hugetlbfs: add arch_hugetlb_valid_size
[linux-2.6-microblaze.git] / arch / powerpc / mm / hugetlbpage.c
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
29 bool 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
35 pte_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
44 static 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                 new = NULL;
57         } else if (IS_ENABLED(CONFIG_PPC_8xx)) {
58                 cachep = NULL;
59                 num_hugepd = 1;
60                 new = pte_alloc_one(mm);
61         } else {
62                 cachep = PGT_CACHE(pdshift - pshift);
63                 num_hugepd = 1;
64                 new = NULL;
65         }
66
67         if (!cachep && !new) {
68                 WARN_ONCE(1, "No page table cache created for hugetlb tables");
69                 return -ENOMEM;
70         }
71
72         if (cachep)
73                 new = kmem_cache_alloc(cachep, pgtable_gfp_flags(mm, GFP_KERNEL));
74
75         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
76         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
77
78         if (!new)
79                 return -ENOMEM;
80
81         /*
82          * Make sure other cpus find the hugepd set only after a
83          * properly initialized page table is visible to them.
84          * For more details look for comment in __pte_alloc().
85          */
86         smp_wmb();
87
88         spin_lock(ptl);
89         /*
90          * We have multiple higher-level entries that point to the same
91          * actual pte location.  Fill in each as we go and backtrack on error.
92          * We need all of these so the DTLB pgtable walk code can find the
93          * right higher-level entry without knowing if it's a hugepage or not.
94          */
95         for (i = 0; i < num_hugepd; i++, hpdp++) {
96                 if (unlikely(!hugepd_none(*hpdp)))
97                         break;
98                 hugepd_populate(hpdp, new, pshift);
99         }
100         /* If we bailed from the for loop early, an error occurred, clean up */
101         if (i < num_hugepd) {
102                 for (i = i - 1 ; i >= 0; i--, hpdp--)
103                         *hpdp = __hugepd(0);
104                 if (cachep)
105                         kmem_cache_free(cachep, new);
106                 else
107                         pte_free(mm, new);
108         } else {
109                 kmemleak_ignore(new);
110         }
111         spin_unlock(ptl);
112         return 0;
113 }
114
115 /*
116  * At this point we do the placement change only for BOOK3S 64. This would
117  * possibly work on other subarchs.
118  */
119 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
120 {
121         pgd_t *pg;
122         pud_t *pu;
123         pmd_t *pm;
124         hugepd_t *hpdp = NULL;
125         unsigned pshift = __ffs(sz);
126         unsigned pdshift = PGDIR_SHIFT;
127         spinlock_t *ptl;
128
129         addr &= ~(sz-1);
130         pg = pgd_offset(mm, addr);
131
132 #ifdef CONFIG_PPC_BOOK3S_64
133         if (pshift == PGDIR_SHIFT)
134                 /* 16GB huge page */
135                 return (pte_t *) pg;
136         else if (pshift > PUD_SHIFT) {
137                 /*
138                  * We need to use hugepd table
139                  */
140                 ptl = &mm->page_table_lock;
141                 hpdp = (hugepd_t *)pg;
142         } else {
143                 pdshift = PUD_SHIFT;
144                 pu = pud_alloc(mm, pg, addr);
145                 if (!pu)
146                         return NULL;
147                 if (pshift == PUD_SHIFT)
148                         return (pte_t *)pu;
149                 else if (pshift > PMD_SHIFT) {
150                         ptl = pud_lockptr(mm, pu);
151                         hpdp = (hugepd_t *)pu;
152                 } else {
153                         pdshift = PMD_SHIFT;
154                         pm = pmd_alloc(mm, pu, addr);
155                         if (!pm)
156                                 return NULL;
157                         if (pshift == PMD_SHIFT)
158                                 /* 16MB hugepage */
159                                 return (pte_t *)pm;
160                         else {
161                                 ptl = pmd_lockptr(mm, pm);
162                                 hpdp = (hugepd_t *)pm;
163                         }
164                 }
165         }
166 #else
167         if (pshift >= PGDIR_SHIFT) {
168                 ptl = &mm->page_table_lock;
169                 hpdp = (hugepd_t *)pg;
170         } else {
171                 pdshift = PUD_SHIFT;
172                 pu = pud_alloc(mm, pg, addr);
173                 if (!pu)
174                         return NULL;
175                 if (pshift >= PUD_SHIFT) {
176                         ptl = pud_lockptr(mm, pu);
177                         hpdp = (hugepd_t *)pu;
178                 } else {
179                         pdshift = PMD_SHIFT;
180                         pm = pmd_alloc(mm, pu, addr);
181                         if (!pm)
182                                 return NULL;
183                         ptl = pmd_lockptr(mm, pm);
184                         hpdp = (hugepd_t *)pm;
185                 }
186         }
187 #endif
188         if (!hpdp)
189                 return NULL;
190
191         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
192
193         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
194                                                   pdshift, pshift, ptl))
195                 return NULL;
196
197         return hugepte_offset(*hpdp, addr, pdshift);
198 }
199
200 #ifdef CONFIG_PPC_BOOK3S_64
201 /*
202  * Tracks gpages after the device tree is scanned and before the
203  * huge_boot_pages list is ready on pseries.
204  */
205 #define MAX_NUMBER_GPAGES       1024
206 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
207 __initdata static unsigned nr_gpages;
208
209 /*
210  * Build list of addresses of gigantic pages.  This function is used in early
211  * boot before the buddy allocator is setup.
212  */
213 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
214 {
215         if (!addr)
216                 return;
217         while (number_of_pages > 0) {
218                 gpage_freearray[nr_gpages] = addr;
219                 nr_gpages++;
220                 number_of_pages--;
221                 addr += page_size;
222         }
223 }
224
225 int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
226 {
227         struct huge_bootmem_page *m;
228         if (nr_gpages == 0)
229                 return 0;
230         m = phys_to_virt(gpage_freearray[--nr_gpages]);
231         gpage_freearray[nr_gpages] = 0;
232         list_add(&m->list, &huge_boot_pages);
233         m->hstate = hstate;
234         return 1;
235 }
236 #endif
237
238
239 int __init alloc_bootmem_huge_page(struct hstate *h)
240 {
241
242 #ifdef CONFIG_PPC_BOOK3S_64
243         if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
244                 return pseries_alloc_bootmem_huge_page(h);
245 #endif
246         return __alloc_bootmem_huge_page(h);
247 }
248
249 #ifndef CONFIG_PPC_BOOK3S_64
250 #define HUGEPD_FREELIST_SIZE \
251         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
252
253 struct hugepd_freelist {
254         struct rcu_head rcu;
255         unsigned int index;
256         void *ptes[0];
257 };
258
259 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
260
261 static void hugepd_free_rcu_callback(struct rcu_head *head)
262 {
263         struct hugepd_freelist *batch =
264                 container_of(head, struct hugepd_freelist, rcu);
265         unsigned int i;
266
267         for (i = 0; i < batch->index; i++)
268                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
269
270         free_page((unsigned long)batch);
271 }
272
273 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
274 {
275         struct hugepd_freelist **batchp;
276
277         batchp = &get_cpu_var(hugepd_freelist_cur);
278
279         if (atomic_read(&tlb->mm->mm_users) < 2 ||
280             mm_is_thread_local(tlb->mm)) {
281                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
282                 put_cpu_var(hugepd_freelist_cur);
283                 return;
284         }
285
286         if (*batchp == NULL) {
287                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
288                 (*batchp)->index = 0;
289         }
290
291         (*batchp)->ptes[(*batchp)->index++] = hugepte;
292         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
293                 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
294                 *batchp = NULL;
295         }
296         put_cpu_var(hugepd_freelist_cur);
297 }
298 #else
299 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
300 #endif
301
302 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
303                               unsigned long start, unsigned long end,
304                               unsigned long floor, unsigned long ceiling)
305 {
306         pte_t *hugepte = hugepd_page(*hpdp);
307         int i;
308
309         unsigned long pdmask = ~((1UL << pdshift) - 1);
310         unsigned int num_hugepd = 1;
311         unsigned int shift = hugepd_shift(*hpdp);
312
313         /* Note: On fsl the hpdp may be the first of several */
314         if (shift > pdshift)
315                 num_hugepd = 1 << (shift - pdshift);
316
317         start &= pdmask;
318         if (start < floor)
319                 return;
320         if (ceiling) {
321                 ceiling &= pdmask;
322                 if (! ceiling)
323                         return;
324         }
325         if (end - 1 > ceiling - 1)
326                 return;
327
328         for (i = 0; i < num_hugepd; i++, hpdp++)
329                 *hpdp = __hugepd(0);
330
331         if (shift >= pdshift)
332                 hugepd_free(tlb, hugepte);
333         else if (IS_ENABLED(CONFIG_PPC_8xx))
334                 pgtable_free_tlb(tlb, hugepte, 0);
335         else
336                 pgtable_free_tlb(tlb, hugepte,
337                                  get_hugepd_cache_index(pdshift - shift));
338 }
339
340 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
341                                    unsigned long addr, unsigned long end,
342                                    unsigned long floor, unsigned long ceiling)
343 {
344         pmd_t *pmd;
345         unsigned long next;
346         unsigned long start;
347
348         start = addr;
349         do {
350                 unsigned long more;
351
352                 pmd = pmd_offset(pud, addr);
353                 next = pmd_addr_end(addr, end);
354                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
355                         /*
356                          * if it is not hugepd pointer, we should already find
357                          * it cleared.
358                          */
359                         WARN_ON(!pmd_none_or_clear_bad(pmd));
360                         continue;
361                 }
362                 /*
363                  * Increment next by the size of the huge mapping since
364                  * there may be more than one entry at this level for a
365                  * single hugepage, but all of them point to
366                  * the same kmem cache that holds the hugepte.
367                  */
368                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
369                 if (more > next)
370                         next = more;
371
372                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
373                                   addr, next, floor, ceiling);
374         } while (addr = next, addr != end);
375
376         start &= PUD_MASK;
377         if (start < floor)
378                 return;
379         if (ceiling) {
380                 ceiling &= PUD_MASK;
381                 if (!ceiling)
382                         return;
383         }
384         if (end - 1 > ceiling - 1)
385                 return;
386
387         pmd = pmd_offset(pud, start);
388         pud_clear(pud);
389         pmd_free_tlb(tlb, pmd, start);
390         mm_dec_nr_pmds(tlb->mm);
391 }
392
393 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
394                                    unsigned long addr, unsigned long end,
395                                    unsigned long floor, unsigned long ceiling)
396 {
397         pud_t *pud;
398         unsigned long next;
399         unsigned long start;
400
401         start = addr;
402         do {
403                 pud = pud_offset(pgd, addr);
404                 next = pud_addr_end(addr, end);
405                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
406                         if (pud_none_or_clear_bad(pud))
407                                 continue;
408                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
409                                                ceiling);
410                 } else {
411                         unsigned long more;
412                         /*
413                          * Increment next by the size of the huge mapping since
414                          * there may be more than one entry at this level for a
415                          * single hugepage, but all of them point to
416                          * the same kmem cache that holds the hugepte.
417                          */
418                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
419                         if (more > next)
420                                 next = more;
421
422                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
423                                           addr, next, floor, ceiling);
424                 }
425         } while (addr = next, addr != end);
426
427         start &= PGDIR_MASK;
428         if (start < floor)
429                 return;
430         if (ceiling) {
431                 ceiling &= PGDIR_MASK;
432                 if (!ceiling)
433                         return;
434         }
435         if (end - 1 > ceiling - 1)
436                 return;
437
438         pud = pud_offset(pgd, start);
439         pgd_clear(pgd);
440         pud_free_tlb(tlb, pud, start);
441         mm_dec_nr_puds(tlb->mm);
442 }
443
444 /*
445  * This function frees user-level page tables of a process.
446  */
447 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
448                             unsigned long addr, unsigned long end,
449                             unsigned long floor, unsigned long ceiling)
450 {
451         pgd_t *pgd;
452         unsigned long next;
453
454         /*
455          * Because there are a number of different possible pagetable
456          * layouts for hugepage ranges, we limit knowledge of how
457          * things should be laid out to the allocation path
458          * (huge_pte_alloc(), above).  Everything else works out the
459          * structure as it goes from information in the hugepd
460          * pointers.  That means that we can't here use the
461          * optimization used in the normal page free_pgd_range(), of
462          * checking whether we're actually covering a large enough
463          * range to have to do anything at the top level of the walk
464          * instead of at the bottom.
465          *
466          * To make sense of this, you should probably go read the big
467          * block comment at the top of the normal free_pgd_range(),
468          * too.
469          */
470
471         do {
472                 next = pgd_addr_end(addr, end);
473                 pgd = pgd_offset(tlb->mm, addr);
474                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
475                         if (pgd_none_or_clear_bad(pgd))
476                                 continue;
477                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
478                 } else {
479                         unsigned long more;
480                         /*
481                          * Increment next by the size of the huge mapping since
482                          * there may be more than one entry at the pgd level
483                          * for a single hugepage, but all of them point to the
484                          * same kmem cache that holds the hugepte.
485                          */
486                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
487                         if (more > next)
488                                 next = more;
489
490                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
491                                           addr, next, floor, ceiling);
492                 }
493         } while (addr = next, addr != end);
494 }
495
496 struct page *follow_huge_pd(struct vm_area_struct *vma,
497                             unsigned long address, hugepd_t hpd,
498                             int flags, int pdshift)
499 {
500         pte_t *ptep;
501         spinlock_t *ptl;
502         struct page *page = NULL;
503         unsigned long mask;
504         int shift = hugepd_shift(hpd);
505         struct mm_struct *mm = vma->vm_mm;
506
507 retry:
508         /*
509          * hugepage directory entries are protected by mm->page_table_lock
510          * Use this instead of huge_pte_lockptr
511          */
512         ptl = &mm->page_table_lock;
513         spin_lock(ptl);
514
515         ptep = hugepte_offset(hpd, address, pdshift);
516         if (pte_present(*ptep)) {
517                 mask = (1UL << shift) - 1;
518                 page = pte_page(*ptep);
519                 page += ((address & mask) >> PAGE_SHIFT);
520                 if (flags & FOLL_GET)
521                         get_page(page);
522         } else {
523                 if (is_hugetlb_entry_migration(*ptep)) {
524                         spin_unlock(ptl);
525                         __migration_entry_wait(mm, ptep, ptl);
526                         goto retry;
527                 }
528         }
529         spin_unlock(ptl);
530         return page;
531 }
532
533 #ifdef CONFIG_PPC_MM_SLICES
534 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
535                                         unsigned long len, unsigned long pgoff,
536                                         unsigned long flags)
537 {
538         struct hstate *hstate = hstate_file(file);
539         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
540
541 #ifdef CONFIG_PPC_RADIX_MMU
542         if (radix_enabled())
543                 return radix__hugetlb_get_unmapped_area(file, addr, len,
544                                                        pgoff, flags);
545 #endif
546         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
547 }
548 #endif
549
550 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
551 {
552         /* With radix we don't use slice, so derive it from vma*/
553         if (IS_ENABLED(CONFIG_PPC_MM_SLICES) && !radix_enabled()) {
554                 unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
555
556                 return 1UL << mmu_psize_to_shift(psize);
557         }
558         return vma_kernel_pagesize(vma);
559 }
560
561 bool __init arch_hugetlb_valid_size(unsigned long size)
562 {
563         int shift = __ffs(size);
564         int mmu_psize;
565
566         /* Check that it is a page size supported by the hardware and
567          * that it fits within pagetable and slice limits. */
568         if (size <= PAGE_SIZE || !is_power_of_2(size))
569                 return false;
570
571         mmu_psize = check_and_get_huge_psize(shift);
572         if (mmu_psize < 0)
573                 return false;
574
575         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
576
577         return true;
578 }
579
580 static int __init add_huge_page_size(unsigned long long size)
581 {
582         int shift = __ffs(size);
583
584         if (!arch_hugetlb_valid_size((unsigned long)size))
585                 return -EINVAL;
586
587         if (!size_to_hstate(size))
588                 hugetlb_add_hstate(shift - PAGE_SHIFT);
589         return 0;
590 }
591
592 static int __init hugepage_setup_sz(char *str)
593 {
594         unsigned long long size;
595
596         size = memparse(str, &str);
597
598         if (add_huge_page_size(size) != 0) {
599                 hugetlb_bad_size();
600                 pr_err("Invalid huge page size specified(%llu)\n", size);
601         }
602
603         return 1;
604 }
605 __setup("hugepagesz=", hugepage_setup_sz);
606
607 static int __init hugetlbpage_init(void)
608 {
609         bool configured = false;
610         int psize;
611
612         if (hugetlb_disabled) {
613                 pr_info("HugeTLB support is disabled!\n");
614                 return 0;
615         }
616
617         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
618             !mmu_has_feature(MMU_FTR_16M_PAGE))
619                 return -ENODEV;
620
621         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
622                 unsigned shift;
623                 unsigned pdshift;
624
625                 if (!mmu_psize_defs[psize].shift)
626                         continue;
627
628                 shift = mmu_psize_to_shift(psize);
629
630 #ifdef CONFIG_PPC_BOOK3S_64
631                 if (shift > PGDIR_SHIFT)
632                         continue;
633                 else if (shift > PUD_SHIFT)
634                         pdshift = PGDIR_SHIFT;
635                 else if (shift > PMD_SHIFT)
636                         pdshift = PUD_SHIFT;
637                 else
638                         pdshift = PMD_SHIFT;
639 #else
640                 if (shift < PUD_SHIFT)
641                         pdshift = PMD_SHIFT;
642                 else if (shift < PGDIR_SHIFT)
643                         pdshift = PUD_SHIFT;
644                 else
645                         pdshift = PGDIR_SHIFT;
646 #endif
647
648                 if (add_huge_page_size(1ULL << shift) < 0)
649                         continue;
650                 /*
651                  * if we have pdshift and shift value same, we don't
652                  * use pgt cache for hugepd.
653                  */
654                 if (pdshift > shift) {
655                         if (!IS_ENABLED(CONFIG_PPC_8xx))
656                                 pgtable_cache_add(pdshift - shift);
657                 } else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) ||
658                            IS_ENABLED(CONFIG_PPC_8xx)) {
659                         pgtable_cache_add(PTE_T_ORDER);
660                 }
661
662                 configured = true;
663         }
664
665         if (configured) {
666                 if (IS_ENABLED(CONFIG_HUGETLB_PAGE_SIZE_VARIABLE))
667                         hugetlbpage_init_default();
668         } else
669                 pr_info("Failed to initialize. Disabling HugeTLB");
670
671         return 0;
672 }
673
674 arch_initcall(hugetlbpage_init);
675
676 void flush_dcache_icache_hugepage(struct page *page)
677 {
678         int i;
679         void *start;
680
681         BUG_ON(!PageCompound(page));
682
683         for (i = 0; i < compound_nr(page); i++) {
684                 if (!PageHighMem(page)) {
685                         __flush_dcache_icache(page_address(page+i));
686                 } else {
687                         start = kmap_atomic(page+i);
688                         __flush_dcache_icache(start);
689                         kunmap_atomic(start);
690                 }
691         }
692 }