Loading...
1/*
2 * PPC Huge TLB Page Support for Kernel.
3 *
4 * Copyright (C) 2003 David Gibson, IBM Corporation.
5 * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
6 *
7 * Based on the IA-32 version:
8 * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9 */
10
11#include <linux/mm.h>
12#include <linux/io.h>
13#include <linux/slab.h>
14#include <linux/hugetlb.h>
15#include <linux/export.h>
16#include <linux/of_fdt.h>
17#include <linux/memblock.h>
18#include <linux/moduleparam.h>
19#include <linux/swap.h>
20#include <linux/swapops.h>
21#include <linux/kmemleak.h>
22#include <asm/pgtable.h>
23#include <asm/pgalloc.h>
24#include <asm/tlb.h>
25#include <asm/setup.h>
26#include <asm/hugetlb.h>
27#include <asm/pte-walk.h>
28
29bool hugetlb_disabled = false;
30
31#define hugepd_none(hpd) (hpd_val(hpd) == 0)
32
33#define PTE_T_ORDER (__builtin_ffs(sizeof(pte_t)) - __builtin_ffs(sizeof(void *)))
34
35pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
36{
37 /*
38 * Only called for hugetlbfs pages, hence can ignore THP and the
39 * irq disabled walk.
40 */
41 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
42}
43
44static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
45 unsigned long address, unsigned int pdshift,
46 unsigned int pshift, spinlock_t *ptl)
47{
48 struct kmem_cache *cachep;
49 pte_t *new;
50 int i;
51 int num_hugepd;
52
53 if (pshift >= pdshift) {
54 cachep = PGT_CACHE(PTE_T_ORDER);
55 num_hugepd = 1 << (pshift - pdshift);
56 } else if (IS_ENABLED(CONFIG_PPC_8xx)) {
57 cachep = PGT_CACHE(PTE_INDEX_SIZE);
58 num_hugepd = 1;
59 } else {
60 cachep = PGT_CACHE(pdshift - pshift);
61 num_hugepd = 1;
62 }
63
64 if (!cachep) {
65 WARN_ONCE(1, "No page table cache created for hugetlb tables");
66 return -ENOMEM;
67 }
68
69 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
70
71 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
72 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
73
74 if (!new)
75 return -ENOMEM;
76
77 /*
78 * Make sure other cpus find the hugepd set only after a
79 * properly initialized page table is visible to them.
80 * For more details look for comment in __pte_alloc().
81 */
82 smp_wmb();
83
84 spin_lock(ptl);
85 /*
86 * We have multiple higher-level entries that point to the same
87 * actual pte location. Fill in each as we go and backtrack on error.
88 * We need all of these so the DTLB pgtable walk code can find the
89 * right higher-level entry without knowing if it's a hugepage or not.
90 */
91 for (i = 0; i < num_hugepd; i++, hpdp++) {
92 if (unlikely(!hugepd_none(*hpdp)))
93 break;
94 hugepd_populate(hpdp, new, pshift);
95 }
96 /* If we bailed from the for loop early, an error occurred, clean up */
97 if (i < num_hugepd) {
98 for (i = i - 1 ; i >= 0; i--, hpdp--)
99 *hpdp = __hugepd(0);
100 kmem_cache_free(cachep, new);
101 } else {
102 kmemleak_ignore(new);
103 }
104 spin_unlock(ptl);
105 return 0;
106}
107
108/*
109 * At this point we do the placement change only for BOOK3S 64. This would
110 * possibly work on other subarchs.
111 */
112pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
113{
114 pgd_t *pg;
115 pud_t *pu;
116 pmd_t *pm;
117 hugepd_t *hpdp = NULL;
118 unsigned pshift = __ffs(sz);
119 unsigned pdshift = PGDIR_SHIFT;
120 spinlock_t *ptl;
121
122 addr &= ~(sz-1);
123 pg = pgd_offset(mm, addr);
124
125#ifdef CONFIG_PPC_BOOK3S_64
126 if (pshift == PGDIR_SHIFT)
127 /* 16GB huge page */
128 return (pte_t *) pg;
129 else if (pshift > PUD_SHIFT) {
130 /*
131 * We need to use hugepd table
132 */
133 ptl = &mm->page_table_lock;
134 hpdp = (hugepd_t *)pg;
135 } else {
136 pdshift = PUD_SHIFT;
137 pu = pud_alloc(mm, pg, addr);
138 if (!pu)
139 return NULL;
140 if (pshift == PUD_SHIFT)
141 return (pte_t *)pu;
142 else if (pshift > PMD_SHIFT) {
143 ptl = pud_lockptr(mm, pu);
144 hpdp = (hugepd_t *)pu;
145 } else {
146 pdshift = PMD_SHIFT;
147 pm = pmd_alloc(mm, pu, addr);
148 if (!pm)
149 return NULL;
150 if (pshift == PMD_SHIFT)
151 /* 16MB hugepage */
152 return (pte_t *)pm;
153 else {
154 ptl = pmd_lockptr(mm, pm);
155 hpdp = (hugepd_t *)pm;
156 }
157 }
158 }
159#else
160 if (pshift >= PGDIR_SHIFT) {
161 ptl = &mm->page_table_lock;
162 hpdp = (hugepd_t *)pg;
163 } else {
164 pdshift = PUD_SHIFT;
165 pu = pud_alloc(mm, pg, addr);
166 if (!pu)
167 return NULL;
168 if (pshift >= PUD_SHIFT) {
169 ptl = pud_lockptr(mm, pu);
170 hpdp = (hugepd_t *)pu;
171 } else {
172 pdshift = PMD_SHIFT;
173 pm = pmd_alloc(mm, pu, addr);
174 if (!pm)
175 return NULL;
176 ptl = pmd_lockptr(mm, pm);
177 hpdp = (hugepd_t *)pm;
178 }
179 }
180#endif
181 if (!hpdp)
182 return NULL;
183
184 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
185
186 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
187 pdshift, pshift, ptl))
188 return NULL;
189
190 return hugepte_offset(*hpdp, addr, pdshift);
191}
192
193#ifdef CONFIG_PPC_BOOK3S_64
194/*
195 * Tracks gpages after the device tree is scanned and before the
196 * huge_boot_pages list is ready on pseries.
197 */
198#define MAX_NUMBER_GPAGES 1024
199__initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
200__initdata static unsigned nr_gpages;
201
202/*
203 * Build list of addresses of gigantic pages. This function is used in early
204 * boot before the buddy allocator is setup.
205 */
206void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
207{
208 if (!addr)
209 return;
210 while (number_of_pages > 0) {
211 gpage_freearray[nr_gpages] = addr;
212 nr_gpages++;
213 number_of_pages--;
214 addr += page_size;
215 }
216}
217
218int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
219{
220 struct huge_bootmem_page *m;
221 if (nr_gpages == 0)
222 return 0;
223 m = phys_to_virt(gpage_freearray[--nr_gpages]);
224 gpage_freearray[nr_gpages] = 0;
225 list_add(&m->list, &huge_boot_pages);
226 m->hstate = hstate;
227 return 1;
228}
229#endif
230
231
232int __init alloc_bootmem_huge_page(struct hstate *h)
233{
234
235#ifdef CONFIG_PPC_BOOK3S_64
236 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
237 return pseries_alloc_bootmem_huge_page(h);
238#endif
239 return __alloc_bootmem_huge_page(h);
240}
241
242#ifndef CONFIG_PPC_BOOK3S_64
243#define HUGEPD_FREELIST_SIZE \
244 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
245
246struct hugepd_freelist {
247 struct rcu_head rcu;
248 unsigned int index;
249 void *ptes[0];
250};
251
252static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
253
254static void hugepd_free_rcu_callback(struct rcu_head *head)
255{
256 struct hugepd_freelist *batch =
257 container_of(head, struct hugepd_freelist, rcu);
258 unsigned int i;
259
260 for (i = 0; i < batch->index; i++)
261 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
262
263 free_page((unsigned long)batch);
264}
265
266static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
267{
268 struct hugepd_freelist **batchp;
269
270 batchp = &get_cpu_var(hugepd_freelist_cur);
271
272 if (atomic_read(&tlb->mm->mm_users) < 2 ||
273 mm_is_thread_local(tlb->mm)) {
274 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
275 put_cpu_var(hugepd_freelist_cur);
276 return;
277 }
278
279 if (*batchp == NULL) {
280 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
281 (*batchp)->index = 0;
282 }
283
284 (*batchp)->ptes[(*batchp)->index++] = hugepte;
285 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
286 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
287 *batchp = NULL;
288 }
289 put_cpu_var(hugepd_freelist_cur);
290}
291#else
292static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
293#endif
294
295static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
296 unsigned long start, unsigned long end,
297 unsigned long floor, unsigned long ceiling)
298{
299 pte_t *hugepte = hugepd_page(*hpdp);
300 int i;
301
302 unsigned long pdmask = ~((1UL << pdshift) - 1);
303 unsigned int num_hugepd = 1;
304 unsigned int shift = hugepd_shift(*hpdp);
305
306 /* Note: On fsl the hpdp may be the first of several */
307 if (shift > pdshift)
308 num_hugepd = 1 << (shift - pdshift);
309
310 start &= pdmask;
311 if (start < floor)
312 return;
313 if (ceiling) {
314 ceiling &= pdmask;
315 if (! ceiling)
316 return;
317 }
318 if (end - 1 > ceiling - 1)
319 return;
320
321 for (i = 0; i < num_hugepd; i++, hpdp++)
322 *hpdp = __hugepd(0);
323
324 if (shift >= pdshift)
325 hugepd_free(tlb, hugepte);
326 else if (IS_ENABLED(CONFIG_PPC_8xx))
327 pgtable_free_tlb(tlb, hugepte,
328 get_hugepd_cache_index(PTE_INDEX_SIZE));
329 else
330 pgtable_free_tlb(tlb, hugepte,
331 get_hugepd_cache_index(pdshift - shift));
332}
333
334static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
335 unsigned long addr, unsigned long end,
336 unsigned long floor, unsigned long ceiling)
337{
338 pmd_t *pmd;
339 unsigned long next;
340 unsigned long start;
341
342 start = addr;
343 do {
344 unsigned long more;
345
346 pmd = pmd_offset(pud, addr);
347 next = pmd_addr_end(addr, end);
348 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
349 /*
350 * if it is not hugepd pointer, we should already find
351 * it cleared.
352 */
353 WARN_ON(!pmd_none_or_clear_bad(pmd));
354 continue;
355 }
356 /*
357 * Increment next by the size of the huge mapping since
358 * there may be more than one entry at this level for a
359 * single hugepage, but all of them point to
360 * the same kmem cache that holds the hugepte.
361 */
362 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
363 if (more > next)
364 next = more;
365
366 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
367 addr, next, floor, ceiling);
368 } while (addr = next, addr != end);
369
370 start &= PUD_MASK;
371 if (start < floor)
372 return;
373 if (ceiling) {
374 ceiling &= PUD_MASK;
375 if (!ceiling)
376 return;
377 }
378 if (end - 1 > ceiling - 1)
379 return;
380
381 pmd = pmd_offset(pud, start);
382 pud_clear(pud);
383 pmd_free_tlb(tlb, pmd, start);
384 mm_dec_nr_pmds(tlb->mm);
385}
386
387static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
388 unsigned long addr, unsigned long end,
389 unsigned long floor, unsigned long ceiling)
390{
391 pud_t *pud;
392 unsigned long next;
393 unsigned long start;
394
395 start = addr;
396 do {
397 pud = pud_offset(pgd, addr);
398 next = pud_addr_end(addr, end);
399 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
400 if (pud_none_or_clear_bad(pud))
401 continue;
402 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
403 ceiling);
404 } else {
405 unsigned long more;
406 /*
407 * Increment next by the size of the huge mapping since
408 * there may be more than one entry at this level for a
409 * single hugepage, but all of them point to
410 * the same kmem cache that holds the hugepte.
411 */
412 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
413 if (more > next)
414 next = more;
415
416 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
417 addr, next, floor, ceiling);
418 }
419 } while (addr = next, addr != end);
420
421 start &= PGDIR_MASK;
422 if (start < floor)
423 return;
424 if (ceiling) {
425 ceiling &= PGDIR_MASK;
426 if (!ceiling)
427 return;
428 }
429 if (end - 1 > ceiling - 1)
430 return;
431
432 pud = pud_offset(pgd, start);
433 pgd_clear(pgd);
434 pud_free_tlb(tlb, pud, start);
435 mm_dec_nr_puds(tlb->mm);
436}
437
438/*
439 * This function frees user-level page tables of a process.
440 */
441void hugetlb_free_pgd_range(struct mmu_gather *tlb,
442 unsigned long addr, unsigned long end,
443 unsigned long floor, unsigned long ceiling)
444{
445 pgd_t *pgd;
446 unsigned long next;
447
448 /*
449 * Because there are a number of different possible pagetable
450 * layouts for hugepage ranges, we limit knowledge of how
451 * things should be laid out to the allocation path
452 * (huge_pte_alloc(), above). Everything else works out the
453 * structure as it goes from information in the hugepd
454 * pointers. That means that we can't here use the
455 * optimization used in the normal page free_pgd_range(), of
456 * checking whether we're actually covering a large enough
457 * range to have to do anything at the top level of the walk
458 * instead of at the bottom.
459 *
460 * To make sense of this, you should probably go read the big
461 * block comment at the top of the normal free_pgd_range(),
462 * too.
463 */
464
465 do {
466 next = pgd_addr_end(addr, end);
467 pgd = pgd_offset(tlb->mm, addr);
468 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
469 if (pgd_none_or_clear_bad(pgd))
470 continue;
471 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
472 } else {
473 unsigned long more;
474 /*
475 * Increment next by the size of the huge mapping since
476 * there may be more than one entry at the pgd level
477 * for a single hugepage, but all of them point to the
478 * same kmem cache that holds the hugepte.
479 */
480 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
481 if (more > next)
482 next = more;
483
484 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
485 addr, next, floor, ceiling);
486 }
487 } while (addr = next, addr != end);
488}
489
490struct page *follow_huge_pd(struct vm_area_struct *vma,
491 unsigned long address, hugepd_t hpd,
492 int flags, int pdshift)
493{
494 pte_t *ptep;
495 spinlock_t *ptl;
496 struct page *page = NULL;
497 unsigned long mask;
498 int shift = hugepd_shift(hpd);
499 struct mm_struct *mm = vma->vm_mm;
500
501retry:
502 /*
503 * hugepage directory entries are protected by mm->page_table_lock
504 * Use this instead of huge_pte_lockptr
505 */
506 ptl = &mm->page_table_lock;
507 spin_lock(ptl);
508
509 ptep = hugepte_offset(hpd, address, pdshift);
510 if (pte_present(*ptep)) {
511 mask = (1UL << shift) - 1;
512 page = pte_page(*ptep);
513 page += ((address & mask) >> PAGE_SHIFT);
514 if (flags & FOLL_GET)
515 get_page(page);
516 } else {
517 if (is_hugetlb_entry_migration(*ptep)) {
518 spin_unlock(ptl);
519 __migration_entry_wait(mm, ptep, ptl);
520 goto retry;
521 }
522 }
523 spin_unlock(ptl);
524 return page;
525}
526
527#ifdef CONFIG_PPC_MM_SLICES
528unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
529 unsigned long len, unsigned long pgoff,
530 unsigned long flags)
531{
532 struct hstate *hstate = hstate_file(file);
533 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
534
535#ifdef CONFIG_PPC_RADIX_MMU
536 if (radix_enabled())
537 return radix__hugetlb_get_unmapped_area(file, addr, len,
538 pgoff, flags);
539#endif
540 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
541}
542#endif
543
544unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
545{
546 /* With radix we don't use slice, so derive it from vma*/
547 if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
548 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
549
550 return 1UL << mmu_psize_to_shift(psize);
551 }
552 return vma_kernel_pagesize(vma);
553}
554
555static int __init add_huge_page_size(unsigned long long size)
556{
557 int shift = __ffs(size);
558 int mmu_psize;
559
560 /* Check that it is a page size supported by the hardware and
561 * that it fits within pagetable and slice limits. */
562 if (size <= PAGE_SIZE || !is_power_of_2(size))
563 return -EINVAL;
564
565 mmu_psize = check_and_get_huge_psize(shift);
566 if (mmu_psize < 0)
567 return -EINVAL;
568
569 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
570
571 /* Return if huge page size has already been setup */
572 if (size_to_hstate(size))
573 return 0;
574
575 hugetlb_add_hstate(shift - PAGE_SHIFT);
576
577 return 0;
578}
579
580static int __init hugepage_setup_sz(char *str)
581{
582 unsigned long long size;
583
584 size = memparse(str, &str);
585
586 if (add_huge_page_size(size) != 0) {
587 hugetlb_bad_size();
588 pr_err("Invalid huge page size specified(%llu)\n", size);
589 }
590
591 return 1;
592}
593__setup("hugepagesz=", hugepage_setup_sz);
594
595static int __init hugetlbpage_init(void)
596{
597 bool configured = false;
598 int psize;
599
600 if (hugetlb_disabled) {
601 pr_info("HugeTLB support is disabled!\n");
602 return 0;
603 }
604
605 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
606 !mmu_has_feature(MMU_FTR_16M_PAGE))
607 return -ENODEV;
608
609 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
610 unsigned shift;
611 unsigned pdshift;
612
613 if (!mmu_psize_defs[psize].shift)
614 continue;
615
616 shift = mmu_psize_to_shift(psize);
617
618#ifdef CONFIG_PPC_BOOK3S_64
619 if (shift > PGDIR_SHIFT)
620 continue;
621 else if (shift > PUD_SHIFT)
622 pdshift = PGDIR_SHIFT;
623 else if (shift > PMD_SHIFT)
624 pdshift = PUD_SHIFT;
625 else
626 pdshift = PMD_SHIFT;
627#else
628 if (shift < PUD_SHIFT)
629 pdshift = PMD_SHIFT;
630 else if (shift < PGDIR_SHIFT)
631 pdshift = PUD_SHIFT;
632 else
633 pdshift = PGDIR_SHIFT;
634#endif
635
636 if (add_huge_page_size(1ULL << shift) < 0)
637 continue;
638 /*
639 * if we have pdshift and shift value same, we don't
640 * use pgt cache for hugepd.
641 */
642 if (pdshift > shift && IS_ENABLED(CONFIG_PPC_8xx))
643 pgtable_cache_add(PTE_INDEX_SIZE);
644 else if (pdshift > shift)
645 pgtable_cache_add(pdshift - shift);
646 else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) || IS_ENABLED(CONFIG_PPC_8xx))
647 pgtable_cache_add(PTE_T_ORDER);
648
649 configured = true;
650 }
651
652 if (configured) {
653 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
654 hugetlbpage_init_default();
655 } else
656 pr_info("Failed to initialize. Disabling HugeTLB");
657
658 return 0;
659}
660
661arch_initcall(hugetlbpage_init);
662
663void flush_dcache_icache_hugepage(struct page *page)
664{
665 int i;
666 void *start;
667
668 BUG_ON(!PageCompound(page));
669
670 for (i = 0; i < compound_nr(page); i++) {
671 if (!PageHighMem(page)) {
672 __flush_dcache_icache(page_address(page+i));
673 } else {
674 start = kmap_atomic(page+i);
675 __flush_dcache_icache(start);
676 kunmap_atomic(start);
677 }
678 }
679}
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/mm.h>
11#include <linux/io.h>
12#include <linux/slab.h>
13#include <linux/hugetlb.h>
14#include <asm/pgtable.h>
15#include <asm/pgalloc.h>
16#include <asm/tlb.h>
17
18#define PAGE_SHIFT_64K 16
19#define PAGE_SHIFT_16M 24
20#define PAGE_SHIFT_16G 34
21
22#define MAX_NUMBER_GPAGES 1024
23
24/* Tracks the 16G pages after the device tree is scanned and before the
25 * huge_boot_pages list is ready. */
26static unsigned long gpage_freearray[MAX_NUMBER_GPAGES];
27static unsigned nr_gpages;
28
29/* Flag to mark huge PD pointers. This means pmd_bad() and pud_bad()
30 * will choke on pointers to hugepte tables, which is handy for
31 * catching screwups early. */
32
33static inline int shift_to_mmu_psize(unsigned int shift)
34{
35 int psize;
36
37 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize)
38 if (mmu_psize_defs[psize].shift == shift)
39 return psize;
40 return -1;
41}
42
43static inline unsigned int mmu_psize_to_shift(unsigned int mmu_psize)
44{
45 if (mmu_psize_defs[mmu_psize].shift)
46 return mmu_psize_defs[mmu_psize].shift;
47 BUG();
48}
49
50#define hugepd_none(hpd) ((hpd).pd == 0)
51
52static inline pte_t *hugepd_page(hugepd_t hpd)
53{
54 BUG_ON(!hugepd_ok(hpd));
55 return (pte_t *)((hpd.pd & ~HUGEPD_SHIFT_MASK) | 0xc000000000000000);
56}
57
58static inline unsigned int hugepd_shift(hugepd_t hpd)
59{
60 return hpd.pd & HUGEPD_SHIFT_MASK;
61}
62
63static inline pte_t *hugepte_offset(hugepd_t *hpdp, unsigned long addr, unsigned pdshift)
64{
65 unsigned long idx = (addr & ((1UL << pdshift) - 1)) >> hugepd_shift(*hpdp);
66 pte_t *dir = hugepd_page(*hpdp);
67
68 return dir + idx;
69}
70
71pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
72{
73 pgd_t *pg;
74 pud_t *pu;
75 pmd_t *pm;
76 hugepd_t *hpdp = NULL;
77 unsigned pdshift = PGDIR_SHIFT;
78
79 if (shift)
80 *shift = 0;
81
82 pg = pgdir + pgd_index(ea);
83 if (is_hugepd(pg)) {
84 hpdp = (hugepd_t *)pg;
85 } else if (!pgd_none(*pg)) {
86 pdshift = PUD_SHIFT;
87 pu = pud_offset(pg, ea);
88 if (is_hugepd(pu))
89 hpdp = (hugepd_t *)pu;
90 else if (!pud_none(*pu)) {
91 pdshift = PMD_SHIFT;
92 pm = pmd_offset(pu, ea);
93 if (is_hugepd(pm))
94 hpdp = (hugepd_t *)pm;
95 else if (!pmd_none(*pm)) {
96 return pte_offset_map(pm, ea);
97 }
98 }
99 }
100
101 if (!hpdp)
102 return NULL;
103
104 if (shift)
105 *shift = hugepd_shift(*hpdp);
106 return hugepte_offset(hpdp, ea, pdshift);
107}
108
109pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
110{
111 return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
112}
113
114static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
115 unsigned long address, unsigned pdshift, unsigned pshift)
116{
117 pte_t *new = kmem_cache_zalloc(PGT_CACHE(pdshift - pshift),
118 GFP_KERNEL|__GFP_REPEAT);
119
120 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
121 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
122
123 if (! new)
124 return -ENOMEM;
125
126 spin_lock(&mm->page_table_lock);
127 if (!hugepd_none(*hpdp))
128 kmem_cache_free(PGT_CACHE(pdshift - pshift), new);
129 else
130 hpdp->pd = ((unsigned long)new & ~0x8000000000000000) | pshift;
131 spin_unlock(&mm->page_table_lock);
132 return 0;
133}
134
135pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
136{
137 pgd_t *pg;
138 pud_t *pu;
139 pmd_t *pm;
140 hugepd_t *hpdp = NULL;
141 unsigned pshift = __ffs(sz);
142 unsigned pdshift = PGDIR_SHIFT;
143
144 addr &= ~(sz-1);
145
146 pg = pgd_offset(mm, addr);
147 if (pshift >= PUD_SHIFT) {
148 hpdp = (hugepd_t *)pg;
149 } else {
150 pdshift = PUD_SHIFT;
151 pu = pud_alloc(mm, pg, addr);
152 if (pshift >= PMD_SHIFT) {
153 hpdp = (hugepd_t *)pu;
154 } else {
155 pdshift = PMD_SHIFT;
156 pm = pmd_alloc(mm, pu, addr);
157 hpdp = (hugepd_t *)pm;
158 }
159 }
160
161 if (!hpdp)
162 return NULL;
163
164 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
165
166 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
167 return NULL;
168
169 return hugepte_offset(hpdp, addr, pdshift);
170}
171
172/* Build list of addresses of gigantic pages. This function is used in early
173 * boot before the buddy or bootmem allocator is setup.
174 */
175void add_gpage(unsigned long addr, unsigned long page_size,
176 unsigned long number_of_pages)
177{
178 if (!addr)
179 return;
180 while (number_of_pages > 0) {
181 gpage_freearray[nr_gpages] = addr;
182 nr_gpages++;
183 number_of_pages--;
184 addr += page_size;
185 }
186}
187
188/* Moves the gigantic page addresses from the temporary list to the
189 * huge_boot_pages list.
190 */
191int alloc_bootmem_huge_page(struct hstate *hstate)
192{
193 struct huge_bootmem_page *m;
194 if (nr_gpages == 0)
195 return 0;
196 m = phys_to_virt(gpage_freearray[--nr_gpages]);
197 gpage_freearray[nr_gpages] = 0;
198 list_add(&m->list, &huge_boot_pages);
199 m->hstate = hstate;
200 return 1;
201}
202
203int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
204{
205 return 0;
206}
207
208static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
209 unsigned long start, unsigned long end,
210 unsigned long floor, unsigned long ceiling)
211{
212 pte_t *hugepte = hugepd_page(*hpdp);
213 unsigned shift = hugepd_shift(*hpdp);
214 unsigned long pdmask = ~((1UL << pdshift) - 1);
215
216 start &= pdmask;
217 if (start < floor)
218 return;
219 if (ceiling) {
220 ceiling &= pdmask;
221 if (! ceiling)
222 return;
223 }
224 if (end - 1 > ceiling - 1)
225 return;
226
227 hpdp->pd = 0;
228 tlb->need_flush = 1;
229 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
230}
231
232static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
233 unsigned long addr, unsigned long end,
234 unsigned long floor, unsigned long ceiling)
235{
236 pmd_t *pmd;
237 unsigned long next;
238 unsigned long start;
239
240 start = addr;
241 pmd = pmd_offset(pud, addr);
242 do {
243 next = pmd_addr_end(addr, end);
244 if (pmd_none(*pmd))
245 continue;
246 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
247 addr, next, floor, ceiling);
248 } while (pmd++, addr = next, addr != end);
249
250 start &= PUD_MASK;
251 if (start < floor)
252 return;
253 if (ceiling) {
254 ceiling &= PUD_MASK;
255 if (!ceiling)
256 return;
257 }
258 if (end - 1 > ceiling - 1)
259 return;
260
261 pmd = pmd_offset(pud, start);
262 pud_clear(pud);
263 pmd_free_tlb(tlb, pmd, start);
264}
265
266static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
267 unsigned long addr, unsigned long end,
268 unsigned long floor, unsigned long ceiling)
269{
270 pud_t *pud;
271 unsigned long next;
272 unsigned long start;
273
274 start = addr;
275 pud = pud_offset(pgd, addr);
276 do {
277 next = pud_addr_end(addr, end);
278 if (!is_hugepd(pud)) {
279 if (pud_none_or_clear_bad(pud))
280 continue;
281 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
282 ceiling);
283 } else {
284 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
285 addr, next, floor, ceiling);
286 }
287 } while (pud++, addr = next, addr != end);
288
289 start &= PGDIR_MASK;
290 if (start < floor)
291 return;
292 if (ceiling) {
293 ceiling &= PGDIR_MASK;
294 if (!ceiling)
295 return;
296 }
297 if (end - 1 > ceiling - 1)
298 return;
299
300 pud = pud_offset(pgd, start);
301 pgd_clear(pgd);
302 pud_free_tlb(tlb, pud, start);
303}
304
305/*
306 * This function frees user-level page tables of a process.
307 *
308 * Must be called with pagetable lock held.
309 */
310void hugetlb_free_pgd_range(struct mmu_gather *tlb,
311 unsigned long addr, unsigned long end,
312 unsigned long floor, unsigned long ceiling)
313{
314 pgd_t *pgd;
315 unsigned long next;
316
317 /*
318 * Because there are a number of different possible pagetable
319 * layouts for hugepage ranges, we limit knowledge of how
320 * things should be laid out to the allocation path
321 * (huge_pte_alloc(), above). Everything else works out the
322 * structure as it goes from information in the hugepd
323 * pointers. That means that we can't here use the
324 * optimization used in the normal page free_pgd_range(), of
325 * checking whether we're actually covering a large enough
326 * range to have to do anything at the top level of the walk
327 * instead of at the bottom.
328 *
329 * To make sense of this, you should probably go read the big
330 * block comment at the top of the normal free_pgd_range(),
331 * too.
332 */
333
334 pgd = pgd_offset(tlb->mm, addr);
335 do {
336 next = pgd_addr_end(addr, end);
337 if (!is_hugepd(pgd)) {
338 if (pgd_none_or_clear_bad(pgd))
339 continue;
340 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
341 } else {
342 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
343 addr, next, floor, ceiling);
344 }
345 } while (pgd++, addr = next, addr != end);
346}
347
348struct page *
349follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
350{
351 pte_t *ptep;
352 struct page *page;
353 unsigned shift;
354 unsigned long mask;
355
356 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
357
358 /* Verify it is a huge page else bail. */
359 if (!ptep || !shift)
360 return ERR_PTR(-EINVAL);
361
362 mask = (1UL << shift) - 1;
363 page = pte_page(*ptep);
364 if (page)
365 page += (address & mask) / PAGE_SIZE;
366
367 return page;
368}
369
370int pmd_huge(pmd_t pmd)
371{
372 return 0;
373}
374
375int pud_huge(pud_t pud)
376{
377 return 0;
378}
379
380struct page *
381follow_huge_pmd(struct mm_struct *mm, unsigned long address,
382 pmd_t *pmd, int write)
383{
384 BUG();
385 return NULL;
386}
387
388static noinline int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
389 unsigned long end, int write, struct page **pages, int *nr)
390{
391 unsigned long mask;
392 unsigned long pte_end;
393 struct page *head, *page;
394 pte_t pte;
395 int refs;
396
397 pte_end = (addr + sz) & ~(sz-1);
398 if (pte_end < end)
399 end = pte_end;
400
401 pte = *ptep;
402 mask = _PAGE_PRESENT | _PAGE_USER;
403 if (write)
404 mask |= _PAGE_RW;
405
406 if ((pte_val(pte) & mask) != mask)
407 return 0;
408
409 /* hugepages are never "special" */
410 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
411
412 refs = 0;
413 head = pte_page(pte);
414
415 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
416 do {
417 VM_BUG_ON(compound_head(page) != head);
418 pages[*nr] = page;
419 (*nr)++;
420 page++;
421 refs++;
422 } while (addr += PAGE_SIZE, addr != end);
423
424 if (!page_cache_add_speculative(head, refs)) {
425 *nr -= refs;
426 return 0;
427 }
428
429 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
430 /* Could be optimized better */
431 while (*nr) {
432 put_page(page);
433 (*nr)--;
434 }
435 }
436
437 return 1;
438}
439
440static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
441 unsigned long sz)
442{
443 unsigned long __boundary = (addr + sz) & ~(sz-1);
444 return (__boundary - 1 < end - 1) ? __boundary : end;
445}
446
447int gup_hugepd(hugepd_t *hugepd, unsigned pdshift,
448 unsigned long addr, unsigned long end,
449 int write, struct page **pages, int *nr)
450{
451 pte_t *ptep;
452 unsigned long sz = 1UL << hugepd_shift(*hugepd);
453 unsigned long next;
454
455 ptep = hugepte_offset(hugepd, addr, pdshift);
456 do {
457 next = hugepte_addr_end(addr, end, sz);
458 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
459 return 0;
460 } while (ptep++, addr = next, addr != end);
461
462 return 1;
463}
464
465unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
466 unsigned long len, unsigned long pgoff,
467 unsigned long flags)
468{
469 struct hstate *hstate = hstate_file(file);
470 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
471
472 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1, 0);
473}
474
475unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
476{
477 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
478
479 return 1UL << mmu_psize_to_shift(psize);
480}
481
482static int __init add_huge_page_size(unsigned long long size)
483{
484 int shift = __ffs(size);
485 int mmu_psize;
486
487 /* Check that it is a page size supported by the hardware and
488 * that it fits within pagetable and slice limits. */
489 if (!is_power_of_2(size)
490 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
491 return -EINVAL;
492
493 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
494 return -EINVAL;
495
496#ifdef CONFIG_SPU_FS_64K_LS
497 /* Disable support for 64K huge pages when 64K SPU local store
498 * support is enabled as the current implementation conflicts.
499 */
500 if (shift == PAGE_SHIFT_64K)
501 return -EINVAL;
502#endif /* CONFIG_SPU_FS_64K_LS */
503
504 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
505
506 /* Return if huge page size has already been setup */
507 if (size_to_hstate(size))
508 return 0;
509
510 hugetlb_add_hstate(shift - PAGE_SHIFT);
511
512 return 0;
513}
514
515static int __init hugepage_setup_sz(char *str)
516{
517 unsigned long long size;
518
519 size = memparse(str, &str);
520
521 if (add_huge_page_size(size) != 0)
522 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
523
524 return 1;
525}
526__setup("hugepagesz=", hugepage_setup_sz);
527
528static int __init hugetlbpage_init(void)
529{
530 int psize;
531
532 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
533 return -ENODEV;
534
535 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
536 unsigned shift;
537 unsigned pdshift;
538
539 if (!mmu_psize_defs[psize].shift)
540 continue;
541
542 shift = mmu_psize_to_shift(psize);
543
544 if (add_huge_page_size(1ULL << shift) < 0)
545 continue;
546
547 if (shift < PMD_SHIFT)
548 pdshift = PMD_SHIFT;
549 else if (shift < PUD_SHIFT)
550 pdshift = PUD_SHIFT;
551 else
552 pdshift = PGDIR_SHIFT;
553
554 pgtable_cache_add(pdshift - shift, NULL);
555 if (!PGT_CACHE(pdshift - shift))
556 panic("hugetlbpage_init(): could not create "
557 "pgtable cache for %d bit pagesize\n", shift);
558 }
559
560 /* Set default large page size. Currently, we pick 16M or 1M
561 * depending on what is available
562 */
563 if (mmu_psize_defs[MMU_PAGE_16M].shift)
564 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
565 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
566 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
567
568 return 0;
569}
570
571module_init(hugetlbpage_init);
572
573void flush_dcache_icache_hugepage(struct page *page)
574{
575 int i;
576
577 BUG_ON(!PageCompound(page));
578
579 for (i = 0; i < (1UL << compound_order(page)); i++)
580 __flush_dcache_icache(page_address(page+i));
581}