xref: /linux/arch/powerpc/mm/hugetlbpage.c (revision 10accd2e6890b57db8e717e9aee91b791f90fe14)
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 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
57 {
58 	/* Only called for hugetlbfs pages, hence can ignore THP */
59 	return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL, NULL);
60 }
61 
62 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
63 			   unsigned long address, unsigned pdshift, unsigned pshift)
64 {
65 	struct kmem_cache *cachep;
66 	pte_t *new;
67 
68 #ifdef CONFIG_PPC_FSL_BOOK3E
69 	int i;
70 	int num_hugepd = 1 << (pshift - pdshift);
71 	cachep = hugepte_cache;
72 #else
73 	cachep = PGT_CACHE(pdshift - pshift);
74 #endif
75 
76 	new = kmem_cache_zalloc(cachep, GFP_KERNEL);
77 
78 	BUG_ON(pshift > HUGEPD_SHIFT_MASK);
79 	BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
80 
81 	if (! new)
82 		return -ENOMEM;
83 
84 	/*
85 	 * Make sure other cpus find the hugepd set only after a
86 	 * properly initialized page table is visible to them.
87 	 * For more details look for comment in __pte_alloc().
88 	 */
89 	smp_wmb();
90 
91 	spin_lock(&mm->page_table_lock);
92 #ifdef CONFIG_PPC_FSL_BOOK3E
93 	/*
94 	 * We have multiple higher-level entries that point to the same
95 	 * actual pte location.  Fill in each as we go and backtrack on error.
96 	 * We need all of these so the DTLB pgtable walk code can find the
97 	 * right higher-level entry without knowing if it's a hugepage or not.
98 	 */
99 	for (i = 0; i < num_hugepd; i++, hpdp++) {
100 		if (unlikely(!hugepd_none(*hpdp)))
101 			break;
102 		else
103 			/* We use the old format for PPC_FSL_BOOK3E */
104 			hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
105 	}
106 	/* If we bailed from the for loop early, an error occurred, clean up */
107 	if (i < num_hugepd) {
108 		for (i = i - 1 ; i >= 0; i--, hpdp--)
109 			hpdp->pd = 0;
110 		kmem_cache_free(cachep, new);
111 	}
112 #else
113 	if (!hugepd_none(*hpdp))
114 		kmem_cache_free(cachep, new);
115 	else {
116 #ifdef CONFIG_PPC_BOOK3S_64
117 		hpdp->pd = __pa(new) | (shift_to_mmu_psize(pshift) << 2);
118 #else
119 		hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
120 #endif
121 	}
122 #endif
123 	spin_unlock(&mm->page_table_lock);
124 	return 0;
125 }
126 
127 /*
128  * These macros define how to determine which level of the page table holds
129  * the hpdp.
130  */
131 #ifdef CONFIG_PPC_FSL_BOOK3E
132 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
133 #define HUGEPD_PUD_SHIFT PUD_SHIFT
134 #else
135 #define HUGEPD_PGD_SHIFT PUD_SHIFT
136 #define HUGEPD_PUD_SHIFT PMD_SHIFT
137 #endif
138 
139 #ifdef CONFIG_PPC_BOOK3S_64
140 /*
141  * At this point we do the placement change only for BOOK3S 64. This would
142  * possibly work on other subarchs.
143  */
144 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
145 {
146 	pgd_t *pg;
147 	pud_t *pu;
148 	pmd_t *pm;
149 	hugepd_t *hpdp = NULL;
150 	unsigned pshift = __ffs(sz);
151 	unsigned pdshift = PGDIR_SHIFT;
152 
153 	addr &= ~(sz-1);
154 	pg = pgd_offset(mm, addr);
155 
156 	if (pshift == PGDIR_SHIFT)
157 		/* 16GB huge page */
158 		return (pte_t *) pg;
159 	else if (pshift > PUD_SHIFT)
160 		/*
161 		 * We need to use hugepd table
162 		 */
163 		hpdp = (hugepd_t *)pg;
164 	else {
165 		pdshift = PUD_SHIFT;
166 		pu = pud_alloc(mm, pg, addr);
167 		if (pshift == PUD_SHIFT)
168 			return (pte_t *)pu;
169 		else if (pshift > PMD_SHIFT)
170 			hpdp = (hugepd_t *)pu;
171 		else {
172 			pdshift = PMD_SHIFT;
173 			pm = pmd_alloc(mm, pu, addr);
174 			if (pshift == PMD_SHIFT)
175 				/* 16MB hugepage */
176 				return (pte_t *)pm;
177 			else
178 				hpdp = (hugepd_t *)pm;
179 		}
180 	}
181 	if (!hpdp)
182 		return NULL;
183 
184 	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
185 
186 	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
187 		return NULL;
188 
189 	return hugepte_offset(*hpdp, addr, pdshift);
190 }
191 
192 #else
193 
194 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
195 {
196 	pgd_t *pg;
197 	pud_t *pu;
198 	pmd_t *pm;
199 	hugepd_t *hpdp = NULL;
200 	unsigned pshift = __ffs(sz);
201 	unsigned pdshift = PGDIR_SHIFT;
202 
203 	addr &= ~(sz-1);
204 
205 	pg = pgd_offset(mm, addr);
206 
207 	if (pshift >= HUGEPD_PGD_SHIFT) {
208 		hpdp = (hugepd_t *)pg;
209 	} else {
210 		pdshift = PUD_SHIFT;
211 		pu = pud_alloc(mm, pg, addr);
212 		if (pshift >= HUGEPD_PUD_SHIFT) {
213 			hpdp = (hugepd_t *)pu;
214 		} else {
215 			pdshift = PMD_SHIFT;
216 			pm = pmd_alloc(mm, pu, addr);
217 			hpdp = (hugepd_t *)pm;
218 		}
219 	}
220 
221 	if (!hpdp)
222 		return NULL;
223 
224 	BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
225 
226 	if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
227 		return NULL;
228 
229 	return hugepte_offset(*hpdp, addr, pdshift);
230 }
231 #endif
232 
233 #ifdef CONFIG_PPC_FSL_BOOK3E
234 /* Build list of addresses of gigantic pages.  This function is used in early
235  * boot before the buddy allocator is setup.
236  */
237 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
238 {
239 	unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
240 	int i;
241 
242 	if (addr == 0)
243 		return;
244 
245 	gpage_freearray[idx].nr_gpages = number_of_pages;
246 
247 	for (i = 0; i < number_of_pages; i++) {
248 		gpage_freearray[idx].gpage_list[i] = addr;
249 		addr += page_size;
250 	}
251 }
252 
253 /*
254  * Moves the gigantic page addresses from the temporary list to the
255  * huge_boot_pages list.
256  */
257 int alloc_bootmem_huge_page(struct hstate *hstate)
258 {
259 	struct huge_bootmem_page *m;
260 	int idx = shift_to_mmu_psize(huge_page_shift(hstate));
261 	int nr_gpages = gpage_freearray[idx].nr_gpages;
262 
263 	if (nr_gpages == 0)
264 		return 0;
265 
266 #ifdef CONFIG_HIGHMEM
267 	/*
268 	 * If gpages can be in highmem we can't use the trick of storing the
269 	 * data structure in the page; allocate space for this
270 	 */
271 	m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
272 	m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
273 #else
274 	m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
275 #endif
276 
277 	list_add(&m->list, &huge_boot_pages);
278 	gpage_freearray[idx].nr_gpages = nr_gpages;
279 	gpage_freearray[idx].gpage_list[nr_gpages] = 0;
280 	m->hstate = hstate;
281 
282 	return 1;
283 }
284 /*
285  * Scan the command line hugepagesz= options for gigantic pages; store those in
286  * a list that we use to allocate the memory once all options are parsed.
287  */
288 
289 unsigned long gpage_npages[MMU_PAGE_COUNT];
290 
291 static int __init do_gpage_early_setup(char *param, char *val,
292 				       const char *unused, void *arg)
293 {
294 	static phys_addr_t size;
295 	unsigned long npages;
296 
297 	/*
298 	 * The hugepagesz and hugepages cmdline options are interleaved.  We
299 	 * use the size variable to keep track of whether or not this was done
300 	 * properly and skip over instances where it is incorrect.  Other
301 	 * command-line parsing code will issue warnings, so we don't need to.
302 	 *
303 	 */
304 	if ((strcmp(param, "default_hugepagesz") == 0) ||
305 	    (strcmp(param, "hugepagesz") == 0)) {
306 		size = memparse(val, NULL);
307 	} else if (strcmp(param, "hugepages") == 0) {
308 		if (size != 0) {
309 			if (sscanf(val, "%lu", &npages) <= 0)
310 				npages = 0;
311 			if (npages > MAX_NUMBER_GPAGES) {
312 				pr_warn("MMU: %lu pages requested for page "
313 					"size %llu KB, limiting to "
314 					__stringify(MAX_NUMBER_GPAGES) "\n",
315 					npages, size / 1024);
316 				npages = MAX_NUMBER_GPAGES;
317 			}
318 			gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
319 			size = 0;
320 		}
321 	}
322 	return 0;
323 }
324 
325 
326 /*
327  * This function allocates physical space for pages that are larger than the
328  * buddy allocator can handle.  We want to allocate these in highmem because
329  * the amount of lowmem is limited.  This means that this function MUST be
330  * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
331  * allocate to grab highmem.
332  */
333 void __init reserve_hugetlb_gpages(void)
334 {
335 	static __initdata char cmdline[COMMAND_LINE_SIZE];
336 	phys_addr_t size, base;
337 	int i;
338 
339 	strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
340 	parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
341 			NULL, &do_gpage_early_setup);
342 
343 	/*
344 	 * Walk gpage list in reverse, allocating larger page sizes first.
345 	 * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
346 	 * When we reach the point in the list where pages are no longer
347 	 * considered gpages, we're done.
348 	 */
349 	for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
350 		if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
351 			continue;
352 		else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
353 			break;
354 
355 		size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
356 		base = memblock_alloc_base(size * gpage_npages[i], size,
357 					   MEMBLOCK_ALLOC_ANYWHERE);
358 		add_gpage(base, size, gpage_npages[i]);
359 	}
360 }
361 
362 #else /* !PPC_FSL_BOOK3E */
363 
364 /* Build list of addresses of gigantic pages.  This function is used in early
365  * boot before the buddy allocator is setup.
366  */
367 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
368 {
369 	if (!addr)
370 		return;
371 	while (number_of_pages > 0) {
372 		gpage_freearray[nr_gpages] = addr;
373 		nr_gpages++;
374 		number_of_pages--;
375 		addr += page_size;
376 	}
377 }
378 
379 /* Moves the gigantic page addresses from the temporary list to the
380  * huge_boot_pages list.
381  */
382 int alloc_bootmem_huge_page(struct hstate *hstate)
383 {
384 	struct huge_bootmem_page *m;
385 	if (nr_gpages == 0)
386 		return 0;
387 	m = phys_to_virt(gpage_freearray[--nr_gpages]);
388 	gpage_freearray[nr_gpages] = 0;
389 	list_add(&m->list, &huge_boot_pages);
390 	m->hstate = hstate;
391 	return 1;
392 }
393 #endif
394 
395 #ifdef CONFIG_PPC_FSL_BOOK3E
396 #define HUGEPD_FREELIST_SIZE \
397 	((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
398 
399 struct hugepd_freelist {
400 	struct rcu_head	rcu;
401 	unsigned int index;
402 	void *ptes[0];
403 };
404 
405 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
406 
407 static void hugepd_free_rcu_callback(struct rcu_head *head)
408 {
409 	struct hugepd_freelist *batch =
410 		container_of(head, struct hugepd_freelist, rcu);
411 	unsigned int i;
412 
413 	for (i = 0; i < batch->index; i++)
414 		kmem_cache_free(hugepte_cache, batch->ptes[i]);
415 
416 	free_page((unsigned long)batch);
417 }
418 
419 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
420 {
421 	struct hugepd_freelist **batchp;
422 
423 	batchp = &get_cpu_var(hugepd_freelist_cur);
424 
425 	if (atomic_read(&tlb->mm->mm_users) < 2 ||
426 	    cpumask_equal(mm_cpumask(tlb->mm),
427 			  cpumask_of(smp_processor_id()))) {
428 		kmem_cache_free(hugepte_cache, hugepte);
429 		put_cpu_var(hugepd_freelist_cur);
430 		return;
431 	}
432 
433 	if (*batchp == NULL) {
434 		*batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
435 		(*batchp)->index = 0;
436 	}
437 
438 	(*batchp)->ptes[(*batchp)->index++] = hugepte;
439 	if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
440 		call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
441 		*batchp = NULL;
442 	}
443 	put_cpu_var(hugepd_freelist_cur);
444 }
445 #endif
446 
447 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
448 			      unsigned long start, unsigned long end,
449 			      unsigned long floor, unsigned long ceiling)
450 {
451 	pte_t *hugepte = hugepd_page(*hpdp);
452 	int i;
453 
454 	unsigned long pdmask = ~((1UL << pdshift) - 1);
455 	unsigned int num_hugepd = 1;
456 
457 #ifdef CONFIG_PPC_FSL_BOOK3E
458 	/* Note: On fsl the hpdp may be the first of several */
459 	num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
460 #else
461 	unsigned int shift = hugepd_shift(*hpdp);
462 #endif
463 
464 	start &= pdmask;
465 	if (start < floor)
466 		return;
467 	if (ceiling) {
468 		ceiling &= pdmask;
469 		if (! ceiling)
470 			return;
471 	}
472 	if (end - 1 > ceiling - 1)
473 		return;
474 
475 	for (i = 0; i < num_hugepd; i++, hpdp++)
476 		hpdp->pd = 0;
477 
478 #ifdef CONFIG_PPC_FSL_BOOK3E
479 	hugepd_free(tlb, hugepte);
480 #else
481 	pgtable_free_tlb(tlb, hugepte, pdshift - shift);
482 #endif
483 }
484 
485 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
486 				   unsigned long addr, unsigned long end,
487 				   unsigned long floor, unsigned long ceiling)
488 {
489 	pmd_t *pmd;
490 	unsigned long next;
491 	unsigned long start;
492 
493 	start = addr;
494 	do {
495 		pmd = pmd_offset(pud, addr);
496 		next = pmd_addr_end(addr, end);
497 		if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
498 			/*
499 			 * if it is not hugepd pointer, we should already find
500 			 * it cleared.
501 			 */
502 			WARN_ON(!pmd_none_or_clear_bad(pmd));
503 			continue;
504 		}
505 #ifdef CONFIG_PPC_FSL_BOOK3E
506 		/*
507 		 * Increment next by the size of the huge mapping since
508 		 * there may be more than one entry at this level for a
509 		 * single hugepage, but all of them point to
510 		 * the same kmem cache that holds the hugepte.
511 		 */
512 		next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
513 #endif
514 		free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
515 				  addr, next, floor, ceiling);
516 	} while (addr = next, addr != end);
517 
518 	start &= PUD_MASK;
519 	if (start < floor)
520 		return;
521 	if (ceiling) {
522 		ceiling &= PUD_MASK;
523 		if (!ceiling)
524 			return;
525 	}
526 	if (end - 1 > ceiling - 1)
527 		return;
528 
529 	pmd = pmd_offset(pud, start);
530 	pud_clear(pud);
531 	pmd_free_tlb(tlb, pmd, start);
532 	mm_dec_nr_pmds(tlb->mm);
533 }
534 
535 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
536 				   unsigned long addr, unsigned long end,
537 				   unsigned long floor, unsigned long ceiling)
538 {
539 	pud_t *pud;
540 	unsigned long next;
541 	unsigned long start;
542 
543 	start = addr;
544 	do {
545 		pud = pud_offset(pgd, addr);
546 		next = pud_addr_end(addr, end);
547 		if (!is_hugepd(__hugepd(pud_val(*pud)))) {
548 			if (pud_none_or_clear_bad(pud))
549 				continue;
550 			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
551 					       ceiling);
552 		} else {
553 #ifdef CONFIG_PPC_FSL_BOOK3E
554 			/*
555 			 * Increment next by the size of the huge mapping since
556 			 * there may be more than one entry at this level for a
557 			 * single hugepage, but all of them point to
558 			 * the same kmem cache that holds the hugepte.
559 			 */
560 			next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
561 #endif
562 			free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
563 					  addr, next, floor, ceiling);
564 		}
565 	} while (addr = next, addr != end);
566 
567 	start &= PGDIR_MASK;
568 	if (start < floor)
569 		return;
570 	if (ceiling) {
571 		ceiling &= PGDIR_MASK;
572 		if (!ceiling)
573 			return;
574 	}
575 	if (end - 1 > ceiling - 1)
576 		return;
577 
578 	pud = pud_offset(pgd, start);
579 	pgd_clear(pgd);
580 	pud_free_tlb(tlb, pud, start);
581 }
582 
583 /*
584  * This function frees user-level page tables of a process.
585  */
586 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
587 			    unsigned long addr, unsigned long end,
588 			    unsigned long floor, unsigned long ceiling)
589 {
590 	pgd_t *pgd;
591 	unsigned long next;
592 
593 	/*
594 	 * Because there are a number of different possible pagetable
595 	 * layouts for hugepage ranges, we limit knowledge of how
596 	 * things should be laid out to the allocation path
597 	 * (huge_pte_alloc(), above).  Everything else works out the
598 	 * structure as it goes from information in the hugepd
599 	 * pointers.  That means that we can't here use the
600 	 * optimization used in the normal page free_pgd_range(), of
601 	 * checking whether we're actually covering a large enough
602 	 * range to have to do anything at the top level of the walk
603 	 * instead of at the bottom.
604 	 *
605 	 * To make sense of this, you should probably go read the big
606 	 * block comment at the top of the normal free_pgd_range(),
607 	 * too.
608 	 */
609 
610 	do {
611 		next = pgd_addr_end(addr, end);
612 		pgd = pgd_offset(tlb->mm, addr);
613 		if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
614 			if (pgd_none_or_clear_bad(pgd))
615 				continue;
616 			hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
617 		} else {
618 #ifdef CONFIG_PPC_FSL_BOOK3E
619 			/*
620 			 * Increment next by the size of the huge mapping since
621 			 * there may be more than one entry at the pgd level
622 			 * for a single hugepage, but all of them point to the
623 			 * same kmem cache that holds the hugepte.
624 			 */
625 			next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
626 #endif
627 			free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
628 					  addr, next, floor, ceiling);
629 		}
630 	} while (addr = next, addr != end);
631 }
632 
633 /*
634  * We are holding mmap_sem, so a parallel huge page collapse cannot run.
635  * To prevent hugepage split, disable irq.
636  */
637 struct page *
638 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
639 {
640 	bool is_thp;
641 	pte_t *ptep, pte;
642 	unsigned shift;
643 	unsigned long mask, flags;
644 	struct page *page = ERR_PTR(-EINVAL);
645 
646 	local_irq_save(flags);
647 	ptep = find_linux_pte_or_hugepte(mm->pgd, address, &is_thp, &shift);
648 	if (!ptep)
649 		goto no_page;
650 	pte = READ_ONCE(*ptep);
651 	/*
652 	 * Verify it is a huge page else bail.
653 	 * Transparent hugepages are handled by generic code. We can skip them
654 	 * here.
655 	 */
656 	if (!shift || is_thp)
657 		goto no_page;
658 
659 	if (!pte_present(pte)) {
660 		page = NULL;
661 		goto no_page;
662 	}
663 	mask = (1UL << shift) - 1;
664 	page = pte_page(pte);
665 	if (page)
666 		page += (address & mask) / PAGE_SIZE;
667 
668 no_page:
669 	local_irq_restore(flags);
670 	return page;
671 }
672 
673 struct page *
674 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
675 		pmd_t *pmd, int write)
676 {
677 	BUG();
678 	return NULL;
679 }
680 
681 struct page *
682 follow_huge_pud(struct mm_struct *mm, unsigned long address,
683 		pud_t *pud, int write)
684 {
685 	BUG();
686 	return NULL;
687 }
688 
689 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
690 				      unsigned long sz)
691 {
692 	unsigned long __boundary = (addr + sz) & ~(sz-1);
693 	return (__boundary - 1 < end - 1) ? __boundary : end;
694 }
695 
696 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
697 		unsigned long end, int write, struct page **pages, int *nr)
698 {
699 	pte_t *ptep;
700 	unsigned long sz = 1UL << hugepd_shift(hugepd);
701 	unsigned long next;
702 
703 	ptep = hugepte_offset(hugepd, addr, pdshift);
704 	do {
705 		next = hugepte_addr_end(addr, end, sz);
706 		if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
707 			return 0;
708 	} while (ptep++, addr = next, addr != end);
709 
710 	return 1;
711 }
712 
713 #ifdef CONFIG_PPC_MM_SLICES
714 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
715 					unsigned long len, unsigned long pgoff,
716 					unsigned long flags)
717 {
718 	struct hstate *hstate = hstate_file(file);
719 	int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
720 
721 	if (radix_enabled())
722 		return radix__hugetlb_get_unmapped_area(file, addr, len,
723 						       pgoff, flags);
724 	return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
725 }
726 #endif
727 
728 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
729 {
730 #ifdef CONFIG_PPC_MM_SLICES
731 	unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
732 	/* With radix we don't use slice, so derive it from vma*/
733 	if (!radix_enabled())
734 		return 1UL << mmu_psize_to_shift(psize);
735 #endif
736 	if (!is_vm_hugetlb_page(vma))
737 		return PAGE_SIZE;
738 
739 	return huge_page_size(hstate_vma(vma));
740 }
741 
742 static inline bool is_power_of_4(unsigned long x)
743 {
744 	if (is_power_of_2(x))
745 		return (__ilog2(x) % 2) ? false : true;
746 	return false;
747 }
748 
749 static int __init add_huge_page_size(unsigned long long size)
750 {
751 	int shift = __ffs(size);
752 	int mmu_psize;
753 
754 	/* Check that it is a page size supported by the hardware and
755 	 * that it fits within pagetable and slice limits. */
756 #ifdef CONFIG_PPC_FSL_BOOK3E
757 	if ((size < PAGE_SIZE) || !is_power_of_4(size))
758 		return -EINVAL;
759 #else
760 	if (!is_power_of_2(size)
761 	    || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
762 		return -EINVAL;
763 #endif
764 
765 	if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
766 		return -EINVAL;
767 
768 	BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
769 
770 	/* Return if huge page size has already been setup */
771 	if (size_to_hstate(size))
772 		return 0;
773 
774 	hugetlb_add_hstate(shift - PAGE_SHIFT);
775 
776 	return 0;
777 }
778 
779 static int __init hugepage_setup_sz(char *str)
780 {
781 	unsigned long long size;
782 
783 	size = memparse(str, &str);
784 
785 	if (add_huge_page_size(size) != 0) {
786 		hugetlb_bad_size();
787 		pr_err("Invalid huge page size specified(%llu)\n", size);
788 	}
789 
790 	return 1;
791 }
792 __setup("hugepagesz=", hugepage_setup_sz);
793 
794 #ifdef CONFIG_PPC_FSL_BOOK3E
795 struct kmem_cache *hugepte_cache;
796 static int __init hugetlbpage_init(void)
797 {
798 	int psize;
799 
800 	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
801 		unsigned shift;
802 
803 		if (!mmu_psize_defs[psize].shift)
804 			continue;
805 
806 		shift = mmu_psize_to_shift(psize);
807 
808 		/* Don't treat normal page sizes as huge... */
809 		if (shift != PAGE_SHIFT)
810 			if (add_huge_page_size(1ULL << shift) < 0)
811 				continue;
812 	}
813 
814 	/*
815 	 * Create a kmem cache for hugeptes.  The bottom bits in the pte have
816 	 * size information encoded in them, so align them to allow this
817 	 */
818 	hugepte_cache =  kmem_cache_create("hugepte-cache", sizeof(pte_t),
819 					   HUGEPD_SHIFT_MASK + 1, 0, NULL);
820 	if (hugepte_cache == NULL)
821 		panic("%s: Unable to create kmem cache for hugeptes\n",
822 		      __func__);
823 
824 	/* Default hpage size = 4M */
825 	if (mmu_psize_defs[MMU_PAGE_4M].shift)
826 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
827 	else
828 		panic("%s: Unable to set default huge page size\n", __func__);
829 
830 
831 	return 0;
832 }
833 #else
834 static int __init hugetlbpage_init(void)
835 {
836 	int psize;
837 
838 	if (!radix_enabled() && !mmu_has_feature(MMU_FTR_16M_PAGE))
839 		return -ENODEV;
840 
841 	for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
842 		unsigned shift;
843 		unsigned pdshift;
844 
845 		if (!mmu_psize_defs[psize].shift)
846 			continue;
847 
848 		shift = mmu_psize_to_shift(psize);
849 
850 		if (add_huge_page_size(1ULL << shift) < 0)
851 			continue;
852 
853 		if (shift < PMD_SHIFT)
854 			pdshift = PMD_SHIFT;
855 		else if (shift < PUD_SHIFT)
856 			pdshift = PUD_SHIFT;
857 		else
858 			pdshift = PGDIR_SHIFT;
859 		/*
860 		 * if we have pdshift and shift value same, we don't
861 		 * use pgt cache for hugepd.
862 		 */
863 		if (pdshift != shift) {
864 			pgtable_cache_add(pdshift - shift, NULL);
865 			if (!PGT_CACHE(pdshift - shift))
866 				panic("hugetlbpage_init(): could not create "
867 				      "pgtable cache for %d bit pagesize\n", shift);
868 		}
869 	}
870 
871 	/* Set default large page size. Currently, we pick 16M or 1M
872 	 * depending on what is available
873 	 */
874 	if (mmu_psize_defs[MMU_PAGE_16M].shift)
875 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
876 	else if (mmu_psize_defs[MMU_PAGE_1M].shift)
877 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
878 	else if (mmu_psize_defs[MMU_PAGE_2M].shift)
879 		HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_2M].shift;
880 
881 
882 	return 0;
883 }
884 #endif
885 arch_initcall(hugetlbpage_init);
886 
887 void flush_dcache_icache_hugepage(struct page *page)
888 {
889 	int i;
890 	void *start;
891 
892 	BUG_ON(!PageCompound(page));
893 
894 	for (i = 0; i < (1UL << compound_order(page)); i++) {
895 		if (!PageHighMem(page)) {
896 			__flush_dcache_icache(page_address(page+i));
897 		} else {
898 			start = kmap_atomic(page+i);
899 			__flush_dcache_icache(start);
900 			kunmap_atomic(start);
901 		}
902 	}
903 }
904 
905 #endif /* CONFIG_HUGETLB_PAGE */
906 
907 /*
908  * We have 4 cases for pgds and pmds:
909  * (1) invalid (all zeroes)
910  * (2) pointer to next table, as normal; bottom 6 bits == 0
911  * (3) leaf pte for huge page _PAGE_PTE set
912  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
913  *
914  * So long as we atomically load page table pointers we are safe against teardown,
915  * we can follow the address down to the the page and take a ref on it.
916  * This function need to be called with interrupts disabled. We use this variant
917  * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
918  */
919 
920 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
921 				   bool *is_thp, unsigned *shift)
922 {
923 	pgd_t pgd, *pgdp;
924 	pud_t pud, *pudp;
925 	pmd_t pmd, *pmdp;
926 	pte_t *ret_pte;
927 	hugepd_t *hpdp = NULL;
928 	unsigned pdshift = PGDIR_SHIFT;
929 
930 	if (shift)
931 		*shift = 0;
932 
933 	if (is_thp)
934 		*is_thp = false;
935 
936 	pgdp = pgdir + pgd_index(ea);
937 	pgd  = READ_ONCE(*pgdp);
938 	/*
939 	 * Always operate on the local stack value. This make sure the
940 	 * value don't get updated by a parallel THP split/collapse,
941 	 * page fault or a page unmap. The return pte_t * is still not
942 	 * stable. So should be checked there for above conditions.
943 	 */
944 	if (pgd_none(pgd))
945 		return NULL;
946 	else if (pgd_huge(pgd)) {
947 		ret_pte = (pte_t *) pgdp;
948 		goto out;
949 	} else if (is_hugepd(__hugepd(pgd_val(pgd))))
950 		hpdp = (hugepd_t *)&pgd;
951 	else {
952 		/*
953 		 * Even if we end up with an unmap, the pgtable will not
954 		 * be freed, because we do an rcu free and here we are
955 		 * irq disabled
956 		 */
957 		pdshift = PUD_SHIFT;
958 		pudp = pud_offset(&pgd, ea);
959 		pud  = READ_ONCE(*pudp);
960 
961 		if (pud_none(pud))
962 			return NULL;
963 		else if (pud_huge(pud)) {
964 			ret_pte = (pte_t *) pudp;
965 			goto out;
966 		} else if (is_hugepd(__hugepd(pud_val(pud))))
967 			hpdp = (hugepd_t *)&pud;
968 		else {
969 			pdshift = PMD_SHIFT;
970 			pmdp = pmd_offset(&pud, ea);
971 			pmd  = READ_ONCE(*pmdp);
972 			/*
973 			 * A hugepage collapse is captured by pmd_none, because
974 			 * it mark the pmd none and do a hpte invalidate.
975 			 */
976 			if (pmd_none(pmd))
977 				return NULL;
978 
979 			if (pmd_trans_huge(pmd)) {
980 				if (is_thp)
981 					*is_thp = true;
982 				ret_pte = (pte_t *) pmdp;
983 				goto out;
984 			}
985 
986 			if (pmd_huge(pmd)) {
987 				ret_pte = (pte_t *) pmdp;
988 				goto out;
989 			} else if (is_hugepd(__hugepd(pmd_val(pmd))))
990 				hpdp = (hugepd_t *)&pmd;
991 			else
992 				return pte_offset_kernel(&pmd, ea);
993 		}
994 	}
995 	if (!hpdp)
996 		return NULL;
997 
998 	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
999 	pdshift = hugepd_shift(*hpdp);
1000 out:
1001 	if (shift)
1002 		*shift = pdshift;
1003 	return ret_pte;
1004 }
1005 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
1006 
1007 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1008 		unsigned long end, int write, struct page **pages, int *nr)
1009 {
1010 	unsigned long mask;
1011 	unsigned long pte_end;
1012 	struct page *head, *page;
1013 	pte_t pte;
1014 	int refs;
1015 
1016 	pte_end = (addr + sz) & ~(sz-1);
1017 	if (pte_end < end)
1018 		end = pte_end;
1019 
1020 	pte = READ_ONCE(*ptep);
1021 	mask = _PAGE_PRESENT | _PAGE_READ;
1022 	if (write)
1023 		mask |= _PAGE_WRITE;
1024 
1025 	if ((pte_val(pte) & mask) != mask)
1026 		return 0;
1027 
1028 	/* hugepages are never "special" */
1029 	VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1030 
1031 	refs = 0;
1032 	head = pte_page(pte);
1033 
1034 	page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1035 	do {
1036 		VM_BUG_ON(compound_head(page) != head);
1037 		pages[*nr] = page;
1038 		(*nr)++;
1039 		page++;
1040 		refs++;
1041 	} while (addr += PAGE_SIZE, addr != end);
1042 
1043 	if (!page_cache_add_speculative(head, refs)) {
1044 		*nr -= refs;
1045 		return 0;
1046 	}
1047 
1048 	if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1049 		/* Could be optimized better */
1050 		*nr -= refs;
1051 		while (refs--)
1052 			put_page(head);
1053 		return 0;
1054 	}
1055 
1056 	return 1;
1057 }
1058