blob: 9531590e6b69e59985bf127e71e07edb6fae8939 [file] [log] [blame]
Tejun Heofbf59bc2009-02-20 16:29:08 +09001/*
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 *
Tejun Heo8d408b42009-02-24 11:57:21 +090051 * - use pcpu_setup_first_chunk() during percpu area initialization to
52 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090053 */
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
Tejun Heofbf59bc2009-02-20 16:29:08 +090070#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
71#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
72
73struct pcpu_chunk {
74 struct list_head list; /* linked to pcpu_slot lists */
75 struct rb_node rb_node; /* key is chunk->vm->addr */
76 int free_size; /* free bytes in the chunk */
77 int contig_hint; /* max contiguous size hint */
78 struct vm_struct *vm; /* mapped vmalloc region */
79 int map_used; /* # of map entries used */
80 int map_alloc; /* # of map entries allocated */
81 int *map; /* allocation map */
Tejun Heo8d408b42009-02-24 11:57:21 +090082 bool immutable; /* no [de]population allowed */
Tejun Heofbf59bc2009-02-20 16:29:08 +090083 struct page *page[]; /* #cpus * UNIT_PAGES */
84};
85
Tejun Heo40150d32009-02-24 12:32:28 +090086static int pcpu_unit_pages __read_mostly;
87static int pcpu_unit_size __read_mostly;
88static int pcpu_chunk_size __read_mostly;
89static int pcpu_nr_slots __read_mostly;
90static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +090091
92/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +090093void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +090094EXPORT_SYMBOL_GPL(pcpu_base_addr);
95
96/* the size of kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +090097static int pcpu_static_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +090098
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 */
113static DEFINE_MUTEX(pcpu_mutex);
114
Tejun Heo40150d32009-02-24 12:32:28 +0900115static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900116static struct rb_root pcpu_addr_root = RB_ROOT; /* chunks by address */
117
Tejun Heod9b55ee2009-02-24 11:57:21 +0900118static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900119{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900120 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900121 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
122}
123
Tejun Heod9b55ee2009-02-24 11:57:21 +0900124static 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
Tejun Heofbf59bc2009-02-20 16:29:08 +0900131static 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
139static int pcpu_page_idx(unsigned int cpu, int page_idx)
140{
Tejun Heod9b55ee2009-02-24 11:57:21 +0900141 return cpu * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900142}
143
144static 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
150static 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
157static 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)
Tejun Heocae3aeb2009-02-21 16:56:23 +0900166 * @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)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900168 *
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 */
176static 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 */
208static 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
220static 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 */
255static 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 */
284static 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
Tejun Heocae3aeb2009-02-21 16:56:23 +0900298 * @head: head size in bytes (can be 0)
299 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900300 *
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 */
312static 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
Tejun Heocae3aeb2009-02-21 16:56:23 +0900354 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900355 * @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 */
364static 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 */
473static 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 */
518static 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
Tejun Heo8d408b42009-02-24 11:57:21 +0900524 /* unmap must not be done on immutable chunk */
525 WARN_ON(chunk->immutable);
526
Tejun Heofbf59bc2009-02-20 16:29:08 +0900527 /*
528 * Each flushing trial can be very expensive, issue flush on
529 * the whole region at once rather than doing it for each cpu.
530 * This could be an overkill but is more scalable.
531 */
532 if (flush)
533 flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start),
534 pcpu_chunk_addr(chunk, last, page_end));
535
536 for_each_possible_cpu(cpu)
537 unmap_kernel_range_noflush(
538 pcpu_chunk_addr(chunk, cpu, page_start),
539 (page_end - page_start) << PAGE_SHIFT);
540
541 /* ditto as flush_cache_vunmap() */
542 if (flush)
543 flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start),
544 pcpu_chunk_addr(chunk, last, page_end));
545}
546
547/**
548 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
549 * @chunk: chunk to depopulate
550 * @off: offset to the area to depopulate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900551 * @size: size of the area to depopulate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900552 * @flush: whether to flush cache and tlb or not
553 *
554 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
555 * from @chunk. If @flush is true, vcache is flushed before unmapping
556 * and tlb after.
557 */
Tejun Heocae3aeb2009-02-21 16:56:23 +0900558static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size,
559 bool flush)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900560{
561 int page_start = PFN_DOWN(off);
562 int page_end = PFN_UP(off + size);
563 int unmap_start = -1;
564 int uninitialized_var(unmap_end);
565 unsigned int cpu;
566 int i;
567
568 for (i = page_start; i < page_end; i++) {
569 for_each_possible_cpu(cpu) {
570 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
571
572 if (!*pagep)
573 continue;
574
575 __free_page(*pagep);
576
577 /*
578 * If it's partial depopulation, it might get
579 * populated or depopulated again. Mark the
580 * page gone.
581 */
582 *pagep = NULL;
583
584 unmap_start = unmap_start < 0 ? i : unmap_start;
585 unmap_end = i + 1;
586 }
587 }
588
589 if (unmap_start >= 0)
590 pcpu_unmap(chunk, unmap_start, unmap_end, flush);
591}
592
593/**
594 * pcpu_map - map pages into a pcpu_chunk
595 * @chunk: chunk of interest
596 * @page_start: page index of the first page to map
597 * @page_end: page index of the last page to map + 1
598 *
599 * For each cpu, map pages [@page_start,@page_end) into @chunk.
600 * vcache is flushed afterwards.
601 */
602static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
603{
604 unsigned int last = num_possible_cpus() - 1;
605 unsigned int cpu;
606 int err;
607
Tejun Heo8d408b42009-02-24 11:57:21 +0900608 /* map must not be done on immutable chunk */
609 WARN_ON(chunk->immutable);
610
Tejun Heofbf59bc2009-02-20 16:29:08 +0900611 for_each_possible_cpu(cpu) {
612 err = map_kernel_range_noflush(
613 pcpu_chunk_addr(chunk, cpu, page_start),
614 (page_end - page_start) << PAGE_SHIFT,
615 PAGE_KERNEL,
616 pcpu_chunk_pagep(chunk, cpu, page_start));
617 if (err < 0)
618 return err;
619 }
620
621 /* flush at once, please read comments in pcpu_unmap() */
622 flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start),
623 pcpu_chunk_addr(chunk, last, page_end));
624 return 0;
625}
626
627/**
628 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
629 * @chunk: chunk of interest
630 * @off: offset to the area to populate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900631 * @size: size of the area to populate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900632 *
633 * For each cpu, populate and map pages [@page_start,@page_end) into
634 * @chunk. The area is cleared on return.
635 */
636static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
637{
638 const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
639 int page_start = PFN_DOWN(off);
640 int page_end = PFN_UP(off + size);
641 int map_start = -1;
Tejun Heo02d51fdf2009-03-01 15:42:36 +0900642 int uninitialized_var(map_end);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900643 unsigned int cpu;
644 int i;
645
646 for (i = page_start; i < page_end; i++) {
647 if (pcpu_chunk_page_occupied(chunk, i)) {
648 if (map_start >= 0) {
649 if (pcpu_map(chunk, map_start, map_end))
650 goto err;
651 map_start = -1;
652 }
653 continue;
654 }
655
656 map_start = map_start < 0 ? i : map_start;
657 map_end = i + 1;
658
659 for_each_possible_cpu(cpu) {
660 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
661
662 *pagep = alloc_pages_node(cpu_to_node(cpu),
663 alloc_mask, 0);
664 if (!*pagep)
665 goto err;
666 }
667 }
668
669 if (map_start >= 0 && pcpu_map(chunk, map_start, map_end))
670 goto err;
671
672 for_each_possible_cpu(cpu)
Tejun Heod9b55ee2009-02-24 11:57:21 +0900673 memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0,
Tejun Heofbf59bc2009-02-20 16:29:08 +0900674 size);
675
676 return 0;
677err:
678 /* likely under heavy memory pressure, give memory back */
679 pcpu_depopulate_chunk(chunk, off, size, true);
680 return -ENOMEM;
681}
682
683static void free_pcpu_chunk(struct pcpu_chunk *chunk)
684{
685 if (!chunk)
686 return;
687 if (chunk->vm)
688 free_vm_area(chunk->vm);
689 pcpu_realloc(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]), 0);
690 kfree(chunk);
691}
692
693static struct pcpu_chunk *alloc_pcpu_chunk(void)
694{
695 struct pcpu_chunk *chunk;
696
697 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
698 if (!chunk)
699 return NULL;
700
701 chunk->map = pcpu_realloc(NULL, 0,
702 PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
703 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
704 chunk->map[chunk->map_used++] = pcpu_unit_size;
705
706 chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL);
707 if (!chunk->vm) {
708 free_pcpu_chunk(chunk);
709 return NULL;
710 }
711
712 INIT_LIST_HEAD(&chunk->list);
713 chunk->free_size = pcpu_unit_size;
714 chunk->contig_hint = pcpu_unit_size;
715
716 return chunk;
717}
718
719/**
720 * __alloc_percpu - allocate percpu area
Tejun Heocae3aeb2009-02-21 16:56:23 +0900721 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900722 * @align: alignment of area (max PAGE_SIZE)
723 *
724 * Allocate percpu area of @size bytes aligned at @align. Might
725 * sleep. Might trigger writeouts.
726 *
727 * RETURNS:
728 * Percpu pointer to the allocated area on success, NULL on failure.
729 */
730void *__alloc_percpu(size_t size, size_t align)
731{
732 void *ptr = NULL;
733 struct pcpu_chunk *chunk;
734 int slot, off;
735
Tejun Heo8d408b42009-02-24 11:57:21 +0900736 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900737 WARN(true, "illegal size (%zu) or align (%zu) for "
738 "percpu allocation\n", size, align);
739 return NULL;
740 }
741
742 mutex_lock(&pcpu_mutex);
743
744 /* allocate area */
745 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
746 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
747 if (size > chunk->contig_hint)
748 continue;
749 off = pcpu_alloc_area(chunk, size, align);
750 if (off >= 0)
751 goto area_found;
752 if (off != -ENOSPC)
753 goto out_unlock;
754 }
755 }
756
757 /* hmmm... no space left, create a new chunk */
758 chunk = alloc_pcpu_chunk();
759 if (!chunk)
760 goto out_unlock;
761 pcpu_chunk_relocate(chunk, -1);
762 pcpu_chunk_addr_insert(chunk);
763
764 off = pcpu_alloc_area(chunk, size, align);
765 if (off < 0)
766 goto out_unlock;
767
768area_found:
769 /* populate, map and clear the area */
770 if (pcpu_populate_chunk(chunk, off, size)) {
771 pcpu_free_area(chunk, off);
772 goto out_unlock;
773 }
774
775 ptr = __addr_to_pcpu_ptr(chunk->vm->addr + off);
776out_unlock:
777 mutex_unlock(&pcpu_mutex);
778 return ptr;
779}
780EXPORT_SYMBOL_GPL(__alloc_percpu);
781
782static void pcpu_kill_chunk(struct pcpu_chunk *chunk)
783{
Tejun Heo8d408b42009-02-24 11:57:21 +0900784 WARN_ON(chunk->immutable);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900785 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false);
786 list_del(&chunk->list);
787 rb_erase(&chunk->rb_node, &pcpu_addr_root);
788 free_pcpu_chunk(chunk);
789}
790
791/**
792 * free_percpu - free percpu area
793 * @ptr: pointer to area to free
794 *
795 * Free percpu area @ptr. Might sleep.
796 */
797void free_percpu(void *ptr)
798{
799 void *addr = __pcpu_ptr_to_addr(ptr);
800 struct pcpu_chunk *chunk;
801 int off;
802
803 if (!ptr)
804 return;
805
806 mutex_lock(&pcpu_mutex);
807
808 chunk = pcpu_chunk_addr_search(addr);
809 off = addr - chunk->vm->addr;
810
811 pcpu_free_area(chunk, off);
812
813 /* the chunk became fully free, kill one if there are other free ones */
814 if (chunk->free_size == pcpu_unit_size) {
815 struct pcpu_chunk *pos;
816
817 list_for_each_entry(pos,
818 &pcpu_slot[pcpu_chunk_slot(chunk)], list)
819 if (pos != chunk) {
820 pcpu_kill_chunk(pos);
821 break;
822 }
823 }
824
825 mutex_unlock(&pcpu_mutex);
826}
827EXPORT_SYMBOL_GPL(free_percpu);
828
829/**
Tejun Heo8d408b42009-02-24 11:57:21 +0900830 * pcpu_setup_first_chunk - initialize the first percpu chunk
831 * @get_page_fn: callback to fetch page pointer
832 * @static_size: the size of static percpu area in bytes
833 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, 0 for auto
Tejun Heo2441d152009-03-06 14:33:59 +0900834 * @dyn_size: free size for dynamic allocation in bytes, 0 for auto
Tejun Heo8d408b42009-02-24 11:57:21 +0900835 * @base_addr: mapped address, NULL for auto
836 * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary
Tejun Heofbf59bc2009-02-20 16:29:08 +0900837 *
Tejun Heo8d408b42009-02-24 11:57:21 +0900838 * Initialize the first percpu chunk which contains the kernel static
839 * perpcu area. This function is to be called from arch percpu area
840 * setup path. The first two parameters are mandatory. The rest are
841 * optional.
842 *
843 * @get_page_fn() should return pointer to percpu page given cpu
844 * number and page number. It should at least return enough pages to
845 * cover the static area. The returned pages for static area should
846 * have been initialized with valid data. If @unit_size is specified,
847 * it can also return pages after the static area. NULL return
848 * indicates end of pages for the cpu. Note that @get_page_fn() must
849 * return the same number of pages for all cpus.
850 *
851 * @unit_size, if non-zero, determines unit size and must be aligned
Tejun Heo2441d152009-03-06 14:33:59 +0900852 * to PAGE_SIZE and equal to or larger than @static_size + @dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +0900853 *
Tejun Heo2441d152009-03-06 14:33:59 +0900854 * @dyn_size determines the number of free bytes after the static
Tejun Heo8d408b42009-02-24 11:57:21 +0900855 * area in the first chunk. If zero, whatever left is available.
856 * Specifying non-zero value make percpu leave the area after
Tejun Heo2441d152009-03-06 14:33:59 +0900857 * @static_size + @dyn_size alone.
Tejun Heo8d408b42009-02-24 11:57:21 +0900858 *
859 * Non-null @base_addr means that the caller already allocated virtual
860 * region for the first chunk and mapped it. percpu must not mess
861 * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL
862 * @populate_pte_fn doesn't make any sense.
863 *
864 * @populate_pte_fn is used to populate the pagetable. NULL means the
865 * caller already populated the pagetable.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900866 *
867 * RETURNS:
868 * The determined pcpu_unit_size which can be used to initialize
869 * percpu access.
870 */
Tejun Heo8d408b42009-02-24 11:57:21 +0900871size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn,
872 size_t static_size, size_t unit_size,
Tejun Heo2441d152009-03-06 14:33:59 +0900873 size_t dyn_size, void *base_addr,
Tejun Heo8d408b42009-02-24 11:57:21 +0900874 pcpu_populate_pte_fn_t populate_pte_fn)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900875{
Tejun Heo2441d152009-03-06 14:33:59 +0900876 static struct vm_struct first_vm;
877 struct pcpu_chunk *schunk;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900878 unsigned int cpu;
Tejun Heo8d408b42009-02-24 11:57:21 +0900879 int nr_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900880 int err, i;
881
Tejun Heo8d408b42009-02-24 11:57:21 +0900882 /* santiy checks */
883 BUG_ON(!static_size);
Tejun Heo2441d152009-03-06 14:33:59 +0900884 BUG_ON(!unit_size && dyn_size);
885 BUG_ON(unit_size && unit_size < static_size + dyn_size);
Tejun Heo8d408b42009-02-24 11:57:21 +0900886 BUG_ON(unit_size & ~PAGE_MASK);
887 BUG_ON(base_addr && !unit_size);
888 BUG_ON(base_addr && populate_pte_fn);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900889
Tejun Heo8d408b42009-02-24 11:57:21 +0900890 if (unit_size)
891 pcpu_unit_pages = unit_size >> PAGE_SHIFT;
892 else
893 pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT,
894 PFN_UP(static_size));
895
896 pcpu_static_size = static_size;
Tejun Heod9b55ee2009-02-24 11:57:21 +0900897 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900898 pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900899 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
Tejun Heocb83b422009-02-24 11:57:20 +0900900 + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900901
Tejun Heod9b55ee2009-02-24 11:57:21 +0900902 /*
903 * Allocate chunk slots. The additional last slot is for
904 * empty chunks.
905 */
906 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900907 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
908 for (i = 0; i < pcpu_nr_slots; i++)
909 INIT_LIST_HEAD(&pcpu_slot[i]);
910
Tejun Heo2441d152009-03-06 14:33:59 +0900911 /* init static chunk */
912 schunk = alloc_bootmem(pcpu_chunk_struct_size);
913 INIT_LIST_HEAD(&schunk->list);
914 schunk->vm = &first_vm;
Tejun Heo8d408b42009-02-24 11:57:21 +0900915
Tejun Heo2441d152009-03-06 14:33:59 +0900916 if (dyn_size)
917 schunk->free_size = dyn_size;
Tejun Heo8d408b42009-02-24 11:57:21 +0900918 else
Tejun Heo2441d152009-03-06 14:33:59 +0900919 schunk->free_size = pcpu_unit_size - pcpu_static_size;
Tejun Heo8d408b42009-02-24 11:57:21 +0900920
Tejun Heo2441d152009-03-06 14:33:59 +0900921 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900922
Tejun Heo8d408b42009-02-24 11:57:21 +0900923 /* allocate vm address */
Tejun Heo2441d152009-03-06 14:33:59 +0900924 first_vm.flags = VM_ALLOC;
925 first_vm.size = pcpu_chunk_size;
Tejun Heo8d408b42009-02-24 11:57:21 +0900926
927 if (!base_addr)
Tejun Heo2441d152009-03-06 14:33:59 +0900928 vm_area_register_early(&first_vm, PAGE_SIZE);
Tejun Heo8d408b42009-02-24 11:57:21 +0900929 else {
930 /*
931 * Pages already mapped. No need to remap into
932 * vmalloc area. In this case the static chunk can't
933 * be mapped or unmapped by percpu and is marked
934 * immutable.
935 */
Tejun Heo2441d152009-03-06 14:33:59 +0900936 first_vm.addr = base_addr;
937 schunk->immutable = true;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900938 }
939
Tejun Heo8d408b42009-02-24 11:57:21 +0900940 /* assign pages */
941 nr_pages = -1;
942 for_each_possible_cpu(cpu) {
943 for (i = 0; i < pcpu_unit_pages; i++) {
944 struct page *page = get_page_fn(cpu, i);
945
946 if (!page)
947 break;
Tejun Heo2441d152009-03-06 14:33:59 +0900948 *pcpu_chunk_pagep(schunk, cpu, i) = page;
Tejun Heo8d408b42009-02-24 11:57:21 +0900949 }
950
951 BUG_ON(i < PFN_UP(pcpu_static_size));
952
953 if (nr_pages < 0)
954 nr_pages = i;
955 else
956 BUG_ON(nr_pages != i);
957 }
958
959 /* map them */
960 if (populate_pte_fn) {
961 for_each_possible_cpu(cpu)
962 for (i = 0; i < nr_pages; i++)
Tejun Heo2441d152009-03-06 14:33:59 +0900963 populate_pte_fn(pcpu_chunk_addr(schunk,
Tejun Heo8d408b42009-02-24 11:57:21 +0900964 cpu, i));
965
Tejun Heo2441d152009-03-06 14:33:59 +0900966 err = pcpu_map(schunk, 0, nr_pages);
Tejun Heo8d408b42009-02-24 11:57:21 +0900967 if (err)
968 panic("failed to setup static percpu area, err=%d\n",
969 err);
970 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900971
Tejun Heo2441d152009-03-06 14:33:59 +0900972 /* link the first chunk in */
973 pcpu_chunk_relocate(schunk, -1);
974 pcpu_chunk_addr_insert(schunk);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900975
976 /* we're done */
Tejun Heo2441d152009-03-06 14:33:59 +0900977 pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900978 return pcpu_unit_size;
979}