Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame^] | 1 | /* |
| 2 | * PPC64 (POWER4) Huge TLB Page Support for Kernel. |
| 3 | * |
| 4 | * Copyright (C) 2003 David Gibson, IBM Corporation. |
| 5 | * |
| 6 | * Based on the IA-32 version: |
| 7 | * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com> |
| 8 | */ |
| 9 | |
| 10 | #include <linux/init.h> |
| 11 | #include <linux/fs.h> |
| 12 | #include <linux/mm.h> |
| 13 | #include <linux/hugetlb.h> |
| 14 | #include <linux/pagemap.h> |
| 15 | #include <linux/smp_lock.h> |
| 16 | #include <linux/slab.h> |
| 17 | #include <linux/err.h> |
| 18 | #include <linux/sysctl.h> |
| 19 | #include <asm/mman.h> |
| 20 | #include <asm/pgalloc.h> |
| 21 | #include <asm/tlb.h> |
| 22 | #include <asm/tlbflush.h> |
| 23 | #include <asm/mmu_context.h> |
| 24 | #include <asm/machdep.h> |
| 25 | #include <asm/cputable.h> |
| 26 | #include <asm/tlb.h> |
| 27 | |
| 28 | #include <linux/sysctl.h> |
| 29 | |
| 30 | #define HUGEPGDIR_SHIFT (HPAGE_SHIFT + PAGE_SHIFT - 3) |
| 31 | #define HUGEPGDIR_SIZE (1UL << HUGEPGDIR_SHIFT) |
| 32 | #define HUGEPGDIR_MASK (~(HUGEPGDIR_SIZE-1)) |
| 33 | |
| 34 | #define HUGEPTE_INDEX_SIZE 9 |
| 35 | #define HUGEPGD_INDEX_SIZE 10 |
| 36 | |
| 37 | #define PTRS_PER_HUGEPTE (1 << HUGEPTE_INDEX_SIZE) |
| 38 | #define PTRS_PER_HUGEPGD (1 << HUGEPGD_INDEX_SIZE) |
| 39 | |
| 40 | static inline int hugepgd_index(unsigned long addr) |
| 41 | { |
| 42 | return (addr & ~REGION_MASK) >> HUGEPGDIR_SHIFT; |
| 43 | } |
| 44 | |
| 45 | static pgd_t *hugepgd_offset(struct mm_struct *mm, unsigned long addr) |
| 46 | { |
| 47 | int index; |
| 48 | |
| 49 | if (! mm->context.huge_pgdir) |
| 50 | return NULL; |
| 51 | |
| 52 | |
| 53 | index = hugepgd_index(addr); |
| 54 | BUG_ON(index >= PTRS_PER_HUGEPGD); |
| 55 | return mm->context.huge_pgdir + index; |
| 56 | } |
| 57 | |
| 58 | static inline pte_t *hugepte_offset(pgd_t *dir, unsigned long addr) |
| 59 | { |
| 60 | int index; |
| 61 | |
| 62 | if (pgd_none(*dir)) |
| 63 | return NULL; |
| 64 | |
| 65 | index = (addr >> HPAGE_SHIFT) % PTRS_PER_HUGEPTE; |
| 66 | return (pte_t *)pgd_page(*dir) + index; |
| 67 | } |
| 68 | |
| 69 | static pgd_t *hugepgd_alloc(struct mm_struct *mm, unsigned long addr) |
| 70 | { |
| 71 | BUG_ON(! in_hugepage_area(mm->context, addr)); |
| 72 | |
| 73 | if (! mm->context.huge_pgdir) { |
| 74 | pgd_t *new; |
| 75 | spin_unlock(&mm->page_table_lock); |
| 76 | /* Don't use pgd_alloc(), because we want __GFP_REPEAT */ |
| 77 | new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT); |
| 78 | BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE)); |
| 79 | spin_lock(&mm->page_table_lock); |
| 80 | |
| 81 | /* |
| 82 | * Because we dropped the lock, we should re-check the |
| 83 | * entry, as somebody else could have populated it.. |
| 84 | */ |
| 85 | if (mm->context.huge_pgdir) |
| 86 | pgd_free(new); |
| 87 | else |
| 88 | mm->context.huge_pgdir = new; |
| 89 | } |
| 90 | return hugepgd_offset(mm, addr); |
| 91 | } |
| 92 | |
| 93 | static pte_t *hugepte_alloc(struct mm_struct *mm, pgd_t *dir, |
| 94 | unsigned long addr) |
| 95 | { |
| 96 | if (! pgd_present(*dir)) { |
| 97 | pte_t *new; |
| 98 | |
| 99 | spin_unlock(&mm->page_table_lock); |
| 100 | new = kmem_cache_alloc(zero_cache, GFP_KERNEL | __GFP_REPEAT); |
| 101 | BUG_ON(memcmp(new, empty_zero_page, PAGE_SIZE)); |
| 102 | spin_lock(&mm->page_table_lock); |
| 103 | /* |
| 104 | * Because we dropped the lock, we should re-check the |
| 105 | * entry, as somebody else could have populated it.. |
| 106 | */ |
| 107 | if (pgd_present(*dir)) { |
| 108 | if (new) |
| 109 | kmem_cache_free(zero_cache, new); |
| 110 | } else { |
| 111 | struct page *ptepage; |
| 112 | |
| 113 | if (! new) |
| 114 | return NULL; |
| 115 | ptepage = virt_to_page(new); |
| 116 | ptepage->mapping = (void *) mm; |
| 117 | ptepage->index = addr & HUGEPGDIR_MASK; |
| 118 | pgd_populate(mm, dir, new); |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | return hugepte_offset(dir, addr); |
| 123 | } |
| 124 | |
| 125 | static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr) |
| 126 | { |
| 127 | pgd_t *pgd; |
| 128 | |
| 129 | BUG_ON(! in_hugepage_area(mm->context, addr)); |
| 130 | |
| 131 | pgd = hugepgd_offset(mm, addr); |
| 132 | if (! pgd) |
| 133 | return NULL; |
| 134 | |
| 135 | return hugepte_offset(pgd, addr); |
| 136 | } |
| 137 | |
| 138 | static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr) |
| 139 | { |
| 140 | pgd_t *pgd; |
| 141 | |
| 142 | BUG_ON(! in_hugepage_area(mm->context, addr)); |
| 143 | |
| 144 | pgd = hugepgd_alloc(mm, addr); |
| 145 | if (! pgd) |
| 146 | return NULL; |
| 147 | |
| 148 | return hugepte_alloc(mm, pgd, addr); |
| 149 | } |
| 150 | |
| 151 | static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma, |
| 152 | unsigned long addr, struct page *page, |
| 153 | pte_t *ptep, int write_access) |
| 154 | { |
| 155 | pte_t entry; |
| 156 | |
| 157 | add_mm_counter(mm, rss, HPAGE_SIZE / PAGE_SIZE); |
| 158 | if (write_access) { |
| 159 | entry = |
| 160 | pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot))); |
| 161 | } else { |
| 162 | entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot)); |
| 163 | } |
| 164 | entry = pte_mkyoung(entry); |
| 165 | entry = pte_mkhuge(entry); |
| 166 | |
| 167 | set_pte_at(mm, addr, ptep, entry); |
| 168 | } |
| 169 | |
| 170 | /* |
| 171 | * This function checks for proper alignment of input addr and len parameters. |
| 172 | */ |
| 173 | int is_aligned_hugepage_range(unsigned long addr, unsigned long len) |
| 174 | { |
| 175 | if (len & ~HPAGE_MASK) |
| 176 | return -EINVAL; |
| 177 | if (addr & ~HPAGE_MASK) |
| 178 | return -EINVAL; |
| 179 | if (! (within_hugepage_low_range(addr, len) |
| 180 | || within_hugepage_high_range(addr, len)) ) |
| 181 | return -EINVAL; |
| 182 | return 0; |
| 183 | } |
| 184 | |
| 185 | static void flush_segments(void *parm) |
| 186 | { |
| 187 | u16 segs = (unsigned long) parm; |
| 188 | unsigned long i; |
| 189 | |
| 190 | asm volatile("isync" : : : "memory"); |
| 191 | |
| 192 | for (i = 0; i < 16; i++) { |
| 193 | if (! (segs & (1U << i))) |
| 194 | continue; |
| 195 | asm volatile("slbie %0" : : "r" (i << SID_SHIFT)); |
| 196 | } |
| 197 | |
| 198 | asm volatile("isync" : : : "memory"); |
| 199 | } |
| 200 | |
| 201 | static int prepare_low_seg_for_htlb(struct mm_struct *mm, unsigned long seg) |
| 202 | { |
| 203 | unsigned long start = seg << SID_SHIFT; |
| 204 | unsigned long end = (seg+1) << SID_SHIFT; |
| 205 | struct vm_area_struct *vma; |
| 206 | unsigned long addr; |
| 207 | struct mmu_gather *tlb; |
| 208 | |
| 209 | BUG_ON(seg >= 16); |
| 210 | |
| 211 | /* Check no VMAs are in the region */ |
| 212 | vma = find_vma(mm, start); |
| 213 | if (vma && (vma->vm_start < end)) |
| 214 | return -EBUSY; |
| 215 | |
| 216 | /* Clean up any leftover PTE pages in the region */ |
| 217 | spin_lock(&mm->page_table_lock); |
| 218 | tlb = tlb_gather_mmu(mm, 0); |
| 219 | for (addr = start; addr < end; addr += PMD_SIZE) { |
| 220 | pgd_t *pgd = pgd_offset(mm, addr); |
| 221 | pmd_t *pmd; |
| 222 | struct page *page; |
| 223 | pte_t *pte; |
| 224 | int i; |
| 225 | |
| 226 | if (pgd_none(*pgd)) |
| 227 | continue; |
| 228 | pmd = pmd_offset(pgd, addr); |
| 229 | if (!pmd || pmd_none(*pmd)) |
| 230 | continue; |
| 231 | if (pmd_bad(*pmd)) { |
| 232 | pmd_ERROR(*pmd); |
| 233 | pmd_clear(pmd); |
| 234 | continue; |
| 235 | } |
| 236 | pte = (pte_t *)pmd_page_kernel(*pmd); |
| 237 | /* No VMAs, so there should be no PTEs, check just in case. */ |
| 238 | for (i = 0; i < PTRS_PER_PTE; i++) { |
| 239 | BUG_ON(!pte_none(*pte)); |
| 240 | pte++; |
| 241 | } |
| 242 | page = pmd_page(*pmd); |
| 243 | pmd_clear(pmd); |
| 244 | mm->nr_ptes--; |
| 245 | dec_page_state(nr_page_table_pages); |
| 246 | pte_free_tlb(tlb, page); |
| 247 | } |
| 248 | tlb_finish_mmu(tlb, start, end); |
| 249 | spin_unlock(&mm->page_table_lock); |
| 250 | |
| 251 | return 0; |
| 252 | } |
| 253 | |
| 254 | static int open_low_hpage_segs(struct mm_struct *mm, u16 newsegs) |
| 255 | { |
| 256 | unsigned long i; |
| 257 | |
| 258 | newsegs &= ~(mm->context.htlb_segs); |
| 259 | if (! newsegs) |
| 260 | return 0; /* The segments we want are already open */ |
| 261 | |
| 262 | for (i = 0; i < 16; i++) |
| 263 | if ((1 << i) & newsegs) |
| 264 | if (prepare_low_seg_for_htlb(mm, i) != 0) |
| 265 | return -EBUSY; |
| 266 | |
| 267 | mm->context.htlb_segs |= newsegs; |
| 268 | |
| 269 | /* update the paca copy of the context struct */ |
| 270 | get_paca()->context = mm->context; |
| 271 | |
| 272 | /* the context change must make it to memory before the flush, |
| 273 | * so that further SLB misses do the right thing. */ |
| 274 | mb(); |
| 275 | on_each_cpu(flush_segments, (void *)(unsigned long)newsegs, 0, 1); |
| 276 | |
| 277 | return 0; |
| 278 | } |
| 279 | |
| 280 | int prepare_hugepage_range(unsigned long addr, unsigned long len) |
| 281 | { |
| 282 | if (within_hugepage_high_range(addr, len)) |
| 283 | return 0; |
| 284 | else if ((addr < 0x100000000UL) && ((addr+len) < 0x100000000UL)) { |
| 285 | int err; |
| 286 | /* Yes, we need both tests, in case addr+len overflows |
| 287 | * 64-bit arithmetic */ |
| 288 | err = open_low_hpage_segs(current->mm, |
| 289 | LOW_ESID_MASK(addr, len)); |
| 290 | if (err) |
| 291 | printk(KERN_DEBUG "prepare_hugepage_range(%lx, %lx)" |
| 292 | " failed (segs: 0x%04hx)\n", addr, len, |
| 293 | LOW_ESID_MASK(addr, len)); |
| 294 | return err; |
| 295 | } |
| 296 | |
| 297 | return -EINVAL; |
| 298 | } |
| 299 | |
| 300 | int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src, |
| 301 | struct vm_area_struct *vma) |
| 302 | { |
| 303 | pte_t *src_pte, *dst_pte, entry; |
| 304 | struct page *ptepage; |
| 305 | unsigned long addr = vma->vm_start; |
| 306 | unsigned long end = vma->vm_end; |
| 307 | int err = -ENOMEM; |
| 308 | |
| 309 | while (addr < end) { |
| 310 | dst_pte = huge_pte_alloc(dst, addr); |
| 311 | if (!dst_pte) |
| 312 | goto out; |
| 313 | |
| 314 | src_pte = huge_pte_offset(src, addr); |
| 315 | entry = *src_pte; |
| 316 | |
| 317 | ptepage = pte_page(entry); |
| 318 | get_page(ptepage); |
| 319 | add_mm_counter(dst, rss, HPAGE_SIZE / PAGE_SIZE); |
| 320 | set_pte_at(dst, addr, dst_pte, entry); |
| 321 | |
| 322 | addr += HPAGE_SIZE; |
| 323 | } |
| 324 | |
| 325 | err = 0; |
| 326 | out: |
| 327 | return err; |
| 328 | } |
| 329 | |
| 330 | int |
| 331 | follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma, |
| 332 | struct page **pages, struct vm_area_struct **vmas, |
| 333 | unsigned long *position, int *length, int i) |
| 334 | { |
| 335 | unsigned long vpfn, vaddr = *position; |
| 336 | int remainder = *length; |
| 337 | |
| 338 | WARN_ON(!is_vm_hugetlb_page(vma)); |
| 339 | |
| 340 | vpfn = vaddr/PAGE_SIZE; |
| 341 | while (vaddr < vma->vm_end && remainder) { |
| 342 | if (pages) { |
| 343 | pte_t *pte; |
| 344 | struct page *page; |
| 345 | |
| 346 | pte = huge_pte_offset(mm, vaddr); |
| 347 | |
| 348 | /* hugetlb should be locked, and hence, prefaulted */ |
| 349 | WARN_ON(!pte || pte_none(*pte)); |
| 350 | |
| 351 | page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)]; |
| 352 | |
| 353 | WARN_ON(!PageCompound(page)); |
| 354 | |
| 355 | get_page(page); |
| 356 | pages[i] = page; |
| 357 | } |
| 358 | |
| 359 | if (vmas) |
| 360 | vmas[i] = vma; |
| 361 | |
| 362 | vaddr += PAGE_SIZE; |
| 363 | ++vpfn; |
| 364 | --remainder; |
| 365 | ++i; |
| 366 | } |
| 367 | |
| 368 | *length = remainder; |
| 369 | *position = vaddr; |
| 370 | |
| 371 | return i; |
| 372 | } |
| 373 | |
| 374 | struct page * |
| 375 | follow_huge_addr(struct mm_struct *mm, unsigned long address, int write) |
| 376 | { |
| 377 | pte_t *ptep; |
| 378 | struct page *page; |
| 379 | |
| 380 | if (! in_hugepage_area(mm->context, address)) |
| 381 | return ERR_PTR(-EINVAL); |
| 382 | |
| 383 | ptep = huge_pte_offset(mm, address); |
| 384 | page = pte_page(*ptep); |
| 385 | if (page) |
| 386 | page += (address % HPAGE_SIZE) / PAGE_SIZE; |
| 387 | |
| 388 | return page; |
| 389 | } |
| 390 | |
| 391 | int pmd_huge(pmd_t pmd) |
| 392 | { |
| 393 | return 0; |
| 394 | } |
| 395 | |
| 396 | struct page * |
| 397 | follow_huge_pmd(struct mm_struct *mm, unsigned long address, |
| 398 | pmd_t *pmd, int write) |
| 399 | { |
| 400 | BUG(); |
| 401 | return NULL; |
| 402 | } |
| 403 | |
| 404 | void unmap_hugepage_range(struct vm_area_struct *vma, |
| 405 | unsigned long start, unsigned long end) |
| 406 | { |
| 407 | struct mm_struct *mm = vma->vm_mm; |
| 408 | unsigned long addr; |
| 409 | pte_t *ptep; |
| 410 | struct page *page; |
| 411 | |
| 412 | WARN_ON(!is_vm_hugetlb_page(vma)); |
| 413 | BUG_ON((start % HPAGE_SIZE) != 0); |
| 414 | BUG_ON((end % HPAGE_SIZE) != 0); |
| 415 | |
| 416 | for (addr = start; addr < end; addr += HPAGE_SIZE) { |
| 417 | pte_t pte; |
| 418 | |
| 419 | ptep = huge_pte_offset(mm, addr); |
| 420 | if (!ptep || pte_none(*ptep)) |
| 421 | continue; |
| 422 | |
| 423 | pte = *ptep; |
| 424 | page = pte_page(pte); |
| 425 | pte_clear(mm, addr, ptep); |
| 426 | |
| 427 | put_page(page); |
| 428 | } |
| 429 | add_mm_counter(mm, rss, -((end - start) >> PAGE_SHIFT)); |
| 430 | flush_tlb_pending(); |
| 431 | } |
| 432 | |
| 433 | void hugetlb_free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *prev, |
| 434 | unsigned long start, unsigned long end) |
| 435 | { |
| 436 | /* Because the huge pgtables are only 2 level, they can take |
| 437 | * at most around 4M, much less than one hugepage which the |
| 438 | * process is presumably entitled to use. So we don't bother |
| 439 | * freeing up the pagetables on unmap, and wait until |
| 440 | * destroy_context() to clean up the lot. */ |
| 441 | } |
| 442 | |
| 443 | int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma) |
| 444 | { |
| 445 | struct mm_struct *mm = current->mm; |
| 446 | unsigned long addr; |
| 447 | int ret = 0; |
| 448 | |
| 449 | WARN_ON(!is_vm_hugetlb_page(vma)); |
| 450 | BUG_ON((vma->vm_start % HPAGE_SIZE) != 0); |
| 451 | BUG_ON((vma->vm_end % HPAGE_SIZE) != 0); |
| 452 | |
| 453 | spin_lock(&mm->page_table_lock); |
| 454 | for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) { |
| 455 | unsigned long idx; |
| 456 | pte_t *pte = huge_pte_alloc(mm, addr); |
| 457 | struct page *page; |
| 458 | |
| 459 | if (!pte) { |
| 460 | ret = -ENOMEM; |
| 461 | goto out; |
| 462 | } |
| 463 | if (! pte_none(*pte)) |
| 464 | continue; |
| 465 | |
| 466 | idx = ((addr - vma->vm_start) >> HPAGE_SHIFT) |
| 467 | + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT)); |
| 468 | page = find_get_page(mapping, idx); |
| 469 | if (!page) { |
| 470 | /* charge the fs quota first */ |
| 471 | if (hugetlb_get_quota(mapping)) { |
| 472 | ret = -ENOMEM; |
| 473 | goto out; |
| 474 | } |
| 475 | page = alloc_huge_page(); |
| 476 | if (!page) { |
| 477 | hugetlb_put_quota(mapping); |
| 478 | ret = -ENOMEM; |
| 479 | goto out; |
| 480 | } |
| 481 | ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC); |
| 482 | if (! ret) { |
| 483 | unlock_page(page); |
| 484 | } else { |
| 485 | hugetlb_put_quota(mapping); |
| 486 | free_huge_page(page); |
| 487 | goto out; |
| 488 | } |
| 489 | } |
| 490 | set_huge_pte(mm, vma, addr, page, pte, vma->vm_flags & VM_WRITE); |
| 491 | } |
| 492 | out: |
| 493 | spin_unlock(&mm->page_table_lock); |
| 494 | return ret; |
| 495 | } |
| 496 | |
| 497 | /* Because we have an exclusive hugepage region which lies within the |
| 498 | * normal user address space, we have to take special measures to make |
| 499 | * non-huge mmap()s evade the hugepage reserved regions. */ |
| 500 | unsigned long arch_get_unmapped_area(struct file *filp, unsigned long addr, |
| 501 | unsigned long len, unsigned long pgoff, |
| 502 | unsigned long flags) |
| 503 | { |
| 504 | struct mm_struct *mm = current->mm; |
| 505 | struct vm_area_struct *vma; |
| 506 | unsigned long start_addr; |
| 507 | |
| 508 | if (len > TASK_SIZE) |
| 509 | return -ENOMEM; |
| 510 | |
| 511 | if (addr) { |
| 512 | addr = PAGE_ALIGN(addr); |
| 513 | vma = find_vma(mm, addr); |
| 514 | if (((TASK_SIZE - len) >= addr) |
| 515 | && (!vma || (addr+len) <= vma->vm_start) |
| 516 | && !is_hugepage_only_range(mm, addr,len)) |
| 517 | return addr; |
| 518 | } |
| 519 | start_addr = addr = mm->free_area_cache; |
| 520 | |
| 521 | full_search: |
| 522 | vma = find_vma(mm, addr); |
| 523 | while (TASK_SIZE - len >= addr) { |
| 524 | BUG_ON(vma && (addr >= vma->vm_end)); |
| 525 | |
| 526 | if (touches_hugepage_low_range(mm, addr, len)) { |
| 527 | addr = ALIGN(addr+1, 1<<SID_SHIFT); |
| 528 | vma = find_vma(mm, addr); |
| 529 | continue; |
| 530 | } |
| 531 | if (touches_hugepage_high_range(addr, len)) { |
| 532 | addr = TASK_HPAGE_END; |
| 533 | vma = find_vma(mm, addr); |
| 534 | continue; |
| 535 | } |
| 536 | if (!vma || addr + len <= vma->vm_start) { |
| 537 | /* |
| 538 | * Remember the place where we stopped the search: |
| 539 | */ |
| 540 | mm->free_area_cache = addr + len; |
| 541 | return addr; |
| 542 | } |
| 543 | addr = vma->vm_end; |
| 544 | vma = vma->vm_next; |
| 545 | } |
| 546 | |
| 547 | /* Make sure we didn't miss any holes */ |
| 548 | if (start_addr != TASK_UNMAPPED_BASE) { |
| 549 | start_addr = addr = TASK_UNMAPPED_BASE; |
| 550 | goto full_search; |
| 551 | } |
| 552 | return -ENOMEM; |
| 553 | } |
| 554 | |
| 555 | /* |
| 556 | * This mmap-allocator allocates new areas top-down from below the |
| 557 | * stack's low limit (the base): |
| 558 | * |
| 559 | * Because we have an exclusive hugepage region which lies within the |
| 560 | * normal user address space, we have to take special measures to make |
| 561 | * non-huge mmap()s evade the hugepage reserved regions. |
| 562 | */ |
| 563 | unsigned long |
| 564 | arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0, |
| 565 | const unsigned long len, const unsigned long pgoff, |
| 566 | const unsigned long flags) |
| 567 | { |
| 568 | struct vm_area_struct *vma, *prev_vma; |
| 569 | struct mm_struct *mm = current->mm; |
| 570 | unsigned long base = mm->mmap_base, addr = addr0; |
| 571 | int first_time = 1; |
| 572 | |
| 573 | /* requested length too big for entire address space */ |
| 574 | if (len > TASK_SIZE) |
| 575 | return -ENOMEM; |
| 576 | |
| 577 | /* dont allow allocations above current base */ |
| 578 | if (mm->free_area_cache > base) |
| 579 | mm->free_area_cache = base; |
| 580 | |
| 581 | /* requesting a specific address */ |
| 582 | if (addr) { |
| 583 | addr = PAGE_ALIGN(addr); |
| 584 | vma = find_vma(mm, addr); |
| 585 | if (TASK_SIZE - len >= addr && |
| 586 | (!vma || addr + len <= vma->vm_start) |
| 587 | && !is_hugepage_only_range(mm, addr,len)) |
| 588 | return addr; |
| 589 | } |
| 590 | |
| 591 | try_again: |
| 592 | /* make sure it can fit in the remaining address space */ |
| 593 | if (mm->free_area_cache < len) |
| 594 | goto fail; |
| 595 | |
| 596 | /* either no address requested or cant fit in requested address hole */ |
| 597 | addr = (mm->free_area_cache - len) & PAGE_MASK; |
| 598 | do { |
| 599 | hugepage_recheck: |
| 600 | if (touches_hugepage_low_range(mm, addr, len)) { |
| 601 | addr = (addr & ((~0) << SID_SHIFT)) - len; |
| 602 | goto hugepage_recheck; |
| 603 | } else if (touches_hugepage_high_range(addr, len)) { |
| 604 | addr = TASK_HPAGE_BASE - len; |
| 605 | } |
| 606 | |
| 607 | /* |
| 608 | * Lookup failure means no vma is above this address, |
| 609 | * i.e. return with success: |
| 610 | */ |
| 611 | if (!(vma = find_vma_prev(mm, addr, &prev_vma))) |
| 612 | return addr; |
| 613 | |
| 614 | /* |
| 615 | * new region fits between prev_vma->vm_end and |
| 616 | * vma->vm_start, use it: |
| 617 | */ |
| 618 | if (addr+len <= vma->vm_start && |
| 619 | (!prev_vma || (addr >= prev_vma->vm_end))) |
| 620 | /* remember the address as a hint for next time */ |
| 621 | return (mm->free_area_cache = addr); |
| 622 | else |
| 623 | /* pull free_area_cache down to the first hole */ |
| 624 | if (mm->free_area_cache == vma->vm_end) |
| 625 | mm->free_area_cache = vma->vm_start; |
| 626 | |
| 627 | /* try just below the current vma->vm_start */ |
| 628 | addr = vma->vm_start-len; |
| 629 | } while (len <= vma->vm_start); |
| 630 | |
| 631 | fail: |
| 632 | /* |
| 633 | * if hint left us with no space for the requested |
| 634 | * mapping then try again: |
| 635 | */ |
| 636 | if (first_time) { |
| 637 | mm->free_area_cache = base; |
| 638 | first_time = 0; |
| 639 | goto try_again; |
| 640 | } |
| 641 | /* |
| 642 | * A failed mmap() very likely causes application failure, |
| 643 | * so fall back to the bottom-up function here. This scenario |
| 644 | * can happen with large stack limits and large mmap() |
| 645 | * allocations. |
| 646 | */ |
| 647 | mm->free_area_cache = TASK_UNMAPPED_BASE; |
| 648 | addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags); |
| 649 | /* |
| 650 | * Restore the topdown base: |
| 651 | */ |
| 652 | mm->free_area_cache = base; |
| 653 | |
| 654 | return addr; |
| 655 | } |
| 656 | |
| 657 | static unsigned long htlb_get_low_area(unsigned long len, u16 segmask) |
| 658 | { |
| 659 | unsigned long addr = 0; |
| 660 | struct vm_area_struct *vma; |
| 661 | |
| 662 | vma = find_vma(current->mm, addr); |
| 663 | while (addr + len <= 0x100000000UL) { |
| 664 | BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */ |
| 665 | |
| 666 | if (! __within_hugepage_low_range(addr, len, segmask)) { |
| 667 | addr = ALIGN(addr+1, 1<<SID_SHIFT); |
| 668 | vma = find_vma(current->mm, addr); |
| 669 | continue; |
| 670 | } |
| 671 | |
| 672 | if (!vma || (addr + len) <= vma->vm_start) |
| 673 | return addr; |
| 674 | addr = ALIGN(vma->vm_end, HPAGE_SIZE); |
| 675 | /* Depending on segmask this might not be a confirmed |
| 676 | * hugepage region, so the ALIGN could have skipped |
| 677 | * some VMAs */ |
| 678 | vma = find_vma(current->mm, addr); |
| 679 | } |
| 680 | |
| 681 | return -ENOMEM; |
| 682 | } |
| 683 | |
| 684 | static unsigned long htlb_get_high_area(unsigned long len) |
| 685 | { |
| 686 | unsigned long addr = TASK_HPAGE_BASE; |
| 687 | struct vm_area_struct *vma; |
| 688 | |
| 689 | vma = find_vma(current->mm, addr); |
| 690 | for (vma = find_vma(current->mm, addr); |
| 691 | addr + len <= TASK_HPAGE_END; |
| 692 | vma = vma->vm_next) { |
| 693 | BUG_ON(vma && (addr >= vma->vm_end)); /* invariant */ |
| 694 | BUG_ON(! within_hugepage_high_range(addr, len)); |
| 695 | |
| 696 | if (!vma || (addr + len) <= vma->vm_start) |
| 697 | return addr; |
| 698 | addr = ALIGN(vma->vm_end, HPAGE_SIZE); |
| 699 | /* Because we're in a hugepage region, this alignment |
| 700 | * should not skip us over any VMAs */ |
| 701 | } |
| 702 | |
| 703 | return -ENOMEM; |
| 704 | } |
| 705 | |
| 706 | unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr, |
| 707 | unsigned long len, unsigned long pgoff, |
| 708 | unsigned long flags) |
| 709 | { |
| 710 | if (len & ~HPAGE_MASK) |
| 711 | return -EINVAL; |
| 712 | |
| 713 | if (!cpu_has_feature(CPU_FTR_16M_PAGE)) |
| 714 | return -EINVAL; |
| 715 | |
| 716 | if (test_thread_flag(TIF_32BIT)) { |
| 717 | int lastshift = 0; |
| 718 | u16 segmask, cursegs = current->mm->context.htlb_segs; |
| 719 | |
| 720 | /* First see if we can do the mapping in the existing |
| 721 | * low hpage segments */ |
| 722 | addr = htlb_get_low_area(len, cursegs); |
| 723 | if (addr != -ENOMEM) |
| 724 | return addr; |
| 725 | |
| 726 | for (segmask = LOW_ESID_MASK(0x100000000UL-len, len); |
| 727 | ! lastshift; segmask >>=1) { |
| 728 | if (segmask & 1) |
| 729 | lastshift = 1; |
| 730 | |
| 731 | addr = htlb_get_low_area(len, cursegs | segmask); |
| 732 | if ((addr != -ENOMEM) |
| 733 | && open_low_hpage_segs(current->mm, segmask) == 0) |
| 734 | return addr; |
| 735 | } |
| 736 | printk(KERN_DEBUG "hugetlb_get_unmapped_area() unable to open" |
| 737 | " enough segments\n"); |
| 738 | return -ENOMEM; |
| 739 | } else { |
| 740 | return htlb_get_high_area(len); |
| 741 | } |
| 742 | } |
| 743 | |
| 744 | void hugetlb_mm_free_pgd(struct mm_struct *mm) |
| 745 | { |
| 746 | int i; |
| 747 | pgd_t *pgdir; |
| 748 | |
| 749 | spin_lock(&mm->page_table_lock); |
| 750 | |
| 751 | pgdir = mm->context.huge_pgdir; |
| 752 | if (! pgdir) |
| 753 | goto out; |
| 754 | |
| 755 | mm->context.huge_pgdir = NULL; |
| 756 | |
| 757 | /* cleanup any hugepte pages leftover */ |
| 758 | for (i = 0; i < PTRS_PER_HUGEPGD; i++) { |
| 759 | pgd_t *pgd = pgdir + i; |
| 760 | |
| 761 | if (! pgd_none(*pgd)) { |
| 762 | pte_t *pte = (pte_t *)pgd_page(*pgd); |
| 763 | struct page *ptepage = virt_to_page(pte); |
| 764 | |
| 765 | ptepage->mapping = NULL; |
| 766 | |
| 767 | BUG_ON(memcmp(pte, empty_zero_page, PAGE_SIZE)); |
| 768 | kmem_cache_free(zero_cache, pte); |
| 769 | } |
| 770 | pgd_clear(pgd); |
| 771 | } |
| 772 | |
| 773 | BUG_ON(memcmp(pgdir, empty_zero_page, PAGE_SIZE)); |
| 774 | kmem_cache_free(zero_cache, pgdir); |
| 775 | |
| 776 | out: |
| 777 | spin_unlock(&mm->page_table_lock); |
| 778 | } |
| 779 | |
| 780 | int hash_huge_page(struct mm_struct *mm, unsigned long access, |
| 781 | unsigned long ea, unsigned long vsid, int local) |
| 782 | { |
| 783 | pte_t *ptep; |
| 784 | unsigned long va, vpn; |
| 785 | pte_t old_pte, new_pte; |
| 786 | unsigned long hpteflags, prpn; |
| 787 | long slot; |
| 788 | int err = 1; |
| 789 | |
| 790 | spin_lock(&mm->page_table_lock); |
| 791 | |
| 792 | ptep = huge_pte_offset(mm, ea); |
| 793 | |
| 794 | /* Search the Linux page table for a match with va */ |
| 795 | va = (vsid << 28) | (ea & 0x0fffffff); |
| 796 | vpn = va >> HPAGE_SHIFT; |
| 797 | |
| 798 | /* |
| 799 | * If no pte found or not present, send the problem up to |
| 800 | * do_page_fault |
| 801 | */ |
| 802 | if (unlikely(!ptep || pte_none(*ptep))) |
| 803 | goto out; |
| 804 | |
| 805 | /* BUG_ON(pte_bad(*ptep)); */ |
| 806 | |
| 807 | /* |
| 808 | * Check the user's access rights to the page. If access should be |
| 809 | * prevented then send the problem up to do_page_fault. |
| 810 | */ |
| 811 | if (unlikely(access & ~pte_val(*ptep))) |
| 812 | goto out; |
| 813 | /* |
| 814 | * At this point, we have a pte (old_pte) which can be used to build |
| 815 | * or update an HPTE. There are 2 cases: |
| 816 | * |
| 817 | * 1. There is a valid (present) pte with no associated HPTE (this is |
| 818 | * the most common case) |
| 819 | * 2. There is a valid (present) pte with an associated HPTE. The |
| 820 | * current values of the pp bits in the HPTE prevent access |
| 821 | * because we are doing software DIRTY bit management and the |
| 822 | * page is currently not DIRTY. |
| 823 | */ |
| 824 | |
| 825 | |
| 826 | old_pte = *ptep; |
| 827 | new_pte = old_pte; |
| 828 | |
| 829 | hpteflags = 0x2 | (! (pte_val(new_pte) & _PAGE_RW)); |
| 830 | /* _PAGE_EXEC -> HW_NO_EXEC since it's inverted */ |
| 831 | hpteflags |= ((pte_val(new_pte) & _PAGE_EXEC) ? 0 : HW_NO_EXEC); |
| 832 | |
| 833 | /* Check if pte already has an hpte (case 2) */ |
| 834 | if (unlikely(pte_val(old_pte) & _PAGE_HASHPTE)) { |
| 835 | /* There MIGHT be an HPTE for this pte */ |
| 836 | unsigned long hash, slot; |
| 837 | |
| 838 | hash = hpt_hash(vpn, 1); |
| 839 | if (pte_val(old_pte) & _PAGE_SECONDARY) |
| 840 | hash = ~hash; |
| 841 | slot = (hash & htab_hash_mask) * HPTES_PER_GROUP; |
| 842 | slot += (pte_val(old_pte) & _PAGE_GROUP_IX) >> 12; |
| 843 | |
| 844 | if (ppc_md.hpte_updatepp(slot, hpteflags, va, 1, local) == -1) |
| 845 | pte_val(old_pte) &= ~_PAGE_HPTEFLAGS; |
| 846 | } |
| 847 | |
| 848 | if (likely(!(pte_val(old_pte) & _PAGE_HASHPTE))) { |
| 849 | unsigned long hash = hpt_hash(vpn, 1); |
| 850 | unsigned long hpte_group; |
| 851 | |
| 852 | prpn = pte_pfn(old_pte); |
| 853 | |
| 854 | repeat: |
| 855 | hpte_group = ((hash & htab_hash_mask) * |
| 856 | HPTES_PER_GROUP) & ~0x7UL; |
| 857 | |
| 858 | /* Update the linux pte with the HPTE slot */ |
| 859 | pte_val(new_pte) &= ~_PAGE_HPTEFLAGS; |
| 860 | pte_val(new_pte) |= _PAGE_HASHPTE; |
| 861 | |
| 862 | /* Add in WIMG bits */ |
| 863 | /* XXX We should store these in the pte */ |
| 864 | hpteflags |= _PAGE_COHERENT; |
| 865 | |
| 866 | slot = ppc_md.hpte_insert(hpte_group, va, prpn, 0, |
| 867 | hpteflags, 0, 1); |
| 868 | |
| 869 | /* Primary is full, try the secondary */ |
| 870 | if (unlikely(slot == -1)) { |
| 871 | pte_val(new_pte) |= _PAGE_SECONDARY; |
| 872 | hpte_group = ((~hash & htab_hash_mask) * |
| 873 | HPTES_PER_GROUP) & ~0x7UL; |
| 874 | slot = ppc_md.hpte_insert(hpte_group, va, prpn, |
| 875 | 1, hpteflags, 0, 1); |
| 876 | if (slot == -1) { |
| 877 | if (mftb() & 0x1) |
| 878 | hpte_group = ((hash & htab_hash_mask) * HPTES_PER_GROUP) & ~0x7UL; |
| 879 | |
| 880 | ppc_md.hpte_remove(hpte_group); |
| 881 | goto repeat; |
| 882 | } |
| 883 | } |
| 884 | |
| 885 | if (unlikely(slot == -2)) |
| 886 | panic("hash_huge_page: pte_insert failed\n"); |
| 887 | |
| 888 | pte_val(new_pte) |= (slot<<12) & _PAGE_GROUP_IX; |
| 889 | |
| 890 | /* |
| 891 | * No need to use ldarx/stdcx here because all who |
| 892 | * might be updating the pte will hold the |
| 893 | * page_table_lock |
| 894 | */ |
| 895 | *ptep = new_pte; |
| 896 | } |
| 897 | |
| 898 | err = 0; |
| 899 | |
| 900 | out: |
| 901 | spin_unlock(&mm->page_table_lock); |
| 902 | |
| 903 | return err; |
| 904 | } |