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
 285         ret = opal_get_msg(__pa(&msg), sizeof(msg));
 286         /* No opal message pending. */
 287         if (ret == OPAL_RESOURCE)
 288                 return;
 289
 290         /* check for errors. */
 291         if (ret) {
 292                 pr_warning("%s: Failed to retrive opal message, err=%lld\n",
 293                                 __func__, ret);
 294                 return;
 295         }
 296
 297         /* Sanity check */
 298         if (msg.msg_type > OPAL_MSG_TYPE_MAX) {
 299                 pr_warning("%s: Unknown message type: %u\n",
 300                                 __func__, msg.msg_type);
 301                 return;
 302         }
 303         opal_message_do_notify(msg.msg_type, (void *)&msg);
 304 }
 305
 306 static int opal_message_notify(struct notifier_block *nb,
 307                           unsigned long events, void *change)
 308 {
 309         if (events & OPAL_EVENT_MSG_PENDING)
 310                 opal_handle_message();
 311         return 0;
 312 }
 313
 314 static struct notifier_block opal_message_nb = {
 315         .notifier_call  = opal_message_notify,
 316         .next           = NULL,
 317         .priority       = 0,
 318 };
 319
 320 static int __init opal_message_init(void)
 321 {
 322         int ret, i;
 323
 324         for (i = 0; i < OPAL_MSG_TYPE_MAX; i++)
 325                 ATOMIC_INIT_NOTIFIER_HEAD(&opal_msg_notifier_head[i]);
 326
 327         ret = opal_notifier_register(&opal_message_nb);
 328         if (ret) {
 329                 pr_err("%s: Can't register OPAL event notifier (%d)\n",
 330                        __func__, ret);
 331                 return ret;
 332         }
 333         return 0;
 334 }
 335 early_initcall(opal_message_init);
 336
 337 int opal_get_chars(uint32_t vtermno, char *buf, int count)
 338 {
 339         s64 rc;
 340         __be64 evt, len;
 341
 342         if (!opal.entry)
 343                 return -ENODEV;
 344         opal_poll_events(&evt);
 345         if ((be64_to_cpu(evt) & OPAL_EVENT_CONSOLE_INPUT) == 0)
 346                 return 0;
 347         len = cpu_to_be64(count);
 348         rc = opal_console_read(vtermno, &len, buf);
 349         if (rc == OPAL_SUCCESS)
 350                 return be64_to_cpu(len);
 351         return 0;
 352 }
 353
 354 int opal_put_chars(uint32_t vtermno, const char *data, int total_len)
 355 {
 356         int written = 0;
 357         __be64 olen;
 358         s64 len, rc;
 359         unsigned long flags;
 360         __be64 evt;
 361
 362         if (!opal.entry)
 363                 return -ENODEV;
 364
 365         /* We want put_chars to be atomic to avoid mangling of hvsi
 366          * packets. To do that, we first test for room and return
 367          * -EAGAIN if there isn't enough.
 368          *
 369          * Unfortunately, opal_console_write_buffer_space() doesn't
 370          * appear to work on opal v1, so we just assume there is
 371          * enough room and be done with it
 372          */
 373         spin_lock_irqsave(&opal_write_lock, flags);
 374         if (firmware_has_feature(FW_FEATURE_OPALv2)) {
 375                 rc = opal_console_write_buffer_space(vtermno, &olen);
 376                 len = be64_to_cpu(olen);
 377                 if (rc || len < total_len) {
 378                         spin_unlock_irqrestore(&opal_write_lock, flags);
 379                         /* Closed -> drop characters */
 380                         if (rc)
 381                                 return total_len;
 382                         opal_poll_events(NULL);
 383                         return -EAGAIN;
 384                 }
 385         }
 386
 387         /* We still try to handle partial completions, though they
 388          * should no longer happen.
 389          */
 390         rc = OPAL_BUSY;
 391         while(total_len > 0 && (rc == OPAL_BUSY ||
 392                                 rc == OPAL_BUSY_EVENT || rc == OPAL_SUCCESS)) {
 393                 olen = cpu_to_be64(total_len);
 394                 rc = opal_console_write(vtermno, &olen, data);
 395                 len = be64_to_cpu(olen);
 396
 397                 /* Closed or other error drop */
 398                 if (rc != OPAL_SUCCESS && rc != OPAL_BUSY &&
 399                     rc != OPAL_BUSY_EVENT) {
 400                         written = total_len;
 401                         break;
 402                 }
 403                 if (rc == OPAL_SUCCESS) {
 404                         total_len -= len;
 405                         data += len;
 406                         written += len;
 407                 }
 408                 /* This is a bit nasty but we need that for the console to
 409                  * flush when there aren't any interrupts. We will clean
 410                  * things a bit later to limit that to synchronous path
 411                  * such as the kernel console and xmon/udbg
 412                  */
 413                 do
 414                         opal_poll_events(&evt);
 415                 while(rc == OPAL_SUCCESS &&
 416                         (be64_to_cpu(evt) & OPAL_EVENT_CONSOLE_OUTPUT));
 417         }
 418         spin_unlock_irqrestore(&opal_write_lock, flags);
 419         return written;
 420 }
 421
 422 static int opal_recover_mce(struct pt_regs *regs,
 423                                         struct machine_check_event *evt)
 424 {
 425         int recovered = 0;
 426         uint64_t ea = get_mce_fault_addr(evt);
 427
 428         if (!(regs->msr & MSR_RI)) {
 429                 /* If MSR_RI isn't set, we cannot recover */
 430                 recovered = 0;
 431         } else if (evt->disposition == MCE_DISPOSITION_RECOVERED) {
 432                 /* Platform corrected itself */
 433                 recovered = 1;
 434         } else if (ea && !is_kernel_addr(ea)) {
 435                 /*
 436                  * Faulting address is not in kernel text. We should be fine.
 437                  * We need to find which process uses this address.
 438                  * For now, kill the task if we have received exception when
 439                  * in userspace.
 440                  *
 441                  * TODO: Queue up this address for hwpoisioning later.
 442                  */
 443                 if (user_mode(regs) && !is_global_init(current)) {
 444                         _exception(SIGBUS, regs, BUS_MCEERR_AR, regs->nip);
 445                         recovered = 1;
 446                 } else
 447                         recovered = 0;
 448         } else if (user_mode(regs) && !is_global_init(current) &&
 449                 evt->severity == MCE_SEV_ERROR_SYNC) {
 450                 /*
 451                  * If we have received a synchronous error when in userspace
 452                  * kill the task.
 453                  */
 454                 _exception(SIGBUS, regs, BUS_MCEERR_AR, regs->nip);
 455                 recovered = 1;
 456         }
 457         return recovered;
 458 }
 459
 460 int opal_machine_check(struct pt_regs *regs)
 461 {
 462         struct machine_check_event evt;
 463
 464         if (!get_mce_event(&evt, MCE_EVENT_RELEASE))
 465                 return 0;
 466
 467         /* Print things out */
 468         if (evt.version != MCE_V1) {
 469                 pr_err("Machine Check Exception, Unknown event version %d !\n",
 470                        evt.version);
 471                 return 0;
 472         }
 473         machine_check_print_event_info(&evt);
 474
 475         if (opal_recover_mce(regs, &evt))
 476                 return 1;
 477         return 0;
 478 }
 479
 480 static uint64_t find_recovery_address(uint64_t nip)
 481 {
 482         int i;
 483
 484         for (i = 0; i < mc_recoverable_range_len; i++)
 485                 if ((nip >= mc_recoverable_range[i].start_addr) &&
 486                     (nip < mc_recoverable_range[i].end_addr))
 487                     return mc_recoverable_range[i].recover_addr;
 488         return 0;
 489 }
 490
 491 bool opal_mce_check_early_recovery(struct pt_regs *regs)
 492 {
 493         uint64_t recover_addr = 0;
 494
 495         if (!opal.base || !opal.size)
 496                 goto out;
 497
 498         if ((regs->nip >= opal.base) &&
 499                         (regs->nip <= (opal.base + opal.size)))
 500                 recover_addr = find_recovery_address(regs->nip);
 501
 502         /*
 503          * Setup regs->nip to rfi into fixup address.
 504          */
 505         if (recover_addr)
 506                 regs->nip = recover_addr;
 507
 508 out:
 509         return !!recover_addr;
 510 }
 511
 512 static irqreturn_t opal_interrupt(int irq, void *data)
 513 {
 514         __be64 events;
 515
 516         opal_handle_interrupt(virq_to_hw(irq), &events);
 517
 518         opal_do_notifier(events);
 519
 520         return IRQ_HANDLED;
 521 }
 522
 523 static int opal_sysfs_init(void)
 524 {
 525         opal_kobj = kobject_create_and_add("opal", firmware_kobj);
 526         if (!opal_kobj) {
 527                 pr_warn("kobject_create_and_add opal failed\n");
 528                 return -ENOMEM;
 529         }
 530
 531         return 0;
 532 }
 533
 534 static int __init opal_init(void)
 535 {
 536         struct device_node *np, *consoles;
 537         const __be32 *irqs;
 538         int rc, i, irqlen;
 539
 540         opal_node = of_find_node_by_path("/ibm,opal");
 541         if (!opal_node) {
 542                 pr_warn("opal: Node not found\n");
 543                 return -ENODEV;
 544         }
 545
 546         /* Register OPAL consoles if any ports */
 547         if (firmware_has_feature(FW_FEATURE_OPALv2))
 548                 consoles = of_find_node_by_path("/ibm,opal/consoles");
 549         else
 550                 consoles = of_node_get(opal_node);
 551         if (consoles) {
 552                 for_each_child_of_node(consoles, np) {
 553                         if (strcmp(np->name, "serial"))
 554                                 continue;
 555                         of_platform_device_create(np, NULL, NULL);
 556                 }
 557                 of_node_put(consoles);
 558         }
 559
 560         /* Find all OPAL interrupts and request them */
 561         irqs = of_get_property(opal_node, "opal-interrupts", &irqlen);
 562         pr_debug("opal: Found %d interrupts reserved for OPAL\n",
 563                  irqs ? (irqlen / 4) : 0);
 564         opal_irq_count = irqlen / 4;
 565         opal_irqs = kzalloc(opal_irq_count * sizeof(unsigned int), GFP_KERNEL);
 566         for (i = 0; irqs && i < (irqlen / 4); i++, irqs++) {
 567                 unsigned int hwirq = be32_to_cpup(irqs);
 568                 unsigned int irq = irq_create_mapping(NULL, hwirq);
 569                 if (irq == NO_IRQ) {
 570                         pr_warning("opal: Failed to map irq 0x%x\n", hwirq);
 571                         continue;
 572                 }
 573                 rc = request_irq(irq, opal_interrupt, 0, "opal", NULL);
 574                 if (rc)
 575                         pr_warning("opal: Error %d requesting irq %d"
 576                                    " (0x%x)\n", rc, irq, hwirq);
 577                 opal_irqs[i] = irq;
 578         }
 579
 580         /* Create "opal" kobject under /sys/firmware */
 581         rc = opal_sysfs_init();
 582         if (rc == 0) {
 583                 /* Setup error log interface */
 584                 rc = opal_elog_init();
 585                 /* Setup code update interface */
 586                 opal_flash_init();
 587                 /* Setup platform dump extract interface */
 588                 opal_platform_dump_init();
 589                 /* Setup system parameters interface */
 590                 opal_sys_param_init();
 591         }
 592
 593         return 0;
 594 }
 595 subsys_initcall(opal_init);
 596
 597 void opal_shutdown(void)
 598 {
 599         unsigned int i;
 600         long rc = OPAL_BUSY;
 601
 602         /* First free interrupts, which will also mask them */
 603         for (i = 0; i < opal_irq_count; i++) {
 604                 if (opal_irqs[i])
 605                         free_irq(opal_irqs[i], NULL);
 606                 opal_irqs[i] = 0;
 607         }
 608
 609         /*
 610          * Then sync with OPAL which ensure anything that can
 611          * potentially write to our memory has completed such
 612          * as an ongoing dump retrieval
 613          */
 614         while (rc == OPAL_BUSY || rc == OPAL_BUSY_EVENT) {
 615                 rc = opal_sync_host_reboot();
 616                 if (rc == OPAL_BUSY)
 617                         opal_poll_events(NULL);
 618                 else
 619                         mdelay(10);
 620         }
 621 }