1 // SPDX-License-Identifier: GPL-2.0
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>
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 unnecessary. 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 #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
62 static bool need_page_idle(void)
66 struct page_ext_operations page_idle_ops
= {
67 .need
= need_page_idle
,
71 static struct page_ext_operations
*page_ext_ops
[] = {
72 #ifdef CONFIG_PAGE_OWNER
75 #if defined(CONFIG_PAGE_IDLE_FLAG) && !defined(CONFIG_64BIT)
80 unsigned long page_ext_size
= sizeof(struct page_ext
);
82 static unsigned long total_usage
;
84 static bool __init
invoke_need_callbacks(void)
87 int entries
= ARRAY_SIZE(page_ext_ops
);
90 for (i
= 0; i
< entries
; i
++) {
91 if (page_ext_ops
[i
]->need
&& page_ext_ops
[i
]->need()) {
92 page_ext_ops
[i
]->offset
= page_ext_size
;
93 page_ext_size
+= page_ext_ops
[i
]->size
;
101 static void __init
invoke_init_callbacks(void)
104 int entries
= ARRAY_SIZE(page_ext_ops
);
106 for (i
= 0; i
< entries
; i
++) {
107 if (page_ext_ops
[i
]->init
)
108 page_ext_ops
[i
]->init();
112 #ifndef CONFIG_SPARSEMEM
113 void __init
page_ext_init_flatmem_late(void)
115 invoke_init_callbacks();
119 static inline struct page_ext
*get_entry(void *base
, unsigned long index
)
121 return base
+ page_ext_size
* index
;
124 #ifndef CONFIG_SPARSEMEM
127 void __meminit
pgdat_page_ext_init(struct pglist_data
*pgdat
)
129 pgdat
->node_page_ext
= NULL
;
132 struct page_ext
*lookup_page_ext(const struct page
*page
)
134 unsigned long pfn
= page_to_pfn(page
);
136 struct page_ext
*base
;
138 base
= NODE_DATA(page_to_nid(page
))->node_page_ext
;
140 * The sanity checks the page allocator does upon freeing a
141 * page can reach here before the page_ext arrays are
142 * allocated when feeding a range of pages to the allocator
143 * for the first time during bootup or memory hotplug.
147 index
= pfn
- round_down(node_start_pfn(page_to_nid(page
)),
149 return get_entry(base
, index
);
152 static int __init
alloc_node_page_ext(int nid
)
154 struct page_ext
*base
;
155 unsigned long table_size
;
156 unsigned long nr_pages
;
158 nr_pages
= NODE_DATA(nid
)->node_spanned_pages
;
163 * Need extra space if node range is not aligned with
164 * MAX_ORDER_NR_PAGES. When page allocator's buddy algorithm
165 * checks buddy's status, range could be out of exact node range.
167 if (!IS_ALIGNED(node_start_pfn(nid
), MAX_ORDER_NR_PAGES
) ||
168 !IS_ALIGNED(node_end_pfn(nid
), MAX_ORDER_NR_PAGES
))
169 nr_pages
+= MAX_ORDER_NR_PAGES
;
171 table_size
= page_ext_size
* nr_pages
;
173 base
= memblock_alloc_try_nid(
174 table_size
, PAGE_SIZE
, __pa(MAX_DMA_ADDRESS
),
175 MEMBLOCK_ALLOC_ACCESSIBLE
, nid
);
178 NODE_DATA(nid
)->node_page_ext
= base
;
179 total_usage
+= table_size
;
183 void __init
page_ext_init_flatmem(void)
188 if (!invoke_need_callbacks())
191 for_each_online_node(nid
) {
192 fail
= alloc_node_page_ext(nid
);
196 pr_info("allocated %ld bytes of page_ext\n", total_usage
);
200 pr_crit("allocation of page_ext failed.\n");
201 panic("Out of memory");
204 #else /* CONFIG_FLATMEM */
206 struct page_ext
*lookup_page_ext(const struct page
*page
)
208 unsigned long pfn
= page_to_pfn(page
);
209 struct mem_section
*section
= __pfn_to_section(pfn
);
211 * The sanity checks the page allocator does upon freeing a
212 * page can reach here before the page_ext arrays are
213 * allocated when feeding a range of pages to the allocator
214 * for the first time during bootup or memory hotplug.
216 if (!section
->page_ext
)
218 return get_entry(section
->page_ext
, pfn
);
221 static void *__meminit
alloc_page_ext(size_t size
, int nid
)
223 gfp_t flags
= GFP_KERNEL
| __GFP_ZERO
| __GFP_NOWARN
;
226 addr
= alloc_pages_exact_nid(nid
, size
, flags
);
228 kmemleak_alloc(addr
, size
, 1, flags
);
232 addr
= vzalloc_node(size
, nid
);
237 static int __meminit
init_section_page_ext(unsigned long pfn
, int nid
)
239 struct mem_section
*section
;
240 struct page_ext
*base
;
241 unsigned long table_size
;
243 section
= __pfn_to_section(pfn
);
245 if (section
->page_ext
)
248 table_size
= page_ext_size
* PAGES_PER_SECTION
;
249 base
= alloc_page_ext(table_size
, nid
);
252 * The value stored in section->page_ext is (base - pfn)
253 * and it does not point to the memory block allocated above,
254 * causing kmemleak false positives.
256 kmemleak_not_leak(base
);
259 pr_err("page ext allocation failure\n");
264 * The passed "pfn" may not be aligned to SECTION. For the calculation
265 * we need to apply a mask.
267 pfn
&= PAGE_SECTION_MASK
;
268 section
->page_ext
= (void *)base
- page_ext_size
* pfn
;
269 total_usage
+= table_size
;
273 static void free_page_ext(void *addr
)
275 if (is_vmalloc_addr(addr
)) {
278 struct page
*page
= virt_to_page(addr
);
281 table_size
= page_ext_size
* PAGES_PER_SECTION
;
283 BUG_ON(PageReserved(page
));
285 free_pages_exact(addr
, table_size
);
289 static void __free_page_ext(unsigned long pfn
)
291 struct mem_section
*ms
;
292 struct page_ext
*base
;
294 ms
= __pfn_to_section(pfn
);
295 if (!ms
|| !ms
->page_ext
)
297 base
= get_entry(ms
->page_ext
, pfn
);
302 static int __meminit
online_page_ext(unsigned long start_pfn
,
303 unsigned long nr_pages
,
306 unsigned long start
, end
, pfn
;
309 start
= SECTION_ALIGN_DOWN(start_pfn
);
310 end
= SECTION_ALIGN_UP(start_pfn
+ nr_pages
);
312 if (nid
== NUMA_NO_NODE
) {
314 * In this case, "nid" already exists and contains valid memory.
315 * "start_pfn" passed to us is a pfn which is an arg for
316 * online__pages(), and start_pfn should exist.
318 nid
= pfn_to_nid(start_pfn
);
319 VM_BUG_ON(!node_state(nid
, N_ONLINE
));
322 for (pfn
= start
; !fail
&& pfn
< end
; pfn
+= PAGES_PER_SECTION
)
323 fail
= init_section_page_ext(pfn
, nid
);
328 for (pfn
= start
; pfn
< end
; pfn
+= PAGES_PER_SECTION
)
329 __free_page_ext(pfn
);
334 static int __meminit
offline_page_ext(unsigned long start_pfn
,
335 unsigned long nr_pages
, int nid
)
337 unsigned long start
, end
, pfn
;
339 start
= SECTION_ALIGN_DOWN(start_pfn
);
340 end
= SECTION_ALIGN_UP(start_pfn
+ nr_pages
);
342 for (pfn
= start
; pfn
< end
; pfn
+= PAGES_PER_SECTION
)
343 __free_page_ext(pfn
);
348 static int __meminit
page_ext_callback(struct notifier_block
*self
,
349 unsigned long action
, void *arg
)
351 struct memory_notify
*mn
= arg
;
355 case MEM_GOING_ONLINE
:
356 ret
= online_page_ext(mn
->start_pfn
,
357 mn
->nr_pages
, mn
->status_change_nid
);
360 offline_page_ext(mn
->start_pfn
,
361 mn
->nr_pages
, mn
->status_change_nid
);
363 case MEM_CANCEL_ONLINE
:
364 offline_page_ext(mn
->start_pfn
,
365 mn
->nr_pages
, mn
->status_change_nid
);
367 case MEM_GOING_OFFLINE
:
370 case MEM_CANCEL_OFFLINE
:
374 return notifier_from_errno(ret
);
377 void __init
page_ext_init(void)
382 if (!invoke_need_callbacks())
385 for_each_node_state(nid
, N_MEMORY
) {
386 unsigned long start_pfn
, end_pfn
;
388 start_pfn
= node_start_pfn(nid
);
389 end_pfn
= node_end_pfn(nid
);
391 * start_pfn and end_pfn may not be aligned to SECTION and the
392 * page->flags of out of node pages are not initialized. So we
393 * scan [start_pfn, the biggest section's pfn < end_pfn) here.
395 for (pfn
= start_pfn
; pfn
< end_pfn
;
396 pfn
= ALIGN(pfn
+ 1, PAGES_PER_SECTION
)) {
401 * Nodes's pfns can be overlapping.
402 * We know some arch can have a nodes layout such as
403 * -------------pfn-------------->
404 * N0 | N1 | N2 | N0 | N1 | N2|....
406 if (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
)