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 pushq $0 /* pt_regs->r9 = 0 */ 89 pushq $0 /* pt_regs->r10 = 0 */ 90 pushq $0 /* pt_regs->r11 = 0 */ 91 pushq %rbx /* pt_regs->rbx */ 92 pushq %rbp /* pt_regs->rbp (will be overwritten) */ 93 pushq $0 /* pt_regs->r12 = 0 */ 94 pushq $0 /* pt_regs->r13 = 0 */ 95 pushq $0 /* pt_regs->r14 = 0 */ 96 pushq $0 /* pt_regs->r15 = 0 */ 97 cld 98 99 /* 100 * SYSENTER doesn't filter flags, so we need to clear NT and AC 101 * ourselves. To save a few cycles, we can check whether 102 * either was set instead of doing an unconditional popfq. 103 * This needs to happen before enabling interrupts so that 104 * we don't get preempted with NT set. 105 * 106 * If TF is set, we will single-step all the way to here -- do_debug 107 * will ignore all the traps. (Yes, this is slow, but so is 108 * single-stepping in general. This allows us to avoid having 109 * a more complicated code to handle the case where a user program 110 * forces us to single-step through the SYSENTER entry code.) 111 * 112 * NB.: .Lsysenter_fix_flags is a label with the code under it moved 113 * out-of-line as an optimization: NT is unlikely to be set in the 114 * majority of the cases and instead of polluting the I$ unnecessarily, 115 * we're keeping that code behind a branch which will predict as 116 * not-taken and therefore its instructions won't be fetched. 117 */ 118 testl $X86_EFLAGS_NT|X86_EFLAGS_AC|X86_EFLAGS_TF, EFLAGS(%rsp) 119 jnz .Lsysenter_fix_flags 120.Lsysenter_flags_fixed: 121 122 /* 123 * User mode is traced as though IRQs are on, and SYSENTER 124 * turned them off. 125 */ 126 TRACE_IRQS_OFF 127 128 movq %rsp, %rdi 129 call do_fast_syscall_32 130 /* XEN PV guests always use IRET path */ 131 ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \ 132 "jmp .Lsyscall_32_done", X86_FEATURE_XENPV 133 jmp sysret32_from_system_call 134 135.Lsysenter_fix_flags: 136 pushq $X86_EFLAGS_FIXED 137 popfq 138 jmp .Lsysenter_flags_fixed 139GLOBAL(__end_entry_SYSENTER_compat) 140ENDPROC(entry_SYSENTER_compat) 141 142/* 143 * 32-bit SYSCALL entry. 144 * 145 * 32-bit system calls through the vDSO's __kernel_vsyscall enter here 146 * on 64-bit kernels running on AMD CPUs. 147 * 148 * The SYSCALL instruction, in principle, should *only* occur in the 149 * vDSO. In practice, it appears that this really is the case. 150 * As evidence: 151 * 152 * - The calling convention for SYSCALL has changed several times without 153 * anyone noticing. 154 * 155 * - Prior to the in-kernel X86_BUG_SYSRET_SS_ATTRS fixup, anything 156 * user task that did SYSCALL without immediately reloading SS 157 * would randomly crash. 158 * 159 * - Most programmers do not directly target AMD CPUs, and the 32-bit 160 * SYSCALL instruction does not exist on Intel CPUs. Even on AMD 161 * CPUs, Linux disables the SYSCALL instruction on 32-bit kernels 162 * because the SYSCALL instruction in legacy/native 32-bit mode (as 163 * opposed to compat mode) is sufficiently poorly designed as to be 164 * essentially unusable. 165 * 166 * 32-bit SYSCALL saves RIP to RCX, clears RFLAGS.RF, then saves 167 * RFLAGS to R11, then loads new SS, CS, and RIP from previously 168 * programmed MSRs. RFLAGS gets masked by a value from another MSR 169 * (so CLD and CLAC are not needed). SYSCALL does not save anything on 170 * the stack and does not change RSP. 171 * 172 * Note: RFLAGS saving+masking-with-MSR happens only in Long mode 173 * (in legacy 32-bit mode, IF, RF and VM bits are cleared and that's it). 174 * Don't get confused: RFLAGS saving+masking depends on Long Mode Active bit 175 * (EFER.LMA=1), NOT on bitness of userspace where SYSCALL executes 176 * or target CS descriptor's L bit (SYSCALL does not read segment descriptors). 177 * 178 * Arguments: 179 * eax system call number 180 * ecx return address 181 * ebx arg1 182 * ebp arg2 (note: not saved in the stack frame, should not be touched) 183 * edx arg3 184 * esi arg4 185 * edi arg5 186 * esp user stack 187 * 0(%esp) arg6 188 */ 189ENTRY(entry_SYSCALL_compat) 190 /* Interrupts are off on entry. */ 191 swapgs 192 193 /* Stash user ESP */ 194 movl %esp, %r8d 195 196 /* Use %rsp as scratch reg. User ESP is stashed in r8 */ 197 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp 198 199 /* Switch to the kernel stack */ 200 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 201 202 /* Construct struct pt_regs on stack */ 203 pushq $__USER32_DS /* pt_regs->ss */ 204 pushq %r8 /* pt_regs->sp */ 205 pushq %r11 /* pt_regs->flags */ 206 pushq $__USER32_CS /* pt_regs->cs */ 207 pushq %rcx /* pt_regs->ip */ 208GLOBAL(entry_SYSCALL_compat_after_hwframe) 209 movl %eax, %eax /* discard orig_ax high bits */ 210 pushq %rax /* pt_regs->orig_ax */ 211 pushq %rdi /* pt_regs->di */ 212 pushq %rsi /* pt_regs->si */ 213 pushq %rdx /* pt_regs->dx */ 214 pushq %rbp /* pt_regs->cx (stashed in bp) */ 215 pushq $-ENOSYS /* pt_regs->ax */ 216 pushq $0 /* pt_regs->r8 = 0 */ 217 pushq $0 /* pt_regs->r9 = 0 */ 218 pushq $0 /* pt_regs->r10 = 0 */ 219 pushq $0 /* pt_regs->r11 = 0 */ 220 pushq %rbx /* pt_regs->rbx */ 221 pushq %rbp /* pt_regs->rbp (will be overwritten) */ 222 pushq $0 /* pt_regs->r12 = 0 */ 223 pushq $0 /* pt_regs->r13 = 0 */ 224 pushq $0 /* pt_regs->r14 = 0 */ 225 pushq $0 /* pt_regs->r15 = 0 */ 226 227 /* 228 * User mode is traced as though IRQs are on, and SYSENTER 229 * turned them off. 230 */ 231 TRACE_IRQS_OFF 232 233 movq %rsp, %rdi 234 call do_fast_syscall_32 235 /* XEN PV guests always use IRET path */ 236 ALTERNATIVE "testl %eax, %eax; jz .Lsyscall_32_done", \ 237 "jmp .Lsyscall_32_done", X86_FEATURE_XENPV 238 239 /* Opportunistic SYSRET */ 240sysret32_from_system_call: 241 TRACE_IRQS_ON /* User mode traces as IRQs on. */ 242 movq RBX(%rsp), %rbx /* pt_regs->rbx */ 243 movq RBP(%rsp), %rbp /* pt_regs->rbp */ 244 movq EFLAGS(%rsp), %r11 /* pt_regs->flags (in r11) */ 245 movq RIP(%rsp), %rcx /* pt_regs->ip (in rcx) */ 246 addq $RAX, %rsp /* Skip r8-r15 */ 247 popq %rax /* pt_regs->rax */ 248 popq %rdx /* Skip pt_regs->cx */ 249 popq %rdx /* pt_regs->dx */ 250 popq %rsi /* pt_regs->si */ 251 popq %rdi /* pt_regs->di */ 252 253 /* 254 * USERGS_SYSRET32 does: 255 * GSBASE = user's GS base 256 * EIP = ECX 257 * RFLAGS = R11 258 * CS = __USER32_CS 259 * SS = __USER_DS 260 * 261 * ECX will not match pt_regs->cx, but we're returning to a vDSO 262 * trampoline that will fix up RCX, so this is okay. 263 * 264 * R12-R15 are callee-saved, so they contain whatever was in them 265 * when the system call started, which is already known to user 266 * code. We zero R8-R10 to avoid info leaks. 267 */ 268 movq RSP-ORIG_RAX(%rsp), %rsp 269 270 /* 271 * The original userspace %rsp (RSP-ORIG_RAX(%rsp)) is stored 272 * on the process stack which is not mapped to userspace and 273 * not readable after we SWITCH_TO_USER_CR3. Delay the CR3 274 * switch until after after the last reference to the process 275 * stack. 276 * 277 * %r8/%r9 are zeroed before the sysret, thus safe to clobber. 278 */ 279 SWITCH_TO_USER_CR3_NOSTACK scratch_reg=%r8 scratch_reg2=%r9 280 281 xorq %r8, %r8 282 xorq %r9, %r9 283 xorq %r10, %r10 284 swapgs 285 sysretl 286END(entry_SYSCALL_compat) 287 288/* 289 * 32-bit legacy system call entry. 290 * 291 * 32-bit x86 Linux system calls traditionally used the INT $0x80 292 * instruction. INT $0x80 lands here. 293 * 294 * This entry point can be used by 32-bit and 64-bit programs to perform 295 * 32-bit system calls. Instances of INT $0x80 can be found inline in 296 * various programs and libraries. It is also used by the vDSO's 297 * __kernel_vsyscall fallback for hardware that doesn't support a faster 298 * entry method. Restarted 32-bit system calls also fall back to INT 299 * $0x80 regardless of what instruction was originally used to do the 300 * system call. 301 * 302 * This is considered a slow path. It is not used by most libc 303 * implementations on modern hardware except during process startup. 304 * 305 * Arguments: 306 * eax system call number 307 * ebx arg1 308 * ecx arg2 309 * edx arg3 310 * esi arg4 311 * edi arg5 312 * ebp arg6 313 */ 314ENTRY(entry_INT80_compat) 315 /* 316 * Interrupts are off on entry. 317 */ 318 ASM_CLAC /* Do this early to minimize exposure */ 319 SWAPGS 320 321 /* 322 * User tracing code (ptrace or signal handlers) might assume that 323 * the saved RAX contains a 32-bit number when we're invoking a 32-bit 324 * syscall. Just in case the high bits are nonzero, zero-extend 325 * the syscall number. (This could almost certainly be deleted 326 * with no ill effects.) 327 */ 328 movl %eax, %eax 329 330 pushq %rax /* pt_regs->orig_ax */ 331 332 /* switch to thread stack expects orig_ax to be pushed */ 333 call switch_to_thread_stack 334 335 pushq %rdi /* pt_regs->di */ 336 pushq %rsi /* pt_regs->si */ 337 pushq %rdx /* pt_regs->dx */ 338 pushq %rcx /* pt_regs->cx */ 339 pushq $-ENOSYS /* pt_regs->ax */ 340 pushq $0 /* pt_regs->r8 = 0 */ 341 pushq $0 /* pt_regs->r9 = 0 */ 342 pushq $0 /* pt_regs->r10 = 0 */ 343 pushq $0 /* pt_regs->r11 = 0 */ 344 pushq %rbx /* pt_regs->rbx */ 345 pushq %rbp /* pt_regs->rbp */ 346 pushq %r12 /* pt_regs->r12 */ 347 pushq %r13 /* pt_regs->r13 */ 348 pushq %r14 /* pt_regs->r14 */ 349 pushq %r15 /* pt_regs->r15 */ 350 cld 351 352 /* 353 * User mode is traced as though IRQs are on, and the interrupt 354 * gate turned them off. 355 */ 356 TRACE_IRQS_OFF 357 358 movq %rsp, %rdi 359 call do_int80_syscall_32 360.Lsyscall_32_done: 361 362 /* Go back to user mode. */ 363 TRACE_IRQS_ON 364 jmp swapgs_restore_regs_and_return_to_usermode 365END(entry_INT80_compat) 366 367ENTRY(stub32_clone) 368 /* 369 * The 32-bit clone ABI is: clone(..., int tls_val, int *child_tidptr). 370 * The 64-bit clone ABI is: clone(..., int *child_tidptr, int tls_val). 371 * 372 * The native 64-bit kernel's sys_clone() implements the latter, 373 * so we need to swap arguments here before calling it: 374 */ 375 xchg %r8, %rcx 376 jmp sys_clone 377ENDPROC(stub32_clone) 378