xref: /openbmc/linux/arch/riscv/mm/hugetlbpage.c (revision 8d81cd1a) 
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/hugetlb.h>
3 #include <linux/err.h>
4 
5 #ifdef CONFIG_RISCV_ISA_SVNAPOT
6 pte_t huge_ptep_get(pte_t *ptep)
7 {
8 	unsigned long pte_num;
9 	int i;
10 	pte_t orig_pte = ptep_get(ptep);
11 
12 	if (!pte_present(orig_pte) || !pte_napot(orig_pte))
13 		return orig_pte;
14 
15 	pte_num = napot_pte_num(napot_cont_order(orig_pte));
16 
17 	for (i = 0; i < pte_num; i++, ptep++) {
18 		pte_t pte = ptep_get(ptep);
19 
20 		if (pte_dirty(pte))
21 			orig_pte = pte_mkdirty(orig_pte);
22 
23 		if (pte_young(pte))
24 			orig_pte = pte_mkyoung(orig_pte);
25 	}
26 
27 	return orig_pte;
28 }
29 
30 pte_t *huge_pte_alloc(struct mm_struct *mm,
31 		      struct vm_area_struct *vma,
32 		      unsigned long addr,
33 		      unsigned long sz)
34 {
35 	unsigned long order;
36 	pte_t *pte = NULL;
37 	pgd_t *pgd;
38 	p4d_t *p4d;
39 	pud_t *pud;
40 	pmd_t *pmd;
41 
42 	pgd = pgd_offset(mm, addr);
43 	p4d = p4d_alloc(mm, pgd, addr);
44 	if (!p4d)
45 		return NULL;
46 
47 	pud = pud_alloc(mm, p4d, addr);
48 	if (!pud)
49 		return NULL;
50 
51 	if (sz == PUD_SIZE) {
52 		pte = (pte_t *)pud;
53 		goto out;
54 	}
55 
56 	if (sz == PMD_SIZE) {
57 		if (want_pmd_share(vma, addr) && pud_none(pudp_get(pud)))
58 			pte = huge_pmd_share(mm, vma, addr, pud);
59 		else
60 			pte = (pte_t *)pmd_alloc(mm, pud, addr);
61 		goto out;
62 	}
63 
64 	pmd = pmd_alloc(mm, pud, addr);
65 	if (!pmd)
66 		return NULL;
67 
68 	for_each_napot_order(order) {
69 		if (napot_cont_size(order) == sz) {
70 			pte = pte_alloc_huge(mm, pmd, addr & napot_cont_mask(order));
71 			break;
72 		}
73 	}
74 
75 out:
76 	if (pte) {
77 		pte_t pteval = ptep_get_lockless(pte);
78 
79 		WARN_ON_ONCE(pte_present(pteval) && !pte_huge(pteval));
80 	}
81 	return pte;
82 }
83 
84 pte_t *huge_pte_offset(struct mm_struct *mm,
85 		       unsigned long addr,
86 		       unsigned long sz)
87 {
88 	unsigned long order;
89 	pte_t *pte = NULL;
90 	pgd_t *pgd;
91 	p4d_t *p4d;
92 	pud_t *pud;
93 	pmd_t *pmd;
94 
95 	pgd = pgd_offset(mm, addr);
96 	if (!pgd_present(pgdp_get(pgd)))
97 		return NULL;
98 
99 	p4d = p4d_offset(pgd, addr);
100 	if (!p4d_present(p4dp_get(p4d)))
101 		return NULL;
102 
103 	pud = pud_offset(p4d, addr);
104 	if (sz == PUD_SIZE)
105 		/* must be pud huge, non-present or none */
106 		return (pte_t *)pud;
107 
108 	if (!pud_present(pudp_get(pud)))
109 		return NULL;
110 
111 	pmd = pmd_offset(pud, addr);
112 	if (sz == PMD_SIZE)
113 		/* must be pmd huge, non-present or none */
114 		return (pte_t *)pmd;
115 
116 	if (!pmd_present(pmdp_get(pmd)))
117 		return NULL;
118 
119 	for_each_napot_order(order) {
120 		if (napot_cont_size(order) == sz) {
121 			pte = pte_offset_huge(pmd, addr & napot_cont_mask(order));
122 			break;
123 		}
124 	}
125 	return pte;
126 }
127 
128 unsigned long hugetlb_mask_last_page(struct hstate *h)
129 {
130 	unsigned long hp_size = huge_page_size(h);
131 
132 	switch (hp_size) {
133 #ifndef __PAGETABLE_PMD_FOLDED
134 	case PUD_SIZE:
135 		return P4D_SIZE - PUD_SIZE;
136 #endif
137 	case PMD_SIZE:
138 		return PUD_SIZE - PMD_SIZE;
139 	case napot_cont_size(NAPOT_CONT64KB_ORDER):
140 		return PMD_SIZE - napot_cont_size(NAPOT_CONT64KB_ORDER);
141 	default:
142 		break;
143 	}
144 
145 	return 0UL;
146 }
147 
148 static pte_t get_clear_contig(struct mm_struct *mm,
149 			      unsigned long addr,
150 			      pte_t *ptep,
151 			      unsigned long pte_num)
152 {
153 	pte_t orig_pte = ptep_get(ptep);
154 	unsigned long i;
155 
156 	for (i = 0; i < pte_num; i++, addr += PAGE_SIZE, ptep++) {
157 		pte_t pte = ptep_get_and_clear(mm, addr, ptep);
158 
159 		if (pte_dirty(pte))
160 			orig_pte = pte_mkdirty(orig_pte);
161 
162 		if (pte_young(pte))
163 			orig_pte = pte_mkyoung(orig_pte);
164 	}
165 
166 	return orig_pte;
167 }
168 
169 static pte_t get_clear_contig_flush(struct mm_struct *mm,
170 				    unsigned long addr,
171 				    pte_t *ptep,
172 				    unsigned long pte_num)
173 {
174 	pte_t orig_pte = get_clear_contig(mm, addr, ptep, pte_num);
175 	struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
176 	bool valid = !pte_none(orig_pte);
177 
178 	if (valid)
179 		flush_tlb_range(&vma, addr, addr + (PAGE_SIZE * pte_num));
180 
181 	return orig_pte;
182 }
183 
184 pte_t arch_make_huge_pte(pte_t entry, unsigned int shift, vm_flags_t flags)
185 {
186 	unsigned long order;
187 
188 	for_each_napot_order(order) {
189 		if (shift == napot_cont_shift(order)) {
190 			entry = pte_mknapot(entry, order);
191 			break;
192 		}
193 	}
194 	if (order == NAPOT_ORDER_MAX)
195 		entry = pte_mkhuge(entry);
196 
197 	return entry;
198 }
199 
200 static void clear_flush(struct mm_struct *mm,
201 			unsigned long addr,
202 			pte_t *ptep,
203 			unsigned long pgsize,
204 			unsigned long ncontig)
205 {
206 	struct vm_area_struct vma = TLB_FLUSH_VMA(mm, 0);
207 	unsigned long i, saddr = addr;
208 
209 	for (i = 0; i < ncontig; i++, addr += pgsize, ptep++)
210 		ptep_get_and_clear(mm, addr, ptep);
211 
212 	flush_tlb_range(&vma, saddr, addr);
213 }
214 
215 /*
216  * When dealing with NAPOT mappings, the privileged specification indicates that
217  * "if an update needs to be made, the OS generally should first mark all of the
218  * PTEs invalid, then issue SFENCE.VMA instruction(s) covering all 4 KiB regions
219  * within the range, [...] then update the PTE(s), as described in Section
220  * 4.2.1.". That's the equivalent of the Break-Before-Make approach used by
221  * arm64.
222  */
223 void set_huge_pte_at(struct mm_struct *mm,
224 		     unsigned long addr,
225 		     pte_t *ptep,
226 		     pte_t pte,
227 		     unsigned long sz)
228 {
229 	unsigned long hugepage_shift, pgsize;
230 	int i, pte_num;
231 
232 	if (sz >= PGDIR_SIZE)
233 		hugepage_shift = PGDIR_SHIFT;
234 	else if (sz >= P4D_SIZE)
235 		hugepage_shift = P4D_SHIFT;
236 	else if (sz >= PUD_SIZE)
237 		hugepage_shift = PUD_SHIFT;
238 	else if (sz >= PMD_SIZE)
239 		hugepage_shift = PMD_SHIFT;
240 	else
241 		hugepage_shift = PAGE_SHIFT;
242 
243 	pte_num = sz >> hugepage_shift;
244 	pgsize = 1 << hugepage_shift;
245 
246 	if (!pte_present(pte)) {
247 		for (i = 0; i < pte_num; i++, ptep++, addr += pgsize)
248 			set_ptes(mm, addr, ptep, pte, 1);
249 		return;
250 	}
251 
252 	if (!pte_napot(pte)) {
253 		set_ptes(mm, addr, ptep, pte, 1);
254 		return;
255 	}
256 
257 	clear_flush(mm, addr, ptep, pgsize, pte_num);
258 
259 	for (i = 0; i < pte_num; i++, ptep++, addr += pgsize)
260 		set_pte_at(mm, addr, ptep, pte);
261 }
262 
263 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
264 			       unsigned long addr,
265 			       pte_t *ptep,
266 			       pte_t pte,
267 			       int dirty)
268 {
269 	struct mm_struct *mm = vma->vm_mm;
270 	unsigned long order;
271 	pte_t orig_pte;
272 	int i, pte_num;
273 
274 	if (!pte_napot(pte))
275 		return ptep_set_access_flags(vma, addr, ptep, pte, dirty);
276 
277 	order = napot_cont_order(pte);
278 	pte_num = napot_pte_num(order);
279 	ptep = huge_pte_offset(mm, addr, napot_cont_size(order));
280 	orig_pte = get_clear_contig_flush(mm, addr, ptep, pte_num);
281 
282 	if (pte_dirty(orig_pte))
283 		pte = pte_mkdirty(pte);
284 
285 	if (pte_young(orig_pte))
286 		pte = pte_mkyoung(pte);
287 
288 	for (i = 0; i < pte_num; i++, addr += PAGE_SIZE, ptep++)
289 		set_pte_at(mm, addr, ptep, pte);
290 
291 	return true;
292 }
293 
294 pte_t huge_ptep_get_and_clear(struct mm_struct *mm,
295 			      unsigned long addr,
296 			      pte_t *ptep)
297 {
298 	pte_t orig_pte = ptep_get(ptep);
299 	int pte_num;
300 
301 	if (!pte_napot(orig_pte))
302 		return ptep_get_and_clear(mm, addr, ptep);
303 
304 	pte_num = napot_pte_num(napot_cont_order(orig_pte));
305 
306 	return get_clear_contig(mm, addr, ptep, pte_num);
307 }
308 
309 void huge_ptep_set_wrprotect(struct mm_struct *mm,
310 			     unsigned long addr,
311 			     pte_t *ptep)
312 {
313 	pte_t pte = ptep_get(ptep);
314 	unsigned long order;
315 	pte_t orig_pte;
316 	int i, pte_num;
317 
318 	if (!pte_napot(pte)) {
319 		ptep_set_wrprotect(mm, addr, ptep);
320 		return;
321 	}
322 
323 	order = napot_cont_order(pte);
324 	pte_num = napot_pte_num(order);
325 	ptep = huge_pte_offset(mm, addr, napot_cont_size(order));
326 	orig_pte = get_clear_contig_flush(mm, addr, ptep, pte_num);
327 
328 	orig_pte = pte_wrprotect(orig_pte);
329 
330 	for (i = 0; i < pte_num; i++, addr += PAGE_SIZE, ptep++)
331 		set_pte_at(mm, addr, ptep, orig_pte);
332 }
333 
334 pte_t huge_ptep_clear_flush(struct vm_area_struct *vma,
335 			    unsigned long addr,
336 			    pte_t *ptep)
337 {
338 	pte_t pte = ptep_get(ptep);
339 	int pte_num;
340 
341 	if (!pte_napot(pte))
342 		return ptep_clear_flush(vma, addr, ptep);
343 
344 	pte_num = napot_pte_num(napot_cont_order(pte));
345 
346 	return get_clear_contig_flush(vma->vm_mm, addr, ptep, pte_num);
347 }
348 
349 void huge_pte_clear(struct mm_struct *mm,
350 		    unsigned long addr,
351 		    pte_t *ptep,
352 		    unsigned long sz)
353 {
354 	pte_t pte = ptep_get(ptep);
355 	int i, pte_num;
356 
357 	if (!pte_napot(pte)) {
358 		pte_clear(mm, addr, ptep);
359 		return;
360 	}
361 
362 	pte_num = napot_pte_num(napot_cont_order(pte));
363 	for (i = 0; i < pte_num; i++, addr += PAGE_SIZE, ptep++)
364 		pte_clear(mm, addr, ptep);
365 }
366 
367 static bool is_napot_size(unsigned long size)
368 {
369 	unsigned long order;
370 
371 	if (!has_svnapot())
372 		return false;
373 
374 	for_each_napot_order(order) {
375 		if (size == napot_cont_size(order))
376 			return true;
377 	}
378 	return false;
379 }
380 
381 static __init int napot_hugetlbpages_init(void)
382 {
383 	if (has_svnapot()) {
384 		unsigned long order;
385 
386 		for_each_napot_order(order)
387 			hugetlb_add_hstate(order);
388 	}
389 	return 0;
390 }
391 arch_initcall(napot_hugetlbpages_init);
392 
393 #else
394 
395 static bool is_napot_size(unsigned long size)
396 {
397 	return false;
398 }
399 
400 #endif /*CONFIG_RISCV_ISA_SVNAPOT*/
401 
402 int pud_huge(pud_t pud)
403 {
404 	return pud_leaf(pud);
405 }
406 
407 int pmd_huge(pmd_t pmd)
408 {
409 	return pmd_leaf(pmd);
410 }
411 
412 static bool __hugetlb_valid_size(unsigned long size)
413 {
414 	if (size == HPAGE_SIZE)
415 		return true;
416 	else if (IS_ENABLED(CONFIG_64BIT) && size == PUD_SIZE)
417 		return true;
418 	else if (is_napot_size(size))
419 		return true;
420 	else
421 		return false;
422 }
423 
424 bool __init arch_hugetlb_valid_size(unsigned long size)
425 {
426 	return __hugetlb_valid_size(size);
427 }
428 
429 #ifdef CONFIG_ARCH_ENABLE_HUGEPAGE_MIGRATION
430 bool arch_hugetlb_migration_supported(struct hstate *h)
431 {
432 	return __hugetlb_valid_size(huge_page_size(h));
433 }
434 #endif
435 
436 #ifdef CONFIG_CONTIG_ALLOC
437 static __init int gigantic_pages_init(void)
438 {
439 	/* With CONTIG_ALLOC, we can allocate gigantic pages at runtime */
440 	if (IS_ENABLED(CONFIG_64BIT))
441 		hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
442 	return 0;
443 }
444 arch_initcall(gigantic_pages_init);
445 #endif
446