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         .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 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                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 187                         ret = remove_memory(start_paddr,
 188                                             PAGES_PER_SECTION << PAGE_SHIFT);
 189                         if (ret) {
 190                                 mem->state = old_state;
 191                                 break;
 192                         }
 193                         break;
 194                 default:
 195                         printk(KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 196                                         __FUNCTION__, mem, action, action);
 197                         WARN_ON(1);
 198                         ret = -EINVAL;
 199         }
 200
 201         return ret;
 202 }
 203
 204 static int memory_block_change_state(struct memory_block *mem,
 205                 unsigned long to_state, unsigned long from_state_req)
 206 {
 207         int ret = 0;
 208         down(&mem->state_sem);
 209
 210         if (mem->state != from_state_req) {
 211                 ret = -EINVAL;
 212                 goto out;
 213         }
 214
 215         ret = memory_block_action(mem, to_state);
 216         if (!ret)
 217                 mem->state = to_state;
 218
 219 out:
 220         up(&mem->state_sem);
 221         return ret;
 222 }
 223
 224 static ssize_t
 225 store_mem_state(struct sys_device *dev, const char *buf, size_t count)
 226 {
 227         struct memory_block *mem;
 228         unsigned int phys_section_nr;
 229         int ret = -EINVAL;
 230
 231         mem = container_of(dev, struct memory_block, sysdev);
 232         phys_section_nr = mem->phys_index;
 233
 234         if (!present_section_nr(phys_section_nr))
 235                 goto out;
 236
 237         if (!strncmp(buf, "online", min((int)count, 6)))
 238                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 239         else if(!strncmp(buf, "offline", min((int)count, 7)))
 240                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 241 out:
 242         if (ret)
 243                 return ret;
 244         return count;
 245 }
 246
 247 /*
 248  * phys_device is a bad name for this.  What I really want
 249  * is a way to differentiate between memory ranges that
 250  * are part of physical devices that constitute
 251  * a complete removable unit or fru.
 252  * i.e. do these ranges belong to the same physical device,
 253  * s.t. if I offline all of these sections I can then
 254  * remove the physical device?
 255  */
 256 static ssize_t show_phys_device(struct sys_device *dev, char *buf)
 257 {
 258         struct memory_block *mem =
 259                 container_of(dev, struct memory_block, sysdev);
 260         return sprintf(buf, "%d\n", mem->phys_device);
 261 }
 262
 263 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 264 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 265 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 266
 267 #define mem_create_simple_file(mem, attr_name)  \
 268         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 269 #define mem_remove_simple_file(mem, attr_name)  \
 270         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 271
 272 /*
 273  * Block size attribute stuff
 274  */
 275 static ssize_t
 276 print_block_size(struct class *class, char *buf)
 277 {
 278         return sprintf(buf, "%lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 279 }
 280
 281 static CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 282
 283 static int block_size_init(void)
 284 {
 285         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 286                                 &class_attr_block_size_bytes.attr);
 287 }
 288
 289 /*
 290  * Some architectures will have custom drivers to do this, and
 291  * will not need to do it from userspace.  The fake hot-add code
 292  * as well as ppc64 will do all of their discovery in userspace
 293  * and will require this interface.
 294  */
 295 #ifdef CONFIG_ARCH_MEMORY_PROBE
 296 static ssize_t
 297 memory_probe_store(struct class *class, const char *buf, size_t count)
 298 {
 299         u64 phys_addr;
 300         int nid;
 301         int ret;
 302
 303         phys_addr = simple_strtoull(buf, NULL, 0);
 304
 305         nid = memory_add_physaddr_to_nid(phys_addr);
 306         ret = add_memory(nid, phys_addr, PAGES_PER_SECTION << PAGE_SHIFT);
 307
 308         if (ret)
 309                 count = ret;
 310
 311         return count;
 312 }
 313 static CLASS_ATTR(probe, 0700, NULL, memory_probe_store);
 314
 315 static int memory_probe_init(void)
 316 {
 317         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 318                                 &class_attr_probe.attr);
 319 }
 320 #else
 321 static inline int memory_probe_init(void)
 322 {
 323         return 0;
 324 }
 325 #endif
 326
 327 /*
 328  * Note that phys_device is optional.  It is here to allow for
 329  * differentiation between which *physical* devices each
 330  * section belongs to...
 331  */
 332
 333 static int add_memory_block(unsigned long node_id, struct mem_section *section,
 334                      unsigned long state, int phys_device)
 335 {
 336         struct memory_block *mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 337         int ret = 0;
 338
 339         if (!mem)
 340                 return -ENOMEM;
 341
 342         mem->phys_index = __section_nr(section);
 343         mem->state = state;
 344         init_MUTEX(&mem->state_sem);
 345         mem->phys_device = phys_device;
 346
 347         ret = register_memory(mem, section, NULL);
 348         if (!ret)
 349                 ret = mem_create_simple_file(mem, phys_index);
 350         if (!ret)
 351                 ret = mem_create_simple_file(mem, state);
 352         if (!ret)
 353                 ret = mem_create_simple_file(mem, phys_device);
 354
 355         return ret;
 356 }
 357
 358 /*
 359  * For now, we have a linear search to go find the appropriate
 360  * memory_block corresponding to a particular phys_index. If
 361  * this gets to be a real problem, we can always use a radix
 362  * tree or something here.
 363  *
 364  * This could be made generic for all sysdev classes.
 365  */
 366 static struct memory_block *find_memory_block(struct mem_section *section)
 367 {
 368         struct kobject *kobj;
 369         struct sys_device *sysdev;
 370         struct memory_block *mem;
 371         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 372
 373         /*
 374          * This only works because we know that section == sysdev->id
 375          * slightly redundant with sysdev_register()
 376          */
 377         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, __section_nr(section));
 378
 379         kobj = kset_find_obj(&memory_sysdev_class.kset, name);
 380         if (!kobj)
 381                 return NULL;
 382
 383         sysdev = container_of(kobj, struct sys_device, kobj);
 384         mem = container_of(sysdev, struct memory_block, sysdev);
 385
 386         return mem;
 387 }
 388
 389 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 390                 int phys_device)
 391 {
 392         struct memory_block *mem;
 393
 394         mem = find_memory_block(section);
 395         mem_remove_simple_file(mem, phys_index);
 396         mem_remove_simple_file(mem, state);
 397         mem_remove_simple_file(mem, phys_device);
 398         unregister_memory(mem, section, NULL);
 399
 400         return 0;
 401 }
 402
 403 /*
 404  * need an interface for the VM to add new memory regions,
 405  * but without onlining it.
 406  */
 407 int register_new_memory(struct mem_section *section)
 408 {
 409         return add_memory_block(0, section, MEM_OFFLINE, 0);
 410 }
 411
 412 int unregister_memory_section(struct mem_section *section)
 413 {
 414         if (!present_section(section))
 415                 return -EINVAL;
 416
 417         return remove_memory_block(0, section, 0);
 418 }
 419
 420 /*
 421  * Initialize the sysfs support for memory devices...
 422  */
 423 int __init memory_dev_init(void)
 424 {
 425         unsigned int i;
 426         int ret;
 427         int err;
 428
 429         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 430         ret = sysdev_class_register(&memory_sysdev_class);
 431         if (ret)
 432                 goto out;
 433
 434         /*
 435          * Create entries for memory sections that were found
 436          * during boot and have been initialized
 437          */
 438         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 439                 if (!present_section_nr(i))
 440                         continue;
 441                 err = add_memory_block(0, __nr_to_section(i), MEM_ONLINE, 0);
 442                 if (!ret)
 443                         ret = err;
 444         }
 445
 446         err = memory_probe_init();
 447         if (!ret)
 448                 ret = err;
 449         err = block_size_init();
 450         if (!ret)
 451                 ret = err;
 452 out:
 453         if (ret)
 454                 printk(KERN_ERR "%s() failed: %d\n", __FUNCTION__, ret);
 455         return ret;
 456 }