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