xref: /openbmc/linux/arch/arm64/mm/hugetlbpage.c (revision 18afb028) 
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * arch/arm64/mm/hugetlbpage.c
4  *
5  * Copyright (C) 2013 Linaro Ltd.
6  *
7  * Based on arch/x86/mm/hugetlbpage.c.
8  */
9 
10 #include <linux/init.h>
11 #include <linux/fs.h>
12 #include <linux/mm.h>
13 #include <linux/hugetlb.h>
14 #include <linux/pagemap.h>
15 #include <linux/err.h>
16 #include <linux/sysctl.h>
17 #include <asm/mman.h>
18 #include <asm/tlb.h>
19 #include <asm/tlbflush.h>
20 
21 /*
22  * HugeTLB Support Matrix
23  *
24  * ---------------------------------------------------
25  * | Page Size | CONT PTE |  PMD  | CONT PMD |  PUD  |
26  * ---------------------------------------------------
27  * |     4K    |   64K    |   2M  |    32M   |   1G  |
28  * |    16K    |    2M    |  32M  |     1G   |       |
29  * |    64K    |    2M    | 512M  |    16G   |       |
30  * ---------------------------------------------------
31  */
32 
33 /*
34  * Reserve CMA areas for the largest supported gigantic
35  * huge page when requested. Any other smaller gigantic
36  * huge pages could still be served from those areas.
37  */
38 #ifdef CONFIG_CMA
39 void __init arm64_hugetlb_cma_reserve(void)
40 {
41 	int order;
42 
43 	if (pud_sect_supported())
44 		order = PUD_SHIFT - PAGE_SHIFT;
45 	else
46 		order = CONT_PMD_SHIFT - PAGE_SHIFT;
47 
48 	/*
49 	 * HugeTLB CMA reservation is required for gigantic
50 	 * huge pages which could not be allocated via the
51 	 * page allocator. Just warn if there is any change
52 	 * breaking this assumption.
53 	 */
54 	WARN_ON(order <= MAX_ORDER);
55 	hugetlb_cma_reserve(order);
56 }
57 #endif /* CONFIG_CMA */
58 
59 static bool __hugetlb_valid_size(unsigned long size)
60 {
61 	switch (size) {
62 #ifndef __PAGETABLE_PMD_FOLDED
63 	case PUD_SIZE:
64 		return pud_sect_supported();
65 #endif
66 	case CONT_PMD_SIZE:
67 	case PMD_SIZE:
68 	case CONT_PTE_SIZE:
69 		return true;
70 	}
71 
72 	return false;
73 }
74 
75 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
76 bool arch_hugetlb_migration_supported(struct hstate *h)
77 {
78 	size_t pagesize = huge_page_size(h);
79 
80 	if (!__hugetlb_valid_size(pagesize)) {
81 		pr_warn("%s: unrecognized huge page size 0x%lx\n",
82 			__func__, pagesize);
83 		return false;
84 	}
85 	return true;
86 }
87 #endif
88 
89 int pmd_huge(pmd_t pmd)
90 {
91 	return pmd_val(pmd) && !(pmd_val(pmd) & PMD_TABLE_BIT);
92 }
93 
94 int pud_huge(pud_t pud)
95 {
96 #ifndef __PAGETABLE_PMD_FOLDED
97 	return pud_val(pud) && !(pud_val(pud) & PUD_TABLE_BIT);
98 #else
99 	return 0;
100 #endif
101 }
102 
103 static int find_num_contig(struct mm_struct *mm, unsigned long addr,
104 			   pte_t *ptep, size_t *pgsize)
105 {
106 	pgd_t *pgdp = pgd_offset(mm, addr);
107 	p4d_t *p4dp;
108 	pud_t *pudp;
109 	pmd_t *pmdp;
110 
111 	*pgsize = PAGE_SIZE;
112 	p4dp = p4d_offset(pgdp, addr);
113 	pudp = pud_offset(p4dp, addr);
114 	pmdp = pmd_offset(pudp, addr);
115 	if ((pte_t *)pmdp == ptep) {
116 		*pgsize = PMD_SIZE;
117 		return CONT_PMDS;
118 	}
119 	return CONT_PTES;
120 }
121 
122 static inline int num_contig_ptes(unsigned long size, size_t *pgsize)
123 {
124 	int contig_ptes = 0;
125 
126 	*pgsize = size;
127 
128 	switch (size) {
129 #ifndef __PAGETABLE_PMD_FOLDED
130 	case PUD_SIZE:
131 		if (pud_sect_supported())
132 			contig_ptes = 1;
133 		break;
134 #endif
135 	case PMD_SIZE:
136 		contig_ptes = 1;
137 		break;
138 	case CONT_PMD_SIZE:
139 		*pgsize = PMD_SIZE;
140 		contig_ptes = CONT_PMDS;
141 		break;
142 	case CONT_PTE_SIZE:
143 		*pgsize = PAGE_SIZE;
144 		contig_ptes = CONT_PTES;
145 		break;
146 	}
147 
148 	return contig_ptes;
149 }
150 
151 pte_t huge_ptep_get(pte_t *ptep)
152 {
153 	int ncontig, i;
154 	size_t pgsize;
155 	pte_t orig_pte = ptep_get(ptep);
156 
157 	if (!pte_present(orig_pte) || !pte_cont(orig_pte))
158 		return orig_pte;
159 
160 	ncontig = num_contig_ptes(page_size(pte_page(orig_pte)), &pgsize);
161 	for (i = 0; i < ncontig; i++, ptep++) {
162 		pte_t pte = ptep_get(ptep);
163 
164 		if (pte_dirty(pte))
165 			orig_pte = pte_mkdirty(orig_pte);
166 
167 		if (pte_young(pte))
168 			orig_pte = pte_mkyoung(orig_pte);
169 	}
170 	return orig_pte;
171 }
172 
173 /*
174  * Changing some bits of contiguous entries requires us to follow a
175  * Break-Before-Make approach, breaking the whole contiguous set
176  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
177  * "Misprogramming of the Contiguous bit", page D4-1762.
178  *
179  * This helper performs the break step.
180  */
181 static pte_t get_clear_contig(struct mm_struct *mm,
182 			     unsigned long addr,
183 			     pte_t *ptep,
184 			     unsigned long pgsize,
185 			     unsigned long ncontig)
186 {
187 	pte_t orig_pte = ptep_get(ptep);
188 	unsigned long i;
189 
190 	for (i = 0; i < ncontig; i++, addr += pgsize, ptep++) {
191 		pte_t pte = ptep_get_and_clear(mm, addr, ptep);
192 
193 		/*
194 		 * If HW_AFDBM is enabled, then the HW could turn on
195 		 * the dirty or accessed bit for any page in the set,
196 		 * so check them all.
197 		 */
198 		if (pte_dirty(pte))
199 			orig_pte = pte_mkdirty(orig_pte);
200 
201 		if (pte_young(pte))
202 			orig_pte = pte_mkyoung(orig_pte);
203 	}
204 	return orig_pte;
205 }
206 
207 static pte_t get_clear_contig_flush(struct mm_struct *mm,
208 				    unsigned long addr,
209 				    pte_t *ptep,
210 				    unsigned long pgsize,
211 				    unsigned long ncontig)
212 {
213 	pte_t orig_pte = get_clear_contig(mm, addr, ptep, pgsize, ncontig);
214 	struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
215 
216 	flush_tlb_range(&vma, addr, addr + (pgsize * ncontig));
217 	return orig_pte;
218 }
219 
220 /*
221  * Changing some bits of contiguous entries requires us to follow a
222  * Break-Before-Make approach, breaking the whole contiguous set
223  * before we can change any entries. See ARM DDI 0487A.k_iss10775,
224  * "Misprogramming of the Contiguous bit", page D4-1762.
225  *
226  * This helper performs the break step for use cases where the
227  * original pte is not needed.
228  */
229 static void clear_flush(struct mm_struct *mm,
230 			     unsigned long addr,
231 			     pte_t *ptep,
232 			     unsigned long pgsize,
233 			     unsigned long ncontig)
234 {
235 	struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
236 	unsigned long i, saddr = addr;
237 
238 	for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
239 		ptep_clear(mm, addr, ptep);
240 
241 	flush_tlb_range(&vma, saddr, addr);
242 }
243 
244 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
245 			    pte_t *ptep, pte_t pte, unsigned long sz)
246 {
247 	size_t pgsize;
248 	int i;
249 	int ncontig;
250 	unsigned long pfn, dpfn;
251 	pgprot_t hugeprot;
252 
253 	ncontig = num_contig_ptes(sz, &pgsize);
254 
255 	if (!pte_present(pte)) {
256 		for (i = 0; i < ncontig; i++, ptep++, addr += pgsize)
257 			set_pte_at(mm, addr, ptep, pte);
258 		return;
259 	}
260 
261 	if (!pte_cont(pte)) {
262 		set_pte_at(mm, addr, ptep, pte);
263 		return;
264 	}
265 
266 	pfn = pte_pfn(pte);
267 	dpfn = pgsize >> PAGE_SHIFT;
268 	hugeprot = pte_pgprot(pte);
269 
270 	clear_flush(mm, addr, ptep, pgsize, ncontig);
271 
272 	for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
273 		set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
274 }
275 
276 pte_t *huge_pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
277 		      unsigned long addr, unsigned long sz)
278 {
279 	pgd_t *pgdp;
280 	p4d_t *p4dp;
281 	pud_t *pudp;
282 	pmd_t *pmdp;
283 	pte_t *ptep = NULL;
284 
285 	pgdp = pgd_offset(mm, addr);
286 	p4dp = p4d_offset(pgdp, addr);
287 	pudp = pud_alloc(mm, p4dp, addr);
288 	if (!pudp)
289 		return NULL;
290 
291 	if (sz == PUD_SIZE) {
292 		ptep = (pte_t *)pudp;
293 	} else if (sz == (CONT_PTE_SIZE)) {
294 		pmdp = pmd_alloc(mm, pudp, addr);
295 		if (!pmdp)
296 			return NULL;
297 
298 		WARN_ON(addr & (sz - 1));
299 		ptep = pte_alloc_huge(mm, pmdp, addr);
300 	} else if (sz == PMD_SIZE) {
301 		if (want_pmd_share(vma, addr) && pud_none(READ_ONCE(*pudp)))
302 			ptep = huge_pmd_share(mm, vma, addr, pudp);
303 		else
304 			ptep = (pte_t *)pmd_alloc(mm, pudp, addr);
305 	} else if (sz == (CONT_PMD_SIZE)) {
306 		pmdp = pmd_alloc(mm, pudp, addr);
307 		WARN_ON(addr & (sz - 1));
308 		return (pte_t *)pmdp;
309 	}
310 
311 	return ptep;
312 }
313 
314 pte_t *huge_pte_offset(struct mm_struct *mm,
315 		       unsigned long addr, unsigned long sz)
316 {
317 	pgd_t *pgdp;
318 	p4d_t *p4dp;
319 	pud_t *pudp, pud;
320 	pmd_t *pmdp, pmd;
321 
322 	pgdp = pgd_offset(mm, addr);
323 	if (!pgd_present(READ_ONCE(*pgdp)))
324 		return NULL;
325 
326 	p4dp = p4d_offset(pgdp, addr);
327 	if (!p4d_present(READ_ONCE(*p4dp)))
328 		return NULL;
329 
330 	pudp = pud_offset(p4dp, addr);
331 	pud = READ_ONCE(*pudp);
332 	if (sz != PUD_SIZE && pud_none(pud))
333 		return NULL;
334 	/* hugepage or swap? */
335 	if (pud_huge(pud) || !pud_present(pud))
336 		return (pte_t *)pudp;
337 	/* table; check the next level */
338 
339 	if (sz == CONT_PMD_SIZE)
340 		addr &= CONT_PMD_MASK;
341 
342 	pmdp = pmd_offset(pudp, addr);
343 	pmd = READ_ONCE(*pmdp);
344 	if (!(sz == PMD_SIZE || sz == CONT_PMD_SIZE) &&
345 	    pmd_none(pmd))
346 		return NULL;
347 	if (pmd_huge(pmd) || !pmd_present(pmd))
348 		return (pte_t *)pmdp;
349 
350 	if (sz == CONT_PTE_SIZE)
351 		return pte_offset_huge(pmdp, (addr & CONT_PTE_MASK));
352 
353 	return NULL;
354 }
355 
356 unsigned long hugetlb_mask_last_page(struct hstate *h)
357 {
358 	unsigned long hp_size = huge_page_size(h);
359 
360 	switch (hp_size) {
361 #ifndef __PAGETABLE_PMD_FOLDED
362 	case PUD_SIZE:
363 		return PGDIR_SIZE - PUD_SIZE;
364 #endif
365 	case CONT_PMD_SIZE:
366 		return PUD_SIZE - CONT_PMD_SIZE;
367 	case PMD_SIZE:
368 		return PUD_SIZE - PMD_SIZE;
369 	case CONT_PTE_SIZE:
370 		return PMD_SIZE - CONT_PTE_SIZE;
371 	default:
372 		break;
373 	}
374 
375 	return 0UL;
376 }
377 
378 pte_t arch_make_huge_pte(pte_t entry, unsigned int shift, vm_flags_t flags)
379 {
380 	size_t pagesize = 1UL << shift;
381 
382 	entry = pte_mkhuge(entry);
383 	if (pagesize == CONT_PTE_SIZE) {
384 		entry = pte_mkcont(entry);
385 	} else if (pagesize == CONT_PMD_SIZE) {
386 		entry = pmd_pte(pmd_mkcont(pte_pmd(entry)));
387 	} else if (pagesize != PUD_SIZE && pagesize != PMD_SIZE) {
388 		pr_warn("%s: unrecognized huge page size 0x%lx\n",
389 			__func__, pagesize);
390 	}
391 	return entry;
392 }
393 
394 void huge_pte_clear(struct mm_struct *mm, unsigned long addr,
395 		    pte_t *ptep, unsigned long sz)
396 {
397 	int i, ncontig;
398 	size_t pgsize;
399 
400 	ncontig = num_contig_ptes(sz, &pgsize);
401 
402 	for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
403 		pte_clear(mm, addr, ptep);
404 }
405 
406 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
407 			      unsigned long addr, pte_t *ptep)
408 {
409 	int ncontig;
410 	size_t pgsize;
411 	pte_t orig_pte = ptep_get(ptep);
412 
413 	if (!pte_cont(orig_pte))
414 		return ptep_get_and_clear(mm, addr, ptep);
415 
416 	ncontig = find_num_contig(mm, addr, ptep, &pgsize);
417 
418 	return get_clear_contig(mm, addr, ptep, pgsize, ncontig);
419 }
420 
421 /*
422  * huge_ptep_set_access_flags will update access flags (dirty, accesssed)
423  * and write permission.
424  *
425  * For a contiguous huge pte range we need to check whether or not write
426  * permission has to change only on the first pte in the set. Then for
427  * all the contiguous ptes we need to check whether or not there is a
428  * discrepancy between dirty or young.
429  */
430 static int __cont_access_flags_changed(pte_t *ptep, pte_t pte, int ncontig)
431 {
432 	int i;
433 
434 	if (pte_write(pte) != pte_write(ptep_get(ptep)))
435 		return 1;
436 
437 	for (i = 0; i < ncontig; i++) {
438 		pte_t orig_pte = ptep_get(ptep + i);
439 
440 		if (pte_dirty(pte) != pte_dirty(orig_pte))
441 			return 1;
442 
443 		if (pte_young(pte) != pte_young(orig_pte))
444 			return 1;
445 	}
446 
447 	return 0;
448 }
449 
450 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
451 			       unsigned long addr, pte_t *ptep,
452 			       pte_t pte, int dirty)
453 {
454 	int ncontig, i;
455 	size_t pgsize = 0;
456 	unsigned long pfn = pte_pfn(pte), dpfn;
457 	struct mm_struct *mm = vma->vm_mm;
458 	pgprot_t hugeprot;
459 	pte_t orig_pte;
460 
461 	if (!pte_cont(pte))
462 		return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
463 
464 	ncontig = find_num_contig(mm, addr, ptep, &pgsize);
465 	dpfn = pgsize >> PAGE_SHIFT;
466 
467 	if (!__cont_access_flags_changed(ptep, pte, ncontig))
468 		return 0;
469 
470 	orig_pte = get_clear_contig_flush(mm, addr, ptep, pgsize, ncontig);
471 
472 	/* Make sure we don't lose the dirty or young state */
473 	if (pte_dirty(orig_pte))
474 		pte = pte_mkdirty(pte);
475 
476 	if (pte_young(orig_pte))
477 		pte = pte_mkyoung(pte);
478 
479 	hugeprot = pte_pgprot(pte);
480 	for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
481 		set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
482 
483 	return 1;
484 }
485 
486 void huge_ptep_set_wrprotect(struct mm_struct *mm,
487 			     unsigned long addr, pte_t *ptep)
488 {
489 	unsigned long pfn, dpfn;
490 	pgprot_t hugeprot;
491 	int ncontig, i;
492 	size_t pgsize;
493 	pte_t pte;
494 
495 	if (!pte_cont(READ_ONCE(*ptep))) {
496 		ptep_set_wrprotect(mm, addr, ptep);
497 		return;
498 	}
499 
500 	ncontig = find_num_contig(mm, addr, ptep, &pgsize);
501 	dpfn = pgsize >> PAGE_SHIFT;
502 
503 	pte = get_clear_contig_flush(mm, addr, ptep, pgsize, ncontig);
504 	pte = pte_wrprotect(pte);
505 
506 	hugeprot = pte_pgprot(pte);
507 	pfn = pte_pfn(pte);
508 
509 	for (i = 0; i < ncontig; i++, ptep++, addr += pgsize, pfn += dpfn)
510 		set_pte_at(mm, addr, ptep, pfn_pte(pfn, hugeprot));
511 }
512 
513 pte_t huge_ptep_clear_flush(struct vm_area_struct *vma,
514 			    unsigned long addr, pte_t *ptep)
515 {
516 	struct mm_struct *mm = vma->vm_mm;
517 	size_t pgsize;
518 	int ncontig;
519 
520 	if (!pte_cont(READ_ONCE(*ptep)))
521 		return ptep_clear_flush(vma, addr, ptep);
522 
523 	ncontig = find_num_contig(mm, addr, ptep, &pgsize);
524 	return get_clear_contig_flush(mm, addr, ptep, pgsize, ncontig);
525 }
526 
527 static int __init hugetlbpage_init(void)
528 {
529 	if (pud_sect_supported())
530 		hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
531 
532 	hugetlb_add_hstate(CONT_PMD_SHIFT - PAGE_SHIFT);
533 	hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
534 	hugetlb_add_hstate(CONT_PTE_SHIFT - PAGE_SHIFT);
535 
536 	return 0;
537 }
538 arch_initcall(hugetlbpage_init);
539 
540 bool __init arch_hugetlb_valid_size(unsigned long size)
541 {
542 	return __hugetlb_valid_size(size);
543 }
544 
545 pte_t huge_ptep_modify_prot_start(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep)
546 {
547 	if (IS_ENABLED(CONFIG_ARM64_ERRATUM_2645198) &&
548 	    cpus_have_const_cap(ARM64_WORKAROUND_2645198)) {
549 		/*
550 		 * Break-before-make (BBM) is required for all user space mappings
551 		 * when the permission changes from executable to non-executable
552 		 * in cases where cpu is affected with errata #2645198.
553 		 */
554 		if (pte_user_exec(READ_ONCE(*ptep)))
555 			return huge_ptep_clear_flush(vma, addr, ptep);
556 	}
557 	return huge_ptep_get_and_clear(vma->vm_mm, addr, ptep);
558 }
559 
560 void huge_ptep_modify_prot_commit(struct vm_area_struct *vma, unsigned long addr, pte_t *ptep,
561 				  pte_t old_pte, pte_t pte)
562 {
563 	unsigned long psize = huge_page_size(hstate_vma(vma));
564 
565 	set_huge_pte_at(vma->vm_mm, addr, ptep, pte, psize);
566 }
567