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 * 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/bootmem.h>
19#include <linux/moduleparam.h>
20#include <asm/pgtable.h>
21#include <asm/pgalloc.h>
22#include <asm/tlb.h>
23#include <asm/setup.h>
24#include <asm/hugetlb.h>
25
26#ifdef CONFIG_HUGETLB_PAGE
27
28#define PAGE_SHIFT_64K 16
29#define PAGE_SHIFT_16M 24
30#define PAGE_SHIFT_16G 34
31
32unsigned int HPAGE_SHIFT;
33
34/*
35 * Tracks gpages after the device tree is scanned and before the
36 * huge_boot_pages list is ready. On non-Freescale implementations, this is
37 * just used to track 16G pages and so is a single array. FSL-based
38 * implementations may have more than one gpage size, so we need multiple
39 * arrays
40 */
41#ifdef CONFIG_PPC_FSL_BOOK3E
42#define MAX_NUMBER_GPAGES 128
43struct psize_gpages {
44 u64 gpage_list[MAX_NUMBER_GPAGES];
45 unsigned int nr_gpages;
46};
47static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
48#else
49#define MAX_NUMBER_GPAGES 1024
50static u64 gpage_freearray[MAX_NUMBER_GPAGES];
51static unsigned nr_gpages;
52#endif
53
54#define hugepd_none(hpd) ((hpd).pd == 0)
55
56pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
57{
58 /* Only called for hugetlbfs pages, hence can ignore THP */
59 return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL);
60}
61
62static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
63 unsigned long address, unsigned pdshift, unsigned pshift)
64{
65 struct kmem_cache *cachep;
66 pte_t *new;
67
68#ifdef CONFIG_PPC_FSL_BOOK3E
69 int i;
70 int num_hugepd = 1 << (pshift - pdshift);
71 cachep = hugepte_cache;
72#else
73 cachep = PGT_CACHE(pdshift - pshift);
74#endif
75
76 new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
77
78 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
79 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
80
81 if (! new)
82 return -ENOMEM;
83
84 spin_lock(&mm->page_table_lock);
85#ifdef CONFIG_PPC_FSL_BOOK3E
86 /*
87 * We have multiple higher-level entries that point to the same
88 * actual pte location. Fill in each as we go and backtrack on error.
89 * We need all of these so the DTLB pgtable walk code can find the
90 * right higher-level entry without knowing if it's a hugepage or not.
91 */
92 for (i = 0; i < num_hugepd; i++, hpdp++) {
93 if (unlikely(!hugepd_none(*hpdp)))
94 break;
95 else
96 /* We use the old format for PPC_FSL_BOOK3E */
97 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
98 }
99 /* If we bailed from the for loop early, an error occurred, clean up */
100 if (i < num_hugepd) {
101 for (i = i - 1 ; i >= 0; i--, hpdp--)
102 hpdp->pd = 0;
103 kmem_cache_free(cachep, new);
104 }
105#else
106 if (!hugepd_none(*hpdp))
107 kmem_cache_free(cachep, new);
108 else {
109#ifdef CONFIG_PPC_BOOK3S_64
110 hpdp->pd = __pa(new) | (shift_to_mmu_psize(pshift) << 2);
111#else
112 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
113#endif
114 }
115#endif
116 spin_unlock(&mm->page_table_lock);
117 return 0;
118}
119
120/*
121 * These macros define how to determine which level of the page table holds
122 * the hpdp.
123 */
124#ifdef CONFIG_PPC_FSL_BOOK3E
125#define HUGEPD_PGD_SHIFT PGDIR_SHIFT
126#define HUGEPD_PUD_SHIFT PUD_SHIFT
127#else
128#define HUGEPD_PGD_SHIFT PUD_SHIFT
129#define HUGEPD_PUD_SHIFT PMD_SHIFT
130#endif
131
132#ifdef CONFIG_PPC_BOOK3S_64
133/*
134 * At this point we do the placement change only for BOOK3S 64. This would
135 * possibly work on other subarchs.
136 */
137pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
138{
139 pgd_t *pg;
140 pud_t *pu;
141 pmd_t *pm;
142 hugepd_t *hpdp = NULL;
143 unsigned pshift = __ffs(sz);
144 unsigned pdshift = PGDIR_SHIFT;
145
146 addr &= ~(sz-1);
147 pg = pgd_offset(mm, addr);
148
149 if (pshift == PGDIR_SHIFT)
150 /* 16GB huge page */
151 return (pte_t *) pg;
152 else if (pshift > PUD_SHIFT)
153 /*
154 * We need to use hugepd table
155 */
156 hpdp = (hugepd_t *)pg;
157 else {
158 pdshift = PUD_SHIFT;
159 pu = pud_alloc(mm, pg, addr);
160 if (pshift == PUD_SHIFT)
161 return (pte_t *)pu;
162 else if (pshift > PMD_SHIFT)
163 hpdp = (hugepd_t *)pu;
164 else {
165 pdshift = PMD_SHIFT;
166 pm = pmd_alloc(mm, pu, addr);
167 if (pshift == PMD_SHIFT)
168 /* 16MB hugepage */
169 return (pte_t *)pm;
170 else
171 hpdp = (hugepd_t *)pm;
172 }
173 }
174 if (!hpdp)
175 return NULL;
176
177 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
178
179 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
180 return NULL;
181
182 return hugepte_offset(*hpdp, addr, pdshift);
183}
184
185#else
186
187pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
188{
189 pgd_t *pg;
190 pud_t *pu;
191 pmd_t *pm;
192 hugepd_t *hpdp = NULL;
193 unsigned pshift = __ffs(sz);
194 unsigned pdshift = PGDIR_SHIFT;
195
196 addr &= ~(sz-1);
197
198 pg = pgd_offset(mm, addr);
199
200 if (pshift >= HUGEPD_PGD_SHIFT) {
201 hpdp = (hugepd_t *)pg;
202 } else {
203 pdshift = PUD_SHIFT;
204 pu = pud_alloc(mm, pg, addr);
205 if (pshift >= HUGEPD_PUD_SHIFT) {
206 hpdp = (hugepd_t *)pu;
207 } else {
208 pdshift = PMD_SHIFT;
209 pm = pmd_alloc(mm, pu, addr);
210 hpdp = (hugepd_t *)pm;
211 }
212 }
213
214 if (!hpdp)
215 return NULL;
216
217 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
218
219 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
220 return NULL;
221
222 return hugepte_offset(*hpdp, addr, pdshift);
223}
224#endif
225
226#ifdef CONFIG_PPC_FSL_BOOK3E
227/* Build list of addresses of gigantic pages. This function is used in early
228 * boot before the buddy allocator is setup.
229 */
230void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
231{
232 unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
233 int i;
234
235 if (addr == 0)
236 return;
237
238 gpage_freearray[idx].nr_gpages = number_of_pages;
239
240 for (i = 0; i < number_of_pages; i++) {
241 gpage_freearray[idx].gpage_list[i] = addr;
242 addr += page_size;
243 }
244}
245
246/*
247 * Moves the gigantic page addresses from the temporary list to the
248 * huge_boot_pages list.
249 */
250int alloc_bootmem_huge_page(struct hstate *hstate)
251{
252 struct huge_bootmem_page *m;
253 int idx = shift_to_mmu_psize(huge_page_shift(hstate));
254 int nr_gpages = gpage_freearray[idx].nr_gpages;
255
256 if (nr_gpages == 0)
257 return 0;
258
259#ifdef CONFIG_HIGHMEM
260 /*
261 * If gpages can be in highmem we can't use the trick of storing the
262 * data structure in the page; allocate space for this
263 */
264 m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
265 m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
266#else
267 m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
268#endif
269
270 list_add(&m->list, &huge_boot_pages);
271 gpage_freearray[idx].nr_gpages = nr_gpages;
272 gpage_freearray[idx].gpage_list[nr_gpages] = 0;
273 m->hstate = hstate;
274
275 return 1;
276}
277/*
278 * Scan the command line hugepagesz= options for gigantic pages; store those in
279 * a list that we use to allocate the memory once all options are parsed.
280 */
281
282unsigned long gpage_npages[MMU_PAGE_COUNT];
283
284static int __init do_gpage_early_setup(char *param, char *val,
285 const char *unused, void *arg)
286{
287 static phys_addr_t size;
288 unsigned long npages;
289
290 /*
291 * The hugepagesz and hugepages cmdline options are interleaved. We
292 * use the size variable to keep track of whether or not this was done
293 * properly and skip over instances where it is incorrect. Other
294 * command-line parsing code will issue warnings, so we don't need to.
295 *
296 */
297 if ((strcmp(param, "default_hugepagesz") == 0) ||
298 (strcmp(param, "hugepagesz") == 0)) {
299 size = memparse(val, NULL);
300 } else if (strcmp(param, "hugepages") == 0) {
301 if (size != 0) {
302 if (sscanf(val, "%lu", &npages) <= 0)
303 npages = 0;
304 if (npages > MAX_NUMBER_GPAGES) {
305 pr_warn("MMU: %lu pages requested for page "
306 "size %llu KB, limiting to "
307 __stringify(MAX_NUMBER_GPAGES) "\n",
308 npages, size / 1024);
309 npages = MAX_NUMBER_GPAGES;
310 }
311 gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
312 size = 0;
313 }
314 }
315 return 0;
316}
317
318
319/*
320 * This function allocates physical space for pages that are larger than the
321 * buddy allocator can handle. We want to allocate these in highmem because
322 * the amount of lowmem is limited. This means that this function MUST be
323 * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
324 * allocate to grab highmem.
325 */
326void __init reserve_hugetlb_gpages(void)
327{
328 static __initdata char cmdline[COMMAND_LINE_SIZE];
329 phys_addr_t size, base;
330 int i;
331
332 strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
333 parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
334 NULL, &do_gpage_early_setup);
335
336 /*
337 * Walk gpage list in reverse, allocating larger page sizes first.
338 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
339 * When we reach the point in the list where pages are no longer
340 * considered gpages, we're done.
341 */
342 for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
343 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
344 continue;
345 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
346 break;
347
348 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
349 base = memblock_alloc_base(size * gpage_npages[i], size,
350 MEMBLOCK_ALLOC_ANYWHERE);
351 add_gpage(base, size, gpage_npages[i]);
352 }
353}
354
355#else /* !PPC_FSL_BOOK3E */
356
357/* Build list of addresses of gigantic pages. This function is used in early
358 * boot before the buddy allocator is setup.
359 */
360void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
361{
362 if (!addr)
363 return;
364 while (number_of_pages > 0) {
365 gpage_freearray[nr_gpages] = addr;
366 nr_gpages++;
367 number_of_pages--;
368 addr += page_size;
369 }
370}
371
372/* Moves the gigantic page addresses from the temporary list to the
373 * huge_boot_pages list.
374 */
375int alloc_bootmem_huge_page(struct hstate *hstate)
376{
377 struct huge_bootmem_page *m;
378 if (nr_gpages == 0)
379 return 0;
380 m = phys_to_virt(gpage_freearray[--nr_gpages]);
381 gpage_freearray[nr_gpages] = 0;
382 list_add(&m->list, &huge_boot_pages);
383 m->hstate = hstate;
384 return 1;
385}
386#endif
387
388#ifdef CONFIG_PPC_FSL_BOOK3E
389#define HUGEPD_FREELIST_SIZE \
390 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
391
392struct hugepd_freelist {
393 struct rcu_head rcu;
394 unsigned int index;
395 void *ptes[0];
396};
397
398static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
399
400static void hugepd_free_rcu_callback(struct rcu_head *head)
401{
402 struct hugepd_freelist *batch =
403 container_of(head, struct hugepd_freelist, rcu);
404 unsigned int i;
405
406 for (i = 0; i < batch->index; i++)
407 kmem_cache_free(hugepte_cache, batch->ptes[i]);
408
409 free_page((unsigned long)batch);
410}
411
412static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
413{
414 struct hugepd_freelist **batchp;
415
416 batchp = &get_cpu_var(hugepd_freelist_cur);
417
418 if (atomic_read(&tlb->mm->mm_users) < 2 ||
419 cpumask_equal(mm_cpumask(tlb->mm),
420 cpumask_of(smp_processor_id()))) {
421 kmem_cache_free(hugepte_cache, hugepte);
422 put_cpu_var(hugepd_freelist_cur);
423 return;
424 }
425
426 if (*batchp == NULL) {
427 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
428 (*batchp)->index = 0;
429 }
430
431 (*batchp)->ptes[(*batchp)->index++] = hugepte;
432 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
433 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
434 *batchp = NULL;
435 }
436 put_cpu_var(hugepd_freelist_cur);
437}
438#endif
439
440static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
441 unsigned long start, unsigned long end,
442 unsigned long floor, unsigned long ceiling)
443{
444 pte_t *hugepte = hugepd_page(*hpdp);
445 int i;
446
447 unsigned long pdmask = ~((1UL << pdshift) - 1);
448 unsigned int num_hugepd = 1;
449
450#ifdef CONFIG_PPC_FSL_BOOK3E
451 /* Note: On fsl the hpdp may be the first of several */
452 num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
453#else
454 unsigned int shift = hugepd_shift(*hpdp);
455#endif
456
457 start &= pdmask;
458 if (start < floor)
459 return;
460 if (ceiling) {
461 ceiling &= pdmask;
462 if (! ceiling)
463 return;
464 }
465 if (end - 1 > ceiling - 1)
466 return;
467
468 for (i = 0; i < num_hugepd; i++, hpdp++)
469 hpdp->pd = 0;
470
471#ifdef CONFIG_PPC_FSL_BOOK3E
472 hugepd_free(tlb, hugepte);
473#else
474 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
475#endif
476}
477
478static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
479 unsigned long addr, unsigned long end,
480 unsigned long floor, unsigned long ceiling)
481{
482 pmd_t *pmd;
483 unsigned long next;
484 unsigned long start;
485
486 start = addr;
487 do {
488 pmd = pmd_offset(pud, addr);
489 next = pmd_addr_end(addr, end);
490 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
491 /*
492 * if it is not hugepd pointer, we should already find
493 * it cleared.
494 */
495 WARN_ON(!pmd_none_or_clear_bad(pmd));
496 continue;
497 }
498#ifdef CONFIG_PPC_FSL_BOOK3E
499 /*
500 * Increment next by the size of the huge mapping since
501 * there may be more than one entry at this level for a
502 * single hugepage, but all of them point to
503 * the same kmem cache that holds the hugepte.
504 */
505 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
506#endif
507 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
508 addr, next, floor, ceiling);
509 } while (addr = next, addr != end);
510
511 start &= PUD_MASK;
512 if (start < floor)
513 return;
514 if (ceiling) {
515 ceiling &= PUD_MASK;
516 if (!ceiling)
517 return;
518 }
519 if (end - 1 > ceiling - 1)
520 return;
521
522 pmd = pmd_offset(pud, start);
523 pud_clear(pud);
524 pmd_free_tlb(tlb, pmd, start);
525 mm_dec_nr_pmds(tlb->mm);
526}
527
528static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
529 unsigned long addr, unsigned long end,
530 unsigned long floor, unsigned long ceiling)
531{
532 pud_t *pud;
533 unsigned long next;
534 unsigned long start;
535
536 start = addr;
537 do {
538 pud = pud_offset(pgd, addr);
539 next = pud_addr_end(addr, end);
540 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
541 if (pud_none_or_clear_bad(pud))
542 continue;
543 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
544 ceiling);
545 } else {
546#ifdef CONFIG_PPC_FSL_BOOK3E
547 /*
548 * Increment next by the size of the huge mapping since
549 * there may be more than one entry at this level for a
550 * single hugepage, but all of them point to
551 * the same kmem cache that holds the hugepte.
552 */
553 next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
554#endif
555 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
556 addr, next, floor, ceiling);
557 }
558 } while (addr = next, addr != end);
559
560 start &= PGDIR_MASK;
561 if (start < floor)
562 return;
563 if (ceiling) {
564 ceiling &= PGDIR_MASK;
565 if (!ceiling)
566 return;
567 }
568 if (end - 1 > ceiling - 1)
569 return;
570
571 pud = pud_offset(pgd, start);
572 pgd_clear(pgd);
573 pud_free_tlb(tlb, pud, start);
574}
575
576/*
577 * This function frees user-level page tables of a process.
578 */
579void hugetlb_free_pgd_range(struct mmu_gather *tlb,
580 unsigned long addr, unsigned long end,
581 unsigned long floor, unsigned long ceiling)
582{
583 pgd_t *pgd;
584 unsigned long next;
585
586 /*
587 * Because there are a number of different possible pagetable
588 * layouts for hugepage ranges, we limit knowledge of how
589 * things should be laid out to the allocation path
590 * (huge_pte_alloc(), above). Everything else works out the
591 * structure as it goes from information in the hugepd
592 * pointers. That means that we can't here use the
593 * optimization used in the normal page free_pgd_range(), of
594 * checking whether we're actually covering a large enough
595 * range to have to do anything at the top level of the walk
596 * instead of at the bottom.
597 *
598 * To make sense of this, you should probably go read the big
599 * block comment at the top of the normal free_pgd_range(),
600 * too.
601 */
602
603 do {
604 next = pgd_addr_end(addr, end);
605 pgd = pgd_offset(tlb->mm, addr);
606 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
607 if (pgd_none_or_clear_bad(pgd))
608 continue;
609 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
610 } else {
611#ifdef CONFIG_PPC_FSL_BOOK3E
612 /*
613 * Increment next by the size of the huge mapping since
614 * there may be more than one entry at the pgd level
615 * for a single hugepage, but all of them point to the
616 * same kmem cache that holds the hugepte.
617 */
618 next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
619#endif
620 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
621 addr, next, floor, ceiling);
622 }
623 } while (addr = next, addr != end);
624}
625
626/*
627 * We are holding mmap_sem, so a parallel huge page collapse cannot run.
628 * To prevent hugepage split, disable irq.
629 */
630struct page *
631follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
632{
633 bool is_thp;
634 pte_t *ptep, pte;
635 unsigned shift;
636 unsigned long mask, flags;
637 struct page *page = ERR_PTR(-EINVAL);
638
639 local_irq_save(flags);
640 ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift);
641 if (!ptep)
642 goto no_page;
643 pte = READ_ONCE(*ptep);
644 /*
645 * Verify it is a huge page else bail.
646 * Transparent hugepages are handled by generic code. We can skip them
647 * here.
648 */
649 if (!shift || is_thp)
650 goto no_page;
651
652 if (!pte_present(pte)) {
653 page = NULL;
654 goto no_page;
655 }
656 mask = (1UL << shift) - 1;
657 page = pte_page(pte);
658 if (page)
659 page += (address & mask) / PAGE_SIZE;
660
661no_page:
662 local_irq_restore(flags);
663 return page;
664}
665
666struct page *
667follow_huge_pmd(struct mm_struct *mm, unsigned long address,
668 pmd_t *pmd, int write)
669{
670 BUG();
671 return NULL;
672}
673
674struct page *
675follow_huge_pud(struct mm_struct *mm, unsigned long address,
676 pud_t *pud, int write)
677{
678 BUG();
679 return NULL;
680}
681
682static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
683 unsigned long sz)
684{
685 unsigned long __boundary = (addr + sz) & ~(sz-1);
686 return (__boundary - 1 < end - 1) ? __boundary : end;
687}
688
689int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
690 unsigned long end, int write, struct page **pages, int *nr)
691{
692 pte_t *ptep;
693 unsigned long sz = 1UL << hugepd_shift(hugepd);
694 unsigned long next;
695
696 ptep = hugepte_offset(hugepd, addr, pdshift);
697 do {
698 next = hugepte_addr_end(addr, end, sz);
699 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
700 return 0;
701 } while (ptep++, addr = next, addr != end);
702
703 return 1;
704}
705
706#ifdef CONFIG_PPC_MM_SLICES
707unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
708 unsigned long len, unsigned long pgoff,
709 unsigned long flags)
710{
711 struct hstate *hstate = hstate_file(file);
712 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
713
714 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
715}
716#endif
717
718unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
719{
720#ifdef CONFIG_PPC_MM_SLICES
721 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
722
723 return 1UL << mmu_psize_to_shift(psize);
724#else
725 if (!is_vm_hugetlb_page(vma))
726 return PAGE_SIZE;
727
728 return huge_page_size(hstate_vma(vma));
729#endif
730}
731
732static inline bool is_power_of_4(unsigned long x)
733{
734 if (is_power_of_2(x))
735 return (__ilog2(x) % 2) ? false : true;
736 return false;
737}
738
739static int __init add_huge_page_size(unsigned long long size)
740{
741 int shift = __ffs(size);
742 int mmu_psize;
743
744 /* Check that it is a page size supported by the hardware and
745 * that it fits within pagetable and slice limits. */
746#ifdef CONFIG_PPC_FSL_BOOK3E
747 if ((size < PAGE_SIZE) || !is_power_of_4(size))
748 return -EINVAL;
749#else
750 if (!is_power_of_2(size)
751 || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
752 return -EINVAL;
753#endif
754
755 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
756 return -EINVAL;
757
758 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
759
760 /* Return if huge page size has already been setup */
761 if (size_to_hstate(size))
762 return 0;
763
764 hugetlb_add_hstate(shift - PAGE_SHIFT);
765
766 return 0;
767}
768
769static int __init hugepage_setup_sz(char *str)
770{
771 unsigned long long size;
772
773 size = memparse(str, &str);
774
775 if (add_huge_page_size(size) != 0)
776 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
777
778 return 1;
779}
780__setup("hugepagesz=", hugepage_setup_sz);
781
782#ifdef CONFIG_PPC_FSL_BOOK3E
783struct kmem_cache *hugepte_cache;
784static int __init hugetlbpage_init(void)
785{
786 int psize;
787
788 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
789 unsigned shift;
790
791 if (!mmu_psize_defs[psize].shift)
792 continue;
793
794 shift = mmu_psize_to_shift(psize);
795
796 /* Don't treat normal page sizes as huge... */
797 if (shift != PAGE_SHIFT)
798 if (add_huge_page_size(1ULL << shift) < 0)
799 continue;
800 }
801
802 /*
803 * Create a kmem cache for hugeptes. The bottom bits in the pte have
804 * size information encoded in them, so align them to allow this
805 */
806 hugepte_cache = kmem_cache_create("hugepte-cache", sizeof(pte_t),
807 HUGEPD_SHIFT_MASK + 1, 0, NULL);
808 if (hugepte_cache == NULL)
809 panic("%s: Unable to create kmem cache for hugeptes\n",
810 __func__);
811
812 /* Default hpage size = 4M */
813 if (mmu_psize_defs[MMU_PAGE_4M].shift)
814 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
815 else
816 panic("%s: Unable to set default huge page size\n", __func__);
817
818
819 return 0;
820}
821#else
822static int __init hugetlbpage_init(void)
823{
824 int psize;
825
826 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
827 return -ENODEV;
828
829 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
830 unsigned shift;
831 unsigned pdshift;
832
833 if (!mmu_psize_defs[psize].shift)
834 continue;
835
836 shift = mmu_psize_to_shift(psize);
837
838 if (add_huge_page_size(1ULL << shift) < 0)
839 continue;
840
841 if (shift < PMD_SHIFT)
842 pdshift = PMD_SHIFT;
843 else if (shift < PUD_SHIFT)
844 pdshift = PUD_SHIFT;
845 else
846 pdshift = PGDIR_SHIFT;
847 /*
848 * if we have pdshift and shift value same, we don't
849 * use pgt cache for hugepd.
850 */
851 if (pdshift != shift) {
852 pgtable_cache_add(pdshift - shift, NULL);
853 if (!PGT_CACHE(pdshift - shift))
854 panic("hugetlbpage_init(): could not create "
855 "pgtable cache for %d bit pagesize\n", shift);
856 }
857 }
858
859 /* Set default large page size. Currently, we pick 16M or 1M
860 * depending on what is available
861 */
862 if (mmu_psize_defs[MMU_PAGE_16M].shift)
863 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
864 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
865 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
866
867 return 0;
868}
869#endif
870arch_initcall(hugetlbpage_init);
871
872void flush_dcache_icache_hugepage(struct page *page)
873{
874 int i;
875 void *start;
876
877 BUG_ON(!PageCompound(page));
878
879 for (i = 0; i < (1UL << compound_order(page)); i++) {
880 if (!PageHighMem(page)) {
881 __flush_dcache_icache(page_address(page+i));
882 } else {
883 start = kmap_atomic(page+i);
884 __flush_dcache_icache(start);
885 kunmap_atomic(start);
886 }
887 }
888}
889
890#endif /* CONFIG_HUGETLB_PAGE */
891
892/*
893 * We have 4 cases for pgds and pmds:
894 * (1) invalid (all zeroes)
895 * (2) pointer to next table, as normal; bottom 6 bits == 0
896 * (3) leaf pte for huge page _PAGE_PTE set
897 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
898 *
899 * So long as we atomically load page table pointers we are safe against teardown,
900 * we can follow the address down to the the page and take a ref on it.
901 * This function need to be called with interrupts disabled. We use this variant
902 * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
903 */
904
905pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
906 bool *is_thp, unsigned *shift)
907{
908 pgd_t pgd, *pgdp;
909 pud_t pud, *pudp;
910 pmd_t pmd, *pmdp;
911 pte_t *ret_pte;
912 hugepd_t *hpdp = NULL;
913 unsigned pdshift = PGDIR_SHIFT;
914
915 if (shift)
916 *shift = 0;
917
918 if (is_thp)
919 *is_thp = false;
920
921 pgdp = pgdir + pgd_index(ea);
922 pgd = READ_ONCE(*pgdp);
923 /*
924 * Always operate on the local stack value. This make sure the
925 * value don't get updated by a parallel THP split/collapse,
926 * page fault or a page unmap. The return pte_t * is still not
927 * stable. So should be checked there for above conditions.
928 */
929 if (pgd_none(pgd))
930 return NULL;
931 else if (pgd_huge(pgd)) {
932 ret_pte = (pte_t *) pgdp;
933 goto out;
934 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
935 hpdp = (hugepd_t *)&pgd;
936 else {
937 /*
938 * Even if we end up with an unmap, the pgtable will not
939 * be freed, because we do an rcu free and here we are
940 * irq disabled
941 */
942 pdshift = PUD_SHIFT;
943 pudp = pud_offset(&pgd, ea);
944 pud = READ_ONCE(*pudp);
945
946 if (pud_none(pud))
947 return NULL;
948 else if (pud_huge(pud)) {
949 ret_pte = (pte_t *) pudp;
950 goto out;
951 } else if (is_hugepd(__hugepd(pud_val(pud))))
952 hpdp = (hugepd_t *)&pud;
953 else {
954 pdshift = PMD_SHIFT;
955 pmdp = pmd_offset(&pud, ea);
956 pmd = READ_ONCE(*pmdp);
957 /*
958 * A hugepage collapse is captured by pmd_none, because
959 * it mark the pmd none and do a hpte invalidate.
960 */
961 if (pmd_none(pmd))
962 return NULL;
963
964 if (pmd_trans_huge(pmd)) {
965 if (is_thp)
966 *is_thp = true;
967 ret_pte = (pte_t *) pmdp;
968 goto out;
969 }
970
971 if (pmd_huge(pmd)) {
972 ret_pte = (pte_t *) pmdp;
973 goto out;
974 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
975 hpdp = (hugepd_t *)&pmd;
976 else
977 return pte_offset_kernel(&pmd, ea);
978 }
979 }
980 if (!hpdp)
981 return NULL;
982
983 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
984 pdshift = hugepd_shift(*hpdp);
985out:
986 if (shift)
987 *shift = pdshift;
988 return ret_pte;
989}
990EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
991
992int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
993 unsigned long end, int write, struct page **pages, int *nr)
994{
995 unsigned long mask;
996 unsigned long pte_end;
997 struct page *head, *page;
998 pte_t pte;
999 int refs;
1000
1001 pte_end = (addr + sz) & ~(sz-1);
1002 if (pte_end < end)
1003 end = pte_end;
1004
1005 pte = READ_ONCE(*ptep);
1006 mask = _PAGE_PRESENT | _PAGE_USER;
1007 if (write)
1008 mask |= _PAGE_RW;
1009
1010 if ((pte_val(pte) & mask) != mask)
1011 return 0;
1012
1013 /* hugepages are never "special" */
1014 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1015
1016 refs = 0;
1017 head = pte_page(pte);
1018
1019 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1020 do {
1021 VM_BUG_ON(compound_head(page) != head);
1022 pages[*nr] = page;
1023 (*nr)++;
1024 page++;
1025 refs++;
1026 } while (addr += PAGE_SIZE, addr != end);
1027
1028 if (!page_cache_add_speculative(head, refs)) {
1029 *nr -= refs;
1030 return 0;
1031 }
1032
1033 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1034 /* Could be optimized better */
1035 *nr -= refs;
1036 while (refs--)
1037 put_page(head);
1038 return 0;
1039 }
1040
1041 return 1;
1042}