blob: 1aa5d8fbca121d434f3cb812e10569f1b8e05f82 [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
Tejun Heoe0100982009-03-10 16:27:48 +090049 * regular address to percpu pointer and back if they need to be
50 * different from the default
Tejun Heofbf59bc2009-02-20 16:29:08 +090051 *
Tejun Heo8d408b42009-02-24 11:57:21 +090052 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
Tejun Heofbf59bc2009-02-20 16:29:08 +090054 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
58#include <linux/list.h>
59#include <linux/mm.h>
60#include <linux/module.h>
61#include <linux/mutex.h>
62#include <linux/percpu.h>
63#include <linux/pfn.h>
64#include <linux/rbtree.h>
65#include <linux/slab.h>
Tejun Heoccea34b2009-03-07 00:44:13 +090066#include <linux/spinlock.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090067#include <linux/vmalloc.h>
Tejun Heoa56dbdd2009-03-07 00:44:11 +090068#include <linux/workqueue.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090069
70#include <asm/cacheflush.h>
Tejun Heoe0100982009-03-10 16:27:48 +090071#include <asm/sections.h>
Tejun Heofbf59bc2009-02-20 16:29:08 +090072#include <asm/tlbflush.h>
73
Tejun Heofbf59bc2009-02-20 16:29:08 +090074#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
75#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
76
Tejun Heoe0100982009-03-10 16:27:48 +090077/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
78#ifndef __addr_to_pcpu_ptr
79#define __addr_to_pcpu_ptr(addr) \
80 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \
81 + (unsigned long)__per_cpu_start)
82#endif
83#ifndef __pcpu_ptr_to_addr
84#define __pcpu_ptr_to_addr(ptr) \
85 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \
86 - (unsigned long)__per_cpu_start)
87#endif
88
Tejun Heofbf59bc2009-02-20 16:29:08 +090089struct pcpu_chunk {
90 struct list_head list; /* linked to pcpu_slot lists */
91 struct rb_node rb_node; /* key is chunk->vm->addr */
92 int free_size; /* free bytes in the chunk */
93 int contig_hint; /* max contiguous size hint */
94 struct vm_struct *vm; /* mapped vmalloc region */
95 int map_used; /* # of map entries used */
96 int map_alloc; /* # of map entries allocated */
97 int *map; /* allocation map */
Tejun Heo8d408b42009-02-24 11:57:21 +090098 bool immutable; /* no [de]population allowed */
Tejun Heo3e24aa52009-03-06 14:33:59 +090099 struct page **page; /* points to page array */
100 struct page *page_ar[]; /* #cpus * UNIT_PAGES */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900101};
102
Tejun Heo40150d32009-02-24 12:32:28 +0900103static int pcpu_unit_pages __read_mostly;
104static int pcpu_unit_size __read_mostly;
105static int pcpu_chunk_size __read_mostly;
106static int pcpu_nr_slots __read_mostly;
107static size_t pcpu_chunk_struct_size __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900108
109/* the address of the first chunk which starts with the kernel static area */
Tejun Heo40150d32009-02-24 12:32:28 +0900110void *pcpu_base_addr __read_mostly;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900111EXPORT_SYMBOL_GPL(pcpu_base_addr);
112
Tejun Heoedcb4632009-03-06 14:33:59 +0900113/* optional reserved chunk, only accessible for reserved allocations */
114static struct pcpu_chunk *pcpu_reserved_chunk;
115/* offset limit of the reserved chunk */
116static int pcpu_reserved_chunk_limit;
117
Tejun Heofbf59bc2009-02-20 16:29:08 +0900118/*
Tejun Heoccea34b2009-03-07 00:44:13 +0900119 * Synchronization rules.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900120 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900121 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
122 * protects allocation/reclaim paths, chunks and chunk->page arrays.
123 * The latter is a spinlock and protects the index data structures -
124 * chunk slots, rbtree, chunks and area maps in chunks.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900125 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900126 * During allocation, pcpu_alloc_mutex is kept locked all the time and
127 * pcpu_lock is grabbed and released as necessary. All actual memory
128 * allocations are done using GFP_KERNEL with pcpu_lock released.
129 *
130 * Free path accesses and alters only the index data structures, so it
131 * can be safely called from atomic context. When memory needs to be
132 * returned to the system, free path schedules reclaim_work which
133 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
134 * reclaimed, release both locks and frees the chunks. Note that it's
135 * necessary to grab both locks to remove a chunk from circulation as
136 * allocation path might be referencing the chunk with only
137 * pcpu_alloc_mutex locked.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900138 */
Tejun Heoccea34b2009-03-07 00:44:13 +0900139static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
140static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900141
Tejun Heo40150d32009-02-24 12:32:28 +0900142static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900143static struct rb_root pcpu_addr_root = RB_ROOT; /* chunks by address */
144
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900145/* reclaim work to release fully free chunks, scheduled from free path */
146static void pcpu_reclaim(struct work_struct *work);
147static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
148
Tejun Heod9b55ee2009-02-24 11:57:21 +0900149static int __pcpu_size_to_slot(int size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900150{
Tejun Heocae3aeb2009-02-21 16:56:23 +0900151 int highbit = fls(size); /* size is in bytes */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900152 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
153}
154
Tejun Heod9b55ee2009-02-24 11:57:21 +0900155static int pcpu_size_to_slot(int size)
156{
157 if (size == pcpu_unit_size)
158 return pcpu_nr_slots - 1;
159 return __pcpu_size_to_slot(size);
160}
161
Tejun Heofbf59bc2009-02-20 16:29:08 +0900162static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
163{
164 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
165 return 0;
166
167 return pcpu_size_to_slot(chunk->free_size);
168}
169
170static int pcpu_page_idx(unsigned int cpu, int page_idx)
171{
Tejun Heod9b55ee2009-02-24 11:57:21 +0900172 return cpu * pcpu_unit_pages + page_idx;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900173}
174
175static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk,
176 unsigned int cpu, int page_idx)
177{
178 return &chunk->page[pcpu_page_idx(cpu, page_idx)];
179}
180
181static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
182 unsigned int cpu, int page_idx)
183{
184 return (unsigned long)chunk->vm->addr +
185 (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT);
186}
187
188static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk,
189 int page_idx)
190{
191 return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL;
192}
193
194/**
Tejun Heo1880d932009-03-07 00:44:09 +0900195 * pcpu_mem_alloc - allocate memory
196 * @size: bytes to allocate
Tejun Heofbf59bc2009-02-20 16:29:08 +0900197 *
Tejun Heo1880d932009-03-07 00:44:09 +0900198 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
199 * kzalloc() is used; otherwise, vmalloc() is used. The returned
200 * memory is always zeroed.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900201 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900202 * CONTEXT:
203 * Does GFP_KERNEL allocation.
204 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900205 * RETURNS:
Tejun Heo1880d932009-03-07 00:44:09 +0900206 * Pointer to the allocated area on success, NULL on failure.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900207 */
Tejun Heo1880d932009-03-07 00:44:09 +0900208static void *pcpu_mem_alloc(size_t size)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900209{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900210 if (size <= PAGE_SIZE)
Tejun Heo1880d932009-03-07 00:44:09 +0900211 return kzalloc(size, GFP_KERNEL);
212 else {
213 void *ptr = vmalloc(size);
214 if (ptr)
215 memset(ptr, 0, size);
216 return ptr;
217 }
218}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900219
Tejun Heo1880d932009-03-07 00:44:09 +0900220/**
221 * pcpu_mem_free - free memory
222 * @ptr: memory to free
223 * @size: size of the area
224 *
225 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
226 */
227static void pcpu_mem_free(void *ptr, size_t size)
228{
229 if (size <= PAGE_SIZE)
230 kfree(ptr);
231 else
232 vfree(ptr);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900233}
234
235/**
236 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
237 * @chunk: chunk of interest
238 * @oslot: the previous slot it was on
239 *
240 * This function is called after an allocation or free changed @chunk.
241 * New slot according to the changed state is determined and @chunk is
Tejun Heoedcb4632009-03-06 14:33:59 +0900242 * moved to the slot. Note that the reserved chunk is never put on
243 * chunk slots.
Tejun Heoccea34b2009-03-07 00:44:13 +0900244 *
245 * CONTEXT:
246 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900247 */
248static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
249{
250 int nslot = pcpu_chunk_slot(chunk);
251
Tejun Heoedcb4632009-03-06 14:33:59 +0900252 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900253 if (oslot < nslot)
254 list_move(&chunk->list, &pcpu_slot[nslot]);
255 else
256 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
257 }
258}
259
260static struct rb_node **pcpu_chunk_rb_search(void *addr,
261 struct rb_node **parentp)
262{
263 struct rb_node **p = &pcpu_addr_root.rb_node;
264 struct rb_node *parent = NULL;
265 struct pcpu_chunk *chunk;
266
267 while (*p) {
268 parent = *p;
269 chunk = rb_entry(parent, struct pcpu_chunk, rb_node);
270
271 if (addr < chunk->vm->addr)
272 p = &(*p)->rb_left;
273 else if (addr > chunk->vm->addr)
274 p = &(*p)->rb_right;
275 else
276 break;
277 }
278
279 if (parentp)
280 *parentp = parent;
281 return p;
282}
283
284/**
285 * pcpu_chunk_addr_search - search for chunk containing specified address
286 * @addr: address to search for
287 *
288 * Look for chunk which might contain @addr. More specifically, it
289 * searchs for the chunk with the highest start address which isn't
290 * beyond @addr.
291 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900292 * CONTEXT:
293 * pcpu_lock.
294 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900295 * RETURNS:
296 * The address of the found chunk.
297 */
298static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
299{
300 struct rb_node *n, *parent;
301 struct pcpu_chunk *chunk;
302
Tejun Heoedcb4632009-03-06 14:33:59 +0900303 /* is it in the reserved chunk? */
304 if (pcpu_reserved_chunk) {
305 void *start = pcpu_reserved_chunk->vm->addr;
306
307 if (addr >= start && addr < start + pcpu_reserved_chunk_limit)
308 return pcpu_reserved_chunk;
309 }
310
311 /* nah... search the regular ones */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900312 n = *pcpu_chunk_rb_search(addr, &parent);
313 if (!n) {
314 /* no exactly matching chunk, the parent is the closest */
315 n = parent;
316 BUG_ON(!n);
317 }
318 chunk = rb_entry(n, struct pcpu_chunk, rb_node);
319
320 if (addr < chunk->vm->addr) {
321 /* the parent was the next one, look for the previous one */
322 n = rb_prev(n);
323 BUG_ON(!n);
324 chunk = rb_entry(n, struct pcpu_chunk, rb_node);
325 }
326
327 return chunk;
328}
329
330/**
331 * pcpu_chunk_addr_insert - insert chunk into address rb tree
332 * @new: chunk to insert
333 *
334 * Insert @new into address rb tree.
Tejun Heoccea34b2009-03-07 00:44:13 +0900335 *
336 * CONTEXT:
337 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900338 */
339static void pcpu_chunk_addr_insert(struct pcpu_chunk *new)
340{
341 struct rb_node **p, *parent;
342
343 p = pcpu_chunk_rb_search(new->vm->addr, &parent);
344 BUG_ON(*p);
345 rb_link_node(&new->rb_node, parent, p);
346 rb_insert_color(&new->rb_node, &pcpu_addr_root);
347}
348
349/**
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900350 * pcpu_extend_area_map - extend area map for allocation
351 * @chunk: target chunk
352 *
353 * Extend area map of @chunk so that it can accomodate an allocation.
354 * A single allocation can split an area into three areas, so this
355 * function makes sure that @chunk->map has at least two extra slots.
356 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900357 * CONTEXT:
358 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
359 * if area map is extended.
360 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900361 * RETURNS:
362 * 0 if noop, 1 if successfully extended, -errno on failure.
363 */
364static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
365{
366 int new_alloc;
367 int *new;
368 size_t size;
369
370 /* has enough? */
371 if (chunk->map_alloc >= chunk->map_used + 2)
372 return 0;
373
Tejun Heoccea34b2009-03-07 00:44:13 +0900374 spin_unlock_irq(&pcpu_lock);
375
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900376 new_alloc = PCPU_DFL_MAP_ALLOC;
377 while (new_alloc < chunk->map_used + 2)
378 new_alloc *= 2;
379
380 new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
Tejun Heoccea34b2009-03-07 00:44:13 +0900381 if (!new) {
382 spin_lock_irq(&pcpu_lock);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900383 return -ENOMEM;
Tejun Heoccea34b2009-03-07 00:44:13 +0900384 }
385
386 /*
387 * Acquire pcpu_lock and switch to new area map. Only free
388 * could have happened inbetween, so map_used couldn't have
389 * grown.
390 */
391 spin_lock_irq(&pcpu_lock);
392 BUG_ON(new_alloc < chunk->map_used + 2);
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900393
394 size = chunk->map_alloc * sizeof(chunk->map[0]);
395 memcpy(new, chunk->map, size);
396
397 /*
398 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
399 * one of the first chunks and still using static map.
400 */
401 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
402 pcpu_mem_free(chunk->map, size);
403
404 chunk->map_alloc = new_alloc;
405 chunk->map = new;
406 return 0;
407}
408
409/**
Tejun Heofbf59bc2009-02-20 16:29:08 +0900410 * pcpu_split_block - split a map block
411 * @chunk: chunk of interest
412 * @i: index of map block to split
Tejun Heocae3aeb2009-02-21 16:56:23 +0900413 * @head: head size in bytes (can be 0)
414 * @tail: tail size in bytes (can be 0)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900415 *
416 * Split the @i'th map block into two or three blocks. If @head is
417 * non-zero, @head bytes block is inserted before block @i moving it
418 * to @i+1 and reducing its size by @head bytes.
419 *
420 * If @tail is non-zero, the target block, which can be @i or @i+1
421 * depending on @head, is reduced by @tail bytes and @tail byte block
422 * is inserted after the target block.
423 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900424 * @chunk->map must have enough free slots to accomodate the split.
Tejun Heoccea34b2009-03-07 00:44:13 +0900425 *
426 * CONTEXT:
427 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900428 */
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900429static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
430 int head, int tail)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900431{
432 int nr_extra = !!head + !!tail;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900433
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900434 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900435
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900436 /* insert new subblocks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900437 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
438 sizeof(chunk->map[0]) * (chunk->map_used - i));
439 chunk->map_used += nr_extra;
440
441 if (head) {
442 chunk->map[i + 1] = chunk->map[i] - head;
443 chunk->map[i++] = head;
444 }
445 if (tail) {
446 chunk->map[i++] -= tail;
447 chunk->map[i] = tail;
448 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900449}
450
451/**
452 * pcpu_alloc_area - allocate area from a pcpu_chunk
453 * @chunk: chunk of interest
Tejun Heocae3aeb2009-02-21 16:56:23 +0900454 * @size: wanted size in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900455 * @align: wanted align
456 *
457 * Try to allocate @size bytes area aligned at @align from @chunk.
458 * Note that this function only allocates the offset. It doesn't
459 * populate or map the area.
460 *
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900461 * @chunk->map must have at least two free slots.
462 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900463 * CONTEXT:
464 * pcpu_lock.
465 *
Tejun Heofbf59bc2009-02-20 16:29:08 +0900466 * RETURNS:
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900467 * Allocated offset in @chunk on success, -1 if no matching area is
468 * found.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900469 */
470static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
471{
472 int oslot = pcpu_chunk_slot(chunk);
473 int max_contig = 0;
474 int i, off;
475
Tejun Heofbf59bc2009-02-20 16:29:08 +0900476 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
477 bool is_last = i + 1 == chunk->map_used;
478 int head, tail;
479
480 /* extra for alignment requirement */
481 head = ALIGN(off, align) - off;
482 BUG_ON(i == 0 && head != 0);
483
484 if (chunk->map[i] < 0)
485 continue;
486 if (chunk->map[i] < head + size) {
487 max_contig = max(chunk->map[i], max_contig);
488 continue;
489 }
490
491 /*
492 * If head is small or the previous block is free,
493 * merge'em. Note that 'small' is defined as smaller
494 * than sizeof(int), which is very small but isn't too
495 * uncommon for percpu allocations.
496 */
497 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
498 if (chunk->map[i - 1] > 0)
499 chunk->map[i - 1] += head;
500 else {
501 chunk->map[i - 1] -= head;
502 chunk->free_size -= head;
503 }
504 chunk->map[i] -= head;
505 off += head;
506 head = 0;
507 }
508
509 /* if tail is small, just keep it around */
510 tail = chunk->map[i] - head - size;
511 if (tail < sizeof(int))
512 tail = 0;
513
514 /* split if warranted */
515 if (head || tail) {
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900516 pcpu_split_block(chunk, i, head, tail);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900517 if (head) {
518 i++;
519 off += head;
520 max_contig = max(chunk->map[i - 1], max_contig);
521 }
522 if (tail)
523 max_contig = max(chunk->map[i + 1], max_contig);
524 }
525
526 /* update hint and mark allocated */
527 if (is_last)
528 chunk->contig_hint = max_contig; /* fully scanned */
529 else
530 chunk->contig_hint = max(chunk->contig_hint,
531 max_contig);
532
533 chunk->free_size -= chunk->map[i];
534 chunk->map[i] = -chunk->map[i];
535
536 pcpu_chunk_relocate(chunk, oslot);
537 return off;
538 }
539
540 chunk->contig_hint = max_contig; /* fully scanned */
541 pcpu_chunk_relocate(chunk, oslot);
542
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900543 /* tell the upper layer that this chunk has no matching area */
544 return -1;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900545}
546
547/**
548 * pcpu_free_area - free area to a pcpu_chunk
549 * @chunk: chunk of interest
550 * @freeme: offset of area to free
551 *
552 * Free area starting from @freeme to @chunk. Note that this function
553 * only modifies the allocation map. It doesn't depopulate or unmap
554 * the area.
Tejun Heoccea34b2009-03-07 00:44:13 +0900555 *
556 * CONTEXT:
557 * pcpu_lock.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900558 */
559static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
560{
561 int oslot = pcpu_chunk_slot(chunk);
562 int i, off;
563
564 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
565 if (off == freeme)
566 break;
567 BUG_ON(off != freeme);
568 BUG_ON(chunk->map[i] > 0);
569
570 chunk->map[i] = -chunk->map[i];
571 chunk->free_size += chunk->map[i];
572
573 /* merge with previous? */
574 if (i > 0 && chunk->map[i - 1] >= 0) {
575 chunk->map[i - 1] += chunk->map[i];
576 chunk->map_used--;
577 memmove(&chunk->map[i], &chunk->map[i + 1],
578 (chunk->map_used - i) * sizeof(chunk->map[0]));
579 i--;
580 }
581 /* merge with next? */
582 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
583 chunk->map[i] += chunk->map[i + 1];
584 chunk->map_used--;
585 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
586 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
587 }
588
589 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
590 pcpu_chunk_relocate(chunk, oslot);
591}
592
593/**
594 * pcpu_unmap - unmap pages out of a pcpu_chunk
595 * @chunk: chunk of interest
596 * @page_start: page index of the first page to unmap
597 * @page_end: page index of the last page to unmap + 1
598 * @flush: whether to flush cache and tlb or not
599 *
600 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
601 * If @flush is true, vcache is flushed before unmapping and tlb
602 * after.
603 */
604static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end,
605 bool flush)
606{
607 unsigned int last = num_possible_cpus() - 1;
608 unsigned int cpu;
609
Tejun Heo8d408b42009-02-24 11:57:21 +0900610 /* unmap must not be done on immutable chunk */
611 WARN_ON(chunk->immutable);
612
Tejun Heofbf59bc2009-02-20 16:29:08 +0900613 /*
614 * Each flushing trial can be very expensive, issue flush on
615 * the whole region at once rather than doing it for each cpu.
616 * This could be an overkill but is more scalable.
617 */
618 if (flush)
619 flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start),
620 pcpu_chunk_addr(chunk, last, page_end));
621
622 for_each_possible_cpu(cpu)
623 unmap_kernel_range_noflush(
624 pcpu_chunk_addr(chunk, cpu, page_start),
625 (page_end - page_start) << PAGE_SHIFT);
626
627 /* ditto as flush_cache_vunmap() */
628 if (flush)
629 flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start),
630 pcpu_chunk_addr(chunk, last, page_end));
631}
632
633/**
634 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
635 * @chunk: chunk to depopulate
636 * @off: offset to the area to depopulate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900637 * @size: size of the area to depopulate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900638 * @flush: whether to flush cache and tlb or not
639 *
640 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
641 * from @chunk. If @flush is true, vcache is flushed before unmapping
642 * and tlb after.
Tejun Heoccea34b2009-03-07 00:44:13 +0900643 *
644 * CONTEXT:
645 * pcpu_alloc_mutex.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900646 */
Tejun Heocae3aeb2009-02-21 16:56:23 +0900647static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size,
648 bool flush)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900649{
650 int page_start = PFN_DOWN(off);
651 int page_end = PFN_UP(off + size);
652 int unmap_start = -1;
653 int uninitialized_var(unmap_end);
654 unsigned int cpu;
655 int i;
656
657 for (i = page_start; i < page_end; i++) {
658 for_each_possible_cpu(cpu) {
659 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
660
661 if (!*pagep)
662 continue;
663
664 __free_page(*pagep);
665
666 /*
667 * If it's partial depopulation, it might get
668 * populated or depopulated again. Mark the
669 * page gone.
670 */
671 *pagep = NULL;
672
673 unmap_start = unmap_start < 0 ? i : unmap_start;
674 unmap_end = i + 1;
675 }
676 }
677
678 if (unmap_start >= 0)
679 pcpu_unmap(chunk, unmap_start, unmap_end, flush);
680}
681
682/**
683 * pcpu_map - map pages into a pcpu_chunk
684 * @chunk: chunk of interest
685 * @page_start: page index of the first page to map
686 * @page_end: page index of the last page to map + 1
687 *
688 * For each cpu, map pages [@page_start,@page_end) into @chunk.
689 * vcache is flushed afterwards.
690 */
691static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end)
692{
693 unsigned int last = num_possible_cpus() - 1;
694 unsigned int cpu;
695 int err;
696
Tejun Heo8d408b42009-02-24 11:57:21 +0900697 /* map must not be done on immutable chunk */
698 WARN_ON(chunk->immutable);
699
Tejun Heofbf59bc2009-02-20 16:29:08 +0900700 for_each_possible_cpu(cpu) {
701 err = map_kernel_range_noflush(
702 pcpu_chunk_addr(chunk, cpu, page_start),
703 (page_end - page_start) << PAGE_SHIFT,
704 PAGE_KERNEL,
705 pcpu_chunk_pagep(chunk, cpu, page_start));
706 if (err < 0)
707 return err;
708 }
709
710 /* flush at once, please read comments in pcpu_unmap() */
711 flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start),
712 pcpu_chunk_addr(chunk, last, page_end));
713 return 0;
714}
715
716/**
717 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
718 * @chunk: chunk of interest
719 * @off: offset to the area to populate
Tejun Heocae3aeb2009-02-21 16:56:23 +0900720 * @size: size of the area to populate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900721 *
722 * For each cpu, populate and map pages [@page_start,@page_end) into
723 * @chunk. The area is cleared on return.
Tejun Heoccea34b2009-03-07 00:44:13 +0900724 *
725 * CONTEXT:
726 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900727 */
728static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
729{
730 const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
731 int page_start = PFN_DOWN(off);
732 int page_end = PFN_UP(off + size);
733 int map_start = -1;
Tejun Heo02d51fdf2009-03-01 15:42:36 +0900734 int uninitialized_var(map_end);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900735 unsigned int cpu;
736 int i;
737
738 for (i = page_start; i < page_end; i++) {
739 if (pcpu_chunk_page_occupied(chunk, i)) {
740 if (map_start >= 0) {
741 if (pcpu_map(chunk, map_start, map_end))
742 goto err;
743 map_start = -1;
744 }
745 continue;
746 }
747
748 map_start = map_start < 0 ? i : map_start;
749 map_end = i + 1;
750
751 for_each_possible_cpu(cpu) {
752 struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i);
753
754 *pagep = alloc_pages_node(cpu_to_node(cpu),
755 alloc_mask, 0);
756 if (!*pagep)
757 goto err;
758 }
759 }
760
761 if (map_start >= 0 && pcpu_map(chunk, map_start, map_end))
762 goto err;
763
764 for_each_possible_cpu(cpu)
Tejun Heod9b55ee2009-02-24 11:57:21 +0900765 memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0,
Tejun Heofbf59bc2009-02-20 16:29:08 +0900766 size);
767
768 return 0;
769err:
770 /* likely under heavy memory pressure, give memory back */
771 pcpu_depopulate_chunk(chunk, off, size, true);
772 return -ENOMEM;
773}
774
775static void free_pcpu_chunk(struct pcpu_chunk *chunk)
776{
777 if (!chunk)
778 return;
779 if (chunk->vm)
780 free_vm_area(chunk->vm);
Tejun Heo1880d932009-03-07 00:44:09 +0900781 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900782 kfree(chunk);
783}
784
785static struct pcpu_chunk *alloc_pcpu_chunk(void)
786{
787 struct pcpu_chunk *chunk;
788
789 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
790 if (!chunk)
791 return NULL;
792
Tejun Heo1880d932009-03-07 00:44:09 +0900793 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
Tejun Heofbf59bc2009-02-20 16:29:08 +0900794 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
795 chunk->map[chunk->map_used++] = pcpu_unit_size;
Tejun Heo3e24aa52009-03-06 14:33:59 +0900796 chunk->page = chunk->page_ar;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900797
798 chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL);
799 if (!chunk->vm) {
800 free_pcpu_chunk(chunk);
801 return NULL;
802 }
803
804 INIT_LIST_HEAD(&chunk->list);
805 chunk->free_size = pcpu_unit_size;
806 chunk->contig_hint = pcpu_unit_size;
807
808 return chunk;
809}
810
811/**
Tejun Heoedcb4632009-03-06 14:33:59 +0900812 * pcpu_alloc - the percpu allocator
Tejun Heocae3aeb2009-02-21 16:56:23 +0900813 * @size: size of area to allocate in bytes
Tejun Heofbf59bc2009-02-20 16:29:08 +0900814 * @align: alignment of area (max PAGE_SIZE)
Tejun Heoedcb4632009-03-06 14:33:59 +0900815 * @reserved: allocate from the reserved chunk if available
Tejun Heofbf59bc2009-02-20 16:29:08 +0900816 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900817 * Allocate percpu area of @size bytes aligned at @align.
818 *
819 * CONTEXT:
820 * Does GFP_KERNEL allocation.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900821 *
822 * RETURNS:
823 * Percpu pointer to the allocated area on success, NULL on failure.
824 */
Tejun Heoedcb4632009-03-06 14:33:59 +0900825static void *pcpu_alloc(size_t size, size_t align, bool reserved)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900826{
Tejun Heofbf59bc2009-02-20 16:29:08 +0900827 struct pcpu_chunk *chunk;
828 int slot, off;
829
Tejun Heo8d408b42009-02-24 11:57:21 +0900830 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
Tejun Heofbf59bc2009-02-20 16:29:08 +0900831 WARN(true, "illegal size (%zu) or align (%zu) for "
832 "percpu allocation\n", size, align);
833 return NULL;
834 }
835
Tejun Heoccea34b2009-03-07 00:44:13 +0900836 mutex_lock(&pcpu_alloc_mutex);
837 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900838
Tejun Heoedcb4632009-03-06 14:33:59 +0900839 /* serve reserved allocations from the reserved chunk if available */
840 if (reserved && pcpu_reserved_chunk) {
841 chunk = pcpu_reserved_chunk;
Tejun Heo9f7dcf22009-03-07 00:44:09 +0900842 if (size > chunk->contig_hint ||
843 pcpu_extend_area_map(chunk) < 0)
Tejun Heoccea34b2009-03-07 00:44:13 +0900844 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900845 off = pcpu_alloc_area(chunk, size, align);
846 if (off >= 0)
847 goto area_found;
Tejun Heoccea34b2009-03-07 00:44:13 +0900848 goto fail_unlock;
Tejun Heoedcb4632009-03-06 14:33:59 +0900849 }
850
Tejun Heoccea34b2009-03-07 00:44:13 +0900851restart:
Tejun Heoedcb4632009-03-06 14:33:59 +0900852 /* search through normal chunks */
Tejun Heofbf59bc2009-02-20 16:29:08 +0900853 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
854 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
855 if (size > chunk->contig_hint)
856 continue;
Tejun Heoccea34b2009-03-07 00:44:13 +0900857
858 switch (pcpu_extend_area_map(chunk)) {
859 case 0:
860 break;
861 case 1:
862 goto restart; /* pcpu_lock dropped, restart */
863 default:
864 goto fail_unlock;
865 }
866
Tejun Heofbf59bc2009-02-20 16:29:08 +0900867 off = pcpu_alloc_area(chunk, size, align);
868 if (off >= 0)
869 goto area_found;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900870 }
871 }
872
873 /* hmmm... no space left, create a new chunk */
Tejun Heoccea34b2009-03-07 00:44:13 +0900874 spin_unlock_irq(&pcpu_lock);
875
Tejun Heofbf59bc2009-02-20 16:29:08 +0900876 chunk = alloc_pcpu_chunk();
877 if (!chunk)
Tejun Heoccea34b2009-03-07 00:44:13 +0900878 goto fail_unlock_mutex;
879
880 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900881 pcpu_chunk_relocate(chunk, -1);
882 pcpu_chunk_addr_insert(chunk);
Tejun Heoccea34b2009-03-07 00:44:13 +0900883 goto restart;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900884
885area_found:
Tejun Heoccea34b2009-03-07 00:44:13 +0900886 spin_unlock_irq(&pcpu_lock);
887
Tejun Heofbf59bc2009-02-20 16:29:08 +0900888 /* populate, map and clear the area */
889 if (pcpu_populate_chunk(chunk, off, size)) {
Tejun Heoccea34b2009-03-07 00:44:13 +0900890 spin_lock_irq(&pcpu_lock);
Tejun Heofbf59bc2009-02-20 16:29:08 +0900891 pcpu_free_area(chunk, off);
Tejun Heoccea34b2009-03-07 00:44:13 +0900892 goto fail_unlock;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900893 }
894
Tejun Heoccea34b2009-03-07 00:44:13 +0900895 mutex_unlock(&pcpu_alloc_mutex);
896
897 return __addr_to_pcpu_ptr(chunk->vm->addr + off);
898
899fail_unlock:
900 spin_unlock_irq(&pcpu_lock);
901fail_unlock_mutex:
902 mutex_unlock(&pcpu_alloc_mutex);
903 return NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900904}
Tejun Heoedcb4632009-03-06 14:33:59 +0900905
906/**
907 * __alloc_percpu - allocate dynamic percpu area
908 * @size: size of area to allocate in bytes
909 * @align: alignment of area (max PAGE_SIZE)
910 *
911 * Allocate percpu area of @size bytes aligned at @align. Might
912 * sleep. Might trigger writeouts.
913 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900914 * CONTEXT:
915 * Does GFP_KERNEL allocation.
916 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900917 * RETURNS:
918 * Percpu pointer to the allocated area on success, NULL on failure.
919 */
920void *__alloc_percpu(size_t size, size_t align)
921{
922 return pcpu_alloc(size, align, false);
923}
Tejun Heofbf59bc2009-02-20 16:29:08 +0900924EXPORT_SYMBOL_GPL(__alloc_percpu);
925
Tejun Heoedcb4632009-03-06 14:33:59 +0900926/**
927 * __alloc_reserved_percpu - allocate reserved percpu area
928 * @size: size of area to allocate in bytes
929 * @align: alignment of area (max PAGE_SIZE)
930 *
931 * Allocate percpu area of @size bytes aligned at @align from reserved
932 * percpu area if arch has set it up; otherwise, allocation is served
933 * from the same dynamic area. Might sleep. Might trigger writeouts.
934 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900935 * CONTEXT:
936 * Does GFP_KERNEL allocation.
937 *
Tejun Heoedcb4632009-03-06 14:33:59 +0900938 * RETURNS:
939 * Percpu pointer to the allocated area on success, NULL on failure.
940 */
941void *__alloc_reserved_percpu(size_t size, size_t align)
942{
943 return pcpu_alloc(size, align, true);
944}
945
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900946/**
947 * pcpu_reclaim - reclaim fully free chunks, workqueue function
948 * @work: unused
949 *
950 * Reclaim all fully free chunks except for the first one.
Tejun Heoccea34b2009-03-07 00:44:13 +0900951 *
952 * CONTEXT:
953 * workqueue context.
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900954 */
955static void pcpu_reclaim(struct work_struct *work)
Tejun Heofbf59bc2009-02-20 16:29:08 +0900956{
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900957 LIST_HEAD(todo);
958 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
959 struct pcpu_chunk *chunk, *next;
960
Tejun Heoccea34b2009-03-07 00:44:13 +0900961 mutex_lock(&pcpu_alloc_mutex);
962 spin_lock_irq(&pcpu_lock);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900963
964 list_for_each_entry_safe(chunk, next, head, list) {
965 WARN_ON(chunk->immutable);
966
967 /* spare the first one */
968 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
969 continue;
970
971 rb_erase(&chunk->rb_node, &pcpu_addr_root);
972 list_move(&chunk->list, &todo);
973 }
974
Tejun Heoccea34b2009-03-07 00:44:13 +0900975 spin_unlock_irq(&pcpu_lock);
976 mutex_unlock(&pcpu_alloc_mutex);
Tejun Heoa56dbdd2009-03-07 00:44:11 +0900977
978 list_for_each_entry_safe(chunk, next, &todo, list) {
979 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false);
980 free_pcpu_chunk(chunk);
981 }
Tejun Heofbf59bc2009-02-20 16:29:08 +0900982}
983
984/**
985 * free_percpu - free percpu area
986 * @ptr: pointer to area to free
987 *
Tejun Heoccea34b2009-03-07 00:44:13 +0900988 * Free percpu area @ptr.
989 *
990 * CONTEXT:
991 * Can be called from atomic context.
Tejun Heofbf59bc2009-02-20 16:29:08 +0900992 */
993void free_percpu(void *ptr)
994{
995 void *addr = __pcpu_ptr_to_addr(ptr);
996 struct pcpu_chunk *chunk;
Tejun Heoccea34b2009-03-07 00:44:13 +0900997 unsigned long flags;
Tejun Heofbf59bc2009-02-20 16:29:08 +0900998 int off;
999
1000 if (!ptr)
1001 return;
1002
Tejun Heoccea34b2009-03-07 00:44:13 +09001003 spin_lock_irqsave(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001004
1005 chunk = pcpu_chunk_addr_search(addr);
1006 off = addr - chunk->vm->addr;
1007
1008 pcpu_free_area(chunk, off);
1009
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001010 /* if there are more than one fully free chunks, wake up grim reaper */
Tejun Heofbf59bc2009-02-20 16:29:08 +09001011 if (chunk->free_size == pcpu_unit_size) {
1012 struct pcpu_chunk *pos;
1013
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001014 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001015 if (pos != chunk) {
Tejun Heoa56dbdd2009-03-07 00:44:11 +09001016 schedule_work(&pcpu_reclaim_work);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001017 break;
1018 }
1019 }
1020
Tejun Heoccea34b2009-03-07 00:44:13 +09001021 spin_unlock_irqrestore(&pcpu_lock, flags);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001022}
1023EXPORT_SYMBOL_GPL(free_percpu);
1024
1025/**
Tejun Heo8d408b42009-02-24 11:57:21 +09001026 * pcpu_setup_first_chunk - initialize the first percpu chunk
1027 * @get_page_fn: callback to fetch page pointer
1028 * @static_size: the size of static percpu area in bytes
Tejun Heoedcb4632009-03-06 14:33:59 +09001029 * @reserved_size: the size of reserved percpu area in bytes
Tejun Heocafe8812009-03-06 14:33:59 +09001030 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
Tejun Heo6074d5b2009-03-10 16:27:48 +09001031 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto
Tejun Heo8d408b42009-02-24 11:57:21 +09001032 * @base_addr: mapped address, NULL for auto
1033 * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary
Tejun Heofbf59bc2009-02-20 16:29:08 +09001034 *
Tejun Heo8d408b42009-02-24 11:57:21 +09001035 * Initialize the first percpu chunk which contains the kernel static
1036 * perpcu area. This function is to be called from arch percpu area
1037 * setup path. The first two parameters are mandatory. The rest are
1038 * optional.
1039 *
1040 * @get_page_fn() should return pointer to percpu page given cpu
1041 * number and page number. It should at least return enough pages to
1042 * cover the static area. The returned pages for static area should
1043 * have been initialized with valid data. If @unit_size is specified,
1044 * it can also return pages after the static area. NULL return
1045 * indicates end of pages for the cpu. Note that @get_page_fn() must
1046 * return the same number of pages for all cpus.
1047 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001048 * @reserved_size, if non-zero, specifies the amount of bytes to
1049 * reserve after the static area in the first chunk. This reserves
1050 * the first chunk such that it's available only through reserved
1051 * percpu allocation. This is primarily used to serve module percpu
1052 * static areas on architectures where the addressing model has
1053 * limited offset range for symbol relocations to guarantee module
1054 * percpu symbols fall inside the relocatable range.
1055 *
Tejun Heo6074d5b2009-03-10 16:27:48 +09001056 * @dyn_size, if non-negative, determines the number of bytes
1057 * available for dynamic allocation in the first chunk. Specifying
1058 * non-negative value makes percpu leave alone the area beyond
1059 * @static_size + @reserved_size + @dyn_size.
1060 *
Tejun Heocafe8812009-03-06 14:33:59 +09001061 * @unit_size, if non-negative, specifies unit size and must be
1062 * aligned to PAGE_SIZE and equal to or larger than @static_size +
Tejun Heo6074d5b2009-03-10 16:27:48 +09001063 * @reserved_size + if non-negative, @dyn_size.
Tejun Heo8d408b42009-02-24 11:57:21 +09001064 *
1065 * Non-null @base_addr means that the caller already allocated virtual
1066 * region for the first chunk and mapped it. percpu must not mess
1067 * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL
1068 * @populate_pte_fn doesn't make any sense.
1069 *
1070 * @populate_pte_fn is used to populate the pagetable. NULL means the
1071 * caller already populated the pagetable.
Tejun Heofbf59bc2009-02-20 16:29:08 +09001072 *
Tejun Heoedcb4632009-03-06 14:33:59 +09001073 * If the first chunk ends up with both reserved and dynamic areas, it
1074 * is served by two chunks - one to serve the core static and reserved
1075 * areas and the other for the dynamic area. They share the same vm
1076 * and page map but uses different area allocation map to stay away
1077 * from each other. The latter chunk is circulated in the chunk slots
1078 * and available for dynamic allocation like any other chunks.
1079 *
Tejun Heofbf59bc2009-02-20 16:29:08 +09001080 * RETURNS:
1081 * The determined pcpu_unit_size which can be used to initialize
1082 * percpu access.
1083 */
Tejun Heo8d408b42009-02-24 11:57:21 +09001084size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn,
Tejun Heoedcb4632009-03-06 14:33:59 +09001085 size_t static_size, size_t reserved_size,
Tejun Heo6074d5b2009-03-10 16:27:48 +09001086 ssize_t dyn_size, ssize_t unit_size,
Tejun Heocafe8812009-03-06 14:33:59 +09001087 void *base_addr,
Tejun Heo8d408b42009-02-24 11:57:21 +09001088 pcpu_populate_pte_fn_t populate_pte_fn)
Tejun Heofbf59bc2009-02-20 16:29:08 +09001089{
Tejun Heo2441d152009-03-06 14:33:59 +09001090 static struct vm_struct first_vm;
Tejun Heoedcb4632009-03-06 14:33:59 +09001091 static int smap[2], dmap[2];
Tejun Heo6074d5b2009-03-10 16:27:48 +09001092 size_t size_sum = static_size + reserved_size +
1093 (dyn_size >= 0 ? dyn_size : 0);
Tejun Heoedcb4632009-03-06 14:33:59 +09001094 struct pcpu_chunk *schunk, *dchunk = NULL;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001095 unsigned int cpu;
Tejun Heo8d408b42009-02-24 11:57:21 +09001096 int nr_pages;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001097 int err, i;
1098
Tejun Heo8d408b42009-02-24 11:57:21 +09001099 /* santiy checks */
Tejun Heoedcb4632009-03-06 14:33:59 +09001100 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1101 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
Tejun Heo8d408b42009-02-24 11:57:21 +09001102 BUG_ON(!static_size);
Tejun Heocafe8812009-03-06 14:33:59 +09001103 if (unit_size >= 0) {
Tejun Heo6074d5b2009-03-10 16:27:48 +09001104 BUG_ON(unit_size < size_sum);
Tejun Heocafe8812009-03-06 14:33:59 +09001105 BUG_ON(unit_size & ~PAGE_MASK);
Tejun Heo6074d5b2009-03-10 16:27:48 +09001106 BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE);
1107 } else
Tejun Heocafe8812009-03-06 14:33:59 +09001108 BUG_ON(base_addr);
Tejun Heo8d408b42009-02-24 11:57:21 +09001109 BUG_ON(base_addr && populate_pte_fn);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001110
Tejun Heocafe8812009-03-06 14:33:59 +09001111 if (unit_size >= 0)
Tejun Heo8d408b42009-02-24 11:57:21 +09001112 pcpu_unit_pages = unit_size >> PAGE_SHIFT;
1113 else
1114 pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT,
Tejun Heo6074d5b2009-03-10 16:27:48 +09001115 PFN_UP(size_sum));
Tejun Heo8d408b42009-02-24 11:57:21 +09001116
Tejun Heod9b55ee2009-02-24 11:57:21 +09001117 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001118 pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001119 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk)
Tejun Heocb83b422009-02-24 11:57:20 +09001120 + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001121
Tejun Heocafe8812009-03-06 14:33:59 +09001122 if (dyn_size < 0)
Tejun Heoedcb4632009-03-06 14:33:59 +09001123 dyn_size = pcpu_unit_size - static_size - reserved_size;
Tejun Heocafe8812009-03-06 14:33:59 +09001124
Tejun Heod9b55ee2009-02-24 11:57:21 +09001125 /*
1126 * Allocate chunk slots. The additional last slot is for
1127 * empty chunks.
1128 */
1129 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001130 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1131 for (i = 0; i < pcpu_nr_slots; i++)
1132 INIT_LIST_HEAD(&pcpu_slot[i]);
1133
Tejun Heoedcb4632009-03-06 14:33:59 +09001134 /*
1135 * Initialize static chunk. If reserved_size is zero, the
1136 * static chunk covers static area + dynamic allocation area
1137 * in the first chunk. If reserved_size is not zero, it
1138 * covers static area + reserved area (mostly used for module
1139 * static percpu allocation).
1140 */
Tejun Heo2441d152009-03-06 14:33:59 +09001141 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1142 INIT_LIST_HEAD(&schunk->list);
1143 schunk->vm = &first_vm;
Tejun Heo61ace7f2009-03-06 14:33:59 +09001144 schunk->map = smap;
1145 schunk->map_alloc = ARRAY_SIZE(smap);
Tejun Heo3e24aa52009-03-06 14:33:59 +09001146 schunk->page = schunk->page_ar;
Tejun Heoedcb4632009-03-06 14:33:59 +09001147
1148 if (reserved_size) {
1149 schunk->free_size = reserved_size;
1150 pcpu_reserved_chunk = schunk; /* not for dynamic alloc */
1151 } else {
1152 schunk->free_size = dyn_size;
1153 dyn_size = 0; /* dynamic area covered */
1154 }
Tejun Heo2441d152009-03-06 14:33:59 +09001155 schunk->contig_hint = schunk->free_size;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001156
Tejun Heo61ace7f2009-03-06 14:33:59 +09001157 schunk->map[schunk->map_used++] = -static_size;
1158 if (schunk->free_size)
1159 schunk->map[schunk->map_used++] = schunk->free_size;
1160
Tejun Heoedcb4632009-03-06 14:33:59 +09001161 pcpu_reserved_chunk_limit = static_size + schunk->free_size;
1162
1163 /* init dynamic chunk if necessary */
1164 if (dyn_size) {
1165 dchunk = alloc_bootmem(sizeof(struct pcpu_chunk));
1166 INIT_LIST_HEAD(&dchunk->list);
1167 dchunk->vm = &first_vm;
1168 dchunk->map = dmap;
1169 dchunk->map_alloc = ARRAY_SIZE(dmap);
1170 dchunk->page = schunk->page_ar; /* share page map with schunk */
1171
1172 dchunk->contig_hint = dchunk->free_size = dyn_size;
1173 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1174 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1175 }
1176
Tejun Heo8d408b42009-02-24 11:57:21 +09001177 /* allocate vm address */
Tejun Heo2441d152009-03-06 14:33:59 +09001178 first_vm.flags = VM_ALLOC;
1179 first_vm.size = pcpu_chunk_size;
Tejun Heo8d408b42009-02-24 11:57:21 +09001180
1181 if (!base_addr)
Tejun Heo2441d152009-03-06 14:33:59 +09001182 vm_area_register_early(&first_vm, PAGE_SIZE);
Tejun Heo8d408b42009-02-24 11:57:21 +09001183 else {
1184 /*
1185 * Pages already mapped. No need to remap into
Tejun Heoedcb4632009-03-06 14:33:59 +09001186 * vmalloc area. In this case the first chunks can't
1187 * be mapped or unmapped by percpu and are marked
Tejun Heo8d408b42009-02-24 11:57:21 +09001188 * immutable.
1189 */
Tejun Heo2441d152009-03-06 14:33:59 +09001190 first_vm.addr = base_addr;
1191 schunk->immutable = true;
Tejun Heoedcb4632009-03-06 14:33:59 +09001192 if (dchunk)
1193 dchunk->immutable = true;
Tejun Heofbf59bc2009-02-20 16:29:08 +09001194 }
1195
Tejun Heo8d408b42009-02-24 11:57:21 +09001196 /* assign pages */
1197 nr_pages = -1;
1198 for_each_possible_cpu(cpu) {
1199 for (i = 0; i < pcpu_unit_pages; i++) {
1200 struct page *page = get_page_fn(cpu, i);
1201
1202 if (!page)
1203 break;
Tejun Heo2441d152009-03-06 14:33:59 +09001204 *pcpu_chunk_pagep(schunk, cpu, i) = page;
Tejun Heo8d408b42009-02-24 11:57:21 +09001205 }
1206
Tejun Heo61ace7f2009-03-06 14:33:59 +09001207 BUG_ON(i < PFN_UP(static_size));
Tejun Heo8d408b42009-02-24 11:57:21 +09001208
1209 if (nr_pages < 0)
1210 nr_pages = i;
1211 else
1212 BUG_ON(nr_pages != i);
1213 }
1214
1215 /* map them */
1216 if (populate_pte_fn) {
1217 for_each_possible_cpu(cpu)
1218 for (i = 0; i < nr_pages; i++)
Tejun Heo2441d152009-03-06 14:33:59 +09001219 populate_pte_fn(pcpu_chunk_addr(schunk,
Tejun Heo8d408b42009-02-24 11:57:21 +09001220 cpu, i));
1221
Tejun Heo2441d152009-03-06 14:33:59 +09001222 err = pcpu_map(schunk, 0, nr_pages);
Tejun Heo8d408b42009-02-24 11:57:21 +09001223 if (err)
1224 panic("failed to setup static percpu area, err=%d\n",
1225 err);
1226 }
Tejun Heofbf59bc2009-02-20 16:29:08 +09001227
Tejun Heo2441d152009-03-06 14:33:59 +09001228 /* link the first chunk in */
Tejun Heoedcb4632009-03-06 14:33:59 +09001229 if (!dchunk) {
1230 pcpu_chunk_relocate(schunk, -1);
1231 pcpu_chunk_addr_insert(schunk);
1232 } else {
1233 pcpu_chunk_relocate(dchunk, -1);
1234 pcpu_chunk_addr_insert(dchunk);
1235 }
Tejun Heofbf59bc2009-02-20 16:29:08 +09001236
1237 /* we're done */
Tejun Heo2441d152009-03-06 14:33:59 +09001238 pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0);
Tejun Heofbf59bc2009-02-20 16:29:08 +09001239 return pcpu_unit_size;
1240}
Tejun Heo66c3a752009-03-10 16:27:48 +09001241
1242/*
1243 * Embedding first chunk setup helper.
1244 */
1245static void *pcpue_ptr __initdata;
1246static size_t pcpue_size __initdata;
1247static size_t pcpue_unit_size __initdata;
1248
1249static struct page * __init pcpue_get_page(unsigned int cpu, int pageno)
1250{
1251 size_t off = (size_t)pageno << PAGE_SHIFT;
1252
1253 if (off >= pcpue_size)
1254 return NULL;
1255
1256 return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size + off);
1257}
1258
1259/**
1260 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1261 * @static_size: the size of static percpu area in bytes
1262 * @reserved_size: the size of reserved percpu area in bytes
1263 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1264 * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto
1265 *
1266 * This is a helper to ease setting up embedded first percpu chunk and
1267 * can be called where pcpu_setup_first_chunk() is expected.
1268 *
1269 * If this function is used to setup the first chunk, it is allocated
1270 * as a contiguous area using bootmem allocator and used as-is without
1271 * being mapped into vmalloc area. This enables the first chunk to
1272 * piggy back on the linear physical mapping which often uses larger
1273 * page size.
1274 *
1275 * When @dyn_size is positive, dynamic area might be larger than
1276 * specified to fill page alignment. Also, when @dyn_size is auto,
1277 * @dyn_size does not fill the whole first chunk but only what's
1278 * necessary for page alignment after static and reserved areas.
1279 *
1280 * If the needed size is smaller than the minimum or specified unit
1281 * size, the leftover is returned to the bootmem allocator.
1282 *
1283 * RETURNS:
1284 * The determined pcpu_unit_size which can be used to initialize
1285 * percpu access on success, -errno on failure.
1286 */
1287ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size,
1288 ssize_t dyn_size, ssize_t unit_size)
1289{
1290 unsigned int cpu;
1291
1292 /* determine parameters and allocate */
1293 pcpue_size = PFN_ALIGN(static_size + reserved_size +
1294 (dyn_size >= 0 ? dyn_size : 0));
1295 if (dyn_size != 0)
1296 dyn_size = pcpue_size - static_size - reserved_size;
1297
1298 if (unit_size >= 0) {
1299 BUG_ON(unit_size < pcpue_size);
1300 pcpue_unit_size = unit_size;
1301 } else
1302 pcpue_unit_size = max_t(size_t, pcpue_size, PCPU_MIN_UNIT_SIZE);
1303
1304 pcpue_ptr = __alloc_bootmem_nopanic(
1305 num_possible_cpus() * pcpue_unit_size,
1306 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
1307 if (!pcpue_ptr)
1308 return -ENOMEM;
1309
1310 /* return the leftover and copy */
1311 for_each_possible_cpu(cpu) {
1312 void *ptr = pcpue_ptr + cpu * pcpue_unit_size;
1313
1314 free_bootmem(__pa(ptr + pcpue_size),
1315 pcpue_unit_size - pcpue_size);
1316 memcpy(ptr, __per_cpu_load, static_size);
1317 }
1318
1319 /* we're ready, commit */
1320 pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n",
1321 pcpue_size >> PAGE_SHIFT, pcpue_ptr, static_size);
1322
1323 return pcpu_setup_first_chunk(pcpue_get_page, static_size,
1324 reserved_size, dyn_size,
1325 pcpue_unit_size, pcpue_ptr, NULL);
1326}