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/topology.h>
  17 #include <linux/capability.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
  32 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
  33 {
  34         return MEMORY_CLASS_NAME;
  35 }
  36
  37 static int memory_uevent(struct kset *kset, struct kobject *kobj, char **envp,
  38                         int num_envp, char *buffer, int buffer_size)
  39 {
  40         int retval = 0;
  41
  42         return retval;
  43 }
  44
  45 static struct kset_uevent_ops memory_uevent_ops = {
  46         .name           = memory_uevent_name,
  47         .uevent         = memory_uevent,
  48 };
  49
  50 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  51
  52 int register_memory_notifier(struct notifier_block *nb)
  53 {
  54         return blocking_notifier_chain_register(&memory_chain, nb);
  55 }
  56
  57 void unregister_memory_notifier(struct notifier_block *nb)
  58 {
  59         blocking_notifier_chain_unregister(&memory_chain, nb);
  60 }
  61
  62 /*
  63  * register_memory - Setup a sysfs device for a memory block
  64  */
  65 int register_memory(struct memory_block *memory, struct mem_section *section,
  66                 struct node *root)
  67 {
  68         int error;
  69
  70         memory->sysdev.cls = &memory_sysdev_class;
  71         memory->sysdev.id = __section_nr(section);
  72
  73         error = sysdev_register(&memory->sysdev);
  74
  75         if (root && !error)
  76                 error = sysfs_create_link(&root->sysdev.kobj,
  77                                           &memory->sysdev.kobj,
  78                                           kobject_name(&memory->sysdev.kobj));
  79
  80         return error;
  81 }
  82
  83 static void
  84 unregister_memory(struct memory_block *memory, struct mem_section *section,
  85                 struct node *root)
  86 {
  87         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
  88         BUG_ON(memory->sysdev.id != __section_nr(section));
  89
  90         sysdev_unregister(&memory->sysdev);
  91         if (root)
  92                 sysfs_remove_link(&root->sysdev.kobj,
  93                                   kobject_name(&memory->sysdev.kobj));
  94 }
  95
  96 /*
  97  * use this as the physical section index that this memsection
  98  * uses.
  99  */
 100
 101 static ssize_t show_mem_phys_index(struct sys_device *dev, char *buf)
 102 {
 103         struct memory_block *mem =
 104                 container_of(dev, struct memory_block, sysdev);
 105         return sprintf(buf, "%08lx\n", mem->phys_index);
 106 }
 107
 108 /*
 109  * online, offline, going offline, etc.
 110  */
 111 static ssize_t show_mem_state(struct sys_device *dev, char *buf)
 112 {
 113         struct memory_block *mem =
 114                 container_of(dev, struct memory_block, sysdev);
 115         ssize_t len = 0;
 116
 117         /*
 118          * We can probably put these states in a nice little array
 119          * so that they're not open-coded
 120          */
 121         switch (mem->state) {
 122                 case MEM_ONLINE:
 123                         len = sprintf(buf, "online\n");
 124                         break;
 125                 case MEM_OFFLINE:
 126                         len = sprintf(buf, "offline\n");
 127                         break;
 128                 case MEM_GOING_OFFLINE:
 129                         len = sprintf(buf, "going-offline\n");
 130                         break;
 131                 default:
 132                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 133                                         mem->state);
 134                         WARN_ON(1);
 135                         break;
 136         }
 137
 138         return len;
 139 }
 140
 141 static inline int memory_notify(unsigned long val, void *v)
 142 {
 143         return blocking_notifier_call_chain(&memory_chain, val, v);
 144 }
 145
 146 /*
 147  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 148  * OK to have direct references to sparsemem variables in here.
 149  */
 150 static int
 151 memory_block_action(struct memory_block *mem, unsigned long action)
 152 {
 153         int i;
 154         unsigned long psection;
 155         unsigned long start_pfn, start_paddr;
 156         struct page *first_page;
 157         int ret;
 158         int old_state = mem->state;
 159
 160         psection = mem->phys_index;
 161         first_page = pfn_to_page(psection << PFN_SECTION_SHIFT);
 162
 163         /*
 164          * The probe routines leave the pages reserved, just
 165          * as the bootmem code does.  Make sure they're still
 166          * that way.
 167          */
 168         if (action == MEM_ONLINE) {
 169                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 170                         if (PageReserved(first_page+i))
 171                                 continue;
 172
 173                         printk(KERN_WARNING "section number %ld page number %d "
 174                                 "not reserved, was it already online? \n",
 175                                 psection, i);
 176                         return -EBUSY;
 177                 }
 178         }
 179
 180         switch (action) {
 181                 case MEM_ONLINE:
 182                         start_pfn = page_to_pfn(first_page);
 183                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 184                         break;
 185                 case MEM_OFFLINE:
 186                         mem->state = MEM_GOING_OFFLINE;
 187                         memory_notify(MEM_GOING_OFFLINE, NULL);
 188                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 189                         ret = remove_memory(start_paddr,
 190                                             PAGES_PER_SECTION << PAGE_SHIFT);
 191                         if (ret) {
 192                                 mem->state = old_state;
 193                                 break;
 194                         }
 195                         memory_notify(MEM_MAPPING_INVALID, NULL);
 196                         break;
 197                 default:
 198                         printk(KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 199                                         __FUNCTION__, mem, action, action);
 200                         WARN_ON(1);
 201                         ret = -EINVAL;
 202         }
 203         /*
 204          * For now, only notify on successful memory operations
 205          */
 206         if (!ret)
 207                 memory_notify(action, NULL);
 208
 209         return ret;
 210 }
 211
 212 static int memory_block_change_state(struct memory_block *mem,
 213                 unsigned long to_state, unsigned long from_state_req)
 214 {
 215         int ret = 0;
 216         down(&mem->state_sem);
 217
 218         if (mem->state != from_state_req) {
 219                 ret = -EINVAL;
 220                 goto out;
 221         }
 222
 223         ret = memory_block_action(mem, to_state);
 224         if (!ret)
 225                 mem->state = to_state;
 226
 227 out:
 228         up(&mem->state_sem);
 229         return ret;
 230 }
 231
 232 static ssize_t
 233 store_mem_state(struct sys_device *dev, const char *buf, size_t count)
 234 {
 235         struct memory_block *mem;
 236         unsigned int phys_section_nr;
 237         int ret = -EINVAL;
 238
 239         mem = container_of(dev, struct memory_block, sysdev);
 240         phys_section_nr = mem->phys_index;
 241
 242         if (!valid_section_nr(phys_section_nr))
 243                 goto out;
 244
 245         if (!strncmp(buf, "online", min((int)count, 6)))
 246                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 247         else if(!strncmp(buf, "offline", min((int)count, 7)))
 248                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 249 out:
 250         if (ret)
 251                 return ret;
 252         return count;
 253 }
 254
 255 /*
 256  * phys_device is a bad name for this.  What I really want
 257  * is a way to differentiate between memory ranges that
 258  * are part of physical devices that constitute
 259  * a complete removable unit or fru.
 260  * i.e. do these ranges belong to the same physical device,
 261  * s.t. if I offline all of these sections I can then
 262  * remove the physical device?
 263  */
 264 static ssize_t show_phys_device(struct sys_device *dev, char *buf)
 265 {
 266         struct memory_block *mem =
 267                 container_of(dev, struct memory_block, sysdev);
 268         return sprintf(buf, "%d\n", mem->phys_device);
 269 }
 270
 271 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 272 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 273 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 274
 275 #define mem_create_simple_file(mem, attr_name)  \
 276         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 277 #define mem_remove_simple_file(mem, attr_name)  \
 278         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 279
 280 /*
 281  * Block size attribute stuff
 282  */
 283 static ssize_t
 284 print_block_size(struct class *class, char *buf)
 285 {
 286         return sprintf(buf, "%lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 287 }
 288
 289 static CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 290
 291 static int block_size_init(void)
 292 {
 293         sysfs_create_file(&memory_sysdev_class.kset.kobj,
 294                 &class_attr_block_size_bytes.attr);
 295         return 0;
 296 }
 297
 298 /*
 299  * Some architectures will have custom drivers to do this, and
 300  * will not need to do it from userspace.  The fake hot-add code
 301  * as well as ppc64 will do all of their discovery in userspace
 302  * and will require this interface.
 303  */
 304 #ifdef CONFIG_ARCH_MEMORY_PROBE
 305 static ssize_t
 306 memory_probe_store(struct class *class, const char *buf, size_t count)
 307 {
 308         u64 phys_addr;
 309         int nid;
 310         int ret;
 311
 312         phys_addr = simple_strtoull(buf, NULL, 0);
 313
 314         nid = memory_add_physaddr_to_nid(phys_addr);
 315         ret = add_memory(nid, 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.uevent_ops = &memory_uevent_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 }