arch/s390/kernel/nmi.c

   1 /*
   2  *   Machine check handler
   3  *
   4  *    Copyright IBM Corp. 2000, 2009
   5  *    Author(s): Ingo Adlung <adlung@de.ibm.com>,
   6  *               Martin Schwidefsky <schwidefsky@de.ibm.com>,
   7  *               Cornelia Huck <cornelia.huck@de.ibm.com>,
   8  *               Heiko Carstens <heiko.carstens@de.ibm.com>,
   9  */
  10
  11 #include <linux/kernel_stat.h>
  12 #include <linux/init.h>
  13 #include <linux/errno.h>
  14 #include <linux/hardirq.h>
  15 #include <linux/time.h>
  16 #include <linux/module.h>
  17 #include <asm/lowcore.h>
  18 #include <asm/smp.h>
  19 #include <asm/etr.h>
  20 #include <asm/cputime.h>
  21 #include <asm/nmi.h>
  22 #include <asm/crw.h>
  23
  24 struct mcck_struct {
  25         int kill_task;
  26         int channel_report;
  27         int warning;
  28         unsigned long long mcck_code;
  29 };
  30
  31 static DEFINE_PER_CPU(struct mcck_struct, cpu_mcck);
  32
  33 static void s390_handle_damage(char *msg)
  34 {
  35         smp_send_stop();
  36         disabled_wait((unsigned long) __builtin_return_address(0));
  37         while (1);
  38 }
  39
  40 /*
  41  * Main machine check handler function. Will be called with interrupts enabled
  42  * or disabled and machine checks enabled or disabled.
  43  */
  44 void s390_handle_mcck(void)
  45 {
  46         unsigned long flags;
  47         struct mcck_struct mcck;
  48
  49         /*
  50          * Disable machine checks and get the current state of accumulated
  51          * machine checks. Afterwards delete the old state and enable machine
  52          * checks again.
  53          */
  54         local_irq_save(flags);
  55         local_mcck_disable();
  56         /*
  57          * Ummm... Does this make sense at all? Copying the percpu struct
  58          * and then zapping it one statement later?
  59          */
  60         memcpy(&mcck, this_cpu_ptr(&cpu_mcck), sizeof(mcck));
  61         memset(&mcck, 0, sizeof(struct mcck_struct));
  62         clear_cpu_flag(CIF_MCCK_PENDING);
  63         local_mcck_enable();
  64         local_irq_restore(flags);
  65
  66         if (mcck.channel_report)
  67                 crw_handle_channel_report();
  68         /*
  69          * A warning may remain for a prolonged period on the bare iron.
  70          * (actually until the machine is powered off, or the problem is gone)
  71          * So we just stop listening for the WARNING MCH and avoid continuously
  72          * being interrupted.  One caveat is however, that we must do this per
  73          * processor and cannot use the smp version of ctl_clear_bit().
  74          * On VM we only get one interrupt per virtally presented machinecheck.
  75          * Though one suffices, we may get one interrupt per (virtual) cpu.
  76          */
  77         if (mcck.warning) {     /* WARNING pending ? */
  78                 static int mchchk_wng_posted = 0;
  79
  80                 /* Use single cpu clear, as we cannot handle smp here. */
  81                 __ctl_clear_bit(14, 24);        /* Disable WARNING MCH */
  82                 if (xchg(&mchchk_wng_posted, 1) == 0)
  83                         kill_cad_pid(SIGPWR, 1);
  84         }
  85         if (mcck.kill_task) {
  86                 local_irq_enable();
  87                 printk(KERN_EMERG "mcck: Terminating task because of machine "
  88                        "malfunction (code 0x%016llx).\n", mcck.mcck_code);
  89                 printk(KERN_EMERG "mcck: task: %s, pid: %d.\n",
  90                        current->comm, current->pid);
  91                 do_exit(SIGSEGV);
  92         }
  93 }
  94 EXPORT_SYMBOL_GPL(s390_handle_mcck);
  95
  96 /*
  97  * returns 0 if all registers could be validated
  98  * returns 1 otherwise
  99  */
 100 static int notrace s390_revalidate_registers(struct mci *mci)
 101 {
 102         int kill_task;
 103         u64 zero;
 104         void *fpt_save_area, *fpt_creg_save_area;
 105
 106         kill_task = 0;
 107         zero = 0;
 108
 109         if (!mci->gr) {
 110                 /*
 111                  * General purpose registers couldn't be restored and have
 112                  * unknown contents. Process needs to be terminated.
 113                  */
 114                 kill_task = 1;
 115         }
 116         if (!mci->fp) {
 117                 /*
 118                  * Floating point registers can't be restored and
 119                  * therefore the process needs to be terminated.
 120                  */
 121                 kill_task = 1;
 122         }
 123 #ifndef CONFIG_64BIT
 124         asm volatile(
 125                 "       ld      0,0(%0)\n"
 126                 "       ld      2,8(%0)\n"
 127                 "       ld      4,16(%0)\n"
 128                 "       ld      6,24(%0)"
 129                 : : "a" (&S390_lowcore.floating_pt_save_area));
 130 #endif
 131
 132         if (MACHINE_HAS_IEEE) {
 133 #ifdef CONFIG_64BIT
 134                 fpt_save_area = &S390_lowcore.floating_pt_save_area;
 135                 fpt_creg_save_area = &S390_lowcore.fpt_creg_save_area;
 136 #else
 137                 fpt_save_area = (void *) S390_lowcore.extended_save_area_addr;
 138                 fpt_creg_save_area = fpt_save_area + 128;
 139 #endif
 140                 if (!mci->fc) {
 141                         /*
 142                          * Floating point control register can't be restored.
 143                          * Task will be terminated.
 144                          */
 145                         asm volatile("lfpc 0(%0)" : : "a" (&zero), "m" (zero));
 146                         kill_task = 1;
 147
 148                 } else
 149                         asm volatile("lfpc 0(%0)" : : "a" (fpt_creg_save_area));
 150
 151                 asm volatile(
 152                         "       ld      0,0(%0)\n"
 153                         "       ld      1,8(%0)\n"
 154                         "       ld      2,16(%0)\n"
 155                         "       ld      3,24(%0)\n"
 156                         "       ld      4,32(%0)\n"
 157                         "       ld      5,40(%0)\n"
 158                         "       ld      6,48(%0)\n"
 159                         "       ld      7,56(%0)\n"
 160                         "       ld      8,64(%0)\n"
 161                         "       ld      9,72(%0)\n"
 162                         "       ld      10,80(%0)\n"
 163                         "       ld      11,88(%0)\n"
 164                         "       ld      12,96(%0)\n"
 165                         "       ld      13,104(%0)\n"
 166                         "       ld      14,112(%0)\n"
 167                         "       ld      15,120(%0)\n"
 168                         : : "a" (fpt_save_area));
 169         }
 170         /* Revalidate access registers */
 171         asm volatile(
 172                 "       lam     0,15,0(%0)"
 173                 : : "a" (&S390_lowcore.access_regs_save_area));
 174         if (!mci->ar) {
 175                 /*
 176                  * Access registers have unknown contents.
 177                  * Terminating task.
 178                  */
 179                 kill_task = 1;
 180         }
 181         /* Revalidate control registers */
 182         if (!mci->cr) {
 183                 /*
 184                  * Control registers have unknown contents.
 185                  * Can't recover and therefore stopping machine.
 186                  */
 187                 s390_handle_damage("invalid control registers.");
 188         } else {
 189 #ifdef CONFIG_64BIT
 190                 asm volatile(
 191                         "       lctlg   0,15,0(%0)"
 192                         : : "a" (&S390_lowcore.cregs_save_area));
 193 #else
 194                 asm volatile(
 195                         "       lctl    0,15,0(%0)"
 196                         : : "a" (&S390_lowcore.cregs_save_area));
 197 #endif
 198         }
 199         /*
 200          * We don't even try to revalidate the TOD register, since we simply
 201          * can't write something sensible into that register.
 202          */
 203 #ifdef CONFIG_64BIT
 204         /*
 205          * See if we can revalidate the TOD programmable register with its
 206          * old contents (should be zero) otherwise set it to zero.
 207          */
 208         if (!mci->pr)
 209                 asm volatile(
 210                         "       sr      0,0\n"
 211                         "       sckpf"
 212                         : : : "0", "cc");
 213         else
 214                 asm volatile(
 215                         "       l       0,0(%0)\n"
 216                         "       sckpf"
 217                         : : "a" (&S390_lowcore.tod_progreg_save_area)
 218                         : "0", "cc");
 219 #endif
 220         /* Revalidate clock comparator register */
 221         set_clock_comparator(S390_lowcore.clock_comparator);
 222         /* Check if old PSW is valid */
 223         if (!mci->wp)
 224                 /*
 225                  * Can't tell if we come from user or kernel mode
 226                  * -> stopping machine.
 227                  */
 228                 s390_handle_damage("old psw invalid.");
 229
 230         if (!mci->ms || !mci->pm || !mci->ia)
 231                 kill_task = 1;
 232
 233         return kill_task;
 234 }
 235
 236 #define MAX_IPD_COUNT   29
 237 #define MAX_IPD_TIME    (5 * 60 * USEC_PER_SEC) /* 5 minutes */
 238
 239 #define ED_STP_ISLAND   6       /* External damage STP island check */
 240 #define ED_STP_SYNC     7       /* External damage STP sync check */
 241 #define ED_ETR_SYNC     12      /* External damage ETR sync check */
 242 #define ED_ETR_SWITCH   13      /* External damage ETR switch to local */
 243
 244 /*
 245  * machine check handler.
 246  */
 247 void notrace s390_do_machine_check(struct pt_regs *regs)
 248 {
 249         static int ipd_count;
 250         static DEFINE_SPINLOCK(ipd_lock);
 251         static unsigned long long last_ipd;
 252         struct mcck_struct *mcck;
 253         unsigned long long tmp;
 254         struct mci *mci;
 255         int umode;
 256
 257         nmi_enter();
 258         inc_irq_stat(NMI_NMI);
 259         mci = (struct mci *) &S390_lowcore.mcck_interruption_code;
 260         mcck = this_cpu_ptr(&cpu_mcck);
 261         umode = user_mode(regs);
 262
 263         if (mci->sd) {
 264                 /* System damage -> stopping machine */
 265                 s390_handle_damage("received system damage machine check.");
 266         }
 267         if (mci->pd) {
 268                 if (mci->b) {
 269                         /* Processing backup -> verify if we can survive this */
 270                         u64 z_mcic, o_mcic, t_mcic;
 271 #ifdef CONFIG_64BIT
 272                         z_mcic = (1ULL<<63 | 1ULL<<59 | 1ULL<<29);
 273                         o_mcic = (1ULL<<43 | 1ULL<<42 | 1ULL<<41 | 1ULL<<40 |
 274                                   1ULL<<36 | 1ULL<<35 | 1ULL<<34 | 1ULL<<32 |
 275                                   1ULL<<30 | 1ULL<<21 | 1ULL<<20 | 1ULL<<17 |
 276                                   1ULL<<16);
 277 #else
 278                         z_mcic = (1ULL<<63 | 1ULL<<59 | 1ULL<<57 | 1ULL<<50 |
 279                                   1ULL<<29);
 280                         o_mcic = (1ULL<<43 | 1ULL<<42 | 1ULL<<41 | 1ULL<<40 |
 281                                   1ULL<<36 | 1ULL<<35 | 1ULL<<34 | 1ULL<<32 |
 282                                   1ULL<<30 | 1ULL<<20 | 1ULL<<17 | 1ULL<<16);
 283 #endif
 284                         t_mcic = *(u64 *)mci;
 285
 286                         if (((t_mcic & z_mcic) != 0) ||
 287                             ((t_mcic & o_mcic) != o_mcic)) {
 288                                 s390_handle_damage("processing backup machine "
 289                                                    "check with damage.");
 290                         }
 291
 292                         /*
 293                          * Nullifying exigent condition, therefore we might
 294                          * retry this instruction.
 295                          */
 296                         spin_lock(&ipd_lock);
 297                         tmp = get_tod_clock();
 298                         if (((tmp - last_ipd) >> 12) < MAX_IPD_TIME)
 299                                 ipd_count++;
 300                         else
 301                                 ipd_count = 1;
 302                         last_ipd = tmp;
 303                         if (ipd_count == MAX_IPD_COUNT)
 304                                 s390_handle_damage("too many ipd retries.");
 305                         spin_unlock(&ipd_lock);
 306                 } else {
 307                         /* Processing damage -> stopping machine */
 308                         s390_handle_damage("received instruction processing "
 309                                            "damage machine check.");
 310                 }
 311         }
 312         if (s390_revalidate_registers(mci)) {
 313                 if (umode) {
 314                         /*
 315                          * Couldn't restore all register contents while in
 316                          * user mode -> mark task for termination.
 317                          */
 318                         mcck->kill_task = 1;
 319                         mcck->mcck_code = *(unsigned long long *) mci;
 320                         set_cpu_flag(CIF_MCCK_PENDING);
 321                 } else {
 322                         /*
 323                          * Couldn't restore all register contents while in
 324                          * kernel mode -> stopping machine.
 325                          */
 326                         s390_handle_damage("unable to revalidate registers.");
 327                 }
 328         }
 329         if (mci->cd) {
 330                 /* Timing facility damage */
 331                 s390_handle_damage("TOD clock damaged");
 332         }
 333         if (mci->ed && mci->ec) {
 334                 /* External damage */
 335                 if (S390_lowcore.external_damage_code & (1U << ED_ETR_SYNC))
 336                         etr_sync_check();
 337                 if (S390_lowcore.external_damage_code & (1U << ED_ETR_SWITCH))
 338                         etr_switch_to_local();
 339                 if (S390_lowcore.external_damage_code & (1U << ED_STP_SYNC))
 340                         stp_sync_check();
 341                 if (S390_lowcore.external_damage_code & (1U << ED_STP_ISLAND))
 342                         stp_island_check();
 343         }
 344         if (mci->se)
 345                 /* Storage error uncorrected */
 346                 s390_handle_damage("received storage error uncorrected "
 347                                    "machine check.");
 348         if (mci->ke)
 349                 /* Storage key-error uncorrected */
 350                 s390_handle_damage("received storage key-error uncorrected "
 351                                    "machine check.");
 352         if (mci->ds && mci->fa)
 353                 /* Storage degradation */
 354                 s390_handle_damage("received storage degradation machine "
 355                                    "check.");
 356         if (mci->cp) {
 357                 /* Channel report word pending */
 358                 mcck->channel_report = 1;
 359                 set_cpu_flag(CIF_MCCK_PENDING);
 360         }
 361         if (mci->w) {
 362                 /* Warning pending */
 363                 mcck->warning = 1;
 364                 set_cpu_flag(CIF_MCCK_PENDING);
 365         }
 366         nmi_exit();
 367 }
 368
 369 static int __init machine_check_init(void)
 370 {
 371         ctl_set_bit(14, 25);    /* enable external damage MCH */
 372         ctl_set_bit(14, 27);    /* enable system recovery MCH */
 373         ctl_set_bit(14, 24);    /* enable warning MCH */
 374         return 0;
 375 }
 376 arch_initcall(machine_check_init);