mm/page_ext.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/mm.h>
   3 #include <linux/mmzone.h>
   4 #include <linux/bootmem.h>
   5 #include <linux/page_ext.h>
   6 #include <linux/memory.h>
   7 #include <linux/vmalloc.h>
   8 #include <linux/kmemleak.h>
   9 #include <linux/page_owner.h>
  10 #include <linux/page_idle.h>
  11
  12 /*
  13  * struct page extension
  14  *
  15  * This is the feature to manage memory for extended data per page.
  16  *
  17  * Until now, we must modify struct page itself to store extra data per page.
  18  * This requires rebuilding the kernel and it is really time consuming process.
  19  * And, sometimes, rebuild is impossible due to third party module dependency.
  20  * At last, enlarging struct page could cause un-wanted system behaviour change.
  21  *
  22  * This feature is intended to overcome above mentioned problems. This feature
  23  * allocates memory for extended data per page in certain place rather than
  24  * the struct page itself. This memory can be accessed by the accessor
  25  * functions provided by this code. During the boot process, it checks whether
  26  * allocation of huge chunk of memory is needed or not. If not, it avoids
  27  * allocating memory at all. With this advantage, we can include this feature
  28  * into the kernel in default and can avoid rebuild and solve related problems.
  29  *
  30  * To help these things to work well, there are two callbacks for clients. One
  31  * is the need callback which is mandatory if user wants to avoid useless
  32  * memory allocation at boot-time. The other is optional, init callback, which
  33  * is used to do proper initialization after memory is allocated.
  34  *
  35  * The need callback is used to decide whether extended memory allocation is
  36  * needed or not. Sometimes users want to deactivate some features in this
  37  * boot and extra memory would be unneccessary. In this case, to avoid
  38  * allocating huge chunk of memory, each clients represent their need of
  39  * extra memory through the need callback. If one of the need callbacks
  40  * returns true, it means that someone needs extra memory so that
  41  * page extension core should allocates memory for page extension. If
  42  * none of need callbacks return true, memory isn't needed at all in this boot
  43  * and page extension core can skip to allocate memory. As result,
  44  * none of memory is wasted.
  45  *
  46  * When need callback returns true, page_ext checks if there is a request for
  47  * extra memory through size in struct page_ext_operations. If it is non-zero,
  48  * extra space is allocated for each page_ext entry and offset is returned to
  49  * user through offset in struct page_ext_operations.
  50  *
  51  * The init callback is used to do proper initialization after page extension
  52  * is completely initialized. In sparse memory system, extra memory is
  53  * allocated some time later than memmap is allocated. In other words, lifetime
  54  * of memory for page extension isn't same with memmap for struct page.
  55  * Therefore, clients can't store extra data until page extension is
  56  * initialized, even if pages are allocated and used freely. This could
  57  * cause inadequate state of extra data per page, so, to prevent it, client
  58  * can utilize this callback to initialize the state of it correctly.
  59  */
  60
  61 static struct page_ext_operations *page_ext_ops[] = {
  62         &debug_guardpage_ops,
  63 #ifdef CONFIG_PAGE_OWNER
  64         &page_owner_ops,
  65 #endif
  66 #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
  67         &page_idle_ops,
  68 #endif
  69 };
  70
  71 static unsigned long total_usage;
  72 static unsigned long extra_mem;
  73
  74 static bool __init invoke_need_callbacks(void)
  75 {
  76         int i;
  77         int entries = ARRAY_SIZE(page_ext_ops);
  78         bool need = false;
  79
  80         for (i = 0; i < entries; i++) {
  81                 if (page_ext_ops[i]->need && page_ext_ops[i]->need()) {
  82                         page_ext_ops[i]->offset = sizeof(struct page_ext) +
  83                                                 extra_mem;
  84                         extra_mem += page_ext_ops[i]->size;
  85                         need = true;
  86                 }
  87         }
  88
  89         return need;
  90 }
  91
  92 static void __init invoke_init_callbacks(void)
  93 {
  94         int i;
  95         int entries = ARRAY_SIZE(page_ext_ops);
  96
  97         for (i = 0; i < entries; i++) {
  98                 if (page_ext_ops[i]->init)
  99                         page_ext_ops[i]->init();
 100         }
 101 }
 102
 103 static unsigned long get_entry_size(void)
 104 {
 105         return sizeof(struct page_ext) + extra_mem;
 106 }
 107
 108 static inline struct page_ext *get_entry(void *base, unsigned long index)
 109 {
 110         return base + get_entry_size() * index;
 111 }
 112
 113 #if !defined(CONFIG_SPARSEMEM)
 114
 115
 116 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
 117 {
 118         pgdat->node_page_ext = NULL;
 119 }
 120
 121 struct page_ext *lookup_page_ext(struct page *page)
 122 {
 123         unsigned long pfn = page_to_pfn(page);
 124         unsigned long index;
 125         struct page_ext *base;
 126
 127         base = NODE_DATA(page_to_nid(page))->node_page_ext;
 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         index = pfn - round_down(node_start_pfn(page_to_nid(page)),
 137                                         MAX_ORDER_NR_PAGES);
 138         return get_entry(base, index);
 139 }
 140
 141 static int __init alloc_node_page_ext(int nid)
 142 {
 143         struct page_ext *base;
 144         unsigned long table_size;
 145         unsigned long nr_pages;
 146
 147         nr_pages = NODE_DATA(nid)->node_spanned_pages;
 148         if (!nr_pages)
 149                 return 0;
 150
 151         /*
 152          * Need extra space if node range is not aligned with
 153          * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
 154          * checks buddy's status, range could be out of exact node range.
 155          */
 156         if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
 157                 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
 158                 nr_pages += MAX_ORDER_NR_PAGES;
 159
 160         table_size = get_entry_size() * nr_pages;
 161
 162         base = memblock_virt_alloc_try_nid_nopanic(
 163                         table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 164                         BOOTMEM_ALLOC_ACCESSIBLE, nid);
 165         if (!base)
 166                 return -ENOMEM;
 167         NODE_DATA(nid)->node_page_ext = base;
 168         total_usage += table_size;
 169         return 0;
 170 }
 171
 172 void __init page_ext_init_flatmem(void)
 173 {
 174
 175         int nid, fail;
 176
 177         if (!invoke_need_callbacks())
 178                 return;
 179
 180         for_each_online_node(nid)  {
 181                 fail = alloc_node_page_ext(nid);
 182                 if (fail)
 183                         goto fail;
 184         }
 185         pr_info("allocated %ld bytes of page_ext\n", total_usage);
 186         invoke_init_callbacks();
 187         return;
 188
 189 fail:
 190         pr_crit("allocation of page_ext failed.\n");
 191         panic("Out of memory");
 192 }
 193
 194 #else /* CONFIG_FLAT_NODE_MEM_MAP */
 195
 196 struct page_ext *lookup_page_ext(struct page *page)
 197 {
 198         unsigned long pfn = page_to_pfn(page);
 199         struct mem_section *section = __pfn_to_section(pfn);
 200         /*
 201          * The sanity checks the page allocator does upon freeing a
 202          * page can reach here before the page_ext arrays are
 203          * allocated when feeding a range of pages to the allocator
 204          * for the first time during bootup or memory hotplug.
 205          */
 206         if (!section->page_ext)
 207                 return NULL;
 208         return get_entry(section->page_ext, pfn);
 209 }
 210
 211 static void *__meminit alloc_page_ext(size_t size, int nid)
 212 {
 213         gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
 214         void *addr = NULL;
 215
 216         addr = alloc_pages_exact_nid(nid, size, flags);
 217         if (addr) {
 218                 kmemleak_alloc(addr, size, 1, flags);
 219                 return addr;
 220         }
 221
 222         addr = vzalloc_node(size, nid);
 223
 224         return addr;
 225 }
 226
 227 static int __meminit init_section_page_ext(unsigned long pfn, int nid)
 228 {
 229         struct mem_section *section;
 230         struct page_ext *base;
 231         unsigned long table_size;
 232
 233         section = __pfn_to_section(pfn);
 234
 235         if (section->page_ext)
 236                 return 0;
 237
 238         table_size = get_entry_size() * PAGES_PER_SECTION;
 239         base = alloc_page_ext(table_size, nid);
 240
 241         /*
 242          * The value stored in section->page_ext is (base - pfn)
 243          * and it does not point to the memory block allocated above,
 244          * causing kmemleak false positives.
 245          */
 246         kmemleak_not_leak(base);
 247
 248         if (!base) {
 249                 pr_err("page ext allocation failure\n");
 250                 return -ENOMEM;
 251         }
 252
 253         /*
 254          * The passed "pfn" may not be aligned to SECTION.  For the calculation
 255          * we need to apply a mask.
 256          */
 257         pfn &= PAGE_SECTION_MASK;
 258         section->page_ext = (void *)base - get_entry_size() * pfn;
 259         total_usage += table_size;
 260         return 0;
 261 }
 262 #ifdef CONFIG_MEMORY_HOTPLUG
 263 static void free_page_ext(void *addr)
 264 {
 265         if (is_vmalloc_addr(addr)) {
 266                 vfree(addr);
 267         } else {
 268                 struct page *page = virt_to_page(addr);
 269                 size_t table_size;
 270
 271                 table_size = get_entry_size() * PAGES_PER_SECTION;
 272
 273                 BUG_ON(PageReserved(page));
 274                 free_pages_exact(addr, table_size);
 275         }
 276 }
 277
 278 static void __free_page_ext(unsigned long pfn)
 279 {
 280         struct mem_section *ms;
 281         struct page_ext *base;
 282
 283         ms = __pfn_to_section(pfn);
 284         if (!ms || !ms->page_ext)
 285                 return;
 286         base = get_entry(ms->page_ext, pfn);
 287         free_page_ext(base);
 288         ms->page_ext = NULL;
 289 }
 290
 291 static int __meminit online_page_ext(unsigned long start_pfn,
 292                                 unsigned long nr_pages,
 293                                 int nid)
 294 {
 295         unsigned long start, end, pfn;
 296         int fail = 0;
 297
 298         start = SECTION_ALIGN_DOWN(start_pfn);
 299         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 300
 301         if (nid == -1) {
 302                 /*
 303                  * In this case, "nid" already exists and contains valid memory.
 304                  * "start_pfn" passed to us is a pfn which is an arg for
 305                  * online__pages(), and start_pfn should exist.
 306                  */
 307                 nid = pfn_to_nid(start_pfn);
 308                 VM_BUG_ON(!node_state(nid, N_ONLINE));
 309         }
 310
 311         for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
 312                 if (!pfn_present(pfn))
 313                         continue;
 314                 fail = init_section_page_ext(pfn, nid);
 315         }
 316         if (!fail)
 317                 return 0;
 318
 319         /* rollback */
 320         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 321                 __free_page_ext(pfn);
 322
 323         return -ENOMEM;
 324 }
 325
 326 static int __meminit offline_page_ext(unsigned long start_pfn,
 327                                 unsigned long nr_pages, int nid)
 328 {
 329         unsigned long start, end, pfn;
 330
 331         start = SECTION_ALIGN_DOWN(start_pfn);
 332         end = SECTION_ALIGN_UP(start_pfn + nr_pages);
 333
 334         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
 335                 __free_page_ext(pfn);
 336         return 0;
 337
 338 }
 339
 340 static int __meminit page_ext_callback(struct notifier_block *self,
 341                                unsigned long action, void *arg)
 342 {
 343         struct memory_notify *mn = arg;
 344         int ret = 0;
 345
 346         switch (action) {
 347         case MEM_GOING_ONLINE:
 348                 ret = online_page_ext(mn->start_pfn,
 349                                    mn->nr_pages, mn->status_change_nid);
 350                 break;
 351         case MEM_OFFLINE:
 352                 offline_page_ext(mn->start_pfn,
 353                                 mn->nr_pages, mn->status_change_nid);
 354                 break;
 355         case MEM_CANCEL_ONLINE:
 356                 offline_page_ext(mn->start_pfn,
 357                                 mn->nr_pages, mn->status_change_nid);
 358                 break;
 359         case MEM_GOING_OFFLINE:
 360                 break;
 361         case MEM_ONLINE:
 362         case MEM_CANCEL_OFFLINE:
 363                 break;
 364         }
 365
 366         return notifier_from_errno(ret);
 367 }
 368
 369 #endif
 370
 371 void __init page_ext_init(void)
 372 {
 373         unsigned long pfn;
 374         int nid;
 375
 376         if (!invoke_need_callbacks())
 377                 return;
 378
 379         for_each_node_state(nid, N_MEMORY) {
 380                 unsigned long start_pfn, end_pfn;
 381
 382                 start_pfn = node_start_pfn(nid);
 383                 end_pfn = node_end_pfn(nid);
 384                 /*
 385                  * start_pfn and end_pfn may not be aligned to SECTION and the
 386                  * page->flags of out of node pages are not initialized.  So we
 387                  * scan [start_pfn, the biggest section's pfn < end_pfn) here.
 388                  */
 389                 for (pfn = start_pfn; pfn < end_pfn;
 390                         pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
 391
 392                         if (!pfn_valid(pfn))
 393                                 continue;
 394                         /*
 395                          * Nodes's pfns can be overlapping.
 396                          * We know some arch can have a nodes layout such as
 397                          * -------------pfn-------------->
 398                          * N0 | N1 | N2 | N0 | N1 | N2|....
 399                          *
 400                          * Take into account DEFERRED_STRUCT_PAGE_INIT.
 401                          */
 402                         if (early_pfn_to_nid(pfn) != nid)
 403                                 continue;
 404                         if (init_section_page_ext(pfn, nid))
 405                                 goto oom;
 406                         cond_resched();
 407                 }
 408         }
 409         hotplug_memory_notifier(page_ext_callback, 0);
 410         pr_info("allocated %ld bytes of page_ext\n", total_usage);
 411         invoke_init_callbacks();
 412         return;
 413
 414 oom:
 415         panic("Out of memory");
 416 }
 417
 418 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
 419 {
 420 }
 421
 422 #endif