arch/powerpc/platforms/powernv/opal.c

   1 /*
   2  * PowerNV OPAL high level interfaces
   3  *
   4  * Copyright 2011 IBM Corp.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #undef DEBUG
  13
  14 #include <linux/types.h>
  15 #include <linux/of.h>
  16 #include <linux/of_fdt.h>
  17 #include <linux/of_platform.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/notifier.h>
  20 #include <linux/slab.h>
  21 #include <linux/sched.h>
  22 #include <linux/kobject.h>
  23 #include <linux/delay.h>
  24 #include <linux/memblock.h>
  25 #include <asm/opal.h>
  26 #include <asm/firmware.h>
  27 #include <asm/mce.h>
  28
  29 #include "powernv.h"
  30
  31 /* /sys/firmware/opal */
  32 struct kobject *opal_kobj;
  33
  34 struct opal {
  35         u64 base;
  36         u64 entry;
  37         u64 size;
  38 } opal;
  39
  40 struct mcheck_recoverable_range {
  41         u64 start_addr;
  42         u64 end_addr;
  43         u64 recover_addr;
  44 };
  45
  46 static struct mcheck_recoverable_range *mc_recoverable_range;
  47 static int mc_recoverable_range_len;
  48
  49 static struct device_node *opal_node;
  50 static DEFINE_SPINLOCK(opal_write_lock);
  51 extern u64 opal_mc_secondary_handler[];
  52 static unsigned int *opal_irqs;
  53 static unsigned int opal_irq_count;
  54 static ATOMIC_NOTIFIER_HEAD(opal_notifier_head);
  55 static struct atomic_notifier_head opal_msg_notifier_head[OPAL_MSG_TYPE_MAX];
  56 static DEFINE_SPINLOCK(opal_notifier_lock);
  57 static uint64_t last_notified_mask = 0x0ul;
  58 static atomic_t opal_notifier_hold = ATOMIC_INIT(0);
  59
  60 int __init early_init_dt_scan_opal(unsigned long node,
  61                                    const char *uname, int depth, void *data)
  62 {
  63         const void *basep, *entryp, *sizep;
  64         unsigned long basesz, entrysz, runtimesz;
  65
  66         if (depth != 1 || strcmp(uname, "ibm,opal") != 0)
  67                 return 0;
  68
  69         basep  = of_get_flat_dt_prop(node, "opal-base-address", &basesz);
  70         entryp = of_get_flat_dt_prop(node, "opal-entry-address", &entrysz);
  71         sizep = of_get_flat_dt_prop(node, "opal-runtime-size", &runtimesz);
  72
  73         if (!basep || !entryp || !sizep)
  74                 return 1;
  75
  76         opal.base = of_read_number(basep, basesz/4);
  77         opal.entry = of_read_number(entryp, entrysz/4);
  78         opal.size = of_read_number(sizep, runtimesz/4);
  79
  80         pr_debug("OPAL Base  = 0x%llx (basep=%p basesz=%ld)\n",
  81                  opal.base, basep, basesz);
  82         pr_debug("OPAL Entry = 0x%llx (entryp=%p basesz=%ld)\n",
  83                  opal.entry, entryp, entrysz);
  84         pr_debug("OPAL Entry = 0x%llx (sizep=%p runtimesz=%ld)\n",
  85                  opal.size, sizep, runtimesz);
  86
  87         powerpc_firmware_features |= FW_FEATURE_OPAL;
  88         if (of_flat_dt_is_compatible(node, "ibm,opal-v3")) {
  89                 powerpc_firmware_features |= FW_FEATURE_OPALv2;
  90                 powerpc_firmware_features |= FW_FEATURE_OPALv3;
  91                 printk("OPAL V3 detected !\n");
  92         } else if (of_flat_dt_is_compatible(node, "ibm,opal-v2")) {
  93                 powerpc_firmware_features |= FW_FEATURE_OPALv2;
  94                 printk("OPAL V2 detected !\n");
  95         } else {
  96                 printk("OPAL V1 detected !\n");
  97         }
  98
  99         return 1;
 100 }
 101
 102 int __init early_init_dt_scan_recoverable_ranges(unsigned long node,
 103                                    const char *uname, int depth, void *data)
 104 {
 105         unsigned long i, size;
 106         const __be32 *prop;
 107
 108         if (depth != 1 || strcmp(uname, "ibm,opal") != 0)
 109                 return 0;
 110
 111         prop = of_get_flat_dt_prop(node, "mcheck-recoverable-ranges", &size);
 112
 113         if (!prop)
 114                 return 1;
 115
 116         pr_debug("Found machine check recoverable ranges.\n");
 117
 118         /*
 119          * Allocate a buffer to hold the MC recoverable ranges. We would be
 120          * accessing them in real mode, hence it needs to be within
 121          * RMO region.
 122          */
 123         mc_recoverable_range =__va(memblock_alloc_base(size, __alignof__(u64),
 124                                                         ppc64_rma_size));
 125         memset(mc_recoverable_range, 0, size);
 126
 127         /*
 128          * Each recoverable address entry is an (start address,len,
 129          * recover address) pair, * 2 cells each, totalling 4 cells per entry.
 130          */
 131         for (i = 0; i < size / (sizeof(*prop) * 5); i++) {
 132                 mc_recoverable_range[i].start_addr =
 133                                         of_read_number(prop + (i * 5) + 0, 2);
 134                 mc_recoverable_range[i].end_addr =
 135                                         mc_recoverable_range[i].start_addr +
 136                                         of_read_number(prop + (i * 5) + 2, 1);
 137                 mc_recoverable_range[i].recover_addr =
 138                                         of_read_number(prop + (i * 5) + 3, 2);
 139
 140                 pr_debug("Machine check recoverable range: %llx..%llx: %llx\n",
 141                                 mc_recoverable_range[i].start_addr,
 142                                 mc_recoverable_range[i].end_addr,
 143                                 mc_recoverable_range[i].recover_addr);
 144         }
 145         mc_recoverable_range_len = i;
 146         return 1;
 147 }
 148
 149 static int __init opal_register_exception_handlers(void)
 150 {
 151 #ifdef __BIG_ENDIAN__
 152         u64 glue;
 153
 154         if (!(powerpc_firmware_features & FW_FEATURE_OPAL))
 155                 return -ENODEV;
 156
 157         /* Hookup some exception handlers except machine check. We use the
 158          * fwnmi area at 0x7000 to provide the glue space to OPAL
 159          */
 160         glue = 0x7000;
 161         opal_register_exception_handler(OPAL_HYPERVISOR_MAINTENANCE_HANDLER,
 162                                         0, glue);
 163         glue += 128;
 164         opal_register_exception_handler(OPAL_SOFTPATCH_HANDLER, 0, glue);
 165 #endif
 166
 167         return 0;
 168 }
 169
 170 early_initcall(opal_register_exception_handlers);
 171
 172 int opal_notifier_register(struct notifier_block *nb)
 173 {
 174         if (!nb) {
 175                 pr_warning("%s: Invalid argument (%p)\n",
 176                            __func__, nb);
 177                 return -EINVAL;
 178         }
 179
 180         atomic_notifier_chain_register(&opal_notifier_head, nb);
 181         return 0;
 182 }
 183 EXPORT_SYMBOL_GPL(opal_notifier_register);
 184
 185 int opal_notifier_unregister(struct notifier_block *nb)
 186 {
 187         if (!nb) {
 188                 pr_warning("%s: Invalid argument (%p)\n",
 189                            __func__, nb);
 190                 return -EINVAL;
 191         }
 192
 193         atomic_notifier_chain_unregister(&opal_notifier_head, nb);
 194         return 0;
 195 }
 196 EXPORT_SYMBOL_GPL(opal_notifier_unregister);
 197
 198 static void opal_do_notifier(uint64_t events)
 199 {
 200         unsigned long flags;
 201         uint64_t changed_mask;
 202
 203         if (atomic_read(&opal_notifier_hold))
 204                 return;
 205
 206         spin_lock_irqsave(&opal_notifier_lock, flags);
 207         changed_mask = last_notified_mask ^ events;
 208         last_notified_mask = events;
 209         spin_unlock_irqrestore(&opal_notifier_lock, flags);
 210
 211         /*
 212          * We feed with the event bits and changed bits for
 213          * enough information to the callback.
 214          */
 215         atomic_notifier_call_chain(&opal_notifier_head,
 216                                    events, (void *)changed_mask);
 217 }
 218
 219 void opal_notifier_update_evt(uint64_t evt_mask,
 220                               uint64_t evt_val)
 221 {
 222         unsigned long flags;
 223
 224         spin_lock_irqsave(&opal_notifier_lock, flags);
 225         last_notified_mask &= ~evt_mask;
 226         last_notified_mask |= evt_val;
 227         spin_unlock_irqrestore(&opal_notifier_lock, flags);
 228 }
 229
 230 void opal_notifier_enable(void)
 231 {
 232         int64_t rc;
 233         uint64_t evt = 0;
 234
 235         atomic_set(&opal_notifier_hold, 0);
 236
 237         /* Process pending events */
 238         rc = opal_poll_events(&evt);
 239         if (rc == OPAL_SUCCESS && evt)
 240                 opal_do_notifier(evt);
 241 }
 242
 243 void opal_notifier_disable(void)
 244 {
 245         atomic_set(&opal_notifier_hold, 1);
 246 }
 247
 248 /*
 249  * Opal message notifier based on message type. Allow subscribers to get
 250  * notified for specific messgae type.
 251  */
 252 int opal_message_notifier_register(enum OpalMessageType msg_type,
 253                                         struct notifier_block *nb)
 254 {
 255         if (!nb) {
 256                 pr_warning("%s: Invalid argument (%p)\n",
 257                            __func__, nb);
 258                 return -EINVAL;
 259         }
 260         if (msg_type > OPAL_MSG_TYPE_MAX) {
 261                 pr_warning("%s: Invalid message type argument (%d)\n",
 262                            __func__, msg_type);
 263                 return -EINVAL;
 264         }
 265         return atomic_notifier_chain_register(
 266                                 &opal_msg_notifier_head[msg_type], nb);
 267 }
 268
 269 static void opal_message_do_notify(uint32_t msg_type, void *msg)
 270 {
 271         /* notify subscribers */
 272         atomic_notifier_call_chain(&opal_msg_notifier_head[msg_type],
 273                                         msg_type, msg);
 274 }
 275
 276 static void opal_handle_message(void)
 277 {
 278         s64 ret;
 279         /*
 280          * TODO: pre-allocate a message buffer depending on opal-msg-size
 281          * value in /proc/device-tree.
 282          */
 283         static struct opal_msg msg;
 284         u32 type;
 285
 286         ret = opal_get_msg(__pa(&msg), sizeof(msg));
 287         /* No opal message pending. */
 288         if (ret == OPAL_RESOURCE)
 289                 return;
 290
 291         /* check for errors. */
 292         if (ret) {
 293                 pr_warning("%s: Failed to retrive opal message, err=%lld\n",
 294                                 __func__, ret);
 295                 return;
 296         }
 297
 298         type = be32_to_cpu(msg.msg_type);
 299
 300         /* Sanity check */
 301         if (type > OPAL_MSG_TYPE_MAX) {
 302                 pr_warning("%s: Unknown message type: %u\n", __func__, type);
 303                 return;
 304         }
 305         opal_message_do_notify(type, (void *)&msg);
 306 }
 307
 308 static int opal_message_notify(struct notifier_block *nb,
 309                           unsigned long events, void *change)
 310 {
 311         if (events & OPAL_EVENT_MSG_PENDING)
 312                 opal_handle_message();
 313         return 0;
 314 }
 315
 316 static struct notifier_block opal_message_nb = {
 317         .notifier_call  = opal_message_notify,
 318         .next           = NULL,
 319         .priority       = 0,
 320 };
 321
 322 static int __init opal_message_init(void)
 323 {
 324         int ret, i;
 325
 326         for (i = 0; i < OPAL_MSG_TYPE_MAX; i++)
 327                 ATOMIC_INIT_NOTIFIER_HEAD(&opal_msg_notifier_head[i]);
 328
 329         ret = opal_notifier_register(&opal_message_nb);
 330         if (ret) {
 331                 pr_err("%s: Can't register OPAL event notifier (%d)\n",
 332                        __func__, ret);
 333                 return ret;
 334         }
 335         return 0;
 336 }
 337 early_initcall(opal_message_init);
 338
 339 int opal_get_chars(uint32_t vtermno, char *buf, int count)
 340 {
 341         s64 rc;
 342         __be64 evt, len;
 343
 344         if (!opal.entry)
 345                 return -ENODEV;
 346         opal_poll_events(&evt);
 347         if ((be64_to_cpu(evt) & OPAL_EVENT_CONSOLE_INPUT) == 0)
 348                 return 0;
 349         len = cpu_to_be64(count);
 350         rc = opal_console_read(vtermno, &len, buf);
 351         if (rc == OPAL_SUCCESS)
 352                 return be64_to_cpu(len);
 353         return 0;
 354 }
 355
 356 int opal_put_chars(uint32_t vtermno, const char *data, int total_len)
 357 {
 358         int written = 0;
 359         __be64 olen;
 360         s64 len, rc;
 361         unsigned long flags;
 362         __be64 evt;
 363
 364         if (!opal.entry)
 365                 return -ENODEV;
 366
 367         /* We want put_chars to be atomic to avoid mangling of hvsi
 368          * packets. To do that, we first test for room and return
 369          * -EAGAIN if there isn't enough.
 370          *
 371          * Unfortunately, opal_console_write_buffer_space() doesn't
 372          * appear to work on opal v1, so we just assume there is
 373          * enough room and be done with it
 374          */
 375         spin_lock_irqsave(&opal_write_lock, flags);
 376         if (firmware_has_feature(FW_FEATURE_OPALv2)) {
 377                 rc = opal_console_write_buffer_space(vtermno, &olen);
 378                 len = be64_to_cpu(olen);
 379                 if (rc || len < total_len) {
 380                         spin_unlock_irqrestore(&opal_write_lock, flags);
 381                         /* Closed -> drop characters */
 382                         if (rc)
 383                                 return total_len;
 384                         opal_poll_events(NULL);
 385                         return -EAGAIN;
 386                 }
 387         }
 388
 389         /* We still try to handle partial completions, though they
 390          * should no longer happen.
 391          */
 392         rc = OPAL_BUSY;
 393         while(total_len > 0 && (rc == OPAL_BUSY ||
 394                                 rc == OPAL_BUSY_EVENT || rc == OPAL_SUCCESS)) {
 395                 olen = cpu_to_be64(total_len);
 396                 rc = opal_console_write(vtermno, &olen, data);
 397                 len = be64_to_cpu(olen);
 398
 399                 /* Closed or other error drop */
 400                 if (rc != OPAL_SUCCESS && rc != OPAL_BUSY &&
 401                     rc != OPAL_BUSY_EVENT) {
 402                         written = total_len;
 403                         break;
 404                 }
 405                 if (rc == OPAL_SUCCESS) {
 406                         total_len -= len;
 407                         data += len;
 408                         written += len;
 409                 }
 410                 /* This is a bit nasty but we need that for the console to
 411                  * flush when there aren't any interrupts. We will clean
 412                  * things a bit later to limit that to synchronous path
 413                  * such as the kernel console and xmon/udbg
 414                  */
 415                 do
 416                         opal_poll_events(&evt);
 417                 while(rc == OPAL_SUCCESS &&
 418                         (be64_to_cpu(evt) & OPAL_EVENT_CONSOLE_OUTPUT));
 419         }
 420         spin_unlock_irqrestore(&opal_write_lock, flags);
 421         return written;
 422 }
 423
 424 static int opal_recover_mce(struct pt_regs *regs,
 425                                         struct machine_check_event *evt)
 426 {
 427         int recovered = 0;
 428         uint64_t ea = get_mce_fault_addr(evt);
 429
 430         if (!(regs->msr & MSR_RI)) {
 431                 /* If MSR_RI isn't set, we cannot recover */
 432                 recovered = 0;
 433         } else if (evt->disposition == MCE_DISPOSITION_RECOVERED) {
 434                 /* Platform corrected itself */
 435                 recovered = 1;
 436         } else if (ea && !is_kernel_addr(ea)) {
 437                 /*
 438                  * Faulting address is not in kernel text. We should be fine.
 439                  * We need to find which process uses this address.
 440                  * For now, kill the task if we have received exception when
 441                  * in userspace.
 442                  *
 443                  * TODO: Queue up this address for hwpoisioning later.
 444                  */
 445                 if (user_mode(regs) && !is_global_init(current)) {
 446                         _exception(SIGBUS, regs, BUS_MCEERR_AR, regs->nip);
 447                         recovered = 1;
 448                 } else
 449                         recovered = 0;
 450         } else if (user_mode(regs) && !is_global_init(current) &&
 451                 evt->severity == MCE_SEV_ERROR_SYNC) {
 452                 /*
 453                  * If we have received a synchronous error when in userspace
 454                  * kill the task.
 455                  */
 456                 _exception(SIGBUS, regs, BUS_MCEERR_AR, regs->nip);
 457                 recovered = 1;
 458         }
 459         return recovered;
 460 }
 461
 462 int opal_machine_check(struct pt_regs *regs)
 463 {
 464         struct machine_check_event evt;
 465
 466         if (!get_mce_event(&evt, MCE_EVENT_RELEASE))
 467                 return 0;
 468
 469         /* Print things out */
 470         if (evt.version != MCE_V1) {
 471                 pr_err("Machine Check Exception, Unknown event version %d !\n",
 472                        evt.version);
 473                 return 0;
 474         }
 475         machine_check_print_event_info(&evt);
 476
 477         if (opal_recover_mce(regs, &evt))
 478                 return 1;
 479         return 0;
 480 }
 481
 482 static uint64_t find_recovery_address(uint64_t nip)
 483 {
 484         int i;
 485
 486         for (i = 0; i < mc_recoverable_range_len; i++)
 487                 if ((nip >= mc_recoverable_range[i].start_addr) &&
 488                     (nip < mc_recoverable_range[i].end_addr))
 489                     return mc_recoverable_range[i].recover_addr;
 490         return 0;
 491 }
 492
 493 bool opal_mce_check_early_recovery(struct pt_regs *regs)
 494 {
 495         uint64_t recover_addr = 0;
 496
 497         if (!opal.base || !opal.size)
 498                 goto out;
 499
 500         if ((regs->nip >= opal.base) &&
 501                         (regs->nip <= (opal.base + opal.size)))
 502                 recover_addr = find_recovery_address(regs->nip);
 503
 504         /*
 505          * Setup regs->nip to rfi into fixup address.
 506          */
 507         if (recover_addr)
 508                 regs->nip = recover_addr;
 509
 510 out:
 511         return !!recover_addr;
 512 }
 513
 514 static irqreturn_t opal_interrupt(int irq, void *data)
 515 {
 516         __be64 events;
 517
 518         opal_handle_interrupt(virq_to_hw(irq), &events);
 519
 520         opal_do_notifier(events);
 521
 522         return IRQ_HANDLED;
 523 }
 524
 525 static int opal_sysfs_init(void)
 526 {
 527         opal_kobj = kobject_create_and_add("opal", firmware_kobj);
 528         if (!opal_kobj) {
 529                 pr_warn("kobject_create_and_add opal failed\n");
 530                 return -ENOMEM;
 531         }
 532
 533         return 0;
 534 }
 535
 536 static int __init opal_init(void)
 537 {
 538         struct device_node *np, *consoles;
 539         const __be32 *irqs;
 540         int rc, i, irqlen;
 541
 542         opal_node = of_find_node_by_path("/ibm,opal");
 543         if (!opal_node) {
 544                 pr_warn("opal: Node not found\n");
 545                 return -ENODEV;
 546         }
 547
 548         /* Register OPAL consoles if any ports */
 549         if (firmware_has_feature(FW_FEATURE_OPALv2))
 550                 consoles = of_find_node_by_path("/ibm,opal/consoles");
 551         else
 552                 consoles = of_node_get(opal_node);
 553         if (consoles) {
 554                 for_each_child_of_node(consoles, np) {
 555                         if (strcmp(np->name, "serial"))
 556                                 continue;
 557                         of_platform_device_create(np, NULL, NULL);
 558                 }
 559                 of_node_put(consoles);
 560         }
 561
 562         /* Find all OPAL interrupts and request them */
 563         irqs = of_get_property(opal_node, "opal-interrupts", &irqlen);
 564         pr_debug("opal: Found %d interrupts reserved for OPAL\n",
 565                  irqs ? (irqlen / 4) : 0);
 566         opal_irq_count = irqlen / 4;
 567         opal_irqs = kzalloc(opal_irq_count * sizeof(unsigned int), GFP_KERNEL);
 568         for (i = 0; irqs && i < (irqlen / 4); i++, irqs++) {
 569                 unsigned int hwirq = be32_to_cpup(irqs);
 570                 unsigned int irq = irq_create_mapping(NULL, hwirq);
 571                 if (irq == NO_IRQ) {
 572                         pr_warning("opal: Failed to map irq 0x%x\n", hwirq);
 573                         continue;
 574                 }
 575                 rc = request_irq(irq, opal_interrupt, 0, "opal", NULL);
 576                 if (rc)
 577                         pr_warning("opal: Error %d requesting irq %d"
 578                                    " (0x%x)\n", rc, irq, hwirq);
 579                 opal_irqs[i] = irq;
 580         }
 581
 582         /* Create "opal" kobject under /sys/firmware */
 583         rc = opal_sysfs_init();
 584         if (rc == 0) {
 585                 /* Setup error log interface */
 586                 rc = opal_elog_init();
 587                 /* Setup code update interface */
 588                 opal_flash_init();
 589                 /* Setup platform dump extract interface */
 590                 opal_platform_dump_init();
 591                 /* Setup system parameters interface */
 592                 opal_sys_param_init();
 593         }
 594
 595         return 0;
 596 }
 597 subsys_initcall(opal_init);
 598
 599 void opal_shutdown(void)
 600 {
 601         unsigned int i;
 602         long rc = OPAL_BUSY;
 603
 604         /* First free interrupts, which will also mask them */
 605         for (i = 0; i < opal_irq_count; i++) {
 606                 if (opal_irqs[i])
 607                         free_irq(opal_irqs[i], NULL);
 608                 opal_irqs[i] = 0;
 609         }
 610
 611         /*
 612          * Then sync with OPAL which ensure anything that can
 613          * potentially write to our memory has completed such
 614          * as an ongoing dump retrieval
 615          */
 616         while (rc == OPAL_BUSY || rc == OPAL_BUSY_EVENT) {
 617                 rc = opal_sync_host_reboot();
 618                 if (rc == OPAL_BUSY)
 619                         opal_poll_events(NULL);
 620                 else
 621                         mdelay(10);
 622         }
 623 }