Pileus Git - ~andy/linux/blob - arch/powerpc/kernel/fadump.c

   1 /*
   2  * Firmware Assisted dump: A robust mechanism to get reliable kernel crash
   3  * dump with assistance from firmware. This approach does not use kexec,
   4  * instead firmware assists in booting the kdump kernel while preserving
   5  * memory contents. The most of the code implementation has been adapted
   6  * from phyp assisted dump implementation written by Linas Vepstas and
   7  * Manish Ahuja
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  22  *
  23  * Copyright 2011 IBM Corporation
  24  * Author: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
  25  */
  26
  27 #undef DEBUG
  28 #define pr_fmt(fmt) "fadump: " fmt
  29
  30 #include <linux/string.h>
  31 #include <linux/memblock.h>
  32 #include <linux/delay.h>
  33 #include <linux/debugfs.h>
  34 #include <linux/seq_file.h>
  35
  36 #include <asm/page.h>
  37 #include <asm/prom.h>
  38 #include <asm/rtas.h>
  39 #include <asm/fadump.h>
  40
  41 static struct fw_dump fw_dump;
  42 static struct fadump_mem_struct fdm;
  43 static const struct fadump_mem_struct *fdm_active;
  44
  45 static DEFINE_MUTEX(fadump_mutex);
  46
  47 /* Scan the Firmware Assisted dump configuration details. */
  48 int __init early_init_dt_scan_fw_dump(unsigned long node,
  49                         const char *uname, int depth, void *data)
  50 {
  51         __be32 *sections;
  52         int i, num_sections;
  53         unsigned long size;
  54         const int *token;
  55
  56         if (depth != 1 || strcmp(uname, "rtas") != 0)
  57                 return 0;
  58
  59         /*
  60          * Check if Firmware Assisted dump is supported. if yes, check
  61          * if dump has been initiated on last reboot.
  62          */
  63         token = of_get_flat_dt_prop(node, "ibm,configure-kernel-dump", NULL);
  64         if (!token)
  65                 return 0;
  66
  67         fw_dump.fadump_supported = 1;
  68         fw_dump.ibm_configure_kernel_dump = *token;
  69
  70         /*
  71          * The 'ibm,kernel-dump' rtas node is present only if there is
  72          * dump data waiting for us.
  73          */
  74         fdm_active = of_get_flat_dt_prop(node, "ibm,kernel-dump", NULL);
  75         if (fdm_active)
  76                 fw_dump.dump_active = 1;
  77
  78         /* Get the sizes required to store dump data for the firmware provided
  79          * dump sections.
  80          * For each dump section type supported, a 32bit cell which defines
  81          * the ID of a supported section followed by two 32 bit cells which
  82          * gives teh size of the section in bytes.
  83          */
  84         sections = of_get_flat_dt_prop(node, "ibm,configure-kernel-dump-sizes",
  85                                         &size);
  86
  87         if (!sections)
  88                 return 0;
  89
  90         num_sections = size / (3 * sizeof(u32));
  91
  92         for (i = 0; i < num_sections; i++, sections += 3) {
  93                 u32 type = (u32)of_read_number(sections, 1);
  94
  95                 switch (type) {
  96                 case FADUMP_CPU_STATE_DATA:
  97                         fw_dump.cpu_state_data_size =
  98                                         of_read_ulong(&sections[1], 2);
  99                         break;
 100                 case FADUMP_HPTE_REGION:
 101                         fw_dump.hpte_region_size =
 102                                         of_read_ulong(&sections[1], 2);
 103                         break;
 104                 }
 105         }
 106         return 1;
 107 }
 108
 109 int is_fadump_active(void)
 110 {
 111         return fw_dump.dump_active;
 112 }
 113
 114 /* Print firmware assisted dump configurations for debugging purpose. */
 115 static void fadump_show_config(void)
 116 {
 117         pr_debug("Support for firmware-assisted dump (fadump): %s\n",
 118                         (fw_dump.fadump_supported ? "present" : "no support"));
 119
 120         if (!fw_dump.fadump_supported)
 121                 return;
 122
 123         pr_debug("Fadump enabled    : %s\n",
 124                                 (fw_dump.fadump_enabled ? "yes" : "no"));
 125         pr_debug("Dump Active       : %s\n",
 126                                 (fw_dump.dump_active ? "yes" : "no"));
 127         pr_debug("Dump section sizes:\n");
 128         pr_debug("    CPU state data size: %lx\n", fw_dump.cpu_state_data_size);
 129         pr_debug("    HPTE region size   : %lx\n", fw_dump.hpte_region_size);
 130         pr_debug("Boot memory size  : %lx\n", fw_dump.boot_memory_size);
 131 }
 132
 133 static unsigned long init_fadump_mem_struct(struct fadump_mem_struct *fdm,
 134                                 unsigned long addr)
 135 {
 136         if (!fdm)
 137                 return 0;
 138
 139         memset(fdm, 0, sizeof(struct fadump_mem_struct));
 140         addr = addr & PAGE_MASK;
 141
 142         fdm->header.dump_format_version = 0x00000001;
 143         fdm->header.dump_num_sections = 3;
 144         fdm->header.dump_status_flag = 0;
 145         fdm->header.offset_first_dump_section =
 146                 (u32)offsetof(struct fadump_mem_struct, cpu_state_data);
 147
 148         /*
 149          * Fields for disk dump option.
 150          * We are not using disk dump option, hence set these fields to 0.
 151          */
 152         fdm->header.dd_block_size = 0;
 153         fdm->header.dd_block_offset = 0;
 154         fdm->header.dd_num_blocks = 0;
 155         fdm->header.dd_offset_disk_path = 0;
 156
 157         /* set 0 to disable an automatic dump-reboot. */
 158         fdm->header.max_time_auto = 0;
 159
 160         /* Kernel dump sections */
 161         /* cpu state data section. */
 162         fdm->cpu_state_data.request_flag = FADUMP_REQUEST_FLAG;
 163         fdm->cpu_state_data.source_data_type = FADUMP_CPU_STATE_DATA;
 164         fdm->cpu_state_data.source_address = 0;
 165         fdm->cpu_state_data.source_len = fw_dump.cpu_state_data_size;
 166         fdm->cpu_state_data.destination_address = addr;
 167         addr += fw_dump.cpu_state_data_size;
 168
 169         /* hpte region section */
 170         fdm->hpte_region.request_flag = FADUMP_REQUEST_FLAG;
 171         fdm->hpte_region.source_data_type = FADUMP_HPTE_REGION;
 172         fdm->hpte_region.source_address = 0;
 173         fdm->hpte_region.source_len = fw_dump.hpte_region_size;
 174         fdm->hpte_region.destination_address = addr;
 175         addr += fw_dump.hpte_region_size;
 176
 177         /* RMA region section */
 178         fdm->rmr_region.request_flag = FADUMP_REQUEST_FLAG;
 179         fdm->rmr_region.source_data_type = FADUMP_REAL_MODE_REGION;
 180         fdm->rmr_region.source_address = RMA_START;
 181         fdm->rmr_region.source_len = fw_dump.boot_memory_size;
 182         fdm->rmr_region.destination_address = addr;
 183         addr += fw_dump.boot_memory_size;
 184
 185         return addr;
 186 }
 187
 188 /**
 189  * fadump_calculate_reserve_size(): reserve variable boot area 5% of System RAM
 190  *
 191  * Function to find the largest memory size we need to reserve during early
 192  * boot process. This will be the size of the memory that is required for a
 193  * kernel to boot successfully.
 194  *
 195  * This function has been taken from phyp-assisted dump feature implementation.
 196  *
 197  * returns larger of 256MB or 5% rounded down to multiples of 256MB.
 198  *
 199  * TODO: Come up with better approach to find out more accurate memory size
 200  * that is required for a kernel to boot successfully.
 201  *
 202  */
 203 static inline unsigned long fadump_calculate_reserve_size(void)
 204 {
 205         unsigned long size;
 206
 207         /*
 208          * Check if the size is specified through fadump_reserve_mem= cmdline
 209          * option. If yes, then use that.
 210          */
 211         if (fw_dump.reserve_bootvar)
 212                 return fw_dump.reserve_bootvar;
 213
 214         /* divide by 20 to get 5% of value */
 215         size = memblock_end_of_DRAM() / 20;
 216
 217         /* round it down in multiples of 256 */
 218         size = size & ~0x0FFFFFFFUL;
 219
 220         /* Truncate to memory_limit. We don't want to over reserve the memory.*/
 221         if (memory_limit && size > memory_limit)
 222                 size = memory_limit;
 223
 224         return (size > MIN_BOOT_MEM ? size : MIN_BOOT_MEM);
 225 }
 226
 227 /*
 228  * Calculate the total memory size required to be reserved for
 229  * firmware-assisted dump registration.
 230  */
 231 static unsigned long get_fadump_area_size(void)
 232 {
 233         unsigned long size = 0;
 234
 235         size += fw_dump.cpu_state_data_size;
 236         size += fw_dump.hpte_region_size;
 237         size += fw_dump.boot_memory_size;
 238
 239         size = PAGE_ALIGN(size);
 240         return size;
 241 }
 242
 243 int __init fadump_reserve_mem(void)
 244 {
 245         unsigned long base, size, memory_boundary;
 246
 247         if (!fw_dump.fadump_enabled)
 248                 return 0;
 249
 250         if (!fw_dump.fadump_supported) {
 251                 printk(KERN_INFO "Firmware-assisted dump is not supported on"
 252                                 " this hardware\n");
 253                 fw_dump.fadump_enabled = 0;
 254                 return 0;
 255         }
 256         /*
 257          * Initialize boot memory size
 258          * If dump is active then we have already calculated the size during
 259          * first kernel.
 260          */
 261         if (fdm_active)
 262                 fw_dump.boot_memory_size = fdm_active->rmr_region.source_len;
 263         else
 264                 fw_dump.boot_memory_size = fadump_calculate_reserve_size();
 265
 266         /*
 267          * Calculate the memory boundary.
 268          * If memory_limit is less than actual memory boundary then reserve
 269          * the memory for fadump beyond the memory_limit and adjust the
 270          * memory_limit accordingly, so that the running kernel can run with
 271          * specified memory_limit.
 272          */
 273         if (memory_limit && memory_limit < memblock_end_of_DRAM()) {
 274                 size = get_fadump_area_size();
 275                 if ((memory_limit + size) < memblock_end_of_DRAM())
 276                         memory_limit += size;
 277                 else
 278                         memory_limit = memblock_end_of_DRAM();
 279                 printk(KERN_INFO "Adjusted memory_limit for firmware-assisted"
 280                                 " dump, now %#016llx\n",
 281                                 (unsigned long long)memory_limit);
 282         }
 283         if (memory_limit)
 284                 memory_boundary = memory_limit;
 285         else
 286                 memory_boundary = memblock_end_of_DRAM();
 287
 288         if (fw_dump.dump_active) {
 289                 printk(KERN_INFO "Firmware-assisted dump is active.\n");
 290                 /*
 291                  * If last boot has crashed then reserve all the memory
 292                  * above boot_memory_size so that we don't touch it until
 293                  * dump is written to disk by userspace tool. This memory
 294                  * will be released for general use once the dump is saved.
 295                  */
 296                 base = fw_dump.boot_memory_size;
 297                 size = memory_boundary - base;
 298                 memblock_reserve(base, size);
 299                 printk(KERN_INFO "Reserved %ldMB of memory at %ldMB "
 300                                 "for saving crash dump\n",
 301                                 (unsigned long)(size >> 20),
 302                                 (unsigned long)(base >> 20));
 303         } else {
 304                 /* Reserve the memory at the top of memory. */
 305                 size = get_fadump_area_size();
 306                 base = memory_boundary - size;
 307                 memblock_reserve(base, size);
 308                 printk(KERN_INFO "Reserved %ldMB of memory at %ldMB "
 309                                 "for firmware-assisted dump\n",
 310                                 (unsigned long)(size >> 20),
 311                                 (unsigned long)(base >> 20));
 312         }
 313         fw_dump.reserve_dump_area_start = base;
 314         fw_dump.reserve_dump_area_size = size;
 315         return 1;
 316 }
 317
 318 /* Look for fadump= cmdline option. */
 319 static int __init early_fadump_param(char *p)
 320 {
 321         if (!p)
 322                 return 1;
 323
 324         if (strncmp(p, "on", 2) == 0)
 325                 fw_dump.fadump_enabled = 1;
 326         else if (strncmp(p, "off", 3) == 0)
 327                 fw_dump.fadump_enabled = 0;
 328
 329         return 0;
 330 }
 331 early_param("fadump", early_fadump_param);
 332
 333 /* Look for fadump_reserve_mem= cmdline option */
 334 static int __init early_fadump_reserve_mem(char *p)
 335 {
 336         if (p)
 337                 fw_dump.reserve_bootvar = memparse(p, &p);
 338         return 0;
 339 }
 340 early_param("fadump_reserve_mem", early_fadump_reserve_mem);
 341
 342 static void register_fw_dump(struct fadump_mem_struct *fdm)
 343 {
 344         int rc;
 345         unsigned int wait_time;
 346
 347         pr_debug("Registering for firmware-assisted kernel dump...\n");
 348
 349         /* TODO: Add upper time limit for the delay */
 350         do {
 351                 rc = rtas_call(fw_dump.ibm_configure_kernel_dump, 3, 1, NULL,
 352                         FADUMP_REGISTER, fdm,
 353                         sizeof(struct fadump_mem_struct));
 354
 355                 wait_time = rtas_busy_delay_time(rc);
 356                 if (wait_time)
 357                         mdelay(wait_time);
 358
 359         } while (wait_time);
 360
 361         switch (rc) {
 362         case -1:
 363                 printk(KERN_ERR "Failed to register firmware-assisted kernel"
 364                         " dump. Hardware Error(%d).\n", rc);
 365                 break;
 366         case -3:
 367                 printk(KERN_ERR "Failed to register firmware-assisted kernel"
 368                         " dump. Parameter Error(%d).\n", rc);
 369                 break;
 370         case -9:
 371                 printk(KERN_ERR "firmware-assisted kernel dump is already "
 372                         " registered.");
 373                 fw_dump.dump_registered = 1;
 374                 break;
 375         case 0:
 376                 printk(KERN_INFO "firmware-assisted kernel dump registration"
 377                         " is successful\n");
 378                 fw_dump.dump_registered = 1;
 379                 break;
 380         }
 381 }
 382
 383 static void register_fadump(void)
 384 {
 385         /*
 386          * If no memory is reserved then we can not register for firmware-
 387          * assisted dump.
 388          */
 389         if (!fw_dump.reserve_dump_area_size)
 390                 return;
 391
 392         /* register the future kernel dump with firmware. */
 393         register_fw_dump(&fdm);
 394 }
 395
 396 static int fadump_unregister_dump(struct fadump_mem_struct *fdm)
 397 {
 398         int rc = 0;
 399         unsigned int wait_time;
 400
 401         pr_debug("Un-register firmware-assisted dump\n");
 402
 403         /* TODO: Add upper time limit for the delay */
 404         do {
 405                 rc = rtas_call(fw_dump.ibm_configure_kernel_dump, 3, 1, NULL,
 406                         FADUMP_UNREGISTER, fdm,
 407                         sizeof(struct fadump_mem_struct));
 408
 409                 wait_time = rtas_busy_delay_time(rc);
 410                 if (wait_time)
 411                         mdelay(wait_time);
 412         } while (wait_time);
 413
 414         if (rc) {
 415                 printk(KERN_ERR "Failed to un-register firmware-assisted dump."
 416                         " unexpected error(%d).\n", rc);
 417                 return rc;
 418         }
 419         fw_dump.dump_registered = 0;
 420         return 0;
 421 }
 422
 423 static ssize_t fadump_enabled_show(struct kobject *kobj,
 424                                         struct kobj_attribute *attr,
 425                                         char *buf)
 426 {
 427         return sprintf(buf, "%d\n", fw_dump.fadump_enabled);
 428 }
 429
 430 static ssize_t fadump_register_show(struct kobject *kobj,
 431                                         struct kobj_attribute *attr,
 432                                         char *buf)
 433 {
 434         return sprintf(buf, "%d\n", fw_dump.dump_registered);
 435 }
 436
 437 static ssize_t fadump_register_store(struct kobject *kobj,
 438                                         struct kobj_attribute *attr,
 439                                         const char *buf, size_t count)
 440 {
 441         int ret = 0;
 442
 443         if (!fw_dump.fadump_enabled || fdm_active)
 444                 return -EPERM;
 445
 446         mutex_lock(&fadump_mutex);
 447
 448         switch (buf[0]) {
 449         case '0':
 450                 if (fw_dump.dump_registered == 0) {
 451                         ret = -EINVAL;
 452                         goto unlock_out;
 453                 }
 454                 /* Un-register Firmware-assisted dump */
 455                 fadump_unregister_dump(&fdm);
 456                 break;
 457         case '1':
 458                 if (fw_dump.dump_registered == 1) {
 459                         ret = -EINVAL;
 460                         goto unlock_out;
 461                 }
 462                 /* Register Firmware-assisted dump */
 463                 register_fadump();
 464                 break;
 465         default:
 466                 ret = -EINVAL;
 467                 break;
 468         }
 469
 470 unlock_out:
 471         mutex_unlock(&fadump_mutex);
 472         return ret < 0 ? ret : count;
 473 }
 474
 475 static int fadump_region_show(struct seq_file *m, void *private)
 476 {
 477         const struct fadump_mem_struct *fdm_ptr;
 478
 479         if (!fw_dump.fadump_enabled)
 480                 return 0;
 481
 482         if (fdm_active)
 483                 fdm_ptr = fdm_active;
 484         else
 485                 fdm_ptr = &fdm;
 486
 487         seq_printf(m,
 488                         "CPU : [%#016llx-%#016llx] %#llx bytes, "
 489                         "Dumped: %#llx\n",
 490                         fdm_ptr->cpu_state_data.destination_address,
 491                         fdm_ptr->cpu_state_data.destination_address +
 492                         fdm_ptr->cpu_state_data.source_len - 1,
 493                         fdm_ptr->cpu_state_data.source_len,
 494                         fdm_ptr->cpu_state_data.bytes_dumped);
 495         seq_printf(m,
 496                         "HPTE: [%#016llx-%#016llx] %#llx bytes, "
 497                         "Dumped: %#llx\n",
 498                         fdm_ptr->hpte_region.destination_address,
 499                         fdm_ptr->hpte_region.destination_address +
 500                         fdm_ptr->hpte_region.source_len - 1,
 501                         fdm_ptr->hpte_region.source_len,
 502                         fdm_ptr->hpte_region.bytes_dumped);
 503         seq_printf(m,
 504                         "DUMP: [%#016llx-%#016llx] %#llx bytes, "
 505                         "Dumped: %#llx\n",
 506                         fdm_ptr->rmr_region.destination_address,
 507                         fdm_ptr->rmr_region.destination_address +
 508                         fdm_ptr->rmr_region.source_len - 1,
 509                         fdm_ptr->rmr_region.source_len,
 510                         fdm_ptr->rmr_region.bytes_dumped);
 511
 512         if (!fdm_active ||
 513                 (fw_dump.reserve_dump_area_start ==
 514                 fdm_ptr->cpu_state_data.destination_address))
 515                 return 0;
 516
 517         /* Dump is active. Show reserved memory region. */
 518         seq_printf(m,
 519                         "    : [%#016llx-%#016llx] %#llx bytes, "
 520                         "Dumped: %#llx\n",
 521                         (unsigned long long)fw_dump.reserve_dump_area_start,
 522                         fdm_ptr->cpu_state_data.destination_address - 1,
 523                         fdm_ptr->cpu_state_data.destination_address -
 524                         fw_dump.reserve_dump_area_start,
 525                         fdm_ptr->cpu_state_data.destination_address -
 526                         fw_dump.reserve_dump_area_start);
 527         return 0;
 528 }
 529
 530 static struct kobj_attribute fadump_attr = __ATTR(fadump_enabled,
 531                                                 0444, fadump_enabled_show,
 532                                                 NULL);
 533 static struct kobj_attribute fadump_register_attr = __ATTR(fadump_registered,
 534                                                 0644, fadump_register_show,
 535                                                 fadump_register_store);
 536
 537 static int fadump_region_open(struct inode *inode, struct file *file)
 538 {
 539         return single_open(file, fadump_region_show, inode->i_private);
 540 }
 541
 542 static const struct file_operations fadump_region_fops = {
 543         .open    = fadump_region_open,
 544         .read    = seq_read,
 545         .llseek  = seq_lseek,
 546         .release = single_release,
 547 };
 548
 549 static void fadump_init_files(void)
 550 {
 551         struct dentry *debugfs_file;
 552         int rc = 0;
 553
 554         rc = sysfs_create_file(kernel_kobj, &fadump_attr.attr);
 555         if (rc)
 556                 printk(KERN_ERR "fadump: unable to create sysfs file"
 557                         " fadump_enabled (%d)\n", rc);
 558
 559         rc = sysfs_create_file(kernel_kobj, &fadump_register_attr.attr);
 560         if (rc)
 561                 printk(KERN_ERR "fadump: unable to create sysfs file"
 562                         " fadump_registered (%d)\n", rc);
 563
 564         debugfs_file = debugfs_create_file("fadump_region", 0444,
 565                                         powerpc_debugfs_root, NULL,
 566                                         &fadump_region_fops);
 567         if (!debugfs_file)
 568                 printk(KERN_ERR "fadump: unable to create debugfs file"
 569                                 " fadump_region\n");
 570         return;
 571 }
 572
 573 /*
 574  * Prepare for firmware-assisted dump.
 575  */
 576 int __init setup_fadump(void)
 577 {
 578         if (!fw_dump.fadump_enabled)
 579                 return 0;
 580
 581         if (!fw_dump.fadump_supported) {
 582                 printk(KERN_ERR "Firmware-assisted dump is not supported on"
 583                         " this hardware\n");
 584                 return 0;
 585         }
 586
 587         fadump_show_config();
 588         /* Initialize the kernel dump memory structure for FAD registration. */
 589         if (fw_dump.reserve_dump_area_size)
 590                 init_fadump_mem_struct(&fdm, fw_dump.reserve_dump_area_start);
 591         fadump_init_files();
 592
 593         return 1;
 594 }
 595 subsys_initcall(setup_fadump);