xref: /openbmc/linux/arch/x86/mm/pgtable.c (revision 4800cd83)
1 #include <linux/mm.h>
2 #include <linux/gfp.h>
3 #include <asm/pgalloc.h>
4 #include <asm/pgtable.h>
5 #include <asm/tlb.h>
6 #include <asm/fixmap.h>
7 
8 #define PGALLOC_GFP GFP_KERNEL | __GFP_NOTRACK | __GFP_REPEAT | __GFP_ZERO
9 
10 #ifdef CONFIG_HIGHPTE
11 #define PGALLOC_USER_GFP __GFP_HIGHMEM
12 #else
13 #define PGALLOC_USER_GFP 0
14 #endif
15 
16 gfp_t __userpte_alloc_gfp = PGALLOC_GFP | PGALLOC_USER_GFP;
17 
18 pte_t *pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
19 {
20 	return (pte_t *)__get_free_page(PGALLOC_GFP);
21 }
22 
23 pgtable_t pte_alloc_one(struct mm_struct *mm, unsigned long address)
24 {
25 	struct page *pte;
26 
27 	pte = alloc_pages(__userpte_alloc_gfp, 0);
28 	if (pte)
29 		pgtable_page_ctor(pte);
30 	return pte;
31 }
32 
33 static int __init setup_userpte(char *arg)
34 {
35 	if (!arg)
36 		return -EINVAL;
37 
38 	/*
39 	 * "userpte=nohigh" disables allocation of user pagetables in
40 	 * high memory.
41 	 */
42 	if (strcmp(arg, "nohigh") == 0)
43 		__userpte_alloc_gfp &= ~__GFP_HIGHMEM;
44 	else
45 		return -EINVAL;
46 	return 0;
47 }
48 early_param("userpte", setup_userpte);
49 
50 void ___pte_free_tlb(struct mmu_gather *tlb, struct page *pte)
51 {
52 	pgtable_page_dtor(pte);
53 	paravirt_release_pte(page_to_pfn(pte));
54 	tlb_remove_page(tlb, pte);
55 }
56 
57 #if PAGETABLE_LEVELS > 2
58 void ___pmd_free_tlb(struct mmu_gather *tlb, pmd_t *pmd)
59 {
60 	paravirt_release_pmd(__pa(pmd) >> PAGE_SHIFT);
61 	tlb_remove_page(tlb, virt_to_page(pmd));
62 }
63 
64 #if PAGETABLE_LEVELS > 3
65 void ___pud_free_tlb(struct mmu_gather *tlb, pud_t *pud)
66 {
67 	paravirt_release_pud(__pa(pud) >> PAGE_SHIFT);
68 	tlb_remove_page(tlb, virt_to_page(pud));
69 }
70 #endif	/* PAGETABLE_LEVELS > 3 */
71 #endif	/* PAGETABLE_LEVELS > 2 */
72 
73 static inline void pgd_list_add(pgd_t *pgd)
74 {
75 	struct page *page = virt_to_page(pgd);
76 
77 	list_add(&page->lru, &pgd_list);
78 }
79 
80 static inline void pgd_list_del(pgd_t *pgd)
81 {
82 	struct page *page = virt_to_page(pgd);
83 
84 	list_del(&page->lru);
85 }
86 
87 #define UNSHARED_PTRS_PER_PGD				\
88 	(SHARED_KERNEL_PMD ? KERNEL_PGD_BOUNDARY : PTRS_PER_PGD)
89 
90 
91 static void pgd_set_mm(pgd_t *pgd, struct mm_struct *mm)
92 {
93 	BUILD_BUG_ON(sizeof(virt_to_page(pgd)->index) < sizeof(mm));
94 	virt_to_page(pgd)->index = (pgoff_t)mm;
95 }
96 
97 struct mm_struct *pgd_page_get_mm(struct page *page)
98 {
99 	return (struct mm_struct *)page->index;
100 }
101 
102 static void pgd_ctor(struct mm_struct *mm, pgd_t *pgd)
103 {
104 	/* If the pgd points to a shared pagetable level (either the
105 	   ptes in non-PAE, or shared PMD in PAE), then just copy the
106 	   references from swapper_pg_dir. */
107 	if (PAGETABLE_LEVELS == 2 ||
108 	    (PAGETABLE_LEVELS == 3 && SHARED_KERNEL_PMD) ||
109 	    PAGETABLE_LEVELS == 4) {
110 		clone_pgd_range(pgd + KERNEL_PGD_BOUNDARY,
111 				swapper_pg_dir + KERNEL_PGD_BOUNDARY,
112 				KERNEL_PGD_PTRS);
113 	}
114 
115 	/* list required to sync kernel mapping updates */
116 	if (!SHARED_KERNEL_PMD) {
117 		pgd_set_mm(pgd, mm);
118 		pgd_list_add(pgd);
119 	}
120 }
121 
122 static void pgd_dtor(pgd_t *pgd)
123 {
124 	if (SHARED_KERNEL_PMD)
125 		return;
126 
127 	spin_lock(&pgd_lock);
128 	pgd_list_del(pgd);
129 	spin_unlock(&pgd_lock);
130 }
131 
132 /*
133  * List of all pgd's needed for non-PAE so it can invalidate entries
134  * in both cached and uncached pgd's; not needed for PAE since the
135  * kernel pmd is shared. If PAE were not to share the pmd a similar
136  * tactic would be needed. This is essentially codepath-based locking
137  * against pageattr.c; it is the unique case in which a valid change
138  * of kernel pagetables can't be lazily synchronized by vmalloc faults.
139  * vmalloc faults work because attached pagetables are never freed.
140  * -- wli
141  */
142 
143 #ifdef CONFIG_X86_PAE
144 /*
145  * In PAE mode, we need to do a cr3 reload (=tlb flush) when
146  * updating the top-level pagetable entries to guarantee the
147  * processor notices the update.  Since this is expensive, and
148  * all 4 top-level entries are used almost immediately in a
149  * new process's life, we just pre-populate them here.
150  *
151  * Also, if we're in a paravirt environment where the kernel pmd is
152  * not shared between pagetables (!SHARED_KERNEL_PMDS), we allocate
153  * and initialize the kernel pmds here.
154  */
155 #define PREALLOCATED_PMDS	UNSHARED_PTRS_PER_PGD
156 
157 void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmd)
158 {
159 	paravirt_alloc_pmd(mm, __pa(pmd) >> PAGE_SHIFT);
160 
161 	/* Note: almost everything apart from _PAGE_PRESENT is
162 	   reserved at the pmd (PDPT) level. */
163 	set_pud(pudp, __pud(__pa(pmd) | _PAGE_PRESENT));
164 
165 	/*
166 	 * According to Intel App note "TLBs, Paging-Structure Caches,
167 	 * and Their Invalidation", April 2007, document 317080-001,
168 	 * section 8.1: in PAE mode we explicitly have to flush the
169 	 * TLB via cr3 if the top-level pgd is changed...
170 	 */
171 	if (mm == current->active_mm)
172 		write_cr3(read_cr3());
173 }
174 #else  /* !CONFIG_X86_PAE */
175 
176 /* No need to prepopulate any pagetable entries in non-PAE modes. */
177 #define PREALLOCATED_PMDS	0
178 
179 #endif	/* CONFIG_X86_PAE */
180 
181 static void free_pmds(pmd_t *pmds[])
182 {
183 	int i;
184 
185 	for(i = 0; i < PREALLOCATED_PMDS; i++)
186 		if (pmds[i])
187 			free_page((unsigned long)pmds[i]);
188 }
189 
190 static int preallocate_pmds(pmd_t *pmds[])
191 {
192 	int i;
193 	bool failed = false;
194 
195 	for(i = 0; i < PREALLOCATED_PMDS; i++) {
196 		pmd_t *pmd = (pmd_t *)__get_free_page(PGALLOC_GFP);
197 		if (pmd == NULL)
198 			failed = true;
199 		pmds[i] = pmd;
200 	}
201 
202 	if (failed) {
203 		free_pmds(pmds);
204 		return -ENOMEM;
205 	}
206 
207 	return 0;
208 }
209 
210 /*
211  * Mop up any pmd pages which may still be attached to the pgd.
212  * Normally they will be freed by munmap/exit_mmap, but any pmd we
213  * preallocate which never got a corresponding vma will need to be
214  * freed manually.
215  */
216 static void pgd_mop_up_pmds(struct mm_struct *mm, pgd_t *pgdp)
217 {
218 	int i;
219 
220 	for(i = 0; i < PREALLOCATED_PMDS; i++) {
221 		pgd_t pgd = pgdp[i];
222 
223 		if (pgd_val(pgd) != 0) {
224 			pmd_t *pmd = (pmd_t *)pgd_page_vaddr(pgd);
225 
226 			pgdp[i] = native_make_pgd(0);
227 
228 			paravirt_release_pmd(pgd_val(pgd) >> PAGE_SHIFT);
229 			pmd_free(mm, pmd);
230 		}
231 	}
232 }
233 
234 static void pgd_prepopulate_pmd(struct mm_struct *mm, pgd_t *pgd, pmd_t *pmds[])
235 {
236 	pud_t *pud;
237 	unsigned long addr;
238 	int i;
239 
240 	if (PREALLOCATED_PMDS == 0) /* Work around gcc-3.4.x bug */
241 		return;
242 
243 	pud = pud_offset(pgd, 0);
244 
245  	for (addr = i = 0; i < PREALLOCATED_PMDS;
246 	     i++, pud++, addr += PUD_SIZE) {
247 		pmd_t *pmd = pmds[i];
248 
249 		if (i >= KERNEL_PGD_BOUNDARY)
250 			memcpy(pmd, (pmd_t *)pgd_page_vaddr(swapper_pg_dir[i]),
251 			       sizeof(pmd_t) * PTRS_PER_PMD);
252 
253 		pud_populate(mm, pud, pmd);
254 	}
255 }
256 
257 pgd_t *pgd_alloc(struct mm_struct *mm)
258 {
259 	pgd_t *pgd;
260 	pmd_t *pmds[PREALLOCATED_PMDS];
261 
262 	pgd = (pgd_t *)__get_free_page(PGALLOC_GFP);
263 
264 	if (pgd == NULL)
265 		goto out;
266 
267 	mm->pgd = pgd;
268 
269 	if (preallocate_pmds(pmds) != 0)
270 		goto out_free_pgd;
271 
272 	if (paravirt_pgd_alloc(mm) != 0)
273 		goto out_free_pmds;
274 
275 	/*
276 	 * Make sure that pre-populating the pmds is atomic with
277 	 * respect to anything walking the pgd_list, so that they
278 	 * never see a partially populated pgd.
279 	 */
280 	spin_lock(&pgd_lock);
281 
282 	pgd_ctor(mm, pgd);
283 	pgd_prepopulate_pmd(mm, pgd, pmds);
284 
285 	spin_unlock(&pgd_lock);
286 
287 	return pgd;
288 
289 out_free_pmds:
290 	free_pmds(pmds);
291 out_free_pgd:
292 	free_page((unsigned long)pgd);
293 out:
294 	return NULL;
295 }
296 
297 void pgd_free(struct mm_struct *mm, pgd_t *pgd)
298 {
299 	pgd_mop_up_pmds(mm, pgd);
300 	pgd_dtor(pgd);
301 	paravirt_pgd_free(mm, pgd);
302 	free_page((unsigned long)pgd);
303 }
304 
305 int ptep_set_access_flags(struct vm_area_struct *vma,
306 			  unsigned long address, pte_t *ptep,
307 			  pte_t entry, int dirty)
308 {
309 	int changed = !pte_same(*ptep, entry);
310 
311 	if (changed && dirty) {
312 		*ptep = entry;
313 		pte_update_defer(vma->vm_mm, address, ptep);
314 		flush_tlb_page(vma, address);
315 	}
316 
317 	return changed;
318 }
319 
320 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
321 int pmdp_set_access_flags(struct vm_area_struct *vma,
322 			  unsigned long address, pmd_t *pmdp,
323 			  pmd_t entry, int dirty)
324 {
325 	int changed = !pmd_same(*pmdp, entry);
326 
327 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
328 
329 	if (changed && dirty) {
330 		*pmdp = entry;
331 		pmd_update_defer(vma->vm_mm, address, pmdp);
332 		flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
333 	}
334 
335 	return changed;
336 }
337 #endif
338 
339 int ptep_test_and_clear_young(struct vm_area_struct *vma,
340 			      unsigned long addr, pte_t *ptep)
341 {
342 	int ret = 0;
343 
344 	if (pte_young(*ptep))
345 		ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
346 					 (unsigned long *) &ptep->pte);
347 
348 	if (ret)
349 		pte_update(vma->vm_mm, addr, ptep);
350 
351 	return ret;
352 }
353 
354 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
355 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
356 			      unsigned long addr, pmd_t *pmdp)
357 {
358 	int ret = 0;
359 
360 	if (pmd_young(*pmdp))
361 		ret = test_and_clear_bit(_PAGE_BIT_ACCESSED,
362 					 (unsigned long *)pmdp);
363 
364 	if (ret)
365 		pmd_update(vma->vm_mm, addr, pmdp);
366 
367 	return ret;
368 }
369 #endif
370 
371 int ptep_clear_flush_young(struct vm_area_struct *vma,
372 			   unsigned long address, pte_t *ptep)
373 {
374 	int young;
375 
376 	young = ptep_test_and_clear_young(vma, address, ptep);
377 	if (young)
378 		flush_tlb_page(vma, address);
379 
380 	return young;
381 }
382 
383 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
384 int pmdp_clear_flush_young(struct vm_area_struct *vma,
385 			   unsigned long address, pmd_t *pmdp)
386 {
387 	int young;
388 
389 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
390 
391 	young = pmdp_test_and_clear_young(vma, address, pmdp);
392 	if (young)
393 		flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
394 
395 	return young;
396 }
397 
398 void pmdp_splitting_flush(struct vm_area_struct *vma,
399 			  unsigned long address, pmd_t *pmdp)
400 {
401 	int set;
402 	VM_BUG_ON(address & ~HPAGE_PMD_MASK);
403 	set = !test_and_set_bit(_PAGE_BIT_SPLITTING,
404 				(unsigned long *)pmdp);
405 	if (set) {
406 		pmd_update(vma->vm_mm, address, pmdp);
407 		/* need tlb flush only to serialize against gup-fast */
408 		flush_tlb_range(vma, address, address + HPAGE_PMD_SIZE);
409 	}
410 }
411 #endif
412 
413 /**
414  * reserve_top_address - reserves a hole in the top of kernel address space
415  * @reserve - size of hole to reserve
416  *
417  * Can be used to relocate the fixmap area and poke a hole in the top
418  * of kernel address space to make room for a hypervisor.
419  */
420 void __init reserve_top_address(unsigned long reserve)
421 {
422 #ifdef CONFIG_X86_32
423 	BUG_ON(fixmaps_set > 0);
424 	printk(KERN_INFO "Reserving virtual address space above 0x%08x\n",
425 	       (int)-reserve);
426 	__FIXADDR_TOP = -reserve - PAGE_SIZE;
427 #endif
428 }
429 
430 int fixmaps_set;
431 
432 void __native_set_fixmap(enum fixed_addresses idx, pte_t pte)
433 {
434 	unsigned long address = __fix_to_virt(idx);
435 
436 	if (idx >= __end_of_fixed_addresses) {
437 		BUG();
438 		return;
439 	}
440 	set_pte_vaddr(address, pte);
441 	fixmaps_set++;
442 }
443 
444 void native_set_fixmap(enum fixed_addresses idx, phys_addr_t phys,
445 		       pgprot_t flags)
446 {
447 	__native_set_fixmap(idx, pfn_pte(phys >> PAGE_SHIFT, flags));
448 }
449