1/*
2 * mm/percpu-vm.c - vmalloc area based chunk allocation
3 *
4 * Copyright (C) 2010		SUSE Linux Products GmbH
5 * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * Chunks are mapped into vmalloc areas and populated page by page.
10 * This is the default chunk allocator.
11 */
12
13static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
14				    unsigned int cpu, int page_idx)
15{
16	/* must not be used on pre-mapped chunk */
17	WARN_ON(chunk->immutable);
18
19	return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
20}
21
22/**
23 * pcpu_get_pages - get temp pages array
24 * @chunk: chunk of interest
25 *
26 * Returns pointer to array of pointers to struct page which can be indexed
27 * with pcpu_page_idx().  Note that there is only one array and accesses
28 * should be serialized by pcpu_alloc_mutex.
29 *
30 * RETURNS:
31 * Pointer to temp pages array on success.
32 */
33static struct page **pcpu_get_pages(struct pcpu_chunk *chunk_alloc)
34{
35	static struct page **pages;
36	size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
37
38	lockdep_assert_held(&pcpu_alloc_mutex);
39
40	if (!pages)
41		pages = pcpu_mem_zalloc(pages_size);
42	return pages;
43}
44
45/**
46 * pcpu_free_pages - free pages which were allocated for @chunk
47 * @chunk: chunk pages were allocated for
48 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
49 * @page_start: page index of the first page to be freed
50 * @page_end: page index of the last page to be freed + 1
51 *
52 * Free pages [@page_start and @page_end) in @pages for all units.
53 * The pages were allocated for @chunk.
54 */
55static void pcpu_free_pages(struct pcpu_chunk *chunk,
56			    struct page **pages, int page_start, int page_end)
57{
58	unsigned int cpu;
59	int i;
60
61	for_each_possible_cpu(cpu) {
62		for (i = page_start; i < page_end; i++) {
63			struct page *page = pages[pcpu_page_idx(cpu, i)];
64
65			if (page)
66				__free_page(page);
67		}
68	}
69}
70
71/**
72 * pcpu_alloc_pages - allocates pages for @chunk
73 * @chunk: target chunk
74 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
75 * @page_start: page index of the first page to be allocated
76 * @page_end: page index of the last page to be allocated + 1
77 *
78 * Allocate pages [@page_start,@page_end) into @pages for all units.
79 * The allocation is for @chunk.  Percpu core doesn't care about the
80 * content of @pages and will pass it verbatim to pcpu_map_pages().
81 */
82static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
83			    struct page **pages, int page_start, int page_end)
84{
85	const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
86	unsigned int cpu, tcpu;
87	int i;
88
89	for_each_possible_cpu(cpu) {
90		for (i = page_start; i < page_end; i++) {
91			struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
92
93			*pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
94			if (!*pagep)
95				goto err;
96		}
97	}
98	return 0;
99
100err:
101	while (--i >= page_start)
102		__free_page(pages[pcpu_page_idx(cpu, i)]);
103
104	for_each_possible_cpu(tcpu) {
105		if (tcpu == cpu)
106			break;
107		for (i = page_start; i < page_end; i++)
108			__free_page(pages[pcpu_page_idx(tcpu, i)]);
109	}
110	return -ENOMEM;
111}
112
113/**
114 * pcpu_pre_unmap_flush - flush cache prior to unmapping
115 * @chunk: chunk the regions to be flushed belongs to
116 * @page_start: page index of the first page to be flushed
117 * @page_end: page index of the last page to be flushed + 1
118 *
119 * Pages in [@page_start,@page_end) of @chunk are about to be
120 * unmapped.  Flush cache.  As each flushing trial can be very
121 * expensive, issue flush on the whole region at once rather than
122 * doing it for each cpu.  This could be an overkill but is more
123 * scalable.
124 */
125static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
126				 int page_start, int page_end)
127{
128	flush_cache_vunmap(
129		pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
130		pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
131}
132
133static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
134{
135	unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
136}
137
138/**
139 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
140 * @chunk: chunk of interest
141 * @pages: pages array which can be used to pass information to free
142 * @page_start: page index of the first page to unmap
143 * @page_end: page index of the last page to unmap + 1
144 *
145 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
146 * Corresponding elements in @pages were cleared by the caller and can
147 * be used to carry information to pcpu_free_pages() which will be
148 * called after all unmaps are finished.  The caller should call
149 * proper pre/post flush functions.
150 */
151static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
152			     struct page **pages, int page_start, int page_end)
153{
154	unsigned int cpu;
155	int i;
156
157	for_each_possible_cpu(cpu) {
158		for (i = page_start; i < page_end; i++) {
159			struct page *page;
160
161			page = pcpu_chunk_page(chunk, cpu, i);
162			WARN_ON(!page);
163			pages[pcpu_page_idx(cpu, i)] = page;
164		}
165		__pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
166				   page_end - page_start);
167	}
168}
169
170/**
171 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
172 * @chunk: pcpu_chunk the regions to be flushed belong to
173 * @page_start: page index of the first page to be flushed
174 * @page_end: page index of the last page to be flushed + 1
175 *
176 * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
177 * TLB for the regions.  This can be skipped if the area is to be
178 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
179 *
180 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
181 * for the whole region.
182 */
183static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
184				      int page_start, int page_end)
185{
186	flush_tlb_kernel_range(
187		pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
188		pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
189}
190
191static int __pcpu_map_pages(unsigned long addr, struct page **pages,
192			    int nr_pages)
193{
194	return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
195					PAGE_KERNEL, pages);
196}
197
198/**
199 * pcpu_map_pages - map pages into a pcpu_chunk
200 * @chunk: chunk of interest
201 * @pages: pages array containing pages to be mapped
202 * @page_start: page index of the first page to map
203 * @page_end: page index of the last page to map + 1
204 *
205 * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
206 * caller is responsible for calling pcpu_post_map_flush() after all
207 * mappings are complete.
208 *
209 * This function is responsible for setting up whatever is necessary for
210 * reverse lookup (addr -> chunk).
211 */
212static int pcpu_map_pages(struct pcpu_chunk *chunk,
213			  struct page **pages, int page_start, int page_end)
214{
215	unsigned int cpu, tcpu;
216	int i, err;
217
218	for_each_possible_cpu(cpu) {
219		err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
220				       &pages[pcpu_page_idx(cpu, page_start)],
221				       page_end - page_start);
222		if (err < 0)
223			goto err;
224
225		for (i = page_start; i < page_end; i++)
226			pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
227					    chunk);
228	}
229	return 0;
230err:
231	for_each_possible_cpu(tcpu) {
232		if (tcpu == cpu)
233			break;
234		__pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
235				   page_end - page_start);
236	}
237	pcpu_post_unmap_tlb_flush(chunk, page_start, page_end);
238	return err;
239}
240
241/**
242 * pcpu_post_map_flush - flush cache after mapping
243 * @chunk: pcpu_chunk the regions to be flushed belong to
244 * @page_start: page index of the first page to be flushed
245 * @page_end: page index of the last page to be flushed + 1
246 *
247 * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
248 * cache.
249 *
250 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
251 * for the whole region.
252 */
253static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
254				int page_start, int page_end)
255{
256	flush_cache_vmap(
257		pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
258		pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
259}
260
261/**
262 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
263 * @chunk: chunk of interest
264 * @page_start: the start page
265 * @page_end: the end page
266 *
267 * For each cpu, populate and map pages [@page_start,@page_end) into
268 * @chunk.
269 *
270 * CONTEXT:
271 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
272 */
273static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
274			       int page_start, int page_end)
275{
276	struct page **pages;
277
278	pages = pcpu_get_pages(chunk);
279	if (!pages)
280		return -ENOMEM;
281
282	if (pcpu_alloc_pages(chunk, pages, page_start, page_end))
283		return -ENOMEM;
284
285	if (pcpu_map_pages(chunk, pages, page_start, page_end)) {
286		pcpu_free_pages(chunk, pages, page_start, page_end);
287		return -ENOMEM;
288	}
289	pcpu_post_map_flush(chunk, page_start, page_end);
290
291	return 0;
292}
293
294/**
295 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
296 * @chunk: chunk to depopulate
297 * @page_start: the start page
298 * @page_end: the end page
299 *
300 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
301 * from @chunk.
302 *
303 * CONTEXT:
304 * pcpu_alloc_mutex.
305 */
306static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
307				  int page_start, int page_end)
308{
309	struct page **pages;
310
311	/*
312	 * If control reaches here, there must have been at least one
313	 * successful population attempt so the temp pages array must
314	 * be available now.
315	 */
316	pages = pcpu_get_pages(chunk);
317	BUG_ON(!pages);
318
319	/* unmap and free */
320	pcpu_pre_unmap_flush(chunk, page_start, page_end);
321
322	pcpu_unmap_pages(chunk, pages, page_start, page_end);
323
324	/* no need to flush tlb, vmalloc will handle it lazily */
325
326	pcpu_free_pages(chunk, pages, page_start, page_end);
327}
328
329static struct pcpu_chunk *pcpu_create_chunk(void)
330{
331	struct pcpu_chunk *chunk;
332	struct vm_struct **vms;
333
334	chunk = pcpu_alloc_chunk();
335	if (!chunk)
336		return NULL;
337
338	vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
339				pcpu_nr_groups, pcpu_atom_size);
340	if (!vms) {
341		pcpu_free_chunk(chunk);
342		return NULL;
343	}
344
345	chunk->data = vms;
346	chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
347	return chunk;
348}
349
350static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
351{
352	if (chunk && chunk->data)
353		pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
354	pcpu_free_chunk(chunk);
355}
356
357static struct page *pcpu_addr_to_page(void *addr)
358{
359	return vmalloc_to_page(addr);
360}
361
362static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
363{
364	/* no extra restriction */
365	return 0;
366}
367