arch/powerpc/platforms/pseries/phyp_dump.c

   1 /*
   2  * Hypervisor-assisted dump
   3  *
   4  * Linas Vepstas, Manish Ahuja 2008
   5  * Copyright 2008 IBM Corp.
   6  *
   7  *      This program is free software; you can redistribute it and/or
   8  *      modify it under the terms of the GNU General Public License
   9  *      as published by the Free Software Foundation; either version
  10  *      2 of the License, or (at your option) any later version.
  11  *
  12  */
  13
  14 #include <linux/init.h>
  15 #include <linux/kobject.h>
  16 #include <linux/mm.h>
  17 #include <linux/of.h>
  18 #include <linux/pfn.h>
  19 #include <linux/swap.h>
  20 #include <linux/sysfs.h>
  21
  22 #include <asm/page.h>
  23 #include <asm/phyp_dump.h>
  24 #include <asm/machdep.h>
  25 #include <asm/prom.h>
  26 #include <asm/rtas.h>
  27
  28 /* Variables, used to communicate data between early boot and late boot */
  29 static struct phyp_dump phyp_dump_vars;
  30 struct phyp_dump *phyp_dump_info = &phyp_dump_vars;
  31
  32 static int ibm_configure_kernel_dump;
  33 /* ------------------------------------------------- */
  34 /* RTAS interfaces to declare the dump regions */
  35
  36 struct dump_section {
  37         u32 dump_flags;
  38         u16 source_type;
  39         u16 error_flags;
  40         u64 source_address;
  41         u64 source_length;
  42         u64 length_copied;
  43         u64 destination_address;
  44 };
  45
  46 struct phyp_dump_header {
  47         u32 version;
  48         u16 num_of_sections;
  49         u16 status;
  50
  51         u32 first_offset_section;
  52         u32 dump_disk_section;
  53         u64 block_num_dd;
  54         u64 num_of_blocks_dd;
  55         u32 offset_dd;
  56         u32 maxtime_to_auto;
  57         /* No dump disk path string used */
  58
  59         struct dump_section cpu_data;
  60         struct dump_section hpte_data;
  61         struct dump_section kernel_data;
  62 };
  63
  64 /* The dump header *must be* in low memory, so .bss it */
  65 static struct phyp_dump_header phdr;
  66
  67 #define NUM_DUMP_SECTIONS       3
  68 #define DUMP_HEADER_VERSION     0x1
  69 #define DUMP_REQUEST_FLAG       0x1
  70 #define DUMP_SOURCE_CPU         0x0001
  71 #define DUMP_SOURCE_HPTE        0x0002
  72 #define DUMP_SOURCE_RMO         0x0011
  73
  74 /**
  75  * init_dump_header() - initialize the header declaring a dump
  76  * Returns: length of dump save area.
  77  *
  78  * When the hypervisor saves crashed state, it needs to put
  79  * it somewhere. The dump header tells the hypervisor where
  80  * the data can be saved.
  81  */
  82 static unsigned long init_dump_header(struct phyp_dump_header *ph)
  83 {
  84         unsigned long addr_offset = 0;
  85
  86         /* Set up the dump header */
  87         ph->version = DUMP_HEADER_VERSION;
  88         ph->num_of_sections = NUM_DUMP_SECTIONS;
  89         ph->status = 0;
  90
  91         ph->first_offset_section =
  92                 (u32)offsetof(struct phyp_dump_header, cpu_data);
  93         ph->dump_disk_section = 0;
  94         ph->block_num_dd = 0;
  95         ph->num_of_blocks_dd = 0;
  96         ph->offset_dd = 0;
  97
  98         ph->maxtime_to_auto = 0; /* disabled */
  99
 100         /* The first two sections are mandatory */
 101         ph->cpu_data.dump_flags = DUMP_REQUEST_FLAG;
 102         ph->cpu_data.source_type = DUMP_SOURCE_CPU;
 103         ph->cpu_data.source_address = 0;
 104         ph->cpu_data.source_length = phyp_dump_info->cpu_state_size;
 105         ph->cpu_data.destination_address = addr_offset;
 106         addr_offset += phyp_dump_info->cpu_state_size;
 107
 108         ph->hpte_data.dump_flags = DUMP_REQUEST_FLAG;
 109         ph->hpte_data.source_type = DUMP_SOURCE_HPTE;
 110         ph->hpte_data.source_address = 0;
 111         ph->hpte_data.source_length = phyp_dump_info->hpte_region_size;
 112         ph->hpte_data.destination_address = addr_offset;
 113         addr_offset += phyp_dump_info->hpte_region_size;
 114
 115         /* This section describes the low kernel region */
 116         ph->kernel_data.dump_flags = DUMP_REQUEST_FLAG;
 117         ph->kernel_data.source_type = DUMP_SOURCE_RMO;
 118         ph->kernel_data.source_address = PHYP_DUMP_RMR_START;
 119         ph->kernel_data.source_length = PHYP_DUMP_RMR_END;
 120         ph->kernel_data.destination_address = addr_offset;
 121         addr_offset += ph->kernel_data.source_length;
 122
 123         return addr_offset;
 124 }
 125
 126 static void print_dump_header(const struct phyp_dump_header *ph)
 127 {
 128 #ifdef DEBUG
 129         printk(KERN_INFO "dump header:\n");
 130         /* setup some ph->sections required */
 131         printk(KERN_INFO "version = %d\n", ph->version);
 132         printk(KERN_INFO "Sections = %d\n", ph->num_of_sections);
 133         printk(KERN_INFO "Status = 0x%x\n", ph->status);
 134
 135         /* No ph->disk, so all should be set to 0 */
 136         printk(KERN_INFO "Offset to first section 0x%x\n",
 137                 ph->first_offset_section);
 138         printk(KERN_INFO "dump disk sections should be zero\n");
 139         printk(KERN_INFO "dump disk section = %d\n", ph->dump_disk_section);
 140         printk(KERN_INFO "block num = %ld\n", ph->block_num_dd);
 141         printk(KERN_INFO "number of blocks = %ld\n", ph->num_of_blocks_dd);
 142         printk(KERN_INFO "dump disk offset = %d\n", ph->offset_dd);
 143         printk(KERN_INFO "Max auto time= %d\n", ph->maxtime_to_auto);
 144
 145         /*set cpu state and hpte states as well scratch pad area */
 146         printk(KERN_INFO " CPU AREA \n");
 147         printk(KERN_INFO "cpu dump_flags =%d\n", ph->cpu_data.dump_flags);
 148         printk(KERN_INFO "cpu source_type =%d\n", ph->cpu_data.source_type);
 149         printk(KERN_INFO "cpu error_flags =%d\n", ph->cpu_data.error_flags);
 150         printk(KERN_INFO "cpu source_address =%lx\n",
 151                 ph->cpu_data.source_address);
 152         printk(KERN_INFO "cpu source_length =%lx\n",
 153                 ph->cpu_data.source_length);
 154         printk(KERN_INFO "cpu length_copied =%lx\n",
 155                 ph->cpu_data.length_copied);
 156
 157         printk(KERN_INFO " HPTE AREA \n");
 158         printk(KERN_INFO "HPTE dump_flags =%d\n", ph->hpte_data.dump_flags);
 159         printk(KERN_INFO "HPTE source_type =%d\n", ph->hpte_data.source_type);
 160         printk(KERN_INFO "HPTE error_flags =%d\n", ph->hpte_data.error_flags);
 161         printk(KERN_INFO "HPTE source_address =%lx\n",
 162                 ph->hpte_data.source_address);
 163         printk(KERN_INFO "HPTE source_length =%lx\n",
 164                 ph->hpte_data.source_length);
 165         printk(KERN_INFO "HPTE length_copied =%lx\n",
 166                 ph->hpte_data.length_copied);
 167
 168         printk(KERN_INFO " SRSD AREA \n");
 169         printk(KERN_INFO "SRSD dump_flags =%d\n", ph->kernel_data.dump_flags);
 170         printk(KERN_INFO "SRSD source_type =%d\n", ph->kernel_data.source_type);
 171         printk(KERN_INFO "SRSD error_flags =%d\n", ph->kernel_data.error_flags);
 172         printk(KERN_INFO "SRSD source_address =%lx\n",
 173                 ph->kernel_data.source_address);
 174         printk(KERN_INFO "SRSD source_length =%lx\n",
 175                 ph->kernel_data.source_length);
 176         printk(KERN_INFO "SRSD length_copied =%lx\n",
 177                 ph->kernel_data.length_copied);
 178 #endif
 179 }
 180
 181 static void register_dump_area(struct phyp_dump_header *ph, unsigned long addr)
 182 {
 183         int rc;
 184         ph->cpu_data.destination_address += addr;
 185         ph->hpte_data.destination_address += addr;
 186         ph->kernel_data.destination_address += addr;
 187
 188         do {
 189                 rc = rtas_call(ibm_configure_kernel_dump, 3, 1, NULL,
 190                                 1, ph, sizeof(struct phyp_dump_header));
 191         } while (rtas_busy_delay(rc));
 192
 193         if (rc) {
 194                 printk(KERN_ERR "phyp-dump: unexpected error (%d) on "
 195                                                 "register\n", rc);
 196                 print_dump_header(ph);
 197         }
 198 }
 199
 200 /* ------------------------------------------------- */
 201 /**
 202  * release_memory_range -- release memory previously lmb_reserved
 203  * @start_pfn: starting physical frame number
 204  * @nr_pages: number of pages to free.
 205  *
 206  * This routine will release memory that had been previously
 207  * lmb_reserved in early boot. The released memory becomes
 208  * available for genreal use.
 209  */
 210 static void
 211 release_memory_range(unsigned long start_pfn, unsigned long nr_pages)
 212 {
 213         struct page *rpage;
 214         unsigned long end_pfn;
 215         long i;
 216
 217         end_pfn = start_pfn + nr_pages;
 218
 219         for (i = start_pfn; i <= end_pfn; i++) {
 220                 rpage = pfn_to_page(i);
 221                 if (PageReserved(rpage)) {
 222                         ClearPageReserved(rpage);
 223                         init_page_count(rpage);
 224                         __free_page(rpage);
 225                         totalram_pages++;
 226                 }
 227         }
 228 }
 229
 230 /* ------------------------------------------------- */
 231 /**
 232  * sysfs_release_region -- sysfs interface to release memory range.
 233  *
 234  * Usage:
 235  *   "echo <start addr> <length> > /sys/kernel/release_region"
 236  *
 237  * Example:
 238  *   "echo 0x40000000 0x10000000 > /sys/kernel/release_region"
 239  *
 240  * will release 256MB starting at 1GB.
 241  */
 242 static ssize_t store_release_region(struct kobject *kobj,
 243                                 struct kobj_attribute *attr,
 244                                 const char *buf, size_t count)
 245 {
 246         unsigned long start_addr, length, end_addr;
 247         unsigned long start_pfn, nr_pages;
 248         ssize_t ret;
 249
 250         ret = sscanf(buf, "%lx %lx", &start_addr, &length);
 251         if (ret != 2)
 252                 return -EINVAL;
 253
 254         /* Range-check - don't free any reserved memory that
 255          * wasn't reserved for phyp-dump */
 256         if (start_addr < phyp_dump_info->init_reserve_start)
 257                 start_addr = phyp_dump_info->init_reserve_start;
 258
 259         end_addr = phyp_dump_info->init_reserve_start +
 260                         phyp_dump_info->init_reserve_size;
 261         if (start_addr+length > end_addr)
 262                 length = end_addr - start_addr;
 263
 264         /* Release the region of memory assed in by user */
 265         start_pfn = PFN_DOWN(start_addr);
 266         nr_pages = PFN_DOWN(length);
 267         release_memory_range(start_pfn, nr_pages);
 268
 269         return count;
 270 }
 271
 272 static struct kobj_attribute rr = __ATTR(release_region, 0600,
 273                                          NULL, store_release_region);
 274
 275 static int __init phyp_dump_setup(void)
 276 {
 277         struct device_node *rtas;
 278         const struct phyp_dump_header *dump_header = NULL;
 279         unsigned long dump_area_start;
 280         unsigned long dump_area_length;
 281         int header_len = 0;
 282         int rc;
 283
 284         /* If no memory was reserved in early boot, there is nothing to do */
 285         if (phyp_dump_info->init_reserve_size == 0)
 286                 return 0;
 287
 288         /* Return if phyp dump not supported */
 289         if (!phyp_dump_info->phyp_dump_configured)
 290                 return -ENOSYS;
 291
 292         /* Is there dump data waiting for us? If there isn't,
 293          * then register a new dump area, and release all of
 294          * the rest of the reserved ram.
 295          *
 296          * The /rtas/ibm,kernel-dump rtas node is present only
 297          * if there is dump data waiting for us.
 298          */
 299         rtas = of_find_node_by_path("/rtas");
 300         if (rtas) {
 301                 dump_header = of_get_property(rtas, "ibm,kernel-dump",
 302                                                 &header_len);
 303                 of_node_put(rtas);
 304         }
 305
 306         print_dump_header(dump_header);
 307         dump_area_length = init_dump_header(&phdr);
 308         /* align down */
 309         dump_area_start = phyp_dump_info->init_reserve_start & PAGE_MASK;
 310
 311         if (dump_header == NULL) {
 312                 register_dump_area(&phdr, dump_area_start);
 313                 return 0;
 314         }
 315
 316         /* Should we create a dump_subsys, analogous to s390/ipl.c ? */
 317         rc = sysfs_create_file(kernel_kobj, &rr.attr);
 318         if (rc)
 319                 printk(KERN_ERR "phyp-dump: unable to create sysfs file (%d)\n",
 320                                                                         rc);
 321
 322         /* ToDo: re-register the dump area, for next time. */
 323         return 0;
 324 }
 325 machine_subsys_initcall(pseries, phyp_dump_setup);
 326
 327 int __init early_init_dt_scan_phyp_dump(unsigned long node,
 328                 const char *uname, int depth, void *data)
 329 {
 330         const unsigned int *sizes;
 331
 332         phyp_dump_info->phyp_dump_configured = 0;
 333         phyp_dump_info->phyp_dump_is_active = 0;
 334
 335         if (depth != 1 || strcmp(uname, "rtas") != 0)
 336                 return 0;
 337
 338         if (of_get_flat_dt_prop(node, "ibm,configure-kernel-dump", NULL))
 339                 phyp_dump_info->phyp_dump_configured++;
 340
 341         if (of_get_flat_dt_prop(node, "ibm,dump-kernel", NULL))
 342                 phyp_dump_info->phyp_dump_is_active++;
 343
 344         sizes = of_get_flat_dt_prop(node, "ibm,configure-kernel-dump-sizes",
 345                                     NULL);
 346         if (!sizes)
 347                 return 0;
 348
 349         if (sizes[0] == 1)
 350                 phyp_dump_info->cpu_state_size = *((unsigned long *)&sizes[1]);
 351
 352         if (sizes[3] == 2)
 353                 phyp_dump_info->hpte_region_size =
 354                                                 *((unsigned long *)&sizes[4]);
 355         return 1;
 356 }