1/* SPDX-License-Identifier: GPL-2.0 */ 2/* 3 * linux/arch/x86_64/entry.S 4 * 5 * Copyright (C) 1991, 1992 Linus Torvalds 6 * Copyright (C) 2000, 2001, 2002 Andi Kleen SuSE Labs 7 * Copyright (C) 2000 Pavel Machek <pavel@suse.cz> 8 * 9 * entry.S contains the system-call and fault low-level handling routines. 10 * 11 * Some of this is documented in Documentation/arch/x86/entry_64.rst 12 * 13 * A note on terminology: 14 * - iret frame: Architecture defined interrupt frame from SS to RIP 15 * at the top of the kernel process stack. 16 * 17 * Some macro usage: 18 * - SYM_FUNC_START/END:Define functions in the symbol table. 19 * - idtentry: Define exception entry points. 20 */ 21#include <linux/export.h> 22#include <linux/linkage.h> 23#include <asm/segment.h> 24#include <asm/cache.h> 25#include <asm/errno.h> 26#include <asm/asm-offsets.h> 27#include <asm/msr.h> 28#include <asm/unistd.h> 29#include <asm/thread_info.h> 30#include <asm/hw_irq.h> 31#include <asm/page_types.h> 32#include <asm/irqflags.h> 33#include <asm/paravirt.h> 34#include <asm/percpu.h> 35#include <asm/asm.h> 36#include <asm/smap.h> 37#include <asm/pgtable_types.h> 38#include <asm/frame.h> 39#include <asm/trapnr.h> 40#include <asm/nospec-branch.h> 41#include <asm/fsgsbase.h> 42#include <linux/err.h> 43 44#include "calling.h" 45 46.code64 47.section .entry.text, "ax" 48 49/* 50 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers. 51 * 52 * This is the only entry point used for 64-bit system calls. The 53 * hardware interface is reasonably well designed and the register to 54 * argument mapping Linux uses fits well with the registers that are 55 * available when SYSCALL is used. 56 * 57 * SYSCALL instructions can be found inlined in libc implementations as 58 * well as some other programs and libraries. There are also a handful 59 * of SYSCALL instructions in the vDSO used, for example, as a 60 * clock_gettimeofday fallback. 61 * 62 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11, 63 * then loads new ss, cs, and rip from previously programmed MSRs. 64 * rflags gets masked by a value from another MSR (so CLD and CLAC 65 * are not needed). SYSCALL does not save anything on the stack 66 * and does not change rsp. 67 * 68 * Registers on entry: 69 * rax system call number 70 * rcx return address 71 * r11 saved rflags (note: r11 is callee-clobbered register in C ABI) 72 * rdi arg0 73 * rsi arg1 74 * rdx arg2 75 * r10 arg3 (needs to be moved to rcx to conform to C ABI) 76 * r8 arg4 77 * r9 arg5 78 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI) 79 * 80 * Only called from user space. 81 * 82 * When user can change pt_regs->foo always force IRET. That is because 83 * it deals with uncanonical addresses better. SYSRET has trouble 84 * with them due to bugs in both AMD and Intel CPUs. 85 */ 86 87SYM_CODE_START(entry_SYSCALL_64) 88 UNWIND_HINT_ENTRY 89 ENDBR 90 91 swapgs 92 /* tss.sp2 is scratch space. */ 93 movq %rsp, PER_CPU_VAR(cpu_tss_rw + TSS_sp2) 94 SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp 95 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 96 97SYM_INNER_LABEL(entry_SYSCALL_64_safe_stack, SYM_L_GLOBAL) 98 ANNOTATE_NOENDBR 99 100 /* Construct struct pt_regs on stack */ 101 pushq $__USER_DS /* pt_regs->ss */ 102 pushq PER_CPU_VAR(cpu_tss_rw + TSS_sp2) /* pt_regs->sp */ 103 pushq %r11 /* pt_regs->flags */ 104 pushq $__USER_CS /* pt_regs->cs */ 105 pushq %rcx /* pt_regs->ip */ 106SYM_INNER_LABEL(entry_SYSCALL_64_after_hwframe, SYM_L_GLOBAL) 107 pushq %rax /* pt_regs->orig_ax */ 108 109 PUSH_AND_CLEAR_REGS rax=$-ENOSYS 110 111 /* IRQs are off. */ 112 movq %rsp, %rdi 113 /* Sign extend the lower 32bit as syscall numbers are treated as int */ 114 movslq %eax, %rsi 115 116 /* clobbers %rax, make sure it is after saving the syscall nr */ 117 IBRS_ENTER 118 UNTRAIN_RET 119 CLEAR_BRANCH_HISTORY 120 121 call do_syscall_64 /* returns with IRQs disabled */ 122 123 /* 124 * Try to use SYSRET instead of IRET if we're returning to 125 * a completely clean 64-bit userspace context. If we're not, 126 * go to the slow exit path. 127 * In the Xen PV case we must use iret anyway. 128 */ 129 130 ALTERNATIVE "testb %al, %al; jz swapgs_restore_regs_and_return_to_usermode", \ 131 "jmp swapgs_restore_regs_and_return_to_usermode", X86_FEATURE_XENPV 132 133 /* 134 * We win! This label is here just for ease of understanding 135 * perf profiles. Nothing jumps here. 136 */ 137syscall_return_via_sysret: 138 IBRS_EXIT 139 POP_REGS pop_rdi=0 140 141 /* 142 * Now all regs are restored except RSP and RDI. 143 * Save old stack pointer and switch to trampoline stack. 144 */ 145 movq %rsp, %rdi 146 movq PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp 147 UNWIND_HINT_END_OF_STACK 148 149 pushq RSP-RDI(%rdi) /* RSP */ 150 pushq (%rdi) /* RDI */ 151 152 /* 153 * We are on the trampoline stack. All regs except RDI are live. 154 * We can do future final exit work right here. 155 */ 156 STACKLEAK_ERASE_NOCLOBBER 157 158 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 159 160 popq %rdi 161 popq %rsp 162SYM_INNER_LABEL(entry_SYSRETQ_unsafe_stack, SYM_L_GLOBAL) 163 ANNOTATE_NOENDBR 164 swapgs 165 CLEAR_CPU_BUFFERS 166 sysretq 167SYM_INNER_LABEL(entry_SYSRETQ_end, SYM_L_GLOBAL) 168 ANNOTATE_NOENDBR 169 int3 170SYM_CODE_END(entry_SYSCALL_64) 171 172/* 173 * %rdi: prev task 174 * %rsi: next task 175 */ 176.pushsection .text, "ax" 177SYM_FUNC_START(__switch_to_asm) 178 ANNOTATE_NOENDBR 179 /* 180 * Save callee-saved registers 181 * This must match the order in inactive_task_frame 182 */ 183 pushq %rbp 184 pushq %rbx 185 pushq %r12 186 pushq %r13 187 pushq %r14 188 pushq %r15 189 190 /* switch stack */ 191 movq %rsp, TASK_threadsp(%rdi) 192 movq TASK_threadsp(%rsi), %rsp 193 194#ifdef CONFIG_STACKPROTECTOR 195 movq TASK_stack_canary(%rsi), %rbx 196 movq %rbx, PER_CPU_VAR(__stack_chk_guard) 197#endif 198 199 /* 200 * When switching from a shallower to a deeper call stack 201 * the RSB may either underflow or use entries populated 202 * with userspace addresses. On CPUs where those concerns 203 * exist, overwrite the RSB with entries which capture 204 * speculative execution to prevent attack. 205 */ 206 FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW 207 208 /* restore callee-saved registers */ 209 popq %r15 210 popq %r14 211 popq %r13 212 popq %r12 213 popq %rbx 214 popq %rbp 215 216 jmp __switch_to 217SYM_FUNC_END(__switch_to_asm) 218.popsection 219 220/* 221 * A newly forked process directly context switches into this address. 222 * 223 * rax: prev task we switched from 224 * rbx: kernel thread func (NULL for user thread) 225 * r12: kernel thread arg 226 */ 227.pushsection .text, "ax" 228SYM_CODE_START(ret_from_fork_asm) 229 /* 230 * This is the start of the kernel stack; even through there's a 231 * register set at the top, the regset isn't necessarily coherent 232 * (consider kthreads) and one cannot unwind further. 233 * 234 * This ensures stack unwinds of kernel threads terminate in a known 235 * good state. 236 */ 237 UNWIND_HINT_END_OF_STACK 238 ANNOTATE_NOENDBR // copy_thread 239 CALL_DEPTH_ACCOUNT 240 241 movq %rax, %rdi /* prev */ 242 movq %rsp, %rsi /* regs */ 243 movq %rbx, %rdx /* fn */ 244 movq %r12, %rcx /* fn_arg */ 245 call ret_from_fork 246 247 /* 248 * Set the stack state to what is expected for the target function 249 * -- at this point the register set should be a valid user set 250 * and unwind should work normally. 251 */ 252 UNWIND_HINT_REGS 253 254#ifdef CONFIG_X86_FRED 255 ALTERNATIVE "jmp swapgs_restore_regs_and_return_to_usermode", \ 256 "jmp asm_fred_exit_user", X86_FEATURE_FRED 257#else 258 jmp swapgs_restore_regs_and_return_to_usermode 259#endif 260SYM_CODE_END(ret_from_fork_asm) 261.popsection 262 263.macro DEBUG_ENTRY_ASSERT_IRQS_OFF 264#ifdef CONFIG_DEBUG_ENTRY 265 pushq %rax 266 SAVE_FLAGS 267 testl $X86_EFLAGS_IF, %eax 268 jz .Lokay_\@ 269 ud2 270.Lokay_\@: 271 popq %rax 272#endif 273.endm 274 275SYM_CODE_START(xen_error_entry) 276 ANNOTATE_NOENDBR 277 UNWIND_HINT_FUNC 278 PUSH_AND_CLEAR_REGS save_ret=1 279 ENCODE_FRAME_POINTER 8 280 UNTRAIN_RET_FROM_CALL 281 RET 282SYM_CODE_END(xen_error_entry) 283 284/** 285 * idtentry_body - Macro to emit code calling the C function 286 * @cfunc: C function to be called 287 * @has_error_code: Hardware pushed error code on stack 288 */ 289.macro idtentry_body cfunc has_error_code:req 290 291 /* 292 * Call error_entry() and switch to the task stack if from userspace. 293 * 294 * When in XENPV, it is already in the task stack, and it can't fault 295 * for native_iret() nor native_load_gs_index() since XENPV uses its 296 * own pvops for IRET and load_gs_index(). And it doesn't need to 297 * switch the CR3. So it can skip invoking error_entry(). 298 */ 299 ALTERNATIVE "call error_entry; movq %rax, %rsp", \ 300 "call xen_error_entry", X86_FEATURE_XENPV 301 302 ENCODE_FRAME_POINTER 303 UNWIND_HINT_REGS 304 305 movq %rsp, %rdi /* pt_regs pointer into 1st argument*/ 306 307 .if \has_error_code == 1 308 movq ORIG_RAX(%rsp), %rsi /* get error code into 2nd argument*/ 309 movq $-1, ORIG_RAX(%rsp) /* no syscall to restart */ 310 .endif 311 312 /* For some configurations \cfunc ends up being a noreturn. */ 313 ANNOTATE_REACHABLE 314 call \cfunc 315 316 jmp error_return 317.endm 318 319/** 320 * idtentry - Macro to generate entry stubs for simple IDT entries 321 * @vector: Vector number 322 * @asmsym: ASM symbol for the entry point 323 * @cfunc: C function to be called 324 * @has_error_code: Hardware pushed error code on stack 325 * 326 * The macro emits code to set up the kernel context for straight forward 327 * and simple IDT entries. No IST stack, no paranoid entry checks. 328 */ 329.macro idtentry vector asmsym cfunc has_error_code:req 330SYM_CODE_START(\asmsym) 331 332 .if \vector == X86_TRAP_BP 333 /* #BP advances %rip to the next instruction */ 334 UNWIND_HINT_IRET_ENTRY offset=\has_error_code*8 signal=0 335 .else 336 UNWIND_HINT_IRET_ENTRY offset=\has_error_code*8 337 .endif 338 339 ENDBR 340 ASM_CLAC 341 cld 342 343 .if \has_error_code == 0 344 pushq $-1 /* ORIG_RAX: no syscall to restart */ 345 .endif 346 347 .if \vector == X86_TRAP_BP 348 /* 349 * If coming from kernel space, create a 6-word gap to allow the 350 * int3 handler to emulate a call instruction. 351 */ 352 testb $3, CS-ORIG_RAX(%rsp) 353 jnz .Lfrom_usermode_no_gap_\@ 354 .rept 6 355 pushq 5*8(%rsp) 356 .endr 357 UNWIND_HINT_IRET_REGS offset=8 358.Lfrom_usermode_no_gap_\@: 359 .endif 360 361 idtentry_body \cfunc \has_error_code 362 363_ASM_NOKPROBE(\asmsym) 364SYM_CODE_END(\asmsym) 365.endm 366 367/* 368 * Interrupt entry/exit. 369 * 370 + The interrupt stubs push (vector) onto the stack, which is the error_code 371 * position of idtentry exceptions, and jump to one of the two idtentry points 372 * (common/spurious). 373 * 374 * common_interrupt is a hotpath, align it to a cache line 375 */ 376.macro idtentry_irq vector cfunc 377 .p2align CONFIG_X86_L1_CACHE_SHIFT 378 idtentry \vector asm_\cfunc \cfunc has_error_code=1 379.endm 380 381/** 382 * idtentry_mce_db - Macro to generate entry stubs for #MC and #DB 383 * @vector: Vector number 384 * @asmsym: ASM symbol for the entry point 385 * @cfunc: C function to be called 386 * 387 * The macro emits code to set up the kernel context for #MC and #DB 388 * 389 * If the entry comes from user space it uses the normal entry path 390 * including the return to user space work and preemption checks on 391 * exit. 392 * 393 * If hits in kernel mode then it needs to go through the paranoid 394 * entry as the exception can hit any random state. No preemption 395 * check on exit to keep the paranoid path simple. 396 */ 397.macro idtentry_mce_db vector asmsym cfunc 398SYM_CODE_START(\asmsym) 399 UNWIND_HINT_IRET_ENTRY 400 ENDBR 401 ASM_CLAC 402 cld 403 404 pushq $-1 /* ORIG_RAX: no syscall to restart */ 405 406 /* 407 * If the entry is from userspace, switch stacks and treat it as 408 * a normal entry. 409 */ 410 testb $3, CS-ORIG_RAX(%rsp) 411 jnz .Lfrom_usermode_switch_stack_\@ 412 413 /* paranoid_entry returns GS information for paranoid_exit in EBX. */ 414 call paranoid_entry 415 416 UNWIND_HINT_REGS 417 418 movq %rsp, %rdi /* pt_regs pointer */ 419 420 call \cfunc 421 422 jmp paranoid_exit 423 424 /* Switch to the regular task stack and use the noist entry point */ 425.Lfrom_usermode_switch_stack_\@: 426 idtentry_body noist_\cfunc, has_error_code=0 427 428_ASM_NOKPROBE(\asmsym) 429SYM_CODE_END(\asmsym) 430.endm 431 432#ifdef CONFIG_AMD_MEM_ENCRYPT 433/** 434 * idtentry_vc - Macro to generate entry stub for #VC 435 * @vector: Vector number 436 * @asmsym: ASM symbol for the entry point 437 * @cfunc: C function to be called 438 * 439 * The macro emits code to set up the kernel context for #VC. The #VC handler 440 * runs on an IST stack and needs to be able to cause nested #VC exceptions. 441 * 442 * To make this work the #VC entry code tries its best to pretend it doesn't use 443 * an IST stack by switching to the task stack if coming from user-space (which 444 * includes early SYSCALL entry path) or back to the stack in the IRET frame if 445 * entered from kernel-mode. 446 * 447 * If entered from kernel-mode the return stack is validated first, and if it is 448 * not safe to use (e.g. because it points to the entry stack) the #VC handler 449 * will switch to a fall-back stack (VC2) and call a special handler function. 450 * 451 * The macro is only used for one vector, but it is planned to be extended in 452 * the future for the #HV exception. 453 */ 454.macro idtentry_vc vector asmsym cfunc 455SYM_CODE_START(\asmsym) 456 UNWIND_HINT_IRET_ENTRY 457 ENDBR 458 ASM_CLAC 459 cld 460 461 /* 462 * If the entry is from userspace, switch stacks and treat it as 463 * a normal entry. 464 */ 465 testb $3, CS-ORIG_RAX(%rsp) 466 jnz .Lfrom_usermode_switch_stack_\@ 467 468 /* 469 * paranoid_entry returns SWAPGS flag for paranoid_exit in EBX. 470 * EBX == 0 -> SWAPGS, EBX == 1 -> no SWAPGS 471 */ 472 call paranoid_entry 473 474 UNWIND_HINT_REGS 475 476 /* 477 * Switch off the IST stack to make it free for nested exceptions. The 478 * vc_switch_off_ist() function will switch back to the interrupted 479 * stack if it is safe to do so. If not it switches to the VC fall-back 480 * stack. 481 */ 482 movq %rsp, %rdi /* pt_regs pointer */ 483 call vc_switch_off_ist 484 movq %rax, %rsp /* Switch to new stack */ 485 486 ENCODE_FRAME_POINTER 487 UNWIND_HINT_REGS 488 489 /* Update pt_regs */ 490 movq ORIG_RAX(%rsp), %rsi /* get error code into 2nd argument*/ 491 movq $-1, ORIG_RAX(%rsp) /* no syscall to restart */ 492 493 movq %rsp, %rdi /* pt_regs pointer */ 494 495 call kernel_\cfunc 496 497 /* 498 * No need to switch back to the IST stack. The current stack is either 499 * identical to the stack in the IRET frame or the VC fall-back stack, 500 * so it is definitely mapped even with PTI enabled. 501 */ 502 jmp paranoid_exit 503 504 /* Switch to the regular task stack */ 505.Lfrom_usermode_switch_stack_\@: 506 idtentry_body user_\cfunc, has_error_code=1 507 508_ASM_NOKPROBE(\asmsym) 509SYM_CODE_END(\asmsym) 510.endm 511#endif 512 513/* 514 * Double fault entry. Straight paranoid. No checks from which context 515 * this comes because for the espfix induced #DF this would do the wrong 516 * thing. 517 */ 518.macro idtentry_df vector asmsym cfunc 519SYM_CODE_START(\asmsym) 520 UNWIND_HINT_IRET_ENTRY offset=8 521 ENDBR 522 ASM_CLAC 523 cld 524 525 /* paranoid_entry returns GS information for paranoid_exit in EBX. */ 526 call paranoid_entry 527 UNWIND_HINT_REGS 528 529 movq %rsp, %rdi /* pt_regs pointer into first argument */ 530 movq ORIG_RAX(%rsp), %rsi /* get error code into 2nd argument*/ 531 movq $-1, ORIG_RAX(%rsp) /* no syscall to restart */ 532 533 /* For some configurations \cfunc ends up being a noreturn. */ 534 ANNOTATE_REACHABLE 535 call \cfunc 536 537 jmp paranoid_exit 538 539_ASM_NOKPROBE(\asmsym) 540SYM_CODE_END(\asmsym) 541.endm 542 543/* 544 * Include the defines which emit the idt entries which are shared 545 * shared between 32 and 64 bit and emit the __irqentry_text_* markers 546 * so the stacktrace boundary checks work. 547 */ 548 __ALIGN 549 .globl __irqentry_text_start 550__irqentry_text_start: 551 552#include <asm/idtentry.h> 553 554 __ALIGN 555 .globl __irqentry_text_end 556__irqentry_text_end: 557 ANNOTATE_NOENDBR 558 559SYM_CODE_START_LOCAL(common_interrupt_return) 560SYM_INNER_LABEL(swapgs_restore_regs_and_return_to_usermode, SYM_L_GLOBAL) 561 IBRS_EXIT 562#ifdef CONFIG_XEN_PV 563 ALTERNATIVE "", "jmp xenpv_restore_regs_and_return_to_usermode", X86_FEATURE_XENPV 564#endif 565#ifdef CONFIG_MITIGATION_PAGE_TABLE_ISOLATION 566 ALTERNATIVE "", "jmp .Lpti_restore_regs_and_return_to_usermode", X86_FEATURE_PTI 567#endif 568 569 STACKLEAK_ERASE 570 POP_REGS 571 add $8, %rsp /* orig_ax */ 572 UNWIND_HINT_IRET_REGS 573 574.Lswapgs_and_iret: 575 swapgs 576 CLEAR_CPU_BUFFERS 577 /* Assert that the IRET frame indicates user mode. */ 578 testb $3, 8(%rsp) 579 jnz .Lnative_iret 580 ud2 581 582#ifdef CONFIG_MITIGATION_PAGE_TABLE_ISOLATION 583.Lpti_restore_regs_and_return_to_usermode: 584 POP_REGS pop_rdi=0 585 586 /* 587 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS. 588 * Save old stack pointer and switch to trampoline stack. 589 */ 590 movq %rsp, %rdi 591 movq PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp 592 UNWIND_HINT_END_OF_STACK 593 594 /* Copy the IRET frame to the trampoline stack. */ 595 pushq 6*8(%rdi) /* SS */ 596 pushq 5*8(%rdi) /* RSP */ 597 pushq 4*8(%rdi) /* EFLAGS */ 598 pushq 3*8(%rdi) /* CS */ 599 pushq 2*8(%rdi) /* RIP */ 600 601 /* Push user RDI on the trampoline stack. */ 602 pushq (%rdi) 603 604 /* 605 * We are on the trampoline stack. All regs except RDI are live. 606 * We can do future final exit work right here. 607 */ 608 STACKLEAK_ERASE_NOCLOBBER 609 610 push %rax 611 SWITCH_TO_USER_CR3 scratch_reg=%rdi scratch_reg2=%rax 612 pop %rax 613 614 /* Restore RDI. */ 615 popq %rdi 616 jmp .Lswapgs_and_iret 617#endif 618 619SYM_INNER_LABEL(restore_regs_and_return_to_kernel, SYM_L_GLOBAL) 620#ifdef CONFIG_DEBUG_ENTRY 621 /* Assert that pt_regs indicates kernel mode. */ 622 testb $3, CS(%rsp) 623 jz 1f 624 ud2 6251: 626#endif 627 POP_REGS 628 addq $8, %rsp /* skip regs->orig_ax */ 629 /* 630 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization 631 * when returning from IPI handler. 632 */ 633#ifdef CONFIG_XEN_PV 634SYM_INNER_LABEL(early_xen_iret_patch, SYM_L_GLOBAL) 635 ANNOTATE_NOENDBR 636 .byte 0xe9 637 .long .Lnative_iret - (. + 4) 638#endif 639 640.Lnative_iret: 641 UNWIND_HINT_IRET_REGS 642 /* 643 * Are we returning to a stack segment from the LDT? Note: in 644 * 64-bit mode SS:RSP on the exception stack is always valid. 645 */ 646#ifdef CONFIG_X86_ESPFIX64 647 testb $4, (SS-RIP)(%rsp) 648 jnz native_irq_return_ldt 649#endif 650 651SYM_INNER_LABEL(native_irq_return_iret, SYM_L_GLOBAL) 652 ANNOTATE_NOENDBR // exc_double_fault 653 /* 654 * This may fault. Non-paranoid faults on return to userspace are 655 * handled by fixup_bad_iret. These include #SS, #GP, and #NP. 656 * Double-faults due to espfix64 are handled in exc_double_fault. 657 * Other faults here are fatal. 658 */ 659 iretq 660 661#ifdef CONFIG_X86_ESPFIX64 662native_irq_return_ldt: 663 /* 664 * We are running with user GSBASE. All GPRs contain their user 665 * values. We have a percpu ESPFIX stack that is eight slots 666 * long (see ESPFIX_STACK_SIZE). espfix_waddr points to the bottom 667 * of the ESPFIX stack. 668 * 669 * We clobber RAX and RDI in this code. We stash RDI on the 670 * normal stack and RAX on the ESPFIX stack. 671 * 672 * The ESPFIX stack layout we set up looks like this: 673 * 674 * --- top of ESPFIX stack --- 675 * SS 676 * RSP 677 * RFLAGS 678 * CS 679 * RIP <-- RSP points here when we're done 680 * RAX <-- espfix_waddr points here 681 * --- bottom of ESPFIX stack --- 682 */ 683 684 pushq %rdi /* Stash user RDI */ 685 swapgs /* to kernel GS */ 686 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi /* to kernel CR3 */ 687 688 movq PER_CPU_VAR(espfix_waddr), %rdi 689 movq %rax, (0*8)(%rdi) /* user RAX */ 690 movq (1*8)(%rsp), %rax /* user RIP */ 691 movq %rax, (1*8)(%rdi) 692 movq (2*8)(%rsp), %rax /* user CS */ 693 movq %rax, (2*8)(%rdi) 694 movq (3*8)(%rsp), %rax /* user RFLAGS */ 695 movq %rax, (3*8)(%rdi) 696 movq (5*8)(%rsp), %rax /* user SS */ 697 movq %rax, (5*8)(%rdi) 698 movq (4*8)(%rsp), %rax /* user RSP */ 699 movq %rax, (4*8)(%rdi) 700 /* Now RAX == RSP. */ 701 702 andl $0xffff0000, %eax /* RAX = (RSP & 0xffff0000) */ 703 704 /* 705 * espfix_stack[31:16] == 0. The page tables are set up such that 706 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of 707 * espfix_waddr for any X. That is, there are 65536 RO aliases of 708 * the same page. Set up RSP so that RSP[31:16] contains the 709 * respective 16 bits of the /userspace/ RSP and RSP nonetheless 710 * still points to an RO alias of the ESPFIX stack. 711 */ 712 orq PER_CPU_VAR(espfix_stack), %rax 713 714 SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi 715 swapgs /* to user GS */ 716 popq %rdi /* Restore user RDI */ 717 718 movq %rax, %rsp 719 UNWIND_HINT_IRET_REGS offset=8 720 721 /* 722 * At this point, we cannot write to the stack any more, but we can 723 * still read. 724 */ 725 popq %rax /* Restore user RAX */ 726 727 CLEAR_CPU_BUFFERS 728 729 /* 730 * RSP now points to an ordinary IRET frame, except that the page 731 * is read-only and RSP[31:16] are preloaded with the userspace 732 * values. We can now IRET back to userspace. 733 */ 734 jmp native_irq_return_iret 735#endif 736SYM_CODE_END(common_interrupt_return) 737_ASM_NOKPROBE(common_interrupt_return) 738 739/* 740 * Reload gs selector with exception handling 741 * di: new selector 742 * 743 * Is in entry.text as it shouldn't be instrumented. 744 */ 745SYM_FUNC_START(asm_load_gs_index) 746 ANNOTATE_NOENDBR 747 FRAME_BEGIN 748 swapgs 749.Lgs_change: 750 ANNOTATE_NOENDBR // error_entry 751 movl %edi, %gs 7522: ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE 753 swapgs 754 FRAME_END 755 RET 756 757 /* running with kernelgs */ 758.Lbad_gs: 759 swapgs /* switch back to user gs */ 760.macro ZAP_GS 761 /* This can't be a string because the preprocessor needs to see it. */ 762 movl $__USER_DS, %eax 763 movl %eax, %gs 764.endm 765 ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG 766 xorl %eax, %eax 767 movl %eax, %gs 768 jmp 2b 769 770 _ASM_EXTABLE(.Lgs_change, .Lbad_gs) 771 772SYM_FUNC_END(asm_load_gs_index) 773EXPORT_SYMBOL(asm_load_gs_index) 774 775#ifdef CONFIG_XEN_PV 776/* 777 * A note on the "critical region" in our callback handler. 778 * We want to avoid stacking callback handlers due to events occurring 779 * during handling of the last event. To do this, we keep events disabled 780 * until we've done all processing. HOWEVER, we must enable events before 781 * popping the stack frame (can't be done atomically) and so it would still 782 * be possible to get enough handler activations to overflow the stack. 783 * Although unlikely, bugs of that kind are hard to track down, so we'd 784 * like to avoid the possibility. 785 * So, on entry to the handler we detect whether we interrupted an 786 * existing activation in its critical region -- if so, we pop the current 787 * activation and restart the handler using the previous one. 788 * 789 * C calling convention: exc_xen_hypervisor_callback(struct *pt_regs) 790 */ 791 __FUNC_ALIGN 792SYM_CODE_START_LOCAL_NOALIGN(exc_xen_hypervisor_callback) 793 794/* 795 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will 796 * see the correct pointer to the pt_regs 797 */ 798 UNWIND_HINT_FUNC 799 movq %rdi, %rsp /* we don't return, adjust the stack frame */ 800 UNWIND_HINT_REGS 801 802 call xen_pv_evtchn_do_upcall 803 804 jmp error_return 805SYM_CODE_END(exc_xen_hypervisor_callback) 806 807/* 808 * Hypervisor uses this for application faults while it executes. 809 * We get here for two reasons: 810 * 1. Fault while reloading DS, ES, FS or GS 811 * 2. Fault while executing IRET 812 * Category 1 we do not need to fix up as Xen has already reloaded all segment 813 * registers that could be reloaded and zeroed the others. 814 * Category 2 we fix up by killing the current process. We cannot use the 815 * normal Linux return path in this case because if we use the IRET hypercall 816 * to pop the stack frame we end up in an infinite loop of failsafe callbacks. 817 * We distinguish between categories by comparing each saved segment register 818 * with its current contents: any discrepancy means we in category 1. 819 */ 820 __FUNC_ALIGN 821SYM_CODE_START_NOALIGN(xen_failsafe_callback) 822 UNWIND_HINT_UNDEFINED 823 ENDBR 824 movl %ds, %ecx 825 cmpw %cx, 0x10(%rsp) 826 jne 1f 827 movl %es, %ecx 828 cmpw %cx, 0x18(%rsp) 829 jne 1f 830 movl %fs, %ecx 831 cmpw %cx, 0x20(%rsp) 832 jne 1f 833 movl %gs, %ecx 834 cmpw %cx, 0x28(%rsp) 835 jne 1f 836 /* All segments match their saved values => Category 2 (Bad IRET). */ 837 movq (%rsp), %rcx 838 movq 8(%rsp), %r11 839 addq $0x30, %rsp 840 pushq $0 /* RIP */ 841 UNWIND_HINT_IRET_REGS offset=8 842 jmp asm_exc_general_protection 8431: /* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */ 844 movq (%rsp), %rcx 845 movq 8(%rsp), %r11 846 addq $0x30, %rsp 847 UNWIND_HINT_IRET_REGS 848 pushq $-1 /* orig_ax = -1 => not a system call */ 849 PUSH_AND_CLEAR_REGS 850 ENCODE_FRAME_POINTER 851 jmp error_return 852SYM_CODE_END(xen_failsafe_callback) 853#endif /* CONFIG_XEN_PV */ 854 855/* 856 * Save all registers in pt_regs. Return GSBASE related information 857 * in EBX depending on the availability of the FSGSBASE instructions: 858 * 859 * FSGSBASE R/EBX 860 * N 0 -> SWAPGS on exit 861 * 1 -> no SWAPGS on exit 862 * 863 * Y GSBASE value at entry, must be restored in paranoid_exit 864 * 865 * R14 - old CR3 866 * R15 - old SPEC_CTRL 867 */ 868SYM_CODE_START(paranoid_entry) 869 ANNOTATE_NOENDBR 870 UNWIND_HINT_FUNC 871 PUSH_AND_CLEAR_REGS save_ret=1 872 ENCODE_FRAME_POINTER 8 873 874 /* 875 * Always stash CR3 in %r14. This value will be restored, 876 * verbatim, at exit. Needed if paranoid_entry interrupted 877 * another entry that already switched to the user CR3 value 878 * but has not yet returned to userspace. 879 * 880 * This is also why CS (stashed in the "iret frame" by the 881 * hardware at entry) can not be used: this may be a return 882 * to kernel code, but with a user CR3 value. 883 * 884 * Switching CR3 does not depend on kernel GSBASE so it can 885 * be done before switching to the kernel GSBASE. This is 886 * required for FSGSBASE because the kernel GSBASE has to 887 * be retrieved from a kernel internal table. 888 */ 889 SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14 890 891 /* 892 * Handling GSBASE depends on the availability of FSGSBASE. 893 * 894 * Without FSGSBASE the kernel enforces that negative GSBASE 895 * values indicate kernel GSBASE. With FSGSBASE no assumptions 896 * can be made about the GSBASE value when entering from user 897 * space. 898 */ 899 ALTERNATIVE "jmp .Lparanoid_entry_checkgs", "", X86_FEATURE_FSGSBASE 900 901 /* 902 * Read the current GSBASE and store it in %rbx unconditionally, 903 * retrieve and set the current CPUs kernel GSBASE. The stored value 904 * has to be restored in paranoid_exit unconditionally. 905 * 906 * The unconditional write to GS base below ensures that no subsequent 907 * loads based on a mispredicted GS base can happen, therefore no LFENCE 908 * is needed here. 909 */ 910 SAVE_AND_SET_GSBASE scratch_reg=%rax save_reg=%rbx 911 jmp .Lparanoid_gsbase_done 912 913.Lparanoid_entry_checkgs: 914 /* EBX = 1 -> kernel GSBASE active, no restore required */ 915 movl $1, %ebx 916 917 /* 918 * The kernel-enforced convention is a negative GSBASE indicates 919 * a kernel value. No SWAPGS needed on entry and exit. 920 */ 921 movl $MSR_GS_BASE, %ecx 922 rdmsr 923 testl %edx, %edx 924 js .Lparanoid_kernel_gsbase 925 926 /* EBX = 0 -> SWAPGS required on exit */ 927 xorl %ebx, %ebx 928 swapgs 929.Lparanoid_kernel_gsbase: 930 FENCE_SWAPGS_KERNEL_ENTRY 931.Lparanoid_gsbase_done: 932 933 /* 934 * Once we have CR3 and %GS setup save and set SPEC_CTRL. Just like 935 * CR3 above, keep the old value in a callee saved register. 936 */ 937 IBRS_ENTER save_reg=%r15 938 UNTRAIN_RET_FROM_CALL 939 940 RET 941SYM_CODE_END(paranoid_entry) 942 943/* 944 * "Paranoid" exit path from exception stack. This is invoked 945 * only on return from non-NMI IST interrupts that came 946 * from kernel space. 947 * 948 * We may be returning to very strange contexts (e.g. very early 949 * in syscall entry), so checking for preemption here would 950 * be complicated. Fortunately, there's no good reason to try 951 * to handle preemption here. 952 * 953 * R/EBX contains the GSBASE related information depending on the 954 * availability of the FSGSBASE instructions: 955 * 956 * FSGSBASE R/EBX 957 * N 0 -> SWAPGS on exit 958 * 1 -> no SWAPGS on exit 959 * 960 * Y User space GSBASE, must be restored unconditionally 961 * 962 * R14 - old CR3 963 * R15 - old SPEC_CTRL 964 */ 965SYM_CODE_START_LOCAL(paranoid_exit) 966 UNWIND_HINT_REGS 967 968 /* 969 * Must restore IBRS state before both CR3 and %GS since we need access 970 * to the per-CPU x86_spec_ctrl_shadow variable. 971 */ 972 IBRS_EXIT save_reg=%r15 973 974 /* 975 * The order of operations is important. PARANOID_RESTORE_CR3 requires 976 * kernel GSBASE. 977 * 978 * NB to anyone to try to optimize this code: this code does 979 * not execute at all for exceptions from user mode. Those 980 * exceptions go through error_return instead. 981 */ 982 PARANOID_RESTORE_CR3 scratch_reg=%rax save_reg=%r14 983 984 /* Handle the three GSBASE cases */ 985 ALTERNATIVE "jmp .Lparanoid_exit_checkgs", "", X86_FEATURE_FSGSBASE 986 987 /* With FSGSBASE enabled, unconditionally restore GSBASE */ 988 wrgsbase %rbx 989 jmp restore_regs_and_return_to_kernel 990 991.Lparanoid_exit_checkgs: 992 /* On non-FSGSBASE systems, conditionally do SWAPGS */ 993 testl %ebx, %ebx 994 jnz restore_regs_and_return_to_kernel 995 996 /* We are returning to a context with user GSBASE */ 997 swapgs 998 jmp restore_regs_and_return_to_kernel 999SYM_CODE_END(paranoid_exit) 1000 1001/* 1002 * Switch GS and CR3 if needed. 1003 */ 1004SYM_CODE_START(error_entry) 1005 ANNOTATE_NOENDBR 1006 UNWIND_HINT_FUNC 1007 1008 PUSH_AND_CLEAR_REGS save_ret=1 1009 ENCODE_FRAME_POINTER 8 1010 1011 testb $3, CS+8(%rsp) 1012 jz .Lerror_kernelspace 1013 1014 /* 1015 * We entered from user mode or we're pretending to have entered 1016 * from user mode due to an IRET fault. 1017 */ 1018 swapgs 1019 FENCE_SWAPGS_USER_ENTRY 1020 /* We have user CR3. Change to kernel CR3. */ 1021 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1022 IBRS_ENTER 1023 UNTRAIN_RET_FROM_CALL 1024 1025 leaq 8(%rsp), %rdi /* arg0 = pt_regs pointer */ 1026 /* Put us onto the real thread stack. */ 1027 jmp sync_regs 1028 1029 /* 1030 * There are two places in the kernel that can potentially fault with 1031 * usergs. Handle them here. B stepping K8s sometimes report a 1032 * truncated RIP for IRET exceptions returning to compat mode. Check 1033 * for these here too. 1034 */ 1035.Lerror_kernelspace: 1036 leaq native_irq_return_iret(%rip), %rcx 1037 cmpq %rcx, RIP+8(%rsp) 1038 je .Lerror_bad_iret 1039 movl %ecx, %eax /* zero extend */ 1040 cmpq %rax, RIP+8(%rsp) 1041 je .Lbstep_iret 1042 cmpq $.Lgs_change, RIP+8(%rsp) 1043 jne .Lerror_entry_done_lfence 1044 1045 /* 1046 * hack: .Lgs_change can fail with user gsbase. If this happens, fix up 1047 * gsbase and proceed. We'll fix up the exception and land in 1048 * .Lgs_change's error handler with kernel gsbase. 1049 */ 1050 swapgs 1051 1052 /* 1053 * Issue an LFENCE to prevent GS speculation, regardless of whether it is a 1054 * kernel or user gsbase. 1055 */ 1056.Lerror_entry_done_lfence: 1057 FENCE_SWAPGS_KERNEL_ENTRY 1058 CALL_DEPTH_ACCOUNT 1059 leaq 8(%rsp), %rax /* return pt_regs pointer */ 1060 VALIDATE_UNRET_END 1061 RET 1062 1063.Lbstep_iret: 1064 /* Fix truncated RIP */ 1065 movq %rcx, RIP+8(%rsp) 1066 /* fall through */ 1067 1068.Lerror_bad_iret: 1069 /* 1070 * We came from an IRET to user mode, so we have user 1071 * gsbase and CR3. Switch to kernel gsbase and CR3: 1072 */ 1073 swapgs 1074 FENCE_SWAPGS_USER_ENTRY 1075 SWITCH_TO_KERNEL_CR3 scratch_reg=%rax 1076 IBRS_ENTER 1077 UNTRAIN_RET_FROM_CALL 1078 1079 /* 1080 * Pretend that the exception came from user mode: set up pt_regs 1081 * as if we faulted immediately after IRET. 1082 */ 1083 leaq 8(%rsp), %rdi /* arg0 = pt_regs pointer */ 1084 call fixup_bad_iret 1085 mov %rax, %rdi 1086 jmp sync_regs 1087SYM_CODE_END(error_entry) 1088 1089SYM_CODE_START_LOCAL(error_return) 1090 UNWIND_HINT_REGS 1091 DEBUG_ENTRY_ASSERT_IRQS_OFF 1092 testb $3, CS(%rsp) 1093 jz restore_regs_and_return_to_kernel 1094 jmp swapgs_restore_regs_and_return_to_usermode 1095SYM_CODE_END(error_return) 1096 1097/* 1098 * Runs on exception stack. Xen PV does not go through this path at all, 1099 * so we can use real assembly here. 1100 * 1101 * Registers: 1102 * %r14: Used to save/restore the CR3 of the interrupted context 1103 * when MITIGATION_PAGE_TABLE_ISOLATION is in use. Do not clobber. 1104 */ 1105SYM_CODE_START(asm_exc_nmi) 1106 UNWIND_HINT_IRET_ENTRY 1107 ENDBR 1108 1109 /* 1110 * We allow breakpoints in NMIs. If a breakpoint occurs, then 1111 * the iretq it performs will take us out of NMI context. 1112 * This means that we can have nested NMIs where the next 1113 * NMI is using the top of the stack of the previous NMI. We 1114 * can't let it execute because the nested NMI will corrupt the 1115 * stack of the previous NMI. NMI handlers are not re-entrant 1116 * anyway. 1117 * 1118 * To handle this case we do the following: 1119 * Check a special location on the stack that contains a 1120 * variable that is set when NMIs are executing. 1121 * The interrupted task's stack is also checked to see if it 1122 * is an NMI stack. 1123 * If the variable is not set and the stack is not the NMI 1124 * stack then: 1125 * o Set the special variable on the stack 1126 * o Copy the interrupt frame into an "outermost" location on the 1127 * stack 1128 * o Copy the interrupt frame into an "iret" location on the stack 1129 * o Continue processing the NMI 1130 * If the variable is set or the previous stack is the NMI stack: 1131 * o Modify the "iret" location to jump to the repeat_nmi 1132 * o return back to the first NMI 1133 * 1134 * Now on exit of the first NMI, we first clear the stack variable 1135 * The NMI stack will tell any nested NMIs at that point that it is 1136 * nested. Then we pop the stack normally with iret, and if there was 1137 * a nested NMI that updated the copy interrupt stack frame, a 1138 * jump will be made to the repeat_nmi code that will handle the second 1139 * NMI. 1140 * 1141 * However, espfix prevents us from directly returning to userspace 1142 * with a single IRET instruction. Similarly, IRET to user mode 1143 * can fault. We therefore handle NMIs from user space like 1144 * other IST entries. 1145 */ 1146 1147 ASM_CLAC 1148 cld 1149 1150 /* Use %rdx as our temp variable throughout */ 1151 pushq %rdx 1152 1153 testb $3, CS-RIP+8(%rsp) 1154 jz .Lnmi_from_kernel 1155 1156 /* 1157 * NMI from user mode. We need to run on the thread stack, but we 1158 * can't go through the normal entry paths: NMIs are masked, and 1159 * we don't want to enable interrupts, because then we'll end 1160 * up in an awkward situation in which IRQs are on but NMIs 1161 * are off. 1162 * 1163 * We also must not push anything to the stack before switching 1164 * stacks lest we corrupt the "NMI executing" variable. 1165 */ 1166 1167 swapgs 1168 FENCE_SWAPGS_USER_ENTRY 1169 SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx 1170 movq %rsp, %rdx 1171 movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp 1172 UNWIND_HINT_IRET_REGS base=%rdx offset=8 1173 pushq 5*8(%rdx) /* pt_regs->ss */ 1174 pushq 4*8(%rdx) /* pt_regs->rsp */ 1175 pushq 3*8(%rdx) /* pt_regs->flags */ 1176 pushq 2*8(%rdx) /* pt_regs->cs */ 1177 pushq 1*8(%rdx) /* pt_regs->rip */ 1178 UNWIND_HINT_IRET_REGS 1179 pushq $-1 /* pt_regs->orig_ax */ 1180 PUSH_AND_CLEAR_REGS rdx=(%rdx) 1181 ENCODE_FRAME_POINTER 1182 1183 IBRS_ENTER 1184 UNTRAIN_RET 1185 1186 /* 1187 * At this point we no longer need to worry about stack damage 1188 * due to nesting -- we're on the normal thread stack and we're 1189 * done with the NMI stack. 1190 */ 1191 1192 movq %rsp, %rdi 1193 call exc_nmi 1194 1195 /* 1196 * Return back to user mode. We must *not* do the normal exit 1197 * work, because we don't want to enable interrupts. 1198 */ 1199 jmp swapgs_restore_regs_and_return_to_usermode 1200 1201.Lnmi_from_kernel: 1202 /* 1203 * Here's what our stack frame will look like: 1204 * +---------------------------------------------------------+ 1205 * | original SS | 1206 * | original Return RSP | 1207 * | original RFLAGS | 1208 * | original CS | 1209 * | original RIP | 1210 * +---------------------------------------------------------+ 1211 * | temp storage for rdx | 1212 * +---------------------------------------------------------+ 1213 * | "NMI executing" variable | 1214 * +---------------------------------------------------------+ 1215 * | iret SS } Copied from "outermost" frame | 1216 * | iret Return RSP } on each loop iteration; overwritten | 1217 * | iret RFLAGS } by a nested NMI to force another | 1218 * | iret CS } iteration if needed. | 1219 * | iret RIP } | 1220 * +---------------------------------------------------------+ 1221 * | outermost SS } initialized in first_nmi; | 1222 * | outermost Return RSP } will not be changed before | 1223 * | outermost RFLAGS } NMI processing is done. | 1224 * | outermost CS } Copied to "iret" frame on each | 1225 * | outermost RIP } iteration. | 1226 * +---------------------------------------------------------+ 1227 * | pt_regs | 1228 * +---------------------------------------------------------+ 1229 * 1230 * The "original" frame is used by hardware. Before re-enabling 1231 * NMIs, we need to be done with it, and we need to leave enough 1232 * space for the asm code here. 1233 * 1234 * We return by executing IRET while RSP points to the "iret" frame. 1235 * That will either return for real or it will loop back into NMI 1236 * processing. 1237 * 1238 * The "outermost" frame is copied to the "iret" frame on each 1239 * iteration of the loop, so each iteration starts with the "iret" 1240 * frame pointing to the final return target. 1241 */ 1242 1243 /* 1244 * Determine whether we're a nested NMI. 1245 * 1246 * If we interrupted kernel code between repeat_nmi and 1247 * end_repeat_nmi, then we are a nested NMI. We must not 1248 * modify the "iret" frame because it's being written by 1249 * the outer NMI. That's okay; the outer NMI handler is 1250 * about to call exc_nmi() anyway, so we can just resume 1251 * the outer NMI. 1252 */ 1253 1254 movq $repeat_nmi, %rdx 1255 cmpq 8(%rsp), %rdx 1256 ja 1f 1257 movq $end_repeat_nmi, %rdx 1258 cmpq 8(%rsp), %rdx 1259 ja nested_nmi_out 12601: 1261 1262 /* 1263 * Now check "NMI executing". If it's set, then we're nested. 1264 * This will not detect if we interrupted an outer NMI just 1265 * before IRET. 1266 */ 1267 cmpl $1, -8(%rsp) 1268 je nested_nmi 1269 1270 /* 1271 * Now test if the previous stack was an NMI stack. This covers 1272 * the case where we interrupt an outer NMI after it clears 1273 * "NMI executing" but before IRET. We need to be careful, though: 1274 * there is one case in which RSP could point to the NMI stack 1275 * despite there being no NMI active: naughty userspace controls 1276 * RSP at the very beginning of the SYSCALL targets. We can 1277 * pull a fast one on naughty userspace, though: we program 1278 * SYSCALL to mask DF, so userspace cannot cause DF to be set 1279 * if it controls the kernel's RSP. We set DF before we clear 1280 * "NMI executing". 1281 */ 1282 lea 6*8(%rsp), %rdx 1283 /* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */ 1284 cmpq %rdx, 4*8(%rsp) 1285 /* If the stack pointer is above the NMI stack, this is a normal NMI */ 1286 ja first_nmi 1287 1288 subq $EXCEPTION_STKSZ, %rdx 1289 cmpq %rdx, 4*8(%rsp) 1290 /* If it is below the NMI stack, it is a normal NMI */ 1291 jb first_nmi 1292 1293 /* Ah, it is within the NMI stack. */ 1294 1295 testb $(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp) 1296 jz first_nmi /* RSP was user controlled. */ 1297 1298 /* This is a nested NMI. */ 1299 1300nested_nmi: 1301 /* 1302 * Modify the "iret" frame to point to repeat_nmi, forcing another 1303 * iteration of NMI handling. 1304 */ 1305 subq $8, %rsp 1306 leaq -10*8(%rsp), %rdx 1307 pushq $__KERNEL_DS 1308 pushq %rdx 1309 pushfq 1310 pushq $__KERNEL_CS 1311 pushq $repeat_nmi 1312 1313 /* Put stack back */ 1314 addq $(6*8), %rsp 1315 1316nested_nmi_out: 1317 popq %rdx 1318 1319 /* We are returning to kernel mode, so this cannot result in a fault. */ 1320 iretq 1321 1322first_nmi: 1323 /* Restore rdx. */ 1324 movq (%rsp), %rdx 1325 1326 /* Make room for "NMI executing". */ 1327 pushq $0 1328 1329 /* Leave room for the "iret" frame */ 1330 subq $(5*8), %rsp 1331 1332 /* Copy the "original" frame to the "outermost" frame */ 1333 .rept 5 1334 pushq 11*8(%rsp) 1335 .endr 1336 UNWIND_HINT_IRET_REGS 1337 1338 /* Everything up to here is safe from nested NMIs */ 1339 1340#ifdef CONFIG_DEBUG_ENTRY 1341 /* 1342 * For ease of testing, unmask NMIs right away. Disabled by 1343 * default because IRET is very expensive. 1344 */ 1345 pushq $0 /* SS */ 1346 pushq %rsp /* RSP (minus 8 because of the previous push) */ 1347 addq $8, (%rsp) /* Fix up RSP */ 1348 pushfq /* RFLAGS */ 1349 pushq $__KERNEL_CS /* CS */ 1350 pushq $1f /* RIP */ 1351 iretq /* continues at repeat_nmi below */ 1352 UNWIND_HINT_IRET_REGS 13531: 1354#endif 1355 1356repeat_nmi: 1357 ANNOTATE_NOENDBR // this code 1358 /* 1359 * If there was a nested NMI, the first NMI's iret will return 1360 * here. But NMIs are still enabled and we can take another 1361 * nested NMI. The nested NMI checks the interrupted RIP to see 1362 * if it is between repeat_nmi and end_repeat_nmi, and if so 1363 * it will just return, as we are about to repeat an NMI anyway. 1364 * This makes it safe to copy to the stack frame that a nested 1365 * NMI will update. 1366 * 1367 * RSP is pointing to "outermost RIP". gsbase is unknown, but, if 1368 * we're repeating an NMI, gsbase has the same value that it had on 1369 * the first iteration. paranoid_entry will load the kernel 1370 * gsbase if needed before we call exc_nmi(). "NMI executing" 1371 * is zero. 1372 */ 1373 movq $1, 10*8(%rsp) /* Set "NMI executing". */ 1374 1375 /* 1376 * Copy the "outermost" frame to the "iret" frame. NMIs that nest 1377 * here must not modify the "iret" frame while we're writing to 1378 * it or it will end up containing garbage. 1379 */ 1380 addq $(10*8), %rsp 1381 .rept 5 1382 pushq -6*8(%rsp) 1383 .endr 1384 subq $(5*8), %rsp 1385end_repeat_nmi: 1386 ANNOTATE_NOENDBR // this code 1387 1388 /* 1389 * Everything below this point can be preempted by a nested NMI. 1390 * If this happens, then the inner NMI will change the "iret" 1391 * frame to point back to repeat_nmi. 1392 */ 1393 pushq $-1 /* ORIG_RAX: no syscall to restart */ 1394 1395 /* 1396 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit 1397 * as we should not be calling schedule in NMI context. 1398 * Even with normal interrupts enabled. An NMI should not be 1399 * setting NEED_RESCHED or anything that normal interrupts and 1400 * exceptions might do. 1401 */ 1402 call paranoid_entry 1403 UNWIND_HINT_REGS 1404 1405 movq %rsp, %rdi 1406 call exc_nmi 1407 1408 /* Always restore stashed SPEC_CTRL value (see paranoid_entry) */ 1409 IBRS_EXIT save_reg=%r15 1410 1411 PARANOID_RESTORE_CR3 scratch_reg=%r15 save_reg=%r14 1412 1413 /* 1414 * The above invocation of paranoid_entry stored the GSBASE 1415 * related information in R/EBX depending on the availability 1416 * of FSGSBASE. 1417 * 1418 * If FSGSBASE is enabled, restore the saved GSBASE value 1419 * unconditionally, otherwise take the conditional SWAPGS path. 1420 */ 1421 ALTERNATIVE "jmp nmi_no_fsgsbase", "", X86_FEATURE_FSGSBASE 1422 1423 wrgsbase %rbx 1424 jmp nmi_restore 1425 1426nmi_no_fsgsbase: 1427 /* EBX == 0 -> invoke SWAPGS */ 1428 testl %ebx, %ebx 1429 jnz nmi_restore 1430 1431nmi_swapgs: 1432 swapgs 1433 1434nmi_restore: 1435 POP_REGS 1436 1437 /* 1438 * Skip orig_ax and the "outermost" frame to point RSP at the "iret" 1439 * at the "iret" frame. 1440 */ 1441 addq $6*8, %rsp 1442 1443 /* 1444 * Clear "NMI executing". Set DF first so that we can easily 1445 * distinguish the remaining code between here and IRET from 1446 * the SYSCALL entry and exit paths. 1447 * 1448 * We arguably should just inspect RIP instead, but I (Andy) wrote 1449 * this code when I had the misapprehension that Xen PV supported 1450 * NMIs, and Xen PV would break that approach. 1451 */ 1452 std 1453 movq $0, 5*8(%rsp) /* clear "NMI executing" */ 1454 1455 /* 1456 * Skip CLEAR_CPU_BUFFERS here, since it only helps in rare cases like 1457 * NMI in kernel after user state is restored. For an unprivileged user 1458 * these conditions are hard to meet. 1459 */ 1460 1461 /* 1462 * iretq reads the "iret" frame and exits the NMI stack in a 1463 * single instruction. We are returning to kernel mode, so this 1464 * cannot result in a fault. Similarly, we don't need to worry 1465 * about espfix64 on the way back to kernel mode. 1466 */ 1467 iretq 1468SYM_CODE_END(asm_exc_nmi) 1469 1470/* 1471 * This handles SYSCALL from 32-bit code. There is no way to program 1472 * MSRs to fully disable 32-bit SYSCALL. 1473 */ 1474SYM_CODE_START(entry_SYSCALL32_ignore) 1475 UNWIND_HINT_END_OF_STACK 1476 ENDBR 1477 mov $-ENOSYS, %eax 1478 CLEAR_CPU_BUFFERS 1479 sysretl 1480SYM_CODE_END(entry_SYSCALL32_ignore) 1481 1482.pushsection .text, "ax" 1483 __FUNC_ALIGN 1484SYM_CODE_START_NOALIGN(rewind_stack_and_make_dead) 1485 UNWIND_HINT_FUNC 1486 /* Prevent any naive code from trying to unwind to our caller. */ 1487 xorl %ebp, %ebp 1488 1489 movq PER_CPU_VAR(cpu_current_top_of_stack), %rax 1490 leaq -PTREGS_SIZE(%rax), %rsp 1491 UNWIND_HINT_REGS 1492 1493 call make_task_dead 1494SYM_CODE_END(rewind_stack_and_make_dead) 1495.popsection 1496 1497/* 1498 * This sequence executes branches in order to remove user branch information 1499 * from the branch history tracker in the Branch Predictor, therefore removing 1500 * user influence on subsequent BTB lookups. 1501 * 1502 * It should be used on parts prior to Alder Lake. Newer parts should use the 1503 * BHI_DIS_S hardware control instead. If a pre-Alder Lake part is being 1504 * virtualized on newer hardware the VMM should protect against BHI attacks by 1505 * setting BHI_DIS_S for the guests. 1506 * 1507 * CALLs/RETs are necessary to prevent Loop Stream Detector(LSD) from engaging 1508 * and not clearing the branch history. The call tree looks like: 1509 * 1510 * call 1 1511 * call 2 1512 * call 2 1513 * call 2 1514 * call 2 1515 * call 2 1516 * ret 1517 * ret 1518 * ret 1519 * ret 1520 * ret 1521 * ret 1522 * 1523 * This means that the stack is non-constant and ORC can't unwind it with %rsp 1524 * alone. Therefore we unconditionally set up the frame pointer, which allows 1525 * ORC to unwind properly. 1526 * 1527 * The alignment is for performance and not for safety, and may be safely 1528 * refactored in the future if needed. The .skips are for safety, to ensure 1529 * that all RETs are in the second half of a cacheline to mitigate Indirect 1530 * Target Selection, rather than taking the slowpath via its_return_thunk. 1531 */ 1532SYM_FUNC_START(clear_bhb_loop) 1533 ANNOTATE_NOENDBR 1534 push %rbp 1535 mov %rsp, %rbp 1536 movl $5, %ecx 1537 ANNOTATE_INTRA_FUNCTION_CALL 1538 call 1f 1539 jmp 5f 1540 .align 64, 0xcc 1541 /* 1542 * Shift instructions so that the RET is in the upper half of the 1543 * cacheline and don't take the slowpath to its_return_thunk. 1544 */ 1545 .skip 32 - (.Lret1 - 1f), 0xcc 1546 ANNOTATE_INTRA_FUNCTION_CALL 15471: call 2f 1548.Lret1: RET 1549 .align 64, 0xcc 1550 /* 1551 * As above shift instructions for RET at .Lret2 as well. 1552 * 1553 * This should be ideally be: .skip 32 - (.Lret2 - 2f), 0xcc 1554 * but some Clang versions (e.g. 18) don't like this. 1555 */ 1556 .skip 32 - 18, 0xcc 15572: movl $5, %eax 15583: jmp 4f 1559 nop 15604: sub $1, %eax 1561 jnz 3b 1562 sub $1, %ecx 1563 jnz 1b 1564.Lret2: RET 15655: lfence 1566 pop %rbp 1567 RET 1568SYM_FUNC_END(clear_bhb_loop) 1569EXPORT_SYMBOL_GPL(clear_bhb_loop) 1570STACK_FRAME_NON_STANDARD(clear_bhb_loop) 1571