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 return (void *)fix_to_virt(FIX_HOLE); 78 #endif 79 } 80 81 /* 82 * Used for LDT copy/destruction. 83 */ 84 static inline void init_new_context_ldt(struct mm_struct *mm) 85 { 86 mm->context.ldt = NULL; 87 init_rwsem(&mm->context.ldt_usr_sem); 88 } 89 int ldt_dup_context(struct mm_struct *oldmm, struct mm_struct *mm); 90 void destroy_context_ldt(struct mm_struct *mm); 91 void ldt_arch_exit_mmap(struct mm_struct *mm); 92 #else /* CONFIG_MODIFY_LDT_SYSCALL */ 93 static inline void init_new_context_ldt(struct mm_struct *mm) { } 94 static inline int ldt_dup_context(struct mm_struct *oldmm, 95 struct mm_struct *mm) 96 { 97 return 0; 98 } 99 static inline void destroy_context_ldt(struct mm_struct *mm) { } 100 static inline void ldt_arch_exit_mmap(struct mm_struct *mm) { } 101 #endif 102 103 static inline void load_mm_ldt(struct mm_struct *mm) 104 { 105 #ifdef CONFIG_MODIFY_LDT_SYSCALL 106 struct ldt_struct *ldt; 107 108 /* READ_ONCE synchronizes with smp_store_release */ 109 ldt = READ_ONCE(mm->context.ldt); 110 111 /* 112 * Any change to mm->context.ldt is followed by an IPI to all 113 * CPUs with the mm active. The LDT will not be freed until 114 * after the IPI is handled by all such CPUs. This means that, 115 * if the ldt_struct changes before we return, the values we see 116 * will be safe, and the new values will be loaded before we run 117 * any user code. 118 * 119 * NB: don't try to convert this to use RCU without extreme care. 120 * We would still need IRQs off, because we don't want to change 121 * the local LDT after an IPI loaded a newer value than the one 122 * that we can see. 123 */ 124 125 if (unlikely(ldt)) { 126 if (static_cpu_has(X86_FEATURE_PTI)) { 127 if (WARN_ON_ONCE((unsigned long)ldt->slot > 1)) { 128 /* 129 * Whoops -- either the new LDT isn't mapped 130 * (if slot == -1) or is mapped into a bogus 131 * slot (if slot > 1). 132 */ 133 clear_LDT(); 134 return; 135 } 136 137 /* 138 * If page table isolation is enabled, ldt->entries 139 * will not be mapped in the userspace pagetables. 140 * Tell the CPU to access the LDT through the alias 141 * at ldt_slot_va(ldt->slot). 142 */ 143 set_ldt(ldt_slot_va(ldt->slot), ldt->nr_entries); 144 } else { 145 set_ldt(ldt->entries, ldt->nr_entries); 146 } 147 } else { 148 clear_LDT(); 149 } 150 #else 151 clear_LDT(); 152 #endif 153 } 154 155 static inline void switch_ldt(struct mm_struct *prev, struct mm_struct *next) 156 { 157 #ifdef CONFIG_MODIFY_LDT_SYSCALL 158 /* 159 * Load the LDT if either the old or new mm had an LDT. 160 * 161 * An mm will never go from having an LDT to not having an LDT. Two 162 * mms never share an LDT, so we don't gain anything by checking to 163 * see whether the LDT changed. There's also no guarantee that 164 * prev->context.ldt actually matches LDTR, but, if LDTR is non-NULL, 165 * then prev->context.ldt will also be non-NULL. 166 * 167 * If we really cared, we could optimize the case where prev == next 168 * and we're exiting lazy mode. Most of the time, if this happens, 169 * we don't actually need to reload LDTR, but modify_ldt() is mostly 170 * used by legacy code and emulators where we don't need this level of 171 * performance. 172 * 173 * This uses | instead of || because it generates better code. 174 */ 175 if (unlikely((unsigned long)prev->context.ldt | 176 (unsigned long)next->context.ldt)) 177 load_mm_ldt(next); 178 #endif 179 180 DEBUG_LOCKS_WARN_ON(preemptible()); 181 } 182 183 void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk); 184 185 static inline int init_new_context(struct task_struct *tsk, 186 struct mm_struct *mm) 187 { 188 mutex_init(&mm->context.lock); 189 190 mm->context.ctx_id = atomic64_inc_return(&last_mm_ctx_id); 191 atomic64_set(&mm->context.tlb_gen, 0); 192 193 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS 194 if (cpu_feature_enabled(X86_FEATURE_OSPKE)) { 195 /* pkey 0 is the default and always allocated */ 196 mm->context.pkey_allocation_map = 0x1; 197 /* -1 means unallocated or invalid */ 198 mm->context.execute_only_pkey = -1; 199 } 200 #endif 201 init_new_context_ldt(mm); 202 return 0; 203 } 204 static inline void destroy_context(struct mm_struct *mm) 205 { 206 destroy_context_ldt(mm); 207 } 208 209 extern void switch_mm(struct mm_struct *prev, struct mm_struct *next, 210 struct task_struct *tsk); 211 212 extern void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next, 213 struct task_struct *tsk); 214 #define switch_mm_irqs_off switch_mm_irqs_off 215 216 #define activate_mm(prev, next) \ 217 do { \ 218 paravirt_activate_mm((prev), (next)); \ 219 switch_mm((prev), (next), NULL); \ 220 } while (0); 221 222 #ifdef CONFIG_X86_32 223 #define deactivate_mm(tsk, mm) \ 224 do { \ 225 lazy_load_gs(0); \ 226 } while (0) 227 #else 228 #define deactivate_mm(tsk, mm) \ 229 do { \ 230 load_gs_index(0); \ 231 loadsegment(fs, 0); \ 232 } while (0) 233 #endif 234 235 static inline int arch_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm) 236 { 237 paravirt_arch_dup_mmap(oldmm, mm); 238 return ldt_dup_context(oldmm, mm); 239 } 240 241 static inline void arch_exit_mmap(struct mm_struct *mm) 242 { 243 paravirt_arch_exit_mmap(mm); 244 ldt_arch_exit_mmap(mm); 245 } 246 247 #ifdef CONFIG_X86_64 248 static inline bool is_64bit_mm(struct mm_struct *mm) 249 { 250 return !IS_ENABLED(CONFIG_IA32_EMULATION) || 251 !(mm->context.ia32_compat == TIF_IA32); 252 } 253 #else 254 static inline bool is_64bit_mm(struct mm_struct *mm) 255 { 256 return false; 257 } 258 #endif 259 260 static inline void arch_bprm_mm_init(struct mm_struct *mm, 261 struct vm_area_struct *vma) 262 { 263 mpx_mm_init(mm); 264 } 265 266 static inline void arch_unmap(struct mm_struct *mm, struct vm_area_struct *vma, 267 unsigned long start, unsigned long end) 268 { 269 /* 270 * mpx_notify_unmap() goes and reads a rarely-hot 271 * cacheline in the mm_struct. That can be expensive 272 * enough to be seen in profiles. 273 * 274 * The mpx_notify_unmap() call and its contents have been 275 * observed to affect munmap() performance on hardware 276 * where MPX is not present. 277 * 278 * The unlikely() optimizes for the fast case: no MPX 279 * in the CPU, or no MPX use in the process. Even if 280 * we get this wrong (in the unlikely event that MPX 281 * is widely enabled on some system) the overhead of 282 * MPX itself (reading bounds tables) is expected to 283 * overwhelm the overhead of getting this unlikely() 284 * consistently wrong. 285 */ 286 if (unlikely(cpu_feature_enabled(X86_FEATURE_MPX))) 287 mpx_notify_unmap(mm, vma, start, end); 288 } 289 290 #ifdef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS 291 static inline int vma_pkey(struct vm_area_struct *vma) 292 { 293 unsigned long vma_pkey_mask = VM_PKEY_BIT0 | VM_PKEY_BIT1 | 294 VM_PKEY_BIT2 | VM_PKEY_BIT3; 295 296 return (vma->vm_flags & vma_pkey_mask) >> VM_PKEY_SHIFT; 297 } 298 #else 299 static inline int vma_pkey(struct vm_area_struct *vma) 300 { 301 return 0; 302 } 303 #endif 304 305 /* 306 * We only want to enforce protection keys on the current process 307 * because we effectively have no access to PKRU for other 308 * processes or any way to tell *which * PKRU in a threaded 309 * process we could use. 310 * 311 * So do not enforce things if the VMA is not from the current 312 * mm, or if we are in a kernel thread. 313 */ 314 static inline bool vma_is_foreign(struct vm_area_struct *vma) 315 { 316 if (!current->mm) 317 return true; 318 /* 319 * Should PKRU be enforced on the access to this VMA? If 320 * the VMA is from another process, then PKRU has no 321 * relevance and should not be enforced. 322 */ 323 if (current->mm != vma->vm_mm) 324 return true; 325 326 return false; 327 } 328 329 static inline bool arch_vma_access_permitted(struct vm_area_struct *vma, 330 bool write, bool execute, bool foreign) 331 { 332 /* pkeys never affect instruction fetches */ 333 if (execute) 334 return true; 335 /* allow access if the VMA is not one from this process */ 336 if (foreign || vma_is_foreign(vma)) 337 return true; 338 return __pkru_allows_pkey(vma_pkey(vma), write); 339 } 340 341 /* 342 * This can be used from process context to figure out what the value of 343 * CR3 is without needing to do a (slow) __read_cr3(). 344 * 345 * It's intended to be used for code like KVM that sneakily changes CR3 346 * and needs to restore it. It needs to be used very carefully. 347 */ 348 static inline unsigned long __get_current_cr3_fast(void) 349 { 350 unsigned long cr3 = build_cr3(this_cpu_read(cpu_tlbstate.loaded_mm)->pgd, 351 this_cpu_read(cpu_tlbstate.loaded_mm_asid)); 352 353 /* For now, be very restrictive about when this can be called. */ 354 VM_WARN_ON(in_nmi() || preemptible()); 355 356 VM_BUG_ON(cr3 != __read_cr3()); 357 return cr3; 358 } 359 360 #endif /* _ASM_X86_MMU_CONTEXT_H */ 361