Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 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> |
Joonsoo Kim | 48c96a3 | 2014-12-12 16:56:01 -0800 | [diff] [blame] | 8 | #include <linux/page_owner.h> |
Vladimir Davydov | 33c3fc7 | 2015-09-09 15:35:45 -0700 | [diff] [blame] | 9 | #include <linux/page_idle.h> |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 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 | * |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 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 | * |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 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[] = { |
Joonsoo Kim | e30825f | 2014-12-12 16:55:49 -0800 | [diff] [blame] | 61 | &debug_guardpage_ops, |
Joonsoo Kim | 48c96a3 | 2014-12-12 16:56:01 -0800 | [diff] [blame] | 62 | #ifdef CONFIG_PAGE_OWNER |
| 63 | &page_owner_ops, |
| 64 | #endif |
Vladimir Davydov | 33c3fc7 | 2015-09-09 15:35:45 -0700 | [diff] [blame] | 65 | #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT) |
| 66 | &page_idle_ops, |
| 67 | #endif |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 68 | }; |
| 69 | |
| 70 | static unsigned long total_usage; |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 71 | static unsigned long extra_mem; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 72 | |
| 73 | static bool __init invoke_need_callbacks(void) |
| 74 | { |
| 75 | int i; |
| 76 | int entries = ARRAY_SIZE(page_ext_ops); |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 77 | bool need = false; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 78 | |
| 79 | for (i = 0; i < entries; i++) { |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 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 | } |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 86 | } |
| 87 | |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 88 | return need; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 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 | |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 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 | |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 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); |
Joonsoo Kim | 0b06bb3 | 2016-10-07 16:58:24 -0700 | [diff] [blame] | 123 | unsigned long index; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 124 | struct page_ext *base; |
| 125 | |
| 126 | base = NODE_DATA(page_to_nid(page))->node_page_ext; |
Laura Abbott | 1414c7f | 2016-03-15 14:56:30 -0700 | [diff] [blame] | 127 | #if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING) |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 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. |
Laura Abbott | 1414c7f | 2016-03-15 14:56:30 -0700 | [diff] [blame] | 133 | * |
| 134 | * This check is also necessary for ensuring page poisoning |
| 135 | * works as expected when enabled |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 136 | */ |
| 137 | if (unlikely(!base)) |
| 138 | return NULL; |
| 139 | #endif |
Joonsoo Kim | 0b06bb3 | 2016-10-07 16:58:24 -0700 | [diff] [blame] | 140 | index = pfn - round_down(node_start_pfn(page_to_nid(page)), |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 141 | MAX_ORDER_NR_PAGES); |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 142 | return get_entry(base, index); |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 143 | } |
| 144 | |
| 145 | static int __init alloc_node_page_ext(int nid) |
| 146 | { |
| 147 | struct page_ext *base; |
| 148 | unsigned long table_size; |
| 149 | unsigned long nr_pages; |
| 150 | |
| 151 | nr_pages = NODE_DATA(nid)->node_spanned_pages; |
| 152 | if (!nr_pages) |
| 153 | return 0; |
| 154 | |
| 155 | /* |
| 156 | * Need extra space if node range is not aligned with |
| 157 | * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm |
| 158 | * checks buddy's status, range could be out of exact node range. |
| 159 | */ |
| 160 | if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) || |
| 161 | !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES)) |
| 162 | nr_pages += MAX_ORDER_NR_PAGES; |
| 163 | |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 164 | table_size = get_entry_size() * nr_pages; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 165 | |
| 166 | base = memblock_virt_alloc_try_nid_nopanic( |
| 167 | table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS), |
| 168 | BOOTMEM_ALLOC_ACCESSIBLE, nid); |
| 169 | if (!base) |
| 170 | return -ENOMEM; |
| 171 | NODE_DATA(nid)->node_page_ext = base; |
| 172 | total_usage += table_size; |
| 173 | return 0; |
| 174 | } |
| 175 | |
| 176 | void __init page_ext_init_flatmem(void) |
| 177 | { |
| 178 | |
| 179 | int nid, fail; |
| 180 | |
| 181 | if (!invoke_need_callbacks()) |
| 182 | return; |
| 183 | |
| 184 | for_each_online_node(nid) { |
| 185 | fail = alloc_node_page_ext(nid); |
| 186 | if (fail) |
| 187 | goto fail; |
| 188 | } |
| 189 | pr_info("allocated %ld bytes of page_ext\n", total_usage); |
| 190 | invoke_init_callbacks(); |
| 191 | return; |
| 192 | |
| 193 | fail: |
| 194 | pr_crit("allocation of page_ext failed.\n"); |
| 195 | panic("Out of memory"); |
| 196 | } |
| 197 | |
| 198 | #else /* CONFIG_FLAT_NODE_MEM_MAP */ |
| 199 | |
| 200 | struct page_ext *lookup_page_ext(struct page *page) |
| 201 | { |
| 202 | unsigned long pfn = page_to_pfn(page); |
| 203 | struct mem_section *section = __pfn_to_section(pfn); |
Laura Abbott | 1414c7f | 2016-03-15 14:56:30 -0700 | [diff] [blame] | 204 | #if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING) |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 205 | /* |
| 206 | * The sanity checks the page allocator does upon freeing a |
| 207 | * page can reach here before the page_ext arrays are |
| 208 | * allocated when feeding a range of pages to the allocator |
| 209 | * for the first time during bootup or memory hotplug. |
Laura Abbott | 1414c7f | 2016-03-15 14:56:30 -0700 | [diff] [blame] | 210 | * |
| 211 | * This check is also necessary for ensuring page poisoning |
| 212 | * works as expected when enabled |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 213 | */ |
| 214 | if (!section->page_ext) |
| 215 | return NULL; |
| 216 | #endif |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 217 | return get_entry(section->page_ext, pfn); |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 218 | } |
| 219 | |
| 220 | static void *__meminit alloc_page_ext(size_t size, int nid) |
| 221 | { |
| 222 | gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN; |
| 223 | void *addr = NULL; |
| 224 | |
| 225 | addr = alloc_pages_exact_nid(nid, size, flags); |
| 226 | if (addr) { |
| 227 | kmemleak_alloc(addr, size, 1, flags); |
| 228 | return addr; |
| 229 | } |
| 230 | |
| 231 | if (node_state(nid, N_HIGH_MEMORY)) |
| 232 | addr = vzalloc_node(size, nid); |
| 233 | else |
| 234 | addr = vzalloc(size); |
| 235 | |
| 236 | return addr; |
| 237 | } |
| 238 | |
| 239 | static int __meminit init_section_page_ext(unsigned long pfn, int nid) |
| 240 | { |
| 241 | struct mem_section *section; |
| 242 | struct page_ext *base; |
| 243 | unsigned long table_size; |
| 244 | |
| 245 | section = __pfn_to_section(pfn); |
| 246 | |
| 247 | if (section->page_ext) |
| 248 | return 0; |
| 249 | |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 250 | table_size = get_entry_size() * PAGES_PER_SECTION; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 251 | base = alloc_page_ext(table_size, nid); |
| 252 | |
| 253 | /* |
| 254 | * The value stored in section->page_ext is (base - pfn) |
| 255 | * and it does not point to the memory block allocated above, |
| 256 | * causing kmemleak false positives. |
| 257 | */ |
| 258 | kmemleak_not_leak(base); |
| 259 | |
| 260 | if (!base) { |
| 261 | pr_err("page ext allocation failure\n"); |
| 262 | return -ENOMEM; |
| 263 | } |
| 264 | |
| 265 | /* |
| 266 | * The passed "pfn" may not be aligned to SECTION. For the calculation |
| 267 | * we need to apply a mask. |
| 268 | */ |
| 269 | pfn &= PAGE_SECTION_MASK; |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 270 | section->page_ext = (void *)base - get_entry_size() * pfn; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 271 | total_usage += table_size; |
| 272 | return 0; |
| 273 | } |
| 274 | #ifdef CONFIG_MEMORY_HOTPLUG |
| 275 | static void free_page_ext(void *addr) |
| 276 | { |
| 277 | if (is_vmalloc_addr(addr)) { |
| 278 | vfree(addr); |
| 279 | } else { |
| 280 | struct page *page = virt_to_page(addr); |
| 281 | size_t table_size; |
| 282 | |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 283 | table_size = get_entry_size() * PAGES_PER_SECTION; |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 284 | |
| 285 | BUG_ON(PageReserved(page)); |
Qian Cai | 7880fc29 | 2019-03-05 15:49:46 -0800 | [diff] [blame] | 286 | kmemleak_free(addr); |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 287 | free_pages_exact(addr, table_size); |
| 288 | } |
| 289 | } |
| 290 | |
| 291 | static void __free_page_ext(unsigned long pfn) |
| 292 | { |
| 293 | struct mem_section *ms; |
| 294 | struct page_ext *base; |
| 295 | |
| 296 | ms = __pfn_to_section(pfn); |
| 297 | if (!ms || !ms->page_ext) |
| 298 | return; |
Joonsoo Kim | 980ac16 | 2016-10-07 16:58:27 -0700 | [diff] [blame] | 299 | base = get_entry(ms->page_ext, pfn); |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 300 | free_page_ext(base); |
| 301 | ms->page_ext = NULL; |
| 302 | } |
| 303 | |
| 304 | static int __meminit online_page_ext(unsigned long start_pfn, |
| 305 | unsigned long nr_pages, |
| 306 | int nid) |
| 307 | { |
| 308 | unsigned long start, end, pfn; |
| 309 | int fail = 0; |
| 310 | |
| 311 | start = SECTION_ALIGN_DOWN(start_pfn); |
| 312 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); |
| 313 | |
| 314 | if (nid == -1) { |
| 315 | /* |
| 316 | * In this case, "nid" already exists and contains valid memory. |
| 317 | * "start_pfn" passed to us is a pfn which is an arg for |
| 318 | * online__pages(), and start_pfn should exist. |
| 319 | */ |
| 320 | nid = pfn_to_nid(start_pfn); |
| 321 | VM_BUG_ON(!node_state(nid, N_ONLINE)); |
| 322 | } |
| 323 | |
| 324 | for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) { |
| 325 | if (!pfn_present(pfn)) |
| 326 | continue; |
| 327 | fail = init_section_page_ext(pfn, nid); |
| 328 | } |
| 329 | if (!fail) |
| 330 | return 0; |
| 331 | |
| 332 | /* rollback */ |
| 333 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) |
| 334 | __free_page_ext(pfn); |
| 335 | |
| 336 | return -ENOMEM; |
| 337 | } |
| 338 | |
| 339 | static int __meminit offline_page_ext(unsigned long start_pfn, |
| 340 | unsigned long nr_pages, int nid) |
| 341 | { |
| 342 | unsigned long start, end, pfn; |
| 343 | |
| 344 | start = SECTION_ALIGN_DOWN(start_pfn); |
| 345 | end = SECTION_ALIGN_UP(start_pfn + nr_pages); |
| 346 | |
| 347 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) |
| 348 | __free_page_ext(pfn); |
| 349 | return 0; |
| 350 | |
| 351 | } |
| 352 | |
| 353 | static int __meminit page_ext_callback(struct notifier_block *self, |
| 354 | unsigned long action, void *arg) |
| 355 | { |
| 356 | struct memory_notify *mn = arg; |
| 357 | int ret = 0; |
| 358 | |
| 359 | switch (action) { |
| 360 | case MEM_GOING_ONLINE: |
| 361 | ret = online_page_ext(mn->start_pfn, |
| 362 | mn->nr_pages, mn->status_change_nid); |
| 363 | break; |
| 364 | case MEM_OFFLINE: |
| 365 | offline_page_ext(mn->start_pfn, |
| 366 | mn->nr_pages, mn->status_change_nid); |
| 367 | break; |
| 368 | case MEM_CANCEL_ONLINE: |
| 369 | offline_page_ext(mn->start_pfn, |
| 370 | mn->nr_pages, mn->status_change_nid); |
| 371 | break; |
| 372 | case MEM_GOING_OFFLINE: |
| 373 | break; |
| 374 | case MEM_ONLINE: |
| 375 | case MEM_CANCEL_OFFLINE: |
| 376 | break; |
| 377 | } |
| 378 | |
| 379 | return notifier_from_errno(ret); |
| 380 | } |
| 381 | |
| 382 | #endif |
| 383 | |
| 384 | void __init page_ext_init(void) |
| 385 | { |
| 386 | unsigned long pfn; |
| 387 | int nid; |
| 388 | |
| 389 | if (!invoke_need_callbacks()) |
| 390 | return; |
| 391 | |
| 392 | for_each_node_state(nid, N_MEMORY) { |
| 393 | unsigned long start_pfn, end_pfn; |
| 394 | |
| 395 | start_pfn = node_start_pfn(nid); |
| 396 | end_pfn = node_end_pfn(nid); |
| 397 | /* |
| 398 | * start_pfn and end_pfn may not be aligned to SECTION and the |
| 399 | * page->flags of out of node pages are not initialized. So we |
| 400 | * scan [start_pfn, the biggest section's pfn < end_pfn) here. |
| 401 | */ |
| 402 | for (pfn = start_pfn; pfn < end_pfn; |
| 403 | pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) { |
| 404 | |
| 405 | if (!pfn_valid(pfn)) |
| 406 | continue; |
| 407 | /* |
| 408 | * Nodes's pfns can be overlapping. |
| 409 | * We know some arch can have a nodes layout such as |
| 410 | * -------------pfn--------------> |
| 411 | * N0 | N1 | N2 | N0 | N1 | N2|.... |
Yang Shi | fe53ca5 | 2016-05-27 14:27:30 -0700 | [diff] [blame] | 412 | * |
| 413 | * Take into account DEFERRED_STRUCT_PAGE_INIT. |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 414 | */ |
Yang Shi | fe53ca5 | 2016-05-27 14:27:30 -0700 | [diff] [blame] | 415 | if (early_pfn_to_nid(pfn) != nid) |
Joonsoo Kim | eefa864 | 2014-12-12 16:55:46 -0800 | [diff] [blame] | 416 | continue; |
| 417 | if (init_section_page_ext(pfn, nid)) |
| 418 | goto oom; |
| 419 | } |
| 420 | } |
| 421 | hotplug_memory_notifier(page_ext_callback, 0); |
| 422 | pr_info("allocated %ld bytes of page_ext\n", total_usage); |
| 423 | invoke_init_callbacks(); |
| 424 | return; |
| 425 | |
| 426 | oom: |
| 427 | panic("Out of memory"); |
| 428 | } |
| 429 | |
| 430 | void __meminit pgdat_page_ext_init(struct pglist_data *pgdat) |
| 431 | { |
| 432 | } |
| 433 | |
| 434 | #endif |