Commit | Line | Data |
---|---|---|
3df33f12 AK |
1 | /* |
2 | * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or | |
5 | * modify it under the terms of the GNU General Public License | |
6 | * as published by the Free Software Foundation; either version | |
7 | * 2 of the License, or (at your option) any later version. | |
8 | */ | |
9 | ||
10 | #include <linux/sched.h> | |
589ee628 | 11 | #include <linux/mm_types.h> |
59879d54 | 12 | #include <linux/memblock.h> |
fa4531f7 | 13 | #include <misc/cxl-base.h> |
589ee628 | 14 | |
3df33f12 AK |
15 | #include <asm/pgalloc.h> |
16 | #include <asm/tlb.h> | |
59879d54 AK |
17 | #include <asm/trace.h> |
18 | #include <asm/powernv.h> | |
3df33f12 AK |
19 | |
20 | #include "mmu_decl.h" | |
21 | #include <trace/events/thp.h> | |
22 | ||
8a6c697b AK |
23 | unsigned long __pmd_frag_nr; |
24 | EXPORT_SYMBOL(__pmd_frag_nr); | |
25 | unsigned long __pmd_frag_size_shift; | |
26 | EXPORT_SYMBOL(__pmd_frag_size_shift); | |
27 | ||
eea8148c ME |
28 | int (*register_process_table)(unsigned long base, unsigned long page_size, |
29 | unsigned long tbl_size); | |
30 | ||
3df33f12 AK |
31 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
32 | /* | |
33 | * This is called when relaxing access to a hugepage. It's also called in the page | |
34 | * fault path when we don't hit any of the major fault cases, ie, a minor | |
35 | * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have | |
36 | * handled those two for us, we additionally deal with missing execute | |
37 | * permission here on some processors | |
38 | */ | |
39 | int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address, | |
40 | pmd_t *pmdp, pmd_t entry, int dirty) | |
41 | { | |
42 | int changed; | |
43 | #ifdef CONFIG_DEBUG_VM | |
ebd31197 | 44 | WARN_ON(!pmd_trans_huge(*pmdp) && !pmd_devmap(*pmdp)); |
af60a4cf | 45 | assert_spin_locked(pmd_lockptr(vma->vm_mm, pmdp)); |
3df33f12 AK |
46 | #endif |
47 | changed = !pmd_same(*(pmdp), entry); | |
48 | if (changed) { | |
e4c1112c AK |
49 | /* |
50 | * We can use MMU_PAGE_2M here, because only radix | |
51 | * path look at the psize. | |
52 | */ | |
53 | __ptep_set_access_flags(vma, pmdp_ptep(pmdp), | |
54 | pmd_pte(entry), address, MMU_PAGE_2M); | |
3df33f12 AK |
55 | } |
56 | return changed; | |
57 | } | |
58 | ||
59 | int pmdp_test_and_clear_young(struct vm_area_struct *vma, | |
60 | unsigned long address, pmd_t *pmdp) | |
61 | { | |
62 | return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp); | |
63 | } | |
64 | /* | |
65 | * set a new huge pmd. We should not be called for updating | |
66 | * an existing pmd entry. That should go via pmd_hugepage_update. | |
67 | */ | |
68 | void set_pmd_at(struct mm_struct *mm, unsigned long addr, | |
69 | pmd_t *pmdp, pmd_t pmd) | |
70 | { | |
71 | #ifdef CONFIG_DEBUG_VM | |
da7ad366 AK |
72 | /* |
73 | * Make sure hardware valid bit is not set. We don't do | |
74 | * tlb flush for this update. | |
75 | */ | |
76 | WARN_ON(pte_val(pmd_pte(*pmdp)) & _PAGE_PRESENT); | |
af60a4cf | 77 | assert_spin_locked(pmd_lockptr(mm, pmdp)); |
ae28f17b | 78 | WARN_ON(!(pmd_large(pmd) || pmd_devmap(pmd))); |
3df33f12 AK |
79 | #endif |
80 | trace_hugepage_set_pmd(addr, pmd_val(pmd)); | |
81 | return set_pte_at(mm, addr, pmdp_ptep(pmdp), pmd_pte(pmd)); | |
82 | } | |
fa4531f7 AK |
83 | |
84 | static void do_nothing(void *unused) | |
85 | { | |
86 | ||
87 | } | |
88 | /* | |
89 | * Serialize against find_current_mm_pte which does lock-less | |
90 | * lookup in page tables with local interrupts disabled. For huge pages | |
91 | * it casts pmd_t to pte_t. Since format of pte_t is different from | |
92 | * pmd_t we want to prevent transit from pmd pointing to page table | |
93 | * to pmd pointing to huge page (and back) while interrupts are disabled. | |
94 | * We clear pmd to possibly replace it with page table pointer in | |
95 | * different code paths. So make sure we wait for the parallel | |
96 | * find_current_mm_pte to finish. | |
97 | */ | |
98 | void serialize_against_pte_lookup(struct mm_struct *mm) | |
99 | { | |
100 | smp_mb(); | |
0f4bc093 | 101 | smp_call_function_many(mm_cpumask(mm), do_nothing, NULL, 1); |
fa4531f7 AK |
102 | } |
103 | ||
3df33f12 AK |
104 | /* |
105 | * We use this to invalidate a pmdp entry before switching from a | |
106 | * hugepte to regular pmd entry. | |
107 | */ | |
8cc931e0 | 108 | pmd_t pmdp_invalidate(struct vm_area_struct *vma, unsigned long address, |
3df33f12 AK |
109 | pmd_t *pmdp) |
110 | { | |
8cc931e0 AK |
111 | unsigned long old_pmd; |
112 | ||
da7ad366 | 113 | old_pmd = pmd_hugepage_update(vma->vm_mm, address, pmdp, _PAGE_PRESENT, _PAGE_INVALID); |
d8e91e93 | 114 | flush_pmd_tlb_range(vma, address, address + HPAGE_PMD_SIZE); |
3df33f12 AK |
115 | /* |
116 | * This ensures that generic code that rely on IRQ disabling | |
117 | * to prevent a parallel THP split work as expected. | |
118 | */ | |
fa4531f7 | 119 | serialize_against_pte_lookup(vma->vm_mm); |
8cc931e0 | 120 | return __pmd(old_pmd); |
3df33f12 AK |
121 | } |
122 | ||
123 | static pmd_t pmd_set_protbits(pmd_t pmd, pgprot_t pgprot) | |
124 | { | |
125 | return __pmd(pmd_val(pmd) | pgprot_val(pgprot)); | |
126 | } | |
127 | ||
128 | pmd_t pfn_pmd(unsigned long pfn, pgprot_t pgprot) | |
129 | { | |
130 | unsigned long pmdv; | |
131 | ||
132 | pmdv = (pfn << PAGE_SHIFT) & PTE_RPN_MASK; | |
133 | return pmd_set_protbits(__pmd(pmdv), pgprot); | |
134 | } | |
135 | ||
136 | pmd_t mk_pmd(struct page *page, pgprot_t pgprot) | |
137 | { | |
138 | return pfn_pmd(page_to_pfn(page), pgprot); | |
139 | } | |
140 | ||
141 | pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot) | |
142 | { | |
143 | unsigned long pmdv; | |
144 | ||
145 | pmdv = pmd_val(pmd); | |
146 | pmdv &= _HPAGE_CHG_MASK; | |
147 | return pmd_set_protbits(__pmd(pmdv), newprot); | |
148 | } | |
149 | ||
150 | /* | |
151 | * This is called at the end of handling a user page fault, when the | |
152 | * fault has been handled by updating a HUGE PMD entry in the linux page tables. | |
153 | * We use it to preload an HPTE into the hash table corresponding to | |
154 | * the updated linux HUGE PMD entry. | |
155 | */ | |
156 | void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr, | |
157 | pmd_t *pmd) | |
158 | { | |
68662f85 NP |
159 | if (radix_enabled()) |
160 | prefetch((void *)addr); | |
3df33f12 AK |
161 | } |
162 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ | |
fe036a06 BH |
163 | |
164 | /* For use by kexec */ | |
165 | void mmu_cleanup_all(void) | |
166 | { | |
167 | if (radix_enabled()) | |
168 | radix__mmu_cleanup_all(); | |
169 | else if (mmu_hash_ops.hpte_clear_all) | |
170 | mmu_hash_ops.hpte_clear_all(); | |
171 | } | |
32b53c01 RA |
172 | |
173 | #ifdef CONFIG_MEMORY_HOTPLUG | |
f437c517 | 174 | int __meminit create_section_mapping(unsigned long start, unsigned long end, int nid) |
32b53c01 RA |
175 | { |
176 | if (radix_enabled()) | |
29ab6c47 | 177 | return radix__create_section_mapping(start, end, nid); |
32b53c01 | 178 | |
29ab6c47 | 179 | return hash__create_section_mapping(start, end, nid); |
32b53c01 RA |
180 | } |
181 | ||
bde709a7 | 182 | int __meminit remove_section_mapping(unsigned long start, unsigned long end) |
32b53c01 RA |
183 | { |
184 | if (radix_enabled()) | |
4b5d62ca | 185 | return radix__remove_section_mapping(start, end); |
32b53c01 RA |
186 | |
187 | return hash__remove_section_mapping(start, end); | |
188 | } | |
189 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
59879d54 AK |
190 | |
191 | void __init mmu_partition_table_init(void) | |
192 | { | |
193 | unsigned long patb_size = 1UL << PATB_SIZE_SHIFT; | |
194 | unsigned long ptcr; | |
195 | ||
196 | BUILD_BUG_ON_MSG((PATB_SIZE_SHIFT > 36), "Partition table size too large."); | |
197 | partition_tb = __va(memblock_alloc_base(patb_size, patb_size, | |
198 | MEMBLOCK_ALLOC_ANYWHERE)); | |
199 | ||
200 | /* Initialize the Partition Table with no entries */ | |
201 | memset((void *)partition_tb, 0, patb_size); | |
202 | ||
203 | /* | |
204 | * update partition table control register, | |
205 | * 64 K size. | |
206 | */ | |
207 | ptcr = __pa(partition_tb) | (PATB_SIZE_SHIFT - 12); | |
208 | mtspr(SPRN_PTCR, ptcr); | |
209 | powernv_set_nmmu_ptcr(ptcr); | |
210 | } | |
211 | ||
212 | void mmu_partition_table_set_entry(unsigned int lpid, unsigned long dw0, | |
213 | unsigned long dw1) | |
214 | { | |
215 | unsigned long old = be64_to_cpu(partition_tb[lpid].patb0); | |
216 | ||
217 | partition_tb[lpid].patb0 = cpu_to_be64(dw0); | |
218 | partition_tb[lpid].patb1 = cpu_to_be64(dw1); | |
219 | ||
220 | /* | |
221 | * Global flush of TLBs and partition table caches for this lpid. | |
222 | * The type of flush (hash or radix) depends on what the previous | |
223 | * use of this partition ID was, not the new use. | |
224 | */ | |
225 | asm volatile("ptesync" : : : "memory"); | |
226 | if (old & PATB_HR) { | |
227 | asm volatile(PPC_TLBIE_5(%0,%1,2,0,1) : : | |
228 | "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid)); | |
229 | asm volatile(PPC_TLBIE_5(%0,%1,2,1,1) : : | |
230 | "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid)); | |
231 | trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 1); | |
232 | } else { | |
233 | asm volatile(PPC_TLBIE_5(%0,%1,2,0,0) : : | |
234 | "r" (TLBIEL_INVAL_SET_LPID), "r" (lpid)); | |
235 | trace_tlbie(lpid, 0, TLBIEL_INVAL_SET_LPID, lpid, 2, 0, 0); | |
236 | } | |
237 | /* do we need fixup here ?*/ | |
238 | asm volatile("eieio; tlbsync; ptesync" : : : "memory"); | |
239 | } | |
240 | EXPORT_SYMBOL_GPL(mmu_partition_table_set_entry); | |
1c7ec8a4 | 241 | |
8a6c697b AK |
242 | static pmd_t *get_pmd_from_cache(struct mm_struct *mm) |
243 | { | |
244 | void *pmd_frag, *ret; | |
245 | ||
246 | spin_lock(&mm->page_table_lock); | |
247 | ret = mm->context.pmd_frag; | |
248 | if (ret) { | |
249 | pmd_frag = ret + PMD_FRAG_SIZE; | |
250 | /* | |
251 | * If we have taken up all the fragments mark PTE page NULL | |
252 | */ | |
253 | if (((unsigned long)pmd_frag & ~PAGE_MASK) == 0) | |
254 | pmd_frag = NULL; | |
255 | mm->context.pmd_frag = pmd_frag; | |
256 | } | |
257 | spin_unlock(&mm->page_table_lock); | |
258 | return (pmd_t *)ret; | |
259 | } | |
260 | ||
261 | static pmd_t *__alloc_for_pmdcache(struct mm_struct *mm) | |
262 | { | |
263 | void *ret = NULL; | |
264 | struct page *page; | |
265 | gfp_t gfp = GFP_KERNEL_ACCOUNT | __GFP_ZERO; | |
266 | ||
267 | if (mm == &init_mm) | |
268 | gfp &= ~__GFP_ACCOUNT; | |
269 | page = alloc_page(gfp); | |
270 | if (!page) | |
271 | return NULL; | |
272 | if (!pgtable_pmd_page_ctor(page)) { | |
273 | __free_pages(page, 0); | |
274 | return NULL; | |
275 | } | |
276 | ||
4231aba0 NP |
277 | atomic_set(&page->pt_frag_refcount, 1); |
278 | ||
8a6c697b AK |
279 | ret = page_address(page); |
280 | /* | |
281 | * if we support only one fragment just return the | |
282 | * allocated page. | |
283 | */ | |
284 | if (PMD_FRAG_NR == 1) | |
285 | return ret; | |
286 | ||
287 | spin_lock(&mm->page_table_lock); | |
288 | /* | |
289 | * If we find pgtable_page set, we return | |
290 | * the allocated page with single fragement | |
291 | * count. | |
292 | */ | |
293 | if (likely(!mm->context.pmd_frag)) { | |
4231aba0 | 294 | atomic_set(&page->pt_frag_refcount, PMD_FRAG_NR); |
8a6c697b AK |
295 | mm->context.pmd_frag = ret + PMD_FRAG_SIZE; |
296 | } | |
297 | spin_unlock(&mm->page_table_lock); | |
298 | ||
299 | return (pmd_t *)ret; | |
300 | } | |
301 | ||
302 | pmd_t *pmd_fragment_alloc(struct mm_struct *mm, unsigned long vmaddr) | |
303 | { | |
304 | pmd_t *pmd; | |
305 | ||
306 | pmd = get_pmd_from_cache(mm); | |
307 | if (pmd) | |
308 | return pmd; | |
309 | ||
310 | return __alloc_for_pmdcache(mm); | |
311 | } | |
312 | ||
313 | void pmd_fragment_free(unsigned long *pmd) | |
314 | { | |
315 | struct page *page = virt_to_page(pmd); | |
316 | ||
4231aba0 NP |
317 | BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0); |
318 | if (atomic_dec_and_test(&page->pt_frag_refcount)) { | |
8a6c697b | 319 | pgtable_pmd_page_dtor(page); |
4231aba0 | 320 | __free_page(page); |
8a6c697b AK |
321 | } |
322 | } | |
323 | ||
70234676 AK |
324 | static pte_t *get_pte_from_cache(struct mm_struct *mm) |
325 | { | |
326 | void *pte_frag, *ret; | |
327 | ||
328 | spin_lock(&mm->page_table_lock); | |
329 | ret = mm->context.pte_frag; | |
330 | if (ret) { | |
331 | pte_frag = ret + PTE_FRAG_SIZE; | |
332 | /* | |
333 | * If we have taken up all the fragments mark PTE page NULL | |
334 | */ | |
335 | if (((unsigned long)pte_frag & ~PAGE_MASK) == 0) | |
336 | pte_frag = NULL; | |
337 | mm->context.pte_frag = pte_frag; | |
338 | } | |
339 | spin_unlock(&mm->page_table_lock); | |
340 | return (pte_t *)ret; | |
341 | } | |
342 | ||
343 | static pte_t *__alloc_for_ptecache(struct mm_struct *mm, int kernel) | |
344 | { | |
345 | void *ret = NULL; | |
346 | struct page *page; | |
347 | ||
348 | if (!kernel) { | |
349 | page = alloc_page(PGALLOC_GFP | __GFP_ACCOUNT); | |
350 | if (!page) | |
351 | return NULL; | |
352 | if (!pgtable_page_ctor(page)) { | |
353 | __free_page(page); | |
354 | return NULL; | |
355 | } | |
356 | } else { | |
357 | page = alloc_page(PGALLOC_GFP); | |
358 | if (!page) | |
359 | return NULL; | |
360 | } | |
361 | ||
4231aba0 | 362 | atomic_set(&page->pt_frag_refcount, 1); |
1c7ec8a4 | 363 | |
70234676 | 364 | ret = page_address(page); |
1c7ec8a4 AK |
365 | /* |
366 | * if we support only one fragment just return the | |
367 | * allocated page. | |
368 | */ | |
369 | if (PTE_FRAG_NR == 1) | |
370 | return ret; | |
70234676 AK |
371 | spin_lock(&mm->page_table_lock); |
372 | /* | |
373 | * If we find pgtable_page set, we return | |
374 | * the allocated page with single fragement | |
375 | * count. | |
376 | */ | |
377 | if (likely(!mm->context.pte_frag)) { | |
4231aba0 | 378 | atomic_set(&page->pt_frag_refcount, PTE_FRAG_NR); |
70234676 AK |
379 | mm->context.pte_frag = ret + PTE_FRAG_SIZE; |
380 | } | |
381 | spin_unlock(&mm->page_table_lock); | |
382 | ||
383 | return (pte_t *)ret; | |
384 | } | |
385 | ||
386 | pte_t *pte_fragment_alloc(struct mm_struct *mm, unsigned long vmaddr, int kernel) | |
387 | { | |
388 | pte_t *pte; | |
389 | ||
390 | pte = get_pte_from_cache(mm); | |
391 | if (pte) | |
392 | return pte; | |
393 | ||
394 | return __alloc_for_ptecache(mm, kernel); | |
395 | } | |
396 | ||
70234676 AK |
397 | void pte_fragment_free(unsigned long *table, int kernel) |
398 | { | |
399 | struct page *page = virt_to_page(table); | |
400 | ||
4231aba0 NP |
401 | BUG_ON(atomic_read(&page->pt_frag_refcount) <= 0); |
402 | if (atomic_dec_and_test(&page->pt_frag_refcount)) { | |
70234676 AK |
403 | if (!kernel) |
404 | pgtable_page_dtor(page); | |
4231aba0 | 405 | __free_page(page); |
70234676 AK |
406 | } |
407 | } | |
408 | ||
0c4d2680 AK |
409 | static inline void pgtable_free(void *table, int index) |
410 | { | |
411 | switch (index) { | |
412 | case PTE_INDEX: | |
413 | pte_fragment_free(table, 0); | |
414 | break; | |
415 | case PMD_INDEX: | |
738f9645 | 416 | pmd_fragment_free(table); |
0c4d2680 AK |
417 | break; |
418 | case PUD_INDEX: | |
419 | kmem_cache_free(PGT_CACHE(PUD_CACHE_INDEX), table); | |
420 | break; | |
fadd03c6 AK |
421 | #if defined(CONFIG_PPC_4K_PAGES) && defined(CONFIG_HUGETLB_PAGE) |
422 | /* 16M hugepd directory at pud level */ | |
423 | case HTLB_16M_INDEX: | |
424 | BUILD_BUG_ON(H_16M_CACHE_INDEX <= 0); | |
425 | kmem_cache_free(PGT_CACHE(H_16M_CACHE_INDEX), table); | |
426 | break; | |
427 | /* 16G hugepd directory at the pgd level */ | |
428 | case HTLB_16G_INDEX: | |
429 | BUILD_BUG_ON(H_16G_CACHE_INDEX <= 0); | |
430 | kmem_cache_free(PGT_CACHE(H_16G_CACHE_INDEX), table); | |
431 | break; | |
432 | #endif | |
0c4d2680 AK |
433 | /* We don't free pgd table via RCU callback */ |
434 | default: | |
435 | BUG(); | |
436 | } | |
437 | } | |
438 | ||
70234676 | 439 | #ifdef CONFIG_SMP |
0c4d2680 | 440 | void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index) |
70234676 AK |
441 | { |
442 | unsigned long pgf = (unsigned long)table; | |
443 | ||
0c4d2680 AK |
444 | BUG_ON(index > MAX_PGTABLE_INDEX_SIZE); |
445 | pgf |= index; | |
70234676 AK |
446 | tlb_remove_table(tlb, (void *)pgf); |
447 | } | |
448 | ||
449 | void __tlb_remove_table(void *_table) | |
450 | { | |
451 | void *table = (void *)((unsigned long)_table & ~MAX_PGTABLE_INDEX_SIZE); | |
0c4d2680 | 452 | unsigned int index = (unsigned long)_table & MAX_PGTABLE_INDEX_SIZE; |
70234676 | 453 | |
0c4d2680 | 454 | return pgtable_free(table, index); |
70234676 AK |
455 | } |
456 | #else | |
0c4d2680 | 457 | void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int index) |
70234676 | 458 | { |
0c4d2680 | 459 | return pgtable_free(table, index); |
70234676 AK |
460 | } |
461 | #endif | |
a2dc009a AK |
462 | |
463 | #ifdef CONFIG_PROC_FS | |
464 | atomic_long_t direct_pages_count[MMU_PAGE_COUNT]; | |
465 | ||
466 | void arch_report_meminfo(struct seq_file *m) | |
467 | { | |
468 | /* | |
469 | * Hash maps the memory with one size mmu_linear_psize. | |
470 | * So don't bother to print these on hash | |
471 | */ | |
472 | if (!radix_enabled()) | |
473 | return; | |
474 | seq_printf(m, "DirectMap4k: %8lu kB\n", | |
475 | atomic_long_read(&direct_pages_count[MMU_PAGE_4K]) << 2); | |
476 | seq_printf(m, "DirectMap64k: %8lu kB\n", | |
477 | atomic_long_read(&direct_pages_count[MMU_PAGE_64K]) << 6); | |
478 | seq_printf(m, "DirectMap2M: %8lu kB\n", | |
479 | atomic_long_read(&direct_pages_count[MMU_PAGE_2M]) << 11); | |
480 | seq_printf(m, "DirectMap1G: %8lu kB\n", | |
481 | atomic_long_read(&direct_pages_count[MMU_PAGE_1G]) << 20); | |
482 | } | |
483 | #endif /* CONFIG_PROC_FS */ |