1 /* 1 /*
2 * MIPS Huge TLB Page Support for Kernel. !! 2 * SPARC64 Huge TLB page support.
3 * 3 *
4 * This file is subject to the terms and condi !! 4 * Copyright (C) 2002, 2003 David S. Miller (davem@redhat.com)
5 * License. See the file "COPYING" in the mai <<
6 * for more details. <<
7 * <<
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@ <<
9 * Copyright 2005, Embedded Alley Solutions, I <<
10 * Matt Porter <mporter@embeddedalley.com> <<
11 * Copyright (C) 2008, 2009 Cavium Networks, I <<
12 */ 5 */
13 6
>> 7 #include <linux/config.h>
>> 8 #include <linux/init.h>
14 #include <linux/fs.h> 9 #include <linux/fs.h>
15 #include <linux/mm.h> 10 #include <linux/mm.h>
16 #include <linux/hugetlb.h> 11 #include <linux/hugetlb.h>
17 #include <linux/pagemap.h> 12 #include <linux/pagemap.h>
18 #include <linux/err.h> !! 13 #include <linux/smp_lock.h>
>> 14 #include <linux/slab.h>
19 #include <linux/sysctl.h> 15 #include <linux/sysctl.h>
>> 16
20 #include <asm/mman.h> 17 #include <asm/mman.h>
>> 18 #include <asm/pgalloc.h>
21 #include <asm/tlb.h> 19 #include <asm/tlb.h>
22 #include <asm/tlbflush.h> 20 #include <asm/tlbflush.h>
>> 21 #include <asm/cacheflush.h>
>> 22
>> 23 static long htlbpagemem;
>> 24 int htlbpage_max;
>> 25 static long htlbzone_pages;
>> 26
>> 27 static struct list_head hugepage_freelists[MAX_NUMNODES];
>> 28 static spinlock_t htlbpage_lock = SPIN_LOCK_UNLOCKED;
>> 29
>> 30 static void enqueue_huge_page(struct page *page)
>> 31 {
>> 32 list_add(&page->list,
>> 33 &hugepage_freelists[page_zone(page)->zone_pgdat->node_id]);
>> 34 }
23 35
24 pte_t *huge_pte_alloc(struct mm_struct *mm, st !! 36 static struct page *dequeue_huge_page(void)
25 unsigned long addr, unsi !! 37 {
>> 38 int nid = numa_node_id();
>> 39 struct page *page = NULL;
>> 40
>> 41 if (list_empty(&hugepage_freelists[nid])) {
>> 42 for (nid = 0; nid < MAX_NUMNODES; ++nid)
>> 43 if (!list_empty(&hugepage_freelists[nid]))
>> 44 break;
>> 45 }
>> 46 if (nid >= 0 && nid < MAX_NUMNODES &&
>> 47 !list_empty(&hugepage_freelists[nid])) {
>> 48 page = list_entry(hugepage_freelists[nid].next,
>> 49 struct page, list);
>> 50 list_del(&page->list);
>> 51 }
>> 52 return page;
>> 53 }
>> 54
>> 55 static struct page *alloc_fresh_huge_page(void)
>> 56 {
>> 57 static int nid = 0;
>> 58 struct page *page;
>> 59 page = alloc_pages_node(nid, GFP_HIGHUSER, HUGETLB_PAGE_ORDER);
>> 60 nid = (nid + 1) % numnodes;
>> 61 return page;
>> 62 }
>> 63
>> 64 static void free_huge_page(struct page *page);
>> 65
>> 66 static struct page *alloc_hugetlb_page(void)
>> 67 {
>> 68 struct page *page;
>> 69
>> 70 spin_lock(&htlbpage_lock);
>> 71 page = dequeue_huge_page();
>> 72 if (!page) {
>> 73 spin_unlock(&htlbpage_lock);
>> 74 return NULL;
>> 75 }
>> 76 htlbpagemem--;
>> 77 spin_unlock(&htlbpage_lock);
>> 78 set_page_count(page, 1);
>> 79 page->lru.prev = (void *)free_huge_page;
>> 80 memset(page_address(page), 0, HPAGE_SIZE);
>> 81 return page;
>> 82 }
>> 83
>> 84 static pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr)
26 { 85 {
27 pgd_t *pgd; 86 pgd_t *pgd;
28 p4d_t *p4d; !! 87 pmd_t *pmd;
29 pud_t *pud; <<
30 pte_t *pte = NULL; 88 pte_t *pte = NULL;
31 89
32 pgd = pgd_offset(mm, addr); 90 pgd = pgd_offset(mm, addr);
33 p4d = p4d_alloc(mm, pgd, addr); !! 91 if (pgd) {
34 pud = pud_alloc(mm, p4d, addr); !! 92 pmd = pmd_alloc(mm, pgd, addr);
35 if (pud) !! 93 if (pmd)
36 pte = (pte_t *)pmd_alloc(mm, p !! 94 pte = pte_alloc_map(mm, pmd, addr);
37 !! 95 }
38 return pte; 96 return pte;
39 } 97 }
40 98
41 pte_t *huge_pte_offset(struct mm_struct *mm, u !! 99 static pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
42 unsigned long sz) <<
43 { 100 {
44 pgd_t *pgd; 101 pgd_t *pgd;
45 p4d_t *p4d; !! 102 pmd_t *pmd;
46 pud_t *pud; !! 103 pte_t *pte = NULL;
47 pmd_t *pmd = NULL; <<
48 104
49 pgd = pgd_offset(mm, addr); 105 pgd = pgd_offset(mm, addr);
50 if (pgd_present(*pgd)) { !! 106 if (pgd) {
51 p4d = p4d_offset(pgd, addr); !! 107 pmd = pmd_offset(pgd, addr);
52 if (p4d_present(*p4d)) { !! 108 if (pmd)
53 pud = pud_offset(p4d, !! 109 pte = pte_offset_map(pmd, addr);
54 if (pud_present(*pud)) !! 110 }
55 pmd = pmd_offs !! 111 return pte;
>> 112 }
>> 113
>> 114 #define mk_pte_huge(entry) do { pte_val(entry) |= _PAGE_SZHUGE; } while (0)
>> 115
>> 116 static void set_huge_pte(struct mm_struct *mm, struct vm_area_struct *vma,
>> 117 struct page *page, pte_t * page_table, int write_access)
>> 118 {
>> 119 unsigned long i;
>> 120 pte_t entry;
>> 121
>> 122 mm->rss += (HPAGE_SIZE / PAGE_SIZE);
>> 123
>> 124 if (write_access)
>> 125 entry = pte_mkwrite(pte_mkdirty(mk_pte(page,
>> 126 vma->vm_page_prot)));
>> 127 else
>> 128 entry = pte_wrprotect(mk_pte(page, vma->vm_page_prot));
>> 129 entry = pte_mkyoung(entry);
>> 130 mk_pte_huge(entry);
>> 131
>> 132 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
>> 133 set_pte(page_table, entry);
>> 134 page_table++;
>> 135
>> 136 pte_val(entry) += PAGE_SIZE;
>> 137 }
>> 138 }
>> 139
>> 140 /*
>> 141 * This function checks for proper alignment of input addr and len parameters.
>> 142 */
>> 143 int is_aligned_hugepage_range(unsigned long addr, unsigned long len)
>> 144 {
>> 145 if (len & ~HPAGE_MASK)
>> 146 return -EINVAL;
>> 147 if (addr & ~HPAGE_MASK)
>> 148 return -EINVAL;
>> 149 return 0;
>> 150 }
>> 151
>> 152 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
>> 153 struct vm_area_struct *vma)
>> 154 {
>> 155 pte_t *src_pte, *dst_pte, entry;
>> 156 struct page *ptepage;
>> 157 unsigned long addr = vma->vm_start;
>> 158 unsigned long end = vma->vm_end;
>> 159 int i;
>> 160
>> 161 while (addr < end) {
>> 162 dst_pte = huge_pte_alloc(dst, addr);
>> 163 if (!dst_pte)
>> 164 goto nomem;
>> 165 src_pte = huge_pte_offset(src, addr);
>> 166 BUG_ON(!src_pte || pte_none(*src_pte));
>> 167 entry = *src_pte;
>> 168 ptepage = pte_page(entry);
>> 169 get_page(ptepage);
>> 170 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
>> 171 set_pte(dst_pte, entry);
>> 172 pte_val(entry) += PAGE_SIZE;
>> 173 dst_pte++;
>> 174 }
>> 175 dst->rss += (HPAGE_SIZE / PAGE_SIZE);
>> 176 addr += HPAGE_SIZE;
>> 177 }
>> 178 return 0;
>> 179
>> 180 nomem:
>> 181 return -ENOMEM;
>> 182 }
>> 183
>> 184 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
>> 185 struct page **pages, struct vm_area_struct **vmas,
>> 186 unsigned long *position, int *length, int i)
>> 187 {
>> 188 unsigned long vaddr = *position;
>> 189 int remainder = *length;
>> 190
>> 191 WARN_ON(!is_vm_hugetlb_page(vma));
>> 192
>> 193 while (vaddr < vma->vm_end && remainder) {
>> 194 if (pages) {
>> 195 pte_t *pte;
>> 196 struct page *page;
>> 197
>> 198 pte = huge_pte_offset(mm, vaddr);
>> 199
>> 200 /* hugetlb should be locked, and hence, prefaulted */
>> 201 BUG_ON(!pte || pte_none(*pte));
>> 202
>> 203 page = pte_page(*pte);
>> 204
>> 205 WARN_ON(!PageCompound(page));
>> 206
>> 207 get_page(page);
>> 208 pages[i] = page;
>> 209 }
>> 210
>> 211 if (vmas)
>> 212 vmas[i] = vma;
>> 213
>> 214 vaddr += PAGE_SIZE;
>> 215 --remainder;
>> 216 ++i;
>> 217 }
>> 218
>> 219 *length = remainder;
>> 220 *position = vaddr;
>> 221
>> 222 return i;
>> 223 }
>> 224
>> 225 struct page *follow_huge_addr(struct mm_struct *mm,
>> 226 struct vm_area_struct *vma,
>> 227 unsigned long address, int write)
>> 228 {
>> 229 return NULL;
>> 230 }
>> 231
>> 232 struct vm_area_struct *hugepage_vma(struct mm_struct *mm, unsigned long addr)
>> 233 {
>> 234 return NULL;
>> 235 }
>> 236
>> 237 int pmd_huge(pmd_t pmd)
>> 238 {
>> 239 return 0;
>> 240 }
>> 241
>> 242 struct page *follow_huge_pmd(struct mm_struct *mm, unsigned long address,
>> 243 pmd_t *pmd, int write)
>> 244 {
>> 245 return NULL;
>> 246 }
>> 247
>> 248 static void free_huge_page(struct page *page)
>> 249 {
>> 250 BUG_ON(page_count(page));
>> 251 BUG_ON(page->mapping);
>> 252
>> 253 INIT_LIST_HEAD(&page->list);
>> 254
>> 255 spin_lock(&htlbpage_lock);
>> 256 enqueue_huge_page(page);
>> 257 htlbpagemem++;
>> 258 spin_unlock(&htlbpage_lock);
>> 259 }
>> 260
>> 261 void huge_page_release(struct page *page)
>> 262 {
>> 263 if (!put_page_testzero(page))
>> 264 return;
>> 265
>> 266 free_huge_page(page);
>> 267 }
>> 268
>> 269 void unmap_hugepage_range(struct vm_area_struct *vma,
>> 270 unsigned long start, unsigned long end)
>> 271 {
>> 272 struct mm_struct *mm = vma->vm_mm;
>> 273 unsigned long address;
>> 274 pte_t *pte;
>> 275 struct page *page;
>> 276 int i;
>> 277
>> 278 BUG_ON(start & (HPAGE_SIZE - 1));
>> 279 BUG_ON(end & (HPAGE_SIZE - 1));
>> 280
>> 281 for (address = start; address < end; address += HPAGE_SIZE) {
>> 282 pte = huge_pte_offset(mm, address);
>> 283 BUG_ON(!pte);
>> 284 if (pte_none(*pte))
>> 285 continue;
>> 286 page = pte_page(*pte);
>> 287 huge_page_release(page);
>> 288 for (i = 0; i < (1 << HUGETLB_PAGE_ORDER); i++) {
>> 289 pte_clear(pte);
>> 290 pte++;
56 } 291 }
57 } 292 }
58 return (pte_t *) pmd; !! 293 mm->rss -= (end - start) >> PAGE_SHIFT;
>> 294 flush_tlb_range(vma, start, end);
59 } 295 }
>> 296
>> 297 void zap_hugepage_range(struct vm_area_struct *vma,
>> 298 unsigned long start, unsigned long length)
>> 299 {
>> 300 struct mm_struct *mm = vma->vm_mm;
>> 301
>> 302 spin_lock(&mm->page_table_lock);
>> 303 unmap_hugepage_range(vma, start, start + length);
>> 304 spin_unlock(&mm->page_table_lock);
>> 305 }
>> 306
>> 307 int hugetlb_prefault(struct address_space *mapping, struct vm_area_struct *vma)
>> 308 {
>> 309 struct mm_struct *mm = current->mm;
>> 310 unsigned long addr;
>> 311 int ret = 0;
>> 312
>> 313 BUG_ON(vma->vm_start & ~HPAGE_MASK);
>> 314 BUG_ON(vma->vm_end & ~HPAGE_MASK);
>> 315
>> 316 spin_lock(&mm->page_table_lock);
>> 317 for (addr = vma->vm_start; addr < vma->vm_end; addr += HPAGE_SIZE) {
>> 318 unsigned long idx;
>> 319 pte_t *pte = huge_pte_alloc(mm, addr);
>> 320 struct page *page;
>> 321
>> 322 if (!pte) {
>> 323 ret = -ENOMEM;
>> 324 goto out;
>> 325 }
>> 326 if (!pte_none(*pte))
>> 327 continue;
>> 328
>> 329 idx = ((addr - vma->vm_start) >> HPAGE_SHIFT)
>> 330 + (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
>> 331 page = find_get_page(mapping, idx);
>> 332 if (!page) {
>> 333 /* charge the fs quota first */
>> 334 if (hugetlb_get_quota(mapping)) {
>> 335 ret = -ENOMEM;
>> 336 goto out;
>> 337 }
>> 338 page = alloc_hugetlb_page();
>> 339 if (!page) {
>> 340 hugetlb_put_quota(mapping);
>> 341 ret = -ENOMEM;
>> 342 goto out;
>> 343 }
>> 344 ret = add_to_page_cache(page, mapping, idx, GFP_ATOMIC);
>> 345 unlock_page(page);
>> 346 if (ret) {
>> 347 hugetlb_put_quota(mapping);
>> 348 free_huge_page(page);
>> 349 goto out;
>> 350 }
>> 351 }
>> 352 set_huge_pte(mm, vma, page, pte, vma->vm_flags & VM_WRITE);
>> 353 }
>> 354 out:
>> 355 spin_unlock(&mm->page_table_lock);
>> 356 return ret;
>> 357 }
>> 358
>> 359 static void update_and_free_page(struct page *page)
>> 360 {
>> 361 int j;
>> 362 struct page *map;
>> 363
>> 364 map = page;
>> 365 htlbzone_pages--;
>> 366 for (j = 0; j < (HPAGE_SIZE / PAGE_SIZE); j++) {
>> 367 map->flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
>> 368 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
>> 369 1 << PG_private | 1<< PG_writeback);
>> 370 set_page_count(map, 0);
>> 371 map++;
>> 372 }
>> 373 set_page_count(page, 1);
>> 374 __free_pages(page, HUGETLB_PAGE_ORDER);
>> 375 }
>> 376
>> 377 static int try_to_free_low(int count)
>> 378 {
>> 379 struct list_head *p;
>> 380 struct page *page, *map;
>> 381
>> 382 map = NULL;
>> 383 spin_lock(&htlbpage_lock);
>> 384 /* all lowmem is on node 0 */
>> 385 list_for_each(p, &hugepage_freelists[0]) {
>> 386 if (map) {
>> 387 list_del(&map->list);
>> 388 update_and_free_page(map);
>> 389 htlbpagemem--;
>> 390 map = NULL;
>> 391 if (++count == 0)
>> 392 break;
>> 393 }
>> 394 page = list_entry(p, struct page, list);
>> 395 if (!PageHighMem(page))
>> 396 map = page;
>> 397 }
>> 398 if (map) {
>> 399 list_del(&map->list);
>> 400 update_and_free_page(map);
>> 401 htlbpagemem--;
>> 402 count++;
>> 403 }
>> 404 spin_unlock(&htlbpage_lock);
>> 405 return count;
>> 406 }
>> 407
>> 408 static int set_hugetlb_mem_size(int count)
>> 409 {
>> 410 int lcount;
>> 411 struct page *page;
>> 412
>> 413 if (count < 0)
>> 414 lcount = count;
>> 415 else
>> 416 lcount = count - htlbzone_pages;
>> 417
>> 418 if (lcount == 0)
>> 419 return (int)htlbzone_pages;
>> 420 if (lcount > 0) { /* Increase the mem size. */
>> 421 while (lcount--) {
>> 422 page = alloc_fresh_huge_page();
>> 423 if (page == NULL)
>> 424 break;
>> 425 spin_lock(&htlbpage_lock);
>> 426 enqueue_huge_page(page);
>> 427 htlbpagemem++;
>> 428 htlbzone_pages++;
>> 429 spin_unlock(&htlbpage_lock);
>> 430 }
>> 431 return (int) htlbzone_pages;
>> 432 }
>> 433 /* Shrink the memory size. */
>> 434 lcount = try_to_free_low(lcount);
>> 435 while (lcount++) {
>> 436 page = alloc_hugetlb_page();
>> 437 if (page == NULL)
>> 438 break;
>> 439 spin_lock(&htlbpage_lock);
>> 440 update_and_free_page(page);
>> 441 spin_unlock(&htlbpage_lock);
>> 442 }
>> 443 return (int) htlbzone_pages;
>> 444 }
>> 445
>> 446 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
>> 447 struct file *file, void *buffer, size_t *length)
>> 448 {
>> 449 proc_dointvec(table, write, file, buffer, length);
>> 450 htlbpage_max = set_hugetlb_mem_size(htlbpage_max);
>> 451 return 0;
>> 452 }
>> 453
>> 454 static int __init hugetlb_setup(char *s)
>> 455 {
>> 456 if (sscanf(s, "%d", &htlbpage_max) <= 0)
>> 457 htlbpage_max = 0;
>> 458 return 1;
>> 459 }
>> 460 __setup("hugepages=", hugetlb_setup);
>> 461
>> 462 static int __init hugetlb_init(void)
>> 463 {
>> 464 int i;
>> 465 struct page *page;
>> 466
>> 467 for (i = 0; i < MAX_NUMNODES; ++i)
>> 468 INIT_LIST_HEAD(&hugepage_freelists[i]);
>> 469
>> 470 for (i = 0; i < htlbpage_max; ++i) {
>> 471 page = alloc_fresh_huge_page();
>> 472 if (!page)
>> 473 break;
>> 474 spin_lock(&htlbpage_lock);
>> 475 enqueue_huge_page(page);
>> 476 spin_unlock(&htlbpage_lock);
>> 477 }
>> 478 htlbpage_max = htlbpagemem = htlbzone_pages = i;
>> 479 printk("Total HugeTLB memory allocated, %ld\n", htlbpagemem);
>> 480 return 0;
>> 481 }
>> 482 module_init(hugetlb_init);
>> 483
>> 484 int hugetlb_report_meminfo(char *buf)
>> 485 {
>> 486 return sprintf(buf,
>> 487 "HugePages_Total: %5lu\n"
>> 488 "HugePages_Free: %5lu\n"
>> 489 "Hugepagesize: %5lu kB\n",
>> 490 htlbzone_pages,
>> 491 htlbpagemem,
>> 492 HPAGE_SIZE/1024);
>> 493 }
>> 494
>> 495 int is_hugepage_mem_enough(size_t size)
>> 496 {
>> 497 return (size + ~HPAGE_MASK)/HPAGE_SIZE <= htlbpagemem;
>> 498 }
>> 499
>> 500 /*
>> 501 * We cannot handle pagefaults against hugetlb pages at all. They cause
>> 502 * handle_mm_fault() to try to instantiate regular-sized pages in the
>> 503 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
>> 504 * this far.
>> 505 */
>> 506 static struct page *hugetlb_nopage(struct vm_area_struct *vma,
>> 507 unsigned long address, int unused)
>> 508 {
>> 509 BUG();
>> 510 return NULL;
>> 511 }
>> 512
>> 513 struct vm_operations_struct hugetlb_vm_ops = {
>> 514 .nopage = hugetlb_nopage,
>> 515 };
60 516
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.