1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * common.c - C code for kernel entry and exit 4 * Copyright (c) 2015 Andrew Lutomirski 5 * 6 * Based on asm and ptrace code by many authors. The code here originated 7 * in ptrace.c and signal.c. 8 */ 9 10 #include <linux/kernel.h> 11 #include <linux/sched.h> 12 #include <linux/sched/task_stack.h> 13 #include <linux/mm.h> 14 #include <linux/smp.h> 15 #include <linux/errno.h> 16 #include <linux/ptrace.h> 17 #include <linux/tracehook.h> 18 #include <linux/audit.h> 19 #include <linux/seccomp.h> 20 #include <linux/signal.h> 21 #include <linux/export.h> 22 #include <linux/context_tracking.h> 23 #include <linux/user-return-notifier.h> 24 #include <linux/nospec.h> 25 #include <linux/uprobes.h> 26 #include <linux/livepatch.h> 27 #include <linux/syscalls.h> 28 #include <linux/uaccess.h> 29 30 #include <asm/desc.h> 31 #include <asm/traps.h> 32 #include <asm/vdso.h> 33 #include <asm/cpufeature.h> 34 #include <asm/fpu/api.h> 35 #include <asm/nospec-branch.h> 36 37 #define CREATE_TRACE_POINTS 38 #include <trace/events/syscalls.h> 39 40 #ifdef CONFIG_CONTEXT_TRACKING 41 /* Called on entry from user mode with IRQs off. */ 42 __visible inline void enter_from_user_mode(void) 43 { 44 CT_WARN_ON(ct_state() != CONTEXT_USER); 45 user_exit_irqoff(); 46 } 47 #else 48 static inline void enter_from_user_mode(void) {} 49 #endif 50 51 static void do_audit_syscall_entry(struct pt_regs *regs, u32 arch) 52 { 53 #ifdef CONFIG_X86_64 54 if (arch == AUDIT_ARCH_X86_64) { 55 audit_syscall_entry(regs->orig_ax, regs->di, 56 regs->si, regs->dx, regs->r10); 57 } else 58 #endif 59 { 60 audit_syscall_entry(regs->orig_ax, regs->bx, 61 regs->cx, regs->dx, regs->si); 62 } 63 } 64 65 /* 66 * Returns the syscall nr to run (which should match regs->orig_ax) or -1 67 * to skip the syscall. 68 */ 69 static long syscall_trace_enter(struct pt_regs *regs) 70 { 71 u32 arch = in_ia32_syscall() ? AUDIT_ARCH_I386 : AUDIT_ARCH_X86_64; 72 73 struct thread_info *ti = current_thread_info(); 74 unsigned long ret = 0; 75 bool emulated = false; 76 u32 work; 77 78 if (IS_ENABLED(CONFIG_DEBUG_ENTRY)) 79 BUG_ON(regs != task_pt_regs(current)); 80 81 work = READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY; 82 83 if (unlikely(work & _TIF_SYSCALL_EMU)) 84 emulated = true; 85 86 if ((emulated || (work & _TIF_SYSCALL_TRACE)) && 87 tracehook_report_syscall_entry(regs)) 88 return -1L; 89 90 if (emulated) 91 return -1L; 92 93 #ifdef CONFIG_SECCOMP 94 /* 95 * Do seccomp after ptrace, to catch any tracer changes. 96 */ 97 if (work & _TIF_SECCOMP) { 98 struct seccomp_data sd; 99 100 sd.arch = arch; 101 sd.nr = regs->orig_ax; 102 sd.instruction_pointer = regs->ip; 103 #ifdef CONFIG_X86_64 104 if (arch == AUDIT_ARCH_X86_64) { 105 sd.args[0] = regs->di; 106 sd.args[1] = regs->si; 107 sd.args[2] = regs->dx; 108 sd.args[3] = regs->r10; 109 sd.args[4] = regs->r8; 110 sd.args[5] = regs->r9; 111 } else 112 #endif 113 { 114 sd.args[0] = regs->bx; 115 sd.args[1] = regs->cx; 116 sd.args[2] = regs->dx; 117 sd.args[3] = regs->si; 118 sd.args[4] = regs->di; 119 sd.args[5] = regs->bp; 120 } 121 122 ret = __secure_computing(&sd); 123 if (ret == -1) 124 return ret; 125 } 126 #endif 127 128 if (unlikely(test_thread_flag(TIF_SYSCALL_TRACEPOINT))) 129 trace_sys_enter(regs, regs->orig_ax); 130 131 do_audit_syscall_entry(regs, arch); 132 133 return ret ?: regs->orig_ax; 134 } 135 136 #define EXIT_TO_USERMODE_LOOP_FLAGS \ 137 (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_UPROBE | \ 138 _TIF_NEED_RESCHED | _TIF_USER_RETURN_NOTIFY | _TIF_PATCH_PENDING) 139 140 static void exit_to_usermode_loop(struct pt_regs *regs, u32 cached_flags) 141 { 142 /* 143 * In order to return to user mode, we need to have IRQs off with 144 * none of EXIT_TO_USERMODE_LOOP_FLAGS set. Several of these flags 145 * can be set at any time on preemptible kernels if we have IRQs on, 146 * so we need to loop. Disabling preemption wouldn't help: doing the 147 * work to clear some of the flags can sleep. 148 */ 149 while (true) { 150 /* We have work to do. */ 151 local_irq_enable(); 152 153 if (cached_flags & _TIF_NEED_RESCHED) 154 schedule(); 155 156 if (cached_flags & _TIF_UPROBE) 157 uprobe_notify_resume(regs); 158 159 if (cached_flags & _TIF_PATCH_PENDING) 160 klp_update_patch_state(current); 161 162 /* deal with pending signal delivery */ 163 if (cached_flags & _TIF_SIGPENDING) 164 do_signal(regs); 165 166 if (cached_flags & _TIF_NOTIFY_RESUME) { 167 clear_thread_flag(TIF_NOTIFY_RESUME); 168 tracehook_notify_resume(regs); 169 rseq_handle_notify_resume(NULL, regs); 170 } 171 172 if (cached_flags & _TIF_USER_RETURN_NOTIFY) 173 fire_user_return_notifiers(); 174 175 /* Disable IRQs and retry */ 176 local_irq_disable(); 177 178 cached_flags = READ_ONCE(current_thread_info()->flags); 179 180 if (!(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS)) 181 break; 182 } 183 } 184 185 /* Called with IRQs disabled. */ 186 __visible inline void prepare_exit_to_usermode(struct pt_regs *regs) 187 { 188 struct thread_info *ti = current_thread_info(); 189 u32 cached_flags; 190 191 addr_limit_user_check(); 192 193 lockdep_assert_irqs_disabled(); 194 lockdep_sys_exit(); 195 196 cached_flags = READ_ONCE(ti->flags); 197 198 if (unlikely(cached_flags & EXIT_TO_USERMODE_LOOP_FLAGS)) 199 exit_to_usermode_loop(regs, cached_flags); 200 201 /* Reload ti->flags; we may have rescheduled above. */ 202 cached_flags = READ_ONCE(ti->flags); 203 204 fpregs_assert_state_consistent(); 205 if (unlikely(cached_flags & _TIF_NEED_FPU_LOAD)) 206 switch_fpu_return(); 207 208 #ifdef CONFIG_COMPAT 209 /* 210 * Compat syscalls set TS_COMPAT. Make sure we clear it before 211 * returning to user mode. We need to clear it *after* signal 212 * handling, because syscall restart has a fixup for compat 213 * syscalls. The fixup is exercised by the ptrace_syscall_32 214 * selftest. 215 * 216 * We also need to clear TS_REGS_POKED_I386: the 32-bit tracer 217 * special case only applies after poking regs and before the 218 * very next return to user mode. 219 */ 220 ti->status &= ~(TS_COMPAT|TS_I386_REGS_POKED); 221 #endif 222 223 user_enter_irqoff(); 224 225 mds_user_clear_cpu_buffers(); 226 } 227 228 #define SYSCALL_EXIT_WORK_FLAGS \ 229 (_TIF_SYSCALL_TRACE | _TIF_SYSCALL_AUDIT | \ 230 _TIF_SINGLESTEP | _TIF_SYSCALL_TRACEPOINT) 231 232 static void syscall_slow_exit_work(struct pt_regs *regs, u32 cached_flags) 233 { 234 bool step; 235 236 audit_syscall_exit(regs); 237 238 if (cached_flags & _TIF_SYSCALL_TRACEPOINT) 239 trace_sys_exit(regs, regs->ax); 240 241 /* 242 * If TIF_SYSCALL_EMU is set, we only get here because of 243 * TIF_SINGLESTEP (i.e. this is PTRACE_SYSEMU_SINGLESTEP). 244 * We already reported this syscall instruction in 245 * syscall_trace_enter(). 246 */ 247 step = unlikely( 248 (cached_flags & (_TIF_SINGLESTEP | _TIF_SYSCALL_EMU)) 249 == _TIF_SINGLESTEP); 250 if (step || cached_flags & _TIF_SYSCALL_TRACE) 251 tracehook_report_syscall_exit(regs, step); 252 } 253 254 /* 255 * Called with IRQs on and fully valid regs. Returns with IRQs off in a 256 * state such that we can immediately switch to user mode. 257 */ 258 __visible inline void syscall_return_slowpath(struct pt_regs *regs) 259 { 260 struct thread_info *ti = current_thread_info(); 261 u32 cached_flags = READ_ONCE(ti->flags); 262 263 CT_WARN_ON(ct_state() != CONTEXT_KERNEL); 264 265 if (IS_ENABLED(CONFIG_PROVE_LOCKING) && 266 WARN(irqs_disabled(), "syscall %ld left IRQs disabled", regs->orig_ax)) 267 local_irq_enable(); 268 269 rseq_syscall(regs); 270 271 /* 272 * First do one-time work. If these work items are enabled, we 273 * want to run them exactly once per syscall exit with IRQs on. 274 */ 275 if (unlikely(cached_flags & SYSCALL_EXIT_WORK_FLAGS)) 276 syscall_slow_exit_work(regs, cached_flags); 277 278 local_irq_disable(); 279 prepare_exit_to_usermode(regs); 280 } 281 282 #ifdef CONFIG_X86_64 283 __visible void do_syscall_64(unsigned long nr, struct pt_regs *regs) 284 { 285 struct thread_info *ti; 286 287 enter_from_user_mode(); 288 local_irq_enable(); 289 ti = current_thread_info(); 290 if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY) 291 nr = syscall_trace_enter(regs); 292 293 /* 294 * NB: Native and x32 syscalls are dispatched from the same 295 * table. The only functional difference is the x32 bit in 296 * regs->orig_ax, which changes the behavior of some syscalls. 297 */ 298 nr &= __SYSCALL_MASK; 299 if (likely(nr < NR_syscalls)) { 300 nr = array_index_nospec(nr, NR_syscalls); 301 regs->ax = sys_call_table[nr](regs); 302 } 303 304 syscall_return_slowpath(regs); 305 } 306 #endif 307 308 #if defined(CONFIG_X86_32) || defined(CONFIG_IA32_EMULATION) 309 /* 310 * Does a 32-bit syscall. Called with IRQs on in CONTEXT_KERNEL. Does 311 * all entry and exit work and returns with IRQs off. This function is 312 * extremely hot in workloads that use it, and it's usually called from 313 * do_fast_syscall_32, so forcibly inline it to improve performance. 314 */ 315 static __always_inline void do_syscall_32_irqs_on(struct pt_regs *regs) 316 { 317 struct thread_info *ti = current_thread_info(); 318 unsigned int nr = (unsigned int)regs->orig_ax; 319 320 #ifdef CONFIG_IA32_EMULATION 321 ti->status |= TS_COMPAT; 322 #endif 323 324 if (READ_ONCE(ti->flags) & _TIF_WORK_SYSCALL_ENTRY) { 325 /* 326 * Subtlety here: if ptrace pokes something larger than 327 * 2^32-1 into orig_ax, this truncates it. This may or 328 * may not be necessary, but it matches the old asm 329 * behavior. 330 */ 331 nr = syscall_trace_enter(regs); 332 } 333 334 if (likely(nr < IA32_NR_syscalls)) { 335 nr = array_index_nospec(nr, IA32_NR_syscalls); 336 #ifdef CONFIG_IA32_EMULATION 337 regs->ax = ia32_sys_call_table[nr](regs); 338 #else 339 /* 340 * It's possible that a 32-bit syscall implementation 341 * takes a 64-bit parameter but nonetheless assumes that 342 * the high bits are zero. Make sure we zero-extend all 343 * of the args. 344 */ 345 regs->ax = ia32_sys_call_table[nr]( 346 (unsigned int)regs->bx, (unsigned int)regs->cx, 347 (unsigned int)regs->dx, (unsigned int)regs->si, 348 (unsigned int)regs->di, (unsigned int)regs->bp); 349 #endif /* CONFIG_IA32_EMULATION */ 350 } 351 352 syscall_return_slowpath(regs); 353 } 354 355 /* Handles int $0x80 */ 356 __visible void do_int80_syscall_32(struct pt_regs *regs) 357 { 358 enter_from_user_mode(); 359 local_irq_enable(); 360 do_syscall_32_irqs_on(regs); 361 } 362 363 /* Returns 0 to return using IRET or 1 to return using SYSEXIT/SYSRETL. */ 364 __visible long do_fast_syscall_32(struct pt_regs *regs) 365 { 366 /* 367 * Called using the internal vDSO SYSENTER/SYSCALL32 calling 368 * convention. Adjust regs so it looks like we entered using int80. 369 */ 370 371 unsigned long landing_pad = (unsigned long)current->mm->context.vdso + 372 vdso_image_32.sym_int80_landing_pad; 373 374 /* 375 * SYSENTER loses EIP, and even SYSCALL32 needs us to skip forward 376 * so that 'regs->ip -= 2' lands back on an int $0x80 instruction. 377 * Fix it up. 378 */ 379 regs->ip = landing_pad; 380 381 enter_from_user_mode(); 382 383 local_irq_enable(); 384 385 /* Fetch EBP from where the vDSO stashed it. */ 386 if ( 387 #ifdef CONFIG_X86_64 388 /* 389 * Micro-optimization: the pointer we're following is explicitly 390 * 32 bits, so it can't be out of range. 391 */ 392 __get_user(*(u32 *)®s->bp, 393 (u32 __user __force *)(unsigned long)(u32)regs->sp) 394 #else 395 get_user(*(u32 *)®s->bp, 396 (u32 __user __force *)(unsigned long)(u32)regs->sp) 397 #endif 398 ) { 399 400 /* User code screwed up. */ 401 local_irq_disable(); 402 regs->ax = -EFAULT; 403 prepare_exit_to_usermode(regs); 404 return 0; /* Keep it simple: use IRET. */ 405 } 406 407 /* Now this is just like a normal syscall. */ 408 do_syscall_32_irqs_on(regs); 409 410 #ifdef CONFIG_X86_64 411 /* 412 * Opportunistic SYSRETL: if possible, try to return using SYSRETL. 413 * SYSRETL is available on all 64-bit CPUs, so we don't need to 414 * bother with SYSEXIT. 415 * 416 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP, 417 * because the ECX fixup above will ensure that this is essentially 418 * never the case. 419 */ 420 return regs->cs == __USER32_CS && regs->ss == __USER_DS && 421 regs->ip == landing_pad && 422 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF)) == 0; 423 #else 424 /* 425 * Opportunistic SYSEXIT: if possible, try to return using SYSEXIT. 426 * 427 * Unlike 64-bit opportunistic SYSRET, we can't check that CX == IP, 428 * because the ECX fixup above will ensure that this is essentially 429 * never the case. 430 * 431 * We don't allow syscalls at all from VM86 mode, but we still 432 * need to check VM, because we might be returning from sys_vm86. 433 */ 434 return static_cpu_has(X86_FEATURE_SEP) && 435 regs->cs == __USER_CS && regs->ss == __USER_DS && 436 regs->ip == landing_pad && 437 (regs->flags & (X86_EFLAGS_RF | X86_EFLAGS_TF | X86_EFLAGS_VM)) == 0; 438 #endif 439 } 440 #endif 441