 595					page_shift);
 596		if (err)
 597			return err;
 598
 599		addr += 1UL << page_shift;
 600	}
 601
 602	return 0;
 603}
 604
 605int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 606		pgprot_t prot, struct page **pages, unsigned int page_shift)
 607{
 608	int ret = kmsan_vmap_pages_range_noflush(addr, end, prot, pages,
 609						 page_shift);
 610
 611	if (ret)
 612		return ret;
 613	return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 614}
 615
 616/**
 617 * vmap_pages_range - map pages to a kernel virtual address
 618 * @addr: start of the VM area to map
 619 * @end: end of the VM area to map (non-inclusive)
 620 * @prot: page protection flags to use
 621 * @pages: pages to map (always PAGE_SIZE pages)
 622 * @page_shift: maximum shift that the pages may be mapped with, @pages must
 623 * be aligned and contiguous up to at least this shift.
 624 *
 625 * RETURNS:
 626 * 0 on success, -errno on failure.
 627 */
 628static int vmap_pages_range(unsigned long addr, unsigned long end,
 629		pgprot_t prot, struct page **pages, unsigned int page_shift)
 630{
 631	int err;
 632
 633	err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 634	flush_cache_vmap(addr, end);
 635	return err;
 636}
 637
 638int is_vmalloc_or_module_addr(const void *x)
 639{
 640	/*
 641	 * ARM, x86-64 and sparc64 put modules in a special place,
 642	 * and fall back on vmalloc() if that fails. Others
 643	 * just put it in the vmalloc space.
 644	 */
 645#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 646	unsigned long addr = (unsigned long)kasan_reset_tag(x);
 647	if (addr >= MODULES_VADDR && addr < MODULES_END)
 648		return 1;
 649#endif
 650	return is_vmalloc_addr(x);
 651}
 652EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr);
 653
 654/*
 655 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
 656 * return the tail page that corresponds to the base page address, which
 657 * matches small vmap mappings.
 658 */
 659struct page *vmalloc_to_page(const void *vmalloc_addr)
 660{
 661	unsigned long addr = (unsigned long) vmalloc_addr;
 662	struct page *page = NULL;
 663	pgd_t *pgd = pgd_offset_k(addr);
 664	p4d_t *p4d;
 665	pud_t *pud;
 666	pmd_t *pmd;
 667	pte_t *ptep, pte;
 668
 669	/*
 670	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 671	 * architectures that do not vmalloc module space
 672	 */
 673	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 674
 675	if (pgd_none(*pgd))
 676		return NULL;
 677	if (WARN_ON_ONCE(pgd_leaf(*pgd)))
 678		return NULL; /* XXX: no allowance for huge pgd */
 679	if (WARN_ON_ONCE(pgd_bad(*pgd)))
 680		return NULL;
 681
 682	p4d = p4d_offset(pgd, addr);
 683	if (p4d_none(*p4d))
 684		return NULL;
 685	if (p4d_leaf(*p4d))
 686		return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
 687	if (WARN_ON_ONCE(p4d_bad(*p4d)))
 688		return NULL;
 689
 690	pud = pud_offset(p4d, addr);
 691	if (pud_none(*pud))
 692		return NULL;
 693	if (pud_leaf(*pud))
 694		return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 695	if (WARN_ON_ONCE(pud_bad(*pud)))
 696		return NULL;
 697
 698	pmd = pmd_offset(pud, addr);
 699	if (pmd_none(*pmd))
 700		return NULL;
 701	if (pmd_leaf(*pmd))
 702		return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 703	if (WARN_ON_ONCE(pmd_bad(*pmd)))
 704		return NULL;
 705
 706	ptep = pte_offset_kernel(pmd, addr);
 707	pte = ptep_get(ptep);
 708	if (pte_present(pte))
 709		page = pte_page(pte);
 
 710
 711	return page;
 712}
 713EXPORT_SYMBOL(vmalloc_to_page);
 714
 715/*
 716 * Map a vmalloc()-space virtual address to the physical page frame number.
 717 */
 718unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 719{
 720	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 721}
 722EXPORT_SYMBOL(vmalloc_to_pfn);
 723
 724
 725/*** Global kva allocator ***/
 726
 727#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
 728#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
 729
 730
 731static DEFINE_SPINLOCK(vmap_area_lock);
 732static DEFINE_SPINLOCK(free_vmap_area_lock);
 733/* Export for kexec only */
 734LIST_HEAD(vmap_area_list);
 735static struct rb_root vmap_area_root = RB_ROOT;
 736static bool vmap_initialized __read_mostly;
 737
 738static struct rb_root purge_vmap_area_root = RB_ROOT;
 739static LIST_HEAD(purge_vmap_area_list);
 740static DEFINE_SPINLOCK(purge_vmap_area_lock);
 741
 742/*
 743 * This kmem_cache is used for vmap_area objects. Instead of
 744 * allocating from slab we reuse an object from this cache to
 745 * make things faster. Especially in "no edge" splitting of
 746 * free block.
 747 */
 748static struct kmem_cache *vmap_area_cachep;
 749
 750/*
 751 * This linked list is used in pair with free_vmap_area_root.
 752 * It gives O(1) access to prev/next to perform fast coalescing.
 753 */
 754static LIST_HEAD(free_vmap_area_list);
 755
 756/*
 757 * This augment red-black tree represents the free vmap space.
 758 * All vmap_area objects in this tree are sorted by va->va_start
 759 * address. It is used for allocation and merging when a vmap
 760 * object is released.
 761 *
 762 * Each vmap_area node contains a maximum available free block
 763 * of its sub-tree, right or left. Therefore it is possible to
 764 * find a lowest match of free area.
 765 */
 766static struct rb_root free_vmap_area_root = RB_ROOT;
 767
 768/*
 769 * Preload a CPU with one object for "no edge" split case. The
 770 * aim is to get rid of allocations from the atomic context, thus
 771 * to use more permissive allocation masks.
 772 */
 773static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
 774
 775static __always_inline unsigned long
 776va_size(struct vmap_area *va)
 777{
 778	return (va->va_end - va->va_start);
 779}
 780
 781static __always_inline unsigned long
 782get_subtree_max_size(struct rb_node *node)
 783{
 784	struct vmap_area *va;
 785
 786	va = rb_entry_safe(node, struct vmap_area, rb_node);
 787	return va ? va->subtree_max_size : 0;
 788}
 789
 790RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
 791	struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
 792
 793static void reclaim_and_purge_vmap_areas(void);
 794static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
 795static void drain_vmap_area_work(struct work_struct *work);
 796static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
 797
 798static atomic_long_t nr_vmalloc_pages;
 799
 800unsigned long vmalloc_nr_pages(void)
 801{
 802	return atomic_long_read(&nr_vmalloc_pages);
 803}
 804
 805/* Look up the first VA which satisfies addr < va_end, NULL if none. */
 806static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
 807{
 808	struct vmap_area *va = NULL;
 809	struct rb_node *n = vmap_area_root.rb_node;
 810
 811	addr = (unsigned long)kasan_reset_tag((void *)addr);
 812
 813	while (n) {
 814		struct vmap_area *tmp;
 815
 816		tmp = rb_entry(n, struct vmap_area, rb_node);
 817		if (tmp->va_end > addr) {
 818			va = tmp;
 819			if (tmp->va_start <= addr)
 820				break;
 821
 822			n = n->rb_left;
 823		} else
 824			n = n->rb_right;
 825	}
 826
 827	return va;
 828}
 829
 830static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
 831{
 832	struct rb_node *n = root->rb_node;
 833
 834	addr = (unsigned long)kasan_reset_tag((void *)addr);
 835
 836	while (n) {
 837		struct vmap_area *va;
 838
 839		va = rb_entry(n, struct vmap_area, rb_node);
 840		if (addr < va->va_start)
 841			n = n->rb_left;
 842		else if (addr >= va->va_end)
 843			n = n->rb_right;
 844		else
 845			return va;
 846	}
 847
 848	return NULL;
 849}
 850
 851/*
 852 * This function returns back addresses of parent node
 853 * and its left or right link for further processing.
 854 *
 855 * Otherwise NULL is returned. In that case all further
 856 * steps regarding inserting of conflicting overlap range
 857 * have to be declined and actually considered as a bug.
 858 */
 859static __always_inline struct rb_node **
 860find_va_links(struct vmap_area *va,
 861	struct rb_root *root, struct rb_node *from,
 862	struct rb_node **parent)
 863{
 864	struct vmap_area *tmp_va;
 865	struct rb_node **link;
 866
 867	if (root) {
 868		link = &root->rb_node;
 869		if (unlikely(!*link)) {
 870			*parent = NULL;
 871			return link;
 872		}
 873	} else {
 874		link = &from;
 875	}
 876
 877	/*
 878	 * Go to the bottom of the tree. When we hit the last point
 879	 * we end up with parent rb_node and correct direction, i name
 880	 * it link, where the new va->rb_node will be attached to.
 881	 */
 882	do {
 883		tmp_va = rb_entry(*link, struct vmap_area, rb_node);
 884
 885		/*
 886		 * During the traversal we also do some sanity check.
 887		 * Trigger the BUG() if there are sides(left/right)
 888		 * or full overlaps.
 889		 */
 890		if (va->va_end <= tmp_va->va_start)
 891			link = &(*link)->rb_left;
 892		else if (va->va_start >= tmp_va->va_end)
 893			link = &(*link)->rb_right;
 894		else {
 895			WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
 896				va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
 897
 898			return NULL;
 899		}
 900	} while (*link);
 901
 902	*parent = &tmp_va->rb_node;
 903	return link;
 904}
 905
 906static __always_inline struct list_head *
 907get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
 908{
 909	struct list_head *list;
 910
 911	if (unlikely(!parent))
 912		/*
 913		 * The red-black tree where we try to find VA neighbors
 914		 * before merging or inserting is empty, i.e. it means
 915		 * there is no free vmap space. Normally it does not
 916		 * happen but we handle this case anyway.
 917		 */
 918		return NULL;
 919
 920	list = &rb_entry(parent, struct vmap_area, rb_node)->list;
 921	return (&parent->rb_right == link ? list->next : list);
 922}
 923
 924static __always_inline void
 925__link_va(struct vmap_area *va, struct rb_root *root,
 926	struct rb_node *parent, struct rb_node **link,
 927	struct list_head *head, bool augment)
 928{
 929	/*
 930	 * VA is still not in the list, but we can
 931	 * identify its future previous list_head node.
 932	 */
 933	if (likely(parent)) {
 934		head = &rb_entry(parent, struct vmap_area, rb_node)->list;
 935		if (&parent->rb_right != link)
 936			head = head->prev;
 937	}
 938
 939	/* Insert to the rb-tree */
 940	rb_link_node(&va->rb_node, parent, link);
 941	if (augment) {
 942		/*
 943		 * Some explanation here. Just perform simple insertion
 944		 * to the tree. We do not set va->subtree_max_size to
 945		 * its current size before calling rb_insert_augmented().
 946		 * It is because we populate the tree from the bottom
 947		 * to parent levels when the node _is_ in the tree.
 948		 *
 949		 * Therefore we set subtree_max_size to zero after insertion,
 950		 * to let __augment_tree_propagate_from() puts everything to
 951		 * the correct order later on.
 952		 */
 953		rb_insert_augmented(&va->rb_node,
 954			root, &free_vmap_area_rb_augment_cb);
 955		va->subtree_max_size = 0;
 956	} else {
 957		rb_insert_color(&va->rb_node, root);
 958	}
 959
 960	/* Address-sort this list */
 961	list_add(&va->list, head);
 962}
 963
 964static __always_inline void
 965link_va(struct vmap_area *va, struct rb_root *root,
 966	struct rb_node *parent, struct rb_node **link,
 967	struct list_head *head)
 968{
 969	__link_va(va, root, parent, link, head, false);
 970}
 971
 972static __always_inline void
 973link_va_augment(struct vmap_area *va, struct rb_root *root,
 974	struct rb_node *parent, struct rb_node **link,
 975	struct list_head *head)
 976{
 977	__link_va(va, root, parent, link, head, true);
 978}
 979
 980static __always_inline void
 981__unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
 982{
 983	if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
 984		return;
 985
 986	if (augment)
 987		rb_erase_augmented(&va->rb_node,
 988			root, &free_vmap_area_rb_augment_cb);
 989	else
 990		rb_erase(&va->rb_node, root);
 991
 992	list_del_init(&va->list);
 993	RB_CLEAR_NODE(&va->rb_node);
 994}
 995
 996static __always_inline void
 997unlink_va(struct vmap_area *va, struct rb_root *root)
 998{
 999	__unlink_va(va, root, false);
1000}
1001
1002static __always_inline void
1003unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1004{
1005	__unlink_va(va, root, true);
1006}
1007
1008#if DEBUG_AUGMENT_PROPAGATE_CHECK
1009/*
1010 * Gets called when remove the node and rotate.
1011 */
1012static __always_inline unsigned long
1013compute_subtree_max_size(struct vmap_area *va)
1014{
1015	return max3(va_size(va),
1016		get_subtree_max_size(va->rb_node.rb_left),
1017		get_subtree_max_size(va->rb_node.rb_right));
1018}
1019
1020static void
1021augment_tree_propagate_check(void)
1022{
1023	struct vmap_area *va;
1024	unsigned long computed_size;
1025
1026	list_for_each_entry(va, &free_vmap_area_list, list) {
1027		computed_size = compute_subtree_max_size(va);
1028		if (computed_size != va->subtree_max_size)
1029			pr_emerg("tree is corrupted: %lu, %lu\n",
1030				va_size(va), va->subtree_max_size);
1031	}
1032}
1033#endif
1034
1035/*
1036 * This function populates subtree_max_size from bottom to upper
1037 * levels starting from VA point. The propagation must be done
1038 * when VA size is modified by changing its va_start/va_end. Or
1039 * in case of newly inserting of VA to the tree.
1040 *
1041 * It means that __augment_tree_propagate_from() must be called:
1042 * - After VA has been inserted to the tree(free path);
1043 * - After VA has been shrunk(allocation path);
1044 * - After VA has been increased(merging path).
1045 *
1046 * Please note that, it does not mean that upper parent nodes
1047 * and their subtree_max_size are recalculated all the time up
1048 * to the root node.
1049 *
1050 *       4--8
1051 *        /\
1052 *       /  \
1053 *      /    \
1054 *    2--2  8--8
1055 *
1056 * For example if we modify the node 4, shrinking it to 2, then
1057 * no any modification is required. If we shrink the node 2 to 1
1058 * its subtree_max_size is updated only, and set to 1. If we shrink
1059 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1060 * node becomes 4--6.
1061 */
1062static __always_inline void
1063augment_tree_propagate_from(struct vmap_area *va)
1064{
1065	/*
1066	 * Populate the tree from bottom towards the root until
1067	 * the calculated maximum available size of checked node
1068	 * is equal to its current one.
1069	 */
1070	free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1071
1072#if DEBUG_AUGMENT_PROPAGATE_CHECK
1073	augment_tree_propagate_check();
1074#endif
1075}
1076
1077static void
1078insert_vmap_area(struct vmap_area *va,
1079	struct rb_root *root, struct list_head *head)
1080{
1081	struct rb_node **link;
1082	struct rb_node *parent;
1083
1084	link = find_va_links(va, root, NULL, &parent);
1085	if (link)
1086		link_va(va, root, parent, link, head);
1087}
1088
1089static void
1090insert_vmap_area_augment(struct vmap_area *va,
1091	struct rb_node *from, struct rb_root *root,
1092	struct list_head *head)
1093{
1094	struct rb_node **link;
1095	struct rb_node *parent;
1096
1097	if (from)
1098		link = find_va_links(va, NULL, from, &parent);
1099	else
1100		link = find_va_links(va, root, NULL, &parent);
1101
1102	if (link) {
1103		link_va_augment(va, root, parent, link, head);
1104		augment_tree_propagate_from(va);
1105	}
1106}
1107
1108/*
1109 * Merge de-allocated chunk of VA memory with previous
1110 * and next free blocks. If coalesce is not done a new
1111 * free area is inserted. If VA has been merged, it is
1112 * freed.
1113 *
1114 * Please note, it can return NULL in case of overlap
1115 * ranges, followed by WARN() report. Despite it is a
1116 * buggy behaviour, a system can be alive and keep
1117 * ongoing.
1118 */
1119static __always_inline struct vmap_area *
1120__merge_or_add_vmap_area(struct vmap_area *va,
1121	struct rb_root *root, struct list_head *head, bool augment)
1122{
1123	struct vmap_area *sibling;
1124	struct list_head *next;
1125	struct rb_node **link;
1126	struct rb_node *parent;
1127	bool merged = false;
1128
1129	/*
1130	 * Find a place in the tree where VA potentially will be
1131	 * inserted, unless it is merged with its sibling/siblings.
1132	 */
1133	link = find_va_links(va, root, NULL, &parent);
1134	if (!link)
1135		return NULL;
1136
1137	/*
1138	 * Get next node of VA to check if merging can be done.
1139	 */
1140	next = get_va_next_sibling(parent, link);
1141	if (unlikely(next == NULL))
1142		goto insert;
1143
1144	/*
1145	 * start            end
1146	 * |                |
1147	 * |<------VA------>|<-----Next----->|
1148	 *                  |                |
1149	 *                  start            end
1150	 */
1151	if (next != head) {
1152		sibling = list_entry(next, struct vmap_area, list);
1153		if (sibling->va_start == va->va_end) {
1154			sibling->va_start = va->va_start;
1155
1156			/* Free vmap_area object. */
1157			kmem_cache_free(vmap_area_cachep, va);
1158
1159			/* Point to the new merged area. */
1160			va = sibling;
1161			merged = true;
1162		}
1163	}
1164
1165	/*
1166	 * start            end
1167	 * |                |
1168	 * |<-----Prev----->|<------VA------>|
1169	 *                  |                |
1170	 *                  start            end
1171	 */
1172	if (next->prev != head) {
1173		sibling = list_entry(next->prev, struct vmap_area, list);
1174		if (sibling->va_end == va->va_start) {
1175			/*
1176			 * If both neighbors are coalesced, it is important
1177			 * to unlink the "next" node first, followed by merging
1178			 * with "previous" one. Otherwise the tree might not be
1179			 * fully populated if a sibling's augmented value is
1180			 * "normalized" because of rotation operations.
1181			 */
1182			if (merged)
1183				__unlink_va(va, root, augment);
1184
1185			sibling->va_end = va->va_end;
1186
1187			/* Free vmap_area object. */
1188			kmem_cache_free(vmap_area_cachep, va);
1189
1190			/* Point to the new merged area. */
1191			va = sibling;
1192			merged = true;
1193		}
1194	}
1195
1196insert:
1197	if (!merged)
1198		__link_va(va, root, parent, link, head, augment);
1199
1200	return va;
1201}
1202
1203static __always_inline struct vmap_area *
1204merge_or_add_vmap_area(struct vmap_area *va,
1205	struct rb_root *root, struct list_head *head)
1206{
1207	return __merge_or_add_vmap_area(va, root, head, false);
1208}
1209
1210static __always_inline struct vmap_area *
1211merge_or_add_vmap_area_augment(struct vmap_area *va,
1212	struct rb_root *root, struct list_head *head)
1213{
1214	va = __merge_or_add_vmap_area(va, root, head, true);
1215	if (va)
1216		augment_tree_propagate_from(va);
1217
1218	return va;
1219}
1220
1221static __always_inline bool
1222is_within_this_va(struct vmap_area *va, unsigned long size,
1223	unsigned long align, unsigned long vstart)
1224{
1225	unsigned long nva_start_addr;
1226
1227	if (va->va_start > vstart)
1228		nva_start_addr = ALIGN(va->va_start, align);
1229	else
1230		nva_start_addr = ALIGN(vstart, align);
1231
1232	/* Can be overflowed due to big size or alignment. */
1233	if (nva_start_addr + size < nva_start_addr ||
1234			nva_start_addr < vstart)
1235		return false;
1236
1237	return (nva_start_addr + size <= va->va_end);
1238}
1239
1240/*
1241 * Find the first free block(lowest start address) in the tree,
1242 * that will accomplish the request corresponding to passing
1243 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1244 * a search length is adjusted to account for worst case alignment
1245 * overhead.
1246 */
1247static __always_inline struct vmap_area *
1248find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1249	unsigned long align, unsigned long vstart, bool adjust_search_size)
1250{
1251	struct vmap_area *va;
1252	struct rb_node *node;
1253	unsigned long length;
1254
1255	/* Start from the root. */
1256	node = root->rb_node;
1257
1258	/* Adjust the search size for alignment overhead. */
1259	length = adjust_search_size ? size + align - 1 : size;
1260
1261	while (node) {
1262		va = rb_entry(node, struct vmap_area, rb_node);
1263
1264		if (get_subtree_max_size(node->rb_left) >= length &&
1265				vstart < va->va_start) {
1266			node = node->rb_left;
1267		} else {
1268			if (is_within_this_va(va, size, align, vstart))
1269				return va;
1270
1271			/*
1272			 * Does not make sense to go deeper towards the right
1273			 * sub-tree if it does not have a free block that is
1274			 * equal or bigger to the requested search length.
1275			 */
1276			if (get_subtree_max_size(node->rb_right) >= length) {
1277				node = node->rb_right;
1278				continue;
1279			}
1280
1281			/*
1282			 * OK. We roll back and find the first right sub-tree,
1283			 * that will satisfy the search criteria. It can happen
1284			 * due to "vstart" restriction or an alignment overhead
1285			 * that is bigger then PAGE_SIZE.
1286			 */
1287			while ((node = rb_parent(node))) {
1288				va = rb_entry(node, struct vmap_area, rb_node);
1289				if (is_within_this_va(va, size, align, vstart))
1290					return va;
1291
1292				if (get_subtree_max_size(node->rb_right) >= length &&
1293						vstart <= va->va_start) {
1294					/*
1295					 * Shift the vstart forward. Please note, we update it with
1296					 * parent's start address adding "1" because we do not want
1297					 * to enter same sub-tree after it has already been checked
1298					 * and no suitable free block found there.
1299					 */
1300					vstart = va->va_start + 1;
1301					node = node->rb_right;
1302					break;
1303				}
1304			}
1305		}
1306	}
1307
1308	return NULL;
1309}
1310
1311#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1312#include <linux/random.h>
1313
1314static struct vmap_area *
1315find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1316	unsigned long align, unsigned long vstart)
1317{
1318	struct vmap_area *va;
1319
1320	list_for_each_entry(va, head, list) {
1321		if (!is_within_this_va(va, size, align, vstart))
1322			continue;
1323
1324		return va;
1325	}
1326
1327	return NULL;
1328}
1329
1330static void
1331find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1332			     unsigned long size, unsigned long align)
1333{
1334	struct vmap_area *va_1, *va_2;
1335	unsigned long vstart;
1336	unsigned int rnd;
1337
1338	get_random_bytes(&rnd, sizeof(rnd));
1339	vstart = VMALLOC_START + rnd;
1340
1341	va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1342	va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1343
1344	if (va_1 != va_2)
1345		pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1346			va_1, va_2, vstart);
1347}
1348#endif
1349
1350enum fit_type {
1351	NOTHING_FIT = 0,
1352	FL_FIT_TYPE = 1,	/* full fit */
1353	LE_FIT_TYPE = 2,	/* left edge fit */
1354	RE_FIT_TYPE = 3,	/* right edge fit */
1355	NE_FIT_TYPE = 4		/* no edge fit */
1356};
1357
1358static __always_inline enum fit_type
1359classify_va_fit_type(struct vmap_area *va,
1360	unsigned long nva_start_addr, unsigned long size)
1361{
1362	enum fit_type type;
1363
1364	/* Check if it is within VA. */
1365	if (nva_start_addr < va->va_start ||
1366			nva_start_addr + size > va->va_end)
1367		return NOTHING_FIT;
1368
1369	/* Now classify. */
1370	if (va->va_start == nva_start_addr) {
1371		if (va->va_end == nva_start_addr + size)
1372			type = FL_FIT_TYPE;
1373		else
1374			type = LE_FIT_TYPE;
1375	} else if (va->va_end == nva_start_addr + size) {
1376		type = RE_FIT_TYPE;
1377	} else {
1378		type = NE_FIT_TYPE;
1379	}
1380
1381	return type;
1382}
1383
1384static __always_inline int
1385adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1386		      struct vmap_area *va, unsigned long nva_start_addr,
1387		      unsigned long size)
1388{
1389	struct vmap_area *lva = NULL;
1390	enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1391
1392	if (type == FL_FIT_TYPE) {
1393		/*
1394		 * No need to split VA, it fully fits.
1395		 *
1396		 * |               |
1397		 * V      NVA      V
1398		 * |---------------|
1399		 */
1400		unlink_va_augment(va, root);
1401		kmem_cache_free(vmap_area_cachep, va);
1402	} else if (type == LE_FIT_TYPE) {
1403		/*
1404		 * Split left edge of fit VA.
1405		 *
1406		 * |       |
1407		 * V  NVA  V   R
1408		 * |-------|-------|
1409		 */
1410		va->va_start += size;
1411	} else if (type == RE_FIT_TYPE) {
1412		/*
1413		 * Split right edge of fit VA.
1414		 *
1415		 *         |       |
1416		 *     L   V  NVA  V
1417		 * |-------|-------|
1418		 */
1419		va->va_end = nva_start_addr;
1420	} else if (type == NE_FIT_TYPE) {
1421		/*
1422		 * Split no edge of fit VA.
1423		 *
1424		 *     |       |
1425		 *   L V  NVA  V R
1426		 * |---|-------|---|
1427		 */
1428		lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1429		if (unlikely(!lva)) {
1430			/*
1431			 * For percpu allocator we do not do any pre-allocation
1432			 * and leave it as it is. The reason is it most likely
1433			 * never ends up with NE_FIT_TYPE splitting. In case of
1434			 * percpu allocations offsets and sizes are aligned to
1435			 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1436			 * are its main fitting cases.
1437			 *
1438			 * There are a few exceptions though, as an example it is
1439			 * a first allocation (early boot up) when we have "one"
1440			 * big free space that has to be split.
1441			 *
1442			 * Also we can hit this path in case of regular "vmap"
1443			 * allocations, if "this" current CPU was not preloaded.
1444			 * See the comment in alloc_vmap_area() why. If so, then
1445			 * GFP_NOWAIT is used instead to get an extra object for
1446			 * split purpose. That is rare and most time does not
1447			 * occur.
1448			 *
1449			 * What happens if an allocation gets failed. Basically,
1450			 * an "overflow" path is triggered to purge lazily freed
1451			 * areas to free some memory, then, the "retry" path is
1452			 * triggered to repeat one more time. See more details
1453			 * in alloc_vmap_area() function.
1454			 */
1455			lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1456			if (!lva)
1457				return -1;
1458		}
1459
1460		/*
1461		 * Build the remainder.
1462		 */
1463		lva->va_start = va->va_start;
1464		lva->va_end = nva_start_addr;
1465
1466		/*
1467		 * Shrink this VA to remaining size.
1468		 */
1469		va->va_start = nva_start_addr + size;
1470	} else {
1471		return -1;
1472	}
1473
1474	if (type != FL_FIT_TYPE) {
1475		augment_tree_propagate_from(va);
1476
1477		if (lva)	/* type == NE_FIT_TYPE */
1478			insert_vmap_area_augment(lva, &va->rb_node, root, head);
1479	}
1480
1481	return 0;
1482}
1483
1484/*
1485 * Returns a start address of the newly allocated area, if success.
1486 * Otherwise a vend is returned that indicates failure.
1487 */
1488static __always_inline unsigned long
1489__alloc_vmap_area(struct rb_root *root, struct list_head *head,
1490	unsigned long size, unsigned long align,
1491	unsigned long vstart, unsigned long vend)
1492{
1493	bool adjust_search_size = true;
1494	unsigned long nva_start_addr;
1495	struct vmap_area *va;
1496	int ret;
1497
1498	/*
1499	 * Do not adjust when:
1500	 *   a) align <= PAGE_SIZE, because it does not make any sense.
1501	 *      All blocks(their start addresses) are at least PAGE_SIZE
1502	 *      aligned anyway;
1503	 *   b) a short range where a requested size corresponds to exactly
1504	 *      specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1505	 *      With adjusted search length an allocation would not succeed.
1506	 */
1507	if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1508		adjust_search_size = false;
1509
1510	va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1511	if (unlikely(!va))
1512		return vend;
1513
1514	if (va->va_start > vstart)
1515		nva_start_addr = ALIGN(va->va_start, align);
1516	else
1517		nva_start_addr = ALIGN(vstart, align);
1518
1519	/* Check the "vend" restriction. */
1520	if (nva_start_addr + size > vend)
1521		return vend;
1522
1523	/* Update the free vmap_area. */
1524	ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1525	if (WARN_ON_ONCE(ret))
1526		return vend;
1527
1528#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1529	find_vmap_lowest_match_check(root, head, size, align);
1530#endif
1531
1532	return nva_start_addr;
1533}
1534
1535/*
1536 * Free a region of KVA allocated by alloc_vmap_area
1537 */
1538static void free_vmap_area(struct vmap_area *va)
1539{
1540	/*
1541	 * Remove from the busy tree/list.
1542	 */
1543	spin_lock(&vmap_area_lock);
1544	unlink_va(va, &vmap_area_root);
1545	spin_unlock(&vmap_area_lock);
1546
1547	/*
1548	 * Insert/Merge it back to the free tree/list.
1549	 */
1550	spin_lock(&free_vmap_area_lock);
1551	merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1552	spin_unlock(&free_vmap_area_lock);
1553}
1554
1555static inline void
1556preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1557{
1558	struct vmap_area *va = NULL;
1559
1560	/*
1561	 * Preload this CPU with one extra vmap_area object. It is used
1562	 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1563	 * a CPU that does an allocation is preloaded.
1564	 *
1565	 * We do it in non-atomic context, thus it allows us to use more
1566	 * permissive allocation masks to be more stable under low memory
1567	 * condition and high memory pressure.
1568	 */
1569	if (!this_cpu_read(ne_fit_preload_node))
1570		va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1571
1572	spin_lock(lock);
1573
1574	if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
1575		kmem_cache_free(vmap_area_cachep, va);
1576}
1577
1578/*
1579 * Allocate a region of KVA of the specified size and alignment, within the
1580 * vstart and vend.
1581 */
1582static struct vmap_area *alloc_vmap_area(unsigned long size,
1583				unsigned long align,
1584				unsigned long vstart, unsigned long vend,
1585				int node, gfp_t gfp_mask,
1586				unsigned long va_flags)
1587{
1588	struct vmap_area *va;
1589	unsigned long freed;
1590	unsigned long addr;
1591	int purged = 0;
1592	int ret;
1593
1594	if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align)))
1595		return ERR_PTR(-EINVAL);
 
1596
1597	if (unlikely(!vmap_initialized))
1598		return ERR_PTR(-EBUSY);
1599
1600	might_sleep();
1601	gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1602
1603	va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1604	if (unlikely(!va))
1605		return ERR_PTR(-ENOMEM);
1606
1607	/*
1608	 * Only scan the relevant parts containing pointers to other objects
1609	 * to avoid false negatives.
1610	 */
1611	kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1612
1613retry:
1614	preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1615	addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1616		size, align, vstart, vend);
1617	spin_unlock(&free_vmap_area_lock);
1618
1619	trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
1620
1621	/*
1622	 * If an allocation fails, the "vend" address is
1623	 * returned. Therefore trigger the overflow path.
1624	 */
1625	if (unlikely(addr == vend))
1626		goto overflow;
1627
1628	va->va_start = addr;
1629	va->va_end = addr + size;
1630	va->vm = NULL;
1631	va->flags = va_flags;
1632
1633	spin_lock(&vmap_area_lock);
1634	insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1635	spin_unlock(&vmap_area_lock);
1636
1637	BUG_ON(!IS_ALIGNED(va->va_start, align));
1638	BUG_ON(va->va_start < vstart);
1639	BUG_ON(va->va_end > vend);
1640
1641	ret = kasan_populate_vmalloc(addr, size);
1642	if (ret) {
1643		free_vmap_area(va);
1644		return ERR_PTR(ret);
1645	}
1646
1647	return va;
1648
1649overflow:
1650	if (!purged) {
1651		reclaim_and_purge_vmap_areas();
1652		purged = 1;
1653		goto retry;
1654	}
1655
1656	freed = 0;
1657	blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1658
1659	if (freed > 0) {
1660		purged = 0;
1661		goto retry;
1662	}
1663
1664	if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1665		pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1666			size);
1667
1668	kmem_cache_free(vmap_area_cachep, va);
1669	return ERR_PTR(-EBUSY);
1670}
1671
1672int register_vmap_purge_notifier(struct notifier_block *nb)
1673{
1674	return blocking_notifier_chain_register(&vmap_notify_list, nb);
1675}
1676EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1677
1678int unregister_vmap_purge_notifier(struct notifier_block *nb)
1679{
1680	return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1681}
1682EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1683
1684/*
1685 * lazy_max_pages is the maximum amount of virtual address space we gather up
1686 * before attempting to purge with a TLB flush.
1687 *
1688 * There is a tradeoff here: a larger number will cover more kernel page tables
1689 * and take slightly longer to purge, but it will linearly reduce the number of
1690 * global TLB flushes that must be performed. It would seem natural to scale
1691 * this number up linearly with the number of CPUs (because vmapping activity
1692 * could also scale linearly with the number of CPUs), however it is likely
1693 * that in practice, workloads might be constrained in other ways that mean
1694 * vmap activity will not scale linearly with CPUs. Also, I want to be
1695 * conservative and not introduce a big latency on huge systems, so go with
1696 * a less aggressive log scale. It will still be an improvement over the old
1697 * code, and it will be simple to change the scale factor if we find that it
1698 * becomes a problem on bigger systems.
1699 */
1700static unsigned long lazy_max_pages(void)
1701{
1702	unsigned int log;
1703
1704	log = fls(num_online_cpus());
1705
1706	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1707}
1708
1709static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1710
1711/*
1712 * Serialize vmap purging.  There is no actual critical section protected
1713 * by this lock, but we want to avoid concurrent calls for performance
1714 * reasons and to make the pcpu_get_vm_areas more deterministic.
1715 */
1716static DEFINE_MUTEX(vmap_purge_lock);
1717
1718/* for per-CPU blocks */
1719static void purge_fragmented_blocks_allcpus(void);
1720
1721/*
1722 * Purges all lazily-freed vmap areas.
1723 */
1724static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1725{
1726	unsigned long resched_threshold;
1727	unsigned int num_purged_areas = 0;
1728	struct list_head local_purge_list;
1729	struct vmap_area *va, *n_va;
1730
1731	lockdep_assert_held(&vmap_purge_lock);
1732
1733	spin_lock(&purge_vmap_area_lock);
1734	purge_vmap_area_root = RB_ROOT;
1735	list_replace_init(&purge_vmap_area_list, &local_purge_list);
1736	spin_unlock(&purge_vmap_area_lock);
1737
1738	if (unlikely(list_empty(&local_purge_list)))
1739		goto out;
1740
1741	start = min(start,
1742		list_first_entry(&local_purge_list,
1743			struct vmap_area, list)->va_start);
1744
1745	end = max(end,
1746		list_last_entry(&local_purge_list,
1747			struct vmap_area, list)->va_end);
1748
1749	flush_tlb_kernel_range(start, end);
1750	resched_threshold = lazy_max_pages() << 1;
1751
1752	spin_lock(&free_vmap_area_lock);
1753	list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1754		unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1755		unsigned long orig_start = va->va_start;
1756		unsigned long orig_end = va->va_end;
1757
1758		/*
1759		 * Finally insert or merge lazily-freed area. It is
1760		 * detached and there is no need to "unlink" it from
1761		 * anything.
1762		 */
1763		va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1764				&free_vmap_area_list);
1765
1766		if (!va)
1767			continue;
1768
1769		if (is_vmalloc_or_module_addr((void *)orig_start))
1770			kasan_release_vmalloc(orig_start, orig_end,
1771					      va->va_start, va->va_end);
1772
1773		atomic_long_sub(nr, &vmap_lazy_nr);
1774		num_purged_areas++;
1775
1776		if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1777			cond_resched_lock(&free_vmap_area_lock);
1778	}
1779	spin_unlock(&free_vmap_area_lock);
1780
1781out:
1782	trace_purge_vmap_area_lazy(start, end, num_purged_areas);
1783	return num_purged_areas > 0;
1784}
1785
1786/*
1787 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
1788 */
1789static void reclaim_and_purge_vmap_areas(void)
1790
1791{
1792	mutex_lock(&vmap_purge_lock);
1793	purge_fragmented_blocks_allcpus();
1794	__purge_vmap_area_lazy(ULONG_MAX, 0);
1795	mutex_unlock(&vmap_purge_lock);
1796}
1797
1798static void drain_vmap_area_work(struct work_struct *work)
1799{
1800	unsigned long nr_lazy;
1801
1802	do {
1803		mutex_lock(&vmap_purge_lock);
1804		__purge_vmap_area_lazy(ULONG_MAX, 0);
1805		mutex_unlock(&vmap_purge_lock);
1806
1807		/* Recheck if further work is required. */
1808		nr_lazy = atomic_long_read(&vmap_lazy_nr);
1809	} while (nr_lazy > lazy_max_pages());
1810}
1811
1812/*
1813 * Free a vmap area, caller ensuring that the area has been unmapped,
1814 * unlinked and flush_cache_vunmap had been called for the correct
1815 * range previously.
1816 */
1817static void free_vmap_area_noflush(struct vmap_area *va)
1818{
1819	unsigned long nr_lazy_max = lazy_max_pages();
1820	unsigned long va_start = va->va_start;
1821	unsigned long nr_lazy;
1822
1823	if (WARN_ON_ONCE(!list_empty(&va->list)))
1824		return;
 
1825
1826	nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1827				PAGE_SHIFT, &vmap_lazy_nr);
1828
1829	/*
1830	 * Merge or place it to the purge tree/list.
1831	 */
1832	spin_lock(&purge_vmap_area_lock);
1833	merge_or_add_vmap_area(va,
1834		&purge_vmap_area_root, &purge_vmap_area_list);
1835	spin_unlock(&purge_vmap_area_lock);
1836
1837	trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
1838
1839	/* After this point, we may free va at any time */
1840	if (unlikely(nr_lazy > nr_lazy_max))
1841		schedule_work(&drain_vmap_work);
1842}
1843
1844/*
1845 * Free and unmap a vmap area
1846 */
1847static void free_unmap_vmap_area(struct vmap_area *va)
1848{
1849	flush_cache_vunmap(va->va_start, va->va_end);
1850	vunmap_range_noflush(va->va_start, va->va_end);
1851	if (debug_pagealloc_enabled_static())
1852		flush_tlb_kernel_range(va->va_start, va->va_end);
1853
1854	free_vmap_area_noflush(va);
1855}
1856
1857struct vmap_area *find_vmap_area(unsigned long addr)
1858{
1859	struct vmap_area *va;
1860
1861	spin_lock(&vmap_area_lock);
1862	va = __find_vmap_area(addr, &vmap_area_root);
1863	spin_unlock(&vmap_area_lock);
1864
1865	return va;
1866}
1867
1868static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
1869{
1870	struct vmap_area *va;
1871
1872	spin_lock(&vmap_area_lock);
1873	va = __find_vmap_area(addr, &vmap_area_root);
1874	if (va)
1875		unlink_va(va, &vmap_area_root);
1876	spin_unlock(&vmap_area_lock);
1877
1878	return va;
1879}
1880
1881/*** Per cpu kva allocator ***/
1882
1883/*
1884 * vmap space is limited especially on 32 bit architectures. Ensure there is
1885 * room for at least 16 percpu vmap blocks per CPU.
1886 */
1887/*
1888 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1889 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
1890 * instead (we just need a rough idea)
1891 */
1892#if BITS_PER_LONG == 32
1893#define VMALLOC_SPACE		(128UL*1024*1024)
1894#else
1895#define VMALLOC_SPACE		(128UL*1024*1024*1024)
1896#endif
1897
1898#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
1899#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
1900#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
1901#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
1902#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
1903#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
1904#define VMAP_BBMAP_BITS		\
1905		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
1906		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
1907			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1908
1909#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
1910
1911/*
1912 * Purge threshold to prevent overeager purging of fragmented blocks for
1913 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
1914 */
1915#define VMAP_PURGE_THRESHOLD	(VMAP_BBMAP_BITS / 4)
1916
1917#define VMAP_RAM		0x1 /* indicates vm_map_ram area*/
1918#define VMAP_BLOCK		0x2 /* mark out the vmap_block sub-type*/
1919#define VMAP_FLAGS_MASK		0x3
1920
1921struct vmap_block_queue {
1922	spinlock_t lock;
1923	struct list_head free;
1924
1925	/*
1926	 * An xarray requires an extra memory dynamically to
1927	 * be allocated. If it is an issue, we can use rb-tree
1928	 * instead.
1929	 */
1930	struct xarray vmap_blocks;
1931};
1932
1933struct vmap_block {
1934	spinlock_t lock;
1935	struct vmap_area *va;
1936	unsigned long free, dirty;
1937	DECLARE_BITMAP(used_map, VMAP_BBMAP_BITS);
1938	unsigned long dirty_min, dirty_max; /*< dirty range */
1939	struct list_head free_list;
1940	struct rcu_head rcu_head;
1941	struct list_head purge;
1942};
1943
1944/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1945static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1946
1947/*
1948 * In order to fast access to any "vmap_block" associated with a
1949 * specific address, we use a hash.
1950 *
1951 * A per-cpu vmap_block_queue is used in both ways, to serialize
1952 * an access to free block chains among CPUs(alloc path) and it
1953 * also acts as a vmap_block hash(alloc/free paths). It means we
1954 * overload it, since we already have the per-cpu array which is
1955 * used as a hash table. When used as a hash a 'cpu' passed to
1956 * per_cpu() is not actually a CPU but rather a hash index.
1957 *
1958 * A hash function is addr_to_vb_xa() which hashes any address
1959 * to a specific index(in a hash) it belongs to. This then uses a
1960 * per_cpu() macro to access an array with generated index.
1961 *
1962 * An example:
1963 *
1964 *  CPU_1  CPU_2  CPU_0
1965 *    |      |      |
1966 *    V      V      V
1967 * 0     10     20     30     40     50     60
1968 * |------|------|------|------|------|------|...<vmap address space>
1969 *   CPU0   CPU1   CPU2   CPU0   CPU1   CPU2
1970 *
1971 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
1972 *   it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
1973 *
1974 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
1975 *   it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
1976 *
1977 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
1978 *   it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
1979 *
1980 * This technique almost always avoids lock contention on insert/remove,
1981 * however xarray spinlocks protect against any contention that remains.
1982 */
1983static struct xarray *
1984addr_to_vb_xa(unsigned long addr)
1985{
1986	int index = (addr / VMAP_BLOCK_SIZE) % num_possible_cpus();
1987
1988	return &per_cpu(vmap_block_queue, index).vmap_blocks;
1989}
1990
1991/*
1992 * We should probably have a fallback mechanism to allocate virtual memory
1993 * out of partially filled vmap blocks. However vmap block sizing should be
1994 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1995 * big problem.
1996 */
1997
1998static unsigned long addr_to_vb_idx(unsigned long addr)
1999{
2000	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
2001	addr /= VMAP_BLOCK_SIZE;
2002	return addr;
2003}
2004
2005static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
2006{
2007	unsigned long addr;
2008
2009	addr = va_start + (pages_off << PAGE_SHIFT);
2010	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
2011	return (void *)addr;
2012}
2013
2014/**
2015 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2016 *                  block. Of course pages number can't exceed VMAP_BBMAP_BITS
2017 * @order:    how many 2^order pages should be occupied in newly allocated block
2018 * @gfp_mask: flags for the page level allocator
2019 *
2020 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2021 */
2022static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
2023{
2024	struct vmap_block_queue *vbq;
2025	struct vmap_block *vb;
2026	struct vmap_area *va;
2027	struct xarray *xa;
2028	unsigned long vb_idx;
2029	int node, err;
2030	void *vaddr;
2031
2032	node = numa_node_id();
2033
2034	vb = kmalloc_node(sizeof(struct vmap_block),
2035			gfp_mask & GFP_RECLAIM_MASK, node);
2036	if (unlikely(!vb))
2037		return ERR_PTR(-ENOMEM);
2038
2039	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
2040					VMALLOC_START, VMALLOC_END,
2041					node, gfp_mask,
2042					VMAP_RAM|VMAP_BLOCK);
2043	if (IS_ERR(va)) {
2044		kfree(vb);
2045		return ERR_CAST(va);
2046	}
2047
2048	vaddr = vmap_block_vaddr(va->va_start, 0);
2049	spin_lock_init(&vb->lock);
2050	vb->va = va;
2051	/* At least something should be left free */
2052	BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
2053	bitmap_zero(vb->used_map, VMAP_BBMAP_BITS);
2054	vb->free = VMAP_BBMAP_BITS - (1UL << order);
2055	vb->dirty = 0;
2056	vb->dirty_min = VMAP_BBMAP_BITS;
2057	vb->dirty_max = 0;
2058	bitmap_set(vb->used_map, 0, (1UL << order));
2059	INIT_LIST_HEAD(&vb->free_list);
2060
2061	xa = addr_to_vb_xa(va->va_start);
2062	vb_idx = addr_to_vb_idx(va->va_start);
2063	err = xa_insert(xa, vb_idx, vb, gfp_mask);
2064	if (err) {
2065		kfree(vb);
2066		free_vmap_area(va);
2067		return ERR_PTR(err);
2068	}
2069
2070	vbq = raw_cpu_ptr(&vmap_block_queue);
2071	spin_lock(&vbq->lock);
2072	list_add_tail_rcu(&vb->free_list, &vbq->free);
2073	spin_unlock(&vbq->lock);
2074
2075	return vaddr;
2076}
2077
2078static void free_vmap_block(struct vmap_block *vb)
2079{
2080	struct vmap_block *tmp;
2081	struct xarray *xa;
2082
2083	xa = addr_to_vb_xa(vb->va->va_start);
2084	tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start));
2085	BUG_ON(tmp != vb);
2086
2087	spin_lock(&vmap_area_lock);
2088	unlink_va(vb->va, &vmap_area_root);
2089	spin_unlock(&vmap_area_lock);
2090
2091	free_vmap_area_noflush(vb->va);
2092	kfree_rcu(vb, rcu_head);
2093}
2094
2095static bool purge_fragmented_block(struct vmap_block *vb,
2096		struct vmap_block_queue *vbq, struct list_head *purge_list,
2097		bool force_purge)
2098{
2099	if (vb->free + vb->dirty != VMAP_BBMAP_BITS ||
2100	    vb->dirty == VMAP_BBMAP_BITS)
2101		return false;
2102
2103	/* Don't overeagerly purge usable blocks unless requested */
2104	if (!(force_purge || vb->free < VMAP_PURGE_THRESHOLD))
2105		return false;
2106
2107	/* prevent further allocs after releasing lock */
2108	WRITE_ONCE(vb->free, 0);
2109	/* prevent purging it again */
2110	WRITE_ONCE(vb->dirty, VMAP_BBMAP_BITS);
2111	vb->dirty_min = 0;
2112	vb->dirty_max = VMAP_BBMAP_BITS;
2113	spin_lock(&vbq->lock);
2114	list_del_rcu(&vb->free_list);
2115	spin_unlock(&vbq->lock);
2116	list_add_tail(&vb->purge, purge_list);
2117	return true;
2118}
2119
2120static void free_purged_blocks(struct list_head *purge_list)
2121{
2122	struct vmap_block *vb, *n_vb;
2123
2124	list_for_each_entry_safe(vb, n_vb, purge_list, purge) {
2125		list_del(&vb->purge);
2126		free_vmap_block(vb);
2127	}
2128}
2129
2130static void purge_fragmented_blocks(int cpu)
2131{
2132	LIST_HEAD(purge);
2133	struct vmap_block *vb;
 
2134	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2135
2136	rcu_read_lock();
2137	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2138		unsigned long free = READ_ONCE(vb->free);
2139		unsigned long dirty = READ_ONCE(vb->dirty);
2140
2141		if (free + dirty != VMAP_BBMAP_BITS ||
2142		    dirty == VMAP_BBMAP_BITS)
2143			continue;
2144
2145		spin_lock(&vb->lock);
2146		purge_fragmented_block(vb, vbq, &purge, true);
2147		spin_unlock(&vb->lock);
 
 
 
 
 
 
 
 
 
 
2148	}
2149	rcu_read_unlock();
2150	free_purged_blocks(&purge);
 
 
 
 
2151}
2152
2153static void purge_fragmented_blocks_allcpus(void)
2154{
2155	int cpu;
2156
2157	for_each_possible_cpu(cpu)
2158		purge_fragmented_blocks(cpu);
2159}
2160
2161static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2162{
2163	struct vmap_block_queue *vbq;
2164	struct vmap_block *vb;
2165	void *vaddr = NULL;
2166	unsigned int order;
2167
2168	BUG_ON(offset_in_page(size));
2169	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2170	if (WARN_ON(size == 0)) {
2171		/*
2172		 * Allocating 0 bytes isn't what caller wants since
2173		 * get_order(0) returns funny result. Just warn and terminate
2174		 * early.
2175		 */
2176		return NULL;
2177	}
2178	order = get_order(size);
2179
2180	rcu_read_lock();
2181	vbq = raw_cpu_ptr(&vmap_block_queue);
2182	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2183		unsigned long pages_off;
2184
2185		if (READ_ONCE(vb->free) < (1UL << order))
2186			continue;
2187
2188		spin_lock(&vb->lock);
2189		if (vb->free < (1UL << order)) {
2190			spin_unlock(&vb->lock);
2191			continue;
2192		}
2193
2194		pages_off = VMAP_BBMAP_BITS - vb->free;
2195		vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2196		WRITE_ONCE(vb->free, vb->free - (1UL << order));
2197		bitmap_set(vb->used_map, pages_off, (1UL << order));
2198		if (vb->free == 0) {
2199			spin_lock(&vbq->lock);
2200			list_del_rcu(&vb->free_list);
2201			spin_unlock(&vbq->lock);
2202		}
2203
2204		spin_unlock(&vb->lock);
2205		break;
2206	}
2207
2208	rcu_read_unlock();
2209
2210	/* Allocate new block if nothing was found */
2211	if (!vaddr)
2212		vaddr = new_vmap_block(order, gfp_mask);
2213
2214	return vaddr;
2215}
2216
2217static void vb_free(unsigned long addr, unsigned long size)
2218{
2219	unsigned long offset;
2220	unsigned int order;
2221	struct vmap_block *vb;
2222	struct xarray *xa;
2223
2224	BUG_ON(offset_in_page(size));
2225	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2226
2227	flush_cache_vunmap(addr, addr + size);
2228
2229	order = get_order(size);
2230	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2231
2232	xa = addr_to_vb_xa(addr);
2233	vb = xa_load(xa, addr_to_vb_idx(addr));
2234
2235	spin_lock(&vb->lock);
2236	bitmap_clear(vb->used_map, offset, (1UL << order));
2237	spin_unlock(&vb->lock);
2238
2239	vunmap_range_noflush(addr, addr + size);
2240
2241	if (debug_pagealloc_enabled_static())
2242		flush_tlb_kernel_range(addr, addr + size);
2243
2244	spin_lock(&vb->lock);
2245
2246	/* Expand the not yet TLB flushed dirty range */
2247	vb->dirty_min = min(vb->dirty_min, offset);
2248	vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2249
2250	WRITE_ONCE(vb->dirty, vb->dirty + (1UL << order));
2251	if (vb->dirty == VMAP_BBMAP_BITS) {
2252		BUG_ON(vb->free);
2253		spin_unlock(&vb->lock);
2254		free_vmap_block(vb);
2255	} else
2256		spin_unlock(&vb->lock);
2257}
2258
2259static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2260{
2261	LIST_HEAD(purge_list);
2262	int cpu;
2263
2264	if (unlikely(!vmap_initialized))
2265		return;
2266
2267	mutex_lock(&vmap_purge_lock);
2268
2269	for_each_possible_cpu(cpu) {
2270		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2271		struct vmap_block *vb;
2272		unsigned long idx;
2273
2274		rcu_read_lock();
2275		xa_for_each(&vbq->vmap_blocks, idx, vb) {
2276			spin_lock(&vb->lock);
2277
2278			/*
2279			 * Try to purge a fragmented block first. If it's
2280			 * not purgeable, check whether there is dirty
2281			 * space to be flushed.
2282			 */
2283			if (!purge_fragmented_block(vb, vbq, &purge_list, false) &&
2284			    vb->dirty_max && vb->dirty != VMAP_BBMAP_BITS) {
2285				unsigned long va_start = vb->va->va_start;
2286				unsigned long s, e;
2287
2288				s = va_start + (vb->dirty_min << PAGE_SHIFT);
2289				e = va_start + (vb->dirty_max << PAGE_SHIFT);
2290
2291				start = min(s, start);
2292				end   = max(e, end);
2293
2294				/* Prevent that this is flushed again */
2295				vb->dirty_min = VMAP_BBMAP_BITS;
2296				vb->dirty_max = 0;
2297
2298				flush = 1;
2299			}
2300			spin_unlock(&vb->lock);
2301		}
2302		rcu_read_unlock();
2303	}
2304	free_purged_blocks(&purge_list);
2305
 
 
2306	if (!__purge_vmap_area_lazy(start, end) && flush)
2307		flush_tlb_kernel_range(start, end);
2308	mutex_unlock(&vmap_purge_lock);
2309}
2310
2311/**
2312 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2313 *
2314 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2315 * to amortize TLB flushing overheads. What this means is that any page you
2316 * have now, may, in a former life, have been mapped into kernel virtual
2317 * address by the vmap layer and so there might be some CPUs with TLB entries
2318 * still referencing that page (additional to the regular 1:1 kernel mapping).
2319 *
2320 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2321 * be sure that none of the pages we have control over will have any aliases
2322 * from the vmap layer.
2323 */
2324void vm_unmap_aliases(void)
2325{
2326	unsigned long start = ULONG_MAX, end = 0;
2327	int flush = 0;
2328
2329	_vm_unmap_aliases(start, end, flush);
2330}
2331EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2332
2333/**
2334 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2335 * @mem: the pointer returned by vm_map_ram
2336 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2337 */
2338void vm_unmap_ram(const void *mem, unsigned int count)
2339{
2340	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2341	unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2342	struct vmap_area *va;
2343
2344	might_sleep();
2345	BUG_ON(!addr);
2346	BUG_ON(addr < VMALLOC_START);
2347	BUG_ON(addr > VMALLOC_END);
2348	BUG_ON(!PAGE_ALIGNED(addr));
2349
2350	kasan_poison_vmalloc(mem, size);
2351
2352	if (likely(count <= VMAP_MAX_ALLOC)) {
2353		debug_check_no_locks_freed(mem, size);
2354		vb_free(addr, size);
2355		return;
2356	}
2357
2358	va = find_unlink_vmap_area(addr);
2359	if (WARN_ON_ONCE(!va))
2360		return;
2361
2362	debug_check_no_locks_freed((void *)va->va_start,
2363				    (va->va_end - va->va_start));
2364	free_unmap_vmap_area(va);
2365}
2366EXPORT_SYMBOL(vm_unmap_ram);
2367
2368/**
2369 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2370 * @pages: an array of pointers to the pages to be mapped
2371 * @count: number of pages
2372 * @node: prefer to allocate data structures on this node
2373 *
2374 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2375 * faster than vmap so it's good.  But if you mix long-life and short-life
2376 * objects with vm_map_ram(), it could consume lots of address space through
2377 * fragmentation (especially on a 32bit machine).  You could see failures in
2378 * the end.  Please use this function for short-lived objects.
2379 *
2380 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2381 */
2382void *vm_map_ram(struct page **pages, unsigned int count, int node)
2383{
2384	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2385	unsigned long addr;
2386	void *mem;
2387
2388	if (likely(count <= VMAP_MAX_ALLOC)) {
2389		mem = vb_alloc(size, GFP_KERNEL);
2390		if (IS_ERR(mem))
2391			return NULL;
2392		addr = (unsigned long)mem;
2393	} else {
2394		struct vmap_area *va;
2395		va = alloc_vmap_area(size, PAGE_SIZE,
2396				VMALLOC_START, VMALLOC_END,
2397				node, GFP_KERNEL, VMAP_RAM);
2398		if (IS_ERR(va))
2399			return NULL;
2400
2401		addr = va->va_start;
2402		mem = (void *)addr;
2403	}
2404
2405	if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2406				pages, PAGE_SHIFT) < 0) {
2407		vm_unmap_ram(mem, count);
2408		return NULL;
2409	}
2410
2411	/*
2412	 * Mark the pages as accessible, now that they are mapped.
2413	 * With hardware tag-based KASAN, marking is skipped for
2414	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2415	 */
2416	mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2417
2418	return mem;
2419}
2420EXPORT_SYMBOL(vm_map_ram);
2421
2422static struct vm_struct *vmlist __initdata;
2423
2424static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2425{
2426#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2427	return vm->page_order;
2428#else
2429	return 0;
2430#endif
2431}
2432
2433static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2434{
2435#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2436	vm->page_order = order;
2437#else
2438	BUG_ON(order != 0);
2439#endif
2440}
2441
2442/**
2443 * vm_area_add_early - add vmap area early during boot
2444 * @vm: vm_struct to add
2445 *
2446 * This function is used to add fixed kernel vm area to vmlist before
2447 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
2448 * should contain proper values and the other fields should be zero.
2449 *
2450 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2451 */
2452void __init vm_area_add_early(struct vm_struct *vm)
2453{
2454	struct vm_struct *tmp, **p;
2455
2456	BUG_ON(vmap_initialized);
2457	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2458		if (tmp->addr >= vm->addr) {
2459			BUG_ON(tmp->addr < vm->addr + vm->size);
2460			break;
2461		} else
2462			BUG_ON(tmp->addr + tmp->size > vm->addr);
2463	}
2464	vm->next = *p;
2465	*p = vm;
2466}
2467
2468/**
2469 * vm_area_register_early - register vmap area early during boot
2470 * @vm: vm_struct to register
2471 * @align: requested alignment
2472 *
2473 * This function is used to register kernel vm area before
2474 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
2475 * proper values on entry and other fields should be zero.  On return,
2476 * vm->addr contains the allocated address.
2477 *
2478 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2479 */
2480void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2481{
2482	unsigned long addr = ALIGN(VMALLOC_START, align);
2483	struct vm_struct *cur, **p;
2484
2485	BUG_ON(vmap_initialized);
2486
2487	for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2488		if ((unsigned long)cur->addr - addr >= vm->size)
2489			break;
2490		addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2491	}
2492
2493	BUG_ON(addr > VMALLOC_END - vm->size);
2494	vm->addr = (void *)addr;
2495	vm->next = *p;
2496	*p = vm;
2497	kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2498}
2499
2500static void vmap_init_free_space(void)
2501{
2502	unsigned long vmap_start = 1;
2503	const unsigned long vmap_end = ULONG_MAX;
2504	struct vmap_area *busy, *free;
2505
2506	/*
2507	 *     B     F     B     B     B     F
2508	 * -|-----|.....|-----|-----|-----|.....|-
2509	 *  |           The KVA space           |
2510	 *  |<--------------------------------->|
2511	 */
2512	list_for_each_entry(busy, &vmap_area_list, list) {
2513		if (busy->va_start - vmap_start > 0) {
2514			free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2515			if (!WARN_ON_ONCE(!free)) {
2516				free->va_start = vmap_start;
2517				free->va_end = busy->va_start;
2518
2519				insert_vmap_area_augment(free, NULL,
2520					&free_vmap_area_root,
2521						&free_vmap_area_list);
2522			}
2523		}
2524
2525		vmap_start = busy->va_end;
2526	}
2527
2528	if (vmap_end - vmap_start > 0) {
2529		free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2530		if (!WARN_ON_ONCE(!free)) {
2531			free->va_start = vmap_start;
2532			free->va_end = vmap_end;
2533
2534			insert_vmap_area_augment(free, NULL,
2535				&free_vmap_area_root,
2536					&free_vmap_area_list);
2537		}
2538	}
2539}
2540
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2541static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2542	struct vmap_area *va, unsigned long flags, const void *caller)
2543{
2544	vm->flags = flags;
2545	vm->addr = (void *)va->va_start;
2546	vm->size = va->va_end - va->va_start;
2547	vm->caller = caller;
2548	va->vm = vm;
2549}
2550
2551static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2552			      unsigned long flags, const void *caller)
2553{
2554	spin_lock(&vmap_area_lock);
2555	setup_vmalloc_vm_locked(vm, va, flags, caller);
2556	spin_unlock(&vmap_area_lock);
2557}
2558
2559static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2560{
2561	/*
2562	 * Before removing VM_UNINITIALIZED,
2563	 * we should make sure that vm has proper values.
2564	 * Pair with smp_rmb() in show_numa_info().
2565	 */
2566	smp_wmb();
2567	vm->flags &= ~VM_UNINITIALIZED;
2568}
2569
2570static struct vm_struct *__get_vm_area_node(unsigned long size,
2571		unsigned long align, unsigned long shift, unsigned long flags,
2572		unsigned long start, unsigned long end, int node,
2573		gfp_t gfp_mask, const void *caller)
2574{
2575	struct vmap_area *va;
2576	struct vm_struct *area;
2577	unsigned long requested_size = size;
2578
2579	BUG_ON(in_interrupt());
2580	size = ALIGN(size, 1ul << shift);
2581	if (unlikely(!size))
2582		return NULL;
2583
2584	if (flags & VM_IOREMAP)
2585		align = 1ul << clamp_t(int, get_count_order_long(size),
2586				       PAGE_SHIFT, IOREMAP_MAX_ORDER);
2587
2588	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2589	if (unlikely(!area))
2590		return NULL;
2591
2592	if (!(flags & VM_NO_GUARD))
2593		size += PAGE_SIZE;
2594
2595	va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0);
2596	if (IS_ERR(va)) {
2597		kfree(area);
2598		return NULL;
2599	}
2600
2601	setup_vmalloc_vm(area, va, flags, caller);
2602
2603	/*
2604	 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2605	 * best-effort approach, as they can be mapped outside of vmalloc code.
2606	 * For VM_ALLOC mappings, the pages are marked as accessible after
2607	 * getting mapped in __vmalloc_node_range().
2608	 * With hardware tag-based KASAN, marking is skipped for
2609	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2610	 */
2611	if (!(flags & VM_ALLOC))
2612		area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2613						    KASAN_VMALLOC_PROT_NORMAL);
2614
2615	return area;
2616}
2617
2618struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2619				       unsigned long start, unsigned long end,
2620				       const void *caller)
2621{
2622	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2623				  NUMA_NO_NODE, GFP_KERNEL, caller);
2624}
2625
2626/**
2627 * get_vm_area - reserve a contiguous kernel virtual area
2628 * @size:	 size of the area
2629 * @flags:	 %VM_IOREMAP for I/O mappings or VM_ALLOC
2630 *
2631 * Search an area of @size in the kernel virtual mapping area,
2632 * and reserved it for out purposes.  Returns the area descriptor
2633 * on success or %NULL on failure.
2634 *
2635 * Return: the area descriptor on success or %NULL on failure.
2636 */
2637struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2638{
2639	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2640				  VMALLOC_START, VMALLOC_END,
2641				  NUMA_NO_NODE, GFP_KERNEL,
2642				  __builtin_return_address(0));
2643}
2644
2645struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2646				const void *caller)
2647{
2648	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2649				  VMALLOC_START, VMALLOC_END,
2650				  NUMA_NO_NODE, GFP_KERNEL, caller);
2651}
2652
2653/**
2654 * find_vm_area - find a continuous kernel virtual area
2655 * @addr:	  base address
2656 *
2657 * Search for the kernel VM area starting at @addr, and return it.
2658 * It is up to the caller to do all required locking to keep the returned
2659 * pointer valid.
2660 *
2661 * Return: the area descriptor on success or %NULL on failure.
2662 */
2663struct vm_struct *find_vm_area(const void *addr)
2664{
2665	struct vmap_area *va;
2666
2667	va = find_vmap_area((unsigned long)addr);
2668	if (!va)
2669		return NULL;
2670
2671	return va->vm;
2672}
2673
2674/**
2675 * remove_vm_area - find and remove a continuous kernel virtual area
2676 * @addr:	    base address
2677 *
2678 * Search for the kernel VM area starting at @addr, and remove it.
2679 * This function returns the found VM area, but using it is NOT safe
2680 * on SMP machines, except for its size or flags.
2681 *
2682 * Return: the area descriptor on success or %NULL on failure.
2683 */
2684struct vm_struct *remove_vm_area(const void *addr)
2685{
2686	struct vmap_area *va;
2687	struct vm_struct *vm;
2688
2689	might_sleep();
2690
2691	if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2692			addr))
2693		return NULL;
 
2694
2695	va = find_unlink_vmap_area((unsigned long)addr);
2696	if (!va || !va->vm)
2697		return NULL;
2698	vm = va->vm;
2699
2700	debug_check_no_locks_freed(vm->addr, get_vm_area_size(vm));
2701	debug_check_no_obj_freed(vm->addr, get_vm_area_size(vm));
2702	kasan_free_module_shadow(vm);
2703	kasan_poison_vmalloc(vm->addr, get_vm_area_size(vm));
2704
2705	free_unmap_vmap_area(va);
2706	return vm;
 
 
 
2707}
2708
2709static inline void set_area_direct_map(const struct vm_struct *area,
2710				       int (*set_direct_map)(struct page *page))
2711{
2712	int i;
2713
2714	/* HUGE_VMALLOC passes small pages to set_direct_map */
2715	for (i = 0; i < area->nr_pages; i++)
2716		if (page_address(area->pages[i]))
2717			set_direct_map(area->pages[i]);
2718}
2719
2720/*
2721 * Flush the vm mapping and reset the direct map.
2722 */
2723static void vm_reset_perms(struct vm_struct *area)
2724{
2725	unsigned long start = ULONG_MAX, end = 0;
2726	unsigned int page_order = vm_area_page_order(area);
 
2727	int flush_dmap = 0;
2728	int i;
2729
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2730	/*
2731	 * Find the start and end range of the direct mappings to make sure that
 
2732	 * the vm_unmap_aliases() flush includes the direct map.
2733	 */
2734	for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2735		unsigned long addr = (unsigned long)page_address(area->pages[i]);
2736
2737		if (addr) {
2738			unsigned long page_size;
2739
2740			page_size = PAGE_SIZE << page_order;
2741			start = min(addr, start);
2742			end = max(addr + page_size, end);
2743			flush_dmap = 1;
2744		}
2745	}
2746
2747	/*
2748	 * Set direct map to something invalid so that it won't be cached if
2749	 * there are any accesses after the TLB flush, then flush the TLB and
2750	 * reset the direct map permissions to the default.
2751	 */
2752	set_area_direct_map(area, set_direct_map_invalid_noflush);
2753	_vm_unmap_aliases(start, end, flush_dmap);
2754	set_area_direct_map(area, set_direct_map_default_noflush);
2755}
2756
2757static void delayed_vfree_work(struct work_struct *w)
2758{
2759	struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
2760	struct llist_node *t, *llnode;
2761
2762	llist_for_each_safe(llnode, t, llist_del_all(&p->list))
2763		vfree(llnode);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2764}
2765
2766/**
2767 * vfree_atomic - release memory allocated by vmalloc()
2768 * @addr:	  memory base address
2769 *
2770 * This one is just like vfree() but can be called in any atomic context
2771 * except NMIs.
2772 */
2773void vfree_atomic(const void *addr)
2774{
2775	struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2776
2777	BUG_ON(in_nmi());
 
2778	kmemleak_free(addr);
2779
2780	/*
2781	 * Use raw_cpu_ptr() because this can be called from preemptible
2782	 * context. Preemption is absolutely fine here, because the llist_add()
2783	 * implementation is lockless, so it works even if we are adding to
2784	 * another cpu's list. schedule_work() should be fine with this too.
2785	 */
2786	if (addr && llist_add((struct llist_node *)addr, &p->list))
2787		schedule_work(&p->wq);
 
 
 
2788}
2789
2790/**
2791 * vfree - Release memory allocated by vmalloc()
2792 * @addr:  Memory base address
2793 *
2794 * Free the virtually continuous memory area starting at @addr, as obtained
2795 * from one of the vmalloc() family of APIs.  This will usually also free the
2796 * physical memory underlying the virtual allocation, but that memory is
2797 * reference counted, so it will not be freed until the last user goes away.
2798 *
2799 * If @addr is NULL, no operation is performed.
2800 *
2801 * Context:
2802 * May sleep if called *not* from interrupt context.
2803 * Must not be called in NMI context (strictly speaking, it could be
2804 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2805 * conventions for vfree() arch-dependent would be a really bad idea).
2806 */
2807void vfree(const void *addr)
2808{
2809	struct vm_struct *vm;
2810	int i;
2811
2812	if (unlikely(in_interrupt())) {
2813		vfree_atomic(addr);
2814		return;
2815	}
2816
2817	BUG_ON(in_nmi());
 
2818	kmemleak_free(addr);
2819	might_sleep();
2820
2821	if (!addr)
2822		return;
2823
2824	vm = remove_vm_area(addr);
2825	if (unlikely(!vm)) {
2826		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2827				addr);
2828		return;
2829	}
2830
2831	if (unlikely(vm->flags & VM_FLUSH_RESET_PERMS))
2832		vm_reset_perms(vm);
2833	for (i = 0; i < vm->nr_pages; i++) {
2834		struct page *page = vm->pages[i];
2835
2836		BUG_ON(!page);
2837		mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2838		/*
2839		 * High-order allocs for huge vmallocs are split, so
2840		 * can be freed as an array of order-0 allocations
2841		 */
2842		__free_page(page);
2843		cond_resched();
2844	}
2845	atomic_long_sub(vm->nr_pages, &nr_vmalloc_pages);
2846	kvfree(vm->pages);
2847	kfree(vm);
2848}
2849EXPORT_SYMBOL(vfree);
2850
2851/**
2852 * vunmap - release virtual mapping obtained by vmap()
2853 * @addr:   memory base address
2854 *
2855 * Free the virtually contiguous memory area starting at @addr,
2856 * which was created from the page array passed to vmap().
2857 *
2858 * Must not be called in interrupt context.
2859 */
2860void vunmap(const void *addr)
2861{
2862	struct vm_struct *vm;
2863
2864	BUG_ON(in_interrupt());
2865	might_sleep();
2866
2867	if (!addr)
2868		return;
2869	vm = remove_vm_area(addr);
2870	if (unlikely(!vm)) {
2871		WARN(1, KERN_ERR "Trying to vunmap() nonexistent vm area (%p)\n",
2872				addr);
2873		return;
2874	}
2875	kfree(vm);
2876}
2877EXPORT_SYMBOL(vunmap);
2878
2879/**
2880 * vmap - map an array of pages into virtually contiguous space
2881 * @pages: array of page pointers
2882 * @count: number of pages to map
2883 * @flags: vm_area->flags
2884 * @prot: page protection for the mapping
2885 *
2886 * Maps @count pages from @pages into contiguous kernel virtual space.
2887 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2888 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2889 * are transferred from the caller to vmap(), and will be freed / dropped when
2890 * vfree() is called on the return value.
2891 *
2892 * Return: the address of the area or %NULL on failure
2893 */
2894void *vmap(struct page **pages, unsigned int count,
2895	   unsigned long flags, pgprot_t prot)
2896{
2897	struct vm_struct *area;
2898	unsigned long addr;
2899	unsigned long size;		/* In bytes */
2900
2901	might_sleep();
2902
2903	if (WARN_ON_ONCE(flags & VM_FLUSH_RESET_PERMS))
2904		return NULL;
2905
2906	/*
2907	 * Your top guard is someone else's bottom guard. Not having a top
2908	 * guard compromises someone else's mappings too.
2909	 */
2910	if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2911		flags &= ~VM_NO_GUARD;
2912
2913	if (count > totalram_pages())
2914		return NULL;
2915
2916	size = (unsigned long)count << PAGE_SHIFT;
2917	area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2918	if (!area)
2919		return NULL;
2920
2921	addr = (unsigned long)area->addr;
2922	if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2923				pages, PAGE_SHIFT) < 0) {
2924		vunmap(area->addr);
2925		return NULL;
2926	}
2927
2928	if (flags & VM_MAP_PUT_PAGES) {
2929		area->pages = pages;
2930		area->nr_pages = count;
2931	}
2932	return area->addr;
2933}
2934EXPORT_SYMBOL(vmap);
2935
2936#ifdef CONFIG_VMAP_PFN
2937struct vmap_pfn_data {
2938	unsigned long	*pfns;
2939	pgprot_t	prot;
2940	unsigned int	idx;
2941};
2942
2943static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2944{
2945	struct vmap_pfn_data *data = private;
2946	unsigned long pfn = data->pfns[data->idx];
2947	pte_t ptent;
2948
2949	if (WARN_ON_ONCE(pfn_valid(pfn)))
2950		return -EINVAL;
2951
2952	ptent = pte_mkspecial(pfn_pte(pfn, data->prot));
2953	set_pte_at(&init_mm, addr, pte, ptent);
2954
2955	data->idx++;
2956	return 0;
2957}
2958
2959/**
2960 * vmap_pfn - map an array of PFNs into virtually contiguous space
2961 * @pfns: array of PFNs
2962 * @count: number of pages to map
2963 * @prot: page protection for the mapping
2964 *
2965 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2966 * the start address of the mapping.
2967 */
2968void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2969{
2970	struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2971	struct vm_struct *area;
2972
2973	area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2974			__builtin_return_address(0));
2975	if (!area)
2976		return NULL;
2977	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2978			count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2979		free_vm_area(area);
2980		return NULL;
2981	}
2982
2983	flush_cache_vmap((unsigned long)area->addr,
2984			 (unsigned long)area->addr + count * PAGE_SIZE);
2985
2986	return area->addr;
2987}
2988EXPORT_SYMBOL_GPL(vmap_pfn);
2989#endif /* CONFIG_VMAP_PFN */
2990
2991static inline unsigned int
2992vm_area_alloc_pages(gfp_t gfp, int nid,
2993		unsigned int order, unsigned int nr_pages, struct page **pages)
2994{
2995	unsigned int nr_allocated = 0;
2996	gfp_t alloc_gfp = gfp;
2997	bool nofail = false;
2998	struct page *page;
2999	int i;
3000
3001	/*
3002	 * For order-0 pages we make use of bulk allocator, if
3003	 * the page array is partly or not at all populated due
3004	 * to fails, fallback to a single page allocator that is
3005	 * more permissive.
3006	 */
3007	if (!order) {
3008		/* bulk allocator doesn't support nofail req. officially */
3009		gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
3010
3011		while (nr_allocated < nr_pages) {
3012			unsigned int nr, nr_pages_request;
3013
3014			/*
3015			 * A maximum allowed request is hard-coded and is 100
3016			 * pages per call. That is done in order to prevent a
3017			 * long preemption off scenario in the bulk-allocator
3018			 * so the range is [1:100].
3019			 */
3020			nr_pages_request = min(100U, nr_pages - nr_allocated);
3021
3022			/* memory allocation should consider mempolicy, we can't
3023			 * wrongly use nearest node when nid == NUMA_NO_NODE,
3024			 * otherwise memory may be allocated in only one node,
3025			 * but mempolicy wants to alloc memory by interleaving.
3026			 */
3027			if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
3028				nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
3029							nr_pages_request,
3030							pages + nr_allocated);
3031
3032			else
3033				nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
3034							nr_pages_request,
3035							pages + nr_allocated);
3036
3037			nr_allocated += nr;
3038			cond_resched();
3039
3040			/*
3041			 * If zero or pages were obtained partly,
3042			 * fallback to a single page allocator.
3043			 */
3044			if (nr != nr_pages_request)
3045				break;
3046		}
3047	} else if (gfp & __GFP_NOFAIL) {
3048		/*
3049		 * Higher order nofail allocations are really expensive and
3050		 * potentially dangerous (pre-mature OOM, disruptive reclaim
3051		 * and compaction etc.
3052		 */
3053		alloc_gfp &= ~__GFP_NOFAIL;
3054		nofail = true;
3055	}
3056
3057	/* High-order pages or fallback path if "bulk" fails. */
 
3058	while (nr_allocated < nr_pages) {
3059		if (fatal_signal_pending(current))
3060			break;
3061
3062		if (nid == NUMA_NO_NODE)
3063			page = alloc_pages(alloc_gfp, order);
3064		else
3065			page = alloc_pages_node(nid, alloc_gfp, order);
3066		if (unlikely(!page)) {
3067			if (!nofail)
3068				break;
3069
3070			/* fall back to the zero order allocations */
3071			alloc_gfp |= __GFP_NOFAIL;
3072			order = 0;
3073			continue;
3074		}
3075
3076		/*
3077		 * Higher order allocations must be able to be treated as
3078		 * indepdenent small pages by callers (as they can with
3079		 * small-page vmallocs). Some drivers do their own refcounting
3080		 * on vmalloc_to_page() pages, some use page->mapping,
3081		 * page->lru, etc.
3082		 */
3083		if (order)
3084			split_page(page, order);
3085
3086		/*
3087		 * Careful, we allocate and map page-order pages, but
3088		 * tracking is done per PAGE_SIZE page so as to keep the
3089		 * vm_struct APIs independent of the physical/mapped size.
3090		 */
3091		for (i = 0; i < (1U << order); i++)
3092			pages[nr_allocated + i] = page + i;
3093
3094		cond_resched();
3095		nr_allocated += 1U << order;
3096	}
3097
3098	return nr_allocated;
3099}
3100
3101static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3102				 pgprot_t prot, unsigned int page_shift,
3103				 int node)
3104{
3105	const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3106	bool nofail = gfp_mask & __GFP_NOFAIL;
3107	unsigned long addr = (unsigned long)area->addr;
3108	unsigned long size = get_vm_area_size(area);
3109	unsigned long array_size;
3110	unsigned int nr_small_pages = size >> PAGE_SHIFT;
3111	unsigned int page_order;
3112	unsigned int flags;
3113	int ret;
3114
3115	array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3116
3117	if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3118		gfp_mask |= __GFP_HIGHMEM;
3119
3120	/* Please note that the recursion is strictly bounded. */
3121	if (array_size > PAGE_SIZE) {
3122		area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3123					area->caller);
3124	} else {
3125		area->pages = kmalloc_node(array_size, nested_gfp, node);
3126	}
3127
3128	if (!area->pages) {
3129		warn_alloc(gfp_mask, NULL,
3130			"vmalloc error: size %lu, failed to allocated page array size %lu",
3131			nr_small_pages * PAGE_SIZE, array_size);
3132		free_vm_area(area);
3133		return NULL;
3134	}
3135
3136	set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3137	page_order = vm_area_page_order(area);
3138
3139	area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
3140		node, page_order, nr_small_pages, area->pages);
3141
3142	atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3143	if (gfp_mask & __GFP_ACCOUNT) {
3144		int i;
3145
3146		for (i = 0; i < area->nr_pages; i++)
3147			mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3148	}
3149
3150	/*
3151	 * If not enough pages were obtained to accomplish an
3152	 * allocation request, free them via vfree() if any.
3153	 */
3154	if (area->nr_pages != nr_small_pages) {
3155		/*
3156		 * vm_area_alloc_pages() can fail due to insufficient memory but
3157		 * also:-
3158		 *
3159		 * - a pending fatal signal
3160		 * - insufficient huge page-order pages
3161		 *
3162		 * Since we always retry allocations at order-0 in the huge page
3163		 * case a warning for either is spurious.
3164		 */
3165		if (!fatal_signal_pending(current) && page_order == 0)
3166			warn_alloc(gfp_mask, NULL,
3167				"vmalloc error: size %lu, failed to allocate pages",
3168				area->nr_pages * PAGE_SIZE);
3169		goto fail;
3170	}
3171
3172	/*
3173	 * page tables allocations ignore external gfp mask, enforce it
3174	 * by the scope API
3175	 */
3176	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3177		flags = memalloc_nofs_save();
3178	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3179		flags = memalloc_noio_save();
3180
3181	do {
3182		ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3183			page_shift);
3184		if (nofail && (ret < 0))
3185			schedule_timeout_uninterruptible(1);
3186	} while (nofail && (ret < 0));
3187
3188	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3189		memalloc_nofs_restore(flags);
3190	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3191		memalloc_noio_restore(flags);
3192
3193	if (ret < 0) {
3194		warn_alloc(gfp_mask, NULL,
3195			"vmalloc error: size %lu, failed to map pages",
3196			area->nr_pages * PAGE_SIZE);
3197		goto fail;
3198	}
3199
3200	return area->addr;
3201
3202fail:
3203	vfree(area->addr);
3204	return NULL;
3205}
3206
3207/**
3208 * __vmalloc_node_range - allocate virtually contiguous memory
3209 * @size:		  allocation size
3210 * @align:		  desired alignment
3211 * @start:		  vm area range start
3212 * @end:		  vm area range end
3213 * @gfp_mask:		  flags for the page level allocator
3214 * @prot:		  protection mask for the allocated pages
3215 * @vm_flags:		  additional vm area flags (e.g. %VM_NO_GUARD)
3216 * @node:		  node to use for allocation or NUMA_NO_NODE
3217 * @caller:		  caller's return address
3218 *
3219 * Allocate enough pages to cover @size from the page level
3220 * allocator with @gfp_mask flags. Please note that the full set of gfp
3221 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3222 * supported.
3223 * Zone modifiers are not supported. From the reclaim modifiers
3224 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3225 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3226 * __GFP_RETRY_MAYFAIL are not supported).
3227 *
3228 * __GFP_NOWARN can be used to suppress failures messages.
3229 *
3230 * Map them into contiguous kernel virtual space, using a pagetable
3231 * protection of @prot.
3232 *
3233 * Return: the address of the area or %NULL on failure
3234 */
3235void *__vmalloc_node_range(unsigned long size, unsigned long align,
3236			unsigned long start, unsigned long end, gfp_t gfp_mask,
3237			pgprot_t prot, unsigned long vm_flags, int node,
3238			const void *caller)
3239{
3240	struct vm_struct *area;
3241	void *ret;
3242	kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3243	unsigned long real_size = size;
3244	unsigned long real_align = align;
3245	unsigned int shift = PAGE_SHIFT;
3246
3247	if (WARN_ON_ONCE(!size))
3248		return NULL;
3249
3250	if ((size >> PAGE_SHIFT) > totalram_pages()) {
3251		warn_alloc(gfp_mask, NULL,
3252			"vmalloc error: size %lu, exceeds total pages",
3253			real_size);
3254		return NULL;
3255	}
3256
3257	if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3258		unsigned long size_per_node;
3259
3260		/*
3261		 * Try huge pages. Only try for PAGE_KERNEL allocations,
3262		 * others like modules don't yet expect huge pages in
3263		 * their allocations due to apply_to_page_range not
3264		 * supporting them.
3265		 */
3266
3267		size_per_node = size;
3268		if (node == NUMA_NO_NODE)
3269			size_per_node /= num_online_nodes();
3270		if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3271			shift = PMD_SHIFT;
3272		else
3273			shift = arch_vmap_pte_supported_shift(size_per_node);
3274
3275		align = max(real_align, 1UL << shift);
3276		size = ALIGN(real_size, 1UL << shift);
3277	}
3278
3279again:
3280	area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3281				  VM_UNINITIALIZED | vm_flags, start, end, node,
3282				  gfp_mask, caller);
3283	if (!area) {
3284		bool nofail = gfp_mask & __GFP_NOFAIL;
3285		warn_alloc(gfp_mask, NULL,
3286			"vmalloc error: size %lu, vm_struct allocation failed%s",
3287			real_size, (nofail) ? ". Retrying." : "");
3288		if (nofail) {
3289			schedule_timeout_uninterruptible(1);
3290			goto again;
3291		}
3292		goto fail;
3293	}
3294
3295	/*
3296	 * Prepare arguments for __vmalloc_area_node() and
3297	 * kasan_unpoison_vmalloc().
3298	 */
3299	if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3300		if (kasan_hw_tags_enabled()) {
3301			/*
3302			 * Modify protection bits to allow tagging.
3303			 * This must be done before mapping.
3304			 */
3305			prot = arch_vmap_pgprot_tagged(prot);
3306
3307			/*
3308			 * Skip page_alloc poisoning and zeroing for physical
3309			 * pages backing VM_ALLOC mapping. Memory is instead
3310			 * poisoned and zeroed by kasan_unpoison_vmalloc().
3311			 */
3312			gfp_mask |= __GFP_SKIP_KASAN | __GFP_SKIP_ZERO;
3313		}
3314
3315		/* Take note that the mapping is PAGE_KERNEL. */
3316		kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3317	}
3318
3319	/* Allocate physical pages and map them into vmalloc space. */
3320	ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3321	if (!ret)
3322		goto fail;
3323
3324	/*
3325	 * Mark the pages as accessible, now that they are mapped.
3326	 * The condition for setting KASAN_VMALLOC_INIT should complement the
3327	 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3328	 * to make sure that memory is initialized under the same conditions.
3329	 * Tag-based KASAN modes only assign tags to normal non-executable
3330	 * allocations, see __kasan_unpoison_vmalloc().
3331	 */
3332	kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3333	if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3334	    (gfp_mask & __GFP_SKIP_ZERO))
3335		kasan_flags |= KASAN_VMALLOC_INIT;
3336	/* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3337	area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3338
3339	/*
3340	 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3341	 * flag. It means that vm_struct is not fully initialized.
3342	 * Now, it is fully initialized, so remove this flag here.
3343	 */
3344	clear_vm_uninitialized_flag(area);
3345
3346	size = PAGE_ALIGN(size);
3347	if (!(vm_flags & VM_DEFER_KMEMLEAK))
3348		kmemleak_vmalloc(area, size, gfp_mask);
3349
3350	return area->addr;
3351
3352fail:
3353	if (shift > PAGE_SHIFT) {
3354		shift = PAGE_SHIFT;
3355		align = real_align;
3356		size = real_size;
3357		goto again;
3358	}
3359
3360	return NULL;
3361}
3362
3363/**
3364 * __vmalloc_node - allocate virtually contiguous memory
3365 * @size:	    allocation size
3366 * @align:	    desired alignment
3367 * @gfp_mask:	    flags for the page level allocator
3368 * @node:	    node to use for allocation or NUMA_NO_NODE
3369 * @caller:	    caller's return address
3370 *
3371 * Allocate enough pages to cover @size from the page level allocator with
3372 * @gfp_mask flags.  Map them into contiguous kernel virtual space.
3373 *
3374 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3375 * and __GFP_NOFAIL are not supported
3376 *
3377 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3378 * with mm people.
3379 *
3380 * Return: pointer to the allocated memory or %NULL on error
3381 */
3382void *__vmalloc_node(unsigned long size, unsigned long align,
3383			    gfp_t gfp_mask, int node, const void *caller)
3384{
3385	return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3386				gfp_mask, PAGE_KERNEL, 0, node, caller);
3387}
3388/*
3389 * This is only for performance analysis of vmalloc and stress purpose.
3390 * It is required by vmalloc test module, therefore do not use it other
3391 * than that.
3392 */
3393#ifdef CONFIG_TEST_VMALLOC_MODULE
3394EXPORT_SYMBOL_GPL(__vmalloc_node);
3395#endif
3396
3397void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3398{
3399	return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3400				__builtin_return_address(0));
3401}
3402EXPORT_SYMBOL(__vmalloc);
3403
3404/**
3405 * vmalloc - allocate virtually contiguous memory
3406 * @size:    allocation size
3407 *
3408 * Allocate enough pages to cover @size from the page level
3409 * allocator and map them into contiguous kernel virtual space.
3410 *
3411 * For tight control over page level allocator and protection flags
3412 * use __vmalloc() instead.
3413 *
3414 * Return: pointer to the allocated memory or %NULL on error
3415 */
3416void *vmalloc(unsigned long size)
3417{
3418	return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3419				__builtin_return_address(0));
3420}
3421EXPORT_SYMBOL(vmalloc);
3422
3423/**
3424 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3425 * @size:      allocation size
3426 * @gfp_mask:  flags for the page level allocator
3427 *
3428 * Allocate enough pages to cover @size from the page level
3429 * allocator and map them into contiguous kernel virtual space.
3430 * If @size is greater than or equal to PMD_SIZE, allow using
3431 * huge pages for the memory
3432 *
3433 * Return: pointer to the allocated memory or %NULL on error
3434 */
3435void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3436{
3437	return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3438				    gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3439				    NUMA_NO_NODE, __builtin_return_address(0));
3440}
3441EXPORT_SYMBOL_GPL(vmalloc_huge);
3442
3443/**
3444 * vzalloc - allocate virtually contiguous memory with zero fill
3445 * @size:    allocation size
3446 *
3447 * Allocate enough pages to cover @size from the page level
3448 * allocator and map them into contiguous kernel virtual space.
3449 * The memory allocated is set to zero.
3450 *
3451 * For tight control over page level allocator and protection flags
3452 * use __vmalloc() instead.
3453 *
3454 * Return: pointer to the allocated memory or %NULL on error
3455 */
3456void *vzalloc(unsigned long size)
3457{
3458	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3459				__builtin_return_address(0));
3460}
3461EXPORT_SYMBOL(vzalloc);
3462
3463/**
3464 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3465 * @size: allocation size
3466 *
3467 * The resulting memory area is zeroed so it can be mapped to userspace
3468 * without leaking data.
3469 *
3470 * Return: pointer to the allocated memory or %NULL on error
3471 */
3472void *vmalloc_user(unsigned long size)
3473{
3474	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
3475				    GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3476				    VM_USERMAP, NUMA_NO_NODE,
3477				    __builtin_return_address(0));
3478}
3479EXPORT_SYMBOL(vmalloc_user);
3480
3481/**
3482 * vmalloc_node - allocate memory on a specific node
3483 * @size:	  allocation size
3484 * @node:	  numa node
3485 *
3486 * Allocate enough pages to cover @size from the page level
3487 * allocator and map them into contiguous kernel virtual space.
3488 *
3489 * For tight control over page level allocator and protection flags
3490 * use __vmalloc() instead.
3491 *
3492 * Return: pointer to the allocated memory or %NULL on error
3493 */
3494void *vmalloc_node(unsigned long size, int node)
3495{
3496	return __vmalloc_node(size, 1, GFP_KERNEL, node,
3497			__builtin_return_address(0));
3498}
3499EXPORT_SYMBOL(vmalloc_node);
3500
3501/**
3502 * vzalloc_node - allocate memory on a specific node with zero fill
3503 * @size:	allocation size
3504 * @node:	numa node
3505 *
3506 * Allocate enough pages to cover @size from the page level
3507 * allocator and map them into contiguous kernel virtual space.
3508 * The memory allocated is set to zero.
3509 *
3510 * Return: pointer to the allocated memory or %NULL on error
3511 */
3512void *vzalloc_node(unsigned long size, int node)
3513{
3514	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3515				__builtin_return_address(0));
3516}
3517EXPORT_SYMBOL(vzalloc_node);
3518
3519#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3520#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3521#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3522#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3523#else
3524/*
3525 * 64b systems should always have either DMA or DMA32 zones. For others
3526 * GFP_DMA32 should do the right thing and use the normal zone.
3527 */
3528#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3529#endif
3530
3531/**
3532 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3533 * @size:	allocation size
3534 *
3535 * Allocate enough 32bit PA addressable pages to cover @size from the
3536 * page level allocator and map them into contiguous kernel virtual space.
3537 *
3538 * Return: pointer to the allocated memory or %NULL on error
3539 */
3540void *vmalloc_32(unsigned long size)
3541{
3542	return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3543			__builtin_return_address(0));
3544}
3545EXPORT_SYMBOL(vmalloc_32);
3546
3547/**
3548 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3549 * @size:	     allocation size
3550 *
3551 * The resulting memory area is 32bit addressable and zeroed so it can be
3552 * mapped to userspace without leaking data.
3553 *
3554 * Return: pointer to the allocated memory or %NULL on error
3555 */
3556void *vmalloc_32_user(unsigned long size)
3557{
3558	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
3559				    GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3560				    VM_USERMAP, NUMA_NO_NODE,
3561				    __builtin_return_address(0));
3562}
3563EXPORT_SYMBOL(vmalloc_32_user);
3564
3565/*
3566 * Atomically zero bytes in the iterator.
3567 *
3568 * Returns the number of zeroed bytes.
3569 */
3570static size_t zero_iter(struct iov_iter *iter, size_t count)
3571{
3572	size_t remains = count;
3573
3574	while (remains > 0) {
3575		size_t num, copied;
3576
3577		num = min_t(size_t, remains, PAGE_SIZE);
3578		copied = copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num, iter);
3579		remains -= copied;
3580
3581		if (copied < num)
3582			break;
3583	}
3584
3585	return count - remains;
3586}
3587
3588/*
3589 * small helper routine, copy contents to iter from addr.
3590 * If the page is not present, fill zero.
3591 *
3592 * Returns the number of copied bytes.
3593 */
3594static size_t aligned_vread_iter(struct iov_iter *iter,
3595				 const char *addr, size_t count)
3596{
3597	size_t remains = count;
3598	struct page *page;
3599
3600	while (remains > 0) {
3601		unsigned long offset, length;
3602		size_t copied = 0;
3603
3604		offset = offset_in_page(addr);
3605		length = PAGE_SIZE - offset;
3606		if (length > remains)
3607			length = remains;
3608		page = vmalloc_to_page(addr);
3609		/*
3610		 * To do safe access to this _mapped_ area, we need lock. But
3611		 * adding lock here means that we need to add overhead of
3612		 * vmalloc()/vfree() calls for this _debug_ interface, rarely
3613		 * used. Instead of that, we'll use an local mapping via
3614		 * copy_page_to_iter_nofault() and accept a small overhead in
3615		 * this access function.
3616		 */
3617		if (page)
3618			copied = copy_page_to_iter_nofault(page, offset,
3619							   length, iter);
3620		else
3621			copied = zero_iter(iter, length);
3622
3623		addr += copied;
3624		remains -= copied;
3625
3626		if (copied != length)
3627			break;
3628	}
3629
3630	return count - remains;
3631}
3632
3633/*
3634 * Read from a vm_map_ram region of memory.
3635 *
3636 * Returns the number of copied bytes.
3637 */
3638static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
3639				  size_t count, unsigned long flags)
3640{
3641	char *start;
3642	struct vmap_block *vb;
3643	struct xarray *xa;
3644	unsigned long offset;
3645	unsigned int rs, re;
3646	size_t remains, n;
3647
3648	/*
3649	 * If it's area created by vm_map_ram() interface directly, but
3650	 * not further subdividing and delegating management to vmap_block,
3651	 * handle it here.
3652	 */
3653	if (!(flags & VMAP_BLOCK))
3654		return aligned_vread_iter(iter, addr, count);
3655
3656	remains = count;
3657
3658	/*
3659	 * Area is split into regions and tracked with vmap_block, read out
3660	 * each region and zero fill the hole between regions.
3661	 */
3662	xa = addr_to_vb_xa((unsigned long) addr);
3663	vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr));
3664	if (!vb)
3665		goto finished_zero;
3666
3667	spin_lock(&vb->lock);
3668	if (bitmap_empty(vb->used_map, VMAP_BBMAP_BITS)) {
3669		spin_unlock(&vb->lock);
3670		goto finished_zero;
3671	}
3672
3673	for_each_set_bitrange(rs, re, vb->used_map, VMAP_BBMAP_BITS) {
3674		size_t copied;
3675
3676		if (remains == 0)
3677			goto finished;
3678
3679		start = vmap_block_vaddr(vb->va->va_start, rs);
3680
3681		if (addr < start) {
3682			size_t to_zero = min_t(size_t, start - addr, remains);
3683			size_t zeroed = zero_iter(iter, to_zero);
3684
3685			addr += zeroed;
3686			remains -= zeroed;
3687
3688			if (remains == 0 || zeroed != to_zero)
3689				goto finished;
3690		}
3691
3692		/*it could start reading from the middle of used region*/
3693		offset = offset_in_page(addr);
3694		n = ((re - rs + 1) << PAGE_SHIFT) - offset;
3695		if (n > remains)
3696			n = remains;
3697
3698		copied = aligned_vread_iter(iter, start + offset, n);
3699
3700		addr += copied;
3701		remains -= copied;
3702
3703		if (copied != n)
3704			goto finished;
 
 
3705	}
3706
3707	spin_unlock(&vb->lock);
3708
3709finished_zero:
3710	/* zero-fill the left dirty or free regions */
3711	return count - remains + zero_iter(iter, remains);
3712finished:
3713	/* We couldn't copy/zero everything */
3714	spin_unlock(&vb->lock);
3715	return count - remains;
3716}
3717
3718/**
3719 * vread_iter() - read vmalloc area in a safe way to an iterator.
3720 * @iter:         the iterator to which data should be written.
3721 * @addr:         vm address.
3722 * @count:        number of bytes to be read.
3723 *
3724 * This function checks that addr is a valid vmalloc'ed area, and
3725 * copy data from that area to a given buffer. If the given memory range
3726 * of [addr...addr+count) includes some valid address, data is copied to
3727 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3728 * IOREMAP area is treated as memory hole and no copy is done.
3729 *
3730 * If [addr...addr+count) doesn't includes any intersects with alive
3731 * vm_struct area, returns 0. @buf should be kernel's buffer.
3732 *
3733 * Note: In usual ops, vread() is never necessary because the caller
3734 * should know vmalloc() area is valid and can use memcpy().
3735 * This is for routines which have to access vmalloc area without
3736 * any information, as /proc/kcore.
3737 *
3738 * Return: number of bytes for which addr and buf should be increased
3739 * (same number as @count) or %0 if [addr...addr+count) doesn't
3740 * include any intersection with valid vmalloc area
3741 */
3742long vread_iter(struct iov_iter *iter, const char *addr, size_t count)
3743{
3744	struct vmap_area *va;
3745	struct vm_struct *vm;
3746	char *vaddr;
3747	size_t n, size, flags, remains;
 
3748
3749	addr = kasan_reset_tag(addr);
3750
3751	/* Don't allow overflow */
3752	if ((unsigned long) addr + count < count)
3753		count = -(unsigned long) addr;
3754
3755	remains = count;
3756
3757	spin_lock(&vmap_area_lock);
3758	va = find_vmap_area_exceed_addr((unsigned long)addr);
3759	if (!va)
3760		goto finished_zero;
3761
3762	/* no intersects with alive vmap_area */
3763	if ((unsigned long)addr + remains <= va->va_start)
3764		goto finished_zero;
3765
3766	list_for_each_entry_from(va, &vmap_area_list, list) {
3767		size_t copied;
3768
3769		if (remains == 0)
3770			goto finished;
3771
3772		vm = va->vm;
3773		flags = va->flags & VMAP_FLAGS_MASK;
3774		/*
3775		 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
3776		 * be set together with VMAP_RAM.
3777		 */
3778		WARN_ON(flags == VMAP_BLOCK);
3779
3780		if (!vm && !flags)
3781			continue;
3782
3783		if (vm && (vm->flags & VM_UNINITIALIZED))
3784			continue;
3785
3786		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3787		smp_rmb();
3788
3789		vaddr = (char *) va->va_start;
3790		size = vm ? get_vm_area_size(vm) : va_size(va);
3791
3792		if (addr >= vaddr + size)
3793			continue;
3794
3795		if (addr < vaddr) {
3796			size_t to_zero = min_t(size_t, vaddr - addr, remains);
3797			size_t zeroed = zero_iter(iter, to_zero);
3798
3799			addr += zeroed;
3800			remains -= zeroed;
3801
3802			if (remains == 0 || zeroed != to_zero)
3803				goto finished;
 
 
 
 
3804		}
3805
3806		n = vaddr + size - addr;
3807		if (n > remains)
3808			n = remains;
3809
3810		if (flags & VMAP_RAM)
3811			copied = vmap_ram_vread_iter(iter, addr, n, flags);
3812		else if (!(vm && (vm->flags & VM_IOREMAP)))
3813			copied = aligned_vread_iter(iter, addr, n);
3814		else /* IOREMAP area is treated as memory hole */
3815			copied = zero_iter(iter, n);
3816
3817		addr += copied;
3818		remains -= copied;
3819
3820		if (copied != n)
3821			goto finished;
3822	}
3823
3824finished_zero:
3825	spin_unlock(&vmap_area_lock);
3826	/* zero-fill memory holes */
3827	return count - remains + zero_iter(iter, remains);
3828finished:
3829	/* Nothing remains, or We couldn't copy/zero everything. */
3830	spin_unlock(&vmap_area_lock);
3831
3832	return count - remains;
 
 
 
 
 
 
3833}
3834
3835/**
3836 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3837 * @vma:		vma to cover
3838 * @uaddr:		target user address to start at
3839 * @kaddr:		virtual address of vmalloc kernel memory
3840 * @pgoff:		offset from @kaddr to start at
3841 * @size:		size of map area
3842 *
3843 * Returns:	0 for success, -Exxx on failure
3844 *
3845 * This function checks that @kaddr is a valid vmalloc'ed area,
3846 * and that it is big enough to cover the range starting at
3847 * @uaddr in @vma. Will return failure if that criteria isn't
3848 * met.
3849 *
3850 * Similar to remap_pfn_range() (see mm/memory.c)
3851 */
3852int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3853				void *kaddr, unsigned long pgoff,
3854				unsigned long size)
3855{
3856	struct vm_struct *area;
3857	unsigned long off;
3858	unsigned long end_index;
3859
3860	if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3861		return -EINVAL;
3862
3863	size = PAGE_ALIGN(size);
3864
3865	if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3866		return -EINVAL;
3867
3868	area = find_vm_area(kaddr);
3869	if (!area)
3870		return -EINVAL;
3871
3872	if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3873		return -EINVAL;
3874
3875	if (check_add_overflow(size, off, &end_index) ||
3876	    end_index > get_vm_area_size(area))
3877		return -EINVAL;
3878	kaddr += off;
3879
3880	do {
3881		struct page *page = vmalloc_to_page(kaddr);
3882		int ret;
3883
3884		ret = vm_insert_page(vma, uaddr, page);
3885		if (ret)
3886			return ret;
3887
3888		uaddr += PAGE_SIZE;
3889		kaddr += PAGE_SIZE;
3890		size -= PAGE_SIZE;
3891	} while (size > 0);
3892
3893	vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP);
3894
3895	return 0;
3896}
3897
3898/**
3899 * remap_vmalloc_range - map vmalloc pages to userspace
3900 * @vma:		vma to cover (map full range of vma)
3901 * @addr:		vmalloc memory
3902 * @pgoff:		number of pages into addr before first page to map
3903 *
3904 * Returns:	0 for success, -Exxx on failure
3905 *
3906 * This function checks that addr is a valid vmalloc'ed area, and
3907 * that it is big enough to cover the vma. Will return failure if
3908 * that criteria isn't met.
3909 *
3910 * Similar to remap_pfn_range() (see mm/memory.c)
3911 */
3912int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3913						unsigned long pgoff)
3914{
3915	return remap_vmalloc_range_partial(vma, vma->vm_start,
3916					   addr, pgoff,
3917					   vma->vm_end - vma->vm_start);
3918}
3919EXPORT_SYMBOL(remap_vmalloc_range);
3920
3921void free_vm_area(struct vm_struct *area)
3922{
3923	struct vm_struct *ret;
3924	ret = remove_vm_area(area->addr);
3925	BUG_ON(ret != area);
3926	kfree(area);
3927}
3928EXPORT_SYMBOL_GPL(free_vm_area);
3929
3930#ifdef CONFIG_SMP
3931static struct vmap_area *node_to_va(struct rb_node *n)
3932{
3933	return rb_entry_safe(n, struct vmap_area, rb_node);
3934}
3935
3936/**
3937 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3938 * @addr: target address
3939 *
3940 * Returns: vmap_area if it is found. If there is no such area
3941 *   the first highest(reverse order) vmap_area is returned
3942 *   i.e. va->va_start < addr && va->va_end < addr or NULL
3943 *   if there are no any areas before @addr.
3944 */
3945static struct vmap_area *
3946pvm_find_va_enclose_addr(unsigned long addr)
3947{
3948	struct vmap_area *va, *tmp;
3949	struct rb_node *n;
3950
3951	n = free_vmap_area_root.rb_node;
3952	va = NULL;
3953
3954	while (n) {
3955		tmp = rb_entry(n, struct vmap_area, rb_node);
3956		if (tmp->va_start <= addr) {
3957			va = tmp;
3958			if (tmp->va_end >= addr)
3959				break;
3960
3961			n = n->rb_right;
3962		} else {
3963			n = n->rb_left;
3964		}
3965	}
3966
3967	return va;
3968}
3969
3970/**
3971 * pvm_determine_end_from_reverse - find the highest aligned address
3972 * of free block below VMALLOC_END
3973 * @va:
3974 *   in - the VA we start the search(reverse order);
3975 *   out - the VA with the highest aligned end address.
3976 * @align: alignment for required highest address
3977 *
3978 * Returns: determined end address within vmap_area
3979 */
3980static unsigned long
3981pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3982{
3983	unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3984	unsigned long addr;
3985
3986	if (likely(*va)) {
3987		list_for_each_entry_from_reverse((*va),
3988				&free_vmap_area_list, list) {
3989			addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3990			if ((*va)->va_start < addr)
3991				return addr;
3992		}
3993	}
3994
3995	return 0;
3996}
3997
3998/**
3999 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
4000 * @offsets: array containing offset of each area
4001 * @sizes: array containing size of each area
4002 * @nr_vms: the number of areas to allocate
4003 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4004 *
4005 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4006 *	    vm_structs on success, %NULL on failure
4007 *
4008 * Percpu allocator wants to use congruent vm areas so that it can
4009 * maintain the offsets among percpu areas.  This function allocates
4010 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
4011 * be scattered pretty far, distance between two areas easily going up
4012 * to gigabytes.  To avoid interacting with regular vmallocs, these
4013 * areas are allocated from top.
4014 *
4015 * Despite its complicated look, this allocator is rather simple. It
4016 * does everything top-down and scans free blocks from the end looking
4017 * for matching base. While scanning, if any of the areas do not fit the
4018 * base address is pulled down to fit the area. Scanning is repeated till
4019 * all the areas fit and then all necessary data structures are inserted
4020 * and the result is returned.
4021 */
4022struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
4023				     const size_t *sizes, int nr_vms,
4024				     size_t align)
4025{
4026	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
4027	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4028	struct vmap_area **vas, *va;
4029	struct vm_struct **vms;
4030	int area, area2, last_area, term_area;
4031	unsigned long base, start, size, end, last_end, orig_start, orig_end;
4032	bool purged = false;
4033
4034	/* verify parameters and allocate data structures */
4035	BUG_ON(offset_in_page(align) || !is_power_of_2(align));
4036	for (last_area = 0, area = 0; area < nr_vms; area++) {
4037		start = offsets[area];
4038		end = start + sizes[area];
4039
4040		/* is everything aligned properly? */
4041		BUG_ON(!IS_ALIGNED(offsets[area], align));
4042		BUG_ON(!IS_ALIGNED(sizes[area], align));
4043
4044		/* detect the area with the highest address */
4045		if (start > offsets[last_area])
4046			last_area = area;
4047
4048		for (area2 = area + 1; area2 < nr_vms; area2++) {
4049			unsigned long start2 = offsets[area2];
4050			unsigned long end2 = start2 + sizes[area2];
4051
4052			BUG_ON(start2 < end && start < end2);
4053		}
4054	}
4055	last_end = offsets[last_area] + sizes[last_area];
4056
4057	if (vmalloc_end - vmalloc_start < last_end) {
4058		WARN_ON(true);
4059		return NULL;
4060	}
4061
4062	vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
4063	vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
4064	if (!vas || !vms)
4065		goto err_free2;
4066
4067	for (area = 0; area < nr_vms; area++) {
4068		vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
4069		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
4070		if (!vas[area] || !vms[area])
4071			goto err_free;
4072	}
4073retry:
4074	spin_lock(&free_vmap_area_lock);
4075
4076	/* start scanning - we scan from the top, begin with the last area */
4077	area = term_area = last_area;
4078	start = offsets[area];
4079	end = start + sizes[area];
4080
4081	va = pvm_find_va_enclose_addr(vmalloc_end);
4082	base = pvm_determine_end_from_reverse(&va, align) - end;
4083
4084	while (true) {
4085		/*
4086		 * base might have underflowed, add last_end before
4087		 * comparing.
4088		 */
4089		if (base + last_end < vmalloc_start + last_end)
4090			goto overflow;
4091
4092		/*
4093		 * Fitting base has not been found.
4094		 */
4095		if (va == NULL)
4096			goto overflow;
4097
4098		/*
4099		 * If required width exceeds current VA block, move
4100		 * base downwards and then recheck.
4101		 */
4102		if (base + end > va->va_end) {
4103			base = pvm_determine_end_from_reverse(&va, align) - end;
4104			term_area = area;
4105			continue;
4106		}
4107
4108		/*
4109		 * If this VA does not fit, move base downwards and recheck.
4110		 */
4111		if (base + start < va->va_start) {
4112			va = node_to_va(rb_prev(&va->rb_node));
4113			base = pvm_determine_end_from_reverse(&va, align) - end;
4114			term_area = area;
4115			continue;
4116		}
4117
4118		/*
4119		 * This area fits, move on to the previous one.  If
4120		 * the previous one is the terminal one, we're done.
4121		 */
4122		area = (area + nr_vms - 1) % nr_vms;
4123		if (area == term_area)
4124			break;
4125
4126		start = offsets[area];
4127		end = start + sizes[area];
4128		va = pvm_find_va_enclose_addr(base + end);
4129	}
4130
4131	/* we've found a fitting base, insert all va's */
4132	for (area = 0; area < nr_vms; area++) {
4133		int ret;
4134
4135		start = base + offsets[area];
4136		size = sizes[area];
4137
4138		va = pvm_find_va_enclose_addr(start);
4139		if (WARN_ON_ONCE(va == NULL))
4140			/* It is a BUG(), but trigger recovery instead. */
4141			goto recovery;
4142
4143		ret = adjust_va_to_fit_type(&free_vmap_area_root,
4144					    &free_vmap_area_list,
4145					    va, start, size);
4146		if (WARN_ON_ONCE(unlikely(ret)))
4147			/* It is a BUG(), but trigger recovery instead. */
4148			goto recovery;
4149
4150		/* Allocated area. */
4151		va = vas[area];
4152		va->va_start = start;
4153		va->va_end = start + size;
4154	}
4155
4156	spin_unlock(&free_vmap_area_lock);
4157
4158	/* populate the kasan shadow space */
4159	for (area = 0; area < nr_vms; area++) {
4160		if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
4161			goto err_free_shadow;
4162	}
4163
4164	/* insert all vm's */
4165	spin_lock(&vmap_area_lock);
4166	for (area = 0; area < nr_vms; area++) {
4167		insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
4168
4169		setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
4170				 pcpu_get_vm_areas);
4171	}
4172	spin_unlock(&vmap_area_lock);
4173
4174	/*
4175	 * Mark allocated areas as accessible. Do it now as a best-effort
4176	 * approach, as they can be mapped outside of vmalloc code.
4177	 * With hardware tag-based KASAN, marking is skipped for
4178	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4179	 */
4180	for (area = 0; area < nr_vms; area++)
4181		vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
4182				vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
4183
4184	kfree(vas);
4185	return vms;
4186
4187recovery:
4188	/*
4189	 * Remove previously allocated areas. There is no
4190	 * need in removing these areas from the busy tree,
4191	 * because they are inserted only on the final step
4192	 * and when pcpu_get_vm_areas() is success.
4193	 */
4194	while (area--) {
4195		orig_start = vas[area]->va_start;
4196		orig_end = vas[area]->va_end;
4197		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4198				&free_vmap_area_list);
4199		if (va)
4200			kasan_release_vmalloc(orig_start, orig_end,
4201				va->va_start, va->va_end);
4202		vas[area] = NULL;
4203	}
4204
4205overflow:
4206	spin_unlock(&free_vmap_area_lock);
4207	if (!purged) {
4208		reclaim_and_purge_vmap_areas();
4209		purged = true;
4210
4211		/* Before "retry", check if we recover. */
4212		for (area = 0; area < nr_vms; area++) {
4213			if (vas[area])
4214				continue;
4215
4216			vas[area] = kmem_cache_zalloc(
4217				vmap_area_cachep, GFP_KERNEL);
4218			if (!vas[area])
4219				goto err_free;
4220		}
4221
4222		goto retry;
4223	}
4224
4225err_free:
4226	for (area = 0; area < nr_vms; area++) {
4227		if (vas[area])
4228			kmem_cache_free(vmap_area_cachep, vas[area]);
4229
4230		kfree(vms[area]);
4231	}
4232err_free2:
4233	kfree(vas);
4234	kfree(vms);
4235	return NULL;
4236
4237err_free_shadow:
4238	spin_lock(&free_vmap_area_lock);
4239	/*
4240	 * We release all the vmalloc shadows, even the ones for regions that
4241	 * hadn't been successfully added. This relies on kasan_release_vmalloc
4242	 * being able to tolerate this case.
4243	 */
4244	for (area = 0; area < nr_vms; area++) {
4245		orig_start = vas[area]->va_start;
4246		orig_end = vas[area]->va_end;
4247		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4248				&free_vmap_area_list);
4249		if (va)
4250			kasan_release_vmalloc(orig_start, orig_end,
4251				va->va_start, va->va_end);
4252		vas[area] = NULL;
4253		kfree(vms[area]);
4254	}
4255	spin_unlock(&free_vmap_area_lock);
4256	kfree(vas);
4257	kfree(vms);
4258	return NULL;
4259}
4260
4261/**
4262 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4263 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4264 * @nr_vms: the number of allocated areas
4265 *
4266 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4267 */
4268void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4269{
4270	int i;
4271
4272	for (i = 0; i < nr_vms; i++)
4273		free_vm_area(vms[i]);
4274	kfree(vms);
4275}
4276#endif	/* CONFIG_SMP */
4277
4278#ifdef CONFIG_PRINTK
4279bool vmalloc_dump_obj(void *object)
4280{
4281	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
4282	const void *caller;
4283	struct vm_struct *vm;
4284	struct vmap_area *va;
4285	unsigned long addr;
4286	unsigned int nr_pages;
4287
4288	if (!spin_trylock(&vmap_area_lock))
4289		return false;
4290	va = __find_vmap_area((unsigned long)objp, &vmap_area_root);
4291	if (!va) {
4292		spin_unlock(&vmap_area_lock);
4293		return false;
4294	}
4295
4296	vm = va->vm;
4297	if (!vm) {
4298		spin_unlock(&vmap_area_lock);
4299		return false;
4300	}
4301	addr = (unsigned long)vm->addr;
4302	caller = vm->caller;
4303	nr_pages = vm->nr_pages;
4304	spin_unlock(&vmap_area_lock);
4305	pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4306		nr_pages, addr, caller);
4307	return true;
4308}
4309#endif
4310
4311#ifdef CONFIG_PROC_FS
4312static void *s_start(struct seq_file *m, loff_t *pos)
4313	__acquires(&vmap_purge_lock)
4314	__acquires(&vmap_area_lock)
4315{
4316	mutex_lock(&vmap_purge_lock);
4317	spin_lock(&vmap_area_lock);
4318
4319	return seq_list_start(&vmap_area_list, *pos);
4320}
4321
4322static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4323{
4324	return seq_list_next(p, &vmap_area_list, pos);
4325}
4326
4327static void s_stop(struct seq_file *m, void *p)
4328	__releases(&vmap_area_lock)
4329	__releases(&vmap_purge_lock)
4330{
4331	spin_unlock(&vmap_area_lock);
4332	mutex_unlock(&vmap_purge_lock);
4333}
4334
4335static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4336{
4337	if (IS_ENABLED(CONFIG_NUMA)) {
4338		unsigned int nr, *counters = m->private;
4339		unsigned int step = 1U << vm_area_page_order(v);
4340
4341		if (!counters)
4342			return;
4343
4344		if (v->flags & VM_UNINITIALIZED)
4345			return;
4346		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4347		smp_rmb();
4348
4349		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4350
4351		for (nr = 0; nr < v->nr_pages; nr += step)
4352			counters[page_to_nid(v->pages[nr])] += step;
4353		for_each_node_state(nr, N_HIGH_MEMORY)
4354			if (counters[nr])
4355				seq_printf(m, " N%u=%u", nr, counters[nr]);
4356	}
4357}
4358
4359static void show_purge_info(struct seq_file *m)
4360{
4361	struct vmap_area *va;
4362
4363	spin_lock(&purge_vmap_area_lock);
4364	list_for_each_entry(va, &purge_vmap_area_list, list) {
4365		seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4366			(void *)va->va_start, (void *)va->va_end,
4367			va->va_end - va->va_start);
4368	}
4369	spin_unlock(&purge_vmap_area_lock);
4370}
4371
4372static int s_show(struct seq_file *m, void *p)
4373{
4374	struct vmap_area *va;
4375	struct vm_struct *v;
4376
4377	va = list_entry(p, struct vmap_area, list);
4378
 
 
 
 
4379	if (!va->vm) {
4380		if (va->flags & VMAP_RAM)
4381			seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4382				(void *)va->va_start, (void *)va->va_end,
4383				va->va_end - va->va_start);
4384
4385		goto final;
4386	}
4387
4388	v = va->vm;
4389
4390	seq_printf(m, "0x%pK-0x%pK %7ld",
4391		v->addr, v->addr + v->size, v->size);
4392
4393	if (v->caller)
4394		seq_printf(m, " %pS", v->caller);
4395
4396	if (v->nr_pages)
4397		seq_printf(m, " pages=%d", v->nr_pages);
4398
4399	if (v->phys_addr)
4400		seq_printf(m, " phys=%pa", &v->phys_addr);
4401
4402	if (v->flags & VM_IOREMAP)
4403		seq_puts(m, " ioremap");
4404
4405	if (v->flags & VM_ALLOC)
4406		seq_puts(m, " vmalloc");
4407
4408	if (v->flags & VM_MAP)
4409		seq_puts(m, " vmap");
4410
4411	if (v->flags & VM_USERMAP)
4412		seq_puts(m, " user");
4413
4414	if (v->flags & VM_DMA_COHERENT)
4415		seq_puts(m, " dma-coherent");
4416
4417	if (is_vmalloc_addr(v->pages))
4418		seq_puts(m, " vpages");
4419
4420	show_numa_info(m, v);
4421	seq_putc(m, '\n');
4422
4423	/*
4424	 * As a final step, dump "unpurged" areas.
4425	 */
4426final:
4427	if (list_is_last(&va->list, &vmap_area_list))
4428		show_purge_info(m);
4429
4430	return 0;
4431}
4432
4433static const struct seq_operations vmalloc_op = {
4434	.start = s_start,
4435	.next = s_next,
4436	.stop = s_stop,
4437	.show = s_show,
4438};
4439
4440static int __init proc_vmalloc_init(void)
4441{
4442	if (IS_ENABLED(CONFIG_NUMA))
4443		proc_create_seq_private("vmallocinfo", 0400, NULL,
4444				&vmalloc_op,
4445				nr_node_ids * sizeof(unsigned int), NULL);
4446	else
4447		proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4448	return 0;
4449}
4450module_init(proc_vmalloc_init);
4451
4452#endif
4453
4454void __init vmalloc_init(void)
4455{
4456	struct vmap_area *va;
4457	struct vm_struct *tmp;
4458	int i;
4459
4460	/*
4461	 * Create the cache for vmap_area objects.
4462	 */
4463	vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
4464
4465	for_each_possible_cpu(i) {
4466		struct vmap_block_queue *vbq;
4467		struct vfree_deferred *p;
4468
4469		vbq = &per_cpu(vmap_block_queue, i);
4470		spin_lock_init(&vbq->lock);
4471		INIT_LIST_HEAD(&vbq->free);
4472		p = &per_cpu(vfree_deferred, i);
4473		init_llist_head(&p->list);
4474		INIT_WORK(&p->wq, delayed_vfree_work);
4475		xa_init(&vbq->vmap_blocks);
4476	}
4477
4478	/* Import existing vmlist entries. */
4479	for (tmp = vmlist; tmp; tmp = tmp->next) {
4480		va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
4481		if (WARN_ON_ONCE(!va))
4482			continue;
4483
4484		va->va_start = (unsigned long)tmp->addr;
4485		va->va_end = va->va_start + tmp->size;
4486		va->vm = tmp;
4487		insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
4488	}
4489
4490	/*
4491	 * Now we can initialize a free vmap space.
4492	 */
4493	vmap_init_free_space();
4494	vmap_initialized = true;
4495}
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