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