Linux Audio

Check our new training course

Loading...
v6.2
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 *  Copyright (C) 1993  Linus Torvalds
   4 *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   5 *  SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
   6 *  Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
   7 *  Numa awareness, Christoph Lameter, SGI, June 2005
   8 *  Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
   9 */
  10
  11#include <linux/vmalloc.h>
  12#include <linux/mm.h>
  13#include <linux/module.h>
  14#include <linux/highmem.h>
  15#include <linux/sched/signal.h>
  16#include <linux/slab.h>
  17#include <linux/spinlock.h>
  18#include <linux/interrupt.h>
  19#include <linux/proc_fs.h>
  20#include <linux/seq_file.h>
  21#include <linux/set_memory.h>
  22#include <linux/debugobjects.h>
  23#include <linux/kallsyms.h>
  24#include <linux/list.h>
  25#include <linux/notifier.h>
  26#include <linux/rbtree.h>
  27#include <linux/xarray.h>
  28#include <linux/io.h>
  29#include <linux/rcupdate.h>
  30#include <linux/pfn.h>
  31#include <linux/kmemleak.h>
  32#include <linux/atomic.h>
  33#include <linux/compiler.h>
  34#include <linux/memcontrol.h>
  35#include <linux/llist.h>
 
  36#include <linux/bitops.h>
  37#include <linux/rbtree_augmented.h>
  38#include <linux/overflow.h>
  39#include <linux/pgtable.h>
  40#include <linux/uaccess.h>
  41#include <linux/hugetlb.h>
  42#include <linux/sched/mm.h>
  43#include <asm/tlbflush.h>
  44#include <asm/shmparam.h>
 
  45
  46#define CREATE_TRACE_POINTS
  47#include <trace/events/vmalloc.h>
  48
  49#include "internal.h"
  50#include "pgalloc-track.h"
  51
  52#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
  53static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
  54
  55static int __init set_nohugeiomap(char *str)
  56{
  57	ioremap_max_page_shift = PAGE_SHIFT;
  58	return 0;
  59}
  60early_param("nohugeiomap", set_nohugeiomap);
  61#else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
  62static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
  63#endif	/* CONFIG_HAVE_ARCH_HUGE_VMAP */
  64
  65#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
  66static bool __ro_after_init vmap_allow_huge = true;
  67
  68static int __init set_nohugevmalloc(char *str)
  69{
  70	vmap_allow_huge = false;
  71	return 0;
  72}
  73early_param("nohugevmalloc", set_nohugevmalloc);
  74#else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
  75static const bool vmap_allow_huge = false;
  76#endif	/* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
  77
  78bool is_vmalloc_addr(const void *x)
  79{
  80	unsigned long addr = (unsigned long)kasan_reset_tag(x);
  81
  82	return addr >= VMALLOC_START && addr < VMALLOC_END;
  83}
  84EXPORT_SYMBOL(is_vmalloc_addr);
  85
  86struct vfree_deferred {
  87	struct llist_head list;
  88	struct work_struct wq;
  89};
  90static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
  91
  92static void __vunmap(const void *, int);
  93
  94static void free_work(struct work_struct *w)
  95{
  96	struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
  97	struct llist_node *t, *llnode;
  98
  99	llist_for_each_safe(llnode, t, llist_del_all(&p->list))
 100		__vunmap((void *)llnode, 1);
 101}
 102
 103/*** Page table manipulation functions ***/
 104static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 105			phys_addr_t phys_addr, pgprot_t prot,
 106			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 107{
 108	pte_t *pte;
 109	u64 pfn;
 
 110	unsigned long size = PAGE_SIZE;
 111
 112	pfn = phys_addr >> PAGE_SHIFT;
 113	pte = pte_alloc_kernel_track(pmd, addr, mask);
 114	if (!pte)
 115		return -ENOMEM;
 116	do {
 117		BUG_ON(!pte_none(*pte));
 
 
 
 
 
 
 118
 119#ifdef CONFIG_HUGETLB_PAGE
 120		size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
 121		if (size != PAGE_SIZE) {
 122			pte_t entry = pfn_pte(pfn, prot);
 123
 124			entry = arch_make_huge_pte(entry, ilog2(size), 0);
 125			set_huge_pte_at(&init_mm, addr, pte, entry);
 126			pfn += PFN_DOWN(size);
 127			continue;
 128		}
 129#endif
 130		set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
 131		pfn++;
 132	} while (pte += PFN_DOWN(size), addr += size, addr != end);
 133	*mask |= PGTBL_PTE_MODIFIED;
 134	return 0;
 135}
 136
 137static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
 138			phys_addr_t phys_addr, pgprot_t prot,
 139			unsigned int max_page_shift)
 140{
 141	if (max_page_shift < PMD_SHIFT)
 142		return 0;
 143
 144	if (!arch_vmap_pmd_supported(prot))
 145		return 0;
 146
 147	if ((end - addr) != PMD_SIZE)
 148		return 0;
 149
 150	if (!IS_ALIGNED(addr, PMD_SIZE))
 151		return 0;
 152
 153	if (!IS_ALIGNED(phys_addr, PMD_SIZE))
 154		return 0;
 155
 156	if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
 157		return 0;
 158
 159	return pmd_set_huge(pmd, phys_addr, prot);
 160}
 161
 162static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 163			phys_addr_t phys_addr, pgprot_t prot,
 164			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 165{
 166	pmd_t *pmd;
 167	unsigned long next;
 168
 169	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 170	if (!pmd)
 171		return -ENOMEM;
 172	do {
 173		next = pmd_addr_end(addr, end);
 174
 175		if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
 176					max_page_shift)) {
 177			*mask |= PGTBL_PMD_MODIFIED;
 178			continue;
 179		}
 180
 181		if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
 182			return -ENOMEM;
 183	} while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
 184	return 0;
 185}
 186
 187static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
 188			phys_addr_t phys_addr, pgprot_t prot,
 189			unsigned int max_page_shift)
 190{
 191	if (max_page_shift < PUD_SHIFT)
 192		return 0;
 193
 194	if (!arch_vmap_pud_supported(prot))
 195		return 0;
 196
 197	if ((end - addr) != PUD_SIZE)
 198		return 0;
 199
 200	if (!IS_ALIGNED(addr, PUD_SIZE))
 201		return 0;
 202
 203	if (!IS_ALIGNED(phys_addr, PUD_SIZE))
 204		return 0;
 205
 206	if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
 207		return 0;
 208
 209	return pud_set_huge(pud, phys_addr, prot);
 210}
 211
 212static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 213			phys_addr_t phys_addr, pgprot_t prot,
 214			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 215{
 216	pud_t *pud;
 217	unsigned long next;
 218
 219	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 220	if (!pud)
 221		return -ENOMEM;
 222	do {
 223		next = pud_addr_end(addr, end);
 224
 225		if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
 226					max_page_shift)) {
 227			*mask |= PGTBL_PUD_MODIFIED;
 228			continue;
 229		}
 230
 231		if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
 232					max_page_shift, mask))
 233			return -ENOMEM;
 234	} while (pud++, phys_addr += (next - addr), addr = next, addr != end);
 235	return 0;
 236}
 237
 238static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
 239			phys_addr_t phys_addr, pgprot_t prot,
 240			unsigned int max_page_shift)
 241{
 242	if (max_page_shift < P4D_SHIFT)
 243		return 0;
 244
 245	if (!arch_vmap_p4d_supported(prot))
 246		return 0;
 247
 248	if ((end - addr) != P4D_SIZE)
 249		return 0;
 250
 251	if (!IS_ALIGNED(addr, P4D_SIZE))
 252		return 0;
 253
 254	if (!IS_ALIGNED(phys_addr, P4D_SIZE))
 255		return 0;
 256
 257	if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
 258		return 0;
 259
 260	return p4d_set_huge(p4d, phys_addr, prot);
 261}
 262
 263static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 264			phys_addr_t phys_addr, pgprot_t prot,
 265			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 266{
 267	p4d_t *p4d;
 268	unsigned long next;
 269
 270	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 271	if (!p4d)
 272		return -ENOMEM;
 273	do {
 274		next = p4d_addr_end(addr, end);
 275
 276		if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
 277					max_page_shift)) {
 278			*mask |= PGTBL_P4D_MODIFIED;
 279			continue;
 280		}
 281
 282		if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
 283					max_page_shift, mask))
 284			return -ENOMEM;
 285	} while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
 286	return 0;
 287}
 288
 289static int vmap_range_noflush(unsigned long addr, unsigned long end,
 290			phys_addr_t phys_addr, pgprot_t prot,
 291			unsigned int max_page_shift)
 292{
 293	pgd_t *pgd;
 294	unsigned long start;
 295	unsigned long next;
 296	int err;
 297	pgtbl_mod_mask mask = 0;
 298
 299	might_sleep();
 300	BUG_ON(addr >= end);
 301
 302	start = addr;
 303	pgd = pgd_offset_k(addr);
 304	do {
 305		next = pgd_addr_end(addr, end);
 306		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
 307					max_page_shift, &mask);
 308		if (err)
 309			break;
 310	} while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
 311
 312	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 313		arch_sync_kernel_mappings(start, end);
 314
 315	return err;
 316}
 317
 318int ioremap_page_range(unsigned long addr, unsigned long end,
 319		phys_addr_t phys_addr, pgprot_t prot)
 320{
 321	int err;
 322
 323	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
 324				 ioremap_max_page_shift);
 325	flush_cache_vmap(addr, end);
 326	if (!err)
 327		kmsan_ioremap_page_range(addr, end, phys_addr, prot,
 328					 ioremap_max_page_shift);
 329	return err;
 330}
 331
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 332static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 333			     pgtbl_mod_mask *mask)
 334{
 335	pte_t *pte;
 336
 337	pte = pte_offset_kernel(pmd, addr);
 338	do {
 339		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
 340		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
 341	} while (pte++, addr += PAGE_SIZE, addr != end);
 342	*mask |= PGTBL_PTE_MODIFIED;
 343}
 344
 345static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 346			     pgtbl_mod_mask *mask)
 347{
 348	pmd_t *pmd;
 349	unsigned long next;
 350	int cleared;
 351
 352	pmd = pmd_offset(pud, addr);
 353	do {
 354		next = pmd_addr_end(addr, end);
 355
 356		cleared = pmd_clear_huge(pmd);
 357		if (cleared || pmd_bad(*pmd))
 358			*mask |= PGTBL_PMD_MODIFIED;
 359
 360		if (cleared)
 361			continue;
 362		if (pmd_none_or_clear_bad(pmd))
 363			continue;
 364		vunmap_pte_range(pmd, addr, next, mask);
 365
 366		cond_resched();
 367	} while (pmd++, addr = next, addr != end);
 368}
 369
 370static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 371			     pgtbl_mod_mask *mask)
 372{
 373	pud_t *pud;
 374	unsigned long next;
 375	int cleared;
 376
 377	pud = pud_offset(p4d, addr);
 378	do {
 379		next = pud_addr_end(addr, end);
 380
 381		cleared = pud_clear_huge(pud);
 382		if (cleared || pud_bad(*pud))
 383			*mask |= PGTBL_PUD_MODIFIED;
 384
 385		if (cleared)
 386			continue;
 387		if (pud_none_or_clear_bad(pud))
 388			continue;
 389		vunmap_pmd_range(pud, addr, next, mask);
 390	} while (pud++, addr = next, addr != end);
 391}
 392
 393static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 394			     pgtbl_mod_mask *mask)
 395{
 396	p4d_t *p4d;
 397	unsigned long next;
 398
 399	p4d = p4d_offset(pgd, addr);
 400	do {
 401		next = p4d_addr_end(addr, end);
 402
 403		p4d_clear_huge(p4d);
 404		if (p4d_bad(*p4d))
 405			*mask |= PGTBL_P4D_MODIFIED;
 406
 407		if (p4d_none_or_clear_bad(p4d))
 408			continue;
 409		vunmap_pud_range(p4d, addr, next, mask);
 410	} while (p4d++, addr = next, addr != end);
 411}
 412
 413/*
 414 * vunmap_range_noflush is similar to vunmap_range, but does not
 415 * flush caches or TLBs.
 416 *
 417 * The caller is responsible for calling flush_cache_vmap() before calling
 418 * this function, and flush_tlb_kernel_range after it has returned
 419 * successfully (and before the addresses are expected to cause a page fault
 420 * or be re-mapped for something else, if TLB flushes are being delayed or
 421 * coalesced).
 422 *
 423 * This is an internal function only. Do not use outside mm/.
 424 */
 425void __vunmap_range_noflush(unsigned long start, unsigned long end)
 426{
 427	unsigned long next;
 428	pgd_t *pgd;
 429	unsigned long addr = start;
 430	pgtbl_mod_mask mask = 0;
 431
 432	BUG_ON(addr >= end);
 433	pgd = pgd_offset_k(addr);
 434	do {
 435		next = pgd_addr_end(addr, end);
 436		if (pgd_bad(*pgd))
 437			mask |= PGTBL_PGD_MODIFIED;
 438		if (pgd_none_or_clear_bad(pgd))
 439			continue;
 440		vunmap_p4d_range(pgd, addr, next, &mask);
 441	} while (pgd++, addr = next, addr != end);
 442
 443	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 444		arch_sync_kernel_mappings(start, end);
 445}
 446
 447void vunmap_range_noflush(unsigned long start, unsigned long end)
 448{
 449	kmsan_vunmap_range_noflush(start, end);
 450	__vunmap_range_noflush(start, end);
 451}
 452
 453/**
 454 * vunmap_range - unmap kernel virtual addresses
 455 * @addr: start of the VM area to unmap
 456 * @end: end of the VM area to unmap (non-inclusive)
 457 *
 458 * Clears any present PTEs in the virtual address range, flushes TLBs and
 459 * caches. Any subsequent access to the address before it has been re-mapped
 460 * is a kernel bug.
 461 */
 462void vunmap_range(unsigned long addr, unsigned long end)
 463{
 464	flush_cache_vunmap(addr, end);
 465	vunmap_range_noflush(addr, end);
 466	flush_tlb_kernel_range(addr, end);
 467}
 468
 469static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
 470		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 471		pgtbl_mod_mask *mask)
 472{
 473	pte_t *pte;
 474
 475	/*
 476	 * nr is a running index into the array which helps higher level
 477	 * callers keep track of where we're up to.
 478	 */
 479
 480	pte = pte_alloc_kernel_track(pmd, addr, mask);
 481	if (!pte)
 482		return -ENOMEM;
 483	do {
 484		struct page *page = pages[*nr];
 485
 486		if (WARN_ON(!pte_none(*pte)))
 487			return -EBUSY;
 488		if (WARN_ON(!page))
 489			return -ENOMEM;
 490		if (WARN_ON(!pfn_valid(page_to_pfn(page))))
 491			return -EINVAL;
 492
 493		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 494		(*nr)++;
 495	} while (pte++, addr += PAGE_SIZE, addr != end);
 496	*mask |= PGTBL_PTE_MODIFIED;
 497	return 0;
 498}
 499
 500static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
 501		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 502		pgtbl_mod_mask *mask)
 503{
 504	pmd_t *pmd;
 505	unsigned long next;
 506
 507	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 508	if (!pmd)
 509		return -ENOMEM;
 510	do {
 511		next = pmd_addr_end(addr, end);
 512		if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
 513			return -ENOMEM;
 514	} while (pmd++, addr = next, addr != end);
 515	return 0;
 516}
 517
 518static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
 519		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 520		pgtbl_mod_mask *mask)
 521{
 522	pud_t *pud;
 523	unsigned long next;
 524
 525	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 526	if (!pud)
 527		return -ENOMEM;
 528	do {
 529		next = pud_addr_end(addr, end);
 530		if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
 531			return -ENOMEM;
 532	} while (pud++, addr = next, addr != end);
 533	return 0;
 534}
 535
 536static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
 537		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 538		pgtbl_mod_mask *mask)
 539{
 540	p4d_t *p4d;
 541	unsigned long next;
 542
 543	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 544	if (!p4d)
 545		return -ENOMEM;
 546	do {
 547		next = p4d_addr_end(addr, end);
 548		if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
 549			return -ENOMEM;
 550	} while (p4d++, addr = next, addr != end);
 551	return 0;
 552}
 553
 554static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
 555		pgprot_t prot, struct page **pages)
 556{
 557	unsigned long start = addr;
 558	pgd_t *pgd;
 559	unsigned long next;
 560	int err = 0;
 561	int nr = 0;
 562	pgtbl_mod_mask mask = 0;
 563
 564	BUG_ON(addr >= end);
 565	pgd = pgd_offset_k(addr);
 566	do {
 567		next = pgd_addr_end(addr, end);
 568		if (pgd_bad(*pgd))
 569			mask |= PGTBL_PGD_MODIFIED;
 570		err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
 571		if (err)
 572			return err;
 573	} while (pgd++, addr = next, addr != end);
 574
 575	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 576		arch_sync_kernel_mappings(start, end);
 577
 578	return 0;
 579}
 580
 581/*
 582 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
 583 * flush caches.
 584 *
 585 * The caller is responsible for calling flush_cache_vmap() after this
 586 * function returns successfully and before the addresses are accessed.
 587 *
 588 * This is an internal function only. Do not use outside mm/.
 589 */
 590int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 591		pgprot_t prot, struct page **pages, unsigned int page_shift)
 592{
 593	unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
 594
 595	WARN_ON(page_shift < PAGE_SHIFT);
 596
 597	if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
 598			page_shift == PAGE_SHIFT)
 599		return vmap_small_pages_range_noflush(addr, end, prot, pages);
 600
 601	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
 602		int err;
 603
 604		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
 605					page_to_phys(pages[i]), prot,
 606					page_shift);
 607		if (err)
 608			return err;
 609
 610		addr += 1UL << page_shift;
 611	}
 612
 613	return 0;
 614}
 615
 616int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 617		pgprot_t prot, struct page **pages, unsigned int page_shift)
 618{
 619	kmsan_vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 
 
 
 
 620	return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 621}
 622
 623/**
 624 * vmap_pages_range - map pages to a kernel virtual address
 625 * @addr: start of the VM area to map
 626 * @end: end of the VM area to map (non-inclusive)
 627 * @prot: page protection flags to use
 628 * @pages: pages to map (always PAGE_SIZE pages)
 629 * @page_shift: maximum shift that the pages may be mapped with, @pages must
 630 * be aligned and contiguous up to at least this shift.
 631 *
 632 * RETURNS:
 633 * 0 on success, -errno on failure.
 634 */
 635static int vmap_pages_range(unsigned long addr, unsigned long end,
 636		pgprot_t prot, struct page **pages, unsigned int page_shift)
 637{
 638	int err;
 639
 640	err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 641	flush_cache_vmap(addr, end);
 642	return err;
 643}
 644
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 645int is_vmalloc_or_module_addr(const void *x)
 646{
 647	/*
 648	 * ARM, x86-64 and sparc64 put modules in a special place,
 649	 * and fall back on vmalloc() if that fails. Others
 650	 * just put it in the vmalloc space.
 651	 */
 652#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
 653	unsigned long addr = (unsigned long)kasan_reset_tag(x);
 654	if (addr >= MODULES_VADDR && addr < MODULES_END)
 655		return 1;
 656#endif
 657	return is_vmalloc_addr(x);
 658}
 
 659
 660/*
 661 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
 662 * return the tail page that corresponds to the base page address, which
 663 * matches small vmap mappings.
 664 */
 665struct page *vmalloc_to_page(const void *vmalloc_addr)
 666{
 667	unsigned long addr = (unsigned long) vmalloc_addr;
 668	struct page *page = NULL;
 669	pgd_t *pgd = pgd_offset_k(addr);
 670	p4d_t *p4d;
 671	pud_t *pud;
 672	pmd_t *pmd;
 673	pte_t *ptep, pte;
 674
 675	/*
 676	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 677	 * architectures that do not vmalloc module space
 678	 */
 679	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 680
 681	if (pgd_none(*pgd))
 682		return NULL;
 683	if (WARN_ON_ONCE(pgd_leaf(*pgd)))
 684		return NULL; /* XXX: no allowance for huge pgd */
 685	if (WARN_ON_ONCE(pgd_bad(*pgd)))
 686		return NULL;
 687
 688	p4d = p4d_offset(pgd, addr);
 689	if (p4d_none(*p4d))
 690		return NULL;
 691	if (p4d_leaf(*p4d))
 692		return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
 693	if (WARN_ON_ONCE(p4d_bad(*p4d)))
 694		return NULL;
 695
 696	pud = pud_offset(p4d, addr);
 697	if (pud_none(*pud))
 698		return NULL;
 699	if (pud_leaf(*pud))
 700		return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 701	if (WARN_ON_ONCE(pud_bad(*pud)))
 702		return NULL;
 703
 704	pmd = pmd_offset(pud, addr);
 705	if (pmd_none(*pmd))
 706		return NULL;
 707	if (pmd_leaf(*pmd))
 708		return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 709	if (WARN_ON_ONCE(pmd_bad(*pmd)))
 710		return NULL;
 711
 712	ptep = pte_offset_map(pmd, addr);
 713	pte = *ptep;
 714	if (pte_present(pte))
 715		page = pte_page(pte);
 716	pte_unmap(ptep);
 717
 718	return page;
 719}
 720EXPORT_SYMBOL(vmalloc_to_page);
 721
 722/*
 723 * Map a vmalloc()-space virtual address to the physical page frame number.
 724 */
 725unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 726{
 727	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 728}
 729EXPORT_SYMBOL(vmalloc_to_pfn);
 730
 731
 732/*** Global kva allocator ***/
 733
 734#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
 735#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
 736
 737
 738static DEFINE_SPINLOCK(vmap_area_lock);
 739static DEFINE_SPINLOCK(free_vmap_area_lock);
 740/* Export for kexec only */
 741LIST_HEAD(vmap_area_list);
 742static struct rb_root vmap_area_root = RB_ROOT;
 743static bool vmap_initialized __read_mostly;
 744
 745static struct rb_root purge_vmap_area_root = RB_ROOT;
 746static LIST_HEAD(purge_vmap_area_list);
 747static DEFINE_SPINLOCK(purge_vmap_area_lock);
 748
 749/*
 750 * This kmem_cache is used for vmap_area objects. Instead of
 751 * allocating from slab we reuse an object from this cache to
 752 * make things faster. Especially in "no edge" splitting of
 753 * free block.
 754 */
 755static struct kmem_cache *vmap_area_cachep;
 756
 757/*
 758 * This linked list is used in pair with free_vmap_area_root.
 759 * It gives O(1) access to prev/next to perform fast coalescing.
 760 */
 761static LIST_HEAD(free_vmap_area_list);
 762
 763/*
 764 * This augment red-black tree represents the free vmap space.
 765 * All vmap_area objects in this tree are sorted by va->va_start
 766 * address. It is used for allocation and merging when a vmap
 767 * object is released.
 768 *
 769 * Each vmap_area node contains a maximum available free block
 770 * of its sub-tree, right or left. Therefore it is possible to
 771 * find a lowest match of free area.
 772 */
 773static struct rb_root free_vmap_area_root = RB_ROOT;
 774
 775/*
 776 * Preload a CPU with one object for "no edge" split case. The
 777 * aim is to get rid of allocations from the atomic context, thus
 778 * to use more permissive allocation masks.
 779 */
 780static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
 781
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 782static __always_inline unsigned long
 783va_size(struct vmap_area *va)
 784{
 785	return (va->va_end - va->va_start);
 786}
 787
 788static __always_inline unsigned long
 789get_subtree_max_size(struct rb_node *node)
 790{
 791	struct vmap_area *va;
 792
 793	va = rb_entry_safe(node, struct vmap_area, rb_node);
 794	return va ? va->subtree_max_size : 0;
 795}
 796
 797RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
 798	struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
 799
 800static void purge_vmap_area_lazy(void);
 801static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
 802static void drain_vmap_area_work(struct work_struct *work);
 803static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
 804
 805static atomic_long_t nr_vmalloc_pages;
 806
 807unsigned long vmalloc_nr_pages(void)
 808{
 809	return atomic_long_read(&nr_vmalloc_pages);
 810}
 811
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 812/* Look up the first VA which satisfies addr < va_end, NULL if none. */
 813static struct vmap_area *find_vmap_area_exceed_addr(unsigned long addr)
 
 814{
 815	struct vmap_area *va = NULL;
 816	struct rb_node *n = vmap_area_root.rb_node;
 817
 818	addr = (unsigned long)kasan_reset_tag((void *)addr);
 819
 820	while (n) {
 821		struct vmap_area *tmp;
 822
 823		tmp = rb_entry(n, struct vmap_area, rb_node);
 824		if (tmp->va_end > addr) {
 825			va = tmp;
 826			if (tmp->va_start <= addr)
 827				break;
 828
 829			n = n->rb_left;
 830		} else
 831			n = n->rb_right;
 832	}
 833
 834	return va;
 835}
 836
 837static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
 
 
 
 
 
 
 
 
 838{
 839	struct rb_node *n = root->rb_node;
 
 
 840
 841	addr = (unsigned long)kasan_reset_tag((void *)addr);
 
 
 
 
 
 
 
 
 
 
 
 842
 843	while (n) {
 844		struct vmap_area *va;
 
 
 
 
 
 845
 846		va = rb_entry(n, struct vmap_area, rb_node);
 847		if (addr < va->va_start)
 848			n = n->rb_left;
 849		else if (addr >= va->va_end)
 850			n = n->rb_right;
 851		else
 852			return va;
 
 853	}
 854
 855	return NULL;
 856}
 857
 858/*
 859 * This function returns back addresses of parent node
 860 * and its left or right link for further processing.
 861 *
 862 * Otherwise NULL is returned. In that case all further
 863 * steps regarding inserting of conflicting overlap range
 864 * have to be declined and actually considered as a bug.
 865 */
 866static __always_inline struct rb_node **
 867find_va_links(struct vmap_area *va,
 868	struct rb_root *root, struct rb_node *from,
 869	struct rb_node **parent)
 870{
 871	struct vmap_area *tmp_va;
 872	struct rb_node **link;
 873
 874	if (root) {
 875		link = &root->rb_node;
 876		if (unlikely(!*link)) {
 877			*parent = NULL;
 878			return link;
 879		}
 880	} else {
 881		link = &from;
 882	}
 883
 884	/*
 885	 * Go to the bottom of the tree. When we hit the last point
 886	 * we end up with parent rb_node and correct direction, i name
 887	 * it link, where the new va->rb_node will be attached to.
 888	 */
 889	do {
 890		tmp_va = rb_entry(*link, struct vmap_area, rb_node);
 891
 892		/*
 893		 * During the traversal we also do some sanity check.
 894		 * Trigger the BUG() if there are sides(left/right)
 895		 * or full overlaps.
 896		 */
 897		if (va->va_end <= tmp_va->va_start)
 898			link = &(*link)->rb_left;
 899		else if (va->va_start >= tmp_va->va_end)
 900			link = &(*link)->rb_right;
 901		else {
 902			WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
 903				va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
 904
 905			return NULL;
 906		}
 907	} while (*link);
 908
 909	*parent = &tmp_va->rb_node;
 910	return link;
 911}
 912
 913static __always_inline struct list_head *
 914get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
 915{
 916	struct list_head *list;
 917
 918	if (unlikely(!parent))
 919		/*
 920		 * The red-black tree where we try to find VA neighbors
 921		 * before merging or inserting is empty, i.e. it means
 922		 * there is no free vmap space. Normally it does not
 923		 * happen but we handle this case anyway.
 924		 */
 925		return NULL;
 926
 927	list = &rb_entry(parent, struct vmap_area, rb_node)->list;
 928	return (&parent->rb_right == link ? list->next : list);
 929}
 930
 931static __always_inline void
 932__link_va(struct vmap_area *va, struct rb_root *root,
 933	struct rb_node *parent, struct rb_node **link,
 934	struct list_head *head, bool augment)
 935{
 936	/*
 937	 * VA is still not in the list, but we can
 938	 * identify its future previous list_head node.
 939	 */
 940	if (likely(parent)) {
 941		head = &rb_entry(parent, struct vmap_area, rb_node)->list;
 942		if (&parent->rb_right != link)
 943			head = head->prev;
 944	}
 945
 946	/* Insert to the rb-tree */
 947	rb_link_node(&va->rb_node, parent, link);
 948	if (augment) {
 949		/*
 950		 * Some explanation here. Just perform simple insertion
 951		 * to the tree. We do not set va->subtree_max_size to
 952		 * its current size before calling rb_insert_augmented().
 953		 * It is because we populate the tree from the bottom
 954		 * to parent levels when the node _is_ in the tree.
 955		 *
 956		 * Therefore we set subtree_max_size to zero after insertion,
 957		 * to let __augment_tree_propagate_from() puts everything to
 958		 * the correct order later on.
 959		 */
 960		rb_insert_augmented(&va->rb_node,
 961			root, &free_vmap_area_rb_augment_cb);
 962		va->subtree_max_size = 0;
 963	} else {
 964		rb_insert_color(&va->rb_node, root);
 965	}
 966
 967	/* Address-sort this list */
 968	list_add(&va->list, head);
 969}
 970
 971static __always_inline void
 972link_va(struct vmap_area *va, struct rb_root *root,
 973	struct rb_node *parent, struct rb_node **link,
 974	struct list_head *head)
 975{
 976	__link_va(va, root, parent, link, head, false);
 977}
 978
 979static __always_inline void
 980link_va_augment(struct vmap_area *va, struct rb_root *root,
 981	struct rb_node *parent, struct rb_node **link,
 982	struct list_head *head)
 983{
 984	__link_va(va, root, parent, link, head, true);
 985}
 986
 987static __always_inline void
 988__unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
 989{
 990	if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
 991		return;
 992
 993	if (augment)
 994		rb_erase_augmented(&va->rb_node,
 995			root, &free_vmap_area_rb_augment_cb);
 996	else
 997		rb_erase(&va->rb_node, root);
 998
 999	list_del_init(&va->list);
1000	RB_CLEAR_NODE(&va->rb_node);
1001}
1002
1003static __always_inline void
1004unlink_va(struct vmap_area *va, struct rb_root *root)
1005{
1006	__unlink_va(va, root, false);
1007}
1008
1009static __always_inline void
1010unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1011{
1012	__unlink_va(va, root, true);
1013}
1014
1015#if DEBUG_AUGMENT_PROPAGATE_CHECK
1016/*
1017 * Gets called when remove the node and rotate.
1018 */
1019static __always_inline unsigned long
1020compute_subtree_max_size(struct vmap_area *va)
1021{
1022	return max3(va_size(va),
1023		get_subtree_max_size(va->rb_node.rb_left),
1024		get_subtree_max_size(va->rb_node.rb_right));
1025}
1026
1027static void
1028augment_tree_propagate_check(void)
1029{
1030	struct vmap_area *va;
1031	unsigned long computed_size;
1032
1033	list_for_each_entry(va, &free_vmap_area_list, list) {
1034		computed_size = compute_subtree_max_size(va);
1035		if (computed_size != va->subtree_max_size)
1036			pr_emerg("tree is corrupted: %lu, %lu\n",
1037				va_size(va), va->subtree_max_size);
1038	}
1039}
1040#endif
1041
1042/*
1043 * This function populates subtree_max_size from bottom to upper
1044 * levels starting from VA point. The propagation must be done
1045 * when VA size is modified by changing its va_start/va_end. Or
1046 * in case of newly inserting of VA to the tree.
1047 *
1048 * It means that __augment_tree_propagate_from() must be called:
1049 * - After VA has been inserted to the tree(free path);
1050 * - After VA has been shrunk(allocation path);
1051 * - After VA has been increased(merging path).
1052 *
1053 * Please note that, it does not mean that upper parent nodes
1054 * and their subtree_max_size are recalculated all the time up
1055 * to the root node.
1056 *
1057 *       4--8
1058 *        /\
1059 *       /  \
1060 *      /    \
1061 *    2--2  8--8
1062 *
1063 * For example if we modify the node 4, shrinking it to 2, then
1064 * no any modification is required. If we shrink the node 2 to 1
1065 * its subtree_max_size is updated only, and set to 1. If we shrink
1066 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1067 * node becomes 4--6.
1068 */
1069static __always_inline void
1070augment_tree_propagate_from(struct vmap_area *va)
1071{
1072	/*
1073	 * Populate the tree from bottom towards the root until
1074	 * the calculated maximum available size of checked node
1075	 * is equal to its current one.
1076	 */
1077	free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1078
1079#if DEBUG_AUGMENT_PROPAGATE_CHECK
1080	augment_tree_propagate_check();
1081#endif
1082}
1083
1084static void
1085insert_vmap_area(struct vmap_area *va,
1086	struct rb_root *root, struct list_head *head)
1087{
1088	struct rb_node **link;
1089	struct rb_node *parent;
1090
1091	link = find_va_links(va, root, NULL, &parent);
1092	if (link)
1093		link_va(va, root, parent, link, head);
1094}
1095
1096static void
1097insert_vmap_area_augment(struct vmap_area *va,
1098	struct rb_node *from, struct rb_root *root,
1099	struct list_head *head)
1100{
1101	struct rb_node **link;
1102	struct rb_node *parent;
1103
1104	if (from)
1105		link = find_va_links(va, NULL, from, &parent);
1106	else
1107		link = find_va_links(va, root, NULL, &parent);
1108
1109	if (link) {
1110		link_va_augment(va, root, parent, link, head);
1111		augment_tree_propagate_from(va);
1112	}
1113}
1114
1115/*
1116 * Merge de-allocated chunk of VA memory with previous
1117 * and next free blocks. If coalesce is not done a new
1118 * free area is inserted. If VA has been merged, it is
1119 * freed.
1120 *
1121 * Please note, it can return NULL in case of overlap
1122 * ranges, followed by WARN() report. Despite it is a
1123 * buggy behaviour, a system can be alive and keep
1124 * ongoing.
1125 */
1126static __always_inline struct vmap_area *
1127__merge_or_add_vmap_area(struct vmap_area *va,
1128	struct rb_root *root, struct list_head *head, bool augment)
1129{
1130	struct vmap_area *sibling;
1131	struct list_head *next;
1132	struct rb_node **link;
1133	struct rb_node *parent;
1134	bool merged = false;
1135
1136	/*
1137	 * Find a place in the tree where VA potentially will be
1138	 * inserted, unless it is merged with its sibling/siblings.
1139	 */
1140	link = find_va_links(va, root, NULL, &parent);
1141	if (!link)
1142		return NULL;
1143
1144	/*
1145	 * Get next node of VA to check if merging can be done.
1146	 */
1147	next = get_va_next_sibling(parent, link);
1148	if (unlikely(next == NULL))
1149		goto insert;
1150
1151	/*
1152	 * start            end
1153	 * |                |
1154	 * |<------VA------>|<-----Next----->|
1155	 *                  |                |
1156	 *                  start            end
1157	 */
1158	if (next != head) {
1159		sibling = list_entry(next, struct vmap_area, list);
1160		if (sibling->va_start == va->va_end) {
1161			sibling->va_start = va->va_start;
1162
1163			/* Free vmap_area object. */
1164			kmem_cache_free(vmap_area_cachep, va);
1165
1166			/* Point to the new merged area. */
1167			va = sibling;
1168			merged = true;
1169		}
1170	}
1171
1172	/*
1173	 * start            end
1174	 * |                |
1175	 * |<-----Prev----->|<------VA------>|
1176	 *                  |                |
1177	 *                  start            end
1178	 */
1179	if (next->prev != head) {
1180		sibling = list_entry(next->prev, struct vmap_area, list);
1181		if (sibling->va_end == va->va_start) {
1182			/*
1183			 * If both neighbors are coalesced, it is important
1184			 * to unlink the "next" node first, followed by merging
1185			 * with "previous" one. Otherwise the tree might not be
1186			 * fully populated if a sibling's augmented value is
1187			 * "normalized" because of rotation operations.
1188			 */
1189			if (merged)
1190				__unlink_va(va, root, augment);
1191
1192			sibling->va_end = va->va_end;
1193
1194			/* Free vmap_area object. */
1195			kmem_cache_free(vmap_area_cachep, va);
1196
1197			/* Point to the new merged area. */
1198			va = sibling;
1199			merged = true;
1200		}
1201	}
1202
1203insert:
1204	if (!merged)
1205		__link_va(va, root, parent, link, head, augment);
1206
1207	return va;
1208}
1209
1210static __always_inline struct vmap_area *
1211merge_or_add_vmap_area(struct vmap_area *va,
1212	struct rb_root *root, struct list_head *head)
1213{
1214	return __merge_or_add_vmap_area(va, root, head, false);
1215}
1216
1217static __always_inline struct vmap_area *
1218merge_or_add_vmap_area_augment(struct vmap_area *va,
1219	struct rb_root *root, struct list_head *head)
1220{
1221	va = __merge_or_add_vmap_area(va, root, head, true);
1222	if (va)
1223		augment_tree_propagate_from(va);
1224
1225	return va;
1226}
1227
1228static __always_inline bool
1229is_within_this_va(struct vmap_area *va, unsigned long size,
1230	unsigned long align, unsigned long vstart)
1231{
1232	unsigned long nva_start_addr;
1233
1234	if (va->va_start > vstart)
1235		nva_start_addr = ALIGN(va->va_start, align);
1236	else
1237		nva_start_addr = ALIGN(vstart, align);
1238
1239	/* Can be overflowed due to big size or alignment. */
1240	if (nva_start_addr + size < nva_start_addr ||
1241			nva_start_addr < vstart)
1242		return false;
1243
1244	return (nva_start_addr + size <= va->va_end);
1245}
1246
1247/*
1248 * Find the first free block(lowest start address) in the tree,
1249 * that will accomplish the request corresponding to passing
1250 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1251 * a search length is adjusted to account for worst case alignment
1252 * overhead.
1253 */
1254static __always_inline struct vmap_area *
1255find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1256	unsigned long align, unsigned long vstart, bool adjust_search_size)
1257{
1258	struct vmap_area *va;
1259	struct rb_node *node;
1260	unsigned long length;
1261
1262	/* Start from the root. */
1263	node = root->rb_node;
1264
1265	/* Adjust the search size for alignment overhead. */
1266	length = adjust_search_size ? size + align - 1 : size;
1267
1268	while (node) {
1269		va = rb_entry(node, struct vmap_area, rb_node);
1270
1271		if (get_subtree_max_size(node->rb_left) >= length &&
1272				vstart < va->va_start) {
1273			node = node->rb_left;
1274		} else {
1275			if (is_within_this_va(va, size, align, vstart))
1276				return va;
1277
1278			/*
1279			 * Does not make sense to go deeper towards the right
1280			 * sub-tree if it does not have a free block that is
1281			 * equal or bigger to the requested search length.
1282			 */
1283			if (get_subtree_max_size(node->rb_right) >= length) {
1284				node = node->rb_right;
1285				continue;
1286			}
1287
1288			/*
1289			 * OK. We roll back and find the first right sub-tree,
1290			 * that will satisfy the search criteria. It can happen
1291			 * due to "vstart" restriction or an alignment overhead
1292			 * that is bigger then PAGE_SIZE.
1293			 */
1294			while ((node = rb_parent(node))) {
1295				va = rb_entry(node, struct vmap_area, rb_node);
1296				if (is_within_this_va(va, size, align, vstart))
1297					return va;
1298
1299				if (get_subtree_max_size(node->rb_right) >= length &&
1300						vstart <= va->va_start) {
1301					/*
1302					 * Shift the vstart forward. Please note, we update it with
1303					 * parent's start address adding "1" because we do not want
1304					 * to enter same sub-tree after it has already been checked
1305					 * and no suitable free block found there.
1306					 */
1307					vstart = va->va_start + 1;
1308					node = node->rb_right;
1309					break;
1310				}
1311			}
1312		}
1313	}
1314
1315	return NULL;
1316}
1317
1318#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1319#include <linux/random.h>
1320
1321static struct vmap_area *
1322find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1323	unsigned long align, unsigned long vstart)
1324{
1325	struct vmap_area *va;
1326
1327	list_for_each_entry(va, head, list) {
1328		if (!is_within_this_va(va, size, align, vstart))
1329			continue;
1330
1331		return va;
1332	}
1333
1334	return NULL;
1335}
1336
1337static void
1338find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1339			     unsigned long size, unsigned long align)
1340{
1341	struct vmap_area *va_1, *va_2;
1342	unsigned long vstart;
1343	unsigned int rnd;
1344
1345	get_random_bytes(&rnd, sizeof(rnd));
1346	vstart = VMALLOC_START + rnd;
1347
1348	va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1349	va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1350
1351	if (va_1 != va_2)
1352		pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1353			va_1, va_2, vstart);
1354}
1355#endif
1356
1357enum fit_type {
1358	NOTHING_FIT = 0,
1359	FL_FIT_TYPE = 1,	/* full fit */
1360	LE_FIT_TYPE = 2,	/* left edge fit */
1361	RE_FIT_TYPE = 3,	/* right edge fit */
1362	NE_FIT_TYPE = 4		/* no edge fit */
1363};
1364
1365static __always_inline enum fit_type
1366classify_va_fit_type(struct vmap_area *va,
1367	unsigned long nva_start_addr, unsigned long size)
1368{
1369	enum fit_type type;
1370
1371	/* Check if it is within VA. */
1372	if (nva_start_addr < va->va_start ||
1373			nva_start_addr + size > va->va_end)
1374		return NOTHING_FIT;
1375
1376	/* Now classify. */
1377	if (va->va_start == nva_start_addr) {
1378		if (va->va_end == nva_start_addr + size)
1379			type = FL_FIT_TYPE;
1380		else
1381			type = LE_FIT_TYPE;
1382	} else if (va->va_end == nva_start_addr + size) {
1383		type = RE_FIT_TYPE;
1384	} else {
1385		type = NE_FIT_TYPE;
1386	}
1387
1388	return type;
1389}
1390
1391static __always_inline int
1392adjust_va_to_fit_type(struct rb_root *root, struct list_head *head,
1393		      struct vmap_area *va, unsigned long nva_start_addr,
1394		      unsigned long size)
1395{
1396	struct vmap_area *lva = NULL;
1397	enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1398
1399	if (type == FL_FIT_TYPE) {
1400		/*
1401		 * No need to split VA, it fully fits.
1402		 *
1403		 * |               |
1404		 * V      NVA      V
1405		 * |---------------|
1406		 */
1407		unlink_va_augment(va, root);
1408		kmem_cache_free(vmap_area_cachep, va);
1409	} else if (type == LE_FIT_TYPE) {
1410		/*
1411		 * Split left edge of fit VA.
1412		 *
1413		 * |       |
1414		 * V  NVA  V   R
1415		 * |-------|-------|
1416		 */
1417		va->va_start += size;
1418	} else if (type == RE_FIT_TYPE) {
1419		/*
1420		 * Split right edge of fit VA.
1421		 *
1422		 *         |       |
1423		 *     L   V  NVA  V
1424		 * |-------|-------|
1425		 */
1426		va->va_end = nva_start_addr;
1427	} else if (type == NE_FIT_TYPE) {
1428		/*
1429		 * Split no edge of fit VA.
1430		 *
1431		 *     |       |
1432		 *   L V  NVA  V R
1433		 * |---|-------|---|
1434		 */
1435		lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1436		if (unlikely(!lva)) {
1437			/*
1438			 * For percpu allocator we do not do any pre-allocation
1439			 * and leave it as it is. The reason is it most likely
1440			 * never ends up with NE_FIT_TYPE splitting. In case of
1441			 * percpu allocations offsets and sizes are aligned to
1442			 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1443			 * are its main fitting cases.
1444			 *
1445			 * There are a few exceptions though, as an example it is
1446			 * a first allocation (early boot up) when we have "one"
1447			 * big free space that has to be split.
1448			 *
1449			 * Also we can hit this path in case of regular "vmap"
1450			 * allocations, if "this" current CPU was not preloaded.
1451			 * See the comment in alloc_vmap_area() why. If so, then
1452			 * GFP_NOWAIT is used instead to get an extra object for
1453			 * split purpose. That is rare and most time does not
1454			 * occur.
1455			 *
1456			 * What happens if an allocation gets failed. Basically,
1457			 * an "overflow" path is triggered to purge lazily freed
1458			 * areas to free some memory, then, the "retry" path is
1459			 * triggered to repeat one more time. See more details
1460			 * in alloc_vmap_area() function.
1461			 */
1462			lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1463			if (!lva)
1464				return -1;
1465		}
1466
1467		/*
1468		 * Build the remainder.
1469		 */
1470		lva->va_start = va->va_start;
1471		lva->va_end = nva_start_addr;
1472
1473		/*
1474		 * Shrink this VA to remaining size.
1475		 */
1476		va->va_start = nva_start_addr + size;
1477	} else {
1478		return -1;
1479	}
1480
1481	if (type != FL_FIT_TYPE) {
1482		augment_tree_propagate_from(va);
1483
1484		if (lva)	/* type == NE_FIT_TYPE */
1485			insert_vmap_area_augment(lva, &va->rb_node, root, head);
1486	}
1487
1488	return 0;
1489}
1490
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1491/*
1492 * Returns a start address of the newly allocated area, if success.
1493 * Otherwise a vend is returned that indicates failure.
1494 */
1495static __always_inline unsigned long
1496__alloc_vmap_area(struct rb_root *root, struct list_head *head,
1497	unsigned long size, unsigned long align,
1498	unsigned long vstart, unsigned long vend)
1499{
1500	bool adjust_search_size = true;
1501	unsigned long nva_start_addr;
1502	struct vmap_area *va;
1503	int ret;
1504
1505	/*
1506	 * Do not adjust when:
1507	 *   a) align <= PAGE_SIZE, because it does not make any sense.
1508	 *      All blocks(their start addresses) are at least PAGE_SIZE
1509	 *      aligned anyway;
1510	 *   b) a short range where a requested size corresponds to exactly
1511	 *      specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1512	 *      With adjusted search length an allocation would not succeed.
1513	 */
1514	if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1515		adjust_search_size = false;
1516
1517	va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1518	if (unlikely(!va))
1519		return vend;
1520
1521	if (va->va_start > vstart)
1522		nva_start_addr = ALIGN(va->va_start, align);
1523	else
1524		nva_start_addr = ALIGN(vstart, align);
1525
1526	/* Check the "vend" restriction. */
1527	if (nva_start_addr + size > vend)
1528		return vend;
1529
1530	/* Update the free vmap_area. */
1531	ret = adjust_va_to_fit_type(root, head, va, nva_start_addr, size);
1532	if (WARN_ON_ONCE(ret))
1533		return vend;
1534
1535#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1536	find_vmap_lowest_match_check(root, head, size, align);
1537#endif
1538
1539	return nva_start_addr;
1540}
1541
1542/*
1543 * Free a region of KVA allocated by alloc_vmap_area
1544 */
1545static void free_vmap_area(struct vmap_area *va)
1546{
 
 
1547	/*
1548	 * Remove from the busy tree/list.
1549	 */
1550	spin_lock(&vmap_area_lock);
1551	unlink_va(va, &vmap_area_root);
1552	spin_unlock(&vmap_area_lock);
1553
1554	/*
1555	 * Insert/Merge it back to the free tree/list.
1556	 */
1557	spin_lock(&free_vmap_area_lock);
1558	merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1559	spin_unlock(&free_vmap_area_lock);
1560}
1561
1562static inline void
1563preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1564{
1565	struct vmap_area *va = NULL;
1566
1567	/*
1568	 * Preload this CPU with one extra vmap_area object. It is used
1569	 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1570	 * a CPU that does an allocation is preloaded.
1571	 *
1572	 * We do it in non-atomic context, thus it allows us to use more
1573	 * permissive allocation masks to be more stable under low memory
1574	 * condition and high memory pressure.
1575	 */
1576	if (!this_cpu_read(ne_fit_preload_node))
1577		va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1578
1579	spin_lock(lock);
1580
1581	if (va && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, va))
 
1582		kmem_cache_free(vmap_area_cachep, va);
1583}
1584
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1585/*
1586 * Allocate a region of KVA of the specified size and alignment, within the
1587 * vstart and vend.
1588 */
1589static struct vmap_area *alloc_vmap_area(unsigned long size,
1590				unsigned long align,
1591				unsigned long vstart, unsigned long vend,
1592				int node, gfp_t gfp_mask)
 
1593{
 
1594	struct vmap_area *va;
1595	unsigned long freed;
1596	unsigned long addr;
 
1597	int purged = 0;
1598	int ret;
1599
1600	BUG_ON(!size);
1601	BUG_ON(offset_in_page(size));
1602	BUG_ON(!is_power_of_2(align));
1603
1604	if (unlikely(!vmap_initialized))
1605		return ERR_PTR(-EBUSY);
1606
1607	might_sleep();
1608	gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1609
1610	va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1611	if (unlikely(!va))
1612		return ERR_PTR(-ENOMEM);
1613
1614	/*
1615	 * Only scan the relevant parts containing pointers to other objects
1616	 * to avoid false negatives.
 
 
 
 
1617	 */
1618	kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
 
 
 
 
 
 
 
 
 
 
 
 
 
1619
1620retry:
1621	preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
1622	addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
1623		size, align, vstart, vend);
1624	spin_unlock(&free_vmap_area_lock);
 
 
1625
1626	trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
1627
1628	/*
1629	 * If an allocation fails, the "vend" address is
1630	 * returned. Therefore trigger the overflow path.
1631	 */
1632	if (unlikely(addr == vend))
1633		goto overflow;
1634
1635	va->va_start = addr;
1636	va->va_end = addr + size;
1637	va->vm = NULL;
 
 
 
 
 
 
 
 
 
1638
1639	spin_lock(&vmap_area_lock);
1640	insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1641	spin_unlock(&vmap_area_lock);
1642
1643	BUG_ON(!IS_ALIGNED(va->va_start, align));
1644	BUG_ON(va->va_start < vstart);
1645	BUG_ON(va->va_end > vend);
1646
1647	ret = kasan_populate_vmalloc(addr, size);
1648	if (ret) {
1649		free_vmap_area(va);
1650		return ERR_PTR(ret);
1651	}
1652
1653	return va;
1654
1655overflow:
1656	if (!purged) {
1657		purge_vmap_area_lazy();
1658		purged = 1;
1659		goto retry;
1660	}
1661
1662	freed = 0;
1663	blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1664
1665	if (freed > 0) {
1666		purged = 0;
1667		goto retry;
1668	}
1669
1670	if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1671		pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1672			size);
1673
1674	kmem_cache_free(vmap_area_cachep, va);
1675	return ERR_PTR(-EBUSY);
1676}
1677
1678int register_vmap_purge_notifier(struct notifier_block *nb)
1679{
1680	return blocking_notifier_chain_register(&vmap_notify_list, nb);
1681}
1682EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1683
1684int unregister_vmap_purge_notifier(struct notifier_block *nb)
1685{
1686	return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1687}
1688EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1689
1690/*
1691 * lazy_max_pages is the maximum amount of virtual address space we gather up
1692 * before attempting to purge with a TLB flush.
1693 *
1694 * There is a tradeoff here: a larger number will cover more kernel page tables
1695 * and take slightly longer to purge, but it will linearly reduce the number of
1696 * global TLB flushes that must be performed. It would seem natural to scale
1697 * this number up linearly with the number of CPUs (because vmapping activity
1698 * could also scale linearly with the number of CPUs), however it is likely
1699 * that in practice, workloads might be constrained in other ways that mean
1700 * vmap activity will not scale linearly with CPUs. Also, I want to be
1701 * conservative and not introduce a big latency on huge systems, so go with
1702 * a less aggressive log scale. It will still be an improvement over the old
1703 * code, and it will be simple to change the scale factor if we find that it
1704 * becomes a problem on bigger systems.
1705 */
1706static unsigned long lazy_max_pages(void)
1707{
1708	unsigned int log;
1709
1710	log = fls(num_online_cpus());
1711
1712	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1713}
1714
1715static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1716
1717/*
1718 * Serialize vmap purging.  There is no actual critical section protected
1719 * by this lock, but we want to avoid concurrent calls for performance
1720 * reasons and to make the pcpu_get_vm_areas more deterministic.
1721 */
1722static DEFINE_MUTEX(vmap_purge_lock);
1723
1724/* for per-CPU blocks */
1725static void purge_fragmented_blocks_allcpus(void);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1726
1727/*
1728 * Purges all lazily-freed vmap areas.
1729 */
1730static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
 
1731{
1732	unsigned long resched_threshold;
1733	unsigned int num_purged_areas = 0;
1734	struct list_head local_purge_list;
1735	struct vmap_area *va, *n_va;
 
1736
1737	lockdep_assert_held(&vmap_purge_lock);
1738
1739	spin_lock(&purge_vmap_area_lock);
1740	purge_vmap_area_root = RB_ROOT;
1741	list_replace_init(&purge_vmap_area_list, &local_purge_list);
1742	spin_unlock(&purge_vmap_area_lock);
1743
1744	if (unlikely(list_empty(&local_purge_list)))
1745		goto out;
1746
1747	start = min(start,
1748		list_first_entry(&local_purge_list,
1749			struct vmap_area, list)->va_start);
1750
1751	end = max(end,
1752		list_last_entry(&local_purge_list,
1753			struct vmap_area, list)->va_end);
1754
1755	flush_tlb_kernel_range(start, end);
1756	resched_threshold = lazy_max_pages() << 1;
 
 
1757
1758	spin_lock(&free_vmap_area_lock);
1759	list_for_each_entry_safe(va, n_va, &local_purge_list, list) {
1760		unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1761		unsigned long orig_start = va->va_start;
1762		unsigned long orig_end = va->va_end;
1763
1764		/*
1765		 * Finally insert or merge lazily-freed area. It is
1766		 * detached and there is no need to "unlink" it from
1767		 * anything.
1768		 */
1769		va = merge_or_add_vmap_area_augment(va, &free_vmap_area_root,
1770				&free_vmap_area_list);
1771
1772		if (!va)
1773			continue;
1774
1775		if (is_vmalloc_or_module_addr((void *)orig_start))
1776			kasan_release_vmalloc(orig_start, orig_end,
1777					      va->va_start, va->va_end);
1778
1779		atomic_long_sub(nr, &vmap_lazy_nr);
1780		num_purged_areas++;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1781
1782		if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1783			cond_resched_lock(&free_vmap_area_lock);
 
 
 
 
 
 
1784	}
1785	spin_unlock(&free_vmap_area_lock);
1786
1787out:
1788	trace_purge_vmap_area_lazy(start, end, num_purged_areas);
1789	return num_purged_areas > 0;
1790}
1791
1792/*
1793 * Kick off a purge of the outstanding lazy areas.
1794 */
1795static void purge_vmap_area_lazy(void)
 
1796{
1797	mutex_lock(&vmap_purge_lock);
1798	purge_fragmented_blocks_allcpus();
1799	__purge_vmap_area_lazy(ULONG_MAX, 0);
1800	mutex_unlock(&vmap_purge_lock);
1801}
1802
1803static void drain_vmap_area_work(struct work_struct *work)
1804{
1805	unsigned long nr_lazy;
1806
1807	do {
1808		mutex_lock(&vmap_purge_lock);
1809		__purge_vmap_area_lazy(ULONG_MAX, 0);
1810		mutex_unlock(&vmap_purge_lock);
1811
1812		/* Recheck if further work is required. */
1813		nr_lazy = atomic_long_read(&vmap_lazy_nr);
1814	} while (nr_lazy > lazy_max_pages());
1815}
1816
1817/*
1818 * Free a vmap area, caller ensuring that the area has been unmapped
1819 * and flush_cache_vunmap had been called for the correct range
1820 * previously.
1821 */
1822static void free_vmap_area_noflush(struct vmap_area *va)
1823{
1824	unsigned long nr_lazy_max = lazy_max_pages();
1825	unsigned long va_start = va->va_start;
 
 
1826	unsigned long nr_lazy;
1827
1828	spin_lock(&vmap_area_lock);
1829	unlink_va(va, &vmap_area_root);
1830	spin_unlock(&vmap_area_lock);
1831
1832	nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1833				PAGE_SHIFT, &vmap_lazy_nr);
1834
1835	/*
1836	 * Merge or place it to the purge tree/list.
 
1837	 */
1838	spin_lock(&purge_vmap_area_lock);
1839	merge_or_add_vmap_area(va,
1840		&purge_vmap_area_root, &purge_vmap_area_list);
1841	spin_unlock(&purge_vmap_area_lock);
 
 
1842
1843	trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
1844
1845	/* After this point, we may free va at any time */
1846	if (unlikely(nr_lazy > nr_lazy_max))
1847		schedule_work(&drain_vmap_work);
1848}
1849
1850/*
1851 * Free and unmap a vmap area
1852 */
1853static void free_unmap_vmap_area(struct vmap_area *va)
1854{
1855	flush_cache_vunmap(va->va_start, va->va_end);
1856	vunmap_range_noflush(va->va_start, va->va_end);
1857	if (debug_pagealloc_enabled_static())
1858		flush_tlb_kernel_range(va->va_start, va->va_end);
1859
1860	free_vmap_area_noflush(va);
1861}
1862
1863struct vmap_area *find_vmap_area(unsigned long addr)
1864{
 
1865	struct vmap_area *va;
 
1866
1867	spin_lock(&vmap_area_lock);
1868	va = __find_vmap_area(addr, &vmap_area_root);
1869	spin_unlock(&vmap_area_lock);
1870
1871	return va;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1872}
1873
1874/*** Per cpu kva allocator ***/
1875
1876/*
1877 * vmap space is limited especially on 32 bit architectures. Ensure there is
1878 * room for at least 16 percpu vmap blocks per CPU.
1879 */
1880/*
1881 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1882 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
1883 * instead (we just need a rough idea)
1884 */
1885#if BITS_PER_LONG == 32
1886#define VMALLOC_SPACE		(128UL*1024*1024)
1887#else
1888#define VMALLOC_SPACE		(128UL*1024*1024*1024)
1889#endif
1890
1891#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
1892#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
1893#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
1894#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
1895#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
1896#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
1897#define VMAP_BBMAP_BITS		\
1898		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
1899		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
1900			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1901
1902#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
1903
 
 
 
 
 
 
 
 
 
 
1904struct vmap_block_queue {
1905	spinlock_t lock;
1906	struct list_head free;
 
 
 
 
 
 
 
1907};
1908
1909struct vmap_block {
1910	spinlock_t lock;
1911	struct vmap_area *va;
1912	unsigned long free, dirty;
 
1913	unsigned long dirty_min, dirty_max; /*< dirty range */
1914	struct list_head free_list;
1915	struct rcu_head rcu_head;
1916	struct list_head purge;
 
1917};
1918
1919/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1920static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1921
1922/*
1923 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1924 * in the free path. Could get rid of this if we change the API to return a
1925 * "cookie" from alloc, to be passed to free. But no big deal yet.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1926 */
1927static DEFINE_XARRAY(vmap_blocks);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1928
1929/*
1930 * We should probably have a fallback mechanism to allocate virtual memory
1931 * out of partially filled vmap blocks. However vmap block sizing should be
1932 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1933 * big problem.
1934 */
1935
1936static unsigned long addr_to_vb_idx(unsigned long addr)
1937{
1938	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1939	addr /= VMAP_BLOCK_SIZE;
1940	return addr;
1941}
1942
1943static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1944{
1945	unsigned long addr;
1946
1947	addr = va_start + (pages_off << PAGE_SHIFT);
1948	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1949	return (void *)addr;
1950}
1951
1952/**
1953 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1954 *                  block. Of course pages number can't exceed VMAP_BBMAP_BITS
1955 * @order:    how many 2^order pages should be occupied in newly allocated block
1956 * @gfp_mask: flags for the page level allocator
1957 *
1958 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1959 */
1960static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1961{
1962	struct vmap_block_queue *vbq;
1963	struct vmap_block *vb;
1964	struct vmap_area *va;
 
1965	unsigned long vb_idx;
1966	int node, err;
1967	void *vaddr;
1968
1969	node = numa_node_id();
1970
1971	vb = kmalloc_node(sizeof(struct vmap_block),
1972			gfp_mask & GFP_RECLAIM_MASK, node);
1973	if (unlikely(!vb))
1974		return ERR_PTR(-ENOMEM);
1975
1976	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1977					VMALLOC_START, VMALLOC_END,
1978					node, gfp_mask);
 
1979	if (IS_ERR(va)) {
1980		kfree(vb);
1981		return ERR_CAST(va);
1982	}
1983
1984	vaddr = vmap_block_vaddr(va->va_start, 0);
1985	spin_lock_init(&vb->lock);
1986	vb->va = va;
1987	/* At least something should be left free */
1988	BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
 
1989	vb->free = VMAP_BBMAP_BITS - (1UL << order);
1990	vb->dirty = 0;
1991	vb->dirty_min = VMAP_BBMAP_BITS;
1992	vb->dirty_max = 0;
 
1993	INIT_LIST_HEAD(&vb->free_list);
 
1994
 
1995	vb_idx = addr_to_vb_idx(va->va_start);
1996	err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1997	if (err) {
1998		kfree(vb);
1999		free_vmap_area(va);
2000		return ERR_PTR(err);
2001	}
2002
2003	vbq = raw_cpu_ptr(&vmap_block_queue);
 
 
 
 
 
 
2004	spin_lock(&vbq->lock);
2005	list_add_tail_rcu(&vb->free_list, &vbq->free);
2006	spin_unlock(&vbq->lock);
2007
2008	return vaddr;
2009}
2010
2011static void free_vmap_block(struct vmap_block *vb)
2012{
 
2013	struct vmap_block *tmp;
 
2014
2015	tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
 
2016	BUG_ON(tmp != vb);
2017
 
 
 
 
 
2018	free_vmap_area_noflush(vb->va);
2019	kfree_rcu(vb, rcu_head);
2020}
2021
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2022static void purge_fragmented_blocks(int cpu)
2023{
2024	LIST_HEAD(purge);
2025	struct vmap_block *vb;
2026	struct vmap_block *n_vb;
2027	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2028
2029	rcu_read_lock();
2030	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
 
 
2031
2032		if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
 
2033			continue;
2034
2035		spin_lock(&vb->lock);
2036		if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
2037			vb->free = 0; /* prevent further allocs after releasing lock */
2038			vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
2039			vb->dirty_min = 0;
2040			vb->dirty_max = VMAP_BBMAP_BITS;
2041			spin_lock(&vbq->lock);
2042			list_del_rcu(&vb->free_list);
2043			spin_unlock(&vbq->lock);
2044			spin_unlock(&vb->lock);
2045			list_add_tail(&vb->purge, &purge);
2046		} else
2047			spin_unlock(&vb->lock);
2048	}
2049	rcu_read_unlock();
2050
2051	list_for_each_entry_safe(vb, n_vb, &purge, purge) {
2052		list_del(&vb->purge);
2053		free_vmap_block(vb);
2054	}
2055}
2056
2057static void purge_fragmented_blocks_allcpus(void)
2058{
2059	int cpu;
2060
2061	for_each_possible_cpu(cpu)
2062		purge_fragmented_blocks(cpu);
2063}
2064
2065static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2066{
2067	struct vmap_block_queue *vbq;
2068	struct vmap_block *vb;
2069	void *vaddr = NULL;
2070	unsigned int order;
2071
2072	BUG_ON(offset_in_page(size));
2073	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2074	if (WARN_ON(size == 0)) {
2075		/*
2076		 * Allocating 0 bytes isn't what caller wants since
2077		 * get_order(0) returns funny result. Just warn and terminate
2078		 * early.
2079		 */
2080		return NULL;
2081	}
2082	order = get_order(size);
2083
2084	rcu_read_lock();
2085	vbq = raw_cpu_ptr(&vmap_block_queue);
2086	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2087		unsigned long pages_off;
2088
 
 
 
2089		spin_lock(&vb->lock);
2090		if (vb->free < (1UL << order)) {
2091			spin_unlock(&vb->lock);
2092			continue;
2093		}
2094
2095		pages_off = VMAP_BBMAP_BITS - vb->free;
2096		vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2097		vb->free -= 1UL << order;
 
2098		if (vb->free == 0) {
2099			spin_lock(&vbq->lock);
2100			list_del_rcu(&vb->free_list);
2101			spin_unlock(&vbq->lock);
2102		}
2103
2104		spin_unlock(&vb->lock);
2105		break;
2106	}
2107
2108	rcu_read_unlock();
2109
2110	/* Allocate new block if nothing was found */
2111	if (!vaddr)
2112		vaddr = new_vmap_block(order, gfp_mask);
2113
2114	return vaddr;
2115}
2116
2117static void vb_free(unsigned long addr, unsigned long size)
2118{
2119	unsigned long offset;
2120	unsigned int order;
2121	struct vmap_block *vb;
 
2122
2123	BUG_ON(offset_in_page(size));
2124	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2125
2126	flush_cache_vunmap(addr, addr + size);
2127
2128	order = get_order(size);
2129	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2130	vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
 
 
 
 
 
 
2131
2132	vunmap_range_noflush(addr, addr + size);
2133
2134	if (debug_pagealloc_enabled_static())
2135		flush_tlb_kernel_range(addr, addr + size);
2136
2137	spin_lock(&vb->lock);
2138
2139	/* Expand dirty range */
2140	vb->dirty_min = min(vb->dirty_min, offset);
2141	vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2142
2143	vb->dirty += 1UL << order;
2144	if (vb->dirty == VMAP_BBMAP_BITS) {
2145		BUG_ON(vb->free);
2146		spin_unlock(&vb->lock);
2147		free_vmap_block(vb);
2148	} else
2149		spin_unlock(&vb->lock);
2150}
2151
2152static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2153{
 
2154	int cpu;
2155
2156	if (unlikely(!vmap_initialized))
2157		return;
2158
2159	might_sleep();
2160
2161	for_each_possible_cpu(cpu) {
2162		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2163		struct vmap_block *vb;
 
2164
2165		rcu_read_lock();
2166		list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2167			spin_lock(&vb->lock);
2168			if (vb->dirty && vb->dirty != VMAP_BBMAP_BITS) {
 
 
 
 
 
 
 
2169				unsigned long va_start = vb->va->va_start;
2170				unsigned long s, e;
2171
2172				s = va_start + (vb->dirty_min << PAGE_SHIFT);
2173				e = va_start + (vb->dirty_max << PAGE_SHIFT);
2174
2175				start = min(s, start);
2176				end   = max(e, end);
2177
 
 
 
 
2178				flush = 1;
2179			}
2180			spin_unlock(&vb->lock);
2181		}
2182		rcu_read_unlock();
2183	}
 
2184
2185	mutex_lock(&vmap_purge_lock);
2186	purge_fragmented_blocks_allcpus();
2187	if (!__purge_vmap_area_lazy(start, end) && flush)
2188		flush_tlb_kernel_range(start, end);
2189	mutex_unlock(&vmap_purge_lock);
2190}
2191
2192/**
2193 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2194 *
2195 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2196 * to amortize TLB flushing overheads. What this means is that any page you
2197 * have now, may, in a former life, have been mapped into kernel virtual
2198 * address by the vmap layer and so there might be some CPUs with TLB entries
2199 * still referencing that page (additional to the regular 1:1 kernel mapping).
2200 *
2201 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2202 * be sure that none of the pages we have control over will have any aliases
2203 * from the vmap layer.
2204 */
2205void vm_unmap_aliases(void)
2206{
2207	unsigned long start = ULONG_MAX, end = 0;
2208	int flush = 0;
2209
2210	_vm_unmap_aliases(start, end, flush);
2211}
2212EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2213
2214/**
2215 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2216 * @mem: the pointer returned by vm_map_ram
2217 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2218 */
2219void vm_unmap_ram(const void *mem, unsigned int count)
2220{
2221	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2222	unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2223	struct vmap_area *va;
2224
2225	might_sleep();
2226	BUG_ON(!addr);
2227	BUG_ON(addr < VMALLOC_START);
2228	BUG_ON(addr > VMALLOC_END);
2229	BUG_ON(!PAGE_ALIGNED(addr));
2230
2231	kasan_poison_vmalloc(mem, size);
2232
2233	if (likely(count <= VMAP_MAX_ALLOC)) {
2234		debug_check_no_locks_freed(mem, size);
2235		vb_free(addr, size);
2236		return;
2237	}
2238
2239	va = find_vmap_area(addr);
2240	BUG_ON(!va);
2241	debug_check_no_locks_freed((void *)va->va_start,
2242				    (va->va_end - va->va_start));
 
2243	free_unmap_vmap_area(va);
2244}
2245EXPORT_SYMBOL(vm_unmap_ram);
2246
2247/**
2248 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2249 * @pages: an array of pointers to the pages to be mapped
2250 * @count: number of pages
2251 * @node: prefer to allocate data structures on this node
2252 *
2253 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2254 * faster than vmap so it's good.  But if you mix long-life and short-life
2255 * objects with vm_map_ram(), it could consume lots of address space through
2256 * fragmentation (especially on a 32bit machine).  You could see failures in
2257 * the end.  Please use this function for short-lived objects.
2258 *
2259 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2260 */
2261void *vm_map_ram(struct page **pages, unsigned int count, int node)
2262{
2263	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2264	unsigned long addr;
2265	void *mem;
2266
2267	if (likely(count <= VMAP_MAX_ALLOC)) {
2268		mem = vb_alloc(size, GFP_KERNEL);
2269		if (IS_ERR(mem))
2270			return NULL;
2271		addr = (unsigned long)mem;
2272	} else {
2273		struct vmap_area *va;
2274		va = alloc_vmap_area(size, PAGE_SIZE,
2275				VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
 
 
2276		if (IS_ERR(va))
2277			return NULL;
2278
2279		addr = va->va_start;
2280		mem = (void *)addr;
2281	}
2282
2283	if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2284				pages, PAGE_SHIFT) < 0) {
2285		vm_unmap_ram(mem, count);
2286		return NULL;
2287	}
2288
2289	/*
2290	 * Mark the pages as accessible, now that they are mapped.
2291	 * With hardware tag-based KASAN, marking is skipped for
2292	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2293	 */
2294	mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
2295
2296	return mem;
2297}
2298EXPORT_SYMBOL(vm_map_ram);
2299
2300static struct vm_struct *vmlist __initdata;
2301
2302static inline unsigned int vm_area_page_order(struct vm_struct *vm)
2303{
2304#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2305	return vm->page_order;
2306#else
2307	return 0;
2308#endif
2309}
2310
 
 
 
 
 
2311static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
2312{
2313#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
2314	vm->page_order = order;
2315#else
2316	BUG_ON(order != 0);
2317#endif
2318}
2319
2320/**
2321 * vm_area_add_early - add vmap area early during boot
2322 * @vm: vm_struct to add
2323 *
2324 * This function is used to add fixed kernel vm area to vmlist before
2325 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
2326 * should contain proper values and the other fields should be zero.
2327 *
2328 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2329 */
2330void __init vm_area_add_early(struct vm_struct *vm)
2331{
2332	struct vm_struct *tmp, **p;
2333
2334	BUG_ON(vmap_initialized);
2335	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
2336		if (tmp->addr >= vm->addr) {
2337			BUG_ON(tmp->addr < vm->addr + vm->size);
2338			break;
2339		} else
2340			BUG_ON(tmp->addr + tmp->size > vm->addr);
2341	}
2342	vm->next = *p;
2343	*p = vm;
2344}
2345
2346/**
2347 * vm_area_register_early - register vmap area early during boot
2348 * @vm: vm_struct to register
2349 * @align: requested alignment
2350 *
2351 * This function is used to register kernel vm area before
2352 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
2353 * proper values on entry and other fields should be zero.  On return,
2354 * vm->addr contains the allocated address.
2355 *
2356 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
2357 */
2358void __init vm_area_register_early(struct vm_struct *vm, size_t align)
2359{
2360	unsigned long addr = ALIGN(VMALLOC_START, align);
2361	struct vm_struct *cur, **p;
2362
2363	BUG_ON(vmap_initialized);
2364
2365	for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
2366		if ((unsigned long)cur->addr - addr >= vm->size)
2367			break;
2368		addr = ALIGN((unsigned long)cur->addr + cur->size, align);
2369	}
2370
2371	BUG_ON(addr > VMALLOC_END - vm->size);
2372	vm->addr = (void *)addr;
2373	vm->next = *p;
2374	*p = vm;
2375	kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
2376}
2377
2378static void vmap_init_free_space(void)
2379{
2380	unsigned long vmap_start = 1;
2381	const unsigned long vmap_end = ULONG_MAX;
2382	struct vmap_area *busy, *free;
2383
2384	/*
2385	 *     B     F     B     B     B     F
2386	 * -|-----|.....|-----|-----|-----|.....|-
2387	 *  |           The KVA space           |
2388	 *  |<--------------------------------->|
2389	 */
2390	list_for_each_entry(busy, &vmap_area_list, list) {
2391		if (busy->va_start - vmap_start > 0) {
2392			free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2393			if (!WARN_ON_ONCE(!free)) {
2394				free->va_start = vmap_start;
2395				free->va_end = busy->va_start;
2396
2397				insert_vmap_area_augment(free, NULL,
2398					&free_vmap_area_root,
2399						&free_vmap_area_list);
2400			}
2401		}
2402
2403		vmap_start = busy->va_end;
2404	}
2405
2406	if (vmap_end - vmap_start > 0) {
2407		free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2408		if (!WARN_ON_ONCE(!free)) {
2409			free->va_start = vmap_start;
2410			free->va_end = vmap_end;
2411
2412			insert_vmap_area_augment(free, NULL,
2413				&free_vmap_area_root,
2414					&free_vmap_area_list);
2415		}
2416	}
2417}
2418
2419void __init vmalloc_init(void)
2420{
2421	struct vmap_area *va;
2422	struct vm_struct *tmp;
2423	int i;
2424
2425	/*
2426	 * Create the cache for vmap_area objects.
2427	 */
2428	vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
2429
2430	for_each_possible_cpu(i) {
2431		struct vmap_block_queue *vbq;
2432		struct vfree_deferred *p;
2433
2434		vbq = &per_cpu(vmap_block_queue, i);
2435		spin_lock_init(&vbq->lock);
2436		INIT_LIST_HEAD(&vbq->free);
2437		p = &per_cpu(vfree_deferred, i);
2438		init_llist_head(&p->list);
2439		INIT_WORK(&p->wq, free_work);
2440	}
2441
2442	/* Import existing vmlist entries. */
2443	for (tmp = vmlist; tmp; tmp = tmp->next) {
2444		va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
2445		if (WARN_ON_ONCE(!va))
2446			continue;
2447
2448		va->va_start = (unsigned long)tmp->addr;
2449		va->va_end = va->va_start + tmp->size;
2450		va->vm = tmp;
2451		insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2452	}
2453
2454	/*
2455	 * Now we can initialize a free vmap space.
2456	 */
2457	vmap_init_free_space();
2458	vmap_initialized = true;
2459}
2460
2461static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2462	struct vmap_area *va, unsigned long flags, const void *caller)
2463{
2464	vm->flags = flags;
2465	vm->addr = (void *)va->va_start;
2466	vm->size = va->va_end - va->va_start;
2467	vm->caller = caller;
2468	va->vm = vm;
2469}
2470
2471static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2472			      unsigned long flags, const void *caller)
2473{
2474	spin_lock(&vmap_area_lock);
2475	setup_vmalloc_vm_locked(vm, va, flags, caller);
2476	spin_unlock(&vmap_area_lock);
2477}
2478
2479static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2480{
2481	/*
2482	 * Before removing VM_UNINITIALIZED,
2483	 * we should make sure that vm has proper values.
2484	 * Pair with smp_rmb() in show_numa_info().
2485	 */
2486	smp_wmb();
2487	vm->flags &= ~VM_UNINITIALIZED;
2488}
2489
2490static struct vm_struct *__get_vm_area_node(unsigned long size,
2491		unsigned long align, unsigned long shift, unsigned long flags,
2492		unsigned long start, unsigned long end, int node,
2493		gfp_t gfp_mask, const void *caller)
2494{
2495	struct vmap_area *va;
2496	struct vm_struct *area;
2497	unsigned long requested_size = size;
2498
2499	BUG_ON(in_interrupt());
2500	size = ALIGN(size, 1ul << shift);
2501	if (unlikely(!size))
2502		return NULL;
2503
2504	if (flags & VM_IOREMAP)
2505		align = 1ul << clamp_t(int, get_count_order_long(size),
2506				       PAGE_SHIFT, IOREMAP_MAX_ORDER);
2507
2508	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2509	if (unlikely(!area))
2510		return NULL;
2511
2512	if (!(flags & VM_NO_GUARD))
2513		size += PAGE_SIZE;
2514
2515	va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
 
 
 
2516	if (IS_ERR(va)) {
2517		kfree(area);
2518		return NULL;
2519	}
2520
2521	setup_vmalloc_vm(area, va, flags, caller);
2522
2523	/*
2524	 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
2525	 * best-effort approach, as they can be mapped outside of vmalloc code.
2526	 * For VM_ALLOC mappings, the pages are marked as accessible after
2527	 * getting mapped in __vmalloc_node_range().
2528	 * With hardware tag-based KASAN, marking is skipped for
2529	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
2530	 */
2531	if (!(flags & VM_ALLOC))
2532		area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
2533						    KASAN_VMALLOC_PROT_NORMAL);
2534
2535	return area;
2536}
2537
2538struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2539				       unsigned long start, unsigned long end,
2540				       const void *caller)
2541{
2542	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
2543				  NUMA_NO_NODE, GFP_KERNEL, caller);
2544}
2545
2546/**
2547 * get_vm_area - reserve a contiguous kernel virtual area
2548 * @size:	 size of the area
2549 * @flags:	 %VM_IOREMAP for I/O mappings or VM_ALLOC
2550 *
2551 * Search an area of @size in the kernel virtual mapping area,
2552 * and reserved it for out purposes.  Returns the area descriptor
2553 * on success or %NULL on failure.
2554 *
2555 * Return: the area descriptor on success or %NULL on failure.
2556 */
2557struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2558{
2559	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2560				  VMALLOC_START, VMALLOC_END,
2561				  NUMA_NO_NODE, GFP_KERNEL,
2562				  __builtin_return_address(0));
2563}
2564
2565struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2566				const void *caller)
2567{
2568	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
2569				  VMALLOC_START, VMALLOC_END,
2570				  NUMA_NO_NODE, GFP_KERNEL, caller);
2571}
2572
2573/**
2574 * find_vm_area - find a continuous kernel virtual area
2575 * @addr:	  base address
2576 *
2577 * Search for the kernel VM area starting at @addr, and return it.
2578 * It is up to the caller to do all required locking to keep the returned
2579 * pointer valid.
2580 *
2581 * Return: the area descriptor on success or %NULL on failure.
2582 */
2583struct vm_struct *find_vm_area(const void *addr)
2584{
2585	struct vmap_area *va;
2586
2587	va = find_vmap_area((unsigned long)addr);
2588	if (!va)
2589		return NULL;
2590
2591	return va->vm;
2592}
2593
2594/**
2595 * remove_vm_area - find and remove a continuous kernel virtual area
2596 * @addr:	    base address
2597 *
2598 * Search for the kernel VM area starting at @addr, and remove it.
2599 * This function returns the found VM area, but using it is NOT safe
2600 * on SMP machines, except for its size or flags.
2601 *
2602 * Return: the area descriptor on success or %NULL on failure.
2603 */
2604struct vm_struct *remove_vm_area(const void *addr)
2605{
2606	struct vmap_area *va;
 
2607
2608	might_sleep();
2609
2610	spin_lock(&vmap_area_lock);
2611	va = __find_vmap_area((unsigned long)addr, &vmap_area_root);
2612	if (va && va->vm) {
2613		struct vm_struct *vm = va->vm;
2614
2615		va->vm = NULL;
2616		spin_unlock(&vmap_area_lock);
2617
2618		kasan_free_module_shadow(vm);
2619		free_unmap_vmap_area(va);
 
 
2620
2621		return vm;
2622	}
 
 
2623
2624	spin_unlock(&vmap_area_lock);
2625	return NULL;
2626}
2627
2628static inline void set_area_direct_map(const struct vm_struct *area,
2629				       int (*set_direct_map)(struct page *page))
2630{
2631	int i;
2632
2633	/* HUGE_VMALLOC passes small pages to set_direct_map */
2634	for (i = 0; i < area->nr_pages; i++)
2635		if (page_address(area->pages[i]))
2636			set_direct_map(area->pages[i]);
2637}
2638
2639/* Handle removing and resetting vm mappings related to the vm_struct. */
2640static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
 
 
2641{
2642	unsigned long start = ULONG_MAX, end = 0;
2643	unsigned int page_order = vm_area_page_order(area);
2644	int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2645	int flush_dmap = 0;
2646	int i;
2647
2648	remove_vm_area(area->addr);
2649
2650	/* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2651	if (!flush_reset)
2652		return;
2653
2654	/*
2655	 * If not deallocating pages, just do the flush of the VM area and
2656	 * return.
2657	 */
2658	if (!deallocate_pages) {
2659		vm_unmap_aliases();
2660		return;
2661	}
2662
2663	/*
2664	 * If execution gets here, flush the vm mapping and reset the direct
2665	 * map. Find the start and end range of the direct mappings to make sure
2666	 * the vm_unmap_aliases() flush includes the direct map.
2667	 */
2668	for (i = 0; i < area->nr_pages; i += 1U << page_order) {
2669		unsigned long addr = (unsigned long)page_address(area->pages[i]);
 
2670		if (addr) {
2671			unsigned long page_size;
2672
2673			page_size = PAGE_SIZE << page_order;
2674			start = min(addr, start);
2675			end = max(addr + page_size, end);
2676			flush_dmap = 1;
2677		}
2678	}
2679
2680	/*
2681	 * Set direct map to something invalid so that it won't be cached if
2682	 * there are any accesses after the TLB flush, then flush the TLB and
2683	 * reset the direct map permissions to the default.
2684	 */
2685	set_area_direct_map(area, set_direct_map_invalid_noflush);
2686	_vm_unmap_aliases(start, end, flush_dmap);
2687	set_area_direct_map(area, set_direct_map_default_noflush);
2688}
2689
2690static void __vunmap(const void *addr, int deallocate_pages)
2691{
2692	struct vm_struct *area;
2693
2694	if (!addr)
2695		return;
2696
2697	if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2698			addr))
2699		return;
2700
2701	area = find_vm_area(addr);
2702	if (unlikely(!area)) {
2703		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2704				addr);
2705		return;
2706	}
2707
2708	debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2709	debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2710
2711	kasan_poison_vmalloc(area->addr, get_vm_area_size(area));
2712
2713	vm_remove_mappings(area, deallocate_pages);
2714
2715	if (deallocate_pages) {
2716		int i;
2717
2718		for (i = 0; i < area->nr_pages; i++) {
2719			struct page *page = area->pages[i];
2720
2721			BUG_ON(!page);
2722			mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
2723			/*
2724			 * High-order allocs for huge vmallocs are split, so
2725			 * can be freed as an array of order-0 allocations
2726			 */
2727			__free_pages(page, 0);
2728			cond_resched();
2729		}
2730		atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2731
2732		kvfree(area->pages);
2733	}
2734
2735	kfree(area);
2736}
2737
2738static inline void __vfree_deferred(const void *addr)
2739{
2740	/*
2741	 * Use raw_cpu_ptr() because this can be called from preemptible
2742	 * context. Preemption is absolutely fine here, because the llist_add()
2743	 * implementation is lockless, so it works even if we are adding to
2744	 * another cpu's list. schedule_work() should be fine with this too.
2745	 */
2746	struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2747
2748	if (llist_add((struct llist_node *)addr, &p->list))
2749		schedule_work(&p->wq);
2750}
2751
2752/**
2753 * vfree_atomic - release memory allocated by vmalloc()
2754 * @addr:	  memory base address
2755 *
2756 * This one is just like vfree() but can be called in any atomic context
2757 * except NMIs.
2758 */
2759void vfree_atomic(const void *addr)
2760{
2761	BUG_ON(in_nmi());
2762
 
2763	kmemleak_free(addr);
2764
2765	if (!addr)
2766		return;
2767	__vfree_deferred(addr);
2768}
2769
2770static void __vfree(const void *addr)
2771{
2772	if (unlikely(in_interrupt()))
2773		__vfree_deferred(addr);
2774	else
2775		__vunmap(addr, 1);
2776}
2777
2778/**
2779 * vfree - Release memory allocated by vmalloc()
2780 * @addr:  Memory base address
2781 *
2782 * Free the virtually continuous memory area starting at @addr, as obtained
2783 * from one of the vmalloc() family of APIs.  This will usually also free the
2784 * physical memory underlying the virtual allocation, but that memory is
2785 * reference counted, so it will not be freed until the last user goes away.
2786 *
2787 * If @addr is NULL, no operation is performed.
2788 *
2789 * Context:
2790 * May sleep if called *not* from interrupt context.
2791 * Must not be called in NMI context (strictly speaking, it could be
2792 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2793 * conventions for vfree() arch-dependent would be a really bad idea).
2794 */
2795void vfree(const void *addr)
2796{
2797	BUG_ON(in_nmi());
 
2798
2799	kmemleak_free(addr);
 
 
 
2800
2801	might_sleep_if(!in_interrupt());
 
 
2802
2803	if (!addr)
2804		return;
2805
2806	__vfree(addr);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2807}
2808EXPORT_SYMBOL(vfree);
2809
2810/**
2811 * vunmap - release virtual mapping obtained by vmap()
2812 * @addr:   memory base address
2813 *
2814 * Free the virtually contiguous memory area starting at @addr,
2815 * which was created from the page array passed to vmap().
2816 *
2817 * Must not be called in interrupt context.
2818 */
2819void vunmap(const void *addr)
2820{
 
 
2821	BUG_ON(in_interrupt());
2822	might_sleep();
2823	if (addr)
2824		__vunmap(addr, 0);
 
 
 
 
 
 
 
 
2825}
2826EXPORT_SYMBOL(vunmap);
2827
2828/**
2829 * vmap - map an array of pages into virtually contiguous space
2830 * @pages: array of page pointers
2831 * @count: number of pages to map
2832 * @flags: vm_area->flags
2833 * @prot: page protection for the mapping
2834 *
2835 * Maps @count pages from @pages into contiguous kernel virtual space.
2836 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
2837 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
2838 * are transferred from the caller to vmap(), and will be freed / dropped when
2839 * vfree() is called on the return value.
2840 *
2841 * Return: the address of the area or %NULL on failure
2842 */
2843void *vmap(struct page **pages, unsigned int count,
2844	   unsigned long flags, pgprot_t prot)
2845{
2846	struct vm_struct *area;
2847	unsigned long addr;
2848	unsigned long size;		/* In bytes */
2849
2850	might_sleep();
2851
 
 
 
2852	/*
2853	 * Your top guard is someone else's bottom guard. Not having a top
2854	 * guard compromises someone else's mappings too.
2855	 */
2856	if (WARN_ON_ONCE(flags & VM_NO_GUARD))
2857		flags &= ~VM_NO_GUARD;
2858
2859	if (count > totalram_pages())
2860		return NULL;
2861
2862	size = (unsigned long)count << PAGE_SHIFT;
2863	area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2864	if (!area)
2865		return NULL;
2866
2867	addr = (unsigned long)area->addr;
2868	if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
2869				pages, PAGE_SHIFT) < 0) {
2870		vunmap(area->addr);
2871		return NULL;
2872	}
2873
2874	if (flags & VM_MAP_PUT_PAGES) {
2875		area->pages = pages;
2876		area->nr_pages = count;
2877	}
2878	return area->addr;
2879}
2880EXPORT_SYMBOL(vmap);
2881
2882#ifdef CONFIG_VMAP_PFN
2883struct vmap_pfn_data {
2884	unsigned long	*pfns;
2885	pgprot_t	prot;
2886	unsigned int	idx;
2887};
2888
2889static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
2890{
2891	struct vmap_pfn_data *data = private;
 
 
2892
2893	if (WARN_ON_ONCE(pfn_valid(data->pfns[data->idx])))
2894		return -EINVAL;
2895	*pte = pte_mkspecial(pfn_pte(data->pfns[data->idx++], data->prot));
 
 
 
 
2896	return 0;
2897}
2898
2899/**
2900 * vmap_pfn - map an array of PFNs into virtually contiguous space
2901 * @pfns: array of PFNs
2902 * @count: number of pages to map
2903 * @prot: page protection for the mapping
2904 *
2905 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
2906 * the start address of the mapping.
2907 */
2908void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
2909{
2910	struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
2911	struct vm_struct *area;
2912
2913	area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
2914			__builtin_return_address(0));
2915	if (!area)
2916		return NULL;
2917	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
2918			count * PAGE_SIZE, vmap_pfn_apply, &data)) {
2919		free_vm_area(area);
2920		return NULL;
2921	}
 
 
 
 
2922	return area->addr;
2923}
2924EXPORT_SYMBOL_GPL(vmap_pfn);
2925#endif /* CONFIG_VMAP_PFN */
2926
2927static inline unsigned int
2928vm_area_alloc_pages(gfp_t gfp, int nid,
2929		unsigned int order, unsigned int nr_pages, struct page **pages)
2930{
2931	unsigned int nr_allocated = 0;
2932	struct page *page;
2933	int i;
2934
2935	/*
2936	 * For order-0 pages we make use of bulk allocator, if
2937	 * the page array is partly or not at all populated due
2938	 * to fails, fallback to a single page allocator that is
2939	 * more permissive.
2940	 */
2941	if (!order) {
2942		gfp_t bulk_gfp = gfp & ~__GFP_NOFAIL;
2943
2944		while (nr_allocated < nr_pages) {
2945			unsigned int nr, nr_pages_request;
2946
2947			/*
2948			 * A maximum allowed request is hard-coded and is 100
2949			 * pages per call. That is done in order to prevent a
2950			 * long preemption off scenario in the bulk-allocator
2951			 * so the range is [1:100].
2952			 */
2953			nr_pages_request = min(100U, nr_pages - nr_allocated);
2954
2955			/* memory allocation should consider mempolicy, we can't
2956			 * wrongly use nearest node when nid == NUMA_NO_NODE,
2957			 * otherwise memory may be allocated in only one node,
2958			 * but mempolicy wants to alloc memory by interleaving.
2959			 */
2960			if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
2961				nr = alloc_pages_bulk_array_mempolicy(bulk_gfp,
2962							nr_pages_request,
2963							pages + nr_allocated);
2964
2965			else
2966				nr = alloc_pages_bulk_array_node(bulk_gfp, nid,
2967							nr_pages_request,
2968							pages + nr_allocated);
2969
2970			nr_allocated += nr;
2971			cond_resched();
2972
2973			/*
2974			 * If zero or pages were obtained partly,
2975			 * fallback to a single page allocator.
2976			 */
2977			if (nr != nr_pages_request)
2978				break;
2979		}
2980	}
2981
2982	/* High-order pages or fallback path if "bulk" fails. */
2983
2984	while (nr_allocated < nr_pages) {
2985		if (fatal_signal_pending(current))
2986			break;
2987
2988		if (nid == NUMA_NO_NODE)
2989			page = alloc_pages(gfp, order);
2990		else
2991			page = alloc_pages_node(nid, gfp, order);
 
2992		if (unlikely(!page))
2993			break;
 
2994		/*
2995		 * Higher order allocations must be able to be treated as
2996		 * indepdenent small pages by callers (as they can with
2997		 * small-page vmallocs). Some drivers do their own refcounting
2998		 * on vmalloc_to_page() pages, some use page->mapping,
2999		 * page->lru, etc.
3000		 */
3001		if (order)
3002			split_page(page, order);
3003
3004		/*
3005		 * Careful, we allocate and map page-order pages, but
3006		 * tracking is done per PAGE_SIZE page so as to keep the
3007		 * vm_struct APIs independent of the physical/mapped size.
3008		 */
3009		for (i = 0; i < (1U << order); i++)
3010			pages[nr_allocated + i] = page + i;
3011
3012		cond_resched();
3013		nr_allocated += 1U << order;
3014	}
3015
3016	return nr_allocated;
3017}
3018
3019static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3020				 pgprot_t prot, unsigned int page_shift,
3021				 int node)
3022{
3023	const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3024	bool nofail = gfp_mask & __GFP_NOFAIL;
3025	unsigned long addr = (unsigned long)area->addr;
3026	unsigned long size = get_vm_area_size(area);
3027	unsigned long array_size;
3028	unsigned int nr_small_pages = size >> PAGE_SHIFT;
3029	unsigned int page_order;
3030	unsigned int flags;
3031	int ret;
3032
3033	array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3034	gfp_mask |= __GFP_NOWARN;
3035	if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3036		gfp_mask |= __GFP_HIGHMEM;
3037
3038	/* Please note that the recursion is strictly bounded. */
3039	if (array_size > PAGE_SIZE) {
3040		area->pages = __vmalloc_node(array_size, 1, nested_gfp, node,
3041					area->caller);
3042	} else {
3043		area->pages = kmalloc_node(array_size, nested_gfp, node);
3044	}
3045
3046	if (!area->pages) {
3047		warn_alloc(gfp_mask, NULL,
3048			"vmalloc error: size %lu, failed to allocated page array size %lu",
3049			nr_small_pages * PAGE_SIZE, array_size);
3050		free_vm_area(area);
3051		return NULL;
3052	}
3053
3054	set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3055	page_order = vm_area_page_order(area);
3056
3057	area->nr_pages = vm_area_alloc_pages(gfp_mask | __GFP_NOWARN,
 
 
 
 
 
 
 
 
 
3058		node, page_order, nr_small_pages, area->pages);
3059
3060	atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3061	if (gfp_mask & __GFP_ACCOUNT) {
3062		int i;
3063
3064		for (i = 0; i < area->nr_pages; i++)
3065			mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3066	}
3067
3068	/*
3069	 * If not enough pages were obtained to accomplish an
3070	 * allocation request, free them via __vfree() if any.
3071	 */
3072	if (area->nr_pages != nr_small_pages) {
3073		warn_alloc(gfp_mask, NULL,
3074			"vmalloc error: size %lu, page order %u, failed to allocate pages",
3075			area->nr_pages * PAGE_SIZE, page_order);
 
 
 
 
 
 
 
 
 
 
 
3076		goto fail;
3077	}
3078
3079	/*
3080	 * page tables allocations ignore external gfp mask, enforce it
3081	 * by the scope API
3082	 */
3083	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3084		flags = memalloc_nofs_save();
3085	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3086		flags = memalloc_noio_save();
3087
3088	do {
3089		ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3090			page_shift);
3091		if (nofail && (ret < 0))
3092			schedule_timeout_uninterruptible(1);
3093	} while (nofail && (ret < 0));
3094
3095	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3096		memalloc_nofs_restore(flags);
3097	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3098		memalloc_noio_restore(flags);
3099
3100	if (ret < 0) {
3101		warn_alloc(gfp_mask, NULL,
3102			"vmalloc error: size %lu, failed to map pages",
3103			area->nr_pages * PAGE_SIZE);
3104		goto fail;
3105	}
3106
3107	return area->addr;
3108
3109fail:
3110	__vfree(area->addr);
3111	return NULL;
3112}
3113
3114/**
3115 * __vmalloc_node_range - allocate virtually contiguous memory
3116 * @size:		  allocation size
3117 * @align:		  desired alignment
3118 * @start:		  vm area range start
3119 * @end:		  vm area range end
3120 * @gfp_mask:		  flags for the page level allocator
3121 * @prot:		  protection mask for the allocated pages
3122 * @vm_flags:		  additional vm area flags (e.g. %VM_NO_GUARD)
3123 * @node:		  node to use for allocation or NUMA_NO_NODE
3124 * @caller:		  caller's return address
3125 *
3126 * Allocate enough pages to cover @size from the page level
3127 * allocator with @gfp_mask flags. Please note that the full set of gfp
3128 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3129 * supported.
3130 * Zone modifiers are not supported. From the reclaim modifiers
3131 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3132 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3133 * __GFP_RETRY_MAYFAIL are not supported).
3134 *
3135 * __GFP_NOWARN can be used to suppress failures messages.
3136 *
3137 * Map them into contiguous kernel virtual space, using a pagetable
3138 * protection of @prot.
3139 *
3140 * Return: the address of the area or %NULL on failure
3141 */
3142void *__vmalloc_node_range(unsigned long size, unsigned long align,
3143			unsigned long start, unsigned long end, gfp_t gfp_mask,
3144			pgprot_t prot, unsigned long vm_flags, int node,
3145			const void *caller)
3146{
3147	struct vm_struct *area;
3148	void *ret;
3149	kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3150	unsigned long real_size = size;
3151	unsigned long real_align = align;
3152	unsigned int shift = PAGE_SHIFT;
3153
3154	if (WARN_ON_ONCE(!size))
3155		return NULL;
3156
3157	if ((size >> PAGE_SHIFT) > totalram_pages()) {
3158		warn_alloc(gfp_mask, NULL,
3159			"vmalloc error: size %lu, exceeds total pages",
3160			real_size);
3161		return NULL;
3162	}
3163
3164	if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
3165		unsigned long size_per_node;
3166
3167		/*
3168		 * Try huge pages. Only try for PAGE_KERNEL allocations,
3169		 * others like modules don't yet expect huge pages in
3170		 * their allocations due to apply_to_page_range not
3171		 * supporting them.
3172		 */
3173
3174		size_per_node = size;
3175		if (node == NUMA_NO_NODE)
3176			size_per_node /= num_online_nodes();
3177		if (arch_vmap_pmd_supported(prot) && size_per_node >= PMD_SIZE)
3178			shift = PMD_SHIFT;
3179		else
3180			shift = arch_vmap_pte_supported_shift(size_per_node);
3181
3182		align = max(real_align, 1UL << shift);
3183		size = ALIGN(real_size, 1UL << shift);
3184	}
3185
3186again:
3187	area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3188				  VM_UNINITIALIZED | vm_flags, start, end, node,
3189				  gfp_mask, caller);
3190	if (!area) {
3191		bool nofail = gfp_mask & __GFP_NOFAIL;
3192		warn_alloc(gfp_mask, NULL,
3193			"vmalloc error: size %lu, vm_struct allocation failed%s",
3194			real_size, (nofail) ? ". Retrying." : "");
3195		if (nofail) {
3196			schedule_timeout_uninterruptible(1);
3197			goto again;
3198		}
3199		goto fail;
3200	}
3201
3202	/*
3203	 * Prepare arguments for __vmalloc_area_node() and
3204	 * kasan_unpoison_vmalloc().
3205	 */
3206	if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3207		if (kasan_hw_tags_enabled()) {
3208			/*
3209			 * Modify protection bits to allow tagging.
3210			 * This must be done before mapping.
3211			 */
3212			prot = arch_vmap_pgprot_tagged(prot);
3213
3214			/*
3215			 * Skip page_alloc poisoning and zeroing for physical
3216			 * pages backing VM_ALLOC mapping. Memory is instead
3217			 * poisoned and zeroed by kasan_unpoison_vmalloc().
3218			 */
3219			gfp_mask |= __GFP_SKIP_KASAN_UNPOISON | __GFP_SKIP_ZERO;
3220		}
3221
3222		/* Take note that the mapping is PAGE_KERNEL. */
3223		kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3224	}
3225
3226	/* Allocate physical pages and map them into vmalloc space. */
3227	ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3228	if (!ret)
3229		goto fail;
3230
3231	/*
3232	 * Mark the pages as accessible, now that they are mapped.
3233	 * The condition for setting KASAN_VMALLOC_INIT should complement the
3234	 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3235	 * to make sure that memory is initialized under the same conditions.
3236	 * Tag-based KASAN modes only assign tags to normal non-executable
3237	 * allocations, see __kasan_unpoison_vmalloc().
3238	 */
3239	kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3240	if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3241	    (gfp_mask & __GFP_SKIP_ZERO))
3242		kasan_flags |= KASAN_VMALLOC_INIT;
3243	/* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3244	area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3245
3246	/*
3247	 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3248	 * flag. It means that vm_struct is not fully initialized.
3249	 * Now, it is fully initialized, so remove this flag here.
3250	 */
3251	clear_vm_uninitialized_flag(area);
3252
3253	size = PAGE_ALIGN(size);
3254	if (!(vm_flags & VM_DEFER_KMEMLEAK))
3255		kmemleak_vmalloc(area, size, gfp_mask);
3256
3257	return area->addr;
3258
3259fail:
3260	if (shift > PAGE_SHIFT) {
3261		shift = PAGE_SHIFT;
3262		align = real_align;
3263		size = real_size;
3264		goto again;
3265	}
3266
3267	return NULL;
3268}
3269
3270/**
3271 * __vmalloc_node - allocate virtually contiguous memory
3272 * @size:	    allocation size
3273 * @align:	    desired alignment
3274 * @gfp_mask:	    flags for the page level allocator
3275 * @node:	    node to use for allocation or NUMA_NO_NODE
3276 * @caller:	    caller's return address
3277 *
3278 * Allocate enough pages to cover @size from the page level allocator with
3279 * @gfp_mask flags.  Map them into contiguous kernel virtual space.
3280 *
3281 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3282 * and __GFP_NOFAIL are not supported
3283 *
3284 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3285 * with mm people.
3286 *
3287 * Return: pointer to the allocated memory or %NULL on error
3288 */
3289void *__vmalloc_node(unsigned long size, unsigned long align,
3290			    gfp_t gfp_mask, int node, const void *caller)
3291{
3292	return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
3293				gfp_mask, PAGE_KERNEL, 0, node, caller);
3294}
3295/*
3296 * This is only for performance analysis of vmalloc and stress purpose.
3297 * It is required by vmalloc test module, therefore do not use it other
3298 * than that.
3299 */
3300#ifdef CONFIG_TEST_VMALLOC_MODULE
3301EXPORT_SYMBOL_GPL(__vmalloc_node);
3302#endif
3303
3304void *__vmalloc(unsigned long size, gfp_t gfp_mask)
3305{
3306	return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
3307				__builtin_return_address(0));
3308}
3309EXPORT_SYMBOL(__vmalloc);
3310
3311/**
3312 * vmalloc - allocate virtually contiguous memory
3313 * @size:    allocation size
3314 *
3315 * Allocate enough pages to cover @size from the page level
3316 * allocator and map them into contiguous kernel virtual space.
3317 *
3318 * For tight control over page level allocator and protection flags
3319 * use __vmalloc() instead.
3320 *
3321 * Return: pointer to the allocated memory or %NULL on error
3322 */
3323void *vmalloc(unsigned long size)
3324{
3325	return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3326				__builtin_return_address(0));
3327}
3328EXPORT_SYMBOL(vmalloc);
3329
3330/**
3331 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3332 * @size:      allocation size
3333 * @gfp_mask:  flags for the page level allocator
3334 *
3335 * Allocate enough pages to cover @size from the page level
3336 * allocator and map them into contiguous kernel virtual space.
3337 * If @size is greater than or equal to PMD_SIZE, allow using
3338 * huge pages for the memory
3339 *
3340 * Return: pointer to the allocated memory or %NULL on error
3341 */
3342void *vmalloc_huge(unsigned long size, gfp_t gfp_mask)
3343{
3344	return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
3345				    gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3346				    NUMA_NO_NODE, __builtin_return_address(0));
3347}
3348EXPORT_SYMBOL_GPL(vmalloc_huge);
3349
3350/**
3351 * vzalloc - allocate virtually contiguous memory with zero fill
3352 * @size:    allocation size
3353 *
3354 * Allocate enough pages to cover @size from the page level
3355 * allocator and map them into contiguous kernel virtual space.
3356 * The memory allocated is set to zero.
3357 *
3358 * For tight control over page level allocator and protection flags
3359 * use __vmalloc() instead.
3360 *
3361 * Return: pointer to the allocated memory or %NULL on error
3362 */
3363void *vzalloc(unsigned long size)
3364{
3365	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3366				__builtin_return_address(0));
3367}
3368EXPORT_SYMBOL(vzalloc);
3369
3370/**
3371 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3372 * @size: allocation size
3373 *
3374 * The resulting memory area is zeroed so it can be mapped to userspace
3375 * without leaking data.
3376 *
3377 * Return: pointer to the allocated memory or %NULL on error
3378 */
3379void *vmalloc_user(unsigned long size)
3380{
3381	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
3382				    GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
3383				    VM_USERMAP, NUMA_NO_NODE,
3384				    __builtin_return_address(0));
3385}
3386EXPORT_SYMBOL(vmalloc_user);
3387
3388/**
3389 * vmalloc_node - allocate memory on a specific node
3390 * @size:	  allocation size
3391 * @node:	  numa node
3392 *
3393 * Allocate enough pages to cover @size from the page level
3394 * allocator and map them into contiguous kernel virtual space.
3395 *
3396 * For tight control over page level allocator and protection flags
3397 * use __vmalloc() instead.
3398 *
3399 * Return: pointer to the allocated memory or %NULL on error
3400 */
3401void *vmalloc_node(unsigned long size, int node)
3402{
3403	return __vmalloc_node(size, 1, GFP_KERNEL, node,
3404			__builtin_return_address(0));
3405}
3406EXPORT_SYMBOL(vmalloc_node);
3407
3408/**
3409 * vzalloc_node - allocate memory on a specific node with zero fill
3410 * @size:	allocation size
3411 * @node:	numa node
3412 *
3413 * Allocate enough pages to cover @size from the page level
3414 * allocator and map them into contiguous kernel virtual space.
3415 * The memory allocated is set to zero.
3416 *
3417 * Return: pointer to the allocated memory or %NULL on error
3418 */
3419void *vzalloc_node(unsigned long size, int node)
3420{
3421	return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
3422				__builtin_return_address(0));
3423}
3424EXPORT_SYMBOL(vzalloc_node);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3425
3426#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
3427#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3428#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
3429#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
3430#else
3431/*
3432 * 64b systems should always have either DMA or DMA32 zones. For others
3433 * GFP_DMA32 should do the right thing and use the normal zone.
3434 */
3435#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
3436#endif
3437
3438/**
3439 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
3440 * @size:	allocation size
3441 *
3442 * Allocate enough 32bit PA addressable pages to cover @size from the
3443 * page level allocator and map them into contiguous kernel virtual space.
3444 *
3445 * Return: pointer to the allocated memory or %NULL on error
3446 */
3447void *vmalloc_32(unsigned long size)
3448{
3449	return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
3450			__builtin_return_address(0));
3451}
3452EXPORT_SYMBOL(vmalloc_32);
3453
3454/**
3455 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
3456 * @size:	     allocation size
3457 *
3458 * The resulting memory area is 32bit addressable and zeroed so it can be
3459 * mapped to userspace without leaking data.
3460 *
3461 * Return: pointer to the allocated memory or %NULL on error
3462 */
3463void *vmalloc_32_user(unsigned long size)
3464{
3465	return __vmalloc_node_range(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
3466				    GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
3467				    VM_USERMAP, NUMA_NO_NODE,
3468				    __builtin_return_address(0));
3469}
3470EXPORT_SYMBOL(vmalloc_32_user);
3471
3472/*
3473 * small helper routine , copy contents to buf from addr.
3474 * If the page is not present, fill zero.
 
3475 */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3476
3477static int aligned_vread(char *buf, char *addr, unsigned long count)
 
 
 
 
 
 
 
3478{
3479	struct page *p;
3480	int copied = 0;
3481
3482	while (count) {
3483		unsigned long offset, length;
 
3484
3485		offset = offset_in_page(addr);
3486		length = PAGE_SIZE - offset;
3487		if (length > count)
3488			length = count;
3489		p = vmalloc_to_page(addr);
3490		/*
3491		 * To do safe access to this _mapped_ area, we need
3492		 * lock. But adding lock here means that we need to add
3493		 * overhead of vmalloc()/vfree() calls for this _debug_
3494		 * interface, rarely used. Instead of that, we'll use
3495		 * kmap() and get small overhead in this access function.
 
3496		 */
3497		if (p) {
3498			/* We can expect USER0 is not used -- see vread() */
3499			void *map = kmap_atomic(p);
3500			memcpy(buf, map + offset, length);
3501			kunmap_atomic(map);
3502		} else
3503			memset(buf, 0, length);
3504
3505		addr += length;
3506		buf += length;
3507		copied += length;
3508		count -= length;
 
3509	}
3510	return copied;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3511}
3512
3513/**
3514 * vread() - read vmalloc area in a safe way.
3515 * @buf:     buffer for reading data
3516 * @addr:    vm address.
3517 * @count:   number of bytes to be read.
3518 *
3519 * This function checks that addr is a valid vmalloc'ed area, and
3520 * copy data from that area to a given buffer. If the given memory range
3521 * of [addr...addr+count) includes some valid address, data is copied to
3522 * proper area of @buf. If there are memory holes, they'll be zero-filled.
3523 * IOREMAP area is treated as memory hole and no copy is done.
3524 *
3525 * If [addr...addr+count) doesn't includes any intersects with alive
3526 * vm_struct area, returns 0. @buf should be kernel's buffer.
3527 *
3528 * Note: In usual ops, vread() is never necessary because the caller
3529 * should know vmalloc() area is valid and can use memcpy().
3530 * This is for routines which have to access vmalloc area without
3531 * any information, as /proc/kcore.
3532 *
3533 * Return: number of bytes for which addr and buf should be increased
3534 * (same number as @count) or %0 if [addr...addr+count) doesn't
3535 * include any intersection with valid vmalloc area
3536 */
3537long vread(char *buf, char *addr, unsigned long count)
3538{
 
3539	struct vmap_area *va;
3540	struct vm_struct *vm;
3541	char *vaddr, *buf_start = buf;
3542	unsigned long buflen = count;
3543	unsigned long n;
3544
3545	addr = kasan_reset_tag(addr);
3546
3547	/* Don't allow overflow */
3548	if ((unsigned long) addr + count < count)
3549		count = -(unsigned long) addr;
3550
3551	spin_lock(&vmap_area_lock);
3552	va = find_vmap_area_exceed_addr((unsigned long)addr);
3553	if (!va)
3554		goto finished;
 
3555
3556	/* no intersects with alive vmap_area */
3557	if ((unsigned long)addr + count <= va->va_start)
3558		goto finished;
3559
3560	list_for_each_entry_from(va, &vmap_area_list, list) {
3561		if (!count)
3562			break;
3563
3564		if (!va->vm)
3565			continue;
3566
3567		vm = va->vm;
3568		vaddr = (char *) vm->addr;
3569		if (addr >= vaddr + get_vm_area_size(vm))
3570			continue;
3571		while (addr < vaddr) {
3572			if (count == 0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3573				goto finished;
3574			*buf = '\0';
3575			buf++;
3576			addr++;
3577			count--;
3578		}
3579		n = vaddr + get_vm_area_size(vm) - addr;
3580		if (n > count)
3581			n = count;
3582		if (!(vm->flags & VM_IOREMAP))
3583			aligned_vread(buf, addr, n);
3584		else /* IOREMAP area is treated as memory hole */
3585			memset(buf, 0, n);
3586		buf += n;
3587		addr += n;
3588		count -= n;
3589	}
3590finished:
3591	spin_unlock(&vmap_area_lock);
3592
3593	if (buf == buf_start)
3594		return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3595	/* zero-fill memory holes */
3596	if (buf != buf_start + buflen)
3597		memset(buf, 0, buflen - (buf - buf_start));
 
 
 
3598
3599	return buflen;
3600}
3601
3602/**
3603 * remap_vmalloc_range_partial - map vmalloc pages to userspace
3604 * @vma:		vma to cover
3605 * @uaddr:		target user address to start at
3606 * @kaddr:		virtual address of vmalloc kernel memory
3607 * @pgoff:		offset from @kaddr to start at
3608 * @size:		size of map area
3609 *
3610 * Returns:	0 for success, -Exxx on failure
3611 *
3612 * This function checks that @kaddr is a valid vmalloc'ed area,
3613 * and that it is big enough to cover the range starting at
3614 * @uaddr in @vma. Will return failure if that criteria isn't
3615 * met.
3616 *
3617 * Similar to remap_pfn_range() (see mm/memory.c)
3618 */
3619int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
3620				void *kaddr, unsigned long pgoff,
3621				unsigned long size)
3622{
3623	struct vm_struct *area;
3624	unsigned long off;
3625	unsigned long end_index;
3626
3627	if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
3628		return -EINVAL;
3629
3630	size = PAGE_ALIGN(size);
3631
3632	if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
3633		return -EINVAL;
3634
3635	area = find_vm_area(kaddr);
3636	if (!area)
3637		return -EINVAL;
3638
3639	if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3640		return -EINVAL;
3641
3642	if (check_add_overflow(size, off, &end_index) ||
3643	    end_index > get_vm_area_size(area))
3644		return -EINVAL;
3645	kaddr += off;
3646
3647	do {
3648		struct page *page = vmalloc_to_page(kaddr);
3649		int ret;
3650
3651		ret = vm_insert_page(vma, uaddr, page);
3652		if (ret)
3653			return ret;
3654
3655		uaddr += PAGE_SIZE;
3656		kaddr += PAGE_SIZE;
3657		size -= PAGE_SIZE;
3658	} while (size > 0);
3659
3660	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3661
3662	return 0;
3663}
3664
3665/**
3666 * remap_vmalloc_range - map vmalloc pages to userspace
3667 * @vma:		vma to cover (map full range of vma)
3668 * @addr:		vmalloc memory
3669 * @pgoff:		number of pages into addr before first page to map
3670 *
3671 * Returns:	0 for success, -Exxx on failure
3672 *
3673 * This function checks that addr is a valid vmalloc'ed area, and
3674 * that it is big enough to cover the vma. Will return failure if
3675 * that criteria isn't met.
3676 *
3677 * Similar to remap_pfn_range() (see mm/memory.c)
3678 */
3679int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3680						unsigned long pgoff)
3681{
3682	return remap_vmalloc_range_partial(vma, vma->vm_start,
3683					   addr, pgoff,
3684					   vma->vm_end - vma->vm_start);
3685}
3686EXPORT_SYMBOL(remap_vmalloc_range);
3687
3688void free_vm_area(struct vm_struct *area)
3689{
3690	struct vm_struct *ret;
3691	ret = remove_vm_area(area->addr);
3692	BUG_ON(ret != area);
3693	kfree(area);
3694}
3695EXPORT_SYMBOL_GPL(free_vm_area);
3696
3697#ifdef CONFIG_SMP
3698static struct vmap_area *node_to_va(struct rb_node *n)
3699{
3700	return rb_entry_safe(n, struct vmap_area, rb_node);
3701}
3702
3703/**
3704 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3705 * @addr: target address
3706 *
3707 * Returns: vmap_area if it is found. If there is no such area
3708 *   the first highest(reverse order) vmap_area is returned
3709 *   i.e. va->va_start < addr && va->va_end < addr or NULL
3710 *   if there are no any areas before @addr.
3711 */
3712static struct vmap_area *
3713pvm_find_va_enclose_addr(unsigned long addr)
3714{
3715	struct vmap_area *va, *tmp;
3716	struct rb_node *n;
3717
3718	n = free_vmap_area_root.rb_node;
3719	va = NULL;
3720
3721	while (n) {
3722		tmp = rb_entry(n, struct vmap_area, rb_node);
3723		if (tmp->va_start <= addr) {
3724			va = tmp;
3725			if (tmp->va_end >= addr)
3726				break;
3727
3728			n = n->rb_right;
3729		} else {
3730			n = n->rb_left;
3731		}
3732	}
3733
3734	return va;
3735}
3736
3737/**
3738 * pvm_determine_end_from_reverse - find the highest aligned address
3739 * of free block below VMALLOC_END
3740 * @va:
3741 *   in - the VA we start the search(reverse order);
3742 *   out - the VA with the highest aligned end address.
3743 * @align: alignment for required highest address
3744 *
3745 * Returns: determined end address within vmap_area
3746 */
3747static unsigned long
3748pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3749{
3750	unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3751	unsigned long addr;
3752
3753	if (likely(*va)) {
3754		list_for_each_entry_from_reverse((*va),
3755				&free_vmap_area_list, list) {
3756			addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3757			if ((*va)->va_start < addr)
3758				return addr;
3759		}
3760	}
3761
3762	return 0;
3763}
3764
3765/**
3766 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3767 * @offsets: array containing offset of each area
3768 * @sizes: array containing size of each area
3769 * @nr_vms: the number of areas to allocate
3770 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3771 *
3772 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3773 *	    vm_structs on success, %NULL on failure
3774 *
3775 * Percpu allocator wants to use congruent vm areas so that it can
3776 * maintain the offsets among percpu areas.  This function allocates
3777 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
3778 * be scattered pretty far, distance between two areas easily going up
3779 * to gigabytes.  To avoid interacting with regular vmallocs, these
3780 * areas are allocated from top.
3781 *
3782 * Despite its complicated look, this allocator is rather simple. It
3783 * does everything top-down and scans free blocks from the end looking
3784 * for matching base. While scanning, if any of the areas do not fit the
3785 * base address is pulled down to fit the area. Scanning is repeated till
3786 * all the areas fit and then all necessary data structures are inserted
3787 * and the result is returned.
3788 */
3789struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3790				     const size_t *sizes, int nr_vms,
3791				     size_t align)
3792{
3793	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3794	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3795	struct vmap_area **vas, *va;
3796	struct vm_struct **vms;
3797	int area, area2, last_area, term_area;
3798	unsigned long base, start, size, end, last_end, orig_start, orig_end;
3799	bool purged = false;
3800
3801	/* verify parameters and allocate data structures */
3802	BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3803	for (last_area = 0, area = 0; area < nr_vms; area++) {
3804		start = offsets[area];
3805		end = start + sizes[area];
3806
3807		/* is everything aligned properly? */
3808		BUG_ON(!IS_ALIGNED(offsets[area], align));
3809		BUG_ON(!IS_ALIGNED(sizes[area], align));
3810
3811		/* detect the area with the highest address */
3812		if (start > offsets[last_area])
3813			last_area = area;
3814
3815		for (area2 = area + 1; area2 < nr_vms; area2++) {
3816			unsigned long start2 = offsets[area2];
3817			unsigned long end2 = start2 + sizes[area2];
3818
3819			BUG_ON(start2 < end && start < end2);
3820		}
3821	}
3822	last_end = offsets[last_area] + sizes[last_area];
3823
3824	if (vmalloc_end - vmalloc_start < last_end) {
3825		WARN_ON(true);
3826		return NULL;
3827	}
3828
3829	vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3830	vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3831	if (!vas || !vms)
3832		goto err_free2;
3833
3834	for (area = 0; area < nr_vms; area++) {
3835		vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3836		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3837		if (!vas[area] || !vms[area])
3838			goto err_free;
3839	}
3840retry:
3841	spin_lock(&free_vmap_area_lock);
3842
3843	/* start scanning - we scan from the top, begin with the last area */
3844	area = term_area = last_area;
3845	start = offsets[area];
3846	end = start + sizes[area];
3847
3848	va = pvm_find_va_enclose_addr(vmalloc_end);
3849	base = pvm_determine_end_from_reverse(&va, align) - end;
3850
3851	while (true) {
3852		/*
3853		 * base might have underflowed, add last_end before
3854		 * comparing.
3855		 */
3856		if (base + last_end < vmalloc_start + last_end)
3857			goto overflow;
3858
3859		/*
3860		 * Fitting base has not been found.
3861		 */
3862		if (va == NULL)
3863			goto overflow;
3864
3865		/*
3866		 * If required width exceeds current VA block, move
3867		 * base downwards and then recheck.
3868		 */
3869		if (base + end > va->va_end) {
3870			base = pvm_determine_end_from_reverse(&va, align) - end;
3871			term_area = area;
3872			continue;
3873		}
3874
3875		/*
3876		 * If this VA does not fit, move base downwards and recheck.
3877		 */
3878		if (base + start < va->va_start) {
3879			va = node_to_va(rb_prev(&va->rb_node));
3880			base = pvm_determine_end_from_reverse(&va, align) - end;
3881			term_area = area;
3882			continue;
3883		}
3884
3885		/*
3886		 * This area fits, move on to the previous one.  If
3887		 * the previous one is the terminal one, we're done.
3888		 */
3889		area = (area + nr_vms - 1) % nr_vms;
3890		if (area == term_area)
3891			break;
3892
3893		start = offsets[area];
3894		end = start + sizes[area];
3895		va = pvm_find_va_enclose_addr(base + end);
3896	}
3897
3898	/* we've found a fitting base, insert all va's */
3899	for (area = 0; area < nr_vms; area++) {
3900		int ret;
3901
3902		start = base + offsets[area];
3903		size = sizes[area];
3904
3905		va = pvm_find_va_enclose_addr(start);
3906		if (WARN_ON_ONCE(va == NULL))
3907			/* It is a BUG(), but trigger recovery instead. */
3908			goto recovery;
3909
3910		ret = adjust_va_to_fit_type(&free_vmap_area_root,
3911					    &free_vmap_area_list,
3912					    va, start, size);
3913		if (WARN_ON_ONCE(unlikely(ret)))
3914			/* It is a BUG(), but trigger recovery instead. */
3915			goto recovery;
3916
3917		/* Allocated area. */
3918		va = vas[area];
3919		va->va_start = start;
3920		va->va_end = start + size;
3921	}
3922
3923	spin_unlock(&free_vmap_area_lock);
3924
3925	/* populate the kasan shadow space */
3926	for (area = 0; area < nr_vms; area++) {
3927		if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3928			goto err_free_shadow;
3929	}
3930
3931	/* insert all vm's */
3932	spin_lock(&vmap_area_lock);
3933	for (area = 0; area < nr_vms; area++) {
3934		insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3935
3936		setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
 
 
3937				 pcpu_get_vm_areas);
 
3938	}
3939	spin_unlock(&vmap_area_lock);
3940
3941	/*
3942	 * Mark allocated areas as accessible. Do it now as a best-effort
3943	 * approach, as they can be mapped outside of vmalloc code.
3944	 * With hardware tag-based KASAN, marking is skipped for
3945	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3946	 */
3947	for (area = 0; area < nr_vms; area++)
3948		vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
3949				vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
3950
3951	kfree(vas);
3952	return vms;
3953
3954recovery:
3955	/*
3956	 * Remove previously allocated areas. There is no
3957	 * need in removing these areas from the busy tree,
3958	 * because they are inserted only on the final step
3959	 * and when pcpu_get_vm_areas() is success.
3960	 */
3961	while (area--) {
3962		orig_start = vas[area]->va_start;
3963		orig_end = vas[area]->va_end;
3964		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
3965				&free_vmap_area_list);
3966		if (va)
3967			kasan_release_vmalloc(orig_start, orig_end,
3968				va->va_start, va->va_end);
 
3969		vas[area] = NULL;
3970	}
3971
3972overflow:
3973	spin_unlock(&free_vmap_area_lock);
3974	if (!purged) {
3975		purge_vmap_area_lazy();
3976		purged = true;
3977
3978		/* Before "retry", check if we recover. */
3979		for (area = 0; area < nr_vms; area++) {
3980			if (vas[area])
3981				continue;
3982
3983			vas[area] = kmem_cache_zalloc(
3984				vmap_area_cachep, GFP_KERNEL);
3985			if (!vas[area])
3986				goto err_free;
3987		}
3988
3989		goto retry;
3990	}
3991
3992err_free:
3993	for (area = 0; area < nr_vms; area++) {
3994		if (vas[area])
3995			kmem_cache_free(vmap_area_cachep, vas[area]);
3996
3997		kfree(vms[area]);
3998	}
3999err_free2:
4000	kfree(vas);
4001	kfree(vms);
4002	return NULL;
4003
4004err_free_shadow:
4005	spin_lock(&free_vmap_area_lock);
4006	/*
4007	 * We release all the vmalloc shadows, even the ones for regions that
4008	 * hadn't been successfully added. This relies on kasan_release_vmalloc
4009	 * being able to tolerate this case.
4010	 */
4011	for (area = 0; area < nr_vms; area++) {
4012		orig_start = vas[area]->va_start;
4013		orig_end = vas[area]->va_end;
4014		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4015				&free_vmap_area_list);
4016		if (va)
4017			kasan_release_vmalloc(orig_start, orig_end,
4018				va->va_start, va->va_end);
 
4019		vas[area] = NULL;
4020		kfree(vms[area]);
4021	}
4022	spin_unlock(&free_vmap_area_lock);
4023	kfree(vas);
4024	kfree(vms);
4025	return NULL;
4026}
4027
4028/**
4029 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4030 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4031 * @nr_vms: the number of allocated areas
4032 *
4033 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4034 */
4035void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4036{
4037	int i;
4038
4039	for (i = 0; i < nr_vms; i++)
4040		free_vm_area(vms[i]);
4041	kfree(vms);
4042}
4043#endif	/* CONFIG_SMP */
4044
4045#ifdef CONFIG_PRINTK
4046bool vmalloc_dump_obj(void *object)
4047{
 
4048	struct vm_struct *vm;
4049	void *objp = (void *)PAGE_ALIGN((unsigned long)object);
 
 
 
 
 
 
4050
4051	vm = find_vm_area(objp);
4052	if (!vm)
4053		return false;
4054	pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4055		vm->nr_pages, (unsigned long)vm->addr, vm->caller);
4056	return true;
4057}
4058#endif
4059
4060#ifdef CONFIG_PROC_FS
4061static void *s_start(struct seq_file *m, loff_t *pos)
4062	__acquires(&vmap_purge_lock)
4063	__acquires(&vmap_area_lock)
4064{
4065	mutex_lock(&vmap_purge_lock);
4066	spin_lock(&vmap_area_lock);
4067
4068	return seq_list_start(&vmap_area_list, *pos);
4069}
 
 
 
4070
4071static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4072{
4073	return seq_list_next(p, &vmap_area_list, pos);
4074}
4075
4076static void s_stop(struct seq_file *m, void *p)
4077	__releases(&vmap_area_lock)
4078	__releases(&vmap_purge_lock)
4079{
4080	spin_unlock(&vmap_area_lock);
4081	mutex_unlock(&vmap_purge_lock);
4082}
 
4083
 
4084static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4085{
4086	if (IS_ENABLED(CONFIG_NUMA)) {
4087		unsigned int nr, *counters = m->private;
4088		unsigned int step = 1U << vm_area_page_order(v);
4089
4090		if (!counters)
4091			return;
4092
4093		if (v->flags & VM_UNINITIALIZED)
4094			return;
4095		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4096		smp_rmb();
4097
4098		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4099
4100		for (nr = 0; nr < v->nr_pages; nr += step)
4101			counters[page_to_nid(v->pages[nr])] += step;
4102		for_each_node_state(nr, N_HIGH_MEMORY)
4103			if (counters[nr])
4104				seq_printf(m, " N%u=%u", nr, counters[nr]);
4105	}
4106}
4107
4108static void show_purge_info(struct seq_file *m)
4109{
 
4110	struct vmap_area *va;
 
4111
4112	spin_lock(&purge_vmap_area_lock);
4113	list_for_each_entry(va, &purge_vmap_area_list, list) {
4114		seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4115			(void *)va->va_start, (void *)va->va_end,
4116			va->va_end - va->va_start);
 
 
 
 
 
4117	}
4118	spin_unlock(&purge_vmap_area_lock);
4119}
4120
4121static int s_show(struct seq_file *m, void *p)
4122{
 
4123	struct vmap_area *va;
4124	struct vm_struct *v;
 
4125
4126	va = list_entry(p, struct vmap_area, list);
 
4127
4128	/*
4129	 * s_show can encounter race with remove_vm_area, !vm on behalf
4130	 * of vmap area is being tear down or vm_map_ram allocation.
4131	 */
4132	if (!va->vm) {
4133		seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4134			(void *)va->va_start, (void *)va->va_end,
4135			va->va_end - va->va_start);
4136
4137		goto final;
4138	}
4139
4140	v = va->vm;
4141
4142	seq_printf(m, "0x%pK-0x%pK %7ld",
4143		v->addr, v->addr + v->size, v->size);
4144
4145	if (v->caller)
4146		seq_printf(m, " %pS", v->caller);
4147
4148	if (v->nr_pages)
4149		seq_printf(m, " pages=%d", v->nr_pages);
4150
4151	if (v->phys_addr)
4152		seq_printf(m, " phys=%pa", &v->phys_addr);
4153
4154	if (v->flags & VM_IOREMAP)
4155		seq_puts(m, " ioremap");
4156
4157	if (v->flags & VM_ALLOC)
4158		seq_puts(m, " vmalloc");
4159
4160	if (v->flags & VM_MAP)
4161		seq_puts(m, " vmap");
4162
4163	if (v->flags & VM_USERMAP)
4164		seq_puts(m, " user");
4165
4166	if (v->flags & VM_DMA_COHERENT)
4167		seq_puts(m, " dma-coherent");
4168
4169	if (is_vmalloc_addr(v->pages))
4170		seq_puts(m, " vpages");
4171
4172	show_numa_info(m, v);
4173	seq_putc(m, '\n');
 
 
 
 
 
 
4174
4175	/*
4176	 * As a final step, dump "unpurged" areas.
4177	 */
4178final:
4179	if (list_is_last(&va->list, &vmap_area_list))
4180		show_purge_info(m);
4181
4182	return 0;
4183}
4184
4185static const struct seq_operations vmalloc_op = {
4186	.start = s_start,
4187	.next = s_next,
4188	.stop = s_stop,
4189	.show = s_show,
4190};
4191
4192static int __init proc_vmalloc_init(void)
4193{
 
 
4194	if (IS_ENABLED(CONFIG_NUMA))
4195		proc_create_seq_private("vmallocinfo", 0400, NULL,
4196				&vmalloc_op,
4197				nr_node_ids * sizeof(unsigned int), NULL);
4198	else
4199		proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
4200	return 0;
4201}
4202module_init(proc_vmalloc_init);
4203
4204#endif
v6.13.7
   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#include <linux/page_owner.h>
  46
  47#define CREATE_TRACE_POINTS
  48#include <trace/events/vmalloc.h>
  49
  50#include "internal.h"
  51#include "pgalloc-track.h"
  52
  53#ifdef CONFIG_HAVE_ARCH_HUGE_VMAP
  54static unsigned int __ro_after_init ioremap_max_page_shift = BITS_PER_LONG - 1;
  55
  56static int __init set_nohugeiomap(char *str)
  57{
  58	ioremap_max_page_shift = PAGE_SHIFT;
  59	return 0;
  60}
  61early_param("nohugeiomap", set_nohugeiomap);
  62#else /* CONFIG_HAVE_ARCH_HUGE_VMAP */
  63static const unsigned int ioremap_max_page_shift = PAGE_SHIFT;
  64#endif	/* CONFIG_HAVE_ARCH_HUGE_VMAP */
  65
  66#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
  67static bool __ro_after_init vmap_allow_huge = true;
  68
  69static int __init set_nohugevmalloc(char *str)
  70{
  71	vmap_allow_huge = false;
  72	return 0;
  73}
  74early_param("nohugevmalloc", set_nohugevmalloc);
  75#else /* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
  76static const bool vmap_allow_huge = false;
  77#endif	/* CONFIG_HAVE_ARCH_HUGE_VMALLOC */
  78
  79bool is_vmalloc_addr(const void *x)
  80{
  81	unsigned long addr = (unsigned long)kasan_reset_tag(x);
  82
  83	return addr >= VMALLOC_START && addr < VMALLOC_END;
  84}
  85EXPORT_SYMBOL(is_vmalloc_addr);
  86
  87struct vfree_deferred {
  88	struct llist_head list;
  89	struct work_struct wq;
  90};
  91static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
  92
 
 
 
 
 
 
 
 
 
 
 
  93/*** Page table manipulation functions ***/
  94static int vmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
  95			phys_addr_t phys_addr, pgprot_t prot,
  96			unsigned int max_page_shift, pgtbl_mod_mask *mask)
  97{
  98	pte_t *pte;
  99	u64 pfn;
 100	struct page *page;
 101	unsigned long size = PAGE_SIZE;
 102
 103	pfn = phys_addr >> PAGE_SHIFT;
 104	pte = pte_alloc_kernel_track(pmd, addr, mask);
 105	if (!pte)
 106		return -ENOMEM;
 107	do {
 108		if (unlikely(!pte_none(ptep_get(pte)))) {
 109			if (pfn_valid(pfn)) {
 110				page = pfn_to_page(pfn);
 111				dump_page(page, "remapping already mapped page");
 112			}
 113			BUG();
 114		}
 115
 116#ifdef CONFIG_HUGETLB_PAGE
 117		size = arch_vmap_pte_range_map_size(addr, end, pfn, max_page_shift);
 118		if (size != PAGE_SIZE) {
 119			pte_t entry = pfn_pte(pfn, prot);
 120
 121			entry = arch_make_huge_pte(entry, ilog2(size), 0);
 122			set_huge_pte_at(&init_mm, addr, pte, entry, size);
 123			pfn += PFN_DOWN(size);
 124			continue;
 125		}
 126#endif
 127		set_pte_at(&init_mm, addr, pte, pfn_pte(pfn, prot));
 128		pfn++;
 129	} while (pte += PFN_DOWN(size), addr += size, addr != end);
 130	*mask |= PGTBL_PTE_MODIFIED;
 131	return 0;
 132}
 133
 134static int vmap_try_huge_pmd(pmd_t *pmd, unsigned long addr, unsigned long end,
 135			phys_addr_t phys_addr, pgprot_t prot,
 136			unsigned int max_page_shift)
 137{
 138	if (max_page_shift < PMD_SHIFT)
 139		return 0;
 140
 141	if (!arch_vmap_pmd_supported(prot))
 142		return 0;
 143
 144	if ((end - addr) != PMD_SIZE)
 145		return 0;
 146
 147	if (!IS_ALIGNED(addr, PMD_SIZE))
 148		return 0;
 149
 150	if (!IS_ALIGNED(phys_addr, PMD_SIZE))
 151		return 0;
 152
 153	if (pmd_present(*pmd) && !pmd_free_pte_page(pmd, addr))
 154		return 0;
 155
 156	return pmd_set_huge(pmd, phys_addr, prot);
 157}
 158
 159static int vmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 160			phys_addr_t phys_addr, pgprot_t prot,
 161			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 162{
 163	pmd_t *pmd;
 164	unsigned long next;
 165
 166	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 167	if (!pmd)
 168		return -ENOMEM;
 169	do {
 170		next = pmd_addr_end(addr, end);
 171
 172		if (vmap_try_huge_pmd(pmd, addr, next, phys_addr, prot,
 173					max_page_shift)) {
 174			*mask |= PGTBL_PMD_MODIFIED;
 175			continue;
 176		}
 177
 178		if (vmap_pte_range(pmd, addr, next, phys_addr, prot, max_page_shift, mask))
 179			return -ENOMEM;
 180	} while (pmd++, phys_addr += (next - addr), addr = next, addr != end);
 181	return 0;
 182}
 183
 184static int vmap_try_huge_pud(pud_t *pud, unsigned long addr, unsigned long end,
 185			phys_addr_t phys_addr, pgprot_t prot,
 186			unsigned int max_page_shift)
 187{
 188	if (max_page_shift < PUD_SHIFT)
 189		return 0;
 190
 191	if (!arch_vmap_pud_supported(prot))
 192		return 0;
 193
 194	if ((end - addr) != PUD_SIZE)
 195		return 0;
 196
 197	if (!IS_ALIGNED(addr, PUD_SIZE))
 198		return 0;
 199
 200	if (!IS_ALIGNED(phys_addr, PUD_SIZE))
 201		return 0;
 202
 203	if (pud_present(*pud) && !pud_free_pmd_page(pud, addr))
 204		return 0;
 205
 206	return pud_set_huge(pud, phys_addr, prot);
 207}
 208
 209static int vmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 210			phys_addr_t phys_addr, pgprot_t prot,
 211			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 212{
 213	pud_t *pud;
 214	unsigned long next;
 215
 216	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 217	if (!pud)
 218		return -ENOMEM;
 219	do {
 220		next = pud_addr_end(addr, end);
 221
 222		if (vmap_try_huge_pud(pud, addr, next, phys_addr, prot,
 223					max_page_shift)) {
 224			*mask |= PGTBL_PUD_MODIFIED;
 225			continue;
 226		}
 227
 228		if (vmap_pmd_range(pud, addr, next, phys_addr, prot,
 229					max_page_shift, mask))
 230			return -ENOMEM;
 231	} while (pud++, phys_addr += (next - addr), addr = next, addr != end);
 232	return 0;
 233}
 234
 235static int vmap_try_huge_p4d(p4d_t *p4d, unsigned long addr, unsigned long end,
 236			phys_addr_t phys_addr, pgprot_t prot,
 237			unsigned int max_page_shift)
 238{
 239	if (max_page_shift < P4D_SHIFT)
 240		return 0;
 241
 242	if (!arch_vmap_p4d_supported(prot))
 243		return 0;
 244
 245	if ((end - addr) != P4D_SIZE)
 246		return 0;
 247
 248	if (!IS_ALIGNED(addr, P4D_SIZE))
 249		return 0;
 250
 251	if (!IS_ALIGNED(phys_addr, P4D_SIZE))
 252		return 0;
 253
 254	if (p4d_present(*p4d) && !p4d_free_pud_page(p4d, addr))
 255		return 0;
 256
 257	return p4d_set_huge(p4d, phys_addr, prot);
 258}
 259
 260static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 261			phys_addr_t phys_addr, pgprot_t prot,
 262			unsigned int max_page_shift, pgtbl_mod_mask *mask)
 263{
 264	p4d_t *p4d;
 265	unsigned long next;
 266
 267	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 268	if (!p4d)
 269		return -ENOMEM;
 270	do {
 271		next = p4d_addr_end(addr, end);
 272
 273		if (vmap_try_huge_p4d(p4d, addr, next, phys_addr, prot,
 274					max_page_shift)) {
 275			*mask |= PGTBL_P4D_MODIFIED;
 276			continue;
 277		}
 278
 279		if (vmap_pud_range(p4d, addr, next, phys_addr, prot,
 280					max_page_shift, mask))
 281			return -ENOMEM;
 282	} while (p4d++, phys_addr += (next - addr), addr = next, addr != end);
 283	return 0;
 284}
 285
 286static int vmap_range_noflush(unsigned long addr, unsigned long end,
 287			phys_addr_t phys_addr, pgprot_t prot,
 288			unsigned int max_page_shift)
 289{
 290	pgd_t *pgd;
 291	unsigned long start;
 292	unsigned long next;
 293	int err;
 294	pgtbl_mod_mask mask = 0;
 295
 296	might_sleep();
 297	BUG_ON(addr >= end);
 298
 299	start = addr;
 300	pgd = pgd_offset_k(addr);
 301	do {
 302		next = pgd_addr_end(addr, end);
 303		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
 304					max_page_shift, &mask);
 305		if (err)
 306			break;
 307	} while (pgd++, phys_addr += (next - addr), addr = next, addr != end);
 308
 309	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 310		arch_sync_kernel_mappings(start, end);
 311
 312	return err;
 313}
 314
 315int vmap_page_range(unsigned long addr, unsigned long end,
 316		    phys_addr_t phys_addr, pgprot_t prot)
 317{
 318	int err;
 319
 320	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
 321				 ioremap_max_page_shift);
 322	flush_cache_vmap(addr, end);
 323	if (!err)
 324		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
 325					       ioremap_max_page_shift);
 326	return err;
 327}
 328
 329int ioremap_page_range(unsigned long addr, unsigned long end,
 330		phys_addr_t phys_addr, pgprot_t prot)
 331{
 332	struct vm_struct *area;
 333
 334	area = find_vm_area((void *)addr);
 335	if (!area || !(area->flags & VM_IOREMAP)) {
 336		WARN_ONCE(1, "vm_area at addr %lx is not marked as VM_IOREMAP\n", addr);
 337		return -EINVAL;
 338	}
 339	if (addr != (unsigned long)area->addr ||
 340	    (void *)end != area->addr + get_vm_area_size(area)) {
 341		WARN_ONCE(1, "ioremap request [%lx,%lx) doesn't match vm_area [%lx, %lx)\n",
 342			  addr, end, (long)area->addr,
 343			  (long)area->addr + get_vm_area_size(area));
 344		return -ERANGE;
 345	}
 346	return vmap_page_range(addr, end, phys_addr, prot);
 347}
 348
 349static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
 350			     pgtbl_mod_mask *mask)
 351{
 352	pte_t *pte;
 353
 354	pte = pte_offset_kernel(pmd, addr);
 355	do {
 356		pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
 357		WARN_ON(!pte_none(ptent) && !pte_present(ptent));
 358	} while (pte++, addr += PAGE_SIZE, addr != end);
 359	*mask |= PGTBL_PTE_MODIFIED;
 360}
 361
 362static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
 363			     pgtbl_mod_mask *mask)
 364{
 365	pmd_t *pmd;
 366	unsigned long next;
 367	int cleared;
 368
 369	pmd = pmd_offset(pud, addr);
 370	do {
 371		next = pmd_addr_end(addr, end);
 372
 373		cleared = pmd_clear_huge(pmd);
 374		if (cleared || pmd_bad(*pmd))
 375			*mask |= PGTBL_PMD_MODIFIED;
 376
 377		if (cleared)
 378			continue;
 379		if (pmd_none_or_clear_bad(pmd))
 380			continue;
 381		vunmap_pte_range(pmd, addr, next, mask);
 382
 383		cond_resched();
 384	} while (pmd++, addr = next, addr != end);
 385}
 386
 387static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
 388			     pgtbl_mod_mask *mask)
 389{
 390	pud_t *pud;
 391	unsigned long next;
 392	int cleared;
 393
 394	pud = pud_offset(p4d, addr);
 395	do {
 396		next = pud_addr_end(addr, end);
 397
 398		cleared = pud_clear_huge(pud);
 399		if (cleared || pud_bad(*pud))
 400			*mask |= PGTBL_PUD_MODIFIED;
 401
 402		if (cleared)
 403			continue;
 404		if (pud_none_or_clear_bad(pud))
 405			continue;
 406		vunmap_pmd_range(pud, addr, next, mask);
 407	} while (pud++, addr = next, addr != end);
 408}
 409
 410static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 411			     pgtbl_mod_mask *mask)
 412{
 413	p4d_t *p4d;
 414	unsigned long next;
 415
 416	p4d = p4d_offset(pgd, addr);
 417	do {
 418		next = p4d_addr_end(addr, end);
 419
 420		p4d_clear_huge(p4d);
 421		if (p4d_bad(*p4d))
 422			*mask |= PGTBL_P4D_MODIFIED;
 423
 424		if (p4d_none_or_clear_bad(p4d))
 425			continue;
 426		vunmap_pud_range(p4d, addr, next, mask);
 427	} while (p4d++, addr = next, addr != end);
 428}
 429
 430/*
 431 * vunmap_range_noflush is similar to vunmap_range, but does not
 432 * flush caches or TLBs.
 433 *
 434 * The caller is responsible for calling flush_cache_vmap() before calling
 435 * this function, and flush_tlb_kernel_range after it has returned
 436 * successfully (and before the addresses are expected to cause a page fault
 437 * or be re-mapped for something else, if TLB flushes are being delayed or
 438 * coalesced).
 439 *
 440 * This is an internal function only. Do not use outside mm/.
 441 */
 442void __vunmap_range_noflush(unsigned long start, unsigned long end)
 443{
 444	unsigned long next;
 445	pgd_t *pgd;
 446	unsigned long addr = start;
 447	pgtbl_mod_mask mask = 0;
 448
 449	BUG_ON(addr >= end);
 450	pgd = pgd_offset_k(addr);
 451	do {
 452		next = pgd_addr_end(addr, end);
 453		if (pgd_bad(*pgd))
 454			mask |= PGTBL_PGD_MODIFIED;
 455		if (pgd_none_or_clear_bad(pgd))
 456			continue;
 457		vunmap_p4d_range(pgd, addr, next, &mask);
 458	} while (pgd++, addr = next, addr != end);
 459
 460	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 461		arch_sync_kernel_mappings(start, end);
 462}
 463
 464void vunmap_range_noflush(unsigned long start, unsigned long end)
 465{
 466	kmsan_vunmap_range_noflush(start, end);
 467	__vunmap_range_noflush(start, end);
 468}
 469
 470/**
 471 * vunmap_range - unmap kernel virtual addresses
 472 * @addr: start of the VM area to unmap
 473 * @end: end of the VM area to unmap (non-inclusive)
 474 *
 475 * Clears any present PTEs in the virtual address range, flushes TLBs and
 476 * caches. Any subsequent access to the address before it has been re-mapped
 477 * is a kernel bug.
 478 */
 479void vunmap_range(unsigned long addr, unsigned long end)
 480{
 481	flush_cache_vunmap(addr, end);
 482	vunmap_range_noflush(addr, end);
 483	flush_tlb_kernel_range(addr, end);
 484}
 485
 486static int vmap_pages_pte_range(pmd_t *pmd, unsigned long addr,
 487		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 488		pgtbl_mod_mask *mask)
 489{
 490	pte_t *pte;
 491
 492	/*
 493	 * nr is a running index into the array which helps higher level
 494	 * callers keep track of where we're up to.
 495	 */
 496
 497	pte = pte_alloc_kernel_track(pmd, addr, mask);
 498	if (!pte)
 499		return -ENOMEM;
 500	do {
 501		struct page *page = pages[*nr];
 502
 503		if (WARN_ON(!pte_none(ptep_get(pte))))
 504			return -EBUSY;
 505		if (WARN_ON(!page))
 506			return -ENOMEM;
 507		if (WARN_ON(!pfn_valid(page_to_pfn(page))))
 508			return -EINVAL;
 509
 510		set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 511		(*nr)++;
 512	} while (pte++, addr += PAGE_SIZE, addr != end);
 513	*mask |= PGTBL_PTE_MODIFIED;
 514	return 0;
 515}
 516
 517static int vmap_pages_pmd_range(pud_t *pud, unsigned long addr,
 518		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 519		pgtbl_mod_mask *mask)
 520{
 521	pmd_t *pmd;
 522	unsigned long next;
 523
 524	pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
 525	if (!pmd)
 526		return -ENOMEM;
 527	do {
 528		next = pmd_addr_end(addr, end);
 529		if (vmap_pages_pte_range(pmd, addr, next, prot, pages, nr, mask))
 530			return -ENOMEM;
 531	} while (pmd++, addr = next, addr != end);
 532	return 0;
 533}
 534
 535static int vmap_pages_pud_range(p4d_t *p4d, unsigned long addr,
 536		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 537		pgtbl_mod_mask *mask)
 538{
 539	pud_t *pud;
 540	unsigned long next;
 541
 542	pud = pud_alloc_track(&init_mm, p4d, addr, mask);
 543	if (!pud)
 544		return -ENOMEM;
 545	do {
 546		next = pud_addr_end(addr, end);
 547		if (vmap_pages_pmd_range(pud, addr, next, prot, pages, nr, mask))
 548			return -ENOMEM;
 549	} while (pud++, addr = next, addr != end);
 550	return 0;
 551}
 552
 553static int vmap_pages_p4d_range(pgd_t *pgd, unsigned long addr,
 554		unsigned long end, pgprot_t prot, struct page **pages, int *nr,
 555		pgtbl_mod_mask *mask)
 556{
 557	p4d_t *p4d;
 558	unsigned long next;
 559
 560	p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
 561	if (!p4d)
 562		return -ENOMEM;
 563	do {
 564		next = p4d_addr_end(addr, end);
 565		if (vmap_pages_pud_range(p4d, addr, next, prot, pages, nr, mask))
 566			return -ENOMEM;
 567	} while (p4d++, addr = next, addr != end);
 568	return 0;
 569}
 570
 571static int vmap_small_pages_range_noflush(unsigned long addr, unsigned long end,
 572		pgprot_t prot, struct page **pages)
 573{
 574	unsigned long start = addr;
 575	pgd_t *pgd;
 576	unsigned long next;
 577	int err = 0;
 578	int nr = 0;
 579	pgtbl_mod_mask mask = 0;
 580
 581	BUG_ON(addr >= end);
 582	pgd = pgd_offset_k(addr);
 583	do {
 584		next = pgd_addr_end(addr, end);
 585		if (pgd_bad(*pgd))
 586			mask |= PGTBL_PGD_MODIFIED;
 587		err = vmap_pages_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
 588		if (err)
 589			break;
 590	} while (pgd++, addr = next, addr != end);
 591
 592	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
 593		arch_sync_kernel_mappings(start, end);
 594
 595	return err;
 596}
 597
 598/*
 599 * vmap_pages_range_noflush is similar to vmap_pages_range, but does not
 600 * flush caches.
 601 *
 602 * The caller is responsible for calling flush_cache_vmap() after this
 603 * function returns successfully and before the addresses are accessed.
 604 *
 605 * This is an internal function only. Do not use outside mm/.
 606 */
 607int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 608		pgprot_t prot, struct page **pages, unsigned int page_shift)
 609{
 610	unsigned int i, nr = (end - addr) >> PAGE_SHIFT;
 611
 612	WARN_ON(page_shift < PAGE_SHIFT);
 613
 614	if (!IS_ENABLED(CONFIG_HAVE_ARCH_HUGE_VMALLOC) ||
 615			page_shift == PAGE_SHIFT)
 616		return vmap_small_pages_range_noflush(addr, end, prot, pages);
 617
 618	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
 619		int err;
 620
 621		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
 622					page_to_phys(pages[i]), prot,
 623					page_shift);
 624		if (err)
 625			return err;
 626
 627		addr += 1UL << page_shift;
 628	}
 629
 630	return 0;
 631}
 632
 633int vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 634		pgprot_t prot, struct page **pages, unsigned int page_shift)
 635{
 636	int ret = kmsan_vmap_pages_range_noflush(addr, end, prot, pages,
 637						 page_shift);
 638
 639	if (ret)
 640		return ret;
 641	return __vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 642}
 643
 644/**
 645 * vmap_pages_range - map pages to a kernel virtual address
 646 * @addr: start of the VM area to map
 647 * @end: end of the VM area to map (non-inclusive)
 648 * @prot: page protection flags to use
 649 * @pages: pages to map (always PAGE_SIZE pages)
 650 * @page_shift: maximum shift that the pages may be mapped with, @pages must
 651 * be aligned and contiguous up to at least this shift.
 652 *
 653 * RETURNS:
 654 * 0 on success, -errno on failure.
 655 */
 656int vmap_pages_range(unsigned long addr, unsigned long end,
 657		pgprot_t prot, struct page **pages, unsigned int page_shift)
 658{
 659	int err;
 660
 661	err = vmap_pages_range_noflush(addr, end, prot, pages, page_shift);
 662	flush_cache_vmap(addr, end);
 663	return err;
 664}
 665
 666static int check_sparse_vm_area(struct vm_struct *area, unsigned long start,
 667				unsigned long end)
 668{
 669	might_sleep();
 670	if (WARN_ON_ONCE(area->flags & VM_FLUSH_RESET_PERMS))
 671		return -EINVAL;
 672	if (WARN_ON_ONCE(area->flags & VM_NO_GUARD))
 673		return -EINVAL;
 674	if (WARN_ON_ONCE(!(area->flags & VM_SPARSE)))
 675		return -EINVAL;
 676	if ((end - start) >> PAGE_SHIFT > totalram_pages())
 677		return -E2BIG;
 678	if (start < (unsigned long)area->addr ||
 679	    (void *)end > area->addr + get_vm_area_size(area))
 680		return -ERANGE;
 681	return 0;
 682}
 683
 684/**
 685 * vm_area_map_pages - map pages inside given sparse vm_area
 686 * @area: vm_area
 687 * @start: start address inside vm_area
 688 * @end: end address inside vm_area
 689 * @pages: pages to map (always PAGE_SIZE pages)
 690 */
 691int vm_area_map_pages(struct vm_struct *area, unsigned long start,
 692		      unsigned long end, struct page **pages)
 693{
 694	int err;
 695
 696	err = check_sparse_vm_area(area, start, end);
 697	if (err)
 698		return err;
 699
 700	return vmap_pages_range(start, end, PAGE_KERNEL, pages, PAGE_SHIFT);
 701}
 702
 703/**
 704 * vm_area_unmap_pages - unmap pages inside given sparse vm_area
 705 * @area: vm_area
 706 * @start: start address inside vm_area
 707 * @end: end address inside vm_area
 708 */
 709void vm_area_unmap_pages(struct vm_struct *area, unsigned long start,
 710			 unsigned long end)
 711{
 712	if (check_sparse_vm_area(area, start, end))
 713		return;
 714
 715	vunmap_range(start, end);
 716}
 717
 718int is_vmalloc_or_module_addr(const void *x)
 719{
 720	/*
 721	 * ARM, x86-64 and sparc64 put modules in a special place,
 722	 * and fall back on vmalloc() if that fails. Others
 723	 * just put it in the vmalloc space.
 724	 */
 725#if defined(CONFIG_EXECMEM) && defined(MODULES_VADDR)
 726	unsigned long addr = (unsigned long)kasan_reset_tag(x);
 727	if (addr >= MODULES_VADDR && addr < MODULES_END)
 728		return 1;
 729#endif
 730	return is_vmalloc_addr(x);
 731}
 732EXPORT_SYMBOL_GPL(is_vmalloc_or_module_addr);
 733
 734/*
 735 * Walk a vmap address to the struct page it maps. Huge vmap mappings will
 736 * return the tail page that corresponds to the base page address, which
 737 * matches small vmap mappings.
 738 */
 739struct page *vmalloc_to_page(const void *vmalloc_addr)
 740{
 741	unsigned long addr = (unsigned long) vmalloc_addr;
 742	struct page *page = NULL;
 743	pgd_t *pgd = pgd_offset_k(addr);
 744	p4d_t *p4d;
 745	pud_t *pud;
 746	pmd_t *pmd;
 747	pte_t *ptep, pte;
 748
 749	/*
 750	 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
 751	 * architectures that do not vmalloc module space
 752	 */
 753	VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
 754
 755	if (pgd_none(*pgd))
 756		return NULL;
 757	if (WARN_ON_ONCE(pgd_leaf(*pgd)))
 758		return NULL; /* XXX: no allowance for huge pgd */
 759	if (WARN_ON_ONCE(pgd_bad(*pgd)))
 760		return NULL;
 761
 762	p4d = p4d_offset(pgd, addr);
 763	if (p4d_none(*p4d))
 764		return NULL;
 765	if (p4d_leaf(*p4d))
 766		return p4d_page(*p4d) + ((addr & ~P4D_MASK) >> PAGE_SHIFT);
 767	if (WARN_ON_ONCE(p4d_bad(*p4d)))
 768		return NULL;
 769
 770	pud = pud_offset(p4d, addr);
 771	if (pud_none(*pud))
 772		return NULL;
 773	if (pud_leaf(*pud))
 774		return pud_page(*pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
 775	if (WARN_ON_ONCE(pud_bad(*pud)))
 776		return NULL;
 777
 778	pmd = pmd_offset(pud, addr);
 779	if (pmd_none(*pmd))
 780		return NULL;
 781	if (pmd_leaf(*pmd))
 782		return pmd_page(*pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
 783	if (WARN_ON_ONCE(pmd_bad(*pmd)))
 784		return NULL;
 785
 786	ptep = pte_offset_kernel(pmd, addr);
 787	pte = ptep_get(ptep);
 788	if (pte_present(pte))
 789		page = pte_page(pte);
 
 790
 791	return page;
 792}
 793EXPORT_SYMBOL(vmalloc_to_page);
 794
 795/*
 796 * Map a vmalloc()-space virtual address to the physical page frame number.
 797 */
 798unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
 799{
 800	return page_to_pfn(vmalloc_to_page(vmalloc_addr));
 801}
 802EXPORT_SYMBOL(vmalloc_to_pfn);
 803
 804
 805/*** Global kva allocator ***/
 806
 807#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
 808#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
 809
 810
 
 811static DEFINE_SPINLOCK(free_vmap_area_lock);
 
 
 
 812static bool vmap_initialized __read_mostly;
 813
 
 
 
 
 814/*
 815 * This kmem_cache is used for vmap_area objects. Instead of
 816 * allocating from slab we reuse an object from this cache to
 817 * make things faster. Especially in "no edge" splitting of
 818 * free block.
 819 */
 820static struct kmem_cache *vmap_area_cachep;
 821
 822/*
 823 * This linked list is used in pair with free_vmap_area_root.
 824 * It gives O(1) access to prev/next to perform fast coalescing.
 825 */
 826static LIST_HEAD(free_vmap_area_list);
 827
 828/*
 829 * This augment red-black tree represents the free vmap space.
 830 * All vmap_area objects in this tree are sorted by va->va_start
 831 * address. It is used for allocation and merging when a vmap
 832 * object is released.
 833 *
 834 * Each vmap_area node contains a maximum available free block
 835 * of its sub-tree, right or left. Therefore it is possible to
 836 * find a lowest match of free area.
 837 */
 838static struct rb_root free_vmap_area_root = RB_ROOT;
 839
 840/*
 841 * Preload a CPU with one object for "no edge" split case. The
 842 * aim is to get rid of allocations from the atomic context, thus
 843 * to use more permissive allocation masks.
 844 */
 845static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
 846
 847/*
 848 * This structure defines a single, solid model where a list and
 849 * rb-tree are part of one entity protected by the lock. Nodes are
 850 * sorted in ascending order, thus for O(1) access to left/right
 851 * neighbors a list is used as well as for sequential traversal.
 852 */
 853struct rb_list {
 854	struct rb_root root;
 855	struct list_head head;
 856	spinlock_t lock;
 857};
 858
 859/*
 860 * A fast size storage contains VAs up to 1M size. A pool consists
 861 * of linked between each other ready to go VAs of certain sizes.
 862 * An index in the pool-array corresponds to number of pages + 1.
 863 */
 864#define MAX_VA_SIZE_PAGES 256
 865
 866struct vmap_pool {
 867	struct list_head head;
 868	unsigned long len;
 869};
 870
 871/*
 872 * An effective vmap-node logic. Users make use of nodes instead
 873 * of a global heap. It allows to balance an access and mitigate
 874 * contention.
 875 */
 876static struct vmap_node {
 877	/* Simple size segregated storage. */
 878	struct vmap_pool pool[MAX_VA_SIZE_PAGES];
 879	spinlock_t pool_lock;
 880	bool skip_populate;
 881
 882	/* Bookkeeping data of this node. */
 883	struct rb_list busy;
 884	struct rb_list lazy;
 885
 886	/*
 887	 * Ready-to-free areas.
 888	 */
 889	struct list_head purge_list;
 890	struct work_struct purge_work;
 891	unsigned long nr_purged;
 892} single;
 893
 894/*
 895 * Initial setup consists of one single node, i.e. a balancing
 896 * is fully disabled. Later on, after vmap is initialized these
 897 * parameters are updated based on a system capacity.
 898 */
 899static struct vmap_node *vmap_nodes = &single;
 900static __read_mostly unsigned int nr_vmap_nodes = 1;
 901static __read_mostly unsigned int vmap_zone_size = 1;
 902
 903static inline unsigned int
 904addr_to_node_id(unsigned long addr)
 905{
 906	return (addr / vmap_zone_size) % nr_vmap_nodes;
 907}
 908
 909static inline struct vmap_node *
 910addr_to_node(unsigned long addr)
 911{
 912	return &vmap_nodes[addr_to_node_id(addr)];
 913}
 914
 915static inline struct vmap_node *
 916id_to_node(unsigned int id)
 917{
 918	return &vmap_nodes[id % nr_vmap_nodes];
 919}
 920
 921/*
 922 * We use the value 0 to represent "no node", that is why
 923 * an encoded value will be the node-id incremented by 1.
 924 * It is always greater then 0. A valid node_id which can
 925 * be encoded is [0:nr_vmap_nodes - 1]. If a passed node_id
 926 * is not valid 0 is returned.
 927 */
 928static unsigned int
 929encode_vn_id(unsigned int node_id)
 930{
 931	/* Can store U8_MAX [0:254] nodes. */
 932	if (node_id < nr_vmap_nodes)
 933		return (node_id + 1) << BITS_PER_BYTE;
 934
 935	/* Warn and no node encoded. */
 936	WARN_ONCE(1, "Encode wrong node id (%u)\n", node_id);
 937	return 0;
 938}
 939
 940/*
 941 * Returns an encoded node-id, the valid range is within
 942 * [0:nr_vmap_nodes-1] values. Otherwise nr_vmap_nodes is
 943 * returned if extracted data is wrong.
 944 */
 945static unsigned int
 946decode_vn_id(unsigned int val)
 947{
 948	unsigned int node_id = (val >> BITS_PER_BYTE) - 1;
 949
 950	/* Can store U8_MAX [0:254] nodes. */
 951	if (node_id < nr_vmap_nodes)
 952		return node_id;
 953
 954	/* If it was _not_ zero, warn. */
 955	WARN_ONCE(node_id != UINT_MAX,
 956		"Decode wrong node id (%d)\n", node_id);
 957
 958	return nr_vmap_nodes;
 959}
 960
 961static bool
 962is_vn_id_valid(unsigned int node_id)
 963{
 964	if (node_id < nr_vmap_nodes)
 965		return true;
 966
 967	return false;
 968}
 969
 970static __always_inline unsigned long
 971va_size(struct vmap_area *va)
 972{
 973	return (va->va_end - va->va_start);
 974}
 975
 976static __always_inline unsigned long
 977get_subtree_max_size(struct rb_node *node)
 978{
 979	struct vmap_area *va;
 980
 981	va = rb_entry_safe(node, struct vmap_area, rb_node);
 982	return va ? va->subtree_max_size : 0;
 983}
 984
 985RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
 986	struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
 987
 988static void reclaim_and_purge_vmap_areas(void);
 989static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
 990static void drain_vmap_area_work(struct work_struct *work);
 991static DECLARE_WORK(drain_vmap_work, drain_vmap_area_work);
 992
 993static atomic_long_t nr_vmalloc_pages;
 994
 995unsigned long vmalloc_nr_pages(void)
 996{
 997	return atomic_long_read(&nr_vmalloc_pages);
 998}
 999
1000static struct vmap_area *__find_vmap_area(unsigned long addr, struct rb_root *root)
1001{
1002	struct rb_node *n = root->rb_node;
1003
1004	addr = (unsigned long)kasan_reset_tag((void *)addr);
1005
1006	while (n) {
1007		struct vmap_area *va;
1008
1009		va = rb_entry(n, struct vmap_area, rb_node);
1010		if (addr < va->va_start)
1011			n = n->rb_left;
1012		else if (addr >= va->va_end)
1013			n = n->rb_right;
1014		else
1015			return va;
1016	}
1017
1018	return NULL;
1019}
1020
1021/* Look up the first VA which satisfies addr < va_end, NULL if none. */
1022static struct vmap_area *
1023__find_vmap_area_exceed_addr(unsigned long addr, struct rb_root *root)
1024{
1025	struct vmap_area *va = NULL;
1026	struct rb_node *n = root->rb_node;
1027
1028	addr = (unsigned long)kasan_reset_tag((void *)addr);
1029
1030	while (n) {
1031		struct vmap_area *tmp;
1032
1033		tmp = rb_entry(n, struct vmap_area, rb_node);
1034		if (tmp->va_end > addr) {
1035			va = tmp;
1036			if (tmp->va_start <= addr)
1037				break;
1038
1039			n = n->rb_left;
1040		} else
1041			n = n->rb_right;
1042	}
1043
1044	return va;
1045}
1046
1047/*
1048 * Returns a node where a first VA, that satisfies addr < va_end, resides.
1049 * If success, a node is locked. A user is responsible to unlock it when a
1050 * VA is no longer needed to be accessed.
1051 *
1052 * Returns NULL if nothing found.
1053 */
1054static struct vmap_node *
1055find_vmap_area_exceed_addr_lock(unsigned long addr, struct vmap_area **va)
1056{
1057	unsigned long va_start_lowest;
1058	struct vmap_node *vn;
1059	int i;
1060
1061repeat:
1062	for (i = 0, va_start_lowest = 0; i < nr_vmap_nodes; i++) {
1063		vn = &vmap_nodes[i];
1064
1065		spin_lock(&vn->busy.lock);
1066		*va = __find_vmap_area_exceed_addr(addr, &vn->busy.root);
1067
1068		if (*va)
1069			if (!va_start_lowest || (*va)->va_start < va_start_lowest)
1070				va_start_lowest = (*va)->va_start;
1071		spin_unlock(&vn->busy.lock);
1072	}
1073
1074	/*
1075	 * Check if found VA exists, it might have gone away.  In this case we
1076	 * repeat the search because a VA has been removed concurrently and we
1077	 * need to proceed to the next one, which is a rare case.
1078	 */
1079	if (va_start_lowest) {
1080		vn = addr_to_node(va_start_lowest);
1081
1082		spin_lock(&vn->busy.lock);
1083		*va = __find_vmap_area(va_start_lowest, &vn->busy.root);
1084
1085		if (*va)
1086			return vn;
1087
1088		spin_unlock(&vn->busy.lock);
1089		goto repeat;
1090	}
1091
1092	return NULL;
1093}
1094
1095/*
1096 * This function returns back addresses of parent node
1097 * and its left or right link for further processing.
1098 *
1099 * Otherwise NULL is returned. In that case all further
1100 * steps regarding inserting of conflicting overlap range
1101 * have to be declined and actually considered as a bug.
1102 */
1103static __always_inline struct rb_node **
1104find_va_links(struct vmap_area *va,
1105	struct rb_root *root, struct rb_node *from,
1106	struct rb_node **parent)
1107{
1108	struct vmap_area *tmp_va;
1109	struct rb_node **link;
1110
1111	if (root) {
1112		link = &root->rb_node;
1113		if (unlikely(!*link)) {
1114			*parent = NULL;
1115			return link;
1116		}
1117	} else {
1118		link = &from;
1119	}
1120
1121	/*
1122	 * Go to the bottom of the tree. When we hit the last point
1123	 * we end up with parent rb_node and correct direction, i name
1124	 * it link, where the new va->rb_node will be attached to.
1125	 */
1126	do {
1127		tmp_va = rb_entry(*link, struct vmap_area, rb_node);
1128
1129		/*
1130		 * During the traversal we also do some sanity check.
1131		 * Trigger the BUG() if there are sides(left/right)
1132		 * or full overlaps.
1133		 */
1134		if (va->va_end <= tmp_va->va_start)
1135			link = &(*link)->rb_left;
1136		else if (va->va_start >= tmp_va->va_end)
1137			link = &(*link)->rb_right;
1138		else {
1139			WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
1140				va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
1141
1142			return NULL;
1143		}
1144	} while (*link);
1145
1146	*parent = &tmp_va->rb_node;
1147	return link;
1148}
1149
1150static __always_inline struct list_head *
1151get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
1152{
1153	struct list_head *list;
1154
1155	if (unlikely(!parent))
1156		/*
1157		 * The red-black tree where we try to find VA neighbors
1158		 * before merging or inserting is empty, i.e. it means
1159		 * there is no free vmap space. Normally it does not
1160		 * happen but we handle this case anyway.
1161		 */
1162		return NULL;
1163
1164	list = &rb_entry(parent, struct vmap_area, rb_node)->list;
1165	return (&parent->rb_right == link ? list->next : list);
1166}
1167
1168static __always_inline void
1169__link_va(struct vmap_area *va, struct rb_root *root,
1170	struct rb_node *parent, struct rb_node **link,
1171	struct list_head *head, bool augment)
1172{
1173	/*
1174	 * VA is still not in the list, but we can
1175	 * identify its future previous list_head node.
1176	 */
1177	if (likely(parent)) {
1178		head = &rb_entry(parent, struct vmap_area, rb_node)->list;
1179		if (&parent->rb_right != link)
1180			head = head->prev;
1181	}
1182
1183	/* Insert to the rb-tree */
1184	rb_link_node(&va->rb_node, parent, link);
1185	if (augment) {
1186		/*
1187		 * Some explanation here. Just perform simple insertion
1188		 * to the tree. We do not set va->subtree_max_size to
1189		 * its current size before calling rb_insert_augmented().
1190		 * It is because we populate the tree from the bottom
1191		 * to parent levels when the node _is_ in the tree.
1192		 *
1193		 * Therefore we set subtree_max_size to zero after insertion,
1194		 * to let __augment_tree_propagate_from() puts everything to
1195		 * the correct order later on.
1196		 */
1197		rb_insert_augmented(&va->rb_node,
1198			root, &free_vmap_area_rb_augment_cb);
1199		va->subtree_max_size = 0;
1200	} else {
1201		rb_insert_color(&va->rb_node, root);
1202	}
1203
1204	/* Address-sort this list */
1205	list_add(&va->list, head);
1206}
1207
1208static __always_inline void
1209link_va(struct vmap_area *va, struct rb_root *root,
1210	struct rb_node *parent, struct rb_node **link,
1211	struct list_head *head)
1212{
1213	__link_va(va, root, parent, link, head, false);
1214}
1215
1216static __always_inline void
1217link_va_augment(struct vmap_area *va, struct rb_root *root,
1218	struct rb_node *parent, struct rb_node **link,
1219	struct list_head *head)
1220{
1221	__link_va(va, root, parent, link, head, true);
1222}
1223
1224static __always_inline void
1225__unlink_va(struct vmap_area *va, struct rb_root *root, bool augment)
1226{
1227	if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
1228		return;
1229
1230	if (augment)
1231		rb_erase_augmented(&va->rb_node,
1232			root, &free_vmap_area_rb_augment_cb);
1233	else
1234		rb_erase(&va->rb_node, root);
1235
1236	list_del_init(&va->list);
1237	RB_CLEAR_NODE(&va->rb_node);
1238}
1239
1240static __always_inline void
1241unlink_va(struct vmap_area *va, struct rb_root *root)
1242{
1243	__unlink_va(va, root, false);
1244}
1245
1246static __always_inline void
1247unlink_va_augment(struct vmap_area *va, struct rb_root *root)
1248{
1249	__unlink_va(va, root, true);
1250}
1251
1252#if DEBUG_AUGMENT_PROPAGATE_CHECK
1253/*
1254 * Gets called when remove the node and rotate.
1255 */
1256static __always_inline unsigned long
1257compute_subtree_max_size(struct vmap_area *va)
1258{
1259	return max3(va_size(va),
1260		get_subtree_max_size(va->rb_node.rb_left),
1261		get_subtree_max_size(va->rb_node.rb_right));
1262}
1263
1264static void
1265augment_tree_propagate_check(void)
1266{
1267	struct vmap_area *va;
1268	unsigned long computed_size;
1269
1270	list_for_each_entry(va, &free_vmap_area_list, list) {
1271		computed_size = compute_subtree_max_size(va);
1272		if (computed_size != va->subtree_max_size)
1273			pr_emerg("tree is corrupted: %lu, %lu\n",
1274				va_size(va), va->subtree_max_size);
1275	}
1276}
1277#endif
1278
1279/*
1280 * This function populates subtree_max_size from bottom to upper
1281 * levels starting from VA point. The propagation must be done
1282 * when VA size is modified by changing its va_start/va_end. Or
1283 * in case of newly inserting of VA to the tree.
1284 *
1285 * It means that __augment_tree_propagate_from() must be called:
1286 * - After VA has been inserted to the tree(free path);
1287 * - After VA has been shrunk(allocation path);
1288 * - After VA has been increased(merging path).
1289 *
1290 * Please note that, it does not mean that upper parent nodes
1291 * and their subtree_max_size are recalculated all the time up
1292 * to the root node.
1293 *
1294 *       4--8
1295 *        /\
1296 *       /  \
1297 *      /    \
1298 *    2--2  8--8
1299 *
1300 * For example if we modify the node 4, shrinking it to 2, then
1301 * no any modification is required. If we shrink the node 2 to 1
1302 * its subtree_max_size is updated only, and set to 1. If we shrink
1303 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
1304 * node becomes 4--6.
1305 */
1306static __always_inline void
1307augment_tree_propagate_from(struct vmap_area *va)
1308{
1309	/*
1310	 * Populate the tree from bottom towards the root until
1311	 * the calculated maximum available size of checked node
1312	 * is equal to its current one.
1313	 */
1314	free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
1315
1316#if DEBUG_AUGMENT_PROPAGATE_CHECK
1317	augment_tree_propagate_check();
1318#endif
1319}
1320
1321static void
1322insert_vmap_area(struct vmap_area *va,
1323	struct rb_root *root, struct list_head *head)
1324{
1325	struct rb_node **link;
1326	struct rb_node *parent;
1327
1328	link = find_va_links(va, root, NULL, &parent);
1329	if (link)
1330		link_va(va, root, parent, link, head);
1331}
1332
1333static void
1334insert_vmap_area_augment(struct vmap_area *va,
1335	struct rb_node *from, struct rb_root *root,
1336	struct list_head *head)
1337{
1338	struct rb_node **link;
1339	struct rb_node *parent;
1340
1341	if (from)
1342		link = find_va_links(va, NULL, from, &parent);
1343	else
1344		link = find_va_links(va, root, NULL, &parent);
1345
1346	if (link) {
1347		link_va_augment(va, root, parent, link, head);
1348		augment_tree_propagate_from(va);
1349	}
1350}
1351
1352/*
1353 * Merge de-allocated chunk of VA memory with previous
1354 * and next free blocks. If coalesce is not done a new
1355 * free area is inserted. If VA has been merged, it is
1356 * freed.
1357 *
1358 * Please note, it can return NULL in case of overlap
1359 * ranges, followed by WARN() report. Despite it is a
1360 * buggy behaviour, a system can be alive and keep
1361 * ongoing.
1362 */
1363static __always_inline struct vmap_area *
1364__merge_or_add_vmap_area(struct vmap_area *va,
1365	struct rb_root *root, struct list_head *head, bool augment)
1366{
1367	struct vmap_area *sibling;
1368	struct list_head *next;
1369	struct rb_node **link;
1370	struct rb_node *parent;
1371	bool merged = false;
1372
1373	/*
1374	 * Find a place in the tree where VA potentially will be
1375	 * inserted, unless it is merged with its sibling/siblings.
1376	 */
1377	link = find_va_links(va, root, NULL, &parent);
1378	if (!link)
1379		return NULL;
1380
1381	/*
1382	 * Get next node of VA to check if merging can be done.
1383	 */
1384	next = get_va_next_sibling(parent, link);
1385	if (unlikely(next == NULL))
1386		goto insert;
1387
1388	/*
1389	 * start            end
1390	 * |                |
1391	 * |<------VA------>|<-----Next----->|
1392	 *                  |                |
1393	 *                  start            end
1394	 */
1395	if (next != head) {
1396		sibling = list_entry(next, struct vmap_area, list);
1397		if (sibling->va_start == va->va_end) {
1398			sibling->va_start = va->va_start;
1399
1400			/* Free vmap_area object. */
1401			kmem_cache_free(vmap_area_cachep, va);
1402
1403			/* Point to the new merged area. */
1404			va = sibling;
1405			merged = true;
1406		}
1407	}
1408
1409	/*
1410	 * start            end
1411	 * |                |
1412	 * |<-----Prev----->|<------VA------>|
1413	 *                  |                |
1414	 *                  start            end
1415	 */
1416	if (next->prev != head) {
1417		sibling = list_entry(next->prev, struct vmap_area, list);
1418		if (sibling->va_end == va->va_start) {
1419			/*
1420			 * If both neighbors are coalesced, it is important
1421			 * to unlink the "next" node first, followed by merging
1422			 * with "previous" one. Otherwise the tree might not be
1423			 * fully populated if a sibling's augmented value is
1424			 * "normalized" because of rotation operations.
1425			 */
1426			if (merged)
1427				__unlink_va(va, root, augment);
1428
1429			sibling->va_end = va->va_end;
1430
1431			/* Free vmap_area object. */
1432			kmem_cache_free(vmap_area_cachep, va);
1433
1434			/* Point to the new merged area. */
1435			va = sibling;
1436			merged = true;
1437		}
1438	}
1439
1440insert:
1441	if (!merged)
1442		__link_va(va, root, parent, link, head, augment);
1443
1444	return va;
1445}
1446
1447static __always_inline struct vmap_area *
1448merge_or_add_vmap_area(struct vmap_area *va,
1449	struct rb_root *root, struct list_head *head)
1450{
1451	return __merge_or_add_vmap_area(va, root, head, false);
1452}
1453
1454static __always_inline struct vmap_area *
1455merge_or_add_vmap_area_augment(struct vmap_area *va,
1456	struct rb_root *root, struct list_head *head)
1457{
1458	va = __merge_or_add_vmap_area(va, root, head, true);
1459	if (va)
1460		augment_tree_propagate_from(va);
1461
1462	return va;
1463}
1464
1465static __always_inline bool
1466is_within_this_va(struct vmap_area *va, unsigned long size,
1467	unsigned long align, unsigned long vstart)
1468{
1469	unsigned long nva_start_addr;
1470
1471	if (va->va_start > vstart)
1472		nva_start_addr = ALIGN(va->va_start, align);
1473	else
1474		nva_start_addr = ALIGN(vstart, align);
1475
1476	/* Can be overflowed due to big size or alignment. */
1477	if (nva_start_addr + size < nva_start_addr ||
1478			nva_start_addr < vstart)
1479		return false;
1480
1481	return (nva_start_addr + size <= va->va_end);
1482}
1483
1484/*
1485 * Find the first free block(lowest start address) in the tree,
1486 * that will accomplish the request corresponding to passing
1487 * parameters. Please note, with an alignment bigger than PAGE_SIZE,
1488 * a search length is adjusted to account for worst case alignment
1489 * overhead.
1490 */
1491static __always_inline struct vmap_area *
1492find_vmap_lowest_match(struct rb_root *root, unsigned long size,
1493	unsigned long align, unsigned long vstart, bool adjust_search_size)
1494{
1495	struct vmap_area *va;
1496	struct rb_node *node;
1497	unsigned long length;
1498
1499	/* Start from the root. */
1500	node = root->rb_node;
1501
1502	/* Adjust the search size for alignment overhead. */
1503	length = adjust_search_size ? size + align - 1 : size;
1504
1505	while (node) {
1506		va = rb_entry(node, struct vmap_area, rb_node);
1507
1508		if (get_subtree_max_size(node->rb_left) >= length &&
1509				vstart < va->va_start) {
1510			node = node->rb_left;
1511		} else {
1512			if (is_within_this_va(va, size, align, vstart))
1513				return va;
1514
1515			/*
1516			 * Does not make sense to go deeper towards the right
1517			 * sub-tree if it does not have a free block that is
1518			 * equal or bigger to the requested search length.
1519			 */
1520			if (get_subtree_max_size(node->rb_right) >= length) {
1521				node = node->rb_right;
1522				continue;
1523			}
1524
1525			/*
1526			 * OK. We roll back and find the first right sub-tree,
1527			 * that will satisfy the search criteria. It can happen
1528			 * due to "vstart" restriction or an alignment overhead
1529			 * that is bigger then PAGE_SIZE.
1530			 */
1531			while ((node = rb_parent(node))) {
1532				va = rb_entry(node, struct vmap_area, rb_node);
1533				if (is_within_this_va(va, size, align, vstart))
1534					return va;
1535
1536				if (get_subtree_max_size(node->rb_right) >= length &&
1537						vstart <= va->va_start) {
1538					/*
1539					 * Shift the vstart forward. Please note, we update it with
1540					 * parent's start address adding "1" because we do not want
1541					 * to enter same sub-tree after it has already been checked
1542					 * and no suitable free block found there.
1543					 */
1544					vstart = va->va_start + 1;
1545					node = node->rb_right;
1546					break;
1547				}
1548			}
1549		}
1550	}
1551
1552	return NULL;
1553}
1554
1555#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1556#include <linux/random.h>
1557
1558static struct vmap_area *
1559find_vmap_lowest_linear_match(struct list_head *head, unsigned long size,
1560	unsigned long align, unsigned long vstart)
1561{
1562	struct vmap_area *va;
1563
1564	list_for_each_entry(va, head, list) {
1565		if (!is_within_this_va(va, size, align, vstart))
1566			continue;
1567
1568		return va;
1569	}
1570
1571	return NULL;
1572}
1573
1574static void
1575find_vmap_lowest_match_check(struct rb_root *root, struct list_head *head,
1576			     unsigned long size, unsigned long align)
1577{
1578	struct vmap_area *va_1, *va_2;
1579	unsigned long vstart;
1580	unsigned int rnd;
1581
1582	get_random_bytes(&rnd, sizeof(rnd));
1583	vstart = VMALLOC_START + rnd;
1584
1585	va_1 = find_vmap_lowest_match(root, size, align, vstart, false);
1586	va_2 = find_vmap_lowest_linear_match(head, size, align, vstart);
1587
1588	if (va_1 != va_2)
1589		pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
1590			va_1, va_2, vstart);
1591}
1592#endif
1593
1594enum fit_type {
1595	NOTHING_FIT = 0,
1596	FL_FIT_TYPE = 1,	/* full fit */
1597	LE_FIT_TYPE = 2,	/* left edge fit */
1598	RE_FIT_TYPE = 3,	/* right edge fit */
1599	NE_FIT_TYPE = 4		/* no edge fit */
1600};
1601
1602static __always_inline enum fit_type
1603classify_va_fit_type(struct vmap_area *va,
1604	unsigned long nva_start_addr, unsigned long size)
1605{
1606	enum fit_type type;
1607
1608	/* Check if it is within VA. */
1609	if (nva_start_addr < va->va_start ||
1610			nva_start_addr + size > va->va_end)
1611		return NOTHING_FIT;
1612
1613	/* Now classify. */
1614	if (va->va_start == nva_start_addr) {
1615		if (va->va_end == nva_start_addr + size)
1616			type = FL_FIT_TYPE;
1617		else
1618			type = LE_FIT_TYPE;
1619	} else if (va->va_end == nva_start_addr + size) {
1620		type = RE_FIT_TYPE;
1621	} else {
1622		type = NE_FIT_TYPE;
1623	}
1624
1625	return type;
1626}
1627
1628static __always_inline int
1629va_clip(struct rb_root *root, struct list_head *head,
1630		struct vmap_area *va, unsigned long nva_start_addr,
1631		unsigned long size)
1632{
1633	struct vmap_area *lva = NULL;
1634	enum fit_type type = classify_va_fit_type(va, nva_start_addr, size);
1635
1636	if (type == FL_FIT_TYPE) {
1637		/*
1638		 * No need to split VA, it fully fits.
1639		 *
1640		 * |               |
1641		 * V      NVA      V
1642		 * |---------------|
1643		 */
1644		unlink_va_augment(va, root);
1645		kmem_cache_free(vmap_area_cachep, va);
1646	} else if (type == LE_FIT_TYPE) {
1647		/*
1648		 * Split left edge of fit VA.
1649		 *
1650		 * |       |
1651		 * V  NVA  V   R
1652		 * |-------|-------|
1653		 */
1654		va->va_start += size;
1655	} else if (type == RE_FIT_TYPE) {
1656		/*
1657		 * Split right edge of fit VA.
1658		 *
1659		 *         |       |
1660		 *     L   V  NVA  V
1661		 * |-------|-------|
1662		 */
1663		va->va_end = nva_start_addr;
1664	} else if (type == NE_FIT_TYPE) {
1665		/*
1666		 * Split no edge of fit VA.
1667		 *
1668		 *     |       |
1669		 *   L V  NVA  V R
1670		 * |---|-------|---|
1671		 */
1672		lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1673		if (unlikely(!lva)) {
1674			/*
1675			 * For percpu allocator we do not do any pre-allocation
1676			 * and leave it as it is. The reason is it most likely
1677			 * never ends up with NE_FIT_TYPE splitting. In case of
1678			 * percpu allocations offsets and sizes are aligned to
1679			 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1680			 * are its main fitting cases.
1681			 *
1682			 * There are a few exceptions though, as an example it is
1683			 * a first allocation (early boot up) when we have "one"
1684			 * big free space that has to be split.
1685			 *
1686			 * Also we can hit this path in case of regular "vmap"
1687			 * allocations, if "this" current CPU was not preloaded.
1688			 * See the comment in alloc_vmap_area() why. If so, then
1689			 * GFP_NOWAIT is used instead to get an extra object for
1690			 * split purpose. That is rare and most time does not
1691			 * occur.
1692			 *
1693			 * What happens if an allocation gets failed. Basically,
1694			 * an "overflow" path is triggered to purge lazily freed
1695			 * areas to free some memory, then, the "retry" path is
1696			 * triggered to repeat one more time. See more details
1697			 * in alloc_vmap_area() function.
1698			 */
1699			lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1700			if (!lva)
1701				return -1;
1702		}
1703
1704		/*
1705		 * Build the remainder.
1706		 */
1707		lva->va_start = va->va_start;
1708		lva->va_end = nva_start_addr;
1709
1710		/*
1711		 * Shrink this VA to remaining size.
1712		 */
1713		va->va_start = nva_start_addr + size;
1714	} else {
1715		return -1;
1716	}
1717
1718	if (type != FL_FIT_TYPE) {
1719		augment_tree_propagate_from(va);
1720
1721		if (lva)	/* type == NE_FIT_TYPE */
1722			insert_vmap_area_augment(lva, &va->rb_node, root, head);
1723	}
1724
1725	return 0;
1726}
1727
1728static unsigned long
1729va_alloc(struct vmap_area *va,
1730		struct rb_root *root, struct list_head *head,
1731		unsigned long size, unsigned long align,
1732		unsigned long vstart, unsigned long vend)
1733{
1734	unsigned long nva_start_addr;
1735	int ret;
1736
1737	if (va->va_start > vstart)
1738		nva_start_addr = ALIGN(va->va_start, align);
1739	else
1740		nva_start_addr = ALIGN(vstart, align);
1741
1742	/* Check the "vend" restriction. */
1743	if (nva_start_addr + size > vend)
1744		return vend;
1745
1746	/* Update the free vmap_area. */
1747	ret = va_clip(root, head, va, nva_start_addr, size);
1748	if (WARN_ON_ONCE(ret))
1749		return vend;
1750
1751	return nva_start_addr;
1752}
1753
1754/*
1755 * Returns a start address of the newly allocated area, if success.
1756 * Otherwise a vend is returned that indicates failure.
1757 */
1758static __always_inline unsigned long
1759__alloc_vmap_area(struct rb_root *root, struct list_head *head,
1760	unsigned long size, unsigned long align,
1761	unsigned long vstart, unsigned long vend)
1762{
1763	bool adjust_search_size = true;
1764	unsigned long nva_start_addr;
1765	struct vmap_area *va;
 
1766
1767	/*
1768	 * Do not adjust when:
1769	 *   a) align <= PAGE_SIZE, because it does not make any sense.
1770	 *      All blocks(their start addresses) are at least PAGE_SIZE
1771	 *      aligned anyway;
1772	 *   b) a short range where a requested size corresponds to exactly
1773	 *      specified [vstart:vend] interval and an alignment > PAGE_SIZE.
1774	 *      With adjusted search length an allocation would not succeed.
1775	 */
1776	if (align <= PAGE_SIZE || (align > PAGE_SIZE && (vend - vstart) == size))
1777		adjust_search_size = false;
1778
1779	va = find_vmap_lowest_match(root, size, align, vstart, adjust_search_size);
1780	if (unlikely(!va))
1781		return vend;
1782
1783	nva_start_addr = va_alloc(va, root, head, size, align, vstart, vend);
1784	if (nva_start_addr == vend)
 
 
 
 
 
 
 
 
 
 
1785		return vend;
1786
1787#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1788	find_vmap_lowest_match_check(root, head, size, align);
1789#endif
1790
1791	return nva_start_addr;
1792}
1793
1794/*
1795 * Free a region of KVA allocated by alloc_vmap_area
1796 */
1797static void free_vmap_area(struct vmap_area *va)
1798{
1799	struct vmap_node *vn = addr_to_node(va->va_start);
1800
1801	/*
1802	 * Remove from the busy tree/list.
1803	 */
1804	spin_lock(&vn->busy.lock);
1805	unlink_va(va, &vn->busy.root);
1806	spin_unlock(&vn->busy.lock);
1807
1808	/*
1809	 * Insert/Merge it back to the free tree/list.
1810	 */
1811	spin_lock(&free_vmap_area_lock);
1812	merge_or_add_vmap_area_augment(va, &free_vmap_area_root, &free_vmap_area_list);
1813	spin_unlock(&free_vmap_area_lock);
1814}
1815
1816static inline void
1817preload_this_cpu_lock(spinlock_t *lock, gfp_t gfp_mask, int node)
1818{
1819	struct vmap_area *va = NULL, *tmp;
1820
1821	/*
1822	 * Preload this CPU with one extra vmap_area object. It is used
1823	 * when fit type of free area is NE_FIT_TYPE. It guarantees that
1824	 * a CPU that does an allocation is preloaded.
1825	 *
1826	 * We do it in non-atomic context, thus it allows us to use more
1827	 * permissive allocation masks to be more stable under low memory
1828	 * condition and high memory pressure.
1829	 */
1830	if (!this_cpu_read(ne_fit_preload_node))
1831		va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1832
1833	spin_lock(lock);
1834
1835	tmp = NULL;
1836	if (va && !__this_cpu_try_cmpxchg(ne_fit_preload_node, &tmp, va))
1837		kmem_cache_free(vmap_area_cachep, va);
1838}
1839
1840static struct vmap_pool *
1841size_to_va_pool(struct vmap_node *vn, unsigned long size)
1842{
1843	unsigned int idx = (size - 1) / PAGE_SIZE;
1844
1845	if (idx < MAX_VA_SIZE_PAGES)
1846		return &vn->pool[idx];
1847
1848	return NULL;
1849}
1850
1851static bool
1852node_pool_add_va(struct vmap_node *n, struct vmap_area *va)
1853{
1854	struct vmap_pool *vp;
1855
1856	vp = size_to_va_pool(n, va_size(va));
1857	if (!vp)
1858		return false;
1859
1860	spin_lock(&n->pool_lock);
1861	list_add(&va->list, &vp->head);
1862	WRITE_ONCE(vp->len, vp->len + 1);
1863	spin_unlock(&n->pool_lock);
1864
1865	return true;
1866}
1867
1868static struct vmap_area *
1869node_pool_del_va(struct vmap_node *vn, unsigned long size,
1870		unsigned long align, unsigned long vstart,
1871		unsigned long vend)
1872{
1873	struct vmap_area *va = NULL;
1874	struct vmap_pool *vp;
1875	int err = 0;
1876
1877	vp = size_to_va_pool(vn, size);
1878	if (!vp || list_empty(&vp->head))
1879		return NULL;
1880
1881	spin_lock(&vn->pool_lock);
1882	if (!list_empty(&vp->head)) {
1883		va = list_first_entry(&vp->head, struct vmap_area, list);
1884
1885		if (IS_ALIGNED(va->va_start, align)) {
1886			/*
1887			 * Do some sanity check and emit a warning
1888			 * if one of below checks detects an error.
1889			 */
1890			err |= (va_size(va) != size);
1891			err |= (va->va_start < vstart);
1892			err |= (va->va_end > vend);
1893
1894			if (!WARN_ON_ONCE(err)) {
1895				list_del_init(&va->list);
1896				WRITE_ONCE(vp->len, vp->len - 1);
1897			} else {
1898				va = NULL;
1899			}
1900		} else {
1901			list_move_tail(&va->list, &vp->head);
1902			va = NULL;
1903		}
1904	}
1905	spin_unlock(&vn->pool_lock);
1906
1907	return va;
1908}
1909
1910static struct vmap_area *
1911node_alloc(unsigned long size, unsigned long align,
1912		unsigned long vstart, unsigned long vend,
1913		unsigned long *addr, unsigned int *vn_id)
1914{
1915	struct vmap_area *va;
1916
1917	*vn_id = 0;
1918	*addr = vend;
1919
1920	/*
1921	 * Fallback to a global heap if not vmalloc or there
1922	 * is only one node.
1923	 */
1924	if (vstart != VMALLOC_START || vend != VMALLOC_END ||
1925			nr_vmap_nodes == 1)
1926		return NULL;
1927
1928	*vn_id = raw_smp_processor_id() % nr_vmap_nodes;
1929	va = node_pool_del_va(id_to_node(*vn_id), size, align, vstart, vend);
1930	*vn_id = encode_vn_id(*vn_id);
1931
1932	if (va)
1933		*addr = va->va_start;
1934
1935	return va;
1936}
1937
1938static inline void setup_vmalloc_vm(struct vm_struct *vm,
1939	struct vmap_area *va, unsigned long flags, const void *caller)
1940{
1941	vm->flags = flags;
1942	vm->addr = (void *)va->va_start;
1943	vm->size = va_size(va);
1944	vm->caller = caller;
1945	va->vm = vm;
1946}
1947
1948/*
1949 * Allocate a region of KVA of the specified size and alignment, within the
1950 * vstart and vend. If vm is passed in, the two will also be bound.
1951 */
1952static struct vmap_area *alloc_vmap_area(unsigned long size,
1953				unsigned long align,
1954				unsigned long vstart, unsigned long vend,
1955				int node, gfp_t gfp_mask,
1956				unsigned long va_flags, struct vm_struct *vm)
1957{
1958	struct vmap_node *vn;
1959	struct vmap_area *va;
1960	unsigned long freed;
1961	unsigned long addr;
1962	unsigned int vn_id;
1963	int purged = 0;
1964	int ret;
1965
1966	if (unlikely(!size || offset_in_page(size) || !is_power_of_2(align)))
1967		return ERR_PTR(-EINVAL);
 
1968
1969	if (unlikely(!vmap_initialized))
1970		return ERR_PTR(-EBUSY);
1971
1972	might_sleep();
 
 
 
 
 
1973
1974	/*
1975	 * If a VA is obtained from a global heap(if it fails here)
1976	 * it is anyway marked with this "vn_id" so it is returned
1977	 * to this pool's node later. Such way gives a possibility
1978	 * to populate pools based on users demand.
1979	 *
1980	 * On success a ready to go VA is returned.
1981	 */
1982	va = node_alloc(size, align, vstart, vend, &addr, &vn_id);
1983	if (!va) {
1984		gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1985
1986		va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1987		if (unlikely(!va))
1988			return ERR_PTR(-ENOMEM);
1989
1990		/*
1991		 * Only scan the relevant parts containing pointers to other objects
1992		 * to avoid false negatives.
1993		 */
1994		kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1995	}
1996
1997retry:
1998	if (addr == vend) {
1999		preload_this_cpu_lock(&free_vmap_area_lock, gfp_mask, node);
2000		addr = __alloc_vmap_area(&free_vmap_area_root, &free_vmap_area_list,
2001			size, align, vstart, vend);
2002		spin_unlock(&free_vmap_area_lock);
2003	}
2004
2005	trace_alloc_vmap_area(addr, size, align, vstart, vend, addr == vend);
2006
2007	/*
2008	 * If an allocation fails, the "vend" address is
2009	 * returned. Therefore trigger the overflow path.
2010	 */
2011	if (unlikely(addr == vend))
2012		goto overflow;
2013
2014	va->va_start = addr;
2015	va->va_end = addr + size;
2016	va->vm = NULL;
2017	va->flags = (va_flags | vn_id);
2018
2019	if (vm) {
2020		vm->addr = (void *)va->va_start;
2021		vm->size = va_size(va);
2022		va->vm = vm;
2023	}
2024
2025	vn = addr_to_node(va->va_start);
2026
2027	spin_lock(&vn->busy.lock);
2028	insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
2029	spin_unlock(&vn->busy.lock);
2030
2031	BUG_ON(!IS_ALIGNED(va->va_start, align));
2032	BUG_ON(va->va_start < vstart);
2033	BUG_ON(va->va_end > vend);
2034
2035	ret = kasan_populate_vmalloc(addr, size);
2036	if (ret) {
2037		free_vmap_area(va);
2038		return ERR_PTR(ret);
2039	}
2040
2041	return va;
2042
2043overflow:
2044	if (!purged) {
2045		reclaim_and_purge_vmap_areas();
2046		purged = 1;
2047		goto retry;
2048	}
2049
2050	freed = 0;
2051	blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
2052
2053	if (freed > 0) {
2054		purged = 0;
2055		goto retry;
2056	}
2057
2058	if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
2059		pr_warn("vmalloc_node_range for size %lu failed: Address range restricted to %#lx - %#lx\n",
2060				size, vstart, vend);
2061
2062	kmem_cache_free(vmap_area_cachep, va);
2063	return ERR_PTR(-EBUSY);
2064}
2065
2066int register_vmap_purge_notifier(struct notifier_block *nb)
2067{
2068	return blocking_notifier_chain_register(&vmap_notify_list, nb);
2069}
2070EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
2071
2072int unregister_vmap_purge_notifier(struct notifier_block *nb)
2073{
2074	return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
2075}
2076EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
2077
2078/*
2079 * lazy_max_pages is the maximum amount of virtual address space we gather up
2080 * before attempting to purge with a TLB flush.
2081 *
2082 * There is a tradeoff here: a larger number will cover more kernel page tables
2083 * and take slightly longer to purge, but it will linearly reduce the number of
2084 * global TLB flushes that must be performed. It would seem natural to scale
2085 * this number up linearly with the number of CPUs (because vmapping activity
2086 * could also scale linearly with the number of CPUs), however it is likely
2087 * that in practice, workloads might be constrained in other ways that mean
2088 * vmap activity will not scale linearly with CPUs. Also, I want to be
2089 * conservative and not introduce a big latency on huge systems, so go with
2090 * a less aggressive log scale. It will still be an improvement over the old
2091 * code, and it will be simple to change the scale factor if we find that it
2092 * becomes a problem on bigger systems.
2093 */
2094static unsigned long lazy_max_pages(void)
2095{
2096	unsigned int log;
2097
2098	log = fls(num_online_cpus());
2099
2100	return log * (32UL * 1024 * 1024 / PAGE_SIZE);
2101}
2102
2103static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
2104
2105/*
2106 * Serialize vmap purging.  There is no actual critical section protected
2107 * by this lock, but we want to avoid concurrent calls for performance
2108 * reasons and to make the pcpu_get_vm_areas more deterministic.
2109 */
2110static DEFINE_MUTEX(vmap_purge_lock);
2111
2112/* for per-CPU blocks */
2113static void purge_fragmented_blocks_allcpus(void);
2114static cpumask_t purge_nodes;
2115
2116static void
2117reclaim_list_global(struct list_head *head)
2118{
2119	struct vmap_area *va, *n;
2120
2121	if (list_empty(head))
2122		return;
2123
2124	spin_lock(&free_vmap_area_lock);
2125	list_for_each_entry_safe(va, n, head, list)
2126		merge_or_add_vmap_area_augment(va,
2127			&free_vmap_area_root, &free_vmap_area_list);
2128	spin_unlock(&free_vmap_area_lock);
2129}
2130
2131static void
2132decay_va_pool_node(struct vmap_node *vn, bool full_decay)
2133{
2134	LIST_HEAD(decay_list);
2135	struct rb_root decay_root = RB_ROOT;
2136	struct vmap_area *va, *nva;
2137	unsigned long n_decay;
2138	int i;
2139
2140	for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
2141		LIST_HEAD(tmp_list);
2142
2143		if (list_empty(&vn->pool[i].head))
2144			continue;
2145
2146		/* Detach the pool, so no-one can access it. */
2147		spin_lock(&vn->pool_lock);
2148		list_replace_init(&vn->pool[i].head, &tmp_list);
2149		spin_unlock(&vn->pool_lock);
2150
2151		if (full_decay)
2152			WRITE_ONCE(vn->pool[i].len, 0);
2153
2154		/* Decay a pool by ~25% out of left objects. */
2155		n_decay = vn->pool[i].len >> 2;
2156
2157		list_for_each_entry_safe(va, nva, &tmp_list, list) {
2158			list_del_init(&va->list);
2159			merge_or_add_vmap_area(va, &decay_root, &decay_list);
2160
2161			if (!full_decay) {
2162				WRITE_ONCE(vn->pool[i].len, vn->pool[i].len - 1);
2163
2164				if (!--n_decay)
2165					break;
2166			}
2167		}
2168
2169		/*
2170		 * Attach the pool back if it has been partly decayed.
2171		 * Please note, it is supposed that nobody(other contexts)
2172		 * can populate the pool therefore a simple list replace
2173		 * operation takes place here.
2174		 */
2175		if (!full_decay && !list_empty(&tmp_list)) {
2176			spin_lock(&vn->pool_lock);
2177			list_replace_init(&tmp_list, &vn->pool[i].head);
2178			spin_unlock(&vn->pool_lock);
2179		}
2180	}
2181
2182	reclaim_list_global(&decay_list);
2183}
2184
2185static void
2186kasan_release_vmalloc_node(struct vmap_node *vn)
2187{
2188	struct vmap_area *va;
2189	unsigned long start, end;
2190
2191	start = list_first_entry(&vn->purge_list, struct vmap_area, list)->va_start;
2192	end = list_last_entry(&vn->purge_list, struct vmap_area, list)->va_end;
2193
2194	list_for_each_entry(va, &vn->purge_list, list) {
2195		if (is_vmalloc_or_module_addr((void *) va->va_start))
2196			kasan_release_vmalloc(va->va_start, va->va_end,
2197				va->va_start, va->va_end,
2198				KASAN_VMALLOC_PAGE_RANGE);
2199	}
2200
2201	kasan_release_vmalloc(start, end, start, end, KASAN_VMALLOC_TLB_FLUSH);
2202}
2203
2204static void purge_vmap_node(struct work_struct *work)
2205{
2206	struct vmap_node *vn = container_of(work,
2207		struct vmap_node, purge_work);
2208	unsigned long nr_purged_pages = 0;
2209	struct vmap_area *va, *n_va;
2210	LIST_HEAD(local_list);
2211
2212	if (IS_ENABLED(CONFIG_KASAN_VMALLOC))
2213		kasan_release_vmalloc_node(vn);
2214
2215	vn->nr_purged = 0;
2216
2217	list_for_each_entry_safe(va, n_va, &vn->purge_list, list) {
2218		unsigned long nr = va_size(va) >> PAGE_SHIFT;
2219		unsigned int vn_id = decode_vn_id(va->flags);
2220
2221		list_del_init(&va->list);
2222
2223		nr_purged_pages += nr;
2224		vn->nr_purged++;
2225
2226		if (is_vn_id_valid(vn_id) && !vn->skip_populate)
2227			if (node_pool_add_va(vn, va))
2228				continue;
2229
2230		/* Go back to global. */
2231		list_add(&va->list, &local_list);
2232	}
2233
2234	atomic_long_sub(nr_purged_pages, &vmap_lazy_nr);
2235
2236	reclaim_list_global(&local_list);
2237}
2238
2239/*
2240 * Purges all lazily-freed vmap areas.
2241 */
2242static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end,
2243		bool full_pool_decay)
2244{
2245	unsigned long nr_purged_areas = 0;
2246	unsigned int nr_purge_helpers;
2247	unsigned int nr_purge_nodes;
2248	struct vmap_node *vn;
2249	int i;
2250
2251	lockdep_assert_held(&vmap_purge_lock);
2252
2253	/*
2254	 * Use cpumask to mark which node has to be processed.
2255	 */
2256	purge_nodes = CPU_MASK_NONE;
2257
2258	for (i = 0; i < nr_vmap_nodes; i++) {
2259		vn = &vmap_nodes[i];
2260
2261		INIT_LIST_HEAD(&vn->purge_list);
2262		vn->skip_populate = full_pool_decay;
2263		decay_va_pool_node(vn, full_pool_decay);
2264
2265		if (RB_EMPTY_ROOT(&vn->lazy.root))
2266			continue;
 
2267
2268		spin_lock(&vn->lazy.lock);
2269		WRITE_ONCE(vn->lazy.root.rb_node, NULL);
2270		list_replace_init(&vn->lazy.head, &vn->purge_list);
2271		spin_unlock(&vn->lazy.lock);
2272
2273		start = min(start, list_first_entry(&vn->purge_list,
2274			struct vmap_area, list)->va_start);
 
 
 
2275
2276		end = max(end, list_last_entry(&vn->purge_list,
2277			struct vmap_area, list)->va_end);
 
 
 
 
 
2278
2279		cpumask_set_cpu(i, &purge_nodes);
2280	}
2281
2282	nr_purge_nodes = cpumask_weight(&purge_nodes);
2283	if (nr_purge_nodes > 0) {
2284		flush_tlb_kernel_range(start, end);
2285
2286		/* One extra worker is per a lazy_max_pages() full set minus one. */
2287		nr_purge_helpers = atomic_long_read(&vmap_lazy_nr) / lazy_max_pages();
2288		nr_purge_helpers = clamp(nr_purge_helpers, 1U, nr_purge_nodes) - 1;
2289
2290		for_each_cpu(i, &purge_nodes) {
2291			vn = &vmap_nodes[i];
2292
2293			if (nr_purge_helpers > 0) {
2294				INIT_WORK(&vn->purge_work, purge_vmap_node);
2295
2296				if (cpumask_test_cpu(i, cpu_online_mask))
2297					schedule_work_on(i, &vn->purge_work);
2298				else
2299					schedule_work(&vn->purge_work);
2300
2301				nr_purge_helpers--;
2302			} else {
2303				vn->purge_work.func = NULL;
2304				purge_vmap_node(&vn->purge_work);
2305				nr_purged_areas += vn->nr_purged;
2306			}
2307		}
2308
2309		for_each_cpu(i, &purge_nodes) {
2310			vn = &vmap_nodes[i];
2311
2312			if (vn->purge_work.func) {
2313				flush_work(&vn->purge_work);
2314				nr_purged_areas += vn->nr_purged;
2315			}
2316		}
2317	}
 
2318
2319	trace_purge_vmap_area_lazy(start, end, nr_purged_areas);
2320	return nr_purged_areas > 0;
 
2321}
2322
2323/*
2324 * Reclaim vmap areas by purging fragmented blocks and purge_vmap_area_list.
2325 */
2326static void reclaim_and_purge_vmap_areas(void)
2327
2328{
2329	mutex_lock(&vmap_purge_lock);
2330	purge_fragmented_blocks_allcpus();
2331	__purge_vmap_area_lazy(ULONG_MAX, 0, true);
2332	mutex_unlock(&vmap_purge_lock);
2333}
2334
2335static void drain_vmap_area_work(struct work_struct *work)
2336{
2337	mutex_lock(&vmap_purge_lock);
2338	__purge_vmap_area_lazy(ULONG_MAX, 0, false);
2339	mutex_unlock(&vmap_purge_lock);
 
 
 
 
 
 
 
2340}
2341
2342/*
2343 * Free a vmap area, caller ensuring that the area has been unmapped,
2344 * unlinked and flush_cache_vunmap had been called for the correct
2345 * range previously.
2346 */
2347static void free_vmap_area_noflush(struct vmap_area *va)
2348{
2349	unsigned long nr_lazy_max = lazy_max_pages();
2350	unsigned long va_start = va->va_start;
2351	unsigned int vn_id = decode_vn_id(va->flags);
2352	struct vmap_node *vn;
2353	unsigned long nr_lazy;
2354
2355	if (WARN_ON_ONCE(!list_empty(&va->list)))
2356		return;
 
2357
2358	nr_lazy = atomic_long_add_return(va_size(va) >> PAGE_SHIFT,
2359					 &vmap_lazy_nr);
2360
2361	/*
2362	 * If it was request by a certain node we would like to
2363	 * return it to that node, i.e. its pool for later reuse.
2364	 */
2365	vn = is_vn_id_valid(vn_id) ?
2366		id_to_node(vn_id):addr_to_node(va->va_start);
2367
2368	spin_lock(&vn->lazy.lock);
2369	insert_vmap_area(va, &vn->lazy.root, &vn->lazy.head);
2370	spin_unlock(&vn->lazy.lock);
2371
2372	trace_free_vmap_area_noflush(va_start, nr_lazy, nr_lazy_max);
2373
2374	/* After this point, we may free va at any time */
2375	if (unlikely(nr_lazy > nr_lazy_max))
2376		schedule_work(&drain_vmap_work);
2377}
2378
2379/*
2380 * Free and unmap a vmap area
2381 */
2382static void free_unmap_vmap_area(struct vmap_area *va)
2383{
2384	flush_cache_vunmap(va->va_start, va->va_end);
2385	vunmap_range_noflush(va->va_start, va->va_end);
2386	if (debug_pagealloc_enabled_static())
2387		flush_tlb_kernel_range(va->va_start, va->va_end);
2388
2389	free_vmap_area_noflush(va);
2390}
2391
2392struct vmap_area *find_vmap_area(unsigned long addr)
2393{
2394	struct vmap_node *vn;
2395	struct vmap_area *va;
2396	int i, j;
2397
2398	if (unlikely(!vmap_initialized))
2399		return NULL;
 
2400
2401	/*
2402	 * An addr_to_node_id(addr) converts an address to a node index
2403	 * where a VA is located. If VA spans several zones and passed
2404	 * addr is not the same as va->va_start, what is not common, we
2405	 * may need to scan extra nodes. See an example:
2406	 *
2407	 *      <----va---->
2408	 * -|-----|-----|-----|-----|-
2409	 *     1     2     0     1
2410	 *
2411	 * VA resides in node 1 whereas it spans 1, 2 an 0. If passed
2412	 * addr is within 2 or 0 nodes we should do extra work.
2413	 */
2414	i = j = addr_to_node_id(addr);
2415	do {
2416		vn = &vmap_nodes[i];
2417
2418		spin_lock(&vn->busy.lock);
2419		va = __find_vmap_area(addr, &vn->busy.root);
2420		spin_unlock(&vn->busy.lock);
2421
2422		if (va)
2423			return va;
2424	} while ((i = (i + 1) % nr_vmap_nodes) != j);
2425
2426	return NULL;
2427}
2428
2429static struct vmap_area *find_unlink_vmap_area(unsigned long addr)
2430{
2431	struct vmap_node *vn;
2432	struct vmap_area *va;
2433	int i, j;
2434
2435	/*
2436	 * Check the comment in the find_vmap_area() about the loop.
2437	 */
2438	i = j = addr_to_node_id(addr);
2439	do {
2440		vn = &vmap_nodes[i];
2441
2442		spin_lock(&vn->busy.lock);
2443		va = __find_vmap_area(addr, &vn->busy.root);
2444		if (va)
2445			unlink_va(va, &vn->busy.root);
2446		spin_unlock(&vn->busy.lock);
2447
2448		if (va)
2449			return va;
2450	} while ((i = (i + 1) % nr_vmap_nodes) != j);
2451
2452	return NULL;
2453}
2454
2455/*** Per cpu kva allocator ***/
2456
2457/*
2458 * vmap space is limited especially on 32 bit architectures. Ensure there is
2459 * room for at least 16 percpu vmap blocks per CPU.
2460 */
2461/*
2462 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
2463 * to #define VMALLOC_SPACE		(VMALLOC_END-VMALLOC_START). Guess
2464 * instead (we just need a rough idea)
2465 */
2466#if BITS_PER_LONG == 32
2467#define VMALLOC_SPACE		(128UL*1024*1024)
2468#else
2469#define VMALLOC_SPACE		(128UL*1024*1024*1024)
2470#endif
2471
2472#define VMALLOC_PAGES		(VMALLOC_SPACE / PAGE_SIZE)
2473#define VMAP_MAX_ALLOC		BITS_PER_LONG	/* 256K with 4K pages */
2474#define VMAP_BBMAP_BITS_MAX	1024	/* 4MB with 4K pages */
2475#define VMAP_BBMAP_BITS_MIN	(VMAP_MAX_ALLOC*2)
2476#define VMAP_MIN(x, y)		((x) < (y) ? (x) : (y)) /* can't use min() */
2477#define VMAP_MAX(x, y)		((x) > (y) ? (x) : (y)) /* can't use max() */
2478#define VMAP_BBMAP_BITS		\
2479		VMAP_MIN(VMAP_BBMAP_BITS_MAX,	\
2480		VMAP_MAX(VMAP_BBMAP_BITS_MIN,	\
2481			VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
2482
2483#define VMAP_BLOCK_SIZE		(VMAP_BBMAP_BITS * PAGE_SIZE)
2484
2485/*
2486 * Purge threshold to prevent overeager purging of fragmented blocks for
2487 * regular operations: Purge if vb->free is less than 1/4 of the capacity.
2488 */
2489#define VMAP_PURGE_THRESHOLD	(VMAP_BBMAP_BITS / 4)
2490
2491#define VMAP_RAM		0x1 /* indicates vm_map_ram area*/
2492#define VMAP_BLOCK		0x2 /* mark out the vmap_block sub-type*/
2493#define VMAP_FLAGS_MASK		0x3
2494
2495struct vmap_block_queue {
2496	spinlock_t lock;
2497	struct list_head free;
2498
2499	/*
2500	 * An xarray requires an extra memory dynamically to
2501	 * be allocated. If it is an issue, we can use rb-tree
2502	 * instead.
2503	 */
2504	struct xarray vmap_blocks;
2505};
2506
2507struct vmap_block {
2508	spinlock_t lock;
2509	struct vmap_area *va;
2510	unsigned long free, dirty;
2511	DECLARE_BITMAP(used_map, VMAP_BBMAP_BITS);
2512	unsigned long dirty_min, dirty_max; /*< dirty range */
2513	struct list_head free_list;
2514	struct rcu_head rcu_head;
2515	struct list_head purge;
2516	unsigned int cpu;
2517};
2518
2519/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
2520static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
2521
2522/*
2523 * In order to fast access to any "vmap_block" associated with a
2524 * specific address, we use a hash.
2525 *
2526 * A per-cpu vmap_block_queue is used in both ways, to serialize
2527 * an access to free block chains among CPUs(alloc path) and it
2528 * also acts as a vmap_block hash(alloc/free paths). It means we
2529 * overload it, since we already have the per-cpu array which is
2530 * used as a hash table. When used as a hash a 'cpu' passed to
2531 * per_cpu() is not actually a CPU but rather a hash index.
2532 *
2533 * A hash function is addr_to_vb_xa() which hashes any address
2534 * to a specific index(in a hash) it belongs to. This then uses a
2535 * per_cpu() macro to access an array with generated index.
2536 *
2537 * An example:
2538 *
2539 *  CPU_1  CPU_2  CPU_0
2540 *    |      |      |
2541 *    V      V      V
2542 * 0     10     20     30     40     50     60
2543 * |------|------|------|------|------|------|...<vmap address space>
2544 *   CPU0   CPU1   CPU2   CPU0   CPU1   CPU2
2545 *
2546 * - CPU_1 invokes vm_unmap_ram(6), 6 belongs to CPU0 zone, thus
2547 *   it access: CPU0/INDEX0 -> vmap_blocks -> xa_lock;
2548 *
2549 * - CPU_2 invokes vm_unmap_ram(11), 11 belongs to CPU1 zone, thus
2550 *   it access: CPU1/INDEX1 -> vmap_blocks -> xa_lock;
2551 *
2552 * - CPU_0 invokes vm_unmap_ram(20), 20 belongs to CPU2 zone, thus
2553 *   it access: CPU2/INDEX2 -> vmap_blocks -> xa_lock.
2554 *
2555 * This technique almost always avoids lock contention on insert/remove,
2556 * however xarray spinlocks protect against any contention that remains.
2557 */
2558static struct xarray *
2559addr_to_vb_xa(unsigned long addr)
2560{
2561	int index = (addr / VMAP_BLOCK_SIZE) % nr_cpu_ids;
2562
2563	/*
2564	 * Please note, nr_cpu_ids points on a highest set
2565	 * possible bit, i.e. we never invoke cpumask_next()
2566	 * if an index points on it which is nr_cpu_ids - 1.
2567	 */
2568	if (!cpu_possible(index))
2569		index = cpumask_next(index, cpu_possible_mask);
2570
2571	return &per_cpu(vmap_block_queue, index).vmap_blocks;
2572}
2573
2574/*
2575 * We should probably have a fallback mechanism to allocate virtual memory
2576 * out of partially filled vmap blocks. However vmap block sizing should be
2577 * fairly reasonable according to the vmalloc size, so it shouldn't be a
2578 * big problem.
2579 */
2580
2581static unsigned long addr_to_vb_idx(unsigned long addr)
2582{
2583	addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
2584	addr /= VMAP_BLOCK_SIZE;
2585	return addr;
2586}
2587
2588static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
2589{
2590	unsigned long addr;
2591
2592	addr = va_start + (pages_off << PAGE_SHIFT);
2593	BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
2594	return (void *)addr;
2595}
2596
2597/**
2598 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
2599 *                  block. Of course pages number can't exceed VMAP_BBMAP_BITS
2600 * @order:    how many 2^order pages should be occupied in newly allocated block
2601 * @gfp_mask: flags for the page level allocator
2602 *
2603 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
2604 */
2605static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
2606{
2607	struct vmap_block_queue *vbq;
2608	struct vmap_block *vb;
2609	struct vmap_area *va;
2610	struct xarray *xa;
2611	unsigned long vb_idx;
2612	int node, err;
2613	void *vaddr;
2614
2615	node = numa_node_id();
2616
2617	vb = kmalloc_node(sizeof(struct vmap_block),
2618			gfp_mask & GFP_RECLAIM_MASK, node);
2619	if (unlikely(!vb))
2620		return ERR_PTR(-ENOMEM);
2621
2622	va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
2623					VMALLOC_START, VMALLOC_END,
2624					node, gfp_mask,
2625					VMAP_RAM|VMAP_BLOCK, NULL);
2626	if (IS_ERR(va)) {
2627		kfree(vb);
2628		return ERR_CAST(va);
2629	}
2630
2631	vaddr = vmap_block_vaddr(va->va_start, 0);
2632	spin_lock_init(&vb->lock);
2633	vb->va = va;
2634	/* At least something should be left free */
2635	BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
2636	bitmap_zero(vb->used_map, VMAP_BBMAP_BITS);
2637	vb->free = VMAP_BBMAP_BITS - (1UL << order);
2638	vb->dirty = 0;
2639	vb->dirty_min = VMAP_BBMAP_BITS;
2640	vb->dirty_max = 0;
2641	bitmap_set(vb->used_map, 0, (1UL << order));
2642	INIT_LIST_HEAD(&vb->free_list);
2643	vb->cpu = raw_smp_processor_id();
2644
2645	xa = addr_to_vb_xa(va->va_start);
2646	vb_idx = addr_to_vb_idx(va->va_start);
2647	err = xa_insert(xa, vb_idx, vb, gfp_mask);
2648	if (err) {
2649		kfree(vb);
2650		free_vmap_area(va);
2651		return ERR_PTR(err);
2652	}
2653	/*
2654	 * list_add_tail_rcu could happened in another core
2655	 * rather than vb->cpu due to task migration, which
2656	 * is safe as list_add_tail_rcu will ensure the list's
2657	 * integrity together with list_for_each_rcu from read
2658	 * side.
2659	 */
2660	vbq = per_cpu_ptr(&vmap_block_queue, vb->cpu);
2661	spin_lock(&vbq->lock);
2662	list_add_tail_rcu(&vb->free_list, &vbq->free);
2663	spin_unlock(&vbq->lock);
2664
2665	return vaddr;
2666}
2667
2668static void free_vmap_block(struct vmap_block *vb)
2669{
2670	struct vmap_node *vn;
2671	struct vmap_block *tmp;
2672	struct xarray *xa;
2673
2674	xa = addr_to_vb_xa(vb->va->va_start);
2675	tmp = xa_erase(xa, addr_to_vb_idx(vb->va->va_start));
2676	BUG_ON(tmp != vb);
2677
2678	vn = addr_to_node(vb->va->va_start);
2679	spin_lock(&vn->busy.lock);
2680	unlink_va(vb->va, &vn->busy.root);
2681	spin_unlock(&vn->busy.lock);
2682
2683	free_vmap_area_noflush(vb->va);
2684	kfree_rcu(vb, rcu_head);
2685}
2686
2687static bool purge_fragmented_block(struct vmap_block *vb,
2688		struct list_head *purge_list, bool force_purge)
2689{
2690	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, vb->cpu);
2691
2692	if (vb->free + vb->dirty != VMAP_BBMAP_BITS ||
2693	    vb->dirty == VMAP_BBMAP_BITS)
2694		return false;
2695
2696	/* Don't overeagerly purge usable blocks unless requested */
2697	if (!(force_purge || vb->free < VMAP_PURGE_THRESHOLD))
2698		return false;
2699
2700	/* prevent further allocs after releasing lock */
2701	WRITE_ONCE(vb->free, 0);
2702	/* prevent purging it again */
2703	WRITE_ONCE(vb->dirty, VMAP_BBMAP_BITS);
2704	vb->dirty_min = 0;
2705	vb->dirty_max = VMAP_BBMAP_BITS;
2706	spin_lock(&vbq->lock);
2707	list_del_rcu(&vb->free_list);
2708	spin_unlock(&vbq->lock);
2709	list_add_tail(&vb->purge, purge_list);
2710	return true;
2711}
2712
2713static void free_purged_blocks(struct list_head *purge_list)
2714{
2715	struct vmap_block *vb, *n_vb;
2716
2717	list_for_each_entry_safe(vb, n_vb, purge_list, purge) {
2718		list_del(&vb->purge);
2719		free_vmap_block(vb);
2720	}
2721}
2722
2723static void purge_fragmented_blocks(int cpu)
2724{
2725	LIST_HEAD(purge);
2726	struct vmap_block *vb;
 
2727	struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2728
2729	rcu_read_lock();
2730	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2731		unsigned long free = READ_ONCE(vb->free);
2732		unsigned long dirty = READ_ONCE(vb->dirty);
2733
2734		if (free + dirty != VMAP_BBMAP_BITS ||
2735		    dirty == VMAP_BBMAP_BITS)
2736			continue;
2737
2738		spin_lock(&vb->lock);
2739		purge_fragmented_block(vb, &purge, true);
2740		spin_unlock(&vb->lock);
 
 
 
 
 
 
 
 
 
 
2741	}
2742	rcu_read_unlock();
2743	free_purged_blocks(&purge);
 
 
 
 
2744}
2745
2746static void purge_fragmented_blocks_allcpus(void)
2747{
2748	int cpu;
2749
2750	for_each_possible_cpu(cpu)
2751		purge_fragmented_blocks(cpu);
2752}
2753
2754static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
2755{
2756	struct vmap_block_queue *vbq;
2757	struct vmap_block *vb;
2758	void *vaddr = NULL;
2759	unsigned int order;
2760
2761	BUG_ON(offset_in_page(size));
2762	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2763	if (WARN_ON(size == 0)) {
2764		/*
2765		 * Allocating 0 bytes isn't what caller wants since
2766		 * get_order(0) returns funny result. Just warn and terminate
2767		 * early.
2768		 */
2769		return ERR_PTR(-EINVAL);
2770	}
2771	order = get_order(size);
2772
2773	rcu_read_lock();
2774	vbq = raw_cpu_ptr(&vmap_block_queue);
2775	list_for_each_entry_rcu(vb, &vbq->free, free_list) {
2776		unsigned long pages_off;
2777
2778		if (READ_ONCE(vb->free) < (1UL << order))
2779			continue;
2780
2781		spin_lock(&vb->lock);
2782		if (vb->free < (1UL << order)) {
2783			spin_unlock(&vb->lock);
2784			continue;
2785		}
2786
2787		pages_off = VMAP_BBMAP_BITS - vb->free;
2788		vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
2789		WRITE_ONCE(vb->free, vb->free - (1UL << order));
2790		bitmap_set(vb->used_map, pages_off, (1UL << order));
2791		if (vb->free == 0) {
2792			spin_lock(&vbq->lock);
2793			list_del_rcu(&vb->free_list);
2794			spin_unlock(&vbq->lock);
2795		}
2796
2797		spin_unlock(&vb->lock);
2798		break;
2799	}
2800
2801	rcu_read_unlock();
2802
2803	/* Allocate new block if nothing was found */
2804	if (!vaddr)
2805		vaddr = new_vmap_block(order, gfp_mask);
2806
2807	return vaddr;
2808}
2809
2810static void vb_free(unsigned long addr, unsigned long size)
2811{
2812	unsigned long offset;
2813	unsigned int order;
2814	struct vmap_block *vb;
2815	struct xarray *xa;
2816
2817	BUG_ON(offset_in_page(size));
2818	BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
2819
2820	flush_cache_vunmap(addr, addr + size);
2821
2822	order = get_order(size);
2823	offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
2824
2825	xa = addr_to_vb_xa(addr);
2826	vb = xa_load(xa, addr_to_vb_idx(addr));
2827
2828	spin_lock(&vb->lock);
2829	bitmap_clear(vb->used_map, offset, (1UL << order));
2830	spin_unlock(&vb->lock);
2831
2832	vunmap_range_noflush(addr, addr + size);
2833
2834	if (debug_pagealloc_enabled_static())
2835		flush_tlb_kernel_range(addr, addr + size);
2836
2837	spin_lock(&vb->lock);
2838
2839	/* Expand the not yet TLB flushed dirty range */
2840	vb->dirty_min = min(vb->dirty_min, offset);
2841	vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
2842
2843	WRITE_ONCE(vb->dirty, vb->dirty + (1UL << order));
2844	if (vb->dirty == VMAP_BBMAP_BITS) {
2845		BUG_ON(vb->free);
2846		spin_unlock(&vb->lock);
2847		free_vmap_block(vb);
2848	} else
2849		spin_unlock(&vb->lock);
2850}
2851
2852static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
2853{
2854	LIST_HEAD(purge_list);
2855	int cpu;
2856
2857	if (unlikely(!vmap_initialized))
2858		return;
2859
2860	mutex_lock(&vmap_purge_lock);
2861
2862	for_each_possible_cpu(cpu) {
2863		struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
2864		struct vmap_block *vb;
2865		unsigned long idx;
2866
2867		rcu_read_lock();
2868		xa_for_each(&vbq->vmap_blocks, idx, vb) {
2869			spin_lock(&vb->lock);
2870
2871			/*
2872			 * Try to purge a fragmented block first. If it's
2873			 * not purgeable, check whether there is dirty
2874			 * space to be flushed.
2875			 */
2876			if (!purge_fragmented_block(vb, &purge_list, false) &&
2877			    vb->dirty_max && vb->dirty != VMAP_BBMAP_BITS) {
2878				unsigned long va_start = vb->va->va_start;
2879				unsigned long s, e;
2880
2881				s = va_start + (vb->dirty_min << PAGE_SHIFT);
2882				e = va_start + (vb->dirty_max << PAGE_SHIFT);
2883
2884				start = min(s, start);
2885				end   = max(e, end);
2886
2887				/* Prevent that this is flushed again */
2888				vb->dirty_min = VMAP_BBMAP_BITS;
2889				vb->dirty_max = 0;
2890
2891				flush = 1;
2892			}
2893			spin_unlock(&vb->lock);
2894		}
2895		rcu_read_unlock();
2896	}
2897	free_purged_blocks(&purge_list);
2898
2899	if (!__purge_vmap_area_lazy(start, end, false) && flush)
 
 
2900		flush_tlb_kernel_range(start, end);
2901	mutex_unlock(&vmap_purge_lock);
2902}
2903
2904/**
2905 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
2906 *
2907 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
2908 * to amortize TLB flushing overheads. What this means is that any page you
2909 * have now, may, in a former life, have been mapped into kernel virtual
2910 * address by the vmap layer and so there might be some CPUs with TLB entries
2911 * still referencing that page (additional to the regular 1:1 kernel mapping).
2912 *
2913 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
2914 * be sure that none of the pages we have control over will have any aliases
2915 * from the vmap layer.
2916 */
2917void vm_unmap_aliases(void)
2918{
2919	unsigned long start = ULONG_MAX, end = 0;
2920	int flush = 0;
2921
2922	_vm_unmap_aliases(start, end, flush);
2923}
2924EXPORT_SYMBOL_GPL(vm_unmap_aliases);
2925
2926/**
2927 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
2928 * @mem: the pointer returned by vm_map_ram
2929 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
2930 */
2931void vm_unmap_ram(const void *mem, unsigned int count)
2932{
2933	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2934	unsigned long addr = (unsigned long)kasan_reset_tag(mem);
2935	struct vmap_area *va;
2936
2937	might_sleep();
2938	BUG_ON(!addr);
2939	BUG_ON(addr < VMALLOC_START);
2940	BUG_ON(addr > VMALLOC_END);
2941	BUG_ON(!PAGE_ALIGNED(addr));
2942
2943	kasan_poison_vmalloc(mem, size);
2944
2945	if (likely(count <= VMAP_MAX_ALLOC)) {
2946		debug_check_no_locks_freed(mem, size);
2947		vb_free(addr, size);
2948		return;
2949	}
2950
2951	va = find_unlink_vmap_area(addr);
2952	if (WARN_ON_ONCE(!va))
2953		return;
2954
2955	debug_check_no_locks_freed((void *)va->va_start, va_size(va));
2956	free_unmap_vmap_area(va);
2957}
2958EXPORT_SYMBOL(vm_unmap_ram);
2959
2960/**
2961 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
2962 * @pages: an array of pointers to the pages to be mapped
2963 * @count: number of pages
2964 * @node: prefer to allocate data structures on this node
2965 *
2966 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
2967 * faster than vmap so it's good.  But if you mix long-life and short-life
2968 * objects with vm_map_ram(), it could consume lots of address space through
2969 * fragmentation (especially on a 32bit machine).  You could see failures in
2970 * the end.  Please use this function for short-lived objects.
2971 *
2972 * Returns: a pointer to the address that has been mapped, or %NULL on failure
2973 */
2974void *vm_map_ram(struct page **pages, unsigned int count, int node)
2975{
2976	unsigned long size = (unsigned long)count << PAGE_SHIFT;
2977	unsigned long addr;
2978	void *mem;
2979
2980	if (likely(count <= VMAP_MAX_ALLOC)) {
2981		mem = vb_alloc(size, GFP_KERNEL);
2982		if (IS_ERR(mem))
2983			return NULL;
2984		addr = (unsigned long)mem;
2985	} else {
2986		struct vmap_area *va;
2987		va = alloc_vmap_area(size, PAGE_SIZE,
2988				VMALLOC_START, VMALLOC_END,
2989				node, GFP_KERNEL, VMAP_RAM,
2990				NULL);
2991		if (IS_ERR(va))
2992			return NULL;
2993
2994		addr = va->va_start;
2995		mem = (void *)addr;
2996	}
2997
2998	if (vmap_pages_range(addr, addr + size, PAGE_KERNEL,
2999				pages, PAGE_SHIFT) < 0) {
3000		vm_unmap_ram(mem, count);
3001		return NULL;
3002	}
3003
3004	/*
3005	 * Mark the pages as accessible, now that they are mapped.
3006	 * With hardware tag-based KASAN, marking is skipped for
3007	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3008	 */
3009	mem = kasan_unpoison_vmalloc(mem, size, KASAN_VMALLOC_PROT_NORMAL);
3010
3011	return mem;
3012}
3013EXPORT_SYMBOL(vm_map_ram);
3014
3015static struct vm_struct *vmlist __initdata;
3016
3017static inline unsigned int vm_area_page_order(struct vm_struct *vm)
3018{
3019#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
3020	return vm->page_order;
3021#else
3022	return 0;
3023#endif
3024}
3025
3026unsigned int get_vm_area_page_order(struct vm_struct *vm)
3027{
3028	return vm_area_page_order(vm);
3029}
3030
3031static inline void set_vm_area_page_order(struct vm_struct *vm, unsigned int order)
3032{
3033#ifdef CONFIG_HAVE_ARCH_HUGE_VMALLOC
3034	vm->page_order = order;
3035#else
3036	BUG_ON(order != 0);
3037#endif
3038}
3039
3040/**
3041 * vm_area_add_early - add vmap area early during boot
3042 * @vm: vm_struct to add
3043 *
3044 * This function is used to add fixed kernel vm area to vmlist before
3045 * vmalloc_init() is called.  @vm->addr, @vm->size, and @vm->flags
3046 * should contain proper values and the other fields should be zero.
3047 *
3048 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3049 */
3050void __init vm_area_add_early(struct vm_struct *vm)
3051{
3052	struct vm_struct *tmp, **p;
3053
3054	BUG_ON(vmap_initialized);
3055	for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
3056		if (tmp->addr >= vm->addr) {
3057			BUG_ON(tmp->addr < vm->addr + vm->size);
3058			break;
3059		} else
3060			BUG_ON(tmp->addr + tmp->size > vm->addr);
3061	}
3062	vm->next = *p;
3063	*p = vm;
3064}
3065
3066/**
3067 * vm_area_register_early - register vmap area early during boot
3068 * @vm: vm_struct to register
3069 * @align: requested alignment
3070 *
3071 * This function is used to register kernel vm area before
3072 * vmalloc_init() is called.  @vm->size and @vm->flags should contain
3073 * proper values on entry and other fields should be zero.  On return,
3074 * vm->addr contains the allocated address.
3075 *
3076 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
3077 */
3078void __init vm_area_register_early(struct vm_struct *vm, size_t align)
3079{
3080	unsigned long addr = ALIGN(VMALLOC_START, align);
3081	struct vm_struct *cur, **p;
3082
3083	BUG_ON(vmap_initialized);
3084
3085	for (p = &vmlist; (cur = *p) != NULL; p = &cur->next) {
3086		if ((unsigned long)cur->addr - addr >= vm->size)
3087			break;
3088		addr = ALIGN((unsigned long)cur->addr + cur->size, align);
3089	}
3090
3091	BUG_ON(addr > VMALLOC_END - vm->size);
3092	vm->addr = (void *)addr;
3093	vm->next = *p;
3094	*p = vm;
3095	kasan_populate_early_vm_area_shadow(vm->addr, vm->size);
3096}
3097
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3098static void clear_vm_uninitialized_flag(struct vm_struct *vm)
3099{
3100	/*
3101	 * Before removing VM_UNINITIALIZED,
3102	 * we should make sure that vm has proper values.
3103	 * Pair with smp_rmb() in show_numa_info().
3104	 */
3105	smp_wmb();
3106	vm->flags &= ~VM_UNINITIALIZED;
3107}
3108
3109struct vm_struct *__get_vm_area_node(unsigned long size,
3110		unsigned long align, unsigned long shift, unsigned long flags,
3111		unsigned long start, unsigned long end, int node,
3112		gfp_t gfp_mask, const void *caller)
3113{
3114	struct vmap_area *va;
3115	struct vm_struct *area;
3116	unsigned long requested_size = size;
3117
3118	BUG_ON(in_interrupt());
3119	size = ALIGN(size, 1ul << shift);
3120	if (unlikely(!size))
3121		return NULL;
3122
3123	if (flags & VM_IOREMAP)
3124		align = 1ul << clamp_t(int, get_count_order_long(size),
3125				       PAGE_SHIFT, IOREMAP_MAX_ORDER);
3126
3127	area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
3128	if (unlikely(!area))
3129		return NULL;
3130
3131	if (!(flags & VM_NO_GUARD))
3132		size += PAGE_SIZE;
3133
3134	area->flags = flags;
3135	area->caller = caller;
3136
3137	va = alloc_vmap_area(size, align, start, end, node, gfp_mask, 0, area);
3138	if (IS_ERR(va)) {
3139		kfree(area);
3140		return NULL;
3141	}
3142
 
 
3143	/*
3144	 * Mark pages for non-VM_ALLOC mappings as accessible. Do it now as a
3145	 * best-effort approach, as they can be mapped outside of vmalloc code.
3146	 * For VM_ALLOC mappings, the pages are marked as accessible after
3147	 * getting mapped in __vmalloc_node_range().
3148	 * With hardware tag-based KASAN, marking is skipped for
3149	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
3150	 */
3151	if (!(flags & VM_ALLOC))
3152		area->addr = kasan_unpoison_vmalloc(area->addr, requested_size,
3153						    KASAN_VMALLOC_PROT_NORMAL);
3154
3155	return area;
3156}
3157
3158struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
3159				       unsigned long start, unsigned long end,
3160				       const void *caller)
3161{
3162	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags, start, end,
3163				  NUMA_NO_NODE, GFP_KERNEL, caller);
3164}
3165
3166/**
3167 * get_vm_area - reserve a contiguous kernel virtual area
3168 * @size:	 size of the area
3169 * @flags:	 %VM_IOREMAP for I/O mappings or VM_ALLOC
3170 *
3171 * Search an area of @size in the kernel virtual mapping area,
3172 * and reserved it for out purposes.  Returns the area descriptor
3173 * on success or %NULL on failure.
3174 *
3175 * Return: the area descriptor on success or %NULL on failure.
3176 */
3177struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
3178{
3179	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3180				  VMALLOC_START, VMALLOC_END,
3181				  NUMA_NO_NODE, GFP_KERNEL,
3182				  __builtin_return_address(0));
3183}
3184
3185struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
3186				const void *caller)
3187{
3188	return __get_vm_area_node(size, 1, PAGE_SHIFT, flags,
3189				  VMALLOC_START, VMALLOC_END,
3190				  NUMA_NO_NODE, GFP_KERNEL, caller);
3191}
3192
3193/**
3194 * find_vm_area - find a continuous kernel virtual area
3195 * @addr:	  base address
3196 *
3197 * Search for the kernel VM area starting at @addr, and return it.
3198 * It is up to the caller to do all required locking to keep the returned
3199 * pointer valid.
3200 *
3201 * Return: the area descriptor on success or %NULL on failure.
3202 */
3203struct vm_struct *find_vm_area(const void *addr)
3204{
3205	struct vmap_area *va;
3206
3207	va = find_vmap_area((unsigned long)addr);
3208	if (!va)
3209		return NULL;
3210
3211	return va->vm;
3212}
3213
3214/**
3215 * remove_vm_area - find and remove a continuous kernel virtual area
3216 * @addr:	    base address
3217 *
3218 * Search for the kernel VM area starting at @addr, and remove it.
3219 * This function returns the found VM area, but using it is NOT safe
3220 * on SMP machines, except for its size or flags.
3221 *
3222 * Return: the area descriptor on success or %NULL on failure.
3223 */
3224struct vm_struct *remove_vm_area(const void *addr)
3225{
3226	struct vmap_area *va;
3227	struct vm_struct *vm;
3228
3229	might_sleep();
3230
3231	if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
3232			addr))
3233		return NULL;
 
 
 
 
3234
3235	va = find_unlink_vmap_area((unsigned long)addr);
3236	if (!va || !va->vm)
3237		return NULL;
3238	vm = va->vm;
3239
3240	debug_check_no_locks_freed(vm->addr, get_vm_area_size(vm));
3241	debug_check_no_obj_freed(vm->addr, get_vm_area_size(vm));
3242	kasan_free_module_shadow(vm);
3243	kasan_poison_vmalloc(vm->addr, get_vm_area_size(vm));
3244
3245	free_unmap_vmap_area(va);
3246	return vm;
3247}
3248
3249static inline void set_area_direct_map(const struct vm_struct *area,
3250				       int (*set_direct_map)(struct page *page))
3251{
3252	int i;
3253
3254	/* HUGE_VMALLOC passes small pages to set_direct_map */
3255	for (i = 0; i < area->nr_pages; i++)
3256		if (page_address(area->pages[i]))
3257			set_direct_map(area->pages[i]);
3258}
3259
3260/*
3261 * Flush the vm mapping and reset the direct map.
3262 */
3263static void vm_reset_perms(struct vm_struct *area)
3264{
3265	unsigned long start = ULONG_MAX, end = 0;
3266	unsigned int page_order = vm_area_page_order(area);
 
3267	int flush_dmap = 0;
3268	int i;
3269
 
 
 
 
 
 
3270	/*
3271	 * Find the start and end range of the direct mappings to make sure that
 
 
 
 
 
 
 
 
 
 
3272	 * the vm_unmap_aliases() flush includes the direct map.
3273	 */
3274	for (i = 0; i < area->nr_pages; i += 1U << page_order) {
3275		unsigned long addr = (unsigned long)page_address(area->pages[i]);
3276
3277		if (addr) {
3278			unsigned long page_size;
3279
3280			page_size = PAGE_SIZE << page_order;
3281			start = min(addr, start);
3282			end = max(addr + page_size, end);
3283			flush_dmap = 1;
3284		}
3285	}
3286
3287	/*
3288	 * Set direct map to something invalid so that it won't be cached if
3289	 * there are any accesses after the TLB flush, then flush the TLB and
3290	 * reset the direct map permissions to the default.
3291	 */
3292	set_area_direct_map(area, set_direct_map_invalid_noflush);
3293	_vm_unmap_aliases(start, end, flush_dmap);
3294	set_area_direct_map(area, set_direct_map_default_noflush);
3295}
3296
3297static void delayed_vfree_work(struct work_struct *w)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3298{
3299	struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
3300	struct llist_node *t, *llnode;
 
 
 
 
 
3301
3302	llist_for_each_safe(llnode, t, llist_del_all(&p->list))
3303		vfree(llnode);
3304}
3305
3306/**
3307 * vfree_atomic - release memory allocated by vmalloc()
3308 * @addr:	  memory base address
3309 *
3310 * This one is just like vfree() but can be called in any atomic context
3311 * except NMIs.
3312 */
3313void vfree_atomic(const void *addr)
3314{
3315	struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
3316
3317	BUG_ON(in_nmi());
3318	kmemleak_free(addr);
3319
3320	/*
3321	 * Use raw_cpu_ptr() because this can be called from preemptible
3322	 * context. Preemption is absolutely fine here, because the llist_add()
3323	 * implementation is lockless, so it works even if we are adding to
3324	 * another cpu's list. schedule_work() should be fine with this too.
3325	 */
3326	if (addr && llist_add((struct llist_node *)addr, &p->list))
3327		schedule_work(&p->wq);
 
 
 
3328}
3329
3330/**
3331 * vfree - Release memory allocated by vmalloc()
3332 * @addr:  Memory base address
3333 *
3334 * Free the virtually continuous memory area starting at @addr, as obtained
3335 * from one of the vmalloc() family of APIs.  This will usually also free the
3336 * physical memory underlying the virtual allocation, but that memory is
3337 * reference counted, so it will not be freed until the last user goes away.
3338 *
3339 * If @addr is NULL, no operation is performed.
3340 *
3341 * Context:
3342 * May sleep if called *not* from interrupt context.
3343 * Must not be called in NMI context (strictly speaking, it could be
3344 * if we have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
3345 * conventions for vfree() arch-dependent would be a really bad idea).
3346 */
3347void vfree(const void *addr)
3348{
3349	struct vm_struct *vm;
3350	int i;
3351
3352	if (unlikely(in_interrupt())) {
3353		vfree_atomic(addr);
3354		return;
3355	}
3356
3357	BUG_ON(in_nmi());
3358	kmemleak_free(addr);
3359	might_sleep();
3360
3361	if (!addr)
3362		return;
3363
3364	vm = remove_vm_area(addr);
3365	if (unlikely(!vm)) {
3366		WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
3367				addr);
3368		return;
3369	}
3370
3371	if (unlikely(vm->flags & VM_FLUSH_RESET_PERMS))
3372		vm_reset_perms(vm);
3373	for (i = 0; i < vm->nr_pages; i++) {
3374		struct page *page = vm->pages[i];
3375
3376		BUG_ON(!page);
3377		if (!(vm->flags & VM_MAP_PUT_PAGES))
3378			mod_memcg_page_state(page, MEMCG_VMALLOC, -1);
3379		/*
3380		 * High-order allocs for huge vmallocs are split, so
3381		 * can be freed as an array of order-0 allocations
3382		 */
3383		__free_page(page);
3384		cond_resched();
3385	}
3386	if (!(vm->flags & VM_MAP_PUT_PAGES))
3387		atomic_long_sub(vm->nr_pages, &nr_vmalloc_pages);
3388	kvfree(vm->pages);
3389	kfree(vm);
3390}
3391EXPORT_SYMBOL(vfree);
3392
3393/**
3394 * vunmap - release virtual mapping obtained by vmap()
3395 * @addr:   memory base address
3396 *
3397 * Free the virtually contiguous memory area starting at @addr,
3398 * which was created from the page array passed to vmap().
3399 *
3400 * Must not be called in interrupt context.
3401 */
3402void vunmap(const void *addr)
3403{
3404	struct vm_struct *vm;
3405
3406	BUG_ON(in_interrupt());
3407	might_sleep();
3408
3409	if (!addr)
3410		return;
3411	vm = remove_vm_area(addr);
3412	if (unlikely(!vm)) {
3413		WARN(1, KERN_ERR "Trying to vunmap() nonexistent vm area (%p)\n",
3414				addr);
3415		return;
3416	}
3417	kfree(vm);
3418}
3419EXPORT_SYMBOL(vunmap);
3420
3421/**
3422 * vmap - map an array of pages into virtually contiguous space
3423 * @pages: array of page pointers
3424 * @count: number of pages to map
3425 * @flags: vm_area->flags
3426 * @prot: page protection for the mapping
3427 *
3428 * Maps @count pages from @pages into contiguous kernel virtual space.
3429 * If @flags contains %VM_MAP_PUT_PAGES the ownership of the pages array itself
3430 * (which must be kmalloc or vmalloc memory) and one reference per pages in it
3431 * are transferred from the caller to vmap(), and will be freed / dropped when
3432 * vfree() is called on the return value.
3433 *
3434 * Return: the address of the area or %NULL on failure
3435 */
3436void *vmap(struct page **pages, unsigned int count,
3437	   unsigned long flags, pgprot_t prot)
3438{
3439	struct vm_struct *area;
3440	unsigned long addr;
3441	unsigned long size;		/* In bytes */
3442
3443	might_sleep();
3444
3445	if (WARN_ON_ONCE(flags & VM_FLUSH_RESET_PERMS))
3446		return NULL;
3447
3448	/*
3449	 * Your top guard is someone else's bottom guard. Not having a top
3450	 * guard compromises someone else's mappings too.
3451	 */
3452	if (WARN_ON_ONCE(flags & VM_NO_GUARD))
3453		flags &= ~VM_NO_GUARD;
3454
3455	if (count > totalram_pages())
3456		return NULL;
3457
3458	size = (unsigned long)count << PAGE_SHIFT;
3459	area = get_vm_area_caller(size, flags, __builtin_return_address(0));
3460	if (!area)
3461		return NULL;
3462
3463	addr = (unsigned long)area->addr;
3464	if (vmap_pages_range(addr, addr + size, pgprot_nx(prot),
3465				pages, PAGE_SHIFT) < 0) {
3466		vunmap(area->addr);
3467		return NULL;
3468	}
3469
3470	if (flags & VM_MAP_PUT_PAGES) {
3471		area->pages = pages;
3472		area->nr_pages = count;
3473	}
3474	return area->addr;
3475}
3476EXPORT_SYMBOL(vmap);
3477
3478#ifdef CONFIG_VMAP_PFN
3479struct vmap_pfn_data {
3480	unsigned long	*pfns;
3481	pgprot_t	prot;
3482	unsigned int	idx;
3483};
3484
3485static int vmap_pfn_apply(pte_t *pte, unsigned long addr, void *private)
3486{
3487	struct vmap_pfn_data *data = private;
3488	unsigned long pfn = data->pfns[data->idx];
3489	pte_t ptent;
3490
3491	if (WARN_ON_ONCE(pfn_valid(pfn)))
3492		return -EINVAL;
3493
3494	ptent = pte_mkspecial(pfn_pte(pfn, data->prot));
3495	set_pte_at(&init_mm, addr, pte, ptent);
3496
3497	data->idx++;
3498	return 0;
3499}
3500
3501/**
3502 * vmap_pfn - map an array of PFNs into virtually contiguous space
3503 * @pfns: array of PFNs
3504 * @count: number of pages to map
3505 * @prot: page protection for the mapping
3506 *
3507 * Maps @count PFNs from @pfns into contiguous kernel virtual space and returns
3508 * the start address of the mapping.
3509 */
3510void *vmap_pfn(unsigned long *pfns, unsigned int count, pgprot_t prot)
3511{
3512	struct vmap_pfn_data data = { .pfns = pfns, .prot = pgprot_nx(prot) };
3513	struct vm_struct *area;
3514
3515	area = get_vm_area_caller(count * PAGE_SIZE, VM_IOREMAP,
3516			__builtin_return_address(0));
3517	if (!area)
3518		return NULL;
3519	if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3520			count * PAGE_SIZE, vmap_pfn_apply, &data)) {
3521		free_vm_area(area);
3522		return NULL;
3523	}
3524
3525	flush_cache_vmap((unsigned long)area->addr,
3526			 (unsigned long)area->addr + count * PAGE_SIZE);
3527
3528	return area->addr;
3529}
3530EXPORT_SYMBOL_GPL(vmap_pfn);
3531#endif /* CONFIG_VMAP_PFN */
3532
3533static inline unsigned int
3534vm_area_alloc_pages(gfp_t gfp, int nid,
3535		unsigned int order, unsigned int nr_pages, struct page **pages)
3536{
3537	unsigned int nr_allocated = 0;
3538	struct page *page;
3539	int i;
3540
3541	/*
3542	 * For order-0 pages we make use of bulk allocator, if
3543	 * the page array is partly or not at all populated due
3544	 * to fails, fallback to a single page allocator that is
3545	 * more permissive.
3546	 */
3547	if (!order) {
 
 
3548		while (nr_allocated < nr_pages) {
3549			unsigned int nr, nr_pages_request;
3550
3551			/*
3552			 * A maximum allowed request is hard-coded and is 100
3553			 * pages per call. That is done in order to prevent a
3554			 * long preemption off scenario in the bulk-allocator
3555			 * so the range is [1:100].
3556			 */
3557			nr_pages_request = min(100U, nr_pages - nr_allocated);
3558
3559			/* memory allocation should consider mempolicy, we can't
3560			 * wrongly use nearest node when nid == NUMA_NO_NODE,
3561			 * otherwise memory may be allocated in only one node,
3562			 * but mempolicy wants to alloc memory by interleaving.
3563			 */
3564			if (IS_ENABLED(CONFIG_NUMA) && nid == NUMA_NO_NODE)
3565				nr = alloc_pages_bulk_array_mempolicy_noprof(gfp,
3566							nr_pages_request,
3567							pages + nr_allocated);
 
3568			else
3569				nr = alloc_pages_bulk_array_node_noprof(gfp, nid,
3570							nr_pages_request,
3571							pages + nr_allocated);
3572
3573			nr_allocated += nr;
3574			cond_resched();
3575
3576			/*
3577			 * If zero or pages were obtained partly,
3578			 * fallback to a single page allocator.
3579			 */
3580			if (nr != nr_pages_request)
3581				break;
3582		}
3583	}
3584
3585	/* High-order pages or fallback path if "bulk" fails. */
 
3586	while (nr_allocated < nr_pages) {
3587		if (!(gfp & __GFP_NOFAIL) && fatal_signal_pending(current))
3588			break;
3589
3590		if (nid == NUMA_NO_NODE)
3591			page = alloc_pages_noprof(gfp, order);
3592		else
3593			page = alloc_pages_node_noprof(nid, gfp, order);
3594
3595		if (unlikely(!page))
3596			break;
3597
3598		/*
3599		 * High-order allocations must be able to be treated as
3600		 * independent small pages by callers (as they can with
3601		 * small-page vmallocs). Some drivers do their own refcounting
3602		 * on vmalloc_to_page() pages, some use page->mapping,
3603		 * page->lru, etc.
3604		 */
3605		if (order)
3606			split_page(page, order);
3607
3608		/*
3609		 * Careful, we allocate and map page-order pages, but
3610		 * tracking is done per PAGE_SIZE page so as to keep the
3611		 * vm_struct APIs independent of the physical/mapped size.
3612		 */
3613		for (i = 0; i < (1U << order); i++)
3614			pages[nr_allocated + i] = page + i;
3615
3616		cond_resched();
3617		nr_allocated += 1U << order;
3618	}
3619
3620	return nr_allocated;
3621}
3622
3623static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
3624				 pgprot_t prot, unsigned int page_shift,
3625				 int node)
3626{
3627	const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
3628	bool nofail = gfp_mask & __GFP_NOFAIL;
3629	unsigned long addr = (unsigned long)area->addr;
3630	unsigned long size = get_vm_area_size(area);
3631	unsigned long array_size;
3632	unsigned int nr_small_pages = size >> PAGE_SHIFT;
3633	unsigned int page_order;
3634	unsigned int flags;
3635	int ret;
3636
3637	array_size = (unsigned long)nr_small_pages * sizeof(struct page *);
3638
3639	if (!(gfp_mask & (GFP_DMA | GFP_DMA32)))
3640		gfp_mask |= __GFP_HIGHMEM;
3641
3642	/* Please note that the recursion is strictly bounded. */
3643	if (array_size > PAGE_SIZE) {
3644		area->pages = __vmalloc_node_noprof(array_size, 1, nested_gfp, node,
3645					area->caller);
3646	} else {
3647		area->pages = kmalloc_node_noprof(array_size, nested_gfp, node);
3648	}
3649
3650	if (!area->pages) {
3651		warn_alloc(gfp_mask, NULL,
3652			"vmalloc error: size %lu, failed to allocated page array size %lu",
3653			nr_small_pages * PAGE_SIZE, array_size);
3654		free_vm_area(area);
3655		return NULL;
3656	}
3657
3658	set_vm_area_page_order(area, page_shift - PAGE_SHIFT);
3659	page_order = vm_area_page_order(area);
3660
3661	/*
3662	 * High-order nofail allocations are really expensive and
3663	 * potentially dangerous (pre-mature OOM, disruptive reclaim
3664	 * and compaction etc.
3665	 *
3666	 * Please note, the __vmalloc_node_range_noprof() falls-back
3667	 * to order-0 pages if high-order attempt is unsuccessful.
3668	 */
3669	area->nr_pages = vm_area_alloc_pages((page_order ?
3670		gfp_mask & ~__GFP_NOFAIL : gfp_mask) | __GFP_NOWARN,
3671		node, page_order, nr_small_pages, area->pages);
3672
3673	atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
3674	if (gfp_mask & __GFP_ACCOUNT) {
3675		int i;
3676
3677		for (i = 0; i < area->nr_pages; i++)
3678			mod_memcg_page_state(area->pages[i], MEMCG_VMALLOC, 1);
3679	}
3680
3681	/*
3682	 * If not enough pages were obtained to accomplish an
3683	 * allocation request, free them via vfree() if any.
3684	 */
3685	if (area->nr_pages != nr_small_pages) {
3686		/*
3687		 * vm_area_alloc_pages() can fail due to insufficient memory but
3688		 * also:-
3689		 *
3690		 * - a pending fatal signal
3691		 * - insufficient huge page-order pages
3692		 *
3693		 * Since we always retry allocations at order-0 in the huge page
3694		 * case a warning for either is spurious.
3695		 */
3696		if (!fatal_signal_pending(current) && page_order == 0)
3697			warn_alloc(gfp_mask, NULL,
3698				"vmalloc error: size %lu, failed to allocate pages",
3699				area->nr_pages * PAGE_SIZE);
3700		goto fail;
3701	}
3702
3703	/*
3704	 * page tables allocations ignore external gfp mask, enforce it
3705	 * by the scope API
3706	 */
3707	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3708		flags = memalloc_nofs_save();
3709	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3710		flags = memalloc_noio_save();
3711
3712	do {
3713		ret = vmap_pages_range(addr, addr + size, prot, area->pages,
3714			page_shift);
3715		if (nofail && (ret < 0))
3716			schedule_timeout_uninterruptible(1);
3717	} while (nofail && (ret < 0));
3718
3719	if ((gfp_mask & (__GFP_FS | __GFP_IO)) == __GFP_IO)
3720		memalloc_nofs_restore(flags);
3721	else if ((gfp_mask & (__GFP_FS | __GFP_IO)) == 0)
3722		memalloc_noio_restore(flags);
3723
3724	if (ret < 0) {
3725		warn_alloc(gfp_mask, NULL,
3726			"vmalloc error: size %lu, failed to map pages",
3727			area->nr_pages * PAGE_SIZE);
3728		goto fail;
3729	}
3730
3731	return area->addr;
3732
3733fail:
3734	vfree(area->addr);
3735	return NULL;
3736}
3737
3738/**
3739 * __vmalloc_node_range - allocate virtually contiguous memory
3740 * @size:		  allocation size
3741 * @align:		  desired alignment
3742 * @start:		  vm area range start
3743 * @end:		  vm area range end
3744 * @gfp_mask:		  flags for the page level allocator
3745 * @prot:		  protection mask for the allocated pages
3746 * @vm_flags:		  additional vm area flags (e.g. %VM_NO_GUARD)
3747 * @node:		  node to use for allocation or NUMA_NO_NODE
3748 * @caller:		  caller's return address
3749 *
3750 * Allocate enough pages to cover @size from the page level
3751 * allocator with @gfp_mask flags. Please note that the full set of gfp
3752 * flags are not supported. GFP_KERNEL, GFP_NOFS and GFP_NOIO are all
3753 * supported.
3754 * Zone modifiers are not supported. From the reclaim modifiers
3755 * __GFP_DIRECT_RECLAIM is required (aka GFP_NOWAIT is not supported)
3756 * and only __GFP_NOFAIL is supported (i.e. __GFP_NORETRY and
3757 * __GFP_RETRY_MAYFAIL are not supported).
3758 *
3759 * __GFP_NOWARN can be used to suppress failures messages.
3760 *
3761 * Map them into contiguous kernel virtual space, using a pagetable
3762 * protection of @prot.
3763 *
3764 * Return: the address of the area or %NULL on failure
3765 */
3766void *__vmalloc_node_range_noprof(unsigned long size, unsigned long align,
3767			unsigned long start, unsigned long end, gfp_t gfp_mask,
3768			pgprot_t prot, unsigned long vm_flags, int node,
3769			const void *caller)
3770{
3771	struct vm_struct *area;
3772	void *ret;
3773	kasan_vmalloc_flags_t kasan_flags = KASAN_VMALLOC_NONE;
3774	unsigned long real_size = size;
3775	unsigned long real_align = align;
3776	unsigned int shift = PAGE_SHIFT;
3777
3778	if (WARN_ON_ONCE(!size))
3779		return NULL;
3780
3781	if ((size >> PAGE_SHIFT) > totalram_pages()) {
3782		warn_alloc(gfp_mask, NULL,
3783			"vmalloc error: size %lu, exceeds total pages",
3784			real_size);
3785		return NULL;
3786	}
3787
3788	if (vmap_allow_huge && (vm_flags & VM_ALLOW_HUGE_VMAP)) {
 
 
3789		/*
3790		 * Try huge pages. Only try for PAGE_KERNEL allocations,
3791		 * others like modules don't yet expect huge pages in
3792		 * their allocations due to apply_to_page_range not
3793		 * supporting them.
3794		 */
3795
3796		if (arch_vmap_pmd_supported(prot) && size >= PMD_SIZE)
 
 
 
3797			shift = PMD_SHIFT;
3798		else
3799			shift = arch_vmap_pte_supported_shift(size);
3800
3801		align = max(real_align, 1UL << shift);
3802		size = ALIGN(real_size, 1UL << shift);
3803	}
3804
3805again:
3806	area = __get_vm_area_node(real_size, align, shift, VM_ALLOC |
3807				  VM_UNINITIALIZED | vm_flags, start, end, node,
3808				  gfp_mask, caller);
3809	if (!area) {
3810		bool nofail = gfp_mask & __GFP_NOFAIL;
3811		warn_alloc(gfp_mask, NULL,
3812			"vmalloc error: size %lu, vm_struct allocation failed%s",
3813			real_size, (nofail) ? ". Retrying." : "");
3814		if (nofail) {
3815			schedule_timeout_uninterruptible(1);
3816			goto again;
3817		}
3818		goto fail;
3819	}
3820
3821	/*
3822	 * Prepare arguments for __vmalloc_area_node() and
3823	 * kasan_unpoison_vmalloc().
3824	 */
3825	if (pgprot_val(prot) == pgprot_val(PAGE_KERNEL)) {
3826		if (kasan_hw_tags_enabled()) {
3827			/*
3828			 * Modify protection bits to allow tagging.
3829			 * This must be done before mapping.
3830			 */
3831			prot = arch_vmap_pgprot_tagged(prot);
3832
3833			/*
3834			 * Skip page_alloc poisoning and zeroing for physical
3835			 * pages backing VM_ALLOC mapping. Memory is instead
3836			 * poisoned and zeroed by kasan_unpoison_vmalloc().
3837			 */
3838			gfp_mask |= __GFP_SKIP_KASAN | __GFP_SKIP_ZERO;
3839		}
3840
3841		/* Take note that the mapping is PAGE_KERNEL. */
3842		kasan_flags |= KASAN_VMALLOC_PROT_NORMAL;
3843	}
3844
3845	/* Allocate physical pages and map them into vmalloc space. */
3846	ret = __vmalloc_area_node(area, gfp_mask, prot, shift, node);
3847	if (!ret)
3848		goto fail;
3849
3850	/*
3851	 * Mark the pages as accessible, now that they are mapped.
3852	 * The condition for setting KASAN_VMALLOC_INIT should complement the
3853	 * one in post_alloc_hook() with regards to the __GFP_SKIP_ZERO check
3854	 * to make sure that memory is initialized under the same conditions.
3855	 * Tag-based KASAN modes only assign tags to normal non-executable
3856	 * allocations, see __kasan_unpoison_vmalloc().
3857	 */
3858	kasan_flags |= KASAN_VMALLOC_VM_ALLOC;
3859	if (!want_init_on_free() && want_init_on_alloc(gfp_mask) &&
3860	    (gfp_mask & __GFP_SKIP_ZERO))
3861		kasan_flags |= KASAN_VMALLOC_INIT;
3862	/* KASAN_VMALLOC_PROT_NORMAL already set if required. */
3863	area->addr = kasan_unpoison_vmalloc(area->addr, real_size, kasan_flags);
3864
3865	/*
3866	 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
3867	 * flag. It means that vm_struct is not fully initialized.
3868	 * Now, it is fully initialized, so remove this flag here.
3869	 */
3870	clear_vm_uninitialized_flag(area);
3871
3872	size = PAGE_ALIGN(size);
3873	if (!(vm_flags & VM_DEFER_KMEMLEAK))
3874		kmemleak_vmalloc(area, size, gfp_mask);
3875
3876	return area->addr;
3877
3878fail:
3879	if (shift > PAGE_SHIFT) {
3880		shift = PAGE_SHIFT;
3881		align = real_align;
3882		size = real_size;
3883		goto again;
3884	}
3885
3886	return NULL;
3887}
3888
3889/**
3890 * __vmalloc_node - allocate virtually contiguous memory
3891 * @size:	    allocation size
3892 * @align:	    desired alignment
3893 * @gfp_mask:	    flags for the page level allocator
3894 * @node:	    node to use for allocation or NUMA_NO_NODE
3895 * @caller:	    caller's return address
3896 *
3897 * Allocate enough pages to cover @size from the page level allocator with
3898 * @gfp_mask flags.  Map them into contiguous kernel virtual space.
3899 *
3900 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
3901 * and __GFP_NOFAIL are not supported
3902 *
3903 * Any use of gfp flags outside of GFP_KERNEL should be consulted
3904 * with mm people.
3905 *
3906 * Return: pointer to the allocated memory or %NULL on error
3907 */
3908void *__vmalloc_node_noprof(unsigned long size, unsigned long align,
3909			    gfp_t gfp_mask, int node, const void *caller)
3910{
3911	return __vmalloc_node_range_noprof(size, align, VMALLOC_START, VMALLOC_END,
3912				gfp_mask, PAGE_KERNEL, 0, node, caller);
3913}
3914/*
3915 * This is only for performance analysis of vmalloc and stress purpose.
3916 * It is required by vmalloc test module, therefore do not use it other
3917 * than that.
3918 */
3919#ifdef CONFIG_TEST_VMALLOC_MODULE
3920EXPORT_SYMBOL_GPL(__vmalloc_node_noprof);
3921#endif
3922
3923void *__vmalloc_noprof(unsigned long size, gfp_t gfp_mask)
3924{
3925	return __vmalloc_node_noprof(size, 1, gfp_mask, NUMA_NO_NODE,
3926				__builtin_return_address(0));
3927}
3928EXPORT_SYMBOL(__vmalloc_noprof);
3929
3930/**
3931 * vmalloc - allocate virtually contiguous memory
3932 * @size:    allocation size
3933 *
3934 * Allocate enough pages to cover @size from the page level
3935 * allocator and map them into contiguous kernel virtual space.
3936 *
3937 * For tight control over page level allocator and protection flags
3938 * use __vmalloc() instead.
3939 *
3940 * Return: pointer to the allocated memory or %NULL on error
3941 */
3942void *vmalloc_noprof(unsigned long size)
3943{
3944	return __vmalloc_node_noprof(size, 1, GFP_KERNEL, NUMA_NO_NODE,
3945				__builtin_return_address(0));
3946}
3947EXPORT_SYMBOL(vmalloc_noprof);
3948
3949/**
3950 * vmalloc_huge - allocate virtually contiguous memory, allow huge pages
3951 * @size:      allocation size
3952 * @gfp_mask:  flags for the page level allocator
3953 *
3954 * Allocate enough pages to cover @size from the page level
3955 * allocator and map them into contiguous kernel virtual space.
3956 * If @size is greater than or equal to PMD_SIZE, allow using
3957 * huge pages for the memory
3958 *
3959 * Return: pointer to the allocated memory or %NULL on error
3960 */
3961void *vmalloc_huge_noprof(unsigned long size, gfp_t gfp_mask)
3962{
3963	return __vmalloc_node_range_noprof(size, 1, VMALLOC_START, VMALLOC_END,
3964				    gfp_mask, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
3965				    NUMA_NO_NODE, __builtin_return_address(0));
3966}
3967EXPORT_SYMBOL_GPL(vmalloc_huge_noprof);
3968
3969/**
3970 * vzalloc - allocate virtually contiguous memory with zero fill
3971 * @size:    allocation size
3972 *
3973 * Allocate enough pages to cover @size from the page level
3974 * allocator and map them into contiguous kernel virtual space.
3975 * The memory allocated is set to zero.
3976 *
3977 * For tight control over page level allocator and protection flags
3978 * use __vmalloc() instead.
3979 *
3980 * Return: pointer to the allocated memory or %NULL on error
3981 */
3982void *vzalloc_noprof(unsigned long size)
3983{
3984	return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
3985				__builtin_return_address(0));
3986}
3987EXPORT_SYMBOL(vzalloc_noprof);
3988
3989/**
3990 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
3991 * @size: allocation size
3992 *
3993 * The resulting memory area is zeroed so it can be mapped to userspace
3994 * without leaking data.
3995 *
3996 * Return: pointer to the allocated memory or %NULL on error
3997 */
3998void *vmalloc_user_noprof(unsigned long size)
3999{
4000	return __vmalloc_node_range_noprof(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
4001				    GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
4002				    VM_USERMAP, NUMA_NO_NODE,
4003				    __builtin_return_address(0));
4004}
4005EXPORT_SYMBOL(vmalloc_user_noprof);
4006
4007/**
4008 * vmalloc_node - allocate memory on a specific node
4009 * @size:	  allocation size
4010 * @node:	  numa node
4011 *
4012 * Allocate enough pages to cover @size from the page level
4013 * allocator and map them into contiguous kernel virtual space.
4014 *
4015 * For tight control over page level allocator and protection flags
4016 * use __vmalloc() instead.
4017 *
4018 * Return: pointer to the allocated memory or %NULL on error
4019 */
4020void *vmalloc_node_noprof(unsigned long size, int node)
4021{
4022	return __vmalloc_node_noprof(size, 1, GFP_KERNEL, node,
4023			__builtin_return_address(0));
4024}
4025EXPORT_SYMBOL(vmalloc_node_noprof);
4026
4027/**
4028 * vzalloc_node - allocate memory on a specific node with zero fill
4029 * @size:	allocation size
4030 * @node:	numa node
4031 *
4032 * Allocate enough pages to cover @size from the page level
4033 * allocator and map them into contiguous kernel virtual space.
4034 * The memory allocated is set to zero.
4035 *
4036 * Return: pointer to the allocated memory or %NULL on error
4037 */
4038void *vzalloc_node_noprof(unsigned long size, int node)
4039{
4040	return __vmalloc_node_noprof(size, 1, GFP_KERNEL | __GFP_ZERO, node,
4041				__builtin_return_address(0));
4042}
4043EXPORT_SYMBOL(vzalloc_node_noprof);
4044
4045/**
4046 * vrealloc - reallocate virtually contiguous memory; contents remain unchanged
4047 * @p: object to reallocate memory for
4048 * @size: the size to reallocate
4049 * @flags: the flags for the page level allocator
4050 *
4051 * If @p is %NULL, vrealloc() behaves exactly like vmalloc(). If @size is 0 and
4052 * @p is not a %NULL pointer, the object pointed to is freed.
4053 *
4054 * If __GFP_ZERO logic is requested, callers must ensure that, starting with the
4055 * initial memory allocation, every subsequent call to this API for the same
4056 * memory allocation is flagged with __GFP_ZERO. Otherwise, it is possible that
4057 * __GFP_ZERO is not fully honored by this API.
4058 *
4059 * In any case, the contents of the object pointed to are preserved up to the
4060 * lesser of the new and old sizes.
4061 *
4062 * This function must not be called concurrently with itself or vfree() for the
4063 * same memory allocation.
4064 *
4065 * Return: pointer to the allocated memory; %NULL if @size is zero or in case of
4066 *         failure
4067 */
4068void *vrealloc_noprof(const void *p, size_t size, gfp_t flags)
4069{
4070	size_t old_size = 0;
4071	void *n;
4072
4073	if (!size) {
4074		vfree(p);
4075		return NULL;
4076	}
4077
4078	if (p) {
4079		struct vm_struct *vm;
4080
4081		vm = find_vm_area(p);
4082		if (unlikely(!vm)) {
4083			WARN(1, "Trying to vrealloc() nonexistent vm area (%p)\n", p);
4084			return NULL;
4085		}
4086
4087		old_size = get_vm_area_size(vm);
4088	}
4089
4090	/*
4091	 * TODO: Shrink the vm_area, i.e. unmap and free unused pages. What
4092	 * would be a good heuristic for when to shrink the vm_area?
4093	 */
4094	if (size <= old_size) {
4095		/* Zero out spare memory. */
4096		if (want_init_on_alloc(flags))
4097			memset((void *)p + size, 0, old_size - size);
4098		kasan_poison_vmalloc(p + size, old_size - size);
4099		kasan_unpoison_vmalloc(p, size, KASAN_VMALLOC_PROT_NORMAL);
4100		return (void *)p;
4101	}
4102
4103	/* TODO: Grow the vm_area, i.e. allocate and map additional pages. */
4104	n = __vmalloc_noprof(size, flags);
4105	if (!n)
4106		return NULL;
4107
4108	if (p) {
4109		memcpy(n, p, old_size);
4110		vfree(p);
4111	}
4112
4113	return n;
4114}
4115
4116#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
4117#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4118#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
4119#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
4120#else
4121/*
4122 * 64b systems should always have either DMA or DMA32 zones. For others
4123 * GFP_DMA32 should do the right thing and use the normal zone.
4124 */
4125#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
4126#endif
4127
4128/**
4129 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
4130 * @size:	allocation size
4131 *
4132 * Allocate enough 32bit PA addressable pages to cover @size from the
4133 * page level allocator and map them into contiguous kernel virtual space.
4134 *
4135 * Return: pointer to the allocated memory or %NULL on error
4136 */
4137void *vmalloc_32_noprof(unsigned long size)
4138{
4139	return __vmalloc_node_noprof(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
4140			__builtin_return_address(0));
4141}
4142EXPORT_SYMBOL(vmalloc_32_noprof);
4143
4144/**
4145 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
4146 * @size:	     allocation size
4147 *
4148 * The resulting memory area is 32bit addressable and zeroed so it can be
4149 * mapped to userspace without leaking data.
4150 *
4151 * Return: pointer to the allocated memory or %NULL on error
4152 */
4153void *vmalloc_32_user_noprof(unsigned long size)
4154{
4155	return __vmalloc_node_range_noprof(size, SHMLBA,  VMALLOC_START, VMALLOC_END,
4156				    GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
4157				    VM_USERMAP, NUMA_NO_NODE,
4158				    __builtin_return_address(0));
4159}
4160EXPORT_SYMBOL(vmalloc_32_user_noprof);
4161
4162/*
4163 * Atomically zero bytes in the iterator.
4164 *
4165 * Returns the number of zeroed bytes.
4166 */
4167static size_t zero_iter(struct iov_iter *iter, size_t count)
4168{
4169	size_t remains = count;
4170
4171	while (remains > 0) {
4172		size_t num, copied;
4173
4174		num = min_t(size_t, remains, PAGE_SIZE);
4175		copied = copy_page_to_iter_nofault(ZERO_PAGE(0), 0, num, iter);
4176		remains -= copied;
4177
4178		if (copied < num)
4179			break;
4180	}
4181
4182	return count - remains;
4183}
4184
4185/*
4186 * small helper routine, copy contents to iter from addr.
4187 * If the page is not present, fill zero.
4188 *
4189 * Returns the number of copied bytes.
4190 */
4191static size_t aligned_vread_iter(struct iov_iter *iter,
4192				 const char *addr, size_t count)
4193{
4194	size_t remains = count;
4195	struct page *page;
4196
4197	while (remains > 0) {
4198		unsigned long offset, length;
4199		size_t copied = 0;
4200
4201		offset = offset_in_page(addr);
4202		length = PAGE_SIZE - offset;
4203		if (length > remains)
4204			length = remains;
4205		page = vmalloc_to_page(addr);
4206		/*
4207		 * To do safe access to this _mapped_ area, we need lock. But
4208		 * adding lock here means that we need to add overhead of
4209		 * vmalloc()/vfree() calls for this _debug_ interface, rarely
4210		 * used. Instead of that, we'll use an local mapping via
4211		 * copy_page_to_iter_nofault() and accept a small overhead in
4212		 * this access function.
4213		 */
4214		if (page)
4215			copied = copy_page_to_iter_nofault(page, offset,
4216							   length, iter);
4217		else
4218			copied = zero_iter(iter, length);
 
 
4219
4220		addr += copied;
4221		remains -= copied;
4222
4223		if (copied != length)
4224			break;
4225	}
4226
4227	return count - remains;
4228}
4229
4230/*
4231 * Read from a vm_map_ram region of memory.
4232 *
4233 * Returns the number of copied bytes.
4234 */
4235static size_t vmap_ram_vread_iter(struct iov_iter *iter, const char *addr,
4236				  size_t count, unsigned long flags)
4237{
4238	char *start;
4239	struct vmap_block *vb;
4240	struct xarray *xa;
4241	unsigned long offset;
4242	unsigned int rs, re;
4243	size_t remains, n;
4244
4245	/*
4246	 * If it's area created by vm_map_ram() interface directly, but
4247	 * not further subdividing and delegating management to vmap_block,
4248	 * handle it here.
4249	 */
4250	if (!(flags & VMAP_BLOCK))
4251		return aligned_vread_iter(iter, addr, count);
4252
4253	remains = count;
4254
4255	/*
4256	 * Area is split into regions and tracked with vmap_block, read out
4257	 * each region and zero fill the hole between regions.
4258	 */
4259	xa = addr_to_vb_xa((unsigned long) addr);
4260	vb = xa_load(xa, addr_to_vb_idx((unsigned long)addr));
4261	if (!vb)
4262		goto finished_zero;
4263
4264	spin_lock(&vb->lock);
4265	if (bitmap_empty(vb->used_map, VMAP_BBMAP_BITS)) {
4266		spin_unlock(&vb->lock);
4267		goto finished_zero;
4268	}
4269
4270	for_each_set_bitrange(rs, re, vb->used_map, VMAP_BBMAP_BITS) {
4271		size_t copied;
4272
4273		if (remains == 0)
4274			goto finished;
4275
4276		start = vmap_block_vaddr(vb->va->va_start, rs);
4277
4278		if (addr < start) {
4279			size_t to_zero = min_t(size_t, start - addr, remains);
4280			size_t zeroed = zero_iter(iter, to_zero);
4281
4282			addr += zeroed;
4283			remains -= zeroed;
4284
4285			if (remains == 0 || zeroed != to_zero)
4286				goto finished;
4287		}
4288
4289		/*it could start reading from the middle of used region*/
4290		offset = offset_in_page(addr);
4291		n = ((re - rs + 1) << PAGE_SHIFT) - offset;
4292		if (n > remains)
4293			n = remains;
4294
4295		copied = aligned_vread_iter(iter, start + offset, n);
4296
4297		addr += copied;
4298		remains -= copied;
4299
4300		if (copied != n)
4301			goto finished;
4302	}
4303
4304	spin_unlock(&vb->lock);
4305
4306finished_zero:
4307	/* zero-fill the left dirty or free regions */
4308	return count - remains + zero_iter(iter, remains);
4309finished:
4310	/* We couldn't copy/zero everything */
4311	spin_unlock(&vb->lock);
4312	return count - remains;
4313}
4314
4315/**
4316 * vread_iter() - read vmalloc area in a safe way to an iterator.
4317 * @iter:         the iterator to which data should be written.
4318 * @addr:         vm address.
4319 * @count:        number of bytes to be read.
4320 *
4321 * This function checks that addr is a valid vmalloc'ed area, and
4322 * copy data from that area to a given buffer. If the given memory range
4323 * of [addr...addr+count) includes some valid address, data is copied to
4324 * proper area of @buf. If there are memory holes, they'll be zero-filled.
4325 * IOREMAP area is treated as memory hole and no copy is done.
4326 *
4327 * If [addr...addr+count) doesn't includes any intersects with alive
4328 * vm_struct area, returns 0. @buf should be kernel's buffer.
4329 *
4330 * Note: In usual ops, vread() is never necessary because the caller
4331 * should know vmalloc() area is valid and can use memcpy().
4332 * This is for routines which have to access vmalloc area without
4333 * any information, as /proc/kcore.
4334 *
4335 * Return: number of bytes for which addr and buf should be increased
4336 * (same number as @count) or %0 if [addr...addr+count) doesn't
4337 * include any intersection with valid vmalloc area
4338 */
4339long vread_iter(struct iov_iter *iter, const char *addr, size_t count)
4340{
4341	struct vmap_node *vn;
4342	struct vmap_area *va;
4343	struct vm_struct *vm;
4344	char *vaddr;
4345	size_t n, size, flags, remains;
4346	unsigned long next;
4347
4348	addr = kasan_reset_tag(addr);
4349
4350	/* Don't allow overflow */
4351	if ((unsigned long) addr + count < count)
4352		count = -(unsigned long) addr;
4353
4354	remains = count;
4355
4356	vn = find_vmap_area_exceed_addr_lock((unsigned long) addr, &va);
4357	if (!vn)
4358		goto finished_zero;
4359
4360	/* no intersects with alive vmap_area */
4361	if ((unsigned long)addr + remains <= va->va_start)
4362		goto finished_zero;
4363
4364	do {
4365		size_t copied;
 
4366
4367		if (remains == 0)
4368			goto finished;
4369
4370		vm = va->vm;
4371		flags = va->flags & VMAP_FLAGS_MASK;
4372		/*
4373		 * VMAP_BLOCK indicates a sub-type of vm_map_ram area, need
4374		 * be set together with VMAP_RAM.
4375		 */
4376		WARN_ON(flags == VMAP_BLOCK);
4377
4378		if (!vm && !flags)
4379			goto next_va;
4380
4381		if (vm && (vm->flags & VM_UNINITIALIZED))
4382			goto next_va;
4383
4384		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4385		smp_rmb();
4386
4387		vaddr = (char *) va->va_start;
4388		size = vm ? get_vm_area_size(vm) : va_size(va);
4389
4390		if (addr >= vaddr + size)
4391			goto next_va;
4392
4393		if (addr < vaddr) {
4394			size_t to_zero = min_t(size_t, vaddr - addr, remains);
4395			size_t zeroed = zero_iter(iter, to_zero);
4396
4397			addr += zeroed;
4398			remains -= zeroed;
4399
4400			if (remains == 0 || zeroed != to_zero)
4401				goto finished;
 
 
 
 
4402		}
 
 
 
 
 
 
 
 
 
 
 
 
 
4403
4404		n = vaddr + size - addr;
4405		if (n > remains)
4406			n = remains;
4407
4408		if (flags & VMAP_RAM)
4409			copied = vmap_ram_vread_iter(iter, addr, n, flags);
4410		else if (!(vm && (vm->flags & (VM_IOREMAP | VM_SPARSE))))
4411			copied = aligned_vread_iter(iter, addr, n);
4412		else /* IOREMAP | SPARSE area is treated as memory hole */
4413			copied = zero_iter(iter, n);
4414
4415		addr += copied;
4416		remains -= copied;
4417
4418		if (copied != n)
4419			goto finished;
4420
4421	next_va:
4422		next = va->va_end;
4423		spin_unlock(&vn->busy.lock);
4424	} while ((vn = find_vmap_area_exceed_addr_lock(next, &va)));
4425
4426finished_zero:
4427	if (vn)
4428		spin_unlock(&vn->busy.lock);
4429
4430	/* zero-fill memory holes */
4431	return count - remains + zero_iter(iter, remains);
4432finished:
4433	/* Nothing remains, or We couldn't copy/zero everything. */
4434	if (vn)
4435		spin_unlock(&vn->busy.lock);
4436
4437	return count - remains;
4438}
4439
4440/**
4441 * remap_vmalloc_range_partial - map vmalloc pages to userspace
4442 * @vma:		vma to cover
4443 * @uaddr:		target user address to start at
4444 * @kaddr:		virtual address of vmalloc kernel memory
4445 * @pgoff:		offset from @kaddr to start at
4446 * @size:		size of map area
4447 *
4448 * Returns:	0 for success, -Exxx on failure
4449 *
4450 * This function checks that @kaddr is a valid vmalloc'ed area,
4451 * and that it is big enough to cover the range starting at
4452 * @uaddr in @vma. Will return failure if that criteria isn't
4453 * met.
4454 *
4455 * Similar to remap_pfn_range() (see mm/memory.c)
4456 */
4457int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
4458				void *kaddr, unsigned long pgoff,
4459				unsigned long size)
4460{
4461	struct vm_struct *area;
4462	unsigned long off;
4463	unsigned long end_index;
4464
4465	if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
4466		return -EINVAL;
4467
4468	size = PAGE_ALIGN(size);
4469
4470	if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
4471		return -EINVAL;
4472
4473	area = find_vm_area(kaddr);
4474	if (!area)
4475		return -EINVAL;
4476
4477	if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
4478		return -EINVAL;
4479
4480	if (check_add_overflow(size, off, &end_index) ||
4481	    end_index > get_vm_area_size(area))
4482		return -EINVAL;
4483	kaddr += off;
4484
4485	do {
4486		struct page *page = vmalloc_to_page(kaddr);
4487		int ret;
4488
4489		ret = vm_insert_page(vma, uaddr, page);
4490		if (ret)
4491			return ret;
4492
4493		uaddr += PAGE_SIZE;
4494		kaddr += PAGE_SIZE;
4495		size -= PAGE_SIZE;
4496	} while (size > 0);
4497
4498	vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP);
4499
4500	return 0;
4501}
4502
4503/**
4504 * remap_vmalloc_range - map vmalloc pages to userspace
4505 * @vma:		vma to cover (map full range of vma)
4506 * @addr:		vmalloc memory
4507 * @pgoff:		number of pages into addr before first page to map
4508 *
4509 * Returns:	0 for success, -Exxx on failure
4510 *
4511 * This function checks that addr is a valid vmalloc'ed area, and
4512 * that it is big enough to cover the vma. Will return failure if
4513 * that criteria isn't met.
4514 *
4515 * Similar to remap_pfn_range() (see mm/memory.c)
4516 */
4517int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
4518						unsigned long pgoff)
4519{
4520	return remap_vmalloc_range_partial(vma, vma->vm_start,
4521					   addr, pgoff,
4522					   vma->vm_end - vma->vm_start);
4523}
4524EXPORT_SYMBOL(remap_vmalloc_range);
4525
4526void free_vm_area(struct vm_struct *area)
4527{
4528	struct vm_struct *ret;
4529	ret = remove_vm_area(area->addr);
4530	BUG_ON(ret != area);
4531	kfree(area);
4532}
4533EXPORT_SYMBOL_GPL(free_vm_area);
4534
4535#ifdef CONFIG_SMP
4536static struct vmap_area *node_to_va(struct rb_node *n)
4537{
4538	return rb_entry_safe(n, struct vmap_area, rb_node);
4539}
4540
4541/**
4542 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
4543 * @addr: target address
4544 *
4545 * Returns: vmap_area if it is found. If there is no such area
4546 *   the first highest(reverse order) vmap_area is returned
4547 *   i.e. va->va_start < addr && va->va_end < addr or NULL
4548 *   if there are no any areas before @addr.
4549 */
4550static struct vmap_area *
4551pvm_find_va_enclose_addr(unsigned long addr)
4552{
4553	struct vmap_area *va, *tmp;
4554	struct rb_node *n;
4555
4556	n = free_vmap_area_root.rb_node;
4557	va = NULL;
4558
4559	while (n) {
4560		tmp = rb_entry(n, struct vmap_area, rb_node);
4561		if (tmp->va_start <= addr) {
4562			va = tmp;
4563			if (tmp->va_end >= addr)
4564				break;
4565
4566			n = n->rb_right;
4567		} else {
4568			n = n->rb_left;
4569		}
4570	}
4571
4572	return va;
4573}
4574
4575/**
4576 * pvm_determine_end_from_reverse - find the highest aligned address
4577 * of free block below VMALLOC_END
4578 * @va:
4579 *   in - the VA we start the search(reverse order);
4580 *   out - the VA with the highest aligned end address.
4581 * @align: alignment for required highest address
4582 *
4583 * Returns: determined end address within vmap_area
4584 */
4585static unsigned long
4586pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
4587{
4588	unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4589	unsigned long addr;
4590
4591	if (likely(*va)) {
4592		list_for_each_entry_from_reverse((*va),
4593				&free_vmap_area_list, list) {
4594			addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
4595			if ((*va)->va_start < addr)
4596				return addr;
4597		}
4598	}
4599
4600	return 0;
4601}
4602
4603/**
4604 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
4605 * @offsets: array containing offset of each area
4606 * @sizes: array containing size of each area
4607 * @nr_vms: the number of areas to allocate
4608 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
4609 *
4610 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
4611 *	    vm_structs on success, %NULL on failure
4612 *
4613 * Percpu allocator wants to use congruent vm areas so that it can
4614 * maintain the offsets among percpu areas.  This function allocates
4615 * congruent vmalloc areas for it with GFP_KERNEL.  These areas tend to
4616 * be scattered pretty far, distance between two areas easily going up
4617 * to gigabytes.  To avoid interacting with regular vmallocs, these
4618 * areas are allocated from top.
4619 *
4620 * Despite its complicated look, this allocator is rather simple. It
4621 * does everything top-down and scans free blocks from the end looking
4622 * for matching base. While scanning, if any of the areas do not fit the
4623 * base address is pulled down to fit the area. Scanning is repeated till
4624 * all the areas fit and then all necessary data structures are inserted
4625 * and the result is returned.
4626 */
4627struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
4628				     const size_t *sizes, int nr_vms,
4629				     size_t align)
4630{
4631	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
4632	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
4633	struct vmap_area **vas, *va;
4634	struct vm_struct **vms;
4635	int area, area2, last_area, term_area;
4636	unsigned long base, start, size, end, last_end, orig_start, orig_end;
4637	bool purged = false;
4638
4639	/* verify parameters and allocate data structures */
4640	BUG_ON(offset_in_page(align) || !is_power_of_2(align));
4641	for (last_area = 0, area = 0; area < nr_vms; area++) {
4642		start = offsets[area];
4643		end = start + sizes[area];
4644
4645		/* is everything aligned properly? */
4646		BUG_ON(!IS_ALIGNED(offsets[area], align));
4647		BUG_ON(!IS_ALIGNED(sizes[area], align));
4648
4649		/* detect the area with the highest address */
4650		if (start > offsets[last_area])
4651			last_area = area;
4652
4653		for (area2 = area + 1; area2 < nr_vms; area2++) {
4654			unsigned long start2 = offsets[area2];
4655			unsigned long end2 = start2 + sizes[area2];
4656
4657			BUG_ON(start2 < end && start < end2);
4658		}
4659	}
4660	last_end = offsets[last_area] + sizes[last_area];
4661
4662	if (vmalloc_end - vmalloc_start < last_end) {
4663		WARN_ON(true);
4664		return NULL;
4665	}
4666
4667	vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
4668	vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
4669	if (!vas || !vms)
4670		goto err_free2;
4671
4672	for (area = 0; area < nr_vms; area++) {
4673		vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
4674		vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
4675		if (!vas[area] || !vms[area])
4676			goto err_free;
4677	}
4678retry:
4679	spin_lock(&free_vmap_area_lock);
4680
4681	/* start scanning - we scan from the top, begin with the last area */
4682	area = term_area = last_area;
4683	start = offsets[area];
4684	end = start + sizes[area];
4685
4686	va = pvm_find_va_enclose_addr(vmalloc_end);
4687	base = pvm_determine_end_from_reverse(&va, align) - end;
4688
4689	while (true) {
4690		/*
4691		 * base might have underflowed, add last_end before
4692		 * comparing.
4693		 */
4694		if (base + last_end < vmalloc_start + last_end)
4695			goto overflow;
4696
4697		/*
4698		 * Fitting base has not been found.
4699		 */
4700		if (va == NULL)
4701			goto overflow;
4702
4703		/*
4704		 * If required width exceeds current VA block, move
4705		 * base downwards and then recheck.
4706		 */
4707		if (base + end > va->va_end) {
4708			base = pvm_determine_end_from_reverse(&va, align) - end;
4709			term_area = area;
4710			continue;
4711		}
4712
4713		/*
4714		 * If this VA does not fit, move base downwards and recheck.
4715		 */
4716		if (base + start < va->va_start) {
4717			va = node_to_va(rb_prev(&va->rb_node));
4718			base = pvm_determine_end_from_reverse(&va, align) - end;
4719			term_area = area;
4720			continue;
4721		}
4722
4723		/*
4724		 * This area fits, move on to the previous one.  If
4725		 * the previous one is the terminal one, we're done.
4726		 */
4727		area = (area + nr_vms - 1) % nr_vms;
4728		if (area == term_area)
4729			break;
4730
4731		start = offsets[area];
4732		end = start + sizes[area];
4733		va = pvm_find_va_enclose_addr(base + end);
4734	}
4735
4736	/* we've found a fitting base, insert all va's */
4737	for (area = 0; area < nr_vms; area++) {
4738		int ret;
4739
4740		start = base + offsets[area];
4741		size = sizes[area];
4742
4743		va = pvm_find_va_enclose_addr(start);
4744		if (WARN_ON_ONCE(va == NULL))
4745			/* It is a BUG(), but trigger recovery instead. */
4746			goto recovery;
4747
4748		ret = va_clip(&free_vmap_area_root,
4749			&free_vmap_area_list, va, start, size);
 
4750		if (WARN_ON_ONCE(unlikely(ret)))
4751			/* It is a BUG(), but trigger recovery instead. */
4752			goto recovery;
4753
4754		/* Allocated area. */
4755		va = vas[area];
4756		va->va_start = start;
4757		va->va_end = start + size;
4758	}
4759
4760	spin_unlock(&free_vmap_area_lock);
4761
4762	/* populate the kasan shadow space */
4763	for (area = 0; area < nr_vms; area++) {
4764		if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
4765			goto err_free_shadow;
4766	}
4767
4768	/* insert all vm's */
 
4769	for (area = 0; area < nr_vms; area++) {
4770		struct vmap_node *vn = addr_to_node(vas[area]->va_start);
4771
4772		spin_lock(&vn->busy.lock);
4773		insert_vmap_area(vas[area], &vn->busy.root, &vn->busy.head);
4774		setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
4775				 pcpu_get_vm_areas);
4776		spin_unlock(&vn->busy.lock);
4777	}
 
4778
4779	/*
4780	 * Mark allocated areas as accessible. Do it now as a best-effort
4781	 * approach, as they can be mapped outside of vmalloc code.
4782	 * With hardware tag-based KASAN, marking is skipped for
4783	 * non-VM_ALLOC mappings, see __kasan_unpoison_vmalloc().
4784	 */
4785	for (area = 0; area < nr_vms; area++)
4786		vms[area]->addr = kasan_unpoison_vmalloc(vms[area]->addr,
4787				vms[area]->size, KASAN_VMALLOC_PROT_NORMAL);
4788
4789	kfree(vas);
4790	return vms;
4791
4792recovery:
4793	/*
4794	 * Remove previously allocated areas. There is no
4795	 * need in removing these areas from the busy tree,
4796	 * because they are inserted only on the final step
4797	 * and when pcpu_get_vm_areas() is success.
4798	 */
4799	while (area--) {
4800		orig_start = vas[area]->va_start;
4801		orig_end = vas[area]->va_end;
4802		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4803				&free_vmap_area_list);
4804		if (va)
4805			kasan_release_vmalloc(orig_start, orig_end,
4806				va->va_start, va->va_end,
4807				KASAN_VMALLOC_PAGE_RANGE | KASAN_VMALLOC_TLB_FLUSH);
4808		vas[area] = NULL;
4809	}
4810
4811overflow:
4812	spin_unlock(&free_vmap_area_lock);
4813	if (!purged) {
4814		reclaim_and_purge_vmap_areas();
4815		purged = true;
4816
4817		/* Before "retry", check if we recover. */
4818		for (area = 0; area < nr_vms; area++) {
4819			if (vas[area])
4820				continue;
4821
4822			vas[area] = kmem_cache_zalloc(
4823				vmap_area_cachep, GFP_KERNEL);
4824			if (!vas[area])
4825				goto err_free;
4826		}
4827
4828		goto retry;
4829	}
4830
4831err_free:
4832	for (area = 0; area < nr_vms; area++) {
4833		if (vas[area])
4834			kmem_cache_free(vmap_area_cachep, vas[area]);
4835
4836		kfree(vms[area]);
4837	}
4838err_free2:
4839	kfree(vas);
4840	kfree(vms);
4841	return NULL;
4842
4843err_free_shadow:
4844	spin_lock(&free_vmap_area_lock);
4845	/*
4846	 * We release all the vmalloc shadows, even the ones for regions that
4847	 * hadn't been successfully added. This relies on kasan_release_vmalloc
4848	 * being able to tolerate this case.
4849	 */
4850	for (area = 0; area < nr_vms; area++) {
4851		orig_start = vas[area]->va_start;
4852		orig_end = vas[area]->va_end;
4853		va = merge_or_add_vmap_area_augment(vas[area], &free_vmap_area_root,
4854				&free_vmap_area_list);
4855		if (va)
4856			kasan_release_vmalloc(orig_start, orig_end,
4857				va->va_start, va->va_end,
4858				KASAN_VMALLOC_PAGE_RANGE | KASAN_VMALLOC_TLB_FLUSH);
4859		vas[area] = NULL;
4860		kfree(vms[area]);
4861	}
4862	spin_unlock(&free_vmap_area_lock);
4863	kfree(vas);
4864	kfree(vms);
4865	return NULL;
4866}
4867
4868/**
4869 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
4870 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
4871 * @nr_vms: the number of allocated areas
4872 *
4873 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
4874 */
4875void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
4876{
4877	int i;
4878
4879	for (i = 0; i < nr_vms; i++)
4880		free_vm_area(vms[i]);
4881	kfree(vms);
4882}
4883#endif	/* CONFIG_SMP */
4884
4885#ifdef CONFIG_PRINTK
4886bool vmalloc_dump_obj(void *object)
4887{
4888	const void *caller;
4889	struct vm_struct *vm;
4890	struct vmap_area *va;
4891	struct vmap_node *vn;
4892	unsigned long addr;
4893	unsigned int nr_pages;
4894
4895	addr = PAGE_ALIGN((unsigned long) object);
4896	vn = addr_to_node(addr);
4897
4898	if (!spin_trylock(&vn->busy.lock))
 
4899		return false;
 
 
 
 
 
4900
4901	va = __find_vmap_area(addr, &vn->busy.root);
4902	if (!va || !va->vm) {
4903		spin_unlock(&vn->busy.lock);
4904		return false;
4905	}
 
 
4906
4907	vm = va->vm;
4908	addr = (unsigned long) vm->addr;
4909	caller = vm->caller;
4910	nr_pages = vm->nr_pages;
4911	spin_unlock(&vn->busy.lock);
4912
4913	pr_cont(" %u-page vmalloc region starting at %#lx allocated at %pS\n",
4914		nr_pages, addr, caller);
 
 
4915
4916	return true;
 
 
 
 
 
4917}
4918#endif
4919
4920#ifdef CONFIG_PROC_FS
4921static void show_numa_info(struct seq_file *m, struct vm_struct *v)
4922{
4923	if (IS_ENABLED(CONFIG_NUMA)) {
4924		unsigned int nr, *counters = m->private;
4925		unsigned int step = 1U << vm_area_page_order(v);
4926
4927		if (!counters)
4928			return;
4929
4930		if (v->flags & VM_UNINITIALIZED)
4931			return;
4932		/* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
4933		smp_rmb();
4934
4935		memset(counters, 0, nr_node_ids * sizeof(unsigned int));
4936
4937		for (nr = 0; nr < v->nr_pages; nr += step)
4938			counters[page_to_nid(v->pages[nr])] += step;
4939		for_each_node_state(nr, N_HIGH_MEMORY)
4940			if (counters[nr])
4941				seq_printf(m, " N%u=%u", nr, counters[nr]);
4942	}
4943}
4944
4945static void show_purge_info(struct seq_file *m)
4946{
4947	struct vmap_node *vn;
4948	struct vmap_area *va;
4949	int i;
4950
4951	for (i = 0; i < nr_vmap_nodes; i++) {
4952		vn = &vmap_nodes[i];
4953
4954		spin_lock(&vn->lazy.lock);
4955		list_for_each_entry(va, &vn->lazy.head, list) {
4956			seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
4957				(void *)va->va_start, (void *)va->va_end,
4958				va_size(va));
4959		}
4960		spin_unlock(&vn->lazy.lock);
4961	}
 
4962}
4963
4964static int vmalloc_info_show(struct seq_file *m, void *p)
4965{
4966	struct vmap_node *vn;
4967	struct vmap_area *va;
4968	struct vm_struct *v;
4969	int i;
4970
4971	for (i = 0; i < nr_vmap_nodes; i++) {
4972		vn = &vmap_nodes[i];
4973
4974		spin_lock(&vn->busy.lock);
4975		list_for_each_entry(va, &vn->busy.head, list) {
4976			if (!va->vm) {
4977				if (va->flags & VMAP_RAM)
4978					seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
4979						(void *)va->va_start, (void *)va->va_end,
4980						va_size(va));
 
4981
4982				continue;
4983			}
4984
4985			v = va->vm;
4986
4987			seq_printf(m, "0x%pK-0x%pK %7ld",
4988				v->addr, v->addr + v->size, v->size);
4989
4990			if (v->caller)
4991				seq_printf(m, " %pS", v->caller);
4992
4993			if (v->nr_pages)
4994				seq_printf(m, " pages=%d", v->nr_pages);
4995
4996			if (v->phys_addr)
4997				seq_printf(m, " phys=%pa", &v->phys_addr);
4998
4999			if (v->flags & VM_IOREMAP)
5000				seq_puts(m, " ioremap");
5001
5002			if (v->flags & VM_SPARSE)
5003				seq_puts(m, " sparse");
5004
5005			if (v->flags & VM_ALLOC)
5006				seq_puts(m, " vmalloc");
5007
5008			if (v->flags & VM_MAP)
5009				seq_puts(m, " vmap");
5010
5011			if (v->flags & VM_USERMAP)
5012				seq_puts(m, " user");
5013
5014			if (v->flags & VM_DMA_COHERENT)
5015				seq_puts(m, " dma-coherent");
5016
5017			if (is_vmalloc_addr(v->pages))
5018				seq_puts(m, " vpages");
5019
5020			show_numa_info(m, v);
5021			seq_putc(m, '\n');
5022		}
5023		spin_unlock(&vn->busy.lock);
5024	}
5025
5026	/*
5027	 * As a final step, dump "unpurged" areas.
5028	 */
5029	show_purge_info(m);
 
 
 
5030	return 0;
5031}
5032
 
 
 
 
 
 
 
5033static int __init proc_vmalloc_init(void)
5034{
5035	void *priv_data = NULL;
5036
5037	if (IS_ENABLED(CONFIG_NUMA))
5038		priv_data = kmalloc(nr_node_ids * sizeof(unsigned int), GFP_KERNEL);
5039
5040	proc_create_single_data("vmallocinfo",
5041		0400, NULL, vmalloc_info_show, priv_data);
5042
5043	return 0;
5044}
5045module_init(proc_vmalloc_init);
5046
5047#endif
5048
5049static void __init vmap_init_free_space(void)
5050{
5051	unsigned long vmap_start = 1;
5052	const unsigned long vmap_end = ULONG_MAX;
5053	struct vmap_area *free;
5054	struct vm_struct *busy;
5055
5056	/*
5057	 *     B     F     B     B     B     F
5058	 * -|-----|.....|-----|-----|-----|.....|-
5059	 *  |           The KVA space           |
5060	 *  |<--------------------------------->|
5061	 */
5062	for (busy = vmlist; busy; busy = busy->next) {
5063		if ((unsigned long) busy->addr - vmap_start > 0) {
5064			free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5065			if (!WARN_ON_ONCE(!free)) {
5066				free->va_start = vmap_start;
5067				free->va_end = (unsigned long) busy->addr;
5068
5069				insert_vmap_area_augment(free, NULL,
5070					&free_vmap_area_root,
5071						&free_vmap_area_list);
5072			}
5073		}
5074
5075		vmap_start = (unsigned long) busy->addr + busy->size;
5076	}
5077
5078	if (vmap_end - vmap_start > 0) {
5079		free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5080		if (!WARN_ON_ONCE(!free)) {
5081			free->va_start = vmap_start;
5082			free->va_end = vmap_end;
5083
5084			insert_vmap_area_augment(free, NULL,
5085				&free_vmap_area_root,
5086					&free_vmap_area_list);
5087		}
5088	}
5089}
5090
5091static void vmap_init_nodes(void)
5092{
5093	struct vmap_node *vn;
5094	int i, n;
5095
5096#if BITS_PER_LONG == 64
5097	/*
5098	 * A high threshold of max nodes is fixed and bound to 128,
5099	 * thus a scale factor is 1 for systems where number of cores
5100	 * are less or equal to specified threshold.
5101	 *
5102	 * As for NUMA-aware notes. For bigger systems, for example
5103	 * NUMA with multi-sockets, where we can end-up with thousands
5104	 * of cores in total, a "sub-numa-clustering" should be added.
5105	 *
5106	 * In this case a NUMA domain is considered as a single entity
5107	 * with dedicated sub-nodes in it which describe one group or
5108	 * set of cores. Therefore a per-domain purging is supposed to
5109	 * be added as well as a per-domain balancing.
5110	 */
5111	n = clamp_t(unsigned int, num_possible_cpus(), 1, 128);
5112
5113	if (n > 1) {
5114		vn = kmalloc_array(n, sizeof(*vn), GFP_NOWAIT | __GFP_NOWARN);
5115		if (vn) {
5116			/* Node partition is 16 pages. */
5117			vmap_zone_size = (1 << 4) * PAGE_SIZE;
5118			nr_vmap_nodes = n;
5119			vmap_nodes = vn;
5120		} else {
5121			pr_err("Failed to allocate an array. Disable a node layer\n");
5122		}
5123	}
5124#endif
5125
5126	for (n = 0; n < nr_vmap_nodes; n++) {
5127		vn = &vmap_nodes[n];
5128		vn->busy.root = RB_ROOT;
5129		INIT_LIST_HEAD(&vn->busy.head);
5130		spin_lock_init(&vn->busy.lock);
5131
5132		vn->lazy.root = RB_ROOT;
5133		INIT_LIST_HEAD(&vn->lazy.head);
5134		spin_lock_init(&vn->lazy.lock);
5135
5136		for (i = 0; i < MAX_VA_SIZE_PAGES; i++) {
5137			INIT_LIST_HEAD(&vn->pool[i].head);
5138			WRITE_ONCE(vn->pool[i].len, 0);
5139		}
5140
5141		spin_lock_init(&vn->pool_lock);
5142	}
5143}
5144
5145static unsigned long
5146vmap_node_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
5147{
5148	unsigned long count;
5149	struct vmap_node *vn;
5150	int i, j;
5151
5152	for (count = 0, i = 0; i < nr_vmap_nodes; i++) {
5153		vn = &vmap_nodes[i];
5154
5155		for (j = 0; j < MAX_VA_SIZE_PAGES; j++)
5156			count += READ_ONCE(vn->pool[j].len);
5157	}
5158
5159	return count ? count : SHRINK_EMPTY;
5160}
5161
5162static unsigned long
5163vmap_node_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
5164{
5165	int i;
5166
5167	for (i = 0; i < nr_vmap_nodes; i++)
5168		decay_va_pool_node(&vmap_nodes[i], true);
5169
5170	return SHRINK_STOP;
5171}
5172
5173void __init vmalloc_init(void)
5174{
5175	struct shrinker *vmap_node_shrinker;
5176	struct vmap_area *va;
5177	struct vmap_node *vn;
5178	struct vm_struct *tmp;
5179	int i;
5180
5181	/*
5182	 * Create the cache for vmap_area objects.
5183	 */
5184	vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
5185
5186	for_each_possible_cpu(i) {
5187		struct vmap_block_queue *vbq;
5188		struct vfree_deferred *p;
5189
5190		vbq = &per_cpu(vmap_block_queue, i);
5191		spin_lock_init(&vbq->lock);
5192		INIT_LIST_HEAD(&vbq->free);
5193		p = &per_cpu(vfree_deferred, i);
5194		init_llist_head(&p->list);
5195		INIT_WORK(&p->wq, delayed_vfree_work);
5196		xa_init(&vbq->vmap_blocks);
5197	}
5198
5199	/*
5200	 * Setup nodes before importing vmlist.
5201	 */
5202	vmap_init_nodes();
5203
5204	/* Import existing vmlist entries. */
5205	for (tmp = vmlist; tmp; tmp = tmp->next) {
5206		va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
5207		if (WARN_ON_ONCE(!va))
5208			continue;
5209
5210		va->va_start = (unsigned long)tmp->addr;
5211		va->va_end = va->va_start + tmp->size;
5212		va->vm = tmp;
5213
5214		vn = addr_to_node(va->va_start);
5215		insert_vmap_area(va, &vn->busy.root, &vn->busy.head);
5216	}
5217
5218	/*
5219	 * Now we can initialize a free vmap space.
5220	 */
5221	vmap_init_free_space();
5222	vmap_initialized = true;
5223
5224	vmap_node_shrinker = shrinker_alloc(0, "vmap-node");
5225	if (!vmap_node_shrinker) {
5226		pr_err("Failed to allocate vmap-node shrinker!\n");
5227		return;
5228	}
5229
5230	vmap_node_shrinker->count_objects = vmap_node_shrink_count;
5231	vmap_node_shrinker->scan_objects = vmap_node_shrink_scan;
5232	shrinker_register(vmap_node_shrinker);
5233}