xref: /openbmc/linux/arch/x86/kernel/doublefault_32.c (revision bef7a78d)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/mm.h>
3 #include <linux/sched.h>
4 #include <linux/sched/debug.h>
5 #include <linux/init_task.h>
6 #include <linux/fs.h>
7 
8 #include <linux/uaccess.h>
9 #include <asm/processor.h>
10 #include <asm/desc.h>
11 #include <asm/traps.h>
12 
13 #define ptr_ok(x) ((x) > PAGE_OFFSET && (x) < PAGE_OFFSET + MAXMEM)
14 
15 #define TSS(x) this_cpu_read(cpu_tss_rw.x86_tss.x)
16 
17 static void set_df_gdt_entry(unsigned int cpu);
18 
19 /*
20  * Called by double_fault with CR0.TS and EFLAGS.NT cleared.  The CPU thinks
21  * we're running the doublefault task.  Cannot return.
22  */
23 asmlinkage noinstr void __noreturn doublefault_shim(void)
24 {
25 	unsigned long cr2;
26 	struct pt_regs regs;
27 
28 	BUILD_BUG_ON(sizeof(struct doublefault_stack) != PAGE_SIZE);
29 
30 	cr2 = native_read_cr2();
31 
32 	/* Reset back to the normal kernel task. */
33 	force_reload_TR();
34 	set_df_gdt_entry(smp_processor_id());
35 
36 	trace_hardirqs_off();
37 
38 	/*
39 	 * Fill in pt_regs.  A downside of doing this in C is that the unwinder
40 	 * won't see it (no ENCODE_FRAME_POINTER), so a nested stack dump
41 	 * won't successfully unwind to the source of the double fault.
42 	 * The main dump from exc_double_fault() is fine, though, since it
43 	 * uses these regs directly.
44 	 *
45 	 * If anyone ever cares, this could be moved to asm.
46 	 */
47 	regs.ss		= TSS(ss);
48 	regs.__ssh	= 0;
49 	regs.sp		= TSS(sp);
50 	regs.flags	= TSS(flags);
51 	regs.cs		= TSS(cs);
52 	/* We won't go through the entry asm, so we can leave __csh as 0. */
53 	regs.__csh	= 0;
54 	regs.ip		= TSS(ip);
55 	regs.orig_ax	= 0;
56 	regs.gs		= TSS(gs);
57 	regs.__gsh	= 0;
58 	regs.fs		= TSS(fs);
59 	regs.__fsh	= 0;
60 	regs.es		= TSS(es);
61 	regs.__esh	= 0;
62 	regs.ds		= TSS(ds);
63 	regs.__dsh	= 0;
64 	regs.ax		= TSS(ax);
65 	regs.bp		= TSS(bp);
66 	regs.di		= TSS(di);
67 	regs.si		= TSS(si);
68 	regs.dx		= TSS(dx);
69 	regs.cx		= TSS(cx);
70 	regs.bx		= TSS(bx);
71 
72 	exc_double_fault(&regs, 0, cr2);
73 
74 	/*
75 	 * x86_32 does not save the original CR3 anywhere on a task switch.
76 	 * This means that, even if we wanted to return, we would need to find
77 	 * some way to reconstruct CR3.  We could make a credible guess based
78 	 * on cpu_tlbstate, but that would be racy and would not account for
79 	 * PTI.
80 	 *
81 	 * Instead, don't bother.  We can return through
82 	 * rewind_stack_do_exit() instead.
83 	 */
84 	panic("cannot return from double fault\n");
85 }
86 
87 DEFINE_PER_CPU_PAGE_ALIGNED(struct doublefault_stack, doublefault_stack) = {
88 	.tss = {
89                 /*
90                  * No sp0 or ss0 -- we never run CPL != 0 with this TSS
91                  * active.  sp is filled in later.
92                  */
93 		.ldt		= 0,
94 	.io_bitmap_base	= IO_BITMAP_OFFSET_INVALID,
95 
96 		.ip		= (unsigned long) asm_exc_double_fault,
97 		.flags		= X86_EFLAGS_FIXED,
98 		.es		= __USER_DS,
99 		.cs		= __KERNEL_CS,
100 		.ss		= __KERNEL_DS,
101 		.ds		= __USER_DS,
102 		.fs		= __KERNEL_PERCPU,
103 #ifndef CONFIG_X86_32_LAZY_GS
104 		.gs		= __KERNEL_STACK_CANARY,
105 #endif
106 
107 		.__cr3		= __pa_nodebug(swapper_pg_dir),
108 	},
109 };
110 
111 static void set_df_gdt_entry(unsigned int cpu)
112 {
113 	/* Set up doublefault TSS pointer in the GDT */
114 	__set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS,
115 		       &get_cpu_entry_area(cpu)->doublefault_stack.tss);
116 
117 }
118 
119 void doublefault_init_cpu_tss(void)
120 {
121 	unsigned int cpu = smp_processor_id();
122 	struct cpu_entry_area *cea = get_cpu_entry_area(cpu);
123 
124 	/*
125 	 * The linker isn't smart enough to initialize percpu variables that
126 	 * point to other places in percpu space.
127 	 */
128         this_cpu_write(doublefault_stack.tss.sp,
129                        (unsigned long)&cea->doublefault_stack.stack +
130                        sizeof(doublefault_stack.stack));
131 
132 	set_df_gdt_entry(cpu);
133 }
134