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