1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * MMU context allocation for 64-bit kernels.
4 *
5 * Copyright (C) 2004 Anton Blanchard, IBM Corp. <anton@samba.org>
6 */
7
8 #include <linux/sched.h>
9 #include <linux/kernel.h>
10 #include <linux/errno.h>
11 #include <linux/string.h>
12 #include <linux/types.h>
13 #include <linux/mm.h>
14 #include <linux/pkeys.h>
15 #include <linux/spinlock.h>
16 #include <linux/idr.h>
17 #include <linux/export.h>
18 #include <linux/gfp.h>
19 #include <linux/slab.h>
20
21 #include <asm/mmu_context.h>
22 #include <asm/pgalloc.h>
23
24 static DEFINE_IDA(mmu_context_ida);
25
alloc_context_id(int min_id,int max_id)26 static int alloc_context_id(int min_id, int max_id)
27 {
28 return ida_alloc_range(&mmu_context_ida, min_id, max_id, GFP_KERNEL);
29 }
30
hash__reserve_context_id(int id)31 void hash__reserve_context_id(int id)
32 {
33 int result = ida_alloc_range(&mmu_context_ida, id, id, GFP_KERNEL);
34
35 WARN(result != id, "mmu: Failed to reserve context id %d (rc %d)\n", id, result);
36 }
37
hash__alloc_context_id(void)38 int hash__alloc_context_id(void)
39 {
40 unsigned long max;
41
42 if (mmu_has_feature(MMU_FTR_68_BIT_VA))
43 max = MAX_USER_CONTEXT;
44 else
45 max = MAX_USER_CONTEXT_65BIT_VA;
46
47 return alloc_context_id(MIN_USER_CONTEXT, max);
48 }
49 EXPORT_SYMBOL_GPL(hash__alloc_context_id);
50
51 void slb_setup_new_exec(void);
52
realloc_context_ids(mm_context_t * ctx)53 static int realloc_context_ids(mm_context_t *ctx)
54 {
55 int i, id;
56
57 /*
58 * id 0 (aka. ctx->id) is special, we always allocate a new one, even if
59 * there wasn't one allocated previously (which happens in the exec
60 * case where ctx is newly allocated).
61 *
62 * We have to be a bit careful here. We must keep the existing ids in
63 * the array, so that we can test if they're non-zero to decide if we
64 * need to allocate a new one. However in case of error we must free the
65 * ids we've allocated but *not* any of the existing ones (or risk a
66 * UAF). That's why we decrement i at the start of the error handling
67 * loop, to skip the id that we just tested but couldn't reallocate.
68 */
69 for (i = 0; i < ARRAY_SIZE(ctx->extended_id); i++) {
70 if (i == 0 || ctx->extended_id[i]) {
71 id = hash__alloc_context_id();
72 if (id < 0)
73 goto error;
74
75 ctx->extended_id[i] = id;
76 }
77 }
78
79 /* The caller expects us to return id */
80 return ctx->id;
81
82 error:
83 for (i--; i >= 0; i--) {
84 if (ctx->extended_id[i])
85 ida_free(&mmu_context_ida, ctx->extended_id[i]);
86 }
87
88 return id;
89 }
90
hash__init_new_context(struct mm_struct * mm)91 static int hash__init_new_context(struct mm_struct *mm)
92 {
93 int index;
94
95 mm->context.hash_context = kmalloc(sizeof(struct hash_mm_context),
96 GFP_KERNEL);
97 if (!mm->context.hash_context)
98 return -ENOMEM;
99
100 /*
101 * The old code would re-promote on fork, we don't do that when using
102 * slices as it could cause problem promoting slices that have been
103 * forced down to 4K.
104 *
105 * For book3s we have MMU_NO_CONTEXT set to be ~0. Hence check
106 * explicitly against context.id == 0. This ensures that we properly
107 * initialize context slice details for newly allocated mm's (which will
108 * have id == 0) and don't alter context slice inherited via fork (which
109 * will have id != 0).
110 *
111 * We should not be calling init_new_context() on init_mm. Hence a
112 * check against 0 is OK.
113 */
114 if (mm->context.id == 0) {
115 memset(mm->context.hash_context, 0, sizeof(struct hash_mm_context));
116 slice_init_new_context_exec(mm);
117 } else {
118 /* This is fork. Copy hash_context details from current->mm */
119 memcpy(mm->context.hash_context, current->mm->context.hash_context, sizeof(struct hash_mm_context));
120 #ifdef CONFIG_PPC_SUBPAGE_PROT
121 /* inherit subpage prot detalis if we have one. */
122 if (current->mm->context.hash_context->spt) {
123 mm->context.hash_context->spt = kmalloc(sizeof(struct subpage_prot_table),
124 GFP_KERNEL);
125 if (!mm->context.hash_context->spt) {
126 kfree(mm->context.hash_context);
127 return -ENOMEM;
128 }
129 }
130 #endif
131 }
132
133 index = realloc_context_ids(&mm->context);
134 if (index < 0) {
135 #ifdef CONFIG_PPC_SUBPAGE_PROT
136 kfree(mm->context.hash_context->spt);
137 #endif
138 kfree(mm->context.hash_context);
139 return index;
140 }
141
142 pkey_mm_init(mm);
143 return index;
144 }
145
hash__setup_new_exec(void)146 void hash__setup_new_exec(void)
147 {
148 slice_setup_new_exec();
149
150 slb_setup_new_exec();
151 }
152
radix__init_new_context(struct mm_struct * mm)153 static int radix__init_new_context(struct mm_struct *mm)
154 {
155 unsigned long rts_field;
156 int index, max_id;
157
158 max_id = (1 << mmu_pid_bits) - 1;
159 index = alloc_context_id(mmu_base_pid, max_id);
160 if (index < 0)
161 return index;
162
163 /*
164 * set the process table entry,
165 */
166 rts_field = radix__get_tree_size();
167 process_tb[index].prtb0 = cpu_to_be64(rts_field | __pa(mm->pgd) | RADIX_PGD_INDEX_SIZE);
168
169 /*
170 * Order the above store with subsequent update of the PID
171 * register (at which point HW can start loading/caching
172 * the entry) and the corresponding load by the MMU from
173 * the L2 cache.
174 */
175 asm volatile("ptesync;isync" : : : "memory");
176
177 mm->context.hash_context = NULL;
178
179 return index;
180 }
181
init_new_context(struct task_struct * tsk,struct mm_struct * mm)182 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
183 {
184 int index;
185
186 if (radix_enabled())
187 index = radix__init_new_context(mm);
188 else
189 index = hash__init_new_context(mm);
190
191 if (index < 0)
192 return index;
193
194 mm->context.id = index;
195
196 mm->context.pte_frag = NULL;
197 mm->context.pmd_frag = NULL;
198 #ifdef CONFIG_SPAPR_TCE_IOMMU
199 mm_iommu_init(mm);
200 #endif
201 atomic_set(&mm->context.active_cpus, 0);
202 atomic_set(&mm->context.copros, 0);
203
204 return 0;
205 }
206
__destroy_context(int context_id)207 void __destroy_context(int context_id)
208 {
209 ida_free(&mmu_context_ida, context_id);
210 }
211 EXPORT_SYMBOL_GPL(__destroy_context);
212
destroy_contexts(mm_context_t * ctx)213 static void destroy_contexts(mm_context_t *ctx)
214 {
215 int index, context_id;
216
217 for (index = 0; index < ARRAY_SIZE(ctx->extended_id); index++) {
218 context_id = ctx->extended_id[index];
219 if (context_id)
220 ida_free(&mmu_context_ida, context_id);
221 }
222 kfree(ctx->hash_context);
223 }
224
pmd_frag_destroy(void * pmd_frag)225 static void pmd_frag_destroy(void *pmd_frag)
226 {
227 int count;
228 struct page *page;
229
230 page = virt_to_page(pmd_frag);
231 /* drop all the pending references */
232 count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT;
233 /* We allow PTE_FRAG_NR fragments from a PTE page */
234 if (atomic_sub_and_test(PMD_FRAG_NR - count, &page->pt_frag_refcount)) {
235 pgtable_pmd_page_dtor(page);
236 __free_page(page);
237 }
238 }
239
destroy_pagetable_cache(struct mm_struct * mm)240 static void destroy_pagetable_cache(struct mm_struct *mm)
241 {
242 void *frag;
243
244 frag = mm->context.pte_frag;
245 if (frag)
246 pte_frag_destroy(frag);
247
248 frag = mm->context.pmd_frag;
249 if (frag)
250 pmd_frag_destroy(frag);
251 return;
252 }
253
destroy_context(struct mm_struct * mm)254 void destroy_context(struct mm_struct *mm)
255 {
256 #ifdef CONFIG_SPAPR_TCE_IOMMU
257 WARN_ON_ONCE(!list_empty(&mm->context.iommu_group_mem_list));
258 #endif
259 /*
260 * For tasks which were successfully initialized we end up calling
261 * arch_exit_mmap() which clears the process table entry. And
262 * arch_exit_mmap() is called before the required fullmm TLB flush
263 * which does a RIC=2 flush. Hence for an initialized task, we do clear
264 * any cached process table entries.
265 *
266 * The condition below handles the error case during task init. We have
267 * set the process table entry early and if we fail a task
268 * initialization, we need to ensure the process table entry is zeroed.
269 * We need not worry about process table entry caches because the task
270 * never ran with the PID value.
271 */
272 if (radix_enabled())
273 process_tb[mm->context.id].prtb0 = 0;
274 else
275 subpage_prot_free(mm);
276 destroy_contexts(&mm->context);
277 mm->context.id = MMU_NO_CONTEXT;
278 }
279
arch_exit_mmap(struct mm_struct * mm)280 void arch_exit_mmap(struct mm_struct *mm)
281 {
282 destroy_pagetable_cache(mm);
283
284 if (radix_enabled()) {
285 /*
286 * Radix doesn't have a valid bit in the process table
287 * entries. However we know that at least P9 implementation
288 * will avoid caching an entry with an invalid RTS field,
289 * and 0 is invalid. So this will do.
290 *
291 * This runs before the "fullmm" tlb flush in exit_mmap,
292 * which does a RIC=2 tlbie to clear the process table
293 * entry. See the "fullmm" comments in tlb-radix.c.
294 *
295 * No barrier required here after the store because
296 * this process will do the invalidate, which starts with
297 * ptesync.
298 */
299 process_tb[mm->context.id].prtb0 = 0;
300 }
301 }
302
303 #ifdef CONFIG_PPC_RADIX_MMU
radix__switch_mmu_context(struct mm_struct * prev,struct mm_struct * next)304 void radix__switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
305 {
306 mtspr(SPRN_PID, next->context.id);
307 isync();
308 }
309 #endif
310