1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef _ASM_X86_MMU_CONTEXT_H 3 #define _ASM_X86_MMU_CONTEXT_H 4 5 #include <asm/desc.h> 6 #include <linux/atomic.h> 7 #include <linux/mm_types.h> 8 #include <linux/pkeys.h> 9 10 #include <trace/events/tlb.h> 11 12 #include <asm/pgalloc.h> 13 #include <asm/tlbflush.h> 14 #include <asm/paravirt.h> 15 #include <asm/mpx.h> 16 17 extern atomic64_t last_mm_ctx_id; 18 19 #ifndef CONFIG_PARAVIRT 20 static inline void paravirt_activate_mm(struct mm_struct *prev, 21 struct mm_struct *next) 22 { 23 } 24 #endif /* !CONFIG_PARAVIRT */ 25 26 #ifdef CONFIG_PERF_EVENTS 27 extern struct static_key rdpmc_always_available; 28 29 static inline void load_mm_cr4(struct mm_struct *mm) 30 { 31 if (static_key_false(&rdpmc_always_available) || 32 atomic_read(&mm->context.perf_rdpmc_allowed)) 33 cr4_set_bits(X86_CR4_PCE); 34 else 35 cr4_clear_bits(X86_CR4_PCE); 36 } 37 #else 38 static inline void load_mm_cr4(struct mm_struct *mm) {} 39 #endif 40 41 #ifdef CONFIG_MODIFY_LDT_SYSCALL 42 /* 43 * ldt_structs can be allocated, used, and freed, but they are never 44 * modified while live. 45 */ 46 struct ldt_struct { 47 /* 48 * Xen requires page-aligned LDTs with special permissions. This is 49 * needed to prevent us from installing evil descriptors such as 50 * call gates. On native, we could merge the ldt_struct and LDT 51 * allocations, but it's not worth trying to optimize. 52 */ 53 struct desc_struct *entries; 54 unsigned int nr_entries; 55 56 /* 57 * If PTI is in use, then the entries array is not mapped while we're 58 * in user mode. The whole array will be aliased at the addressed 59 * given by ldt_slot_va(slot). We use two slots so that we can allocate 60 * and map, and enable a new LDT without invalidating the mapping 61 * of an older, still-in-use LDT. 62 * 63 * slot will be -1 if this LDT doesn't have an alias mapping. 64 */ 65 int slot; 66 }; 67 68 /* This is a multiple of PAGE_SIZE. */ 69 #define LDT_SLOT_STRIDE (LDT_ENTRIES * LDT_ENTRY_SIZE) 70 71 static inline void *ldt_slot_va(int slot) 72 { 73 #ifdef CONFIG_X86_64 74 return (void *)(LDT_BASE_ADDR + LDT_SLOT_STRIDE * slot); 75 #else 76 BUG(); 77 #endif 78 } 79 80 /* 81 * Used for LDT copy/destruction. 82 */ 83 static inline void init_new_context_ldt(struct mm_struct *mm) 84 { 85 mm->context.ldt = NULL; 86 init_rwsem(&mm->context.ldt_usr_sem); 87 } 88 int ldt_dup_context(struct mm_struct *oldmm, struct mm_struct *mm); 89 void destroy_context_ldt(struct mm_struct *mm); 90 void ldt_arch_exit_mmap(struct mm_struct *mm); 91 #else /* CONFIG_MODIFY_LDT_SYSCALL */ 92 static inline void init_new_context_ldt(struct mm_struct *mm) { } 93 static inline int ldt_dup_context(struct mm_struct *oldmm, 94 struct mm_struct *mm) 95 { 96 return 0; 97 } 98 static inline void destroy_context_ldt(struct mm_struct *mm) { } 99 static inline void ldt_arch_exit_mmap(struct mm_struct *mm) { } 100 #endif 101 102 static inline void load_mm_ldt(struct mm_struct *mm) 103 { 104 #ifdef CONFIG_MODIFY_LDT_SYSCALL 105 struct ldt_struct *ldt; 106 107 /* READ_ONCE synchronizes with smp_store_release */ 108 ldt = READ_ONCE(mm->context.ldt); 109 110 /* 111 * Any change to mm->context.ldt is followed by an IPI to all 112 * CPUs with the mm active. The LDT will not be freed until 113 * after the IPI is handled by all such CPUs. This means that, 114 * if the ldt_struct changes before we return, the values we see 115 * will be safe, and the new values will be loaded before we run 116 * any user code. 117 * 118 * NB: don't try to convert this to use RCU without extreme care. 119 * We would still need IRQs off, because we don't want to change 120 * the local LDT after an IPI loaded a newer value than the one 121 * that we can see. 122 */ 123 124 if (unlikely(ldt)) { 125 if (static_cpu_has(X86_FEATURE_PTI)) { 126 if (WARN_ON_ONCE((unsigned long)ldt->slot > 1)) { 127 /* 128 * Whoops -- either the new LDT isn't mapped 129 * (if slot == -1) or is mapped into a bogus 130 * slot (if slot > 1). 131 */ 132 clear_LDT(); 133 return; 134 } 135 136 /* 137 * If page table isolation is enabled, ldt->entries 138 * will not be mapped in the userspace pagetables. 139 * Tell the CPU to access the LDT through the alias 140 * at ldt_slot_va(ldt->slot). 141 */ 142 set_ldt(ldt_slot_va(ldt->slot), ldt->nr_entries); 143 } else { 144 set_ldt(ldt->entries, ldt->nr_entries); 145 } 146 } else { 147 clear_LDT(); 148 } 149 #else 150 clear_LDT(); 151 #endif 152 } 153 154 static inline void switch_ldt(struct mm_struct *prev, struct mm_struct *next) 155 { 156 #ifdef CONFIG_MODIFY_LDT_SYSCALL 157 /* 158 * Load the LDT if either the old or new mm had an LDT. 159 * 160 * An mm will never go from having an LDT to not having an LDT. Two 161 * mms never share an LDT, so we don't gain anything by checking to 162 * see whether the LDT changed. There's also no guarantee that 163 * prev->context.ldt actually matches LDTR, but, if LDTR is non-NULL, 164 * then prev->context.ldt will also be non-NULL. 165 * 166 * If we really cared, we could optimize the case where prev == next 167 * and we're exiting lazy mode. Most of the time, if this happens, 168 * we don't actually need to reload LDTR, but modify_ldt() is mostly 169 * used by legacy code and emulators where we don't need this level of 170 * performance. 171 * 172 * This uses | instead of || because it generates better code. 173 */ 174 if (unlikely((unsigned long)prev->context.ldt | 175 (unsigned long)next->context.ldt)) 176 load_mm_ldt(next); 177 #endif 178 179 DEBUG_LOCKS_WARN_ON(preemptible()); 180 } 181 182 void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk); 183 184 static inline int init_new_context(struct task_struct *tsk, 185 struct mm_struct *mm) 186 { 187 mutex_init(&mm->context.lock); 188 189 mm->context.ctx_id = atomic64_inc_return(&last_mm_ctx_id); 190 atomic64_set(&mm->context.tlb_gen, 0); 191 192 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS 193 if (cpu_feature_enabled(X86_FEATURE_OSPKE)) { 194 /* pkey 0 is the default and always allocated */ 195 mm->context.pkey_allocation_map = 0x1; 196 /* -1 means unallocated or invalid */ 197 mm->context.execute_only_pkey = -1; 198 } 199 #endif 200 init_new_context_ldt(mm); 201 return 0; 202 } 203 static inline void destroy_context(struct mm_struct *mm) 204 { 205 destroy_context_ldt(mm); 206 } 207 208 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next, 209 struct task_struct *tsk); 210 211 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next, 212 struct task_struct *tsk); 213 #define switch_mm_irqs_off switch_mm_irqs_off 214 215 #define activate_mm(prev, next) \ 216 do { \ 217 paravirt_activate_mm((prev), (next)); \ 218 switch_mm((prev), (next), NULL); \ 219 } while (0); 220 221 #ifdef CONFIG_X86_32 222 #define deactivate_mm(tsk, mm) \ 223 do { \ 224 lazy_load_gs(0); \ 225 } while (0) 226 #else 227 #define deactivate_mm(tsk, mm) \ 228 do { \ 229 load_gs_index(0); \ 230 loadsegment(fs, 0); \ 231 } while (0) 232 #endif 233 234 static inline int arch_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm) 235 { 236 paravirt_arch_dup_mmap(oldmm, mm); 237 return ldt_dup_context(oldmm, mm); 238 } 239 240 static inline void arch_exit_mmap(struct mm_struct *mm) 241 { 242 paravirt_arch_exit_mmap(mm); 243 ldt_arch_exit_mmap(mm); 244 } 245 246 #ifdef CONFIG_X86_64 247 static inline bool is_64bit_mm(struct mm_struct *mm) 248 { 249 return !IS_ENABLED(CONFIG_IA32_EMULATION) || 250 !(mm->context.ia32_compat == TIF_IA32); 251 } 252 #else 253 static inline bool is_64bit_mm(struct mm_struct *mm) 254 { 255 return false; 256 } 257 #endif 258 259 static inline void arch_bprm_mm_init(struct mm_struct *mm, 260 struct vm_area_struct *vma) 261 { 262 mpx_mm_init(mm); 263 } 264 265 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma, 266 unsigned long start, unsigned long end) 267 { 268 /* 269 * mpx_notify_unmap() goes and reads a rarely-hot 270 * cacheline in the mm_struct. That can be expensive 271 * enough to be seen in profiles. 272 * 273 * The mpx_notify_unmap() call and its contents have been 274 * observed to affect munmap() performance on hardware 275 * where MPX is not present. 276 * 277 * The unlikely() optimizes for the fast case: no MPX 278 * in the CPU, or no MPX use in the process. Even if 279 * we get this wrong (in the unlikely event that MPX 280 * is widely enabled on some system) the overhead of 281 * MPX itself (reading bounds tables) is expected to 282 * overwhelm the overhead of getting this unlikely() 283 * consistently wrong. 284 */ 285 if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX))) 286 mpx_notify_unmap(mm, vma, start, end); 287 } 288 289 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS 290 static inline int vma_pkey(struct vm_area_struct *vma) 291 { 292 unsigned long vma_pkey_mask = VM_PKEY_BIT0 | VM_PKEY_BIT1 | 293 VM_PKEY_BIT2 | VM_PKEY_BIT3; 294 295 return (vma->vm_flags & vma_pkey_mask) >> VM_PKEY_SHIFT; 296 } 297 #else 298 static inline int vma_pkey(struct vm_area_struct *vma) 299 { 300 return 0; 301 } 302 #endif 303 304 /* 305 * We only want to enforce protection keys on the current process 306 * because we effectively have no access to PKRU for other 307 * processes or any way to tell *which * PKRU in a threaded 308 * process we could use. 309 * 310 * So do not enforce things if the VMA is not from the current 311 * mm, or if we are in a kernel thread. 312 */ 313 static inline bool vma_is_foreign(struct vm_area_struct *vma) 314 { 315 if (!current->mm) 316 return true; 317 /* 318 * Should PKRU be enforced on the access to this VMA? If 319 * the VMA is from another process, then PKRU has no 320 * relevance and should not be enforced. 321 */ 322 if (current->mm != vma->vm_mm) 323 return true; 324 325 return false; 326 } 327 328 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma, 329 bool write, bool execute, bool foreign) 330 { 331 /* pkeys never affect instruction fetches */ 332 if (execute) 333 return true; 334 /* allow access if the VMA is not one from this process */ 335 if (foreign || vma_is_foreign(vma)) 336 return true; 337 return __pkru_allows_pkey(vma_pkey(vma), write); 338 } 339 340 /* 341 * This can be used from process context to figure out what the value of 342 * CR3 is without needing to do a (slow) __read_cr3(). 343 * 344 * It's intended to be used for code like KVM that sneakily changes CR3 345 * and needs to restore it. It needs to be used very carefully. 346 */ 347 static inline unsigned long __get_current_cr3_fast(void) 348 { 349 unsigned long cr3 = build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd, 350 this_cpu_read(cpu_tlbstate.loaded_mm_asid)); 351 352 /* For now, be very restrictive about when this can be called. */ 353 VM_WARN_ON(in_nmi() || preemptible()); 354 355 VM_BUG_ON(cr3 != __read_cr3()); 356 return cr3; 357 } 358 359 #endif /* _ASM_X86_MMU_CONTEXT_H */ 360