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 */ 49ENTRY(entry_SYSENTER_compat) 50 /* Interrupts are off on entry. */ 51 SWAPGS 52 53 /* We are about to clobber %rsp anyway, clobbering here is OK */ 54 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp 55 56 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 57 58 /* 59 * User tracing code (ptrace or signal handlers) might assume that 60 * the saved RAX contains a 32-bit number when we're invoking a 32-bit 61 * syscall. Just in case the high bits are nonzero, zero-extend 62 * the syscall number. (This could almost certainly be deleted 63 * with no ill effects.) 64 */ 65 movl %eax, %eax 66 67 /* Construct struct pt_regs on stack */ 68 pushq $__USER32_DS /* pt_regs->ss */ 69 pushq %rbp /* pt_regs->sp (stashed in bp) */ 70 71 /* 72 * Push flags. This is nasty. First, interrupts are currently 73 * off, but we need pt_regs->flags to have IF set. Second, even 74 * if TF was set when SYSENTER started, it's clear by now. We fix 75 * that later using TIF_SINGLESTEP. 76 */ 77 pushfq /* pt_regs->flags (except IF = 0) */ 78 orl $X86_EFLAGS_IF, (%rsp) /* Fix saved flags */ 79 pushq $__USER32_CS /* pt_regs->cs */ 80 pushq $0 /* pt_regs->ip = 0 (placeholder) */ 81 pushq %rax /* pt_regs->orig_ax */ 82 pushq %rdi /* pt_regs->di */ 83 pushq %rsi /* pt_regs->si */ 84 pushq %rdx /* pt_regs->dx */ 85 pushq %rcx /* pt_regs->cx */ 86 pushq $-ENOSYS /* pt_regs->ax */ 87 pushq $0 /* pt_regs->r8 = 0 */ 88 xorl %r8d, %r8d /* nospec r8 */ 89 pushq $0 /* pt_regs->r9 = 0 */ 90 xorl %r9d, %r9d /* nospec r9 */ 91 pushq $0 /* pt_regs->r10 = 0 */ 92 xorl %r10d, %r10d /* nospec r10 */ 93 pushq $0 /* pt_regs->r11 = 0 */ 94 xorl %r11d, %r11d /* nospec r11 */ 95 pushq %rbx /* pt_regs->rbx */ 96 xorl %ebx, %ebx /* nospec rbx */ 97 pushq %rbp /* pt_regs->rbp (will be overwritten) */ 98 xorl %ebp, %ebp /* nospec rbp */ 99 pushq $0 /* pt_regs->r12 = 0 */ 100 xorl %r12d, %r12d /* nospec r12 */ 101 pushq $0 /* pt_regs->r13 = 0 */ 102 xorl %r13d, %r13d /* nospec r13 */ 103 pushq $0 /* pt_regs->r14 = 0 */ 104 xorl %r14d, %r14d /* nospec r14 */ 105 pushq $0 /* pt_regs->r15 = 0 */ 106 xorl %r15d, %r15d /* nospec r15 */ 107 cld 108 109 /* 110 * SYSENTER doesn't filter flags, so we need to clear NT and AC 111 * ourselves. To save a few cycles, we can check whether 112 * either was set instead of doing an unconditional popfq. 113 * This needs to happen before enabling interrupts so that 114 * we don't get preempted with NT set. 115 * 116 * If TF is set, we will single-step all the way to here -- do_debug 117 * will ignore all the traps. (Yes, this is slow, but so is 118 * single-stepping in general. This allows us to avoid having 119 * a more complicated code to handle the case where a user program 120 * forces us to single-step through the SYSENTER entry code.) 121 * 122 * NB.: .Lsysenter_fix_flags is a label with the code under it moved 123 * out-of-line as an optimization: NT is unlikely to be set in the 124 * majority of the cases and instead of polluting the I$ unnecessarily, 125 * we're keeping that code behind a branch which will predict as 126 * not-taken and therefore its instructions won't be fetched. 127 */ 128 testl $X86_EFLAGS_NT|X86_EFLAGS_AC|X86_EFLAGS_TF, EFLAGS(%rsp) 129 jnz .Lsysenter_fix_flags 130.Lsysenter_flags_fixed: 131 132 /* 133 * User mode is traced as though IRQs are on, and SYSENTER 134 * turned them off. 135 */ 136 TRACE_IRQS_OFF 137 138 movq %rsp, %rdi 139 call do_fast_syscall_32 140 /* XEN PV guests always use IRET path */ 141 ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \ 142 "jmp .Lsyscall_32_done", 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 149GLOBAL(__end_entry_SYSENTER_compat) 150ENDPROC(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 */ 199ENTRY(entry_SYSCALL_compat) 200 /* Interrupts are off on entry. */ 201 swapgs 202 203 /* Stash user ESP */ 204 movl %esp, %r8d 205 206 /* Use %rsp as scratch reg. User ESP is stashed in r8 */ 207 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp 208 209 /* Switch to the kernel stack */ 210 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 211 212 /* Construct struct pt_regs on stack */ 213 pushq $__USER32_DS /* pt_regs->ss */ 214 pushq %r8 /* pt_regs->sp */ 215 pushq %r11 /* pt_regs->flags */ 216 pushq $__USER32_CS /* pt_regs->cs */ 217 pushq %rcx /* pt_regs->ip */ 218GLOBAL(entry_SYSCALL_compat_after_hwframe) 219 movl %eax, %eax /* discard orig_ax high bits */ 220 pushq %rax /* pt_regs->orig_ax */ 221 pushq %rdi /* pt_regs->di */ 222 pushq %rsi /* pt_regs->si */ 223 pushq %rdx /* pt_regs->dx */ 224 pushq %rbp /* pt_regs->cx (stashed in bp) */ 225 pushq $-ENOSYS /* pt_regs->ax */ 226 pushq $0 /* pt_regs->r8 = 0 */ 227 xorl %r8d, %r8d /* nospec r8 */ 228 pushq $0 /* pt_regs->r9 = 0 */ 229 xorl %r9d, %r9d /* nospec r9 */ 230 pushq $0 /* pt_regs->r10 = 0 */ 231 xorl %r10d, %r10d /* nospec r10 */ 232 pushq $0 /* pt_regs->r11 = 0 */ 233 xorl %r11d, %r11d /* nospec r11 */ 234 pushq %rbx /* pt_regs->rbx */ 235 xorl %ebx, %ebx /* nospec rbx */ 236 pushq %rbp /* pt_regs->rbp (will be overwritten) */ 237 xorl %ebp, %ebp /* nospec rbp */ 238 pushq $0 /* pt_regs->r12 = 0 */ 239 xorl %r12d, %r12d /* nospec r12 */ 240 pushq $0 /* pt_regs->r13 = 0 */ 241 xorl %r13d, %r13d /* nospec r13 */ 242 pushq $0 /* pt_regs->r14 = 0 */ 243 xorl %r14d, %r14d /* nospec r14 */ 244 pushq $0 /* pt_regs->r15 = 0 */ 245 xorl %r15d, %r15d /* nospec r15 */ 246 247 /* 248 * User mode is traced as though IRQs are on, and SYSENTER 249 * turned them off. 250 */ 251 TRACE_IRQS_OFF 252 253 movq %rsp, %rdi 254 call do_fast_syscall_32 255 /* XEN PV guests always use IRET path */ 256 ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \ 257 "jmp .Lsyscall_32_done", X86_FEATURE_XENPV 258 259 /* Opportunistic SYSRET */ 260sysret32_from_system_call: 261 TRACE_IRQS_ON /* User mode traces as IRQs on. */ 262 movq RBX(%rsp), %rbx /* pt_regs->rbx */ 263 movq RBP(%rsp), %rbp /* pt_regs->rbp */ 264 movq EFLAGS(%rsp), %r11 /* pt_regs->flags (in r11) */ 265 movq RIP(%rsp), %rcx /* pt_regs->ip (in rcx) */ 266 addq $RAX, %rsp /* Skip r8-r15 */ 267 popq %rax /* pt_regs->rax */ 268 popq %rdx /* Skip pt_regs->cx */ 269 popq %rdx /* pt_regs->dx */ 270 popq %rsi /* pt_regs->si */ 271 popq %rdi /* pt_regs->di */ 272 273 /* 274 * USERGS_SYSRET32 does: 275 * GSBASE = user's GS base 276 * EIP = ECX 277 * RFLAGS = R11 278 * CS = __USER32_CS 279 * SS = __USER_DS 280 * 281 * ECX will not match pt_regs->cx, but we're returning to a vDSO 282 * trampoline that will fix up RCX, so this is okay. 283 * 284 * R12-R15 are callee-saved, so they contain whatever was in them 285 * when the system call started, which is already known to user 286 * code. We zero R8-R10 to avoid info leaks. 287 */ 288 movq RSP-ORIG_RAX(%rsp), %rsp 289 290 /* 291 * The original userspace %rsp (RSP-ORIG_RAX(%rsp)) is stored 292 * on the process stack which is not mapped to userspace and 293 * not readable after we SWITCH_TO_USER_CR3. Delay the CR3 294 * switch until after after the last reference to the process 295 * stack. 296 * 297 * %r8/%r9 are zeroed before the sysret, thus safe to clobber. 298 */ 299 SWITCH_TO_USER_CR3_NOSTACK scratch_reg=%r8 scratch_reg2=%r9 300 301 xorl %r8d, %r8d 302 xorl %r9d, %r9d 303 xorl %r10d, %r10d 304 swapgs 305 sysretl 306END(entry_SYSCALL_compat) 307 308/* 309 * 32-bit legacy system call entry. 310 * 311 * 32-bit x86 Linux system calls traditionally used the INT $0x80 312 * instruction. INT $0x80 lands here. 313 * 314 * This entry point can be used by 32-bit and 64-bit programs to perform 315 * 32-bit system calls. Instances of INT $0x80 can be found inline in 316 * various programs and libraries. It is also used by the vDSO's 317 * __kernel_vsyscall fallback for hardware that doesn't support a faster 318 * entry method. Restarted 32-bit system calls also fall back to INT 319 * $0x80 regardless of what instruction was originally used to do the 320 * system call. 321 * 322 * This is considered a slow path. It is not used by most libc 323 * implementations on modern hardware except during process startup. 324 * 325 * Arguments: 326 * eax system call number 327 * ebx arg1 328 * ecx arg2 329 * edx arg3 330 * esi arg4 331 * edi arg5 332 * ebp arg6 333 */ 334ENTRY(entry_INT80_compat) 335 /* 336 * Interrupts are off on entry. 337 */ 338 ASM_CLAC /* Do this early to minimize exposure */ 339 SWAPGS 340 341 /* 342 * User tracing code (ptrace or signal handlers) might assume that 343 * the saved RAX contains a 32-bit number when we're invoking a 32-bit 344 * syscall. Just in case the high bits are nonzero, zero-extend 345 * the syscall number. (This could almost certainly be deleted 346 * with no ill effects.) 347 */ 348 movl %eax, %eax 349 350 /* switch to thread stack expects orig_ax and rdi to be pushed */ 351 pushq %rax /* pt_regs->orig_ax */ 352 pushq %rdi /* pt_regs->di */ 353 354 /* Need to switch before accessing the thread stack. */ 355 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi 356 movq %rsp, %rdi 357 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 358 359 pushq 6*8(%rdi) /* regs->ss */ 360 pushq 5*8(%rdi) /* regs->rsp */ 361 pushq 4*8(%rdi) /* regs->eflags */ 362 pushq 3*8(%rdi) /* regs->cs */ 363 pushq 2*8(%rdi) /* regs->ip */ 364 pushq 1*8(%rdi) /* regs->orig_ax */ 365 366 movq (%rdi), %rdi /* restore %rdi */ 367 368 pushq %rdi /* pt_regs->di */ 369 pushq %rsi /* pt_regs->si */ 370 pushq %rdx /* pt_regs->dx */ 371 pushq %rcx /* pt_regs->cx */ 372 pushq $-ENOSYS /* pt_regs->ax */ 373 pushq $0 /* pt_regs->r8 = 0 */ 374 xorl %r8d, %r8d /* nospec r8 */ 375 pushq $0 /* pt_regs->r9 = 0 */ 376 xorl %r9d, %r9d /* nospec r9 */ 377 pushq $0 /* pt_regs->r10 = 0 */ 378 xorl %r10d, %r10d /* nospec r10 */ 379 pushq $0 /* pt_regs->r11 = 0 */ 380 xorl %r11d, %r11d /* nospec r11 */ 381 pushq %rbx /* pt_regs->rbx */ 382 xorl %ebx, %ebx /* nospec rbx */ 383 pushq %rbp /* pt_regs->rbp */ 384 xorl %ebp, %ebp /* nospec rbp */ 385 pushq %r12 /* pt_regs->r12 */ 386 xorl %r12d, %r12d /* nospec r12 */ 387 pushq %r13 /* pt_regs->r13 */ 388 xorl %r13d, %r13d /* nospec r13 */ 389 pushq %r14 /* pt_regs->r14 */ 390 xorl %r14d, %r14d /* nospec r14 */ 391 pushq %r15 /* pt_regs->r15 */ 392 xorl %r15d, %r15d /* nospec r15 */ 393 cld 394 395 /* 396 * User mode is traced as though IRQs are on, and the interrupt 397 * gate turned them off. 398 */ 399 TRACE_IRQS_OFF 400 401 movq %rsp, %rdi 402 call do_int80_syscall_32 403.Lsyscall_32_done: 404 405 /* Go back to user mode. */ 406 TRACE_IRQS_ON 407 jmp swapgs_restore_regs_and_return_to_usermode 408END(entry_INT80_compat) 409 410ENTRY(stub32_clone) 411 /* 412 * The 32-bit clone ABI is: clone(..., int tls_val, int *child_tidptr). 413 * The 64-bit clone ABI is: clone(..., int *child_tidptr, int tls_val). 414 * 415 * The native 64-bit kernel's sys_clone() implements the latter, 416 * so we need to swap arguments here before calling it: 417 */ 418 xchg %r8, %rcx 419 jmp sys_clone 420ENDPROC(stub32_clone) 421