xref: /openbmc/linux/arch/powerpc/mm/pgtable.c (revision a44e4f3a)
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/pgalloc.h>
27 #include <asm/tlbflush.h>
28 #include <asm/tlb.h>
29 #include <asm/hugetlb.h>
30 
31 static inline int is_exec_fault(void)
32 {
33 	return current->thread.regs && TRAP(current->thread.regs) == 0x400;
34 }
35 
36 /* We only try to do i/d cache coherency on stuff that looks like
37  * reasonably "normal" PTEs. We currently require a PTE to be present
38  * and we avoid _PAGE_SPECIAL and cache inhibited pte. We also only do that
39  * on userspace PTEs
40  */
41 static inline int pte_looks_normal(pte_t pte)
42 {
43 
44 	if (pte_present(pte) && !pte_special(pte)) {
45 		if (pte_ci(pte))
46 			return 0;
47 		if (pte_user(pte))
48 			return 1;
49 	}
50 	return 0;
51 }
52 
53 static struct page *maybe_pte_to_page(pte_t pte)
54 {
55 	unsigned long pfn = pte_pfn(pte);
56 	struct page *page;
57 
58 	if (unlikely(!pfn_valid(pfn)))
59 		return NULL;
60 	page = pfn_to_page(pfn);
61 	if (PageReserved(page))
62 		return NULL;
63 	return page;
64 }
65 
66 #ifdef CONFIG_PPC_BOOK3S
67 
68 /* Server-style MMU handles coherency when hashing if HW exec permission
69  * is supposed per page (currently 64-bit only). If not, then, we always
70  * flush the cache for valid PTEs in set_pte. Embedded CPU without HW exec
71  * support falls into the same category.
72  */
73 
74 static pte_t set_pte_filter_hash(pte_t pte)
75 {
76 	if (radix_enabled())
77 		return pte;
78 
79 	pte = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
80 	if (pte_looks_normal(pte) && !(cpu_has_feature(CPU_FTR_COHERENT_ICACHE) ||
81 				       cpu_has_feature(CPU_FTR_NOEXECUTE))) {
82 		struct page *pg = maybe_pte_to_page(pte);
83 		if (!pg)
84 			return pte;
85 		if (!test_bit(PG_arch_1, &pg->flags)) {
86 			flush_dcache_icache_page(pg);
87 			set_bit(PG_arch_1, &pg->flags);
88 		}
89 	}
90 	return pte;
91 }
92 
93 #else /* CONFIG_PPC_BOOK3S */
94 
95 static pte_t set_pte_filter_hash(pte_t pte) { return pte; }
96 
97 #endif /* CONFIG_PPC_BOOK3S */
98 
99 /* Embedded type MMU with HW exec support. This is a bit more complicated
100  * as we don't have two bits to spare for _PAGE_EXEC and _PAGE_HWEXEC so
101  * instead we "filter out" the exec permission for non clean pages.
102  */
103 static pte_t set_pte_filter(pte_t pte)
104 {
105 	struct page *pg;
106 
107 	if (mmu_has_feature(MMU_FTR_HPTE_TABLE))
108 		return set_pte_filter_hash(pte);
109 
110 	/* No exec permission in the first place, move on */
111 	if (!pte_exec(pte) || !pte_looks_normal(pte))
112 		return pte;
113 
114 	/* If you set _PAGE_EXEC on weird pages you're on your own */
115 	pg = maybe_pte_to_page(pte);
116 	if (unlikely(!pg))
117 		return pte;
118 
119 	/* If the page clean, we move on */
120 	if (test_bit(PG_arch_1, &pg->flags))
121 		return pte;
122 
123 	/* If it's an exec fault, we flush the cache and make it clean */
124 	if (is_exec_fault()) {
125 		flush_dcache_icache_page(pg);
126 		set_bit(PG_arch_1, &pg->flags);
127 		return pte;
128 	}
129 
130 	/* Else, we filter out _PAGE_EXEC */
131 	return pte_exprotect(pte);
132 }
133 
134 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma,
135 				     int dirty)
136 {
137 	struct page *pg;
138 
139 	if (mmu_has_feature(MMU_FTR_HPTE_TABLE))
140 		return pte;
141 
142 	/* So here, we only care about exec faults, as we use them
143 	 * to recover lost _PAGE_EXEC and perform I$/D$ coherency
144 	 * if necessary. Also if _PAGE_EXEC is already set, same deal,
145 	 * we just bail out
146 	 */
147 	if (dirty || pte_exec(pte) || !is_exec_fault())
148 		return pte;
149 
150 #ifdef CONFIG_DEBUG_VM
151 	/* So this is an exec fault, _PAGE_EXEC is not set. If it was
152 	 * an error we would have bailed out earlier in do_page_fault()
153 	 * but let's make sure of it
154 	 */
155 	if (WARN_ON(!(vma->vm_flags & VM_EXEC)))
156 		return pte;
157 #endif /* CONFIG_DEBUG_VM */
158 
159 	/* If you set _PAGE_EXEC on weird pages you're on your own */
160 	pg = maybe_pte_to_page(pte);
161 	if (unlikely(!pg))
162 		goto bail;
163 
164 	/* If the page is already clean, we move on */
165 	if (test_bit(PG_arch_1, &pg->flags))
166 		goto bail;
167 
168 	/* Clean the page and set PG_arch_1 */
169 	flush_dcache_icache_page(pg);
170 	set_bit(PG_arch_1, &pg->flags);
171 
172  bail:
173 	return pte_mkexec(pte);
174 }
175 
176 /*
177  * set_pte stores a linux PTE into the linux page table.
178  */
179 void set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep,
180 		pte_t pte)
181 {
182 	/*
183 	 * Make sure hardware valid bit is not set. We don't do
184 	 * tlb flush for this update.
185 	 */
186 	VM_WARN_ON(pte_hw_valid(*ptep) && !pte_protnone(*ptep));
187 
188 	/* Add the pte bit when trying to set a pte */
189 	pte = pte_mkpte(pte);
190 
191 	/* Note: mm->context.id might not yet have been assigned as
192 	 * this context might not have been activated yet when this
193 	 * is called.
194 	 */
195 	pte = set_pte_filter(pte);
196 
197 	/* Perform the setting of the PTE */
198 	__set_pte_at(mm, addr, ptep, pte, 0);
199 }
200 
201 /*
202  * This is called when relaxing access to a PTE. It's also called in the page
203  * fault path when we don't hit any of the major fault cases, ie, a minor
204  * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
205  * handled those two for us, we additionally deal with missing execute
206  * permission here on some processors
207  */
208 int ptep_set_access_flags(struct vm_area_struct *vma, unsigned long address,
209 			  pte_t *ptep, pte_t entry, int dirty)
210 {
211 	int changed;
212 	entry = set_access_flags_filter(entry, vma, dirty);
213 	changed = !pte_same(*(ptep), entry);
214 	if (changed) {
215 		assert_pte_locked(vma->vm_mm, address);
216 		__ptep_set_access_flags(vma, ptep, entry,
217 					address, mmu_virtual_psize);
218 	}
219 	return changed;
220 }
221 
222 #ifdef CONFIG_HUGETLB_PAGE
223 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
224 			       unsigned long addr, pte_t *ptep,
225 			       pte_t pte, int dirty)
226 {
227 #ifdef HUGETLB_NEED_PRELOAD
228 	/*
229 	 * The "return 1" forces a call of update_mmu_cache, which will write a
230 	 * TLB entry.  Without this, platforms that don't do a write of the TLB
231 	 * entry in the TLB miss handler asm will fault ad infinitum.
232 	 */
233 	ptep_set_access_flags(vma, addr, ptep, pte, dirty);
234 	return 1;
235 #else
236 	int changed, psize;
237 
238 	pte = set_access_flags_filter(pte, vma, dirty);
239 	changed = !pte_same(*(ptep), pte);
240 	if (changed) {
241 
242 #ifdef CONFIG_PPC_BOOK3S_64
243 		struct hstate *h = hstate_vma(vma);
244 
245 		psize = hstate_get_psize(h);
246 #ifdef CONFIG_DEBUG_VM
247 		assert_spin_locked(huge_pte_lockptr(h, vma->vm_mm, ptep));
248 #endif
249 
250 #else
251 		/*
252 		 * Not used on non book3s64 platforms. But 8xx
253 		 * can possibly use tsize derived from hstate.
254 		 */
255 		psize = 0;
256 #endif
257 		__ptep_set_access_flags(vma, ptep, pte, addr, psize);
258 	}
259 	return changed;
260 #endif
261 }
262 #endif /* CONFIG_HUGETLB_PAGE */
263 
264 #ifdef CONFIG_DEBUG_VM
265 void assert_pte_locked(struct mm_struct *mm, unsigned long addr)
266 {
267 	pgd_t *pgd;
268 	pud_t *pud;
269 	pmd_t *pmd;
270 
271 	if (mm == &init_mm)
272 		return;
273 	pgd = mm->pgd + pgd_index(addr);
274 	BUG_ON(pgd_none(*pgd));
275 	pud = pud_offset(pgd, addr);
276 	BUG_ON(pud_none(*pud));
277 	pmd = pmd_offset(pud, addr);
278 	/*
279 	 * khugepaged to collapse normal pages to hugepage, first set
280 	 * pmd to none to force page fault/gup to take mmap_sem. After
281 	 * pmd is set to none, we do a pte_clear which does this assertion
282 	 * so if we find pmd none, return.
283 	 */
284 	if (pmd_none(*pmd))
285 		return;
286 	BUG_ON(!pmd_present(*pmd));
287 	assert_spin_locked(pte_lockptr(mm, pmd));
288 }
289 #endif /* CONFIG_DEBUG_VM */
290 
291 unsigned long vmalloc_to_phys(void *va)
292 {
293 	unsigned long pfn = vmalloc_to_pfn(va);
294 
295 	BUG_ON(!pfn);
296 	return __pa(pfn_to_kaddr(pfn)) + offset_in_page(va);
297 }
298 EXPORT_SYMBOL_GPL(vmalloc_to_phys);
299 
300 /*
301  * We have 4 cases for pgds and pmds:
302  * (1) invalid (all zeroes)
303  * (2) pointer to next table, as normal; bottom 6 bits == 0
304  * (3) leaf pte for huge page _PAGE_PTE set
305  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
306  *
307  * So long as we atomically load page table pointers we are safe against teardown,
308  * we can follow the address down to the the page and take a ref on it.
309  * This function need to be called with interrupts disabled. We use this variant
310  * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
311  */
312 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
313 			bool *is_thp, unsigned *hpage_shift)
314 {
315 	pgd_t pgd, *pgdp;
316 	pud_t pud, *pudp;
317 	pmd_t pmd, *pmdp;
318 	pte_t *ret_pte;
319 	hugepd_t *hpdp = NULL;
320 	unsigned pdshift = PGDIR_SHIFT;
321 
322 	if (hpage_shift)
323 		*hpage_shift = 0;
324 
325 	if (is_thp)
326 		*is_thp = false;
327 
328 	pgdp = pgdir + pgd_index(ea);
329 	pgd  = READ_ONCE(*pgdp);
330 	/*
331 	 * Always operate on the local stack value. This make sure the
332 	 * value don't get updated by a parallel THP split/collapse,
333 	 * page fault or a page unmap. The return pte_t * is still not
334 	 * stable. So should be checked there for above conditions.
335 	 */
336 	if (pgd_none(pgd))
337 		return NULL;
338 
339 	if (pgd_is_leaf(pgd)) {
340 		ret_pte = (pte_t *)pgdp;
341 		goto out;
342 	}
343 
344 	if (is_hugepd(__hugepd(pgd_val(pgd)))) {
345 		hpdp = (hugepd_t *)&pgd;
346 		goto out_huge;
347 	}
348 
349 	/*
350 	 * Even if we end up with an unmap, the pgtable will not
351 	 * be freed, because we do an rcu free and here we are
352 	 * irq disabled
353 	 */
354 	pdshift = PUD_SHIFT;
355 	pudp = pud_offset(&pgd, ea);
356 	pud  = READ_ONCE(*pudp);
357 
358 	if (pud_none(pud))
359 		return NULL;
360 
361 	if (pud_is_leaf(pud)) {
362 		ret_pte = (pte_t *)pudp;
363 		goto out;
364 	}
365 
366 	if (is_hugepd(__hugepd(pud_val(pud)))) {
367 		hpdp = (hugepd_t *)&pud;
368 		goto out_huge;
369 	}
370 
371 	pdshift = PMD_SHIFT;
372 	pmdp = pmd_offset(&pud, ea);
373 	pmd  = READ_ONCE(*pmdp);
374 
375 	/*
376 	 * A hugepage collapse is captured by this condition, see
377 	 * pmdp_collapse_flush.
378 	 */
379 	if (pmd_none(pmd))
380 		return NULL;
381 
382 #ifdef CONFIG_PPC_BOOK3S_64
383 	/*
384 	 * A hugepage split is captured by this condition, see
385 	 * pmdp_invalidate.
386 	 *
387 	 * Huge page modification can be caught here too.
388 	 */
389 	if (pmd_is_serializing(pmd))
390 		return NULL;
391 #endif
392 
393 	if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
394 		if (is_thp)
395 			*is_thp = true;
396 		ret_pte = (pte_t *)pmdp;
397 		goto out;
398 	}
399 
400 	if (pmd_is_leaf(pmd)) {
401 		ret_pte = (pte_t *)pmdp;
402 		goto out;
403 	}
404 
405 	if (is_hugepd(__hugepd(pmd_val(pmd)))) {
406 		hpdp = (hugepd_t *)&pmd;
407 		goto out_huge;
408 	}
409 
410 	return pte_offset_kernel(&pmd, ea);
411 
412 out_huge:
413 	if (!hpdp)
414 		return NULL;
415 
416 	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
417 	pdshift = hugepd_shift(*hpdp);
418 out:
419 	if (hpage_shift)
420 		*hpage_shift = pdshift;
421 	return ret_pte;
422 }
423 EXPORT_SYMBOL_GPL(__find_linux_pte);
424