xref: /openbmc/linux/arch/powerpc/mm/pgtable.c (revision 27ab1c1c)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * This file contains common routines for dealing with free of page tables
4  * Along with common page table handling code
5  *
6  *  Derived from arch/powerpc/mm/tlb_64.c:
7  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
8  *
9  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
10  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
11  *    Copyright (C) 1996 Paul Mackerras
12  *
13  *  Derived from "arch/i386/mm/init.c"
14  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
15  *
16  *  Dave Engebretsen <engebret@us.ibm.com>
17  *      Rework for PPC64 port.
18  */
19 
20 #include <linux/kernel.h>
21 #include <linux/gfp.h>
22 #include <linux/mm.h>
23 #include <linux/percpu.h>
24 #include <linux/hardirq.h>
25 #include <linux/hugetlb.h>
26 #include <asm/tlbflush.h>
27 #include <asm/tlb.h>
28 #include <asm/hugetlb.h>
29 
30 static inline int is_exec_fault(void)
31 {
32 	return current->thread.regs && TRAP(current->thread.regs) == 0x400;
33 }
34 
35 /* We only try to do i/d cache coherency on stuff that looks like
36  * reasonably "normal" PTEs. We currently require a PTE to be present
37  * and we avoid _PAGE_SPECIAL and cache inhibited pte. We also only do that
38  * on userspace PTEs
39  */
40 static inline int pte_looks_normal(pte_t pte)
41 {
42 
43 	if (pte_present(pte) && !pte_special(pte)) {
44 		if (pte_ci(pte))
45 			return 0;
46 		if (pte_user(pte))
47 			return 1;
48 	}
49 	return 0;
50 }
51 
52 static struct page *maybe_pte_to_page(pte_t pte)
53 {
54 	unsigned long pfn = pte_pfn(pte);
55 	struct page *page;
56 
57 	if (unlikely(!pfn_valid(pfn)))
58 		return NULL;
59 	page = pfn_to_page(pfn);
60 	if (PageReserved(page))
61 		return NULL;
62 	return page;
63 }
64 
65 #ifdef CONFIG_PPC_BOOK3S
66 
67 /* Server-style MMU handles coherency when hashing if HW exec permission
68  * is supposed per page (currently 64-bit only). If not, then, we always
69  * flush the cache for valid PTEs in set_pte. Embedded CPU without HW exec
70  * support falls into the same category.
71  */
72 
73 static pte_t set_pte_filter_hash(pte_t pte)
74 {
75 	if (radix_enabled())
76 		return pte;
77 
78 	pte = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
79 	if (pte_looks_normal(pte) && !(cpu_has_feature(CPU_FTR_COHERENT_ICACHE) ||
80 				       cpu_has_feature(CPU_FTR_NOEXECUTE))) {
81 		struct page *pg = maybe_pte_to_page(pte);
82 		if (!pg)
83 			return pte;
84 		if (!test_bit(PG_arch_1, &pg->flags)) {
85 			flush_dcache_icache_page(pg);
86 			set_bit(PG_arch_1, &pg->flags);
87 		}
88 	}
89 	return pte;
90 }
91 
92 #else /* CONFIG_PPC_BOOK3S */
93 
94 static pte_t set_pte_filter_hash(pte_t pte) { return pte; }
95 
96 #endif /* CONFIG_PPC_BOOK3S */
97 
98 /* Embedded type MMU with HW exec support. This is a bit more complicated
99  * as we don't have two bits to spare for _PAGE_EXEC and _PAGE_HWEXEC so
100  * instead we "filter out" the exec permission for non clean pages.
101  */
102 static inline pte_t set_pte_filter(pte_t pte)
103 {
104 	struct page *pg;
105 
106 	if (mmu_has_feature(MMU_FTR_HPTE_TABLE))
107 		return set_pte_filter_hash(pte);
108 
109 	/* No exec permission in the first place, move on */
110 	if (!pte_exec(pte) || !pte_looks_normal(pte))
111 		return pte;
112 
113 	/* If you set _PAGE_EXEC on weird pages you're on your own */
114 	pg = maybe_pte_to_page(pte);
115 	if (unlikely(!pg))
116 		return pte;
117 
118 	/* If the page clean, we move on */
119 	if (test_bit(PG_arch_1, &pg->flags))
120 		return pte;
121 
122 	/* If it's an exec fault, we flush the cache and make it clean */
123 	if (is_exec_fault()) {
124 		flush_dcache_icache_page(pg);
125 		set_bit(PG_arch_1, &pg->flags);
126 		return pte;
127 	}
128 
129 	/* Else, we filter out _PAGE_EXEC */
130 	return pte_exprotect(pte);
131 }
132 
133 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma,
134 				     int dirty)
135 {
136 	struct page *pg;
137 
138 	if (mmu_has_feature(MMU_FTR_HPTE_TABLE))
139 		return pte;
140 
141 	/* So here, we only care about exec faults, as we use them
142 	 * to recover lost _PAGE_EXEC and perform I$/D$ coherency
143 	 * if necessary. Also if _PAGE_EXEC is already set, same deal,
144 	 * we just bail out
145 	 */
146 	if (dirty || pte_exec(pte) || !is_exec_fault())
147 		return pte;
148 
149 #ifdef CONFIG_DEBUG_VM
150 	/* So this is an exec fault, _PAGE_EXEC is not set. If it was
151 	 * an error we would have bailed out earlier in do_page_fault()
152 	 * but let's make sure of it
153 	 */
154 	if (WARN_ON(!(vma->vm_flags & VM_EXEC)))
155 		return pte;
156 #endif /* CONFIG_DEBUG_VM */
157 
158 	/* If you set _PAGE_EXEC on weird pages you're on your own */
159 	pg = maybe_pte_to_page(pte);
160 	if (unlikely(!pg))
161 		goto bail;
162 
163 	/* If the page is already clean, we move on */
164 	if (test_bit(PG_arch_1, &pg->flags))
165 		goto bail;
166 
167 	/* Clean the page and set PG_arch_1 */
168 	flush_dcache_icache_page(pg);
169 	set_bit(PG_arch_1, &pg->flags);
170 
171  bail:
172 	return pte_mkexec(pte);
173 }
174 
175 /*
176  * set_pte stores a linux PTE into the linux page table.
177  */
178 void set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep,
179 		pte_t pte)
180 {
181 	/*
182 	 * Make sure hardware valid bit is not set. We don't do
183 	 * tlb flush for this update.
184 	 */
185 	VM_WARN_ON(pte_hw_valid(*ptep) && !pte_protnone(*ptep));
186 
187 	/* Note: mm->context.id might not yet have been assigned as
188 	 * this context might not have been activated yet when this
189 	 * is called.
190 	 */
191 	pte = set_pte_filter(pte);
192 
193 	/* Perform the setting of the PTE */
194 	__set_pte_at(mm, addr, ptep, pte, 0);
195 }
196 
197 /*
198  * This is called when relaxing access to a PTE. It's also called in the page
199  * fault path when we don't hit any of the major fault cases, ie, a minor
200  * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
201  * handled those two for us, we additionally deal with missing execute
202  * permission here on some processors
203  */
204 int ptep_set_access_flags(struct vm_area_struct *vma, unsigned long address,
205 			  pte_t *ptep, pte_t entry, int dirty)
206 {
207 	int changed;
208 	entry = set_access_flags_filter(entry, vma, dirty);
209 	changed = !pte_same(*(ptep), entry);
210 	if (changed) {
211 		assert_pte_locked(vma->vm_mm, address);
212 		__ptep_set_access_flags(vma, ptep, entry,
213 					address, mmu_virtual_psize);
214 	}
215 	return changed;
216 }
217 
218 #ifdef CONFIG_HUGETLB_PAGE
219 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
220 			       unsigned long addr, pte_t *ptep,
221 			       pte_t pte, int dirty)
222 {
223 #ifdef HUGETLB_NEED_PRELOAD
224 	/*
225 	 * The "return 1" forces a call of update_mmu_cache, which will write a
226 	 * TLB entry.  Without this, platforms that don't do a write of the TLB
227 	 * entry in the TLB miss handler asm will fault ad infinitum.
228 	 */
229 	ptep_set_access_flags(vma, addr, ptep, pte, dirty);
230 	return 1;
231 #else
232 	int changed, psize;
233 
234 	pte = set_access_flags_filter(pte, vma, dirty);
235 	changed = !pte_same(*(ptep), pte);
236 	if (changed) {
237 
238 #ifdef CONFIG_PPC_BOOK3S_64
239 		struct hstate *h = hstate_vma(vma);
240 
241 		psize = hstate_get_psize(h);
242 #ifdef CONFIG_DEBUG_VM
243 		assert_spin_locked(huge_pte_lockptr(h, vma->vm_mm, ptep));
244 #endif
245 
246 #else
247 		/*
248 		 * Not used on non book3s64 platforms.
249 		 * 8xx compares it with mmu_virtual_psize to
250 		 * know if it is a huge page or not.
251 		 */
252 		psize = MMU_PAGE_COUNT;
253 #endif
254 		__ptep_set_access_flags(vma, ptep, pte, addr, psize);
255 	}
256 	return changed;
257 #endif
258 }
259 
260 #if defined(CONFIG_PPC_8xx)
261 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
262 {
263 	pmd_t *pmd = pmd_off(mm, addr);
264 	pte_basic_t val;
265 	pte_basic_t *entry = &ptep->pte;
266 	int num, i;
267 
268 	/*
269 	 * Make sure hardware valid bit is not set. We don't do
270 	 * tlb flush for this update.
271 	 */
272 	VM_WARN_ON(pte_hw_valid(*ptep) && !pte_protnone(*ptep));
273 
274 	pte = set_pte_filter(pte);
275 
276 	val = pte_val(pte);
277 
278 	num = number_of_cells_per_pte(pmd, val, 1);
279 
280 	for (i = 0; i < num; i++, entry++, val += SZ_4K)
281 		*entry = val;
282 }
283 #endif
284 #endif /* CONFIG_HUGETLB_PAGE */
285 
286 #ifdef CONFIG_DEBUG_VM
287 void assert_pte_locked(struct mm_struct *mm, unsigned long addr)
288 {
289 	pgd_t *pgd;
290 	p4d_t *p4d;
291 	pud_t *pud;
292 	pmd_t *pmd;
293 
294 	if (mm == &init_mm)
295 		return;
296 	pgd = mm->pgd + pgd_index(addr);
297 	BUG_ON(pgd_none(*pgd));
298 	p4d = p4d_offset(pgd, addr);
299 	BUG_ON(p4d_none(*p4d));
300 	pud = pud_offset(p4d, addr);
301 	BUG_ON(pud_none(*pud));
302 	pmd = pmd_offset(pud, addr);
303 	/*
304 	 * khugepaged to collapse normal pages to hugepage, first set
305 	 * pmd to none to force page fault/gup to take mmap_lock. After
306 	 * pmd is set to none, we do a pte_clear which does this assertion
307 	 * so if we find pmd none, return.
308 	 */
309 	if (pmd_none(*pmd))
310 		return;
311 	BUG_ON(!pmd_present(*pmd));
312 	assert_spin_locked(pte_lockptr(mm, pmd));
313 }
314 #endif /* CONFIG_DEBUG_VM */
315 
316 unsigned long vmalloc_to_phys(void *va)
317 {
318 	unsigned long pfn = vmalloc_to_pfn(va);
319 
320 	BUG_ON(!pfn);
321 	return __pa(pfn_to_kaddr(pfn)) + offset_in_page(va);
322 }
323 EXPORT_SYMBOL_GPL(vmalloc_to_phys);
324 
325 /*
326  * We have 4 cases for pgds and pmds:
327  * (1) invalid (all zeroes)
328  * (2) pointer to next table, as normal; bottom 6 bits == 0
329  * (3) leaf pte for huge page _PAGE_PTE set
330  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
331  *
332  * So long as we atomically load page table pointers we are safe against teardown,
333  * we can follow the address down to the the page and take a ref on it.
334  * This function need to be called with interrupts disabled. We use this variant
335  * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
336  */
337 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
338 			bool *is_thp, unsigned *hpage_shift)
339 {
340 	pgd_t *pgdp;
341 	p4d_t p4d, *p4dp;
342 	pud_t pud, *pudp;
343 	pmd_t pmd, *pmdp;
344 	pte_t *ret_pte;
345 	hugepd_t *hpdp = NULL;
346 	unsigned pdshift;
347 
348 	if (hpage_shift)
349 		*hpage_shift = 0;
350 
351 	if (is_thp)
352 		*is_thp = false;
353 
354 	/*
355 	 * Always operate on the local stack value. This make sure the
356 	 * value don't get updated by a parallel THP split/collapse,
357 	 * page fault or a page unmap. The return pte_t * is still not
358 	 * stable. So should be checked there for above conditions.
359 	 * Top level is an exception because it is folded into p4d.
360 	 */
361 	pgdp = pgdir + pgd_index(ea);
362 	p4dp = p4d_offset(pgdp, ea);
363 	p4d  = READ_ONCE(*p4dp);
364 	pdshift = P4D_SHIFT;
365 
366 	if (p4d_none(p4d))
367 		return NULL;
368 
369 	if (p4d_is_leaf(p4d)) {
370 		ret_pte = (pte_t *)p4dp;
371 		goto out;
372 	}
373 
374 	if (is_hugepd(__hugepd(p4d_val(p4d)))) {
375 		hpdp = (hugepd_t *)&p4d;
376 		goto out_huge;
377 	}
378 
379 	/*
380 	 * Even if we end up with an unmap, the pgtable will not
381 	 * be freed, because we do an rcu free and here we are
382 	 * irq disabled
383 	 */
384 	pdshift = PUD_SHIFT;
385 	pudp = pud_offset(&p4d, ea);
386 	pud  = READ_ONCE(*pudp);
387 
388 	if (pud_none(pud))
389 		return NULL;
390 
391 	if (pud_is_leaf(pud)) {
392 		ret_pte = (pte_t *)pudp;
393 		goto out;
394 	}
395 
396 	if (is_hugepd(__hugepd(pud_val(pud)))) {
397 		hpdp = (hugepd_t *)&pud;
398 		goto out_huge;
399 	}
400 
401 	pdshift = PMD_SHIFT;
402 	pmdp = pmd_offset(&pud, ea);
403 	pmd  = READ_ONCE(*pmdp);
404 
405 	/*
406 	 * A hugepage collapse is captured by this condition, see
407 	 * pmdp_collapse_flush.
408 	 */
409 	if (pmd_none(pmd))
410 		return NULL;
411 
412 #ifdef CONFIG_PPC_BOOK3S_64
413 	/*
414 	 * A hugepage split is captured by this condition, see
415 	 * pmdp_invalidate.
416 	 *
417 	 * Huge page modification can be caught here too.
418 	 */
419 	if (pmd_is_serializing(pmd))
420 		return NULL;
421 #endif
422 
423 	if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
424 		if (is_thp)
425 			*is_thp = true;
426 		ret_pte = (pte_t *)pmdp;
427 		goto out;
428 	}
429 
430 	if (pmd_is_leaf(pmd)) {
431 		ret_pte = (pte_t *)pmdp;
432 		goto out;
433 	}
434 
435 	if (is_hugepd(__hugepd(pmd_val(pmd)))) {
436 		hpdp = (hugepd_t *)&pmd;
437 		goto out_huge;
438 	}
439 
440 	return pte_offset_kernel(&pmd, ea);
441 
442 out_huge:
443 	if (!hpdp)
444 		return NULL;
445 
446 	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
447 	pdshift = hugepd_shift(*hpdp);
448 out:
449 	if (hpage_shift)
450 		*hpage_shift = pdshift;
451 	return ret_pte;
452 }
453 EXPORT_SYMBOL_GPL(__find_linux_pte);
454