Commit | Line | Data |
---|---|---|
eee24b5a AK |
1 | /* |
2 | * Copyright 2005, Paul Mackerras, IBM Corporation. | |
3 | * Copyright 2009, Benjamin Herrenschmidt, IBM Corporation. | |
4 | * Copyright 2015-2016, Aneesh Kumar K.V, IBM Corporation. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public License | |
8 | * as published by the Free Software Foundation; either version | |
9 | * 2 of the License, or (at your option) any later version. | |
10 | */ | |
11 | ||
12 | #include <linux/sched.h> | |
589ee628 IM |
13 | #include <linux/mm_types.h> |
14 | ||
eee24b5a AK |
15 | #include <asm/pgalloc.h> |
16 | #include <asm/tlb.h> | |
17 | ||
18 | #include "mmu_decl.h" | |
19 | ||
6a1ea362 AK |
20 | #define CREATE_TRACE_POINTS |
21 | #include <trace/events/thp.h> | |
22 | ||
eee24b5a AK |
23 | #ifdef CONFIG_SPARSEMEM_VMEMMAP |
24 | /* | |
25 | * On hash-based CPUs, the vmemmap is bolted in the hash table. | |
26 | * | |
27 | */ | |
31a14fae AK |
28 | int __meminit hash__vmemmap_create_mapping(unsigned long start, |
29 | unsigned long page_size, | |
30 | unsigned long phys) | |
eee24b5a AK |
31 | { |
32 | int rc = htab_bolt_mapping(start, start + page_size, phys, | |
33 | pgprot_val(PAGE_KERNEL), | |
34 | mmu_vmemmap_psize, mmu_kernel_ssize); | |
35 | if (rc < 0) { | |
36 | int rc2 = htab_remove_mapping(start, start + page_size, | |
37 | mmu_vmemmap_psize, | |
38 | mmu_kernel_ssize); | |
39 | BUG_ON(rc2 && (rc2 != -ENOENT)); | |
40 | } | |
41 | return rc; | |
42 | } | |
43 | ||
44 | #ifdef CONFIG_MEMORY_HOTPLUG | |
31a14fae AK |
45 | void hash__vmemmap_remove_mapping(unsigned long start, |
46 | unsigned long page_size) | |
eee24b5a AK |
47 | { |
48 | int rc = htab_remove_mapping(start, start + page_size, | |
49 | mmu_vmemmap_psize, | |
50 | mmu_kernel_ssize); | |
51 | BUG_ON((rc < 0) && (rc != -ENOENT)); | |
52 | WARN_ON(rc == -ENOENT); | |
53 | } | |
54 | #endif | |
55 | #endif /* CONFIG_SPARSEMEM_VMEMMAP */ | |
56 | ||
57 | /* | |
58 | * map_kernel_page currently only called by __ioremap | |
59 | * map_kernel_page adds an entry to the ioremap page table | |
60 | * and adds an entry to the HPT, possibly bolting it | |
61 | */ | |
31a14fae | 62 | int hash__map_kernel_page(unsigned long ea, unsigned long pa, unsigned long flags) |
eee24b5a AK |
63 | { |
64 | pgd_t *pgdp; | |
65 | pud_t *pudp; | |
66 | pmd_t *pmdp; | |
67 | pte_t *ptep; | |
68 | ||
dd1842a2 | 69 | BUILD_BUG_ON(TASK_SIZE_USER64 > H_PGTABLE_RANGE); |
eee24b5a AK |
70 | if (slab_is_available()) { |
71 | pgdp = pgd_offset_k(ea); | |
72 | pudp = pud_alloc(&init_mm, pgdp, ea); | |
73 | if (!pudp) | |
74 | return -ENOMEM; | |
75 | pmdp = pmd_alloc(&init_mm, pudp, ea); | |
76 | if (!pmdp) | |
77 | return -ENOMEM; | |
78 | ptep = pte_alloc_kernel(pmdp, ea); | |
79 | if (!ptep) | |
80 | return -ENOMEM; | |
81 | set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT, | |
82 | __pgprot(flags))); | |
83 | } else { | |
84 | /* | |
85 | * If the mm subsystem is not fully up, we cannot create a | |
86 | * linux page table entry for this mapping. Simply bolt an | |
87 | * entry in the hardware page table. | |
88 | * | |
89 | */ | |
90 | if (htab_bolt_mapping(ea, ea + PAGE_SIZE, pa, flags, | |
91 | mmu_io_psize, mmu_kernel_ssize)) { | |
92 | printk(KERN_ERR "Failed to do bolted mapping IO " | |
93 | "memory at %016lx !\n", pa); | |
94 | return -ENOMEM; | |
95 | } | |
96 | } | |
97 | ||
98 | smp_wmb(); | |
99 | return 0; | |
100 | } | |
6a1ea362 AK |
101 | |
102 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
103 | ||
3df33f12 AK |
104 | unsigned long hash__pmd_hugepage_update(struct mm_struct *mm, unsigned long addr, |
105 | pmd_t *pmdp, unsigned long clr, | |
106 | unsigned long set) | |
6a1ea362 AK |
107 | { |
108 | __be64 old_be, tmp; | |
109 | unsigned long old; | |
110 | ||
111 | #ifdef CONFIG_DEBUG_VM | |
112 | WARN_ON(!pmd_trans_huge(*pmdp)); | |
113 | assert_spin_locked(&mm->page_table_lock); | |
114 | #endif | |
115 | ||
116 | __asm__ __volatile__( | |
117 | "1: ldarx %0,0,%3\n\ | |
118 | and. %1,%0,%6\n\ | |
119 | bne- 1b \n\ | |
120 | andc %1,%0,%4 \n\ | |
121 | or %1,%1,%7\n\ | |
122 | stdcx. %1,0,%3 \n\ | |
123 | bne- 1b" | |
124 | : "=&r" (old_be), "=&r" (tmp), "=m" (*pmdp) | |
125 | : "r" (pmdp), "r" (cpu_to_be64(clr)), "m" (*pmdp), | |
126 | "r" (cpu_to_be64(H_PAGE_BUSY)), "r" (cpu_to_be64(set)) | |
127 | : "cc" ); | |
128 | ||
129 | old = be64_to_cpu(old_be); | |
130 | ||
131 | trace_hugepage_update(addr, old, clr, set); | |
132 | if (old & H_PAGE_HASHPTE) | |
133 | hpte_do_hugepage_flush(mm, addr, pmdp, old); | |
134 | return old; | |
135 | } | |
136 | ||
3df33f12 AK |
137 | pmd_t hash__pmdp_collapse_flush(struct vm_area_struct *vma, unsigned long address, |
138 | pmd_t *pmdp) | |
6a1ea362 AK |
139 | { |
140 | pmd_t pmd; | |
141 | ||
142 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
143 | VM_BUG_ON(pmd_trans_huge(*pmdp)); | |
144 | ||
145 | pmd = *pmdp; | |
146 | pmd_clear(pmdp); | |
147 | /* | |
148 | * Wait for all pending hash_page to finish. This is needed | |
149 | * in case of subpage collapse. When we collapse normal pages | |
150 | * to hugepage, we first clear the pmd, then invalidate all | |
151 | * the PTE entries. The assumption here is that any low level | |
152 | * page fault will see a none pmd and take the slow path that | |
153 | * will wait on mmap_sem. But we could very well be in a | |
154 | * hash_page with local ptep pointer value. Such a hash page | |
155 | * can result in adding new HPTE entries for normal subpages. | |
156 | * That means we could be modifying the page content as we | |
157 | * copy them to a huge page. So wait for parallel hash_page | |
158 | * to finish before invalidating HPTE entries. We can do this | |
159 | * by sending an IPI to all the cpus and executing a dummy | |
160 | * function there. | |
161 | */ | |
162 | kick_all_cpus_sync(); | |
163 | /* | |
164 | * Now invalidate the hpte entries in the range | |
165 | * covered by pmd. This make sure we take a | |
166 | * fault and will find the pmd as none, which will | |
167 | * result in a major fault which takes mmap_sem and | |
168 | * hence wait for collapse to complete. Without this | |
169 | * the __collapse_huge_page_copy can result in copying | |
170 | * the old content. | |
171 | */ | |
172 | flush_tlb_pmd_range(vma->vm_mm, &pmd, address); | |
173 | return pmd; | |
174 | } | |
175 | ||
6a1ea362 AK |
176 | /* |
177 | * We want to put the pgtable in pmd and use pgtable for tracking | |
178 | * the base page size hptes | |
179 | */ | |
3df33f12 AK |
180 | void hash__pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp, |
181 | pgtable_t pgtable) | |
6a1ea362 AK |
182 | { |
183 | pgtable_t *pgtable_slot; | |
184 | assert_spin_locked(&mm->page_table_lock); | |
185 | /* | |
186 | * we store the pgtable in the second half of PMD | |
187 | */ | |
188 | pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD; | |
189 | *pgtable_slot = pgtable; | |
190 | /* | |
191 | * expose the deposited pgtable to other cpus. | |
192 | * before we set the hugepage PTE at pmd level | |
193 | * hash fault code looks at the deposted pgtable | |
194 | * to store hash index values. | |
195 | */ | |
196 | smp_wmb(); | |
197 | } | |
198 | ||
3df33f12 | 199 | pgtable_t hash__pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp) |
6a1ea362 AK |
200 | { |
201 | pgtable_t pgtable; | |
202 | pgtable_t *pgtable_slot; | |
203 | ||
204 | assert_spin_locked(&mm->page_table_lock); | |
205 | pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD; | |
206 | pgtable = *pgtable_slot; | |
207 | /* | |
208 | * Once we withdraw, mark the entry NULL. | |
209 | */ | |
210 | *pgtable_slot = NULL; | |
211 | /* | |
212 | * We store HPTE information in the deposited PTE fragment. | |
213 | * zero out the content on withdraw. | |
214 | */ | |
215 | memset(pgtable, 0, PTE_FRAG_SIZE); | |
216 | return pgtable; | |
217 | } | |
218 | ||
3df33f12 AK |
219 | void hash__pmdp_huge_split_prepare(struct vm_area_struct *vma, |
220 | unsigned long address, pmd_t *pmdp) | |
6a1ea362 AK |
221 | { |
222 | VM_BUG_ON(address & ~HPAGE_PMD_MASK); | |
223 | VM_BUG_ON(REGION_ID(address) != USER_REGION_ID); | |
224 | ||
225 | /* | |
226 | * We can't mark the pmd none here, because that will cause a race | |
227 | * against exit_mmap. We need to continue mark pmd TRANS HUGE, while | |
228 | * we spilt, but at the same time we wan't rest of the ppc64 code | |
229 | * not to insert hash pte on this, because we will be modifying | |
230 | * the deposited pgtable in the caller of this function. Hence | |
231 | * clear the _PAGE_USER so that we move the fault handling to | |
232 | * higher level function and that will serialize against ptl. | |
233 | * We need to flush existing hash pte entries here even though, | |
234 | * the translation is still valid, because we will withdraw | |
235 | * pgtable_t after this. | |
236 | */ | |
237 | pmd_hugepage_update(vma->vm_mm, address, pmdp, 0, _PAGE_PRIVILEGED); | |
238 | } | |
239 | ||
6a1ea362 AK |
240 | /* |
241 | * A linux hugepage PMD was changed and the corresponding hash table entries | |
242 | * neesd to be flushed. | |
243 | */ | |
244 | void hpte_do_hugepage_flush(struct mm_struct *mm, unsigned long addr, | |
245 | pmd_t *pmdp, unsigned long old_pmd) | |
246 | { | |
247 | int ssize; | |
248 | unsigned int psize; | |
249 | unsigned long vsid; | |
250 | unsigned long flags = 0; | |
251 | const struct cpumask *tmp; | |
252 | ||
253 | /* get the base page size,vsid and segment size */ | |
254 | #ifdef CONFIG_DEBUG_VM | |
255 | psize = get_slice_psize(mm, addr); | |
256 | BUG_ON(psize == MMU_PAGE_16M); | |
257 | #endif | |
258 | if (old_pmd & H_PAGE_COMBO) | |
259 | psize = MMU_PAGE_4K; | |
260 | else | |
261 | psize = MMU_PAGE_64K; | |
262 | ||
263 | if (!is_kernel_addr(addr)) { | |
264 | ssize = user_segment_size(addr); | |
265 | vsid = get_vsid(mm->context.id, addr, ssize); | |
266 | WARN_ON(vsid == 0); | |
267 | } else { | |
268 | vsid = get_kernel_vsid(addr, mmu_kernel_ssize); | |
269 | ssize = mmu_kernel_ssize; | |
270 | } | |
271 | ||
272 | tmp = cpumask_of(smp_processor_id()); | |
273 | if (cpumask_equal(mm_cpumask(mm), tmp)) | |
274 | flags |= HPTE_LOCAL_UPDATE; | |
275 | ||
276 | return flush_hash_hugepage(vsid, addr, pmdp, psize, ssize, flags); | |
277 | } | |
278 | ||
3df33f12 AK |
279 | pmd_t hash__pmdp_huge_get_and_clear(struct mm_struct *mm, |
280 | unsigned long addr, pmd_t *pmdp) | |
6a1ea362 AK |
281 | { |
282 | pmd_t old_pmd; | |
283 | pgtable_t pgtable; | |
284 | unsigned long old; | |
285 | pgtable_t *pgtable_slot; | |
286 | ||
287 | old = pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0); | |
288 | old_pmd = __pmd(old); | |
289 | /* | |
290 | * We have pmd == none and we are holding page_table_lock. | |
291 | * So we can safely go and clear the pgtable hash | |
292 | * index info. | |
293 | */ | |
294 | pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD; | |
295 | pgtable = *pgtable_slot; | |
296 | /* | |
297 | * Let's zero out old valid and hash index details | |
298 | * hash fault look at them. | |
299 | */ | |
300 | memset(pgtable, 0, PTE_FRAG_SIZE); | |
301 | /* | |
302 | * Serialize against find_linux_pte_or_hugepte which does lock-less | |
303 | * lookup in page tables with local interrupts disabled. For huge pages | |
304 | * it casts pmd_t to pte_t. Since format of pte_t is different from | |
305 | * pmd_t we want to prevent transit from pmd pointing to page table | |
306 | * to pmd pointing to huge page (and back) while interrupts are disabled. | |
307 | * We clear pmd to possibly replace it with page table pointer in | |
308 | * different code paths. So make sure we wait for the parallel | |
309 | * find_linux_pte_or_hugepage to finish. | |
310 | */ | |
311 | kick_all_cpus_sync(); | |
312 | return old_pmd; | |
313 | } | |
314 | ||
3df33f12 | 315 | int hash__has_transparent_hugepage(void) |
6a1ea362 AK |
316 | { |
317 | ||
318 | if (!mmu_has_feature(MMU_FTR_16M_PAGE)) | |
319 | return 0; | |
320 | /* | |
321 | * We support THP only if PMD_SIZE is 16MB. | |
322 | */ | |
323 | if (mmu_psize_defs[MMU_PAGE_16M].shift != PMD_SHIFT) | |
324 | return 0; | |
325 | /* | |
326 | * We need to make sure that we support 16MB hugepage in a segement | |
327 | * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE | |
328 | * of 64K. | |
329 | */ | |
330 | /* | |
331 | * If we have 64K HPTE, we will be using that by default | |
332 | */ | |
333 | if (mmu_psize_defs[MMU_PAGE_64K].shift && | |
334 | (mmu_psize_defs[MMU_PAGE_64K].penc[MMU_PAGE_16M] == -1)) | |
335 | return 0; | |
336 | /* | |
337 | * Ok we only have 4K HPTE | |
338 | */ | |
339 | if (mmu_psize_defs[MMU_PAGE_4K].penc[MMU_PAGE_16M] == -1) | |
340 | return 0; | |
341 | ||
342 | return 1; | |
343 | } | |
344 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ |