drivers/base/memory.c

   1 /*
   2  * drivers/base/memory.c - basic Memory class support
   3  *
   4  * Written by Matt Tolentino <matthew.e.tolentino@intel.com>
   5  *            Dave Hansen <haveblue@us.ibm.com>
   6  *
   7  * This file provides the necessary infrastructure to represent
   8  * a SPARSEMEM-memory-model system's physical memory in /sysfs.
   9  * All arch-independent code that assumes MEMORY_HOTPLUG requires
  10  * SPARSEMEM should be contained here, or in mm/memory_hotplug.c.
  11  */
  12
  13 #include <linux/sysdev.h>
  14 #include <linux/module.h>
  15 #include <linux/init.h>
  16 #include <linux/sched.h>        /* capable() */
  17 #include <linux/topology.h>
  18 #include <linux/device.h>
  19 #include <linux/memory.h>
  20 #include <linux/kobject.h>
  21 #include <linux/memory_hotplug.h>
  22 #include <linux/mm.h>
  23 #include <asm/atomic.h>
  24 #include <asm/uaccess.h>
  25
  26 #define MEMORY_CLASS_NAME       "memory"
  27
  28 static struct sysdev_class memory_sysdev_class = {
  29         set_kset_name(MEMORY_CLASS_NAME),
  30 };
  31 EXPORT_SYMBOL(memory_sysdev_class);
  32
  33 static char *memory_hotplug_name(struct kset *kset, struct kobject *kobj)
  34 {
  35         return MEMORY_CLASS_NAME;
  36 }
  37
  38 static int memory_hotplug(struct kset *kset, struct kobject *kobj, char **envp,
  39                         int num_envp, char *buffer, int buffer_size)
  40 {
  41         int retval = 0;
  42
  43         return retval;
  44 }
  45
  46 static struct kset_hotplug_ops memory_hotplug_ops = {
  47         .name           = memory_hotplug_name,
  48         .hotplug        = memory_hotplug,
  49 };
  50
  51 static struct notifier_block *memory_chain;
  52
  53 static int register_memory_notifier(struct notifier_block *nb)
  54 {
  55         return notifier_chain_register(&memory_chain, nb);
  56 }
  57
  58 static void unregister_memory_notifier(struct notifier_block *nb)
  59 {
  60         notifier_chain_unregister(&memory_chain, nb);
  61 }
  62
  63 /*
  64  * register_memory - Setup a sysfs device for a memory block
  65  */
  66 static int
  67 register_memory(struct memory_block *memory, struct mem_section *section,
  68                 struct node *root)
  69 {
  70         int error;
  71
  72         memory->sysdev.cls = &memory_sysdev_class;
  73         memory->sysdev.id = __section_nr(section);
  74
  75         error = sysdev_register(&memory->sysdev);
  76
  77         if (root && !error)
  78                 error = sysfs_create_link(&root->sysdev.kobj,
  79                                           &memory->sysdev.kobj,
  80                                           kobject_name(&memory->sysdev.kobj));
  81
  82         return error;
  83 }
  84
  85 static void
  86 unregister_memory(struct memory_block *memory, struct mem_section *section,
  87                 struct node *root)
  88 {
  89         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
  90         BUG_ON(memory->sysdev.id != __section_nr(section));
  91
  92         sysdev_unregister(&memory->sysdev);
  93         if (root)
  94                 sysfs_remove_link(&root->sysdev.kobj,
  95                                   kobject_name(&memory->sysdev.kobj));
  96 }
  97
  98 /*
  99  * use this as the physical section index that this memsection
 100  * uses.
 101  */
 102
 103 static ssize_t show_mem_phys_index(struct sys_device *dev, char *buf)
 104 {
 105         struct memory_block *mem =
 106                 container_of(dev, struct memory_block, sysdev);
 107         return sprintf(buf, "%08lx\n", mem->phys_index);
 108 }
 109
 110 /*
 111  * online, offline, going offline, etc.
 112  */
 113 static ssize_t show_mem_state(struct sys_device *dev, char *buf)
 114 {
 115         struct memory_block *mem =
 116                 container_of(dev, struct memory_block, sysdev);
 117         ssize_t len = 0;
 118
 119         /*
 120          * We can probably put these states in a nice little array
 121          * so that they're not open-coded
 122          */
 123         switch (mem->state) {
 124                 case MEM_ONLINE:
 125                         len = sprintf(buf, "online\n");
 126                         break;
 127                 case MEM_OFFLINE:
 128                         len = sprintf(buf, "offline\n");
 129                         break;
 130                 case MEM_GOING_OFFLINE:
 131                         len = sprintf(buf, "going-offline\n");
 132                         break;
 133                 default:
 134                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 135                                         mem->state);
 136                         WARN_ON(1);
 137                         break;
 138         }
 139
 140         return len;
 141 }
 142
 143 static inline int memory_notify(unsigned long val, void *v)
 144 {
 145         return notifier_call_chain(&memory_chain, val, v);
 146 }
 147
 148 /*
 149  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 150  * OK to have direct references to sparsemem variables in here.
 151  */
 152 static int
 153 memory_block_action(struct memory_block *mem, unsigned long action)
 154 {
 155         int i;
 156         unsigned long psection;
 157         unsigned long start_pfn, start_paddr;
 158         struct page *first_page;
 159         int ret;
 160         int old_state = mem->state;
 161
 162         psection = mem->phys_index;
 163         first_page = pfn_to_page(psection << PFN_SECTION_SHIFT);
 164
 165         /*
 166          * The probe routines leave the pages reserved, just
 167          * as the bootmem code does.  Make sure they're still
 168          * that way.
 169          */
 170         if (action == MEM_ONLINE) {
 171                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 172                         if (PageReserved(first_page+i))
 173                                 continue;
 174
 175                         printk(KERN_WARNING "section number %ld page number %d "
 176                                 "not reserved, was it already online? \n",
 177                                 psection, i);
 178                         return -EBUSY;
 179                 }
 180         }
 181
 182         switch (action) {
 183                 case MEM_ONLINE:
 184                         start_pfn = page_to_pfn(first_page);
 185                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 186                         break;
 187                 case MEM_OFFLINE:
 188                         mem->state = MEM_GOING_OFFLINE;
 189                         memory_notify(MEM_GOING_OFFLINE, NULL);
 190                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 191                         ret = remove_memory(start_paddr,
 192                                             PAGES_PER_SECTION << PAGE_SHIFT);
 193                         if (ret) {
 194                                 mem->state = old_state;
 195                                 break;
 196                         }
 197                         memory_notify(MEM_MAPPING_INVALID, NULL);
 198                         break;
 199                 default:
 200                         printk(KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 201                                         __FUNCTION__, mem, action, action);
 202                         WARN_ON(1);
 203                         ret = -EINVAL;
 204         }
 205         /*
 206          * For now, only notify on successful memory operations
 207          */
 208         if (!ret)
 209                 memory_notify(action, NULL);
 210
 211         return ret;
 212 }
 213
 214 static int memory_block_change_state(struct memory_block *mem,
 215                 unsigned long to_state, unsigned long from_state_req)
 216 {
 217         int ret = 0;
 218         down(&mem->state_sem);
 219
 220         if (mem->state != from_state_req) {
 221                 ret = -EINVAL;
 222                 goto out;
 223         }
 224
 225         ret = memory_block_action(mem, to_state);
 226         if (!ret)
 227                 mem->state = to_state;
 228
 229 out:
 230         up(&mem->state_sem);
 231         return ret;
 232 }
 233
 234 static ssize_t
 235 store_mem_state(struct sys_device *dev, const char *buf, size_t count)
 236 {
 237         struct memory_block *mem;
 238         unsigned int phys_section_nr;
 239         int ret = -EINVAL;
 240
 241         mem = container_of(dev, struct memory_block, sysdev);
 242         phys_section_nr = mem->phys_index;
 243
 244         if (!valid_section_nr(phys_section_nr))
 245                 goto out;
 246
 247         if (!strncmp(buf, "online", min((int)count, 6)))
 248                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 249         else if(!strncmp(buf, "offline", min((int)count, 7)))
 250                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 251 out:
 252         if (ret)
 253                 return ret;
 254         return count;
 255 }
 256
 257 /*
 258  * phys_device is a bad name for this.  What I really want
 259  * is a way to differentiate between memory ranges that
 260  * are part of physical devices that constitute
 261  * a complete removable unit or fru.
 262  * i.e. do these ranges belong to the same physical device,
 263  * s.t. if I offline all of these sections I can then
 264  * remove the physical device?
 265  */
 266 static ssize_t show_phys_device(struct sys_device *dev, char *buf)
 267 {
 268         struct memory_block *mem =
 269                 container_of(dev, struct memory_block, sysdev);
 270         return sprintf(buf, "%d\n", mem->phys_device);
 271 }
 272
 273 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 274 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 275 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 276
 277 #define mem_create_simple_file(mem, attr_name)  \
 278         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 279 #define mem_remove_simple_file(mem, attr_name)  \
 280         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 281
 282 /*
 283  * Block size attribute stuff
 284  */
 285 static ssize_t
 286 print_block_size(struct class *class, char *buf)
 287 {
 288         return sprintf(buf, "%lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 289 }
 290
 291 static CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 292
 293 static int block_size_init(void)
 294 {
 295         sysfs_create_file(&memory_sysdev_class.kset.kobj,
 296                 &class_attr_block_size_bytes.attr);
 297         return 0;
 298 }
 299
 300 /*
 301  * Some architectures will have custom drivers to do this, and
 302  * will not need to do it from userspace.  The fake hot-add code
 303  * as well as ppc64 will do all of their discovery in userspace
 304  * and will require this interface.
 305  */
 306 #ifdef CONFIG_ARCH_MEMORY_PROBE
 307 static ssize_t
 308 memory_probe_store(struct class *class, const char __user *buf, size_t count)
 309 {
 310         u64 phys_addr;
 311         int ret;
 312
 313         phys_addr = simple_strtoull(buf, NULL, 0);
 314
 315         ret = add_memory(phys_addr, PAGES_PER_SECTION << PAGE_SHIFT);
 316
 317         if (ret)
 318                 count = ret;
 319
 320         return count;
 321 }
 322 static CLASS_ATTR(probe, 0700, NULL, memory_probe_store);
 323
 324 static int memory_probe_init(void)
 325 {
 326         sysfs_create_file(&memory_sysdev_class.kset.kobj,
 327                 &class_attr_probe.attr);
 328         return 0;
 329 }
 330 #else
 331 #define memory_probe_init(...)  do {} while (0)
 332 #endif
 333
 334 /*
 335  * Note that phys_device is optional.  It is here to allow for
 336  * differentiation between which *physical* devices each
 337  * section belongs to...
 338  */
 339
 340 static int add_memory_block(unsigned long node_id, struct mem_section *section,
 341                      unsigned long state, int phys_device)
 342 {
 343         struct memory_block *mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 344         int ret = 0;
 345
 346         if (!mem)
 347                 return -ENOMEM;
 348
 349         mem->phys_index = __section_nr(section);
 350         mem->state = state;
 351         init_MUTEX(&mem->state_sem);
 352         mem->phys_device = phys_device;
 353
 354         ret = register_memory(mem, section, NULL);
 355         if (!ret)
 356                 ret = mem_create_simple_file(mem, phys_index);
 357         if (!ret)
 358                 ret = mem_create_simple_file(mem, state);
 359         if (!ret)
 360                 ret = mem_create_simple_file(mem, phys_device);
 361
 362         return ret;
 363 }
 364
 365 /*
 366  * For now, we have a linear search to go find the appropriate
 367  * memory_block corresponding to a particular phys_index. If
 368  * this gets to be a real problem, we can always use a radix
 369  * tree or something here.
 370  *
 371  * This could be made generic for all sysdev classes.
 372  */
 373 static struct memory_block *find_memory_block(struct mem_section *section)
 374 {
 375         struct kobject *kobj;
 376         struct sys_device *sysdev;
 377         struct memory_block *mem;
 378         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 379
 380         /*
 381          * This only works because we know that section == sysdev->id
 382          * slightly redundant with sysdev_register()
 383          */
 384         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, __section_nr(section));
 385
 386         kobj = kset_find_obj(&memory_sysdev_class.kset, name);
 387         if (!kobj)
 388                 return NULL;
 389
 390         sysdev = container_of(kobj, struct sys_device, kobj);
 391         mem = container_of(sysdev, struct memory_block, sysdev);
 392
 393         return mem;
 394 }
 395
 396 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 397                 int phys_device)
 398 {
 399         struct memory_block *mem;
 400
 401         mem = find_memory_block(section);
 402         mem_remove_simple_file(mem, phys_index);
 403         mem_remove_simple_file(mem, state);
 404         mem_remove_simple_file(mem, phys_device);
 405         unregister_memory(mem, section, NULL);
 406
 407         return 0;
 408 }
 409
 410 /*
 411  * need an interface for the VM to add new memory regions,
 412  * but without onlining it.
 413  */
 414 int register_new_memory(struct mem_section *section)
 415 {
 416         return add_memory_block(0, section, MEM_OFFLINE, 0);
 417 }
 418
 419 int unregister_memory_section(struct mem_section *section)
 420 {
 421         if (!valid_section(section))
 422                 return -EINVAL;
 423
 424         return remove_memory_block(0, section, 0);
 425 }
 426
 427 /*
 428  * Initialize the sysfs support for memory devices...
 429  */
 430 int __init memory_dev_init(void)
 431 {
 432         unsigned int i;
 433         int ret;
 434
 435         memory_sysdev_class.kset.hotplug_ops = &memory_hotplug_ops;
 436         ret = sysdev_class_register(&memory_sysdev_class);
 437
 438         /*
 439          * Create entries for memory sections that were found
 440          * during boot and have been initialized
 441          */
 442         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 443                 if (!valid_section_nr(i))
 444                         continue;
 445                 add_memory_block(0, __nr_to_section(i), MEM_ONLINE, 0);
 446         }
 447
 448         memory_probe_init();
 449         block_size_init();
 450
 451         return ret;
 452 }