1 // SPDX-License-Identifier: GPL-2.0
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>
13 * struct page extension
15 * This is the feature to manage memory for extended data per page.
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.
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.
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.
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.
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.
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.
61 static struct page_ext_operations *page_ext_ops[] = {
63 #ifdef CONFIG_PAGE_OWNER
66 #if defined(CONFIG_IDLE_PAGE_TRACKING) && !defined(CONFIG_64BIT)
71 static unsigned long total_usage;
72 static unsigned long extra_mem;
74 static bool __init invoke_need_callbacks(void)
77 int entries = ARRAY_SIZE(page_ext_ops);
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) +
84 extra_mem += page_ext_ops[i]->size;
92 static void __init invoke_init_callbacks(void)
95 int entries = ARRAY_SIZE(page_ext_ops);
97 for (i = 0; i < entries; i++) {
98 if (page_ext_ops[i]->init)
99 page_ext_ops[i]->init();
103 static unsigned long get_entry_size(void)
105 return sizeof(struct page_ext) + extra_mem;
108 static inline struct page_ext *get_entry(void *base, unsigned long index)
110 return base + get_entry_size() * index;
113 #if !defined(CONFIG_SPARSEMEM)
116 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
118 pgdat->node_page_ext = NULL;
121 struct page_ext *lookup_page_ext(struct page *page)
123 unsigned long pfn = page_to_pfn(page);
125 struct page_ext *base;
127 base = NODE_DATA(page_to_nid(page))->node_page_ext;
128 #if defined(CONFIG_DEBUG_VM)
130 * The sanity checks the page allocator does upon freeing a
131 * page can reach here before the page_ext arrays are
132 * allocated when feeding a range of pages to the allocator
133 * for the first time during bootup or memory hotplug.
138 index = pfn - round_down(node_start_pfn(page_to_nid(page)),
140 return get_entry(base, index);
143 static int __init alloc_node_page_ext(int nid)
145 struct page_ext *base;
146 unsigned long table_size;
147 unsigned long nr_pages;
149 nr_pages = NODE_DATA(nid)->node_spanned_pages;
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.
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;
162 table_size = get_entry_size() * nr_pages;
164 base = memblock_virt_alloc_try_nid_nopanic(
165 table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
166 BOOTMEM_ALLOC_ACCESSIBLE, nid);
169 NODE_DATA(nid)->node_page_ext = base;
170 total_usage += table_size;
174 void __init page_ext_init_flatmem(void)
179 if (!invoke_need_callbacks())
182 for_each_online_node(nid) {
183 fail = alloc_node_page_ext(nid);
187 pr_info("allocated %ld bytes of page_ext\n", total_usage);
188 invoke_init_callbacks();
192 pr_crit("allocation of page_ext failed.\n");
193 panic("Out of memory");
196 #else /* CONFIG_FLAT_NODE_MEM_MAP */
198 struct page_ext *lookup_page_ext(struct page *page)
200 unsigned long pfn = page_to_pfn(page);
201 struct mem_section *section = __pfn_to_section(pfn);
202 #if defined(CONFIG_DEBUG_VM)
204 * The sanity checks the page allocator does upon freeing a
205 * page can reach here before the page_ext arrays are
206 * allocated when feeding a range of pages to the allocator
207 * for the first time during bootup or memory hotplug.
209 if (!section->page_ext)
212 return get_entry(section->page_ext, pfn);
215 static void *__meminit alloc_page_ext(size_t size, int nid)
217 gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
220 addr = alloc_pages_exact_nid(nid, size, flags);
222 kmemleak_alloc(addr, size, 1, flags);
226 addr = vzalloc_node(size, nid);
231 static int __meminit init_section_page_ext(unsigned long pfn, int nid)
233 struct mem_section *section;
234 struct page_ext *base;
235 unsigned long table_size;
237 section = __pfn_to_section(pfn);
239 if (section->page_ext)
242 table_size = get_entry_size() * PAGES_PER_SECTION;
243 base = alloc_page_ext(table_size, nid);
246 * The value stored in section->page_ext is (base - pfn)
247 * and it does not point to the memory block allocated above,
248 * causing kmemleak false positives.
250 kmemleak_not_leak(base);
253 pr_err("page ext allocation failure\n");
258 * The passed "pfn" may not be aligned to SECTION. For the calculation
259 * we need to apply a mask.
261 pfn &= PAGE_SECTION_MASK;
262 section->page_ext = (void *)base - get_entry_size() * pfn;
263 total_usage += table_size;
266 #ifdef CONFIG_MEMORY_HOTPLUG
267 static void free_page_ext(void *addr)
269 if (is_vmalloc_addr(addr)) {
272 struct page *page = virt_to_page(addr);
275 table_size = get_entry_size() * PAGES_PER_SECTION;
277 BUG_ON(PageReserved(page));
278 free_pages_exact(addr, table_size);
282 static void __free_page_ext(unsigned long pfn)
284 struct mem_section *ms;
285 struct page_ext *base;
287 ms = __pfn_to_section(pfn);
288 if (!ms || !ms->page_ext)
290 base = get_entry(ms->page_ext, pfn);
295 static int __meminit online_page_ext(unsigned long start_pfn,
296 unsigned long nr_pages,
299 unsigned long start, end, pfn;
302 start = SECTION_ALIGN_DOWN(start_pfn);
303 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
307 * In this case, "nid" already exists and contains valid memory.
308 * "start_pfn" passed to us is a pfn which is an arg for
309 * online__pages(), and start_pfn should exist.
311 nid = pfn_to_nid(start_pfn);
312 VM_BUG_ON(!node_state(nid, N_ONLINE));
315 for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
316 if (!pfn_present(pfn))
318 fail = init_section_page_ext(pfn, nid);
324 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
325 __free_page_ext(pfn);
330 static int __meminit offline_page_ext(unsigned long start_pfn,
331 unsigned long nr_pages, int nid)
333 unsigned long start, end, pfn;
335 start = SECTION_ALIGN_DOWN(start_pfn);
336 end = SECTION_ALIGN_UP(start_pfn + nr_pages);
338 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
339 __free_page_ext(pfn);
344 static int __meminit page_ext_callback(struct notifier_block *self,
345 unsigned long action, void *arg)
347 struct memory_notify *mn = arg;
351 case MEM_GOING_ONLINE:
352 ret = online_page_ext(mn->start_pfn,
353 mn->nr_pages, mn->status_change_nid);
356 offline_page_ext(mn->start_pfn,
357 mn->nr_pages, mn->status_change_nid);
359 case MEM_CANCEL_ONLINE:
360 offline_page_ext(mn->start_pfn,
361 mn->nr_pages, mn->status_change_nid);
363 case MEM_GOING_OFFLINE:
366 case MEM_CANCEL_OFFLINE:
370 return notifier_from_errno(ret);
375 void __init page_ext_init(void)
380 if (!invoke_need_callbacks())
383 for_each_node_state(nid, N_MEMORY) {
384 unsigned long start_pfn, end_pfn;
386 start_pfn = node_start_pfn(nid);
387 end_pfn = node_end_pfn(nid);
389 * start_pfn and end_pfn may not be aligned to SECTION and the
390 * page->flags of out of node pages are not initialized. So we
391 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
393 for (pfn = start_pfn; pfn < end_pfn;
394 pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
399 * Nodes's pfns can be overlapping.
400 * We know some arch can have a nodes layout such as
401 * -------------pfn-------------->
402 * N0 | N1 | N2 | N0 | N1 | N2|....
404 * Take into account DEFERRED_STRUCT_PAGE_INIT.
406 if (early_pfn_to_nid(pfn) != nid)
408 if (init_section_page_ext(pfn, nid))
413 hotplug_memory_notifier(page_ext_callback, 0);
414 pr_info("allocated %ld bytes of page_ext\n", total_usage);
415 invoke_init_callbacks();
419 panic("Out of memory");
422 void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)