mm/page_ext.c

   1 #include <linux/mm.h>
   2 #include <linux/mmzone.h>
   3 #include <linux/bootmem.h>
   4 #include <linux/page_ext.h>
   5 #include <linux/memory.h>
   6 #include <linux/vmalloc.h>
   7 #include <linux/kmemleak.h>
   8 #include <linux/page_owner.h>
   9 #include <linux/page_idle.h>
  10
  11 /*
  12  * struct page extension
  13  *
  14  * This is the feature to manage memory for extended data per page.
  15  *
  16  * Until now, we must modify struct page itself to store extra data per page.
  17  * This requires rebuilding the kernel and it is really time consuming process.
  18  * And, sometimes, rebuild is impossible due to third party module dependency.
  19  * At last, enlarging struct page could cause un-wanted system behaviour change.
  20  *
  21  * This feature is intended to overcome above mentioned problems. This feature
  22  * allocates memory for extended data per page in certain place rather than
  23  * the struct page itself. This memory can be accessed by the accessor
  24  * functions provided by this code. During the boot process, it checks whether
  25  * allocation of huge chunk of memory is needed or not. If not, it avoids
  26  * allocating memory at all. With this advantage, we can include this feature
  27  * into the kernel in default and can avoid rebuild and solve related problems.
  28  *
  29  * To help these things to work well, there are two callbacks for clients. One
  30  * is the need callback which is mandatory if user wants to avoid useless
  31  * memory allocation at boot-time. The other is optional, init callback, which
  32  * is used to do proper initialization after memory is allocated.
  33  *
  34  * The need callback is used to decide whether extended memory allocation is
  35  * needed or not. Sometimes users want to deactivate some features in this
  36  * boot and extra memory would be unneccessary. In this case, to avoid
  37  * allocating huge chunk of memory, each clients represent their need of
  38  * extra memory through the need callback. If one of the need callbacks
  39  * returns true, it means that someone needs extra memory so that
  40  * page extension core should allocates memory for page extension. If
  41  * none of need callbacks return true, memory isn't needed at all in this boot
  42  * and page extension core can skip to allocate memory. As result,
  43  * none of memory is wasted.
  44  *
  45  * When need callback returns true, page_ext checks if there is a request for
  46  * extra memory through size in struct page_ext_operations. If it is non-zero,
  47  * extra space is allocated for each page_ext entry and offset is returned to
  48  * user through offset in struct page_ext_operations.
  49  *
  50  * The init callback is used to do proper initialization after page extension
  51  * is completely initialized. In sparse memory system, extra memory is
  52  * allocated some time later than memmap is allocated. In other words, lifetime
  53  * of memory for page extension isn't same with memmap for struct page.
  54  * Therefore, clients can't store extra data until page extension is
  55  * initialized, even if pages are allocated and used freely. This could
  56  * cause inadequate state of extra data per page, so, to prevent it, client
  57  * can utilize this callback to initialize the state of it correctly.
  58  */
  59
  60 static struct page_ext_operations *page_ext_ops[] = {
  61         &debug_guardpage_ops,
  62 #ifdef CONFIG_PAGE_OWNER
  63         &page_owner_ops,
  64 #endif
  65 #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
  66         &page_idle_ops,
  67 #endif
  68 };
  69
  70 static unsigned long total_usage;
  71 static unsigned long extra_mem;
  72
  73 static bool __init invoke_need_callbacks(void)
  74 {
  75         int i;
  76         int entries = ARRAY_SIZE(page_ext_ops);
  77         bool need = false;
  78
  79         for (i = 0; i < entries; i++) {
  80                 if (page_ext_ops[i]->need && page_ext_ops[i]->need()) {
  81                         page_ext_ops[i]->offset = sizeof(struct page_ext) +
  82                                                 extra_mem;
  83                         extra_mem += page_ext_ops[i]->size;
  84                         need = true;
  85                 }
  86         }
  87
  88         return need;
  89 }
  90
  91 static void __init invoke_init_callbacks(void)
  92 {
  93         int i;
  94         int entries = ARRAY_SIZE(page_ext_ops);
  95
  96         for (i = 0; i < entries; i++) {
  97                 if (page_ext_ops[i]->init)
  98                         page_ext_ops[i]->init();
  99         }
 100 }
 101
 102 static unsigned long get_entry_size(void)
 103 {
 104         return sizeof(struct page_ext) + extra_mem;
 105 }
 106
 107 static inline struct page_ext *get_entry(void *base, unsigned long index)
 108 {
 109         return base + get_entry_size() * index;
 110 }
 111
 112 #if !defined(CONFIG_SPARSEMEM)
 113
 114
 115 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
 116 {
 117         pgdat->node_page_ext = NULL;
 118 }
 119
 120 struct page_ext *lookup_page_ext(struct page *page)
 121 {
 122         unsigned long pfn = page_to_pfn(page);
 123         unsigned long index;
 124         struct page_ext *base;
 125
 126         base = NODE_DATA(page_to_nid(page))->node_page_ext;
 127 #if defined(CONFIG_DEBUG_VM)
 128         /*
 129          * The sanity checks the page allocator does upon freeing a
 130          * page can reach here before the page_ext arrays are
 131          * allocated when feeding a range of pages to the allocator
 132          * for the first time during bootup or memory hotplug.
 133          */
 134         if (unlikely(!base))
 135                 return NULL;
 136 #endif
 137         index = pfn - round_down(node_start_pfn(page_to_nid(page)),
 138                                         MAX_ORDER_NR_PAGES);
 139         return get_entry(base, index);
 140 }
 141
 142 static int __init alloc_node_page_ext(int nid)
 143 {
 144         struct page_ext *base;
 145         unsigned long table_size;
 146         unsigned long nr_pages;
 147
 148         nr_pages = NODE_DATA(nid)->node_spanned_pages;
 149         if (!nr_pages)
 150                 return 0;
 151
 152         /*
 153          * Need extra space if node range is not aligned with
 154          * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
 155          * checks buddy's status, range could be out of exact node range.
 156          */
 157         if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
 158                 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
 159                 nr_pages += MAX_ORDER_NR_PAGES;
 160
 161         table_size = get_entry_size() * nr_pages;
 162
 163         base = memblock_virt_alloc_try_nid_nopanic(
 164                         table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 165                         BOOTMEM_ALLOC_ACCESSIBLE, nid);
 166         if (!base)
 167                 return -ENOMEM;
 168         NODE_DATA(nid)->node_page_ext = base;
 169         total_usage += table_size;
 170         return 0;
 171 }
 172
 173 void __init page_ext_init_flatmem(void)
 174 {
 175
 176         int nid, fail;
 177
 178         if (!invoke_need_callbacks())
 179                 return;
 180
 181         for_each_online_node(nid)  {
 182                 fail = alloc_node_page_ext(nid);
 183                 if (fail)
 184                         goto fail;
 185         }
 186         pr_info("allocated %ld bytes of page_ext\n", total_usage);
 187         invoke_init_callbacks();
 188         return;
 189
 190 fail:
 191         pr_crit("allocation of page_ext failed.\n");
 192         panic("Out of memory");
 193 }
 194
 195 #else /* CONFIG_FLAT_NODE_MEM_MAP */
 196
 197 struct page_ext *lookup_page_ext(struct page *page)
 198 {
 199         unsigned long pfn = page_to_pfn(page);
 200         struct mem_section *section = __pfn_to_section(pfn);
 201 #if defined(CONFIG_DEBUG_VM)
 202         /*
 203          * The sanity checks the page allocator does upon freeing a
 204          * page can reach here before the page_ext arrays are
 205          * allocated when feeding a range of pages to the allocator
 206          * for the first time during bootup or memory hotplug.
 207          */
 208         if (!section->page_ext)
 209                 return NULL;
 210 #endif
 211         return get_entry(section->page_ext, pfn);
 212 }
 213
 214 static void *__meminit alloc_page_ext(size_t size, int nid)
 215 {
 216         gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
 217         void *addr = NULL;
 218
 219         addr = alloc_pages_exact_nid(nid, size, flags);
 220         if (addr) {
 221                 kmemleak_alloc(addr, size, 1, flags);
 222                 return addr;
 223         }
 224
 225         addr = vzalloc_node(size, nid);
 226
 227         return addr;
 228 }
 229
 230 static int __meminit init_section_page_ext(unsigned long pfn, int nid)
 231 {
 232         struct mem_section *section;
 233         struct page_ext *base;
 234         unsigned long table_size;
 235
 236         section = __pfn_to_section(pfn);
 237
 238         if (section->page_ext)
 239                 return 0;
 240
 241         table_size = get_entry_size() * PAGES_PER_SECTION;
 242         base = alloc_page_ext(table_size, nid);
 243
 244         /*
 245          * The value stored in section->page_ext is (base - pfn)
 246          * and it does not point to the memory block allocated above,
 247          * causing kmemleak false positives.
 248          */
 249         kmemleak_not_leak(base);
 250
 251         if (!base) {
 252                 pr_err("page ext allocation failure\n");
 253                 return -ENOMEM;
 254         }
 255
 256         /*
 257          * The passed "pfn" may not be aligned to SECTION.  For the calculation
 258          * we need to apply a mask.
 259          */
 260         pfn &= PAGE_SECTION_MASK;
 261         section->page_ext = (void *)base - get_entry_size() * pfn;
 262         total_usage += table_size;
 263         return 0;
 264 }
 265 #ifdef CONFIG_MEMORY_HOTPLUG
 266 static void free_page_ext(void *addr)
 267 {
 268         if (is_vmalloc_addr(addr)) {
 269                 vfree(addr);
 270         } else {
 271                 struct page *page = virt_to_page(addr);
 272                 size_t table_size;
 273
 274                 table_size = get_entry_size() * PAGES_PER_SECTION;
 275
 276                 BUG_ON(PageReserved(page));
 277                 free_pages_exact(addr, table_size);
 278         }
 279 }
 280
 281 static void __free_page_ext(unsigned long pfn)
 282 {
 283         struct mem_section *ms;
 284         struct page_ext *base;
 285
 286         ms = __pfn_to_section(pfn);
 287         if (!ms || !ms->page_ext)
 288                 return;
 289         base = get_entry(ms->page_ext, pfn);
 290         free_page_ext(base);
 291         ms->page_ext = NULL;
 292 }
 293
 294 static int __meminit online_page_ext(unsigned long start_pfn,
 295                                 unsigned long nr_pages,
 296                                 int nid)
 297 {
 298         unsigned long start, end, pfn;
 299         int fail = 0;
 300
 301         start = SECTION_ALIGN_DOWN(start_pfn);
 302         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 303
 304         if (nid == -1) {
 305                 /*
 306                  * In this case, "nid" already exists and contains valid memory.
 307                  * "start_pfn" passed to us is a pfn which is an arg for
 308                  * online__pages(), and start_pfn should exist.
 309                  */
 310                 nid = pfn_to_nid(start_pfn);
 311                 VM_BUG_ON(!node_state(nid, N_ONLINE));
 312         }
 313
 314         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
 315                 if (!pfn_present(pfn))
 316                         continue;
 317                 fail = init_section_page_ext(pfn, nid);
 318         }
 319         if (!fail)
 320                 return 0;
 321
 322         /* rollback */
 323         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 324                 __free_page_ext(pfn);
 325
 326         return -ENOMEM;
 327 }
 328
 329 static int __meminit offline_page_ext(unsigned long start_pfn,
 330                                 unsigned long nr_pages, int nid)
 331 {
 332         unsigned long start, end, pfn;
 333
 334         start = SECTION_ALIGN_DOWN(start_pfn);
 335         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 336
 337         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 338                 __free_page_ext(pfn);
 339         return 0;
 340
 341 }
 342
 343 static int __meminit page_ext_callback(struct notifier_block *self,
 344                                unsigned long action, void *arg)
 345 {
 346         struct memory_notify *mn = arg;
 347         int ret = 0;
 348
 349         switch (action) {
 350         case MEM_GOING_ONLINE:
 351                 ret = online_page_ext(mn->start_pfn,
 352                                    mn->nr_pages, mn->status_change_nid);
 353                 break;
 354         case MEM_OFFLINE:
 355                 offline_page_ext(mn->start_pfn,
 356                                 mn->nr_pages, mn->status_change_nid);
 357                 break;
 358         case MEM_CANCEL_ONLINE:
 359                 offline_page_ext(mn->start_pfn,
 360                                 mn->nr_pages, mn->status_change_nid);
 361                 break;
 362         case MEM_GOING_OFFLINE:
 363                 break;
 364         case MEM_ONLINE:
 365         case MEM_CANCEL_OFFLINE:
 366                 break;
 367         }
 368
 369         return notifier_from_errno(ret);
 370 }
 371
 372 #endif
 373
 374 void __init page_ext_init(void)
 375 {
 376         unsigned long pfn;
 377         int nid;
 378
 379         if (!invoke_need_callbacks())
 380                 return;
 381
 382         for_each_node_state(nid, N_MEMORY) {
 383                 unsigned long start_pfn, end_pfn;
 384
 385                 start_pfn = node_start_pfn(nid);
 386                 end_pfn = node_end_pfn(nid);
 387                 /*
 388                  * start_pfn and end_pfn may not be aligned to SECTION and the
 389                  * page->flags of out of node pages are not initialized.  So we
 390                  * scan [start_pfn, the biggest section's pfn < end_pfn) here.
 391                  */
 392                 for (pfn = start_pfn; pfn < end_pfn;
 393                         pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
 394
 395                         if (!pfn_valid(pfn))
 396                                 continue;
 397                         /*
 398                          * Nodes's pfns can be overlapping.
 399                          * We know some arch can have a nodes layout such as
 400                          * -------------pfn-------------->
 401                          * N0 | N1 | N2 | N0 | N1 | N2|....
 402                          *
 403                          * Take into account DEFERRED_STRUCT_PAGE_INIT.
 404                          */
 405                         if (early_pfn_to_nid(pfn) != nid)
 406                                 continue;
 407                         if (init_section_page_ext(pfn, nid))
 408                                 goto oom;
 409                         cond_resched();
 410                 }
 411         }
 412         hotplug_memory_notifier(page_ext_callback, 0);
 413         pr_info("allocated %ld bytes of page_ext\n", total_usage);
 414         invoke_init_callbacks();
 415         return;
 416
 417 oom:
 418         panic("Out of memory");
 419 }
 420
 421 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
 422 {
 423 }
 424
 425 #endif