| 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 | |
| 29 | static int alloc_context_id(int min_id, int max_id) |
| 30 | { |
| 31 | return ida_alloc_range(&mmu_context_ida, min: min_id, max: max_id, GFP_KERNEL); |
| 32 | } |
| 33 | |
| 34 | #ifdef CONFIG_PPC_64S_HASH_MMU |
| 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 | |
| 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 |
| 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 | |
| 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 | |
| 150 | void hash__setup_new_exec(void) |
| 151 | { |
| 152 | slice_setup_new_exec(); |
| 153 | } |
| 154 | #else |
| 155 | static inline int hash__init_new_context(struct mm_struct *mm) |
| 156 | { |
| 157 | BUILD_BUG(); |
| 158 | return 0; |
| 159 | } |
| 160 | #endif |
| 161 | |
| 162 | static int radix__init_new_context(struct mm_struct *mm) |
| 163 | { |
| 164 | unsigned long rts_field; |
| 165 | int index, max_id; |
| 166 | |
| 167 | max_id = (1 << mmu_pid_bits) - 1; |
| 168 | index = alloc_context_id(min_id: mmu_base_pid, max_id); |
| 169 | if (index < 0) |
| 170 | return index; |
| 171 | |
| 172 | /* |
| 173 | * set the process table entry, |
| 174 | */ |
| 175 | rts_field = radix__get_tree_size(); |
| 176 | process_tb[index].prtb0 = cpu_to_be64(rts_field | __pa(mm->pgd) | RADIX_PGD_INDEX_SIZE); |
| 177 | |
| 178 | /* |
| 179 | * Order the above store with subsequent update of the PID |
| 180 | * register (at which point HW can start loading/caching |
| 181 | * the entry) and the corresponding load by the MMU from |
| 182 | * the L2 cache. |
| 183 | */ |
| 184 | asm volatile("ptesync;isync" : : : "memory" ); |
| 185 | |
| 186 | #ifdef CONFIG_PPC_64S_HASH_MMU |
| 187 | mm->context.hash_context = NULL; |
| 188 | #endif |
| 189 | |
| 190 | return index; |
| 191 | } |
| 192 | |
| 193 | int init_new_context(struct task_struct *tsk, struct mm_struct *mm) |
| 194 | { |
| 195 | int index; |
| 196 | |
| 197 | if (radix_enabled()) |
| 198 | index = radix__init_new_context(mm); |
| 199 | else |
| 200 | index = hash__init_new_context(mm); |
| 201 | |
| 202 | if (index < 0) |
| 203 | return index; |
| 204 | |
| 205 | mm->context.id = index; |
| 206 | |
| 207 | mm->context.pte_frag = NULL; |
| 208 | mm->context.pmd_frag = NULL; |
| 209 | #ifdef CONFIG_SPAPR_TCE_IOMMU |
| 210 | mm_iommu_init(mm); |
| 211 | #endif |
| 212 | atomic_set(v: &mm->context.active_cpus, i: 0); |
| 213 | atomic_set(v: &mm->context.copros, i: 0); |
| 214 | |
| 215 | return 0; |
| 216 | } |
| 217 | |
| 218 | void __destroy_context(int context_id) |
| 219 | { |
| 220 | ida_free(&mmu_context_ida, id: context_id); |
| 221 | } |
| 222 | EXPORT_SYMBOL_GPL(__destroy_context); |
| 223 | |
| 224 | static void destroy_contexts(mm_context_t *ctx) |
| 225 | { |
| 226 | if (radix_enabled()) { |
| 227 | ida_free(&mmu_context_ida, id: ctx->id); |
| 228 | } else { |
| 229 | #ifdef CONFIG_PPC_64S_HASH_MMU |
| 230 | int index, context_id; |
| 231 | |
| 232 | for (index = 0; index < ARRAY_SIZE(ctx->extended_id); index++) { |
| 233 | context_id = ctx->extended_id[index]; |
| 234 | if (context_id) |
| 235 | ida_free(&mmu_context_ida, context_id); |
| 236 | } |
| 237 | kfree(ctx->hash_context); |
| 238 | #else |
| 239 | BUILD_BUG(); // radix_enabled() should be constant true |
| 240 | #endif |
| 241 | } |
| 242 | } |
| 243 | |
| 244 | static void pmd_frag_destroy(void *pmd_frag) |
| 245 | { |
| 246 | int count; |
| 247 | struct ptdesc *ptdesc; |
| 248 | |
| 249 | ptdesc = virt_to_ptdesc(x: pmd_frag); |
| 250 | /* drop all the pending references */ |
| 251 | count = ((unsigned long)pmd_frag & ~PAGE_MASK) >> PMD_FRAG_SIZE_SHIFT; |
| 252 | /* We allow PTE_FRAG_NR fragments from a PTE page */ |
| 253 | if (atomic_sub_and_test(i: PMD_FRAG_NR - count, v: &ptdesc->pt_frag_refcount)) { |
| 254 | pagetable_dtor(ptdesc); |
| 255 | pagetable_free(pt: ptdesc); |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | static void destroy_pagetable_cache(struct mm_struct *mm) |
| 260 | { |
| 261 | void *frag; |
| 262 | |
| 263 | frag = mm->context.pte_frag; |
| 264 | if (frag) |
| 265 | pte_frag_destroy(frag); |
| 266 | |
| 267 | frag = mm->context.pmd_frag; |
| 268 | if (frag) |
| 269 | pmd_frag_destroy(pmd_frag: frag); |
| 270 | return; |
| 271 | } |
| 272 | |
| 273 | void destroy_context(struct mm_struct *mm) |
| 274 | { |
| 275 | #ifdef CONFIG_SPAPR_TCE_IOMMU |
| 276 | WARN_ON_ONCE(!list_empty(&mm->context.iommu_group_mem_list)); |
| 277 | #endif |
| 278 | /* |
| 279 | * For tasks which were successfully initialized we end up calling |
| 280 | * arch_exit_mmap() which clears the process table entry. And |
| 281 | * arch_exit_mmap() is called before the required fullmm TLB flush |
| 282 | * which does a RIC=2 flush. Hence for an initialized task, we do clear |
| 283 | * any cached process table entries. |
| 284 | * |
| 285 | * The condition below handles the error case during task init. We have |
| 286 | * set the process table entry early and if we fail a task |
| 287 | * initialization, we need to ensure the process table entry is zeroed. |
| 288 | * We need not worry about process table entry caches because the task |
| 289 | * never ran with the PID value. |
| 290 | */ |
| 291 | if (radix_enabled()) |
| 292 | process_tb[mm->context.id].prtb0 = 0; |
| 293 | else |
| 294 | subpage_prot_free(mm); |
| 295 | destroy_contexts(ctx: &mm->context); |
| 296 | mm->context.id = MMU_NO_CONTEXT; |
| 297 | } |
| 298 | |
| 299 | void arch_exit_mmap(struct mm_struct *mm) |
| 300 | { |
| 301 | destroy_pagetable_cache(mm); |
| 302 | |
| 303 | if (radix_enabled()) { |
| 304 | /* |
| 305 | * Radix doesn't have a valid bit in the process table |
| 306 | * entries. However we know that at least P9 implementation |
| 307 | * will avoid caching an entry with an invalid RTS field, |
| 308 | * and 0 is invalid. So this will do. |
| 309 | * |
| 310 | * This runs before the "fullmm" tlb flush in exit_mmap, |
| 311 | * which does a RIC=2 tlbie to clear the process table |
| 312 | * entry. See the "fullmm" comments in tlb-radix.c. |
| 313 | * |
| 314 | * No barrier required here after the store because |
| 315 | * this process will do the invalidate, which starts with |
| 316 | * ptesync. |
| 317 | */ |
| 318 | process_tb[mm->context.id].prtb0 = 0; |
| 319 | } |
| 320 | } |
| 321 | |
| 322 | #ifdef CONFIG_PPC_RADIX_MMU |
| 323 | void radix__switch_mmu_context(struct mm_struct *prev, struct mm_struct *next) |
| 324 | { |
| 325 | mtspr(SPRN_PID, next->context.id); |
| 326 | isync(); |
| 327 | } |
| 328 | #endif |
| 329 | |
| 330 | /** |
| 331 | * cleanup_cpu_mmu_context - Clean up MMU details for this CPU (newly offlined) |
| 332 | * |
| 333 | * This clears the CPU from mm_cpumask for all processes, and then flushes the |
| 334 | * local TLB to ensure TLB coherency in case the CPU is onlined again. |
| 335 | * |
| 336 | * KVM guest translations are not necessarily flushed here. If KVM started |
| 337 | * using mm_cpumask or the Linux APIs which do, this would have to be resolved. |
| 338 | */ |
| 339 | #ifdef CONFIG_HOTPLUG_CPU |
| 340 | void cleanup_cpu_mmu_context(void) |
| 341 | { |
| 342 | int cpu = smp_processor_id(); |
| 343 | |
| 344 | clear_tasks_mm_cpumask(cpu); |
| 345 | tlbiel_all(); |
| 346 | } |
| 347 | #endif |
| 348 | |