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  * Show whether the section of memory is likely to be hot-removable
 105  */
 106 static ssize_t show_mem_removable(struct sys_device *dev, char *buf)
 107 {
 108         unsigned long start_pfn;
 109         int ret;
 110         struct memory_block *mem =
 111                 container_of(dev, struct memory_block, sysdev);
 112
 113         start_pfn = section_nr_to_pfn(mem->phys_index);
 114         ret = is_mem_section_removable(start_pfn, PAGES_PER_SECTION);
 115         return sprintf(buf, "%d\n", ret);
 116 }
 117
 118 /*
 119  * online, offline, going offline, etc.
 120  */
 121 static ssize_t show_mem_state(struct sys_device *dev,
 122                         struct sysdev_attribute *attr, char *buf)
 123 {
 124         struct memory_block *mem =
 125                 container_of(dev, struct memory_block, sysdev);
 126         ssize_t len = 0;
 127
 128         /*
 129          * We can probably put these states in a nice little array
 130          * so that they're not open-coded
 131          */
 132         switch (mem->state) {
 133                 case MEM_ONLINE:
 134                         len = sprintf(buf, "online\n");
 135                         break;
 136                 case MEM_OFFLINE:
 137                         len = sprintf(buf, "offline\n");
 138                         break;
 139                 case MEM_GOING_OFFLINE:
 140                         len = sprintf(buf, "going-offline\n");
 141                         break;
 142                 default:
 143                         len = sprintf(buf, "ERROR-UNKNOWN-%ld\n",
 144                                         mem->state);
 145                         WARN_ON(1);
 146                         break;
 147         }
 148
 149         return len;
 150 }
 151
 152 int memory_notify(unsigned long val, void *v)
 153 {
 154         return blocking_notifier_call_chain(&memory_chain, val, v);
 155 }
 156
 157 /*
 158  * MEMORY_HOTPLUG depends on SPARSEMEM in mm/Kconfig, so it is
 159  * OK to have direct references to sparsemem variables in here.
 160  */
 161 static int
 162 memory_block_action(struct memory_block *mem, unsigned long action)
 163 {
 164         int i;
 165         unsigned long psection;
 166         unsigned long start_pfn, start_paddr;
 167         struct page *first_page;
 168         int ret;
 169         int old_state = mem->state;
 170
 171         psection = mem->phys_index;
 172         first_page = pfn_to_page(psection << PFN_SECTION_SHIFT);
 173
 174         /*
 175          * The probe routines leave the pages reserved, just
 176          * as the bootmem code does.  Make sure they're still
 177          * that way.
 178          */
 179         if (action == MEM_ONLINE) {
 180                 for (i = 0; i < PAGES_PER_SECTION; i++) {
 181                         if (PageReserved(first_page+i))
 182                                 continue;
 183
 184                         printk(KERN_WARNING "section number %ld page number %d "
 185                                 "not reserved, was it already online? \n",
 186                                 psection, i);
 187                         return -EBUSY;
 188                 }
 189         }
 190
 191         switch (action) {
 192                 case MEM_ONLINE:
 193                         start_pfn = page_to_pfn(first_page);
 194                         ret = online_pages(start_pfn, PAGES_PER_SECTION);
 195                         break;
 196                 case MEM_OFFLINE:
 197                         mem->state = MEM_GOING_OFFLINE;
 198                         start_paddr = page_to_pfn(first_page) << PAGE_SHIFT;
 199                         ret = remove_memory(start_paddr,
 200                                             PAGES_PER_SECTION << PAGE_SHIFT);
 201                         if (ret) {
 202                                 mem->state = old_state;
 203                                 break;
 204                         }
 205                         break;
 206                 default:
 207                         printk(KERN_WARNING "%s(%p, %ld) unknown action: %ld\n",
 208                                         __func__, mem, action, action);
 209                         WARN_ON(1);
 210                         ret = -EINVAL;
 211         }
 212
 213         return ret;
 214 }
 215
 216 static int memory_block_change_state(struct memory_block *mem,
 217                 unsigned long to_state, unsigned long from_state_req)
 218 {
 219         int ret = 0;
 220         mutex_lock(&mem->state_mutex);
 221
 222         if (mem->state != from_state_req) {
 223                 ret = -EINVAL;
 224                 goto out;
 225         }
 226
 227         ret = memory_block_action(mem, to_state);
 228         if (!ret)
 229                 mem->state = to_state;
 230
 231 out:
 232         mutex_unlock(&mem->state_mutex);
 233         return ret;
 234 }
 235
 236 static ssize_t
 237 store_mem_state(struct sys_device *dev,
 238                 struct sysdev_attribute *attr, const char *buf, size_t count)
 239 {
 240         struct memory_block *mem;
 241         unsigned int phys_section_nr;
 242         int ret = -EINVAL;
 243
 244         mem = container_of(dev, struct memory_block, sysdev);
 245         phys_section_nr = mem->phys_index;
 246
 247         if (!present_section_nr(phys_section_nr))
 248                 goto out;
 249
 250         if (!strncmp(buf, "online", min((int)count, 6)))
 251                 ret = memory_block_change_state(mem, MEM_ONLINE, MEM_OFFLINE);
 252         else if(!strncmp(buf, "offline", min((int)count, 7)))
 253                 ret = memory_block_change_state(mem, MEM_OFFLINE, MEM_ONLINE);
 254 out:
 255         if (ret)
 256                 return ret;
 257         return count;
 258 }
 259
 260 /*
 261  * phys_device is a bad name for this.  What I really want
 262  * is a way to differentiate between memory ranges that
 263  * are part of physical devices that constitute
 264  * a complete removable unit or fru.
 265  * i.e. do these ranges belong to the same physical device,
 266  * s.t. if I offline all of these sections I can then
 267  * remove the physical device?
 268  */
 269 static ssize_t show_phys_device(struct sys_device *dev,
 270                                 struct sysdev_attribute *attr, char *buf)
 271 {
 272         struct memory_block *mem =
 273                 container_of(dev, struct memory_block, sysdev);
 274         return sprintf(buf, "%d\n", mem->phys_device);
 275 }
 276
 277 static SYSDEV_ATTR(phys_index, 0444, show_mem_phys_index, NULL);
 278 static SYSDEV_ATTR(state, 0644, show_mem_state, store_mem_state);
 279 static SYSDEV_ATTR(phys_device, 0444, show_phys_device, NULL);
 280 static SYSDEV_ATTR(removable, 0444, show_mem_removable, NULL);
 281
 282 #define mem_create_simple_file(mem, attr_name)  \
 283         sysdev_create_file(&mem->sysdev, &attr_##attr_name)
 284 #define mem_remove_simple_file(mem, attr_name)  \
 285         sysdev_remove_file(&mem->sysdev, &attr_##attr_name)
 286
 287 /*
 288  * Block size attribute stuff
 289  */
 290 static ssize_t
 291 print_block_size(struct class *class, char *buf)
 292 {
 293         return sprintf(buf, "%lx\n", (unsigned long)PAGES_PER_SECTION * PAGE_SIZE);
 294 }
 295
 296 static CLASS_ATTR(block_size_bytes, 0444, print_block_size, NULL);
 297
 298 static int block_size_init(void)
 299 {
 300         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 301                                 &class_attr_block_size_bytes.attr);
 302 }
 303
 304 /*
 305  * Some architectures will have custom drivers to do this, and
 306  * will not need to do it from userspace.  The fake hot-add code
 307  * as well as ppc64 will do all of their discovery in userspace
 308  * and will require this interface.
 309  */
 310 #ifdef CONFIG_ARCH_MEMORY_PROBE
 311 static ssize_t
 312 memory_probe_store(struct class *class, const char *buf, size_t count)
 313 {
 314         u64 phys_addr;
 315         int nid;
 316         int ret;
 317
 318         phys_addr = simple_strtoull(buf, NULL, 0);
 319
 320         nid = memory_add_physaddr_to_nid(phys_addr);
 321         ret = add_memory(nid, phys_addr, PAGES_PER_SECTION << PAGE_SHIFT);
 322
 323         if (ret)
 324                 count = ret;
 325
 326         return count;
 327 }
 328 static CLASS_ATTR(probe, 0700, NULL, memory_probe_store);
 329
 330 static int memory_probe_init(void)
 331 {
 332         return sysfs_create_file(&memory_sysdev_class.kset.kobj,
 333                                 &class_attr_probe.attr);
 334 }
 335 #else
 336 static inline int memory_probe_init(void)
 337 {
 338         return 0;
 339 }
 340 #endif
 341
 342 /*
 343  * Note that phys_device is optional.  It is here to allow for
 344  * differentiation between which *physical* devices each
 345  * section belongs to...
 346  */
 347
 348 static int add_memory_block(unsigned long node_id, struct mem_section *section,
 349                      unsigned long state, int phys_device)
 350 {
 351         struct memory_block *mem = kzalloc(sizeof(*mem), GFP_KERNEL);
 352         int ret = 0;
 353
 354         if (!mem)
 355                 return -ENOMEM;
 356
 357         mem->phys_index = __section_nr(section);
 358         mem->state = state;
 359         mutex_init(&mem->state_mutex);
 360         mem->phys_device = phys_device;
 361
 362         ret = register_memory(mem, section);
 363         if (!ret)
 364                 ret = mem_create_simple_file(mem, phys_index);
 365         if (!ret)
 366                 ret = mem_create_simple_file(mem, state);
 367         if (!ret)
 368                 ret = mem_create_simple_file(mem, phys_device);
 369         if (!ret)
 370                 ret = mem_create_simple_file(mem, removable);
 371
 372         return ret;
 373 }
 374
 375 /*
 376  * For now, we have a linear search to go find the appropriate
 377  * memory_block corresponding to a particular phys_index. If
 378  * this gets to be a real problem, we can always use a radix
 379  * tree or something here.
 380  *
 381  * This could be made generic for all sysdev classes.
 382  */
 383 static struct memory_block *find_memory_block(struct mem_section *section)
 384 {
 385         struct kobject *kobj;
 386         struct sys_device *sysdev;
 387         struct memory_block *mem;
 388         char name[sizeof(MEMORY_CLASS_NAME) + 9 + 1];
 389
 390         /*
 391          * This only works because we know that section == sysdev->id
 392          * slightly redundant with sysdev_register()
 393          */
 394         sprintf(&name[0], "%s%d", MEMORY_CLASS_NAME, __section_nr(section));
 395
 396         kobj = kset_find_obj(&memory_sysdev_class.kset, name);
 397         if (!kobj)
 398                 return NULL;
 399
 400         sysdev = container_of(kobj, struct sys_device, kobj);
 401         mem = container_of(sysdev, struct memory_block, sysdev);
 402
 403         return mem;
 404 }
 405
 406 int remove_memory_block(unsigned long node_id, struct mem_section *section,
 407                 int phys_device)
 408 {
 409         struct memory_block *mem;
 410
 411         mem = find_memory_block(section);
 412         mem_remove_simple_file(mem, phys_index);
 413         mem_remove_simple_file(mem, state);
 414         mem_remove_simple_file(mem, phys_device);
 415         mem_remove_simple_file(mem, removable);
 416         unregister_memory(mem, section);
 417
 418         return 0;
 419 }
 420
 421 /*
 422  * need an interface for the VM to add new memory regions,
 423  * but without onlining it.
 424  */
 425 int register_new_memory(struct mem_section *section)
 426 {
 427         return add_memory_block(0, section, MEM_OFFLINE, 0);
 428 }
 429
 430 int unregister_memory_section(struct mem_section *section)
 431 {
 432         if (!present_section(section))
 433                 return -EINVAL;
 434
 435         return remove_memory_block(0, section, 0);
 436 }
 437
 438 /*
 439  * Initialize the sysfs support for memory devices...
 440  */
 441 int __init memory_dev_init(void)
 442 {
 443         unsigned int i;
 444         int ret;
 445         int err;
 446
 447         memory_sysdev_class.kset.uevent_ops = &memory_uevent_ops;
 448         ret = sysdev_class_register(&memory_sysdev_class);
 449         if (ret)
 450                 goto out;
 451
 452         /*
 453          * Create entries for memory sections that were found
 454          * during boot and have been initialized
 455          */
 456         for (i = 0; i < NR_MEM_SECTIONS; i++) {
 457                 if (!present_section_nr(i))
 458                         continue;
 459                 err = add_memory_block(0, __nr_to_section(i), MEM_ONLINE, 0);
 460                 if (!ret)
 461                         ret = err;
 462         }
 463
 464         err = memory_probe_init();
 465         if (!ret)
 466                 ret = err;
 467         err = block_size_init();
 468         if (!ret)
 469                 ret = err;
 470 out:
 471         if (ret)
 472                 printk(KERN_ERR "%s() failed: %d\n", __func__, ret);
 473         return ret;
 474 }