xref: /openbmc/linux/arch/powerpc/mm/pgtable.c (revision 240e6d25)
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 #include <asm/pte-walk.h>
30 
31 #ifdef CONFIG_PPC64
32 #define PGD_ALIGN (sizeof(pgd_t) * MAX_PTRS_PER_PGD)
33 #else
34 #define PGD_ALIGN PAGE_SIZE
35 #endif
36 
37 pgd_t swapper_pg_dir[MAX_PTRS_PER_PGD] __section(".bss..page_aligned") __aligned(PGD_ALIGN);
38 
39 static inline int is_exec_fault(void)
40 {
41 	return current->thread.regs && TRAP(current->thread.regs) == 0x400;
42 }
43 
44 /* We only try to do i/d cache coherency on stuff that looks like
45  * reasonably "normal" PTEs. We currently require a PTE to be present
46  * and we avoid _PAGE_SPECIAL and cache inhibited pte. We also only do that
47  * on userspace PTEs
48  */
49 static inline int pte_looks_normal(pte_t pte)
50 {
51 
52 	if (pte_present(pte) && !pte_special(pte)) {
53 		if (pte_ci(pte))
54 			return 0;
55 		if (pte_user(pte))
56 			return 1;
57 	}
58 	return 0;
59 }
60 
61 static struct page *maybe_pte_to_page(pte_t pte)
62 {
63 	unsigned long pfn = pte_pfn(pte);
64 	struct page *page;
65 
66 	if (unlikely(!pfn_valid(pfn)))
67 		return NULL;
68 	page = pfn_to_page(pfn);
69 	if (PageReserved(page))
70 		return NULL;
71 	return page;
72 }
73 
74 #ifdef CONFIG_PPC_BOOK3S
75 
76 /* Server-style MMU handles coherency when hashing if HW exec permission
77  * is supposed per page (currently 64-bit only). If not, then, we always
78  * flush the cache for valid PTEs in set_pte. Embedded CPU without HW exec
79  * support falls into the same category.
80  */
81 
82 static pte_t set_pte_filter_hash(pte_t pte)
83 {
84 	if (radix_enabled())
85 		return pte;
86 
87 	pte = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
88 	if (pte_looks_normal(pte) && !(cpu_has_feature(CPU_FTR_COHERENT_ICACHE) ||
89 				       cpu_has_feature(CPU_FTR_NOEXECUTE))) {
90 		struct page *pg = maybe_pte_to_page(pte);
91 		if (!pg)
92 			return pte;
93 		if (!test_bit(PG_dcache_clean, &pg->flags)) {
94 			flush_dcache_icache_page(pg);
95 			set_bit(PG_dcache_clean, &pg->flags);
96 		}
97 	}
98 	return pte;
99 }
100 
101 #else /* CONFIG_PPC_BOOK3S */
102 
103 static pte_t set_pte_filter_hash(pte_t pte) { return pte; }
104 
105 #endif /* CONFIG_PPC_BOOK3S */
106 
107 /* Embedded type MMU with HW exec support. This is a bit more complicated
108  * as we don't have two bits to spare for _PAGE_EXEC and _PAGE_HWEXEC so
109  * instead we "filter out" the exec permission for non clean pages.
110  */
111 static inline pte_t set_pte_filter(pte_t pte)
112 {
113 	struct page *pg;
114 
115 	if (mmu_has_feature(MMU_FTR_HPTE_TABLE))
116 		return set_pte_filter_hash(pte);
117 
118 	/* No exec permission in the first place, move on */
119 	if (!pte_exec(pte) || !pte_looks_normal(pte))
120 		return pte;
121 
122 	/* If you set _PAGE_EXEC on weird pages you're on your own */
123 	pg = maybe_pte_to_page(pte);
124 	if (unlikely(!pg))
125 		return pte;
126 
127 	/* If the page clean, we move on */
128 	if (test_bit(PG_dcache_clean, &pg->flags))
129 		return pte;
130 
131 	/* If it's an exec fault, we flush the cache and make it clean */
132 	if (is_exec_fault()) {
133 		flush_dcache_icache_page(pg);
134 		set_bit(PG_dcache_clean, &pg->flags);
135 		return pte;
136 	}
137 
138 	/* Else, we filter out _PAGE_EXEC */
139 	return pte_exprotect(pte);
140 }
141 
142 static pte_t set_access_flags_filter(pte_t pte, struct vm_area_struct *vma,
143 				     int dirty)
144 {
145 	struct page *pg;
146 
147 	if (mmu_has_feature(MMU_FTR_HPTE_TABLE))
148 		return pte;
149 
150 	/* So here, we only care about exec faults, as we use them
151 	 * to recover lost _PAGE_EXEC and perform I$/D$ coherency
152 	 * if necessary. Also if _PAGE_EXEC is already set, same deal,
153 	 * we just bail out
154 	 */
155 	if (dirty || pte_exec(pte) || !is_exec_fault())
156 		return pte;
157 
158 #ifdef CONFIG_DEBUG_VM
159 	/* So this is an exec fault, _PAGE_EXEC is not set. If it was
160 	 * an error we would have bailed out earlier in do_page_fault()
161 	 * but let's make sure of it
162 	 */
163 	if (WARN_ON(!(vma->vm_flags & VM_EXEC)))
164 		return pte;
165 #endif /* CONFIG_DEBUG_VM */
166 
167 	/* If you set _PAGE_EXEC on weird pages you're on your own */
168 	pg = maybe_pte_to_page(pte);
169 	if (unlikely(!pg))
170 		goto bail;
171 
172 	/* If the page is already clean, we move on */
173 	if (test_bit(PG_dcache_clean, &pg->flags))
174 		goto bail;
175 
176 	/* Clean the page and set PG_dcache_clean */
177 	flush_dcache_icache_page(pg);
178 	set_bit(PG_dcache_clean, &pg->flags);
179 
180  bail:
181 	return pte_mkexec(pte);
182 }
183 
184 /*
185  * set_pte stores a linux PTE into the linux page table.
186  */
187 void set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep,
188 		pte_t pte)
189 {
190 	/*
191 	 * Make sure hardware valid bit is not set. We don't do
192 	 * tlb flush for this update.
193 	 */
194 	VM_WARN_ON(pte_hw_valid(*ptep) && !pte_protnone(*ptep));
195 
196 	/* Note: mm->context.id might not yet have been assigned as
197 	 * this context might not have been activated yet when this
198 	 * is called.
199 	 */
200 	pte = set_pte_filter(pte);
201 
202 	/* Perform the setting of the PTE */
203 	__set_pte_at(mm, addr, ptep, pte, 0);
204 }
205 
206 /*
207  * This is called when relaxing access to a PTE. It's also called in the page
208  * fault path when we don't hit any of the major fault cases, ie, a minor
209  * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
210  * handled those two for us, we additionally deal with missing execute
211  * permission here on some processors
212  */
213 int ptep_set_access_flags(struct vm_area_struct *vma, unsigned long address,
214 			  pte_t *ptep, pte_t entry, int dirty)
215 {
216 	int changed;
217 	entry = set_access_flags_filter(entry, vma, dirty);
218 	changed = !pte_same(*(ptep), entry);
219 	if (changed) {
220 		assert_pte_locked(vma->vm_mm, address);
221 		__ptep_set_access_flags(vma, ptep, entry,
222 					address, mmu_virtual_psize);
223 	}
224 	return changed;
225 }
226 
227 #ifdef CONFIG_HUGETLB_PAGE
228 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
229 			       unsigned long addr, pte_t *ptep,
230 			       pte_t pte, int dirty)
231 {
232 #ifdef HUGETLB_NEED_PRELOAD
233 	/*
234 	 * The "return 1" forces a call of update_mmu_cache, which will write a
235 	 * TLB entry.  Without this, platforms that don't do a write of the TLB
236 	 * entry in the TLB miss handler asm will fault ad infinitum.
237 	 */
238 	ptep_set_access_flags(vma, addr, ptep, pte, dirty);
239 	return 1;
240 #else
241 	int changed, psize;
242 
243 	pte = set_access_flags_filter(pte, vma, dirty);
244 	changed = !pte_same(*(ptep), pte);
245 	if (changed) {
246 
247 #ifdef CONFIG_PPC_BOOK3S_64
248 		struct hstate *h = hstate_vma(vma);
249 
250 		psize = hstate_get_psize(h);
251 #ifdef CONFIG_DEBUG_VM
252 		assert_spin_locked(huge_pte_lockptr(h, vma->vm_mm, ptep));
253 #endif
254 
255 #else
256 		/*
257 		 * Not used on non book3s64 platforms.
258 		 * 8xx compares it with mmu_virtual_psize to
259 		 * know if it is a huge page or not.
260 		 */
261 		psize = MMU_PAGE_COUNT;
262 #endif
263 		__ptep_set_access_flags(vma, ptep, pte, addr, psize);
264 	}
265 	return changed;
266 #endif
267 }
268 
269 #if defined(CONFIG_PPC_8xx)
270 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
271 {
272 	pmd_t *pmd = pmd_off(mm, addr);
273 	pte_basic_t val;
274 	pte_basic_t *entry = (pte_basic_t *)ptep;
275 	int num, i;
276 
277 	/*
278 	 * Make sure hardware valid bit is not set. We don't do
279 	 * tlb flush for this update.
280 	 */
281 	VM_WARN_ON(pte_hw_valid(*ptep) && !pte_protnone(*ptep));
282 
283 	pte = set_pte_filter(pte);
284 
285 	val = pte_val(pte);
286 
287 	num = number_of_cells_per_pte(pmd, val, 1);
288 
289 	for (i = 0; i < num; i++, entry++, val += SZ_4K)
290 		*entry = val;
291 }
292 #endif
293 #endif /* CONFIG_HUGETLB_PAGE */
294 
295 #ifdef CONFIG_DEBUG_VM
296 void assert_pte_locked(struct mm_struct *mm, unsigned long addr)
297 {
298 	pgd_t *pgd;
299 	p4d_t *p4d;
300 	pud_t *pud;
301 	pmd_t *pmd;
302 
303 	if (mm == &init_mm)
304 		return;
305 	pgd = mm->pgd + pgd_index(addr);
306 	BUG_ON(pgd_none(*pgd));
307 	p4d = p4d_offset(pgd, addr);
308 	BUG_ON(p4d_none(*p4d));
309 	pud = pud_offset(p4d, addr);
310 	BUG_ON(pud_none(*pud));
311 	pmd = pmd_offset(pud, addr);
312 	/*
313 	 * khugepaged to collapse normal pages to hugepage, first set
314 	 * pmd to none to force page fault/gup to take mmap_lock. After
315 	 * pmd is set to none, we do a pte_clear which does this assertion
316 	 * so if we find pmd none, return.
317 	 */
318 	if (pmd_none(*pmd))
319 		return;
320 	BUG_ON(!pmd_present(*pmd));
321 	assert_spin_locked(pte_lockptr(mm, pmd));
322 }
323 #endif /* CONFIG_DEBUG_VM */
324 
325 unsigned long vmalloc_to_phys(void *va)
326 {
327 	unsigned long pfn = vmalloc_to_pfn(va);
328 
329 	BUG_ON(!pfn);
330 	return __pa(pfn_to_kaddr(pfn)) + offset_in_page(va);
331 }
332 EXPORT_SYMBOL_GPL(vmalloc_to_phys);
333 
334 /*
335  * We have 4 cases for pgds and pmds:
336  * (1) invalid (all zeroes)
337  * (2) pointer to next table, as normal; bottom 6 bits == 0
338  * (3) leaf pte for huge page _PAGE_PTE set
339  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
340  *
341  * So long as we atomically load page table pointers we are safe against teardown,
342  * we can follow the address down to the the page and take a ref on it.
343  * This function need to be called with interrupts disabled. We use this variant
344  * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
345  */
346 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
347 			bool *is_thp, unsigned *hpage_shift)
348 {
349 	pgd_t *pgdp;
350 	p4d_t p4d, *p4dp;
351 	pud_t pud, *pudp;
352 	pmd_t pmd, *pmdp;
353 	pte_t *ret_pte;
354 	hugepd_t *hpdp = NULL;
355 	unsigned pdshift;
356 
357 	if (hpage_shift)
358 		*hpage_shift = 0;
359 
360 	if (is_thp)
361 		*is_thp = false;
362 
363 	/*
364 	 * Always operate on the local stack value. This make sure the
365 	 * value don't get updated by a parallel THP split/collapse,
366 	 * page fault or a page unmap. The return pte_t * is still not
367 	 * stable. So should be checked there for above conditions.
368 	 * Top level is an exception because it is folded into p4d.
369 	 */
370 	pgdp = pgdir + pgd_index(ea);
371 	p4dp = p4d_offset(pgdp, ea);
372 	p4d  = READ_ONCE(*p4dp);
373 	pdshift = P4D_SHIFT;
374 
375 	if (p4d_none(p4d))
376 		return NULL;
377 
378 	if (p4d_is_leaf(p4d)) {
379 		ret_pte = (pte_t *)p4dp;
380 		goto out;
381 	}
382 
383 	if (is_hugepd(__hugepd(p4d_val(p4d)))) {
384 		hpdp = (hugepd_t *)&p4d;
385 		goto out_huge;
386 	}
387 
388 	/*
389 	 * Even if we end up with an unmap, the pgtable will not
390 	 * be freed, because we do an rcu free and here we are
391 	 * irq disabled
392 	 */
393 	pdshift = PUD_SHIFT;
394 	pudp = pud_offset(&p4d, ea);
395 	pud  = READ_ONCE(*pudp);
396 
397 	if (pud_none(pud))
398 		return NULL;
399 
400 	if (pud_is_leaf(pud)) {
401 		ret_pte = (pte_t *)pudp;
402 		goto out;
403 	}
404 
405 	if (is_hugepd(__hugepd(pud_val(pud)))) {
406 		hpdp = (hugepd_t *)&pud;
407 		goto out_huge;
408 	}
409 
410 	pdshift = PMD_SHIFT;
411 	pmdp = pmd_offset(&pud, ea);
412 	pmd  = READ_ONCE(*pmdp);
413 
414 	/*
415 	 * A hugepage collapse is captured by this condition, see
416 	 * pmdp_collapse_flush.
417 	 */
418 	if (pmd_none(pmd))
419 		return NULL;
420 
421 #ifdef CONFIG_PPC_BOOK3S_64
422 	/*
423 	 * A hugepage split is captured by this condition, see
424 	 * pmdp_invalidate.
425 	 *
426 	 * Huge page modification can be caught here too.
427 	 */
428 	if (pmd_is_serializing(pmd))
429 		return NULL;
430 #endif
431 
432 	if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
433 		if (is_thp)
434 			*is_thp = true;
435 		ret_pte = (pte_t *)pmdp;
436 		goto out;
437 	}
438 
439 	if (pmd_is_leaf(pmd)) {
440 		ret_pte = (pte_t *)pmdp;
441 		goto out;
442 	}
443 
444 	if (is_hugepd(__hugepd(pmd_val(pmd)))) {
445 		hpdp = (hugepd_t *)&pmd;
446 		goto out_huge;
447 	}
448 
449 	return pte_offset_kernel(&pmd, ea);
450 
451 out_huge:
452 	if (!hpdp)
453 		return NULL;
454 
455 	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
456 	pdshift = hugepd_shift(*hpdp);
457 out:
458 	if (hpage_shift)
459 		*hpage_shift = pdshift;
460 	return ret_pte;
461 }
462 EXPORT_SYMBOL_GPL(__find_linux_pte);
463