1 From 98a0f7841c6cdd3f882548ed3975df7b22cfa482 Mon Sep 17 00:00:00 2001
2 From: Andy Lutomirski <luto@kernel.org>
3 Date: Mon, 4 Dec 2017 15:07:23 +0100
4 Subject: [PATCH 153/241] x86/entry/64: Use a per-CPU trampoline stack for IDT
7 Content-Type: text/plain; charset=UTF-8
8 Content-Transfer-Encoding: 8bit
12 Historically, IDT entries from usermode have always gone directly
13 to the running task's kernel stack. Rearrange it so that we enter on
14 a per-CPU trampoline stack and then manually switch to the task's stack.
15 This touches a couple of extra cachelines, but it gives us a chance
16 to run some code before we touch the kernel stack.
18 The asm isn't exactly beautiful, but I think that fully refactoring
21 Signed-off-by: Andy Lutomirski <luto@kernel.org>
22 Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
23 Reviewed-by: Borislav Petkov <bp@suse.de>
24 Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
25 Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
26 Cc: Borislav Petkov <bp@alien8.de>
27 Cc: Borislav Petkov <bpetkov@suse.de>
28 Cc: Brian Gerst <brgerst@gmail.com>
29 Cc: Dave Hansen <dave.hansen@intel.com>
30 Cc: Dave Hansen <dave.hansen@linux.intel.com>
31 Cc: David Laight <David.Laight@aculab.com>
32 Cc: Denys Vlasenko <dvlasenk@redhat.com>
33 Cc: Eduardo Valentin <eduval@amazon.com>
34 Cc: Greg KH <gregkh@linuxfoundation.org>
35 Cc: H. Peter Anvin <hpa@zytor.com>
36 Cc: Josh Poimboeuf <jpoimboe@redhat.com>
37 Cc: Juergen Gross <jgross@suse.com>
38 Cc: Linus Torvalds <torvalds@linux-foundation.org>
39 Cc: Peter Zijlstra <peterz@infradead.org>
40 Cc: Rik van Riel <riel@redhat.com>
41 Cc: Will Deacon <will.deacon@arm.com>
42 Cc: aliguori@amazon.com
43 Cc: daniel.gruss@iaik.tugraz.at
45 Cc: keescook@google.com
46 Link: https://lkml.kernel.org/r/20171204150606.225330557@linutronix.de
47 Signed-off-by: Ingo Molnar <mingo@kernel.org>
48 (cherry picked from commit 7f2590a110b837af5679d08fc25c6227c5a8c497)
49 Signed-off-by: Andy Whitcroft <apw@canonical.com>
50 Signed-off-by: Kleber Sacilotto de Souza <kleber.souza@canonical.com>
51 (cherry picked from commit bfb2d0ede023853fb8c24d3dae8974cb2f7117c3)
52 Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
54 arch/x86/include/asm/switch_to.h | 4 ++-
55 arch/x86/include/asm/traps.h | 1 -
56 arch/x86/kernel/cpu/common.c | 6 ++--
57 arch/x86/kernel/traps.c | 21 +++++++------
58 arch/x86/entry/entry_64.S | 67 ++++++++++++++++++++++++++++++----------
59 arch/x86/entry/entry_64_compat.S | 5 ++-
60 6 files changed, 72 insertions(+), 32 deletions(-)
62 diff --git a/arch/x86/include/asm/switch_to.h b/arch/x86/include/asm/switch_to.h
63 index 010cd6e4eafc..ca2fc84ad278 100644
64 --- a/arch/x86/include/asm/switch_to.h
65 +++ b/arch/x86/include/asm/switch_to.h
66 @@ -89,10 +89,12 @@ static inline void refresh_sysenter_cs(struct thread_struct *thread)
67 /* This is used when switching tasks or entering/exiting vm86 mode. */
68 static inline void update_sp0(struct task_struct *task)
70 + /* On x86_64, sp0 always points to the entry trampoline stack, which is constant: */
72 load_sp0(task->thread.sp0);
74 - load_sp0(task_top_of_stack(task));
75 + if (static_cpu_has(X86_FEATURE_XENPV))
76 + load_sp0(task_top_of_stack(task));
80 diff --git a/arch/x86/include/asm/traps.h b/arch/x86/include/asm/traps.h
81 index b052a7621ca1..c3b652672d6f 100644
82 --- a/arch/x86/include/asm/traps.h
83 +++ b/arch/x86/include/asm/traps.h
84 @@ -92,7 +92,6 @@ dotraplinkage void do_segment_not_present(struct pt_regs *, long);
85 dotraplinkage void do_stack_segment(struct pt_regs *, long);
87 dotraplinkage void do_double_fault(struct pt_regs *, long);
88 -asmlinkage struct pt_regs *sync_regs(struct pt_regs *);
90 dotraplinkage void do_general_protection(struct pt_regs *, long);
91 dotraplinkage void do_page_fault(struct pt_regs *, unsigned long);
92 diff --git a/arch/x86/kernel/cpu/common.c b/arch/x86/kernel/cpu/common.c
93 index 4a38de4c6ede..404e4b75db6e 100644
94 --- a/arch/x86/kernel/cpu/common.c
95 +++ b/arch/x86/kernel/cpu/common.c
96 @@ -1623,11 +1623,13 @@ void cpu_init(void)
97 setup_cpu_entry_area(cpu);
100 - * Initialize the TSS. Don't bother initializing sp0, as the initial
101 - * task never enters user mode.
102 + * Initialize the TSS. sp0 points to the entry trampoline stack
103 + * regardless of what task is running.
105 set_tss_desc(cpu, &get_cpu_entry_area(cpu)->tss.x86_tss);
107 + load_sp0((unsigned long)&get_cpu_entry_area(cpu)->tss +
108 + offsetofend(struct tss_struct, SYSENTER_stack));
110 load_mm_ldt(&init_mm);
112 diff --git a/arch/x86/kernel/traps.c b/arch/x86/kernel/traps.c
113 index b69db1ee8733..d9debdafe7a6 100644
114 --- a/arch/x86/kernel/traps.c
115 +++ b/arch/x86/kernel/traps.c
116 @@ -631,14 +631,15 @@ NOKPROBE_SYMBOL(do_int3);
120 - * Help handler running on IST stack to switch off the IST stack if the
121 - * interrupted code was in user mode. The actual stack switch is done in
123 + * Help handler running on a per-cpu (IST or entry trampoline) stack
124 + * to switch to the normal thread stack if the interrupted code was in
125 + * user mode. The actual stack switch is done in entry_64.S
127 asmlinkage __visible notrace struct pt_regs *sync_regs(struct pt_regs *eregs)
129 - struct pt_regs *regs = task_pt_regs(current);
131 + struct pt_regs *regs = (struct pt_regs *)this_cpu_read(cpu_current_top_of_stack) - 1;
136 NOKPROBE_SYMBOL(sync_regs);
137 @@ -654,13 +655,13 @@ struct bad_iret_stack *fixup_bad_iret(struct bad_iret_stack *s)
139 * This is called from entry_64.S early in handling a fault
140 * caused by a bad iret to user mode. To handle the fault
141 - * correctly, we want move our stack frame to task_pt_regs
142 - * and we want to pretend that the exception came from the
144 + * correctly, we want to move our stack frame to where it would
145 + * be had we entered directly on the entry stack (rather than
146 + * just below the IRET frame) and we want to pretend that the
147 + * exception came from the IRET target.
149 struct bad_iret_stack *new_stack =
150 - container_of(task_pt_regs(current),
151 - struct bad_iret_stack, regs);
152 + (struct bad_iret_stack *)this_cpu_read(cpu_tss.x86_tss.sp0) - 1;
154 /* Copy the IRET target to the new stack. */
155 memmove(&new_stack->regs.ip, (void *)s->regs.sp, 5*8);
156 diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
157 index 6c73e96daf78..f70fedc58bac 100644
158 --- a/arch/x86/entry/entry_64.S
159 +++ b/arch/x86/entry/entry_64.S
160 @@ -559,6 +559,13 @@ END(irq_entries_start)
161 /* 0(%rsp): ~(interrupt number) */
162 .macro interrupt func
165 + testb $3, CS-ORIG_RAX(%rsp)
168 + call switch_to_thread_stack
171 ALLOC_PT_GPREGS_ON_STACK
174 @@ -568,12 +575,8 @@ END(irq_entries_start)
178 - * IRQ from user mode. Switch to kernel gsbase and inform context
179 - * tracking that we're in kernel mode.
184 + * IRQ from user mode.
186 * We need to tell lockdep that IRQs are off. We can't do this until
187 * we fix gsbase, and we should do it before enter_from_user_mode
188 * (which can take locks). Since TRACE_IRQS_OFF idempotent,
189 @@ -840,6 +843,32 @@ apicinterrupt IRQ_WORK_VECTOR irq_work_interrupt smp_irq_work_interrupt
191 #define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss) + (TSS_ist + ((x) - 1) * 8)
194 + * Switch to the thread stack. This is called with the IRET frame and
195 + * orig_ax on the stack. (That is, RDI..R12 are not on the stack and
196 + * space has not been allocated for them.)
198 +ENTRY(switch_to_thread_stack)
203 + movq PER_CPU_VAR(cpu_current_top_of_stack), %rsp
204 + UNWIND_HINT sp_offset=16 sp_reg=ORC_REG_DI
206 + pushq 7*8(%rdi) /* regs->ss */
207 + pushq 6*8(%rdi) /* regs->rsp */
208 + pushq 5*8(%rdi) /* regs->eflags */
209 + pushq 4*8(%rdi) /* regs->cs */
210 + pushq 3*8(%rdi) /* regs->ip */
211 + pushq 2*8(%rdi) /* regs->orig_ax */
212 + pushq 8(%rdi) /* return address */
217 +END(switch_to_thread_stack)
219 .macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
221 UNWIND_HINT_IRET_REGS offset=\has_error_code*8
222 @@ -857,11 +886,12 @@ ENTRY(\sym)
224 ALLOC_PT_GPREGS_ON_STACK
229 testb $3, CS(%rsp) /* If coming from userspace, switch stacks */
231 + jnz .Lfrom_usermode_switch_stack_\@
238 @@ -903,20 +933,15 @@ ENTRY(\sym)
245 - * Paranoid entry from userspace. Switch stacks and treat it
246 + * Entry from userspace. Switch stacks and treat it
247 * as a normal entry. This means that paranoid handlers
248 * run in real process context if user_mode(regs).
251 +.Lfrom_usermode_switch_stack_\@:
255 - movq %rsp, %rdi /* pt_regs pointer */
257 - movq %rax, %rsp /* switch stack */
259 movq %rsp, %rdi /* pt_regs pointer */
262 @@ -1177,6 +1202,14 @@ ENTRY(error_entry)
265 .Lerror_entry_from_usermode_after_swapgs:
266 + /* Put us onto the real thread stack. */
267 + popq %r12 /* save return addr in %12 */
268 + movq %rsp, %rdi /* arg0 = pt_regs pointer */
270 + movq %rax, %rsp /* switch stack */
271 + ENCODE_FRAME_POINTER
275 * We need to tell lockdep that IRQs are off. We can't do this until
276 * we fix gsbase, and we should do it before enter_from_user_mode
277 diff --git a/arch/x86/entry/entry_64_compat.S b/arch/x86/entry/entry_64_compat.S
278 index 1f76b66518ee..2270601b6218 100644
279 --- a/arch/x86/entry/entry_64_compat.S
280 +++ b/arch/x86/entry/entry_64_compat.S
281 @@ -305,8 +305,11 @@ ENTRY(entry_INT80_compat)
285 - /* Construct struct pt_regs on stack (iret frame is already on stack) */
286 pushq %rax /* pt_regs->orig_ax */
288 + /* switch to thread stack expects orig_ax to be pushed */
289 + call switch_to_thread_stack
291 pushq %rdi /* pt_regs->di */
292 pushq %rsi /* pt_regs->si */
293 pushq %rdx /* pt_regs->dx */