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 <linux/mutex.h>
  24 #include <asm/atomic.h>
  25 #include <asm/uaccess.h>
  26
  27 #define MEMORY_CLASS_NAME       "memory"
  28
  29 static struct sysdev_class memory_sysdev_class = {
  30         .name = MEMORY_CLASS_NAME,
  31 };
  32
  33 static const char *memory_uevent_name(struct kset *kset, struct kobject *kobj)
  34 {
  35         return MEMORY_CLASS_NAME;
  36 }
  37
  38 static int memory_uevent(struct kset *kset, struct kobject *obj, struct kobj_uevent_env *env)
  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 EXPORT_SYMBOL(register_memory_notifier);
  57
  58 void unregister_memory_notifier(struct notifier_block *nb)
  59 {
  60         blocking_notifier_chain_unregister(&memory_chain, nb);
  61 }
  62 EXPORT_SYMBOL(unregister_memory_notifier);
  63
  64 /*
  65  * register_memory - Setup a sysfs device for a memory block
  66  */
  67 static
  68 int register_memory(struct memory_block *memory, struct mem_section *section)
  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         return error;
  77 }
  78
  79 static void
  80 unregister_memory(struct memory_block *memory, struct mem_section *section)
  81 {
  82         BUG_ON(memory->sysdev.cls != &memory_sysdev_class);
  83         BUG_ON(memory->sysdev.id != __section_nr(section));
  84
  85         /* drop the ref. we got in remove_memory_block() */
  86         kobject_put(&memory->sysdev.kobj);
  87         sysdev_unregister(&memory->sysdev);
  88 }
  89
  90 /*
  91  * use this as the physical section index that this memsection
  92  * uses.
  93  */
  94
  95 static ssize_t show_mem_phys_index(struct sys_device *dev,
  96                         struct sysdev_attribute *attr, char *buf)
  97 {
  98         struct memory_block *mem =
  99                 container_of(dev, struct memory_block, sysdev);
 100         return sprintf(buf, "%08lx\n", mem->phys_index);
 101 }
 102
 103 /*
 104  * online, offline, going offline, etc.
 105  */
 106 static ssize_t show_mem_state(struct sys_device *dev,
 107                         struct sysdev_attribute *attr, char *buf)
 108 {
 109         struct memory_block *mem =
 110                 container_of(dev, struct memory_block, sysdev);
 111         ssize_t len = 0;
 112
 113         /*
 114          * We can probably put these states in a nice little array
 115          * so that they're not open-coded
 116          */
 117         switch (mem->state) {
 118                 case MEM_ONLINE:
 119                         len = sprintf(buf, "online\n");
 120                         break;
 121                 case MEM_OFFLINE:
 122                         len = sprintf(buf, "offline\n");
 123                         break;
 124                 case MEM_GOING_OFFLINE:
 125                         len = sprintf(buf, "going-offline\n");
 126                         break;
 127                 default:
 128                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 129                                         mem->state);
 130                         WARN_ON(1);
 131                         break;
 132         }
 133
 134         return len;
 135 }
 136
 137 int memory_notify(unsigned long val, void *v)
 138 {
 139         return blocking_notifier_call_chain(&memory_chain, val, v);
 140 }
 141
 142 /*
 143  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 144  * OK to have direct references to sparsemem variables in here.
 145  */
 146 static int
 147 memory_block_action(struct memory_block *mem, unsigned long action)
 148 {
 149         int i;
 150         unsigned long psection;
 151         unsigned long start_pfn, start_paddr;
 152         struct page *first_page;
 153         int ret;
 154         int old_state = mem->state;
 155
 156         psection = mem->phys_index;
 157         first_page = pfn_to_page(psection << PFN_SECTION_SHIFT);
 158
 159         /*
 160          * The probe routines leave the pages reserved, just
 161          * as the bootmem code does.  Make sure they're still
 162          * that way.
 163          */
 164         if (action == MEM_ONLINE) {
 165                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 166                         if (PageReserved(first_page+i))
 167                                 continue;
 168
 169                         printk(KERN_WARNING "section number %ld page number %d "
 170                                 "not reserved, was it already online? \n",
 171                                 psection, i);
 172                         return -EBUSY;
 173                 }
 174         }
 175
 176         switch (action) {
 177                 case MEM_ONLINE:
 178                         start_pfn = page_to_pfn(first_page);
 179                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 180                         break;
 181                 case MEM_OFFLINE:
 182                         mem->state = MEM_GOING_OFFLINE;
 183                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 184                         ret = remove_memory(start_paddr,
 185                                             PAGES_PER_SECTION << PAGE_SHIFT);
 186                         if (ret) {
 187                                 mem->state = old_state;
 188                                 break;
 189                         }
 190                         break;
 191                 default:
 192                         printk(KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 193                                         __func__, mem, action, action);
 194                         WARN_ON(1);
 195                         ret = -EINVAL;
 196         }
 197
 198         return ret;
 199 }
 200
 201 static int memory_block_change_state(struct memory_block *mem,
 202                 unsigned long to_state, unsigned long from_state_req)
 203 {
 204         int ret = 0;
 205         mutex_lock(&mem->state_mutex);
 206
 207         if (mem->state != from_state_req) {
 208                 ret = -EINVAL;
 209                 goto out;
 210         }
 211
 212         ret = memory_block_action(mem, to_state);
 213         if (!ret)
 214                 mem->state = to_state;
 215
 216 out:
 217         mutex_unlock(&mem->state_mutex);
 218         return ret;
 219 }
 220
 221 static ssize_t
 222 store_mem_state(struct sys_device *dev,
 223                 struct sysdev_attribute *attr, const char *buf, size_t count)
 224 {
 225         struct memory_block *mem;
 226         unsigned int phys_section_nr;
 227         int ret = -EINVAL;
 228
 229         mem = container_of(dev, struct memory_block, sysdev);
 230         phys_section_nr = mem->phys_index;
 231
 232         if (!present_section_nr(phys_section_nr))
 233                 goto out;
 234
 235         if (!strncmp(buf, "online", min((int)count, 6)))
 236                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 237         else if(!strncmp(buf, "offline", min((int)count, 7)))
 238                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 239 out:
 240         if (ret)
 241                 return ret;
 242         return count;
 243 }
 244
 245 /*
 246  * phys_device is a bad name for this.  What I really want
 247  * is a way to differentiate between memory ranges that
 248  * are part of physical devices that constitute
 249  * a complete removable unit or fru.
 250  * i.e. do these ranges belong to the same physical device,
 251  * s.t. if I offline all of these sections I can then
 252  * remove the physical device?
 253  */
 254 static ssize_t show_phys_device(struct sys_device *dev,
 255                                 struct sysdev_attribute *attr, char *buf)
 256 {
 257         struct memory_block *mem =
 258                 container_of(dev, struct memory_block, sysdev);
 259         return sprintf(buf, "%d\n", mem->phys_device);
 260 }
 261
 262 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 263 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 264 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 265
 266 #define mem_create_simple_file(mem, attr_name)  \
 267         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 268 #define mem_remove_simple_file(mem, attr_name)  \
 269         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 270
 271 /*
 272  * Block size attribute stuff
 273  */
 274 static ssize_t
 275 print_block_size(struct class *class, char *buf)
 276 {
 277         return sprintf(buf, "%lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 278 }
 279
 280 static CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 281
 282 static int block_size_init(void)
 283 {
 284         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 285                                 &class_attr_block_size_bytes.attr);
 286 }
 287
 288 /*
 289  * Some architectures will have custom drivers to do this, and
 290  * will not need to do it from userspace.  The fake hot-add code
 291  * as well as ppc64 will do all of their discovery in userspace
 292  * and will require this interface.
 293  */
 294 #ifdef CONFIG_ARCH_MEMORY_PROBE
 295 static ssize_t
 296 memory_probe_store(struct class *class, const char *buf, size_t count)
 297 {
 298         u64 phys_addr;
 299         int nid;
 300         int ret;
 301
 302         phys_addr = simple_strtoull(buf, NULL, 0);
 303
 304         nid = memory_add_physaddr_to_nid(phys_addr);
 305         ret = add_memory(nid, phys_addr, PAGES_PER_SECTION << PAGE_SHIFT);
 306
 307         if (ret)
 308                 count = ret;
 309
 310         return count;
 311 }
 312 static CLASS_ATTR(probe, 0700, NULL, memory_probe_store);
 313
 314 static int memory_probe_init(void)
 315 {
 316         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 317                                 &class_attr_probe.attr);
 318 }
 319 #else
 320 static inline int memory_probe_init(void)
 321 {
 322         return 0;
 323 }
 324 #endif
 325
 326 /*
 327  * Note that phys_device is optional.  It is here to allow for
 328  * differentiation between which *physical* devices each
 329  * section belongs to...
 330  */
 331
 332 static int add_memory_block(unsigned long node_id, struct mem_section *section,
 333                      unsigned long state, int phys_device)
 334 {
 335         struct memory_block *mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 336         int ret = 0;
 337
 338         if (!mem)
 339                 return -ENOMEM;
 340
 341         mem->phys_index = __section_nr(section);
 342         mem->state = state;
 343         mutex_init(&mem->state_mutex);
 344         mem->phys_device = phys_device;
 345
 346         ret = register_memory(mem, section);
 347         if (!ret)
 348                 ret = mem_create_simple_file(mem, phys_index);
 349         if (!ret)
 350                 ret = mem_create_simple_file(mem, state);
 351         if (!ret)
 352                 ret = mem_create_simple_file(mem, phys_device);
 353
 354         return ret;
 355 }
 356
 357 /*
 358  * For now, we have a linear search to go find the appropriate
 359  * memory_block corresponding to a particular phys_index. If
 360  * this gets to be a real problem, we can always use a radix
 361  * tree or something here.
 362  *
 363  * This could be made generic for all sysdev classes.
 364  */
 365 static struct memory_block *find_memory_block(struct mem_section *section)
 366 {
 367         struct kobject *kobj;
 368         struct sys_device *sysdev;
 369         struct memory_block *mem;
 370         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 371
 372         /*
 373          * This only works because we know that section == sysdev->id
 374          * slightly redundant with sysdev_register()
 375          */
 376         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, __section_nr(section));
 377
 378         kobj = kset_find_obj(&memory_sysdev_class.kset, name);
 379         if (!kobj)
 380                 return NULL;
 381
 382         sysdev = container_of(kobj, struct sys_device, kobj);
 383         mem = container_of(sysdev, struct memory_block, sysdev);
 384
 385         return mem;
 386 }
 387
 388 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 389                 int phys_device)
 390 {
 391         struct memory_block *mem;
 392
 393         mem = find_memory_block(section);
 394         mem_remove_simple_file(mem, phys_index);
 395         mem_remove_simple_file(mem, state);
 396         mem_remove_simple_file(mem, phys_device);
 397         unregister_memory(mem, section);
 398
 399         return 0;
 400 }
 401
 402 /*
 403  * need an interface for the VM to add new memory regions,
 404  * but without onlining it.
 405  */
 406 int register_new_memory(struct mem_section *section)
 407 {
 408         return add_memory_block(0, section, MEM_OFFLINE, 0);
 409 }
 410
 411 int unregister_memory_section(struct mem_section *section)
 412 {
 413         if (!present_section(section))
 414                 return -EINVAL;
 415
 416         return remove_memory_block(0, section, 0);
 417 }
 418
 419 /*
 420  * Initialize the sysfs support for memory devices...
 421  */
 422 int __init memory_dev_init(void)
 423 {
 424         unsigned int i;
 425         int ret;
 426         int err;
 427
 428         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 429         ret = sysdev_class_register(&memory_sysdev_class);
 430         if (ret)
 431                 goto out;
 432
 433         /*
 434          * Create entries for memory sections that were found
 435          * during boot and have been initialized
 436          */
 437         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 438                 if (!present_section_nr(i))
 439                         continue;
 440                 err = add_memory_block(0, __nr_to_section(i), MEM_ONLINE, 0);
 441                 if (!ret)
 442                         ret = err;
 443         }
 444
 445         err = memory_probe_init();
 446         if (!ret)
 447                 ret = err;
 448         err = block_size_init();
 449         if (!ret)
 450                 ret = err;
 451 out:
 452         if (ret)
 453                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 454         return ret;
 455 }