1/* SPDX-License-Identifier: GPL-2.0 */ 2/* 3 * Compatibility mode system call entry point for x86-64. 4 * 5 * Copyright 2000-2002 Andi Kleen, SuSE Labs. 6 */ 7#include "calling.h" 8#include <asm/asm-offsets.h> 9#include <asm/current.h> 10#include <asm/errno.h> 11#include <asm/ia32_unistd.h> 12#include <asm/thread_info.h> 13#include <asm/segment.h> 14#include <asm/irqflags.h> 15#include <asm/asm.h> 16#include <asm/smap.h> 17#include <linux/linkage.h> 18#include <linux/err.h> 19 20 .section .entry.text, "ax" 21 22/* 23 * 32-bit SYSENTER entry. 24 * 25 * 32-bit system calls through the vDSO's __kernel_vsyscall enter here 26 * on 64-bit kernels running on Intel CPUs. 27 * 28 * The SYSENTER instruction, in principle, should *only* occur in the 29 * vDSO. In practice, a small number of Android devices were shipped 30 * with a copy of Bionic that inlined a SYSENTER instruction. This 31 * never happened in any of Google's Bionic versions -- it only happened 32 * in a narrow range of Intel-provided versions. 33 * 34 * SYSENTER loads SS, RSP, CS, and RIP from previously programmed MSRs. 35 * IF and VM in RFLAGS are cleared (IOW: interrupts are off). 36 * SYSENTER does not save anything on the stack, 37 * and does not save old RIP (!!!), RSP, or RFLAGS. 38 * 39 * Arguments: 40 * eax system call number 41 * ebx arg1 42 * ecx arg2 43 * edx arg3 44 * esi arg4 45 * edi arg5 46 * ebp user stack 47 * 0(%ebp) arg6 48 */ 49SYM_CODE_START(entry_SYSENTER_compat) 50 UNWIND_HINT_EMPTY 51 ENDBR 52 /* Interrupts are off on entry. */ 53 SWAPGS 54 55 pushq %rax 56 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 57 popq %rax 58 59 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 60 61 /* Construct struct pt_regs on stack */ 62 pushq $__USER32_DS /* pt_regs->ss */ 63 pushq $0 /* pt_regs->sp = 0 (placeholder) */ 64 65 /* 66 * Push flags. This is nasty. First, interrupts are currently 67 * off, but we need pt_regs->flags to have IF set. Second, if TS 68 * was set in usermode, it's still set, and we're singlestepping 69 * through this code. do_SYSENTER_32() will fix up IF. 70 */ 71 pushfq /* pt_regs->flags (except IF = 0) */ 72 pushq $__USER32_CS /* pt_regs->cs */ 73 pushq $0 /* pt_regs->ip = 0 (placeholder) */ 74SYM_INNER_LABEL(entry_SYSENTER_compat_after_hwframe, SYM_L_GLOBAL) 75 76 /* 77 * User tracing code (ptrace or signal handlers) might assume that 78 * the saved RAX contains a 32-bit number when we're invoking a 32-bit 79 * syscall. Just in case the high bits are nonzero, zero-extend 80 * the syscall number. (This could almost certainly be deleted 81 * with no ill effects.) 82 */ 83 movl %eax, %eax 84 85 pushq %rax /* pt_regs->orig_ax */ 86 pushq %rdi /* pt_regs->di */ 87 pushq %rsi /* pt_regs->si */ 88 pushq %rdx /* pt_regs->dx */ 89 pushq %rcx /* pt_regs->cx */ 90 pushq $-ENOSYS /* pt_regs->ax */ 91 pushq $0 /* pt_regs->r8 = 0 */ 92 xorl %r8d, %r8d /* nospec r8 */ 93 pushq $0 /* pt_regs->r9 = 0 */ 94 xorl %r9d, %r9d /* nospec r9 */ 95 pushq $0 /* pt_regs->r10 = 0 */ 96 xorl %r10d, %r10d /* nospec r10 */ 97 pushq $0 /* pt_regs->r11 = 0 */ 98 xorl %r11d, %r11d /* nospec r11 */ 99 pushq %rbx /* pt_regs->rbx */ 100 xorl %ebx, %ebx /* nospec rbx */ 101 pushq %rbp /* pt_regs->rbp (will be overwritten) */ 102 xorl %ebp, %ebp /* nospec rbp */ 103 pushq $0 /* pt_regs->r12 = 0 */ 104 xorl %r12d, %r12d /* nospec r12 */ 105 pushq $0 /* pt_regs->r13 = 0 */ 106 xorl %r13d, %r13d /* nospec r13 */ 107 pushq $0 /* pt_regs->r14 = 0 */ 108 xorl %r14d, %r14d /* nospec r14 */ 109 pushq $0 /* pt_regs->r15 = 0 */ 110 xorl %r15d, %r15d /* nospec r15 */ 111 112 UNWIND_HINT_REGS 113 114 cld 115 116 /* 117 * SYSENTER doesn't filter flags, so we need to clear NT and AC 118 * ourselves. To save a few cycles, we can check whether 119 * either was set instead of doing an unconditional popfq. 120 * This needs to happen before enabling interrupts so that 121 * we don't get preempted with NT set. 122 * 123 * If TF is set, we will single-step all the way to here -- do_debug 124 * will ignore all the traps. (Yes, this is slow, but so is 125 * single-stepping in general. This allows us to avoid having 126 * a more complicated code to handle the case where a user program 127 * forces us to single-step through the SYSENTER entry code.) 128 * 129 * NB.: .Lsysenter_fix_flags is a label with the code under it moved 130 * out-of-line as an optimization: NT is unlikely to be set in the 131 * majority of the cases and instead of polluting the I$ unnecessarily, 132 * we're keeping that code behind a branch which will predict as 133 * not-taken and therefore its instructions won't be fetched. 134 */ 135 testl $X86_EFLAGS_NT|X86_EFLAGS_AC|X86_EFLAGS_TF, EFLAGS(%rsp) 136 jnz .Lsysenter_fix_flags 137.Lsysenter_flags_fixed: 138 139 movq %rsp, %rdi 140 call do_SYSENTER_32 141 /* XEN PV guests always use IRET path */ 142 ALTERNATIVE "testl %eax, %eax; jz swapgs_restore_regs_and_return_to_usermode", \ 143 "jmp swapgs_restore_regs_and_return_to_usermode", X86_FEATURE_XENPV 144 jmp sysret32_from_system_call 145 146.Lsysenter_fix_flags: 147 pushq $X86_EFLAGS_FIXED 148 popfq 149 jmp .Lsysenter_flags_fixed 150SYM_INNER_LABEL(__end_entry_SYSENTER_compat, SYM_L_GLOBAL) 151 ANNOTATE_NOENDBR // is_sysenter_singlestep 152SYM_CODE_END(entry_SYSENTER_compat) 153 154/* 155 * 32-bit SYSCALL entry. 156 * 157 * 32-bit system calls through the vDSO's __kernel_vsyscall enter here 158 * on 64-bit kernels running on AMD CPUs. 159 * 160 * The SYSCALL instruction, in principle, should *only* occur in the 161 * vDSO. In practice, it appears that this really is the case. 162 * As evidence: 163 * 164 * - The calling convention for SYSCALL has changed several times without 165 * anyone noticing. 166 * 167 * - Prior to the in-kernel X86_BUG_SYSRET_SS_ATTRS fixup, anything 168 * user task that did SYSCALL without immediately reloading SS 169 * would randomly crash. 170 * 171 * - Most programmers do not directly target AMD CPUs, and the 32-bit 172 * SYSCALL instruction does not exist on Intel CPUs. Even on AMD 173 * CPUs, Linux disables the SYSCALL instruction on 32-bit kernels 174 * because the SYSCALL instruction in legacy/native 32-bit mode (as 175 * opposed to compat mode) is sufficiently poorly designed as to be 176 * essentially unusable. 177 * 178 * 32-bit SYSCALL saves RIP to RCX, clears RFLAGS.RF, then saves 179 * RFLAGS to R11, then loads new SS, CS, and RIP from previously 180 * programmed MSRs. RFLAGS gets masked by a value from another MSR 181 * (so CLD and CLAC are not needed). SYSCALL does not save anything on 182 * the stack and does not change RSP. 183 * 184 * Note: RFLAGS saving+masking-with-MSR happens only in Long mode 185 * (in legacy 32-bit mode, IF, RF and VM bits are cleared and that's it). 186 * Don't get confused: RFLAGS saving+masking depends on Long Mode Active bit 187 * (EFER.LMA=1), NOT on bitness of userspace where SYSCALL executes 188 * or target CS descriptor's L bit (SYSCALL does not read segment descriptors). 189 * 190 * Arguments: 191 * eax system call number 192 * ecx return address 193 * ebx arg1 194 * ebp arg2 (note: not saved in the stack frame, should not be touched) 195 * edx arg3 196 * esi arg4 197 * edi arg5 198 * esp user stack 199 * 0(%esp) arg6 200 */ 201SYM_CODE_START(entry_SYSCALL_compat) 202 UNWIND_HINT_EMPTY 203 ENDBR 204 /* Interrupts are off on entry. */ 205 swapgs 206 207 /* Stash user ESP */ 208 movl %esp, %r8d 209 210 /* Use %rsp as scratch reg. User ESP is stashed in r8 */ 211 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp 212 213 /* Switch to the kernel stack */ 214 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 215 216SYM_INNER_LABEL(entry_SYSCALL_compat_safe_stack, SYM_L_GLOBAL) 217 ANNOTATE_NOENDBR 218 219 /* Construct struct pt_regs on stack */ 220 pushq $__USER32_DS /* pt_regs->ss */ 221 pushq %r8 /* pt_regs->sp */ 222 pushq %r11 /* pt_regs->flags */ 223 pushq $__USER32_CS /* pt_regs->cs */ 224 pushq %rcx /* pt_regs->ip */ 225SYM_INNER_LABEL(entry_SYSCALL_compat_after_hwframe, SYM_L_GLOBAL) 226 movl %eax, %eax /* discard orig_ax high bits */ 227 pushq %rax /* pt_regs->orig_ax */ 228 pushq %rdi /* pt_regs->di */ 229 pushq %rsi /* pt_regs->si */ 230 xorl %esi, %esi /* nospec si */ 231 pushq %rdx /* pt_regs->dx */ 232 xorl %edx, %edx /* nospec dx */ 233 pushq %rbp /* pt_regs->cx (stashed in bp) */ 234 xorl %ecx, %ecx /* nospec cx */ 235 pushq $-ENOSYS /* pt_regs->ax */ 236 pushq $0 /* pt_regs->r8 = 0 */ 237 xorl %r8d, %r8d /* nospec r8 */ 238 pushq $0 /* pt_regs->r9 = 0 */ 239 xorl %r9d, %r9d /* nospec r9 */ 240 pushq $0 /* pt_regs->r10 = 0 */ 241 xorl %r10d, %r10d /* nospec r10 */ 242 pushq $0 /* pt_regs->r11 = 0 */ 243 xorl %r11d, %r11d /* nospec r11 */ 244 pushq %rbx /* pt_regs->rbx */ 245 xorl %ebx, %ebx /* nospec rbx */ 246 pushq %rbp /* pt_regs->rbp (will be overwritten) */ 247 xorl %ebp, %ebp /* nospec rbp */ 248 pushq $0 /* pt_regs->r12 = 0 */ 249 xorl %r12d, %r12d /* nospec r12 */ 250 pushq $0 /* pt_regs->r13 = 0 */ 251 xorl %r13d, %r13d /* nospec r13 */ 252 pushq $0 /* pt_regs->r14 = 0 */ 253 xorl %r14d, %r14d /* nospec r14 */ 254 pushq $0 /* pt_regs->r15 = 0 */ 255 xorl %r15d, %r15d /* nospec r15 */ 256 257 UNWIND_HINT_REGS 258 259 movq %rsp, %rdi 260 call do_fast_syscall_32 261 /* XEN PV guests always use IRET path */ 262 ALTERNATIVE "testl %eax, %eax; jz swapgs_restore_regs_and_return_to_usermode", \ 263 "jmp swapgs_restore_regs_and_return_to_usermode", X86_FEATURE_XENPV 264 265 /* Opportunistic SYSRET */ 266sysret32_from_system_call: 267 /* 268 * We are not going to return to userspace from the trampoline 269 * stack. So let's erase the thread stack right now. 270 */ 271 STACKLEAK_ERASE 272 273 movq RBX(%rsp), %rbx /* pt_regs->rbx */ 274 movq RBP(%rsp), %rbp /* pt_regs->rbp */ 275 movq EFLAGS(%rsp), %r11 /* pt_regs->flags (in r11) */ 276 movq RIP(%rsp), %rcx /* pt_regs->ip (in rcx) */ 277 addq $RAX, %rsp /* Skip r8-r15 */ 278 popq %rax /* pt_regs->rax */ 279 popq %rdx /* Skip pt_regs->cx */ 280 popq %rdx /* pt_regs->dx */ 281 popq %rsi /* pt_regs->si */ 282 popq %rdi /* pt_regs->di */ 283 284 /* 285 * USERGS_SYSRET32 does: 286 * GSBASE = user's GS base 287 * EIP = ECX 288 * RFLAGS = R11 289 * CS = __USER32_CS 290 * SS = __USER_DS 291 * 292 * ECX will not match pt_regs->cx, but we're returning to a vDSO 293 * trampoline that will fix up RCX, so this is okay. 294 * 295 * R12-R15 are callee-saved, so they contain whatever was in them 296 * when the system call started, which is already known to user 297 * code. We zero R8-R10 to avoid info leaks. 298 */ 299 movq RSP-ORIG_RAX(%rsp), %rsp 300 301 /* 302 * The original userspace %rsp (RSP-ORIG_RAX(%rsp)) is stored 303 * on the process stack which is not mapped to userspace and 304 * not readable after we SWITCH_TO_USER_CR3. Delay the CR3 305 * switch until after after the last reference to the process 306 * stack. 307 * 308 * %r8/%r9 are zeroed before the sysret, thus safe to clobber. 309 */ 310 SWITCH_TO_USER_CR3_NOSTACK scratch_reg=%r8 scratch_reg2=%r9 311 312 xorl %r8d, %r8d 313 xorl %r9d, %r9d 314 xorl %r10d, %r10d 315 swapgs 316 sysretl 317SYM_CODE_END(entry_SYSCALL_compat) 318 319/* 320 * 32-bit legacy system call entry. 321 * 322 * 32-bit x86 Linux system calls traditionally used the INT $0x80 323 * instruction. INT $0x80 lands here. 324 * 325 * This entry point can be used by 32-bit and 64-bit programs to perform 326 * 32-bit system calls. Instances of INT $0x80 can be found inline in 327 * various programs and libraries. It is also used by the vDSO's 328 * __kernel_vsyscall fallback for hardware that doesn't support a faster 329 * entry method. Restarted 32-bit system calls also fall back to INT 330 * $0x80 regardless of what instruction was originally used to do the 331 * system call. 332 * 333 * This is considered a slow path. It is not used by most libc 334 * implementations on modern hardware except during process startup. 335 * 336 * Arguments: 337 * eax system call number 338 * ebx arg1 339 * ecx arg2 340 * edx arg3 341 * esi arg4 342 * edi arg5 343 * ebp arg6 344 */ 345SYM_CODE_START(entry_INT80_compat) 346 UNWIND_HINT_EMPTY 347 ENDBR 348 /* 349 * Interrupts are off on entry. 350 */ 351 ASM_CLAC /* Do this early to minimize exposure */ 352 SWAPGS 353 354 /* 355 * User tracing code (ptrace or signal handlers) might assume that 356 * the saved RAX contains a 32-bit number when we're invoking a 32-bit 357 * syscall. Just in case the high bits are nonzero, zero-extend 358 * the syscall number. (This could almost certainly be deleted 359 * with no ill effects.) 360 */ 361 movl %eax, %eax 362 363 /* switch to thread stack expects orig_ax and rdi to be pushed */ 364 pushq %rax /* pt_regs->orig_ax */ 365 pushq %rdi /* pt_regs->di */ 366 367 /* Need to switch before accessing the thread stack. */ 368 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi 369 370 /* In the Xen PV case we already run on the thread stack. */ 371 ALTERNATIVE "", "jmp .Lint80_keep_stack", X86_FEATURE_XENPV 372 373 movq %rsp, %rdi 374 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 375 376 pushq 6*8(%rdi) /* regs->ss */ 377 pushq 5*8(%rdi) /* regs->rsp */ 378 pushq 4*8(%rdi) /* regs->eflags */ 379 pushq 3*8(%rdi) /* regs->cs */ 380 pushq 2*8(%rdi) /* regs->ip */ 381 pushq 1*8(%rdi) /* regs->orig_ax */ 382 pushq (%rdi) /* pt_regs->di */ 383.Lint80_keep_stack: 384 385 pushq %rsi /* pt_regs->si */ 386 xorl %esi, %esi /* nospec si */ 387 pushq %rdx /* pt_regs->dx */ 388 xorl %edx, %edx /* nospec dx */ 389 pushq %rcx /* pt_regs->cx */ 390 xorl %ecx, %ecx /* nospec cx */ 391 pushq $-ENOSYS /* pt_regs->ax */ 392 pushq %r8 /* pt_regs->r8 */ 393 xorl %r8d, %r8d /* nospec r8 */ 394 pushq %r9 /* pt_regs->r9 */ 395 xorl %r9d, %r9d /* nospec r9 */ 396 pushq %r10 /* pt_regs->r10*/ 397 xorl %r10d, %r10d /* nospec r10 */ 398 pushq %r11 /* pt_regs->r11 */ 399 xorl %r11d, %r11d /* nospec r11 */ 400 pushq %rbx /* pt_regs->rbx */ 401 xorl %ebx, %ebx /* nospec rbx */ 402 pushq %rbp /* pt_regs->rbp */ 403 xorl %ebp, %ebp /* nospec rbp */ 404 pushq %r12 /* pt_regs->r12 */ 405 xorl %r12d, %r12d /* nospec r12 */ 406 pushq %r13 /* pt_regs->r13 */ 407 xorl %r13d, %r13d /* nospec r13 */ 408 pushq %r14 /* pt_regs->r14 */ 409 xorl %r14d, %r14d /* nospec r14 */ 410 pushq %r15 /* pt_regs->r15 */ 411 xorl %r15d, %r15d /* nospec r15 */ 412 413 UNWIND_HINT_REGS 414 415 cld 416 417 movq %rsp, %rdi 418 call do_int80_syscall_32 419 jmp swapgs_restore_regs_and_return_to_usermode 420SYM_CODE_END(entry_INT80_compat) 421