mm: migration: add migrate_entry_wait_huge()
[linux-3.10.git] / mm / percpu-vm.c
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
13 static 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  */
41 static 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)
53                         pages = pcpu_mem_zalloc(pages_size);
54                 if (may_alloc && !bitmap)
55                         bitmap = pcpu_mem_zalloc(bitmap_size);
56                 if (!pages || !bitmap)
57                         return NULL;
58         }
59
60         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  */
77 static 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  */
106 static 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  */
141 static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
142                                  int page_start, int page_end)
143 {
144         flush_cache_vunmap(
145                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
146                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
147 }
148
149 static 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  */
168 static 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         bitmap_clear(populated, page_start, page_end - page_start);
188 }
189
190 /**
191  * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
192  * @chunk: pcpu_chunk the regions to be flushed belong to
193  * @page_start: page index of the first page to be flushed
194  * @page_end: page index of the last page to be flushed + 1
195  *
196  * Pages [@page_start,@page_end) of @chunk have been unmapped.  Flush
197  * TLB for the regions.  This can be skipped if the area is to be
198  * returned to vmalloc as vmalloc will handle TLB flushing lazily.
199  *
200  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
201  * for the whole region.
202  */
203 static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
204                                       int page_start, int page_end)
205 {
206         flush_tlb_kernel_range(
207                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
208                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
209 }
210
211 static int __pcpu_map_pages(unsigned long addr, struct page **pages,
212                             int nr_pages)
213 {
214         return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
215                                         PAGE_KERNEL, pages);
216 }
217
218 /**
219  * pcpu_map_pages - map pages into a pcpu_chunk
220  * @chunk: chunk of interest
221  * @pages: pages array containing pages to be mapped
222  * @populated: populated bitmap
223  * @page_start: page index of the first page to map
224  * @page_end: page index of the last page to map + 1
225  *
226  * For each cpu, map pages [@page_start,@page_end) into @chunk.  The
227  * caller is responsible for calling pcpu_post_map_flush() after all
228  * mappings are complete.
229  *
230  * This function is responsible for setting corresponding bits in
231  * @chunk->populated bitmap and whatever is necessary for reverse
232  * lookup (addr -> chunk).
233  */
234 static int pcpu_map_pages(struct pcpu_chunk *chunk,
235                           struct page **pages, unsigned long *populated,
236                           int page_start, int page_end)
237 {
238         unsigned int cpu, tcpu;
239         int i, err;
240
241         for_each_possible_cpu(cpu) {
242                 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
243                                        &pages[pcpu_page_idx(cpu, page_start)],
244                                        page_end - page_start);
245                 if (err < 0)
246                         goto err;
247         }
248
249         /* mapping successful, link chunk and mark populated */
250         for (i = page_start; i < page_end; i++) {
251                 for_each_possible_cpu(cpu)
252                         pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
253                                             chunk);
254                 __set_bit(i, populated);
255         }
256
257         return 0;
258
259 err:
260         for_each_possible_cpu(tcpu) {
261                 if (tcpu == cpu)
262                         break;
263                 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
264                                    page_end - page_start);
265         }
266         return err;
267 }
268
269 /**
270  * pcpu_post_map_flush - flush cache after mapping
271  * @chunk: pcpu_chunk the regions to be flushed belong to
272  * @page_start: page index of the first page to be flushed
273  * @page_end: page index of the last page to be flushed + 1
274  *
275  * Pages [@page_start,@page_end) of @chunk have been mapped.  Flush
276  * cache.
277  *
278  * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
279  * for the whole region.
280  */
281 static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
282                                 int page_start, int page_end)
283 {
284         flush_cache_vmap(
285                 pcpu_chunk_addr(chunk, pcpu_low_unit_cpu, page_start),
286                 pcpu_chunk_addr(chunk, pcpu_high_unit_cpu, page_end));
287 }
288
289 /**
290  * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
291  * @chunk: chunk of interest
292  * @off: offset to the area to populate
293  * @size: size of the area to populate in bytes
294  *
295  * For each cpu, populate and map pages [@page_start,@page_end) into
296  * @chunk.  The area is cleared on return.
297  *
298  * CONTEXT:
299  * pcpu_alloc_mutex, does GFP_KERNEL allocation.
300  */
301 static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
302 {
303         int page_start = PFN_DOWN(off);
304         int page_end = PFN_UP(off + size);
305         int free_end = page_start, unmap_end = page_start;
306         struct page **pages;
307         unsigned long *populated;
308         unsigned int cpu;
309         int rs, re, rc;
310
311         /* quick path, check whether all pages are already there */
312         rs = page_start;
313         pcpu_next_pop(chunk, &rs, &re, page_end);
314         if (rs == page_start && re == page_end)
315                 goto clear;
316
317         /* need to allocate and map pages, this chunk can't be immutable */
318         WARN_ON(chunk->immutable);
319
320         pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
321         if (!pages)
322                 return -ENOMEM;
323
324         /* alloc and map */
325         pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
326                 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
327                 if (rc)
328                         goto err_free;
329                 free_end = re;
330         }
331
332         pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
333                 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
334                 if (rc)
335                         goto err_unmap;
336                 unmap_end = re;
337         }
338         pcpu_post_map_flush(chunk, page_start, page_end);
339
340         /* commit new bitmap */
341         bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
342 clear:
343         for_each_possible_cpu(cpu)
344                 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
345         return 0;
346
347 err_unmap:
348         pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
349         pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
350                 pcpu_unmap_pages(chunk, pages, populated, rs, re);
351         pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
352 err_free:
353         pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
354                 pcpu_free_pages(chunk, pages, populated, rs, re);
355         return rc;
356 }
357
358 /**
359  * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
360  * @chunk: chunk to depopulate
361  * @off: offset to the area to depopulate
362  * @size: size of the area to depopulate in bytes
363  *
364  * For each cpu, depopulate and unmap pages [@page_start,@page_end)
365  * from @chunk.  If @flush is true, vcache is flushed before unmapping
366  * and tlb after.
367  *
368  * CONTEXT:
369  * pcpu_alloc_mutex.
370  */
371 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
372 {
373         int page_start = PFN_DOWN(off);
374         int page_end = PFN_UP(off + size);
375         struct page **pages;
376         unsigned long *populated;
377         int rs, re;
378
379         /* quick path, check whether it's empty already */
380         rs = page_start;
381         pcpu_next_unpop(chunk, &rs, &re, page_end);
382         if (rs == page_start && re == page_end)
383                 return;
384
385         /* immutable chunks can't be depopulated */
386         WARN_ON(chunk->immutable);
387
388         /*
389          * If control reaches here, there must have been at least one
390          * successful population attempt so the temp pages array must
391          * be available now.
392          */
393         pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
394         BUG_ON(!pages);
395
396         /* unmap and free */
397         pcpu_pre_unmap_flush(chunk, page_start, page_end);
398
399         pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
400                 pcpu_unmap_pages(chunk, pages, populated, rs, re);
401
402         /* no need to flush tlb, vmalloc will handle it lazily */
403
404         pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
405                 pcpu_free_pages(chunk, pages, populated, rs, re);
406
407         /* commit new bitmap */
408         bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
409 }
410
411 static struct pcpu_chunk *pcpu_create_chunk(void)
412 {
413         struct pcpu_chunk *chunk;
414         struct vm_struct **vms;
415
416         chunk = pcpu_alloc_chunk();
417         if (!chunk)
418                 return NULL;
419
420         vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
421                                 pcpu_nr_groups, pcpu_atom_size);
422         if (!vms) {
423                 pcpu_free_chunk(chunk);
424                 return NULL;
425         }
426
427         chunk->data = vms;
428         chunk->base_addr = vms[0]->addr - pcpu_group_offsets[0];
429         return chunk;
430 }
431
432 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
433 {
434         if (chunk && chunk->data)
435                 pcpu_free_vm_areas(chunk->data, pcpu_nr_groups);
436         pcpu_free_chunk(chunk);
437 }
438
439 static struct page *pcpu_addr_to_page(void *addr)
440 {
441         return vmalloc_to_page(addr);
442 }
443
444 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
445 {
446         /* no extra restriction */
447         return 0;
448 }