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 #include <linux/cpu.h>
21
22 #include <asm/mmu_context.h>
23 #include <asm/pgalloc.h>
24
25 #include "internal.h"
26
27 static DEFINE_IDA(mmu_context_ida);
28
alloc_context_id(int min_id,int max_id)29 static int alloc_context_id(int min_id, int max_id)
30 {
31 return ida_alloc_range(&mmu_context_ida, min_id, max_id, GFP_KERNEL);
32 }
33
34 #ifdef CONFIG_PPC_64S_HASH_MMU
hash__reserve_context_id(int id)35 void __init hash__reserve_context_id(int id)
36 {
37 int result = ida_alloc_range(&mmu_context_ida, id, id, GFP_KERNEL);
38
39 WARN(result != id, "mmu: Failed to reserve context id %d (rc %d)\n", id, result);
40 }
41
hash__alloc_context_id(void)42 int hash__alloc_context_id(void)
43 {
44 unsigned long max;
45
46 if (mmu_has_feature(MMU_FTR_68_BIT_VA))
47 max = MAX_USER_CONTEXT;
48 else
49 max = MAX_USER_CONTEXT_65BIT_VA;
50
51 return alloc_context_id(MIN_USER_CONTEXT, max);
52 }
53 EXPORT_SYMBOL_GPL(hash__alloc_context_id);
54 #endif
55
56 #ifdef CONFIG_PPC_64S_HASH_MMU
realloc_context_ids(mm_context_t * ctx)57 static int realloc_context_ids(mm_context_t *ctx)
58 {
59 int i, id;
60
61 /*
62 * id 0 (aka. ctx->id) is special, we always allocate a new one, even if
63 * there wasn't one allocated previously (which happens in the exec
64 * case where ctx is newly allocated).
65 *
66 * We have to be a bit careful here. We must keep the existing ids in
67 * the array, so that we can test if they're non-zero to decide if we
68 * need to allocate a new one. However in case of error we must free the
69 * ids we've allocated but *not* any of the existing ones (or risk a
70 * UAF). That's why we decrement i at the start of the error handling
71 * loop, to skip the id that we just tested but couldn't reallocate.
72 */
73 for (i = 0; i < ARRAY_SIZE(ctx->extended_id); i++) {
74 if (i == 0 || ctx->extended_id[i]) {
75 id = hash__alloc_context_id();
76 if (id < 0)
77 goto error;
78
79 ctx->extended_id[i] = id;
80 }
81 }
82
83 /* The caller expects us to return id */
84 return ctx->id;
85
86 error:
87 for (i--; i >= 0; i--) {
88 if (ctx->extended_id[i])
89 ida_free(&mmu_context_ida, ctx->extended_id[i]);
90 }
91
92 return id;
93 }
94
hash__init_new_context(struct mm_struct * mm)95 static int hash__init_new_context(struct mm_struct *mm)
96 {
97 int index;
98
99 mm->context.hash_context = kmalloc(sizeof(struct hash_mm_context),
100 GFP_KERNEL);
101 if (!mm->context.hash_context)
102 return -ENOMEM;
103
104 /*
105 * The old code would re-promote on fork, we don't do that when using
106 * slices as it could cause problem promoting slices that have been
107 * forced down to 4K.
108 *
109 * For book3s we have MMU_NO_CONTEXT set to be ~0. Hence check
110 * explicitly against context.id == 0. This ensures that we properly
111 * initialize context slice details for newly allocated mm's (which will
112 * have id == 0) and don't alter context slice inherited via fork (which
113 * will have id != 0).
114 *
115 * We should not be calling init_new_context() on init_mm. Hence a
116 * check against 0 is OK.
117 */
118 if (mm->context.id == 0) {
119 memset(mm->context.hash_context, 0, sizeof(struct hash_mm_context));
120 slice_init_new_context_exec(mm);
121 } else {
122 /* This is fork. Copy hash_context details from current->mm */
123 memcpy(mm->context.hash_context, current->mm->context.hash_context, sizeof(struct hash_mm_context));
124 #ifdef CONFIG_PPC_SUBPAGE_PROT
125 /* inherit subpage prot details if we have one. */
126 if (current->mm->context.hash_context->spt) {
127 mm->context.hash_context->spt = kmalloc(sizeof(struct subpage_prot_table),
128 GFP_KERNEL);
129 if (!mm->context.hash_context->spt) {
130 kfree(mm->context.hash_context);
131 return -ENOMEM;
132 }
133 }
134 #endif
135 }
136
137 index = realloc_context_ids(&mm->context);
138 if (index < 0) {
139 #ifdef CONFIG_PPC_SUBPAGE_PROT
140 kfree(mm->context.hash_context->spt);
141 #endif
142 kfree(mm->context.hash_context);
143 return index;
144 }
145
146 pkey_mm_init(mm);
147 return index;
148 }
149
hash__setup_new_exec(void)150 void hash__setup_new_exec(void)
151 {
152 slice_setup_new_exec();
153
154 slb_setup_new_exec();
155 }
156 #else
hash__init_new_context(struct mm_struct * mm)157 static inline int hash__init_new_context(struct mm_struct *mm)
158 {
159 BUILD_BUG();
160 return 0;
161 }
162 #endif
163
radix__init_new_context(struct mm_struct * mm)164 static int radix__init_new_context(struct mm_struct *mm)
165 {
166 unsigned long rts_field;
167 int index, max_id;
168
169 max_id = (1 << mmu_pid_bits) - 1;
170 index = alloc_context_id(mmu_base_pid, max_id);
171 if (index < 0)
172 return index;
173
174 /*
175 * set the process table entry,
176 */
177 rts_field = radix__get_tree_size();
178 process_tb[index].prtb0 = cpu_to_be64(rts_field | __pa(mm->pgd) | RADIX_PGD_INDEX_SIZE);
179
180 /*
181 * Order the above store with subsequent update of the PID
182 * register (at which point HW can start loading/caching
183 * the entry) and the corresponding load by the MMU from
184 * the L2 cache.
185 */
186 asm volatile("ptesync;isync" : : : "memory");
187
188 #ifdef CONFIG_PPC_64S_HASH_MMU
189 mm->context.hash_context = NULL;
190 #endif
191
192 return index;
193 }
194
init_new_context(struct task_struct * tsk,struct mm_struct * mm)195 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
196 {
197 int index;
198
199 if (radix_enabled())
200 index = radix__init_new_context(mm);
201 else
202 index = hash__init_new_context(mm);
203
204 if (index < 0)
205 return index;
206
207 mm->context.id = index;
208
209 mm->context.pte_frag = NULL;
210 mm->context.pmd_frag = NULL;
211 #ifdef CONFIG_SPAPR_TCE_IOMMU
212 mm_iommu_init(mm);
213 #endif
214 atomic_set(&mm->context.active_cpus, 0);
215 atomic_set(&mm->context.copros, 0);
216
217 return 0;
218 }
219
__destroy_context(int context_id)220 void __destroy_context(int context_id)
221 {
222 ida_free(&mmu_context_ida, context_id);
223 }
224 EXPORT_SYMBOL_GPL(__destroy_context);
225
destroy_contexts(mm_context_t * ctx)226 static void destroy_contexts(mm_context_t *ctx)
227 {
228 if (radix_enabled()) {
229 ida_free(&mmu_context_ida, ctx->id);
230 } else {
231 #ifdef CONFIG_PPC_64S_HASH_MMU
232 int index, context_id;
233
234 for (index = 0; index < ARRAY_SIZE(ctx->extended_id); index++) {
235 context_id = ctx->extended_id[index];
236 if (context_id)
237 ida_free(&mmu_context_ida, context_id);
238 }
239 kfree(ctx->hash_context);
240 #else
241 BUILD_BUG(); // radix_enabled() should be constant true
242 #endif
243 }
244 }
245
pmd_frag_destroy(void * pmd_frag)246 static void pmd_frag_destroy(void *pmd_frag)
247 {
248 int count;
249 struct ptdesc *ptdesc;
250
251 ptdesc = virt_to_ptdesc(pmd_frag);
252 /* drop all the pending references */
253 count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT;
254 /* We allow PTE_FRAG_NR fragments from a PTE page */
255 if (atomic_sub_and_test(PMD_FRAG_NR - count, &ptdesc->pt_frag_refcount)) {
256 pagetable_pmd_dtor(ptdesc);
257 pagetable_free(ptdesc);
258 }
259 }
260
destroy_pagetable_cache(struct mm_struct * mm)261 static void destroy_pagetable_cache(struct mm_struct *mm)
262 {
263 void *frag;
264
265 frag = mm->context.pte_frag;
266 if (frag)
267 pte_frag_destroy(frag);
268
269 frag = mm->context.pmd_frag;
270 if (frag)
271 pmd_frag_destroy(frag);
272 return;
273 }
274
destroy_context(struct mm_struct * mm)275 void destroy_context(struct mm_struct *mm)
276 {
277 #ifdef CONFIG_SPAPR_TCE_IOMMU
278 WARN_ON_ONCE(!list_empty(&mm->context.iommu_group_mem_list));
279 #endif
280 /*
281 * For tasks which were successfully initialized we end up calling
282 * arch_exit_mmap() which clears the process table entry. And
283 * arch_exit_mmap() is called before the required fullmm TLB flush
284 * which does a RIC=2 flush. Hence for an initialized task, we do clear
285 * any cached process table entries.
286 *
287 * The condition below handles the error case during task init. We have
288 * set the process table entry early and if we fail a task
289 * initialization, we need to ensure the process table entry is zeroed.
290 * We need not worry about process table entry caches because the task
291 * never ran with the PID value.
292 */
293 if (radix_enabled())
294 process_tb[mm->context.id].prtb0 = 0;
295 else
296 subpage_prot_free(mm);
297 destroy_contexts(&mm->context);
298 mm->context.id = MMU_NO_CONTEXT;
299 }
300
arch_exit_mmap(struct mm_struct * mm)301 void arch_exit_mmap(struct mm_struct *mm)
302 {
303 destroy_pagetable_cache(mm);
304
305 if (radix_enabled()) {
306 /*
307 * Radix doesn't have a valid bit in the process table
308 * entries. However we know that at least P9 implementation
309 * will avoid caching an entry with an invalid RTS field,
310 * and 0 is invalid. So this will do.
311 *
312 * This runs before the "fullmm" tlb flush in exit_mmap,
313 * which does a RIC=2 tlbie to clear the process table
314 * entry. See the "fullmm" comments in tlb-radix.c.
315 *
316 * No barrier required here after the store because
317 * this process will do the invalidate, which starts with
318 * ptesync.
319 */
320 process_tb[mm->context.id].prtb0 = 0;
321 }
322 }
323
324 #ifdef CONFIG_PPC_RADIX_MMU
radix__switch_mmu_context(struct mm_struct * prev,struct mm_struct * next)325 void radix__switch_mmu_context(struct mm_struct *prev, struct mm_struct *next)
326 {
327 mtspr(SPRN_PID, next->context.id);
328 isync();
329 }
330 #endif
331
332 /**
333 * cleanup_cpu_mmu_context - Clean up MMU details for this CPU (newly offlined)
334 *
335 * This clears the CPU from mm_cpumask for all processes, and then flushes the
336 * local TLB to ensure TLB coherency in case the CPU is onlined again.
337 *
338 * KVM guest translations are not necessarily flushed here. If KVM started
339 * using mm_cpumask or the Linux APIs which do, this would have to be resolved.
340 */
341 #ifdef CONFIG_HOTPLUG_CPU
cleanup_cpu_mmu_context(void)342 void cleanup_cpu_mmu_context(void)
343 {
344 int cpu = smp_processor_id();
345
346 clear_tasks_mm_cpumask(cpu);
347 tlbiel_all();
348 }
349 #endif
350