arch/i386/kernel/crash.c

   1 /*
   2  * Architecture specific (i386) functions for kexec based crash dumps.
   3  *
   4  * Created by: Hariprasad Nellitheertha (hari@in.ibm.com)
   5  *
   6  * Copyright (C) IBM Corporation, 2004. All rights reserved.
   7  *
   8  */
   9
  10 #include <linux/init.h>
  11 #include <linux/types.h>
  12 #include <linux/kernel.h>
  13 #include <linux/smp.h>
  14 #include <linux/irq.h>
  15 #include <linux/reboot.h>
  16 #include <linux/kexec.h>
  17 #include <linux/irq.h>
  18 #include <linux/delay.h>
  19 #include <linux/elf.h>
  20 #include <linux/elfcore.h>
  21
  22 #include <asm/processor.h>
  23 #include <asm/hardirq.h>
  24 #include <asm/nmi.h>
  25 #include <asm/hw_irq.h>
  26 #include <asm/apic.h>
  27 #include <mach_ipi.h>
  28
  29
  30 note_buf_t crash_notes[NR_CPUS];
  31 /* This keeps a track of which one is crashing cpu. */
  32 static int crashing_cpu;
  33
  34 static u32 *append_elf_note(u32 *buf, char *name, unsigned type, void *data,
  35                                                                size_t data_len)
  36 {
  37         struct elf_note note;
  38
  39         note.n_namesz = strlen(name) + 1;
  40         note.n_descsz = data_len;
  41         note.n_type   = type;
  42         memcpy(buf, &note, sizeof(note));
  43         buf += (sizeof(note) +3)/4;
  44         memcpy(buf, name, note.n_namesz);
  45         buf += (note.n_namesz + 3)/4;
  46         memcpy(buf, data, note.n_descsz);
  47         buf += (note.n_descsz + 3)/4;
  48
  49         return buf;
  50 }
  51
  52 static void final_note(u32 *buf)
  53 {
  54         struct elf_note note;
  55
  56         note.n_namesz = 0;
  57         note.n_descsz = 0;
  58         note.n_type   = 0;
  59         memcpy(buf, &note, sizeof(note));
  60 }
  61
  62 static void crash_save_this_cpu(struct pt_regs *regs, int cpu)
  63 {
  64         struct elf_prstatus prstatus;
  65         u32 *buf;
  66
  67         if ((cpu < 0) || (cpu >= NR_CPUS))
  68                 return;
  69
  70         /* Using ELF notes here is opportunistic.
  71          * I need a well defined structure format
  72          * for the data I pass, and I need tags
  73          * on the data to indicate what information I have
  74          * squirrelled away.  ELF notes happen to provide
  75          * all of that that no need to invent something new.
  76          */
  77         buf = &crash_notes[cpu][0];
  78         memset(&prstatus, 0, sizeof(prstatus));
  79         prstatus.pr_pid = current->pid;
  80         elf_core_copy_regs(&prstatus.pr_reg, regs);
  81         buf = append_elf_note(buf, "CORE", NT_PRSTATUS, &prstatus,
  82                                 sizeof(prstatus));
  83         final_note(buf);
  84 }
  85
  86 static void crash_get_current_regs(struct pt_regs *regs)
  87 {
  88         __asm__ __volatile__("movl %%ebx,%0" : "=m"(regs->ebx));
  89         __asm__ __volatile__("movl %%ecx,%0" : "=m"(regs->ecx));
  90         __asm__ __volatile__("movl %%edx,%0" : "=m"(regs->edx));
  91         __asm__ __volatile__("movl %%esi,%0" : "=m"(regs->esi));
  92         __asm__ __volatile__("movl %%edi,%0" : "=m"(regs->edi));
  93         __asm__ __volatile__("movl %%ebp,%0" : "=m"(regs->ebp));
  94         __asm__ __volatile__("movl %%eax,%0" : "=m"(regs->eax));
  95         __asm__ __volatile__("movl %%esp,%0" : "=m"(regs->esp));
  96         __asm__ __volatile__("movw %%ss, %%ax;" :"=a"(regs->xss));
  97         __asm__ __volatile__("movw %%cs, %%ax;" :"=a"(regs->xcs));
  98         __asm__ __volatile__("movw %%ds, %%ax;" :"=a"(regs->xds));
  99         __asm__ __volatile__("movw %%es, %%ax;" :"=a"(regs->xes));
 100         __asm__ __volatile__("pushfl; popl %0" :"=m"(regs->eflags));
 101
 102         regs->eip = (unsigned long)current_text_addr();
 103 }
 104
 105 /* CPU does not save ss and esp on stack if execution is already
 106  * running in kernel mode at the time of NMI occurrence. This code
 107  * fixes it.
 108  */
 109 static void crash_setup_regs(struct pt_regs *newregs, struct pt_regs *oldregs)
 110 {
 111         memcpy(newregs, oldregs, sizeof(*newregs));
 112         newregs->esp = (unsigned long)&(oldregs->esp);
 113         __asm__ __volatile__("xorl %eax, %eax;");
 114         __asm__ __volatile__ ("movw %%ss, %%ax;" :"=a"(newregs->xss));
 115 }
 116
 117 /* We may have saved_regs from where the error came from
 118  * or it is NULL if via a direct panic().
 119  */
 120 static void crash_save_self(struct pt_regs *saved_regs)
 121 {
 122         struct pt_regs regs;
 123         int cpu;
 124
 125         cpu = smp_processor_id();
 126         if (saved_regs)
 127                 crash_setup_regs(&regs, saved_regs);
 128         else
 129                 crash_get_current_regs(&regs);
 130         crash_save_this_cpu(&regs, cpu);
 131 }
 132
 133 #ifdef CONFIG_SMP
 134 static atomic_t waiting_for_crash_ipi;
 135
 136 static int crash_nmi_callback(struct pt_regs *regs, int cpu)
 137 {
 138         struct pt_regs fixed_regs;
 139
 140         /* Don't do anything if this handler is invoked on crashing cpu.
 141          * Otherwise, system will completely hang. Crashing cpu can get
 142          * an NMI if system was initially booted with nmi_watchdog parameter.
 143          */
 144         if (cpu == crashing_cpu)
 145                 return 1;
 146         local_irq_disable();
 147
 148         if (!user_mode(regs)) {
 149                 crash_setup_regs(&fixed_regs, regs);
 150                 regs = &fixed_regs;
 151         }
 152         crash_save_this_cpu(regs, cpu);
 153         disable_local_APIC();
 154         atomic_dec(&waiting_for_crash_ipi);
 155         /* Assume hlt works */
 156         __asm__("hlt");
 157         for(;;);
 158
 159         return 1;
 160 }
 161
 162 /*
 163  * By using the NMI code instead of a vector we just sneak thru the
 164  * word generator coming out with just what we want.  AND it does
 165  * not matter if clustered_apic_mode is set or not.
 166  */
 167 static void smp_send_nmi_allbutself(void)
 168 {
 169         send_IPI_allbutself(APIC_DM_NMI);
 170 }
 171
 172 static void nmi_shootdown_cpus(void)
 173 {
 174         unsigned long msecs;
 175
 176         atomic_set(&waiting_for_crash_ipi, num_online_cpus() - 1);
 177         /* Would it be better to replace the trap vector here? */
 178         set_nmi_callback(crash_nmi_callback);
 179         /* Ensure the new callback function is set before sending
 180          * out the NMI
 181          */
 182         wmb();
 183
 184         smp_send_nmi_allbutself();
 185
 186         msecs = 1000; /* Wait at most a second for the other cpus to stop */
 187         while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
 188                 mdelay(1);
 189                 msecs--;
 190         }
 191
 192         /* Leave the nmi callback set */
 193         disable_local_APIC();
 194 }
 195 #else
 196 static void nmi_shootdown_cpus(void)
 197 {
 198         /* There are no cpus to shootdown */
 199 }
 200 #endif
 201
 202 void machine_crash_shutdown(struct pt_regs *regs)
 203 {
 204         /* This function is only called after the system
 205          * has paniced or is otherwise in a critical state.
 206          * The minimum amount of code to allow a kexec'd kernel
 207          * to run successfully needs to happen here.
 208          *
 209          * In practice this means shooting down the other cpus in
 210          * an SMP system.
 211          */
 212         /* The kernel is broken so disable interrupts */
 213         local_irq_disable();
 214
 215         /* Make a note of crashing cpu. Will be used in NMI callback.*/
 216         crashing_cpu = smp_processor_id();
 217         nmi_shootdown_cpus();
 218         lapic_shutdown();
 219 #if defined(CONFIG_X86_IO_APIC)
 220         disable_IO_APIC();
 221 #endif
 222         crash_save_self(regs);
 223 }