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 #include <asm/switch_to.h>
  24 #include <asm/fpu-internal.h>
  25 #include <asm/ctl_reg.h>
  26
  27 struct mcck_struct {
  28         int kill_task;
  29         int channel_report;
  30         int warning;
  31         unsigned int etr_queue : 1;
  32         unsigned int stp_queue : 1;
  33         unsigned long long mcck_code;
  34 };
  35
  36 static DEFINE_PER_CPU(struct mcck_struct, cpu_mcck);
  37
  38 static void s390_handle_damage(char *msg)
  39 {
  40         smp_send_stop();
  41         disabled_wait((unsigned long) __builtin_return_address(0));
  42         while (1);
  43 }
  44
  45 /*
  46  * Main machine check handler function. Will be called with interrupts enabled
  47  * or disabled and machine checks enabled or disabled.
  48  */
  49 void s390_handle_mcck(void)
  50 {
  51         unsigned long flags;
  52         struct mcck_struct mcck;
  53
  54         /*
  55          * Disable machine checks and get the current state of accumulated
  56          * machine checks. Afterwards delete the old state and enable machine
  57          * checks again.
  58          */
  59         local_irq_save(flags);
  60         local_mcck_disable();
  61         mcck = *this_cpu_ptr(&cpu_mcck);
  62         memset(this_cpu_ptr(&cpu_mcck), 0, sizeof(mcck));
  63         clear_cpu_flag(CIF_MCCK_PENDING);
  64         local_mcck_enable();
  65         local_irq_restore(flags);
  66
  67         if (mcck.channel_report)
  68                 crw_handle_channel_report();
  69         /*
  70          * A warning may remain for a prolonged period on the bare iron.
  71          * (actually until the machine is powered off, or the problem is gone)
  72          * So we just stop listening for the WARNING MCH and avoid continuously
  73          * being interrupted.  One caveat is however, that we must do this per
  74          * processor and cannot use the smp version of ctl_clear_bit().
  75          * On VM we only get one interrupt per virtally presented machinecheck.
  76          * Though one suffices, we may get one interrupt per (virtual) cpu.
  77          */
  78         if (mcck.warning) {     /* WARNING pending ? */
  79                 static int mchchk_wng_posted = 0;
  80
  81                 /* Use single cpu clear, as we cannot handle smp here. */
  82                 __ctl_clear_bit(14, 24);        /* Disable WARNING MCH */
  83                 if (xchg(&mchchk_wng_posted, 1) == 0)
  84                         kill_cad_pid(SIGPWR, 1);
  85         }
  86         if (mcck.etr_queue)
  87                 etr_queue_work();
  88         if (mcck.stp_queue)
  89                 stp_queue_work();
  90         if (mcck.kill_task) {
  91                 local_irq_enable();
  92                 printk(KERN_EMERG "mcck: Terminating task because of machine "
  93                        "malfunction (code 0x%016llx).\n", mcck.mcck_code);
  94                 printk(KERN_EMERG "mcck: task: %s, pid: %d.\n",
  95                        current->comm, current->pid);
  96                 do_exit(SIGSEGV);
  97         }
  98 }
  99 EXPORT_SYMBOL_GPL(s390_handle_mcck);
 100
 101 /*
 102  * returns 0 if all registers could be validated
 103  * returns 1 otherwise
 104  */
 105 static int notrace s390_revalidate_registers(struct mci *mci)
 106 {
 107         int kill_task;
 108         u64 zero;
 109         void *fpt_save_area, *fpt_creg_save_area;
 110
 111         kill_task = 0;
 112         zero = 0;
 113
 114         if (!mci->gr) {
 115                 /*
 116                  * General purpose registers couldn't be restored and have
 117                  * unknown contents. Process needs to be terminated.
 118                  */
 119                 kill_task = 1;
 120         }
 121         if (!mci->fp) {
 122                 /*
 123                  * Floating point registers can't be restored and
 124                  * therefore the process needs to be terminated.
 125                  */
 126                 kill_task = 1;
 127         }
 128         fpt_save_area = &S390_lowcore.floating_pt_save_area;
 129         fpt_creg_save_area = &S390_lowcore.fpt_creg_save_area;
 130         if (!mci->fc) {
 131                 /*
 132                  * Floating point control register can't be restored.
 133                  * Task will be terminated.
 134                  */
 135                 asm volatile("lfpc 0(%0)" : : "a" (&zero), "m" (zero));
 136                 kill_task = 1;
 137         } else
 138                 asm volatile("lfpc 0(%0)" : : "a" (fpt_creg_save_area));
 139
 140         if (!MACHINE_HAS_VX) {
 141                 /* Revalidate floating point registers */
 142                 asm volatile(
 143                         "       ld      0,0(%0)\n"
 144                         "       ld      1,8(%0)\n"
 145                         "       ld      2,16(%0)\n"
 146                         "       ld      3,24(%0)\n"
 147                         "       ld      4,32(%0)\n"
 148                         "       ld      5,40(%0)\n"
 149                         "       ld      6,48(%0)\n"
 150                         "       ld      7,56(%0)\n"
 151                         "       ld      8,64(%0)\n"
 152                         "       ld      9,72(%0)\n"
 153                         "       ld      10,80(%0)\n"
 154                         "       ld      11,88(%0)\n"
 155                         "       ld      12,96(%0)\n"
 156                         "       ld      13,104(%0)\n"
 157                         "       ld      14,112(%0)\n"
 158                         "       ld      15,120(%0)\n"
 159                         : : "a" (fpt_save_area));
 160         } else {
 161                 /* Revalidate vector registers */
 162                 union ctlreg0 cr0;
 163
 164                 if (!mci->vr) {
 165                         /*
 166                          * Vector registers can't be restored and therefore
 167                          * the process needs to be terminated.
 168                          */
 169                         kill_task = 1;
 170                 }
 171                 cr0.val = S390_lowcore.cregs_save_area[0];
 172                 cr0.afp = cr0.vx = 1;
 173                 __ctl_load(cr0.val, 0, 0);
 174                 asm volatile(
 175                         "       la      1,%0\n"
 176                         "       .word   0xe70f,0x1000,0x0036\n" /* vlm 0,15,0(1) */
 177                         "       .word   0xe70f,0x1100,0x0c36\n" /* vlm 16,31,256(1) */
 178                         : : "Q" (*(struct vx_array *)
 179                                  &S390_lowcore.vector_save_area) : "1");
 180                 __ctl_load(S390_lowcore.cregs_save_area[0], 0, 0);
 181         }
 182         /* Revalidate access registers */
 183         asm volatile(
 184                 "       lam     0,15,0(%0)"
 185                 : : "a" (&S390_lowcore.access_regs_save_area));
 186         if (!mci->ar) {
 187                 /*
 188                  * Access registers have unknown contents.
 189                  * Terminating task.
 190                  */
 191                 kill_task = 1;
 192         }
 193         /* Revalidate control registers */
 194         if (!mci->cr) {
 195                 /*
 196                  * Control registers have unknown contents.
 197                  * Can't recover and therefore stopping machine.
 198                  */
 199                 s390_handle_damage("invalid control registers.");
 200         } else {
 201                 asm volatile(
 202                         "       lctlg   0,15,0(%0)"
 203                         : : "a" (&S390_lowcore.cregs_save_area));
 204         }
 205         /*
 206          * We don't even try to revalidate the TOD register, since we simply
 207          * can't write something sensible into that register.
 208          */
 209         /*
 210          * See if we can revalidate the TOD programmable register with its
 211          * old contents (should be zero) otherwise set it to zero.
 212          */
 213         if (!mci->pr)
 214                 asm volatile(
 215                         "       sr      0,0\n"
 216                         "       sckpf"
 217                         : : : "0", "cc");
 218         else
 219                 asm volatile(
 220                         "       l       0,0(%0)\n"
 221                         "       sckpf"
 222                         : : "a" (&S390_lowcore.tod_progreg_save_area)
 223                         : "0", "cc");
 224         /* Revalidate clock comparator register */
 225         set_clock_comparator(S390_lowcore.clock_comparator);
 226         /* Check if old PSW is valid */
 227         if (!mci->wp)
 228                 /*
 229                  * Can't tell if we come from user or kernel mode
 230                  * -> stopping machine.
 231                  */
 232                 s390_handle_damage("old psw invalid.");
 233
 234         if (!mci->ms || !mci->pm || !mci->ia)
 235                 kill_task = 1;
 236
 237         return kill_task;
 238 }
 239
 240 #define MAX_IPD_COUNT   29
 241 #define MAX_IPD_TIME    (5 * 60 * USEC_PER_SEC) /* 5 minutes */
 242
 243 #define ED_STP_ISLAND   6       /* External damage STP island check */
 244 #define ED_STP_SYNC     7       /* External damage STP sync check */
 245 #define ED_ETR_SYNC     12      /* External damage ETR sync check */
 246 #define ED_ETR_SWITCH   13      /* External damage ETR switch to local */
 247
 248 /*
 249  * machine check handler.
 250  */
 251 void notrace s390_do_machine_check(struct pt_regs *regs)
 252 {
 253         static int ipd_count;
 254         static DEFINE_SPINLOCK(ipd_lock);
 255         static unsigned long long last_ipd;
 256         struct mcck_struct *mcck;
 257         unsigned long long tmp;
 258         struct mci *mci;
 259         int umode;
 260
 261         nmi_enter();
 262         inc_irq_stat(NMI_NMI);
 263         mci = (struct mci *) &S390_lowcore.mcck_interruption_code;
 264         mcck = this_cpu_ptr(&cpu_mcck);
 265         umode = user_mode(regs);
 266
 267         if (mci->sd) {
 268                 /* System damage -> stopping machine */
 269                 s390_handle_damage("received system damage machine check.");
 270         }
 271         if (mci->pd) {
 272                 if (mci->b) {
 273                         /* Processing backup -> verify if we can survive this */
 274                         u64 z_mcic, o_mcic, t_mcic;
 275                         z_mcic = (1ULL<<63 | 1ULL<<59 | 1ULL<<29);
 276                         o_mcic = (1ULL<<43 | 1ULL<<42 | 1ULL<<41 | 1ULL<<40 |
 277                                   1ULL<<36 | 1ULL<<35 | 1ULL<<34 | 1ULL<<32 |
 278                                   1ULL<<30 | 1ULL<<21 | 1ULL<<20 | 1ULL<<17 |
 279                                   1ULL<<16);
 280                         t_mcic = *(u64 *)mci;
 281
 282                         if (((t_mcic & z_mcic) != 0) ||
 283                             ((t_mcic & o_mcic) != o_mcic)) {
 284                                 s390_handle_damage("processing backup machine "
 285                                                    "check with damage.");
 286                         }
 287
 288                         /*
 289                          * Nullifying exigent condition, therefore we might
 290                          * retry this instruction.
 291                          */
 292                         spin_lock(&ipd_lock);
 293                         tmp = get_tod_clock();
 294                         if (((tmp - last_ipd) >> 12) < MAX_IPD_TIME)
 295                                 ipd_count++;
 296                         else
 297                                 ipd_count = 1;
 298                         last_ipd = tmp;
 299                         if (ipd_count == MAX_IPD_COUNT)
 300                                 s390_handle_damage("too many ipd retries.");
 301                         spin_unlock(&ipd_lock);
 302                 } else {
 303                         /* Processing damage -> stopping machine */
 304                         s390_handle_damage("received instruction processing "
 305                                            "damage machine check.");
 306                 }
 307         }
 308         if (s390_revalidate_registers(mci)) {
 309                 if (umode) {
 310                         /*
 311                          * Couldn't restore all register contents while in
 312                          * user mode -> mark task for termination.
 313                          */
 314                         mcck->kill_task = 1;
 315                         mcck->mcck_code = *(unsigned long long *) mci;
 316                         set_cpu_flag(CIF_MCCK_PENDING);
 317                 } else {
 318                         /*
 319                          * Couldn't restore all register contents while in
 320                          * kernel mode -> stopping machine.
 321                          */
 322                         s390_handle_damage("unable to revalidate registers.");
 323                 }
 324         }
 325         if (mci->cd) {
 326                 /* Timing facility damage */
 327                 s390_handle_damage("TOD clock damaged");
 328         }
 329         if (mci->ed && mci->ec) {
 330                 /* External damage */
 331                 if (S390_lowcore.external_damage_code & (1U << ED_ETR_SYNC))
 332                         mcck->etr_queue |= etr_sync_check();
 333                 if (S390_lowcore.external_damage_code & (1U << ED_ETR_SWITCH))
 334                         mcck->etr_queue |= etr_switch_to_local();
 335                 if (S390_lowcore.external_damage_code & (1U << ED_STP_SYNC))
 336                         mcck->stp_queue |= stp_sync_check();
 337                 if (S390_lowcore.external_damage_code & (1U << ED_STP_ISLAND))
 338                         mcck->stp_queue |= stp_island_check();
 339                 if (mcck->etr_queue || mcck->stp_queue)
 340                         set_cpu_flag(CIF_MCCK_PENDING);
 341         }
 342         if (mci->se)
 343                 /* Storage error uncorrected */
 344                 s390_handle_damage("received storage error uncorrected "
 345                                    "machine check.");
 346         if (mci->ke)
 347                 /* Storage key-error uncorrected */
 348                 s390_handle_damage("received storage key-error uncorrected "
 349                                    "machine check.");
 350         if (mci->ds && mci->fa)
 351                 /* Storage degradation */
 352                 s390_handle_damage("received storage degradation machine "
 353                                    "check.");
 354         if (mci->cp) {
 355                 /* Channel report word pending */
 356                 mcck->channel_report = 1;
 357                 set_cpu_flag(CIF_MCCK_PENDING);
 358         }
 359         if (mci->w) {
 360                 /* Warning pending */
 361                 mcck->warning = 1;
 362                 set_cpu_flag(CIF_MCCK_PENDING);
 363         }
 364         nmi_exit();
 365 }
 366
 367 static int __init machine_check_init(void)
 368 {
 369         ctl_set_bit(14, 25);    /* enable external damage MCH */
 370         ctl_set_bit(14, 27);    /* enable system recovery MCH */
 371         ctl_set_bit(14, 24);    /* enable warning MCH */
 372         return 0;
 373 }
 374 early_initcall(machine_check_init);