arch/x86/kernel/doublefault_32.c

   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 extern void double_fault(void);
  14 #define ptr_ok(x) ((x) > PAGE_OFFSET && (x) < PAGE_OFFSET + MAXMEM)
  15
  16 #define TSS(x) this_cpu_read(cpu_tss_rw.x86_tss.x)
  17
  18 static void set_df_gdt_entry(unsigned int cpu);
  19
  20 /*
  21  * Called by double_fault with CR0.TS and EFLAGS.NT cleared.  The CPU thinks
  22  * we're running the doublefault task.  Cannot return.
  23  */
  24 asmlinkage notrace void __noreturn doublefault_shim(void)
  25 {
  26         unsigned long cr2;
  27         struct pt_regs regs;
  28
  29         BUILD_BUG_ON(sizeof(struct doublefault_stack) != PAGE_SIZE);
  30
  31         cr2 = native_read_cr2();
  32
  33         /* Reset back to the normal kernel task. */
  34         force_reload_TR();
  35         set_df_gdt_entry(smp_processor_id());
  36
  37         trace_hardirqs_off();
  38
  39         /*
  40          * Fill in pt_regs.  A downside of doing this in C is that the unwinder
  41          * won't see it (no ENCODE_FRAME_POINTER), so a nested stack dump
  42          * won't successfully unwind to the source of the double fault.
  43          * The main dump from do_double_fault() is fine, though, since it
  44          * uses these regs directly.
  45          *
  46          * If anyone ever cares, this could be moved to asm.
  47          */
  48         regs.ss         = TSS(ss);
  49         regs.__ssh      = 0;
  50         regs.sp         = TSS(sp);
  51         regs.flags      = TSS(flags);
  52         regs.cs         = TSS(cs);
  53         /* We won't go through the entry asm, so we can leave __csh as 0. */
  54         regs.__csh      = 0;
  55         regs.ip         = TSS(ip);
  56         regs.orig_ax    = 0;
  57         regs.gs         = TSS(gs);
  58         regs.__gsh      = 0;
  59         regs.fs         = TSS(fs);
  60         regs.__fsh      = 0;
  61         regs.es         = TSS(es);
  62         regs.__esh      = 0;
  63         regs.ds         = TSS(ds);
  64         regs.__dsh      = 0;
  65         regs.ax         = TSS(ax);
  66         regs.bp         = TSS(bp);
  67         regs.di         = TSS(di);
  68         regs.si         = TSS(si);
  69         regs.dx         = TSS(dx);
  70         regs.cx         = TSS(cx);
  71         regs.bx         = TSS(bx);
  72
  73         do_double_fault(&regs, 0, cr2);
  74
  75         /*
  76          * x86_32 does not save the original CR3 anywhere on a task switch.
  77          * This means that, even if we wanted to return, we would need to find
  78          * some way to reconstruct CR3.  We could make a credible guess based
  79          * on cpu_tlbstate, but that would be racy and would not account for
  80          * PTI.
  81          *
  82          * Instead, don't bother.  We can return through
  83          * rewind_stack_do_exit() instead.
  84          */
  85         panic("cannot return from double fault\n");
  86 }
  87 NOKPROBE_SYMBOL(doublefault_shim);
  88
  89 DEFINE_PER_CPU_PAGE_ALIGNED(struct doublefault_stack, doublefault_stack) = {
  90         .tss = {
  91                 /*
  92                  * No sp0 or ss0 -- we never run CPL != 0 with this TSS
  93                  * active.  sp is filled in later.
  94                  */
  95                 .ldt            = 0,
  96         .io_bitmap_base = IO_BITMAP_OFFSET_INVALID,
  97
  98                 .ip             = (unsigned long) double_fault,
  99                 .flags          = X86_EFLAGS_FIXED,
 100                 .es             = __USER_DS,
 101                 .cs             = __KERNEL_CS,
 102                 .ss             = __KERNEL_DS,
 103                 .ds             = __USER_DS,
 104                 .fs             = __KERNEL_PERCPU,
 105 #ifndef CONFIG_X86_32_LAZY_GS
 106                 .gs             = __KERNEL_STACK_CANARY,
 107 #endif
 108
 109                 .__cr3          = __pa_nodebug(swapper_pg_dir),
 110         },
 111 };
 112
 113 static void set_df_gdt_entry(unsigned int cpu)
 114 {
 115         /* Set up doublefault TSS pointer in the GDT */
 116         __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS,
 117                        &get_cpu_entry_area(cpu)->doublefault_stack.tss);
 118
 119 }
 120
 121 void doublefault_init_cpu_tss(void)
 122 {
 123         unsigned int cpu = smp_processor_id();
 124         struct cpu_entry_area *cea = get_cpu_entry_area(cpu);
 125
 126         /*
 127          * The linker isn't smart enough to initialize percpu variables that
 128          * point to other places in percpu space.
 129          */
 130         this_cpu_write(doublefault_stack.tss.sp,
 131                        (unsigned long)&cea->doublefault_stack.stack +
 132                        sizeof(doublefault_stack.stack));
 133
 134         set_df_gdt_entry(cpu);
 135 }