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>
8#include <linux/page_owner.h>
9
10/*
11 * struct page extension
12 *
13 * This is the feature to manage memory for extended data per page.
14 *
15 * Until now, we must modify struct page itself to store extra data per page.
16 * This requires rebuilding the kernel and it is really time consuming process.
17 * And, sometimes, rebuild is impossible due to third party module dependency.
18 * At last, enlarging struct page could cause un-wanted system behaviour change.
19 *
20 * This feature is intended to overcome above mentioned problems. This feature
21 * allocates memory for extended data per page in certain place rather than
22 * the struct page itself. This memory can be accessed by the accessor
23 * functions provided by this code. During the boot process, it checks whether
24 * allocation of huge chunk of memory is needed or not. If not, it avoids
25 * allocating memory at all. With this advantage, we can include this feature
26 * into the kernel in default and can avoid rebuild and solve related problems.
27 *
28 * To help these things to work well, there are two callbacks for clients. One
29 * is the need callback which is mandatory if user wants to avoid useless
30 * memory allocation at boot-time. The other is optional, init callback, which
31 * is used to do proper initialization after memory is allocated.
32 *
33 * The need callback is used to decide whether extended memory allocation is
34 * needed or not. Sometimes users want to deactivate some features in this
35 * boot and extra memory would be unneccessary. In this case, to avoid
36 * allocating huge chunk of memory, each clients represent their need of
37 * extra memory through the need callback. If one of the need callbacks
38 * returns true, it means that someone needs extra memory so that
39 * page extension core should allocates memory for page extension. If
40 * none of need callbacks return true, memory isn't needed at all in this boot
41 * and page extension core can skip to allocate memory. As result,
42 * none of memory is wasted.
43 *
44 * The init callback is used to do proper initialization after page extension
45 * is completely initialized. In sparse memory system, extra memory is
46 * allocated some time later than memmap is allocated. In other words, lifetime
47 * of memory for page extension isn't same with memmap for struct page.
48 * Therefore, clients can't store extra data until page extension is
49 * initialized, even if pages are allocated and used freely. This could
50 * cause inadequate state of extra data per page, so, to prevent it, client
51 * can utilize this callback to initialize the state of it correctly.
52 */
53
54static struct page_ext_operations *page_ext_ops[] = {
55	&debug_guardpage_ops,
56#ifdef CONFIG_PAGE_POISONING
57	&page_poisoning_ops,
58#endif
59#ifdef CONFIG_PAGE_OWNER
60	&page_owner_ops,
61#endif
62};
63
64static unsigned long total_usage;
65
66static bool __init invoke_need_callbacks(void)
67{
68	int i;
69	int entries = ARRAY_SIZE(page_ext_ops);
70
71	for (i = 0; i < entries; i++) {
72		if (page_ext_ops[i]->need && page_ext_ops[i]->need())
73			return true;
74	}
75
76	return false;
77}
78
79static void __init invoke_init_callbacks(void)
80{
81	int i;
82	int entries = ARRAY_SIZE(page_ext_ops);
83
84	for (i = 0; i < entries; i++) {
85		if (page_ext_ops[i]->init)
86			page_ext_ops[i]->init();
87	}
88}
89
90#if !defined(CONFIG_SPARSEMEM)
91
92
93void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
94{
95	pgdat->node_page_ext = NULL;
96}
97
98struct page_ext *lookup_page_ext(struct page *page)
99{
100	unsigned long pfn = page_to_pfn(page);
101	unsigned long offset;
102	struct page_ext *base;
103
104	base = NODE_DATA(page_to_nid(page))->node_page_ext;
105#ifdef CONFIG_DEBUG_VM
106	/*
107	 * The sanity checks the page allocator does upon freeing a
108	 * page can reach here before the page_ext arrays are
109	 * allocated when feeding a range of pages to the allocator
110	 * for the first time during bootup or memory hotplug.
111	 */
112	if (unlikely(!base))
113		return NULL;
114#endif
115	offset = pfn - round_down(node_start_pfn(page_to_nid(page)),
116					MAX_ORDER_NR_PAGES);
117	return base + offset;
118}
119
120static int __init alloc_node_page_ext(int nid)
121{
122	struct page_ext *base;
123	unsigned long table_size;
124	unsigned long nr_pages;
125
126	nr_pages = NODE_DATA(nid)->node_spanned_pages;
127	if (!nr_pages)
128		return 0;
129
130	/*
131	 * Need extra space if node range is not aligned with
132	 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
133	 * checks buddy's status, range could be out of exact node range.
134	 */
135	if (!IS_ALIGNED(node_start_pfn(nid), MAX_ORDER_NR_PAGES) ||
136		!IS_ALIGNED(node_end_pfn(nid), MAX_ORDER_NR_PAGES))
137		nr_pages += MAX_ORDER_NR_PAGES;
138
139	table_size = sizeof(struct page_ext) * nr_pages;
140
141	base = memblock_virt_alloc_try_nid_nopanic(
142			table_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
143			BOOTMEM_ALLOC_ACCESSIBLE, nid);
144	if (!base)
145		return -ENOMEM;
146	NODE_DATA(nid)->node_page_ext = base;
147	total_usage += table_size;
148	return 0;
149}
150
151void __init page_ext_init_flatmem(void)
152{
153
154	int nid, fail;
155
156	if (!invoke_need_callbacks())
157		return;
158
159	for_each_online_node(nid)  {
160		fail = alloc_node_page_ext(nid);
161		if (fail)
162			goto fail;
163	}
164	pr_info("allocated %ld bytes of page_ext\n", total_usage);
165	invoke_init_callbacks();
166	return;
167
168fail:
169	pr_crit("allocation of page_ext failed.\n");
170	panic("Out of memory");
171}
172
173#else /* CONFIG_FLAT_NODE_MEM_MAP */
174
175struct page_ext *lookup_page_ext(struct page *page)
176{
177	unsigned long pfn = page_to_pfn(page);
178	struct mem_section *section = __pfn_to_section(pfn);
179#ifdef CONFIG_DEBUG_VM
180	/*
181	 * The sanity checks the page allocator does upon freeing a
182	 * page can reach here before the page_ext arrays are
183	 * allocated when feeding a range of pages to the allocator
184	 * for the first time during bootup or memory hotplug.
185	 */
186	if (!section->page_ext)
187		return NULL;
188#endif
189	return section->page_ext + pfn;
190}
191
192static void *__meminit alloc_page_ext(size_t size, int nid)
193{
194	gfp_t flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN;
195	void *addr = NULL;
196
197	addr = alloc_pages_exact_nid(nid, size, flags);
198	if (addr) {
199		kmemleak_alloc(addr, size, 1, flags);
200		return addr;
201	}
202
203	if (node_state(nid, N_HIGH_MEMORY))
204		addr = vzalloc_node(size, nid);
205	else
206		addr = vzalloc(size);
207
208	return addr;
209}
210
211static int __meminit init_section_page_ext(unsigned long pfn, int nid)
212{
213	struct mem_section *section;
214	struct page_ext *base;
215	unsigned long table_size;
216
217	section = __pfn_to_section(pfn);
218
219	if (section->page_ext)
220		return 0;
221
222	table_size = sizeof(struct page_ext) * PAGES_PER_SECTION;
223	base = alloc_page_ext(table_size, nid);
224
225	/*
226	 * The value stored in section->page_ext is (base - pfn)
227	 * and it does not point to the memory block allocated above,
228	 * causing kmemleak false positives.
229	 */
230	kmemleak_not_leak(base);
231
232	if (!base) {
233		pr_err("page ext allocation failure\n");
234		return -ENOMEM;
235	}
236
237	/*
238	 * The passed "pfn" may not be aligned to SECTION.  For the calculation
239	 * we need to apply a mask.
240	 */
241	pfn &= PAGE_SECTION_MASK;
242	section->page_ext = base - pfn;
243	total_usage += table_size;
244	return 0;
245}
246#ifdef CONFIG_MEMORY_HOTPLUG
247static void free_page_ext(void *addr)
248{
249	if (is_vmalloc_addr(addr)) {
250		vfree(addr);
251	} else {
252		struct page *page = virt_to_page(addr);
253		size_t table_size;
254
255		table_size = sizeof(struct page_ext) * PAGES_PER_SECTION;
256
257		BUG_ON(PageReserved(page));
258		free_pages_exact(addr, table_size);
259	}
260}
261
262static void __free_page_ext(unsigned long pfn)
263{
264	struct mem_section *ms;
265	struct page_ext *base;
266
267	ms = __pfn_to_section(pfn);
268	if (!ms || !ms->page_ext)
269		return;
270	base = ms->page_ext + pfn;
271	free_page_ext(base);
272	ms->page_ext = NULL;
273}
274
275static int __meminit online_page_ext(unsigned long start_pfn,
276				unsigned long nr_pages,
277				int nid)
278{
279	unsigned long start, end, pfn;
280	int fail = 0;
281
282	start = SECTION_ALIGN_DOWN(start_pfn);
283	end = SECTION_ALIGN_UP(start_pfn + nr_pages);
284
285	if (nid == -1) {
286		/*
287		 * In this case, "nid" already exists and contains valid memory.
288		 * "start_pfn" passed to us is a pfn which is an arg for
289		 * online__pages(), and start_pfn should exist.
290		 */
291		nid = pfn_to_nid(start_pfn);
292		VM_BUG_ON(!node_state(nid, N_ONLINE));
293	}
294
295	for (pfn = start; !fail && pfn < end; pfn += PAGES_PER_SECTION) {
296		if (!pfn_present(pfn))
297			continue;
298		fail = init_section_page_ext(pfn, nid);
299	}
300	if (!fail)
301		return 0;
302
303	/* rollback */
304	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
305		__free_page_ext(pfn);
306
307	return -ENOMEM;
308}
309
310static int __meminit offline_page_ext(unsigned long start_pfn,
311				unsigned long nr_pages, int nid)
312{
313	unsigned long start, end, pfn;
314
315	start = SECTION_ALIGN_DOWN(start_pfn);
316	end = SECTION_ALIGN_UP(start_pfn + nr_pages);
317
318	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION)
319		__free_page_ext(pfn);
320	return 0;
321
322}
323
324static int __meminit page_ext_callback(struct notifier_block *self,
325			       unsigned long action, void *arg)
326{
327	struct memory_notify *mn = arg;
328	int ret = 0;
329
330	switch (action) {
331	case MEM_GOING_ONLINE:
332		ret = online_page_ext(mn->start_pfn,
333				   mn->nr_pages, mn->status_change_nid);
334		break;
335	case MEM_OFFLINE:
336		offline_page_ext(mn->start_pfn,
337				mn->nr_pages, mn->status_change_nid);
338		break;
339	case MEM_CANCEL_ONLINE:
340		offline_page_ext(mn->start_pfn,
341				mn->nr_pages, mn->status_change_nid);
342		break;
343	case MEM_GOING_OFFLINE:
344		break;
345	case MEM_ONLINE:
346	case MEM_CANCEL_OFFLINE:
347		break;
348	}
349
350	return notifier_from_errno(ret);
351}
352
353#endif
354
355void __init page_ext_init(void)
356{
357	unsigned long pfn;
358	int nid;
359
360	if (!invoke_need_callbacks())
361		return;
362
363	for_each_node_state(nid, N_MEMORY) {
364		unsigned long start_pfn, end_pfn;
365
366		start_pfn = node_start_pfn(nid);
367		end_pfn = node_end_pfn(nid);
368		/*
369		 * start_pfn and end_pfn may not be aligned to SECTION and the
370		 * page->flags of out of node pages are not initialized.  So we
371		 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
372		 */
373		for (pfn = start_pfn; pfn < end_pfn;
374			pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
375
376			if (!pfn_valid(pfn))
377				continue;
378			/*
379			 * Nodes's pfns can be overlapping.
380			 * We know some arch can have a nodes layout such as
381			 * -------------pfn-------------->
382			 * N0 | N1 | N2 | N0 | N1 | N2|....
383			 */
384			if (pfn_to_nid(pfn) != nid)
385				continue;
386			if (init_section_page_ext(pfn, nid))
387				goto oom;
388		}
389	}
390	hotplug_memory_notifier(page_ext_callback, 0);
391	pr_info("allocated %ld bytes of page_ext\n", total_usage);
392	invoke_init_callbacks();
393	return;
394
395oom:
396	panic("Out of memory");
397}
398
399void __meminit pgdat_page_ext_init(struct pglist_data *pgdat)
400{
401}
402
403#endif
404