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