mm/page_ext.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include <linux/mm.h>
   3 #include <linux/mmzone.h>
   4 #include <linux/memblock.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 #ifdef CONFIG_DEBUG_PAGEALLOC
  63         &debug_guardpage_ops,
  64 #endif
  65 #ifdef CONFIG_PAGE_OWNER
  66         &page_owner_ops,
  67 #endif
  68 #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
  69         &page_idle_ops,
  70 #endif
  71 };
  72
  73 static unsigned long total_usage;
  74 static unsigned long extra_mem;
  75
  76 static bool __init invoke_need_callbacks(void)
  77 {
  78         int i;
  79         int entries = ARRAY_SIZE(page_ext_ops);
  80         bool need = false;
  81
  82         for (i = 0; i < entries; i++) {
  83                 if (page_ext_ops[i]->need && page_ext_ops[i]->need()) {
  84                         page_ext_ops[i]->offset = sizeof(struct page_ext) +
  85                                                 extra_mem;
  86                         extra_mem += page_ext_ops[i]->size;
  87                         need = true;
  88                 }
  89         }
  90
  91         return need;
  92 }
  93
  94 static void __init invoke_init_callbacks(void)
  95 {
  96         int i;
  97         int entries = ARRAY_SIZE(page_ext_ops);
  98
  99         for (i = 0; i < entries; i++) {
 100                 if (page_ext_ops[i]->init)
 101                         page_ext_ops[i]->init();
 102         }
 103 }
 104
 105 static unsigned long get_entry_size(void)
 106 {
 107         return sizeof(struct page_ext) + extra_mem;
 108 }
 109
 110 static inline struct page_ext *get_entry(void *base, unsigned long index)
 111 {
 112         return base + get_entry_size() * index;
 113 }
 114
 115 #if !defined(CONFIG_SPARSEMEM)
 116
 117
 118 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
 119 {
 120         pgdat->node_page_ext = NULL;
 121 }
 122
 123 struct page_ext *lookup_page_ext(const struct page *page)
 124 {
 125         unsigned long pfn = page_to_pfn(page);
 126         unsigned long index;
 127         struct page_ext *base;
 128
 129         base = NODE_DATA(page_to_nid(page))->node_page_ext;
 130         /*
 131          * The sanity checks the page allocator does upon freeing a
 132          * page can reach here before the page_ext arrays are
 133          * allocated when feeding a range of pages to the allocator
 134          * for the first time during bootup or memory hotplug.
 135          */
 136         if (unlikely(!base))
 137                 return NULL;
 138         index = pfn - round_down(node_start_pfn(page_to_nid(page)),
 139                                         MAX_ORDER_NR_PAGES);
 140         return get_entry(base, index);
 141 }
 142
 143 static int __init alloc_node_page_ext(int nid)
 144 {
 145         struct page_ext *base;
 146         unsigned long table_size;
 147         unsigned long nr_pages;
 148
 149         nr_pages = NODE_DATA(nid)->node_spanned_pages;
 150         if (!nr_pages)
 151                 return 0;
 152
 153         /*
 154          * Need extra space if node range is not aligned with
 155          * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
 156          * checks buddy's status, range could be out of exact node range.
 157          */
 158         if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
 159                 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
 160                 nr_pages += MAX_ORDER_NR_PAGES;
 161
 162         table_size = get_entry_size() * nr_pages;
 163
 164         base = memblock_alloc_try_nid(
 165                         table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 166                         MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 167         if (!base)
 168                 return -ENOMEM;
 169         NODE_DATA(nid)->node_page_ext = base;
 170         total_usage += table_size;
 171         return 0;
 172 }
 173
 174 void __init page_ext_init_flatmem(void)
 175 {
 176
 177         int nid, fail;
 178
 179         if (!invoke_need_callbacks())
 180                 return;
 181
 182         for_each_online_node(nid)  {
 183                 fail = alloc_node_page_ext(nid);
 184                 if (fail)
 185                         goto fail;
 186         }
 187         pr_info("allocated %ld bytes of page_ext\n", total_usage);
 188         invoke_init_callbacks();
 189         return;
 190
 191 fail:
 192         pr_crit("allocation of page_ext failed.\n");
 193         panic("Out of memory");
 194 }
 195
 196 #else /* CONFIG_FLAT_NODE_MEM_MAP */
 197
 198 struct page_ext *lookup_page_ext(const struct page *page)
 199 {
 200         unsigned long pfn = page_to_pfn(page);
 201         struct mem_section *section = __pfn_to_section(pfn);
 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         return get_entry(section->page_ext, pfn);
 211 }
 212
 213 static void *__meminit alloc_page_ext(size_t size, int nid)
 214 {
 215         gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
 216         void *addr = NULL;
 217
 218         addr = alloc_pages_exact_nid(nid, size, flags);
 219         if (addr) {
 220                 kmemleak_alloc(addr, size, 1, flags);
 221                 return addr;
 222         }
 223
 224         addr = vzalloc_node(size, nid);
 225
 226         return addr;
 227 }
 228
 229 static int __meminit init_section_page_ext(unsigned long pfn, int nid)
 230 {
 231         struct mem_section *section;
 232         struct page_ext *base;
 233         unsigned long table_size;
 234
 235         section = __pfn_to_section(pfn);
 236
 237         if (section->page_ext)
 238                 return 0;
 239
 240         table_size = get_entry_size() * PAGES_PER_SECTION;
 241         base = alloc_page_ext(table_size, nid);
 242
 243         /*
 244          * The value stored in section->page_ext is (base - pfn)
 245          * and it does not point to the memory block allocated above,
 246          * causing kmemleak false positives.
 247          */
 248         kmemleak_not_leak(base);
 249
 250         if (!base) {
 251                 pr_err("page ext allocation failure\n");
 252                 return -ENOMEM;
 253         }
 254
 255         /*
 256          * The passed "pfn" may not be aligned to SECTION.  For the calculation
 257          * we need to apply a mask.
 258          */
 259         pfn &= PAGE_SECTION_MASK;
 260         section->page_ext = (void *)base - get_entry_size() * pfn;
 261         total_usage += table_size;
 262         return 0;
 263 }
 264 #ifdef CONFIG_MEMORY_HOTPLUG
 265 static void free_page_ext(void *addr)
 266 {
 267         if (is_vmalloc_addr(addr)) {
 268                 vfree(addr);
 269         } else {
 270                 struct page *page = virt_to_page(addr);
 271                 size_t table_size;
 272
 273                 table_size = get_entry_size() * PAGES_PER_SECTION;
 274
 275                 BUG_ON(PageReserved(page));
 276                 kmemleak_free(addr);
 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 == NUMA_NO_NODE) {
 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                         if (pfn_to_nid(pfn) != nid)
 404                                 continue;
 405                         if (init_section_page_ext(pfn, nid))
 406                                 goto oom;
 407                         cond_resched();
 408                 }
 409         }
 410         hotplug_memory_notifier(page_ext_callback, 0);
 411         pr_info("allocated %ld bytes of page_ext\n", total_usage);
 412         invoke_init_callbacks();
 413         return;
 414
 415 oom:
 416         panic("Out of memory");
 417 }
 418
 419 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
 420 {
 421 }
 422
 423 #endif