xref: /openbmc/linux/arch/powerpc/mm/hugetlbpage.c (revision c819e2cf)
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_16M	24
30 #define PAGE_SHIFT_16G	34
31 
32 unsigned int HPAGE_SHIFT;
33 
34 /*
35  * Tracks gpages after the device tree is scanned and before the
36  * huge_boot_pages list is ready.  On non-Freescale implementations, this is
37  * just used to track 16G pages and so is a single array.  FSL-based
38  * implementations may have more than one gpage size, so we need multiple
39  * arrays
40  */
41 #ifdef CONFIG_PPC_FSL_BOOK3E
42 #define MAX_NUMBER_GPAGES	128
43 struct psize_gpages {
44 	u64 gpage_list[MAX_NUMBER_GPAGES];
45 	unsigned int nr_gpages;
46 };
47 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
48 #else
49 #define MAX_NUMBER_GPAGES	1024
50 static u64 gpage_freearray[MAX_NUMBER_GPAGES];
51 static unsigned nr_gpages;
52 #endif
53 
54 #define hugepd_none(hpd)	((hpd).pd == 0)
55 
56 #ifdef CONFIG_PPC_BOOK3S_64
57 /*
58  * At this point we do the placement change only for BOOK3S 64. This would
59  * possibly work on other subarchs.
60  */
61 
62 /*
63  * We have PGD_INDEX_SIZ = 12 and PTE_INDEX_SIZE = 8, so that we can have
64  * 16GB hugepage pte in PGD and 16MB hugepage pte at PMD;
65  *
66  * Defined in such a way that we can optimize away code block at build time
67  * if CONFIG_HUGETLB_PAGE=n.
68  */
69 int pmd_huge(pmd_t pmd)
70 {
71 	/*
72 	 * leaf pte for huge page, bottom two bits != 00
73 	 */
74 	return ((pmd_val(pmd) & 0x3) != 0x0);
75 }
76 
77 int pud_huge(pud_t pud)
78 {
79 	/*
80 	 * leaf pte for huge page, bottom two bits != 00
81 	 */
82 	return ((pud_val(pud) & 0x3) != 0x0);
83 }
84 
85 int pgd_huge(pgd_t pgd)
86 {
87 	/*
88 	 * leaf pte for huge page, bottom two bits != 00
89 	 */
90 	return ((pgd_val(pgd) & 0x3) != 0x0);
91 }
92 #else
93 int pmd_huge(pmd_t pmd)
94 {
95 	return 0;
96 }
97 
98 int pud_huge(pud_t pud)
99 {
100 	return 0;
101 }
102 
103 int pgd_huge(pgd_t pgd)
104 {
105 	return 0;
106 }
107 #endif
108 
109 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
110 {
111 	/* Only called for hugetlbfs pages, hence can ignore THP */
112 	return find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
113 }
114 
115 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
116 			   unsigned long address, unsigned pdshift, unsigned pshift)
117 {
118 	struct kmem_cache *cachep;
119 	pte_t *new;
120 
121 #ifdef CONFIG_PPC_FSL_BOOK3E
122 	int i;
123 	int num_hugepd = 1 << (pshift - pdshift);
124 	cachep = hugepte_cache;
125 #else
126 	cachep = PGT_CACHE(pdshift - pshift);
127 #endif
128 
129 	new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
130 
131 	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
132 	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
133 
134 	if (! new)
135 		return -ENOMEM;
136 
137 	spin_lock(&mm->page_table_lock);
138 #ifdef CONFIG_PPC_FSL_BOOK3E
139 	/*
140 	 * We have multiple higher-level entries that point to the same
141 	 * actual pte location.  Fill in each as we go and backtrack on error.
142 	 * We need all of these so the DTLB pgtable walk code can find the
143 	 * right higher-level entry without knowing if it's a hugepage or not.
144 	 */
145 	for (i = 0; i < num_hugepd; i++, hpdp++) {
146 		if (unlikely(!hugepd_none(*hpdp)))
147 			break;
148 		else
149 			/* We use the old format for PPC_FSL_BOOK3E */
150 			hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
151 	}
152 	/* If we bailed from the for loop early, an error occurred, clean up */
153 	if (i < num_hugepd) {
154 		for (i = i - 1 ; i >= 0; i--, hpdp--)
155 			hpdp->pd = 0;
156 		kmem_cache_free(cachep, new);
157 	}
158 #else
159 	if (!hugepd_none(*hpdp))
160 		kmem_cache_free(cachep, new);
161 	else {
162 #ifdef CONFIG_PPC_BOOK3S_64
163 		hpdp->pd = (unsigned long)new |
164 			    (shift_to_mmu_psize(pshift) << 2);
165 #else
166 		hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
167 #endif
168 	}
169 #endif
170 	spin_unlock(&mm->page_table_lock);
171 	return 0;
172 }
173 
174 /*
175  * These macros define how to determine which level of the page table holds
176  * the hpdp.
177  */
178 #ifdef CONFIG_PPC_FSL_BOOK3E
179 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
180 #define HUGEPD_PUD_SHIFT PUD_SHIFT
181 #else
182 #define HUGEPD_PGD_SHIFT PUD_SHIFT
183 #define HUGEPD_PUD_SHIFT PMD_SHIFT
184 #endif
185 
186 #ifdef CONFIG_PPC_BOOK3S_64
187 /*
188  * At this point we do the placement change only for BOOK3S 64. This would
189  * possibly work on other subarchs.
190  */
191 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
192 {
193 	pgd_t *pg;
194 	pud_t *pu;
195 	pmd_t *pm;
196 	hugepd_t *hpdp = NULL;
197 	unsigned pshift = __ffs(sz);
198 	unsigned pdshift = PGDIR_SHIFT;
199 
200 	addr &= ~(sz-1);
201 	pg = pgd_offset(mm, addr);
202 
203 	if (pshift == PGDIR_SHIFT)
204 		/* 16GB huge page */
205 		return (pte_t *) pg;
206 	else if (pshift > PUD_SHIFT)
207 		/*
208 		 * We need to use hugepd table
209 		 */
210 		hpdp = (hugepd_t *)pg;
211 	else {
212 		pdshift = PUD_SHIFT;
213 		pu = pud_alloc(mm, pg, addr);
214 		if (pshift == PUD_SHIFT)
215 			return (pte_t *)pu;
216 		else if (pshift > PMD_SHIFT)
217 			hpdp = (hugepd_t *)pu;
218 		else {
219 			pdshift = PMD_SHIFT;
220 			pm = pmd_alloc(mm, pu, addr);
221 			if (pshift == PMD_SHIFT)
222 				/* 16MB hugepage */
223 				return (pte_t *)pm;
224 			else
225 				hpdp = (hugepd_t *)pm;
226 		}
227 	}
228 	if (!hpdp)
229 		return NULL;
230 
231 	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
232 
233 	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
234 		return NULL;
235 
236 	return hugepte_offset(*hpdp, addr, pdshift);
237 }
238 
239 #else
240 
241 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
242 {
243 	pgd_t *pg;
244 	pud_t *pu;
245 	pmd_t *pm;
246 	hugepd_t *hpdp = NULL;
247 	unsigned pshift = __ffs(sz);
248 	unsigned pdshift = PGDIR_SHIFT;
249 
250 	addr &= ~(sz-1);
251 
252 	pg = pgd_offset(mm, addr);
253 
254 	if (pshift >= HUGEPD_PGD_SHIFT) {
255 		hpdp = (hugepd_t *)pg;
256 	} else {
257 		pdshift = PUD_SHIFT;
258 		pu = pud_alloc(mm, pg, addr);
259 		if (pshift >= HUGEPD_PUD_SHIFT) {
260 			hpdp = (hugepd_t *)pu;
261 		} else {
262 			pdshift = PMD_SHIFT;
263 			pm = pmd_alloc(mm, pu, addr);
264 			hpdp = (hugepd_t *)pm;
265 		}
266 	}
267 
268 	if (!hpdp)
269 		return NULL;
270 
271 	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
272 
273 	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
274 		return NULL;
275 
276 	return hugepte_offset(*hpdp, addr, pdshift);
277 }
278 #endif
279 
280 #ifdef CONFIG_PPC_FSL_BOOK3E
281 /* Build list of addresses of gigantic pages.  This function is used in early
282  * boot before the buddy allocator is setup.
283  */
284 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
285 {
286 	unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
287 	int i;
288 
289 	if (addr == 0)
290 		return;
291 
292 	gpage_freearray[idx].nr_gpages = number_of_pages;
293 
294 	for (i = 0; i < number_of_pages; i++) {
295 		gpage_freearray[idx].gpage_list[i] = addr;
296 		addr += page_size;
297 	}
298 }
299 
300 /*
301  * Moves the gigantic page addresses from the temporary list to the
302  * huge_boot_pages list.
303  */
304 int alloc_bootmem_huge_page(struct hstate *hstate)
305 {
306 	struct huge_bootmem_page *m;
307 	int idx = shift_to_mmu_psize(huge_page_shift(hstate));
308 	int nr_gpages = gpage_freearray[idx].nr_gpages;
309 
310 	if (nr_gpages == 0)
311 		return 0;
312 
313 #ifdef CONFIG_HIGHMEM
314 	/*
315 	 * If gpages can be in highmem we can't use the trick of storing the
316 	 * data structure in the page; allocate space for this
317 	 */
318 	m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
319 	m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
320 #else
321 	m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
322 #endif
323 
324 	list_add(&m->list, &huge_boot_pages);
325 	gpage_freearray[idx].nr_gpages = nr_gpages;
326 	gpage_freearray[idx].gpage_list[nr_gpages] = 0;
327 	m->hstate = hstate;
328 
329 	return 1;
330 }
331 /*
332  * Scan the command line hugepagesz= options for gigantic pages; store those in
333  * a list that we use to allocate the memory once all options are parsed.
334  */
335 
336 unsigned long gpage_npages[MMU_PAGE_COUNT];
337 
338 static int __init do_gpage_early_setup(char *param, char *val,
339 				       const char *unused)
340 {
341 	static phys_addr_t size;
342 	unsigned long npages;
343 
344 	/*
345 	 * The hugepagesz and hugepages cmdline options are interleaved.  We
346 	 * use the size variable to keep track of whether or not this was done
347 	 * properly and skip over instances where it is incorrect.  Other
348 	 * command-line parsing code will issue warnings, so we don't need to.
349 	 *
350 	 */
351 	if ((strcmp(param, "default_hugepagesz") == 0) ||
352 	    (strcmp(param, "hugepagesz") == 0)) {
353 		size = memparse(val, NULL);
354 	} else if (strcmp(param, "hugepages") == 0) {
355 		if (size != 0) {
356 			if (sscanf(val, "%lu", &npages) <= 0)
357 				npages = 0;
358 			if (npages > MAX_NUMBER_GPAGES) {
359 				pr_warn("MMU: %lu pages requested for page "
360 					"size %llu KB, limiting to "
361 					__stringify(MAX_NUMBER_GPAGES) "\n",
362 					npages, size / 1024);
363 				npages = MAX_NUMBER_GPAGES;
364 			}
365 			gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
366 			size = 0;
367 		}
368 	}
369 	return 0;
370 }
371 
372 
373 /*
374  * This function allocates physical space for pages that are larger than the
375  * buddy allocator can handle.  We want to allocate these in highmem because
376  * the amount of lowmem is limited.  This means that this function MUST be
377  * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
378  * allocate to grab highmem.
379  */
380 void __init reserve_hugetlb_gpages(void)
381 {
382 	static __initdata char cmdline[COMMAND_LINE_SIZE];
383 	phys_addr_t size, base;
384 	int i;
385 
386 	strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
387 	parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
388 			&do_gpage_early_setup);
389 
390 	/*
391 	 * Walk gpage list in reverse, allocating larger page sizes first.
392 	 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
393 	 * When we reach the point in the list where pages are no longer
394 	 * considered gpages, we're done.
395 	 */
396 	for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
397 		if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
398 			continue;
399 		else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
400 			break;
401 
402 		size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
403 		base = memblock_alloc_base(size * gpage_npages[i], size,
404 					   MEMBLOCK_ALLOC_ANYWHERE);
405 		add_gpage(base, size, gpage_npages[i]);
406 	}
407 }
408 
409 #else /* !PPC_FSL_BOOK3E */
410 
411 /* Build list of addresses of gigantic pages.  This function is used in early
412  * boot before the buddy allocator is setup.
413  */
414 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
415 {
416 	if (!addr)
417 		return;
418 	while (number_of_pages > 0) {
419 		gpage_freearray[nr_gpages] = addr;
420 		nr_gpages++;
421 		number_of_pages--;
422 		addr += page_size;
423 	}
424 }
425 
426 /* Moves the gigantic page addresses from the temporary list to the
427  * huge_boot_pages list.
428  */
429 int alloc_bootmem_huge_page(struct hstate *hstate)
430 {
431 	struct huge_bootmem_page *m;
432 	if (nr_gpages == 0)
433 		return 0;
434 	m = phys_to_virt(gpage_freearray[--nr_gpages]);
435 	gpage_freearray[nr_gpages] = 0;
436 	list_add(&m->list, &huge_boot_pages);
437 	m->hstate = hstate;
438 	return 1;
439 }
440 #endif
441 
442 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
443 {
444 	return 0;
445 }
446 
447 #ifdef CONFIG_PPC_FSL_BOOK3E
448 #define HUGEPD_FREELIST_SIZE \
449 	((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
450 
451 struct hugepd_freelist {
452 	struct rcu_head	rcu;
453 	unsigned int index;
454 	void *ptes[0];
455 };
456 
457 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
458 
459 static void hugepd_free_rcu_callback(struct rcu_head *head)
460 {
461 	struct hugepd_freelist *batch =
462 		container_of(head, struct hugepd_freelist, rcu);
463 	unsigned int i;
464 
465 	for (i = 0; i < batch->index; i++)
466 		kmem_cache_free(hugepte_cache, batch->ptes[i]);
467 
468 	free_page((unsigned long)batch);
469 }
470 
471 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
472 {
473 	struct hugepd_freelist **batchp;
474 
475 	batchp = this_cpu_ptr(&hugepd_freelist_cur);
476 
477 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
478 	    cpumask_equal(mm_cpumask(tlb->mm),
479 			  cpumask_of(smp_processor_id()))) {
480 		kmem_cache_free(hugepte_cache, hugepte);
481         put_cpu_var(hugepd_freelist_cur);
482 		return;
483 	}
484 
485 	if (*batchp == NULL) {
486 		*batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
487 		(*batchp)->index = 0;
488 	}
489 
490 	(*batchp)->ptes[(*batchp)->index++] = hugepte;
491 	if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
492 		call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
493 		*batchp = NULL;
494 	}
495 	put_cpu_var(hugepd_freelist_cur);
496 }
497 #endif
498 
499 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
500 			      unsigned long start, unsigned long end,
501 			      unsigned long floor, unsigned long ceiling)
502 {
503 	pte_t *hugepte = hugepd_page(*hpdp);
504 	int i;
505 
506 	unsigned long pdmask = ~((1UL << pdshift) - 1);
507 	unsigned int num_hugepd = 1;
508 
509 #ifdef CONFIG_PPC_FSL_BOOK3E
510 	/* Note: On fsl the hpdp may be the first of several */
511 	num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
512 #else
513 	unsigned int shift = hugepd_shift(*hpdp);
514 #endif
515 
516 	start &= pdmask;
517 	if (start < floor)
518 		return;
519 	if (ceiling) {
520 		ceiling &= pdmask;
521 		if (! ceiling)
522 			return;
523 	}
524 	if (end - 1 > ceiling - 1)
525 		return;
526 
527 	for (i = 0; i < num_hugepd; i++, hpdp++)
528 		hpdp->pd = 0;
529 
530 #ifdef CONFIG_PPC_FSL_BOOK3E
531 	hugepd_free(tlb, hugepte);
532 #else
533 	pgtable_free_tlb(tlb, hugepte, pdshift - shift);
534 #endif
535 }
536 
537 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
538 				   unsigned long addr, unsigned long end,
539 				   unsigned long floor, unsigned long ceiling)
540 {
541 	pmd_t *pmd;
542 	unsigned long next;
543 	unsigned long start;
544 
545 	start = addr;
546 	do {
547 		pmd = pmd_offset(pud, addr);
548 		next = pmd_addr_end(addr, end);
549 		if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
550 			/*
551 			 * if it is not hugepd pointer, we should already find
552 			 * it cleared.
553 			 */
554 			WARN_ON(!pmd_none_or_clear_bad(pmd));
555 			continue;
556 		}
557 #ifdef CONFIG_PPC_FSL_BOOK3E
558 		/*
559 		 * Increment next by the size of the huge mapping since
560 		 * there may be more than one entry at this level for a
561 		 * single hugepage, but all of them point to
562 		 * the same kmem cache that holds the hugepte.
563 		 */
564 		next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
565 #endif
566 		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
567 				  addr, next, floor, ceiling);
568 	} while (addr = next, addr != end);
569 
570 	start &= PUD_MASK;
571 	if (start < floor)
572 		return;
573 	if (ceiling) {
574 		ceiling &= PUD_MASK;
575 		if (!ceiling)
576 			return;
577 	}
578 	if (end - 1 > ceiling - 1)
579 		return;
580 
581 	pmd = pmd_offset(pud, start);
582 	pud_clear(pud);
583 	pmd_free_tlb(tlb, pmd, start);
584 }
585 
586 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
587 				   unsigned long addr, unsigned long end,
588 				   unsigned long floor, unsigned long ceiling)
589 {
590 	pud_t *pud;
591 	unsigned long next;
592 	unsigned long start;
593 
594 	start = addr;
595 	do {
596 		pud = pud_offset(pgd, addr);
597 		next = pud_addr_end(addr, end);
598 		if (!is_hugepd(__hugepd(pud_val(*pud)))) {
599 			if (pud_none_or_clear_bad(pud))
600 				continue;
601 			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
602 					       ceiling);
603 		} else {
604 #ifdef CONFIG_PPC_FSL_BOOK3E
605 			/*
606 			 * Increment next by the size of the huge mapping since
607 			 * there may be more than one entry at this level for a
608 			 * single hugepage, but all of them point to
609 			 * the same kmem cache that holds the hugepte.
610 			 */
611 			next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
612 #endif
613 			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
614 					  addr, next, floor, ceiling);
615 		}
616 	} while (addr = next, addr != end);
617 
618 	start &= PGDIR_MASK;
619 	if (start < floor)
620 		return;
621 	if (ceiling) {
622 		ceiling &= PGDIR_MASK;
623 		if (!ceiling)
624 			return;
625 	}
626 	if (end - 1 > ceiling - 1)
627 		return;
628 
629 	pud = pud_offset(pgd, start);
630 	pgd_clear(pgd);
631 	pud_free_tlb(tlb, pud, start);
632 }
633 
634 /*
635  * This function frees user-level page tables of a process.
636  */
637 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
638 			    unsigned long addr, unsigned long end,
639 			    unsigned long floor, unsigned long ceiling)
640 {
641 	pgd_t *pgd;
642 	unsigned long next;
643 
644 	/*
645 	 * Because there are a number of different possible pagetable
646 	 * layouts for hugepage ranges, we limit knowledge of how
647 	 * things should be laid out to the allocation path
648 	 * (huge_pte_alloc(), above).  Everything else works out the
649 	 * structure as it goes from information in the hugepd
650 	 * pointers.  That means that we can't here use the
651 	 * optimization used in the normal page free_pgd_range(), of
652 	 * checking whether we're actually covering a large enough
653 	 * range to have to do anything at the top level of the walk
654 	 * instead of at the bottom.
655 	 *
656 	 * To make sense of this, you should probably go read the big
657 	 * block comment at the top of the normal free_pgd_range(),
658 	 * too.
659 	 */
660 
661 	do {
662 		next = pgd_addr_end(addr, end);
663 		pgd = pgd_offset(tlb->mm, addr);
664 		if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
665 			if (pgd_none_or_clear_bad(pgd))
666 				continue;
667 			hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
668 		} else {
669 #ifdef CONFIG_PPC_FSL_BOOK3E
670 			/*
671 			 * Increment next by the size of the huge mapping since
672 			 * there may be more than one entry at the pgd level
673 			 * for a single hugepage, but all of them point to the
674 			 * same kmem cache that holds the hugepte.
675 			 */
676 			next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
677 #endif
678 			free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
679 					  addr, next, floor, ceiling);
680 		}
681 	} while (addr = next, addr != end);
682 }
683 
684 struct page *
685 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
686 {
687 	pte_t *ptep;
688 	struct page *page;
689 	unsigned shift;
690 	unsigned long mask;
691 	/*
692 	 * Transparent hugepages are handled by generic code. We can skip them
693 	 * here.
694 	 */
695 	ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
696 
697 	/* Verify it is a huge page else bail. */
698 	if (!ptep || !shift || pmd_trans_huge(*(pmd_t *)ptep))
699 		return ERR_PTR(-EINVAL);
700 
701 	mask = (1UL << shift) - 1;
702 	page = pte_page(*ptep);
703 	if (page)
704 		page += (address & mask) / PAGE_SIZE;
705 
706 	return page;
707 }
708 
709 struct page *
710 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
711 		pmd_t *pmd, int write)
712 {
713 	BUG();
714 	return NULL;
715 }
716 
717 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
718 				      unsigned long sz)
719 {
720 	unsigned long __boundary = (addr + sz) & ~(sz-1);
721 	return (__boundary - 1 < end - 1) ? __boundary : end;
722 }
723 
724 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
725 		unsigned long end, int write, struct page **pages, int *nr)
726 {
727 	pte_t *ptep;
728 	unsigned long sz = 1UL << hugepd_shift(hugepd);
729 	unsigned long next;
730 
731 	ptep = hugepte_offset(hugepd, addr, pdshift);
732 	do {
733 		next = hugepte_addr_end(addr, end, sz);
734 		if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
735 			return 0;
736 	} while (ptep++, addr = next, addr != end);
737 
738 	return 1;
739 }
740 
741 #ifdef CONFIG_PPC_MM_SLICES
742 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
743 					unsigned long len, unsigned long pgoff,
744 					unsigned long flags)
745 {
746 	struct hstate *hstate = hstate_file(file);
747 	int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
748 
749 	return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
750 }
751 #endif
752 
753 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
754 {
755 #ifdef CONFIG_PPC_MM_SLICES
756 	unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
757 
758 	return 1UL << mmu_psize_to_shift(psize);
759 #else
760 	if (!is_vm_hugetlb_page(vma))
761 		return PAGE_SIZE;
762 
763 	return huge_page_size(hstate_vma(vma));
764 #endif
765 }
766 
767 static inline bool is_power_of_4(unsigned long x)
768 {
769 	if (is_power_of_2(x))
770 		return (__ilog2(x) % 2) ? false : true;
771 	return false;
772 }
773 
774 static int __init add_huge_page_size(unsigned long long size)
775 {
776 	int shift = __ffs(size);
777 	int mmu_psize;
778 
779 	/* Check that it is a page size supported by the hardware and
780 	 * that it fits within pagetable and slice limits. */
781 #ifdef CONFIG_PPC_FSL_BOOK3E
782 	if ((size < PAGE_SIZE) || !is_power_of_4(size))
783 		return -EINVAL;
784 #else
785 	if (!is_power_of_2(size)
786 	    || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
787 		return -EINVAL;
788 #endif
789 
790 	if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
791 		return -EINVAL;
792 
793 #ifdef CONFIG_SPU_FS_64K_LS
794 	/* Disable support for 64K huge pages when 64K SPU local store
795 	 * support is enabled as the current implementation conflicts.
796 	 */
797 	if (shift == PAGE_SHIFT_64K)
798 		return -EINVAL;
799 #endif /* CONFIG_SPU_FS_64K_LS */
800 
801 	BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
802 
803 	/* Return if huge page size has already been setup */
804 	if (size_to_hstate(size))
805 		return 0;
806 
807 	hugetlb_add_hstate(shift - PAGE_SHIFT);
808 
809 	return 0;
810 }
811 
812 static int __init hugepage_setup_sz(char *str)
813 {
814 	unsigned long long size;
815 
816 	size = memparse(str, &str);
817 
818 	if (add_huge_page_size(size) != 0)
819 		printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
820 
821 	return 1;
822 }
823 __setup("hugepagesz=", hugepage_setup_sz);
824 
825 #ifdef CONFIG_PPC_FSL_BOOK3E
826 struct kmem_cache *hugepte_cache;
827 static int __init hugetlbpage_init(void)
828 {
829 	int psize;
830 
831 	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
832 		unsigned shift;
833 
834 		if (!mmu_psize_defs[psize].shift)
835 			continue;
836 
837 		shift = mmu_psize_to_shift(psize);
838 
839 		/* Don't treat normal page sizes as huge... */
840 		if (shift != PAGE_SHIFT)
841 			if (add_huge_page_size(1ULL << shift) < 0)
842 				continue;
843 	}
844 
845 	/*
846 	 * Create a kmem cache for hugeptes.  The bottom bits in the pte have
847 	 * size information encoded in them, so align them to allow this
848 	 */
849 	hugepte_cache =  kmem_cache_create("hugepte-cache", sizeof(pte_t),
850 					   HUGEPD_SHIFT_MASK + 1, 0, NULL);
851 	if (hugepte_cache == NULL)
852 		panic("%s: Unable to create kmem cache for hugeptes\n",
853 		      __func__);
854 
855 	/* Default hpage size = 4M */
856 	if (mmu_psize_defs[MMU_PAGE_4M].shift)
857 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
858 	else
859 		panic("%s: Unable to set default huge page size\n", __func__);
860 
861 
862 	return 0;
863 }
864 #else
865 static int __init hugetlbpage_init(void)
866 {
867 	int psize;
868 
869 	if (!mmu_has_feature(MMU_FTR_16M_PAGE))
870 		return -ENODEV;
871 
872 	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
873 		unsigned shift;
874 		unsigned pdshift;
875 
876 		if (!mmu_psize_defs[psize].shift)
877 			continue;
878 
879 		shift = mmu_psize_to_shift(psize);
880 
881 		if (add_huge_page_size(1ULL << shift) < 0)
882 			continue;
883 
884 		if (shift < PMD_SHIFT)
885 			pdshift = PMD_SHIFT;
886 		else if (shift < PUD_SHIFT)
887 			pdshift = PUD_SHIFT;
888 		else
889 			pdshift = PGDIR_SHIFT;
890 		/*
891 		 * if we have pdshift and shift value same, we don't
892 		 * use pgt cache for hugepd.
893 		 */
894 		if (pdshift != shift) {
895 			pgtable_cache_add(pdshift - shift, NULL);
896 			if (!PGT_CACHE(pdshift - shift))
897 				panic("hugetlbpage_init(): could not create "
898 				      "pgtable cache for %d bit pagesize\n", shift);
899 		}
900 	}
901 
902 	/* Set default large page size. Currently, we pick 16M or 1M
903 	 * depending on what is available
904 	 */
905 	if (mmu_psize_defs[MMU_PAGE_16M].shift)
906 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
907 	else if (mmu_psize_defs[MMU_PAGE_1M].shift)
908 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
909 
910 	return 0;
911 }
912 #endif
913 module_init(hugetlbpage_init);
914 
915 void flush_dcache_icache_hugepage(struct page *page)
916 {
917 	int i;
918 	void *start;
919 
920 	BUG_ON(!PageCompound(page));
921 
922 	for (i = 0; i < (1UL << compound_order(page)); i++) {
923 		if (!PageHighMem(page)) {
924 			__flush_dcache_icache(page_address(page+i));
925 		} else {
926 			start = kmap_atomic(page+i);
927 			__flush_dcache_icache(start);
928 			kunmap_atomic(start);
929 		}
930 	}
931 }
932 
933 #endif /* CONFIG_HUGETLB_PAGE */
934 
935 /*
936  * We have 4 cases for pgds and pmds:
937  * (1) invalid (all zeroes)
938  * (2) pointer to next table, as normal; bottom 6 bits == 0
939  * (3) leaf pte for huge page, bottom two bits != 00
940  * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
941  *
942  * So long as we atomically load page table pointers we are safe against teardown,
943  * we can follow the address down to the the page and take a ref on it.
944  */
945 
946 pte_t *find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea, unsigned *shift)
947 {
948 	pgd_t pgd, *pgdp;
949 	pud_t pud, *pudp;
950 	pmd_t pmd, *pmdp;
951 	pte_t *ret_pte;
952 	hugepd_t *hpdp = NULL;
953 	unsigned pdshift = PGDIR_SHIFT;
954 
955 	if (shift)
956 		*shift = 0;
957 
958 	pgdp = pgdir + pgd_index(ea);
959 	pgd  = ACCESS_ONCE(*pgdp);
960 	/*
961 	 * Always operate on the local stack value. This make sure the
962 	 * value don't get updated by a parallel THP split/collapse,
963 	 * page fault or a page unmap. The return pte_t * is still not
964 	 * stable. So should be checked there for above conditions.
965 	 */
966 	if (pgd_none(pgd))
967 		return NULL;
968 	else if (pgd_huge(pgd)) {
969 		ret_pte = (pte_t *) pgdp;
970 		goto out;
971 	} else if (is_hugepd(__hugepd(pgd_val(pgd))))
972 		hpdp = (hugepd_t *)&pgd;
973 	else {
974 		/*
975 		 * Even if we end up with an unmap, the pgtable will not
976 		 * be freed, because we do an rcu free and here we are
977 		 * irq disabled
978 		 */
979 		pdshift = PUD_SHIFT;
980 		pudp = pud_offset(&pgd, ea);
981 		pud  = ACCESS_ONCE(*pudp);
982 
983 		if (pud_none(pud))
984 			return NULL;
985 		else if (pud_huge(pud)) {
986 			ret_pte = (pte_t *) pudp;
987 			goto out;
988 		} else if (is_hugepd(__hugepd(pud_val(pud))))
989 			hpdp = (hugepd_t *)&pud;
990 		else {
991 			pdshift = PMD_SHIFT;
992 			pmdp = pmd_offset(&pud, ea);
993 			pmd  = ACCESS_ONCE(*pmdp);
994 			/*
995 			 * A hugepage collapse is captured by pmd_none, because
996 			 * it mark the pmd none and do a hpte invalidate.
997 			 *
998 			 * A hugepage split is captured by pmd_trans_splitting
999 			 * because we mark the pmd trans splitting and do a
1000 			 * hpte invalidate
1001 			 *
1002 			 */
1003 			if (pmd_none(pmd) || pmd_trans_splitting(pmd))
1004 				return NULL;
1005 
1006 			if (pmd_huge(pmd) || pmd_large(pmd)) {
1007 				ret_pte = (pte_t *) pmdp;
1008 				goto out;
1009 			} else if (is_hugepd(__hugepd(pmd_val(pmd))))
1010 				hpdp = (hugepd_t *)&pmd;
1011 			else
1012 				return pte_offset_kernel(&pmd, ea);
1013 		}
1014 	}
1015 	if (!hpdp)
1016 		return NULL;
1017 
1018 	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
1019 	pdshift = hugepd_shift(*hpdp);
1020 out:
1021 	if (shift)
1022 		*shift = pdshift;
1023 	return ret_pte;
1024 }
1025 EXPORT_SYMBOL_GPL(find_linux_pte_or_hugepte);
1026 
1027 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1028 		unsigned long end, int write, struct page **pages, int *nr)
1029 {
1030 	unsigned long mask;
1031 	unsigned long pte_end;
1032 	struct page *head, *page, *tail;
1033 	pte_t pte;
1034 	int refs;
1035 
1036 	pte_end = (addr + sz) & ~(sz-1);
1037 	if (pte_end < end)
1038 		end = pte_end;
1039 
1040 	pte = ACCESS_ONCE(*ptep);
1041 	mask = _PAGE_PRESENT | _PAGE_USER;
1042 	if (write)
1043 		mask |= _PAGE_RW;
1044 
1045 	if ((pte_val(pte) & mask) != mask)
1046 		return 0;
1047 
1048 	/* hugepages are never "special" */
1049 	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1050 
1051 	refs = 0;
1052 	head = pte_page(pte);
1053 
1054 	page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1055 	tail = page;
1056 	do {
1057 		VM_BUG_ON(compound_head(page) != head);
1058 		pages[*nr] = page;
1059 		(*nr)++;
1060 		page++;
1061 		refs++;
1062 	} while (addr += PAGE_SIZE, addr != end);
1063 
1064 	if (!page_cache_add_speculative(head, refs)) {
1065 		*nr -= refs;
1066 		return 0;
1067 	}
1068 
1069 	if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1070 		/* Could be optimized better */
1071 		*nr -= refs;
1072 		while (refs--)
1073 			put_page(head);
1074 		return 0;
1075 	}
1076 
1077 	/*
1078 	 * Any tail page need their mapcount reference taken before we
1079 	 * return.
1080 	 */
1081 	while (refs--) {
1082 		if (PageTail(tail))
1083 			get_huge_page_tail(tail);
1084 		tail++;
1085 	}
1086 
1087 	return 1;
1088 }
1089