Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/vmalloc.c
4 *
5 * Copyright (C) 1993 Linus Torvalds
6 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
7 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
8 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
9 * Numa awareness, Christoph Lameter, SGI, June 2005
10 * Improving global KVA allocator, Uladzislau Rezki, Sony, May 2019
11 */
12
13#include <linux/vmalloc.h>
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/highmem.h>
17#include <linux/sched/signal.h>
18#include <linux/slab.h>
19#include <linux/spinlock.h>
20#include <linux/interrupt.h>
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/set_memory.h>
24#include <linux/debugobjects.h>
25#include <linux/kallsyms.h>
26#include <linux/list.h>
27#include <linux/notifier.h>
28#include <linux/rbtree.h>
29#include <linux/xarray.h>
30#include <linux/rcupdate.h>
31#include <linux/pfn.h>
32#include <linux/kmemleak.h>
33#include <linux/atomic.h>
34#include <linux/compiler.h>
35#include <linux/llist.h>
36#include <linux/bitops.h>
37#include <linux/rbtree_augmented.h>
38#include <linux/overflow.h>
39
40#include <linux/uaccess.h>
41#include <asm/tlbflush.h>
42#include <asm/shmparam.h>
43
44#include "internal.h"
45#include "pgalloc-track.h"
46
47bool is_vmalloc_addr(const void *x)
48{
49 unsigned long addr = (unsigned long)x;
50
51 return addr >= VMALLOC_START && addr < VMALLOC_END;
52}
53EXPORT_SYMBOL(is_vmalloc_addr);
54
55struct vfree_deferred {
56 struct llist_head list;
57 struct work_struct wq;
58};
59static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
60
61static void __vunmap(const void *, int);
62
63static void free_work(struct work_struct *w)
64{
65 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
66 struct llist_node *t, *llnode;
67
68 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
69 __vunmap((void *)llnode, 1);
70}
71
72/*** Page table manipulation functions ***/
73
74static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
75 pgtbl_mod_mask *mask)
76{
77 pte_t *pte;
78
79 pte = pte_offset_kernel(pmd, addr);
80 do {
81 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
82 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
83 } while (pte++, addr += PAGE_SIZE, addr != end);
84 *mask |= PGTBL_PTE_MODIFIED;
85}
86
87static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
88 pgtbl_mod_mask *mask)
89{
90 pmd_t *pmd;
91 unsigned long next;
92 int cleared;
93
94 pmd = pmd_offset(pud, addr);
95 do {
96 next = pmd_addr_end(addr, end);
97
98 cleared = pmd_clear_huge(pmd);
99 if (cleared || pmd_bad(*pmd))
100 *mask |= PGTBL_PMD_MODIFIED;
101
102 if (cleared)
103 continue;
104 if (pmd_none_or_clear_bad(pmd))
105 continue;
106 vunmap_pte_range(pmd, addr, next, mask);
107
108 cond_resched();
109 } while (pmd++, addr = next, addr != end);
110}
111
112static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
113 pgtbl_mod_mask *mask)
114{
115 pud_t *pud;
116 unsigned long next;
117 int cleared;
118
119 pud = pud_offset(p4d, addr);
120 do {
121 next = pud_addr_end(addr, end);
122
123 cleared = pud_clear_huge(pud);
124 if (cleared || pud_bad(*pud))
125 *mask |= PGTBL_PUD_MODIFIED;
126
127 if (cleared)
128 continue;
129 if (pud_none_or_clear_bad(pud))
130 continue;
131 vunmap_pmd_range(pud, addr, next, mask);
132 } while (pud++, addr = next, addr != end);
133}
134
135static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
136 pgtbl_mod_mask *mask)
137{
138 p4d_t *p4d;
139 unsigned long next;
140 int cleared;
141
142 p4d = p4d_offset(pgd, addr);
143 do {
144 next = p4d_addr_end(addr, end);
145
146 cleared = p4d_clear_huge(p4d);
147 if (cleared || p4d_bad(*p4d))
148 *mask |= PGTBL_P4D_MODIFIED;
149
150 if (cleared)
151 continue;
152 if (p4d_none_or_clear_bad(p4d))
153 continue;
154 vunmap_pud_range(p4d, addr, next, mask);
155 } while (p4d++, addr = next, addr != end);
156}
157
158/**
159 * unmap_kernel_range_noflush - unmap kernel VM area
160 * @start: start of the VM area to unmap
161 * @size: size of the VM area to unmap
162 *
163 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size specify
164 * should have been allocated using get_vm_area() and its friends.
165 *
166 * NOTE:
167 * This function does NOT do any cache flushing. The caller is responsible
168 * for calling flush_cache_vunmap() on to-be-mapped areas before calling this
169 * function and flush_tlb_kernel_range() after.
170 */
171void unmap_kernel_range_noflush(unsigned long start, unsigned long size)
172{
173 unsigned long end = start + size;
174 unsigned long next;
175 pgd_t *pgd;
176 unsigned long addr = start;
177 pgtbl_mod_mask mask = 0;
178
179 BUG_ON(addr >= end);
180 pgd = pgd_offset_k(addr);
181 do {
182 next = pgd_addr_end(addr, end);
183 if (pgd_bad(*pgd))
184 mask |= PGTBL_PGD_MODIFIED;
185 if (pgd_none_or_clear_bad(pgd))
186 continue;
187 vunmap_p4d_range(pgd, addr, next, &mask);
188 } while (pgd++, addr = next, addr != end);
189
190 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
191 arch_sync_kernel_mappings(start, end);
192}
193
194static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
195 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
196 pgtbl_mod_mask *mask)
197{
198 pte_t *pte;
199
200 /*
201 * nr is a running index into the array which helps higher level
202 * callers keep track of where we're up to.
203 */
204
205 pte = pte_alloc_kernel_track(pmd, addr, mask);
206 if (!pte)
207 return -ENOMEM;
208 do {
209 struct page *page = pages[*nr];
210
211 if (WARN_ON(!pte_none(*pte)))
212 return -EBUSY;
213 if (WARN_ON(!page))
214 return -ENOMEM;
215 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
216 (*nr)++;
217 } while (pte++, addr += PAGE_SIZE, addr != end);
218 *mask |= PGTBL_PTE_MODIFIED;
219 return 0;
220}
221
222static int vmap_pmd_range(pud_t *pud, unsigned long addr,
223 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
224 pgtbl_mod_mask *mask)
225{
226 pmd_t *pmd;
227 unsigned long next;
228
229 pmd = pmd_alloc_track(&init_mm, pud, addr, mask);
230 if (!pmd)
231 return -ENOMEM;
232 do {
233 next = pmd_addr_end(addr, end);
234 if (vmap_pte_range(pmd, addr, next, prot, pages, nr, mask))
235 return -ENOMEM;
236 } while (pmd++, addr = next, addr != end);
237 return 0;
238}
239
240static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
241 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
242 pgtbl_mod_mask *mask)
243{
244 pud_t *pud;
245 unsigned long next;
246
247 pud = pud_alloc_track(&init_mm, p4d, addr, mask);
248 if (!pud)
249 return -ENOMEM;
250 do {
251 next = pud_addr_end(addr, end);
252 if (vmap_pmd_range(pud, addr, next, prot, pages, nr, mask))
253 return -ENOMEM;
254 } while (pud++, addr = next, addr != end);
255 return 0;
256}
257
258static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
259 unsigned long end, pgprot_t prot, struct page **pages, int *nr,
260 pgtbl_mod_mask *mask)
261{
262 p4d_t *p4d;
263 unsigned long next;
264
265 p4d = p4d_alloc_track(&init_mm, pgd, addr, mask);
266 if (!p4d)
267 return -ENOMEM;
268 do {
269 next = p4d_addr_end(addr, end);
270 if (vmap_pud_range(p4d, addr, next, prot, pages, nr, mask))
271 return -ENOMEM;
272 } while (p4d++, addr = next, addr != end);
273 return 0;
274}
275
276/**
277 * map_kernel_range_noflush - map kernel VM area with the specified pages
278 * @addr: start of the VM area to map
279 * @size: size of the VM area to map
280 * @prot: page protection flags to use
281 * @pages: pages to map
282 *
283 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size specify should
284 * have been allocated using get_vm_area() and its friends.
285 *
286 * NOTE:
287 * This function does NOT do any cache flushing. The caller is responsible for
288 * calling flush_cache_vmap() on to-be-mapped areas before calling this
289 * function.
290 *
291 * RETURNS:
292 * 0 on success, -errno on failure.
293 */
294int map_kernel_range_noflush(unsigned long addr, unsigned long size,
295 pgprot_t prot, struct page **pages)
296{
297 unsigned long start = addr;
298 unsigned long end = addr + size;
299 unsigned long next;
300 pgd_t *pgd;
301 int err = 0;
302 int nr = 0;
303 pgtbl_mod_mask mask = 0;
304
305 BUG_ON(addr >= end);
306 pgd = pgd_offset_k(addr);
307 do {
308 next = pgd_addr_end(addr, end);
309 if (pgd_bad(*pgd))
310 mask |= PGTBL_PGD_MODIFIED;
311 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr, &mask);
312 if (err)
313 return err;
314 } while (pgd++, addr = next, addr != end);
315
316 if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
317 arch_sync_kernel_mappings(start, end);
318
319 return 0;
320}
321
322int map_kernel_range(unsigned long start, unsigned long size, pgprot_t prot,
323 struct page **pages)
324{
325 int ret;
326
327 ret = map_kernel_range_noflush(start, size, prot, pages);
328 flush_cache_vmap(start, start + size);
329 return ret;
330}
331
332int is_vmalloc_or_module_addr(const void *x)
333{
334 /*
335 * ARM, x86-64 and sparc64 put modules in a special place,
336 * and fall back on vmalloc() if that fails. Others
337 * just put it in the vmalloc space.
338 */
339#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
340 unsigned long addr = (unsigned long)x;
341 if (addr >= MODULES_VADDR && addr < MODULES_END)
342 return 1;
343#endif
344 return is_vmalloc_addr(x);
345}
346
347/*
348 * Walk a vmap address to the struct page it maps.
349 */
350struct page *vmalloc_to_page(const void *vmalloc_addr)
351{
352 unsigned long addr = (unsigned long) vmalloc_addr;
353 struct page *page = NULL;
354 pgd_t *pgd = pgd_offset_k(addr);
355 p4d_t *p4d;
356 pud_t *pud;
357 pmd_t *pmd;
358 pte_t *ptep, pte;
359
360 /*
361 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
362 * architectures that do not vmalloc module space
363 */
364 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
365
366 if (pgd_none(*pgd))
367 return NULL;
368 p4d = p4d_offset(pgd, addr);
369 if (p4d_none(*p4d))
370 return NULL;
371 pud = pud_offset(p4d, addr);
372
373 /*
374 * Don't dereference bad PUD or PMD (below) entries. This will also
375 * identify huge mappings, which we may encounter on architectures
376 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
377 * identified as vmalloc addresses by is_vmalloc_addr(), but are
378 * not [unambiguously] associated with a struct page, so there is
379 * no correct value to return for them.
380 */
381 WARN_ON_ONCE(pud_bad(*pud));
382 if (pud_none(*pud) || pud_bad(*pud))
383 return NULL;
384 pmd = pmd_offset(pud, addr);
385 WARN_ON_ONCE(pmd_bad(*pmd));
386 if (pmd_none(*pmd) || pmd_bad(*pmd))
387 return NULL;
388
389 ptep = pte_offset_map(pmd, addr);
390 pte = *ptep;
391 if (pte_present(pte))
392 page = pte_page(pte);
393 pte_unmap(ptep);
394 return page;
395}
396EXPORT_SYMBOL(vmalloc_to_page);
397
398/*
399 * Map a vmalloc()-space virtual address to the physical page frame number.
400 */
401unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
402{
403 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
404}
405EXPORT_SYMBOL(vmalloc_to_pfn);
406
407
408/*** Global kva allocator ***/
409
410#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
411#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
412
413
414static DEFINE_SPINLOCK(vmap_area_lock);
415static DEFINE_SPINLOCK(free_vmap_area_lock);
416/* Export for kexec only */
417LIST_HEAD(vmap_area_list);
418static LLIST_HEAD(vmap_purge_list);
419static struct rb_root vmap_area_root = RB_ROOT;
420static bool vmap_initialized __read_mostly;
421
422/*
423 * This kmem_cache is used for vmap_area objects. Instead of
424 * allocating from slab we reuse an object from this cache to
425 * make things faster. Especially in "no edge" splitting of
426 * free block.
427 */
428static struct kmem_cache *vmap_area_cachep;
429
430/*
431 * This linked list is used in pair with free_vmap_area_root.
432 * It gives O(1) access to prev/next to perform fast coalescing.
433 */
434static LIST_HEAD(free_vmap_area_list);
435
436/*
437 * This augment red-black tree represents the free vmap space.
438 * All vmap_area objects in this tree are sorted by va->va_start
439 * address. It is used for allocation and merging when a vmap
440 * object is released.
441 *
442 * Each vmap_area node contains a maximum available free block
443 * of its sub-tree, right or left. Therefore it is possible to
444 * find a lowest match of free area.
445 */
446static struct rb_root free_vmap_area_root = RB_ROOT;
447
448/*
449 * Preload a CPU with one object for "no edge" split case. The
450 * aim is to get rid of allocations from the atomic context, thus
451 * to use more permissive allocation masks.
452 */
453static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
454
455static __always_inline unsigned long
456va_size(struct vmap_area *va)
457{
458 return (va->va_end - va->va_start);
459}
460
461static __always_inline unsigned long
462get_subtree_max_size(struct rb_node *node)
463{
464 struct vmap_area *va;
465
466 va = rb_entry_safe(node, struct vmap_area, rb_node);
467 return va ? va->subtree_max_size : 0;
468}
469
470/*
471 * Gets called when remove the node and rotate.
472 */
473static __always_inline unsigned long
474compute_subtree_max_size(struct vmap_area *va)
475{
476 return max3(va_size(va),
477 get_subtree_max_size(va->rb_node.rb_left),
478 get_subtree_max_size(va->rb_node.rb_right));
479}
480
481RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
482 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
483
484static void purge_vmap_area_lazy(void);
485static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
486static unsigned long lazy_max_pages(void);
487
488static atomic_long_t nr_vmalloc_pages;
489
490unsigned long vmalloc_nr_pages(void)
491{
492 return atomic_long_read(&nr_vmalloc_pages);
493}
494
495static struct vmap_area *__find_vmap_area(unsigned long addr)
496{
497 struct rb_node *n = vmap_area_root.rb_node;
498
499 while (n) {
500 struct vmap_area *va;
501
502 va = rb_entry(n, struct vmap_area, rb_node);
503 if (addr < va->va_start)
504 n = n->rb_left;
505 else if (addr >= va->va_end)
506 n = n->rb_right;
507 else
508 return va;
509 }
510
511 return NULL;
512}
513
514/*
515 * This function returns back addresses of parent node
516 * and its left or right link for further processing.
517 *
518 * Otherwise NULL is returned. In that case all further
519 * steps regarding inserting of conflicting overlap range
520 * have to be declined and actually considered as a bug.
521 */
522static __always_inline struct rb_node **
523find_va_links(struct vmap_area *va,
524 struct rb_root *root, struct rb_node *from,
525 struct rb_node **parent)
526{
527 struct vmap_area *tmp_va;
528 struct rb_node **link;
529
530 if (root) {
531 link = &root->rb_node;
532 if (unlikely(!*link)) {
533 *parent = NULL;
534 return link;
535 }
536 } else {
537 link = &from;
538 }
539
540 /*
541 * Go to the bottom of the tree. When we hit the last point
542 * we end up with parent rb_node and correct direction, i name
543 * it link, where the new va->rb_node will be attached to.
544 */
545 do {
546 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
547
548 /*
549 * During the traversal we also do some sanity check.
550 * Trigger the BUG() if there are sides(left/right)
551 * or full overlaps.
552 */
553 if (va->va_start < tmp_va->va_end &&
554 va->va_end <= tmp_va->va_start)
555 link = &(*link)->rb_left;
556 else if (va->va_end > tmp_va->va_start &&
557 va->va_start >= tmp_va->va_end)
558 link = &(*link)->rb_right;
559 else {
560 WARN(1, "vmalloc bug: 0x%lx-0x%lx overlaps with 0x%lx-0x%lx\n",
561 va->va_start, va->va_end, tmp_va->va_start, tmp_va->va_end);
562
563 return NULL;
564 }
565 } while (*link);
566
567 *parent = &tmp_va->rb_node;
568 return link;
569}
570
571static __always_inline struct list_head *
572get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
573{
574 struct list_head *list;
575
576 if (unlikely(!parent))
577 /*
578 * The red-black tree where we try to find VA neighbors
579 * before merging or inserting is empty, i.e. it means
580 * there is no free vmap space. Normally it does not
581 * happen but we handle this case anyway.
582 */
583 return NULL;
584
585 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
586 return (&parent->rb_right == link ? list->next : list);
587}
588
589static __always_inline void
590link_va(struct vmap_area *va, struct rb_root *root,
591 struct rb_node *parent, struct rb_node **link, struct list_head *head)
592{
593 /*
594 * VA is still not in the list, but we can
595 * identify its future previous list_head node.
596 */
597 if (likely(parent)) {
598 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
599 if (&parent->rb_right != link)
600 head = head->prev;
601 }
602
603 /* Insert to the rb-tree */
604 rb_link_node(&va->rb_node, parent, link);
605 if (root == &free_vmap_area_root) {
606 /*
607 * Some explanation here. Just perform simple insertion
608 * to the tree. We do not set va->subtree_max_size to
609 * its current size before calling rb_insert_augmented().
610 * It is because of we populate the tree from the bottom
611 * to parent levels when the node _is_ in the tree.
612 *
613 * Therefore we set subtree_max_size to zero after insertion,
614 * to let __augment_tree_propagate_from() puts everything to
615 * the correct order later on.
616 */
617 rb_insert_augmented(&va->rb_node,
618 root, &free_vmap_area_rb_augment_cb);
619 va->subtree_max_size = 0;
620 } else {
621 rb_insert_color(&va->rb_node, root);
622 }
623
624 /* Address-sort this list */
625 list_add(&va->list, head);
626}
627
628static __always_inline void
629unlink_va(struct vmap_area *va, struct rb_root *root)
630{
631 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
632 return;
633
634 if (root == &free_vmap_area_root)
635 rb_erase_augmented(&va->rb_node,
636 root, &free_vmap_area_rb_augment_cb);
637 else
638 rb_erase(&va->rb_node, root);
639
640 list_del(&va->list);
641 RB_CLEAR_NODE(&va->rb_node);
642}
643
644#if DEBUG_AUGMENT_PROPAGATE_CHECK
645static void
646augment_tree_propagate_check(void)
647{
648 struct vmap_area *va;
649 unsigned long computed_size;
650
651 list_for_each_entry(va, &free_vmap_area_list, list) {
652 computed_size = compute_subtree_max_size(va);
653 if (computed_size != va->subtree_max_size)
654 pr_emerg("tree is corrupted: %lu, %lu\n",
655 va_size(va), va->subtree_max_size);
656 }
657}
658#endif
659
660/*
661 * This function populates subtree_max_size from bottom to upper
662 * levels starting from VA point. The propagation must be done
663 * when VA size is modified by changing its va_start/va_end. Or
664 * in case of newly inserting of VA to the tree.
665 *
666 * It means that __augment_tree_propagate_from() must be called:
667 * - After VA has been inserted to the tree(free path);
668 * - After VA has been shrunk(allocation path);
669 * - After VA has been increased(merging path).
670 *
671 * Please note that, it does not mean that upper parent nodes
672 * and their subtree_max_size are recalculated all the time up
673 * to the root node.
674 *
675 * 4--8
676 * /\
677 * / \
678 * / \
679 * 2--2 8--8
680 *
681 * For example if we modify the node 4, shrinking it to 2, then
682 * no any modification is required. If we shrink the node 2 to 1
683 * its subtree_max_size is updated only, and set to 1. If we shrink
684 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
685 * node becomes 4--6.
686 */
687static __always_inline void
688augment_tree_propagate_from(struct vmap_area *va)
689{
690 /*
691 * Populate the tree from bottom towards the root until
692 * the calculated maximum available size of checked node
693 * is equal to its current one.
694 */
695 free_vmap_area_rb_augment_cb_propagate(&va->rb_node, NULL);
696
697#if DEBUG_AUGMENT_PROPAGATE_CHECK
698 augment_tree_propagate_check();
699#endif
700}
701
702static void
703insert_vmap_area(struct vmap_area *va,
704 struct rb_root *root, struct list_head *head)
705{
706 struct rb_node **link;
707 struct rb_node *parent;
708
709 link = find_va_links(va, root, NULL, &parent);
710 if (link)
711 link_va(va, root, parent, link, head);
712}
713
714static void
715insert_vmap_area_augment(struct vmap_area *va,
716 struct rb_node *from, struct rb_root *root,
717 struct list_head *head)
718{
719 struct rb_node **link;
720 struct rb_node *parent;
721
722 if (from)
723 link = find_va_links(va, NULL, from, &parent);
724 else
725 link = find_va_links(va, root, NULL, &parent);
726
727 if (link) {
728 link_va(va, root, parent, link, head);
729 augment_tree_propagate_from(va);
730 }
731}
732
733/*
734 * Merge de-allocated chunk of VA memory with previous
735 * and next free blocks. If coalesce is not done a new
736 * free area is inserted. If VA has been merged, it is
737 * freed.
738 *
739 * Please note, it can return NULL in case of overlap
740 * ranges, followed by WARN() report. Despite it is a
741 * buggy behaviour, a system can be alive and keep
742 * ongoing.
743 */
744static __always_inline struct vmap_area *
745merge_or_add_vmap_area(struct vmap_area *va,
746 struct rb_root *root, struct list_head *head)
747{
748 struct vmap_area *sibling;
749 struct list_head *next;
750 struct rb_node **link;
751 struct rb_node *parent;
752 bool merged = false;
753
754 /*
755 * Find a place in the tree where VA potentially will be
756 * inserted, unless it is merged with its sibling/siblings.
757 */
758 link = find_va_links(va, root, NULL, &parent);
759 if (!link)
760 return NULL;
761
762 /*
763 * Get next node of VA to check if merging can be done.
764 */
765 next = get_va_next_sibling(parent, link);
766 if (unlikely(next == NULL))
767 goto insert;
768
769 /*
770 * start end
771 * | |
772 * |<------VA------>|<-----Next----->|
773 * | |
774 * start end
775 */
776 if (next != head) {
777 sibling = list_entry(next, struct vmap_area, list);
778 if (sibling->va_start == va->va_end) {
779 sibling->va_start = va->va_start;
780
781 /* Free vmap_area object. */
782 kmem_cache_free(vmap_area_cachep, va);
783
784 /* Point to the new merged area. */
785 va = sibling;
786 merged = true;
787 }
788 }
789
790 /*
791 * start end
792 * | |
793 * |<-----Prev----->|<------VA------>|
794 * | |
795 * start end
796 */
797 if (next->prev != head) {
798 sibling = list_entry(next->prev, struct vmap_area, list);
799 if (sibling->va_end == va->va_start) {
800 /*
801 * If both neighbors are coalesced, it is important
802 * to unlink the "next" node first, followed by merging
803 * with "previous" one. Otherwise the tree might not be
804 * fully populated if a sibling's augmented value is
805 * "normalized" because of rotation operations.
806 */
807 if (merged)
808 unlink_va(va, root);
809
810 sibling->va_end = va->va_end;
811
812 /* Free vmap_area object. */
813 kmem_cache_free(vmap_area_cachep, va);
814
815 /* Point to the new merged area. */
816 va = sibling;
817 merged = true;
818 }
819 }
820
821insert:
822 if (!merged)
823 link_va(va, root, parent, link, head);
824
825 /*
826 * Last step is to check and update the tree.
827 */
828 augment_tree_propagate_from(va);
829 return va;
830}
831
832static __always_inline bool
833is_within_this_va(struct vmap_area *va, unsigned long size,
834 unsigned long align, unsigned long vstart)
835{
836 unsigned long nva_start_addr;
837
838 if (va->va_start > vstart)
839 nva_start_addr = ALIGN(va->va_start, align);
840 else
841 nva_start_addr = ALIGN(vstart, align);
842
843 /* Can be overflowed due to big size or alignment. */
844 if (nva_start_addr + size < nva_start_addr ||
845 nva_start_addr < vstart)
846 return false;
847
848 return (nva_start_addr + size <= va->va_end);
849}
850
851/*
852 * Find the first free block(lowest start address) in the tree,
853 * that will accomplish the request corresponding to passing
854 * parameters.
855 */
856static __always_inline struct vmap_area *
857find_vmap_lowest_match(unsigned long size,
858 unsigned long align, unsigned long vstart)
859{
860 struct vmap_area *va;
861 struct rb_node *node;
862 unsigned long length;
863
864 /* Start from the root. */
865 node = free_vmap_area_root.rb_node;
866
867 /* Adjust the search size for alignment overhead. */
868 length = size + align - 1;
869
870 while (node) {
871 va = rb_entry(node, struct vmap_area, rb_node);
872
873 if (get_subtree_max_size(node->rb_left) >= length &&
874 vstart < va->va_start) {
875 node = node->rb_left;
876 } else {
877 if (is_within_this_va(va, size, align, vstart))
878 return va;
879
880 /*
881 * Does not make sense to go deeper towards the right
882 * sub-tree if it does not have a free block that is
883 * equal or bigger to the requested search length.
884 */
885 if (get_subtree_max_size(node->rb_right) >= length) {
886 node = node->rb_right;
887 continue;
888 }
889
890 /*
891 * OK. We roll back and find the first right sub-tree,
892 * that will satisfy the search criteria. It can happen
893 * only once due to "vstart" restriction.
894 */
895 while ((node = rb_parent(node))) {
896 va = rb_entry(node, struct vmap_area, rb_node);
897 if (is_within_this_va(va, size, align, vstart))
898 return va;
899
900 if (get_subtree_max_size(node->rb_right) >= length &&
901 vstart <= va->va_start) {
902 node = node->rb_right;
903 break;
904 }
905 }
906 }
907 }
908
909 return NULL;
910}
911
912#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
913#include <linux/random.h>
914
915static struct vmap_area *
916find_vmap_lowest_linear_match(unsigned long size,
917 unsigned long align, unsigned long vstart)
918{
919 struct vmap_area *va;
920
921 list_for_each_entry(va, &free_vmap_area_list, list) {
922 if (!is_within_this_va(va, size, align, vstart))
923 continue;
924
925 return va;
926 }
927
928 return NULL;
929}
930
931static void
932find_vmap_lowest_match_check(unsigned long size)
933{
934 struct vmap_area *va_1, *va_2;
935 unsigned long vstart;
936 unsigned int rnd;
937
938 get_random_bytes(&rnd, sizeof(rnd));
939 vstart = VMALLOC_START + rnd;
940
941 va_1 = find_vmap_lowest_match(size, 1, vstart);
942 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
943
944 if (va_1 != va_2)
945 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
946 va_1, va_2, vstart);
947}
948#endif
949
950enum fit_type {
951 NOTHING_FIT = 0,
952 FL_FIT_TYPE = 1, /* full fit */
953 LE_FIT_TYPE = 2, /* left edge fit */
954 RE_FIT_TYPE = 3, /* right edge fit */
955 NE_FIT_TYPE = 4 /* no edge fit */
956};
957
958static __always_inline enum fit_type
959classify_va_fit_type(struct vmap_area *va,
960 unsigned long nva_start_addr, unsigned long size)
961{
962 enum fit_type type;
963
964 /* Check if it is within VA. */
965 if (nva_start_addr < va->va_start ||
966 nva_start_addr + size > va->va_end)
967 return NOTHING_FIT;
968
969 /* Now classify. */
970 if (va->va_start == nva_start_addr) {
971 if (va->va_end == nva_start_addr + size)
972 type = FL_FIT_TYPE;
973 else
974 type = LE_FIT_TYPE;
975 } else if (va->va_end == nva_start_addr + size) {
976 type = RE_FIT_TYPE;
977 } else {
978 type = NE_FIT_TYPE;
979 }
980
981 return type;
982}
983
984static __always_inline int
985adjust_va_to_fit_type(struct vmap_area *va,
986 unsigned long nva_start_addr, unsigned long size,
987 enum fit_type type)
988{
989 struct vmap_area *lva = NULL;
990
991 if (type == FL_FIT_TYPE) {
992 /*
993 * No need to split VA, it fully fits.
994 *
995 * | |
996 * V NVA V
997 * |---------------|
998 */
999 unlink_va(va, &free_vmap_area_root);
1000 kmem_cache_free(vmap_area_cachep, va);
1001 } else if (type == LE_FIT_TYPE) {
1002 /*
1003 * Split left edge of fit VA.
1004 *
1005 * | |
1006 * V NVA V R
1007 * |-------|-------|
1008 */
1009 va->va_start += size;
1010 } else if (type == RE_FIT_TYPE) {
1011 /*
1012 * Split right edge of fit VA.
1013 *
1014 * | |
1015 * L V NVA V
1016 * |-------|-------|
1017 */
1018 va->va_end = nva_start_addr;
1019 } else if (type == NE_FIT_TYPE) {
1020 /*
1021 * Split no edge of fit VA.
1022 *
1023 * | |
1024 * L V NVA V R
1025 * |---|-------|---|
1026 */
1027 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
1028 if (unlikely(!lva)) {
1029 /*
1030 * For percpu allocator we do not do any pre-allocation
1031 * and leave it as it is. The reason is it most likely
1032 * never ends up with NE_FIT_TYPE splitting. In case of
1033 * percpu allocations offsets and sizes are aligned to
1034 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
1035 * are its main fitting cases.
1036 *
1037 * There are a few exceptions though, as an example it is
1038 * a first allocation (early boot up) when we have "one"
1039 * big free space that has to be split.
1040 *
1041 * Also we can hit this path in case of regular "vmap"
1042 * allocations, if "this" current CPU was not preloaded.
1043 * See the comment in alloc_vmap_area() why. If so, then
1044 * GFP_NOWAIT is used instead to get an extra object for
1045 * split purpose. That is rare and most time does not
1046 * occur.
1047 *
1048 * What happens if an allocation gets failed. Basically,
1049 * an "overflow" path is triggered to purge lazily freed
1050 * areas to free some memory, then, the "retry" path is
1051 * triggered to repeat one more time. See more details
1052 * in alloc_vmap_area() function.
1053 */
1054 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
1055 if (!lva)
1056 return -1;
1057 }
1058
1059 /*
1060 * Build the remainder.
1061 */
1062 lva->va_start = va->va_start;
1063 lva->va_end = nva_start_addr;
1064
1065 /*
1066 * Shrink this VA to remaining size.
1067 */
1068 va->va_start = nva_start_addr + size;
1069 } else {
1070 return -1;
1071 }
1072
1073 if (type != FL_FIT_TYPE) {
1074 augment_tree_propagate_from(va);
1075
1076 if (lva) /* type == NE_FIT_TYPE */
1077 insert_vmap_area_augment(lva, &va->rb_node,
1078 &free_vmap_area_root, &free_vmap_area_list);
1079 }
1080
1081 return 0;
1082}
1083
1084/*
1085 * Returns a start address of the newly allocated area, if success.
1086 * Otherwise a vend is returned that indicates failure.
1087 */
1088static __always_inline unsigned long
1089__alloc_vmap_area(unsigned long size, unsigned long align,
1090 unsigned long vstart, unsigned long vend)
1091{
1092 unsigned long nva_start_addr;
1093 struct vmap_area *va;
1094 enum fit_type type;
1095 int ret;
1096
1097 va = find_vmap_lowest_match(size, align, vstart);
1098 if (unlikely(!va))
1099 return vend;
1100
1101 if (va->va_start > vstart)
1102 nva_start_addr = ALIGN(va->va_start, align);
1103 else
1104 nva_start_addr = ALIGN(vstart, align);
1105
1106 /* Check the "vend" restriction. */
1107 if (nva_start_addr + size > vend)
1108 return vend;
1109
1110 /* Classify what we have found. */
1111 type = classify_va_fit_type(va, nva_start_addr, size);
1112 if (WARN_ON_ONCE(type == NOTHING_FIT))
1113 return vend;
1114
1115 /* Update the free vmap_area. */
1116 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1117 if (ret)
1118 return vend;
1119
1120#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1121 find_vmap_lowest_match_check(size);
1122#endif
1123
1124 return nva_start_addr;
1125}
1126
1127/*
1128 * Free a region of KVA allocated by alloc_vmap_area
1129 */
1130static void free_vmap_area(struct vmap_area *va)
1131{
1132 /*
1133 * Remove from the busy tree/list.
1134 */
1135 spin_lock(&vmap_area_lock);
1136 unlink_va(va, &vmap_area_root);
1137 spin_unlock(&vmap_area_lock);
1138
1139 /*
1140 * Insert/Merge it back to the free tree/list.
1141 */
1142 spin_lock(&free_vmap_area_lock);
1143 merge_or_add_vmap_area(va, &free_vmap_area_root, &free_vmap_area_list);
1144 spin_unlock(&free_vmap_area_lock);
1145}
1146
1147/*
1148 * Allocate a region of KVA of the specified size and alignment, within the
1149 * vstart and vend.
1150 */
1151static struct vmap_area *alloc_vmap_area(unsigned long size,
1152 unsigned long align,
1153 unsigned long vstart, unsigned long vend,
1154 int node, gfp_t gfp_mask)
1155{
1156 struct vmap_area *va, *pva;
1157 unsigned long addr;
1158 int purged = 0;
1159 int ret;
1160
1161 BUG_ON(!size);
1162 BUG_ON(offset_in_page(size));
1163 BUG_ON(!is_power_of_2(align));
1164
1165 if (unlikely(!vmap_initialized))
1166 return ERR_PTR(-EBUSY);
1167
1168 might_sleep();
1169 gfp_mask = gfp_mask & GFP_RECLAIM_MASK;
1170
1171 va = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1172 if (unlikely(!va))
1173 return ERR_PTR(-ENOMEM);
1174
1175 /*
1176 * Only scan the relevant parts containing pointers to other objects
1177 * to avoid false negatives.
1178 */
1179 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask);
1180
1181retry:
1182 /*
1183 * Preload this CPU with one extra vmap_area object. It is used
1184 * when fit type of free area is NE_FIT_TYPE. Please note, it
1185 * does not guarantee that an allocation occurs on a CPU that
1186 * is preloaded, instead we minimize the case when it is not.
1187 * It can happen because of cpu migration, because there is a
1188 * race until the below spinlock is taken.
1189 *
1190 * The preload is done in non-atomic context, thus it allows us
1191 * to use more permissive allocation masks to be more stable under
1192 * low memory condition and high memory pressure. In rare case,
1193 * if not preloaded, GFP_NOWAIT is used.
1194 *
1195 * Set "pva" to NULL here, because of "retry" path.
1196 */
1197 pva = NULL;
1198
1199 if (!this_cpu_read(ne_fit_preload_node))
1200 /*
1201 * Even if it fails we do not really care about that.
1202 * Just proceed as it is. If needed "overflow" path
1203 * will refill the cache we allocate from.
1204 */
1205 pva = kmem_cache_alloc_node(vmap_area_cachep, gfp_mask, node);
1206
1207 spin_lock(&free_vmap_area_lock);
1208
1209 if (pva && __this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva))
1210 kmem_cache_free(vmap_area_cachep, pva);
1211
1212 /*
1213 * If an allocation fails, the "vend" address is
1214 * returned. Therefore trigger the overflow path.
1215 */
1216 addr = __alloc_vmap_area(size, align, vstart, vend);
1217 spin_unlock(&free_vmap_area_lock);
1218
1219 if (unlikely(addr == vend))
1220 goto overflow;
1221
1222 va->va_start = addr;
1223 va->va_end = addr + size;
1224 va->vm = NULL;
1225
1226
1227 spin_lock(&vmap_area_lock);
1228 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1229 spin_unlock(&vmap_area_lock);
1230
1231 BUG_ON(!IS_ALIGNED(va->va_start, align));
1232 BUG_ON(va->va_start < vstart);
1233 BUG_ON(va->va_end > vend);
1234
1235 ret = kasan_populate_vmalloc(addr, size);
1236 if (ret) {
1237 free_vmap_area(va);
1238 return ERR_PTR(ret);
1239 }
1240
1241 return va;
1242
1243overflow:
1244 if (!purged) {
1245 purge_vmap_area_lazy();
1246 purged = 1;
1247 goto retry;
1248 }
1249
1250 if (gfpflags_allow_blocking(gfp_mask)) {
1251 unsigned long freed = 0;
1252 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1253 if (freed > 0) {
1254 purged = 0;
1255 goto retry;
1256 }
1257 }
1258
1259 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1260 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1261 size);
1262
1263 kmem_cache_free(vmap_area_cachep, va);
1264 return ERR_PTR(-EBUSY);
1265}
1266
1267int register_vmap_purge_notifier(struct notifier_block *nb)
1268{
1269 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1270}
1271EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1272
1273int unregister_vmap_purge_notifier(struct notifier_block *nb)
1274{
1275 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1276}
1277EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1278
1279/*
1280 * lazy_max_pages is the maximum amount of virtual address space we gather up
1281 * before attempting to purge with a TLB flush.
1282 *
1283 * There is a tradeoff here: a larger number will cover more kernel page tables
1284 * and take slightly longer to purge, but it will linearly reduce the number of
1285 * global TLB flushes that must be performed. It would seem natural to scale
1286 * this number up linearly with the number of CPUs (because vmapping activity
1287 * could also scale linearly with the number of CPUs), however it is likely
1288 * that in practice, workloads might be constrained in other ways that mean
1289 * vmap activity will not scale linearly with CPUs. Also, I want to be
1290 * conservative and not introduce a big latency on huge systems, so go with
1291 * a less aggressive log scale. It will still be an improvement over the old
1292 * code, and it will be simple to change the scale factor if we find that it
1293 * becomes a problem on bigger systems.
1294 */
1295static unsigned long lazy_max_pages(void)
1296{
1297 unsigned int log;
1298
1299 log = fls(num_online_cpus());
1300
1301 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1302}
1303
1304static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1305
1306/*
1307 * Serialize vmap purging. There is no actual criticial section protected
1308 * by this look, but we want to avoid concurrent calls for performance
1309 * reasons and to make the pcpu_get_vm_areas more deterministic.
1310 */
1311static DEFINE_MUTEX(vmap_purge_lock);
1312
1313/* for per-CPU blocks */
1314static void purge_fragmented_blocks_allcpus(void);
1315
1316/*
1317 * called before a call to iounmap() if the caller wants vm_area_struct's
1318 * immediately freed.
1319 */
1320void set_iounmap_nonlazy(void)
1321{
1322 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1323}
1324
1325/*
1326 * Purges all lazily-freed vmap areas.
1327 */
1328static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1329{
1330 unsigned long resched_threshold;
1331 struct llist_node *valist;
1332 struct vmap_area *va;
1333 struct vmap_area *n_va;
1334
1335 lockdep_assert_held(&vmap_purge_lock);
1336
1337 valist = llist_del_all(&vmap_purge_list);
1338 if (unlikely(valist == NULL))
1339 return false;
1340
1341 /*
1342 * TODO: to calculate a flush range without looping.
1343 * The list can be up to lazy_max_pages() elements.
1344 */
1345 llist_for_each_entry(va, valist, purge_list) {
1346 if (va->va_start < start)
1347 start = va->va_start;
1348 if (va->va_end > end)
1349 end = va->va_end;
1350 }
1351
1352 flush_tlb_kernel_range(start, end);
1353 resched_threshold = lazy_max_pages() << 1;
1354
1355 spin_lock(&free_vmap_area_lock);
1356 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1357 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1358 unsigned long orig_start = va->va_start;
1359 unsigned long orig_end = va->va_end;
1360
1361 /*
1362 * Finally insert or merge lazily-freed area. It is
1363 * detached and there is no need to "unlink" it from
1364 * anything.
1365 */
1366 va = merge_or_add_vmap_area(va, &free_vmap_area_root,
1367 &free_vmap_area_list);
1368
1369 if (!va)
1370 continue;
1371
1372 if (is_vmalloc_or_module_addr((void *)orig_start))
1373 kasan_release_vmalloc(orig_start, orig_end,
1374 va->va_start, va->va_end);
1375
1376 atomic_long_sub(nr, &vmap_lazy_nr);
1377
1378 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1379 cond_resched_lock(&free_vmap_area_lock);
1380 }
1381 spin_unlock(&free_vmap_area_lock);
1382 return true;
1383}
1384
1385/*
1386 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1387 * is already purging.
1388 */
1389static void try_purge_vmap_area_lazy(void)
1390{
1391 if (mutex_trylock(&vmap_purge_lock)) {
1392 __purge_vmap_area_lazy(ULONG_MAX, 0);
1393 mutex_unlock(&vmap_purge_lock);
1394 }
1395}
1396
1397/*
1398 * Kick off a purge of the outstanding lazy areas.
1399 */
1400static void purge_vmap_area_lazy(void)
1401{
1402 mutex_lock(&vmap_purge_lock);
1403 purge_fragmented_blocks_allcpus();
1404 __purge_vmap_area_lazy(ULONG_MAX, 0);
1405 mutex_unlock(&vmap_purge_lock);
1406}
1407
1408/*
1409 * Free a vmap area, caller ensuring that the area has been unmapped
1410 * and flush_cache_vunmap had been called for the correct range
1411 * previously.
1412 */
1413static void free_vmap_area_noflush(struct vmap_area *va)
1414{
1415 unsigned long nr_lazy;
1416
1417 spin_lock(&vmap_area_lock);
1418 unlink_va(va, &vmap_area_root);
1419 spin_unlock(&vmap_area_lock);
1420
1421 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1422 PAGE_SHIFT, &vmap_lazy_nr);
1423
1424 /* After this point, we may free va at any time */
1425 llist_add(&va->purge_list, &vmap_purge_list);
1426
1427 if (unlikely(nr_lazy > lazy_max_pages()))
1428 try_purge_vmap_area_lazy();
1429}
1430
1431/*
1432 * Free and unmap a vmap area
1433 */
1434static void free_unmap_vmap_area(struct vmap_area *va)
1435{
1436 flush_cache_vunmap(va->va_start, va->va_end);
1437 unmap_kernel_range_noflush(va->va_start, va->va_end - va->va_start);
1438 if (debug_pagealloc_enabled_static())
1439 flush_tlb_kernel_range(va->va_start, va->va_end);
1440
1441 free_vmap_area_noflush(va);
1442}
1443
1444static struct vmap_area *find_vmap_area(unsigned long addr)
1445{
1446 struct vmap_area *va;
1447
1448 spin_lock(&vmap_area_lock);
1449 va = __find_vmap_area(addr);
1450 spin_unlock(&vmap_area_lock);
1451
1452 return va;
1453}
1454
1455/*** Per cpu kva allocator ***/
1456
1457/*
1458 * vmap space is limited especially on 32 bit architectures. Ensure there is
1459 * room for at least 16 percpu vmap blocks per CPU.
1460 */
1461/*
1462 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1463 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1464 * instead (we just need a rough idea)
1465 */
1466#if BITS_PER_LONG == 32
1467#define VMALLOC_SPACE (128UL*1024*1024)
1468#else
1469#define VMALLOC_SPACE (128UL*1024*1024*1024)
1470#endif
1471
1472#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1473#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1474#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1475#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1476#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1477#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1478#define VMAP_BBMAP_BITS \
1479 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1480 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1481 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1482
1483#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1484
1485struct vmap_block_queue {
1486 spinlock_t lock;
1487 struct list_head free;
1488};
1489
1490struct vmap_block {
1491 spinlock_t lock;
1492 struct vmap_area *va;
1493 unsigned long free, dirty;
1494 unsigned long dirty_min, dirty_max; /*< dirty range */
1495 struct list_head free_list;
1496 struct rcu_head rcu_head;
1497 struct list_head purge;
1498};
1499
1500/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1501static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1502
1503/*
1504 * XArray of vmap blocks, indexed by address, to quickly find a vmap block
1505 * in the free path. Could get rid of this if we change the API to return a
1506 * "cookie" from alloc, to be passed to free. But no big deal yet.
1507 */
1508static DEFINE_XARRAY(vmap_blocks);
1509
1510/*
1511 * We should probably have a fallback mechanism to allocate virtual memory
1512 * out of partially filled vmap blocks. However vmap block sizing should be
1513 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1514 * big problem.
1515 */
1516
1517static unsigned long addr_to_vb_idx(unsigned long addr)
1518{
1519 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1520 addr /= VMAP_BLOCK_SIZE;
1521 return addr;
1522}
1523
1524static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1525{
1526 unsigned long addr;
1527
1528 addr = va_start + (pages_off << PAGE_SHIFT);
1529 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1530 return (void *)addr;
1531}
1532
1533/**
1534 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1535 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1536 * @order: how many 2^order pages should be occupied in newly allocated block
1537 * @gfp_mask: flags for the page level allocator
1538 *
1539 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1540 */
1541static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1542{
1543 struct vmap_block_queue *vbq;
1544 struct vmap_block *vb;
1545 struct vmap_area *va;
1546 unsigned long vb_idx;
1547 int node, err;
1548 void *vaddr;
1549
1550 node = numa_node_id();
1551
1552 vb = kmalloc_node(sizeof(struct vmap_block),
1553 gfp_mask & GFP_RECLAIM_MASK, node);
1554 if (unlikely(!vb))
1555 return ERR_PTR(-ENOMEM);
1556
1557 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1558 VMALLOC_START, VMALLOC_END,
1559 node, gfp_mask);
1560 if (IS_ERR(va)) {
1561 kfree(vb);
1562 return ERR_CAST(va);
1563 }
1564
1565 vaddr = vmap_block_vaddr(va->va_start, 0);
1566 spin_lock_init(&vb->lock);
1567 vb->va = va;
1568 /* At least something should be left free */
1569 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1570 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1571 vb->dirty = 0;
1572 vb->dirty_min = VMAP_BBMAP_BITS;
1573 vb->dirty_max = 0;
1574 INIT_LIST_HEAD(&vb->free_list);
1575
1576 vb_idx = addr_to_vb_idx(va->va_start);
1577 err = xa_insert(&vmap_blocks, vb_idx, vb, gfp_mask);
1578 if (err) {
1579 kfree(vb);
1580 free_vmap_area(va);
1581 return ERR_PTR(err);
1582 }
1583
1584 vbq = &get_cpu_var(vmap_block_queue);
1585 spin_lock(&vbq->lock);
1586 list_add_tail_rcu(&vb->free_list, &vbq->free);
1587 spin_unlock(&vbq->lock);
1588 put_cpu_var(vmap_block_queue);
1589
1590 return vaddr;
1591}
1592
1593static void free_vmap_block(struct vmap_block *vb)
1594{
1595 struct vmap_block *tmp;
1596
1597 tmp = xa_erase(&vmap_blocks, addr_to_vb_idx(vb->va->va_start));
1598 BUG_ON(tmp != vb);
1599
1600 free_vmap_area_noflush(vb->va);
1601 kfree_rcu(vb, rcu_head);
1602}
1603
1604static void purge_fragmented_blocks(int cpu)
1605{
1606 LIST_HEAD(purge);
1607 struct vmap_block *vb;
1608 struct vmap_block *n_vb;
1609 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1610
1611 rcu_read_lock();
1612 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1613
1614 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1615 continue;
1616
1617 spin_lock(&vb->lock);
1618 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1619 vb->free = 0; /* prevent further allocs after releasing lock */
1620 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1621 vb->dirty_min = 0;
1622 vb->dirty_max = VMAP_BBMAP_BITS;
1623 spin_lock(&vbq->lock);
1624 list_del_rcu(&vb->free_list);
1625 spin_unlock(&vbq->lock);
1626 spin_unlock(&vb->lock);
1627 list_add_tail(&vb->purge, &purge);
1628 } else
1629 spin_unlock(&vb->lock);
1630 }
1631 rcu_read_unlock();
1632
1633 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1634 list_del(&vb->purge);
1635 free_vmap_block(vb);
1636 }
1637}
1638
1639static void purge_fragmented_blocks_allcpus(void)
1640{
1641 int cpu;
1642
1643 for_each_possible_cpu(cpu)
1644 purge_fragmented_blocks(cpu);
1645}
1646
1647static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1648{
1649 struct vmap_block_queue *vbq;
1650 struct vmap_block *vb;
1651 void *vaddr = NULL;
1652 unsigned int order;
1653
1654 BUG_ON(offset_in_page(size));
1655 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1656 if (WARN_ON(size == 0)) {
1657 /*
1658 * Allocating 0 bytes isn't what caller wants since
1659 * get_order(0) returns funny result. Just warn and terminate
1660 * early.
1661 */
1662 return NULL;
1663 }
1664 order = get_order(size);
1665
1666 rcu_read_lock();
1667 vbq = &get_cpu_var(vmap_block_queue);
1668 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1669 unsigned long pages_off;
1670
1671 spin_lock(&vb->lock);
1672 if (vb->free < (1UL << order)) {
1673 spin_unlock(&vb->lock);
1674 continue;
1675 }
1676
1677 pages_off = VMAP_BBMAP_BITS - vb->free;
1678 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1679 vb->free -= 1UL << order;
1680 if (vb->free == 0) {
1681 spin_lock(&vbq->lock);
1682 list_del_rcu(&vb->free_list);
1683 spin_unlock(&vbq->lock);
1684 }
1685
1686 spin_unlock(&vb->lock);
1687 break;
1688 }
1689
1690 put_cpu_var(vmap_block_queue);
1691 rcu_read_unlock();
1692
1693 /* Allocate new block if nothing was found */
1694 if (!vaddr)
1695 vaddr = new_vmap_block(order, gfp_mask);
1696
1697 return vaddr;
1698}
1699
1700static void vb_free(unsigned long addr, unsigned long size)
1701{
1702 unsigned long offset;
1703 unsigned int order;
1704 struct vmap_block *vb;
1705
1706 BUG_ON(offset_in_page(size));
1707 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1708
1709 flush_cache_vunmap(addr, addr + size);
1710
1711 order = get_order(size);
1712 offset = (addr & (VMAP_BLOCK_SIZE - 1)) >> PAGE_SHIFT;
1713 vb = xa_load(&vmap_blocks, addr_to_vb_idx(addr));
1714
1715 unmap_kernel_range_noflush(addr, size);
1716
1717 if (debug_pagealloc_enabled_static())
1718 flush_tlb_kernel_range(addr, addr + size);
1719
1720 spin_lock(&vb->lock);
1721
1722 /* Expand dirty range */
1723 vb->dirty_min = min(vb->dirty_min, offset);
1724 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1725
1726 vb->dirty += 1UL << order;
1727 if (vb->dirty == VMAP_BBMAP_BITS) {
1728 BUG_ON(vb->free);
1729 spin_unlock(&vb->lock);
1730 free_vmap_block(vb);
1731 } else
1732 spin_unlock(&vb->lock);
1733}
1734
1735static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1736{
1737 int cpu;
1738
1739 if (unlikely(!vmap_initialized))
1740 return;
1741
1742 might_sleep();
1743
1744 for_each_possible_cpu(cpu) {
1745 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1746 struct vmap_block *vb;
1747
1748 rcu_read_lock();
1749 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1750 spin_lock(&vb->lock);
1751 if (vb->dirty) {
1752 unsigned long va_start = vb->va->va_start;
1753 unsigned long s, e;
1754
1755 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1756 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1757
1758 start = min(s, start);
1759 end = max(e, end);
1760
1761 flush = 1;
1762 }
1763 spin_unlock(&vb->lock);
1764 }
1765 rcu_read_unlock();
1766 }
1767
1768 mutex_lock(&vmap_purge_lock);
1769 purge_fragmented_blocks_allcpus();
1770 if (!__purge_vmap_area_lazy(start, end) && flush)
1771 flush_tlb_kernel_range(start, end);
1772 mutex_unlock(&vmap_purge_lock);
1773}
1774
1775/**
1776 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1777 *
1778 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1779 * to amortize TLB flushing overheads. What this means is that any page you
1780 * have now, may, in a former life, have been mapped into kernel virtual
1781 * address by the vmap layer and so there might be some CPUs with TLB entries
1782 * still referencing that page (additional to the regular 1:1 kernel mapping).
1783 *
1784 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1785 * be sure that none of the pages we have control over will have any aliases
1786 * from the vmap layer.
1787 */
1788void vm_unmap_aliases(void)
1789{
1790 unsigned long start = ULONG_MAX, end = 0;
1791 int flush = 0;
1792
1793 _vm_unmap_aliases(start, end, flush);
1794}
1795EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1796
1797/**
1798 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1799 * @mem: the pointer returned by vm_map_ram
1800 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1801 */
1802void vm_unmap_ram(const void *mem, unsigned int count)
1803{
1804 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1805 unsigned long addr = (unsigned long)mem;
1806 struct vmap_area *va;
1807
1808 might_sleep();
1809 BUG_ON(!addr);
1810 BUG_ON(addr < VMALLOC_START);
1811 BUG_ON(addr > VMALLOC_END);
1812 BUG_ON(!PAGE_ALIGNED(addr));
1813
1814 kasan_poison_vmalloc(mem, size);
1815
1816 if (likely(count <= VMAP_MAX_ALLOC)) {
1817 debug_check_no_locks_freed(mem, size);
1818 vb_free(addr, size);
1819 return;
1820 }
1821
1822 va = find_vmap_area(addr);
1823 BUG_ON(!va);
1824 debug_check_no_locks_freed((void *)va->va_start,
1825 (va->va_end - va->va_start));
1826 free_unmap_vmap_area(va);
1827}
1828EXPORT_SYMBOL(vm_unmap_ram);
1829
1830/**
1831 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1832 * @pages: an array of pointers to the pages to be mapped
1833 * @count: number of pages
1834 * @node: prefer to allocate data structures on this node
1835 *
1836 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1837 * faster than vmap so it's good. But if you mix long-life and short-life
1838 * objects with vm_map_ram(), it could consume lots of address space through
1839 * fragmentation (especially on a 32bit machine). You could see failures in
1840 * the end. Please use this function for short-lived objects.
1841 *
1842 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1843 */
1844void *vm_map_ram(struct page **pages, unsigned int count, int node)
1845{
1846 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1847 unsigned long addr;
1848 void *mem;
1849
1850 if (likely(count <= VMAP_MAX_ALLOC)) {
1851 mem = vb_alloc(size, GFP_KERNEL);
1852 if (IS_ERR(mem))
1853 return NULL;
1854 addr = (unsigned long)mem;
1855 } else {
1856 struct vmap_area *va;
1857 va = alloc_vmap_area(size, PAGE_SIZE,
1858 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1859 if (IS_ERR(va))
1860 return NULL;
1861
1862 addr = va->va_start;
1863 mem = (void *)addr;
1864 }
1865
1866 kasan_unpoison_vmalloc(mem, size);
1867
1868 if (map_kernel_range(addr, size, PAGE_KERNEL, pages) < 0) {
1869 vm_unmap_ram(mem, count);
1870 return NULL;
1871 }
1872 return mem;
1873}
1874EXPORT_SYMBOL(vm_map_ram);
1875
1876static struct vm_struct *vmlist __initdata;
1877
1878/**
1879 * vm_area_add_early - add vmap area early during boot
1880 * @vm: vm_struct to add
1881 *
1882 * This function is used to add fixed kernel vm area to vmlist before
1883 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1884 * should contain proper values and the other fields should be zero.
1885 *
1886 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1887 */
1888void __init vm_area_add_early(struct vm_struct *vm)
1889{
1890 struct vm_struct *tmp, **p;
1891
1892 BUG_ON(vmap_initialized);
1893 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1894 if (tmp->addr >= vm->addr) {
1895 BUG_ON(tmp->addr < vm->addr + vm->size);
1896 break;
1897 } else
1898 BUG_ON(tmp->addr + tmp->size > vm->addr);
1899 }
1900 vm->next = *p;
1901 *p = vm;
1902}
1903
1904/**
1905 * vm_area_register_early - register vmap area early during boot
1906 * @vm: vm_struct to register
1907 * @align: requested alignment
1908 *
1909 * This function is used to register kernel vm area before
1910 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1911 * proper values on entry and other fields should be zero. On return,
1912 * vm->addr contains the allocated address.
1913 *
1914 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1915 */
1916void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1917{
1918 static size_t vm_init_off __initdata;
1919 unsigned long addr;
1920
1921 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1922 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1923
1924 vm->addr = (void *)addr;
1925
1926 vm_area_add_early(vm);
1927}
1928
1929static void vmap_init_free_space(void)
1930{
1931 unsigned long vmap_start = 1;
1932 const unsigned long vmap_end = ULONG_MAX;
1933 struct vmap_area *busy, *free;
1934
1935 /*
1936 * B F B B B F
1937 * -|-----|.....|-----|-----|-----|.....|-
1938 * | The KVA space |
1939 * |<--------------------------------->|
1940 */
1941 list_for_each_entry(busy, &vmap_area_list, list) {
1942 if (busy->va_start - vmap_start > 0) {
1943 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1944 if (!WARN_ON_ONCE(!free)) {
1945 free->va_start = vmap_start;
1946 free->va_end = busy->va_start;
1947
1948 insert_vmap_area_augment(free, NULL,
1949 &free_vmap_area_root,
1950 &free_vmap_area_list);
1951 }
1952 }
1953
1954 vmap_start = busy->va_end;
1955 }
1956
1957 if (vmap_end - vmap_start > 0) {
1958 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1959 if (!WARN_ON_ONCE(!free)) {
1960 free->va_start = vmap_start;
1961 free->va_end = vmap_end;
1962
1963 insert_vmap_area_augment(free, NULL,
1964 &free_vmap_area_root,
1965 &free_vmap_area_list);
1966 }
1967 }
1968}
1969
1970void __init vmalloc_init(void)
1971{
1972 struct vmap_area *va;
1973 struct vm_struct *tmp;
1974 int i;
1975
1976 /*
1977 * Create the cache for vmap_area objects.
1978 */
1979 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1980
1981 for_each_possible_cpu(i) {
1982 struct vmap_block_queue *vbq;
1983 struct vfree_deferred *p;
1984
1985 vbq = &per_cpu(vmap_block_queue, i);
1986 spin_lock_init(&vbq->lock);
1987 INIT_LIST_HEAD(&vbq->free);
1988 p = &per_cpu(vfree_deferred, i);
1989 init_llist_head(&p->list);
1990 INIT_WORK(&p->wq, free_work);
1991 }
1992
1993 /* Import existing vmlist entries. */
1994 for (tmp = vmlist; tmp; tmp = tmp->next) {
1995 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1996 if (WARN_ON_ONCE(!va))
1997 continue;
1998
1999 va->va_start = (unsigned long)tmp->addr;
2000 va->va_end = va->va_start + tmp->size;
2001 va->vm = tmp;
2002 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
2003 }
2004
2005 /*
2006 * Now we can initialize a free vmap space.
2007 */
2008 vmap_init_free_space();
2009 vmap_initialized = true;
2010}
2011
2012/**
2013 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
2014 * @addr: start of the VM area to unmap
2015 * @size: size of the VM area to unmap
2016 *
2017 * Similar to unmap_kernel_range_noflush() but flushes vcache before
2018 * the unmapping and tlb after.
2019 */
2020void unmap_kernel_range(unsigned long addr, unsigned long size)
2021{
2022 unsigned long end = addr + size;
2023
2024 flush_cache_vunmap(addr, end);
2025 unmap_kernel_range_noflush(addr, size);
2026 flush_tlb_kernel_range(addr, end);
2027}
2028
2029static inline void setup_vmalloc_vm_locked(struct vm_struct *vm,
2030 struct vmap_area *va, unsigned long flags, const void *caller)
2031{
2032 vm->flags = flags;
2033 vm->addr = (void *)va->va_start;
2034 vm->size = va->va_end - va->va_start;
2035 vm->caller = caller;
2036 va->vm = vm;
2037}
2038
2039static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2040 unsigned long flags, const void *caller)
2041{
2042 spin_lock(&vmap_area_lock);
2043 setup_vmalloc_vm_locked(vm, va, flags, caller);
2044 spin_unlock(&vmap_area_lock);
2045}
2046
2047static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2048{
2049 /*
2050 * Before removing VM_UNINITIALIZED,
2051 * we should make sure that vm has proper values.
2052 * Pair with smp_rmb() in show_numa_info().
2053 */
2054 smp_wmb();
2055 vm->flags &= ~VM_UNINITIALIZED;
2056}
2057
2058static struct vm_struct *__get_vm_area_node(unsigned long size,
2059 unsigned long align, unsigned long flags, unsigned long start,
2060 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
2061{
2062 struct vmap_area *va;
2063 struct vm_struct *area;
2064 unsigned long requested_size = size;
2065
2066 BUG_ON(in_interrupt());
2067 size = PAGE_ALIGN(size);
2068 if (unlikely(!size))
2069 return NULL;
2070
2071 if (flags & VM_IOREMAP)
2072 align = 1ul << clamp_t(int, get_count_order_long(size),
2073 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2074
2075 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2076 if (unlikely(!area))
2077 return NULL;
2078
2079 if (!(flags & VM_NO_GUARD))
2080 size += PAGE_SIZE;
2081
2082 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2083 if (IS_ERR(va)) {
2084 kfree(area);
2085 return NULL;
2086 }
2087
2088 kasan_unpoison_vmalloc((void *)va->va_start, requested_size);
2089
2090 setup_vmalloc_vm(area, va, flags, caller);
2091
2092 return area;
2093}
2094
2095struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2096 unsigned long start, unsigned long end,
2097 const void *caller)
2098{
2099 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2100 GFP_KERNEL, caller);
2101}
2102
2103/**
2104 * get_vm_area - reserve a contiguous kernel virtual area
2105 * @size: size of the area
2106 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2107 *
2108 * Search an area of @size in the kernel virtual mapping area,
2109 * and reserved it for out purposes. Returns the area descriptor
2110 * on success or %NULL on failure.
2111 *
2112 * Return: the area descriptor on success or %NULL on failure.
2113 */
2114struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2115{
2116 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2117 NUMA_NO_NODE, GFP_KERNEL,
2118 __builtin_return_address(0));
2119}
2120
2121struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2122 const void *caller)
2123{
2124 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2125 NUMA_NO_NODE, GFP_KERNEL, caller);
2126}
2127
2128/**
2129 * find_vm_area - find a continuous kernel virtual area
2130 * @addr: base address
2131 *
2132 * Search for the kernel VM area starting at @addr, and return it.
2133 * It is up to the caller to do all required locking to keep the returned
2134 * pointer valid.
2135 *
2136 * Return: pointer to the found area or %NULL on faulure
2137 */
2138struct vm_struct *find_vm_area(const void *addr)
2139{
2140 struct vmap_area *va;
2141
2142 va = find_vmap_area((unsigned long)addr);
2143 if (!va)
2144 return NULL;
2145
2146 return va->vm;
2147}
2148
2149/**
2150 * remove_vm_area - find and remove a continuous kernel virtual area
2151 * @addr: base address
2152 *
2153 * Search for the kernel VM area starting at @addr, and remove it.
2154 * This function returns the found VM area, but using it is NOT safe
2155 * on SMP machines, except for its size or flags.
2156 *
2157 * Return: pointer to the found area or %NULL on faulure
2158 */
2159struct vm_struct *remove_vm_area(const void *addr)
2160{
2161 struct vmap_area *va;
2162
2163 might_sleep();
2164
2165 spin_lock(&vmap_area_lock);
2166 va = __find_vmap_area((unsigned long)addr);
2167 if (va && va->vm) {
2168 struct vm_struct *vm = va->vm;
2169
2170 va->vm = NULL;
2171 spin_unlock(&vmap_area_lock);
2172
2173 kasan_free_shadow(vm);
2174 free_unmap_vmap_area(va);
2175
2176 return vm;
2177 }
2178
2179 spin_unlock(&vmap_area_lock);
2180 return NULL;
2181}
2182
2183static inline void set_area_direct_map(const struct vm_struct *area,
2184 int (*set_direct_map)(struct page *page))
2185{
2186 int i;
2187
2188 for (i = 0; i < area->nr_pages; i++)
2189 if (page_address(area->pages[i]))
2190 set_direct_map(area->pages[i]);
2191}
2192
2193/* Handle removing and resetting vm mappings related to the vm_struct. */
2194static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2195{
2196 unsigned long start = ULONG_MAX, end = 0;
2197 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2198 int flush_dmap = 0;
2199 int i;
2200
2201 remove_vm_area(area->addr);
2202
2203 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2204 if (!flush_reset)
2205 return;
2206
2207 /*
2208 * If not deallocating pages, just do the flush of the VM area and
2209 * return.
2210 */
2211 if (!deallocate_pages) {
2212 vm_unmap_aliases();
2213 return;
2214 }
2215
2216 /*
2217 * If execution gets here, flush the vm mapping and reset the direct
2218 * map. Find the start and end range of the direct mappings to make sure
2219 * the vm_unmap_aliases() flush includes the direct map.
2220 */
2221 for (i = 0; i < area->nr_pages; i++) {
2222 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2223 if (addr) {
2224 start = min(addr, start);
2225 end = max(addr + PAGE_SIZE, end);
2226 flush_dmap = 1;
2227 }
2228 }
2229
2230 /*
2231 * Set direct map to something invalid so that it won't be cached if
2232 * there are any accesses after the TLB flush, then flush the TLB and
2233 * reset the direct map permissions to the default.
2234 */
2235 set_area_direct_map(area, set_direct_map_invalid_noflush);
2236 _vm_unmap_aliases(start, end, flush_dmap);
2237 set_area_direct_map(area, set_direct_map_default_noflush);
2238}
2239
2240static void __vunmap(const void *addr, int deallocate_pages)
2241{
2242 struct vm_struct *area;
2243
2244 if (!addr)
2245 return;
2246
2247 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2248 addr))
2249 return;
2250
2251 area = find_vm_area(addr);
2252 if (unlikely(!area)) {
2253 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2254 addr);
2255 return;
2256 }
2257
2258 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2259 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2260
2261 kasan_poison_vmalloc(area->addr, area->size);
2262
2263 vm_remove_mappings(area, deallocate_pages);
2264
2265 if (deallocate_pages) {
2266 int i;
2267
2268 for (i = 0; i < area->nr_pages; i++) {
2269 struct page *page = area->pages[i];
2270
2271 BUG_ON(!page);
2272 __free_pages(page, 0);
2273 }
2274 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2275
2276 kvfree(area->pages);
2277 }
2278
2279 kfree(area);
2280 return;
2281}
2282
2283static inline void __vfree_deferred(const void *addr)
2284{
2285 /*
2286 * Use raw_cpu_ptr() because this can be called from preemptible
2287 * context. Preemption is absolutely fine here, because the llist_add()
2288 * implementation is lockless, so it works even if we are adding to
2289 * another cpu's list. schedule_work() should be fine with this too.
2290 */
2291 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2292
2293 if (llist_add((struct llist_node *)addr, &p->list))
2294 schedule_work(&p->wq);
2295}
2296
2297/**
2298 * vfree_atomic - release memory allocated by vmalloc()
2299 * @addr: memory base address
2300 *
2301 * This one is just like vfree() but can be called in any atomic context
2302 * except NMIs.
2303 */
2304void vfree_atomic(const void *addr)
2305{
2306 BUG_ON(in_nmi());
2307
2308 kmemleak_free(addr);
2309
2310 if (!addr)
2311 return;
2312 __vfree_deferred(addr);
2313}
2314
2315static void __vfree(const void *addr)
2316{
2317 if (unlikely(in_interrupt()))
2318 __vfree_deferred(addr);
2319 else
2320 __vunmap(addr, 1);
2321}
2322
2323/**
2324 * vfree - release memory allocated by vmalloc()
2325 * @addr: memory base address
2326 *
2327 * Free the virtually continuous memory area starting at @addr, as
2328 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2329 * NULL, no operation is performed.
2330 *
2331 * Must not be called in NMI context (strictly speaking, only if we don't
2332 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2333 * conventions for vfree() arch-depenedent would be a really bad idea)
2334 *
2335 * May sleep if called *not* from interrupt context.
2336 *
2337 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2338 */
2339void vfree(const void *addr)
2340{
2341 BUG_ON(in_nmi());
2342
2343 kmemleak_free(addr);
2344
2345 might_sleep_if(!in_interrupt());
2346
2347 if (!addr)
2348 return;
2349
2350 __vfree(addr);
2351}
2352EXPORT_SYMBOL(vfree);
2353
2354/**
2355 * vunmap - release virtual mapping obtained by vmap()
2356 * @addr: memory base address
2357 *
2358 * Free the virtually contiguous memory area starting at @addr,
2359 * which was created from the page array passed to vmap().
2360 *
2361 * Must not be called in interrupt context.
2362 */
2363void vunmap(const void *addr)
2364{
2365 BUG_ON(in_interrupt());
2366 might_sleep();
2367 if (addr)
2368 __vunmap(addr, 0);
2369}
2370EXPORT_SYMBOL(vunmap);
2371
2372/**
2373 * vmap - map an array of pages into virtually contiguous space
2374 * @pages: array of page pointers
2375 * @count: number of pages to map
2376 * @flags: vm_area->flags
2377 * @prot: page protection for the mapping
2378 *
2379 * Maps @count pages from @pages into contiguous kernel virtual
2380 * space.
2381 *
2382 * Return: the address of the area or %NULL on failure
2383 */
2384void *vmap(struct page **pages, unsigned int count,
2385 unsigned long flags, pgprot_t prot)
2386{
2387 struct vm_struct *area;
2388 unsigned long size; /* In bytes */
2389
2390 might_sleep();
2391
2392 if (count > totalram_pages())
2393 return NULL;
2394
2395 size = (unsigned long)count << PAGE_SHIFT;
2396 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2397 if (!area)
2398 return NULL;
2399
2400 if (map_kernel_range((unsigned long)area->addr, size, pgprot_nx(prot),
2401 pages) < 0) {
2402 vunmap(area->addr);
2403 return NULL;
2404 }
2405
2406 return area->addr;
2407}
2408EXPORT_SYMBOL(vmap);
2409
2410static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2411 pgprot_t prot, int node)
2412{
2413 struct page **pages;
2414 unsigned int nr_pages, array_size, i;
2415 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2416 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2417 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2418 0 :
2419 __GFP_HIGHMEM;
2420
2421 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2422 array_size = (nr_pages * sizeof(struct page *));
2423
2424 /* Please note that the recursion is strictly bounded. */
2425 if (array_size > PAGE_SIZE) {
2426 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2427 node, area->caller);
2428 } else {
2429 pages = kmalloc_node(array_size, nested_gfp, node);
2430 }
2431
2432 if (!pages) {
2433 remove_vm_area(area->addr);
2434 kfree(area);
2435 return NULL;
2436 }
2437
2438 area->pages = pages;
2439 area->nr_pages = nr_pages;
2440
2441 for (i = 0; i < area->nr_pages; i++) {
2442 struct page *page;
2443
2444 if (node == NUMA_NO_NODE)
2445 page = alloc_page(alloc_mask|highmem_mask);
2446 else
2447 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2448
2449 if (unlikely(!page)) {
2450 /* Successfully allocated i pages, free them in __vunmap() */
2451 area->nr_pages = i;
2452 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2453 goto fail;
2454 }
2455 area->pages[i] = page;
2456 if (gfpflags_allow_blocking(gfp_mask))
2457 cond_resched();
2458 }
2459 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2460
2461 if (map_kernel_range((unsigned long)area->addr, get_vm_area_size(area),
2462 prot, pages) < 0)
2463 goto fail;
2464
2465 return area->addr;
2466
2467fail:
2468 warn_alloc(gfp_mask, NULL,
2469 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2470 (area->nr_pages*PAGE_SIZE), area->size);
2471 __vfree(area->addr);
2472 return NULL;
2473}
2474
2475/**
2476 * __vmalloc_node_range - allocate virtually contiguous memory
2477 * @size: allocation size
2478 * @align: desired alignment
2479 * @start: vm area range start
2480 * @end: vm area range end
2481 * @gfp_mask: flags for the page level allocator
2482 * @prot: protection mask for the allocated pages
2483 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2484 * @node: node to use for allocation or NUMA_NO_NODE
2485 * @caller: caller's return address
2486 *
2487 * Allocate enough pages to cover @size from the page level
2488 * allocator with @gfp_mask flags. Map them into contiguous
2489 * kernel virtual space, using a pagetable protection of @prot.
2490 *
2491 * Return: the address of the area or %NULL on failure
2492 */
2493void *__vmalloc_node_range(unsigned long size, unsigned long align,
2494 unsigned long start, unsigned long end, gfp_t gfp_mask,
2495 pgprot_t prot, unsigned long vm_flags, int node,
2496 const void *caller)
2497{
2498 struct vm_struct *area;
2499 void *addr;
2500 unsigned long real_size = size;
2501
2502 size = PAGE_ALIGN(size);
2503 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2504 goto fail;
2505
2506 area = __get_vm_area_node(real_size, align, VM_ALLOC | VM_UNINITIALIZED |
2507 vm_flags, start, end, node, gfp_mask, caller);
2508 if (!area)
2509 goto fail;
2510
2511 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2512 if (!addr)
2513 return NULL;
2514
2515 /*
2516 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2517 * flag. It means that vm_struct is not fully initialized.
2518 * Now, it is fully initialized, so remove this flag here.
2519 */
2520 clear_vm_uninitialized_flag(area);
2521
2522 kmemleak_vmalloc(area, size, gfp_mask);
2523
2524 return addr;
2525
2526fail:
2527 warn_alloc(gfp_mask, NULL,
2528 "vmalloc: allocation failure: %lu bytes", real_size);
2529 return NULL;
2530}
2531
2532/**
2533 * __vmalloc_node - allocate virtually contiguous memory
2534 * @size: allocation size
2535 * @align: desired alignment
2536 * @gfp_mask: flags for the page level allocator
2537 * @node: node to use for allocation or NUMA_NO_NODE
2538 * @caller: caller's return address
2539 *
2540 * Allocate enough pages to cover @size from the page level allocator with
2541 * @gfp_mask flags. Map them into contiguous kernel virtual space.
2542 *
2543 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2544 * and __GFP_NOFAIL are not supported
2545 *
2546 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2547 * with mm people.
2548 *
2549 * Return: pointer to the allocated memory or %NULL on error
2550 */
2551void *__vmalloc_node(unsigned long size, unsigned long align,
2552 gfp_t gfp_mask, int node, const void *caller)
2553{
2554 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2555 gfp_mask, PAGE_KERNEL, 0, node, caller);
2556}
2557/*
2558 * This is only for performance analysis of vmalloc and stress purpose.
2559 * It is required by vmalloc test module, therefore do not use it other
2560 * than that.
2561 */
2562#ifdef CONFIG_TEST_VMALLOC_MODULE
2563EXPORT_SYMBOL_GPL(__vmalloc_node);
2564#endif
2565
2566void *__vmalloc(unsigned long size, gfp_t gfp_mask)
2567{
2568 return __vmalloc_node(size, 1, gfp_mask, NUMA_NO_NODE,
2569 __builtin_return_address(0));
2570}
2571EXPORT_SYMBOL(__vmalloc);
2572
2573/**
2574 * vmalloc - allocate virtually contiguous memory
2575 * @size: allocation size
2576 *
2577 * Allocate enough pages to cover @size from the page level
2578 * allocator and map them into contiguous kernel virtual space.
2579 *
2580 * For tight control over page level allocator and protection flags
2581 * use __vmalloc() instead.
2582 *
2583 * Return: pointer to the allocated memory or %NULL on error
2584 */
2585void *vmalloc(unsigned long size)
2586{
2587 return __vmalloc_node(size, 1, GFP_KERNEL, NUMA_NO_NODE,
2588 __builtin_return_address(0));
2589}
2590EXPORT_SYMBOL(vmalloc);
2591
2592/**
2593 * vzalloc - allocate virtually contiguous memory with zero fill
2594 * @size: allocation size
2595 *
2596 * Allocate enough pages to cover @size from the page level
2597 * allocator and map them into contiguous kernel virtual space.
2598 * The memory allocated is set to zero.
2599 *
2600 * For tight control over page level allocator and protection flags
2601 * use __vmalloc() instead.
2602 *
2603 * Return: pointer to the allocated memory or %NULL on error
2604 */
2605void *vzalloc(unsigned long size)
2606{
2607 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, NUMA_NO_NODE,
2608 __builtin_return_address(0));
2609}
2610EXPORT_SYMBOL(vzalloc);
2611
2612/**
2613 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2614 * @size: allocation size
2615 *
2616 * The resulting memory area is zeroed so it can be mapped to userspace
2617 * without leaking data.
2618 *
2619 * Return: pointer to the allocated memory or %NULL on error
2620 */
2621void *vmalloc_user(unsigned long size)
2622{
2623 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2624 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2625 VM_USERMAP, NUMA_NO_NODE,
2626 __builtin_return_address(0));
2627}
2628EXPORT_SYMBOL(vmalloc_user);
2629
2630/**
2631 * vmalloc_node - allocate memory on a specific node
2632 * @size: allocation size
2633 * @node: numa node
2634 *
2635 * Allocate enough pages to cover @size from the page level
2636 * allocator and map them into contiguous kernel virtual space.
2637 *
2638 * For tight control over page level allocator and protection flags
2639 * use __vmalloc() instead.
2640 *
2641 * Return: pointer to the allocated memory or %NULL on error
2642 */
2643void *vmalloc_node(unsigned long size, int node)
2644{
2645 return __vmalloc_node(size, 1, GFP_KERNEL, node,
2646 __builtin_return_address(0));
2647}
2648EXPORT_SYMBOL(vmalloc_node);
2649
2650/**
2651 * vzalloc_node - allocate memory on a specific node with zero fill
2652 * @size: allocation size
2653 * @node: numa node
2654 *
2655 * Allocate enough pages to cover @size from the page level
2656 * allocator and map them into contiguous kernel virtual space.
2657 * The memory allocated is set to zero.
2658 *
2659 * Return: pointer to the allocated memory or %NULL on error
2660 */
2661void *vzalloc_node(unsigned long size, int node)
2662{
2663 return __vmalloc_node(size, 1, GFP_KERNEL | __GFP_ZERO, node,
2664 __builtin_return_address(0));
2665}
2666EXPORT_SYMBOL(vzalloc_node);
2667
2668#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2669#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2670#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2671#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2672#else
2673/*
2674 * 64b systems should always have either DMA or DMA32 zones. For others
2675 * GFP_DMA32 should do the right thing and use the normal zone.
2676 */
2677#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2678#endif
2679
2680/**
2681 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2682 * @size: allocation size
2683 *
2684 * Allocate enough 32bit PA addressable pages to cover @size from the
2685 * page level allocator and map them into contiguous kernel virtual space.
2686 *
2687 * Return: pointer to the allocated memory or %NULL on error
2688 */
2689void *vmalloc_32(unsigned long size)
2690{
2691 return __vmalloc_node(size, 1, GFP_VMALLOC32, NUMA_NO_NODE,
2692 __builtin_return_address(0));
2693}
2694EXPORT_SYMBOL(vmalloc_32);
2695
2696/**
2697 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2698 * @size: allocation size
2699 *
2700 * The resulting memory area is 32bit addressable and zeroed so it can be
2701 * mapped to userspace without leaking data.
2702 *
2703 * Return: pointer to the allocated memory or %NULL on error
2704 */
2705void *vmalloc_32_user(unsigned long size)
2706{
2707 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2708 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2709 VM_USERMAP, NUMA_NO_NODE,
2710 __builtin_return_address(0));
2711}
2712EXPORT_SYMBOL(vmalloc_32_user);
2713
2714/*
2715 * small helper routine , copy contents to buf from addr.
2716 * If the page is not present, fill zero.
2717 */
2718
2719static int aligned_vread(char *buf, char *addr, unsigned long count)
2720{
2721 struct page *p;
2722 int copied = 0;
2723
2724 while (count) {
2725 unsigned long offset, length;
2726
2727 offset = offset_in_page(addr);
2728 length = PAGE_SIZE - offset;
2729 if (length > count)
2730 length = count;
2731 p = vmalloc_to_page(addr);
2732 /*
2733 * To do safe access to this _mapped_ area, we need
2734 * lock. But adding lock here means that we need to add
2735 * overhead of vmalloc()/vfree() calles for this _debug_
2736 * interface, rarely used. Instead of that, we'll use
2737 * kmap() and get small overhead in this access function.
2738 */
2739 if (p) {
2740 /*
2741 * we can expect USER0 is not used (see vread/vwrite's
2742 * function description)
2743 */
2744 void *map = kmap_atomic(p);
2745 memcpy(buf, map + offset, length);
2746 kunmap_atomic(map);
2747 } else
2748 memset(buf, 0, length);
2749
2750 addr += length;
2751 buf += length;
2752 copied += length;
2753 count -= length;
2754 }
2755 return copied;
2756}
2757
2758static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2759{
2760 struct page *p;
2761 int copied = 0;
2762
2763 while (count) {
2764 unsigned long offset, length;
2765
2766 offset = offset_in_page(addr);
2767 length = PAGE_SIZE - offset;
2768 if (length > count)
2769 length = count;
2770 p = vmalloc_to_page(addr);
2771 /*
2772 * To do safe access to this _mapped_ area, we need
2773 * lock. But adding lock here means that we need to add
2774 * overhead of vmalloc()/vfree() calles for this _debug_
2775 * interface, rarely used. Instead of that, we'll use
2776 * kmap() and get small overhead in this access function.
2777 */
2778 if (p) {
2779 /*
2780 * we can expect USER0 is not used (see vread/vwrite's
2781 * function description)
2782 */
2783 void *map = kmap_atomic(p);
2784 memcpy(map + offset, buf, length);
2785 kunmap_atomic(map);
2786 }
2787 addr += length;
2788 buf += length;
2789 copied += length;
2790 count -= length;
2791 }
2792 return copied;
2793}
2794
2795/**
2796 * vread() - read vmalloc area in a safe way.
2797 * @buf: buffer for reading data
2798 * @addr: vm address.
2799 * @count: number of bytes to be read.
2800 *
2801 * This function checks that addr is a valid vmalloc'ed area, and
2802 * copy data from that area to a given buffer. If the given memory range
2803 * of [addr...addr+count) includes some valid address, data is copied to
2804 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2805 * IOREMAP area is treated as memory hole and no copy is done.
2806 *
2807 * If [addr...addr+count) doesn't includes any intersects with alive
2808 * vm_struct area, returns 0. @buf should be kernel's buffer.
2809 *
2810 * Note: In usual ops, vread() is never necessary because the caller
2811 * should know vmalloc() area is valid and can use memcpy().
2812 * This is for routines which have to access vmalloc area without
2813 * any information, as /dev/kmem.
2814 *
2815 * Return: number of bytes for which addr and buf should be increased
2816 * (same number as @count) or %0 if [addr...addr+count) doesn't
2817 * include any intersection with valid vmalloc area
2818 */
2819long vread(char *buf, char *addr, unsigned long count)
2820{
2821 struct vmap_area *va;
2822 struct vm_struct *vm;
2823 char *vaddr, *buf_start = buf;
2824 unsigned long buflen = count;
2825 unsigned long n;
2826
2827 /* Don't allow overflow */
2828 if ((unsigned long) addr + count < count)
2829 count = -(unsigned long) addr;
2830
2831 spin_lock(&vmap_area_lock);
2832 list_for_each_entry(va, &vmap_area_list, list) {
2833 if (!count)
2834 break;
2835
2836 if (!va->vm)
2837 continue;
2838
2839 vm = va->vm;
2840 vaddr = (char *) vm->addr;
2841 if (addr >= vaddr + get_vm_area_size(vm))
2842 continue;
2843 while (addr < vaddr) {
2844 if (count == 0)
2845 goto finished;
2846 *buf = '\0';
2847 buf++;
2848 addr++;
2849 count--;
2850 }
2851 n = vaddr + get_vm_area_size(vm) - addr;
2852 if (n > count)
2853 n = count;
2854 if (!(vm->flags & VM_IOREMAP))
2855 aligned_vread(buf, addr, n);
2856 else /* IOREMAP area is treated as memory hole */
2857 memset(buf, 0, n);
2858 buf += n;
2859 addr += n;
2860 count -= n;
2861 }
2862finished:
2863 spin_unlock(&vmap_area_lock);
2864
2865 if (buf == buf_start)
2866 return 0;
2867 /* zero-fill memory holes */
2868 if (buf != buf_start + buflen)
2869 memset(buf, 0, buflen - (buf - buf_start));
2870
2871 return buflen;
2872}
2873
2874/**
2875 * vwrite() - write vmalloc area in a safe way.
2876 * @buf: buffer for source data
2877 * @addr: vm address.
2878 * @count: number of bytes to be read.
2879 *
2880 * This function checks that addr is a valid vmalloc'ed area, and
2881 * copy data from a buffer to the given addr. If specified range of
2882 * [addr...addr+count) includes some valid address, data is copied from
2883 * proper area of @buf. If there are memory holes, no copy to hole.
2884 * IOREMAP area is treated as memory hole and no copy is done.
2885 *
2886 * If [addr...addr+count) doesn't includes any intersects with alive
2887 * vm_struct area, returns 0. @buf should be kernel's buffer.
2888 *
2889 * Note: In usual ops, vwrite() is never necessary because the caller
2890 * should know vmalloc() area is valid and can use memcpy().
2891 * This is for routines which have to access vmalloc area without
2892 * any information, as /dev/kmem.
2893 *
2894 * Return: number of bytes for which addr and buf should be
2895 * increased (same number as @count) or %0 if [addr...addr+count)
2896 * doesn't include any intersection with valid vmalloc area
2897 */
2898long vwrite(char *buf, char *addr, unsigned long count)
2899{
2900 struct vmap_area *va;
2901 struct vm_struct *vm;
2902 char *vaddr;
2903 unsigned long n, buflen;
2904 int copied = 0;
2905
2906 /* Don't allow overflow */
2907 if ((unsigned long) addr + count < count)
2908 count = -(unsigned long) addr;
2909 buflen = count;
2910
2911 spin_lock(&vmap_area_lock);
2912 list_for_each_entry(va, &vmap_area_list, list) {
2913 if (!count)
2914 break;
2915
2916 if (!va->vm)
2917 continue;
2918
2919 vm = va->vm;
2920 vaddr = (char *) vm->addr;
2921 if (addr >= vaddr + get_vm_area_size(vm))
2922 continue;
2923 while (addr < vaddr) {
2924 if (count == 0)
2925 goto finished;
2926 buf++;
2927 addr++;
2928 count--;
2929 }
2930 n = vaddr + get_vm_area_size(vm) - addr;
2931 if (n > count)
2932 n = count;
2933 if (!(vm->flags & VM_IOREMAP)) {
2934 aligned_vwrite(buf, addr, n);
2935 copied++;
2936 }
2937 buf += n;
2938 addr += n;
2939 count -= n;
2940 }
2941finished:
2942 spin_unlock(&vmap_area_lock);
2943 if (!copied)
2944 return 0;
2945 return buflen;
2946}
2947
2948/**
2949 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2950 * @vma: vma to cover
2951 * @uaddr: target user address to start at
2952 * @kaddr: virtual address of vmalloc kernel memory
2953 * @pgoff: offset from @kaddr to start at
2954 * @size: size of map area
2955 *
2956 * Returns: 0 for success, -Exxx on failure
2957 *
2958 * This function checks that @kaddr is a valid vmalloc'ed area,
2959 * and that it is big enough to cover the range starting at
2960 * @uaddr in @vma. Will return failure if that criteria isn't
2961 * met.
2962 *
2963 * Similar to remap_pfn_range() (see mm/memory.c)
2964 */
2965int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2966 void *kaddr, unsigned long pgoff,
2967 unsigned long size)
2968{
2969 struct vm_struct *area;
2970 unsigned long off;
2971 unsigned long end_index;
2972
2973 if (check_shl_overflow(pgoff, PAGE_SHIFT, &off))
2974 return -EINVAL;
2975
2976 size = PAGE_ALIGN(size);
2977
2978 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2979 return -EINVAL;
2980
2981 area = find_vm_area(kaddr);
2982 if (!area)
2983 return -EINVAL;
2984
2985 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
2986 return -EINVAL;
2987
2988 if (check_add_overflow(size, off, &end_index) ||
2989 end_index > get_vm_area_size(area))
2990 return -EINVAL;
2991 kaddr += off;
2992
2993 do {
2994 struct page *page = vmalloc_to_page(kaddr);
2995 int ret;
2996
2997 ret = vm_insert_page(vma, uaddr, page);
2998 if (ret)
2999 return ret;
3000
3001 uaddr += PAGE_SIZE;
3002 kaddr += PAGE_SIZE;
3003 size -= PAGE_SIZE;
3004 } while (size > 0);
3005
3006 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3007
3008 return 0;
3009}
3010EXPORT_SYMBOL(remap_vmalloc_range_partial);
3011
3012/**
3013 * remap_vmalloc_range - map vmalloc pages to userspace
3014 * @vma: vma to cover (map full range of vma)
3015 * @addr: vmalloc memory
3016 * @pgoff: number of pages into addr before first page to map
3017 *
3018 * Returns: 0 for success, -Exxx on failure
3019 *
3020 * This function checks that addr is a valid vmalloc'ed area, and
3021 * that it is big enough to cover the vma. Will return failure if
3022 * that criteria isn't met.
3023 *
3024 * Similar to remap_pfn_range() (see mm/memory.c)
3025 */
3026int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3027 unsigned long pgoff)
3028{
3029 return remap_vmalloc_range_partial(vma, vma->vm_start,
3030 addr, pgoff,
3031 vma->vm_end - vma->vm_start);
3032}
3033EXPORT_SYMBOL(remap_vmalloc_range);
3034
3035static int f(pte_t *pte, unsigned long addr, void *data)
3036{
3037 pte_t ***p = data;
3038
3039 if (p) {
3040 *(*p) = pte;
3041 (*p)++;
3042 }
3043 return 0;
3044}
3045
3046/**
3047 * alloc_vm_area - allocate a range of kernel address space
3048 * @size: size of the area
3049 * @ptes: returns the PTEs for the address space
3050 *
3051 * Returns: NULL on failure, vm_struct on success
3052 *
3053 * This function reserves a range of kernel address space, and
3054 * allocates pagetables to map that range. No actual mappings
3055 * are created.
3056 *
3057 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3058 * allocated for the VM area are returned.
3059 */
3060struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3061{
3062 struct vm_struct *area;
3063
3064 area = get_vm_area_caller(size, VM_IOREMAP,
3065 __builtin_return_address(0));
3066 if (area == NULL)
3067 return NULL;
3068
3069 /*
3070 * This ensures that page tables are constructed for this region
3071 * of kernel virtual address space and mapped into init_mm.
3072 */
3073 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3074 size, f, ptes ? &ptes : NULL)) {
3075 free_vm_area(area);
3076 return NULL;
3077 }
3078
3079 return area;
3080}
3081EXPORT_SYMBOL_GPL(alloc_vm_area);
3082
3083void free_vm_area(struct vm_struct *area)
3084{
3085 struct vm_struct *ret;
3086 ret = remove_vm_area(area->addr);
3087 BUG_ON(ret != area);
3088 kfree(area);
3089}
3090EXPORT_SYMBOL_GPL(free_vm_area);
3091
3092#ifdef CONFIG_SMP
3093static struct vmap_area *node_to_va(struct rb_node *n)
3094{
3095 return rb_entry_safe(n, struct vmap_area, rb_node);
3096}
3097
3098/**
3099 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3100 * @addr: target address
3101 *
3102 * Returns: vmap_area if it is found. If there is no such area
3103 * the first highest(reverse order) vmap_area is returned
3104 * i.e. va->va_start < addr && va->va_end < addr or NULL
3105 * if there are no any areas before @addr.
3106 */
3107static struct vmap_area *
3108pvm_find_va_enclose_addr(unsigned long addr)
3109{
3110 struct vmap_area *va, *tmp;
3111 struct rb_node *n;
3112
3113 n = free_vmap_area_root.rb_node;
3114 va = NULL;
3115
3116 while (n) {
3117 tmp = rb_entry(n, struct vmap_area, rb_node);
3118 if (tmp->va_start <= addr) {
3119 va = tmp;
3120 if (tmp->va_end >= addr)
3121 break;
3122
3123 n = n->rb_right;
3124 } else {
3125 n = n->rb_left;
3126 }
3127 }
3128
3129 return va;
3130}
3131
3132/**
3133 * pvm_determine_end_from_reverse - find the highest aligned address
3134 * of free block below VMALLOC_END
3135 * @va:
3136 * in - the VA we start the search(reverse order);
3137 * out - the VA with the highest aligned end address.
3138 *
3139 * Returns: determined end address within vmap_area
3140 */
3141static unsigned long
3142pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3143{
3144 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3145 unsigned long addr;
3146
3147 if (likely(*va)) {
3148 list_for_each_entry_from_reverse((*va),
3149 &free_vmap_area_list, list) {
3150 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3151 if ((*va)->va_start < addr)
3152 return addr;
3153 }
3154 }
3155
3156 return 0;
3157}
3158
3159/**
3160 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3161 * @offsets: array containing offset of each area
3162 * @sizes: array containing size of each area
3163 * @nr_vms: the number of areas to allocate
3164 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3165 *
3166 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3167 * vm_structs on success, %NULL on failure
3168 *
3169 * Percpu allocator wants to use congruent vm areas so that it can
3170 * maintain the offsets among percpu areas. This function allocates
3171 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3172 * be scattered pretty far, distance between two areas easily going up
3173 * to gigabytes. To avoid interacting with regular vmallocs, these
3174 * areas are allocated from top.
3175 *
3176 * Despite its complicated look, this allocator is rather simple. It
3177 * does everything top-down and scans free blocks from the end looking
3178 * for matching base. While scanning, if any of the areas do not fit the
3179 * base address is pulled down to fit the area. Scanning is repeated till
3180 * all the areas fit and then all necessary data structures are inserted
3181 * and the result is returned.
3182 */
3183struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3184 const size_t *sizes, int nr_vms,
3185 size_t align)
3186{
3187 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3188 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3189 struct vmap_area **vas, *va;
3190 struct vm_struct **vms;
3191 int area, area2, last_area, term_area;
3192 unsigned long base, start, size, end, last_end, orig_start, orig_end;
3193 bool purged = false;
3194 enum fit_type type;
3195
3196 /* verify parameters and allocate data structures */
3197 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3198 for (last_area = 0, area = 0; area < nr_vms; area++) {
3199 start = offsets[area];
3200 end = start + sizes[area];
3201
3202 /* is everything aligned properly? */
3203 BUG_ON(!IS_ALIGNED(offsets[area], align));
3204 BUG_ON(!IS_ALIGNED(sizes[area], align));
3205
3206 /* detect the area with the highest address */
3207 if (start > offsets[last_area])
3208 last_area = area;
3209
3210 for (area2 = area + 1; area2 < nr_vms; area2++) {
3211 unsigned long start2 = offsets[area2];
3212 unsigned long end2 = start2 + sizes[area2];
3213
3214 BUG_ON(start2 < end && start < end2);
3215 }
3216 }
3217 last_end = offsets[last_area] + sizes[last_area];
3218
3219 if (vmalloc_end - vmalloc_start < last_end) {
3220 WARN_ON(true);
3221 return NULL;
3222 }
3223
3224 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3225 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3226 if (!vas || !vms)
3227 goto err_free2;
3228
3229 for (area = 0; area < nr_vms; area++) {
3230 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3231 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3232 if (!vas[area] || !vms[area])
3233 goto err_free;
3234 }
3235retry:
3236 spin_lock(&free_vmap_area_lock);
3237
3238 /* start scanning - we scan from the top, begin with the last area */
3239 area = term_area = last_area;
3240 start = offsets[area];
3241 end = start + sizes[area];
3242
3243 va = pvm_find_va_enclose_addr(vmalloc_end);
3244 base = pvm_determine_end_from_reverse(&va, align) - end;
3245
3246 while (true) {
3247 /*
3248 * base might have underflowed, add last_end before
3249 * comparing.
3250 */
3251 if (base + last_end < vmalloc_start + last_end)
3252 goto overflow;
3253
3254 /*
3255 * Fitting base has not been found.
3256 */
3257 if (va == NULL)
3258 goto overflow;
3259
3260 /*
3261 * If required width exceeds current VA block, move
3262 * base downwards and then recheck.
3263 */
3264 if (base + end > va->va_end) {
3265 base = pvm_determine_end_from_reverse(&va, align) - end;
3266 term_area = area;
3267 continue;
3268 }
3269
3270 /*
3271 * If this VA does not fit, move base downwards and recheck.
3272 */
3273 if (base + start < va->va_start) {
3274 va = node_to_va(rb_prev(&va->rb_node));
3275 base = pvm_determine_end_from_reverse(&va, align) - end;
3276 term_area = area;
3277 continue;
3278 }
3279
3280 /*
3281 * This area fits, move on to the previous one. If
3282 * the previous one is the terminal one, we're done.
3283 */
3284 area = (area + nr_vms - 1) % nr_vms;
3285 if (area == term_area)
3286 break;
3287
3288 start = offsets[area];
3289 end = start + sizes[area];
3290 va = pvm_find_va_enclose_addr(base + end);
3291 }
3292
3293 /* we've found a fitting base, insert all va's */
3294 for (area = 0; area < nr_vms; area++) {
3295 int ret;
3296
3297 start = base + offsets[area];
3298 size = sizes[area];
3299
3300 va = pvm_find_va_enclose_addr(start);
3301 if (WARN_ON_ONCE(va == NULL))
3302 /* It is a BUG(), but trigger recovery instead. */
3303 goto recovery;
3304
3305 type = classify_va_fit_type(va, start, size);
3306 if (WARN_ON_ONCE(type == NOTHING_FIT))
3307 /* It is a BUG(), but trigger recovery instead. */
3308 goto recovery;
3309
3310 ret = adjust_va_to_fit_type(va, start, size, type);
3311 if (unlikely(ret))
3312 goto recovery;
3313
3314 /* Allocated area. */
3315 va = vas[area];
3316 va->va_start = start;
3317 va->va_end = start + size;
3318 }
3319
3320 spin_unlock(&free_vmap_area_lock);
3321
3322 /* populate the kasan shadow space */
3323 for (area = 0; area < nr_vms; area++) {
3324 if (kasan_populate_vmalloc(vas[area]->va_start, sizes[area]))
3325 goto err_free_shadow;
3326
3327 kasan_unpoison_vmalloc((void *)vas[area]->va_start,
3328 sizes[area]);
3329 }
3330
3331 /* insert all vm's */
3332 spin_lock(&vmap_area_lock);
3333 for (area = 0; area < nr_vms; area++) {
3334 insert_vmap_area(vas[area], &vmap_area_root, &vmap_area_list);
3335
3336 setup_vmalloc_vm_locked(vms[area], vas[area], VM_ALLOC,
3337 pcpu_get_vm_areas);
3338 }
3339 spin_unlock(&vmap_area_lock);
3340
3341 kfree(vas);
3342 return vms;
3343
3344recovery:
3345 /*
3346 * Remove previously allocated areas. There is no
3347 * need in removing these areas from the busy tree,
3348 * because they are inserted only on the final step
3349 * and when pcpu_get_vm_areas() is success.
3350 */
3351 while (area--) {
3352 orig_start = vas[area]->va_start;
3353 orig_end = vas[area]->va_end;
3354 va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3355 &free_vmap_area_list);
3356 if (va)
3357 kasan_release_vmalloc(orig_start, orig_end,
3358 va->va_start, va->va_end);
3359 vas[area] = NULL;
3360 }
3361
3362overflow:
3363 spin_unlock(&free_vmap_area_lock);
3364 if (!purged) {
3365 purge_vmap_area_lazy();
3366 purged = true;
3367
3368 /* Before "retry", check if we recover. */
3369 for (area = 0; area < nr_vms; area++) {
3370 if (vas[area])
3371 continue;
3372
3373 vas[area] = kmem_cache_zalloc(
3374 vmap_area_cachep, GFP_KERNEL);
3375 if (!vas[area])
3376 goto err_free;
3377 }
3378
3379 goto retry;
3380 }
3381
3382err_free:
3383 for (area = 0; area < nr_vms; area++) {
3384 if (vas[area])
3385 kmem_cache_free(vmap_area_cachep, vas[area]);
3386
3387 kfree(vms[area]);
3388 }
3389err_free2:
3390 kfree(vas);
3391 kfree(vms);
3392 return NULL;
3393
3394err_free_shadow:
3395 spin_lock(&free_vmap_area_lock);
3396 /*
3397 * We release all the vmalloc shadows, even the ones for regions that
3398 * hadn't been successfully added. This relies on kasan_release_vmalloc
3399 * being able to tolerate this case.
3400 */
3401 for (area = 0; area < nr_vms; area++) {
3402 orig_start = vas[area]->va_start;
3403 orig_end = vas[area]->va_end;
3404 va = merge_or_add_vmap_area(vas[area], &free_vmap_area_root,
3405 &free_vmap_area_list);
3406 if (va)
3407 kasan_release_vmalloc(orig_start, orig_end,
3408 va->va_start, va->va_end);
3409 vas[area] = NULL;
3410 kfree(vms[area]);
3411 }
3412 spin_unlock(&free_vmap_area_lock);
3413 kfree(vas);
3414 kfree(vms);
3415 return NULL;
3416}
3417
3418/**
3419 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3420 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3421 * @nr_vms: the number of allocated areas
3422 *
3423 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3424 */
3425void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3426{
3427 int i;
3428
3429 for (i = 0; i < nr_vms; i++)
3430 free_vm_area(vms[i]);
3431 kfree(vms);
3432}
3433#endif /* CONFIG_SMP */
3434
3435#ifdef CONFIG_PROC_FS
3436static void *s_start(struct seq_file *m, loff_t *pos)
3437 __acquires(&vmap_purge_lock)
3438 __acquires(&vmap_area_lock)
3439{
3440 mutex_lock(&vmap_purge_lock);
3441 spin_lock(&vmap_area_lock);
3442
3443 return seq_list_start(&vmap_area_list, *pos);
3444}
3445
3446static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3447{
3448 return seq_list_next(p, &vmap_area_list, pos);
3449}
3450
3451static void s_stop(struct seq_file *m, void *p)
3452 __releases(&vmap_purge_lock)
3453 __releases(&vmap_area_lock)
3454{
3455 mutex_unlock(&vmap_purge_lock);
3456 spin_unlock(&vmap_area_lock);
3457}
3458
3459static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3460{
3461 if (IS_ENABLED(CONFIG_NUMA)) {
3462 unsigned int nr, *counters = m->private;
3463
3464 if (!counters)
3465 return;
3466
3467 if (v->flags & VM_UNINITIALIZED)
3468 return;
3469 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3470 smp_rmb();
3471
3472 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3473
3474 for (nr = 0; nr < v->nr_pages; nr++)
3475 counters[page_to_nid(v->pages[nr])]++;
3476
3477 for_each_node_state(nr, N_HIGH_MEMORY)
3478 if (counters[nr])
3479 seq_printf(m, " N%u=%u", nr, counters[nr]);
3480 }
3481}
3482
3483static void show_purge_info(struct seq_file *m)
3484{
3485 struct llist_node *head;
3486 struct vmap_area *va;
3487
3488 head = READ_ONCE(vmap_purge_list.first);
3489 if (head == NULL)
3490 return;
3491
3492 llist_for_each_entry(va, head, purge_list) {
3493 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3494 (void *)va->va_start, (void *)va->va_end,
3495 va->va_end - va->va_start);
3496 }
3497}
3498
3499static int s_show(struct seq_file *m, void *p)
3500{
3501 struct vmap_area *va;
3502 struct vm_struct *v;
3503
3504 va = list_entry(p, struct vmap_area, list);
3505
3506 /*
3507 * s_show can encounter race with remove_vm_area, !vm on behalf
3508 * of vmap area is being tear down or vm_map_ram allocation.
3509 */
3510 if (!va->vm) {
3511 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3512 (void *)va->va_start, (void *)va->va_end,
3513 va->va_end - va->va_start);
3514
3515 return 0;
3516 }
3517
3518 v = va->vm;
3519
3520 seq_printf(m, "0x%pK-0x%pK %7ld",
3521 v->addr, v->addr + v->size, v->size);
3522
3523 if (v->caller)
3524 seq_printf(m, " %pS", v->caller);
3525
3526 if (v->nr_pages)
3527 seq_printf(m, " pages=%d", v->nr_pages);
3528
3529 if (v->phys_addr)
3530 seq_printf(m, " phys=%pa", &v->phys_addr);
3531
3532 if (v->flags & VM_IOREMAP)
3533 seq_puts(m, " ioremap");
3534
3535 if (v->flags & VM_ALLOC)
3536 seq_puts(m, " vmalloc");
3537
3538 if (v->flags & VM_MAP)
3539 seq_puts(m, " vmap");
3540
3541 if (v->flags & VM_USERMAP)
3542 seq_puts(m, " user");
3543
3544 if (v->flags & VM_DMA_COHERENT)
3545 seq_puts(m, " dma-coherent");
3546
3547 if (is_vmalloc_addr(v->pages))
3548 seq_puts(m, " vpages");
3549
3550 show_numa_info(m, v);
3551 seq_putc(m, '\n');
3552
3553 /*
3554 * As a final step, dump "unpurged" areas. Note,
3555 * that entire "/proc/vmallocinfo" output will not
3556 * be address sorted, because the purge list is not
3557 * sorted.
3558 */
3559 if (list_is_last(&va->list, &vmap_area_list))
3560 show_purge_info(m);
3561
3562 return 0;
3563}
3564
3565static const struct seq_operations vmalloc_op = {
3566 .start = s_start,
3567 .next = s_next,
3568 .stop = s_stop,
3569 .show = s_show,
3570};
3571
3572static int __init proc_vmalloc_init(void)
3573{
3574 if (IS_ENABLED(CONFIG_NUMA))
3575 proc_create_seq_private("vmallocinfo", 0400, NULL,
3576 &vmalloc_op,
3577 nr_node_ids * sizeof(unsigned int), NULL);
3578 else
3579 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3580 return 0;
3581}
3582module_init(proc_vmalloc_init);
3583
3584#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/vmalloc.c
4 *
5 * Copyright (C) 1993 Linus Torvalds
6 * Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
7 * SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
8 * Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
9 * Numa awareness, Christoph Lameter, SGI, June 2005
10 */
11
12#include <linux/vmalloc.h>
13#include <linux/mm.h>
14#include <linux/module.h>
15#include <linux/highmem.h>
16#include <linux/sched/signal.h>
17#include <linux/slab.h>
18#include <linux/spinlock.h>
19#include <linux/interrupt.h>
20#include <linux/proc_fs.h>
21#include <linux/seq_file.h>
22#include <linux/set_memory.h>
23#include <linux/debugobjects.h>
24#include <linux/kallsyms.h>
25#include <linux/list.h>
26#include <linux/notifier.h>
27#include <linux/rbtree.h>
28#include <linux/radix-tree.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/llist.h>
35#include <linux/bitops.h>
36#include <linux/rbtree_augmented.h>
37
38#include <linux/uaccess.h>
39#include <asm/tlbflush.h>
40#include <asm/shmparam.h>
41
42#include "internal.h"
43
44struct vfree_deferred {
45 struct llist_head list;
46 struct work_struct wq;
47};
48static DEFINE_PER_CPU(struct vfree_deferred, vfree_deferred);
49
50static void __vunmap(const void *, int);
51
52static void free_work(struct work_struct *w)
53{
54 struct vfree_deferred *p = container_of(w, struct vfree_deferred, wq);
55 struct llist_node *t, *llnode;
56
57 llist_for_each_safe(llnode, t, llist_del_all(&p->list))
58 __vunmap((void *)llnode, 1);
59}
60
61/*** Page table manipulation functions ***/
62
63static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
64{
65 pte_t *pte;
66
67 pte = pte_offset_kernel(pmd, addr);
68 do {
69 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
70 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
71 } while (pte++, addr += PAGE_SIZE, addr != end);
72}
73
74static void vunmap_pmd_range(pud_t *pud, unsigned long addr, unsigned long end)
75{
76 pmd_t *pmd;
77 unsigned long next;
78
79 pmd = pmd_offset(pud, addr);
80 do {
81 next = pmd_addr_end(addr, end);
82 if (pmd_clear_huge(pmd))
83 continue;
84 if (pmd_none_or_clear_bad(pmd))
85 continue;
86 vunmap_pte_range(pmd, addr, next);
87 } while (pmd++, addr = next, addr != end);
88}
89
90static void vunmap_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end)
91{
92 pud_t *pud;
93 unsigned long next;
94
95 pud = pud_offset(p4d, addr);
96 do {
97 next = pud_addr_end(addr, end);
98 if (pud_clear_huge(pud))
99 continue;
100 if (pud_none_or_clear_bad(pud))
101 continue;
102 vunmap_pmd_range(pud, addr, next);
103 } while (pud++, addr = next, addr != end);
104}
105
106static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end)
107{
108 p4d_t *p4d;
109 unsigned long next;
110
111 p4d = p4d_offset(pgd, addr);
112 do {
113 next = p4d_addr_end(addr, end);
114 if (p4d_clear_huge(p4d))
115 continue;
116 if (p4d_none_or_clear_bad(p4d))
117 continue;
118 vunmap_pud_range(p4d, addr, next);
119 } while (p4d++, addr = next, addr != end);
120}
121
122static void vunmap_page_range(unsigned long addr, unsigned long end)
123{
124 pgd_t *pgd;
125 unsigned long next;
126
127 BUG_ON(addr >= end);
128 pgd = pgd_offset_k(addr);
129 do {
130 next = pgd_addr_end(addr, end);
131 if (pgd_none_or_clear_bad(pgd))
132 continue;
133 vunmap_p4d_range(pgd, addr, next);
134 } while (pgd++, addr = next, addr != end);
135}
136
137static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
138 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
139{
140 pte_t *pte;
141
142 /*
143 * nr is a running index into the array which helps higher level
144 * callers keep track of where we're up to.
145 */
146
147 pte = pte_alloc_kernel(pmd, addr);
148 if (!pte)
149 return -ENOMEM;
150 do {
151 struct page *page = pages[*nr];
152
153 if (WARN_ON(!pte_none(*pte)))
154 return -EBUSY;
155 if (WARN_ON(!page))
156 return -ENOMEM;
157 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
158 (*nr)++;
159 } while (pte++, addr += PAGE_SIZE, addr != end);
160 return 0;
161}
162
163static int vmap_pmd_range(pud_t *pud, unsigned long addr,
164 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
165{
166 pmd_t *pmd;
167 unsigned long next;
168
169 pmd = pmd_alloc(&init_mm, pud, addr);
170 if (!pmd)
171 return -ENOMEM;
172 do {
173 next = pmd_addr_end(addr, end);
174 if (vmap_pte_range(pmd, addr, next, prot, pages, nr))
175 return -ENOMEM;
176 } while (pmd++, addr = next, addr != end);
177 return 0;
178}
179
180static int vmap_pud_range(p4d_t *p4d, unsigned long addr,
181 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
182{
183 pud_t *pud;
184 unsigned long next;
185
186 pud = pud_alloc(&init_mm, p4d, addr);
187 if (!pud)
188 return -ENOMEM;
189 do {
190 next = pud_addr_end(addr, end);
191 if (vmap_pmd_range(pud, addr, next, prot, pages, nr))
192 return -ENOMEM;
193 } while (pud++, addr = next, addr != end);
194 return 0;
195}
196
197static int vmap_p4d_range(pgd_t *pgd, unsigned long addr,
198 unsigned long end, pgprot_t prot, struct page **pages, int *nr)
199{
200 p4d_t *p4d;
201 unsigned long next;
202
203 p4d = p4d_alloc(&init_mm, pgd, addr);
204 if (!p4d)
205 return -ENOMEM;
206 do {
207 next = p4d_addr_end(addr, end);
208 if (vmap_pud_range(p4d, addr, next, prot, pages, nr))
209 return -ENOMEM;
210 } while (p4d++, addr = next, addr != end);
211 return 0;
212}
213
214/*
215 * Set up page tables in kva (addr, end). The ptes shall have prot "prot", and
216 * will have pfns corresponding to the "pages" array.
217 *
218 * Ie. pte at addr+N*PAGE_SIZE shall point to pfn corresponding to pages[N]
219 */
220static int vmap_page_range_noflush(unsigned long start, unsigned long end,
221 pgprot_t prot, struct page **pages)
222{
223 pgd_t *pgd;
224 unsigned long next;
225 unsigned long addr = start;
226 int err = 0;
227 int nr = 0;
228
229 BUG_ON(addr >= end);
230 pgd = pgd_offset_k(addr);
231 do {
232 next = pgd_addr_end(addr, end);
233 err = vmap_p4d_range(pgd, addr, next, prot, pages, &nr);
234 if (err)
235 return err;
236 } while (pgd++, addr = next, addr != end);
237
238 return nr;
239}
240
241static int vmap_page_range(unsigned long start, unsigned long end,
242 pgprot_t prot, struct page **pages)
243{
244 int ret;
245
246 ret = vmap_page_range_noflush(start, end, prot, pages);
247 flush_cache_vmap(start, end);
248 return ret;
249}
250
251int is_vmalloc_or_module_addr(const void *x)
252{
253 /*
254 * ARM, x86-64 and sparc64 put modules in a special place,
255 * and fall back on vmalloc() if that fails. Others
256 * just put it in the vmalloc space.
257 */
258#if defined(CONFIG_MODULES) && defined(MODULES_VADDR)
259 unsigned long addr = (unsigned long)x;
260 if (addr >= MODULES_VADDR && addr < MODULES_END)
261 return 1;
262#endif
263 return is_vmalloc_addr(x);
264}
265
266/*
267 * Walk a vmap address to the struct page it maps.
268 */
269struct page *vmalloc_to_page(const void *vmalloc_addr)
270{
271 unsigned long addr = (unsigned long) vmalloc_addr;
272 struct page *page = NULL;
273 pgd_t *pgd = pgd_offset_k(addr);
274 p4d_t *p4d;
275 pud_t *pud;
276 pmd_t *pmd;
277 pte_t *ptep, pte;
278
279 /*
280 * XXX we might need to change this if we add VIRTUAL_BUG_ON for
281 * architectures that do not vmalloc module space
282 */
283 VIRTUAL_BUG_ON(!is_vmalloc_or_module_addr(vmalloc_addr));
284
285 if (pgd_none(*pgd))
286 return NULL;
287 p4d = p4d_offset(pgd, addr);
288 if (p4d_none(*p4d))
289 return NULL;
290 pud = pud_offset(p4d, addr);
291
292 /*
293 * Don't dereference bad PUD or PMD (below) entries. This will also
294 * identify huge mappings, which we may encounter on architectures
295 * that define CONFIG_HAVE_ARCH_HUGE_VMAP=y. Such regions will be
296 * identified as vmalloc addresses by is_vmalloc_addr(), but are
297 * not [unambiguously] associated with a struct page, so there is
298 * no correct value to return for them.
299 */
300 WARN_ON_ONCE(pud_bad(*pud));
301 if (pud_none(*pud) || pud_bad(*pud))
302 return NULL;
303 pmd = pmd_offset(pud, addr);
304 WARN_ON_ONCE(pmd_bad(*pmd));
305 if (pmd_none(*pmd) || pmd_bad(*pmd))
306 return NULL;
307
308 ptep = pte_offset_map(pmd, addr);
309 pte = *ptep;
310 if (pte_present(pte))
311 page = pte_page(pte);
312 pte_unmap(ptep);
313 return page;
314}
315EXPORT_SYMBOL(vmalloc_to_page);
316
317/*
318 * Map a vmalloc()-space virtual address to the physical page frame number.
319 */
320unsigned long vmalloc_to_pfn(const void *vmalloc_addr)
321{
322 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
323}
324EXPORT_SYMBOL(vmalloc_to_pfn);
325
326
327/*** Global kva allocator ***/
328
329#define DEBUG_AUGMENT_PROPAGATE_CHECK 0
330#define DEBUG_AUGMENT_LOWEST_MATCH_CHECK 0
331
332
333static DEFINE_SPINLOCK(vmap_area_lock);
334/* Export for kexec only */
335LIST_HEAD(vmap_area_list);
336static LLIST_HEAD(vmap_purge_list);
337static struct rb_root vmap_area_root = RB_ROOT;
338static bool vmap_initialized __read_mostly;
339
340/*
341 * This kmem_cache is used for vmap_area objects. Instead of
342 * allocating from slab we reuse an object from this cache to
343 * make things faster. Especially in "no edge" splitting of
344 * free block.
345 */
346static struct kmem_cache *vmap_area_cachep;
347
348/*
349 * This linked list is used in pair with free_vmap_area_root.
350 * It gives O(1) access to prev/next to perform fast coalescing.
351 */
352static LIST_HEAD(free_vmap_area_list);
353
354/*
355 * This augment red-black tree represents the free vmap space.
356 * All vmap_area objects in this tree are sorted by va->va_start
357 * address. It is used for allocation and merging when a vmap
358 * object is released.
359 *
360 * Each vmap_area node contains a maximum available free block
361 * of its sub-tree, right or left. Therefore it is possible to
362 * find a lowest match of free area.
363 */
364static struct rb_root free_vmap_area_root = RB_ROOT;
365
366/*
367 * Preload a CPU with one object for "no edge" split case. The
368 * aim is to get rid of allocations from the atomic context, thus
369 * to use more permissive allocation masks.
370 */
371static DEFINE_PER_CPU(struct vmap_area *, ne_fit_preload_node);
372
373static __always_inline unsigned long
374va_size(struct vmap_area *va)
375{
376 return (va->va_end - va->va_start);
377}
378
379static __always_inline unsigned long
380get_subtree_max_size(struct rb_node *node)
381{
382 struct vmap_area *va;
383
384 va = rb_entry_safe(node, struct vmap_area, rb_node);
385 return va ? va->subtree_max_size : 0;
386}
387
388/*
389 * Gets called when remove the node and rotate.
390 */
391static __always_inline unsigned long
392compute_subtree_max_size(struct vmap_area *va)
393{
394 return max3(va_size(va),
395 get_subtree_max_size(va->rb_node.rb_left),
396 get_subtree_max_size(va->rb_node.rb_right));
397}
398
399RB_DECLARE_CALLBACKS_MAX(static, free_vmap_area_rb_augment_cb,
400 struct vmap_area, rb_node, unsigned long, subtree_max_size, va_size)
401
402static void purge_vmap_area_lazy(void);
403static BLOCKING_NOTIFIER_HEAD(vmap_notify_list);
404static unsigned long lazy_max_pages(void);
405
406static atomic_long_t nr_vmalloc_pages;
407
408unsigned long vmalloc_nr_pages(void)
409{
410 return atomic_long_read(&nr_vmalloc_pages);
411}
412
413static struct vmap_area *__find_vmap_area(unsigned long addr)
414{
415 struct rb_node *n = vmap_area_root.rb_node;
416
417 while (n) {
418 struct vmap_area *va;
419
420 va = rb_entry(n, struct vmap_area, rb_node);
421 if (addr < va->va_start)
422 n = n->rb_left;
423 else if (addr >= va->va_end)
424 n = n->rb_right;
425 else
426 return va;
427 }
428
429 return NULL;
430}
431
432/*
433 * This function returns back addresses of parent node
434 * and its left or right link for further processing.
435 */
436static __always_inline struct rb_node **
437find_va_links(struct vmap_area *va,
438 struct rb_root *root, struct rb_node *from,
439 struct rb_node **parent)
440{
441 struct vmap_area *tmp_va;
442 struct rb_node **link;
443
444 if (root) {
445 link = &root->rb_node;
446 if (unlikely(!*link)) {
447 *parent = NULL;
448 return link;
449 }
450 } else {
451 link = &from;
452 }
453
454 /*
455 * Go to the bottom of the tree. When we hit the last point
456 * we end up with parent rb_node and correct direction, i name
457 * it link, where the new va->rb_node will be attached to.
458 */
459 do {
460 tmp_va = rb_entry(*link, struct vmap_area, rb_node);
461
462 /*
463 * During the traversal we also do some sanity check.
464 * Trigger the BUG() if there are sides(left/right)
465 * or full overlaps.
466 */
467 if (va->va_start < tmp_va->va_end &&
468 va->va_end <= tmp_va->va_start)
469 link = &(*link)->rb_left;
470 else if (va->va_end > tmp_va->va_start &&
471 va->va_start >= tmp_va->va_end)
472 link = &(*link)->rb_right;
473 else
474 BUG();
475 } while (*link);
476
477 *parent = &tmp_va->rb_node;
478 return link;
479}
480
481static __always_inline struct list_head *
482get_va_next_sibling(struct rb_node *parent, struct rb_node **link)
483{
484 struct list_head *list;
485
486 if (unlikely(!parent))
487 /*
488 * The red-black tree where we try to find VA neighbors
489 * before merging or inserting is empty, i.e. it means
490 * there is no free vmap space. Normally it does not
491 * happen but we handle this case anyway.
492 */
493 return NULL;
494
495 list = &rb_entry(parent, struct vmap_area, rb_node)->list;
496 return (&parent->rb_right == link ? list->next : list);
497}
498
499static __always_inline void
500link_va(struct vmap_area *va, struct rb_root *root,
501 struct rb_node *parent, struct rb_node **link, struct list_head *head)
502{
503 /*
504 * VA is still not in the list, but we can
505 * identify its future previous list_head node.
506 */
507 if (likely(parent)) {
508 head = &rb_entry(parent, struct vmap_area, rb_node)->list;
509 if (&parent->rb_right != link)
510 head = head->prev;
511 }
512
513 /* Insert to the rb-tree */
514 rb_link_node(&va->rb_node, parent, link);
515 if (root == &free_vmap_area_root) {
516 /*
517 * Some explanation here. Just perform simple insertion
518 * to the tree. We do not set va->subtree_max_size to
519 * its current size before calling rb_insert_augmented().
520 * It is because of we populate the tree from the bottom
521 * to parent levels when the node _is_ in the tree.
522 *
523 * Therefore we set subtree_max_size to zero after insertion,
524 * to let __augment_tree_propagate_from() puts everything to
525 * the correct order later on.
526 */
527 rb_insert_augmented(&va->rb_node,
528 root, &free_vmap_area_rb_augment_cb);
529 va->subtree_max_size = 0;
530 } else {
531 rb_insert_color(&va->rb_node, root);
532 }
533
534 /* Address-sort this list */
535 list_add(&va->list, head);
536}
537
538static __always_inline void
539unlink_va(struct vmap_area *va, struct rb_root *root)
540{
541 if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
542 return;
543
544 if (root == &free_vmap_area_root)
545 rb_erase_augmented(&va->rb_node,
546 root, &free_vmap_area_rb_augment_cb);
547 else
548 rb_erase(&va->rb_node, root);
549
550 list_del(&va->list);
551 RB_CLEAR_NODE(&va->rb_node);
552}
553
554#if DEBUG_AUGMENT_PROPAGATE_CHECK
555static void
556augment_tree_propagate_check(struct rb_node *n)
557{
558 struct vmap_area *va;
559 struct rb_node *node;
560 unsigned long size;
561 bool found = false;
562
563 if (n == NULL)
564 return;
565
566 va = rb_entry(n, struct vmap_area, rb_node);
567 size = va->subtree_max_size;
568 node = n;
569
570 while (node) {
571 va = rb_entry(node, struct vmap_area, rb_node);
572
573 if (get_subtree_max_size(node->rb_left) == size) {
574 node = node->rb_left;
575 } else {
576 if (va_size(va) == size) {
577 found = true;
578 break;
579 }
580
581 node = node->rb_right;
582 }
583 }
584
585 if (!found) {
586 va = rb_entry(n, struct vmap_area, rb_node);
587 pr_emerg("tree is corrupted: %lu, %lu\n",
588 va_size(va), va->subtree_max_size);
589 }
590
591 augment_tree_propagate_check(n->rb_left);
592 augment_tree_propagate_check(n->rb_right);
593}
594#endif
595
596/*
597 * This function populates subtree_max_size from bottom to upper
598 * levels starting from VA point. The propagation must be done
599 * when VA size is modified by changing its va_start/va_end. Or
600 * in case of newly inserting of VA to the tree.
601 *
602 * It means that __augment_tree_propagate_from() must be called:
603 * - After VA has been inserted to the tree(free path);
604 * - After VA has been shrunk(allocation path);
605 * - After VA has been increased(merging path).
606 *
607 * Please note that, it does not mean that upper parent nodes
608 * and their subtree_max_size are recalculated all the time up
609 * to the root node.
610 *
611 * 4--8
612 * /\
613 * / \
614 * / \
615 * 2--2 8--8
616 *
617 * For example if we modify the node 4, shrinking it to 2, then
618 * no any modification is required. If we shrink the node 2 to 1
619 * its subtree_max_size is updated only, and set to 1. If we shrink
620 * the node 8 to 6, then its subtree_max_size is set to 6 and parent
621 * node becomes 4--6.
622 */
623static __always_inline void
624augment_tree_propagate_from(struct vmap_area *va)
625{
626 struct rb_node *node = &va->rb_node;
627 unsigned long new_va_sub_max_size;
628
629 while (node) {
630 va = rb_entry(node, struct vmap_area, rb_node);
631 new_va_sub_max_size = compute_subtree_max_size(va);
632
633 /*
634 * If the newly calculated maximum available size of the
635 * subtree is equal to the current one, then it means that
636 * the tree is propagated correctly. So we have to stop at
637 * this point to save cycles.
638 */
639 if (va->subtree_max_size == new_va_sub_max_size)
640 break;
641
642 va->subtree_max_size = new_va_sub_max_size;
643 node = rb_parent(&va->rb_node);
644 }
645
646#if DEBUG_AUGMENT_PROPAGATE_CHECK
647 augment_tree_propagate_check(free_vmap_area_root.rb_node);
648#endif
649}
650
651static void
652insert_vmap_area(struct vmap_area *va,
653 struct rb_root *root, struct list_head *head)
654{
655 struct rb_node **link;
656 struct rb_node *parent;
657
658 link = find_va_links(va, root, NULL, &parent);
659 link_va(va, root, parent, link, head);
660}
661
662static void
663insert_vmap_area_augment(struct vmap_area *va,
664 struct rb_node *from, struct rb_root *root,
665 struct list_head *head)
666{
667 struct rb_node **link;
668 struct rb_node *parent;
669
670 if (from)
671 link = find_va_links(va, NULL, from, &parent);
672 else
673 link = find_va_links(va, root, NULL, &parent);
674
675 link_va(va, root, parent, link, head);
676 augment_tree_propagate_from(va);
677}
678
679/*
680 * Merge de-allocated chunk of VA memory with previous
681 * and next free blocks. If coalesce is not done a new
682 * free area is inserted. If VA has been merged, it is
683 * freed.
684 */
685static __always_inline void
686merge_or_add_vmap_area(struct vmap_area *va,
687 struct rb_root *root, struct list_head *head)
688{
689 struct vmap_area *sibling;
690 struct list_head *next;
691 struct rb_node **link;
692 struct rb_node *parent;
693 bool merged = false;
694
695 /*
696 * Find a place in the tree where VA potentially will be
697 * inserted, unless it is merged with its sibling/siblings.
698 */
699 link = find_va_links(va, root, NULL, &parent);
700
701 /*
702 * Get next node of VA to check if merging can be done.
703 */
704 next = get_va_next_sibling(parent, link);
705 if (unlikely(next == NULL))
706 goto insert;
707
708 /*
709 * start end
710 * | |
711 * |<------VA------>|<-----Next----->|
712 * | |
713 * start end
714 */
715 if (next != head) {
716 sibling = list_entry(next, struct vmap_area, list);
717 if (sibling->va_start == va->va_end) {
718 sibling->va_start = va->va_start;
719
720 /* Check and update the tree if needed. */
721 augment_tree_propagate_from(sibling);
722
723 /* Free vmap_area object. */
724 kmem_cache_free(vmap_area_cachep, va);
725
726 /* Point to the new merged area. */
727 va = sibling;
728 merged = true;
729 }
730 }
731
732 /*
733 * start end
734 * | |
735 * |<-----Prev----->|<------VA------>|
736 * | |
737 * start end
738 */
739 if (next->prev != head) {
740 sibling = list_entry(next->prev, struct vmap_area, list);
741 if (sibling->va_end == va->va_start) {
742 sibling->va_end = va->va_end;
743
744 /* Check and update the tree if needed. */
745 augment_tree_propagate_from(sibling);
746
747 if (merged)
748 unlink_va(va, root);
749
750 /* Free vmap_area object. */
751 kmem_cache_free(vmap_area_cachep, va);
752 return;
753 }
754 }
755
756insert:
757 if (!merged) {
758 link_va(va, root, parent, link, head);
759 augment_tree_propagate_from(va);
760 }
761}
762
763static __always_inline bool
764is_within_this_va(struct vmap_area *va, unsigned long size,
765 unsigned long align, unsigned long vstart)
766{
767 unsigned long nva_start_addr;
768
769 if (va->va_start > vstart)
770 nva_start_addr = ALIGN(va->va_start, align);
771 else
772 nva_start_addr = ALIGN(vstart, align);
773
774 /* Can be overflowed due to big size or alignment. */
775 if (nva_start_addr + size < nva_start_addr ||
776 nva_start_addr < vstart)
777 return false;
778
779 return (nva_start_addr + size <= va->va_end);
780}
781
782/*
783 * Find the first free block(lowest start address) in the tree,
784 * that will accomplish the request corresponding to passing
785 * parameters.
786 */
787static __always_inline struct vmap_area *
788find_vmap_lowest_match(unsigned long size,
789 unsigned long align, unsigned long vstart)
790{
791 struct vmap_area *va;
792 struct rb_node *node;
793 unsigned long length;
794
795 /* Start from the root. */
796 node = free_vmap_area_root.rb_node;
797
798 /* Adjust the search size for alignment overhead. */
799 length = size + align - 1;
800
801 while (node) {
802 va = rb_entry(node, struct vmap_area, rb_node);
803
804 if (get_subtree_max_size(node->rb_left) >= length &&
805 vstart < va->va_start) {
806 node = node->rb_left;
807 } else {
808 if (is_within_this_va(va, size, align, vstart))
809 return va;
810
811 /*
812 * Does not make sense to go deeper towards the right
813 * sub-tree if it does not have a free block that is
814 * equal or bigger to the requested search length.
815 */
816 if (get_subtree_max_size(node->rb_right) >= length) {
817 node = node->rb_right;
818 continue;
819 }
820
821 /*
822 * OK. We roll back and find the first right sub-tree,
823 * that will satisfy the search criteria. It can happen
824 * only once due to "vstart" restriction.
825 */
826 while ((node = rb_parent(node))) {
827 va = rb_entry(node, struct vmap_area, rb_node);
828 if (is_within_this_va(va, size, align, vstart))
829 return va;
830
831 if (get_subtree_max_size(node->rb_right) >= length &&
832 vstart <= va->va_start) {
833 node = node->rb_right;
834 break;
835 }
836 }
837 }
838 }
839
840 return NULL;
841}
842
843#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
844#include <linux/random.h>
845
846static struct vmap_area *
847find_vmap_lowest_linear_match(unsigned long size,
848 unsigned long align, unsigned long vstart)
849{
850 struct vmap_area *va;
851
852 list_for_each_entry(va, &free_vmap_area_list, list) {
853 if (!is_within_this_va(va, size, align, vstart))
854 continue;
855
856 return va;
857 }
858
859 return NULL;
860}
861
862static void
863find_vmap_lowest_match_check(unsigned long size)
864{
865 struct vmap_area *va_1, *va_2;
866 unsigned long vstart;
867 unsigned int rnd;
868
869 get_random_bytes(&rnd, sizeof(rnd));
870 vstart = VMALLOC_START + rnd;
871
872 va_1 = find_vmap_lowest_match(size, 1, vstart);
873 va_2 = find_vmap_lowest_linear_match(size, 1, vstart);
874
875 if (va_1 != va_2)
876 pr_emerg("not lowest: t: 0x%p, l: 0x%p, v: 0x%lx\n",
877 va_1, va_2, vstart);
878}
879#endif
880
881enum fit_type {
882 NOTHING_FIT = 0,
883 FL_FIT_TYPE = 1, /* full fit */
884 LE_FIT_TYPE = 2, /* left edge fit */
885 RE_FIT_TYPE = 3, /* right edge fit */
886 NE_FIT_TYPE = 4 /* no edge fit */
887};
888
889static __always_inline enum fit_type
890classify_va_fit_type(struct vmap_area *va,
891 unsigned long nva_start_addr, unsigned long size)
892{
893 enum fit_type type;
894
895 /* Check if it is within VA. */
896 if (nva_start_addr < va->va_start ||
897 nva_start_addr + size > va->va_end)
898 return NOTHING_FIT;
899
900 /* Now classify. */
901 if (va->va_start == nva_start_addr) {
902 if (va->va_end == nva_start_addr + size)
903 type = FL_FIT_TYPE;
904 else
905 type = LE_FIT_TYPE;
906 } else if (va->va_end == nva_start_addr + size) {
907 type = RE_FIT_TYPE;
908 } else {
909 type = NE_FIT_TYPE;
910 }
911
912 return type;
913}
914
915static __always_inline int
916adjust_va_to_fit_type(struct vmap_area *va,
917 unsigned long nva_start_addr, unsigned long size,
918 enum fit_type type)
919{
920 struct vmap_area *lva = NULL;
921
922 if (type == FL_FIT_TYPE) {
923 /*
924 * No need to split VA, it fully fits.
925 *
926 * | |
927 * V NVA V
928 * |---------------|
929 */
930 unlink_va(va, &free_vmap_area_root);
931 kmem_cache_free(vmap_area_cachep, va);
932 } else if (type == LE_FIT_TYPE) {
933 /*
934 * Split left edge of fit VA.
935 *
936 * | |
937 * V NVA V R
938 * |-------|-------|
939 */
940 va->va_start += size;
941 } else if (type == RE_FIT_TYPE) {
942 /*
943 * Split right edge of fit VA.
944 *
945 * | |
946 * L V NVA V
947 * |-------|-------|
948 */
949 va->va_end = nva_start_addr;
950 } else if (type == NE_FIT_TYPE) {
951 /*
952 * Split no edge of fit VA.
953 *
954 * | |
955 * L V NVA V R
956 * |---|-------|---|
957 */
958 lva = __this_cpu_xchg(ne_fit_preload_node, NULL);
959 if (unlikely(!lva)) {
960 /*
961 * For percpu allocator we do not do any pre-allocation
962 * and leave it as it is. The reason is it most likely
963 * never ends up with NE_FIT_TYPE splitting. In case of
964 * percpu allocations offsets and sizes are aligned to
965 * fixed align request, i.e. RE_FIT_TYPE and FL_FIT_TYPE
966 * are its main fitting cases.
967 *
968 * There are a few exceptions though, as an example it is
969 * a first allocation (early boot up) when we have "one"
970 * big free space that has to be split.
971 */
972 lva = kmem_cache_alloc(vmap_area_cachep, GFP_NOWAIT);
973 if (!lva)
974 return -1;
975 }
976
977 /*
978 * Build the remainder.
979 */
980 lva->va_start = va->va_start;
981 lva->va_end = nva_start_addr;
982
983 /*
984 * Shrink this VA to remaining size.
985 */
986 va->va_start = nva_start_addr + size;
987 } else {
988 return -1;
989 }
990
991 if (type != FL_FIT_TYPE) {
992 augment_tree_propagate_from(va);
993
994 if (lva) /* type == NE_FIT_TYPE */
995 insert_vmap_area_augment(lva, &va->rb_node,
996 &free_vmap_area_root, &free_vmap_area_list);
997 }
998
999 return 0;
1000}
1001
1002/*
1003 * Returns a start address of the newly allocated area, if success.
1004 * Otherwise a vend is returned that indicates failure.
1005 */
1006static __always_inline unsigned long
1007__alloc_vmap_area(unsigned long size, unsigned long align,
1008 unsigned long vstart, unsigned long vend)
1009{
1010 unsigned long nva_start_addr;
1011 struct vmap_area *va;
1012 enum fit_type type;
1013 int ret;
1014
1015 va = find_vmap_lowest_match(size, align, vstart);
1016 if (unlikely(!va))
1017 return vend;
1018
1019 if (va->va_start > vstart)
1020 nva_start_addr = ALIGN(va->va_start, align);
1021 else
1022 nva_start_addr = ALIGN(vstart, align);
1023
1024 /* Check the "vend" restriction. */
1025 if (nva_start_addr + size > vend)
1026 return vend;
1027
1028 /* Classify what we have found. */
1029 type = classify_va_fit_type(va, nva_start_addr, size);
1030 if (WARN_ON_ONCE(type == NOTHING_FIT))
1031 return vend;
1032
1033 /* Update the free vmap_area. */
1034 ret = adjust_va_to_fit_type(va, nva_start_addr, size, type);
1035 if (ret)
1036 return vend;
1037
1038#if DEBUG_AUGMENT_LOWEST_MATCH_CHECK
1039 find_vmap_lowest_match_check(size);
1040#endif
1041
1042 return nva_start_addr;
1043}
1044
1045/*
1046 * Allocate a region of KVA of the specified size and alignment, within the
1047 * vstart and vend.
1048 */
1049static struct vmap_area *alloc_vmap_area(unsigned long size,
1050 unsigned long align,
1051 unsigned long vstart, unsigned long vend,
1052 int node, gfp_t gfp_mask)
1053{
1054 struct vmap_area *va, *pva;
1055 unsigned long addr;
1056 int purged = 0;
1057
1058 BUG_ON(!size);
1059 BUG_ON(offset_in_page(size));
1060 BUG_ON(!is_power_of_2(align));
1061
1062 if (unlikely(!vmap_initialized))
1063 return ERR_PTR(-EBUSY);
1064
1065 might_sleep();
1066
1067 va = kmem_cache_alloc_node(vmap_area_cachep,
1068 gfp_mask & GFP_RECLAIM_MASK, node);
1069 if (unlikely(!va))
1070 return ERR_PTR(-ENOMEM);
1071
1072 /*
1073 * Only scan the relevant parts containing pointers to other objects
1074 * to avoid false negatives.
1075 */
1076 kmemleak_scan_area(&va->rb_node, SIZE_MAX, gfp_mask & GFP_RECLAIM_MASK);
1077
1078retry:
1079 /*
1080 * Preload this CPU with one extra vmap_area object to ensure
1081 * that we have it available when fit type of free area is
1082 * NE_FIT_TYPE.
1083 *
1084 * The preload is done in non-atomic context, thus it allows us
1085 * to use more permissive allocation masks to be more stable under
1086 * low memory condition and high memory pressure.
1087 *
1088 * Even if it fails we do not really care about that. Just proceed
1089 * as it is. "overflow" path will refill the cache we allocate from.
1090 */
1091 preempt_disable();
1092 if (!__this_cpu_read(ne_fit_preload_node)) {
1093 preempt_enable();
1094 pva = kmem_cache_alloc_node(vmap_area_cachep, GFP_KERNEL, node);
1095 preempt_disable();
1096
1097 if (__this_cpu_cmpxchg(ne_fit_preload_node, NULL, pva)) {
1098 if (pva)
1099 kmem_cache_free(vmap_area_cachep, pva);
1100 }
1101 }
1102
1103 spin_lock(&vmap_area_lock);
1104 preempt_enable();
1105
1106 /*
1107 * If an allocation fails, the "vend" address is
1108 * returned. Therefore trigger the overflow path.
1109 */
1110 addr = __alloc_vmap_area(size, align, vstart, vend);
1111 if (unlikely(addr == vend))
1112 goto overflow;
1113
1114 va->va_start = addr;
1115 va->va_end = addr + size;
1116 va->vm = NULL;
1117 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1118
1119 spin_unlock(&vmap_area_lock);
1120
1121 BUG_ON(!IS_ALIGNED(va->va_start, align));
1122 BUG_ON(va->va_start < vstart);
1123 BUG_ON(va->va_end > vend);
1124
1125 return va;
1126
1127overflow:
1128 spin_unlock(&vmap_area_lock);
1129 if (!purged) {
1130 purge_vmap_area_lazy();
1131 purged = 1;
1132 goto retry;
1133 }
1134
1135 if (gfpflags_allow_blocking(gfp_mask)) {
1136 unsigned long freed = 0;
1137 blocking_notifier_call_chain(&vmap_notify_list, 0, &freed);
1138 if (freed > 0) {
1139 purged = 0;
1140 goto retry;
1141 }
1142 }
1143
1144 if (!(gfp_mask & __GFP_NOWARN) && printk_ratelimit())
1145 pr_warn("vmap allocation for size %lu failed: use vmalloc=<size> to increase size\n",
1146 size);
1147
1148 kmem_cache_free(vmap_area_cachep, va);
1149 return ERR_PTR(-EBUSY);
1150}
1151
1152int register_vmap_purge_notifier(struct notifier_block *nb)
1153{
1154 return blocking_notifier_chain_register(&vmap_notify_list, nb);
1155}
1156EXPORT_SYMBOL_GPL(register_vmap_purge_notifier);
1157
1158int unregister_vmap_purge_notifier(struct notifier_block *nb)
1159{
1160 return blocking_notifier_chain_unregister(&vmap_notify_list, nb);
1161}
1162EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
1163
1164static void __free_vmap_area(struct vmap_area *va)
1165{
1166 /*
1167 * Remove from the busy tree/list.
1168 */
1169 unlink_va(va, &vmap_area_root);
1170
1171 /*
1172 * Merge VA with its neighbors, otherwise just add it.
1173 */
1174 merge_or_add_vmap_area(va,
1175 &free_vmap_area_root, &free_vmap_area_list);
1176}
1177
1178/*
1179 * Free a region of KVA allocated by alloc_vmap_area
1180 */
1181static void free_vmap_area(struct vmap_area *va)
1182{
1183 spin_lock(&vmap_area_lock);
1184 __free_vmap_area(va);
1185 spin_unlock(&vmap_area_lock);
1186}
1187
1188/*
1189 * Clear the pagetable entries of a given vmap_area
1190 */
1191static void unmap_vmap_area(struct vmap_area *va)
1192{
1193 vunmap_page_range(va->va_start, va->va_end);
1194}
1195
1196/*
1197 * lazy_max_pages is the maximum amount of virtual address space we gather up
1198 * before attempting to purge with a TLB flush.
1199 *
1200 * There is a tradeoff here: a larger number will cover more kernel page tables
1201 * and take slightly longer to purge, but it will linearly reduce the number of
1202 * global TLB flushes that must be performed. It would seem natural to scale
1203 * this number up linearly with the number of CPUs (because vmapping activity
1204 * could also scale linearly with the number of CPUs), however it is likely
1205 * that in practice, workloads might be constrained in other ways that mean
1206 * vmap activity will not scale linearly with CPUs. Also, I want to be
1207 * conservative and not introduce a big latency on huge systems, so go with
1208 * a less aggressive log scale. It will still be an improvement over the old
1209 * code, and it will be simple to change the scale factor if we find that it
1210 * becomes a problem on bigger systems.
1211 */
1212static unsigned long lazy_max_pages(void)
1213{
1214 unsigned int log;
1215
1216 log = fls(num_online_cpus());
1217
1218 return log * (32UL * 1024 * 1024 / PAGE_SIZE);
1219}
1220
1221static atomic_long_t vmap_lazy_nr = ATOMIC_LONG_INIT(0);
1222
1223/*
1224 * Serialize vmap purging. There is no actual criticial section protected
1225 * by this look, but we want to avoid concurrent calls for performance
1226 * reasons and to make the pcpu_get_vm_areas more deterministic.
1227 */
1228static DEFINE_MUTEX(vmap_purge_lock);
1229
1230/* for per-CPU blocks */
1231static void purge_fragmented_blocks_allcpus(void);
1232
1233/*
1234 * called before a call to iounmap() if the caller wants vm_area_struct's
1235 * immediately freed.
1236 */
1237void set_iounmap_nonlazy(void)
1238{
1239 atomic_long_set(&vmap_lazy_nr, lazy_max_pages()+1);
1240}
1241
1242/*
1243 * Purges all lazily-freed vmap areas.
1244 */
1245static bool __purge_vmap_area_lazy(unsigned long start, unsigned long end)
1246{
1247 unsigned long resched_threshold;
1248 struct llist_node *valist;
1249 struct vmap_area *va;
1250 struct vmap_area *n_va;
1251
1252 lockdep_assert_held(&vmap_purge_lock);
1253
1254 valist = llist_del_all(&vmap_purge_list);
1255 if (unlikely(valist == NULL))
1256 return false;
1257
1258 /*
1259 * First make sure the mappings are removed from all page-tables
1260 * before they are freed.
1261 */
1262 vmalloc_sync_all();
1263
1264 /*
1265 * TODO: to calculate a flush range without looping.
1266 * The list can be up to lazy_max_pages() elements.
1267 */
1268 llist_for_each_entry(va, valist, purge_list) {
1269 if (va->va_start < start)
1270 start = va->va_start;
1271 if (va->va_end > end)
1272 end = va->va_end;
1273 }
1274
1275 flush_tlb_kernel_range(start, end);
1276 resched_threshold = lazy_max_pages() << 1;
1277
1278 spin_lock(&vmap_area_lock);
1279 llist_for_each_entry_safe(va, n_va, valist, purge_list) {
1280 unsigned long nr = (va->va_end - va->va_start) >> PAGE_SHIFT;
1281
1282 /*
1283 * Finally insert or merge lazily-freed area. It is
1284 * detached and there is no need to "unlink" it from
1285 * anything.
1286 */
1287 merge_or_add_vmap_area(va,
1288 &free_vmap_area_root, &free_vmap_area_list);
1289
1290 atomic_long_sub(nr, &vmap_lazy_nr);
1291
1292 if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
1293 cond_resched_lock(&vmap_area_lock);
1294 }
1295 spin_unlock(&vmap_area_lock);
1296 return true;
1297}
1298
1299/*
1300 * Kick off a purge of the outstanding lazy areas. Don't bother if somebody
1301 * is already purging.
1302 */
1303static void try_purge_vmap_area_lazy(void)
1304{
1305 if (mutex_trylock(&vmap_purge_lock)) {
1306 __purge_vmap_area_lazy(ULONG_MAX, 0);
1307 mutex_unlock(&vmap_purge_lock);
1308 }
1309}
1310
1311/*
1312 * Kick off a purge of the outstanding lazy areas.
1313 */
1314static void purge_vmap_area_lazy(void)
1315{
1316 mutex_lock(&vmap_purge_lock);
1317 purge_fragmented_blocks_allcpus();
1318 __purge_vmap_area_lazy(ULONG_MAX, 0);
1319 mutex_unlock(&vmap_purge_lock);
1320}
1321
1322/*
1323 * Free a vmap area, caller ensuring that the area has been unmapped
1324 * and flush_cache_vunmap had been called for the correct range
1325 * previously.
1326 */
1327static void free_vmap_area_noflush(struct vmap_area *va)
1328{
1329 unsigned long nr_lazy;
1330
1331 spin_lock(&vmap_area_lock);
1332 unlink_va(va, &vmap_area_root);
1333 spin_unlock(&vmap_area_lock);
1334
1335 nr_lazy = atomic_long_add_return((va->va_end - va->va_start) >>
1336 PAGE_SHIFT, &vmap_lazy_nr);
1337
1338 /* After this point, we may free va at any time */
1339 llist_add(&va->purge_list, &vmap_purge_list);
1340
1341 if (unlikely(nr_lazy > lazy_max_pages()))
1342 try_purge_vmap_area_lazy();
1343}
1344
1345/*
1346 * Free and unmap a vmap area
1347 */
1348static void free_unmap_vmap_area(struct vmap_area *va)
1349{
1350 flush_cache_vunmap(va->va_start, va->va_end);
1351 unmap_vmap_area(va);
1352 if (debug_pagealloc_enabled())
1353 flush_tlb_kernel_range(va->va_start, va->va_end);
1354
1355 free_vmap_area_noflush(va);
1356}
1357
1358static struct vmap_area *find_vmap_area(unsigned long addr)
1359{
1360 struct vmap_area *va;
1361
1362 spin_lock(&vmap_area_lock);
1363 va = __find_vmap_area(addr);
1364 spin_unlock(&vmap_area_lock);
1365
1366 return va;
1367}
1368
1369/*** Per cpu kva allocator ***/
1370
1371/*
1372 * vmap space is limited especially on 32 bit architectures. Ensure there is
1373 * room for at least 16 percpu vmap blocks per CPU.
1374 */
1375/*
1376 * If we had a constant VMALLOC_START and VMALLOC_END, we'd like to be able
1377 * to #define VMALLOC_SPACE (VMALLOC_END-VMALLOC_START). Guess
1378 * instead (we just need a rough idea)
1379 */
1380#if BITS_PER_LONG == 32
1381#define VMALLOC_SPACE (128UL*1024*1024)
1382#else
1383#define VMALLOC_SPACE (128UL*1024*1024*1024)
1384#endif
1385
1386#define VMALLOC_PAGES (VMALLOC_SPACE / PAGE_SIZE)
1387#define VMAP_MAX_ALLOC BITS_PER_LONG /* 256K with 4K pages */
1388#define VMAP_BBMAP_BITS_MAX 1024 /* 4MB with 4K pages */
1389#define VMAP_BBMAP_BITS_MIN (VMAP_MAX_ALLOC*2)
1390#define VMAP_MIN(x, y) ((x) < (y) ? (x) : (y)) /* can't use min() */
1391#define VMAP_MAX(x, y) ((x) > (y) ? (x) : (y)) /* can't use max() */
1392#define VMAP_BBMAP_BITS \
1393 VMAP_MIN(VMAP_BBMAP_BITS_MAX, \
1394 VMAP_MAX(VMAP_BBMAP_BITS_MIN, \
1395 VMALLOC_PAGES / roundup_pow_of_two(NR_CPUS) / 16))
1396
1397#define VMAP_BLOCK_SIZE (VMAP_BBMAP_BITS * PAGE_SIZE)
1398
1399struct vmap_block_queue {
1400 spinlock_t lock;
1401 struct list_head free;
1402};
1403
1404struct vmap_block {
1405 spinlock_t lock;
1406 struct vmap_area *va;
1407 unsigned long free, dirty;
1408 unsigned long dirty_min, dirty_max; /*< dirty range */
1409 struct list_head free_list;
1410 struct rcu_head rcu_head;
1411 struct list_head purge;
1412};
1413
1414/* Queue of free and dirty vmap blocks, for allocation and flushing purposes */
1415static DEFINE_PER_CPU(struct vmap_block_queue, vmap_block_queue);
1416
1417/*
1418 * Radix tree of vmap blocks, indexed by address, to quickly find a vmap block
1419 * in the free path. Could get rid of this if we change the API to return a
1420 * "cookie" from alloc, to be passed to free. But no big deal yet.
1421 */
1422static DEFINE_SPINLOCK(vmap_block_tree_lock);
1423static RADIX_TREE(vmap_block_tree, GFP_ATOMIC);
1424
1425/*
1426 * We should probably have a fallback mechanism to allocate virtual memory
1427 * out of partially filled vmap blocks. However vmap block sizing should be
1428 * fairly reasonable according to the vmalloc size, so it shouldn't be a
1429 * big problem.
1430 */
1431
1432static unsigned long addr_to_vb_idx(unsigned long addr)
1433{
1434 addr -= VMALLOC_START & ~(VMAP_BLOCK_SIZE-1);
1435 addr /= VMAP_BLOCK_SIZE;
1436 return addr;
1437}
1438
1439static void *vmap_block_vaddr(unsigned long va_start, unsigned long pages_off)
1440{
1441 unsigned long addr;
1442
1443 addr = va_start + (pages_off << PAGE_SHIFT);
1444 BUG_ON(addr_to_vb_idx(addr) != addr_to_vb_idx(va_start));
1445 return (void *)addr;
1446}
1447
1448/**
1449 * new_vmap_block - allocates new vmap_block and occupies 2^order pages in this
1450 * block. Of course pages number can't exceed VMAP_BBMAP_BITS
1451 * @order: how many 2^order pages should be occupied in newly allocated block
1452 * @gfp_mask: flags for the page level allocator
1453 *
1454 * Return: virtual address in a newly allocated block or ERR_PTR(-errno)
1455 */
1456static void *new_vmap_block(unsigned int order, gfp_t gfp_mask)
1457{
1458 struct vmap_block_queue *vbq;
1459 struct vmap_block *vb;
1460 struct vmap_area *va;
1461 unsigned long vb_idx;
1462 int node, err;
1463 void *vaddr;
1464
1465 node = numa_node_id();
1466
1467 vb = kmalloc_node(sizeof(struct vmap_block),
1468 gfp_mask & GFP_RECLAIM_MASK, node);
1469 if (unlikely(!vb))
1470 return ERR_PTR(-ENOMEM);
1471
1472 va = alloc_vmap_area(VMAP_BLOCK_SIZE, VMAP_BLOCK_SIZE,
1473 VMALLOC_START, VMALLOC_END,
1474 node, gfp_mask);
1475 if (IS_ERR(va)) {
1476 kfree(vb);
1477 return ERR_CAST(va);
1478 }
1479
1480 err = radix_tree_preload(gfp_mask);
1481 if (unlikely(err)) {
1482 kfree(vb);
1483 free_vmap_area(va);
1484 return ERR_PTR(err);
1485 }
1486
1487 vaddr = vmap_block_vaddr(va->va_start, 0);
1488 spin_lock_init(&vb->lock);
1489 vb->va = va;
1490 /* At least something should be left free */
1491 BUG_ON(VMAP_BBMAP_BITS <= (1UL << order));
1492 vb->free = VMAP_BBMAP_BITS - (1UL << order);
1493 vb->dirty = 0;
1494 vb->dirty_min = VMAP_BBMAP_BITS;
1495 vb->dirty_max = 0;
1496 INIT_LIST_HEAD(&vb->free_list);
1497
1498 vb_idx = addr_to_vb_idx(va->va_start);
1499 spin_lock(&vmap_block_tree_lock);
1500 err = radix_tree_insert(&vmap_block_tree, vb_idx, vb);
1501 spin_unlock(&vmap_block_tree_lock);
1502 BUG_ON(err);
1503 radix_tree_preload_end();
1504
1505 vbq = &get_cpu_var(vmap_block_queue);
1506 spin_lock(&vbq->lock);
1507 list_add_tail_rcu(&vb->free_list, &vbq->free);
1508 spin_unlock(&vbq->lock);
1509 put_cpu_var(vmap_block_queue);
1510
1511 return vaddr;
1512}
1513
1514static void free_vmap_block(struct vmap_block *vb)
1515{
1516 struct vmap_block *tmp;
1517 unsigned long vb_idx;
1518
1519 vb_idx = addr_to_vb_idx(vb->va->va_start);
1520 spin_lock(&vmap_block_tree_lock);
1521 tmp = radix_tree_delete(&vmap_block_tree, vb_idx);
1522 spin_unlock(&vmap_block_tree_lock);
1523 BUG_ON(tmp != vb);
1524
1525 free_vmap_area_noflush(vb->va);
1526 kfree_rcu(vb, rcu_head);
1527}
1528
1529static void purge_fragmented_blocks(int cpu)
1530{
1531 LIST_HEAD(purge);
1532 struct vmap_block *vb;
1533 struct vmap_block *n_vb;
1534 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1535
1536 rcu_read_lock();
1537 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1538
1539 if (!(vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS))
1540 continue;
1541
1542 spin_lock(&vb->lock);
1543 if (vb->free + vb->dirty == VMAP_BBMAP_BITS && vb->dirty != VMAP_BBMAP_BITS) {
1544 vb->free = 0; /* prevent further allocs after releasing lock */
1545 vb->dirty = VMAP_BBMAP_BITS; /* prevent purging it again */
1546 vb->dirty_min = 0;
1547 vb->dirty_max = VMAP_BBMAP_BITS;
1548 spin_lock(&vbq->lock);
1549 list_del_rcu(&vb->free_list);
1550 spin_unlock(&vbq->lock);
1551 spin_unlock(&vb->lock);
1552 list_add_tail(&vb->purge, &purge);
1553 } else
1554 spin_unlock(&vb->lock);
1555 }
1556 rcu_read_unlock();
1557
1558 list_for_each_entry_safe(vb, n_vb, &purge, purge) {
1559 list_del(&vb->purge);
1560 free_vmap_block(vb);
1561 }
1562}
1563
1564static void purge_fragmented_blocks_allcpus(void)
1565{
1566 int cpu;
1567
1568 for_each_possible_cpu(cpu)
1569 purge_fragmented_blocks(cpu);
1570}
1571
1572static void *vb_alloc(unsigned long size, gfp_t gfp_mask)
1573{
1574 struct vmap_block_queue *vbq;
1575 struct vmap_block *vb;
1576 void *vaddr = NULL;
1577 unsigned int order;
1578
1579 BUG_ON(offset_in_page(size));
1580 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1581 if (WARN_ON(size == 0)) {
1582 /*
1583 * Allocating 0 bytes isn't what caller wants since
1584 * get_order(0) returns funny result. Just warn and terminate
1585 * early.
1586 */
1587 return NULL;
1588 }
1589 order = get_order(size);
1590
1591 rcu_read_lock();
1592 vbq = &get_cpu_var(vmap_block_queue);
1593 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1594 unsigned long pages_off;
1595
1596 spin_lock(&vb->lock);
1597 if (vb->free < (1UL << order)) {
1598 spin_unlock(&vb->lock);
1599 continue;
1600 }
1601
1602 pages_off = VMAP_BBMAP_BITS - vb->free;
1603 vaddr = vmap_block_vaddr(vb->va->va_start, pages_off);
1604 vb->free -= 1UL << order;
1605 if (vb->free == 0) {
1606 spin_lock(&vbq->lock);
1607 list_del_rcu(&vb->free_list);
1608 spin_unlock(&vbq->lock);
1609 }
1610
1611 spin_unlock(&vb->lock);
1612 break;
1613 }
1614
1615 put_cpu_var(vmap_block_queue);
1616 rcu_read_unlock();
1617
1618 /* Allocate new block if nothing was found */
1619 if (!vaddr)
1620 vaddr = new_vmap_block(order, gfp_mask);
1621
1622 return vaddr;
1623}
1624
1625static void vb_free(const void *addr, unsigned long size)
1626{
1627 unsigned long offset;
1628 unsigned long vb_idx;
1629 unsigned int order;
1630 struct vmap_block *vb;
1631
1632 BUG_ON(offset_in_page(size));
1633 BUG_ON(size > PAGE_SIZE*VMAP_MAX_ALLOC);
1634
1635 flush_cache_vunmap((unsigned long)addr, (unsigned long)addr + size);
1636
1637 order = get_order(size);
1638
1639 offset = (unsigned long)addr & (VMAP_BLOCK_SIZE - 1);
1640 offset >>= PAGE_SHIFT;
1641
1642 vb_idx = addr_to_vb_idx((unsigned long)addr);
1643 rcu_read_lock();
1644 vb = radix_tree_lookup(&vmap_block_tree, vb_idx);
1645 rcu_read_unlock();
1646 BUG_ON(!vb);
1647
1648 vunmap_page_range((unsigned long)addr, (unsigned long)addr + size);
1649
1650 if (debug_pagealloc_enabled())
1651 flush_tlb_kernel_range((unsigned long)addr,
1652 (unsigned long)addr + size);
1653
1654 spin_lock(&vb->lock);
1655
1656 /* Expand dirty range */
1657 vb->dirty_min = min(vb->dirty_min, offset);
1658 vb->dirty_max = max(vb->dirty_max, offset + (1UL << order));
1659
1660 vb->dirty += 1UL << order;
1661 if (vb->dirty == VMAP_BBMAP_BITS) {
1662 BUG_ON(vb->free);
1663 spin_unlock(&vb->lock);
1664 free_vmap_block(vb);
1665 } else
1666 spin_unlock(&vb->lock);
1667}
1668
1669static void _vm_unmap_aliases(unsigned long start, unsigned long end, int flush)
1670{
1671 int cpu;
1672
1673 if (unlikely(!vmap_initialized))
1674 return;
1675
1676 might_sleep();
1677
1678 for_each_possible_cpu(cpu) {
1679 struct vmap_block_queue *vbq = &per_cpu(vmap_block_queue, cpu);
1680 struct vmap_block *vb;
1681
1682 rcu_read_lock();
1683 list_for_each_entry_rcu(vb, &vbq->free, free_list) {
1684 spin_lock(&vb->lock);
1685 if (vb->dirty) {
1686 unsigned long va_start = vb->va->va_start;
1687 unsigned long s, e;
1688
1689 s = va_start + (vb->dirty_min << PAGE_SHIFT);
1690 e = va_start + (vb->dirty_max << PAGE_SHIFT);
1691
1692 start = min(s, start);
1693 end = max(e, end);
1694
1695 flush = 1;
1696 }
1697 spin_unlock(&vb->lock);
1698 }
1699 rcu_read_unlock();
1700 }
1701
1702 mutex_lock(&vmap_purge_lock);
1703 purge_fragmented_blocks_allcpus();
1704 if (!__purge_vmap_area_lazy(start, end) && flush)
1705 flush_tlb_kernel_range(start, end);
1706 mutex_unlock(&vmap_purge_lock);
1707}
1708
1709/**
1710 * vm_unmap_aliases - unmap outstanding lazy aliases in the vmap layer
1711 *
1712 * The vmap/vmalloc layer lazily flushes kernel virtual mappings primarily
1713 * to amortize TLB flushing overheads. What this means is that any page you
1714 * have now, may, in a former life, have been mapped into kernel virtual
1715 * address by the vmap layer and so there might be some CPUs with TLB entries
1716 * still referencing that page (additional to the regular 1:1 kernel mapping).
1717 *
1718 * vm_unmap_aliases flushes all such lazy mappings. After it returns, we can
1719 * be sure that none of the pages we have control over will have any aliases
1720 * from the vmap layer.
1721 */
1722void vm_unmap_aliases(void)
1723{
1724 unsigned long start = ULONG_MAX, end = 0;
1725 int flush = 0;
1726
1727 _vm_unmap_aliases(start, end, flush);
1728}
1729EXPORT_SYMBOL_GPL(vm_unmap_aliases);
1730
1731/**
1732 * vm_unmap_ram - unmap linear kernel address space set up by vm_map_ram
1733 * @mem: the pointer returned by vm_map_ram
1734 * @count: the count passed to that vm_map_ram call (cannot unmap partial)
1735 */
1736void vm_unmap_ram(const void *mem, unsigned int count)
1737{
1738 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1739 unsigned long addr = (unsigned long)mem;
1740 struct vmap_area *va;
1741
1742 might_sleep();
1743 BUG_ON(!addr);
1744 BUG_ON(addr < VMALLOC_START);
1745 BUG_ON(addr > VMALLOC_END);
1746 BUG_ON(!PAGE_ALIGNED(addr));
1747
1748 if (likely(count <= VMAP_MAX_ALLOC)) {
1749 debug_check_no_locks_freed(mem, size);
1750 vb_free(mem, size);
1751 return;
1752 }
1753
1754 va = find_vmap_area(addr);
1755 BUG_ON(!va);
1756 debug_check_no_locks_freed((void *)va->va_start,
1757 (va->va_end - va->va_start));
1758 free_unmap_vmap_area(va);
1759}
1760EXPORT_SYMBOL(vm_unmap_ram);
1761
1762/**
1763 * vm_map_ram - map pages linearly into kernel virtual address (vmalloc space)
1764 * @pages: an array of pointers to the pages to be mapped
1765 * @count: number of pages
1766 * @node: prefer to allocate data structures on this node
1767 * @prot: memory protection to use. PAGE_KERNEL for regular RAM
1768 *
1769 * If you use this function for less than VMAP_MAX_ALLOC pages, it could be
1770 * faster than vmap so it's good. But if you mix long-life and short-life
1771 * objects with vm_map_ram(), it could consume lots of address space through
1772 * fragmentation (especially on a 32bit machine). You could see failures in
1773 * the end. Please use this function for short-lived objects.
1774 *
1775 * Returns: a pointer to the address that has been mapped, or %NULL on failure
1776 */
1777void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
1778{
1779 unsigned long size = (unsigned long)count << PAGE_SHIFT;
1780 unsigned long addr;
1781 void *mem;
1782
1783 if (likely(count <= VMAP_MAX_ALLOC)) {
1784 mem = vb_alloc(size, GFP_KERNEL);
1785 if (IS_ERR(mem))
1786 return NULL;
1787 addr = (unsigned long)mem;
1788 } else {
1789 struct vmap_area *va;
1790 va = alloc_vmap_area(size, PAGE_SIZE,
1791 VMALLOC_START, VMALLOC_END, node, GFP_KERNEL);
1792 if (IS_ERR(va))
1793 return NULL;
1794
1795 addr = va->va_start;
1796 mem = (void *)addr;
1797 }
1798 if (vmap_page_range(addr, addr + size, prot, pages) < 0) {
1799 vm_unmap_ram(mem, count);
1800 return NULL;
1801 }
1802 return mem;
1803}
1804EXPORT_SYMBOL(vm_map_ram);
1805
1806static struct vm_struct *vmlist __initdata;
1807
1808/**
1809 * vm_area_add_early - add vmap area early during boot
1810 * @vm: vm_struct to add
1811 *
1812 * This function is used to add fixed kernel vm area to vmlist before
1813 * vmalloc_init() is called. @vm->addr, @vm->size, and @vm->flags
1814 * should contain proper values and the other fields should be zero.
1815 *
1816 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1817 */
1818void __init vm_area_add_early(struct vm_struct *vm)
1819{
1820 struct vm_struct *tmp, **p;
1821
1822 BUG_ON(vmap_initialized);
1823 for (p = &vmlist; (tmp = *p) != NULL; p = &tmp->next) {
1824 if (tmp->addr >= vm->addr) {
1825 BUG_ON(tmp->addr < vm->addr + vm->size);
1826 break;
1827 } else
1828 BUG_ON(tmp->addr + tmp->size > vm->addr);
1829 }
1830 vm->next = *p;
1831 *p = vm;
1832}
1833
1834/**
1835 * vm_area_register_early - register vmap area early during boot
1836 * @vm: vm_struct to register
1837 * @align: requested alignment
1838 *
1839 * This function is used to register kernel vm area before
1840 * vmalloc_init() is called. @vm->size and @vm->flags should contain
1841 * proper values on entry and other fields should be zero. On return,
1842 * vm->addr contains the allocated address.
1843 *
1844 * DO NOT USE THIS FUNCTION UNLESS YOU KNOW WHAT YOU'RE DOING.
1845 */
1846void __init vm_area_register_early(struct vm_struct *vm, size_t align)
1847{
1848 static size_t vm_init_off __initdata;
1849 unsigned long addr;
1850
1851 addr = ALIGN(VMALLOC_START + vm_init_off, align);
1852 vm_init_off = PFN_ALIGN(addr + vm->size) - VMALLOC_START;
1853
1854 vm->addr = (void *)addr;
1855
1856 vm_area_add_early(vm);
1857}
1858
1859static void vmap_init_free_space(void)
1860{
1861 unsigned long vmap_start = 1;
1862 const unsigned long vmap_end = ULONG_MAX;
1863 struct vmap_area *busy, *free;
1864
1865 /*
1866 * B F B B B F
1867 * -|-----|.....|-----|-----|-----|.....|-
1868 * | The KVA space |
1869 * |<--------------------------------->|
1870 */
1871 list_for_each_entry(busy, &vmap_area_list, list) {
1872 if (busy->va_start - vmap_start > 0) {
1873 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1874 if (!WARN_ON_ONCE(!free)) {
1875 free->va_start = vmap_start;
1876 free->va_end = busy->va_start;
1877
1878 insert_vmap_area_augment(free, NULL,
1879 &free_vmap_area_root,
1880 &free_vmap_area_list);
1881 }
1882 }
1883
1884 vmap_start = busy->va_end;
1885 }
1886
1887 if (vmap_end - vmap_start > 0) {
1888 free = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1889 if (!WARN_ON_ONCE(!free)) {
1890 free->va_start = vmap_start;
1891 free->va_end = vmap_end;
1892
1893 insert_vmap_area_augment(free, NULL,
1894 &free_vmap_area_root,
1895 &free_vmap_area_list);
1896 }
1897 }
1898}
1899
1900void __init vmalloc_init(void)
1901{
1902 struct vmap_area *va;
1903 struct vm_struct *tmp;
1904 int i;
1905
1906 /*
1907 * Create the cache for vmap_area objects.
1908 */
1909 vmap_area_cachep = KMEM_CACHE(vmap_area, SLAB_PANIC);
1910
1911 for_each_possible_cpu(i) {
1912 struct vmap_block_queue *vbq;
1913 struct vfree_deferred *p;
1914
1915 vbq = &per_cpu(vmap_block_queue, i);
1916 spin_lock_init(&vbq->lock);
1917 INIT_LIST_HEAD(&vbq->free);
1918 p = &per_cpu(vfree_deferred, i);
1919 init_llist_head(&p->list);
1920 INIT_WORK(&p->wq, free_work);
1921 }
1922
1923 /* Import existing vmlist entries. */
1924 for (tmp = vmlist; tmp; tmp = tmp->next) {
1925 va = kmem_cache_zalloc(vmap_area_cachep, GFP_NOWAIT);
1926 if (WARN_ON_ONCE(!va))
1927 continue;
1928
1929 va->va_start = (unsigned long)tmp->addr;
1930 va->va_end = va->va_start + tmp->size;
1931 va->vm = tmp;
1932 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
1933 }
1934
1935 /*
1936 * Now we can initialize a free vmap space.
1937 */
1938 vmap_init_free_space();
1939 vmap_initialized = true;
1940}
1941
1942/**
1943 * map_kernel_range_noflush - map kernel VM area with the specified pages
1944 * @addr: start of the VM area to map
1945 * @size: size of the VM area to map
1946 * @prot: page protection flags to use
1947 * @pages: pages to map
1948 *
1949 * Map PFN_UP(@size) pages at @addr. The VM area @addr and @size
1950 * specify should have been allocated using get_vm_area() and its
1951 * friends.
1952 *
1953 * NOTE:
1954 * This function does NOT do any cache flushing. The caller is
1955 * responsible for calling flush_cache_vmap() on to-be-mapped areas
1956 * before calling this function.
1957 *
1958 * RETURNS:
1959 * The number of pages mapped on success, -errno on failure.
1960 */
1961int map_kernel_range_noflush(unsigned long addr, unsigned long size,
1962 pgprot_t prot, struct page **pages)
1963{
1964 return vmap_page_range_noflush(addr, addr + size, prot, pages);
1965}
1966
1967/**
1968 * unmap_kernel_range_noflush - unmap kernel VM area
1969 * @addr: start of the VM area to unmap
1970 * @size: size of the VM area to unmap
1971 *
1972 * Unmap PFN_UP(@size) pages at @addr. The VM area @addr and @size
1973 * specify should have been allocated using get_vm_area() and its
1974 * friends.
1975 *
1976 * NOTE:
1977 * This function does NOT do any cache flushing. The caller is
1978 * responsible for calling flush_cache_vunmap() on to-be-mapped areas
1979 * before calling this function and flush_tlb_kernel_range() after.
1980 */
1981void unmap_kernel_range_noflush(unsigned long addr, unsigned long size)
1982{
1983 vunmap_page_range(addr, addr + size);
1984}
1985EXPORT_SYMBOL_GPL(unmap_kernel_range_noflush);
1986
1987/**
1988 * unmap_kernel_range - unmap kernel VM area and flush cache and TLB
1989 * @addr: start of the VM area to unmap
1990 * @size: size of the VM area to unmap
1991 *
1992 * Similar to unmap_kernel_range_noflush() but flushes vcache before
1993 * the unmapping and tlb after.
1994 */
1995void unmap_kernel_range(unsigned long addr, unsigned long size)
1996{
1997 unsigned long end = addr + size;
1998
1999 flush_cache_vunmap(addr, end);
2000 vunmap_page_range(addr, end);
2001 flush_tlb_kernel_range(addr, end);
2002}
2003EXPORT_SYMBOL_GPL(unmap_kernel_range);
2004
2005int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page **pages)
2006{
2007 unsigned long addr = (unsigned long)area->addr;
2008 unsigned long end = addr + get_vm_area_size(area);
2009 int err;
2010
2011 err = vmap_page_range(addr, end, prot, pages);
2012
2013 return err > 0 ? 0 : err;
2014}
2015EXPORT_SYMBOL_GPL(map_vm_area);
2016
2017static void setup_vmalloc_vm(struct vm_struct *vm, struct vmap_area *va,
2018 unsigned long flags, const void *caller)
2019{
2020 spin_lock(&vmap_area_lock);
2021 vm->flags = flags;
2022 vm->addr = (void *)va->va_start;
2023 vm->size = va->va_end - va->va_start;
2024 vm->caller = caller;
2025 va->vm = vm;
2026 spin_unlock(&vmap_area_lock);
2027}
2028
2029static void clear_vm_uninitialized_flag(struct vm_struct *vm)
2030{
2031 /*
2032 * Before removing VM_UNINITIALIZED,
2033 * we should make sure that vm has proper values.
2034 * Pair with smp_rmb() in show_numa_info().
2035 */
2036 smp_wmb();
2037 vm->flags &= ~VM_UNINITIALIZED;
2038}
2039
2040static struct vm_struct *__get_vm_area_node(unsigned long size,
2041 unsigned long align, unsigned long flags, unsigned long start,
2042 unsigned long end, int node, gfp_t gfp_mask, const void *caller)
2043{
2044 struct vmap_area *va;
2045 struct vm_struct *area;
2046
2047 BUG_ON(in_interrupt());
2048 size = PAGE_ALIGN(size);
2049 if (unlikely(!size))
2050 return NULL;
2051
2052 if (flags & VM_IOREMAP)
2053 align = 1ul << clamp_t(int, get_count_order_long(size),
2054 PAGE_SHIFT, IOREMAP_MAX_ORDER);
2055
2056 area = kzalloc_node(sizeof(*area), gfp_mask & GFP_RECLAIM_MASK, node);
2057 if (unlikely(!area))
2058 return NULL;
2059
2060 if (!(flags & VM_NO_GUARD))
2061 size += PAGE_SIZE;
2062
2063 va = alloc_vmap_area(size, align, start, end, node, gfp_mask);
2064 if (IS_ERR(va)) {
2065 kfree(area);
2066 return NULL;
2067 }
2068
2069 setup_vmalloc_vm(area, va, flags, caller);
2070
2071 return area;
2072}
2073
2074struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
2075 unsigned long start, unsigned long end)
2076{
2077 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2078 GFP_KERNEL, __builtin_return_address(0));
2079}
2080EXPORT_SYMBOL_GPL(__get_vm_area);
2081
2082struct vm_struct *__get_vm_area_caller(unsigned long size, unsigned long flags,
2083 unsigned long start, unsigned long end,
2084 const void *caller)
2085{
2086 return __get_vm_area_node(size, 1, flags, start, end, NUMA_NO_NODE,
2087 GFP_KERNEL, caller);
2088}
2089
2090/**
2091 * get_vm_area - reserve a contiguous kernel virtual area
2092 * @size: size of the area
2093 * @flags: %VM_IOREMAP for I/O mappings or VM_ALLOC
2094 *
2095 * Search an area of @size in the kernel virtual mapping area,
2096 * and reserved it for out purposes. Returns the area descriptor
2097 * on success or %NULL on failure.
2098 *
2099 * Return: the area descriptor on success or %NULL on failure.
2100 */
2101struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
2102{
2103 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2104 NUMA_NO_NODE, GFP_KERNEL,
2105 __builtin_return_address(0));
2106}
2107
2108struct vm_struct *get_vm_area_caller(unsigned long size, unsigned long flags,
2109 const void *caller)
2110{
2111 return __get_vm_area_node(size, 1, flags, VMALLOC_START, VMALLOC_END,
2112 NUMA_NO_NODE, GFP_KERNEL, caller);
2113}
2114
2115/**
2116 * find_vm_area - find a continuous kernel virtual area
2117 * @addr: base address
2118 *
2119 * Search for the kernel VM area starting at @addr, and return it.
2120 * It is up to the caller to do all required locking to keep the returned
2121 * pointer valid.
2122 *
2123 * Return: pointer to the found area or %NULL on faulure
2124 */
2125struct vm_struct *find_vm_area(const void *addr)
2126{
2127 struct vmap_area *va;
2128
2129 va = find_vmap_area((unsigned long)addr);
2130 if (!va)
2131 return NULL;
2132
2133 return va->vm;
2134}
2135
2136/**
2137 * remove_vm_area - find and remove a continuous kernel virtual area
2138 * @addr: base address
2139 *
2140 * Search for the kernel VM area starting at @addr, and remove it.
2141 * This function returns the found VM area, but using it is NOT safe
2142 * on SMP machines, except for its size or flags.
2143 *
2144 * Return: pointer to the found area or %NULL on faulure
2145 */
2146struct vm_struct *remove_vm_area(const void *addr)
2147{
2148 struct vmap_area *va;
2149
2150 might_sleep();
2151
2152 spin_lock(&vmap_area_lock);
2153 va = __find_vmap_area((unsigned long)addr);
2154 if (va && va->vm) {
2155 struct vm_struct *vm = va->vm;
2156
2157 va->vm = NULL;
2158 spin_unlock(&vmap_area_lock);
2159
2160 kasan_free_shadow(vm);
2161 free_unmap_vmap_area(va);
2162
2163 return vm;
2164 }
2165
2166 spin_unlock(&vmap_area_lock);
2167 return NULL;
2168}
2169
2170static inline void set_area_direct_map(const struct vm_struct *area,
2171 int (*set_direct_map)(struct page *page))
2172{
2173 int i;
2174
2175 for (i = 0; i < area->nr_pages; i++)
2176 if (page_address(area->pages[i]))
2177 set_direct_map(area->pages[i]);
2178}
2179
2180/* Handle removing and resetting vm mappings related to the vm_struct. */
2181static void vm_remove_mappings(struct vm_struct *area, int deallocate_pages)
2182{
2183 unsigned long start = ULONG_MAX, end = 0;
2184 int flush_reset = area->flags & VM_FLUSH_RESET_PERMS;
2185 int flush_dmap = 0;
2186 int i;
2187
2188 remove_vm_area(area->addr);
2189
2190 /* If this is not VM_FLUSH_RESET_PERMS memory, no need for the below. */
2191 if (!flush_reset)
2192 return;
2193
2194 /*
2195 * If not deallocating pages, just do the flush of the VM area and
2196 * return.
2197 */
2198 if (!deallocate_pages) {
2199 vm_unmap_aliases();
2200 return;
2201 }
2202
2203 /*
2204 * If execution gets here, flush the vm mapping and reset the direct
2205 * map. Find the start and end range of the direct mappings to make sure
2206 * the vm_unmap_aliases() flush includes the direct map.
2207 */
2208 for (i = 0; i < area->nr_pages; i++) {
2209 unsigned long addr = (unsigned long)page_address(area->pages[i]);
2210 if (addr) {
2211 start = min(addr, start);
2212 end = max(addr + PAGE_SIZE, end);
2213 flush_dmap = 1;
2214 }
2215 }
2216
2217 /*
2218 * Set direct map to something invalid so that it won't be cached if
2219 * there are any accesses after the TLB flush, then flush the TLB and
2220 * reset the direct map permissions to the default.
2221 */
2222 set_area_direct_map(area, set_direct_map_invalid_noflush);
2223 _vm_unmap_aliases(start, end, flush_dmap);
2224 set_area_direct_map(area, set_direct_map_default_noflush);
2225}
2226
2227static void __vunmap(const void *addr, int deallocate_pages)
2228{
2229 struct vm_struct *area;
2230
2231 if (!addr)
2232 return;
2233
2234 if (WARN(!PAGE_ALIGNED(addr), "Trying to vfree() bad address (%p)\n",
2235 addr))
2236 return;
2237
2238 area = find_vm_area(addr);
2239 if (unlikely(!area)) {
2240 WARN(1, KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
2241 addr);
2242 return;
2243 }
2244
2245 debug_check_no_locks_freed(area->addr, get_vm_area_size(area));
2246 debug_check_no_obj_freed(area->addr, get_vm_area_size(area));
2247
2248 vm_remove_mappings(area, deallocate_pages);
2249
2250 if (deallocate_pages) {
2251 int i;
2252
2253 for (i = 0; i < area->nr_pages; i++) {
2254 struct page *page = area->pages[i];
2255
2256 BUG_ON(!page);
2257 __free_pages(page, 0);
2258 }
2259 atomic_long_sub(area->nr_pages, &nr_vmalloc_pages);
2260
2261 kvfree(area->pages);
2262 }
2263
2264 kfree(area);
2265 return;
2266}
2267
2268static inline void __vfree_deferred(const void *addr)
2269{
2270 /*
2271 * Use raw_cpu_ptr() because this can be called from preemptible
2272 * context. Preemption is absolutely fine here, because the llist_add()
2273 * implementation is lockless, so it works even if we are adding to
2274 * nother cpu's list. schedule_work() should be fine with this too.
2275 */
2276 struct vfree_deferred *p = raw_cpu_ptr(&vfree_deferred);
2277
2278 if (llist_add((struct llist_node *)addr, &p->list))
2279 schedule_work(&p->wq);
2280}
2281
2282/**
2283 * vfree_atomic - release memory allocated by vmalloc()
2284 * @addr: memory base address
2285 *
2286 * This one is just like vfree() but can be called in any atomic context
2287 * except NMIs.
2288 */
2289void vfree_atomic(const void *addr)
2290{
2291 BUG_ON(in_nmi());
2292
2293 kmemleak_free(addr);
2294
2295 if (!addr)
2296 return;
2297 __vfree_deferred(addr);
2298}
2299
2300static void __vfree(const void *addr)
2301{
2302 if (unlikely(in_interrupt()))
2303 __vfree_deferred(addr);
2304 else
2305 __vunmap(addr, 1);
2306}
2307
2308/**
2309 * vfree - release memory allocated by vmalloc()
2310 * @addr: memory base address
2311 *
2312 * Free the virtually continuous memory area starting at @addr, as
2313 * obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
2314 * NULL, no operation is performed.
2315 *
2316 * Must not be called in NMI context (strictly speaking, only if we don't
2317 * have CONFIG_ARCH_HAVE_NMI_SAFE_CMPXCHG, but making the calling
2318 * conventions for vfree() arch-depenedent would be a really bad idea)
2319 *
2320 * May sleep if called *not* from interrupt context.
2321 *
2322 * NOTE: assumes that the object at @addr has a size >= sizeof(llist_node)
2323 */
2324void vfree(const void *addr)
2325{
2326 BUG_ON(in_nmi());
2327
2328 kmemleak_free(addr);
2329
2330 might_sleep_if(!in_interrupt());
2331
2332 if (!addr)
2333 return;
2334
2335 __vfree(addr);
2336}
2337EXPORT_SYMBOL(vfree);
2338
2339/**
2340 * vunmap - release virtual mapping obtained by vmap()
2341 * @addr: memory base address
2342 *
2343 * Free the virtually contiguous memory area starting at @addr,
2344 * which was created from the page array passed to vmap().
2345 *
2346 * Must not be called in interrupt context.
2347 */
2348void vunmap(const void *addr)
2349{
2350 BUG_ON(in_interrupt());
2351 might_sleep();
2352 if (addr)
2353 __vunmap(addr, 0);
2354}
2355EXPORT_SYMBOL(vunmap);
2356
2357/**
2358 * vmap - map an array of pages into virtually contiguous space
2359 * @pages: array of page pointers
2360 * @count: number of pages to map
2361 * @flags: vm_area->flags
2362 * @prot: page protection for the mapping
2363 *
2364 * Maps @count pages from @pages into contiguous kernel virtual
2365 * space.
2366 *
2367 * Return: the address of the area or %NULL on failure
2368 */
2369void *vmap(struct page **pages, unsigned int count,
2370 unsigned long flags, pgprot_t prot)
2371{
2372 struct vm_struct *area;
2373 unsigned long size; /* In bytes */
2374
2375 might_sleep();
2376
2377 if (count > totalram_pages())
2378 return NULL;
2379
2380 size = (unsigned long)count << PAGE_SHIFT;
2381 area = get_vm_area_caller(size, flags, __builtin_return_address(0));
2382 if (!area)
2383 return NULL;
2384
2385 if (map_vm_area(area, prot, pages)) {
2386 vunmap(area->addr);
2387 return NULL;
2388 }
2389
2390 return area->addr;
2391}
2392EXPORT_SYMBOL(vmap);
2393
2394static void *__vmalloc_node(unsigned long size, unsigned long align,
2395 gfp_t gfp_mask, pgprot_t prot,
2396 int node, const void *caller);
2397static void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
2398 pgprot_t prot, int node)
2399{
2400 struct page **pages;
2401 unsigned int nr_pages, array_size, i;
2402 const gfp_t nested_gfp = (gfp_mask & GFP_RECLAIM_MASK) | __GFP_ZERO;
2403 const gfp_t alloc_mask = gfp_mask | __GFP_NOWARN;
2404 const gfp_t highmem_mask = (gfp_mask & (GFP_DMA | GFP_DMA32)) ?
2405 0 :
2406 __GFP_HIGHMEM;
2407
2408 nr_pages = get_vm_area_size(area) >> PAGE_SHIFT;
2409 array_size = (nr_pages * sizeof(struct page *));
2410
2411 /* Please note that the recursion is strictly bounded. */
2412 if (array_size > PAGE_SIZE) {
2413 pages = __vmalloc_node(array_size, 1, nested_gfp|highmem_mask,
2414 PAGE_KERNEL, node, area->caller);
2415 } else {
2416 pages = kmalloc_node(array_size, nested_gfp, node);
2417 }
2418
2419 if (!pages) {
2420 remove_vm_area(area->addr);
2421 kfree(area);
2422 return NULL;
2423 }
2424
2425 area->pages = pages;
2426 area->nr_pages = nr_pages;
2427
2428 for (i = 0; i < area->nr_pages; i++) {
2429 struct page *page;
2430
2431 if (node == NUMA_NO_NODE)
2432 page = alloc_page(alloc_mask|highmem_mask);
2433 else
2434 page = alloc_pages_node(node, alloc_mask|highmem_mask, 0);
2435
2436 if (unlikely(!page)) {
2437 /* Successfully allocated i pages, free them in __vunmap() */
2438 area->nr_pages = i;
2439 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2440 goto fail;
2441 }
2442 area->pages[i] = page;
2443 if (gfpflags_allow_blocking(gfp_mask|highmem_mask))
2444 cond_resched();
2445 }
2446 atomic_long_add(area->nr_pages, &nr_vmalloc_pages);
2447
2448 if (map_vm_area(area, prot, pages))
2449 goto fail;
2450 return area->addr;
2451
2452fail:
2453 warn_alloc(gfp_mask, NULL,
2454 "vmalloc: allocation failure, allocated %ld of %ld bytes",
2455 (area->nr_pages*PAGE_SIZE), area->size);
2456 __vfree(area->addr);
2457 return NULL;
2458}
2459
2460/**
2461 * __vmalloc_node_range - allocate virtually contiguous memory
2462 * @size: allocation size
2463 * @align: desired alignment
2464 * @start: vm area range start
2465 * @end: vm area range end
2466 * @gfp_mask: flags for the page level allocator
2467 * @prot: protection mask for the allocated pages
2468 * @vm_flags: additional vm area flags (e.g. %VM_NO_GUARD)
2469 * @node: node to use for allocation or NUMA_NO_NODE
2470 * @caller: caller's return address
2471 *
2472 * Allocate enough pages to cover @size from the page level
2473 * allocator with @gfp_mask flags. Map them into contiguous
2474 * kernel virtual space, using a pagetable protection of @prot.
2475 *
2476 * Return: the address of the area or %NULL on failure
2477 */
2478void *__vmalloc_node_range(unsigned long size, unsigned long align,
2479 unsigned long start, unsigned long end, gfp_t gfp_mask,
2480 pgprot_t prot, unsigned long vm_flags, int node,
2481 const void *caller)
2482{
2483 struct vm_struct *area;
2484 void *addr;
2485 unsigned long real_size = size;
2486
2487 size = PAGE_ALIGN(size);
2488 if (!size || (size >> PAGE_SHIFT) > totalram_pages())
2489 goto fail;
2490
2491 area = __get_vm_area_node(size, align, VM_ALLOC | VM_UNINITIALIZED |
2492 vm_flags, start, end, node, gfp_mask, caller);
2493 if (!area)
2494 goto fail;
2495
2496 addr = __vmalloc_area_node(area, gfp_mask, prot, node);
2497 if (!addr)
2498 return NULL;
2499
2500 /*
2501 * In this function, newly allocated vm_struct has VM_UNINITIALIZED
2502 * flag. It means that vm_struct is not fully initialized.
2503 * Now, it is fully initialized, so remove this flag here.
2504 */
2505 clear_vm_uninitialized_flag(area);
2506
2507 kmemleak_vmalloc(area, size, gfp_mask);
2508
2509 return addr;
2510
2511fail:
2512 warn_alloc(gfp_mask, NULL,
2513 "vmalloc: allocation failure: %lu bytes", real_size);
2514 return NULL;
2515}
2516
2517/*
2518 * This is only for performance analysis of vmalloc and stress purpose.
2519 * It is required by vmalloc test module, therefore do not use it other
2520 * than that.
2521 */
2522#ifdef CONFIG_TEST_VMALLOC_MODULE
2523EXPORT_SYMBOL_GPL(__vmalloc_node_range);
2524#endif
2525
2526/**
2527 * __vmalloc_node - allocate virtually contiguous memory
2528 * @size: allocation size
2529 * @align: desired alignment
2530 * @gfp_mask: flags for the page level allocator
2531 * @prot: protection mask for the allocated pages
2532 * @node: node to use for allocation or NUMA_NO_NODE
2533 * @caller: caller's return address
2534 *
2535 * Allocate enough pages to cover @size from the page level
2536 * allocator with @gfp_mask flags. Map them into contiguous
2537 * kernel virtual space, using a pagetable protection of @prot.
2538 *
2539 * Reclaim modifiers in @gfp_mask - __GFP_NORETRY, __GFP_RETRY_MAYFAIL
2540 * and __GFP_NOFAIL are not supported
2541 *
2542 * Any use of gfp flags outside of GFP_KERNEL should be consulted
2543 * with mm people.
2544 *
2545 * Return: pointer to the allocated memory or %NULL on error
2546 */
2547static void *__vmalloc_node(unsigned long size, unsigned long align,
2548 gfp_t gfp_mask, pgprot_t prot,
2549 int node, const void *caller)
2550{
2551 return __vmalloc_node_range(size, align, VMALLOC_START, VMALLOC_END,
2552 gfp_mask, prot, 0, node, caller);
2553}
2554
2555void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
2556{
2557 return __vmalloc_node(size, 1, gfp_mask, prot, NUMA_NO_NODE,
2558 __builtin_return_address(0));
2559}
2560EXPORT_SYMBOL(__vmalloc);
2561
2562static inline void *__vmalloc_node_flags(unsigned long size,
2563 int node, gfp_t flags)
2564{
2565 return __vmalloc_node(size, 1, flags, PAGE_KERNEL,
2566 node, __builtin_return_address(0));
2567}
2568
2569
2570void *__vmalloc_node_flags_caller(unsigned long size, int node, gfp_t flags,
2571 void *caller)
2572{
2573 return __vmalloc_node(size, 1, flags, PAGE_KERNEL, node, caller);
2574}
2575
2576/**
2577 * vmalloc - allocate virtually contiguous memory
2578 * @size: allocation size
2579 *
2580 * Allocate enough pages to cover @size from the page level
2581 * allocator and map them into contiguous kernel virtual space.
2582 *
2583 * For tight control over page level allocator and protection flags
2584 * use __vmalloc() instead.
2585 *
2586 * Return: pointer to the allocated memory or %NULL on error
2587 */
2588void *vmalloc(unsigned long size)
2589{
2590 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2591 GFP_KERNEL);
2592}
2593EXPORT_SYMBOL(vmalloc);
2594
2595/**
2596 * vzalloc - allocate virtually contiguous memory with zero fill
2597 * @size: allocation size
2598 *
2599 * Allocate enough pages to cover @size from the page level
2600 * allocator and map them into contiguous kernel virtual space.
2601 * The memory allocated is set to zero.
2602 *
2603 * For tight control over page level allocator and protection flags
2604 * use __vmalloc() instead.
2605 *
2606 * Return: pointer to the allocated memory or %NULL on error
2607 */
2608void *vzalloc(unsigned long size)
2609{
2610 return __vmalloc_node_flags(size, NUMA_NO_NODE,
2611 GFP_KERNEL | __GFP_ZERO);
2612}
2613EXPORT_SYMBOL(vzalloc);
2614
2615/**
2616 * vmalloc_user - allocate zeroed virtually contiguous memory for userspace
2617 * @size: allocation size
2618 *
2619 * The resulting memory area is zeroed so it can be mapped to userspace
2620 * without leaking data.
2621 *
2622 * Return: pointer to the allocated memory or %NULL on error
2623 */
2624void *vmalloc_user(unsigned long size)
2625{
2626 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2627 GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL,
2628 VM_USERMAP, NUMA_NO_NODE,
2629 __builtin_return_address(0));
2630}
2631EXPORT_SYMBOL(vmalloc_user);
2632
2633/**
2634 * vmalloc_node - allocate memory on a specific node
2635 * @size: allocation size
2636 * @node: numa node
2637 *
2638 * Allocate enough pages to cover @size from the page level
2639 * allocator and map them into contiguous kernel virtual space.
2640 *
2641 * For tight control over page level allocator and protection flags
2642 * use __vmalloc() instead.
2643 *
2644 * Return: pointer to the allocated memory or %NULL on error
2645 */
2646void *vmalloc_node(unsigned long size, int node)
2647{
2648 return __vmalloc_node(size, 1, GFP_KERNEL, PAGE_KERNEL,
2649 node, __builtin_return_address(0));
2650}
2651EXPORT_SYMBOL(vmalloc_node);
2652
2653/**
2654 * vzalloc_node - allocate memory on a specific node with zero fill
2655 * @size: allocation size
2656 * @node: numa node
2657 *
2658 * Allocate enough pages to cover @size from the page level
2659 * allocator and map them into contiguous kernel virtual space.
2660 * The memory allocated is set to zero.
2661 *
2662 * For tight control over page level allocator and protection flags
2663 * use __vmalloc_node() instead.
2664 *
2665 * Return: pointer to the allocated memory or %NULL on error
2666 */
2667void *vzalloc_node(unsigned long size, int node)
2668{
2669 return __vmalloc_node_flags(size, node,
2670 GFP_KERNEL | __GFP_ZERO);
2671}
2672EXPORT_SYMBOL(vzalloc_node);
2673
2674/**
2675 * vmalloc_exec - allocate virtually contiguous, executable memory
2676 * @size: allocation size
2677 *
2678 * Kernel-internal function to allocate enough pages to cover @size
2679 * the page level allocator and map them into contiguous and
2680 * executable kernel virtual space.
2681 *
2682 * For tight control over page level allocator and protection flags
2683 * use __vmalloc() instead.
2684 *
2685 * Return: pointer to the allocated memory or %NULL on error
2686 */
2687void *vmalloc_exec(unsigned long size)
2688{
2689 return __vmalloc_node_range(size, 1, VMALLOC_START, VMALLOC_END,
2690 GFP_KERNEL, PAGE_KERNEL_EXEC, VM_FLUSH_RESET_PERMS,
2691 NUMA_NO_NODE, __builtin_return_address(0));
2692}
2693
2694#if defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA32)
2695#define GFP_VMALLOC32 (GFP_DMA32 | GFP_KERNEL)
2696#elif defined(CONFIG_64BIT) && defined(CONFIG_ZONE_DMA)
2697#define GFP_VMALLOC32 (GFP_DMA | GFP_KERNEL)
2698#else
2699/*
2700 * 64b systems should always have either DMA or DMA32 zones. For others
2701 * GFP_DMA32 should do the right thing and use the normal zone.
2702 */
2703#define GFP_VMALLOC32 GFP_DMA32 | GFP_KERNEL
2704#endif
2705
2706/**
2707 * vmalloc_32 - allocate virtually contiguous memory (32bit addressable)
2708 * @size: allocation size
2709 *
2710 * Allocate enough 32bit PA addressable pages to cover @size from the
2711 * page level allocator and map them into contiguous kernel virtual space.
2712 *
2713 * Return: pointer to the allocated memory or %NULL on error
2714 */
2715void *vmalloc_32(unsigned long size)
2716{
2717 return __vmalloc_node(size, 1, GFP_VMALLOC32, PAGE_KERNEL,
2718 NUMA_NO_NODE, __builtin_return_address(0));
2719}
2720EXPORT_SYMBOL(vmalloc_32);
2721
2722/**
2723 * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
2724 * @size: allocation size
2725 *
2726 * The resulting memory area is 32bit addressable and zeroed so it can be
2727 * mapped to userspace without leaking data.
2728 *
2729 * Return: pointer to the allocated memory or %NULL on error
2730 */
2731void *vmalloc_32_user(unsigned long size)
2732{
2733 return __vmalloc_node_range(size, SHMLBA, VMALLOC_START, VMALLOC_END,
2734 GFP_VMALLOC32 | __GFP_ZERO, PAGE_KERNEL,
2735 VM_USERMAP, NUMA_NO_NODE,
2736 __builtin_return_address(0));
2737}
2738EXPORT_SYMBOL(vmalloc_32_user);
2739
2740/*
2741 * small helper routine , copy contents to buf from addr.
2742 * If the page is not present, fill zero.
2743 */
2744
2745static int aligned_vread(char *buf, char *addr, unsigned long count)
2746{
2747 struct page *p;
2748 int copied = 0;
2749
2750 while (count) {
2751 unsigned long offset, length;
2752
2753 offset = offset_in_page(addr);
2754 length = PAGE_SIZE - offset;
2755 if (length > count)
2756 length = count;
2757 p = vmalloc_to_page(addr);
2758 /*
2759 * To do safe access to this _mapped_ area, we need
2760 * lock. But adding lock here means that we need to add
2761 * overhead of vmalloc()/vfree() calles for this _debug_
2762 * interface, rarely used. Instead of that, we'll use
2763 * kmap() and get small overhead in this access function.
2764 */
2765 if (p) {
2766 /*
2767 * we can expect USER0 is not used (see vread/vwrite's
2768 * function description)
2769 */
2770 void *map = kmap_atomic(p);
2771 memcpy(buf, map + offset, length);
2772 kunmap_atomic(map);
2773 } else
2774 memset(buf, 0, length);
2775
2776 addr += length;
2777 buf += length;
2778 copied += length;
2779 count -= length;
2780 }
2781 return copied;
2782}
2783
2784static int aligned_vwrite(char *buf, char *addr, unsigned long count)
2785{
2786 struct page *p;
2787 int copied = 0;
2788
2789 while (count) {
2790 unsigned long offset, length;
2791
2792 offset = offset_in_page(addr);
2793 length = PAGE_SIZE - offset;
2794 if (length > count)
2795 length = count;
2796 p = vmalloc_to_page(addr);
2797 /*
2798 * To do safe access to this _mapped_ area, we need
2799 * lock. But adding lock here means that we need to add
2800 * overhead of vmalloc()/vfree() calles for this _debug_
2801 * interface, rarely used. Instead of that, we'll use
2802 * kmap() and get small overhead in this access function.
2803 */
2804 if (p) {
2805 /*
2806 * we can expect USER0 is not used (see vread/vwrite's
2807 * function description)
2808 */
2809 void *map = kmap_atomic(p);
2810 memcpy(map + offset, buf, length);
2811 kunmap_atomic(map);
2812 }
2813 addr += length;
2814 buf += length;
2815 copied += length;
2816 count -= length;
2817 }
2818 return copied;
2819}
2820
2821/**
2822 * vread() - read vmalloc area in a safe way.
2823 * @buf: buffer for reading data
2824 * @addr: vm address.
2825 * @count: number of bytes to be read.
2826 *
2827 * This function checks that addr is a valid vmalloc'ed area, and
2828 * copy data from that area to a given buffer. If the given memory range
2829 * of [addr...addr+count) includes some valid address, data is copied to
2830 * proper area of @buf. If there are memory holes, they'll be zero-filled.
2831 * IOREMAP area is treated as memory hole and no copy is done.
2832 *
2833 * If [addr...addr+count) doesn't includes any intersects with alive
2834 * vm_struct area, returns 0. @buf should be kernel's buffer.
2835 *
2836 * Note: In usual ops, vread() is never necessary because the caller
2837 * should know vmalloc() area is valid and can use memcpy().
2838 * This is for routines which have to access vmalloc area without
2839 * any information, as /dev/kmem.
2840 *
2841 * Return: number of bytes for which addr and buf should be increased
2842 * (same number as @count) or %0 if [addr...addr+count) doesn't
2843 * include any intersection with valid vmalloc area
2844 */
2845long vread(char *buf, char *addr, unsigned long count)
2846{
2847 struct vmap_area *va;
2848 struct vm_struct *vm;
2849 char *vaddr, *buf_start = buf;
2850 unsigned long buflen = count;
2851 unsigned long n;
2852
2853 /* Don't allow overflow */
2854 if ((unsigned long) addr + count < count)
2855 count = -(unsigned long) addr;
2856
2857 spin_lock(&vmap_area_lock);
2858 list_for_each_entry(va, &vmap_area_list, list) {
2859 if (!count)
2860 break;
2861
2862 if (!va->vm)
2863 continue;
2864
2865 vm = va->vm;
2866 vaddr = (char *) vm->addr;
2867 if (addr >= vaddr + get_vm_area_size(vm))
2868 continue;
2869 while (addr < vaddr) {
2870 if (count == 0)
2871 goto finished;
2872 *buf = '\0';
2873 buf++;
2874 addr++;
2875 count--;
2876 }
2877 n = vaddr + get_vm_area_size(vm) - addr;
2878 if (n > count)
2879 n = count;
2880 if (!(vm->flags & VM_IOREMAP))
2881 aligned_vread(buf, addr, n);
2882 else /* IOREMAP area is treated as memory hole */
2883 memset(buf, 0, n);
2884 buf += n;
2885 addr += n;
2886 count -= n;
2887 }
2888finished:
2889 spin_unlock(&vmap_area_lock);
2890
2891 if (buf == buf_start)
2892 return 0;
2893 /* zero-fill memory holes */
2894 if (buf != buf_start + buflen)
2895 memset(buf, 0, buflen - (buf - buf_start));
2896
2897 return buflen;
2898}
2899
2900/**
2901 * vwrite() - write vmalloc area in a safe way.
2902 * @buf: buffer for source data
2903 * @addr: vm address.
2904 * @count: number of bytes to be read.
2905 *
2906 * This function checks that addr is a valid vmalloc'ed area, and
2907 * copy data from a buffer to the given addr. If specified range of
2908 * [addr...addr+count) includes some valid address, data is copied from
2909 * proper area of @buf. If there are memory holes, no copy to hole.
2910 * IOREMAP area is treated as memory hole and no copy is done.
2911 *
2912 * If [addr...addr+count) doesn't includes any intersects with alive
2913 * vm_struct area, returns 0. @buf should be kernel's buffer.
2914 *
2915 * Note: In usual ops, vwrite() is never necessary because the caller
2916 * should know vmalloc() area is valid and can use memcpy().
2917 * This is for routines which have to access vmalloc area without
2918 * any information, as /dev/kmem.
2919 *
2920 * Return: number of bytes for which addr and buf should be
2921 * increased (same number as @count) or %0 if [addr...addr+count)
2922 * doesn't include any intersection with valid vmalloc area
2923 */
2924long vwrite(char *buf, char *addr, unsigned long count)
2925{
2926 struct vmap_area *va;
2927 struct vm_struct *vm;
2928 char *vaddr;
2929 unsigned long n, buflen;
2930 int copied = 0;
2931
2932 /* Don't allow overflow */
2933 if ((unsigned long) addr + count < count)
2934 count = -(unsigned long) addr;
2935 buflen = count;
2936
2937 spin_lock(&vmap_area_lock);
2938 list_for_each_entry(va, &vmap_area_list, list) {
2939 if (!count)
2940 break;
2941
2942 if (!va->vm)
2943 continue;
2944
2945 vm = va->vm;
2946 vaddr = (char *) vm->addr;
2947 if (addr >= vaddr + get_vm_area_size(vm))
2948 continue;
2949 while (addr < vaddr) {
2950 if (count == 0)
2951 goto finished;
2952 buf++;
2953 addr++;
2954 count--;
2955 }
2956 n = vaddr + get_vm_area_size(vm) - addr;
2957 if (n > count)
2958 n = count;
2959 if (!(vm->flags & VM_IOREMAP)) {
2960 aligned_vwrite(buf, addr, n);
2961 copied++;
2962 }
2963 buf += n;
2964 addr += n;
2965 count -= n;
2966 }
2967finished:
2968 spin_unlock(&vmap_area_lock);
2969 if (!copied)
2970 return 0;
2971 return buflen;
2972}
2973
2974/**
2975 * remap_vmalloc_range_partial - map vmalloc pages to userspace
2976 * @vma: vma to cover
2977 * @uaddr: target user address to start at
2978 * @kaddr: virtual address of vmalloc kernel memory
2979 * @size: size of map area
2980 *
2981 * Returns: 0 for success, -Exxx on failure
2982 *
2983 * This function checks that @kaddr is a valid vmalloc'ed area,
2984 * and that it is big enough to cover the range starting at
2985 * @uaddr in @vma. Will return failure if that criteria isn't
2986 * met.
2987 *
2988 * Similar to remap_pfn_range() (see mm/memory.c)
2989 */
2990int remap_vmalloc_range_partial(struct vm_area_struct *vma, unsigned long uaddr,
2991 void *kaddr, unsigned long size)
2992{
2993 struct vm_struct *area;
2994
2995 size = PAGE_ALIGN(size);
2996
2997 if (!PAGE_ALIGNED(uaddr) || !PAGE_ALIGNED(kaddr))
2998 return -EINVAL;
2999
3000 area = find_vm_area(kaddr);
3001 if (!area)
3002 return -EINVAL;
3003
3004 if (!(area->flags & (VM_USERMAP | VM_DMA_COHERENT)))
3005 return -EINVAL;
3006
3007 if (kaddr + size > area->addr + get_vm_area_size(area))
3008 return -EINVAL;
3009
3010 do {
3011 struct page *page = vmalloc_to_page(kaddr);
3012 int ret;
3013
3014 ret = vm_insert_page(vma, uaddr, page);
3015 if (ret)
3016 return ret;
3017
3018 uaddr += PAGE_SIZE;
3019 kaddr += PAGE_SIZE;
3020 size -= PAGE_SIZE;
3021 } while (size > 0);
3022
3023 vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
3024
3025 return 0;
3026}
3027EXPORT_SYMBOL(remap_vmalloc_range_partial);
3028
3029/**
3030 * remap_vmalloc_range - map vmalloc pages to userspace
3031 * @vma: vma to cover (map full range of vma)
3032 * @addr: vmalloc memory
3033 * @pgoff: number of pages into addr before first page to map
3034 *
3035 * Returns: 0 for success, -Exxx on failure
3036 *
3037 * This function checks that addr is a valid vmalloc'ed area, and
3038 * that it is big enough to cover the vma. Will return failure if
3039 * that criteria isn't met.
3040 *
3041 * Similar to remap_pfn_range() (see mm/memory.c)
3042 */
3043int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
3044 unsigned long pgoff)
3045{
3046 return remap_vmalloc_range_partial(vma, vma->vm_start,
3047 addr + (pgoff << PAGE_SHIFT),
3048 vma->vm_end - vma->vm_start);
3049}
3050EXPORT_SYMBOL(remap_vmalloc_range);
3051
3052/*
3053 * Implement a stub for vmalloc_sync_all() if the architecture chose not to
3054 * have one.
3055 *
3056 * The purpose of this function is to make sure the vmalloc area
3057 * mappings are identical in all page-tables in the system.
3058 */
3059void __weak vmalloc_sync_all(void)
3060{
3061}
3062
3063
3064static int f(pte_t *pte, unsigned long addr, void *data)
3065{
3066 pte_t ***p = data;
3067
3068 if (p) {
3069 *(*p) = pte;
3070 (*p)++;
3071 }
3072 return 0;
3073}
3074
3075/**
3076 * alloc_vm_area - allocate a range of kernel address space
3077 * @size: size of the area
3078 * @ptes: returns the PTEs for the address space
3079 *
3080 * Returns: NULL on failure, vm_struct on success
3081 *
3082 * This function reserves a range of kernel address space, and
3083 * allocates pagetables to map that range. No actual mappings
3084 * are created.
3085 *
3086 * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
3087 * allocated for the VM area are returned.
3088 */
3089struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
3090{
3091 struct vm_struct *area;
3092
3093 area = get_vm_area_caller(size, VM_IOREMAP,
3094 __builtin_return_address(0));
3095 if (area == NULL)
3096 return NULL;
3097
3098 /*
3099 * This ensures that page tables are constructed for this region
3100 * of kernel virtual address space and mapped into init_mm.
3101 */
3102 if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
3103 size, f, ptes ? &ptes : NULL)) {
3104 free_vm_area(area);
3105 return NULL;
3106 }
3107
3108 return area;
3109}
3110EXPORT_SYMBOL_GPL(alloc_vm_area);
3111
3112void free_vm_area(struct vm_struct *area)
3113{
3114 struct vm_struct *ret;
3115 ret = remove_vm_area(area->addr);
3116 BUG_ON(ret != area);
3117 kfree(area);
3118}
3119EXPORT_SYMBOL_GPL(free_vm_area);
3120
3121#ifdef CONFIG_SMP
3122static struct vmap_area *node_to_va(struct rb_node *n)
3123{
3124 return rb_entry_safe(n, struct vmap_area, rb_node);
3125}
3126
3127/**
3128 * pvm_find_va_enclose_addr - find the vmap_area @addr belongs to
3129 * @addr: target address
3130 *
3131 * Returns: vmap_area if it is found. If there is no such area
3132 * the first highest(reverse order) vmap_area is returned
3133 * i.e. va->va_start < addr && va->va_end < addr or NULL
3134 * if there are no any areas before @addr.
3135 */
3136static struct vmap_area *
3137pvm_find_va_enclose_addr(unsigned long addr)
3138{
3139 struct vmap_area *va, *tmp;
3140 struct rb_node *n;
3141
3142 n = free_vmap_area_root.rb_node;
3143 va = NULL;
3144
3145 while (n) {
3146 tmp = rb_entry(n, struct vmap_area, rb_node);
3147 if (tmp->va_start <= addr) {
3148 va = tmp;
3149 if (tmp->va_end >= addr)
3150 break;
3151
3152 n = n->rb_right;
3153 } else {
3154 n = n->rb_left;
3155 }
3156 }
3157
3158 return va;
3159}
3160
3161/**
3162 * pvm_determine_end_from_reverse - find the highest aligned address
3163 * of free block below VMALLOC_END
3164 * @va:
3165 * in - the VA we start the search(reverse order);
3166 * out - the VA with the highest aligned end address.
3167 *
3168 * Returns: determined end address within vmap_area
3169 */
3170static unsigned long
3171pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
3172{
3173 unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3174 unsigned long addr;
3175
3176 if (likely(*va)) {
3177 list_for_each_entry_from_reverse((*va),
3178 &free_vmap_area_list, list) {
3179 addr = min((*va)->va_end & ~(align - 1), vmalloc_end);
3180 if ((*va)->va_start < addr)
3181 return addr;
3182 }
3183 }
3184
3185 return 0;
3186}
3187
3188/**
3189 * pcpu_get_vm_areas - allocate vmalloc areas for percpu allocator
3190 * @offsets: array containing offset of each area
3191 * @sizes: array containing size of each area
3192 * @nr_vms: the number of areas to allocate
3193 * @align: alignment, all entries in @offsets and @sizes must be aligned to this
3194 *
3195 * Returns: kmalloc'd vm_struct pointer array pointing to allocated
3196 * vm_structs on success, %NULL on failure
3197 *
3198 * Percpu allocator wants to use congruent vm areas so that it can
3199 * maintain the offsets among percpu areas. This function allocates
3200 * congruent vmalloc areas for it with GFP_KERNEL. These areas tend to
3201 * be scattered pretty far, distance between two areas easily going up
3202 * to gigabytes. To avoid interacting with regular vmallocs, these
3203 * areas are allocated from top.
3204 *
3205 * Despite its complicated look, this allocator is rather simple. It
3206 * does everything top-down and scans free blocks from the end looking
3207 * for matching base. While scanning, if any of the areas do not fit the
3208 * base address is pulled down to fit the area. Scanning is repeated till
3209 * all the areas fit and then all necessary data structures are inserted
3210 * and the result is returned.
3211 */
3212struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
3213 const size_t *sizes, int nr_vms,
3214 size_t align)
3215{
3216 const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
3217 const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
3218 struct vmap_area **vas, *va;
3219 struct vm_struct **vms;
3220 int area, area2, last_area, term_area;
3221 unsigned long base, start, size, end, last_end;
3222 bool purged = false;
3223 enum fit_type type;
3224
3225 /* verify parameters and allocate data structures */
3226 BUG_ON(offset_in_page(align) || !is_power_of_2(align));
3227 for (last_area = 0, area = 0; area < nr_vms; area++) {
3228 start = offsets[area];
3229 end = start + sizes[area];
3230
3231 /* is everything aligned properly? */
3232 BUG_ON(!IS_ALIGNED(offsets[area], align));
3233 BUG_ON(!IS_ALIGNED(sizes[area], align));
3234
3235 /* detect the area with the highest address */
3236 if (start > offsets[last_area])
3237 last_area = area;
3238
3239 for (area2 = area + 1; area2 < nr_vms; area2++) {
3240 unsigned long start2 = offsets[area2];
3241 unsigned long end2 = start2 + sizes[area2];
3242
3243 BUG_ON(start2 < end && start < end2);
3244 }
3245 }
3246 last_end = offsets[last_area] + sizes[last_area];
3247
3248 if (vmalloc_end - vmalloc_start < last_end) {
3249 WARN_ON(true);
3250 return NULL;
3251 }
3252
3253 vms = kcalloc(nr_vms, sizeof(vms[0]), GFP_KERNEL);
3254 vas = kcalloc(nr_vms, sizeof(vas[0]), GFP_KERNEL);
3255 if (!vas || !vms)
3256 goto err_free2;
3257
3258 for (area = 0; area < nr_vms; area++) {
3259 vas[area] = kmem_cache_zalloc(vmap_area_cachep, GFP_KERNEL);
3260 vms[area] = kzalloc(sizeof(struct vm_struct), GFP_KERNEL);
3261 if (!vas[area] || !vms[area])
3262 goto err_free;
3263 }
3264retry:
3265 spin_lock(&vmap_area_lock);
3266
3267 /* start scanning - we scan from the top, begin with the last area */
3268 area = term_area = last_area;
3269 start = offsets[area];
3270 end = start + sizes[area];
3271
3272 va = pvm_find_va_enclose_addr(vmalloc_end);
3273 base = pvm_determine_end_from_reverse(&va, align) - end;
3274
3275 while (true) {
3276 /*
3277 * base might have underflowed, add last_end before
3278 * comparing.
3279 */
3280 if (base + last_end < vmalloc_start + last_end)
3281 goto overflow;
3282
3283 /*
3284 * Fitting base has not been found.
3285 */
3286 if (va == NULL)
3287 goto overflow;
3288
3289 /*
3290 * If required width exeeds current VA block, move
3291 * base downwards and then recheck.
3292 */
3293 if (base + end > va->va_end) {
3294 base = pvm_determine_end_from_reverse(&va, align) - end;
3295 term_area = area;
3296 continue;
3297 }
3298
3299 /*
3300 * If this VA does not fit, move base downwards and recheck.
3301 */
3302 if (base + start < va->va_start) {
3303 va = node_to_va(rb_prev(&va->rb_node));
3304 base = pvm_determine_end_from_reverse(&va, align) - end;
3305 term_area = area;
3306 continue;
3307 }
3308
3309 /*
3310 * This area fits, move on to the previous one. If
3311 * the previous one is the terminal one, we're done.
3312 */
3313 area = (area + nr_vms - 1) % nr_vms;
3314 if (area == term_area)
3315 break;
3316
3317 start = offsets[area];
3318 end = start + sizes[area];
3319 va = pvm_find_va_enclose_addr(base + end);
3320 }
3321
3322 /* we've found a fitting base, insert all va's */
3323 for (area = 0; area < nr_vms; area++) {
3324 int ret;
3325
3326 start = base + offsets[area];
3327 size = sizes[area];
3328
3329 va = pvm_find_va_enclose_addr(start);
3330 if (WARN_ON_ONCE(va == NULL))
3331 /* It is a BUG(), but trigger recovery instead. */
3332 goto recovery;
3333
3334 type = classify_va_fit_type(va, start, size);
3335 if (WARN_ON_ONCE(type == NOTHING_FIT))
3336 /* It is a BUG(), but trigger recovery instead. */
3337 goto recovery;
3338
3339 ret = adjust_va_to_fit_type(va, start, size, type);
3340 if (unlikely(ret))
3341 goto recovery;
3342
3343 /* Allocated area. */
3344 va = vas[area];
3345 va->va_start = start;
3346 va->va_end = start + size;
3347
3348 insert_vmap_area(va, &vmap_area_root, &vmap_area_list);
3349 }
3350
3351 spin_unlock(&vmap_area_lock);
3352
3353 /* insert all vm's */
3354 for (area = 0; area < nr_vms; area++)
3355 setup_vmalloc_vm(vms[area], vas[area], VM_ALLOC,
3356 pcpu_get_vm_areas);
3357
3358 kfree(vas);
3359 return vms;
3360
3361recovery:
3362 /* Remove previously inserted areas. */
3363 while (area--) {
3364 __free_vmap_area(vas[area]);
3365 vas[area] = NULL;
3366 }
3367
3368overflow:
3369 spin_unlock(&vmap_area_lock);
3370 if (!purged) {
3371 purge_vmap_area_lazy();
3372 purged = true;
3373
3374 /* Before "retry", check if we recover. */
3375 for (area = 0; area < nr_vms; area++) {
3376 if (vas[area])
3377 continue;
3378
3379 vas[area] = kmem_cache_zalloc(
3380 vmap_area_cachep, GFP_KERNEL);
3381 if (!vas[area])
3382 goto err_free;
3383 }
3384
3385 goto retry;
3386 }
3387
3388err_free:
3389 for (area = 0; area < nr_vms; area++) {
3390 if (vas[area])
3391 kmem_cache_free(vmap_area_cachep, vas[area]);
3392
3393 kfree(vms[area]);
3394 }
3395err_free2:
3396 kfree(vas);
3397 kfree(vms);
3398 return NULL;
3399}
3400
3401/**
3402 * pcpu_free_vm_areas - free vmalloc areas for percpu allocator
3403 * @vms: vm_struct pointer array returned by pcpu_get_vm_areas()
3404 * @nr_vms: the number of allocated areas
3405 *
3406 * Free vm_structs and the array allocated by pcpu_get_vm_areas().
3407 */
3408void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms)
3409{
3410 int i;
3411
3412 for (i = 0; i < nr_vms; i++)
3413 free_vm_area(vms[i]);
3414 kfree(vms);
3415}
3416#endif /* CONFIG_SMP */
3417
3418#ifdef CONFIG_PROC_FS
3419static void *s_start(struct seq_file *m, loff_t *pos)
3420 __acquires(&vmap_area_lock)
3421{
3422 spin_lock(&vmap_area_lock);
3423 return seq_list_start(&vmap_area_list, *pos);
3424}
3425
3426static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3427{
3428 return seq_list_next(p, &vmap_area_list, pos);
3429}
3430
3431static void s_stop(struct seq_file *m, void *p)
3432 __releases(&vmap_area_lock)
3433{
3434 spin_unlock(&vmap_area_lock);
3435}
3436
3437static void show_numa_info(struct seq_file *m, struct vm_struct *v)
3438{
3439 if (IS_ENABLED(CONFIG_NUMA)) {
3440 unsigned int nr, *counters = m->private;
3441
3442 if (!counters)
3443 return;
3444
3445 if (v->flags & VM_UNINITIALIZED)
3446 return;
3447 /* Pair with smp_wmb() in clear_vm_uninitialized_flag() */
3448 smp_rmb();
3449
3450 memset(counters, 0, nr_node_ids * sizeof(unsigned int));
3451
3452 for (nr = 0; nr < v->nr_pages; nr++)
3453 counters[page_to_nid(v->pages[nr])]++;
3454
3455 for_each_node_state(nr, N_HIGH_MEMORY)
3456 if (counters[nr])
3457 seq_printf(m, " N%u=%u", nr, counters[nr]);
3458 }
3459}
3460
3461static void show_purge_info(struct seq_file *m)
3462{
3463 struct llist_node *head;
3464 struct vmap_area *va;
3465
3466 head = READ_ONCE(vmap_purge_list.first);
3467 if (head == NULL)
3468 return;
3469
3470 llist_for_each_entry(va, head, purge_list) {
3471 seq_printf(m, "0x%pK-0x%pK %7ld unpurged vm_area\n",
3472 (void *)va->va_start, (void *)va->va_end,
3473 va->va_end - va->va_start);
3474 }
3475}
3476
3477static int s_show(struct seq_file *m, void *p)
3478{
3479 struct vmap_area *va;
3480 struct vm_struct *v;
3481
3482 va = list_entry(p, struct vmap_area, list);
3483
3484 /*
3485 * s_show can encounter race with remove_vm_area, !vm on behalf
3486 * of vmap area is being tear down or vm_map_ram allocation.
3487 */
3488 if (!va->vm) {
3489 seq_printf(m, "0x%pK-0x%pK %7ld vm_map_ram\n",
3490 (void *)va->va_start, (void *)va->va_end,
3491 va->va_end - va->va_start);
3492
3493 return 0;
3494 }
3495
3496 v = va->vm;
3497
3498 seq_printf(m, "0x%pK-0x%pK %7ld",
3499 v->addr, v->addr + v->size, v->size);
3500
3501 if (v->caller)
3502 seq_printf(m, " %pS", v->caller);
3503
3504 if (v->nr_pages)
3505 seq_printf(m, " pages=%d", v->nr_pages);
3506
3507 if (v->phys_addr)
3508 seq_printf(m, " phys=%pa", &v->phys_addr);
3509
3510 if (v->flags & VM_IOREMAP)
3511 seq_puts(m, " ioremap");
3512
3513 if (v->flags & VM_ALLOC)
3514 seq_puts(m, " vmalloc");
3515
3516 if (v->flags & VM_MAP)
3517 seq_puts(m, " vmap");
3518
3519 if (v->flags & VM_USERMAP)
3520 seq_puts(m, " user");
3521
3522 if (v->flags & VM_DMA_COHERENT)
3523 seq_puts(m, " dma-coherent");
3524
3525 if (is_vmalloc_addr(v->pages))
3526 seq_puts(m, " vpages");
3527
3528 show_numa_info(m, v);
3529 seq_putc(m, '\n');
3530
3531 /*
3532 * As a final step, dump "unpurged" areas. Note,
3533 * that entire "/proc/vmallocinfo" output will not
3534 * be address sorted, because the purge list is not
3535 * sorted.
3536 */
3537 if (list_is_last(&va->list, &vmap_area_list))
3538 show_purge_info(m);
3539
3540 return 0;
3541}
3542
3543static const struct seq_operations vmalloc_op = {
3544 .start = s_start,
3545 .next = s_next,
3546 .stop = s_stop,
3547 .show = s_show,
3548};
3549
3550static int __init proc_vmalloc_init(void)
3551{
3552 if (IS_ENABLED(CONFIG_NUMA))
3553 proc_create_seq_private("vmallocinfo", 0400, NULL,
3554 &vmalloc_op,
3555 nr_node_ids * sizeof(unsigned int), NULL);
3556 else
3557 proc_create_seq("vmallocinfo", 0400, NULL, &vmalloc_op);
3558 return 0;
3559}
3560module_init(proc_vmalloc_init);
3561
3562#endif