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/moduleparam.h>
19#include <linux/swap.h>
20#include <linux/swapops.h>
21#include <linux/kmemleak.h>
22#include <asm/pgalloc.h>
23#include <asm/tlb.h>
24#include <asm/setup.h>
25#include <asm/hugetlb.h>
26#include <asm/pte-walk.h>
27
28bool hugetlb_disabled = false;
29
30#define hugepd_none(hpd) (hpd_val(hpd) == 0)
31
32#define PTE_T_ORDER (__builtin_ffs(sizeof(pte_basic_t)) - \
33 __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 {
57 cachep = PGT_CACHE(pdshift - pshift);
58 num_hugepd = 1;
59 }
60
61 if (!cachep) {
62 WARN_ONCE(1, "No page table cache created for hugetlb tables");
63 return -ENOMEM;
64 }
65
66 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
67
68 BUG_ON(pshift > HUGEPD_SHIFT_MASK);
69 BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
70
71 if (!new)
72 return -ENOMEM;
73
74 /*
75 * Make sure other cpus find the hugepd set only after a
76 * properly initialized page table is visible to them.
77 * For more details look for comment in __pte_alloc().
78 */
79 smp_wmb();
80
81 spin_lock(ptl);
82 /*
83 * We have multiple higher-level entries that point to the same
84 * actual pte location. Fill in each as we go and backtrack on error.
85 * We need all of these so the DTLB pgtable walk code can find the
86 * right higher-level entry without knowing if it's a hugepage or not.
87 */
88 for (i = 0; i < num_hugepd; i++, hpdp++) {
89 if (unlikely(!hugepd_none(*hpdp)))
90 break;
91 hugepd_populate(hpdp, new, pshift);
92 }
93 /* If we bailed from the for loop early, an error occurred, clean up */
94 if (i < num_hugepd) {
95 for (i = i - 1 ; i >= 0; i--, hpdp--)
96 *hpdp = __hugepd(0);
97 kmem_cache_free(cachep, new);
98 } else {
99 kmemleak_ignore(new);
100 }
101 spin_unlock(ptl);
102 return 0;
103}
104
105/*
106 * At this point we do the placement change only for BOOK3S 64. This would
107 * possibly work on other subarchs.
108 */
109pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
110{
111 pgd_t *pg;
112 p4d_t *p4;
113 pud_t *pu;
114 pmd_t *pm;
115 hugepd_t *hpdp = NULL;
116 unsigned pshift = __ffs(sz);
117 unsigned pdshift = PGDIR_SHIFT;
118 spinlock_t *ptl;
119
120 addr &= ~(sz-1);
121 pg = pgd_offset(mm, addr);
122 p4 = p4d_offset(pg, addr);
123
124#ifdef CONFIG_PPC_BOOK3S_64
125 if (pshift == PGDIR_SHIFT)
126 /* 16GB huge page */
127 return (pte_t *) p4;
128 else if (pshift > PUD_SHIFT) {
129 /*
130 * We need to use hugepd table
131 */
132 ptl = &mm->page_table_lock;
133 hpdp = (hugepd_t *)p4;
134 } else {
135 pdshift = PUD_SHIFT;
136 pu = pud_alloc(mm, p4, addr);
137 if (!pu)
138 return NULL;
139 if (pshift == PUD_SHIFT)
140 return (pte_t *)pu;
141 else if (pshift > PMD_SHIFT) {
142 ptl = pud_lockptr(mm, pu);
143 hpdp = (hugepd_t *)pu;
144 } else {
145 pdshift = PMD_SHIFT;
146 pm = pmd_alloc(mm, pu, addr);
147 if (!pm)
148 return NULL;
149 if (pshift == PMD_SHIFT)
150 /* 16MB hugepage */
151 return (pte_t *)pm;
152 else {
153 ptl = pmd_lockptr(mm, pm);
154 hpdp = (hugepd_t *)pm;
155 }
156 }
157 }
158#else
159 if (pshift >= PGDIR_SHIFT) {
160 ptl = &mm->page_table_lock;
161 hpdp = (hugepd_t *)p4;
162 } else {
163 pdshift = PUD_SHIFT;
164 pu = pud_alloc(mm, p4, addr);
165 if (!pu)
166 return NULL;
167 if (pshift >= PUD_SHIFT) {
168 ptl = pud_lockptr(mm, pu);
169 hpdp = (hugepd_t *)pu;
170 } else {
171 pdshift = PMD_SHIFT;
172 pm = pmd_alloc(mm, pu, addr);
173 if (!pm)
174 return NULL;
175 ptl = pmd_lockptr(mm, pm);
176 hpdp = (hugepd_t *)pm;
177 }
178 }
179#endif
180 if (!hpdp)
181 return NULL;
182
183 if (IS_ENABLED(CONFIG_PPC_8xx) && sz == SZ_512K)
184 return pte_alloc_map(mm, (pmd_t *)hpdp, addr);
185
186 BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
187
188 if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
189 pdshift, pshift, ptl))
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#ifndef CONFIG_PPC_BOOK3S_64
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[];
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(PGT_CACHE(PTE_T_ORDER), 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(PGT_CACHE(PTE_T_ORDER), 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(&(*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,
330 get_hugepd_cache_index(pdshift - shift));
331}
332
333static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, unsigned long addr)
334{
335 pgtable_t token = pmd_pgtable(*pmd);
336
337 pmd_clear(pmd);
338 pte_free_tlb(tlb, token, addr);
339 mm_dec_nr_ptes(tlb->mm);
340}
341
342static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
343 unsigned long addr, unsigned long end,
344 unsigned long floor, unsigned long ceiling)
345{
346 pmd_t *pmd;
347 unsigned long next;
348 unsigned long start;
349
350 start = addr;
351 do {
352 unsigned long more;
353
354 pmd = pmd_offset(pud, addr);
355 next = pmd_addr_end(addr, end);
356 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
357 if (pmd_none_or_clear_bad(pmd))
358 continue;
359
360 /*
361 * if it is not hugepd pointer, we should already find
362 * it cleared.
363 */
364 WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
365
366 hugetlb_free_pte_range(tlb, pmd, addr);
367
368 continue;
369 }
370 /*
371 * Increment next by the size of the huge mapping since
372 * there may be more than one entry at this level for a
373 * single hugepage, but all of them point to
374 * the same kmem cache that holds the hugepte.
375 */
376 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
377 if (more > next)
378 next = more;
379
380 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
381 addr, next, floor, ceiling);
382 } while (addr = next, addr != end);
383
384 start &= PUD_MASK;
385 if (start < floor)
386 return;
387 if (ceiling) {
388 ceiling &= PUD_MASK;
389 if (!ceiling)
390 return;
391 }
392 if (end - 1 > ceiling - 1)
393 return;
394
395 pmd = pmd_offset(pud, start);
396 pud_clear(pud);
397 pmd_free_tlb(tlb, pmd, start);
398 mm_dec_nr_pmds(tlb->mm);
399}
400
401static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
402 unsigned long addr, unsigned long end,
403 unsigned long floor, unsigned long ceiling)
404{
405 pud_t *pud;
406 unsigned long next;
407 unsigned long start;
408
409 start = addr;
410 do {
411 pud = pud_offset(p4d, addr);
412 next = pud_addr_end(addr, end);
413 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
414 if (pud_none_or_clear_bad(pud))
415 continue;
416 hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
417 ceiling);
418 } else {
419 unsigned long more;
420 /*
421 * Increment next by the size of the huge mapping since
422 * there may be more than one entry at this level for a
423 * single hugepage, but all of them point to
424 * the same kmem cache that holds the hugepte.
425 */
426 more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
427 if (more > next)
428 next = more;
429
430 free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
431 addr, next, floor, ceiling);
432 }
433 } while (addr = next, addr != end);
434
435 start &= PGDIR_MASK;
436 if (start < floor)
437 return;
438 if (ceiling) {
439 ceiling &= PGDIR_MASK;
440 if (!ceiling)
441 return;
442 }
443 if (end - 1 > ceiling - 1)
444 return;
445
446 pud = pud_offset(p4d, start);
447 p4d_clear(p4d);
448 pud_free_tlb(tlb, pud, start);
449 mm_dec_nr_puds(tlb->mm);
450}
451
452/*
453 * This function frees user-level page tables of a process.
454 */
455void hugetlb_free_pgd_range(struct mmu_gather *tlb,
456 unsigned long addr, unsigned long end,
457 unsigned long floor, unsigned long ceiling)
458{
459 pgd_t *pgd;
460 p4d_t *p4d;
461 unsigned long next;
462
463 /*
464 * Because there are a number of different possible pagetable
465 * layouts for hugepage ranges, we limit knowledge of how
466 * things should be laid out to the allocation path
467 * (huge_pte_alloc(), above). Everything else works out the
468 * structure as it goes from information in the hugepd
469 * pointers. That means that we can't here use the
470 * optimization used in the normal page free_pgd_range(), of
471 * checking whether we're actually covering a large enough
472 * range to have to do anything at the top level of the walk
473 * instead of at the bottom.
474 *
475 * To make sense of this, you should probably go read the big
476 * block comment at the top of the normal free_pgd_range(),
477 * too.
478 */
479
480 do {
481 next = pgd_addr_end(addr, end);
482 pgd = pgd_offset(tlb->mm, addr);
483 p4d = p4d_offset(pgd, addr);
484 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
485 if (p4d_none_or_clear_bad(p4d))
486 continue;
487 hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
488 } else {
489 unsigned long more;
490 /*
491 * Increment next by the size of the huge mapping since
492 * there may be more than one entry at the pgd level
493 * for a single hugepage, but all of them point to the
494 * same kmem cache that holds the hugepte.
495 */
496 more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
497 if (more > next)
498 next = more;
499
500 free_hugepd_range(tlb, (hugepd_t *)p4d, PGDIR_SHIFT,
501 addr, next, floor, ceiling);
502 }
503 } while (addr = next, addr != end);
504}
505
506struct page *follow_huge_pd(struct vm_area_struct *vma,
507 unsigned long address, hugepd_t hpd,
508 int flags, int pdshift)
509{
510 pte_t *ptep;
511 spinlock_t *ptl;
512 struct page *page = NULL;
513 unsigned long mask;
514 int shift = hugepd_shift(hpd);
515 struct mm_struct *mm = vma->vm_mm;
516
517retry:
518 /*
519 * hugepage directory entries are protected by mm->page_table_lock
520 * Use this instead of huge_pte_lockptr
521 */
522 ptl = &mm->page_table_lock;
523 spin_lock(ptl);
524
525 ptep = hugepte_offset(hpd, address, pdshift);
526 if (pte_present(*ptep)) {
527 mask = (1UL << shift) - 1;
528 page = pte_page(*ptep);
529 page += ((address & mask) >> PAGE_SHIFT);
530 if (flags & FOLL_GET)
531 get_page(page);
532 } else {
533 if (is_hugetlb_entry_migration(*ptep)) {
534 spin_unlock(ptl);
535 __migration_entry_wait(mm, ptep, ptl);
536 goto retry;
537 }
538 }
539 spin_unlock(ptl);
540 return page;
541}
542
543#ifdef CONFIG_PPC_MM_SLICES
544unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
545 unsigned long len, unsigned long pgoff,
546 unsigned long flags)
547{
548 struct hstate *hstate = hstate_file(file);
549 int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
550
551#ifdef CONFIG_PPC_RADIX_MMU
552 if (radix_enabled())
553 return radix__hugetlb_get_unmapped_area(file, addr, len,
554 pgoff, flags);
555#endif
556 return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
557}
558#endif
559
560unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
561{
562 /* With radix we don't use slice, so derive it from vma*/
563 if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
564 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
565
566 return 1UL << mmu_psize_to_shift(psize);
567 }
568 return vma_kernel_pagesize(vma);
569}
570
571bool __init arch_hugetlb_valid_size(unsigned long size)
572{
573 int shift = __ffs(size);
574 int mmu_psize;
575
576 /* Check that it is a page size supported by the hardware and
577 * that it fits within pagetable and slice limits. */
578 if (size <= PAGE_SIZE || !is_power_of_2(size))
579 return false;
580
581 mmu_psize = check_and_get_huge_psize(shift);
582 if (mmu_psize < 0)
583 return false;
584
585 BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
586
587 return true;
588}
589
590static int __init add_huge_page_size(unsigned long long size)
591{
592 int shift = __ffs(size);
593
594 if (!arch_hugetlb_valid_size((unsigned long)size))
595 return -EINVAL;
596
597 hugetlb_add_hstate(shift - PAGE_SHIFT);
598 return 0;
599}
600
601static int __init hugetlbpage_init(void)
602{
603 bool configured = false;
604 int psize;
605
606 if (hugetlb_disabled) {
607 pr_info("HugeTLB support is disabled!\n");
608 return 0;
609 }
610
611 if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
612 !mmu_has_feature(MMU_FTR_16M_PAGE))
613 return -ENODEV;
614
615 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
616 unsigned shift;
617 unsigned pdshift;
618
619 if (!mmu_psize_defs[psize].shift)
620 continue;
621
622 shift = mmu_psize_to_shift(psize);
623
624#ifdef CONFIG_PPC_BOOK3S_64
625 if (shift > PGDIR_SHIFT)
626 continue;
627 else if (shift > PUD_SHIFT)
628 pdshift = PGDIR_SHIFT;
629 else if (shift > PMD_SHIFT)
630 pdshift = PUD_SHIFT;
631 else
632 pdshift = PMD_SHIFT;
633#else
634 if (shift < PUD_SHIFT)
635 pdshift = PMD_SHIFT;
636 else if (shift < PGDIR_SHIFT)
637 pdshift = PUD_SHIFT;
638 else
639 pdshift = PGDIR_SHIFT;
640#endif
641
642 if (add_huge_page_size(1ULL << shift) < 0)
643 continue;
644 /*
645 * if we have pdshift and shift value same, we don't
646 * use pgt cache for hugepd.
647 */
648 if (pdshift > shift) {
649 if (!IS_ENABLED(CONFIG_PPC_8xx))
650 pgtable_cache_add(pdshift - shift);
651 } else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) ||
652 IS_ENABLED(CONFIG_PPC_8xx)) {
653 pgtable_cache_add(PTE_T_ORDER);
654 }
655
656 configured = true;
657 }
658
659 if (configured) {
660 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
661 hugetlbpage_init_default();
662 } else
663 pr_info("Failed to initialize. Disabling HugeTLB");
664
665 return 0;
666}
667
668arch_initcall(hugetlbpage_init);
669
670void flush_dcache_icache_hugepage(struct page *page)
671{
672 int i;
673 void *start;
674
675 BUG_ON(!PageCompound(page));
676
677 for (i = 0; i < compound_nr(page); i++) {
678 if (!PageHighMem(page)) {
679 __flush_dcache_icache(page_address(page+i));
680 } else {
681 start = kmap_atomic(page+i);
682 __flush_dcache_icache(start);
683 kunmap_atomic(start);
684 }
685 }
686}
687
688void __init gigantic_hugetlb_cma_reserve(void)
689{
690 unsigned long order = 0;
691
692 if (radix_enabled())
693 order = PUD_SHIFT - PAGE_SHIFT;
694 else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
695 /*
696 * For pseries we do use ibm,expected#pages for reserving 16G pages.
697 */
698 order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
699
700 if (order) {
701 VM_WARN_ON(order < MAX_ORDER);
702 hugetlb_cma_reserve(order);
703 }
704}