xref: /openbmc/linux/arch/powerpc/mm/pgtable.c (revision 2f0754f2)
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 	pte = __pte(pte_val(pte) & ~_PAGE_HPTEFLAGS);
85 	if (pte_looks_normal(pte) && !(cpu_has_feature(CPU_FTR_COHERENT_ICACHE) ||
86 				       cpu_has_feature(CPU_FTR_NOEXECUTE))) {
87 		struct page *pg = maybe_pte_to_page(pte);
88 		if (!pg)
89 			return pte;
90 		if (!test_bit(PG_dcache_clean, &pg->flags)) {
91 			flush_dcache_icache_page(pg);
92 			set_bit(PG_dcache_clean, &pg->flags);
93 		}
94 	}
95 	return pte;
96 }
97 
98 #else /* CONFIG_PPC_BOOK3S */
99 
100 static pte_t set_pte_filter_hash(pte_t pte) { return pte; }
101 
102 #endif /* CONFIG_PPC_BOOK3S */
103 
104 /* Embedded type MMU with HW exec support. This is a bit more complicated
105  * as we don't have two bits to spare for _PAGE_EXEC and _PAGE_HWEXEC so
106  * instead we "filter out" the exec permission for non clean pages.
107  */
108 static inline pte_t set_pte_filter(pte_t pte)
109 {
110 	struct page *pg;
111 
112 	if (radix_enabled())
113 		return pte;
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 (IS_ENABLED(CONFIG_PPC_BOOK3S_64))
148 		return pte;
149 
150 	if (mmu_has_feature(MMU_FTR_HPTE_TABLE))
151 		return pte;
152 
153 	/* So here, we only care about exec faults, as we use them
154 	 * to recover lost _PAGE_EXEC and perform I$/D$ coherency
155 	 * if necessary. Also if _PAGE_EXEC is already set, same deal,
156 	 * we just bail out
157 	 */
158 	if (dirty || pte_exec(pte) || !is_exec_fault())
159 		return pte;
160 
161 #ifdef CONFIG_DEBUG_VM
162 	/* So this is an exec fault, _PAGE_EXEC is not set. If it was
163 	 * an error we would have bailed out earlier in do_page_fault()
164 	 * but let's make sure of it
165 	 */
166 	if (WARN_ON(!(vma->vm_flags & VM_EXEC)))
167 		return pte;
168 #endif /* CONFIG_DEBUG_VM */
169 
170 	/* If you set _PAGE_EXEC on weird pages you're on your own */
171 	pg = maybe_pte_to_page(pte);
172 	if (unlikely(!pg))
173 		goto bail;
174 
175 	/* If the page is already clean, we move on */
176 	if (test_bit(PG_dcache_clean, &pg->flags))
177 		goto bail;
178 
179 	/* Clean the page and set PG_dcache_clean */
180 	flush_dcache_icache_page(pg);
181 	set_bit(PG_dcache_clean, &pg->flags);
182 
183  bail:
184 	return pte_mkexec(pte);
185 }
186 
187 /*
188  * set_pte stores a linux PTE into the linux page table.
189  */
190 void set_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep,
191 		pte_t pte)
192 {
193 	/*
194 	 * Make sure hardware valid bit is not set. We don't do
195 	 * tlb flush for this update.
196 	 */
197 	VM_WARN_ON(pte_hw_valid(*ptep) && !pte_protnone(*ptep));
198 
199 	/* Note: mm->context.id might not yet have been assigned as
200 	 * this context might not have been activated yet when this
201 	 * is called.
202 	 */
203 	pte = set_pte_filter(pte);
204 
205 	/* Perform the setting of the PTE */
206 	__set_pte_at(mm, addr, ptep, pte, 0);
207 }
208 
209 /*
210  * This is called when relaxing access to a PTE. It's also called in the page
211  * fault path when we don't hit any of the major fault cases, ie, a minor
212  * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
213  * handled those two for us, we additionally deal with missing execute
214  * permission here on some processors
215  */
216 int ptep_set_access_flags(struct vm_area_struct *vma, unsigned long address,
217 			  pte_t *ptep, pte_t entry, int dirty)
218 {
219 	int changed;
220 	entry = set_access_flags_filter(entry, vma, dirty);
221 	changed = !pte_same(*(ptep), entry);
222 	if (changed) {
223 		assert_pte_locked(vma->vm_mm, address);
224 		__ptep_set_access_flags(vma, ptep, entry,
225 					address, mmu_virtual_psize);
226 	}
227 	return changed;
228 }
229 
230 #ifdef CONFIG_HUGETLB_PAGE
231 int huge_ptep_set_access_flags(struct vm_area_struct *vma,
232 			       unsigned long addr, pte_t *ptep,
233 			       pte_t pte, int dirty)
234 {
235 #ifdef HUGETLB_NEED_PRELOAD
236 	/*
237 	 * The "return 1" forces a call of update_mmu_cache, which will write a
238 	 * TLB entry.  Without this, platforms that don't do a write of the TLB
239 	 * entry in the TLB miss handler asm will fault ad infinitum.
240 	 */
241 	ptep_set_access_flags(vma, addr, ptep, pte, dirty);
242 	return 1;
243 #else
244 	int changed, psize;
245 
246 	pte = set_access_flags_filter(pte, vma, dirty);
247 	changed = !pte_same(*(ptep), pte);
248 	if (changed) {
249 
250 #ifdef CONFIG_PPC_BOOK3S_64
251 		struct hstate *h = hstate_vma(vma);
252 
253 		psize = hstate_get_psize(h);
254 #ifdef CONFIG_DEBUG_VM
255 		assert_spin_locked(huge_pte_lockptr(h, vma->vm_mm, ptep));
256 #endif
257 
258 #else
259 		/*
260 		 * Not used on non book3s64 platforms.
261 		 * 8xx compares it with mmu_virtual_psize to
262 		 * know if it is a huge page or not.
263 		 */
264 		psize = MMU_PAGE_COUNT;
265 #endif
266 		__ptep_set_access_flags(vma, ptep, pte, addr, psize);
267 	}
268 	return changed;
269 #endif
270 }
271 
272 #if defined(CONFIG_PPC_8xx)
273 void set_huge_pte_at(struct mm_struct *mm, unsigned long addr, pte_t *ptep, pte_t pte)
274 {
275 	pmd_t *pmd = pmd_off(mm, addr);
276 	pte_basic_t val;
277 	pte_basic_t *entry = (pte_basic_t *)ptep;
278 	int num, i;
279 
280 	/*
281 	 * Make sure hardware valid bit is not set. We don't do
282 	 * tlb flush for this update.
283 	 */
284 	VM_WARN_ON(pte_hw_valid(*ptep) && !pte_protnone(*ptep));
285 
286 	pte = set_pte_filter(pte);
287 
288 	val = pte_val(pte);
289 
290 	num = number_of_cells_per_pte(pmd, val, 1);
291 
292 	for (i = 0; i < num; i++, entry++, val += SZ_4K)
293 		*entry = val;
294 }
295 #endif
296 #endif /* CONFIG_HUGETLB_PAGE */
297 
298 #ifdef CONFIG_DEBUG_VM
299 void assert_pte_locked(struct mm_struct *mm, unsigned long addr)
300 {
301 	pgd_t *pgd;
302 	p4d_t *p4d;
303 	pud_t *pud;
304 	pmd_t *pmd;
305 
306 	if (mm == &init_mm)
307 		return;
308 	pgd = mm->pgd + pgd_index(addr);
309 	BUG_ON(pgd_none(*pgd));
310 	p4d = p4d_offset(pgd, addr);
311 	BUG_ON(p4d_none(*p4d));
312 	pud = pud_offset(p4d, addr);
313 	BUG_ON(pud_none(*pud));
314 	pmd = pmd_offset(pud, addr);
315 	/*
316 	 * khugepaged to collapse normal pages to hugepage, first set
317 	 * pmd to none to force page fault/gup to take mmap_lock. After
318 	 * pmd is set to none, we do a pte_clear which does this assertion
319 	 * so if we find pmd none, return.
320 	 */
321 	if (pmd_none(*pmd))
322 		return;
323 	BUG_ON(!pmd_present(*pmd));
324 	assert_spin_locked(pte_lockptr(mm, pmd));
325 }
326 #endif /* CONFIG_DEBUG_VM */
327 
328 unsigned long vmalloc_to_phys(void *va)
329 {
330 	unsigned long pfn = vmalloc_to_pfn(va);
331 
332 	BUG_ON(!pfn);
333 	return __pa(pfn_to_kaddr(pfn)) + offset_in_page(va);
334 }
335 EXPORT_SYMBOL_GPL(vmalloc_to_phys);
336 
337 /*
338  * We have 4 cases for pgds and pmds:
339  * (1) invalid (all zeroes)
340  * (2) pointer to next table, as normal; bottom 6 bits == 0
341  * (3) leaf pte for huge page _PAGE_PTE set
342  * (4) hugepd pointer, _PAGE_PTE = 0 and bits [2..6] indicate size of table
343  *
344  * So long as we atomically load page table pointers we are safe against teardown,
345  * we can follow the address down to the the page and take a ref on it.
346  * This function need to be called with interrupts disabled. We use this variant
347  * when we have MSR[EE] = 0 but the paca->irq_soft_mask = IRQS_ENABLED
348  */
349 pte_t *__find_linux_pte(pgd_t *pgdir, unsigned long ea,
350 			bool *is_thp, unsigned *hpage_shift)
351 {
352 	pgd_t *pgdp;
353 	p4d_t p4d, *p4dp;
354 	pud_t pud, *pudp;
355 	pmd_t pmd, *pmdp;
356 	pte_t *ret_pte;
357 	hugepd_t *hpdp = NULL;
358 	unsigned pdshift;
359 
360 	if (hpage_shift)
361 		*hpage_shift = 0;
362 
363 	if (is_thp)
364 		*is_thp = false;
365 
366 	/*
367 	 * Always operate on the local stack value. This make sure the
368 	 * value don't get updated by a parallel THP split/collapse,
369 	 * page fault or a page unmap. The return pte_t * is still not
370 	 * stable. So should be checked there for above conditions.
371 	 * Top level is an exception because it is folded into p4d.
372 	 */
373 	pgdp = pgdir + pgd_index(ea);
374 	p4dp = p4d_offset(pgdp, ea);
375 	p4d  = READ_ONCE(*p4dp);
376 	pdshift = P4D_SHIFT;
377 
378 	if (p4d_none(p4d))
379 		return NULL;
380 
381 	if (p4d_is_leaf(p4d)) {
382 		ret_pte = (pte_t *)p4dp;
383 		goto out;
384 	}
385 
386 	if (is_hugepd(__hugepd(p4d_val(p4d)))) {
387 		hpdp = (hugepd_t *)&p4d;
388 		goto out_huge;
389 	}
390 
391 	/*
392 	 * Even if we end up with an unmap, the pgtable will not
393 	 * be freed, because we do an rcu free and here we are
394 	 * irq disabled
395 	 */
396 	pdshift = PUD_SHIFT;
397 	pudp = pud_offset(&p4d, ea);
398 	pud  = READ_ONCE(*pudp);
399 
400 	if (pud_none(pud))
401 		return NULL;
402 
403 	if (pud_is_leaf(pud)) {
404 		ret_pte = (pte_t *)pudp;
405 		goto out;
406 	}
407 
408 	if (is_hugepd(__hugepd(pud_val(pud)))) {
409 		hpdp = (hugepd_t *)&pud;
410 		goto out_huge;
411 	}
412 
413 	pdshift = PMD_SHIFT;
414 	pmdp = pmd_offset(&pud, ea);
415 	pmd  = READ_ONCE(*pmdp);
416 
417 	/*
418 	 * A hugepage collapse is captured by this condition, see
419 	 * pmdp_collapse_flush.
420 	 */
421 	if (pmd_none(pmd))
422 		return NULL;
423 
424 #ifdef CONFIG_PPC_BOOK3S_64
425 	/*
426 	 * A hugepage split is captured by this condition, see
427 	 * pmdp_invalidate.
428 	 *
429 	 * Huge page modification can be caught here too.
430 	 */
431 	if (pmd_is_serializing(pmd))
432 		return NULL;
433 #endif
434 
435 	if (pmd_trans_huge(pmd) || pmd_devmap(pmd)) {
436 		if (is_thp)
437 			*is_thp = true;
438 		ret_pte = (pte_t *)pmdp;
439 		goto out;
440 	}
441 
442 	if (pmd_is_leaf(pmd)) {
443 		ret_pte = (pte_t *)pmdp;
444 		goto out;
445 	}
446 
447 	if (is_hugepd(__hugepd(pmd_val(pmd)))) {
448 		hpdp = (hugepd_t *)&pmd;
449 		goto out_huge;
450 	}
451 
452 	return pte_offset_kernel(&pmd, ea);
453 
454 out_huge:
455 	if (!hpdp)
456 		return NULL;
457 
458 	ret_pte = hugepte_offset(*hpdp, ea, pdshift);
459 	pdshift = hugepd_shift(*hpdp);
460 out:
461 	if (hpage_shift)
462 		*hpage_shift = pdshift;
463 	return ret_pte;
464 }
465 EXPORT_SYMBOL_GPL(__find_linux_pte);
466