xref: /openbmc/linux/arch/sparc/mm/hugetlbpage.c (revision 239480ab)
1 /*
2  * SPARC64 Huge TLB page support.
3  *
4  * Copyright (C) 2002, 2003, 2006 David S. Miller (davem@davemloft.net)
5  */
6 
7 #include <linux/fs.h>
8 #include <linux/mm.h>
9 #include <linux/sched/mm.h>
10 #include <linux/hugetlb.h>
11 #include <linux/pagemap.h>
12 #include <linux/sysctl.h>
13 
14 #include <asm/mman.h>
15 #include <asm/pgalloc.h>
16 #include <asm/pgtable.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 <= vma->vm_start))
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_2GB_SHIFT:
147 		hugepage_size = _PAGE_SZ2GB_4V;
148 		pte_val(entry) |= _PAGE_PMD_HUGE;
149 		break;
150 	case HPAGE_256MB_SHIFT:
151 		hugepage_size = _PAGE_SZ256MB_4V;
152 		pte_val(entry) |= _PAGE_PMD_HUGE;
153 		break;
154 	case HPAGE_SHIFT:
155 		pte_val(entry) |= _PAGE_PMD_HUGE;
156 		break;
157 	case HPAGE_64K_SHIFT:
158 		hugepage_size = _PAGE_SZ64K_4V;
159 		break;
160 	default:
161 		WARN_ONCE(1, "unsupported hugepage shift=%u\n", shift);
162 	}
163 
164 	pte_val(entry) = pte_val(entry) | hugepage_size;
165 	return entry;
166 }
167 
168 static pte_t hugepage_shift_to_tte(pte_t entry, unsigned int shift)
169 {
170 	if (tlb_type == hypervisor)
171 		return sun4v_hugepage_shift_to_tte(entry, shift);
172 	else
173 		return sun4u_hugepage_shift_to_tte(entry, shift);
174 }
175 
176 pte_t arch_make_huge_pte(pte_t entry, struct vm_area_struct *vma,
177 			 struct page *page, int writeable)
178 {
179 	unsigned int shift = huge_page_shift(hstate_vma(vma));
180 
181 	return hugepage_shift_to_tte(entry, shift);
182 }
183 
184 static unsigned int sun4v_huge_tte_to_shift(pte_t entry)
185 {
186 	unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4V;
187 	unsigned int shift;
188 
189 	switch (tte_szbits) {
190 	case _PAGE_SZ2GB_4V:
191 		shift = HPAGE_2GB_SHIFT;
192 		break;
193 	case _PAGE_SZ256MB_4V:
194 		shift = HPAGE_256MB_SHIFT;
195 		break;
196 	case _PAGE_SZ4MB_4V:
197 		shift = REAL_HPAGE_SHIFT;
198 		break;
199 	case _PAGE_SZ64K_4V:
200 		shift = HPAGE_64K_SHIFT;
201 		break;
202 	default:
203 		shift = PAGE_SHIFT;
204 		break;
205 	}
206 	return shift;
207 }
208 
209 static unsigned int sun4u_huge_tte_to_shift(pte_t entry)
210 {
211 	unsigned long tte_szbits = pte_val(entry) & _PAGE_SZALL_4U;
212 	unsigned int shift;
213 
214 	switch (tte_szbits) {
215 	case _PAGE_SZ256MB_4U:
216 		shift = HPAGE_256MB_SHIFT;
217 		break;
218 	case _PAGE_SZ4MB_4U:
219 		shift = REAL_HPAGE_SHIFT;
220 		break;
221 	case _PAGE_SZ64K_4U:
222 		shift = HPAGE_64K_SHIFT;
223 		break;
224 	default:
225 		shift = PAGE_SHIFT;
226 		break;
227 	}
228 	return shift;
229 }
230 
231 static unsigned int huge_tte_to_shift(pte_t entry)
232 {
233 	unsigned long shift;
234 
235 	if (tlb_type == hypervisor)
236 		shift = sun4v_huge_tte_to_shift(entry);
237 	else
238 		shift = sun4u_huge_tte_to_shift(entry);
239 
240 	if (shift == PAGE_SHIFT)
241 		WARN_ONCE(1, "tto_to_shift: invalid hugepage tte=0x%lx\n",
242 			  pte_val(entry));
243 
244 	return shift;
245 }
246 
247 static unsigned long huge_tte_to_size(pte_t pte)
248 {
249 	unsigned long size = 1UL << huge_tte_to_shift(pte);
250 
251 	if (size == REAL_HPAGE_SIZE)
252 		size = HPAGE_SIZE;
253 	return size;
254 }
255 
256 pte_t *huge_pte_alloc(struct mm_struct *mm,
257 			unsigned long addr, unsigned long sz)
258 {
259 	pgd_t *pgd;
260 	pud_t *pud;
261 	pmd_t *pmd;
262 	pte_t *pte = NULL;
263 
264 	pgd = pgd_offset(mm, addr);
265 	pud = pud_alloc(mm, pgd, addr);
266 	if (pud) {
267 		pmd = pmd_alloc(mm, pud, addr);
268 		if (!pmd)
269 			return NULL;
270 
271 		if (sz >= PMD_SIZE)
272 			pte = (pte_t *)pmd;
273 		else
274 			pte = pte_alloc_map(mm, pmd, addr);
275 	}
276 
277 	return pte;
278 }
279 
280 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
281 {
282 	pgd_t *pgd;
283 	pud_t *pud;
284 	pmd_t *pmd;
285 	pte_t *pte = NULL;
286 
287 	pgd = pgd_offset(mm, addr);
288 	if (!pgd_none(*pgd)) {
289 		pud = pud_offset(pgd, addr);
290 		if (!pud_none(*pud)) {
291 			pmd = pmd_offset(pud, addr);
292 			if (!pmd_none(*pmd)) {
293 				if (is_hugetlb_pmd(*pmd))
294 					pte = (pte_t *)pmd;
295 				else
296 					pte = pte_offset_map(pmd, addr);
297 			}
298 		}
299 	}
300 
301 	return pte;
302 }
303 
304 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr,
305 		     pte_t *ptep, pte_t entry)
306 {
307 	unsigned int i, nptes, orig_shift, shift;
308 	unsigned long size;
309 	pte_t orig;
310 
311 	size = huge_tte_to_size(entry);
312 	shift = size >= HPAGE_SIZE ? PMD_SHIFT : PAGE_SHIFT;
313 	nptes = size >> shift;
314 
315 	if (!pte_present(*ptep) && pte_present(entry))
316 		mm->context.hugetlb_pte_count += nptes;
317 
318 	addr &= ~(size - 1);
319 	orig = *ptep;
320 	orig_shift = pte_none(orig) ? PAGE_SHIFT : huge_tte_to_shift(orig);
321 
322 	for (i = 0; i < nptes; i++)
323 		ptep[i] = __pte(pte_val(entry) + (i << shift));
324 
325 	maybe_tlb_batch_add(mm, addr, ptep, orig, 0, orig_shift);
326 	/* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
327 	if (size == HPAGE_SIZE)
328 		maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, orig, 0,
329 				    orig_shift);
330 }
331 
332 pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr,
333 			      pte_t *ptep)
334 {
335 	unsigned int i, nptes, hugepage_shift;
336 	unsigned long size;
337 	pte_t entry;
338 
339 	entry = *ptep;
340 	size = huge_tte_to_size(entry);
341 	if (size >= HPAGE_SIZE)
342 		nptes = size >> PMD_SHIFT;
343 	else
344 		nptes = size >> PAGE_SHIFT;
345 
346 	hugepage_shift = pte_none(entry) ? PAGE_SHIFT :
347 		huge_tte_to_shift(entry);
348 
349 	if (pte_present(entry))
350 		mm->context.hugetlb_pte_count -= nptes;
351 
352 	addr &= ~(size - 1);
353 	for (i = 0; i < nptes; i++)
354 		ptep[i] = __pte(0UL);
355 
356 	maybe_tlb_batch_add(mm, addr, ptep, entry, 0, hugepage_shift);
357 	/* An HPAGE_SIZE'ed page is composed of two REAL_HPAGE_SIZE'ed pages */
358 	if (size == HPAGE_SIZE)
359 		maybe_tlb_batch_add(mm, addr + REAL_HPAGE_SIZE, ptep, entry, 0,
360 				    hugepage_shift);
361 
362 	return entry;
363 }
364 
365 int pmd_huge(pmd_t pmd)
366 {
367 	return !pmd_none(pmd) &&
368 		(pmd_val(pmd) & (_PAGE_VALID|_PAGE_PMD_HUGE)) != _PAGE_VALID;
369 }
370 
371 int pud_huge(pud_t pud)
372 {
373 	return 0;
374 }
375 
376 static void hugetlb_free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
377 			   unsigned long addr)
378 {
379 	pgtable_t token = pmd_pgtable(*pmd);
380 
381 	pmd_clear(pmd);
382 	pte_free_tlb(tlb, token, addr);
383 	atomic_long_dec(&tlb->mm->nr_ptes);
384 }
385 
386 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
387 				   unsigned long addr, unsigned long end,
388 				   unsigned long floor, unsigned long ceiling)
389 {
390 	pmd_t *pmd;
391 	unsigned long next;
392 	unsigned long start;
393 
394 	start = addr;
395 	pmd = pmd_offset(pud, addr);
396 	do {
397 		next = pmd_addr_end(addr, end);
398 		if (pmd_none(*pmd))
399 			continue;
400 		if (is_hugetlb_pmd(*pmd))
401 			pmd_clear(pmd);
402 		else
403 			hugetlb_free_pte_range(tlb, pmd, addr);
404 	} while (pmd++, addr = next, addr != end);
405 
406 	start &= PUD_MASK;
407 	if (start < floor)
408 		return;
409 	if (ceiling) {
410 		ceiling &= PUD_MASK;
411 		if (!ceiling)
412 			return;
413 	}
414 	if (end - 1 > ceiling - 1)
415 		return;
416 
417 	pmd = pmd_offset(pud, start);
418 	pud_clear(pud);
419 	pmd_free_tlb(tlb, pmd, start);
420 	mm_dec_nr_pmds(tlb->mm);
421 }
422 
423 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
424 				   unsigned long addr, unsigned long end,
425 				   unsigned long floor, unsigned long ceiling)
426 {
427 	pud_t *pud;
428 	unsigned long next;
429 	unsigned long start;
430 
431 	start = addr;
432 	pud = pud_offset(pgd, addr);
433 	do {
434 		next = pud_addr_end(addr, end);
435 		if (pud_none_or_clear_bad(pud))
436 			continue;
437 		hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
438 				       ceiling);
439 	} while (pud++, addr = next, addr != end);
440 
441 	start &= PGDIR_MASK;
442 	if (start < floor)
443 		return;
444 	if (ceiling) {
445 		ceiling &= PGDIR_MASK;
446 		if (!ceiling)
447 			return;
448 	}
449 	if (end - 1 > ceiling - 1)
450 		return;
451 
452 	pud = pud_offset(pgd, start);
453 	pgd_clear(pgd);
454 	pud_free_tlb(tlb, pud, start);
455 }
456 
457 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
458 			    unsigned long addr, unsigned long end,
459 			    unsigned long floor, unsigned long ceiling)
460 {
461 	pgd_t *pgd;
462 	unsigned long next;
463 
464 	addr &= PMD_MASK;
465 	if (addr < floor) {
466 		addr += PMD_SIZE;
467 		if (!addr)
468 			return;
469 	}
470 	if (ceiling) {
471 		ceiling &= PMD_MASK;
472 		if (!ceiling)
473 			return;
474 	}
475 	if (end - 1 > ceiling - 1)
476 		end -= PMD_SIZE;
477 	if (addr > end - 1)
478 		return;
479 
480 	pgd = pgd_offset(tlb->mm, addr);
481 	do {
482 		next = pgd_addr_end(addr, end);
483 		if (pgd_none_or_clear_bad(pgd))
484 			continue;
485 		hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
486 	} while (pgd++, addr = next, addr != end);
487 }
488