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>
12 * struct page extension
14 * This is the feature to manage memory for extended data per page.
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.
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.
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.
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.
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.
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.
60 static struct page_ext_operations *page_ext_ops[] = {
62 #ifdef CONFIG_PAGE_POISONING
65 #ifdef CONFIG_PAGE_OWNER
68 #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
73 static unsigned long total_usage;
74 static unsigned long extra_mem;
76 static bool __init invoke_need_callbacks(void)
79 int entries = ARRAY_SIZE(page_ext_ops);
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) +
86 extra_mem += page_ext_ops[i]->size;
94 static void __init invoke_init_callbacks(void)
97 int entries = ARRAY_SIZE(page_ext_ops);
99 for (i = 0; i < entries; i++) {
100 if (page_ext_ops[i]->init)
101 page_ext_ops[i]->init();
105 static unsigned long get_entry_size(void)
107 return sizeof(struct page_ext) + extra_mem;
110 static inline struct page_ext *get_entry(void *base, unsigned long index)
112 return base + get_entry_size() * index;
115 #if !defined(CONFIG_SPARSEMEM)
118 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
120 pgdat->node_page_ext = NULL;
123 struct page_ext *lookup_page_ext(struct page *page)
125 unsigned long pfn = page_to_pfn(page);
127 struct page_ext *base;
129 base = NODE_DATA(page_to_nid(page))->node_page_ext;
130 #if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
132 * The sanity checks the page allocator does upon freeing a
133 * page can reach here before the page_ext arrays are
134 * allocated when feeding a range of pages to the allocator
135 * for the first time during bootup or memory hotplug.
137 * This check is also necessary for ensuring page poisoning
138 * works as expected when enabled
143 index = pfn - round_down(node_start_pfn(page_to_nid(page)),
145 return get_entry(base, index);
148 static int __init alloc_node_page_ext(int nid)
150 struct page_ext *base;
151 unsigned long table_size;
152 unsigned long nr_pages;
154 nr_pages = NODE_DATA(nid)->node_spanned_pages;
159 * Need extra space if node range is not aligned with
160 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
161 * checks buddy's status, range could be out of exact node range.
163 if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
164 !IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
165 nr_pages += MAX_ORDER_NR_PAGES;
167 table_size = get_entry_size() * nr_pages;
169 base = memblock_virt_alloc_try_nid_nopanic(
170 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
171 BOOTMEM_ALLOC_ACCESSIBLE, nid);
174 NODE_DATA(nid)->node_page_ext = base;
175 total_usage += table_size;
179 void __init page_ext_init_flatmem(void)
184 if (!invoke_need_callbacks())
187 for_each_online_node(nid) {
188 fail = alloc_node_page_ext(nid);
192 pr_info("allocated %ld bytes of page_ext\n", total_usage);
193 invoke_init_callbacks();
197 pr_crit("allocation of page_ext failed.\n");
198 panic("Out of memory");
201 #else /* CONFIG_FLAT_NODE_MEM_MAP */
203 struct page_ext *lookup_page_ext(struct page *page)
205 unsigned long pfn = page_to_pfn(page);
206 struct mem_section *section = __pfn_to_section(pfn);
207 #if defined(CONFIG_DEBUG_VM) || defined(CONFIG_PAGE_POISONING)
209 * The sanity checks the page allocator does upon freeing a
210 * page can reach here before the page_ext arrays are
211 * allocated when feeding a range of pages to the allocator
212 * for the first time during bootup or memory hotplug.
214 * This check is also necessary for ensuring page poisoning
215 * works as expected when enabled
217 if (!section->page_ext)
220 return get_entry(section->page_ext, pfn);
223 static void *__meminit alloc_page_ext(size_t size, int nid)
225 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
228 addr = alloc_pages_exact_nid(nid, size, flags);
230 kmemleak_alloc(addr, size, 1, flags);
234 if (node_state(nid, N_HIGH_MEMORY))
235 addr = vzalloc_node(size, nid);
237 addr = vzalloc(size);
242 static int __meminit init_section_page_ext(unsigned long pfn, int nid)
244 struct mem_section *section;
245 struct page_ext *base;
246 unsigned long table_size;
248 section = __pfn_to_section(pfn);
250 if (section->page_ext)
253 table_size = get_entry_size() * PAGES_PER_SECTION;
254 base = alloc_page_ext(table_size, nid);
257 * The value stored in section->page_ext is (base - pfn)
258 * and it does not point to the memory block allocated above,
259 * causing kmemleak false positives.
261 kmemleak_not_leak(base);
264 pr_err("page ext allocation failure\n");
269 * The passed "pfn" may not be aligned to SECTION. For the calculation
270 * we need to apply a mask.
272 pfn &= PAGE_SECTION_MASK;
273 section->page_ext = (void *)base - get_entry_size() * pfn;
274 total_usage += table_size;
277 #ifdef CONFIG_MEMORY_HOTPLUG
278 static void free_page_ext(void *addr)
280 if (is_vmalloc_addr(addr)) {
283 struct page *page = virt_to_page(addr);
286 table_size = get_entry_size() * PAGES_PER_SECTION;
288 BUG_ON(PageReserved(page));
289 free_pages_exact(addr, table_size);
293 static void __free_page_ext(unsigned long pfn)
295 struct mem_section *ms;
296 struct page_ext *base;
298 ms = __pfn_to_section(pfn);
299 if (!ms || !ms->page_ext)
301 base = get_entry(ms->page_ext, pfn);
306 static int __meminit online_page_ext(unsigned long start_pfn,
307 unsigned long nr_pages,
310 unsigned long start, end, pfn;
313 start = SECTION_ALIGN_DOWN(start_pfn);
314 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
318 * In this case, "nid" already exists and contains valid memory.
319 * "start_pfn" passed to us is a pfn which is an arg for
320 * online__pages(), and start_pfn should exist.
322 nid = pfn_to_nid(start_pfn);
323 VM_BUG_ON(!node_state(nid, N_ONLINE));
326 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
327 if (!pfn_present(pfn))
329 fail = init_section_page_ext(pfn, nid);
335 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
336 __free_page_ext(pfn);
341 static int __meminit offline_page_ext(unsigned long start_pfn,
342 unsigned long nr_pages, int nid)
344 unsigned long start, end, pfn;
346 start = SECTION_ALIGN_DOWN(start_pfn);
347 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
349 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
350 __free_page_ext(pfn);
355 static int __meminit page_ext_callback(struct notifier_block *self,
356 unsigned long action, void *arg)
358 struct memory_notify *mn = arg;
362 case MEM_GOING_ONLINE:
363 ret = online_page_ext(mn->start_pfn,
364 mn->nr_pages, mn->status_change_nid);
367 offline_page_ext(mn->start_pfn,
368 mn->nr_pages, mn->status_change_nid);
370 case MEM_CANCEL_ONLINE:
371 offline_page_ext(mn->start_pfn,
372 mn->nr_pages, mn->status_change_nid);
374 case MEM_GOING_OFFLINE:
377 case MEM_CANCEL_OFFLINE:
381 return notifier_from_errno(ret);
386 void __init page_ext_init(void)
391 if (!invoke_need_callbacks())
394 for_each_node_state(nid, N_MEMORY) {
395 unsigned long start_pfn, end_pfn;
397 start_pfn = node_start_pfn(nid);
398 end_pfn = node_end_pfn(nid);
400 * start_pfn and end_pfn may not be aligned to SECTION and the
401 * page->flags of out of node pages are not initialized. So we
402 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
404 for (pfn = start_pfn; pfn < end_pfn;
405 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
410 * Nodes's pfns can be overlapping.
411 * We know some arch can have a nodes layout such as
412 * -------------pfn-------------->
413 * N0 | N1 | N2 | N0 | N1 | N2|....
415 * Take into account DEFERRED_STRUCT_PAGE_INIT.
417 if (early_pfn_to_nid(pfn) != nid)
419 if (init_section_page_ext(pfn, nid))
423 hotplug_memory_notifier(page_ext_callback, 0);
424 pr_info("allocated %ld bytes of page_ext\n", total_usage);
425 invoke_init_callbacks();
429 panic("Out of memory");
432 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)