blob: 12a48a88c0d8cb00dd55b2adcea3e536493dee1a [file] [log] [blame]
Tejun Heo9f645532010-04-09 18:57:01 +09001/*
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_and_bitmap - get temp pages array and bitmap
24 * @chunk: chunk of interest
25 * @bitmapp: output parameter for bitmap
26 * @may_alloc: may allocate the array
27 *
28 * Returns pointer to array of pointers to struct page and bitmap,
29 * both of which can be indexed with pcpu_page_idx(). The returned
30 * array is cleared to zero and *@bitmapp is copied from
31 * @chunk->populated. Note that there is only one array and bitmap
32 * and access exclusion is the caller's responsibility.
33 *
34 * CONTEXT:
35 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
36 * Otherwise, don't care.
37 *
38 * RETURNS:
39 * Pointer to temp pages array on success, NULL on failure.
40 */
41static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
42 unsigned long **bitmapp,
43 bool may_alloc)
44{
45 static struct page **pages;
46 static unsigned long *bitmap;
47 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
48 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
49 sizeof(unsigned long);
50
51 if (!pages || !bitmap) {
52 if (may_alloc && !pages)
Bob Liu90459ce2011-08-04 11:02:33 +020053 pages = pcpu_mem_zalloc(pages_size);
Tejun Heo9f645532010-04-09 18:57:01 +090054 if (may_alloc && !bitmap)
Bob Liu90459ce2011-08-04 11:02:33 +020055 bitmap = pcpu_mem_zalloc(bitmap_size);
Tejun Heo9f645532010-04-09 18:57:01 +090056 if (!pages || !bitmap)
57 return NULL;
58 }
59
Tejun Heo9f645532010-04-09 18:57:01 +090060 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
61
62 *bitmapp = bitmap;
63 return pages;
64}
65
66/**
67 * pcpu_free_pages - free pages which were allocated for @chunk
68 * @chunk: chunk pages were allocated for
69 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
70 * @populated: populated bitmap
71 * @page_start: page index of the first page to be freed
72 * @page_end: page index of the last page to be freed + 1
73 *
74 * Free pages [@page_start and @page_end) in @pages for all units.
75 * The pages were allocated for @chunk.
76 */
77static void pcpu_free_pages(struct pcpu_chunk *chunk,
78 struct page **pages, unsigned long *populated,
79 int page_start, int page_end)
80{
81 unsigned int cpu;
82 int i;
83
84 for_each_possible_cpu(cpu) {
85 for (i = page_start; i < page_end; i++) {
86 struct page *page = pages[pcpu_page_idx(cpu, i)];
87
88 if (page)
89 __free_page(page);
90 }
91 }
92}
93
94/**
95 * pcpu_alloc_pages - allocates pages for @chunk
96 * @chunk: target chunk
97 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
98 * @populated: populated bitmap
99 * @page_start: page index of the first page to be allocated
100 * @page_end: page index of the last page to be allocated + 1
101 *
102 * Allocate pages [@page_start,@page_end) into @pages for all units.
103 * The allocation is for @chunk. Percpu core doesn't care about the
104 * content of @pages and will pass it verbatim to pcpu_map_pages().
105 */
106static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
107 struct page **pages, unsigned long *populated,
108 int page_start, int page_end)
109{
110 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
111 unsigned int cpu;
112 int i;
113
114 for_each_possible_cpu(cpu) {
115 for (i = page_start; i < page_end; i++) {
116 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
117
118 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
119 if (!*pagep) {
120 pcpu_free_pages(chunk, pages, populated,
121 page_start, page_end);
122 return -ENOMEM;
123 }
124 }
125 }
126 return 0;
127}
128
129/**
130 * pcpu_pre_unmap_flush - flush cache prior to unmapping
131 * @chunk: chunk the regions to be flushed belongs to
132 * @page_start: page index of the first page to be flushed
133 * @page_end: page index of the last page to be flushed + 1
134 *
135 * Pages in [@page_start,@page_end) of @chunk are about to be
136 * unmapped. Flush cache. As each flushing trial can be very
137 * expensive, issue flush on the whole region at once rather than
138 * doing it for each cpu. This could be an overkill but is more
139 * scalable.
140 */
141static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
142 int page_start, int page_end)
143{
144 flush_cache_vunmap(
Tejun Heoa855b842011-11-18 10:55:35 -0800145 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
146 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
Tejun Heo9f645532010-04-09 18:57:01 +0900147}
148
149static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
150{
151 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
152}
153
154/**
155 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
156 * @chunk: chunk of interest
157 * @pages: pages array which can be used to pass information to free
158 * @populated: populated bitmap
159 * @page_start: page index of the first page to unmap
160 * @page_end: page index of the last page to unmap + 1
161 *
162 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
163 * Corresponding elements in @pages were cleared by the caller and can
164 * be used to carry information to pcpu_free_pages() which will be
165 * called after all unmaps are finished. The caller should call
166 * proper pre/post flush functions.
167 */
168static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
169 struct page **pages, unsigned long *populated,
170 int page_start, int page_end)
171{
172 unsigned int cpu;
173 int i;
174
175 for_each_possible_cpu(cpu) {
176 for (i = page_start; i < page_end; i++) {
177 struct page *page;
178
179 page = pcpu_chunk_page(chunk, cpu, i);
180 WARN_ON(!page);
181 pages[pcpu_page_idx(cpu, i)] = page;
182 }
183 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
184 page_end - page_start);
185 }
186
187 for (i = page_start; i < page_end; i++)
188 __clear_bit(i, populated);
189}
190
191/**
192 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
193 * @chunk: pcpu_chunk the regions to be flushed belong to
194 * @page_start: page index of the first page to be flushed
195 * @page_end: page index of the last page to be flushed + 1
196 *
197 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
198 * TLB for the regions. This can be skipped if the area is to be
199 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
200 *
201 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
202 * for the whole region.
203 */
204static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
205 int page_start, int page_end)
206{
207 flush_tlb_kernel_range(
Tejun Heoa855b842011-11-18 10:55:35 -0800208 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
209 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
Tejun Heo9f645532010-04-09 18:57:01 +0900210}
211
212static int __pcpu_map_pages(unsigned long addr, struct page **pages,
213 int nr_pages)
214{
215 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
216 PAGE_KERNEL, pages);
217}
218
219/**
220 * pcpu_map_pages - map pages into a pcpu_chunk
221 * @chunk: chunk of interest
222 * @pages: pages array containing pages to be mapped
223 * @populated: populated bitmap
224 * @page_start: page index of the first page to map
225 * @page_end: page index of the last page to map + 1
226 *
227 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
228 * caller is responsible for calling pcpu_post_map_flush() after all
229 * mappings are complete.
230 *
231 * This function is responsible for setting corresponding bits in
232 * @chunk->populated bitmap and whatever is necessary for reverse
233 * lookup (addr -> chunk).
234 */
235static int pcpu_map_pages(struct pcpu_chunk *chunk,
236 struct page **pages, unsigned long *populated,
237 int page_start, int page_end)
238{
239 unsigned int cpu, tcpu;
240 int i, err;
241
242 for_each_possible_cpu(cpu) {
243 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
244 &pages[pcpu_page_idx(cpu, page_start)],
245 page_end - page_start);
246 if (err < 0)
247 goto err;
248 }
249
250 /* mapping successful, link chunk and mark populated */
251 for (i = page_start; i < page_end; i++) {
252 for_each_possible_cpu(cpu)
253 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
254 chunk);
255 __set_bit(i, populated);
256 }
257
258 return 0;
259
260err:
261 for_each_possible_cpu(tcpu) {
262 if (tcpu == cpu)
263 break;
264 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
265 page_end - page_start);
266 }
267 return err;
268}
269
270/**
271 * pcpu_post_map_flush - flush cache after mapping
272 * @chunk: pcpu_chunk the regions to be flushed belong to
273 * @page_start: page index of the first page to be flushed
274 * @page_end: page index of the last page to be flushed + 1
275 *
276 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
277 * cache.
278 *
279 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
280 * for the whole region.
281 */
282static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
283 int page_start, int page_end)
284{
285 flush_cache_vmap(
Tejun Heoa855b842011-11-18 10:55:35 -0800286 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
287 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
Tejun Heo9f645532010-04-09 18:57:01 +0900288}
289
290/**
291 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
292 * @chunk: chunk of interest
293 * @off: offset to the area to populate
294 * @size: size of the area to populate in bytes
295 *
296 * For each cpu, populate and map pages [@page_start,@page_end) into
297 * @chunk. The area is cleared on return.
298 *
299 * CONTEXT:
300 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
301 */
302static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
303{
304 int page_start = PFN_DOWN(off);
305 int page_end = PFN_UP(off + size);
306 int free_end = page_start, unmap_end = page_start;
307 struct page **pages;
308 unsigned long *populated;
309 unsigned int cpu;
310 int rs, re, rc;
311
312 /* quick path, check whether all pages are already there */
313 rs = page_start;
314 pcpu_next_pop(chunk, &rs, &re, page_end);
315 if (rs == page_start && re == page_end)
316 goto clear;
317
318 /* need to allocate and map pages, this chunk can't be immutable */
319 WARN_ON(chunk->immutable);
320
321 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
322 if (!pages)
323 return -ENOMEM;
324
325 /* alloc and map */
326 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
327 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
328 if (rc)
329 goto err_free;
330 free_end = re;
331 }
332
333 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
334 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
335 if (rc)
336 goto err_unmap;
337 unmap_end = re;
338 }
339 pcpu_post_map_flush(chunk, page_start, page_end);
340
341 /* commit new bitmap */
342 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
343clear:
344 for_each_possible_cpu(cpu)
345 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
346 return 0;
347
348err_unmap:
349 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
350 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
351 pcpu_unmap_pages(chunk, pages, populated, rs, re);
352 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
353err_free:
354 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
355 pcpu_free_pages(chunk, pages, populated, rs, re);
356 return rc;
357}
358
359/**
360 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
361 * @chunk: chunk to depopulate
362 * @off: offset to the area to depopulate
363 * @size: size of the area to depopulate in bytes
364 * @flush: whether to flush cache and tlb or not
365 *
366 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
367 * from @chunk. If @flush is true, vcache is flushed before unmapping
368 * and tlb after.
369 *
370 * CONTEXT:
371 * pcpu_alloc_mutex.
372 */
373static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
374{
375 int page_start = PFN_DOWN(off);
376 int page_end = PFN_UP(off + size);
377 struct page **pages;
378 unsigned long *populated;
379 int rs, re;
380
381 /* quick path, check whether it's empty already */
382 rs = page_start;
383 pcpu_next_unpop(chunk, &rs, &re, page_end);
384 if (rs == page_start && re == page_end)
385 return;
386
387 /* immutable chunks can't be depopulated */
388 WARN_ON(chunk->immutable);
389
390 /*
391 * If control reaches here, there must have been at least one
392 * successful population attempt so the temp pages array must
393 * be available now.
394 */
395 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
396 BUG_ON(!pages);
397
398 /* unmap and free */
399 pcpu_pre_unmap_flush(chunk, page_start, page_end);
400
401 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
402 pcpu_unmap_pages(chunk, pages, populated, rs, re);
403
404 /* no need to flush tlb, vmalloc will handle it lazily */
405
406 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
407 pcpu_free_pages(chunk, pages, populated, rs, re);
408
409 /* commit new bitmap */
410 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
411}
412
413static struct pcpu_chunk *pcpu_create_chunk(void)
414{
415 struct pcpu_chunk *chunk;
416 struct vm_struct **vms;
417
418 chunk = pcpu_alloc_chunk();
419 if (!chunk)
420 return NULL;
421
422 vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
David Rientjesec3f64f2011-01-13 15:46:01 -0800423 pcpu_nr_groups, pcpu_atom_size);
Tejun Heo9f645532010-04-09 18:57:01 +0900424 if (!vms) {
425 pcpu_free_chunk(chunk);
426 return NULL;
427 }
428
429 chunk->data = vms;
430 chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
431 return chunk;
432}
433
434static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
435{
436 if (chunk && chunk->data)
437 pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
438 pcpu_free_chunk(chunk);
439}
440
441static struct page *pcpu_addr_to_page(void *addr)
442{
443 return vmalloc_to_page(addr);
444}
445
446static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
447{
448 /* no extra restriction */
449 return 0;
450}