xref: /openbmc/linux/arch/sparc/mm/hugetlbpage.c (revision e15a5365)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * SPARC64 Huge TLB page support.
4  *
5  * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
6  */
7 
8 #include <linux/fs.h>
9 #include <linux/mm.h>
10 #include <linux/sched/mm.h>
11 #include <linux/hugetlb.h>
12 #include <linux/pagemap.h>
13 #include <linux/sysctl.h>
14 
15 #include <asm/mman.h>
16 #include <asm/pgalloc.h>
17 #include <asm/tlb.h>
18 #include <asm/tlbflush.h>
19 #include <asm/cacheflush.h>
20 #include <asm/mmu_context.h>
21 
22 /* Slightly simplified from the non-hugepage variant because by
23  * definition we don't have to worry about any page coloring stuff
24  */
25 
26 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *filp,
27 							unsigned long addr,
28 							unsigned long len,
29 							unsigned long pgoff,
30 							unsigned long flags)
31 {
32 	struct hstate *h = hstate_file(filp);
33 	unsigned long task_size = TASK_SIZE;
34 	struct vm_unmapped_area_info info;
35 
36 	if (test_thread_flag(TIF_32BIT))
37 		task_size = STACK_TOP32;
38 
39 	info.flags = 0;
40 	info.length = len;
41 	info.low_limit = TASK_UNMAPPED_BASE;
42 	info.high_limit = min(task_size, VA_EXCLUDE_START);
43 	info.align_mask = PAGE_MASK & ~huge_page_mask(h);
44 	info.align_offset = 0;
45 	addr = vm_unmapped_area(&info);
46 
47 	if ((addr & ~PAGE_MASK) && task_size > VA_EXCLUDE_END) {
48 		VM_BUG_ON(addr != -ENOMEM);
49 		info.low_limit = VA_EXCLUDE_END;
50 		info.high_limit = task_size;
51 		addr = vm_unmapped_area(&info);
52 	}
53 
54 	return addr;
55 }
56 
57 static unsigned long
58 hugetlb_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
59 				  const unsigned long len,
60 				  const unsigned long pgoff,
61 				  const unsigned long flags)
62 {
63 	struct hstate *h = hstate_file(filp);
64 	struct mm_struct *mm = current->mm;
65 	unsigned long addr = addr0;
66 	struct vm_unmapped_area_info info;
67 
68 	/* This should only ever run for 32-bit processes.  */
69 	BUG_ON(!test_thread_flag(TIF_32BIT));
70 
71 	info.flags = VM_UNMAPPED_AREA_TOPDOWN;
72 	info.length = len;
73 	info.low_limit = PAGE_SIZE;
74 	info.high_limit = mm->mmap_base;
75 	info.align_mask = PAGE_MASK & ~huge_page_mask(h);
76 	info.align_offset = 0;
77 	addr = vm_unmapped_area(&info);
78 
79 	/*
80 	 * A failed mmap() very likely causes application failure,
81 	 * so fall back to the bottom-up function here. This scenario
82 	 * can happen with large stack limits and large mmap()
83 	 * allocations.
84 	 */
85 	if (addr & ~PAGE_MASK) {
86 		VM_BUG_ON(addr != -ENOMEM);
87 		info.flags = 0;
88 		info.low_limit = TASK_UNMAPPED_BASE;
89 		info.high_limit = STACK_TOP32;
90 		addr = vm_unmapped_area(&info);
91 	}
92 
93 	return addr;
94 }
95 
96 unsigned long
97 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
98 		unsigned long len, unsigned long pgoff, unsigned long flags)
99 {
100 	struct hstate *h = hstate_file(file);
101 	struct mm_struct *mm = current->mm;
102 	struct vm_area_struct *vma;
103 	unsigned long task_size = TASK_SIZE;
104 
105 	if (test_thread_flag(TIF_32BIT))
106 		task_size = STACK_TOP32;
107 
108 	if (len & ~huge_page_mask(h))
109 		return -EINVAL;
110 	if (len > task_size)
111 		return -ENOMEM;
112 
113 	if (flags & MAP_FIXED) {
114 		if (prepare_hugepage_range(file, addr, len))
115 			return -EINVAL;
116 		return addr;
117 	}
118 
119 	if (addr) {
120 		addr = ALIGN(addr, huge_page_size(h));
121 		vma = find_vma(mm, addr);
122 		if (task_size - len >= addr &&
123 		    (!vma || addr + len <= vm_start_gap(vma)))
124 			return addr;
125 	}
126 	if (mm->get_unmapped_area == arch_get_unmapped_area)
127 		return hugetlb_get_unmapped_area_bottomup(file, addr, len,
128 				pgoff, flags);
129 	else
130 		return hugetlb_get_unmapped_area_topdown(file, addr, len,
131 				pgoff, flags);
132 }
133 
134 static pte_t sun4u_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
135 {
136 	return entry;
137 }
138 
139 static pte_t sun4v_hugepage_shift_to_tte(pte_t entry, unsigned int shift)
140 {
141 	unsigned long hugepage_size = _PAGE_SZ4MB_4V;
142 
143 	pte_val(entry) = pte_val(entry) & ~_PAGE_SZALL_4V;
144 
145 	switch (shift) {
146 	case HPAGE_16GB_SHIFT:
147 		hugepage_size = _PAGE_SZ16GB_4V;
148 		pte_val(entry) |= _PAGE_PUD_HUGE;
149 		break;
150 	case HPAGE_2GB_SHIFT:
151 		hugepage_size = _PAGE_SZ2GB_4V;
152 		pte_val(entry) |= _PAGE_PMD_HUGE;
153 		break;
154 	case HPAGE_256MB_SHIFT:
155 		hugepage_size = _PAGE_SZ256MB_4V;
156 		pte_val(entry) |= _PAGE_PMD_HUGE;
157 		break;
158 	case HPAGE_SHIFT:
159 		pte_val(entry) |= _PAGE_PMD_HUGE;
160 		break;
161 	case HPAGE_64K_SHIFT:
162 		hugepage_size = _PAGE_SZ64K_4V;
163 		break;
164 	default:
165 		WARN_ONCE(1, "unsupported hugepage shift=%u\n", shift);
166 	}
167 
168 	pte_val(entry) = pte_val(entry) | hugepage_size;
169 	return entry;
170 }
171 
172 static pte_t hugepage_shift_to_tte(pte_t entry, unsigned int shift)
173 {
174 	if (tlb_type == hypervisor)
175 		return sun4v_hugepage_shift_to_tte(entry, shift);
176 	else
177 		return sun4u_hugepage_shift_to_tte(entry, shift);
178 }
179 
180 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
181 			 struct page *page, int writeable)
182 {
183 	unsigned int shift = huge_page_shift(hstate_vma(vma));
184 	pte_t pte;
185 
186 	pte = hugepage_shift_to_tte(entry, shift);
187 
188 #ifdef CONFIG_SPARC64
189 	/* If this vma has ADI enabled on it, turn on TTE.mcd
190 	 */
191 	if (vma->vm_flags & VM_SPARC_ADI)
192 		return pte_mkmcd(pte);
193 	else
194 		return pte_mknotmcd(pte);
195 #else
196 	return pte;
197 #endif
198 }
199 
200 static unsigned int sun4v_huge_tte_to_shift(pte_t entry)
201 {
202 	unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V;
203 	unsigned int shift;
204 
205 	switch (tte_szbits) {
206 	case _PAGE_SZ16GB_4V:
207 		shift = HPAGE_16GB_SHIFT;
208 		break;
209 	case _PAGE_SZ2GB_4V:
210 		shift = HPAGE_2GB_SHIFT;
211 		break;
212 	case _PAGE_SZ256MB_4V:
213 		shift = HPAGE_256MB_SHIFT;
214 		break;
215 	case _PAGE_SZ4MB_4V:
216 		shift = REAL_HPAGE_SHIFT;
217 		break;
218 	case _PAGE_SZ64K_4V:
219 		shift = HPAGE_64K_SHIFT;
220 		break;
221 	default:
222 		shift = PAGE_SHIFT;
223 		break;
224 	}
225 	return shift;
226 }
227 
228 static unsigned int sun4u_huge_tte_to_shift(pte_t entry)
229 {
230 	unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U;
231 	unsigned int shift;
232 
233 	switch (tte_szbits) {
234 	case _PAGE_SZ256MB_4U:
235 		shift = HPAGE_256MB_SHIFT;
236 		break;
237 	case _PAGE_SZ4MB_4U:
238 		shift = REAL_HPAGE_SHIFT;
239 		break;
240 	case _PAGE_SZ64K_4U:
241 		shift = HPAGE_64K_SHIFT;
242 		break;
243 	default:
244 		shift = PAGE_SHIFT;
245 		break;
246 	}
247 	return shift;
248 }
249 
250 static unsigned int huge_tte_to_shift(pte_t entry)
251 {
252 	unsigned long shift;
253 
254 	if (tlb_type == hypervisor)
255 		shift = sun4v_huge_tte_to_shift(entry);
256 	else
257 		shift = sun4u_huge_tte_to_shift(entry);
258 
259 	if (shift == PAGE_SHIFT)
260 		WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n",
261 			  pte_val(entry));
262 
263 	return shift;
264 }
265 
266 static unsigned long huge_tte_to_size(pte_t pte)
267 {
268 	unsigned long size = 1UL << huge_tte_to_shift(pte);
269 
270 	if (size == REAL_HPAGE_SIZE)
271 		size = HPAGE_SIZE;
272 	return size;
273 }
274 
275 pte_t *huge_pte_alloc(struct mm_struct *mm,
276 			unsigned long addr, unsigned long sz)
277 {
278 	pgd_t *pgd;
279 	p4d_t *p4d;
280 	pud_t *pud;
281 	pmd_t *pmd;
282 
283 	pgd = pgd_offset(mm, addr);
284 	p4d = p4d_offset(pgd, addr);
285 	pud = pud_alloc(mm, p4d, addr);
286 	if (!pud)
287 		return NULL;
288 	if (sz >= PUD_SIZE)
289 		return (pte_t *)pud;
290 	pmd = pmd_alloc(mm, pud, addr);
291 	if (!pmd)
292 		return NULL;
293 	if (sz >= PMD_SIZE)
294 		return (pte_t *)pmd;
295 	return pte_alloc_map(mm, pmd, addr);
296 }
297 
298 pte_t *huge_pte_offset(struct mm_struct *mm,
299 		       unsigned long addr, unsigned long sz)
300 {
301 	pgd_t *pgd;
302 	p4d_t *p4d;
303 	pud_t *pud;
304 	pmd_t *pmd;
305 
306 	pgd = pgd_offset(mm, addr);
307 	if (pgd_none(*pgd))
308 		return NULL;
309 	p4d = p4d_offset(pgd, addr);
310 	if (p4d_none(*p4d))
311 		return NULL;
312 	pud = pud_offset(p4d, addr);
313 	if (pud_none(*pud))
314 		return NULL;
315 	if (is_hugetlb_pud(*pud))
316 		return (pte_t *)pud;
317 	pmd = pmd_offset(pud, addr);
318 	if (pmd_none(*pmd))
319 		return NULL;
320 	if (is_hugetlb_pmd(*pmd))
321 		return (pte_t *)pmd;
322 	return pte_offset_map(pmd, addr);
323 }
324 
325 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
326 		     pte_t *ptep, pte_t entry)
327 {
328 	unsigned int nptes, orig_shift, shift;
329 	unsigned long i, size;
330 	pte_t orig;
331 
332 	size = huge_tte_to_size(entry);
333 
334 	shift = PAGE_SHIFT;
335 	if (size >= PUD_SIZE)
336 		shift = PUD_SHIFT;
337 	else if (size >= PMD_SIZE)
338 		shift = PMD_SHIFT;
339 	else
340 		shift = PAGE_SHIFT;
341 
342 	nptes = size >> shift;
343 
344 	if (!pte_present(*ptep) && pte_present(entry))
345 		mm->context.hugetlb_pte_count += nptes;
346 
347 	addr &= ~(size - 1);
348 	orig = *ptep;
349 	orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig);
350 
351 	for (i = 0; i < nptes; i++)
352 		ptep[i] = __pte(pte_val(entry) + (i << shift));
353 
354 	maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift);
355 	/* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
356 	if (size == HPAGE_SIZE)
357 		maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0,
358 				    orig_shift);
359 }
360 
361 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
362 			      pte_t *ptep)
363 {
364 	unsigned int i, nptes, orig_shift, shift;
365 	unsigned long size;
366 	pte_t entry;
367 
368 	entry = *ptep;
369 	size = huge_tte_to_size(entry);
370 
371 	shift = PAGE_SHIFT;
372 	if (size >= PUD_SIZE)
373 		shift = PUD_SHIFT;
374 	else if (size >= PMD_SIZE)
375 		shift = PMD_SHIFT;
376 	else
377 		shift = PAGE_SHIFT;
378 
379 	nptes = size >> shift;
380 	orig_shift = pte_none(entry) ? PAGE_SHIFT : huge_tte_to_shift(entry);
381 
382 	if (pte_present(entry))
383 		mm->context.hugetlb_pte_count -= nptes;
384 
385 	addr &= ~(size - 1);
386 	for (i = 0; i < nptes; i++)
387 		ptep[i] = __pte(0UL);
388 
389 	maybe_tlb_batch_add(mm, addr, ptep, entry, 0, orig_shift);
390 	/* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
391 	if (size == HPAGE_SIZE)
392 		maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0,
393 				    orig_shift);
394 
395 	return entry;
396 }
397 
398 int pmd_huge(pmd_t pmd)
399 {
400 	return !pmd_none(pmd) &&
401 		(pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID;
402 }
403 
404 int pud_huge(pud_t pud)
405 {
406 	return !pud_none(pud) &&
407 		(pud_val(pud) & (_PAGE_VALID|_PAGE_PUD_HUGE)) != _PAGE_VALID;
408 }
409 
410 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
411 			   unsigned long addr)
412 {
413 	pgtable_t token = pmd_pgtable(*pmd);
414 
415 	pmd_clear(pmd);
416 	pte_free_tlb(tlb, token, addr);
417 	mm_dec_nr_ptes(tlb->mm);
418 }
419 
420 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
421 				   unsigned long addr, unsigned long end,
422 				   unsigned long floor, unsigned long ceiling)
423 {
424 	pmd_t *pmd;
425 	unsigned long next;
426 	unsigned long start;
427 
428 	start = addr;
429 	pmd = pmd_offset(pud, addr);
430 	do {
431 		next = pmd_addr_end(addr, end);
432 		if (pmd_none(*pmd))
433 			continue;
434 		if (is_hugetlb_pmd(*pmd))
435 			pmd_clear(pmd);
436 		else
437 			hugetlb_free_pte_range(tlb, pmd, addr);
438 	} while (pmd++, addr = next, addr != end);
439 
440 	start &= PUD_MASK;
441 	if (start < floor)
442 		return;
443 	if (ceiling) {
444 		ceiling &= PUD_MASK;
445 		if (!ceiling)
446 			return;
447 	}
448 	if (end - 1 > ceiling - 1)
449 		return;
450 
451 	pmd = pmd_offset(pud, start);
452 	pud_clear(pud);
453 	pmd_free_tlb(tlb, pmd, start);
454 	mm_dec_nr_pmds(tlb->mm);
455 }
456 
457 static void hugetlb_free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
458 				   unsigned long addr, unsigned long end,
459 				   unsigned long floor, unsigned long ceiling)
460 {
461 	pud_t *pud;
462 	unsigned long next;
463 	unsigned long start;
464 
465 	start = addr;
466 	pud = pud_offset(p4d, addr);
467 	do {
468 		next = pud_addr_end(addr, end);
469 		if (pud_none_or_clear_bad(pud))
470 			continue;
471 		if (is_hugetlb_pud(*pud))
472 			pud_clear(pud);
473 		else
474 			hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
475 					       ceiling);
476 	} while (pud++, addr = next, addr != end);
477 
478 	start &= PGDIR_MASK;
479 	if (start < floor)
480 		return;
481 	if (ceiling) {
482 		ceiling &= PGDIR_MASK;
483 		if (!ceiling)
484 			return;
485 	}
486 	if (end - 1 > ceiling - 1)
487 		return;
488 
489 	pud = pud_offset(p4d, start);
490 	p4d_clear(p4d);
491 	pud_free_tlb(tlb, pud, start);
492 	mm_dec_nr_puds(tlb->mm);
493 }
494 
495 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
496 			    unsigned long addr, unsigned long end,
497 			    unsigned long floor, unsigned long ceiling)
498 {
499 	pgd_t *pgd;
500 	p4d_t *p4d;
501 	unsigned long next;
502 
503 	addr &= PMD_MASK;
504 	if (addr < floor) {
505 		addr += PMD_SIZE;
506 		if (!addr)
507 			return;
508 	}
509 	if (ceiling) {
510 		ceiling &= PMD_MASK;
511 		if (!ceiling)
512 			return;
513 	}
514 	if (end - 1 > ceiling - 1)
515 		end -= PMD_SIZE;
516 	if (addr > end - 1)
517 		return;
518 
519 	pgd = pgd_offset(tlb->mm, addr);
520 	p4d = p4d_offset(pgd, addr);
521 	do {
522 		next = p4d_addr_end(addr, end);
523 		if (p4d_none_or_clear_bad(p4d))
524 			continue;
525 		hugetlb_free_pud_range(tlb, p4d, addr, next, floor, ceiling);
526 	} while (p4d++, addr = next, addr != end);
527 }
528