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 *obj, struct kobj_uevent_env *env)
  38 {
  39         int retval = 0;
  40
  41         return retval;
  42 }
  43
  44 static struct kset_uevent_ops memory_uevent_ops = {
  45         .name           = memory_uevent_name,
  46         .uevent         = memory_uevent,
  47 };
  48
  49 static BLOCKING_NOTIFIER_HEAD(memory_chain);
  50
  51 int register_memory_notifier(struct notifier_block *nb)
  52 {
  53         return blocking_notifier_chain_register(&memory_chain, nb);
  54 }
  55
  56 void unregister_memory_notifier(struct notifier_block *nb)
  57 {
  58         blocking_notifier_chain_unregister(&memory_chain, nb);
  59 }
  60
  61 /*
  62  * register_memory - Setup a sysfs device for a memory block
  63  */
  64 int register_memory(struct memory_block *memory, struct mem_section *section,
  65                 struct node *root)
  66 {
  67         int error;
  68
  69         memory->sysdev.cls = &memory_sysdev_class;
  70         memory->sysdev.id = __section_nr(section);
  71
  72         error = sysdev_register(&memory->sysdev);
  73
  74         if (root && !error)
  75                 error = sysfs_create_link(&root->sysdev.kobj,
  76                                           &memory->sysdev.kobj,
  77                                           kobject_name(&memory->sysdev.kobj));
  78
  79         return error;
  80 }
  81
  82 static void
  83 unregister_memory(struct memory_block *memory, struct mem_section *section,
  84                 struct node *root)
  85 {
  86         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
  87         BUG_ON(memory->sysdev.id != __section_nr(section));
  88
  89         sysdev_unregister(&memory->sysdev);
  90         if (root)
  91                 sysfs_remove_link(&root->sysdev.kobj,
  92                                   kobject_name(&memory->sysdev.kobj));
  93 }
  94
  95 /*
  96  * use this as the physical section index that this memsection
  97  * uses.
  98  */
  99
 100 static ssize_t show_mem_phys_index(struct sys_device *dev, char *buf)
 101 {
 102         struct memory_block *mem =
 103                 container_of(dev, struct memory_block, sysdev);
 104         return sprintf(buf, "%08lx\n", mem->phys_index);
 105 }
 106
 107 /*
 108  * online, offline, going offline, etc.
 109  */
 110 static ssize_t show_mem_state(struct sys_device *dev, char *buf)
 111 {
 112         struct memory_block *mem =
 113                 container_of(dev, struct memory_block, sysdev);
 114         ssize_t len = 0;
 115
 116         /*
 117          * We can probably put these states in a nice little array
 118          * so that they're not open-coded
 119          */
 120         switch (mem->state) {
 121                 case MEM_ONLINE:
 122                         len = sprintf(buf, "online\n");
 123                         break;
 124                 case MEM_OFFLINE:
 125                         len = sprintf(buf, "offline\n");
 126                         break;
 127                 case MEM_GOING_OFFLINE:
 128                         len = sprintf(buf, "going-offline\n");
 129                         break;
 130                 default:
 131                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 132                                         mem->state);
 133                         WARN_ON(1);
 134                         break;
 135         }
 136
 137         return len;
 138 }
 139
 140 static inline int memory_notify(unsigned long val, void *v)
 141 {
 142         return blocking_notifier_call_chain(&memory_chain, val, v);
 143 }
 144
 145 /*
 146  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 147  * OK to have direct references to sparsemem variables in here.
 148  */
 149 static int
 150 memory_block_action(struct memory_block *mem, unsigned long action)
 151 {
 152         int i;
 153         unsigned long psection;
 154         unsigned long start_pfn, start_paddr;
 155         struct page *first_page;
 156         int ret;
 157         int old_state = mem->state;
 158
 159         psection = mem->phys_index;
 160         first_page = pfn_to_page(psection << PFN_SECTION_SHIFT);
 161
 162         /*
 163          * The probe routines leave the pages reserved, just
 164          * as the bootmem code does.  Make sure they're still
 165          * that way.
 166          */
 167         if (action == MEM_ONLINE) {
 168                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 169                         if (PageReserved(first_page+i))
 170                                 continue;
 171
 172                         printk(KERN_WARNING "section number %ld page number %d "
 173                                 "not reserved, was it already online? \n",
 174                                 psection, i);
 175                         return -EBUSY;
 176                 }
 177         }
 178
 179         switch (action) {
 180                 case MEM_ONLINE:
 181                         start_pfn = page_to_pfn(first_page);
 182                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 183                         break;
 184                 case MEM_OFFLINE:
 185                         mem->state = MEM_GOING_OFFLINE;
 186                         memory_notify(MEM_GOING_OFFLINE, NULL);
 187                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 188                         ret = remove_memory(start_paddr,
 189                                             PAGES_PER_SECTION << PAGE_SHIFT);
 190                         if (ret) {
 191                                 mem->state = old_state;
 192                                 break;
 193                         }
 194                         memory_notify(MEM_MAPPING_INVALID, NULL);
 195                         break;
 196                 default:
 197                         printk(KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 198                                         __FUNCTION__, mem, action, action);
 199                         WARN_ON(1);
 200                         ret = -EINVAL;
 201         }
 202         /*
 203          * For now, only notify on successful memory operations
 204          */
 205         if (!ret)
 206                 memory_notify(action, NULL);
 207
 208         return ret;
 209 }
 210
 211 static int memory_block_change_state(struct memory_block *mem,
 212                 unsigned long to_state, unsigned long from_state_req)
 213 {
 214         int ret = 0;
 215         down(&mem->state_sem);
 216
 217         if (mem->state != from_state_req) {
 218                 ret = -EINVAL;
 219                 goto out;
 220         }
 221
 222         ret = memory_block_action(mem, to_state);
 223         if (!ret)
 224                 mem->state = to_state;
 225
 226 out:
 227         up(&mem->state_sem);
 228         return ret;
 229 }
 230
 231 static ssize_t
 232 store_mem_state(struct sys_device *dev, const char *buf, size_t count)
 233 {
 234         struct memory_block *mem;
 235         unsigned int phys_section_nr;
 236         int ret = -EINVAL;
 237
 238         mem = container_of(dev, struct memory_block, sysdev);
 239         phys_section_nr = mem->phys_index;
 240
 241         if (!present_section_nr(phys_section_nr))
 242                 goto out;
 243
 244         if (!strncmp(buf, "online", min((int)count, 6)))
 245                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 246         else if(!strncmp(buf, "offline", min((int)count, 7)))
 247                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 248 out:
 249         if (ret)
 250                 return ret;
 251         return count;
 252 }
 253
 254 /*
 255  * phys_device is a bad name for this.  What I really want
 256  * is a way to differentiate between memory ranges that
 257  * are part of physical devices that constitute
 258  * a complete removable unit or fru.
 259  * i.e. do these ranges belong to the same physical device,
 260  * s.t. if I offline all of these sections I can then
 261  * remove the physical device?
 262  */
 263 static ssize_t show_phys_device(struct sys_device *dev, char *buf)
 264 {
 265         struct memory_block *mem =
 266                 container_of(dev, struct memory_block, sysdev);
 267         return sprintf(buf, "%d\n", mem->phys_device);
 268 }
 269
 270 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 271 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 272 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 273
 274 #define mem_create_simple_file(mem, attr_name)  \
 275         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 276 #define mem_remove_simple_file(mem, attr_name)  \
 277         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 278
 279 /*
 280  * Block size attribute stuff
 281  */
 282 static ssize_t
 283 print_block_size(struct class *class, char *buf)
 284 {
 285         return sprintf(buf, "%lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 286 }
 287
 288 static CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 289
 290 static int block_size_init(void)
 291 {
 292         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 293                                 &class_attr_block_size_bytes.attr);
 294 }
 295
 296 /*
 297  * Some architectures will have custom drivers to do this, and
 298  * will not need to do it from userspace.  The fake hot-add code
 299  * as well as ppc64 will do all of their discovery in userspace
 300  * and will require this interface.
 301  */
 302 #ifdef CONFIG_ARCH_MEMORY_PROBE
 303 static ssize_t
 304 memory_probe_store(struct class *class, const char *buf, size_t count)
 305 {
 306         u64 phys_addr;
 307         int nid;
 308         int ret;
 309
 310         phys_addr = simple_strtoull(buf, NULL, 0);
 311
 312         nid = memory_add_physaddr_to_nid(phys_addr);
 313         ret = add_memory(nid, phys_addr, PAGES_PER_SECTION << PAGE_SHIFT);
 314
 315         if (ret)
 316                 count = ret;
 317
 318         return count;
 319 }
 320 static CLASS_ATTR(probe, 0700, NULL, memory_probe_store);
 321
 322 static int memory_probe_init(void)
 323 {
 324         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 325                                 &class_attr_probe.attr);
 326 }
 327 #else
 328 static inline int memory_probe_init(void)
 329 {
 330         return 0;
 331 }
 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 (!present_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         int err;
 435
 436         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 437         ret = sysdev_class_register(&memory_sysdev_class);
 438         if (ret)
 439                 goto out;
 440
 441         /*
 442          * Create entries for memory sections that were found
 443          * during boot and have been initialized
 444          */
 445         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 446                 if (!present_section_nr(i))
 447                         continue;
 448                 err = add_memory_block(0, __nr_to_section(i), MEM_ONLINE, 0);
 449                 if (!ret)
 450                         ret = err;
 451         }
 452
 453         err = memory_probe_init();
 454         if (!ret)
 455                 ret = err;
 456         err = block_size_init();
 457         if (!ret)
 458                 ret = err;
 459 out:
 460         if (ret)
 461                 printk(KERN_ERR "%s() failed: %d\n", __FUNCTION__, ret);
 462         return ret;
 463 }