Merge tag 'pull-18-rc1-work.fd' of git://git.kernel.org/pub/scm/linux/kernel/git...
[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/pgalloc.h>
23 #include <asm/tlb.h>
24 #include <asm/setup.h>
25 #include <asm/hugetlb.h>
26 #include <asm/pte-walk.h>
27
28 bool 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
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         } 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  */
109 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
110                       unsigned long addr, unsigned long sz)
111 {
112         pgd_t *pg;
113         p4d_t *p4;
114         pud_t *pu;
115         pmd_t *pm;
116         hugepd_t *hpdp = NULL;
117         unsigned pshift = __ffs(sz);
118         unsigned pdshift = PGDIR_SHIFT;
119         spinlock_t *ptl;
120
121         addr &= ~(sz-1);
122         pg = pgd_offset(mm, addr);
123         p4 = p4d_offset(pg, addr);
124
125 #ifdef CONFIG_PPC_BOOK3S_64
126         if (pshift == PGDIR_SHIFT)
127                 /* 16GB huge page */
128                 return (pte_t *) p4;
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 *)p4;
135         } else {
136                 pdshift = PUD_SHIFT;
137                 pu = pud_alloc(mm, p4, 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 *)p4;
163         } else {
164                 pdshift = PUD_SHIFT;
165                 pu = pud_alloc(mm, p4, 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         if (IS_ENABLED(CONFIG_PPC_8xx) && pshift < PMD_SHIFT)
185                 return pte_alloc_map(mm, (pmd_t *)hpdp, addr);
186
187         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
188
189         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr,
190                                                   pdshift, pshift, ptl))
191                 return NULL;
192
193         return hugepte_offset(*hpdp, addr, pdshift);
194 }
195
196 #ifdef CONFIG_PPC_BOOK3S_64
197 /*
198  * Tracks gpages after the device tree is scanned and before the
199  * huge_boot_pages list is ready on pseries.
200  */
201 #define MAX_NUMBER_GPAGES       1024
202 __initdata static u64 gpage_freearray[MAX_NUMBER_GPAGES];
203 __initdata static unsigned nr_gpages;
204
205 /*
206  * Build list of addresses of gigantic pages.  This function is used in early
207  * boot before the buddy allocator is setup.
208  */
209 void __init pseries_add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
210 {
211         if (!addr)
212                 return;
213         while (number_of_pages > 0) {
214                 gpage_freearray[nr_gpages] = addr;
215                 nr_gpages++;
216                 number_of_pages--;
217                 addr += page_size;
218         }
219 }
220
221 static int __init pseries_alloc_bootmem_huge_page(struct hstate *hstate)
222 {
223         struct huge_bootmem_page *m;
224         if (nr_gpages == 0)
225                 return 0;
226         m = phys_to_virt(gpage_freearray[--nr_gpages]);
227         gpage_freearray[nr_gpages] = 0;
228         list_add(&m->list, &huge_boot_pages);
229         m->hstate = hstate;
230         return 1;
231 }
232
233 bool __init hugetlb_node_alloc_supported(void)
234 {
235         return false;
236 }
237 #endif
238
239
240 int __init alloc_bootmem_huge_page(struct hstate *h, int nid)
241 {
242
243 #ifdef CONFIG_PPC_BOOK3S_64
244         if (firmware_has_feature(FW_FEATURE_LPAR) && !radix_enabled())
245                 return pseries_alloc_bootmem_huge_page(h);
246 #endif
247         return __alloc_bootmem_huge_page(h, nid);
248 }
249
250 #ifndef CONFIG_PPC_BOOK3S_64
251 #define HUGEPD_FREELIST_SIZE \
252         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
253
254 struct hugepd_freelist {
255         struct rcu_head rcu;
256         unsigned int index;
257         void *ptes[];
258 };
259
260 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
261
262 static void hugepd_free_rcu_callback(struct rcu_head *head)
263 {
264         struct hugepd_freelist *batch =
265                 container_of(head, struct hugepd_freelist, rcu);
266         unsigned int i;
267
268         for (i = 0; i < batch->index; i++)
269                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), batch->ptes[i]);
270
271         free_page((unsigned long)batch);
272 }
273
274 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
275 {
276         struct hugepd_freelist **batchp;
277
278         batchp = &get_cpu_var(hugepd_freelist_cur);
279
280         if (atomic_read(&tlb->mm->mm_users) < 2 ||
281             mm_is_thread_local(tlb->mm)) {
282                 kmem_cache_free(PGT_CACHE(PTE_T_ORDER), hugepte);
283                 put_cpu_var(hugepd_freelist_cur);
284                 return;
285         }
286
287         if (*batchp == NULL) {
288                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
289                 (*batchp)->index = 0;
290         }
291
292         (*batchp)->ptes[(*batchp)->index++] = hugepte;
293         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
294                 call_rcu(&(*batchp)->rcu, hugepd_free_rcu_callback);
295                 *batchp = NULL;
296         }
297         put_cpu_var(hugepd_freelist_cur);
298 }
299 #else
300 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
301 #endif
302
303 /* Return true when the entry to be freed maps more than the area being freed */
304 static bool range_is_outside_limits(unsigned long start, unsigned long end,
305                                     unsigned long floor, unsigned long ceiling,
306                                     unsigned long mask)
307 {
308         if ((start & mask) < floor)
309                 return true;
310         if (ceiling) {
311                 ceiling &= mask;
312                 if (!ceiling)
313                         return true;
314         }
315         return end - 1 > ceiling - 1;
316 }
317
318 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
319                               unsigned long start, unsigned long end,
320                               unsigned long floor, unsigned long ceiling)
321 {
322         pte_t *hugepte = hugepd_page(*hpdp);
323         int i;
324
325         unsigned long pdmask = ~((1UL << pdshift) - 1);
326         unsigned int num_hugepd = 1;
327         unsigned int shift = hugepd_shift(*hpdp);
328
329         /* Note: On fsl the hpdp may be the first of several */
330         if (shift > pdshift)
331                 num_hugepd = 1 << (shift - pdshift);
332
333         if (range_is_outside_limits(start, end, floor, ceiling, pdmask))
334                 return;
335
336         for (i = 0; i < num_hugepd; i++, hpdp++)
337                 *hpdp = __hugepd(0);
338
339         if (shift >= pdshift)
340                 hugepd_free(tlb, hugepte);
341         else
342                 pgtable_free_tlb(tlb, hugepte,
343                                  get_hugepd_cache_index(pdshift - shift));
344 }
345
346 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
347                                    unsigned long addr, unsigned long end,
348                                    unsigned long floor, unsigned long ceiling)
349 {
350         pgtable_t token = pmd_pgtable(*pmd);
351
352         if (range_is_outside_limits(addr, end, floor, ceiling, PMD_MASK))
353                 return;
354
355         pmd_clear(pmd);
356         pte_free_tlb(tlb, token, addr);
357         mm_dec_nr_ptes(tlb->mm);
358 }
359
360 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
361                                    unsigned long addr, unsigned long end,
362                                    unsigned long floor, unsigned long ceiling)
363 {
364         pmd_t *pmd;
365         unsigned long next;
366         unsigned long start;
367
368         start = addr;
369         do {
370                 unsigned long more;
371
372                 pmd = pmd_offset(pud, addr);
373                 next = pmd_addr_end(addr, end);
374                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
375                         if (pmd_none_or_clear_bad(pmd))
376                                 continue;
377
378                         /*
379                          * if it is not hugepd pointer, we should already find
380                          * it cleared.
381                          */
382                         WARN_ON(!IS_ENABLED(CONFIG_PPC_8xx));
383
384                         hugetlb_free_pte_range(tlb, pmd, addr, end, floor, ceiling);
385
386                         continue;
387                 }
388                 /*
389                  * Increment next by the size of the huge mapping since
390                  * there may be more than one entry at this level for a
391                  * single hugepage, but all of them point to
392                  * the same kmem cache that holds the hugepte.
393                  */
394                 more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
395                 if (more > next)
396                         next = more;
397
398                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
399                                   addr, next, floor, ceiling);
400         } while (addr = next, addr != end);
401
402         if (range_is_outside_limits(start, end, floor, ceiling, PUD_MASK))
403                 return;
404
405         pmd = pmd_offset(pud, start & PUD_MASK);
406         pud_clear(pud);
407         pmd_free_tlb(tlb, pmd, start & PUD_MASK);
408         mm_dec_nr_pmds(tlb->mm);
409 }
410
411 static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
412                                    unsigned long addr, unsigned long end,
413                                    unsigned long floor, unsigned long ceiling)
414 {
415         pud_t *pud;
416         unsigned long next;
417         unsigned long start;
418
419         start = addr;
420         do {
421                 pud = pud_offset(p4d, addr);
422                 next = pud_addr_end(addr, end);
423                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
424                         if (pud_none_or_clear_bad(pud))
425                                 continue;
426                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
427                                                ceiling);
428                 } else {
429                         unsigned long more;
430                         /*
431                          * Increment next by the size of the huge mapping since
432                          * there may be more than one entry at this level for a
433                          * single hugepage, but all of them point to
434                          * the same kmem cache that holds the hugepte.
435                          */
436                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
437                         if (more > next)
438                                 next = more;
439
440                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
441                                           addr, next, floor, ceiling);
442                 }
443         } while (addr = next, addr != end);
444
445         if (range_is_outside_limits(start, end, floor, ceiling, PGDIR_MASK))
446                 return;
447
448         pud = pud_offset(p4d, start & PGDIR_MASK);
449         p4d_clear(p4d);
450         pud_free_tlb(tlb, pud, start & PGDIR_MASK);
451         mm_dec_nr_puds(tlb->mm);
452 }
453
454 /*
455  * This function frees user-level page tables of a process.
456  */
457 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
458                             unsigned long addr, unsigned long end,
459                             unsigned long floor, unsigned long ceiling)
460 {
461         pgd_t *pgd;
462         p4d_t *p4d;
463         unsigned long next;
464
465         /*
466          * Because there are a number of different possible pagetable
467          * layouts for hugepage ranges, we limit knowledge of how
468          * things should be laid out to the allocation path
469          * (huge_pte_alloc(), above).  Everything else works out the
470          * structure as it goes from information in the hugepd
471          * pointers.  That means that we can't here use the
472          * optimization used in the normal page free_pgd_range(), of
473          * checking whether we're actually covering a large enough
474          * range to have to do anything at the top level of the walk
475          * instead of at the bottom.
476          *
477          * To make sense of this, you should probably go read the big
478          * block comment at the top of the normal free_pgd_range(),
479          * too.
480          */
481
482         do {
483                 next = pgd_addr_end(addr, end);
484                 pgd = pgd_offset(tlb->mm, addr);
485                 p4d = p4d_offset(pgd, addr);
486                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
487                         if (p4d_none_or_clear_bad(p4d))
488                                 continue;
489                         hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
490                 } else {
491                         unsigned long more;
492                         /*
493                          * Increment next by the size of the huge mapping since
494                          * there may be more than one entry at the pgd level
495                          * for a single hugepage, but all of them point to the
496                          * same kmem cache that holds the hugepte.
497                          */
498                         more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
499                         if (more > next)
500                                 next = more;
501
502                         free_hugepd_range(tlb, (hugepd_t *)p4d, PGDIR_SHIFT,
503                                           addr, next, floor, ceiling);
504                 }
505         } while (addr = next, addr != end);
506 }
507
508 struct page *follow_huge_pd(struct vm_area_struct *vma,
509                             unsigned long address, hugepd_t hpd,
510                             int flags, int pdshift)
511 {
512         pte_t *ptep;
513         spinlock_t *ptl;
514         struct page *page = NULL;
515         unsigned long mask;
516         int shift = hugepd_shift(hpd);
517         struct mm_struct *mm = vma->vm_mm;
518
519 retry:
520         /*
521          * hugepage directory entries are protected by mm->page_table_lock
522          * Use this instead of huge_pte_lockptr
523          */
524         ptl = &mm->page_table_lock;
525         spin_lock(ptl);
526
527         ptep = hugepte_offset(hpd, address, pdshift);
528         if (pte_present(*ptep)) {
529                 mask = (1UL << shift) - 1;
530                 page = pte_page(*ptep);
531                 page += ((address & mask) >> PAGE_SHIFT);
532                 if (flags & FOLL_GET)
533                         get_page(page);
534         } else {
535                 if (is_hugetlb_entry_migration(*ptep)) {
536                         spin_unlock(ptl);
537                         __migration_entry_wait(mm, ptep, ptl);
538                         goto retry;
539                 }
540         }
541         spin_unlock(ptl);
542         return page;
543 }
544
545 bool __init arch_hugetlb_valid_size(unsigned long size)
546 {
547         int shift = __ffs(size);
548         int mmu_psize;
549
550         /* Check that it is a page size supported by the hardware and
551          * that it fits within pagetable and slice limits. */
552         if (size <= PAGE_SIZE || !is_power_of_2(size))
553                 return false;
554
555         mmu_psize = check_and_get_huge_psize(shift);
556         if (mmu_psize < 0)
557                 return false;
558
559         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
560
561         return true;
562 }
563
564 static int __init add_huge_page_size(unsigned long long size)
565 {
566         int shift = __ffs(size);
567
568         if (!arch_hugetlb_valid_size((unsigned long)size))
569                 return -EINVAL;
570
571         hugetlb_add_hstate(shift - PAGE_SHIFT);
572         return 0;
573 }
574
575 static int __init hugetlbpage_init(void)
576 {
577         bool configured = false;
578         int psize;
579
580         if (hugetlb_disabled) {
581                 pr_info("HugeTLB support is disabled!\n");
582                 return 0;
583         }
584
585         if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) && !radix_enabled() &&
586             !mmu_has_feature(MMU_FTR_16M_PAGE))
587                 return -ENODEV;
588
589         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
590                 unsigned shift;
591                 unsigned pdshift;
592
593                 if (!mmu_psize_defs[psize].shift)
594                         continue;
595
596                 shift = mmu_psize_to_shift(psize);
597
598 #ifdef CONFIG_PPC_BOOK3S_64
599                 if (shift > PGDIR_SHIFT)
600                         continue;
601                 else if (shift > PUD_SHIFT)
602                         pdshift = PGDIR_SHIFT;
603                 else if (shift > PMD_SHIFT)
604                         pdshift = PUD_SHIFT;
605                 else
606                         pdshift = PMD_SHIFT;
607 #else
608                 if (shift < PUD_SHIFT)
609                         pdshift = PMD_SHIFT;
610                 else if (shift < PGDIR_SHIFT)
611                         pdshift = PUD_SHIFT;
612                 else
613                         pdshift = PGDIR_SHIFT;
614 #endif
615
616                 if (add_huge_page_size(1ULL << shift) < 0)
617                         continue;
618                 /*
619                  * if we have pdshift and shift value same, we don't
620                  * use pgt cache for hugepd.
621                  */
622                 if (pdshift > shift) {
623                         if (!IS_ENABLED(CONFIG_PPC_8xx))
624                                 pgtable_cache_add(pdshift - shift);
625                 } else if (IS_ENABLED(CONFIG_PPC_FSL_BOOK3E) ||
626                            IS_ENABLED(CONFIG_PPC_8xx)) {
627                         pgtable_cache_add(PTE_T_ORDER);
628                 }
629
630                 configured = true;
631         }
632
633         if (!configured)
634                 pr_info("Failed to initialize. Disabling HugeTLB");
635
636         return 0;
637 }
638
639 arch_initcall(hugetlbpage_init);
640
641 void __init gigantic_hugetlb_cma_reserve(void)
642 {
643         unsigned long order = 0;
644
645         if (radix_enabled())
646                 order = PUD_SHIFT - PAGE_SHIFT;
647         else if (!firmware_has_feature(FW_FEATURE_LPAR) && mmu_psize_defs[MMU_PAGE_16G].shift)
648                 /*
649                  * For pseries we do use ibm,expected#pages for reserving 16G pages.
650                  */
651                 order = mmu_psize_to_shift(MMU_PAGE_16G) - PAGE_SHIFT;
652
653         if (order) {
654                 VM_WARN_ON(order < MAX_ORDER);
655                 hugetlb_cma_reserve(order);
656         }
657 }