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 | |
74d46d6b TH |
11 | * chunk is consisted of nr_cpu_ids units and the first chunk is used |
12 | * for static percpu variables in the kernel image (special boot time | |
13 | * alloc/init handling necessary as these areas need to be brought up | |
14 | * before allocation services are running). Unit grows as necessary | |
15 | * and all units grow or shrink in unison. When a chunk is filled up, | |
16 | * another chunk is allocated. ie. in vmalloc area | |
fbf59bc9 TH |
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 | |
e1b9aa3f | 26 | * percpu base registers pcpu_unit_size apart. |
fbf59bc9 TH |
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. | |
e1b9aa3f CL |
41 | * Chunks can be determined from the address using the index field |
42 | * in the page struct. The index field contains a pointer to the chunk. | |
fbf59bc9 TH |
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 | |
e0100983 TH |
49 | * regular address to percpu pointer and back if they need to be |
50 | * different from the default | |
fbf59bc9 | 51 | * |
8d408b4b TH |
52 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
53 | * setup the first chunk containing the kernel static percpu area | |
fbf59bc9 TH |
54 | */ |
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> | |
fbf59bc9 | 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> | |
e0100983 | 70 | #include <asm/sections.h> |
fbf59bc9 TH |
71 | #include <asm/tlbflush.h> |
72 | ||
fbf59bc9 TH |
73 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
74 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ | |
75 | ||
e0100983 TH |
76 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
77 | #ifndef __addr_to_pcpu_ptr | |
78 | #define __addr_to_pcpu_ptr(addr) \ | |
79 | (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \ | |
80 | + (unsigned long)__per_cpu_start) | |
81 | #endif | |
82 | #ifndef __pcpu_ptr_to_addr | |
83 | #define __pcpu_ptr_to_addr(ptr) \ | |
84 | (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \ | |
85 | - (unsigned long)__per_cpu_start) | |
86 | #endif | |
87 | ||
fbf59bc9 TH |
88 | struct pcpu_chunk { |
89 | struct list_head list; /* linked to pcpu_slot lists */ | |
fbf59bc9 TH |
90 | int free_size; /* free bytes in the chunk */ |
91 | int contig_hint; /* max contiguous size hint */ | |
92 | struct vm_struct *vm; /* mapped vmalloc region */ | |
93 | int map_used; /* # of map entries used */ | |
94 | int map_alloc; /* # of map entries allocated */ | |
95 | int *map; /* allocation map */ | |
8d408b4b | 96 | bool immutable; /* no [de]population allowed */ |
3e24aa58 TH |
97 | struct page **page; /* points to page array */ |
98 | struct page *page_ar[]; /* #cpus * UNIT_PAGES */ | |
fbf59bc9 TH |
99 | }; |
100 | ||
40150d37 TH |
101 | static int pcpu_unit_pages __read_mostly; |
102 | static int pcpu_unit_size __read_mostly; | |
103 | static int pcpu_chunk_size __read_mostly; | |
104 | static int pcpu_nr_slots __read_mostly; | |
105 | static size_t pcpu_chunk_struct_size __read_mostly; | |
fbf59bc9 TH |
106 | |
107 | /* the address of the first chunk which starts with the kernel static area */ | |
40150d37 | 108 | void *pcpu_base_addr __read_mostly; |
fbf59bc9 TH |
109 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
110 | ||
ae9e6bc9 TH |
111 | /* |
112 | * The first chunk which always exists. Note that unlike other | |
113 | * chunks, this one can be allocated and mapped in several different | |
114 | * ways and thus often doesn't live in the vmalloc area. | |
115 | */ | |
116 | static struct pcpu_chunk *pcpu_first_chunk; | |
117 | ||
118 | /* | |
119 | * Optional reserved chunk. This chunk reserves part of the first | |
120 | * chunk and serves it for reserved allocations. The amount of | |
121 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved | |
122 | * area doesn't exist, the following variables contain NULL and 0 | |
123 | * respectively. | |
124 | */ | |
edcb4639 | 125 | static struct pcpu_chunk *pcpu_reserved_chunk; |
edcb4639 TH |
126 | static int pcpu_reserved_chunk_limit; |
127 | ||
fbf59bc9 | 128 | /* |
ccea34b5 TH |
129 | * Synchronization rules. |
130 | * | |
131 | * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former | |
132 | * protects allocation/reclaim paths, chunks and chunk->page arrays. | |
133 | * The latter is a spinlock and protects the index data structures - | |
e1b9aa3f | 134 | * chunk slots, chunks and area maps in chunks. |
ccea34b5 TH |
135 | * |
136 | * During allocation, pcpu_alloc_mutex is kept locked all the time and | |
137 | * pcpu_lock is grabbed and released as necessary. All actual memory | |
138 | * allocations are done using GFP_KERNEL with pcpu_lock released. | |
139 | * | |
140 | * Free path accesses and alters only the index data structures, so it | |
141 | * can be safely called from atomic context. When memory needs to be | |
142 | * returned to the system, free path schedules reclaim_work which | |
143 | * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be | |
144 | * reclaimed, release both locks and frees the chunks. Note that it's | |
145 | * necessary to grab both locks to remove a chunk from circulation as | |
146 | * allocation path might be referencing the chunk with only | |
147 | * pcpu_alloc_mutex locked. | |
fbf59bc9 | 148 | */ |
ccea34b5 TH |
149 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ |
150 | static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ | |
fbf59bc9 | 151 | |
40150d37 | 152 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
fbf59bc9 | 153 | |
a56dbddf TH |
154 | /* reclaim work to release fully free chunks, scheduled from free path */ |
155 | static void pcpu_reclaim(struct work_struct *work); | |
156 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); | |
157 | ||
d9b55eeb | 158 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 159 | { |
cae3aeb8 | 160 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
161 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
162 | } | |
163 | ||
d9b55eeb TH |
164 | static int pcpu_size_to_slot(int size) |
165 | { | |
166 | if (size == pcpu_unit_size) | |
167 | return pcpu_nr_slots - 1; | |
168 | return __pcpu_size_to_slot(size); | |
169 | } | |
170 | ||
fbf59bc9 TH |
171 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
172 | { | |
173 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
174 | return 0; | |
175 | ||
176 | return pcpu_size_to_slot(chunk->free_size); | |
177 | } | |
178 | ||
179 | static int pcpu_page_idx(unsigned int cpu, int page_idx) | |
180 | { | |
d9b55eeb | 181 | return cpu * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
182 | } |
183 | ||
184 | static struct page **pcpu_chunk_pagep(struct pcpu_chunk *chunk, | |
185 | unsigned int cpu, int page_idx) | |
186 | { | |
187 | return &chunk->page[pcpu_page_idx(cpu, page_idx)]; | |
188 | } | |
189 | ||
190 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, | |
191 | unsigned int cpu, int page_idx) | |
192 | { | |
193 | return (unsigned long)chunk->vm->addr + | |
194 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); | |
195 | } | |
196 | ||
197 | static bool pcpu_chunk_page_occupied(struct pcpu_chunk *chunk, | |
198 | int page_idx) | |
199 | { | |
04a13c7c TH |
200 | /* |
201 | * Any possible cpu id can be used here, so there's no need to | |
202 | * worry about preemption or cpu hotplug. | |
203 | */ | |
204 | return *pcpu_chunk_pagep(chunk, raw_smp_processor_id(), | |
205 | page_idx) != NULL; | |
fbf59bc9 TH |
206 | } |
207 | ||
e1b9aa3f CL |
208 | /* set the pointer to a chunk in a page struct */ |
209 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
210 | { | |
211 | page->index = (unsigned long)pcpu; | |
212 | } | |
213 | ||
214 | /* obtain pointer to a chunk from a page struct */ | |
215 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
216 | { | |
217 | return (struct pcpu_chunk *)page->index; | |
218 | } | |
219 | ||
fbf59bc9 | 220 | /** |
1880d93b TH |
221 | * pcpu_mem_alloc - allocate memory |
222 | * @size: bytes to allocate | |
fbf59bc9 | 223 | * |
1880d93b TH |
224 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
225 | * kzalloc() is used; otherwise, vmalloc() is used. The returned | |
226 | * memory is always zeroed. | |
fbf59bc9 | 227 | * |
ccea34b5 TH |
228 | * CONTEXT: |
229 | * Does GFP_KERNEL allocation. | |
230 | * | |
fbf59bc9 | 231 | * RETURNS: |
1880d93b | 232 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 233 | */ |
1880d93b | 234 | static void *pcpu_mem_alloc(size_t size) |
fbf59bc9 | 235 | { |
1880d93b TH |
236 | if (size <= PAGE_SIZE) |
237 | return kzalloc(size, GFP_KERNEL); | |
238 | else { | |
239 | void *ptr = vmalloc(size); | |
240 | if (ptr) | |
241 | memset(ptr, 0, size); | |
242 | return ptr; | |
243 | } | |
244 | } | |
fbf59bc9 | 245 | |
1880d93b TH |
246 | /** |
247 | * pcpu_mem_free - free memory | |
248 | * @ptr: memory to free | |
249 | * @size: size of the area | |
250 | * | |
251 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). | |
252 | */ | |
253 | static void pcpu_mem_free(void *ptr, size_t size) | |
254 | { | |
fbf59bc9 | 255 | if (size <= PAGE_SIZE) |
1880d93b | 256 | kfree(ptr); |
fbf59bc9 | 257 | else |
1880d93b | 258 | vfree(ptr); |
fbf59bc9 TH |
259 | } |
260 | ||
261 | /** | |
262 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
263 | * @chunk: chunk of interest | |
264 | * @oslot: the previous slot it was on | |
265 | * | |
266 | * This function is called after an allocation or free changed @chunk. | |
267 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
268 | * moved to the slot. Note that the reserved chunk is never put on |
269 | * chunk slots. | |
ccea34b5 TH |
270 | * |
271 | * CONTEXT: | |
272 | * pcpu_lock. | |
fbf59bc9 TH |
273 | */ |
274 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
275 | { | |
276 | int nslot = pcpu_chunk_slot(chunk); | |
277 | ||
edcb4639 | 278 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
279 | if (oslot < nslot) |
280 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
281 | else | |
282 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
283 | } | |
284 | } | |
285 | ||
fbf59bc9 | 286 | /** |
e1b9aa3f CL |
287 | * pcpu_chunk_addr_search - determine chunk containing specified address |
288 | * @addr: address for which the chunk needs to be determined. | |
ccea34b5 | 289 | * |
fbf59bc9 TH |
290 | * RETURNS: |
291 | * The address of the found chunk. | |
292 | */ | |
293 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
294 | { | |
ae9e6bc9 | 295 | void *first_start = pcpu_first_chunk->vm->addr; |
fbf59bc9 | 296 | |
ae9e6bc9 TH |
297 | /* is it in the first chunk? */ |
298 | if (addr >= first_start && addr < first_start + pcpu_chunk_size) { | |
299 | /* is it in the reserved area? */ | |
300 | if (addr < first_start + pcpu_reserved_chunk_limit) | |
edcb4639 | 301 | return pcpu_reserved_chunk; |
ae9e6bc9 | 302 | return pcpu_first_chunk; |
edcb4639 TH |
303 | } |
304 | ||
04a13c7c TH |
305 | /* |
306 | * The address is relative to unit0 which might be unused and | |
307 | * thus unmapped. Offset the address to the unit space of the | |
308 | * current processor before looking it up in the vmalloc | |
309 | * space. Note that any possible cpu id can be used here, so | |
310 | * there's no need to worry about preemption or cpu hotplug. | |
311 | */ | |
312 | addr += raw_smp_processor_id() * pcpu_unit_size; | |
e1b9aa3f | 313 | return pcpu_get_page_chunk(vmalloc_to_page(addr)); |
fbf59bc9 TH |
314 | } |
315 | ||
9f7dcf22 TH |
316 | /** |
317 | * pcpu_extend_area_map - extend area map for allocation | |
318 | * @chunk: target chunk | |
319 | * | |
320 | * Extend area map of @chunk so that it can accomodate an allocation. | |
321 | * A single allocation can split an area into three areas, so this | |
322 | * function makes sure that @chunk->map has at least two extra slots. | |
323 | * | |
ccea34b5 TH |
324 | * CONTEXT: |
325 | * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired | |
326 | * if area map is extended. | |
327 | * | |
9f7dcf22 TH |
328 | * RETURNS: |
329 | * 0 if noop, 1 if successfully extended, -errno on failure. | |
330 | */ | |
331 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) | |
332 | { | |
333 | int new_alloc; | |
334 | int *new; | |
335 | size_t size; | |
336 | ||
337 | /* has enough? */ | |
338 | if (chunk->map_alloc >= chunk->map_used + 2) | |
339 | return 0; | |
340 | ||
ccea34b5 TH |
341 | spin_unlock_irq(&pcpu_lock); |
342 | ||
9f7dcf22 TH |
343 | new_alloc = PCPU_DFL_MAP_ALLOC; |
344 | while (new_alloc < chunk->map_used + 2) | |
345 | new_alloc *= 2; | |
346 | ||
347 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); | |
ccea34b5 TH |
348 | if (!new) { |
349 | spin_lock_irq(&pcpu_lock); | |
9f7dcf22 | 350 | return -ENOMEM; |
ccea34b5 TH |
351 | } |
352 | ||
353 | /* | |
354 | * Acquire pcpu_lock and switch to new area map. Only free | |
355 | * could have happened inbetween, so map_used couldn't have | |
356 | * grown. | |
357 | */ | |
358 | spin_lock_irq(&pcpu_lock); | |
359 | BUG_ON(new_alloc < chunk->map_used + 2); | |
9f7dcf22 TH |
360 | |
361 | size = chunk->map_alloc * sizeof(chunk->map[0]); | |
362 | memcpy(new, chunk->map, size); | |
363 | ||
364 | /* | |
365 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is | |
366 | * one of the first chunks and still using static map. | |
367 | */ | |
368 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) | |
369 | pcpu_mem_free(chunk->map, size); | |
370 | ||
371 | chunk->map_alloc = new_alloc; | |
372 | chunk->map = new; | |
373 | return 0; | |
374 | } | |
375 | ||
fbf59bc9 TH |
376 | /** |
377 | * pcpu_split_block - split a map block | |
378 | * @chunk: chunk of interest | |
379 | * @i: index of map block to split | |
cae3aeb8 TH |
380 | * @head: head size in bytes (can be 0) |
381 | * @tail: tail size in bytes (can be 0) | |
fbf59bc9 TH |
382 | * |
383 | * Split the @i'th map block into two or three blocks. If @head is | |
384 | * non-zero, @head bytes block is inserted before block @i moving it | |
385 | * to @i+1 and reducing its size by @head bytes. | |
386 | * | |
387 | * If @tail is non-zero, the target block, which can be @i or @i+1 | |
388 | * depending on @head, is reduced by @tail bytes and @tail byte block | |
389 | * is inserted after the target block. | |
390 | * | |
9f7dcf22 | 391 | * @chunk->map must have enough free slots to accomodate the split. |
ccea34b5 TH |
392 | * |
393 | * CONTEXT: | |
394 | * pcpu_lock. | |
fbf59bc9 | 395 | */ |
9f7dcf22 TH |
396 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
397 | int head, int tail) | |
fbf59bc9 TH |
398 | { |
399 | int nr_extra = !!head + !!tail; | |
1880d93b | 400 | |
9f7dcf22 | 401 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
fbf59bc9 | 402 | |
9f7dcf22 | 403 | /* insert new subblocks */ |
fbf59bc9 TH |
404 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
405 | sizeof(chunk->map[0]) * (chunk->map_used - i)); | |
406 | chunk->map_used += nr_extra; | |
407 | ||
408 | if (head) { | |
409 | chunk->map[i + 1] = chunk->map[i] - head; | |
410 | chunk->map[i++] = head; | |
411 | } | |
412 | if (tail) { | |
413 | chunk->map[i++] -= tail; | |
414 | chunk->map[i] = tail; | |
415 | } | |
fbf59bc9 TH |
416 | } |
417 | ||
418 | /** | |
419 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
420 | * @chunk: chunk of interest | |
cae3aeb8 | 421 | * @size: wanted size in bytes |
fbf59bc9 TH |
422 | * @align: wanted align |
423 | * | |
424 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
425 | * Note that this function only allocates the offset. It doesn't | |
426 | * populate or map the area. | |
427 | * | |
9f7dcf22 TH |
428 | * @chunk->map must have at least two free slots. |
429 | * | |
ccea34b5 TH |
430 | * CONTEXT: |
431 | * pcpu_lock. | |
432 | * | |
fbf59bc9 | 433 | * RETURNS: |
9f7dcf22 TH |
434 | * Allocated offset in @chunk on success, -1 if no matching area is |
435 | * found. | |
fbf59bc9 TH |
436 | */ |
437 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) | |
438 | { | |
439 | int oslot = pcpu_chunk_slot(chunk); | |
440 | int max_contig = 0; | |
441 | int i, off; | |
442 | ||
fbf59bc9 TH |
443 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
444 | bool is_last = i + 1 == chunk->map_used; | |
445 | int head, tail; | |
446 | ||
447 | /* extra for alignment requirement */ | |
448 | head = ALIGN(off, align) - off; | |
449 | BUG_ON(i == 0 && head != 0); | |
450 | ||
451 | if (chunk->map[i] < 0) | |
452 | continue; | |
453 | if (chunk->map[i] < head + size) { | |
454 | max_contig = max(chunk->map[i], max_contig); | |
455 | continue; | |
456 | } | |
457 | ||
458 | /* | |
459 | * If head is small or the previous block is free, | |
460 | * merge'em. Note that 'small' is defined as smaller | |
461 | * than sizeof(int), which is very small but isn't too | |
462 | * uncommon for percpu allocations. | |
463 | */ | |
464 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { | |
465 | if (chunk->map[i - 1] > 0) | |
466 | chunk->map[i - 1] += head; | |
467 | else { | |
468 | chunk->map[i - 1] -= head; | |
469 | chunk->free_size -= head; | |
470 | } | |
471 | chunk->map[i] -= head; | |
472 | off += head; | |
473 | head = 0; | |
474 | } | |
475 | ||
476 | /* if tail is small, just keep it around */ | |
477 | tail = chunk->map[i] - head - size; | |
478 | if (tail < sizeof(int)) | |
479 | tail = 0; | |
480 | ||
481 | /* split if warranted */ | |
482 | if (head || tail) { | |
9f7dcf22 | 483 | pcpu_split_block(chunk, i, head, tail); |
fbf59bc9 TH |
484 | if (head) { |
485 | i++; | |
486 | off += head; | |
487 | max_contig = max(chunk->map[i - 1], max_contig); | |
488 | } | |
489 | if (tail) | |
490 | max_contig = max(chunk->map[i + 1], max_contig); | |
491 | } | |
492 | ||
493 | /* update hint and mark allocated */ | |
494 | if (is_last) | |
495 | chunk->contig_hint = max_contig; /* fully scanned */ | |
496 | else | |
497 | chunk->contig_hint = max(chunk->contig_hint, | |
498 | max_contig); | |
499 | ||
500 | chunk->free_size -= chunk->map[i]; | |
501 | chunk->map[i] = -chunk->map[i]; | |
502 | ||
503 | pcpu_chunk_relocate(chunk, oslot); | |
504 | return off; | |
505 | } | |
506 | ||
507 | chunk->contig_hint = max_contig; /* fully scanned */ | |
508 | pcpu_chunk_relocate(chunk, oslot); | |
509 | ||
9f7dcf22 TH |
510 | /* tell the upper layer that this chunk has no matching area */ |
511 | return -1; | |
fbf59bc9 TH |
512 | } |
513 | ||
514 | /** | |
515 | * pcpu_free_area - free area to a pcpu_chunk | |
516 | * @chunk: chunk of interest | |
517 | * @freeme: offset of area to free | |
518 | * | |
519 | * Free area starting from @freeme to @chunk. Note that this function | |
520 | * only modifies the allocation map. It doesn't depopulate or unmap | |
521 | * the area. | |
ccea34b5 TH |
522 | * |
523 | * CONTEXT: | |
524 | * pcpu_lock. | |
fbf59bc9 TH |
525 | */ |
526 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) | |
527 | { | |
528 | int oslot = pcpu_chunk_slot(chunk); | |
529 | int i, off; | |
530 | ||
531 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) | |
532 | if (off == freeme) | |
533 | break; | |
534 | BUG_ON(off != freeme); | |
535 | BUG_ON(chunk->map[i] > 0); | |
536 | ||
537 | chunk->map[i] = -chunk->map[i]; | |
538 | chunk->free_size += chunk->map[i]; | |
539 | ||
540 | /* merge with previous? */ | |
541 | if (i > 0 && chunk->map[i - 1] >= 0) { | |
542 | chunk->map[i - 1] += chunk->map[i]; | |
543 | chunk->map_used--; | |
544 | memmove(&chunk->map[i], &chunk->map[i + 1], | |
545 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
546 | i--; | |
547 | } | |
548 | /* merge with next? */ | |
549 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { | |
550 | chunk->map[i] += chunk->map[i + 1]; | |
551 | chunk->map_used--; | |
552 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], | |
553 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); | |
554 | } | |
555 | ||
556 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); | |
557 | pcpu_chunk_relocate(chunk, oslot); | |
558 | } | |
559 | ||
560 | /** | |
561 | * pcpu_unmap - unmap pages out of a pcpu_chunk | |
562 | * @chunk: chunk of interest | |
563 | * @page_start: page index of the first page to unmap | |
564 | * @page_end: page index of the last page to unmap + 1 | |
85ae87c1 | 565 | * @flush_tlb: whether to flush tlb or not |
fbf59bc9 TH |
566 | * |
567 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. | |
568 | * If @flush is true, vcache is flushed before unmapping and tlb | |
569 | * after. | |
570 | */ | |
571 | static void pcpu_unmap(struct pcpu_chunk *chunk, int page_start, int page_end, | |
85ae87c1 | 572 | bool flush_tlb) |
fbf59bc9 | 573 | { |
74d46d6b | 574 | unsigned int last = nr_cpu_ids - 1; |
fbf59bc9 TH |
575 | unsigned int cpu; |
576 | ||
8d408b4b TH |
577 | /* unmap must not be done on immutable chunk */ |
578 | WARN_ON(chunk->immutable); | |
579 | ||
fbf59bc9 TH |
580 | /* |
581 | * Each flushing trial can be very expensive, issue flush on | |
582 | * the whole region at once rather than doing it for each cpu. | |
583 | * This could be an overkill but is more scalable. | |
584 | */ | |
85ae87c1 TH |
585 | flush_cache_vunmap(pcpu_chunk_addr(chunk, 0, page_start), |
586 | pcpu_chunk_addr(chunk, last, page_end)); | |
fbf59bc9 TH |
587 | |
588 | for_each_possible_cpu(cpu) | |
589 | unmap_kernel_range_noflush( | |
590 | pcpu_chunk_addr(chunk, cpu, page_start), | |
591 | (page_end - page_start) << PAGE_SHIFT); | |
592 | ||
593 | /* ditto as flush_cache_vunmap() */ | |
85ae87c1 | 594 | if (flush_tlb) |
fbf59bc9 TH |
595 | flush_tlb_kernel_range(pcpu_chunk_addr(chunk, 0, page_start), |
596 | pcpu_chunk_addr(chunk, last, page_end)); | |
597 | } | |
598 | ||
599 | /** | |
600 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk | |
601 | * @chunk: chunk to depopulate | |
602 | * @off: offset to the area to depopulate | |
cae3aeb8 | 603 | * @size: size of the area to depopulate in bytes |
fbf59bc9 TH |
604 | * @flush: whether to flush cache and tlb or not |
605 | * | |
606 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) | |
607 | * from @chunk. If @flush is true, vcache is flushed before unmapping | |
608 | * and tlb after. | |
ccea34b5 TH |
609 | * |
610 | * CONTEXT: | |
611 | * pcpu_alloc_mutex. | |
fbf59bc9 | 612 | */ |
cae3aeb8 TH |
613 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size, |
614 | bool flush) | |
fbf59bc9 TH |
615 | { |
616 | int page_start = PFN_DOWN(off); | |
617 | int page_end = PFN_UP(off + size); | |
618 | int unmap_start = -1; | |
619 | int uninitialized_var(unmap_end); | |
620 | unsigned int cpu; | |
621 | int i; | |
622 | ||
623 | for (i = page_start; i < page_end; i++) { | |
624 | for_each_possible_cpu(cpu) { | |
625 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); | |
626 | ||
627 | if (!*pagep) | |
628 | continue; | |
629 | ||
630 | __free_page(*pagep); | |
631 | ||
632 | /* | |
633 | * If it's partial depopulation, it might get | |
634 | * populated or depopulated again. Mark the | |
635 | * page gone. | |
636 | */ | |
637 | *pagep = NULL; | |
638 | ||
639 | unmap_start = unmap_start < 0 ? i : unmap_start; | |
640 | unmap_end = i + 1; | |
641 | } | |
642 | } | |
643 | ||
644 | if (unmap_start >= 0) | |
645 | pcpu_unmap(chunk, unmap_start, unmap_end, flush); | |
646 | } | |
647 | ||
648 | /** | |
649 | * pcpu_map - map pages into a pcpu_chunk | |
650 | * @chunk: chunk of interest | |
651 | * @page_start: page index of the first page to map | |
652 | * @page_end: page index of the last page to map + 1 | |
653 | * | |
654 | * For each cpu, map pages [@page_start,@page_end) into @chunk. | |
655 | * vcache is flushed afterwards. | |
656 | */ | |
657 | static int pcpu_map(struct pcpu_chunk *chunk, int page_start, int page_end) | |
658 | { | |
74d46d6b | 659 | unsigned int last = nr_cpu_ids - 1; |
fbf59bc9 TH |
660 | unsigned int cpu; |
661 | int err; | |
662 | ||
8d408b4b TH |
663 | /* map must not be done on immutable chunk */ |
664 | WARN_ON(chunk->immutable); | |
665 | ||
fbf59bc9 TH |
666 | for_each_possible_cpu(cpu) { |
667 | err = map_kernel_range_noflush( | |
668 | pcpu_chunk_addr(chunk, cpu, page_start), | |
669 | (page_end - page_start) << PAGE_SHIFT, | |
670 | PAGE_KERNEL, | |
671 | pcpu_chunk_pagep(chunk, cpu, page_start)); | |
672 | if (err < 0) | |
673 | return err; | |
674 | } | |
675 | ||
676 | /* flush at once, please read comments in pcpu_unmap() */ | |
677 | flush_cache_vmap(pcpu_chunk_addr(chunk, 0, page_start), | |
678 | pcpu_chunk_addr(chunk, last, page_end)); | |
679 | return 0; | |
680 | } | |
681 | ||
682 | /** | |
683 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk | |
684 | * @chunk: chunk of interest | |
685 | * @off: offset to the area to populate | |
cae3aeb8 | 686 | * @size: size of the area to populate in bytes |
fbf59bc9 TH |
687 | * |
688 | * For each cpu, populate and map pages [@page_start,@page_end) into | |
689 | * @chunk. The area is cleared on return. | |
ccea34b5 TH |
690 | * |
691 | * CONTEXT: | |
692 | * pcpu_alloc_mutex, does GFP_KERNEL allocation. | |
fbf59bc9 TH |
693 | */ |
694 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) | |
695 | { | |
696 | const gfp_t alloc_mask = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; | |
697 | int page_start = PFN_DOWN(off); | |
698 | int page_end = PFN_UP(off + size); | |
699 | int map_start = -1; | |
02d51fdf | 700 | int uninitialized_var(map_end); |
fbf59bc9 TH |
701 | unsigned int cpu; |
702 | int i; | |
703 | ||
704 | for (i = page_start; i < page_end; i++) { | |
705 | if (pcpu_chunk_page_occupied(chunk, i)) { | |
706 | if (map_start >= 0) { | |
707 | if (pcpu_map(chunk, map_start, map_end)) | |
708 | goto err; | |
709 | map_start = -1; | |
710 | } | |
711 | continue; | |
712 | } | |
713 | ||
714 | map_start = map_start < 0 ? i : map_start; | |
715 | map_end = i + 1; | |
716 | ||
717 | for_each_possible_cpu(cpu) { | |
718 | struct page **pagep = pcpu_chunk_pagep(chunk, cpu, i); | |
719 | ||
720 | *pagep = alloc_pages_node(cpu_to_node(cpu), | |
721 | alloc_mask, 0); | |
722 | if (!*pagep) | |
723 | goto err; | |
e1b9aa3f | 724 | pcpu_set_page_chunk(*pagep, chunk); |
fbf59bc9 TH |
725 | } |
726 | } | |
727 | ||
728 | if (map_start >= 0 && pcpu_map(chunk, map_start, map_end)) | |
729 | goto err; | |
730 | ||
731 | for_each_possible_cpu(cpu) | |
d9b55eeb | 732 | memset(chunk->vm->addr + cpu * pcpu_unit_size + off, 0, |
fbf59bc9 TH |
733 | size); |
734 | ||
735 | return 0; | |
736 | err: | |
737 | /* likely under heavy memory pressure, give memory back */ | |
738 | pcpu_depopulate_chunk(chunk, off, size, true); | |
739 | return -ENOMEM; | |
740 | } | |
741 | ||
742 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) | |
743 | { | |
744 | if (!chunk) | |
745 | return; | |
746 | if (chunk->vm) | |
747 | free_vm_area(chunk->vm); | |
1880d93b | 748 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
fbf59bc9 TH |
749 | kfree(chunk); |
750 | } | |
751 | ||
752 | static struct pcpu_chunk *alloc_pcpu_chunk(void) | |
753 | { | |
754 | struct pcpu_chunk *chunk; | |
755 | ||
756 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); | |
757 | if (!chunk) | |
758 | return NULL; | |
759 | ||
1880d93b | 760 | chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
fbf59bc9 TH |
761 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
762 | chunk->map[chunk->map_used++] = pcpu_unit_size; | |
3e24aa58 | 763 | chunk->page = chunk->page_ar; |
fbf59bc9 | 764 | |
142d44b0 | 765 | chunk->vm = get_vm_area(pcpu_chunk_size, VM_ALLOC); |
fbf59bc9 TH |
766 | if (!chunk->vm) { |
767 | free_pcpu_chunk(chunk); | |
768 | return NULL; | |
769 | } | |
770 | ||
771 | INIT_LIST_HEAD(&chunk->list); | |
772 | chunk->free_size = pcpu_unit_size; | |
773 | chunk->contig_hint = pcpu_unit_size; | |
774 | ||
775 | return chunk; | |
776 | } | |
777 | ||
778 | /** | |
edcb4639 | 779 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 780 | * @size: size of area to allocate in bytes |
fbf59bc9 | 781 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 782 | * @reserved: allocate from the reserved chunk if available |
fbf59bc9 | 783 | * |
ccea34b5 TH |
784 | * Allocate percpu area of @size bytes aligned at @align. |
785 | * | |
786 | * CONTEXT: | |
787 | * Does GFP_KERNEL allocation. | |
fbf59bc9 TH |
788 | * |
789 | * RETURNS: | |
790 | * Percpu pointer to the allocated area on success, NULL on failure. | |
791 | */ | |
edcb4639 | 792 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
fbf59bc9 | 793 | { |
fbf59bc9 TH |
794 | struct pcpu_chunk *chunk; |
795 | int slot, off; | |
796 | ||
8d408b4b | 797 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
fbf59bc9 TH |
798 | WARN(true, "illegal size (%zu) or align (%zu) for " |
799 | "percpu allocation\n", size, align); | |
800 | return NULL; | |
801 | } | |
802 | ||
ccea34b5 TH |
803 | mutex_lock(&pcpu_alloc_mutex); |
804 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 805 | |
edcb4639 TH |
806 | /* serve reserved allocations from the reserved chunk if available */ |
807 | if (reserved && pcpu_reserved_chunk) { | |
808 | chunk = pcpu_reserved_chunk; | |
9f7dcf22 TH |
809 | if (size > chunk->contig_hint || |
810 | pcpu_extend_area_map(chunk) < 0) | |
ccea34b5 | 811 | goto fail_unlock; |
edcb4639 TH |
812 | off = pcpu_alloc_area(chunk, size, align); |
813 | if (off >= 0) | |
814 | goto area_found; | |
ccea34b5 | 815 | goto fail_unlock; |
edcb4639 TH |
816 | } |
817 | ||
ccea34b5 | 818 | restart: |
edcb4639 | 819 | /* search through normal chunks */ |
fbf59bc9 TH |
820 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
821 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
822 | if (size > chunk->contig_hint) | |
823 | continue; | |
ccea34b5 TH |
824 | |
825 | switch (pcpu_extend_area_map(chunk)) { | |
826 | case 0: | |
827 | break; | |
828 | case 1: | |
829 | goto restart; /* pcpu_lock dropped, restart */ | |
830 | default: | |
831 | goto fail_unlock; | |
832 | } | |
833 | ||
fbf59bc9 TH |
834 | off = pcpu_alloc_area(chunk, size, align); |
835 | if (off >= 0) | |
836 | goto area_found; | |
fbf59bc9 TH |
837 | } |
838 | } | |
839 | ||
840 | /* hmmm... no space left, create a new chunk */ | |
ccea34b5 TH |
841 | spin_unlock_irq(&pcpu_lock); |
842 | ||
fbf59bc9 TH |
843 | chunk = alloc_pcpu_chunk(); |
844 | if (!chunk) | |
ccea34b5 TH |
845 | goto fail_unlock_mutex; |
846 | ||
847 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 848 | pcpu_chunk_relocate(chunk, -1); |
ccea34b5 | 849 | goto restart; |
fbf59bc9 TH |
850 | |
851 | area_found: | |
ccea34b5 TH |
852 | spin_unlock_irq(&pcpu_lock); |
853 | ||
fbf59bc9 TH |
854 | /* populate, map and clear the area */ |
855 | if (pcpu_populate_chunk(chunk, off, size)) { | |
ccea34b5 | 856 | spin_lock_irq(&pcpu_lock); |
fbf59bc9 | 857 | pcpu_free_area(chunk, off); |
ccea34b5 | 858 | goto fail_unlock; |
fbf59bc9 TH |
859 | } |
860 | ||
ccea34b5 TH |
861 | mutex_unlock(&pcpu_alloc_mutex); |
862 | ||
863 | return __addr_to_pcpu_ptr(chunk->vm->addr + off); | |
864 | ||
865 | fail_unlock: | |
866 | spin_unlock_irq(&pcpu_lock); | |
867 | fail_unlock_mutex: | |
868 | mutex_unlock(&pcpu_alloc_mutex); | |
869 | return NULL; | |
fbf59bc9 | 870 | } |
edcb4639 TH |
871 | |
872 | /** | |
873 | * __alloc_percpu - allocate dynamic percpu area | |
874 | * @size: size of area to allocate in bytes | |
875 | * @align: alignment of area (max PAGE_SIZE) | |
876 | * | |
877 | * Allocate percpu area of @size bytes aligned at @align. Might | |
878 | * sleep. Might trigger writeouts. | |
879 | * | |
ccea34b5 TH |
880 | * CONTEXT: |
881 | * Does GFP_KERNEL allocation. | |
882 | * | |
edcb4639 TH |
883 | * RETURNS: |
884 | * Percpu pointer to the allocated area on success, NULL on failure. | |
885 | */ | |
886 | void *__alloc_percpu(size_t size, size_t align) | |
887 | { | |
888 | return pcpu_alloc(size, align, false); | |
889 | } | |
fbf59bc9 TH |
890 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
891 | ||
edcb4639 TH |
892 | /** |
893 | * __alloc_reserved_percpu - allocate reserved 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 from reserved | |
898 | * percpu area if arch has set it up; otherwise, allocation is served | |
899 | * from the same dynamic area. Might sleep. Might trigger writeouts. | |
900 | * | |
ccea34b5 TH |
901 | * CONTEXT: |
902 | * Does GFP_KERNEL allocation. | |
903 | * | |
edcb4639 TH |
904 | * RETURNS: |
905 | * Percpu pointer to the allocated area on success, NULL on failure. | |
906 | */ | |
907 | void *__alloc_reserved_percpu(size_t size, size_t align) | |
908 | { | |
909 | return pcpu_alloc(size, align, true); | |
910 | } | |
911 | ||
a56dbddf TH |
912 | /** |
913 | * pcpu_reclaim - reclaim fully free chunks, workqueue function | |
914 | * @work: unused | |
915 | * | |
916 | * Reclaim all fully free chunks except for the first one. | |
ccea34b5 TH |
917 | * |
918 | * CONTEXT: | |
919 | * workqueue context. | |
a56dbddf TH |
920 | */ |
921 | static void pcpu_reclaim(struct work_struct *work) | |
fbf59bc9 | 922 | { |
a56dbddf TH |
923 | LIST_HEAD(todo); |
924 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; | |
925 | struct pcpu_chunk *chunk, *next; | |
926 | ||
ccea34b5 TH |
927 | mutex_lock(&pcpu_alloc_mutex); |
928 | spin_lock_irq(&pcpu_lock); | |
a56dbddf TH |
929 | |
930 | list_for_each_entry_safe(chunk, next, head, list) { | |
931 | WARN_ON(chunk->immutable); | |
932 | ||
933 | /* spare the first one */ | |
934 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) | |
935 | continue; | |
936 | ||
a56dbddf TH |
937 | list_move(&chunk->list, &todo); |
938 | } | |
939 | ||
ccea34b5 TH |
940 | spin_unlock_irq(&pcpu_lock); |
941 | mutex_unlock(&pcpu_alloc_mutex); | |
a56dbddf TH |
942 | |
943 | list_for_each_entry_safe(chunk, next, &todo, list) { | |
944 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size, false); | |
945 | free_pcpu_chunk(chunk); | |
946 | } | |
fbf59bc9 TH |
947 | } |
948 | ||
949 | /** | |
950 | * free_percpu - free percpu area | |
951 | * @ptr: pointer to area to free | |
952 | * | |
ccea34b5 TH |
953 | * Free percpu area @ptr. |
954 | * | |
955 | * CONTEXT: | |
956 | * Can be called from atomic context. | |
fbf59bc9 TH |
957 | */ |
958 | void free_percpu(void *ptr) | |
959 | { | |
960 | void *addr = __pcpu_ptr_to_addr(ptr); | |
961 | struct pcpu_chunk *chunk; | |
ccea34b5 | 962 | unsigned long flags; |
fbf59bc9 TH |
963 | int off; |
964 | ||
965 | if (!ptr) | |
966 | return; | |
967 | ||
ccea34b5 | 968 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
969 | |
970 | chunk = pcpu_chunk_addr_search(addr); | |
971 | off = addr - chunk->vm->addr; | |
972 | ||
973 | pcpu_free_area(chunk, off); | |
974 | ||
a56dbddf | 975 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
976 | if (chunk->free_size == pcpu_unit_size) { |
977 | struct pcpu_chunk *pos; | |
978 | ||
a56dbddf | 979 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 980 | if (pos != chunk) { |
a56dbddf | 981 | schedule_work(&pcpu_reclaim_work); |
fbf59bc9 TH |
982 | break; |
983 | } | |
984 | } | |
985 | ||
ccea34b5 | 986 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
987 | } |
988 | EXPORT_SYMBOL_GPL(free_percpu); | |
989 | ||
990 | /** | |
8d408b4b TH |
991 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
992 | * @get_page_fn: callback to fetch page pointer | |
993 | * @static_size: the size of static percpu area in bytes | |
edcb4639 | 994 | * @reserved_size: the size of reserved percpu area in bytes |
cafe8816 | 995 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
6074d5b0 | 996 | * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto |
8d408b4b TH |
997 | * @base_addr: mapped address, NULL for auto |
998 | * @populate_pte_fn: callback to allocate pagetable, NULL if unnecessary | |
999 | * | |
1000 | * Initialize the first percpu chunk which contains the kernel static | |
1001 | * perpcu area. This function is to be called from arch percpu area | |
1002 | * setup path. The first two parameters are mandatory. The rest are | |
1003 | * optional. | |
1004 | * | |
1005 | * @get_page_fn() should return pointer to percpu page given cpu | |
1006 | * number and page number. It should at least return enough pages to | |
1007 | * cover the static area. The returned pages for static area should | |
1008 | * have been initialized with valid data. If @unit_size is specified, | |
1009 | * it can also return pages after the static area. NULL return | |
1010 | * indicates end of pages for the cpu. Note that @get_page_fn() must | |
1011 | * return the same number of pages for all cpus. | |
1012 | * | |
edcb4639 TH |
1013 | * @reserved_size, if non-zero, specifies the amount of bytes to |
1014 | * reserve after the static area in the first chunk. This reserves | |
1015 | * the first chunk such that it's available only through reserved | |
1016 | * percpu allocation. This is primarily used to serve module percpu | |
1017 | * static areas on architectures where the addressing model has | |
1018 | * limited offset range for symbol relocations to guarantee module | |
1019 | * percpu symbols fall inside the relocatable range. | |
1020 | * | |
6074d5b0 TH |
1021 | * @dyn_size, if non-negative, determines the number of bytes |
1022 | * available for dynamic allocation in the first chunk. Specifying | |
1023 | * non-negative value makes percpu leave alone the area beyond | |
1024 | * @static_size + @reserved_size + @dyn_size. | |
1025 | * | |
cafe8816 TH |
1026 | * @unit_size, if non-negative, specifies unit size and must be |
1027 | * aligned to PAGE_SIZE and equal to or larger than @static_size + | |
6074d5b0 | 1028 | * @reserved_size + if non-negative, @dyn_size. |
8d408b4b TH |
1029 | * |
1030 | * Non-null @base_addr means that the caller already allocated virtual | |
1031 | * region for the first chunk and mapped it. percpu must not mess | |
1032 | * with the chunk. Note that @base_addr with 0 @unit_size or non-NULL | |
1033 | * @populate_pte_fn doesn't make any sense. | |
1034 | * | |
1035 | * @populate_pte_fn is used to populate the pagetable. NULL means the | |
1036 | * caller already populated the pagetable. | |
fbf59bc9 | 1037 | * |
edcb4639 TH |
1038 | * If the first chunk ends up with both reserved and dynamic areas, it |
1039 | * is served by two chunks - one to serve the core static and reserved | |
1040 | * areas and the other for the dynamic area. They share the same vm | |
1041 | * and page map but uses different area allocation map to stay away | |
1042 | * from each other. The latter chunk is circulated in the chunk slots | |
1043 | * and available for dynamic allocation like any other chunks. | |
1044 | * | |
fbf59bc9 TH |
1045 | * RETURNS: |
1046 | * The determined pcpu_unit_size which can be used to initialize | |
1047 | * percpu access. | |
1048 | */ | |
8d408b4b | 1049 | size_t __init pcpu_setup_first_chunk(pcpu_get_page_fn_t get_page_fn, |
edcb4639 | 1050 | size_t static_size, size_t reserved_size, |
6074d5b0 | 1051 | ssize_t dyn_size, ssize_t unit_size, |
cafe8816 | 1052 | void *base_addr, |
8d408b4b | 1053 | pcpu_populate_pte_fn_t populate_pte_fn) |
fbf59bc9 | 1054 | { |
2441d15c | 1055 | static struct vm_struct first_vm; |
edcb4639 | 1056 | static int smap[2], dmap[2]; |
6074d5b0 TH |
1057 | size_t size_sum = static_size + reserved_size + |
1058 | (dyn_size >= 0 ? dyn_size : 0); | |
edcb4639 | 1059 | struct pcpu_chunk *schunk, *dchunk = NULL; |
fbf59bc9 | 1060 | unsigned int cpu; |
8d408b4b | 1061 | int nr_pages; |
fbf59bc9 TH |
1062 | int err, i; |
1063 | ||
8d408b4b | 1064 | /* santiy checks */ |
edcb4639 TH |
1065 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
1066 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); | |
8d408b4b | 1067 | BUG_ON(!static_size); |
cafe8816 | 1068 | if (unit_size >= 0) { |
6074d5b0 | 1069 | BUG_ON(unit_size < size_sum); |
cafe8816 | 1070 | BUG_ON(unit_size & ~PAGE_MASK); |
6074d5b0 TH |
1071 | BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE); |
1072 | } else | |
cafe8816 | 1073 | BUG_ON(base_addr); |
8d408b4b | 1074 | BUG_ON(base_addr && populate_pte_fn); |
fbf59bc9 | 1075 | |
cafe8816 | 1076 | if (unit_size >= 0) |
8d408b4b TH |
1077 | pcpu_unit_pages = unit_size >> PAGE_SHIFT; |
1078 | else | |
1079 | pcpu_unit_pages = max_t(int, PCPU_MIN_UNIT_SIZE >> PAGE_SHIFT, | |
6074d5b0 | 1080 | PFN_UP(size_sum)); |
8d408b4b | 1081 | |
d9b55eeb | 1082 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
74d46d6b | 1083 | pcpu_chunk_size = nr_cpu_ids * pcpu_unit_size; |
fbf59bc9 | 1084 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) |
74d46d6b | 1085 | + nr_cpu_ids * pcpu_unit_pages * sizeof(struct page *); |
fbf59bc9 | 1086 | |
cafe8816 | 1087 | if (dyn_size < 0) |
edcb4639 | 1088 | dyn_size = pcpu_unit_size - static_size - reserved_size; |
cafe8816 | 1089 | |
d9b55eeb TH |
1090 | /* |
1091 | * Allocate chunk slots. The additional last slot is for | |
1092 | * empty chunks. | |
1093 | */ | |
1094 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
fbf59bc9 TH |
1095 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
1096 | for (i = 0; i < pcpu_nr_slots; i++) | |
1097 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1098 | ||
edcb4639 TH |
1099 | /* |
1100 | * Initialize static chunk. If reserved_size is zero, the | |
1101 | * static chunk covers static area + dynamic allocation area | |
1102 | * in the first chunk. If reserved_size is not zero, it | |
1103 | * covers static area + reserved area (mostly used for module | |
1104 | * static percpu allocation). | |
1105 | */ | |
2441d15c TH |
1106 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
1107 | INIT_LIST_HEAD(&schunk->list); | |
1108 | schunk->vm = &first_vm; | |
61ace7fa TH |
1109 | schunk->map = smap; |
1110 | schunk->map_alloc = ARRAY_SIZE(smap); | |
3e24aa58 | 1111 | schunk->page = schunk->page_ar; |
edcb4639 TH |
1112 | |
1113 | if (reserved_size) { | |
1114 | schunk->free_size = reserved_size; | |
ae9e6bc9 TH |
1115 | pcpu_reserved_chunk = schunk; |
1116 | pcpu_reserved_chunk_limit = static_size + reserved_size; | |
edcb4639 TH |
1117 | } else { |
1118 | schunk->free_size = dyn_size; | |
1119 | dyn_size = 0; /* dynamic area covered */ | |
1120 | } | |
2441d15c | 1121 | schunk->contig_hint = schunk->free_size; |
fbf59bc9 | 1122 | |
61ace7fa TH |
1123 | schunk->map[schunk->map_used++] = -static_size; |
1124 | if (schunk->free_size) | |
1125 | schunk->map[schunk->map_used++] = schunk->free_size; | |
1126 | ||
edcb4639 TH |
1127 | /* init dynamic chunk if necessary */ |
1128 | if (dyn_size) { | |
1129 | dchunk = alloc_bootmem(sizeof(struct pcpu_chunk)); | |
1130 | INIT_LIST_HEAD(&dchunk->list); | |
1131 | dchunk->vm = &first_vm; | |
1132 | dchunk->map = dmap; | |
1133 | dchunk->map_alloc = ARRAY_SIZE(dmap); | |
1134 | dchunk->page = schunk->page_ar; /* share page map with schunk */ | |
1135 | ||
1136 | dchunk->contig_hint = dchunk->free_size = dyn_size; | |
1137 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; | |
1138 | dchunk->map[dchunk->map_used++] = dchunk->free_size; | |
1139 | } | |
1140 | ||
8d408b4b | 1141 | /* allocate vm address */ |
2441d15c TH |
1142 | first_vm.flags = VM_ALLOC; |
1143 | first_vm.size = pcpu_chunk_size; | |
8d408b4b TH |
1144 | |
1145 | if (!base_addr) | |
2441d15c | 1146 | vm_area_register_early(&first_vm, PAGE_SIZE); |
8d408b4b TH |
1147 | else { |
1148 | /* | |
1149 | * Pages already mapped. No need to remap into | |
edcb4639 TH |
1150 | * vmalloc area. In this case the first chunks can't |
1151 | * be mapped or unmapped by percpu and are marked | |
8d408b4b TH |
1152 | * immutable. |
1153 | */ | |
2441d15c TH |
1154 | first_vm.addr = base_addr; |
1155 | schunk->immutable = true; | |
edcb4639 TH |
1156 | if (dchunk) |
1157 | dchunk->immutable = true; | |
8d408b4b TH |
1158 | } |
1159 | ||
1160 | /* assign pages */ | |
1161 | nr_pages = -1; | |
fbf59bc9 | 1162 | for_each_possible_cpu(cpu) { |
8d408b4b TH |
1163 | for (i = 0; i < pcpu_unit_pages; i++) { |
1164 | struct page *page = get_page_fn(cpu, i); | |
1165 | ||
1166 | if (!page) | |
1167 | break; | |
2441d15c | 1168 | *pcpu_chunk_pagep(schunk, cpu, i) = page; |
fbf59bc9 | 1169 | } |
8d408b4b | 1170 | |
61ace7fa | 1171 | BUG_ON(i < PFN_UP(static_size)); |
8d408b4b TH |
1172 | |
1173 | if (nr_pages < 0) | |
1174 | nr_pages = i; | |
1175 | else | |
1176 | BUG_ON(nr_pages != i); | |
fbf59bc9 TH |
1177 | } |
1178 | ||
8d408b4b TH |
1179 | /* map them */ |
1180 | if (populate_pte_fn) { | |
1181 | for_each_possible_cpu(cpu) | |
1182 | for (i = 0; i < nr_pages; i++) | |
2441d15c | 1183 | populate_pte_fn(pcpu_chunk_addr(schunk, |
8d408b4b TH |
1184 | cpu, i)); |
1185 | ||
2441d15c | 1186 | err = pcpu_map(schunk, 0, nr_pages); |
8d408b4b TH |
1187 | if (err) |
1188 | panic("failed to setup static percpu area, err=%d\n", | |
1189 | err); | |
1190 | } | |
fbf59bc9 | 1191 | |
2441d15c | 1192 | /* link the first chunk in */ |
ae9e6bc9 TH |
1193 | pcpu_first_chunk = dchunk ?: schunk; |
1194 | pcpu_chunk_relocate(pcpu_first_chunk, -1); | |
fbf59bc9 TH |
1195 | |
1196 | /* we're done */ | |
2441d15c | 1197 | pcpu_base_addr = (void *)pcpu_chunk_addr(schunk, 0, 0); |
fbf59bc9 TH |
1198 | return pcpu_unit_size; |
1199 | } | |
66c3a757 TH |
1200 | |
1201 | /* | |
1202 | * Embedding first chunk setup helper. | |
1203 | */ | |
1204 | static void *pcpue_ptr __initdata; | |
1205 | static size_t pcpue_size __initdata; | |
1206 | static size_t pcpue_unit_size __initdata; | |
1207 | ||
1208 | static struct page * __init pcpue_get_page(unsigned int cpu, int pageno) | |
1209 | { | |
1210 | size_t off = (size_t)pageno << PAGE_SHIFT; | |
1211 | ||
1212 | if (off >= pcpue_size) | |
1213 | return NULL; | |
1214 | ||
1215 | return virt_to_page(pcpue_ptr + cpu * pcpue_unit_size + off); | |
1216 | } | |
1217 | ||
1218 | /** | |
1219 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
1220 | * @static_size: the size of static percpu area in bytes | |
1221 | * @reserved_size: the size of reserved percpu area in bytes | |
1222 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto | |
1223 | * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE, -1 for auto | |
1224 | * | |
1225 | * This is a helper to ease setting up embedded first percpu chunk and | |
1226 | * can be called where pcpu_setup_first_chunk() is expected. | |
1227 | * | |
1228 | * If this function is used to setup the first chunk, it is allocated | |
1229 | * as a contiguous area using bootmem allocator and used as-is without | |
1230 | * being mapped into vmalloc area. This enables the first chunk to | |
1231 | * piggy back on the linear physical mapping which often uses larger | |
1232 | * page size. | |
1233 | * | |
1234 | * When @dyn_size is positive, dynamic area might be larger than | |
1235 | * specified to fill page alignment. Also, when @dyn_size is auto, | |
1236 | * @dyn_size does not fill the whole first chunk but only what's | |
1237 | * necessary for page alignment after static and reserved areas. | |
1238 | * | |
1239 | * If the needed size is smaller than the minimum or specified unit | |
1240 | * size, the leftover is returned to the bootmem allocator. | |
1241 | * | |
1242 | * RETURNS: | |
1243 | * The determined pcpu_unit_size which can be used to initialize | |
1244 | * percpu access on success, -errno on failure. | |
1245 | */ | |
1246 | ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size, | |
1247 | ssize_t dyn_size, ssize_t unit_size) | |
1248 | { | |
fa8a7094 | 1249 | size_t chunk_size; |
66c3a757 TH |
1250 | unsigned int cpu; |
1251 | ||
1252 | /* determine parameters and allocate */ | |
1253 | pcpue_size = PFN_ALIGN(static_size + reserved_size + | |
1254 | (dyn_size >= 0 ? dyn_size : 0)); | |
1255 | if (dyn_size != 0) | |
1256 | dyn_size = pcpue_size - static_size - reserved_size; | |
1257 | ||
1258 | if (unit_size >= 0) { | |
1259 | BUG_ON(unit_size < pcpue_size); | |
1260 | pcpue_unit_size = unit_size; | |
1261 | } else | |
1262 | pcpue_unit_size = max_t(size_t, pcpue_size, PCPU_MIN_UNIT_SIZE); | |
1263 | ||
74d46d6b | 1264 | chunk_size = pcpue_unit_size * nr_cpu_ids; |
fa8a7094 TH |
1265 | |
1266 | pcpue_ptr = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE, | |
1267 | __pa(MAX_DMA_ADDRESS)); | |
1268 | if (!pcpue_ptr) { | |
1269 | pr_warning("PERCPU: failed to allocate %zu bytes for " | |
1270 | "embedding\n", chunk_size); | |
66c3a757 | 1271 | return -ENOMEM; |
fa8a7094 | 1272 | } |
66c3a757 TH |
1273 | |
1274 | /* return the leftover and copy */ | |
74d46d6b | 1275 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) { |
66c3a757 TH |
1276 | void *ptr = pcpue_ptr + cpu * pcpue_unit_size; |
1277 | ||
74d46d6b TH |
1278 | if (cpu_possible(cpu)) { |
1279 | free_bootmem(__pa(ptr + pcpue_size), | |
1280 | pcpue_unit_size - pcpue_size); | |
1281 | memcpy(ptr, __per_cpu_load, static_size); | |
1282 | } else | |
1283 | free_bootmem(__pa(ptr), pcpue_unit_size); | |
66c3a757 TH |
1284 | } |
1285 | ||
1286 | /* we're ready, commit */ | |
1287 | pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n", | |
1288 | pcpue_size >> PAGE_SHIFT, pcpue_ptr, static_size); | |
1289 | ||
1290 | return pcpu_setup_first_chunk(pcpue_get_page, static_size, | |
1291 | reserved_size, dyn_size, | |
1292 | pcpue_unit_size, pcpue_ptr, NULL); | |
1293 | } |