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 <linux/swap.h>
21#include <linux/swapops.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
29
30#ifdef CONFIG_HUGETLB_PAGE
31
32#define PAGE_SHIFT_64K 16
33#define PAGE_SHIFT_512K 19
34#define PAGE_SHIFT_8M 23
35#define PAGE_SHIFT_16M 24
36#define PAGE_SHIFT_16G 34
37
38unsigned int HPAGE_SHIFT;
39EXPORT_SYMBOL(HPAGE_SHIFT);
40
41#define hugepd_none(hpd) (hpd_val(hpd) == 0)
42
43pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz)
44{
45 /*
46 * Only called for hugetlbfs pages, hence can ignore THP and the
47 * irq disabled walk.
48 */
49 return __find_linux_pte(mm->pgd, addr, NULL, NULL);
50}
51
52static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
53 unsigned long address, unsigned pdshift, unsigned pshift)
54{
55 struct kmem_cache *cachep;
56 pte_t *new;
57 int i;
58 int num_hugepd;
59
60 if (pshift >= pdshift) {
61 cachep = hugepte_cache;
62 num_hugepd = 1 << (pshift - pdshift);
63 } else {
64 cachep = PGT_CACHE(pdshift - pshift);
65 num_hugepd = 1;
66 }
67
68 new = kmem_cache_zalloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
69
70 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
71 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
72
73 if (! new)
74 return -ENOMEM;
75
76 /*
77 * Make sure other cpus find the hugepd set only after a
78 * properly initialized page table is visible to them.
79 * For more details look for comment in __pte_alloc().
80 */
81 smp_wmb();
82
83 spin_lock(&mm->page_table_lock);
84
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 else {
95#ifdef CONFIG_PPC_BOOK3S_64
96 *hpdp = __hugepd(__pa(new) |
97 (shift_to_mmu_psize(pshift) << 2));
98#elif defined(CONFIG_PPC_8xx)
99 *hpdp = __hugepd(__pa(new) | _PMD_USER |
100 (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
101 _PMD_PAGE_512K) | _PMD_PRESENT);
102#else
103 /* We use the old format for PPC_FSL_BOOK3E */
104 *hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
105#endif
106 }
107 }
108 /* If we bailed from the for loop early, an error occurred, clean up */
109 if (i < num_hugepd) {
110 for (i = i - 1 ; i >= 0; i--, hpdp--)
111 *hpdp = __hugepd(0);
112 kmem_cache_free(cachep, new);
113 }
114 spin_unlock(&mm->page_table_lock);
115 return 0;
116}
117
118/*
119 * These macros define how to determine which level of the page table holds
120 * the hpdp.
121 */
122#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
123#define HUGEPD_PGD_SHIFT PGDIR_SHIFT
124#define HUGEPD_PUD_SHIFT PUD_SHIFT
125#endif
126
127/*
128 * At this point we do the placement change only for BOOK3S 64. This would
129 * possibly work on other subarchs.
130 */
131pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
132{
133 pgd_t *pg;
134 pud_t *pu;
135 pmd_t *pm;
136 hugepd_t *hpdp = NULL;
137 unsigned pshift = __ffs(sz);
138 unsigned pdshift = PGDIR_SHIFT;
139
140 addr &= ~(sz-1);
141 pg = pgd_offset(mm, addr);
142
143#ifdef CONFIG_PPC_BOOK3S_64
144 if (pshift == PGDIR_SHIFT)
145 /* 16GB huge page */
146 return (pte_t *) pg;
147 else if (pshift > PUD_SHIFT)
148 /*
149 * We need to use hugepd table
150 */
151 hpdp = (hugepd_t *)pg;
152 else {
153 pdshift = PUD_SHIFT;
154 pu = pud_alloc(mm, pg, addr);
155 if (pshift == PUD_SHIFT)
156 return (pte_t *)pu;
157 else if (pshift > PMD_SHIFT)
158 hpdp = (hugepd_t *)pu;
159 else {
160 pdshift = PMD_SHIFT;
161 pm = pmd_alloc(mm, pu, addr);
162 if (pshift == PMD_SHIFT)
163 /* 16MB hugepage */
164 return (pte_t *)pm;
165 else
166 hpdp = (hugepd_t *)pm;
167 }
168 }
169#else
170 if (pshift >= HUGEPD_PGD_SHIFT) {
171 hpdp = (hugepd_t *)pg;
172 } else {
173 pdshift = PUD_SHIFT;
174 pu = pud_alloc(mm, pg, addr);
175 if (pshift >= HUGEPD_PUD_SHIFT) {
176 hpdp = (hugepd_t *)pu;
177 } else {
178 pdshift = PMD_SHIFT;
179 pm = pmd_alloc(mm, pu, addr);
180 hpdp = (hugepd_t *)pm;
181 }
182 }
183#endif
184 if (!hpdp)
185 return NULL;
186
187 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
188
189 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
190 return NULL;
191
192 return hugepte_offset(*hpdp, addr, pdshift);
193}
194
195#ifdef CONFIG_PPC_BOOK3S_64
196/*
197 * Tracks gpages after the device tree is scanned and before the
198 * huge_boot_pages list is ready on pseries.
199 */
200#define MAX_NUMBER_GPAGES 1024
201__initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
202__initdata static unsigned nr_gpages;
203
204/*
205 * Build list of addresses of gigantic pages. This function is used in early
206 * boot before the buddy allocator is setup.
207 */
208void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
209{
210 if (!addr)
211 return;
212 while (number_of_pages > 0) {
213 gpage_freearray[nr_gpages] = addr;
214 nr_gpages++;
215 number_of_pages--;
216 addr += page_size;
217 }
218}
219
220int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
221{
222 struct huge_bootmem_page *m;
223 if (nr_gpages == 0)
224 return 0;
225 m = phys_to_virt(gpage_freearray[--nr_gpages]);
226 gpage_freearray[nr_gpages] = 0;
227 list_add(&m->list, &huge_boot_pages);
228 m->hstate = hstate;
229 return 1;
230}
231#endif
232
233
234int __init alloc_bootmem_huge_page(struct hstate *h)
235{
236
237#ifdef CONFIG_PPC_BOOK3S_64
238 if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
239 return pseries_alloc_bootmem_huge_page(h);
240#endif
241 return __alloc_bootmem_huge_page(h);
242}
243
244#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
245#define HUGEPD_FREELIST_SIZE \
246 ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
247
248struct hugepd_freelist {
249 struct rcu_head rcu;
250 unsigned int index;
251 void *ptes[0];
252};
253
254static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
255
256static void hugepd_free_rcu_callback(struct rcu_head *head)
257{
258 struct hugepd_freelist *batch =
259 container_of(head, struct hugepd_freelist, rcu);
260 unsigned int i;
261
262 for (i = 0; i < batch->index; i++)
263 kmem_cache_free(hugepte_cache, batch->ptes[i]);
264
265 free_page((unsigned long)batch);
266}
267
268static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
269{
270 struct hugepd_freelist **batchp;
271
272 batchp = &get_cpu_var(hugepd_freelist_cur);
273
274 if (atomic_read(&tlb->mm->mm_users) < 2 ||
275 mm_is_thread_local(tlb->mm)) {
276 kmem_cache_free(hugepte_cache, hugepte);
277 put_cpu_var(hugepd_freelist_cur);
278 return;
279 }
280
281 if (*batchp == NULL) {
282 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
283 (*batchp)->index = 0;
284 }
285
286 (*batchp)->ptes[(*batchp)->index++] = hugepte;
287 if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
288 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
289 *batchp = NULL;
290 }
291 put_cpu_var(hugepd_freelist_cur);
292}
293#else
294static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
295#endif
296
297static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
298 unsigned long start, unsigned long end,
299 unsigned long floor, unsigned long ceiling)
300{
301 pte_t *hugepte = hugepd_page(*hpdp);
302 int i;
303
304 unsigned long pdmask = ~((1UL << pdshift) - 1);
305 unsigned int num_hugepd = 1;
306 unsigned int shift = hugepd_shift(*hpdp);
307
308 /* Note: On fsl the hpdp may be the first of several */
309 if (shift > pdshift)
310 num_hugepd = 1 << (shift - pdshift);
311
312 start &= pdmask;
313 if (start < floor)
314 return;
315 if (ceiling) {
316 ceiling &= pdmask;
317 if (! ceiling)
318 return;
319 }
320 if (end - 1 > ceiling - 1)
321 return;
322
323 for (i = 0; i < num_hugepd; i++, hpdp++)
324 *hpdp = __hugepd(0);
325
326 if (shift >= pdshift)
327 hugepd_free(tlb, hugepte);
328 else
329 pgtable_free_tlb(tlb, hugepte, pdshift - shift);
330}
331
332static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
333 unsigned long addr, unsigned long end,
334 unsigned long floor, unsigned long ceiling)
335{
336 pmd_t *pmd;
337 unsigned long next;
338 unsigned long start;
339
340 start = addr;
341 do {
342 unsigned long more;
343
344 pmd = pmd_offset(pud, addr);
345 next = pmd_addr_end(addr, end);
346 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
347 /*
348 * if it is not hugepd pointer, we should already find
349 * it cleared.
350 */
351 WARN_ON(!pmd_none_or_clear_bad(pmd));
352 continue;
353 }
354 /*
355 * Increment next by the size of the huge mapping since
356 * there may be more than one entry at this level for a
357 * single hugepage, but all of them point to
358 * the same kmem cache that holds the hugepte.
359 */
360 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
361 if (more > next)
362 next = more;
363
364 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
365 addr, next, floor, ceiling);
366 } while (addr = next, addr != end);
367
368 start &= PUD_MASK;
369 if (start < floor)
370 return;
371 if (ceiling) {
372 ceiling &= PUD_MASK;
373 if (!ceiling)
374 return;
375 }
376 if (end - 1 > ceiling - 1)
377 return;
378
379 pmd = pmd_offset(pud, start);
380 pud_clear(pud);
381 pmd_free_tlb(tlb, pmd, start);
382 mm_dec_nr_pmds(tlb->mm);
383}
384
385static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
386 unsigned long addr, unsigned long end,
387 unsigned long floor, unsigned long ceiling)
388{
389 pud_t *pud;
390 unsigned long next;
391 unsigned long start;
392
393 start = addr;
394 do {
395 pud = pud_offset(pgd, addr);
396 next = pud_addr_end(addr, end);
397 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
398 if (pud_none_or_clear_bad(pud))
399 continue;
400 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
401 ceiling);
402 } else {
403 unsigned long more;
404 /*
405 * Increment next by the size of the huge mapping since
406 * there may be more than one entry at this level for a
407 * single hugepage, but all of them point to
408 * the same kmem cache that holds the hugepte.
409 */
410 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
411 if (more > next)
412 next = more;
413
414 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
415 addr, next, floor, ceiling);
416 }
417 } while (addr = next, addr != end);
418
419 start &= PGDIR_MASK;
420 if (start < floor)
421 return;
422 if (ceiling) {
423 ceiling &= PGDIR_MASK;
424 if (!ceiling)
425 return;
426 }
427 if (end - 1 > ceiling - 1)
428 return;
429
430 pud = pud_offset(pgd, start);
431 pgd_clear(pgd);
432 pud_free_tlb(tlb, pud, start);
433 mm_dec_nr_puds(tlb->mm);
434}
435
436/*
437 * This function frees user-level page tables of a process.
438 */
439void hugetlb_free_pgd_range(struct mmu_gather *tlb,
440 unsigned long addr, unsigned long end,
441 unsigned long floor, unsigned long ceiling)
442{
443 pgd_t *pgd;
444 unsigned long next;
445
446 /*
447 * Because there are a number of different possible pagetable
448 * layouts for hugepage ranges, we limit knowledge of how
449 * things should be laid out to the allocation path
450 * (huge_pte_alloc(), above). Everything else works out the
451 * structure as it goes from information in the hugepd
452 * pointers. That means that we can't here use the
453 * optimization used in the normal page free_pgd_range(), of
454 * checking whether we're actually covering a large enough
455 * range to have to do anything at the top level of the walk
456 * instead of at the bottom.
457 *
458 * To make sense of this, you should probably go read the big
459 * block comment at the top of the normal free_pgd_range(),
460 * too.
461 */
462
463 do {
464 next = pgd_addr_end(addr, end);
465 pgd = pgd_offset(tlb->mm, addr);
466 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
467 if (pgd_none_or_clear_bad(pgd))
468 continue;
469 hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
470 } else {
471 unsigned long more;
472 /*
473 * Increment next by the size of the huge mapping since
474 * there may be more than one entry at the pgd level
475 * for a single hugepage, but all of them point to the
476 * same kmem cache that holds the hugepte.
477 */
478 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
479 if (more > next)
480 next = more;
481
482 free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
483 addr, next, floor, ceiling);
484 }
485 } while (addr = next, addr != end);
486}
487
488struct page *follow_huge_pd(struct vm_area_struct *vma,
489 unsigned long address, hugepd_t hpd,
490 int flags, int pdshift)
491{
492 pte_t *ptep;
493 spinlock_t *ptl;
494 struct page *page = NULL;
495 unsigned long mask;
496 int shift = hugepd_shift(hpd);
497 struct mm_struct *mm = vma->vm_mm;
498
499retry:
500 ptl = &mm->page_table_lock;
501 spin_lock(ptl);
502
503 ptep = hugepte_offset(hpd, address, pdshift);
504 if (pte_present(*ptep)) {
505 mask = (1UL << shift) - 1;
506 page = pte_page(*ptep);
507 page += ((address & mask) >> PAGE_SHIFT);
508 if (flags & FOLL_GET)
509 get_page(page);
510 } else {
511 if (is_hugetlb_entry_migration(*ptep)) {
512 spin_unlock(ptl);
513 __migration_entry_wait(mm, ptep, ptl);
514 goto retry;
515 }
516 }
517 spin_unlock(ptl);
518 return page;
519}
520
521static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
522 unsigned long sz)
523{
524 unsigned long __boundary = (addr + sz) & ~(sz-1);
525 return (__boundary - 1 < end - 1) ? __boundary : end;
526}
527
528int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
529 unsigned long end, int write, struct page **pages, int *nr)
530{
531 pte_t *ptep;
532 unsigned long sz = 1UL << hugepd_shift(hugepd);
533 unsigned long next;
534
535 ptep = hugepte_offset(hugepd, addr, pdshift);
536 do {
537 next = hugepte_addr_end(addr, end, sz);
538 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
539 return 0;
540 } while (ptep++, addr = next, addr != end);
541
542 return 1;
543}
544
545#ifdef CONFIG_PPC_MM_SLICES
546unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
547 unsigned long len, unsigned long pgoff,
548 unsigned long flags)
549{
550 struct hstate *hstate = hstate_file(file);
551 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
552
553#ifdef CONFIG_PPC_RADIX_MMU
554 if (radix_enabled())
555 return radix__hugetlb_get_unmapped_area(file, addr, len,
556 pgoff, flags);
557#endif
558 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
559}
560#endif
561
562unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
563{
564#ifdef CONFIG_PPC_MM_SLICES
565 /* With radix we don't use slice, so derive it from vma*/
566 if (!radix_enabled()) {
567 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
568
569 return 1UL << mmu_psize_to_shift(psize);
570 }
571#endif
572 return vma_kernel_pagesize(vma);
573}
574
575static inline bool is_power_of_4(unsigned long x)
576{
577 if (is_power_of_2(x))
578 return (__ilog2(x) % 2) ? false : true;
579 return false;
580}
581
582static int __init add_huge_page_size(unsigned long long size)
583{
584 int shift = __ffs(size);
585 int mmu_psize;
586
587 /* Check that it is a page size supported by the hardware and
588 * that it fits within pagetable and slice limits. */
589 if (size <= PAGE_SIZE)
590 return -EINVAL;
591#if defined(CONFIG_PPC_FSL_BOOK3E)
592 if (!is_power_of_4(size))
593 return -EINVAL;
594#elif !defined(CONFIG_PPC_8xx)
595 if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
596 return -EINVAL;
597#endif
598
599 if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
600 return -EINVAL;
601
602#ifdef CONFIG_PPC_BOOK3S_64
603 /*
604 * We need to make sure that for different page sizes reported by
605 * firmware we only add hugetlb support for page sizes that can be
606 * supported by linux page table layout.
607 * For now we have
608 * Radix: 2M
609 * Hash: 16M and 16G
610 */
611 if (radix_enabled()) {
612 if (mmu_psize != MMU_PAGE_2M) {
613 if (cpu_has_feature(CPU_FTR_POWER9_DD1) ||
614 (mmu_psize != MMU_PAGE_1G))
615 return -EINVAL;
616 }
617 } else {
618 if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
619 return -EINVAL;
620 }
621#endif
622
623 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
624
625 /* Return if huge page size has already been setup */
626 if (size_to_hstate(size))
627 return 0;
628
629 hugetlb_add_hstate(shift - PAGE_SHIFT);
630
631 return 0;
632}
633
634static int __init hugepage_setup_sz(char *str)
635{
636 unsigned long long size;
637
638 size = memparse(str, &str);
639
640 if (add_huge_page_size(size) != 0) {
641 hugetlb_bad_size();
642 pr_err("Invalid huge page size specified(%llu)\n", size);
643 }
644
645 return 1;
646}
647__setup("hugepagesz=", hugepage_setup_sz);
648
649struct kmem_cache *hugepte_cache;
650static int __init hugetlbpage_init(void)
651{
652 int psize;
653
654#if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
655 if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
656 return -ENODEV;
657#endif
658 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
659 unsigned shift;
660 unsigned pdshift;
661
662 if (!mmu_psize_defs[psize].shift)
663 continue;
664
665 shift = mmu_psize_to_shift(psize);
666
667#ifdef CONFIG_PPC_BOOK3S_64
668 if (shift > PGDIR_SHIFT)
669 continue;
670 else if (shift > PUD_SHIFT)
671 pdshift = PGDIR_SHIFT;
672 else if (shift > PMD_SHIFT)
673 pdshift = PUD_SHIFT;
674 else
675 pdshift = PMD_SHIFT;
676#else
677 if (shift < HUGEPD_PUD_SHIFT)
678 pdshift = PMD_SHIFT;
679 else if (shift < HUGEPD_PGD_SHIFT)
680 pdshift = PUD_SHIFT;
681 else
682 pdshift = PGDIR_SHIFT;
683#endif
684
685 if (add_huge_page_size(1ULL << shift) < 0)
686 continue;
687 /*
688 * if we have pdshift and shift value same, we don't
689 * use pgt cache for hugepd.
690 */
691 if (pdshift > shift)
692 pgtable_cache_add(pdshift - shift, NULL);
693#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
694 else if (!hugepte_cache) {
695 /*
696 * Create a kmem cache for hugeptes. The bottom bits in
697 * the pte have size information encoded in them, so
698 * align them to allow this
699 */
700 hugepte_cache = kmem_cache_create("hugepte-cache",
701 sizeof(pte_t),
702 HUGEPD_SHIFT_MASK + 1,
703 0, NULL);
704 if (hugepte_cache == NULL)
705 panic("%s: Unable to create kmem cache "
706 "for hugeptes\n", __func__);
707
708 }
709#endif
710 }
711
712#if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
713 /* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
714 if (mmu_psize_defs[MMU_PAGE_4M].shift)
715 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
716 else if (mmu_psize_defs[MMU_PAGE_512K].shift)
717 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
718#else
719 /* Set default large page size. Currently, we pick 16M or 1M
720 * depending on what is available
721 */
722 if (mmu_psize_defs[MMU_PAGE_16M].shift)
723 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
724 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
725 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
726 else if (mmu_psize_defs[MMU_PAGE_2M].shift)
727 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
728#endif
729 return 0;
730}
731
732arch_initcall(hugetlbpage_init);
733
734void flush_dcache_icache_hugepage(struct page *page)
735{
736 int i;
737 void *start;
738
739 BUG_ON(!PageCompound(page));
740
741 for (i = 0; i < (1UL << compound_order(page)); i++) {
742 if (!PageHighMem(page)) {
743 __flush_dcache_icache(page_address(page+i));
744 } else {
745 start = kmap_atomic(page+i);
746 __flush_dcache_icache(start);
747 kunmap_atomic(start);
748 }
749 }
750}
751
752#endif /* CONFIG_HUGETLB_PAGE */
753
754/*
755 * We have 4 cases for pgds and pmds:
756 * (1) invalid (all zeroes)
757 * (2) pointer to next table, as normal; bottom 6 bits == 0
758 * (3) leaf pte for huge page _PAGE_PTE set
759 * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
760 *
761 * So long as we atomically load page table pointers we are safe against teardown,
762 * we can follow the address down to the the page and take a ref on it.
763 * This function need to be called with interrupts disabled. We use this variant
764 * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
765 */
766pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
767 bool *is_thp, unsigned *hpage_shift)
768{
769 pgd_t pgd, *pgdp;
770 pud_t pud, *pudp;
771 pmd_t pmd, *pmdp;
772 pte_t *ret_pte;
773 hugepd_t *hpdp = NULL;
774 unsigned pdshift = PGDIR_SHIFT;
775
776 if (hpage_shift)
777 *hpage_shift = 0;
778
779 if (is_thp)
780 *is_thp = false;
781
782 pgdp = pgdir + pgd_index(ea);
783 pgd = READ_ONCE(*pgdp);
784 /*
785 * Always operate on the local stack value. This make sure the
786 * value don't get updated by a parallel THP split/collapse,
787 * page fault or a page unmap. The return pte_t * is still not
788 * stable. So should be checked there for above conditions.
789 */
790 if (pgd_none(pgd))
791 return NULL;
792 else if (pgd_huge(pgd)) {
793 ret_pte = (pte_t *) pgdp;
794 goto out;
795 } else if (is_hugepd(__hugepd(pgd_val(pgd))))
796 hpdp = (hugepd_t *)&pgd;
797 else {
798 /*
799 * Even if we end up with an unmap, the pgtable will not
800 * be freed, because we do an rcu free and here we are
801 * irq disabled
802 */
803 pdshift = PUD_SHIFT;
804 pudp = pud_offset(&pgd, ea);
805 pud = READ_ONCE(*pudp);
806
807 if (pud_none(pud))
808 return NULL;
809 else if (pud_huge(pud)) {
810 ret_pte = (pte_t *) pudp;
811 goto out;
812 } else if (is_hugepd(__hugepd(pud_val(pud))))
813 hpdp = (hugepd_t *)&pud;
814 else {
815 pdshift = PMD_SHIFT;
816 pmdp = pmd_offset(&pud, ea);
817 pmd = READ_ONCE(*pmdp);
818 /*
819 * A hugepage collapse is captured by pmd_none, because
820 * it mark the pmd none and do a hpte invalidate.
821 */
822 if (pmd_none(pmd))
823 return NULL;
824
825 if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
826 if (is_thp)
827 *is_thp = true;
828 ret_pte = (pte_t *) pmdp;
829 goto out;
830 }
831
832 if (pmd_huge(pmd)) {
833 ret_pte = (pte_t *) pmdp;
834 goto out;
835 } else if (is_hugepd(__hugepd(pmd_val(pmd))))
836 hpdp = (hugepd_t *)&pmd;
837 else
838 return pte_offset_kernel(&pmd, ea);
839 }
840 }
841 if (!hpdp)
842 return NULL;
843
844 ret_pte = hugepte_offset(*hpdp, ea, pdshift);
845 pdshift = hugepd_shift(*hpdp);
846out:
847 if (hpage_shift)
848 *hpage_shift = pdshift;
849 return ret_pte;
850}
851EXPORT_SYMBOL_GPL(__find_linux_pte);
852
853int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
854 unsigned long end, int write, struct page **pages, int *nr)
855{
856 unsigned long pte_end;
857 struct page *head, *page;
858 pte_t pte;
859 int refs;
860
861 pte_end = (addr + sz) & ~(sz-1);
862 if (pte_end < end)
863 end = pte_end;
864
865 pte = READ_ONCE(*ptep);
866
867 if (!pte_access_permitted(pte, write))
868 return 0;
869
870 /* hugepages are never "special" */
871 VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
872
873 refs = 0;
874 head = pte_page(pte);
875
876 page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
877 do {
878 VM_BUG_ON(compound_head(page) != head);
879 pages[*nr] = page;
880 (*nr)++;
881 page++;
882 refs++;
883 } while (addr += PAGE_SIZE, addr != end);
884
885 if (!page_cache_add_speculative(head, refs)) {
886 *nr -= refs;
887 return 0;
888 }
889
890 if (unlikely(pte_val(pte) != pte_val(*ptep))) {
891 /* Could be optimized better */
892 *nr -= refs;
893 while (refs--)
894 put_page(head);
895 return 0;
896 }
897
898 return 1;
899}