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 | * | |
8d408b4b TH |
51 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
52 | * setup the first chunk containing the kernel static percpu area | |
fbf59bc9 TH |
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> | |
ccea34b5 | 65 | #include <linux/spinlock.h> |
fbf59bc9 | 66 | #include <linux/vmalloc.h> |
a56dbddf | 67 | #include <linux/workqueue.h> |
fbf59bc9 TH |
68 | |
69 | #include <asm/cacheflush.h> | |
70 | #include <asm/tlbflush.h> | |
71 | ||
fbf59bc9 TH |
72 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
73 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ | |
74 | ||
75 | struct pcpu_chunk { | |
76 | struct list_head list; /* linked to pcpu_slot lists */ | |
77 | struct rb_node rb_node; /* key is chunk->vm->addr */ | |
78 | int free_size; /* free bytes in the chunk */ | |
79 | int contig_hint; /* max contiguous size hint */ | |
80 | struct vm_struct *vm; /* mapped vmalloc region */ | |
81 | int map_used; /* # of map entries used */ | |
82 | int map_alloc; /* # of map entries allocated */ | |
83 | int *map; /* allocation map */ | |
8d408b4b | 84 | bool immutable; /* no [de]population allowed */ |
3e24aa58 TH |
85 | struct page **page; /* points to page array */ |
86 | struct page *page_ar[]; /* #cpus * UNIT_PAGES */ | |
fbf59bc9 TH |
87 | }; |
88 | ||
40150d37 TH |
89 | static int pcpu_unit_pages __read_mostly; |
90 | static int pcpu_unit_size __read_mostly; | |
91 | static int pcpu_chunk_size __read_mostly; | |
92 | static int pcpu_nr_slots __read_mostly; | |
93 | static size_t pcpu_chunk_struct_size __read_mostly; | |
fbf59bc9 TH |
94 | |
95 | /* the address of the first chunk which starts with the kernel static area */ | |
40150d37 | 96 | void *pcpu_base_addr __read_mostly; |
fbf59bc9 TH |
97 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
98 | ||
edcb4639 TH |
99 | /* optional reserved chunk, only accessible for reserved allocations */ |
100 | static struct pcpu_chunk *pcpu_reserved_chunk; | |
101 | /* offset limit of the reserved chunk */ | |
102 | static int pcpu_reserved_chunk_limit; | |
103 | ||
fbf59bc9 | 104 | /* |
ccea34b5 TH |
105 | * Synchronization rules. |
106 | * | |
107 | * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former | |
108 | * protects allocation/reclaim paths, chunks and chunk->page arrays. | |
109 | * The latter is a spinlock and protects the index data structures - | |
110 | * chunk slots, rbtree, chunks and area maps in chunks. | |
111 | * | |
112 | * During allocation, pcpu_alloc_mutex is kept locked all the time and | |
113 | * pcpu_lock is grabbed and released as necessary. All actual memory | |
114 | * allocations are done using GFP_KERNEL with pcpu_lock released. | |
115 | * | |
116 | * Free path accesses and alters only the index data structures, so it | |
117 | * can be safely called from atomic context. When memory needs to be | |
118 | * returned to the system, free path schedules reclaim_work which | |
119 | * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be | |
120 | * reclaimed, release both locks and frees the chunks. Note that it's | |
121 | * necessary to grab both locks to remove a chunk from circulation as | |
122 | * allocation path might be referencing the chunk with only | |
123 | * pcpu_alloc_mutex locked. | |
fbf59bc9 | 124 | */ |
ccea34b5 TH |
125 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ |
126 | static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ | |
fbf59bc9 | 127 | |
40150d37 | 128 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
fbf59bc9 TH |
129 | static struct rb_root pcpu_addr_root = RB_ROOT; /* chunks by address */ |
130 | ||
a56dbddf TH |
131 | /* reclaim work to release fully free chunks, scheduled from free path */ |
132 | static void pcpu_reclaim(struct work_struct *work); | |
133 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); | |
134 | ||
d9b55eeb | 135 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 136 | { |
cae3aeb8 | 137 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
138 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
139 | } | |
140 | ||
d9b55eeb TH |
141 | static int pcpu_size_to_slot(int size) |
142 | { | |
143 | if (size == pcpu_unit_size) | |
144 | return pcpu_nr_slots - 1; | |
145 | return __pcpu_size_to_slot(size); | |
146 | } | |
147 | ||
fbf59bc9 TH |
148 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
149 | { | |
150 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
151 | return 0; | |
152 | ||
153 | return pcpu_size_to_slot(chunk->free_size); | |
154 | } | |
155 | ||
156 | static int pcpu_page_idx(unsigned int cpu, int page_idx) | |
157 | { | |
d9b55eeb | 158 | return cpu * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
159 | } |
160 | ||
161 | static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk, | |
162 | unsigned int cpu, int page_idx) | |
163 | { | |
164 | return &chunk->page[pcpu_page_idx(cpu, page_idx)]; | |
165 | } | |
166 | ||
167 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, | |
168 | unsigned int cpu, int page_idx) | |
169 | { | |
170 | return (unsigned long)chunk->vm->addr + | |
171 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); | |
172 | } | |
173 | ||
174 | static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk, | |
175 | int page_idx) | |
176 | { | |
177 | return *pcpu_chunk_pagep(chunk, 0, page_idx) != NULL; | |
178 | } | |
179 | ||
180 | /** | |
1880d93b TH |
181 | * pcpu_mem_alloc - allocate memory |
182 | * @size: bytes to allocate | |
fbf59bc9 | 183 | * |
1880d93b TH |
184 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
185 | * kzalloc() is used; otherwise, vmalloc() is used. The returned | |
186 | * memory is always zeroed. | |
fbf59bc9 | 187 | * |
ccea34b5 TH |
188 | * CONTEXT: |
189 | * Does GFP_KERNEL allocation. | |
190 | * | |
fbf59bc9 | 191 | * RETURNS: |
1880d93b | 192 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 193 | */ |
1880d93b | 194 | static void *pcpu_mem_alloc(size_t size) |
fbf59bc9 | 195 | { |
1880d93b TH |
196 | if (size <= PAGE_SIZE) |
197 | return kzalloc(size, GFP_KERNEL); | |
198 | else { | |
199 | void *ptr = vmalloc(size); | |
200 | if (ptr) | |
201 | memset(ptr, 0, size); | |
202 | return ptr; | |
203 | } | |
204 | } | |
fbf59bc9 | 205 | |
1880d93b TH |
206 | /** |
207 | * pcpu_mem_free - free memory | |
208 | * @ptr: memory to free | |
209 | * @size: size of the area | |
210 | * | |
211 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). | |
212 | */ | |
213 | static void pcpu_mem_free(void *ptr, size_t size) | |
214 | { | |
fbf59bc9 | 215 | if (size <= PAGE_SIZE) |
1880d93b | 216 | kfree(ptr); |
fbf59bc9 | 217 | else |
1880d93b | 218 | vfree(ptr); |
fbf59bc9 TH |
219 | } |
220 | ||
221 | /** | |
222 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
223 | * @chunk: chunk of interest | |
224 | * @oslot: the previous slot it was on | |
225 | * | |
226 | * This function is called after an allocation or free changed @chunk. | |
227 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
228 | * moved to the slot. Note that the reserved chunk is never put on |
229 | * chunk slots. | |
ccea34b5 TH |
230 | * |
231 | * CONTEXT: | |
232 | * pcpu_lock. | |
fbf59bc9 TH |
233 | */ |
234 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
235 | { | |
236 | int nslot = pcpu_chunk_slot(chunk); | |
237 | ||
edcb4639 | 238 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
239 | if (oslot < nslot) |
240 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
241 | else | |
242 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
243 | } | |
244 | } | |
245 | ||
246 | static struct rb_node **pcpu_chunk_rb_search(void *addr, | |
247 | struct rb_node **parentp) | |
248 | { | |
249 | struct rb_node **p = &pcpu_addr_root.rb_node; | |
250 | struct rb_node *parent = NULL; | |
251 | struct pcpu_chunk *chunk; | |
252 | ||
253 | while (*p) { | |
254 | parent = *p; | |
255 | chunk = rb_entry(parent, struct pcpu_chunk, rb_node); | |
256 | ||
257 | if (addr < chunk->vm->addr) | |
258 | p = &(*p)->rb_left; | |
259 | else if (addr > chunk->vm->addr) | |
260 | p = &(*p)->rb_right; | |
261 | else | |
262 | break; | |
263 | } | |
264 | ||
265 | if (parentp) | |
266 | *parentp = parent; | |
267 | return p; | |
268 | } | |
269 | ||
270 | /** | |
271 | * pcpu_chunk_addr_search - search for chunk containing specified address | |
272 | * @addr: address to search for | |
273 | * | |
274 | * Look for chunk which might contain @addr. More specifically, it | |
275 | * searchs for the chunk with the highest start address which isn't | |
276 | * beyond @addr. | |
277 | * | |
ccea34b5 TH |
278 | * CONTEXT: |
279 | * pcpu_lock. | |
280 | * | |
fbf59bc9 TH |
281 | * RETURNS: |
282 | * The address of the found chunk. | |
283 | */ | |
284 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
285 | { | |
286 | struct rb_node *n, *parent; | |
287 | struct pcpu_chunk *chunk; | |
288 | ||
edcb4639 TH |
289 | /* is it in the reserved chunk? */ |
290 | if (pcpu_reserved_chunk) { | |
291 | void *start = pcpu_reserved_chunk->vm->addr; | |
292 | ||
293 | if (addr >= start && addr < start + pcpu_reserved_chunk_limit) | |
294 | return pcpu_reserved_chunk; | |
295 | } | |
296 | ||
297 | /* nah... search the regular ones */ | |
fbf59bc9 TH |
298 | n = *pcpu_chunk_rb_search(addr, &parent); |
299 | if (!n) { | |
300 | /* no exactly matching chunk, the parent is the closest */ | |
301 | n = parent; | |
302 | BUG_ON(!n); | |
303 | } | |
304 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); | |
305 | ||
306 | if (addr < chunk->vm->addr) { | |
307 | /* the parent was the next one, look for the previous one */ | |
308 | n = rb_prev(n); | |
309 | BUG_ON(!n); | |
310 | chunk = rb_entry(n, struct pcpu_chunk, rb_node); | |
311 | } | |
312 | ||
313 | return chunk; | |
314 | } | |
315 | ||
316 | /** | |
317 | * pcpu_chunk_addr_insert - insert chunk into address rb tree | |
318 | * @new: chunk to insert | |
319 | * | |
320 | * Insert @new into address rb tree. | |
ccea34b5 TH |
321 | * |
322 | * CONTEXT: | |
323 | * pcpu_lock. | |
fbf59bc9 TH |
324 | */ |
325 | static void pcpu_chunk_addr_insert(struct pcpu_chunk *new) | |
326 | { | |
327 | struct rb_node **p, *parent; | |
328 | ||
329 | p = pcpu_chunk_rb_search(new->vm->addr, &parent); | |
330 | BUG_ON(*p); | |
331 | rb_link_node(&new->rb_node, parent, p); | |
332 | rb_insert_color(&new->rb_node, &pcpu_addr_root); | |
333 | } | |
334 | ||
9f7dcf22 TH |
335 | /** |
336 | * pcpu_extend_area_map - extend area map for allocation | |
337 | * @chunk: target chunk | |
338 | * | |
339 | * Extend area map of @chunk so that it can accomodate an allocation. | |
340 | * A single allocation can split an area into three areas, so this | |
341 | * function makes sure that @chunk->map has at least two extra slots. | |
342 | * | |
ccea34b5 TH |
343 | * CONTEXT: |
344 | * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired | |
345 | * if area map is extended. | |
346 | * | |
9f7dcf22 TH |
347 | * RETURNS: |
348 | * 0 if noop, 1 if successfully extended, -errno on failure. | |
349 | */ | |
350 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) | |
351 | { | |
352 | int new_alloc; | |
353 | int *new; | |
354 | size_t size; | |
355 | ||
356 | /* has enough? */ | |
357 | if (chunk->map_alloc >= chunk->map_used + 2) | |
358 | return 0; | |
359 | ||
ccea34b5 TH |
360 | spin_unlock_irq(&pcpu_lock); |
361 | ||
9f7dcf22 TH |
362 | new_alloc = PCPU_DFL_MAP_ALLOC; |
363 | while (new_alloc < chunk->map_used + 2) | |
364 | new_alloc *= 2; | |
365 | ||
366 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); | |
ccea34b5 TH |
367 | if (!new) { |
368 | spin_lock_irq(&pcpu_lock); | |
9f7dcf22 | 369 | return -ENOMEM; |
ccea34b5 TH |
370 | } |
371 | ||
372 | /* | |
373 | * Acquire pcpu_lock and switch to new area map. Only free | |
374 | * could have happened inbetween, so map_used couldn't have | |
375 | * grown. | |
376 | */ | |
377 | spin_lock_irq(&pcpu_lock); | |
378 | BUG_ON(new_alloc < chunk->map_used + 2); | |
9f7dcf22 TH |
379 | |
380 | size = chunk->map_alloc * sizeof(chunk->map[0]); | |
381 | memcpy(new, chunk->map, size); | |
382 | ||
383 | /* | |
384 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is | |
385 | * one of the first chunks and still using static map. | |
386 | */ | |
387 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) | |
388 | pcpu_mem_free(chunk->map, size); | |
389 | ||
390 | chunk->map_alloc = new_alloc; | |
391 | chunk->map = new; | |
392 | return 0; | |
393 | } | |
394 | ||
fbf59bc9 TH |
395 | /** |
396 | * pcpu_split_block - split a map block | |
397 | * @chunk: chunk of interest | |
398 | * @i: index of map block to split | |
cae3aeb8 TH |
399 | * @head: head size in bytes (can be 0) |
400 | * @tail: tail size in bytes (can be 0) | |
fbf59bc9 TH |
401 | * |
402 | * Split the @i'th map block into two or three blocks. If @head is | |
403 | * non-zero, @head bytes block is inserted before block @i moving it | |
404 | * to @i+1 and reducing its size by @head bytes. | |
405 | * | |
406 | * If @tail is non-zero, the target block, which can be @i or @i+1 | |
407 | * depending on @head, is reduced by @tail bytes and @tail byte block | |
408 | * is inserted after the target block. | |
409 | * | |
9f7dcf22 | 410 | * @chunk->map must have enough free slots to accomodate the split. |
ccea34b5 TH |
411 | * |
412 | * CONTEXT: | |
413 | * pcpu_lock. | |
fbf59bc9 | 414 | */ |
9f7dcf22 TH |
415 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
416 | int head, int tail) | |
fbf59bc9 TH |
417 | { |
418 | int nr_extra = !!head + !!tail; | |
1880d93b | 419 | |
9f7dcf22 | 420 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
fbf59bc9 | 421 | |
9f7dcf22 | 422 | /* insert new subblocks */ |
fbf59bc9 TH |
423 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
424 | sizeof(chunk->map[0]) * (chunk->map_used - i)); | |
425 | chunk->map_used += nr_extra; | |
426 | ||
427 | if (head) { | |
428 | chunk->map[i + 1] = chunk->map[i] - head; | |
429 | chunk->map[i++] = head; | |
430 | } | |
431 | if (tail) { | |
432 | chunk->map[i++] -= tail; | |
433 | chunk->map[i] = tail; | |
434 | } | |
fbf59bc9 TH |
435 | } |
436 | ||
437 | /** | |
438 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
439 | * @chunk: chunk of interest | |
cae3aeb8 | 440 | * @size: wanted size in bytes |
fbf59bc9 TH |
441 | * @align: wanted align |
442 | * | |
443 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
444 | * Note that this function only allocates the offset. It doesn't | |
445 | * populate or map the area. | |
446 | * | |
9f7dcf22 TH |
447 | * @chunk->map must have at least two free slots. |
448 | * | |
ccea34b5 TH |
449 | * CONTEXT: |
450 | * pcpu_lock. | |
451 | * | |
fbf59bc9 | 452 | * RETURNS: |
9f7dcf22 TH |
453 | * Allocated offset in @chunk on success, -1 if no matching area is |
454 | * found. | |
fbf59bc9 TH |
455 | */ |
456 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) | |
457 | { | |
458 | int oslot = pcpu_chunk_slot(chunk); | |
459 | int max_contig = 0; | |
460 | int i, off; | |
461 | ||
fbf59bc9 TH |
462 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
463 | bool is_last = i + 1 == chunk->map_used; | |
464 | int head, tail; | |
465 | ||
466 | /* extra for alignment requirement */ | |
467 | head = ALIGN(off, align) - off; | |
468 | BUG_ON(i == 0 && head != 0); | |
469 | ||
470 | if (chunk->map[i] < 0) | |
471 | continue; | |
472 | if (chunk->map[i] < head + size) { | |
473 | max_contig = max(chunk->map[i], max_contig); | |
474 | continue; | |
475 | } | |
476 | ||
477 | /* | |
478 | * If head is small or the previous block is free, | |
479 | * merge'em. Note that 'small' is defined as smaller | |
480 | * than sizeof(int), which is very small but isn't too | |
481 | * uncommon for percpu allocations. | |
482 | */ | |
483 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { | |
484 | if (chunk->map[i - 1] > 0) | |
485 | chunk->map[i - 1] += head; | |
486 | else { | |
487 | chunk->map[i - 1] -= head; | |
488 | chunk->free_size -= head; | |
489 | } | |
490 | chunk->map[i] -= head; | |
491 | off += head; | |
492 | head = 0; | |
493 | } | |
494 | ||
495 | /* if tail is small, just keep it around */ | |
496 | tail = chunk->map[i] - head - size; | |
497 | if (tail < sizeof(int)) | |
498 | tail = 0; | |
499 | ||
500 | /* split if warranted */ | |
501 | if (head || tail) { | |
9f7dcf22 | 502 | pcpu_split_block(chunk, i, head, tail); |
fbf59bc9 TH |
503 | if (head) { |
504 | i++; | |
505 | off += head; | |
506 | max_contig = max(chunk->map[i - 1], max_contig); | |
507 | } | |
508 | if (tail) | |
509 | max_contig = max(chunk->map[i + 1], max_contig); | |
510 | } | |
511 | ||
512 | /* update hint and mark allocated */ | |
513 | if (is_last) | |
514 | chunk->contig_hint = max_contig; /* fully scanned */ | |
515 | else | |
516 | chunk->contig_hint = max(chunk->contig_hint, | |
517 | max_contig); | |
518 | ||
519 | chunk->free_size -= chunk->map[i]; | |
520 | chunk->map[i] = -chunk->map[i]; | |
521 | ||
522 | pcpu_chunk_relocate(chunk, oslot); | |
523 | return off; | |
524 | } | |
525 | ||
526 | chunk->contig_hint = max_contig; /* fully scanned */ | |
527 | pcpu_chunk_relocate(chunk, oslot); | |
528 | ||
9f7dcf22 TH |
529 | /* tell the upper layer that this chunk has no matching area */ |
530 | return -1; | |
fbf59bc9 TH |
531 | } |
532 | ||
533 | /** | |
534 | * pcpu_free_area - free area to a pcpu_chunk | |
535 | * @chunk: chunk of interest | |
536 | * @freeme: offset of area to free | |
537 | * | |
538 | * Free area starting from @freeme to @chunk. Note that this function | |
539 | * only modifies the allocation map. It doesn't depopulate or unmap | |
540 | * the area. | |
ccea34b5 TH |
541 | * |
542 | * CONTEXT: | |
543 | * pcpu_lock. | |
fbf59bc9 TH |
544 | */ |
545 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) | |
546 | { | |
547 | int oslot = pcpu_chunk_slot(chunk); | |
548 | int i, off; | |
549 | ||
550 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) | |
551 | if (off == freeme) | |
552 | break; | |
553 | BUG_ON(off != freeme); | |
554 | BUG_ON(chunk->map[i] > 0); | |
555 | ||
556 | chunk->map[i] = -chunk->map[i]; | |
557 | chunk->free_size += chunk->map[i]; | |
558 | ||
559 | /* merge with previous? */ | |
560 | if (i > 0 && chunk->map[i - 1] >= 0) { | |
561 | chunk->map[i - 1] += chunk->map[i]; | |
562 | chunk->map_used--; | |
563 | memmove(&chunk->map[i], &chunk->map[i + 1], | |
564 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
565 | i--; | |
566 | } | |
567 | /* merge with next? */ | |
568 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { | |
569 | chunk->map[i] += chunk->map[i + 1]; | |
570 | chunk->map_used--; | |
571 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], | |
572 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); | |
573 | } | |
574 | ||
575 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); | |
576 | pcpu_chunk_relocate(chunk, oslot); | |
577 | } | |
578 | ||
579 | /** | |
580 | * pcpu_unmap - unmap pages out of a pcpu_chunk | |
581 | * @chunk: chunk of interest | |
582 | * @page_start: page index of the first page to unmap | |
583 | * @page_end: page index of the last page to unmap + 1 | |
584 | * @flush: whether to flush cache and tlb or not | |
585 | * | |
586 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. | |
587 | * If @flush is true, vcache is flushed before unmapping and tlb | |
588 | * after. | |
589 | */ | |
590 | static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end, | |
591 | bool flush) | |
592 | { | |
593 | unsigned int last = num_possible_cpus() - 1; | |
594 | unsigned int cpu; | |
595 | ||
8d408b4b TH |
596 | /* unmap must not be done on immutable chunk */ |
597 | WARN_ON(chunk->immutable); | |
598 | ||
fbf59bc9 TH |
599 | /* |
600 | * Each flushing trial can be very expensive, issue flush on | |
601 | * the whole region at once rather than doing it for each cpu. | |
602 | * This could be an overkill but is more scalable. | |
603 | */ | |
604 | if (flush) | |
605 | flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start), | |
606 | pcpu_chunk_addr(chunk, last, page_end)); | |
607 | ||
608 | for_each_possible_cpu(cpu) | |
609 | unmap_kernel_range_noflush( | |
610 | pcpu_chunk_addr(chunk, cpu, page_start), | |
611 | (page_end - page_start) << PAGE_SHIFT); | |
612 | ||
613 | /* ditto as flush_cache_vunmap() */ | |
614 | if (flush) | |
615 | flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start), | |
616 | pcpu_chunk_addr(chunk, last, page_end)); | |
617 | } | |
618 | ||
619 | /** | |
620 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk | |
621 | * @chunk: chunk to depopulate | |
622 | * @off: offset to the area to depopulate | |
cae3aeb8 | 623 | * @size: size of the area to depopulate in bytes |
fbf59bc9 TH |
624 | * @flush: whether to flush cache and tlb or not |
625 | * | |
626 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) | |
627 | * from @chunk. If @flush is true, vcache is flushed before unmapping | |
628 | * and tlb after. | |
ccea34b5 TH |
629 | * |
630 | * CONTEXT: | |
631 | * pcpu_alloc_mutex. | |
fbf59bc9 | 632 | */ |
cae3aeb8 TH |
633 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size, |
634 | bool flush) | |
fbf59bc9 TH |
635 | { |
636 | int page_start = PFN_DOWN(off); | |
637 | int page_end = PFN_UP(off + size); | |
638 | int unmap_start = -1; | |
639 | int uninitialized_var(unmap_end); | |
640 | unsigned int cpu; | |
641 | int i; | |
642 | ||
643 | for (i = page_start; i < page_end; i++) { | |
644 | for_each_possible_cpu(cpu) { | |
645 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); | |
646 | ||
647 | if (!*pagep) | |
648 | continue; | |
649 | ||
650 | __free_page(*pagep); | |
651 | ||
652 | /* | |
653 | * If it's partial depopulation, it might get | |
654 | * populated or depopulated again. Mark the | |
655 | * page gone. | |
656 | */ | |
657 | *pagep = NULL; | |
658 | ||
659 | unmap_start = unmap_start < 0 ? i : unmap_start; | |
660 | unmap_end = i + 1; | |
661 | } | |
662 | } | |
663 | ||
664 | if (unmap_start >= 0) | |
665 | pcpu_unmap(chunk, unmap_start, unmap_end, flush); | |
666 | } | |
667 | ||
668 | /** | |
669 | * pcpu_map - map pages into a pcpu_chunk | |
670 | * @chunk: chunk of interest | |
671 | * @page_start: page index of the first page to map | |
672 | * @page_end: page index of the last page to map + 1 | |
673 | * | |
674 | * For each cpu, map pages [@page_start,@page_end) into @chunk. | |
675 | * vcache is flushed afterwards. | |
676 | */ | |
677 | static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end) | |
678 | { | |
679 | unsigned int last = num_possible_cpus() - 1; | |
680 | unsigned int cpu; | |
681 | int err; | |
682 | ||
8d408b4b TH |
683 | /* map must not be done on immutable chunk */ |
684 | WARN_ON(chunk->immutable); | |
685 | ||
fbf59bc9 TH |
686 | for_each_possible_cpu(cpu) { |
687 | err = map_kernel_range_noflush( | |
688 | pcpu_chunk_addr(chunk, cpu, page_start), | |
689 | (page_end - page_start) << PAGE_SHIFT, | |
690 | PAGE_KERNEL, | |
691 | pcpu_chunk_pagep(chunk, cpu, page_start)); | |
692 | if (err < 0) | |
693 | return err; | |
694 | } | |
695 | ||
696 | /* flush at once, please read comments in pcpu_unmap() */ | |
697 | flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start), | |
698 | pcpu_chunk_addr(chunk, last, page_end)); | |
699 | return 0; | |
700 | } | |
701 | ||
702 | /** | |
703 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk | |
704 | * @chunk: chunk of interest | |
705 | * @off: offset to the area to populate | |
cae3aeb8 | 706 | * @size: size of the area to populate in bytes |
fbf59bc9 TH |
707 | * |
708 | * For each cpu, populate and map pages [@page_start,@page_end) into | |
709 | * @chunk. The area is cleared on return. | |
ccea34b5 TH |
710 | * |
711 | * CONTEXT: | |
712 | * pcpu_alloc_mutex, does GFP_KERNEL allocation. | |
fbf59bc9 TH |
713 | */ |
714 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) | |
715 | { | |
716 | const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; | |
717 | int page_start = PFN_DOWN(off); | |
718 | int page_end = PFN_UP(off + size); | |
719 | int map_start = -1; | |
02d51fdf | 720 | int uninitialized_var(map_end); |
fbf59bc9 TH |
721 | unsigned int cpu; |
722 | int i; | |
723 | ||
724 | for (i = page_start; i < page_end; i++) { | |
725 | if (pcpu_chunk_page_occupied(chunk, i)) { | |
726 | if (map_start >= 0) { | |
727 | if (pcpu_map(chunk, map_start, map_end)) | |
728 | goto err; | |
729 | map_start = -1; | |
730 | } | |
731 | continue; | |
732 | } | |
733 | ||
734 | map_start = map_start < 0 ? i : map_start; | |
735 | map_end = i + 1; | |
736 | ||
737 | for_each_possible_cpu(cpu) { | |
738 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); | |
739 | ||
740 | *pagep = alloc_pages_node(cpu_to_node(cpu), | |
741 | alloc_mask, 0); | |
742 | if (!*pagep) | |
743 | goto err; | |
744 | } | |
745 | } | |
746 | ||
747 | if (map_start >= 0 && pcpu_map(chunk, map_start, map_end)) | |
748 | goto err; | |
749 | ||
750 | for_each_possible_cpu(cpu) | |
d9b55eeb | 751 | memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0, |
fbf59bc9 TH |
752 | size); |
753 | ||
754 | return 0; | |
755 | err: | |
756 | /* likely under heavy memory pressure, give memory back */ | |
757 | pcpu_depopulate_chunk(chunk, off, size, true); | |
758 | return -ENOMEM; | |
759 | } | |
760 | ||
761 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) | |
762 | { | |
763 | if (!chunk) | |
764 | return; | |
765 | if (chunk->vm) | |
766 | free_vm_area(chunk->vm); | |
1880d93b | 767 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
fbf59bc9 TH |
768 | kfree(chunk); |
769 | } | |
770 | ||
771 | static struct pcpu_chunk *alloc_pcpu_chunk(void) | |
772 | { | |
773 | struct pcpu_chunk *chunk; | |
774 | ||
775 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); | |
776 | if (!chunk) | |
777 | return NULL; | |
778 | ||
1880d93b | 779 | chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
fbf59bc9 TH |
780 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
781 | chunk->map[chunk->map_used++] = pcpu_unit_size; | |
3e24aa58 | 782 | chunk->page = chunk->page_ar; |
fbf59bc9 TH |
783 | |
784 | chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL); | |
785 | if (!chunk->vm) { | |
786 | free_pcpu_chunk(chunk); | |
787 | return NULL; | |
788 | } | |
789 | ||
790 | INIT_LIST_HEAD(&chunk->list); | |
791 | chunk->free_size = pcpu_unit_size; | |
792 | chunk->contig_hint = pcpu_unit_size; | |
793 | ||
794 | return chunk; | |
795 | } | |
796 | ||
797 | /** | |
edcb4639 | 798 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 799 | * @size: size of area to allocate in bytes |
fbf59bc9 | 800 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 801 | * @reserved: allocate from the reserved chunk if available |
fbf59bc9 | 802 | * |
ccea34b5 TH |
803 | * Allocate percpu area of @size bytes aligned at @align. |
804 | * | |
805 | * CONTEXT: | |
806 | * Does GFP_KERNEL allocation. | |
fbf59bc9 TH |
807 | * |
808 | * RETURNS: | |
809 | * Percpu pointer to the allocated area on success, NULL on failure. | |
810 | */ | |
edcb4639 | 811 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
fbf59bc9 | 812 | { |
fbf59bc9 TH |
813 | struct pcpu_chunk *chunk; |
814 | int slot, off; | |
815 | ||
8d408b4b | 816 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
fbf59bc9 TH |
817 | WARN(true, "illegal size (%zu) or align (%zu) for " |
818 | "percpu allocation\n", size, align); | |
819 | return NULL; | |
820 | } | |
821 | ||
ccea34b5 TH |
822 | mutex_lock(&pcpu_alloc_mutex); |
823 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 824 | |
edcb4639 TH |
825 | /* serve reserved allocations from the reserved chunk if available */ |
826 | if (reserved && pcpu_reserved_chunk) { | |
827 | chunk = pcpu_reserved_chunk; | |
9f7dcf22 TH |
828 | if (size > chunk->contig_hint || |
829 | pcpu_extend_area_map(chunk) < 0) | |
ccea34b5 | 830 | goto fail_unlock; |
edcb4639 TH |
831 | off = pcpu_alloc_area(chunk, size, align); |
832 | if (off >= 0) | |
833 | goto area_found; | |
ccea34b5 | 834 | goto fail_unlock; |
edcb4639 TH |
835 | } |
836 | ||
ccea34b5 | 837 | restart: |
edcb4639 | 838 | /* search through normal chunks */ |
fbf59bc9 TH |
839 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
840 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
841 | if (size > chunk->contig_hint) | |
842 | continue; | |
ccea34b5 TH |
843 | |
844 | switch (pcpu_extend_area_map(chunk)) { | |
845 | case 0: | |
846 | break; | |
847 | case 1: | |
848 | goto restart; /* pcpu_lock dropped, restart */ | |
849 | default: | |
850 | goto fail_unlock; | |
851 | } | |
852 | ||
fbf59bc9 TH |
853 | off = pcpu_alloc_area(chunk, size, align); |
854 | if (off >= 0) | |
855 | goto area_found; | |
fbf59bc9 TH |
856 | } |
857 | } | |
858 | ||
859 | /* hmmm... no space left, create a new chunk */ | |
ccea34b5 TH |
860 | spin_unlock_irq(&pcpu_lock); |
861 | ||
fbf59bc9 TH |
862 | chunk = alloc_pcpu_chunk(); |
863 | if (!chunk) | |
ccea34b5 TH |
864 | goto fail_unlock_mutex; |
865 | ||
866 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 TH |
867 | pcpu_chunk_relocate(chunk, -1); |
868 | pcpu_chunk_addr_insert(chunk); | |
ccea34b5 | 869 | goto restart; |
fbf59bc9 TH |
870 | |
871 | area_found: | |
ccea34b5 TH |
872 | spin_unlock_irq(&pcpu_lock); |
873 | ||
fbf59bc9 TH |
874 | /* populate, map and clear the area */ |
875 | if (pcpu_populate_chunk(chunk, off, size)) { | |
ccea34b5 | 876 | spin_lock_irq(&pcpu_lock); |
fbf59bc9 | 877 | pcpu_free_area(chunk, off); |
ccea34b5 | 878 | goto fail_unlock; |
fbf59bc9 TH |
879 | } |
880 | ||
ccea34b5 TH |
881 | mutex_unlock(&pcpu_alloc_mutex); |
882 | ||
883 | return __addr_to_pcpu_ptr(chunk->vm->addr + off); | |
884 | ||
885 | fail_unlock: | |
886 | spin_unlock_irq(&pcpu_lock); | |
887 | fail_unlock_mutex: | |
888 | mutex_unlock(&pcpu_alloc_mutex); | |
889 | return NULL; | |
fbf59bc9 | 890 | } |
edcb4639 TH |
891 | |
892 | /** | |
893 | * __alloc_percpu - allocate dynamic percpu area | |
894 | * @size: size of area to allocate in bytes | |
895 | * @align: alignment of area (max PAGE_SIZE) | |
896 | * | |
897 | * Allocate percpu area of @size bytes aligned at @align. Might | |
898 | * sleep. Might trigger writeouts. | |
899 | * | |
ccea34b5 TH |
900 | * CONTEXT: |
901 | * Does GFP_KERNEL allocation. | |
902 | * | |
edcb4639 TH |
903 | * RETURNS: |
904 | * Percpu pointer to the allocated area on success, NULL on failure. | |
905 | */ | |
906 | void *__alloc_percpu(size_t size, size_t align) | |
907 | { | |
908 | return pcpu_alloc(size, align, false); | |
909 | } | |
fbf59bc9 TH |
910 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
911 | ||
edcb4639 TH |
912 | /** |
913 | * __alloc_reserved_percpu - allocate reserved percpu area | |
914 | * @size: size of area to allocate in bytes | |
915 | * @align: alignment of area (max PAGE_SIZE) | |
916 | * | |
917 | * Allocate percpu area of @size bytes aligned at @align from reserved | |
918 | * percpu area if arch has set it up; otherwise, allocation is served | |
919 | * from the same dynamic area. Might sleep. Might trigger writeouts. | |
920 | * | |
ccea34b5 TH |
921 | * CONTEXT: |
922 | * Does GFP_KERNEL allocation. | |
923 | * | |
edcb4639 TH |
924 | * RETURNS: |
925 | * Percpu pointer to the allocated area on success, NULL on failure. | |
926 | */ | |
927 | void *__alloc_reserved_percpu(size_t size, size_t align) | |
928 | { | |
929 | return pcpu_alloc(size, align, true); | |
930 | } | |
931 | ||
a56dbddf TH |
932 | /** |
933 | * pcpu_reclaim - reclaim fully free chunks, workqueue function | |
934 | * @work: unused | |
935 | * | |
936 | * Reclaim all fully free chunks except for the first one. | |
ccea34b5 TH |
937 | * |
938 | * CONTEXT: | |
939 | * workqueue context. | |
a56dbddf TH |
940 | */ |
941 | static void pcpu_reclaim(struct work_struct *work) | |
fbf59bc9 | 942 | { |
a56dbddf TH |
943 | LIST_HEAD(todo); |
944 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; | |
945 | struct pcpu_chunk *chunk, *next; | |
946 | ||
ccea34b5 TH |
947 | mutex_lock(&pcpu_alloc_mutex); |
948 | spin_lock_irq(&pcpu_lock); | |
a56dbddf TH |
949 | |
950 | list_for_each_entry_safe(chunk, next, head, list) { | |
951 | WARN_ON(chunk->immutable); | |
952 | ||
953 | /* spare the first one */ | |
954 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) | |
955 | continue; | |
956 | ||
957 | rb_erase(&chunk->rb_node, &pcpu_addr_root); | |
958 | list_move(&chunk->list, &todo); | |
959 | } | |
960 | ||
ccea34b5 TH |
961 | spin_unlock_irq(&pcpu_lock); |
962 | mutex_unlock(&pcpu_alloc_mutex); | |
a56dbddf TH |
963 | |
964 | list_for_each_entry_safe(chunk, next, &todo, list) { | |
965 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false); | |
966 | free_pcpu_chunk(chunk); | |
967 | } | |
fbf59bc9 TH |
968 | } |
969 | ||
970 | /** | |
971 | * free_percpu - free percpu area | |
972 | * @ptr: pointer to area to free | |
973 | * | |
ccea34b5 TH |
974 | * Free percpu area @ptr. |
975 | * | |
976 | * CONTEXT: | |
977 | * Can be called from atomic context. | |
fbf59bc9 TH |
978 | */ |
979 | void free_percpu(void *ptr) | |
980 | { | |
981 | void *addr = __pcpu_ptr_to_addr(ptr); | |
982 | struct pcpu_chunk *chunk; | |
ccea34b5 | 983 | unsigned long flags; |
fbf59bc9 TH |
984 | int off; |
985 | ||
986 | if (!ptr) | |
987 | return; | |
988 | ||
ccea34b5 | 989 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
990 | |
991 | chunk = pcpu_chunk_addr_search(addr); | |
992 | off = addr - chunk->vm->addr; | |
993 | ||
994 | pcpu_free_area(chunk, off); | |
995 | ||
a56dbddf | 996 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
997 | if (chunk->free_size == pcpu_unit_size) { |
998 | struct pcpu_chunk *pos; | |
999 | ||
a56dbddf | 1000 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1001 | if (pos != chunk) { |
a56dbddf | 1002 | schedule_work(&pcpu_reclaim_work); |
fbf59bc9 TH |
1003 | break; |
1004 | } | |
1005 | } | |
1006 | ||
ccea34b5 | 1007 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1008 | } |
1009 | EXPORT_SYMBOL_GPL(free_percpu); | |
1010 | ||
1011 | /** | |
8d408b4b TH |
1012 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
1013 | * @get_page_fn: callback to fetch page pointer | |
1014 | * @static_size: the size of static percpu area in bytes | |
edcb4639 | 1015 | * @reserved_size: the size of reserved percpu area in bytes |
cafe8816 TH |
1016 | * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto |
1017 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto | |
8d408b4b TH |
1018 | * @base_addr: mapped address, NULL for auto |
1019 | * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary | |
1020 | * | |
1021 | * Initialize the first percpu chunk which contains the kernel static | |
1022 | * perpcu area. This function is to be called from arch percpu area | |
1023 | * setup path. The first two parameters are mandatory. The rest are | |
1024 | * optional. | |
1025 | * | |
1026 | * @get_page_fn() should return pointer to percpu page given cpu | |
1027 | * number and page number. It should at least return enough pages to | |
1028 | * cover the static area. The returned pages for static area should | |
1029 | * have been initialized with valid data. If @unit_size is specified, | |
1030 | * it can also return pages after the static area. NULL return | |
1031 | * indicates end of pages for the cpu. Note that @get_page_fn() must | |
1032 | * return the same number of pages for all cpus. | |
1033 | * | |
edcb4639 TH |
1034 | * @reserved_size, if non-zero, specifies the amount of bytes to |
1035 | * reserve after the static area in the first chunk. This reserves | |
1036 | * the first chunk such that it's available only through reserved | |
1037 | * percpu allocation. This is primarily used to serve module percpu | |
1038 | * static areas on architectures where the addressing model has | |
1039 | * limited offset range for symbol relocations to guarantee module | |
1040 | * percpu symbols fall inside the relocatable range. | |
1041 | * | |
cafe8816 TH |
1042 | * @unit_size, if non-negative, specifies unit size and must be |
1043 | * aligned to PAGE_SIZE and equal to or larger than @static_size + | |
edcb4639 | 1044 | * @reserved_size + @dyn_size. |
8d408b4b | 1045 | * |
cafe8816 TH |
1046 | * @dyn_size, if non-negative, limits the number of bytes available |
1047 | * for dynamic allocation in the first chunk. Specifying non-negative | |
1048 | * value make percpu leave alone the area beyond @static_size + | |
edcb4639 | 1049 | * @reserved_size + @dyn_size. |
8d408b4b TH |
1050 | * |
1051 | * Non-null @base_addr means that the caller already allocated virtual | |
1052 | * region for the first chunk and mapped it. percpu must not mess | |
1053 | * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL | |
1054 | * @populate_pte_fn doesn't make any sense. | |
1055 | * | |
1056 | * @populate_pte_fn is used to populate the pagetable. NULL means the | |
1057 | * caller already populated the pagetable. | |
fbf59bc9 | 1058 | * |
edcb4639 TH |
1059 | * If the first chunk ends up with both reserved and dynamic areas, it |
1060 | * is served by two chunks - one to serve the core static and reserved | |
1061 | * areas and the other for the dynamic area. They share the same vm | |
1062 | * and page map but uses different area allocation map to stay away | |
1063 | * from each other. The latter chunk is circulated in the chunk slots | |
1064 | * and available for dynamic allocation like any other chunks. | |
1065 | * | |
fbf59bc9 TH |
1066 | * RETURNS: |
1067 | * The determined pcpu_unit_size which can be used to initialize | |
1068 | * percpu access. | |
1069 | */ | |
8d408b4b | 1070 | size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn, |
edcb4639 | 1071 | size_t static_size, size_t reserved_size, |
cafe8816 TH |
1072 | ssize_t unit_size, ssize_t dyn_size, |
1073 | void *base_addr, | |
8d408b4b | 1074 | pcpu_populate_pte_fn_t populate_pte_fn) |
fbf59bc9 | 1075 | { |
2441d15c | 1076 | static struct vm_struct first_vm; |
edcb4639 TH |
1077 | static int smap[2], dmap[2]; |
1078 | struct pcpu_chunk *schunk, *dchunk = NULL; | |
fbf59bc9 | 1079 | unsigned int cpu; |
8d408b4b | 1080 | int nr_pages; |
fbf59bc9 TH |
1081 | int err, i; |
1082 | ||
8d408b4b | 1083 | /* santiy checks */ |
edcb4639 TH |
1084 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
1085 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); | |
8d408b4b | 1086 | BUG_ON(!static_size); |
cafe8816 | 1087 | if (unit_size >= 0) { |
edcb4639 | 1088 | BUG_ON(unit_size < static_size + reserved_size + |
cafe8816 TH |
1089 | (dyn_size >= 0 ? dyn_size : 0)); |
1090 | BUG_ON(unit_size & ~PAGE_MASK); | |
1091 | } else { | |
1092 | BUG_ON(dyn_size >= 0); | |
1093 | BUG_ON(base_addr); | |
1094 | } | |
8d408b4b | 1095 | BUG_ON(base_addr && populate_pte_fn); |
fbf59bc9 | 1096 | |
cafe8816 | 1097 | if (unit_size >= 0) |
8d408b4b TH |
1098 | pcpu_unit_pages = unit_size >> PAGE_SHIFT; |
1099 | else | |
1100 | pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT, | |
edcb4639 | 1101 | PFN_UP(static_size + reserved_size)); |
8d408b4b | 1102 | |
d9b55eeb | 1103 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
fbf59bc9 | 1104 | pcpu_chunk_size = num_possible_cpus() * pcpu_unit_size; |
fbf59bc9 | 1105 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) |
cb83b42e | 1106 | + num_possible_cpus() * pcpu_unit_pages * sizeof(struct page *); |
fbf59bc9 | 1107 | |
cafe8816 | 1108 | if (dyn_size < 0) |
edcb4639 | 1109 | dyn_size = pcpu_unit_size - static_size - reserved_size; |
cafe8816 | 1110 | |
d9b55eeb TH |
1111 | /* |
1112 | * Allocate chunk slots. The additional last slot is for | |
1113 | * empty chunks. | |
1114 | */ | |
1115 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
fbf59bc9 TH |
1116 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
1117 | for (i = 0; i < pcpu_nr_slots; i++) | |
1118 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1119 | ||
edcb4639 TH |
1120 | /* |
1121 | * Initialize static chunk. If reserved_size is zero, the | |
1122 | * static chunk covers static area + dynamic allocation area | |
1123 | * in the first chunk. If reserved_size is not zero, it | |
1124 | * covers static area + reserved area (mostly used for module | |
1125 | * static percpu allocation). | |
1126 | */ | |
2441d15c TH |
1127 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
1128 | INIT_LIST_HEAD(&schunk->list); | |
1129 | schunk->vm = &first_vm; | |
61ace7fa TH |
1130 | schunk->map = smap; |
1131 | schunk->map_alloc = ARRAY_SIZE(smap); | |
3e24aa58 | 1132 | schunk->page = schunk->page_ar; |
edcb4639 TH |
1133 | |
1134 | if (reserved_size) { | |
1135 | schunk->free_size = reserved_size; | |
1136 | pcpu_reserved_chunk = schunk; /* not for dynamic alloc */ | |
1137 | } else { | |
1138 | schunk->free_size = dyn_size; | |
1139 | dyn_size = 0; /* dynamic area covered */ | |
1140 | } | |
2441d15c | 1141 | schunk->contig_hint = schunk->free_size; |
fbf59bc9 | 1142 | |
61ace7fa TH |
1143 | schunk->map[schunk->map_used++] = -static_size; |
1144 | if (schunk->free_size) | |
1145 | schunk->map[schunk->map_used++] = schunk->free_size; | |
1146 | ||
edcb4639 TH |
1147 | pcpu_reserved_chunk_limit = static_size + schunk->free_size; |
1148 | ||
1149 | /* init dynamic chunk if necessary */ | |
1150 | if (dyn_size) { | |
1151 | dchunk = alloc_bootmem(sizeof(struct pcpu_chunk)); | |
1152 | INIT_LIST_HEAD(&dchunk->list); | |
1153 | dchunk->vm = &first_vm; | |
1154 | dchunk->map = dmap; | |
1155 | dchunk->map_alloc = ARRAY_SIZE(dmap); | |
1156 | dchunk->page = schunk->page_ar; /* share page map with schunk */ | |
1157 | ||
1158 | dchunk->contig_hint = dchunk->free_size = dyn_size; | |
1159 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; | |
1160 | dchunk->map[dchunk->map_used++] = dchunk->free_size; | |
1161 | } | |
1162 | ||
8d408b4b | 1163 | /* allocate vm address */ |
2441d15c TH |
1164 | first_vm.flags = VM_ALLOC; |
1165 | first_vm.size = pcpu_chunk_size; | |
8d408b4b TH |
1166 | |
1167 | if (!base_addr) | |
2441d15c | 1168 | vm_area_register_early(&first_vm, PAGE_SIZE); |
8d408b4b TH |
1169 | else { |
1170 | /* | |
1171 | * Pages already mapped. No need to remap into | |
edcb4639 TH |
1172 | * vmalloc area. In this case the first chunks can't |
1173 | * be mapped or unmapped by percpu and are marked | |
8d408b4b TH |
1174 | * immutable. |
1175 | */ | |
2441d15c TH |
1176 | first_vm.addr = base_addr; |
1177 | schunk->immutable = true; | |
edcb4639 TH |
1178 | if (dchunk) |
1179 | dchunk->immutable = true; | |
8d408b4b TH |
1180 | } |
1181 | ||
1182 | /* assign pages */ | |
1183 | nr_pages = -1; | |
fbf59bc9 | 1184 | for_each_possible_cpu(cpu) { |
8d408b4b TH |
1185 | for (i = 0; i < pcpu_unit_pages; i++) { |
1186 | struct page *page = get_page_fn(cpu, i); | |
1187 | ||
1188 | if (!page) | |
1189 | break; | |
2441d15c | 1190 | *pcpu_chunk_pagep(schunk, cpu, i) = page; |
fbf59bc9 | 1191 | } |
8d408b4b | 1192 | |
61ace7fa | 1193 | BUG_ON(i < PFN_UP(static_size)); |
8d408b4b TH |
1194 | |
1195 | if (nr_pages < 0) | |
1196 | nr_pages = i; | |
1197 | else | |
1198 | BUG_ON(nr_pages != i); | |
fbf59bc9 TH |
1199 | } |
1200 | ||
8d408b4b TH |
1201 | /* map them */ |
1202 | if (populate_pte_fn) { | |
1203 | for_each_possible_cpu(cpu) | |
1204 | for (i = 0; i < nr_pages; i++) | |
2441d15c | 1205 | populate_pte_fn(pcpu_chunk_addr(schunk, |
8d408b4b TH |
1206 | cpu, i)); |
1207 | ||
2441d15c | 1208 | err = pcpu_map(schunk, 0, nr_pages); |
8d408b4b TH |
1209 | if (err) |
1210 | panic("failed to setup static percpu area, err=%d\n", | |
1211 | err); | |
1212 | } | |
fbf59bc9 | 1213 | |
2441d15c | 1214 | /* link the first chunk in */ |
edcb4639 TH |
1215 | if (!dchunk) { |
1216 | pcpu_chunk_relocate(schunk, -1); | |
1217 | pcpu_chunk_addr_insert(schunk); | |
1218 | } else { | |
1219 | pcpu_chunk_relocate(dchunk, -1); | |
1220 | pcpu_chunk_addr_insert(dchunk); | |
1221 | } | |
fbf59bc9 TH |
1222 | |
1223 | /* we're done */ | |
2441d15c | 1224 | pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0); |
fbf59bc9 TH |
1225 | return pcpu_unit_size; |
1226 | } |