]>
Commit | Line | Data |
---|---|---|
fbf59bc9 TH |
1 | /* |
2 | * linux/mm/percpu.c - percpu memory allocator | |
3 | * | |
4 | * Copyright (C) 2009 SUSE Linux Products GmbH | |
5 | * Copyright (C) 2009 Tejun Heo <tj@kernel.org> | |
6 | * | |
7 | * This file is released under the GPLv2. | |
8 | * | |
9 | * This is percpu allocator which can handle both static and dynamic | |
10 | * areas. Percpu areas are allocated in chunks in vmalloc area. Each | |
11 | * chunk is consisted of num_possible_cpus() units and the first chunk | |
12 | * is used for static percpu variables in the kernel image (special | |
13 | * boot time alloc/init handling necessary as these areas need to be | |
14 | * brought up before allocation services are running). Unit grows as | |
15 | * necessary and all units grow or shrink in unison. When a chunk is | |
16 | * filled up, another chunk is allocated. ie. in vmalloc area | |
17 | * | |
18 | * c0 c1 c2 | |
19 | * ------------------- ------------------- ------------ | |
20 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u | |
21 | * ------------------- ...... ------------------- .... ------------ | |
22 | * | |
23 | * Allocation is done in offset-size areas of single unit space. Ie, | |
24 | * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, | |
25 | * c1:u1, c1:u2 and c1:u3. Percpu access can be done by configuring | |
26 | * percpu base registers UNIT_SIZE apart. | |
27 | * | |
28 | * There are usually many small percpu allocations many of them as | |
29 | * small as 4 bytes. The allocator organizes chunks into lists | |
30 | * according to free size and tries to allocate from the fullest one. | |
31 | * Each chunk keeps the maximum contiguous area size hint which is | |
32 | * guaranteed to be eqaul to or larger than the maximum contiguous | |
33 | * area in the chunk. This helps the allocator not to iterate the | |
34 | * chunk maps unnecessarily. | |
35 | * | |
36 | * Allocation state in each chunk is kept using an array of integers | |
37 | * on chunk->map. A positive value in the map represents a free | |
38 | * region and negative allocated. Allocation inside a chunk is done | |
39 | * by scanning this map sequentially and serving the first matching | |
40 | * entry. This is mostly copied from the percpu_modalloc() allocator. | |
41 | * Chunks are also linked into a rb tree to ease address to chunk | |
42 | * mapping during free. | |
43 | * | |
44 | * To use this allocator, arch code should do the followings. | |
45 | * | |
46 | * - define CONFIG_HAVE_DYNAMIC_PER_CPU_AREA | |
47 | * | |
48 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate | |
49 | * regular address to percpu pointer and back | |
50 | * | |
51 | * - use pcpu_setup_static() during percpu area initialization to | |
52 | * setup kernel static percpu area | |
53 | */ | |
54 | ||
55 | #include <linux/bitmap.h> | |
56 | #include <linux/bootmem.h> | |
57 | #include <linux/list.h> | |
58 | #include <linux/mm.h> | |
59 | #include <linux/module.h> | |
60 | #include <linux/mutex.h> | |
61 | #include <linux/percpu.h> | |
62 | #include <linux/pfn.h> | |
63 | #include <linux/rbtree.h> | |
64 | #include <linux/slab.h> | |
65 | #include <linux/vmalloc.h> | |
66 | ||
67 | #include <asm/cacheflush.h> | |
68 | #include <asm/tlbflush.h> | |
69 | ||
d9b55eeb | 70 | #define PCPU_MIN_UNIT_PAGES 16 /* max alloc size in pages */ |
fbf59bc9 TH |
71 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
72 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ | |
73 | ||
74 | struct pcpu_chunk { | |
75 | struct list_head list; /* linked to pcpu_slot lists */ | |
76 | struct rb_node rb_node; /* key is chunk->vm->addr */ | |
77 | int free_size; /* free bytes in the chunk */ | |
78 | int contig_hint; /* max contiguous size hint */ | |
79 | struct vm_struct *vm; /* mapped vmalloc region */ | |
80 | int map_used; /* # of map entries used */ | |
81 | int map_alloc; /* # of map entries allocated */ | |
82 | int *map; /* allocation map */ | |
83 | struct page *page[]; /* #cpus * UNIT_PAGES */ | |
84 | }; | |
85 | ||
fbf59bc9 | 86 | static int pcpu_unit_pages; |
fbf59bc9 TH |
87 | static int pcpu_unit_size; |
88 | static int pcpu_chunk_size; | |
89 | static int pcpu_nr_slots; | |
90 | static size_t pcpu_chunk_struct_size; | |
91 | ||
92 | /* the address of the first chunk which starts with the kernel static area */ | |
93 | void *pcpu_base_addr; | |
94 | EXPORT_SYMBOL_GPL(pcpu_base_addr); | |
95 | ||
96 | /* the size of kernel static area */ | |
97 | static int pcpu_static_size; | |
98 | ||
99 | /* | |
100 | * One mutex to rule them all. | |
101 | * | |
102 | * The following mutex is grabbed in the outermost public alloc/free | |
103 | * interface functions and released only when the operation is | |
104 | * complete. As such, every function in this file other than the | |
105 | * outermost functions are called under pcpu_mutex. | |
106 | * | |
107 | * It can easily be switched to use spinlock such that only the area | |
108 | * allocation and page population commit are protected with it doing | |
109 | * actual [de]allocation without holding any lock. However, given | |
110 | * what this allocator does, I think it's better to let them run | |
111 | * sequentially. | |
112 | */ | |
113 | static DEFINE_MUTEX(pcpu_mutex); | |
114 | ||
115 | static struct list_head *pcpu_slot; /* chunk list slots */ | |
116 | static struct rb_root pcpu_addr_root = RB_ROOT; /* chunks by address */ | |
117 | ||
d9b55eeb | 118 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 119 | { |
cae3aeb8 | 120 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
121 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
122 | } | |
123 | ||
d9b55eeb TH |
124 | static int pcpu_size_to_slot(int size) |
125 | { | |
126 | if (size == pcpu_unit_size) | |
127 | return pcpu_nr_slots - 1; | |
128 | return __pcpu_size_to_slot(size); | |
129 | } | |
130 | ||
fbf59bc9 TH |
131 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
132 | { | |
133 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
134 | return 0; | |
135 | ||
136 | return pcpu_size_to_slot(chunk->free_size); | |
137 | } | |
138 | ||
139 | static int pcpu_page_idx(unsigned int cpu, int page_idx) | |
140 | { | |
d9b55eeb | 141 | return cpu * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
142 | } |
143 | ||
144 | static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk, | |
145 | unsigned int cpu, int page_idx) | |
146 | { | |
147 | return &chunk->page[pcpu_page_idx(cpu, page_idx)]; | |
148 | } | |
149 | ||
150 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, | |
151 | unsigned int cpu, int page_idx) | |
152 | { | |
153 | return (unsigned long)chunk->vm->addr + | |
154 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); | |
155 | } | |
156 | ||
157 | static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk, | |
158 | int page_idx) | |
159 | { | |
160 | return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL; | |
161 | } | |
162 | ||
163 | /** | |
164 | * pcpu_realloc - versatile realloc | |
165 | * @p: the current pointer (can be NULL for new allocations) | |
cae3aeb8 TH |
166 | * @size: the current size in bytes (can be 0 for new allocations) |
167 | * @new_size: the wanted new size in bytes (can be 0 for free) | |
fbf59bc9 TH |
168 | * |
169 | * More robust realloc which can be used to allocate, resize or free a | |
170 | * memory area of arbitrary size. If the needed size goes over | |
171 | * PAGE_SIZE, kernel VM is used. | |
172 | * | |
173 | * RETURNS: | |
174 | * The new pointer on success, NULL on failure. | |
175 | */ | |
176 | static void *pcpu_realloc(void *p, size_t size, size_t new_size) | |
177 | { | |
178 | void *new; | |
179 | ||
180 | if (new_size <= PAGE_SIZE) | |
181 | new = kmalloc(new_size, GFP_KERNEL); | |
182 | else | |
183 | new = vmalloc(new_size); | |
184 | if (new_size && !new) | |
185 | return NULL; | |
186 | ||
187 | memcpy(new, p, min(size, new_size)); | |
188 | if (new_size > size) | |
189 | memset(new + size, 0, new_size - size); | |
190 | ||
191 | if (size <= PAGE_SIZE) | |
192 | kfree(p); | |
193 | else | |
194 | vfree(p); | |
195 | ||
196 | return new; | |
197 | } | |
198 | ||
199 | /** | |
200 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
201 | * @chunk: chunk of interest | |
202 | * @oslot: the previous slot it was on | |
203 | * | |
204 | * This function is called after an allocation or free changed @chunk. | |
205 | * New slot according to the changed state is determined and @chunk is | |
206 | * moved to the slot. | |
207 | */ | |
208 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
209 | { | |
210 | int nslot = pcpu_chunk_slot(chunk); | |
211 | ||
212 | if (oslot != nslot) { | |
213 | if (oslot < nslot) | |
214 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
215 | else | |
216 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
217 | } | |
218 | } | |
219 | ||
220 | static struct rb_node **pcpu_chunk_rb_search(void *addr, | |
221 | struct rb_node **parentp) | |
222 | { | |
223 | struct rb_node **p = &pcpu_addr_root.rb_node; | |
224 | struct rb_node *parent = NULL; | |
225 | struct pcpu_chunk *chunk; | |
226 | ||
227 | while (*p) { | |
228 | parent = *p; | |
229 | chunk = rb_entry(parent, struct pcpu_chunk, rb_node); | |
230 | ||
231 | if (addr < chunk->vm->addr) | |
232 | p = &(*p)->rb_left; | |
233 | else if (addr > chunk->vm->addr) | |
234 | p = &(*p)->rb_right; | |
235 | else | |
236 | break; | |
237 | } | |
238 | ||
239 | if (parentp) | |
240 | *parentp = parent; | |
241 | return p; | |
242 | } | |
243 | ||
244 | /** | |
245 | * pcpu_chunk_addr_search - search for chunk containing specified address | |
246 | * @addr: address to search for | |
247 | * | |
248 | * Look for chunk which might contain @addr. More specifically, it | |
249 | * searchs for the chunk with the highest start address which isn't | |
250 | * beyond @addr. | |
251 | * | |
252 | * RETURNS: | |
253 | * The address of the found chunk. | |
254 | */ | |
255 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
256 | { | |
257 | struct rb_node *n, *parent; | |
258 | struct pcpu_chunk *chunk; | |
259 | ||
260 | n = *pcpu_chunk_rb_search(addr, &parent); | |
261 | if (!n) { | |
262 | /* no exactly matching chunk, the parent is the closest */ | |
263 | n = parent; | |
264 | BUG_ON(!n); | |
265 | } | |
266 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); | |
267 | ||
268 | if (addr < chunk->vm->addr) { | |
269 | /* the parent was the next one, look for the previous one */ | |
270 | n = rb_prev(n); | |
271 | BUG_ON(!n); | |
272 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); | |
273 | } | |
274 | ||
275 | return chunk; | |
276 | } | |
277 | ||
278 | /** | |
279 | * pcpu_chunk_addr_insert - insert chunk into address rb tree | |
280 | * @new: chunk to insert | |
281 | * | |
282 | * Insert @new into address rb tree. | |
283 | */ | |
284 | static void pcpu_chunk_addr_insert(struct pcpu_chunk *new) | |
285 | { | |
286 | struct rb_node **p, *parent; | |
287 | ||
288 | p = pcpu_chunk_rb_search(new->vm->addr, &parent); | |
289 | BUG_ON(*p); | |
290 | rb_link_node(&new->rb_node, parent, p); | |
291 | rb_insert_color(&new->rb_node, &pcpu_addr_root); | |
292 | } | |
293 | ||
294 | /** | |
295 | * pcpu_split_block - split a map block | |
296 | * @chunk: chunk of interest | |
297 | * @i: index of map block to split | |
cae3aeb8 TH |
298 | * @head: head size in bytes (can be 0) |
299 | * @tail: tail size in bytes (can be 0) | |
fbf59bc9 TH |
300 | * |
301 | * Split the @i'th map block into two or three blocks. If @head is | |
302 | * non-zero, @head bytes block is inserted before block @i moving it | |
303 | * to @i+1 and reducing its size by @head bytes. | |
304 | * | |
305 | * If @tail is non-zero, the target block, which can be @i or @i+1 | |
306 | * depending on @head, is reduced by @tail bytes and @tail byte block | |
307 | * is inserted after the target block. | |
308 | * | |
309 | * RETURNS: | |
310 | * 0 on success, -errno on failure. | |
311 | */ | |
312 | static int pcpu_split_block(struct pcpu_chunk *chunk, int i, int head, int tail) | |
313 | { | |
314 | int nr_extra = !!head + !!tail; | |
315 | int target = chunk->map_used + nr_extra; | |
316 | ||
317 | /* reallocation required? */ | |
318 | if (chunk->map_alloc < target) { | |
319 | int new_alloc = chunk->map_alloc; | |
320 | int *new; | |
321 | ||
322 | while (new_alloc < target) | |
323 | new_alloc *= 2; | |
324 | ||
325 | new = pcpu_realloc(chunk->map, | |
326 | chunk->map_alloc * sizeof(new[0]), | |
327 | new_alloc * sizeof(new[0])); | |
328 | if (!new) | |
329 | return -ENOMEM; | |
330 | ||
331 | chunk->map_alloc = new_alloc; | |
332 | chunk->map = new; | |
333 | } | |
334 | ||
335 | /* insert a new subblock */ | |
336 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], | |
337 | sizeof(chunk->map[0]) * (chunk->map_used - i)); | |
338 | chunk->map_used += nr_extra; | |
339 | ||
340 | if (head) { | |
341 | chunk->map[i + 1] = chunk->map[i] - head; | |
342 | chunk->map[i++] = head; | |
343 | } | |
344 | if (tail) { | |
345 | chunk->map[i++] -= tail; | |
346 | chunk->map[i] = tail; | |
347 | } | |
348 | return 0; | |
349 | } | |
350 | ||
351 | /** | |
352 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
353 | * @chunk: chunk of interest | |
cae3aeb8 | 354 | * @size: wanted size in bytes |
fbf59bc9 TH |
355 | * @align: wanted align |
356 | * | |
357 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
358 | * Note that this function only allocates the offset. It doesn't | |
359 | * populate or map the area. | |
360 | * | |
361 | * RETURNS: | |
362 | * Allocated offset in @chunk on success, -errno on failure. | |
363 | */ | |
364 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) | |
365 | { | |
366 | int oslot = pcpu_chunk_slot(chunk); | |
367 | int max_contig = 0; | |
368 | int i, off; | |
369 | ||
370 | /* | |
371 | * The static chunk initially doesn't have map attached | |
372 | * because kmalloc wasn't available during init. Give it one. | |
373 | */ | |
374 | if (unlikely(!chunk->map)) { | |
375 | chunk->map = pcpu_realloc(NULL, 0, | |
376 | PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); | |
377 | if (!chunk->map) | |
378 | return -ENOMEM; | |
379 | ||
380 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; | |
381 | chunk->map[chunk->map_used++] = -pcpu_static_size; | |
382 | if (chunk->free_size) | |
383 | chunk->map[chunk->map_used++] = chunk->free_size; | |
384 | } | |
385 | ||
386 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { | |
387 | bool is_last = i + 1 == chunk->map_used; | |
388 | int head, tail; | |
389 | ||
390 | /* extra for alignment requirement */ | |
391 | head = ALIGN(off, align) - off; | |
392 | BUG_ON(i == 0 && head != 0); | |
393 | ||
394 | if (chunk->map[i] < 0) | |
395 | continue; | |
396 | if (chunk->map[i] < head + size) { | |
397 | max_contig = max(chunk->map[i], max_contig); | |
398 | continue; | |
399 | } | |
400 | ||
401 | /* | |
402 | * If head is small or the previous block is free, | |
403 | * merge'em. Note that 'small' is defined as smaller | |
404 | * than sizeof(int), which is very small but isn't too | |
405 | * uncommon for percpu allocations. | |
406 | */ | |
407 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { | |
408 | if (chunk->map[i - 1] > 0) | |
409 | chunk->map[i - 1] += head; | |
410 | else { | |
411 | chunk->map[i - 1] -= head; | |
412 | chunk->free_size -= head; | |
413 | } | |
414 | chunk->map[i] -= head; | |
415 | off += head; | |
416 | head = 0; | |
417 | } | |
418 | ||
419 | /* if tail is small, just keep it around */ | |
420 | tail = chunk->map[i] - head - size; | |
421 | if (tail < sizeof(int)) | |
422 | tail = 0; | |
423 | ||
424 | /* split if warranted */ | |
425 | if (head || tail) { | |
426 | if (pcpu_split_block(chunk, i, head, tail)) | |
427 | return -ENOMEM; | |
428 | if (head) { | |
429 | i++; | |
430 | off += head; | |
431 | max_contig = max(chunk->map[i - 1], max_contig); | |
432 | } | |
433 | if (tail) | |
434 | max_contig = max(chunk->map[i + 1], max_contig); | |
435 | } | |
436 | ||
437 | /* update hint and mark allocated */ | |
438 | if (is_last) | |
439 | chunk->contig_hint = max_contig; /* fully scanned */ | |
440 | else | |
441 | chunk->contig_hint = max(chunk->contig_hint, | |
442 | max_contig); | |
443 | ||
444 | chunk->free_size -= chunk->map[i]; | |
445 | chunk->map[i] = -chunk->map[i]; | |
446 | ||
447 | pcpu_chunk_relocate(chunk, oslot); | |
448 | return off; | |
449 | } | |
450 | ||
451 | chunk->contig_hint = max_contig; /* fully scanned */ | |
452 | pcpu_chunk_relocate(chunk, oslot); | |
453 | ||
454 | /* | |
455 | * Tell the upper layer that this chunk has no area left. | |
456 | * Note that this is not an error condition but a notification | |
457 | * to upper layer that it needs to look at other chunks. | |
458 | * -ENOSPC is chosen as it isn't used in memory subsystem and | |
459 | * matches the meaning in a way. | |
460 | */ | |
461 | return -ENOSPC; | |
462 | } | |
463 | ||
464 | /** | |
465 | * pcpu_free_area - free area to a pcpu_chunk | |
466 | * @chunk: chunk of interest | |
467 | * @freeme: offset of area to free | |
468 | * | |
469 | * Free area starting from @freeme to @chunk. Note that this function | |
470 | * only modifies the allocation map. It doesn't depopulate or unmap | |
471 | * the area. | |
472 | */ | |
473 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) | |
474 | { | |
475 | int oslot = pcpu_chunk_slot(chunk); | |
476 | int i, off; | |
477 | ||
478 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) | |
479 | if (off == freeme) | |
480 | break; | |
481 | BUG_ON(off != freeme); | |
482 | BUG_ON(chunk->map[i] > 0); | |
483 | ||
484 | chunk->map[i] = -chunk->map[i]; | |
485 | chunk->free_size += chunk->map[i]; | |
486 | ||
487 | /* merge with previous? */ | |
488 | if (i > 0 && chunk->map[i - 1] >= 0) { | |
489 | chunk->map[i - 1] += chunk->map[i]; | |
490 | chunk->map_used--; | |
491 | memmove(&chunk->map[i], &chunk->map[i + 1], | |
492 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
493 | i--; | |
494 | } | |
495 | /* merge with next? */ | |
496 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { | |
497 | chunk->map[i] += chunk->map[i + 1]; | |
498 | chunk->map_used--; | |
499 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], | |
500 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); | |
501 | } | |
502 | ||
503 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); | |
504 | pcpu_chunk_relocate(chunk, oslot); | |
505 | } | |
506 | ||
507 | /** | |
508 | * pcpu_unmap - unmap pages out of a pcpu_chunk | |
509 | * @chunk: chunk of interest | |
510 | * @page_start: page index of the first page to unmap | |
511 | * @page_end: page index of the last page to unmap + 1 | |
512 | * @flush: whether to flush cache and tlb or not | |
513 | * | |
514 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. | |
515 | * If @flush is true, vcache is flushed before unmapping and tlb | |
516 | * after. | |
517 | */ | |
518 | static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end, | |
519 | bool flush) | |
520 | { | |
521 | unsigned int last = num_possible_cpus() - 1; | |
522 | unsigned int cpu; | |
523 | ||
524 | /* | |
525 | * Each flushing trial can be very expensive, issue flush on | |
526 | * the whole region at once rather than doing it for each cpu. | |
527 | * This could be an overkill but is more scalable. | |
528 | */ | |
529 | if (flush) | |
530 | flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start), | |
531 | pcpu_chunk_addr(chunk, last, page_end)); | |
532 | ||
533 | for_each_possible_cpu(cpu) | |
534 | unmap_kernel_range_noflush( | |
535 | pcpu_chunk_addr(chunk, cpu, page_start), | |
536 | (page_end - page_start) << PAGE_SHIFT); | |
537 | ||
538 | /* ditto as flush_cache_vunmap() */ | |
539 | if (flush) | |
540 | flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start), | |
541 | pcpu_chunk_addr(chunk, last, page_end)); | |
542 | } | |
543 | ||
544 | /** | |
545 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk | |
546 | * @chunk: chunk to depopulate | |
547 | * @off: offset to the area to depopulate | |
cae3aeb8 | 548 | * @size: size of the area to depopulate in bytes |
fbf59bc9 TH |
549 | * @flush: whether to flush cache and tlb or not |
550 | * | |
551 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) | |
552 | * from @chunk. If @flush is true, vcache is flushed before unmapping | |
553 | * and tlb after. | |
554 | */ | |
cae3aeb8 TH |
555 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size, |
556 | bool flush) | |
fbf59bc9 TH |
557 | { |
558 | int page_start = PFN_DOWN(off); | |
559 | int page_end = PFN_UP(off + size); | |
560 | int unmap_start = -1; | |
561 | int uninitialized_var(unmap_end); | |
562 | unsigned int cpu; | |
563 | int i; | |
564 | ||
565 | for (i = page_start; i < page_end; i++) { | |
566 | for_each_possible_cpu(cpu) { | |
567 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); | |
568 | ||
569 | if (!*pagep) | |
570 | continue; | |
571 | ||
572 | __free_page(*pagep); | |
573 | ||
574 | /* | |
575 | * If it's partial depopulation, it might get | |
576 | * populated or depopulated again. Mark the | |
577 | * page gone. | |
578 | */ | |
579 | *pagep = NULL; | |
580 | ||
581 | unmap_start = unmap_start < 0 ? i : unmap_start; | |
582 | unmap_end = i + 1; | |
583 | } | |
584 | } | |
585 | ||
586 | if (unmap_start >= 0) | |
587 | pcpu_unmap(chunk, unmap_start, unmap_end, flush); | |
588 | } | |
589 | ||
590 | /** | |
591 | * pcpu_map - map pages into a pcpu_chunk | |
592 | * @chunk: chunk of interest | |
593 | * @page_start: page index of the first page to map | |
594 | * @page_end: page index of the last page to map + 1 | |
595 | * | |
596 | * For each cpu, map pages [@page_start,@page_end) into @chunk. | |
597 | * vcache is flushed afterwards. | |
598 | */ | |
599 | static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end) | |
600 | { | |
601 | unsigned int last = num_possible_cpus() - 1; | |
602 | unsigned int cpu; | |
603 | int err; | |
604 | ||
605 | for_each_possible_cpu(cpu) { | |
606 | err = map_kernel_range_noflush( | |
607 | pcpu_chunk_addr(chunk, cpu, page_start), | |
608 | (page_end - page_start) << PAGE_SHIFT, | |
609 | PAGE_KERNEL, | |
610 | pcpu_chunk_pagep(chunk, cpu, page_start)); | |
611 | if (err < 0) | |
612 | return err; | |
613 | } | |
614 | ||
615 | /* flush at once, please read comments in pcpu_unmap() */ | |
616 | flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start), | |
617 | pcpu_chunk_addr(chunk, last, page_end)); | |
618 | return 0; | |
619 | } | |
620 | ||
621 | /** | |
622 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk | |
623 | * @chunk: chunk of interest | |
624 | * @off: offset to the area to populate | |
cae3aeb8 | 625 | * @size: size of the area to populate in bytes |
fbf59bc9 TH |
626 | * |
627 | * For each cpu, populate and map pages [@page_start,@page_end) into | |
628 | * @chunk. The area is cleared on return. | |
629 | */ | |
630 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) | |
631 | { | |
632 | const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; | |
633 | int page_start = PFN_DOWN(off); | |
634 | int page_end = PFN_UP(off + size); | |
635 | int map_start = -1; | |
636 | int map_end; | |
637 | unsigned int cpu; | |
638 | int i; | |
639 | ||
640 | for (i = page_start; i < page_end; i++) { | |
641 | if (pcpu_chunk_page_occupied(chunk, i)) { | |
642 | if (map_start >= 0) { | |
643 | if (pcpu_map(chunk, map_start, map_end)) | |
644 | goto err; | |
645 | map_start = -1; | |
646 | } | |
647 | continue; | |
648 | } | |
649 | ||
650 | map_start = map_start < 0 ? i : map_start; | |
651 | map_end = i + 1; | |
652 | ||
653 | for_each_possible_cpu(cpu) { | |
654 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); | |
655 | ||
656 | *pagep = alloc_pages_node(cpu_to_node(cpu), | |
657 | alloc_mask, 0); | |
658 | if (!*pagep) | |
659 | goto err; | |
660 | } | |
661 | } | |
662 | ||
663 | if (map_start >= 0 && pcpu_map(chunk, map_start, map_end)) | |
664 | goto err; | |
665 | ||
666 | for_each_possible_cpu(cpu) | |
d9b55eeb | 667 | memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0, |
fbf59bc9 TH |
668 | size); |
669 | ||
670 | return 0; | |
671 | err: | |
672 | /* likely under heavy memory pressure, give memory back */ | |
673 | pcpu_depopulate_chunk(chunk, off, size, true); | |
674 | return -ENOMEM; | |
675 | } | |
676 | ||
677 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) | |
678 | { | |
679 | if (!chunk) | |
680 | return; | |
681 | if (chunk->vm) | |
682 | free_vm_area(chunk->vm); | |
683 | pcpu_realloc(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]), 0); | |
684 | kfree(chunk); | |
685 | } | |
686 | ||
687 | static struct pcpu_chunk *alloc_pcpu_chunk(void) | |
688 | { | |
689 | struct pcpu_chunk *chunk; | |
690 | ||
691 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); | |
692 | if (!chunk) | |
693 | return NULL; | |
694 | ||
695 | chunk->map = pcpu_realloc(NULL, 0, | |
696 | PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); | |
697 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; | |
698 | chunk->map[chunk->map_used++] = pcpu_unit_size; | |
699 | ||
700 | chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL); | |
701 | if (!chunk->vm) { | |
702 | free_pcpu_chunk(chunk); | |
703 | return NULL; | |
704 | } | |
705 | ||
706 | INIT_LIST_HEAD(&chunk->list); | |
707 | chunk->free_size = pcpu_unit_size; | |
708 | chunk->contig_hint = pcpu_unit_size; | |
709 | ||
710 | return chunk; | |
711 | } | |
712 | ||
713 | /** | |
714 | * __alloc_percpu - allocate percpu area | |
cae3aeb8 | 715 | * @size: size of area to allocate in bytes |
fbf59bc9 TH |
716 | * @align: alignment of area (max PAGE_SIZE) |
717 | * | |
718 | * Allocate percpu area of @size bytes aligned at @align. Might | |
719 | * sleep. Might trigger writeouts. | |
720 | * | |
721 | * RETURNS: | |
722 | * Percpu pointer to the allocated area on success, NULL on failure. | |
723 | */ | |
724 | void *__alloc_percpu(size_t size, size_t align) | |
725 | { | |
726 | void *ptr = NULL; | |
727 | struct pcpu_chunk *chunk; | |
728 | int slot, off; | |
729 | ||
d9b55eeb | 730 | if (unlikely(!size || size > PCPU_MIN_UNIT_PAGES * PAGE_SIZE || |
fbf59bc9 TH |
731 | align > PAGE_SIZE)) { |
732 | WARN(true, "illegal size (%zu) or align (%zu) for " | |
733 | "percpu allocation\n", size, align); | |
734 | return NULL; | |
735 | } | |
736 | ||
737 | mutex_lock(&pcpu_mutex); | |
738 | ||
739 | /* allocate area */ | |
740 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { | |
741 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
742 | if (size > chunk->contig_hint) | |
743 | continue; | |
744 | off = pcpu_alloc_area(chunk, size, align); | |
745 | if (off >= 0) | |
746 | goto area_found; | |
747 | if (off != -ENOSPC) | |
748 | goto out_unlock; | |
749 | } | |
750 | } | |
751 | ||
752 | /* hmmm... no space left, create a new chunk */ | |
753 | chunk = alloc_pcpu_chunk(); | |
754 | if (!chunk) | |
755 | goto out_unlock; | |
756 | pcpu_chunk_relocate(chunk, -1); | |
757 | pcpu_chunk_addr_insert(chunk); | |
758 | ||
759 | off = pcpu_alloc_area(chunk, size, align); | |
760 | if (off < 0) | |
761 | goto out_unlock; | |
762 | ||
763 | area_found: | |
764 | /* populate, map and clear the area */ | |
765 | if (pcpu_populate_chunk(chunk, off, size)) { | |
766 | pcpu_free_area(chunk, off); | |
767 | goto out_unlock; | |
768 | } | |
769 | ||
770 | ptr = __addr_to_pcpu_ptr(chunk->vm->addr + off); | |
771 | out_unlock: | |
772 | mutex_unlock(&pcpu_mutex); | |
773 | return ptr; | |
774 | } | |
775 | EXPORT_SYMBOL_GPL(__alloc_percpu); | |
776 | ||
777 | static void pcpu_kill_chunk(struct pcpu_chunk *chunk) | |
778 | { | |
779 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false); | |
780 | list_del(&chunk->list); | |
781 | rb_erase(&chunk->rb_node, &pcpu_addr_root); | |
782 | free_pcpu_chunk(chunk); | |
783 | } | |
784 | ||
785 | /** | |
786 | * free_percpu - free percpu area | |
787 | * @ptr: pointer to area to free | |
788 | * | |
789 | * Free percpu area @ptr. Might sleep. | |
790 | */ | |
791 | void free_percpu(void *ptr) | |
792 | { | |
793 | void *addr = __pcpu_ptr_to_addr(ptr); | |
794 | struct pcpu_chunk *chunk; | |
795 | int off; | |
796 | ||
797 | if (!ptr) | |
798 | return; | |
799 | ||
800 | mutex_lock(&pcpu_mutex); | |
801 | ||
802 | chunk = pcpu_chunk_addr_search(addr); | |
803 | off = addr - chunk->vm->addr; | |
804 | ||
805 | pcpu_free_area(chunk, off); | |
806 | ||
807 | /* the chunk became fully free, kill one if there are other free ones */ | |
808 | if (chunk->free_size == pcpu_unit_size) { | |
809 | struct pcpu_chunk *pos; | |
810 | ||
811 | list_for_each_entry(pos, | |
812 | &pcpu_slot[pcpu_chunk_slot(chunk)], list) | |
813 | if (pos != chunk) { | |
814 | pcpu_kill_chunk(pos); | |
815 | break; | |
816 | } | |
817 | } | |
818 | ||
819 | mutex_unlock(&pcpu_mutex); | |
820 | } | |
821 | EXPORT_SYMBOL_GPL(free_percpu); | |
822 | ||
823 | /** | |
824 | * pcpu_setup_static - initialize kernel static percpu area | |
825 | * @populate_pte_fn: callback to allocate pagetable | |
826 | * @pages: num_possible_cpus() * PFN_UP(cpu_size) pages | |
cae3aeb8 | 827 | * @cpu_size: the size of static percpu area in bytes |
fbf59bc9 TH |
828 | * |
829 | * Initialize kernel static percpu area. The caller should allocate | |
830 | * all the necessary pages and pass them in @pages. | |
831 | * @populate_pte_fn() is called on each page to be used for percpu | |
832 | * mapping and is responsible for making sure all the necessary page | |
833 | * tables for the page is allocated. | |
834 | * | |
835 | * RETURNS: | |
836 | * The determined pcpu_unit_size which can be used to initialize | |
837 | * percpu access. | |
838 | */ | |
839 | size_t __init pcpu_setup_static(pcpu_populate_pte_fn_t populate_pte_fn, | |
840 | struct page **pages, size_t cpu_size) | |
841 | { | |
842 | static struct vm_struct static_vm; | |
843 | struct pcpu_chunk *static_chunk; | |
844 | int nr_cpu_pages = DIV_ROUND_UP(cpu_size, PAGE_SIZE); | |
845 | unsigned int cpu; | |
846 | int err, i; | |
847 | ||
d9b55eeb | 848 | pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_PAGES, PFN_UP(cpu_size)); |
fbf59bc9 TH |
849 | |
850 | pcpu_static_size = cpu_size; | |
d9b55eeb | 851 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
fbf59bc9 | 852 | pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size; |
fbf59bc9 | 853 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) |
cb83b42e | 854 | + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *); |
fbf59bc9 | 855 | |
d9b55eeb TH |
856 | /* |
857 | * Allocate chunk slots. The additional last slot is for | |
858 | * empty chunks. | |
859 | */ | |
860 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
fbf59bc9 TH |
861 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
862 | for (i = 0; i < pcpu_nr_slots; i++) | |
863 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
864 | ||
865 | /* init and register vm area */ | |
866 | static_vm.flags = VM_ALLOC; | |
867 | static_vm.size = pcpu_chunk_size; | |
c0c0a293 | 868 | vm_area_register_early(&static_vm, PAGE_SIZE); |
fbf59bc9 TH |
869 | |
870 | /* init static_chunk */ | |
871 | static_chunk = alloc_bootmem(pcpu_chunk_struct_size); | |
872 | INIT_LIST_HEAD(&static_chunk->list); | |
873 | static_chunk->vm = &static_vm; | |
874 | static_chunk->free_size = pcpu_unit_size - pcpu_static_size; | |
875 | static_chunk->contig_hint = static_chunk->free_size; | |
876 | ||
877 | /* assign pages and map them */ | |
878 | for_each_possible_cpu(cpu) { | |
879 | for (i = 0; i < nr_cpu_pages; i++) { | |
880 | *pcpu_chunk_pagep(static_chunk, cpu, i) = *pages++; | |
881 | populate_pte_fn(pcpu_chunk_addr(static_chunk, cpu, i)); | |
882 | } | |
883 | } | |
884 | ||
885 | err = pcpu_map(static_chunk, 0, nr_cpu_pages); | |
886 | if (err) | |
887 | panic("failed to setup static percpu area, err=%d\n", err); | |
888 | ||
889 | /* link static_chunk in */ | |
890 | pcpu_chunk_relocate(static_chunk, -1); | |
891 | pcpu_chunk_addr_insert(static_chunk); | |
892 | ||
893 | /* we're done */ | |
894 | pcpu_base_addr = (void *)pcpu_chunk_addr(static_chunk, 0, 0); | |
895 | return pcpu_unit_size; | |
896 | } |