]>
Commit | Line | Data |
---|---|---|
fbf59bc9 | 1 | /* |
88999a89 | 2 | * mm/percpu.c - percpu memory allocator |
fbf59bc9 TH |
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 | |
88999a89 TH |
10 | * areas. Percpu areas are allocated in chunks. Each chunk is |
11 | * consisted of boot-time determined number of units and the first | |
12 | * chunk is used for static percpu variables in the kernel image | |
2f39e637 TH |
13 | * (special boot time alloc/init handling necessary as these areas |
14 | * need to be brought up before allocation services are running). | |
15 | * Unit grows as necessary and all units grow or shrink in unison. | |
88999a89 | 16 | * When a chunk is filled up, another chunk is allocated. |
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, | |
2f39e637 TH |
25 | * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to |
26 | * cpus. On NUMA, the mapping can be non-linear and even sparse. | |
27 | * Percpu access can be done by configuring percpu base registers | |
28 | * according to cpu to unit mapping and pcpu_unit_size. | |
fbf59bc9 | 29 | * |
2f39e637 TH |
30 | * There are usually many small percpu allocations many of them being |
31 | * as small as 4 bytes. The allocator organizes chunks into lists | |
fbf59bc9 TH |
32 | * according to free size and tries to allocate from the fullest one. |
33 | * Each chunk keeps the maximum contiguous area size hint which is | |
4785879e | 34 | * guaranteed to be equal to or larger than the maximum contiguous |
fbf59bc9 TH |
35 | * area in the chunk. This helps the allocator not to iterate the |
36 | * chunk maps unnecessarily. | |
37 | * | |
38 | * Allocation state in each chunk is kept using an array of integers | |
39 | * on chunk->map. A positive value in the map represents a free | |
40 | * region and negative allocated. Allocation inside a chunk is done | |
41 | * by scanning this map sequentially and serving the first matching | |
42 | * entry. This is mostly copied from the percpu_modalloc() allocator. | |
e1b9aa3f CL |
43 | * Chunks can be determined from the address using the index field |
44 | * in the page struct. The index field contains a pointer to the chunk. | |
fbf59bc9 TH |
45 | * |
46 | * To use this allocator, arch code should do the followings. | |
47 | * | |
fbf59bc9 | 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> | |
fd1e8a1f | 58 | #include <linux/err.h> |
fbf59bc9 | 59 | #include <linux/list.h> |
a530b795 | 60 | #include <linux/log2.h> |
fbf59bc9 TH |
61 | #include <linux/mm.h> |
62 | #include <linux/module.h> | |
63 | #include <linux/mutex.h> | |
64 | #include <linux/percpu.h> | |
65 | #include <linux/pfn.h> | |
fbf59bc9 | 66 | #include <linux/slab.h> |
ccea34b5 | 67 | #include <linux/spinlock.h> |
fbf59bc9 | 68 | #include <linux/vmalloc.h> |
a56dbddf | 69 | #include <linux/workqueue.h> |
f528f0b8 | 70 | #include <linux/kmemleak.h> |
fbf59bc9 TH |
71 | |
72 | #include <asm/cacheflush.h> | |
e0100983 | 73 | #include <asm/sections.h> |
fbf59bc9 | 74 | #include <asm/tlbflush.h> |
3b034b0d | 75 | #include <asm/io.h> |
fbf59bc9 | 76 | |
fbf59bc9 TH |
77 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
78 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ | |
9c824b6a TH |
79 | #define PCPU_ATOMIC_MAP_MARGIN_LOW 32 |
80 | #define PCPU_ATOMIC_MAP_MARGIN_HIGH 64 | |
fbf59bc9 | 81 | |
bbddff05 | 82 | #ifdef CONFIG_SMP |
e0100983 TH |
83 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
84 | #ifndef __addr_to_pcpu_ptr | |
85 | #define __addr_to_pcpu_ptr(addr) \ | |
43cf38eb TH |
86 | (void __percpu *)((unsigned long)(addr) - \ |
87 | (unsigned long)pcpu_base_addr + \ | |
88 | (unsigned long)__per_cpu_start) | |
e0100983 TH |
89 | #endif |
90 | #ifndef __pcpu_ptr_to_addr | |
91 | #define __pcpu_ptr_to_addr(ptr) \ | |
43cf38eb TH |
92 | (void __force *)((unsigned long)(ptr) + \ |
93 | (unsigned long)pcpu_base_addr - \ | |
94 | (unsigned long)__per_cpu_start) | |
e0100983 | 95 | #endif |
bbddff05 TH |
96 | #else /* CONFIG_SMP */ |
97 | /* on UP, it's always identity mapped */ | |
98 | #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) | |
99 | #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) | |
100 | #endif /* CONFIG_SMP */ | |
e0100983 | 101 | |
fbf59bc9 TH |
102 | struct pcpu_chunk { |
103 | struct list_head list; /* linked to pcpu_slot lists */ | |
fbf59bc9 TH |
104 | int free_size; /* free bytes in the chunk */ |
105 | int contig_hint; /* max contiguous size hint */ | |
bba174f5 | 106 | void *base_addr; /* base address of this chunk */ |
9c824b6a | 107 | |
723ad1d9 | 108 | int map_used; /* # of map entries used before the sentry */ |
fbf59bc9 TH |
109 | int map_alloc; /* # of map entries allocated */ |
110 | int *map; /* allocation map */ | |
9c824b6a TH |
111 | struct work_struct map_extend_work;/* async ->map[] extension */ |
112 | ||
88999a89 | 113 | void *data; /* chunk data */ |
3d331ad7 | 114 | int first_free; /* no free below this */ |
8d408b4b | 115 | bool immutable; /* no [de]population allowed */ |
b539b87f | 116 | int nr_populated; /* # of populated pages */ |
ce3141a2 | 117 | unsigned long populated[]; /* populated bitmap */ |
fbf59bc9 TH |
118 | }; |
119 | ||
40150d37 TH |
120 | static int pcpu_unit_pages __read_mostly; |
121 | static int pcpu_unit_size __read_mostly; | |
2f39e637 | 122 | static int pcpu_nr_units __read_mostly; |
6563297c | 123 | static int pcpu_atom_size __read_mostly; |
40150d37 TH |
124 | static int pcpu_nr_slots __read_mostly; |
125 | static size_t pcpu_chunk_struct_size __read_mostly; | |
fbf59bc9 | 126 | |
a855b84c TH |
127 | /* cpus with the lowest and highest unit addresses */ |
128 | static unsigned int pcpu_low_unit_cpu __read_mostly; | |
129 | static unsigned int pcpu_high_unit_cpu __read_mostly; | |
2f39e637 | 130 | |
fbf59bc9 | 131 | /* the address of the first chunk which starts with the kernel static area */ |
40150d37 | 132 | void *pcpu_base_addr __read_mostly; |
fbf59bc9 TH |
133 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
134 | ||
fb435d52 TH |
135 | static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ |
136 | const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ | |
2f39e637 | 137 | |
6563297c TH |
138 | /* group information, used for vm allocation */ |
139 | static int pcpu_nr_groups __read_mostly; | |
140 | static const unsigned long *pcpu_group_offsets __read_mostly; | |
141 | static const size_t *pcpu_group_sizes __read_mostly; | |
142 | ||
ae9e6bc9 TH |
143 | /* |
144 | * The first chunk which always exists. Note that unlike other | |
145 | * chunks, this one can be allocated and mapped in several different | |
146 | * ways and thus often doesn't live in the vmalloc area. | |
147 | */ | |
148 | static struct pcpu_chunk *pcpu_first_chunk; | |
149 | ||
150 | /* | |
151 | * Optional reserved chunk. This chunk reserves part of the first | |
152 | * chunk and serves it for reserved allocations. The amount of | |
153 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved | |
154 | * area doesn't exist, the following variables contain NULL and 0 | |
155 | * respectively. | |
156 | */ | |
edcb4639 | 157 | static struct pcpu_chunk *pcpu_reserved_chunk; |
edcb4639 TH |
158 | static int pcpu_reserved_chunk_limit; |
159 | ||
b38d08f3 TH |
160 | static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ |
161 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop */ | |
fbf59bc9 | 162 | |
40150d37 | 163 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
fbf59bc9 | 164 | |
b539b87f TH |
165 | /* |
166 | * The number of empty populated pages, protected by pcpu_lock. The | |
167 | * reserved chunk doesn't contribute to the count. | |
168 | */ | |
169 | static int pcpu_nr_empty_pop_pages; | |
170 | ||
a56dbddf TH |
171 | /* reclaim work to release fully free chunks, scheduled from free path */ |
172 | static void pcpu_reclaim(struct work_struct *work); | |
173 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); | |
174 | ||
020ec653 TH |
175 | static bool pcpu_addr_in_first_chunk(void *addr) |
176 | { | |
177 | void *first_start = pcpu_first_chunk->base_addr; | |
178 | ||
179 | return addr >= first_start && addr < first_start + pcpu_unit_size; | |
180 | } | |
181 | ||
182 | static bool pcpu_addr_in_reserved_chunk(void *addr) | |
183 | { | |
184 | void *first_start = pcpu_first_chunk->base_addr; | |
185 | ||
186 | return addr >= first_start && | |
187 | addr < first_start + pcpu_reserved_chunk_limit; | |
188 | } | |
189 | ||
d9b55eeb | 190 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 191 | { |
cae3aeb8 | 192 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
193 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
194 | } | |
195 | ||
d9b55eeb TH |
196 | static int pcpu_size_to_slot(int size) |
197 | { | |
198 | if (size == pcpu_unit_size) | |
199 | return pcpu_nr_slots - 1; | |
200 | return __pcpu_size_to_slot(size); | |
201 | } | |
202 | ||
fbf59bc9 TH |
203 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
204 | { | |
205 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
206 | return 0; | |
207 | ||
208 | return pcpu_size_to_slot(chunk->free_size); | |
209 | } | |
210 | ||
88999a89 TH |
211 | /* set the pointer to a chunk in a page struct */ |
212 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
213 | { | |
214 | page->index = (unsigned long)pcpu; | |
215 | } | |
216 | ||
217 | /* obtain pointer to a chunk from a page struct */ | |
218 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
219 | { | |
220 | return (struct pcpu_chunk *)page->index; | |
221 | } | |
222 | ||
223 | static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) | |
fbf59bc9 | 224 | { |
2f39e637 | 225 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
226 | } |
227 | ||
9983b6f0 TH |
228 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
229 | unsigned int cpu, int page_idx) | |
fbf59bc9 | 230 | { |
bba174f5 | 231 | return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + |
fb435d52 | 232 | (page_idx << PAGE_SHIFT); |
fbf59bc9 TH |
233 | } |
234 | ||
88999a89 TH |
235 | static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, |
236 | int *rs, int *re, int end) | |
ce3141a2 TH |
237 | { |
238 | *rs = find_next_zero_bit(chunk->populated, end, *rs); | |
239 | *re = find_next_bit(chunk->populated, end, *rs + 1); | |
240 | } | |
241 | ||
88999a89 TH |
242 | static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, |
243 | int *rs, int *re, int end) | |
ce3141a2 TH |
244 | { |
245 | *rs = find_next_bit(chunk->populated, end, *rs); | |
246 | *re = find_next_zero_bit(chunk->populated, end, *rs + 1); | |
247 | } | |
248 | ||
249 | /* | |
250 | * (Un)populated page region iterators. Iterate over (un)populated | |
b595076a | 251 | * page regions between @start and @end in @chunk. @rs and @re should |
ce3141a2 TH |
252 | * be integer variables and will be set to start and end page index of |
253 | * the current region. | |
254 | */ | |
255 | #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ | |
256 | for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ | |
257 | (rs) < (re); \ | |
258 | (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) | |
259 | ||
260 | #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ | |
261 | for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ | |
262 | (rs) < (re); \ | |
263 | (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) | |
264 | ||
fbf59bc9 | 265 | /** |
90459ce0 | 266 | * pcpu_mem_zalloc - allocate memory |
1880d93b | 267 | * @size: bytes to allocate |
fbf59bc9 | 268 | * |
1880d93b | 269 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
90459ce0 | 270 | * kzalloc() is used; otherwise, vzalloc() is used. The returned |
1880d93b | 271 | * memory is always zeroed. |
fbf59bc9 | 272 | * |
ccea34b5 TH |
273 | * CONTEXT: |
274 | * Does GFP_KERNEL allocation. | |
275 | * | |
fbf59bc9 | 276 | * RETURNS: |
1880d93b | 277 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 278 | */ |
90459ce0 | 279 | static void *pcpu_mem_zalloc(size_t size) |
fbf59bc9 | 280 | { |
099a19d9 TH |
281 | if (WARN_ON_ONCE(!slab_is_available())) |
282 | return NULL; | |
283 | ||
1880d93b TH |
284 | if (size <= PAGE_SIZE) |
285 | return kzalloc(size, GFP_KERNEL); | |
7af4c093 JJ |
286 | else |
287 | return vzalloc(size); | |
1880d93b | 288 | } |
fbf59bc9 | 289 | |
1880d93b TH |
290 | /** |
291 | * pcpu_mem_free - free memory | |
292 | * @ptr: memory to free | |
293 | * @size: size of the area | |
294 | * | |
90459ce0 | 295 | * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). |
1880d93b TH |
296 | */ |
297 | static void pcpu_mem_free(void *ptr, size_t size) | |
298 | { | |
fbf59bc9 | 299 | if (size <= PAGE_SIZE) |
1880d93b | 300 | kfree(ptr); |
fbf59bc9 | 301 | else |
1880d93b | 302 | vfree(ptr); |
fbf59bc9 TH |
303 | } |
304 | ||
b539b87f TH |
305 | /** |
306 | * pcpu_count_occupied_pages - count the number of pages an area occupies | |
307 | * @chunk: chunk of interest | |
308 | * @i: index of the area in question | |
309 | * | |
310 | * Count the number of pages chunk's @i'th area occupies. When the area's | |
311 | * start and/or end address isn't aligned to page boundary, the straddled | |
312 | * page is included in the count iff the rest of the page is free. | |
313 | */ | |
314 | static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i) | |
315 | { | |
316 | int off = chunk->map[i] & ~1; | |
317 | int end = chunk->map[i + 1] & ~1; | |
318 | ||
319 | if (!PAGE_ALIGNED(off) && i > 0) { | |
320 | int prev = chunk->map[i - 1]; | |
321 | ||
322 | if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE)) | |
323 | off = round_down(off, PAGE_SIZE); | |
324 | } | |
325 | ||
326 | if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) { | |
327 | int next = chunk->map[i + 1]; | |
328 | int nend = chunk->map[i + 2] & ~1; | |
329 | ||
330 | if (!(next & 1) && nend >= round_up(end, PAGE_SIZE)) | |
331 | end = round_up(end, PAGE_SIZE); | |
332 | } | |
333 | ||
334 | return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0); | |
335 | } | |
336 | ||
fbf59bc9 TH |
337 | /** |
338 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
339 | * @chunk: chunk of interest | |
340 | * @oslot: the previous slot it was on | |
341 | * | |
342 | * This function is called after an allocation or free changed @chunk. | |
343 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
344 | * moved to the slot. Note that the reserved chunk is never put on |
345 | * chunk slots. | |
ccea34b5 TH |
346 | * |
347 | * CONTEXT: | |
348 | * pcpu_lock. | |
fbf59bc9 TH |
349 | */ |
350 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
351 | { | |
352 | int nslot = pcpu_chunk_slot(chunk); | |
353 | ||
edcb4639 | 354 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
355 | if (oslot < nslot) |
356 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
357 | else | |
358 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
359 | } | |
360 | } | |
361 | ||
9f7dcf22 | 362 | /** |
833af842 TH |
363 | * pcpu_need_to_extend - determine whether chunk area map needs to be extended |
364 | * @chunk: chunk of interest | |
9c824b6a | 365 | * @is_atomic: the allocation context |
9f7dcf22 | 366 | * |
9c824b6a TH |
367 | * Determine whether area map of @chunk needs to be extended. If |
368 | * @is_atomic, only the amount necessary for a new allocation is | |
369 | * considered; however, async extension is scheduled if the left amount is | |
370 | * low. If !@is_atomic, it aims for more empty space. Combined, this | |
371 | * ensures that the map is likely to have enough available space to | |
372 | * accomodate atomic allocations which can't extend maps directly. | |
9f7dcf22 | 373 | * |
ccea34b5 | 374 | * CONTEXT: |
833af842 | 375 | * pcpu_lock. |
ccea34b5 | 376 | * |
9f7dcf22 | 377 | * RETURNS: |
833af842 TH |
378 | * New target map allocation length if extension is necessary, 0 |
379 | * otherwise. | |
9f7dcf22 | 380 | */ |
9c824b6a | 381 | static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic) |
9f7dcf22 | 382 | { |
9c824b6a TH |
383 | int margin, new_alloc; |
384 | ||
385 | if (is_atomic) { | |
386 | margin = 3; | |
9f7dcf22 | 387 | |
9c824b6a TH |
388 | if (chunk->map_alloc < |
389 | chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) | |
390 | schedule_work(&chunk->map_extend_work); | |
391 | } else { | |
392 | margin = PCPU_ATOMIC_MAP_MARGIN_HIGH; | |
393 | } | |
394 | ||
395 | if (chunk->map_alloc >= chunk->map_used + margin) | |
9f7dcf22 TH |
396 | return 0; |
397 | ||
398 | new_alloc = PCPU_DFL_MAP_ALLOC; | |
9c824b6a | 399 | while (new_alloc < chunk->map_used + margin) |
9f7dcf22 TH |
400 | new_alloc *= 2; |
401 | ||
833af842 TH |
402 | return new_alloc; |
403 | } | |
404 | ||
405 | /** | |
406 | * pcpu_extend_area_map - extend area map of a chunk | |
407 | * @chunk: chunk of interest | |
408 | * @new_alloc: new target allocation length of the area map | |
409 | * | |
410 | * Extend area map of @chunk to have @new_alloc entries. | |
411 | * | |
412 | * CONTEXT: | |
413 | * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. | |
414 | * | |
415 | * RETURNS: | |
416 | * 0 on success, -errno on failure. | |
417 | */ | |
418 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) | |
419 | { | |
420 | int *old = NULL, *new = NULL; | |
421 | size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); | |
422 | unsigned long flags; | |
423 | ||
90459ce0 | 424 | new = pcpu_mem_zalloc(new_size); |
833af842 | 425 | if (!new) |
9f7dcf22 | 426 | return -ENOMEM; |
ccea34b5 | 427 | |
833af842 TH |
428 | /* acquire pcpu_lock and switch to new area map */ |
429 | spin_lock_irqsave(&pcpu_lock, flags); | |
430 | ||
431 | if (new_alloc <= chunk->map_alloc) | |
432 | goto out_unlock; | |
9f7dcf22 | 433 | |
833af842 | 434 | old_size = chunk->map_alloc * sizeof(chunk->map[0]); |
a002d148 HS |
435 | old = chunk->map; |
436 | ||
437 | memcpy(new, old, old_size); | |
9f7dcf22 | 438 | |
9f7dcf22 TH |
439 | chunk->map_alloc = new_alloc; |
440 | chunk->map = new; | |
833af842 TH |
441 | new = NULL; |
442 | ||
443 | out_unlock: | |
444 | spin_unlock_irqrestore(&pcpu_lock, flags); | |
445 | ||
446 | /* | |
447 | * pcpu_mem_free() might end up calling vfree() which uses | |
448 | * IRQ-unsafe lock and thus can't be called under pcpu_lock. | |
449 | */ | |
450 | pcpu_mem_free(old, old_size); | |
451 | pcpu_mem_free(new, new_size); | |
452 | ||
9f7dcf22 TH |
453 | return 0; |
454 | } | |
455 | ||
9c824b6a TH |
456 | static void pcpu_map_extend_workfn(struct work_struct *work) |
457 | { | |
458 | struct pcpu_chunk *chunk = container_of(work, struct pcpu_chunk, | |
459 | map_extend_work); | |
460 | int new_alloc; | |
461 | ||
462 | spin_lock_irq(&pcpu_lock); | |
463 | new_alloc = pcpu_need_to_extend(chunk, false); | |
464 | spin_unlock_irq(&pcpu_lock); | |
465 | ||
466 | if (new_alloc) | |
467 | pcpu_extend_area_map(chunk, new_alloc); | |
468 | } | |
469 | ||
a16037c8 TH |
470 | /** |
471 | * pcpu_fit_in_area - try to fit the requested allocation in a candidate area | |
472 | * @chunk: chunk the candidate area belongs to | |
473 | * @off: the offset to the start of the candidate area | |
474 | * @this_size: the size of the candidate area | |
475 | * @size: the size of the target allocation | |
476 | * @align: the alignment of the target allocation | |
477 | * @pop_only: only allocate from already populated region | |
478 | * | |
479 | * We're trying to allocate @size bytes aligned at @align. @chunk's area | |
480 | * at @off sized @this_size is a candidate. This function determines | |
481 | * whether the target allocation fits in the candidate area and returns the | |
482 | * number of bytes to pad after @off. If the target area doesn't fit, -1 | |
483 | * is returned. | |
484 | * | |
485 | * If @pop_only is %true, this function only considers the already | |
486 | * populated part of the candidate area. | |
487 | */ | |
488 | static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size, | |
489 | int size, int align, bool pop_only) | |
490 | { | |
491 | int cand_off = off; | |
492 | ||
493 | while (true) { | |
494 | int head = ALIGN(cand_off, align) - off; | |
495 | int page_start, page_end, rs, re; | |
496 | ||
497 | if (this_size < head + size) | |
498 | return -1; | |
499 | ||
500 | if (!pop_only) | |
501 | return head; | |
502 | ||
503 | /* | |
504 | * If the first unpopulated page is beyond the end of the | |
505 | * allocation, the whole allocation is populated; | |
506 | * otherwise, retry from the end of the unpopulated area. | |
507 | */ | |
508 | page_start = PFN_DOWN(head + off); | |
509 | page_end = PFN_UP(head + off + size); | |
510 | ||
511 | rs = page_start; | |
512 | pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size)); | |
513 | if (rs >= page_end) | |
514 | return head; | |
515 | cand_off = re * PAGE_SIZE; | |
516 | } | |
517 | } | |
518 | ||
fbf59bc9 TH |
519 | /** |
520 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
521 | * @chunk: chunk of interest | |
cae3aeb8 | 522 | * @size: wanted size in bytes |
fbf59bc9 | 523 | * @align: wanted align |
a16037c8 | 524 | * @pop_only: allocate only from the populated area |
b539b87f | 525 | * @occ_pages_p: out param for the number of pages the area occupies |
fbf59bc9 TH |
526 | * |
527 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
528 | * Note that this function only allocates the offset. It doesn't | |
529 | * populate or map the area. | |
530 | * | |
9f7dcf22 TH |
531 | * @chunk->map must have at least two free slots. |
532 | * | |
ccea34b5 TH |
533 | * CONTEXT: |
534 | * pcpu_lock. | |
535 | * | |
fbf59bc9 | 536 | * RETURNS: |
9f7dcf22 TH |
537 | * Allocated offset in @chunk on success, -1 if no matching area is |
538 | * found. | |
fbf59bc9 | 539 | */ |
a16037c8 | 540 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align, |
b539b87f | 541 | bool pop_only, int *occ_pages_p) |
fbf59bc9 TH |
542 | { |
543 | int oslot = pcpu_chunk_slot(chunk); | |
544 | int max_contig = 0; | |
545 | int i, off; | |
3d331ad7 | 546 | bool seen_free = false; |
723ad1d9 | 547 | int *p; |
fbf59bc9 | 548 | |
3d331ad7 | 549 | for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) { |
fbf59bc9 | 550 | int head, tail; |
723ad1d9 AV |
551 | int this_size; |
552 | ||
553 | off = *p; | |
554 | if (off & 1) | |
555 | continue; | |
fbf59bc9 | 556 | |
723ad1d9 | 557 | this_size = (p[1] & ~1) - off; |
a16037c8 TH |
558 | |
559 | head = pcpu_fit_in_area(chunk, off, this_size, size, align, | |
560 | pop_only); | |
561 | if (head < 0) { | |
3d331ad7 AV |
562 | if (!seen_free) { |
563 | chunk->first_free = i; | |
564 | seen_free = true; | |
565 | } | |
723ad1d9 | 566 | max_contig = max(this_size, max_contig); |
fbf59bc9 TH |
567 | continue; |
568 | } | |
569 | ||
570 | /* | |
571 | * If head is small or the previous block is free, | |
572 | * merge'em. Note that 'small' is defined as smaller | |
573 | * than sizeof(int), which is very small but isn't too | |
574 | * uncommon for percpu allocations. | |
575 | */ | |
723ad1d9 | 576 | if (head && (head < sizeof(int) || !(p[-1] & 1))) { |
21ddfd38 | 577 | *p = off += head; |
723ad1d9 | 578 | if (p[-1] & 1) |
fbf59bc9 | 579 | chunk->free_size -= head; |
21ddfd38 JZ |
580 | else |
581 | max_contig = max(*p - p[-1], max_contig); | |
723ad1d9 | 582 | this_size -= head; |
fbf59bc9 TH |
583 | head = 0; |
584 | } | |
585 | ||
586 | /* if tail is small, just keep it around */ | |
723ad1d9 AV |
587 | tail = this_size - head - size; |
588 | if (tail < sizeof(int)) { | |
fbf59bc9 | 589 | tail = 0; |
723ad1d9 AV |
590 | size = this_size - head; |
591 | } | |
fbf59bc9 TH |
592 | |
593 | /* split if warranted */ | |
594 | if (head || tail) { | |
706c16f2 AV |
595 | int nr_extra = !!head + !!tail; |
596 | ||
597 | /* insert new subblocks */ | |
723ad1d9 | 598 | memmove(p + nr_extra + 1, p + 1, |
706c16f2 AV |
599 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
600 | chunk->map_used += nr_extra; | |
601 | ||
fbf59bc9 | 602 | if (head) { |
3d331ad7 AV |
603 | if (!seen_free) { |
604 | chunk->first_free = i; | |
605 | seen_free = true; | |
606 | } | |
723ad1d9 AV |
607 | *++p = off += head; |
608 | ++i; | |
706c16f2 AV |
609 | max_contig = max(head, max_contig); |
610 | } | |
611 | if (tail) { | |
723ad1d9 | 612 | p[1] = off + size; |
706c16f2 | 613 | max_contig = max(tail, max_contig); |
fbf59bc9 | 614 | } |
fbf59bc9 TH |
615 | } |
616 | ||
3d331ad7 AV |
617 | if (!seen_free) |
618 | chunk->first_free = i + 1; | |
619 | ||
fbf59bc9 | 620 | /* update hint and mark allocated */ |
723ad1d9 | 621 | if (i + 1 == chunk->map_used) |
fbf59bc9 TH |
622 | chunk->contig_hint = max_contig; /* fully scanned */ |
623 | else | |
624 | chunk->contig_hint = max(chunk->contig_hint, | |
625 | max_contig); | |
626 | ||
723ad1d9 AV |
627 | chunk->free_size -= size; |
628 | *p |= 1; | |
fbf59bc9 | 629 | |
b539b87f | 630 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
fbf59bc9 TH |
631 | pcpu_chunk_relocate(chunk, oslot); |
632 | return off; | |
633 | } | |
634 | ||
635 | chunk->contig_hint = max_contig; /* fully scanned */ | |
636 | pcpu_chunk_relocate(chunk, oslot); | |
637 | ||
9f7dcf22 TH |
638 | /* tell the upper layer that this chunk has no matching area */ |
639 | return -1; | |
fbf59bc9 TH |
640 | } |
641 | ||
642 | /** | |
643 | * pcpu_free_area - free area to a pcpu_chunk | |
644 | * @chunk: chunk of interest | |
645 | * @freeme: offset of area to free | |
b539b87f | 646 | * @occ_pages_p: out param for the number of pages the area occupies |
fbf59bc9 TH |
647 | * |
648 | * Free area starting from @freeme to @chunk. Note that this function | |
649 | * only modifies the allocation map. It doesn't depopulate or unmap | |
650 | * the area. | |
ccea34b5 TH |
651 | * |
652 | * CONTEXT: | |
653 | * pcpu_lock. | |
fbf59bc9 | 654 | */ |
b539b87f TH |
655 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme, |
656 | int *occ_pages_p) | |
fbf59bc9 TH |
657 | { |
658 | int oslot = pcpu_chunk_slot(chunk); | |
723ad1d9 AV |
659 | int off = 0; |
660 | unsigned i, j; | |
661 | int to_free = 0; | |
662 | int *p; | |
663 | ||
664 | freeme |= 1; /* we are searching for <given offset, in use> pair */ | |
665 | ||
666 | i = 0; | |
667 | j = chunk->map_used; | |
668 | while (i != j) { | |
669 | unsigned k = (i + j) / 2; | |
670 | off = chunk->map[k]; | |
671 | if (off < freeme) | |
672 | i = k + 1; | |
673 | else if (off > freeme) | |
674 | j = k; | |
675 | else | |
676 | i = j = k; | |
677 | } | |
fbf59bc9 | 678 | BUG_ON(off != freeme); |
fbf59bc9 | 679 | |
3d331ad7 AV |
680 | if (i < chunk->first_free) |
681 | chunk->first_free = i; | |
682 | ||
723ad1d9 AV |
683 | p = chunk->map + i; |
684 | *p = off &= ~1; | |
685 | chunk->free_size += (p[1] & ~1) - off; | |
fbf59bc9 | 686 | |
b539b87f TH |
687 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
688 | ||
723ad1d9 AV |
689 | /* merge with next? */ |
690 | if (!(p[1] & 1)) | |
691 | to_free++; | |
fbf59bc9 | 692 | /* merge with previous? */ |
723ad1d9 AV |
693 | if (i > 0 && !(p[-1] & 1)) { |
694 | to_free++; | |
fbf59bc9 | 695 | i--; |
723ad1d9 | 696 | p--; |
fbf59bc9 | 697 | } |
723ad1d9 AV |
698 | if (to_free) { |
699 | chunk->map_used -= to_free; | |
700 | memmove(p + 1, p + 1 + to_free, | |
701 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
fbf59bc9 TH |
702 | } |
703 | ||
723ad1d9 | 704 | chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint); |
fbf59bc9 TH |
705 | pcpu_chunk_relocate(chunk, oslot); |
706 | } | |
707 | ||
6081089f TH |
708 | static struct pcpu_chunk *pcpu_alloc_chunk(void) |
709 | { | |
710 | struct pcpu_chunk *chunk; | |
711 | ||
90459ce0 | 712 | chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); |
6081089f TH |
713 | if (!chunk) |
714 | return NULL; | |
715 | ||
90459ce0 BL |
716 | chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC * |
717 | sizeof(chunk->map[0])); | |
6081089f | 718 | if (!chunk->map) { |
5a838c3b | 719 | pcpu_mem_free(chunk, pcpu_chunk_struct_size); |
6081089f TH |
720 | return NULL; |
721 | } | |
722 | ||
723 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; | |
723ad1d9 AV |
724 | chunk->map[0] = 0; |
725 | chunk->map[1] = pcpu_unit_size | 1; | |
726 | chunk->map_used = 1; | |
6081089f TH |
727 | |
728 | INIT_LIST_HEAD(&chunk->list); | |
9c824b6a | 729 | INIT_WORK(&chunk->map_extend_work, pcpu_map_extend_workfn); |
6081089f TH |
730 | chunk->free_size = pcpu_unit_size; |
731 | chunk->contig_hint = pcpu_unit_size; | |
732 | ||
733 | return chunk; | |
734 | } | |
735 | ||
736 | static void pcpu_free_chunk(struct pcpu_chunk *chunk) | |
737 | { | |
738 | if (!chunk) | |
739 | return; | |
740 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); | |
b4916cb1 | 741 | pcpu_mem_free(chunk, pcpu_chunk_struct_size); |
6081089f TH |
742 | } |
743 | ||
b539b87f TH |
744 | /** |
745 | * pcpu_chunk_populated - post-population bookkeeping | |
746 | * @chunk: pcpu_chunk which got populated | |
747 | * @page_start: the start page | |
748 | * @page_end: the end page | |
749 | * | |
750 | * Pages in [@page_start,@page_end) have been populated to @chunk. Update | |
751 | * the bookkeeping information accordingly. Must be called after each | |
752 | * successful population. | |
753 | */ | |
754 | static void pcpu_chunk_populated(struct pcpu_chunk *chunk, | |
755 | int page_start, int page_end) | |
756 | { | |
757 | int nr = page_end - page_start; | |
758 | ||
759 | lockdep_assert_held(&pcpu_lock); | |
760 | ||
761 | bitmap_set(chunk->populated, page_start, nr); | |
762 | chunk->nr_populated += nr; | |
763 | pcpu_nr_empty_pop_pages += nr; | |
764 | } | |
765 | ||
766 | /** | |
767 | * pcpu_chunk_depopulated - post-depopulation bookkeeping | |
768 | * @chunk: pcpu_chunk which got depopulated | |
769 | * @page_start: the start page | |
770 | * @page_end: the end page | |
771 | * | |
772 | * Pages in [@page_start,@page_end) have been depopulated from @chunk. | |
773 | * Update the bookkeeping information accordingly. Must be called after | |
774 | * each successful depopulation. | |
775 | */ | |
776 | static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, | |
777 | int page_start, int page_end) | |
778 | { | |
779 | int nr = page_end - page_start; | |
780 | ||
781 | lockdep_assert_held(&pcpu_lock); | |
782 | ||
783 | bitmap_clear(chunk->populated, page_start, nr); | |
784 | chunk->nr_populated -= nr; | |
785 | pcpu_nr_empty_pop_pages -= nr; | |
786 | } | |
787 | ||
9f645532 TH |
788 | /* |
789 | * Chunk management implementation. | |
790 | * | |
791 | * To allow different implementations, chunk alloc/free and | |
792 | * [de]population are implemented in a separate file which is pulled | |
793 | * into this file and compiled together. The following functions | |
794 | * should be implemented. | |
795 | * | |
796 | * pcpu_populate_chunk - populate the specified range of a chunk | |
797 | * pcpu_depopulate_chunk - depopulate the specified range of a chunk | |
798 | * pcpu_create_chunk - create a new chunk | |
799 | * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop | |
800 | * pcpu_addr_to_page - translate address to physical address | |
801 | * pcpu_verify_alloc_info - check alloc_info is acceptable during init | |
fbf59bc9 | 802 | */ |
9f645532 TH |
803 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); |
804 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); | |
805 | static struct pcpu_chunk *pcpu_create_chunk(void); | |
806 | static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); | |
807 | static struct page *pcpu_addr_to_page(void *addr); | |
808 | static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); | |
fbf59bc9 | 809 | |
b0c9778b TH |
810 | #ifdef CONFIG_NEED_PER_CPU_KM |
811 | #include "percpu-km.c" | |
812 | #else | |
9f645532 | 813 | #include "percpu-vm.c" |
b0c9778b | 814 | #endif |
fbf59bc9 | 815 | |
88999a89 TH |
816 | /** |
817 | * pcpu_chunk_addr_search - determine chunk containing specified address | |
818 | * @addr: address for which the chunk needs to be determined. | |
819 | * | |
820 | * RETURNS: | |
821 | * The address of the found chunk. | |
822 | */ | |
823 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
824 | { | |
825 | /* is it in the first chunk? */ | |
826 | if (pcpu_addr_in_first_chunk(addr)) { | |
827 | /* is it in the reserved area? */ | |
828 | if (pcpu_addr_in_reserved_chunk(addr)) | |
829 | return pcpu_reserved_chunk; | |
830 | return pcpu_first_chunk; | |
831 | } | |
832 | ||
833 | /* | |
834 | * The address is relative to unit0 which might be unused and | |
835 | * thus unmapped. Offset the address to the unit space of the | |
836 | * current processor before looking it up in the vmalloc | |
837 | * space. Note that any possible cpu id can be used here, so | |
838 | * there's no need to worry about preemption or cpu hotplug. | |
839 | */ | |
840 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; | |
9f645532 | 841 | return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); |
88999a89 TH |
842 | } |
843 | ||
fbf59bc9 | 844 | /** |
edcb4639 | 845 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 846 | * @size: size of area to allocate in bytes |
fbf59bc9 | 847 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 848 | * @reserved: allocate from the reserved chunk if available |
5835d96e | 849 | * @gfp: allocation flags |
fbf59bc9 | 850 | * |
5835d96e TH |
851 | * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't |
852 | * contain %GFP_KERNEL, the allocation is atomic. | |
fbf59bc9 TH |
853 | * |
854 | * RETURNS: | |
855 | * Percpu pointer to the allocated area on success, NULL on failure. | |
856 | */ | |
5835d96e TH |
857 | static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, |
858 | gfp_t gfp) | |
fbf59bc9 | 859 | { |
f2badb0c | 860 | static int warn_limit = 10; |
fbf59bc9 | 861 | struct pcpu_chunk *chunk; |
f2badb0c | 862 | const char *err; |
5835d96e | 863 | bool is_atomic = !(gfp & GFP_KERNEL); |
b539b87f | 864 | int occ_pages = 0; |
b38d08f3 | 865 | int slot, off, new_alloc, cpu, ret; |
403a91b1 | 866 | unsigned long flags; |
f528f0b8 | 867 | void __percpu *ptr; |
fbf59bc9 | 868 | |
723ad1d9 AV |
869 | /* |
870 | * We want the lowest bit of offset available for in-use/free | |
2f69fa82 | 871 | * indicator, so force >= 16bit alignment and make size even. |
723ad1d9 AV |
872 | */ |
873 | if (unlikely(align < 2)) | |
874 | align = 2; | |
875 | ||
fb009e3a | 876 | size = ALIGN(size, 2); |
2f69fa82 | 877 | |
8d408b4b | 878 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
fbf59bc9 TH |
879 | WARN(true, "illegal size (%zu) or align (%zu) for " |
880 | "percpu allocation\n", size, align); | |
881 | return NULL; | |
882 | } | |
883 | ||
403a91b1 | 884 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 | 885 | |
edcb4639 TH |
886 | /* serve reserved allocations from the reserved chunk if available */ |
887 | if (reserved && pcpu_reserved_chunk) { | |
888 | chunk = pcpu_reserved_chunk; | |
833af842 TH |
889 | |
890 | if (size > chunk->contig_hint) { | |
891 | err = "alloc from reserved chunk failed"; | |
ccea34b5 | 892 | goto fail_unlock; |
f2badb0c | 893 | } |
833af842 | 894 | |
9c824b6a | 895 | while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) { |
833af842 | 896 | spin_unlock_irqrestore(&pcpu_lock, flags); |
5835d96e TH |
897 | if (is_atomic || |
898 | pcpu_extend_area_map(chunk, new_alloc) < 0) { | |
833af842 | 899 | err = "failed to extend area map of reserved chunk"; |
b38d08f3 | 900 | goto fail; |
833af842 TH |
901 | } |
902 | spin_lock_irqsave(&pcpu_lock, flags); | |
903 | } | |
904 | ||
b539b87f TH |
905 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
906 | &occ_pages); | |
edcb4639 TH |
907 | if (off >= 0) |
908 | goto area_found; | |
833af842 | 909 | |
f2badb0c | 910 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 911 | goto fail_unlock; |
edcb4639 TH |
912 | } |
913 | ||
ccea34b5 | 914 | restart: |
edcb4639 | 915 | /* search through normal chunks */ |
fbf59bc9 TH |
916 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
917 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
918 | if (size > chunk->contig_hint) | |
919 | continue; | |
ccea34b5 | 920 | |
9c824b6a | 921 | new_alloc = pcpu_need_to_extend(chunk, is_atomic); |
833af842 | 922 | if (new_alloc) { |
5835d96e TH |
923 | if (is_atomic) |
924 | continue; | |
833af842 TH |
925 | spin_unlock_irqrestore(&pcpu_lock, flags); |
926 | if (pcpu_extend_area_map(chunk, | |
927 | new_alloc) < 0) { | |
928 | err = "failed to extend area map"; | |
b38d08f3 | 929 | goto fail; |
833af842 TH |
930 | } |
931 | spin_lock_irqsave(&pcpu_lock, flags); | |
932 | /* | |
933 | * pcpu_lock has been dropped, need to | |
934 | * restart cpu_slot list walking. | |
935 | */ | |
936 | goto restart; | |
ccea34b5 TH |
937 | } |
938 | ||
b539b87f TH |
939 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
940 | &occ_pages); | |
fbf59bc9 TH |
941 | if (off >= 0) |
942 | goto area_found; | |
fbf59bc9 TH |
943 | } |
944 | } | |
945 | ||
403a91b1 | 946 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 947 | |
b38d08f3 TH |
948 | /* |
949 | * No space left. Create a new chunk. We don't want multiple | |
950 | * tasks to create chunks simultaneously. Serialize and create iff | |
951 | * there's still no empty chunk after grabbing the mutex. | |
952 | */ | |
5835d96e TH |
953 | if (is_atomic) |
954 | goto fail; | |
955 | ||
b38d08f3 TH |
956 | mutex_lock(&pcpu_alloc_mutex); |
957 | ||
958 | if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { | |
959 | chunk = pcpu_create_chunk(); | |
960 | if (!chunk) { | |
961 | err = "failed to allocate new chunk"; | |
962 | goto fail; | |
963 | } | |
964 | ||
965 | spin_lock_irqsave(&pcpu_lock, flags); | |
966 | pcpu_chunk_relocate(chunk, -1); | |
967 | } else { | |
968 | spin_lock_irqsave(&pcpu_lock, flags); | |
f2badb0c | 969 | } |
ccea34b5 | 970 | |
b38d08f3 | 971 | mutex_unlock(&pcpu_alloc_mutex); |
ccea34b5 | 972 | goto restart; |
fbf59bc9 TH |
973 | |
974 | area_found: | |
403a91b1 | 975 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 976 | |
dca49645 | 977 | /* populate if not all pages are already there */ |
5835d96e | 978 | if (!is_atomic) { |
e04d3208 | 979 | int page_start, page_end, rs, re; |
dca49645 | 980 | |
e04d3208 | 981 | mutex_lock(&pcpu_alloc_mutex); |
dca49645 | 982 | |
e04d3208 TH |
983 | page_start = PFN_DOWN(off); |
984 | page_end = PFN_UP(off + size); | |
b38d08f3 | 985 | |
e04d3208 TH |
986 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
987 | WARN_ON(chunk->immutable); | |
988 | ||
989 | ret = pcpu_populate_chunk(chunk, rs, re); | |
990 | ||
991 | spin_lock_irqsave(&pcpu_lock, flags); | |
992 | if (ret) { | |
993 | mutex_unlock(&pcpu_alloc_mutex); | |
b539b87f | 994 | pcpu_free_area(chunk, off, &occ_pages); |
e04d3208 TH |
995 | err = "failed to populate"; |
996 | goto fail_unlock; | |
997 | } | |
b539b87f | 998 | pcpu_chunk_populated(chunk, rs, re); |
e04d3208 | 999 | spin_unlock_irqrestore(&pcpu_lock, flags); |
dca49645 | 1000 | } |
fbf59bc9 | 1001 | |
e04d3208 TH |
1002 | mutex_unlock(&pcpu_alloc_mutex); |
1003 | } | |
ccea34b5 | 1004 | |
b539b87f TH |
1005 | if (chunk != pcpu_reserved_chunk) |
1006 | pcpu_nr_empty_pop_pages -= occ_pages; | |
1007 | ||
dca49645 TH |
1008 | /* clear the areas and return address relative to base address */ |
1009 | for_each_possible_cpu(cpu) | |
1010 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); | |
1011 | ||
f528f0b8 CM |
1012 | ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); |
1013 | kmemleak_alloc_percpu(ptr, size); | |
1014 | return ptr; | |
ccea34b5 TH |
1015 | |
1016 | fail_unlock: | |
403a91b1 | 1017 | spin_unlock_irqrestore(&pcpu_lock, flags); |
b38d08f3 | 1018 | fail: |
5835d96e TH |
1019 | if (!is_atomic && warn_limit) { |
1020 | pr_warning("PERCPU: allocation failed, size=%zu align=%zu atomic=%d, %s\n", | |
1021 | size, align, is_atomic, err); | |
f2badb0c TH |
1022 | dump_stack(); |
1023 | if (!--warn_limit) | |
1024 | pr_info("PERCPU: limit reached, disable warning\n"); | |
1025 | } | |
ccea34b5 | 1026 | return NULL; |
fbf59bc9 | 1027 | } |
edcb4639 TH |
1028 | |
1029 | /** | |
5835d96e | 1030 | * __alloc_percpu_gfp - allocate dynamic percpu area |
edcb4639 TH |
1031 | * @size: size of area to allocate in bytes |
1032 | * @align: alignment of area (max PAGE_SIZE) | |
5835d96e | 1033 | * @gfp: allocation flags |
edcb4639 | 1034 | * |
5835d96e TH |
1035 | * Allocate zero-filled percpu area of @size bytes aligned at @align. If |
1036 | * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can | |
1037 | * be called from any context but is a lot more likely to fail. | |
ccea34b5 | 1038 | * |
edcb4639 TH |
1039 | * RETURNS: |
1040 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1041 | */ | |
5835d96e TH |
1042 | void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) |
1043 | { | |
1044 | return pcpu_alloc(size, align, false, gfp); | |
1045 | } | |
1046 | EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); | |
1047 | ||
1048 | /** | |
1049 | * __alloc_percpu - allocate dynamic percpu area | |
1050 | * @size: size of area to allocate in bytes | |
1051 | * @align: alignment of area (max PAGE_SIZE) | |
1052 | * | |
1053 | * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). | |
1054 | */ | |
43cf38eb | 1055 | void __percpu *__alloc_percpu(size_t size, size_t align) |
edcb4639 | 1056 | { |
5835d96e | 1057 | return pcpu_alloc(size, align, false, GFP_KERNEL); |
edcb4639 | 1058 | } |
fbf59bc9 TH |
1059 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1060 | ||
edcb4639 TH |
1061 | /** |
1062 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1063 | * @size: size of area to allocate in bytes | |
1064 | * @align: alignment of area (max PAGE_SIZE) | |
1065 | * | |
9329ba97 TH |
1066 | * Allocate zero-filled percpu area of @size bytes aligned at @align |
1067 | * from reserved percpu area if arch has set it up; otherwise, | |
1068 | * allocation is served from the same dynamic area. Might sleep. | |
1069 | * Might trigger writeouts. | |
edcb4639 | 1070 | * |
ccea34b5 TH |
1071 | * CONTEXT: |
1072 | * Does GFP_KERNEL allocation. | |
1073 | * | |
edcb4639 TH |
1074 | * RETURNS: |
1075 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1076 | */ | |
43cf38eb | 1077 | void __percpu *__alloc_reserved_percpu(size_t size, size_t align) |
edcb4639 | 1078 | { |
5835d96e | 1079 | return pcpu_alloc(size, align, true, GFP_KERNEL); |
edcb4639 TH |
1080 | } |
1081 | ||
a56dbddf TH |
1082 | /** |
1083 | * pcpu_reclaim - reclaim fully free chunks, workqueue function | |
1084 | * @work: unused | |
1085 | * | |
1086 | * Reclaim all fully free chunks except for the first one. | |
ccea34b5 TH |
1087 | * |
1088 | * CONTEXT: | |
1089 | * workqueue context. | |
a56dbddf TH |
1090 | */ |
1091 | static void pcpu_reclaim(struct work_struct *work) | |
fbf59bc9 | 1092 | { |
a56dbddf TH |
1093 | LIST_HEAD(todo); |
1094 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; | |
1095 | struct pcpu_chunk *chunk, *next; | |
1096 | ||
ccea34b5 TH |
1097 | mutex_lock(&pcpu_alloc_mutex); |
1098 | spin_lock_irq(&pcpu_lock); | |
a56dbddf TH |
1099 | |
1100 | list_for_each_entry_safe(chunk, next, head, list) { | |
1101 | WARN_ON(chunk->immutable); | |
1102 | ||
1103 | /* spare the first one */ | |
1104 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) | |
1105 | continue; | |
1106 | ||
a56dbddf TH |
1107 | list_move(&chunk->list, &todo); |
1108 | } | |
1109 | ||
ccea34b5 | 1110 | spin_unlock_irq(&pcpu_lock); |
a56dbddf TH |
1111 | |
1112 | list_for_each_entry_safe(chunk, next, &todo, list) { | |
a93ace48 | 1113 | int rs, re; |
dca49645 | 1114 | |
a93ace48 TH |
1115 | pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) { |
1116 | pcpu_depopulate_chunk(chunk, rs, re); | |
b539b87f TH |
1117 | spin_lock_irq(&pcpu_lock); |
1118 | pcpu_chunk_depopulated(chunk, rs, re); | |
1119 | spin_unlock_irq(&pcpu_lock); | |
a93ace48 | 1120 | } |
6081089f | 1121 | pcpu_destroy_chunk(chunk); |
a56dbddf | 1122 | } |
971f3918 TH |
1123 | |
1124 | mutex_unlock(&pcpu_alloc_mutex); | |
fbf59bc9 TH |
1125 | } |
1126 | ||
1127 | /** | |
1128 | * free_percpu - free percpu area | |
1129 | * @ptr: pointer to area to free | |
1130 | * | |
ccea34b5 TH |
1131 | * Free percpu area @ptr. |
1132 | * | |
1133 | * CONTEXT: | |
1134 | * Can be called from atomic context. | |
fbf59bc9 | 1135 | */ |
43cf38eb | 1136 | void free_percpu(void __percpu *ptr) |
fbf59bc9 | 1137 | { |
129182e5 | 1138 | void *addr; |
fbf59bc9 | 1139 | struct pcpu_chunk *chunk; |
ccea34b5 | 1140 | unsigned long flags; |
b539b87f | 1141 | int off, occ_pages; |
fbf59bc9 TH |
1142 | |
1143 | if (!ptr) | |
1144 | return; | |
1145 | ||
f528f0b8 CM |
1146 | kmemleak_free_percpu(ptr); |
1147 | ||
129182e5 AM |
1148 | addr = __pcpu_ptr_to_addr(ptr); |
1149 | ||
ccea34b5 | 1150 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1151 | |
1152 | chunk = pcpu_chunk_addr_search(addr); | |
bba174f5 | 1153 | off = addr - chunk->base_addr; |
fbf59bc9 | 1154 | |
b539b87f TH |
1155 | pcpu_free_area(chunk, off, &occ_pages); |
1156 | ||
1157 | if (chunk != pcpu_reserved_chunk) | |
1158 | pcpu_nr_empty_pop_pages += occ_pages; | |
fbf59bc9 | 1159 | |
a56dbddf | 1160 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
1161 | if (chunk->free_size == pcpu_unit_size) { |
1162 | struct pcpu_chunk *pos; | |
1163 | ||
a56dbddf | 1164 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1165 | if (pos != chunk) { |
a56dbddf | 1166 | schedule_work(&pcpu_reclaim_work); |
fbf59bc9 TH |
1167 | break; |
1168 | } | |
1169 | } | |
1170 | ||
ccea34b5 | 1171 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1172 | } |
1173 | EXPORT_SYMBOL_GPL(free_percpu); | |
1174 | ||
10fad5e4 TH |
1175 | /** |
1176 | * is_kernel_percpu_address - test whether address is from static percpu area | |
1177 | * @addr: address to test | |
1178 | * | |
1179 | * Test whether @addr belongs to in-kernel static percpu area. Module | |
1180 | * static percpu areas are not considered. For those, use | |
1181 | * is_module_percpu_address(). | |
1182 | * | |
1183 | * RETURNS: | |
1184 | * %true if @addr is from in-kernel static percpu area, %false otherwise. | |
1185 | */ | |
1186 | bool is_kernel_percpu_address(unsigned long addr) | |
1187 | { | |
bbddff05 | 1188 | #ifdef CONFIG_SMP |
10fad5e4 TH |
1189 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
1190 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); | |
1191 | unsigned int cpu; | |
1192 | ||
1193 | for_each_possible_cpu(cpu) { | |
1194 | void *start = per_cpu_ptr(base, cpu); | |
1195 | ||
1196 | if ((void *)addr >= start && (void *)addr < start + static_size) | |
1197 | return true; | |
1198 | } | |
bbddff05 TH |
1199 | #endif |
1200 | /* on UP, can't distinguish from other static vars, always false */ | |
10fad5e4 TH |
1201 | return false; |
1202 | } | |
1203 | ||
3b034b0d VG |
1204 | /** |
1205 | * per_cpu_ptr_to_phys - convert translated percpu address to physical address | |
1206 | * @addr: the address to be converted to physical address | |
1207 | * | |
1208 | * Given @addr which is dereferenceable address obtained via one of | |
1209 | * percpu access macros, this function translates it into its physical | |
1210 | * address. The caller is responsible for ensuring @addr stays valid | |
1211 | * until this function finishes. | |
1212 | * | |
67589c71 DY |
1213 | * percpu allocator has special setup for the first chunk, which currently |
1214 | * supports either embedding in linear address space or vmalloc mapping, | |
1215 | * and, from the second one, the backing allocator (currently either vm or | |
1216 | * km) provides translation. | |
1217 | * | |
1218 | * The addr can be tranlated simply without checking if it falls into the | |
1219 | * first chunk. But the current code reflects better how percpu allocator | |
1220 | * actually works, and the verification can discover both bugs in percpu | |
1221 | * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current | |
1222 | * code. | |
1223 | * | |
3b034b0d VG |
1224 | * RETURNS: |
1225 | * The physical address for @addr. | |
1226 | */ | |
1227 | phys_addr_t per_cpu_ptr_to_phys(void *addr) | |
1228 | { | |
9983b6f0 TH |
1229 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
1230 | bool in_first_chunk = false; | |
a855b84c | 1231 | unsigned long first_low, first_high; |
9983b6f0 TH |
1232 | unsigned int cpu; |
1233 | ||
1234 | /* | |
a855b84c | 1235 | * The following test on unit_low/high isn't strictly |
9983b6f0 TH |
1236 | * necessary but will speed up lookups of addresses which |
1237 | * aren't in the first chunk. | |
1238 | */ | |
a855b84c TH |
1239 | first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0); |
1240 | first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu, | |
1241 | pcpu_unit_pages); | |
1242 | if ((unsigned long)addr >= first_low && | |
1243 | (unsigned long)addr < first_high) { | |
9983b6f0 TH |
1244 | for_each_possible_cpu(cpu) { |
1245 | void *start = per_cpu_ptr(base, cpu); | |
1246 | ||
1247 | if (addr >= start && addr < start + pcpu_unit_size) { | |
1248 | in_first_chunk = true; | |
1249 | break; | |
1250 | } | |
1251 | } | |
1252 | } | |
1253 | ||
1254 | if (in_first_chunk) { | |
eac522ef | 1255 | if (!is_vmalloc_addr(addr)) |
020ec653 TH |
1256 | return __pa(addr); |
1257 | else | |
9f57bd4d ES |
1258 | return page_to_phys(vmalloc_to_page(addr)) + |
1259 | offset_in_page(addr); | |
020ec653 | 1260 | } else |
9f57bd4d ES |
1261 | return page_to_phys(pcpu_addr_to_page(addr)) + |
1262 | offset_in_page(addr); | |
3b034b0d VG |
1263 | } |
1264 | ||
fbf59bc9 | 1265 | /** |
fd1e8a1f TH |
1266 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
1267 | * @nr_groups: the number of groups | |
1268 | * @nr_units: the number of units | |
1269 | * | |
1270 | * Allocate ai which is large enough for @nr_groups groups containing | |
1271 | * @nr_units units. The returned ai's groups[0].cpu_map points to the | |
1272 | * cpu_map array which is long enough for @nr_units and filled with | |
1273 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map | |
1274 | * pointer of other groups. | |
1275 | * | |
1276 | * RETURNS: | |
1277 | * Pointer to the allocated pcpu_alloc_info on success, NULL on | |
1278 | * failure. | |
1279 | */ | |
1280 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, | |
1281 | int nr_units) | |
1282 | { | |
1283 | struct pcpu_alloc_info *ai; | |
1284 | size_t base_size, ai_size; | |
1285 | void *ptr; | |
1286 | int unit; | |
1287 | ||
1288 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), | |
1289 | __alignof__(ai->groups[0].cpu_map[0])); | |
1290 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); | |
1291 | ||
999c17e3 | 1292 | ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); |
fd1e8a1f TH |
1293 | if (!ptr) |
1294 | return NULL; | |
1295 | ai = ptr; | |
1296 | ptr += base_size; | |
1297 | ||
1298 | ai->groups[0].cpu_map = ptr; | |
1299 | ||
1300 | for (unit = 0; unit < nr_units; unit++) | |
1301 | ai->groups[0].cpu_map[unit] = NR_CPUS; | |
1302 | ||
1303 | ai->nr_groups = nr_groups; | |
1304 | ai->__ai_size = PFN_ALIGN(ai_size); | |
1305 | ||
1306 | return ai; | |
1307 | } | |
1308 | ||
1309 | /** | |
1310 | * pcpu_free_alloc_info - free percpu allocation info | |
1311 | * @ai: pcpu_alloc_info to free | |
1312 | * | |
1313 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). | |
1314 | */ | |
1315 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) | |
1316 | { | |
999c17e3 | 1317 | memblock_free_early(__pa(ai), ai->__ai_size); |
fd1e8a1f TH |
1318 | } |
1319 | ||
fd1e8a1f TH |
1320 | /** |
1321 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info | |
1322 | * @lvl: loglevel | |
1323 | * @ai: allocation info to dump | |
1324 | * | |
1325 | * Print out information about @ai using loglevel @lvl. | |
1326 | */ | |
1327 | static void pcpu_dump_alloc_info(const char *lvl, | |
1328 | const struct pcpu_alloc_info *ai) | |
033e48fb | 1329 | { |
fd1e8a1f | 1330 | int group_width = 1, cpu_width = 1, width; |
033e48fb | 1331 | char empty_str[] = "--------"; |
fd1e8a1f TH |
1332 | int alloc = 0, alloc_end = 0; |
1333 | int group, v; | |
1334 | int upa, apl; /* units per alloc, allocs per line */ | |
1335 | ||
1336 | v = ai->nr_groups; | |
1337 | while (v /= 10) | |
1338 | group_width++; | |
033e48fb | 1339 | |
fd1e8a1f | 1340 | v = num_possible_cpus(); |
033e48fb | 1341 | while (v /= 10) |
fd1e8a1f TH |
1342 | cpu_width++; |
1343 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; | |
033e48fb | 1344 | |
fd1e8a1f TH |
1345 | upa = ai->alloc_size / ai->unit_size; |
1346 | width = upa * (cpu_width + 1) + group_width + 3; | |
1347 | apl = rounddown_pow_of_two(max(60 / width, 1)); | |
033e48fb | 1348 | |
fd1e8a1f TH |
1349 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
1350 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, | |
1351 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); | |
033e48fb | 1352 | |
fd1e8a1f TH |
1353 | for (group = 0; group < ai->nr_groups; group++) { |
1354 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
1355 | int unit = 0, unit_end = 0; | |
1356 | ||
1357 | BUG_ON(gi->nr_units % upa); | |
1358 | for (alloc_end += gi->nr_units / upa; | |
1359 | alloc < alloc_end; alloc++) { | |
1360 | if (!(alloc % apl)) { | |
cb129820 | 1361 | printk(KERN_CONT "\n"); |
fd1e8a1f TH |
1362 | printk("%spcpu-alloc: ", lvl); |
1363 | } | |
cb129820 | 1364 | printk(KERN_CONT "[%0*d] ", group_width, group); |
fd1e8a1f TH |
1365 | |
1366 | for (unit_end += upa; unit < unit_end; unit++) | |
1367 | if (gi->cpu_map[unit] != NR_CPUS) | |
cb129820 | 1368 | printk(KERN_CONT "%0*d ", cpu_width, |
fd1e8a1f TH |
1369 | gi->cpu_map[unit]); |
1370 | else | |
cb129820 | 1371 | printk(KERN_CONT "%s ", empty_str); |
033e48fb | 1372 | } |
033e48fb | 1373 | } |
cb129820 | 1374 | printk(KERN_CONT "\n"); |
033e48fb | 1375 | } |
033e48fb | 1376 | |
fbf59bc9 | 1377 | /** |
8d408b4b | 1378 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
fd1e8a1f | 1379 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
38a6be52 | 1380 | * @base_addr: mapped address |
8d408b4b TH |
1381 | * |
1382 | * Initialize the first percpu chunk which contains the kernel static | |
1383 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1384 | * setup path. |
8d408b4b | 1385 | * |
fd1e8a1f TH |
1386 | * @ai contains all information necessary to initialize the first |
1387 | * chunk and prime the dynamic percpu allocator. | |
1388 | * | |
1389 | * @ai->static_size is the size of static percpu area. | |
1390 | * | |
1391 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to | |
edcb4639 TH |
1392 | * reserve after the static area in the first chunk. This reserves |
1393 | * the first chunk such that it's available only through reserved | |
1394 | * percpu allocation. This is primarily used to serve module percpu | |
1395 | * static areas on architectures where the addressing model has | |
1396 | * limited offset range for symbol relocations to guarantee module | |
1397 | * percpu symbols fall inside the relocatable range. | |
1398 | * | |
fd1e8a1f TH |
1399 | * @ai->dyn_size determines the number of bytes available for dynamic |
1400 | * allocation in the first chunk. The area between @ai->static_size + | |
1401 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. | |
6074d5b0 | 1402 | * |
fd1e8a1f TH |
1403 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
1404 | * and equal to or larger than @ai->static_size + @ai->reserved_size + | |
1405 | * @ai->dyn_size. | |
8d408b4b | 1406 | * |
fd1e8a1f TH |
1407 | * @ai->atom_size is the allocation atom size and used as alignment |
1408 | * for vm areas. | |
8d408b4b | 1409 | * |
fd1e8a1f TH |
1410 | * @ai->alloc_size is the allocation size and always multiple of |
1411 | * @ai->atom_size. This is larger than @ai->atom_size if | |
1412 | * @ai->unit_size is larger than @ai->atom_size. | |
1413 | * | |
1414 | * @ai->nr_groups and @ai->groups describe virtual memory layout of | |
1415 | * percpu areas. Units which should be colocated are put into the | |
1416 | * same group. Dynamic VM areas will be allocated according to these | |
1417 | * groupings. If @ai->nr_groups is zero, a single group containing | |
1418 | * all units is assumed. | |
8d408b4b | 1419 | * |
38a6be52 TH |
1420 | * The caller should have mapped the first chunk at @base_addr and |
1421 | * copied static data to each unit. | |
fbf59bc9 | 1422 | * |
edcb4639 TH |
1423 | * If the first chunk ends up with both reserved and dynamic areas, it |
1424 | * is served by two chunks - one to serve the core static and reserved | |
1425 | * areas and the other for the dynamic area. They share the same vm | |
1426 | * and page map but uses different area allocation map to stay away | |
1427 | * from each other. The latter chunk is circulated in the chunk slots | |
1428 | * and available for dynamic allocation like any other chunks. | |
1429 | * | |
fbf59bc9 | 1430 | * RETURNS: |
fb435d52 | 1431 | * 0 on success, -errno on failure. |
fbf59bc9 | 1432 | */ |
fb435d52 TH |
1433 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
1434 | void *base_addr) | |
fbf59bc9 | 1435 | { |
635b75fc | 1436 | static char cpus_buf[4096] __initdata; |
099a19d9 TH |
1437 | static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; |
1438 | static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; | |
fd1e8a1f TH |
1439 | size_t dyn_size = ai->dyn_size; |
1440 | size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; | |
edcb4639 | 1441 | struct pcpu_chunk *schunk, *dchunk = NULL; |
6563297c TH |
1442 | unsigned long *group_offsets; |
1443 | size_t *group_sizes; | |
fb435d52 | 1444 | unsigned long *unit_off; |
fbf59bc9 | 1445 | unsigned int cpu; |
fd1e8a1f TH |
1446 | int *unit_map; |
1447 | int group, unit, i; | |
fbf59bc9 | 1448 | |
635b75fc TH |
1449 | cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); |
1450 | ||
1451 | #define PCPU_SETUP_BUG_ON(cond) do { \ | |
1452 | if (unlikely(cond)) { \ | |
1453 | pr_emerg("PERCPU: failed to initialize, %s", #cond); \ | |
1454 | pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ | |
1455 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ | |
1456 | BUG(); \ | |
1457 | } \ | |
1458 | } while (0) | |
1459 | ||
2f39e637 | 1460 | /* sanity checks */ |
635b75fc | 1461 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
bbddff05 | 1462 | #ifdef CONFIG_SMP |
635b75fc | 1463 | PCPU_SETUP_BUG_ON(!ai->static_size); |
0415b00d | 1464 | PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK); |
bbddff05 | 1465 | #endif |
635b75fc | 1466 | PCPU_SETUP_BUG_ON(!base_addr); |
0415b00d | 1467 | PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK); |
635b75fc TH |
1468 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
1469 | PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); | |
1470 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); | |
099a19d9 | 1471 | PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); |
9f645532 | 1472 | PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); |
8d408b4b | 1473 | |
6563297c | 1474 | /* process group information and build config tables accordingly */ |
999c17e3 SS |
1475 | group_offsets = memblock_virt_alloc(ai->nr_groups * |
1476 | sizeof(group_offsets[0]), 0); | |
1477 | group_sizes = memblock_virt_alloc(ai->nr_groups * | |
1478 | sizeof(group_sizes[0]), 0); | |
1479 | unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); | |
1480 | unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); | |
2f39e637 | 1481 | |
fd1e8a1f | 1482 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
ffe0d5a5 | 1483 | unit_map[cpu] = UINT_MAX; |
a855b84c TH |
1484 | |
1485 | pcpu_low_unit_cpu = NR_CPUS; | |
1486 | pcpu_high_unit_cpu = NR_CPUS; | |
2f39e637 | 1487 | |
fd1e8a1f TH |
1488 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
1489 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
2f39e637 | 1490 | |
6563297c TH |
1491 | group_offsets[group] = gi->base_offset; |
1492 | group_sizes[group] = gi->nr_units * ai->unit_size; | |
1493 | ||
fd1e8a1f TH |
1494 | for (i = 0; i < gi->nr_units; i++) { |
1495 | cpu = gi->cpu_map[i]; | |
1496 | if (cpu == NR_CPUS) | |
1497 | continue; | |
8d408b4b | 1498 | |
635b75fc TH |
1499 | PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); |
1500 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); | |
1501 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); | |
fbf59bc9 | 1502 | |
fd1e8a1f | 1503 | unit_map[cpu] = unit + i; |
fb435d52 TH |
1504 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
1505 | ||
a855b84c TH |
1506 | /* determine low/high unit_cpu */ |
1507 | if (pcpu_low_unit_cpu == NR_CPUS || | |
1508 | unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) | |
1509 | pcpu_low_unit_cpu = cpu; | |
1510 | if (pcpu_high_unit_cpu == NR_CPUS || | |
1511 | unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) | |
1512 | pcpu_high_unit_cpu = cpu; | |
fd1e8a1f | 1513 | } |
2f39e637 | 1514 | } |
fd1e8a1f TH |
1515 | pcpu_nr_units = unit; |
1516 | ||
1517 | for_each_possible_cpu(cpu) | |
635b75fc TH |
1518 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
1519 | ||
1520 | /* we're done parsing the input, undefine BUG macro and dump config */ | |
1521 | #undef PCPU_SETUP_BUG_ON | |
bcbea798 | 1522 | pcpu_dump_alloc_info(KERN_DEBUG, ai); |
fd1e8a1f | 1523 | |
6563297c TH |
1524 | pcpu_nr_groups = ai->nr_groups; |
1525 | pcpu_group_offsets = group_offsets; | |
1526 | pcpu_group_sizes = group_sizes; | |
fd1e8a1f | 1527 | pcpu_unit_map = unit_map; |
fb435d52 | 1528 | pcpu_unit_offsets = unit_off; |
2f39e637 TH |
1529 | |
1530 | /* determine basic parameters */ | |
fd1e8a1f | 1531 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
d9b55eeb | 1532 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
6563297c | 1533 | pcpu_atom_size = ai->atom_size; |
ce3141a2 TH |
1534 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
1535 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
cafe8816 | 1536 | |
d9b55eeb TH |
1537 | /* |
1538 | * Allocate chunk slots. The additional last slot is for | |
1539 | * empty chunks. | |
1540 | */ | |
1541 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
999c17e3 SS |
1542 | pcpu_slot = memblock_virt_alloc( |
1543 | pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); | |
fbf59bc9 TH |
1544 | for (i = 0; i < pcpu_nr_slots; i++) |
1545 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1546 | ||
edcb4639 TH |
1547 | /* |
1548 | * Initialize static chunk. If reserved_size is zero, the | |
1549 | * static chunk covers static area + dynamic allocation area | |
1550 | * in the first chunk. If reserved_size is not zero, it | |
1551 | * covers static area + reserved area (mostly used for module | |
1552 | * static percpu allocation). | |
1553 | */ | |
999c17e3 | 1554 | schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0); |
2441d15c | 1555 | INIT_LIST_HEAD(&schunk->list); |
9c824b6a | 1556 | INIT_WORK(&schunk->map_extend_work, pcpu_map_extend_workfn); |
bba174f5 | 1557 | schunk->base_addr = base_addr; |
61ace7fa TH |
1558 | schunk->map = smap; |
1559 | schunk->map_alloc = ARRAY_SIZE(smap); | |
38a6be52 | 1560 | schunk->immutable = true; |
ce3141a2 | 1561 | bitmap_fill(schunk->populated, pcpu_unit_pages); |
b539b87f | 1562 | schunk->nr_populated = pcpu_unit_pages; |
edcb4639 | 1563 | |
fd1e8a1f TH |
1564 | if (ai->reserved_size) { |
1565 | schunk->free_size = ai->reserved_size; | |
ae9e6bc9 | 1566 | pcpu_reserved_chunk = schunk; |
fd1e8a1f | 1567 | pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; |
edcb4639 TH |
1568 | } else { |
1569 | schunk->free_size = dyn_size; | |
1570 | dyn_size = 0; /* dynamic area covered */ | |
1571 | } | |
2441d15c | 1572 | schunk->contig_hint = schunk->free_size; |
fbf59bc9 | 1573 | |
723ad1d9 AV |
1574 | schunk->map[0] = 1; |
1575 | schunk->map[1] = ai->static_size; | |
1576 | schunk->map_used = 1; | |
61ace7fa | 1577 | if (schunk->free_size) |
723ad1d9 AV |
1578 | schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size); |
1579 | else | |
1580 | schunk->map[1] |= 1; | |
61ace7fa | 1581 | |
edcb4639 TH |
1582 | /* init dynamic chunk if necessary */ |
1583 | if (dyn_size) { | |
999c17e3 | 1584 | dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0); |
edcb4639 | 1585 | INIT_LIST_HEAD(&dchunk->list); |
9c824b6a | 1586 | INIT_WORK(&dchunk->map_extend_work, pcpu_map_extend_workfn); |
bba174f5 | 1587 | dchunk->base_addr = base_addr; |
edcb4639 TH |
1588 | dchunk->map = dmap; |
1589 | dchunk->map_alloc = ARRAY_SIZE(dmap); | |
38a6be52 | 1590 | dchunk->immutable = true; |
ce3141a2 | 1591 | bitmap_fill(dchunk->populated, pcpu_unit_pages); |
b539b87f | 1592 | dchunk->nr_populated = pcpu_unit_pages; |
edcb4639 TH |
1593 | |
1594 | dchunk->contig_hint = dchunk->free_size = dyn_size; | |
723ad1d9 AV |
1595 | dchunk->map[0] = 1; |
1596 | dchunk->map[1] = pcpu_reserved_chunk_limit; | |
1597 | dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1; | |
1598 | dchunk->map_used = 2; | |
edcb4639 TH |
1599 | } |
1600 | ||
2441d15c | 1601 | /* link the first chunk in */ |
ae9e6bc9 | 1602 | pcpu_first_chunk = dchunk ?: schunk; |
b539b87f TH |
1603 | pcpu_nr_empty_pop_pages += |
1604 | pcpu_count_occupied_pages(pcpu_first_chunk, 1); | |
ae9e6bc9 | 1605 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
fbf59bc9 TH |
1606 | |
1607 | /* we're done */ | |
bba174f5 | 1608 | pcpu_base_addr = base_addr; |
fb435d52 | 1609 | return 0; |
fbf59bc9 | 1610 | } |
66c3a757 | 1611 | |
bbddff05 TH |
1612 | #ifdef CONFIG_SMP |
1613 | ||
17f3609c | 1614 | const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { |
f58dc01b TH |
1615 | [PCPU_FC_AUTO] = "auto", |
1616 | [PCPU_FC_EMBED] = "embed", | |
1617 | [PCPU_FC_PAGE] = "page", | |
f58dc01b | 1618 | }; |
66c3a757 | 1619 | |
f58dc01b | 1620 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
66c3a757 | 1621 | |
f58dc01b TH |
1622 | static int __init percpu_alloc_setup(char *str) |
1623 | { | |
5479c78a CG |
1624 | if (!str) |
1625 | return -EINVAL; | |
1626 | ||
f58dc01b TH |
1627 | if (0) |
1628 | /* nada */; | |
1629 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK | |
1630 | else if (!strcmp(str, "embed")) | |
1631 | pcpu_chosen_fc = PCPU_FC_EMBED; | |
1632 | #endif | |
1633 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
1634 | else if (!strcmp(str, "page")) | |
1635 | pcpu_chosen_fc = PCPU_FC_PAGE; | |
f58dc01b TH |
1636 | #endif |
1637 | else | |
1638 | pr_warning("PERCPU: unknown allocator %s specified\n", str); | |
66c3a757 | 1639 | |
f58dc01b | 1640 | return 0; |
66c3a757 | 1641 | } |
f58dc01b | 1642 | early_param("percpu_alloc", percpu_alloc_setup); |
66c3a757 | 1643 | |
3c9a024f TH |
1644 | /* |
1645 | * pcpu_embed_first_chunk() is used by the generic percpu setup. | |
1646 | * Build it if needed by the arch config or the generic setup is going | |
1647 | * to be used. | |
1648 | */ | |
08fc4580 TH |
1649 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
1650 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
3c9a024f TH |
1651 | #define BUILD_EMBED_FIRST_CHUNK |
1652 | #endif | |
1653 | ||
1654 | /* build pcpu_page_first_chunk() iff needed by the arch config */ | |
1655 | #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) | |
1656 | #define BUILD_PAGE_FIRST_CHUNK | |
1657 | #endif | |
1658 | ||
1659 | /* pcpu_build_alloc_info() is used by both embed and page first chunk */ | |
1660 | #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) | |
1661 | /** | |
1662 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs | |
1663 | * @reserved_size: the size of reserved percpu area in bytes | |
1664 | * @dyn_size: minimum free size for dynamic allocation in bytes | |
1665 | * @atom_size: allocation atom size | |
1666 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
1667 | * | |
1668 | * This function determines grouping of units, their mappings to cpus | |
1669 | * and other parameters considering needed percpu size, allocation | |
1670 | * atom size and distances between CPUs. | |
1671 | * | |
1672 | * Groups are always mutliples of atom size and CPUs which are of | |
1673 | * LOCAL_DISTANCE both ways are grouped together and share space for | |
1674 | * units in the same group. The returned configuration is guaranteed | |
1675 | * to have CPUs on different nodes on different groups and >=75% usage | |
1676 | * of allocated virtual address space. | |
1677 | * | |
1678 | * RETURNS: | |
1679 | * On success, pointer to the new allocation_info is returned. On | |
1680 | * failure, ERR_PTR value is returned. | |
1681 | */ | |
1682 | static struct pcpu_alloc_info * __init pcpu_build_alloc_info( | |
1683 | size_t reserved_size, size_t dyn_size, | |
1684 | size_t atom_size, | |
1685 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
1686 | { | |
1687 | static int group_map[NR_CPUS] __initdata; | |
1688 | static int group_cnt[NR_CPUS] __initdata; | |
1689 | const size_t static_size = __per_cpu_end - __per_cpu_start; | |
1690 | int nr_groups = 1, nr_units = 0; | |
1691 | size_t size_sum, min_unit_size, alloc_size; | |
1692 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
1693 | int last_allocs, group, unit; | |
1694 | unsigned int cpu, tcpu; | |
1695 | struct pcpu_alloc_info *ai; | |
1696 | unsigned int *cpu_map; | |
1697 | ||
1698 | /* this function may be called multiple times */ | |
1699 | memset(group_map, 0, sizeof(group_map)); | |
1700 | memset(group_cnt, 0, sizeof(group_cnt)); | |
1701 | ||
1702 | /* calculate size_sum and ensure dyn_size is enough for early alloc */ | |
1703 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
1704 | max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); | |
1705 | dyn_size = size_sum - static_size - reserved_size; | |
1706 | ||
1707 | /* | |
1708 | * Determine min_unit_size, alloc_size and max_upa such that | |
1709 | * alloc_size is multiple of atom_size and is the smallest | |
25985edc | 1710 | * which can accommodate 4k aligned segments which are equal to |
3c9a024f TH |
1711 | * or larger than min_unit_size. |
1712 | */ | |
1713 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
1714 | ||
1715 | alloc_size = roundup(min_unit_size, atom_size); | |
1716 | upa = alloc_size / min_unit_size; | |
1717 | while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) | |
1718 | upa--; | |
1719 | max_upa = upa; | |
1720 | ||
1721 | /* group cpus according to their proximity */ | |
1722 | for_each_possible_cpu(cpu) { | |
1723 | group = 0; | |
1724 | next_group: | |
1725 | for_each_possible_cpu(tcpu) { | |
1726 | if (cpu == tcpu) | |
1727 | break; | |
1728 | if (group_map[tcpu] == group && cpu_distance_fn && | |
1729 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || | |
1730 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
1731 | group++; | |
1732 | nr_groups = max(nr_groups, group + 1); | |
1733 | goto next_group; | |
1734 | } | |
1735 | } | |
1736 | group_map[cpu] = group; | |
1737 | group_cnt[group]++; | |
1738 | } | |
1739 | ||
1740 | /* | |
1741 | * Expand unit size until address space usage goes over 75% | |
1742 | * and then as much as possible without using more address | |
1743 | * space. | |
1744 | */ | |
1745 | last_allocs = INT_MAX; | |
1746 | for (upa = max_upa; upa; upa--) { | |
1747 | int allocs = 0, wasted = 0; | |
1748 | ||
1749 | if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) | |
1750 | continue; | |
1751 | ||
1752 | for (group = 0; group < nr_groups; group++) { | |
1753 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); | |
1754 | allocs += this_allocs; | |
1755 | wasted += this_allocs * upa - group_cnt[group]; | |
1756 | } | |
1757 | ||
1758 | /* | |
1759 | * Don't accept if wastage is over 1/3. The | |
1760 | * greater-than comparison ensures upa==1 always | |
1761 | * passes the following check. | |
1762 | */ | |
1763 | if (wasted > num_possible_cpus() / 3) | |
1764 | continue; | |
1765 | ||
1766 | /* and then don't consume more memory */ | |
1767 | if (allocs > last_allocs) | |
1768 | break; | |
1769 | last_allocs = allocs; | |
1770 | best_upa = upa; | |
1771 | } | |
1772 | upa = best_upa; | |
1773 | ||
1774 | /* allocate and fill alloc_info */ | |
1775 | for (group = 0; group < nr_groups; group++) | |
1776 | nr_units += roundup(group_cnt[group], upa); | |
1777 | ||
1778 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); | |
1779 | if (!ai) | |
1780 | return ERR_PTR(-ENOMEM); | |
1781 | cpu_map = ai->groups[0].cpu_map; | |
1782 | ||
1783 | for (group = 0; group < nr_groups; group++) { | |
1784 | ai->groups[group].cpu_map = cpu_map; | |
1785 | cpu_map += roundup(group_cnt[group], upa); | |
1786 | } | |
1787 | ||
1788 | ai->static_size = static_size; | |
1789 | ai->reserved_size = reserved_size; | |
1790 | ai->dyn_size = dyn_size; | |
1791 | ai->unit_size = alloc_size / upa; | |
1792 | ai->atom_size = atom_size; | |
1793 | ai->alloc_size = alloc_size; | |
1794 | ||
1795 | for (group = 0, unit = 0; group_cnt[group]; group++) { | |
1796 | struct pcpu_group_info *gi = &ai->groups[group]; | |
1797 | ||
1798 | /* | |
1799 | * Initialize base_offset as if all groups are located | |
1800 | * back-to-back. The caller should update this to | |
1801 | * reflect actual allocation. | |
1802 | */ | |
1803 | gi->base_offset = unit * ai->unit_size; | |
1804 | ||
1805 | for_each_possible_cpu(cpu) | |
1806 | if (group_map[cpu] == group) | |
1807 | gi->cpu_map[gi->nr_units++] = cpu; | |
1808 | gi->nr_units = roundup(gi->nr_units, upa); | |
1809 | unit += gi->nr_units; | |
1810 | } | |
1811 | BUG_ON(unit != nr_units); | |
1812 | ||
1813 | return ai; | |
1814 | } | |
1815 | #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ | |
1816 | ||
1817 | #if defined(BUILD_EMBED_FIRST_CHUNK) | |
66c3a757 TH |
1818 | /** |
1819 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
66c3a757 | 1820 | * @reserved_size: the size of reserved percpu area in bytes |
4ba6ce25 | 1821 | * @dyn_size: minimum free size for dynamic allocation in bytes |
c8826dd5 TH |
1822 | * @atom_size: allocation atom size |
1823 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
1824 | * @alloc_fn: function to allocate percpu page | |
25985edc | 1825 | * @free_fn: function to free percpu page |
66c3a757 TH |
1826 | * |
1827 | * This is a helper to ease setting up embedded first percpu chunk and | |
1828 | * can be called where pcpu_setup_first_chunk() is expected. | |
1829 | * | |
1830 | * If this function is used to setup the first chunk, it is allocated | |
c8826dd5 TH |
1831 | * by calling @alloc_fn and used as-is without being mapped into |
1832 | * vmalloc area. Allocations are always whole multiples of @atom_size | |
1833 | * aligned to @atom_size. | |
1834 | * | |
1835 | * This enables the first chunk to piggy back on the linear physical | |
1836 | * mapping which often uses larger page size. Please note that this | |
1837 | * can result in very sparse cpu->unit mapping on NUMA machines thus | |
1838 | * requiring large vmalloc address space. Don't use this allocator if | |
1839 | * vmalloc space is not orders of magnitude larger than distances | |
1840 | * between node memory addresses (ie. 32bit NUMA machines). | |
66c3a757 | 1841 | * |
4ba6ce25 | 1842 | * @dyn_size specifies the minimum dynamic area size. |
66c3a757 TH |
1843 | * |
1844 | * If the needed size is smaller than the minimum or specified unit | |
c8826dd5 | 1845 | * size, the leftover is returned using @free_fn. |
66c3a757 TH |
1846 | * |
1847 | * RETURNS: | |
fb435d52 | 1848 | * 0 on success, -errno on failure. |
66c3a757 | 1849 | */ |
4ba6ce25 | 1850 | int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, |
c8826dd5 TH |
1851 | size_t atom_size, |
1852 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, | |
1853 | pcpu_fc_alloc_fn_t alloc_fn, | |
1854 | pcpu_fc_free_fn_t free_fn) | |
66c3a757 | 1855 | { |
c8826dd5 TH |
1856 | void *base = (void *)ULONG_MAX; |
1857 | void **areas = NULL; | |
fd1e8a1f | 1858 | struct pcpu_alloc_info *ai; |
6ea529a2 | 1859 | size_t size_sum, areas_size, max_distance; |
c8826dd5 | 1860 | int group, i, rc; |
66c3a757 | 1861 | |
c8826dd5 TH |
1862 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
1863 | cpu_distance_fn); | |
fd1e8a1f TH |
1864 | if (IS_ERR(ai)) |
1865 | return PTR_ERR(ai); | |
66c3a757 | 1866 | |
fd1e8a1f | 1867 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
c8826dd5 | 1868 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
fa8a7094 | 1869 | |
999c17e3 | 1870 | areas = memblock_virt_alloc_nopanic(areas_size, 0); |
c8826dd5 | 1871 | if (!areas) { |
fb435d52 | 1872 | rc = -ENOMEM; |
c8826dd5 | 1873 | goto out_free; |
fa8a7094 | 1874 | } |
66c3a757 | 1875 | |
c8826dd5 TH |
1876 | /* allocate, copy and determine base address */ |
1877 | for (group = 0; group < ai->nr_groups; group++) { | |
1878 | struct pcpu_group_info *gi = &ai->groups[group]; | |
1879 | unsigned int cpu = NR_CPUS; | |
1880 | void *ptr; | |
1881 | ||
1882 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) | |
1883 | cpu = gi->cpu_map[i]; | |
1884 | BUG_ON(cpu == NR_CPUS); | |
1885 | ||
1886 | /* allocate space for the whole group */ | |
1887 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); | |
1888 | if (!ptr) { | |
1889 | rc = -ENOMEM; | |
1890 | goto out_free_areas; | |
1891 | } | |
f528f0b8 CM |
1892 | /* kmemleak tracks the percpu allocations separately */ |
1893 | kmemleak_free(ptr); | |
c8826dd5 | 1894 | areas[group] = ptr; |
fd1e8a1f | 1895 | |
c8826dd5 | 1896 | base = min(ptr, base); |
42b64281 TH |
1897 | } |
1898 | ||
1899 | /* | |
1900 | * Copy data and free unused parts. This should happen after all | |
1901 | * allocations are complete; otherwise, we may end up with | |
1902 | * overlapping groups. | |
1903 | */ | |
1904 | for (group = 0; group < ai->nr_groups; group++) { | |
1905 | struct pcpu_group_info *gi = &ai->groups[group]; | |
1906 | void *ptr = areas[group]; | |
c8826dd5 TH |
1907 | |
1908 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { | |
1909 | if (gi->cpu_map[i] == NR_CPUS) { | |
1910 | /* unused unit, free whole */ | |
1911 | free_fn(ptr, ai->unit_size); | |
1912 | continue; | |
1913 | } | |
1914 | /* copy and return the unused part */ | |
1915 | memcpy(ptr, __per_cpu_load, ai->static_size); | |
1916 | free_fn(ptr + size_sum, ai->unit_size - size_sum); | |
1917 | } | |
fa8a7094 | 1918 | } |
66c3a757 | 1919 | |
c8826dd5 | 1920 | /* base address is now known, determine group base offsets */ |
6ea529a2 TH |
1921 | max_distance = 0; |
1922 | for (group = 0; group < ai->nr_groups; group++) { | |
c8826dd5 | 1923 | ai->groups[group].base_offset = areas[group] - base; |
1a0c3298 TH |
1924 | max_distance = max_t(size_t, max_distance, |
1925 | ai->groups[group].base_offset); | |
6ea529a2 TH |
1926 | } |
1927 | max_distance += ai->unit_size; | |
1928 | ||
1929 | /* warn if maximum distance is further than 75% of vmalloc space */ | |
8a092171 | 1930 | if (max_distance > VMALLOC_TOTAL * 3 / 4) { |
1a0c3298 | 1931 | pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc " |
787e5b06 | 1932 | "space 0x%lx\n", max_distance, |
8a092171 | 1933 | VMALLOC_TOTAL); |
6ea529a2 TH |
1934 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
1935 | /* and fail if we have fallback */ | |
1936 | rc = -EINVAL; | |
1937 | goto out_free; | |
1938 | #endif | |
1939 | } | |
c8826dd5 | 1940 | |
004018e2 | 1941 | pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
fd1e8a1f TH |
1942 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
1943 | ai->dyn_size, ai->unit_size); | |
d4b95f80 | 1944 | |
fb435d52 | 1945 | rc = pcpu_setup_first_chunk(ai, base); |
c8826dd5 TH |
1946 | goto out_free; |
1947 | ||
1948 | out_free_areas: | |
1949 | for (group = 0; group < ai->nr_groups; group++) | |
f851c8d8 MH |
1950 | if (areas[group]) |
1951 | free_fn(areas[group], | |
1952 | ai->groups[group].nr_units * ai->unit_size); | |
c8826dd5 | 1953 | out_free: |
fd1e8a1f | 1954 | pcpu_free_alloc_info(ai); |
c8826dd5 | 1955 | if (areas) |
999c17e3 | 1956 | memblock_free_early(__pa(areas), areas_size); |
fb435d52 | 1957 | return rc; |
d4b95f80 | 1958 | } |
3c9a024f | 1959 | #endif /* BUILD_EMBED_FIRST_CHUNK */ |
d4b95f80 | 1960 | |
3c9a024f | 1961 | #ifdef BUILD_PAGE_FIRST_CHUNK |
d4b95f80 | 1962 | /** |
00ae4064 | 1963 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
1964 | * @reserved_size: the size of reserved percpu area in bytes |
1965 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
25985edc | 1966 | * @free_fn: function to free percpu page, always called with PAGE_SIZE |
d4b95f80 TH |
1967 | * @populate_pte_fn: function to populate pte |
1968 | * | |
00ae4064 TH |
1969 | * This is a helper to ease setting up page-remapped first percpu |
1970 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
1971 | * |
1972 | * This is the basic allocator. Static percpu area is allocated | |
1973 | * page-by-page into vmalloc area. | |
1974 | * | |
1975 | * RETURNS: | |
fb435d52 | 1976 | * 0 on success, -errno on failure. |
d4b95f80 | 1977 | */ |
fb435d52 TH |
1978 | int __init pcpu_page_first_chunk(size_t reserved_size, |
1979 | pcpu_fc_alloc_fn_t alloc_fn, | |
1980 | pcpu_fc_free_fn_t free_fn, | |
1981 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 1982 | { |
8f05a6a6 | 1983 | static struct vm_struct vm; |
fd1e8a1f | 1984 | struct pcpu_alloc_info *ai; |
00ae4064 | 1985 | char psize_str[16]; |
ce3141a2 | 1986 | int unit_pages; |
d4b95f80 | 1987 | size_t pages_size; |
ce3141a2 | 1988 | struct page **pages; |
fb435d52 | 1989 | int unit, i, j, rc; |
d4b95f80 | 1990 | |
00ae4064 TH |
1991 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
1992 | ||
4ba6ce25 | 1993 | ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); |
fd1e8a1f TH |
1994 | if (IS_ERR(ai)) |
1995 | return PTR_ERR(ai); | |
1996 | BUG_ON(ai->nr_groups != 1); | |
1997 | BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); | |
1998 | ||
1999 | unit_pages = ai->unit_size >> PAGE_SHIFT; | |
d4b95f80 TH |
2000 | |
2001 | /* unaligned allocations can't be freed, round up to page size */ | |
fd1e8a1f TH |
2002 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
2003 | sizeof(pages[0])); | |
999c17e3 | 2004 | pages = memblock_virt_alloc(pages_size, 0); |
d4b95f80 | 2005 | |
8f05a6a6 | 2006 | /* allocate pages */ |
d4b95f80 | 2007 | j = 0; |
fd1e8a1f | 2008 | for (unit = 0; unit < num_possible_cpus(); unit++) |
ce3141a2 | 2009 | for (i = 0; i < unit_pages; i++) { |
fd1e8a1f | 2010 | unsigned int cpu = ai->groups[0].cpu_map[unit]; |
d4b95f80 TH |
2011 | void *ptr; |
2012 | ||
3cbc8565 | 2013 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
d4b95f80 | 2014 | if (!ptr) { |
00ae4064 TH |
2015 | pr_warning("PERCPU: failed to allocate %s page " |
2016 | "for cpu%u\n", psize_str, cpu); | |
d4b95f80 TH |
2017 | goto enomem; |
2018 | } | |
f528f0b8 CM |
2019 | /* kmemleak tracks the percpu allocations separately */ |
2020 | kmemleak_free(ptr); | |
ce3141a2 | 2021 | pages[j++] = virt_to_page(ptr); |
d4b95f80 TH |
2022 | } |
2023 | ||
8f05a6a6 TH |
2024 | /* allocate vm area, map the pages and copy static data */ |
2025 | vm.flags = VM_ALLOC; | |
fd1e8a1f | 2026 | vm.size = num_possible_cpus() * ai->unit_size; |
8f05a6a6 TH |
2027 | vm_area_register_early(&vm, PAGE_SIZE); |
2028 | ||
fd1e8a1f | 2029 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
1d9d3257 | 2030 | unsigned long unit_addr = |
fd1e8a1f | 2031 | (unsigned long)vm.addr + unit * ai->unit_size; |
8f05a6a6 | 2032 | |
ce3141a2 | 2033 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
2034 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
2035 | ||
2036 | /* pte already populated, the following shouldn't fail */ | |
fb435d52 TH |
2037 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
2038 | unit_pages); | |
2039 | if (rc < 0) | |
2040 | panic("failed to map percpu area, err=%d\n", rc); | |
66c3a757 | 2041 | |
8f05a6a6 TH |
2042 | /* |
2043 | * FIXME: Archs with virtual cache should flush local | |
2044 | * cache for the linear mapping here - something | |
2045 | * equivalent to flush_cache_vmap() on the local cpu. | |
2046 | * flush_cache_vmap() can't be used as most supporting | |
2047 | * data structures are not set up yet. | |
2048 | */ | |
2049 | ||
2050 | /* copy static data */ | |
fd1e8a1f | 2051 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
66c3a757 TH |
2052 | } |
2053 | ||
2054 | /* we're ready, commit */ | |
1d9d3257 | 2055 | pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", |
fd1e8a1f TH |
2056 | unit_pages, psize_str, vm.addr, ai->static_size, |
2057 | ai->reserved_size, ai->dyn_size); | |
d4b95f80 | 2058 | |
fb435d52 | 2059 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
d4b95f80 TH |
2060 | goto out_free_ar; |
2061 | ||
2062 | enomem: | |
2063 | while (--j >= 0) | |
ce3141a2 | 2064 | free_fn(page_address(pages[j]), PAGE_SIZE); |
fb435d52 | 2065 | rc = -ENOMEM; |
d4b95f80 | 2066 | out_free_ar: |
999c17e3 | 2067 | memblock_free_early(__pa(pages), pages_size); |
fd1e8a1f | 2068 | pcpu_free_alloc_info(ai); |
fb435d52 | 2069 | return rc; |
d4b95f80 | 2070 | } |
3c9a024f | 2071 | #endif /* BUILD_PAGE_FIRST_CHUNK */ |
d4b95f80 | 2072 | |
bbddff05 | 2073 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
e74e3962 | 2074 | /* |
bbddff05 | 2075 | * Generic SMP percpu area setup. |
e74e3962 TH |
2076 | * |
2077 | * The embedding helper is used because its behavior closely resembles | |
2078 | * the original non-dynamic generic percpu area setup. This is | |
2079 | * important because many archs have addressing restrictions and might | |
2080 | * fail if the percpu area is located far away from the previous | |
2081 | * location. As an added bonus, in non-NUMA cases, embedding is | |
2082 | * generally a good idea TLB-wise because percpu area can piggy back | |
2083 | * on the physical linear memory mapping which uses large page | |
2084 | * mappings on applicable archs. | |
2085 | */ | |
e74e3962 TH |
2086 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
2087 | EXPORT_SYMBOL(__per_cpu_offset); | |
2088 | ||
c8826dd5 TH |
2089 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
2090 | size_t align) | |
2091 | { | |
999c17e3 SS |
2092 | return memblock_virt_alloc_from_nopanic( |
2093 | size, align, __pa(MAX_DMA_ADDRESS)); | |
c8826dd5 | 2094 | } |
66c3a757 | 2095 | |
c8826dd5 TH |
2096 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
2097 | { | |
999c17e3 | 2098 | memblock_free_early(__pa(ptr), size); |
c8826dd5 TH |
2099 | } |
2100 | ||
e74e3962 TH |
2101 | void __init setup_per_cpu_areas(void) |
2102 | { | |
e74e3962 TH |
2103 | unsigned long delta; |
2104 | unsigned int cpu; | |
fb435d52 | 2105 | int rc; |
e74e3962 TH |
2106 | |
2107 | /* | |
2108 | * Always reserve area for module percpu variables. That's | |
2109 | * what the legacy allocator did. | |
2110 | */ | |
fb435d52 | 2111 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
c8826dd5 TH |
2112 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
2113 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); | |
fb435d52 | 2114 | if (rc < 0) |
bbddff05 | 2115 | panic("Failed to initialize percpu areas."); |
e74e3962 TH |
2116 | |
2117 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2118 | for_each_possible_cpu(cpu) | |
fb435d52 | 2119 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
66c3a757 | 2120 | } |
bbddff05 TH |
2121 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
2122 | ||
2123 | #else /* CONFIG_SMP */ | |
2124 | ||
2125 | /* | |
2126 | * UP percpu area setup. | |
2127 | * | |
2128 | * UP always uses km-based percpu allocator with identity mapping. | |
2129 | * Static percpu variables are indistinguishable from the usual static | |
2130 | * variables and don't require any special preparation. | |
2131 | */ | |
2132 | void __init setup_per_cpu_areas(void) | |
2133 | { | |
2134 | const size_t unit_size = | |
2135 | roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, | |
2136 | PERCPU_DYNAMIC_RESERVE)); | |
2137 | struct pcpu_alloc_info *ai; | |
2138 | void *fc; | |
2139 | ||
2140 | ai = pcpu_alloc_alloc_info(1, 1); | |
999c17e3 SS |
2141 | fc = memblock_virt_alloc_from_nopanic(unit_size, |
2142 | PAGE_SIZE, | |
2143 | __pa(MAX_DMA_ADDRESS)); | |
bbddff05 TH |
2144 | if (!ai || !fc) |
2145 | panic("Failed to allocate memory for percpu areas."); | |
100d13c3 CM |
2146 | /* kmemleak tracks the percpu allocations separately */ |
2147 | kmemleak_free(fc); | |
bbddff05 TH |
2148 | |
2149 | ai->dyn_size = unit_size; | |
2150 | ai->unit_size = unit_size; | |
2151 | ai->atom_size = unit_size; | |
2152 | ai->alloc_size = unit_size; | |
2153 | ai->groups[0].nr_units = 1; | |
2154 | ai->groups[0].cpu_map[0] = 0; | |
2155 | ||
2156 | if (pcpu_setup_first_chunk(ai, fc) < 0) | |
2157 | panic("Failed to initialize percpu areas."); | |
3189eddb HL |
2158 | |
2159 | pcpu_free_alloc_info(ai); | |
bbddff05 TH |
2160 | } |
2161 | ||
2162 | #endif /* CONFIG_SMP */ | |
099a19d9 TH |
2163 | |
2164 | /* | |
2165 | * First and reserved chunks are initialized with temporary allocation | |
2166 | * map in initdata so that they can be used before slab is online. | |
2167 | * This function is called after slab is brought up and replaces those | |
2168 | * with properly allocated maps. | |
2169 | */ | |
2170 | void __init percpu_init_late(void) | |
2171 | { | |
2172 | struct pcpu_chunk *target_chunks[] = | |
2173 | { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; | |
2174 | struct pcpu_chunk *chunk; | |
2175 | unsigned long flags; | |
2176 | int i; | |
2177 | ||
2178 | for (i = 0; (chunk = target_chunks[i]); i++) { | |
2179 | int *map; | |
2180 | const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); | |
2181 | ||
2182 | BUILD_BUG_ON(size > PAGE_SIZE); | |
2183 | ||
90459ce0 | 2184 | map = pcpu_mem_zalloc(size); |
099a19d9 TH |
2185 | BUG_ON(!map); |
2186 | ||
2187 | spin_lock_irqsave(&pcpu_lock, flags); | |
2188 | memcpy(map, chunk->map, size); | |
2189 | chunk->map = map; | |
2190 | spin_unlock_irqrestore(&pcpu_lock, flags); | |
2191 | } | |
2192 | } |