xref: /linux/arch/powerpc/mm/hugetlbpage.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 /*
2  * PPC Huge TLB Page Support for Kernel.
3  *
4  * Copyright (C) 2003 David Gibson, IBM Corporation.
5  * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
6  *
7  * Based on the IA-32 version:
8  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
9  */
10 
11 #include <linux/mm.h>
12 #include <linux/io.h>
13 #include <linux/slab.h>
14 #include <linux/hugetlb.h>
15 #include <linux/export.h>
16 #include <linux/of_fdt.h>
17 #include <linux/memblock.h>
18 #include <linux/bootmem.h>
19 #include <linux/moduleparam.h>
20 #include <asm/pgtable.h>
21 #include <asm/pgalloc.h>
22 #include <asm/tlb.h>
23 #include <asm/setup.h>
24 #include <asm/hugetlb.h>
25 
26 #ifdef CONFIG_HUGETLB_PAGE
27 
28 #define PAGE_SHIFT_64K	16
29 #define PAGE_SHIFT_512K	19
30 #define PAGE_SHIFT_8M	23
31 #define PAGE_SHIFT_16M	24
32 #define PAGE_SHIFT_16G	34
33 
34 unsigned int HPAGE_SHIFT;
35 
36 /*
37  * Tracks gpages after the device tree is scanned and before the
38  * huge_boot_pages list is ready.  On non-Freescale implementations, this is
39  * just used to track 16G pages and so is a single array.  FSL-based
40  * implementations may have more than one gpage size, so we need multiple
41  * arrays
42  */
43 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
44 #define MAX_NUMBER_GPAGES	128
45 struct psize_gpages {
46 	u64 gpage_list[MAX_NUMBER_GPAGES];
47 	unsigned int nr_gpages;
48 };
49 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
50 #else
51 #define MAX_NUMBER_GPAGES	1024
52 static u64 gpage_freearray[MAX_NUMBER_GPAGES];
53 static unsigned nr_gpages;
54 #endif
55 
56 #define hugepd_none(hpd)	(hpd_val(hpd) == 0)
57 
58 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
59 {
60 	/* Only called for hugetlbfs pages, hence can ignore THP */
61 	return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL);
62 }
63 
64 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
65 			   unsigned long address, unsigned pdshift, unsigned pshift)
66 {
67 	struct kmem_cache *cachep;
68 	pte_t *new;
69 	int i;
70 	int num_hugepd;
71 
72 	if (pshift >= pdshift) {
73 		cachep = hugepte_cache;
74 		num_hugepd = 1 << (pshift - pdshift);
75 	} else {
76 		cachep = PGT_CACHE(pdshift - pshift);
77 		num_hugepd = 1;
78 	}
79 
80 	new = kmem_cache_zalloc(cachep, GFP_KERNEL);
81 
82 	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
83 	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
84 
85 	if (! new)
86 		return -ENOMEM;
87 
88 	/*
89 	 * Make sure other cpus find the hugepd set only after a
90 	 * properly initialized page table is visible to them.
91 	 * For more details look for comment in __pte_alloc().
92 	 */
93 	smp_wmb();
94 
95 	spin_lock(&mm->page_table_lock);
96 
97 	/*
98 	 * We have multiple higher-level entries that point to the same
99 	 * actual pte location.  Fill in each as we go and backtrack on error.
100 	 * We need all of these so the DTLB pgtable walk code can find the
101 	 * right higher-level entry without knowing if it's a hugepage or not.
102 	 */
103 	for (i = 0; i < num_hugepd; i++, hpdp++) {
104 		if (unlikely(!hugepd_none(*hpdp)))
105 			break;
106 		else {
107 #ifdef CONFIG_PPC_BOOK3S_64
108 			*hpdp = __hugepd(__pa(new) |
109 					 (shift_to_mmu_psize(pshift) << 2));
110 #elif defined(CONFIG_PPC_8xx)
111 			*hpdp = __hugepd(__pa(new) |
112 					 (pshift == PAGE_SHIFT_8M ? _PMD_PAGE_8M :
113 					  _PMD_PAGE_512K) | _PMD_PRESENT);
114 #else
115 			/* We use the old format for PPC_FSL_BOOK3E */
116 			*hpdp = __hugepd(((unsigned long)new & ~PD_HUGE) | pshift);
117 #endif
118 		}
119 	}
120 	/* If we bailed from the for loop early, an error occurred, clean up */
121 	if (i < num_hugepd) {
122 		for (i = i - 1 ; i >= 0; i--, hpdp--)
123 			*hpdp = __hugepd(0);
124 		kmem_cache_free(cachep, new);
125 	}
126 	spin_unlock(&mm->page_table_lock);
127 	return 0;
128 }
129 
130 /*
131  * These macros define how to determine which level of the page table holds
132  * the hpdp.
133  */
134 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
135 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
136 #define HUGEPD_PUD_SHIFT PUD_SHIFT
137 #else
138 #define HUGEPD_PGD_SHIFT PUD_SHIFT
139 #define HUGEPD_PUD_SHIFT PMD_SHIFT
140 #endif
141 
142 /*
143  * At this point we do the placement change only for BOOK3S 64. This would
144  * possibly work on other subarchs.
145  */
146 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
147 {
148 	pgd_t *pg;
149 	pud_t *pu;
150 	pmd_t *pm;
151 	hugepd_t *hpdp = NULL;
152 	unsigned pshift = __ffs(sz);
153 	unsigned pdshift = PGDIR_SHIFT;
154 
155 	addr &= ~(sz-1);
156 	pg = pgd_offset(mm, addr);
157 
158 #ifdef CONFIG_PPC_BOOK3S_64
159 	if (pshift == PGDIR_SHIFT)
160 		/* 16GB huge page */
161 		return (pte_t *) pg;
162 	else if (pshift > PUD_SHIFT)
163 		/*
164 		 * We need to use hugepd table
165 		 */
166 		hpdp = (hugepd_t *)pg;
167 	else {
168 		pdshift = PUD_SHIFT;
169 		pu = pud_alloc(mm, pg, addr);
170 		if (pshift == PUD_SHIFT)
171 			return (pte_t *)pu;
172 		else if (pshift > PMD_SHIFT)
173 			hpdp = (hugepd_t *)pu;
174 		else {
175 			pdshift = PMD_SHIFT;
176 			pm = pmd_alloc(mm, pu, addr);
177 			if (pshift == PMD_SHIFT)
178 				/* 16MB hugepage */
179 				return (pte_t *)pm;
180 			else
181 				hpdp = (hugepd_t *)pm;
182 		}
183 	}
184 #else
185 	if (pshift >= HUGEPD_PGD_SHIFT) {
186 		hpdp = (hugepd_t *)pg;
187 	} else {
188 		pdshift = PUD_SHIFT;
189 		pu = pud_alloc(mm, pg, addr);
190 		if (pshift >= HUGEPD_PUD_SHIFT) {
191 			hpdp = (hugepd_t *)pu;
192 		} else {
193 			pdshift = PMD_SHIFT;
194 			pm = pmd_alloc(mm, pu, addr);
195 			hpdp = (hugepd_t *)pm;
196 		}
197 	}
198 #endif
199 	if (!hpdp)
200 		return NULL;
201 
202 	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
203 
204 	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
205 		return NULL;
206 
207 	return hugepte_offset(*hpdp, addr, pdshift);
208 }
209 
210 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
211 /* Build list of addresses of gigantic pages.  This function is used in early
212  * boot before the buddy allocator is setup.
213  */
214 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
215 {
216 	unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
217 	int i;
218 
219 	if (addr == 0)
220 		return;
221 
222 	gpage_freearray[idx].nr_gpages = number_of_pages;
223 
224 	for (i = 0; i < number_of_pages; i++) {
225 		gpage_freearray[idx].gpage_list[i] = addr;
226 		addr += page_size;
227 	}
228 }
229 
230 /*
231  * Moves the gigantic page addresses from the temporary list to the
232  * huge_boot_pages list.
233  */
234 int alloc_bootmem_huge_page(struct hstate *hstate)
235 {
236 	struct huge_bootmem_page *m;
237 	int idx = shift_to_mmu_psize(huge_page_shift(hstate));
238 	int nr_gpages = gpage_freearray[idx].nr_gpages;
239 
240 	if (nr_gpages == 0)
241 		return 0;
242 
243 #ifdef CONFIG_HIGHMEM
244 	/*
245 	 * If gpages can be in highmem we can't use the trick of storing the
246 	 * data structure in the page; allocate space for this
247 	 */
248 	m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
249 	m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
250 #else
251 	m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
252 #endif
253 
254 	list_add(&m->list, &huge_boot_pages);
255 	gpage_freearray[idx].nr_gpages = nr_gpages;
256 	gpage_freearray[idx].gpage_list[nr_gpages] = 0;
257 	m->hstate = hstate;
258 
259 	return 1;
260 }
261 /*
262  * Scan the command line hugepagesz= options for gigantic pages; store those in
263  * a list that we use to allocate the memory once all options are parsed.
264  */
265 
266 unsigned long gpage_npages[MMU_PAGE_COUNT];
267 
268 static int __init do_gpage_early_setup(char *param, char *val,
269 				       const char *unused, void *arg)
270 {
271 	static phys_addr_t size;
272 	unsigned long npages;
273 
274 	/*
275 	 * The hugepagesz and hugepages cmdline options are interleaved.  We
276 	 * use the size variable to keep track of whether or not this was done
277 	 * properly and skip over instances where it is incorrect.  Other
278 	 * command-line parsing code will issue warnings, so we don't need to.
279 	 *
280 	 */
281 	if ((strcmp(param, "default_hugepagesz") == 0) ||
282 	    (strcmp(param, "hugepagesz") == 0)) {
283 		size = memparse(val, NULL);
284 	} else if (strcmp(param, "hugepages") == 0) {
285 		if (size != 0) {
286 			if (sscanf(val, "%lu", &npages) <= 0)
287 				npages = 0;
288 			if (npages > MAX_NUMBER_GPAGES) {
289 				pr_warn("MMU: %lu pages requested for page "
290 #ifdef CONFIG_PHYS_ADDR_T_64BIT
291 					"size %llu KB, limiting to "
292 #else
293 					"size %u KB, limiting to "
294 #endif
295 					__stringify(MAX_NUMBER_GPAGES) "\n",
296 					npages, size / 1024);
297 				npages = MAX_NUMBER_GPAGES;
298 			}
299 			gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
300 			size = 0;
301 		}
302 	}
303 	return 0;
304 }
305 
306 
307 /*
308  * This function allocates physical space for pages that are larger than the
309  * buddy allocator can handle.  We want to allocate these in highmem because
310  * the amount of lowmem is limited.  This means that this function MUST be
311  * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
312  * allocate to grab highmem.
313  */
314 void __init reserve_hugetlb_gpages(void)
315 {
316 	static __initdata char cmdline[COMMAND_LINE_SIZE];
317 	phys_addr_t size, base;
318 	int i;
319 
320 	strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
321 	parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
322 			NULL, &do_gpage_early_setup);
323 
324 	/*
325 	 * Walk gpage list in reverse, allocating larger page sizes first.
326 	 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
327 	 * When we reach the point in the list where pages are no longer
328 	 * considered gpages, we're done.
329 	 */
330 	for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
331 		if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
332 			continue;
333 		else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
334 			break;
335 
336 		size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
337 		base = memblock_alloc_base(size * gpage_npages[i], size,
338 					   MEMBLOCK_ALLOC_ANYWHERE);
339 		add_gpage(base, size, gpage_npages[i]);
340 	}
341 }
342 
343 #else /* !PPC_FSL_BOOK3E */
344 
345 /* Build list of addresses of gigantic pages.  This function is used in early
346  * boot before the buddy allocator is setup.
347  */
348 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
349 {
350 	if (!addr)
351 		return;
352 	while (number_of_pages > 0) {
353 		gpage_freearray[nr_gpages] = addr;
354 		nr_gpages++;
355 		number_of_pages--;
356 		addr += page_size;
357 	}
358 }
359 
360 /* Moves the gigantic page addresses from the temporary list to the
361  * huge_boot_pages list.
362  */
363 int alloc_bootmem_huge_page(struct hstate *hstate)
364 {
365 	struct huge_bootmem_page *m;
366 	if (nr_gpages == 0)
367 		return 0;
368 	m = phys_to_virt(gpage_freearray[--nr_gpages]);
369 	gpage_freearray[nr_gpages] = 0;
370 	list_add(&m->list, &huge_boot_pages);
371 	m->hstate = hstate;
372 	return 1;
373 }
374 #endif
375 
376 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
377 #define HUGEPD_FREELIST_SIZE \
378 	((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
379 
380 struct hugepd_freelist {
381 	struct rcu_head	rcu;
382 	unsigned int index;
383 	void *ptes[0];
384 };
385 
386 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
387 
388 static void hugepd_free_rcu_callback(struct rcu_head *head)
389 {
390 	struct hugepd_freelist *batch =
391 		container_of(head, struct hugepd_freelist, rcu);
392 	unsigned int i;
393 
394 	for (i = 0; i < batch->index; i++)
395 		kmem_cache_free(hugepte_cache, batch->ptes[i]);
396 
397 	free_page((unsigned long)batch);
398 }
399 
400 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
401 {
402 	struct hugepd_freelist **batchp;
403 
404 	batchp = &get_cpu_var(hugepd_freelist_cur);
405 
406 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
407 	    cpumask_equal(mm_cpumask(tlb->mm),
408 			  cpumask_of(smp_processor_id()))) {
409 		kmem_cache_free(hugepte_cache, hugepte);
410 		put_cpu_var(hugepd_freelist_cur);
411 		return;
412 	}
413 
414 	if (*batchp == NULL) {
415 		*batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
416 		(*batchp)->index = 0;
417 	}
418 
419 	(*batchp)->ptes[(*batchp)->index++] = hugepte;
420 	if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
421 		call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
422 		*batchp = NULL;
423 	}
424 	put_cpu_var(hugepd_freelist_cur);
425 }
426 #else
427 static inline void hugepd_free(struct mmu_gather *tlb, void *hugepte) {}
428 #endif
429 
430 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
431 			      unsigned long start, unsigned long end,
432 			      unsigned long floor, unsigned long ceiling)
433 {
434 	pte_t *hugepte = hugepd_page(*hpdp);
435 	int i;
436 
437 	unsigned long pdmask = ~((1UL << pdshift) - 1);
438 	unsigned int num_hugepd = 1;
439 	unsigned int shift = hugepd_shift(*hpdp);
440 
441 	/* Note: On fsl the hpdp may be the first of several */
442 	if (shift > pdshift)
443 		num_hugepd = 1 << (shift - pdshift);
444 
445 	start &= pdmask;
446 	if (start < floor)
447 		return;
448 	if (ceiling) {
449 		ceiling &= pdmask;
450 		if (! ceiling)
451 			return;
452 	}
453 	if (end - 1 > ceiling - 1)
454 		return;
455 
456 	for (i = 0; i < num_hugepd; i++, hpdp++)
457 		*hpdp = __hugepd(0);
458 
459 	if (shift >= pdshift)
460 		hugepd_free(tlb, hugepte);
461 	else
462 		pgtable_free_tlb(tlb, hugepte, pdshift - shift);
463 }
464 
465 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
466 				   unsigned long addr, unsigned long end,
467 				   unsigned long floor, unsigned long ceiling)
468 {
469 	pmd_t *pmd;
470 	unsigned long next;
471 	unsigned long start;
472 
473 	start = addr;
474 	do {
475 		unsigned long more;
476 
477 		pmd = pmd_offset(pud, addr);
478 		next = pmd_addr_end(addr, end);
479 		if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
480 			/*
481 			 * if it is not hugepd pointer, we should already find
482 			 * it cleared.
483 			 */
484 			WARN_ON(!pmd_none_or_clear_bad(pmd));
485 			continue;
486 		}
487 		/*
488 		 * Increment next by the size of the huge mapping since
489 		 * there may be more than one entry at this level for a
490 		 * single hugepage, but all of them point to
491 		 * the same kmem cache that holds the hugepte.
492 		 */
493 		more = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
494 		if (more > next)
495 			next = more;
496 
497 		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
498 				  addr, next, floor, ceiling);
499 	} while (addr = next, addr != end);
500 
501 	start &= PUD_MASK;
502 	if (start < floor)
503 		return;
504 	if (ceiling) {
505 		ceiling &= PUD_MASK;
506 		if (!ceiling)
507 			return;
508 	}
509 	if (end - 1 > ceiling - 1)
510 		return;
511 
512 	pmd = pmd_offset(pud, start);
513 	pud_clear(pud);
514 	pmd_free_tlb(tlb, pmd, start);
515 	mm_dec_nr_pmds(tlb->mm);
516 }
517 
518 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
519 				   unsigned long addr, unsigned long end,
520 				   unsigned long floor, unsigned long ceiling)
521 {
522 	pud_t *pud;
523 	unsigned long next;
524 	unsigned long start;
525 
526 	start = addr;
527 	do {
528 		pud = pud_offset(pgd, addr);
529 		next = pud_addr_end(addr, end);
530 		if (!is_hugepd(__hugepd(pud_val(*pud)))) {
531 			if (pud_none_or_clear_bad(pud))
532 				continue;
533 			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
534 					       ceiling);
535 		} else {
536 			unsigned long more;
537 			/*
538 			 * Increment next by the size of the huge mapping since
539 			 * there may be more than one entry at this level for a
540 			 * single hugepage, but all of them point to
541 			 * the same kmem cache that holds the hugepte.
542 			 */
543 			more = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
544 			if (more > next)
545 				next = more;
546 
547 			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
548 					  addr, next, floor, ceiling);
549 		}
550 	} while (addr = next, addr != end);
551 
552 	start &= PGDIR_MASK;
553 	if (start < floor)
554 		return;
555 	if (ceiling) {
556 		ceiling &= PGDIR_MASK;
557 		if (!ceiling)
558 			return;
559 	}
560 	if (end - 1 > ceiling - 1)
561 		return;
562 
563 	pud = pud_offset(pgd, start);
564 	pgd_clear(pgd);
565 	pud_free_tlb(tlb, pud, start);
566 }
567 
568 /*
569  * This function frees user-level page tables of a process.
570  */
571 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
572 			    unsigned long addr, unsigned long end,
573 			    unsigned long floor, unsigned long ceiling)
574 {
575 	pgd_t *pgd;
576 	unsigned long next;
577 
578 	/*
579 	 * Because there are a number of different possible pagetable
580 	 * layouts for hugepage ranges, we limit knowledge of how
581 	 * things should be laid out to the allocation path
582 	 * (huge_pte_alloc(), above).  Everything else works out the
583 	 * structure as it goes from information in the hugepd
584 	 * pointers.  That means that we can't here use the
585 	 * optimization used in the normal page free_pgd_range(), of
586 	 * checking whether we're actually covering a large enough
587 	 * range to have to do anything at the top level of the walk
588 	 * instead of at the bottom.
589 	 *
590 	 * To make sense of this, you should probably go read the big
591 	 * block comment at the top of the normal free_pgd_range(),
592 	 * too.
593 	 */
594 
595 	do {
596 		next = pgd_addr_end(addr, end);
597 		pgd = pgd_offset(tlb->mm, addr);
598 		if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
599 			if (pgd_none_or_clear_bad(pgd))
600 				continue;
601 			hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
602 		} else {
603 			unsigned long more;
604 			/*
605 			 * Increment next by the size of the huge mapping since
606 			 * there may be more than one entry at the pgd level
607 			 * for a single hugepage, but all of them point to the
608 			 * same kmem cache that holds the hugepte.
609 			 */
610 			more = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
611 			if (more > next)
612 				next = more;
613 
614 			free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
615 					  addr, next, floor, ceiling);
616 		}
617 	} while (addr = next, addr != end);
618 }
619 
620 /*
621  * We are holding mmap_sem, so a parallel huge page collapse cannot run.
622  * To prevent hugepage split, disable irq.
623  */
624 struct page *
625 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
626 {
627 	bool is_thp;
628 	pte_t *ptep, pte;
629 	unsigned shift;
630 	unsigned long mask, flags;
631 	struct page *page = ERR_PTR(-EINVAL);
632 
633 	local_irq_save(flags);
634 	ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift);
635 	if (!ptep)
636 		goto no_page;
637 	pte = READ_ONCE(*ptep);
638 	/*
639 	 * Verify it is a huge page else bail.
640 	 * Transparent hugepages are handled by generic code. We can skip them
641 	 * here.
642 	 */
643 	if (!shift || is_thp)
644 		goto no_page;
645 
646 	if (!pte_present(pte)) {
647 		page = NULL;
648 		goto no_page;
649 	}
650 	mask = (1UL << shift) - 1;
651 	page = pte_page(pte);
652 	if (page)
653 		page += (address & mask) / PAGE_SIZE;
654 
655 no_page:
656 	local_irq_restore(flags);
657 	return page;
658 }
659 
660 struct page *
661 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
662 		pmd_t *pmd, int write)
663 {
664 	BUG();
665 	return NULL;
666 }
667 
668 struct page *
669 follow_huge_pud(struct mm_struct *mm, unsigned long address,
670 		pud_t *pud, int write)
671 {
672 	BUG();
673 	return NULL;
674 }
675 
676 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
677 				      unsigned long sz)
678 {
679 	unsigned long __boundary = (addr + sz) & ~(sz-1);
680 	return (__boundary - 1 < end - 1) ? __boundary : end;
681 }
682 
683 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
684 		unsigned long end, int write, struct page **pages, int *nr)
685 {
686 	pte_t *ptep;
687 	unsigned long sz = 1UL << hugepd_shift(hugepd);
688 	unsigned long next;
689 
690 	ptep = hugepte_offset(hugepd, addr, pdshift);
691 	do {
692 		next = hugepte_addr_end(addr, end, sz);
693 		if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
694 			return 0;
695 	} while (ptep++, addr = next, addr != end);
696 
697 	return 1;
698 }
699 
700 #ifdef CONFIG_PPC_MM_SLICES
701 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
702 					unsigned long len, unsigned long pgoff,
703 					unsigned long flags)
704 {
705 	struct hstate *hstate = hstate_file(file);
706 	int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
707 
708 	if (radix_enabled())
709 		return radix__hugetlb_get_unmapped_area(file, addr, len,
710 						       pgoff, flags);
711 	return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
712 }
713 #endif
714 
715 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
716 {
717 #ifdef CONFIG_PPC_MM_SLICES
718 	unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
719 	/* With radix we don't use slice, so derive it from vma*/
720 	if (!radix_enabled())
721 		return 1UL << mmu_psize_to_shift(psize);
722 #endif
723 	if (!is_vm_hugetlb_page(vma))
724 		return PAGE_SIZE;
725 
726 	return huge_page_size(hstate_vma(vma));
727 }
728 
729 static inline bool is_power_of_4(unsigned long x)
730 {
731 	if (is_power_of_2(x))
732 		return (__ilog2(x) % 2) ? false : true;
733 	return false;
734 }
735 
736 static int __init add_huge_page_size(unsigned long long size)
737 {
738 	int shift = __ffs(size);
739 	int mmu_psize;
740 
741 	/* Check that it is a page size supported by the hardware and
742 	 * that it fits within pagetable and slice limits. */
743 	if (size <= PAGE_SIZE)
744 		return -EINVAL;
745 #if defined(CONFIG_PPC_FSL_BOOK3E)
746 	if (!is_power_of_4(size))
747 		return -EINVAL;
748 #elif !defined(CONFIG_PPC_8xx)
749 	if (!is_power_of_2(size) || (shift > SLICE_HIGH_SHIFT))
750 		return -EINVAL;
751 #endif
752 
753 	if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
754 		return -EINVAL;
755 
756 #ifdef CONFIG_PPC_BOOK3S_64
757 	/*
758 	 * We need to make sure that for different page sizes reported by
759 	 * firmware we only add hugetlb support for page sizes that can be
760 	 * supported by linux page table layout.
761 	 * For now we have
762 	 * Radix: 2M
763 	 * Hash: 16M and 16G
764 	 */
765 	if (radix_enabled()) {
766 		if (mmu_psize != MMU_PAGE_2M)
767 			return -EINVAL;
768 	} else {
769 		if (mmu_psize != MMU_PAGE_16M && mmu_psize != MMU_PAGE_16G)
770 			return -EINVAL;
771 	}
772 #endif
773 
774 	BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
775 
776 	/* Return if huge page size has already been setup */
777 	if (size_to_hstate(size))
778 		return 0;
779 
780 	hugetlb_add_hstate(shift - PAGE_SHIFT);
781 
782 	return 0;
783 }
784 
785 static int __init hugepage_setup_sz(char *str)
786 {
787 	unsigned long long size;
788 
789 	size = memparse(str, &str);
790 
791 	if (add_huge_page_size(size) != 0) {
792 		hugetlb_bad_size();
793 		pr_err("Invalid huge page size specified(%llu)\n", size);
794 	}
795 
796 	return 1;
797 }
798 __setup("hugepagesz=", hugepage_setup_sz);
799 
800 struct kmem_cache *hugepte_cache;
801 static int __init hugetlbpage_init(void)
802 {
803 	int psize;
804 
805 #if !defined(CONFIG_PPC_FSL_BOOK3E) && !defined(CONFIG_PPC_8xx)
806 	if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
807 		return -ENODEV;
808 #endif
809 	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
810 		unsigned shift;
811 		unsigned pdshift;
812 
813 		if (!mmu_psize_defs[psize].shift)
814 			continue;
815 
816 		shift = mmu_psize_to_shift(psize);
817 
818 		if (add_huge_page_size(1ULL << shift) < 0)
819 			continue;
820 
821 		if (shift < HUGEPD_PUD_SHIFT)
822 			pdshift = PMD_SHIFT;
823 		else if (shift < HUGEPD_PGD_SHIFT)
824 			pdshift = PUD_SHIFT;
825 		else
826 			pdshift = PGDIR_SHIFT;
827 		/*
828 		 * if we have pdshift and shift value same, we don't
829 		 * use pgt cache for hugepd.
830 		 */
831 		if (pdshift > shift)
832 			pgtable_cache_add(pdshift - shift, NULL);
833 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
834 		else if (!hugepte_cache) {
835 			/*
836 			 * Create a kmem cache for hugeptes.  The bottom bits in
837 			 * the pte have size information encoded in them, so
838 			 * align them to allow this
839 			 */
840 			hugepte_cache = kmem_cache_create("hugepte-cache",
841 							  sizeof(pte_t),
842 							  HUGEPD_SHIFT_MASK + 1,
843 							  0, NULL);
844 			if (hugepte_cache == NULL)
845 				panic("%s: Unable to create kmem cache "
846 				      "for hugeptes\n", __func__);
847 
848 		}
849 #endif
850 	}
851 
852 #if defined(CONFIG_PPC_FSL_BOOK3E) || defined(CONFIG_PPC_8xx)
853 	/* Default hpage size = 4M on FSL_BOOK3E and 512k on 8xx */
854 	if (mmu_psize_defs[MMU_PAGE_4M].shift)
855 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
856 	else if (mmu_psize_defs[MMU_PAGE_512K].shift)
857 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_512K].shift;
858 #else
859 	/* Set default large page size. Currently, we pick 16M or 1M
860 	 * depending on what is available
861 	 */
862 	if (mmu_psize_defs[MMU_PAGE_16M].shift)
863 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
864 	else if (mmu_psize_defs[MMU_PAGE_1M].shift)
865 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
866 	else if (mmu_psize_defs[MMU_PAGE_2M].shift)
867 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
868 #endif
869 	return 0;
870 }
871 
872 arch_initcall(hugetlbpage_init);
873 
874 void flush_dcache_icache_hugepage(struct page *page)
875 {
876 	int i;
877 	void *start;
878 
879 	BUG_ON(!PageCompound(page));
880 
881 	for (i = 0; i < (1UL << compound_order(page)); i++) {
882 		if (!PageHighMem(page)) {
883 			__flush_dcache_icache(page_address(page+i));
884 		} else {
885 			start = kmap_atomic(page+i);
886 			__flush_dcache_icache(start);
887 			kunmap_atomic(start);
888 		}
889 	}
890 }
891 
892 #endif /* CONFIG_HUGETLB_PAGE */
893 
894 /*
895  * We have 4 cases for pgds and pmds:
896  * (1) invalid (all zeroes)
897  * (2) pointer to next table, as normal; bottom 6 bits == 0
898  * (3) leaf pte for huge page _PAGE_PTE set
899  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
900  *
901  * So long as we atomically load page table pointers we are safe against teardown,
902  * we can follow the address down to the the page and take a ref on it.
903  * This function need to be called with interrupts disabled. We use this variant
904  * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
905  */
906 
907 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
908 				   bool *is_thp, unsigned *shift)
909 {
910 	pgd_t pgd, *pgdp;
911 	pud_t pud, *pudp;
912 	pmd_t pmd, *pmdp;
913 	pte_t *ret_pte;
914 	hugepd_t *hpdp = NULL;
915 	unsigned pdshift = PGDIR_SHIFT;
916 
917 	if (shift)
918 		*shift = 0;
919 
920 	if (is_thp)
921 		*is_thp = false;
922 
923 	pgdp = pgdir + pgd_index(ea);
924 	pgd  = READ_ONCE(*pgdp);
925 	/*
926 	 * Always operate on the local stack value. This make sure the
927 	 * value don't get updated by a parallel THP split/collapse,
928 	 * page fault or a page unmap. The return pte_t * is still not
929 	 * stable. So should be checked there for above conditions.
930 	 */
931 	if (pgd_none(pgd))
932 		return NULL;
933 	else if (pgd_huge(pgd)) {
934 		ret_pte = (pte_t *) pgdp;
935 		goto out;
936 	} else if (is_hugepd(__hugepd(pgd_val(pgd))))
937 		hpdp = (hugepd_t *)&pgd;
938 	else {
939 		/*
940 		 * Even if we end up with an unmap, the pgtable will not
941 		 * be freed, because we do an rcu free and here we are
942 		 * irq disabled
943 		 */
944 		pdshift = PUD_SHIFT;
945 		pudp = pud_offset(&pgd, ea);
946 		pud  = READ_ONCE(*pudp);
947 
948 		if (pud_none(pud))
949 			return NULL;
950 		else if (pud_huge(pud)) {
951 			ret_pte = (pte_t *) pudp;
952 			goto out;
953 		} else if (is_hugepd(__hugepd(pud_val(pud))))
954 			hpdp = (hugepd_t *)&pud;
955 		else {
956 			pdshift = PMD_SHIFT;
957 			pmdp = pmd_offset(&pud, ea);
958 			pmd  = READ_ONCE(*pmdp);
959 			/*
960 			 * A hugepage collapse is captured by pmd_none, because
961 			 * it mark the pmd none and do a hpte invalidate.
962 			 */
963 			if (pmd_none(pmd))
964 				return NULL;
965 
966 			if (pmd_trans_huge(pmd)) {
967 				if (is_thp)
968 					*is_thp = true;
969 				ret_pte = (pte_t *) pmdp;
970 				goto out;
971 			}
972 
973 			if (pmd_huge(pmd)) {
974 				ret_pte = (pte_t *) pmdp;
975 				goto out;
976 			} else if (is_hugepd(__hugepd(pmd_val(pmd))))
977 				hpdp = (hugepd_t *)&pmd;
978 			else
979 				return pte_offset_kernel(&pmd, ea);
980 		}
981 	}
982 	if (!hpdp)
983 		return NULL;
984 
985 	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
986 	pdshift = hugepd_shift(*hpdp);
987 out:
988 	if (shift)
989 		*shift = pdshift;
990 	return ret_pte;
991 }
992 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
993 
994 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
995 		unsigned long end, int write, struct page **pages, int *nr)
996 {
997 	unsigned long mask;
998 	unsigned long pte_end;
999 	struct page *head, *page;
1000 	pte_t pte;
1001 	int refs;
1002 
1003 	pte_end = (addr + sz) & ~(sz-1);
1004 	if (pte_end < end)
1005 		end = pte_end;
1006 
1007 	pte = READ_ONCE(*ptep);
1008 	mask = _PAGE_PRESENT | _PAGE_READ;
1009 
1010 	/*
1011 	 * On some CPUs like the 8xx, _PAGE_RW hence _PAGE_WRITE is defined
1012 	 * as 0 and _PAGE_RO has to be set when a page is not writable
1013 	 */
1014 	if (write)
1015 		mask |= _PAGE_WRITE;
1016 	else
1017 		mask |= _PAGE_RO;
1018 
1019 	if ((pte_val(pte) & mask) != mask)
1020 		return 0;
1021 
1022 	/* hugepages are never "special" */
1023 	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1024 
1025 	refs = 0;
1026 	head = pte_page(pte);
1027 
1028 	page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1029 	do {
1030 		VM_BUG_ON(compound_head(page) != head);
1031 		pages[*nr] = page;
1032 		(*nr)++;
1033 		page++;
1034 		refs++;
1035 	} while (addr += PAGE_SIZE, addr != end);
1036 
1037 	if (!page_cache_add_speculative(head, refs)) {
1038 		*nr -= refs;
1039 		return 0;
1040 	}
1041 
1042 	if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1043 		/* Could be optimized better */
1044 		*nr -= refs;
1045 		while (refs--)
1046 			put_page(head);
1047 		return 0;
1048 	}
1049 
1050 	return 1;
1051 }
1052