]>
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 | * | |
9c015162 | 7 | * This file is released under the GPLv2 license. |
fbf59bc9 | 8 | * |
9c015162 DZF |
9 | * The percpu allocator handles both static and dynamic areas. Percpu |
10 | * areas are allocated in chunks which are divided into units. There is | |
11 | * a 1-to-1 mapping for units to possible cpus. These units are grouped | |
12 | * based on NUMA properties of the machine. | |
fbf59bc9 TH |
13 | * |
14 | * c0 c1 c2 | |
15 | * ------------------- ------------------- ------------ | |
16 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u | |
17 | * ------------------- ...... ------------------- .... ------------ | |
18 | * | |
9c015162 DZF |
19 | * Allocation is done by offsets into a unit's address space. Ie., an |
20 | * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, | |
21 | * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear | |
22 | * and even sparse. Access is handled by configuring percpu base | |
23 | * registers according to the cpu to unit mappings and offsetting the | |
24 | * base address using pcpu_unit_size. | |
25 | * | |
26 | * There is special consideration for the first chunk which must handle | |
27 | * the static percpu variables in the kernel image as allocation services | |
28 | * are not online yet. In short, the first chunk is structure like so: | |
29 | * | |
30 | * <Static | [Reserved] | Dynamic> | |
31 | * | |
32 | * The static data is copied from the original section managed by the | |
33 | * linker. The reserved section, if non-zero, primarily manages static | |
34 | * percpu variables from kernel modules. Finally, the dynamic section | |
35 | * takes care of normal allocations. | |
fbf59bc9 TH |
36 | * |
37 | * Allocation state in each chunk is kept using an array of integers | |
38 | * on chunk->map. A positive value in the map represents a free | |
39 | * region and negative allocated. Allocation inside a chunk is done | |
40 | * by scanning this map sequentially and serving the first matching | |
41 | * entry. This is mostly copied from the percpu_modalloc() allocator. | |
e1b9aa3f CL |
42 | * Chunks can be determined from the address using the index field |
43 | * in the page struct. The index field contains a pointer to the chunk. | |
fbf59bc9 | 44 | * |
9c015162 DZF |
45 | * These chunks are organized into lists according to free_size and |
46 | * tries to allocate from the fullest chunk first. Each chunk maintains | |
47 | * a maximum contiguous area size hint which is guaranteed to be equal | |
48 | * to or larger than the maximum contiguous area in the chunk. This | |
49 | * helps prevent the allocator from iterating over chunks unnecessarily. | |
50 | * | |
4091fb95 | 51 | * To use this allocator, arch code should do the following: |
fbf59bc9 | 52 | * |
fbf59bc9 | 53 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
e0100983 TH |
54 | * regular address to percpu pointer and back if they need to be |
55 | * different from the default | |
fbf59bc9 | 56 | * |
8d408b4b TH |
57 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
58 | * setup the first chunk containing the kernel static percpu area | |
fbf59bc9 TH |
59 | */ |
60 | ||
870d4b12 JP |
61 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
62 | ||
fbf59bc9 TH |
63 | #include <linux/bitmap.h> |
64 | #include <linux/bootmem.h> | |
fd1e8a1f | 65 | #include <linux/err.h> |
ca460b3c | 66 | #include <linux/lcm.h> |
fbf59bc9 | 67 | #include <linux/list.h> |
a530b795 | 68 | #include <linux/log2.h> |
fbf59bc9 TH |
69 | #include <linux/mm.h> |
70 | #include <linux/module.h> | |
71 | #include <linux/mutex.h> | |
72 | #include <linux/percpu.h> | |
73 | #include <linux/pfn.h> | |
fbf59bc9 | 74 | #include <linux/slab.h> |
ccea34b5 | 75 | #include <linux/spinlock.h> |
fbf59bc9 | 76 | #include <linux/vmalloc.h> |
a56dbddf | 77 | #include <linux/workqueue.h> |
f528f0b8 | 78 | #include <linux/kmemleak.h> |
fbf59bc9 TH |
79 | |
80 | #include <asm/cacheflush.h> | |
e0100983 | 81 | #include <asm/sections.h> |
fbf59bc9 | 82 | #include <asm/tlbflush.h> |
3b034b0d | 83 | #include <asm/io.h> |
fbf59bc9 | 84 | |
df95e795 DZ |
85 | #define CREATE_TRACE_POINTS |
86 | #include <trace/events/percpu.h> | |
87 | ||
8fa3ed80 DZ |
88 | #include "percpu-internal.h" |
89 | ||
40064aec DZF |
90 | /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */ |
91 | #define PCPU_SLOT_BASE_SHIFT 5 | |
92 | ||
1a4d7607 TH |
93 | #define PCPU_EMPTY_POP_PAGES_LOW 2 |
94 | #define PCPU_EMPTY_POP_PAGES_HIGH 4 | |
fbf59bc9 | 95 | |
bbddff05 | 96 | #ifdef CONFIG_SMP |
e0100983 TH |
97 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
98 | #ifndef __addr_to_pcpu_ptr | |
99 | #define __addr_to_pcpu_ptr(addr) \ | |
43cf38eb TH |
100 | (void __percpu *)((unsigned long)(addr) - \ |
101 | (unsigned long)pcpu_base_addr + \ | |
102 | (unsigned long)__per_cpu_start) | |
e0100983 TH |
103 | #endif |
104 | #ifndef __pcpu_ptr_to_addr | |
105 | #define __pcpu_ptr_to_addr(ptr) \ | |
43cf38eb TH |
106 | (void __force *)((unsigned long)(ptr) + \ |
107 | (unsigned long)pcpu_base_addr - \ | |
108 | (unsigned long)__per_cpu_start) | |
e0100983 | 109 | #endif |
bbddff05 TH |
110 | #else /* CONFIG_SMP */ |
111 | /* on UP, it's always identity mapped */ | |
112 | #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) | |
113 | #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) | |
114 | #endif /* CONFIG_SMP */ | |
e0100983 | 115 | |
1328710b DM |
116 | static int pcpu_unit_pages __ro_after_init; |
117 | static int pcpu_unit_size __ro_after_init; | |
118 | static int pcpu_nr_units __ro_after_init; | |
119 | static int pcpu_atom_size __ro_after_init; | |
8fa3ed80 | 120 | int pcpu_nr_slots __ro_after_init; |
1328710b | 121 | static size_t pcpu_chunk_struct_size __ro_after_init; |
fbf59bc9 | 122 | |
a855b84c | 123 | /* cpus with the lowest and highest unit addresses */ |
1328710b DM |
124 | static unsigned int pcpu_low_unit_cpu __ro_after_init; |
125 | static unsigned int pcpu_high_unit_cpu __ro_after_init; | |
2f39e637 | 126 | |
fbf59bc9 | 127 | /* the address of the first chunk which starts with the kernel static area */ |
1328710b | 128 | void *pcpu_base_addr __ro_after_init; |
fbf59bc9 TH |
129 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
130 | ||
1328710b DM |
131 | static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ |
132 | const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ | |
2f39e637 | 133 | |
6563297c | 134 | /* group information, used for vm allocation */ |
1328710b DM |
135 | static int pcpu_nr_groups __ro_after_init; |
136 | static const unsigned long *pcpu_group_offsets __ro_after_init; | |
137 | static const size_t *pcpu_group_sizes __ro_after_init; | |
6563297c | 138 | |
ae9e6bc9 TH |
139 | /* |
140 | * The first chunk which always exists. Note that unlike other | |
141 | * chunks, this one can be allocated and mapped in several different | |
142 | * ways and thus often doesn't live in the vmalloc area. | |
143 | */ | |
8fa3ed80 | 144 | struct pcpu_chunk *pcpu_first_chunk __ro_after_init; |
ae9e6bc9 TH |
145 | |
146 | /* | |
147 | * Optional reserved chunk. This chunk reserves part of the first | |
e2266705 DZF |
148 | * chunk and serves it for reserved allocations. When the reserved |
149 | * region doesn't exist, the following variable is NULL. | |
ae9e6bc9 | 150 | */ |
8fa3ed80 | 151 | struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; |
edcb4639 | 152 | |
8fa3ed80 | 153 | DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ |
6710e594 | 154 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ |
fbf59bc9 | 155 | |
8fa3ed80 | 156 | struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */ |
fbf59bc9 | 157 | |
4f996e23 TH |
158 | /* chunks which need their map areas extended, protected by pcpu_lock */ |
159 | static LIST_HEAD(pcpu_map_extend_chunks); | |
160 | ||
b539b87f TH |
161 | /* |
162 | * The number of empty populated pages, protected by pcpu_lock. The | |
163 | * reserved chunk doesn't contribute to the count. | |
164 | */ | |
6b9b6f39 | 165 | int pcpu_nr_empty_pop_pages; |
b539b87f | 166 | |
1a4d7607 TH |
167 | /* |
168 | * Balance work is used to populate or destroy chunks asynchronously. We | |
169 | * try to keep the number of populated free pages between | |
170 | * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one | |
171 | * empty chunk. | |
172 | */ | |
fe6bd8c3 TH |
173 | static void pcpu_balance_workfn(struct work_struct *work); |
174 | static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); | |
1a4d7607 TH |
175 | static bool pcpu_async_enabled __read_mostly; |
176 | static bool pcpu_atomic_alloc_failed; | |
177 | ||
178 | static void pcpu_schedule_balance_work(void) | |
179 | { | |
180 | if (pcpu_async_enabled) | |
181 | schedule_work(&pcpu_balance_work); | |
182 | } | |
a56dbddf | 183 | |
c0ebfdc3 | 184 | /** |
560f2c23 DZF |
185 | * pcpu_addr_in_chunk - check if the address is served from this chunk |
186 | * @chunk: chunk of interest | |
187 | * @addr: percpu address | |
c0ebfdc3 DZF |
188 | * |
189 | * RETURNS: | |
560f2c23 | 190 | * True if the address is served from this chunk. |
c0ebfdc3 | 191 | */ |
560f2c23 | 192 | static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) |
020ec653 | 193 | { |
c0ebfdc3 DZF |
194 | void *start_addr, *end_addr; |
195 | ||
560f2c23 | 196 | if (!chunk) |
c0ebfdc3 | 197 | return false; |
020ec653 | 198 | |
560f2c23 DZF |
199 | start_addr = chunk->base_addr + chunk->start_offset; |
200 | end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - | |
201 | chunk->end_offset; | |
c0ebfdc3 DZF |
202 | |
203 | return addr >= start_addr && addr < end_addr; | |
020ec653 TH |
204 | } |
205 | ||
d9b55eeb | 206 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 207 | { |
cae3aeb8 | 208 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
209 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
210 | } | |
211 | ||
d9b55eeb TH |
212 | static int pcpu_size_to_slot(int size) |
213 | { | |
214 | if (size == pcpu_unit_size) | |
215 | return pcpu_nr_slots - 1; | |
216 | return __pcpu_size_to_slot(size); | |
217 | } | |
218 | ||
fbf59bc9 TH |
219 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
220 | { | |
40064aec | 221 | if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0) |
fbf59bc9 TH |
222 | return 0; |
223 | ||
40064aec | 224 | return pcpu_size_to_slot(chunk->free_bytes); |
fbf59bc9 TH |
225 | } |
226 | ||
88999a89 TH |
227 | /* set the pointer to a chunk in a page struct */ |
228 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
229 | { | |
230 | page->index = (unsigned long)pcpu; | |
231 | } | |
232 | ||
233 | /* obtain pointer to a chunk from a page struct */ | |
234 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
235 | { | |
236 | return (struct pcpu_chunk *)page->index; | |
237 | } | |
238 | ||
239 | static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) | |
fbf59bc9 | 240 | { |
2f39e637 | 241 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
242 | } |
243 | ||
c0ebfdc3 DZF |
244 | static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) |
245 | { | |
246 | return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); | |
247 | } | |
248 | ||
9983b6f0 TH |
249 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
250 | unsigned int cpu, int page_idx) | |
fbf59bc9 | 251 | { |
c0ebfdc3 DZF |
252 | return (unsigned long)chunk->base_addr + |
253 | pcpu_unit_page_offset(cpu, page_idx); | |
fbf59bc9 TH |
254 | } |
255 | ||
91e914c5 | 256 | static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end) |
ce3141a2 | 257 | { |
91e914c5 DZF |
258 | *rs = find_next_zero_bit(bitmap, end, *rs); |
259 | *re = find_next_bit(bitmap, end, *rs + 1); | |
ce3141a2 TH |
260 | } |
261 | ||
91e914c5 | 262 | static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end) |
ce3141a2 | 263 | { |
91e914c5 DZF |
264 | *rs = find_next_bit(bitmap, end, *rs); |
265 | *re = find_next_zero_bit(bitmap, end, *rs + 1); | |
ce3141a2 TH |
266 | } |
267 | ||
268 | /* | |
91e914c5 DZF |
269 | * Bitmap region iterators. Iterates over the bitmap between |
270 | * [@start, @end) in @chunk. @rs and @re should be integer variables | |
271 | * and will be set to start and end index of the current free region. | |
ce3141a2 | 272 | */ |
91e914c5 DZF |
273 | #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \ |
274 | for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \ | |
275 | (rs) < (re); \ | |
276 | (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end))) | |
ce3141a2 | 277 | |
91e914c5 DZF |
278 | #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \ |
279 | for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \ | |
280 | (rs) < (re); \ | |
281 | (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end))) | |
ce3141a2 | 282 | |
ca460b3c DZF |
283 | /* |
284 | * The following are helper functions to help access bitmaps and convert | |
285 | * between bitmap offsets to address offsets. | |
286 | */ | |
287 | static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) | |
288 | { | |
289 | return chunk->alloc_map + | |
290 | (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); | |
291 | } | |
292 | ||
293 | static unsigned long pcpu_off_to_block_index(int off) | |
294 | { | |
295 | return off / PCPU_BITMAP_BLOCK_BITS; | |
296 | } | |
297 | ||
298 | static unsigned long pcpu_off_to_block_off(int off) | |
299 | { | |
300 | return off & (PCPU_BITMAP_BLOCK_BITS - 1); | |
301 | } | |
302 | ||
fbf59bc9 | 303 | /** |
90459ce0 | 304 | * pcpu_mem_zalloc - allocate memory |
1880d93b | 305 | * @size: bytes to allocate |
fbf59bc9 | 306 | * |
1880d93b | 307 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
90459ce0 | 308 | * kzalloc() is used; otherwise, vzalloc() is used. The returned |
1880d93b | 309 | * memory is always zeroed. |
fbf59bc9 | 310 | * |
ccea34b5 TH |
311 | * CONTEXT: |
312 | * Does GFP_KERNEL allocation. | |
313 | * | |
fbf59bc9 | 314 | * RETURNS: |
1880d93b | 315 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 316 | */ |
90459ce0 | 317 | static void *pcpu_mem_zalloc(size_t size) |
fbf59bc9 | 318 | { |
099a19d9 TH |
319 | if (WARN_ON_ONCE(!slab_is_available())) |
320 | return NULL; | |
321 | ||
1880d93b TH |
322 | if (size <= PAGE_SIZE) |
323 | return kzalloc(size, GFP_KERNEL); | |
7af4c093 JJ |
324 | else |
325 | return vzalloc(size); | |
1880d93b | 326 | } |
fbf59bc9 | 327 | |
1880d93b TH |
328 | /** |
329 | * pcpu_mem_free - free memory | |
330 | * @ptr: memory to free | |
1880d93b | 331 | * |
90459ce0 | 332 | * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). |
1880d93b | 333 | */ |
1d5cfdb0 | 334 | static void pcpu_mem_free(void *ptr) |
1880d93b | 335 | { |
1d5cfdb0 | 336 | kvfree(ptr); |
fbf59bc9 TH |
337 | } |
338 | ||
339 | /** | |
340 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
341 | * @chunk: chunk of interest | |
342 | * @oslot: the previous slot it was on | |
343 | * | |
344 | * This function is called after an allocation or free changed @chunk. | |
345 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
346 | * moved to the slot. Note that the reserved chunk is never put on |
347 | * chunk slots. | |
ccea34b5 TH |
348 | * |
349 | * CONTEXT: | |
350 | * pcpu_lock. | |
fbf59bc9 TH |
351 | */ |
352 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
353 | { | |
354 | int nslot = pcpu_chunk_slot(chunk); | |
355 | ||
edcb4639 | 356 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
357 | if (oslot < nslot) |
358 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
359 | else | |
360 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
361 | } | |
362 | } | |
363 | ||
9f7dcf22 | 364 | /** |
40064aec | 365 | * pcpu_cnt_pop_pages- counts populated backing pages in range |
833af842 | 366 | * @chunk: chunk of interest |
40064aec DZF |
367 | * @bit_off: start offset |
368 | * @bits: size of area to check | |
9f7dcf22 | 369 | * |
40064aec DZF |
370 | * Calculates the number of populated pages in the region |
371 | * [page_start, page_end). This keeps track of how many empty populated | |
372 | * pages are available and decide if async work should be scheduled. | |
ccea34b5 | 373 | * |
9f7dcf22 | 374 | * RETURNS: |
40064aec | 375 | * The nr of populated pages. |
9f7dcf22 | 376 | */ |
40064aec DZF |
377 | static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off, |
378 | int bits) | |
9f7dcf22 | 379 | { |
40064aec DZF |
380 | int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE); |
381 | int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); | |
4f996e23 | 382 | |
40064aec | 383 | if (page_start >= page_end) |
9f7dcf22 TH |
384 | return 0; |
385 | ||
40064aec DZF |
386 | /* |
387 | * bitmap_weight counts the number of bits set in a bitmap up to | |
388 | * the specified number of bits. This is counting the populated | |
389 | * pages up to page_end and then subtracting the populated pages | |
390 | * up to page_start to count the populated pages in | |
391 | * [page_start, page_end). | |
392 | */ | |
393 | return bitmap_weight(chunk->populated, page_end) - | |
394 | bitmap_weight(chunk->populated, page_start); | |
833af842 TH |
395 | } |
396 | ||
397 | /** | |
40064aec | 398 | * pcpu_chunk_update - updates the chunk metadata given a free area |
833af842 | 399 | * @chunk: chunk of interest |
40064aec DZF |
400 | * @bit_off: chunk offset |
401 | * @bits: size of free area | |
833af842 | 402 | * |
13f96637 | 403 | * This updates the chunk's contig hint and starting offset given a free area. |
40064aec DZF |
404 | */ |
405 | static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits) | |
406 | { | |
13f96637 DZF |
407 | if (bits > chunk->contig_bits) { |
408 | chunk->contig_bits_start = bit_off; | |
40064aec | 409 | chunk->contig_bits = bits; |
13f96637 | 410 | } |
40064aec DZF |
411 | } |
412 | ||
413 | /** | |
414 | * pcpu_chunk_refresh_hint - updates metadata about a chunk | |
415 | * @chunk: chunk of interest | |
833af842 | 416 | * |
40064aec | 417 | * Iterates over the chunk to find the largest free area. |
833af842 | 418 | * |
40064aec DZF |
419 | * Updates: |
420 | * chunk->contig_bits | |
13f96637 | 421 | * chunk->contig_bits_start |
40064aec | 422 | * nr_empty_pop_pages |
833af842 | 423 | */ |
40064aec | 424 | static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk) |
833af842 | 425 | { |
40064aec DZF |
426 | int bits, nr_empty_pop_pages; |
427 | int rs, re; /* region start, region end */ | |
833af842 | 428 | |
40064aec DZF |
429 | /* clear metadata */ |
430 | chunk->contig_bits = 0; | |
6710e594 | 431 | |
40064aec | 432 | bits = nr_empty_pop_pages = 0; |
86b442fb | 433 | pcpu_for_each_unpop_region(chunk->alloc_map, rs, re, chunk->first_bit, |
40064aec DZF |
434 | pcpu_chunk_map_bits(chunk)) { |
435 | bits = re - rs; | |
ccea34b5 | 436 | |
40064aec | 437 | pcpu_chunk_update(chunk, rs, bits); |
833af842 | 438 | |
40064aec DZF |
439 | nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, rs, bits); |
440 | } | |
9f7dcf22 | 441 | |
40064aec DZF |
442 | /* |
443 | * Keep track of nr_empty_pop_pages. | |
444 | * | |
445 | * The chunk maintains the previous number of free pages it held, | |
446 | * so the delta is used to update the global counter. The reserved | |
447 | * chunk is not part of the free page count as they are populated | |
448 | * at init and are special to serving reserved allocations. | |
449 | */ | |
450 | if (chunk != pcpu_reserved_chunk) | |
451 | pcpu_nr_empty_pop_pages += | |
452 | (nr_empty_pop_pages - chunk->nr_empty_pop_pages); | |
a002d148 | 453 | |
40064aec DZF |
454 | chunk->nr_empty_pop_pages = nr_empty_pop_pages; |
455 | } | |
9f7dcf22 | 456 | |
ca460b3c DZF |
457 | /** |
458 | * pcpu_block_update - updates a block given a free area | |
459 | * @block: block of interest | |
460 | * @start: start offset in block | |
461 | * @end: end offset in block | |
462 | * | |
463 | * Updates a block given a known free area. The region [start, end) is | |
464 | * expected to be the entirety of the free area within a block. | |
465 | */ | |
466 | static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) | |
467 | { | |
468 | int contig = end - start; | |
469 | ||
470 | block->first_free = min(block->first_free, start); | |
471 | if (start == 0) | |
472 | block->left_free = contig; | |
473 | ||
474 | if (end == PCPU_BITMAP_BLOCK_BITS) | |
475 | block->right_free = contig; | |
476 | ||
477 | if (contig > block->contig_hint) { | |
478 | block->contig_hint_start = start; | |
479 | block->contig_hint = contig; | |
480 | } | |
481 | } | |
482 | ||
483 | /** | |
484 | * pcpu_block_refresh_hint | |
485 | * @chunk: chunk of interest | |
486 | * @index: index of the metadata block | |
487 | * | |
488 | * Scans over the block beginning at first_free and updates the block | |
489 | * metadata accordingly. | |
490 | */ | |
491 | static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) | |
492 | { | |
493 | struct pcpu_block_md *block = chunk->md_blocks + index; | |
494 | unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); | |
495 | int rs, re; /* region start, region end */ | |
496 | ||
497 | /* clear hints */ | |
498 | block->contig_hint = 0; | |
499 | block->left_free = block->right_free = 0; | |
500 | ||
501 | /* iterate over free areas and update the contig hints */ | |
502 | pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free, | |
503 | PCPU_BITMAP_BLOCK_BITS) { | |
504 | pcpu_block_update(block, rs, re); | |
505 | } | |
506 | } | |
507 | ||
508 | /** | |
509 | * pcpu_block_update_hint_alloc - update hint on allocation path | |
510 | * @chunk: chunk of interest | |
511 | * @bit_off: chunk offset | |
512 | * @bits: size of request | |
513 | */ | |
514 | static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, | |
515 | int bits) | |
516 | { | |
517 | struct pcpu_block_md *s_block, *e_block, *block; | |
518 | int s_index, e_index; /* block indexes of the freed allocation */ | |
519 | int s_off, e_off; /* block offsets of the freed allocation */ | |
520 | ||
521 | /* | |
522 | * Calculate per block offsets. | |
523 | * The calculation uses an inclusive range, but the resulting offsets | |
524 | * are [start, end). e_index always points to the last block in the | |
525 | * range. | |
526 | */ | |
527 | s_index = pcpu_off_to_block_index(bit_off); | |
528 | e_index = pcpu_off_to_block_index(bit_off + bits - 1); | |
529 | s_off = pcpu_off_to_block_off(bit_off); | |
530 | e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; | |
531 | ||
532 | s_block = chunk->md_blocks + s_index; | |
533 | e_block = chunk->md_blocks + e_index; | |
534 | ||
535 | /* | |
536 | * Update s_block. | |
537 | */ | |
538 | pcpu_block_refresh_hint(chunk, s_index); | |
539 | ||
540 | /* | |
541 | * Update e_block. | |
542 | */ | |
543 | if (s_index != e_index) { | |
544 | pcpu_block_refresh_hint(chunk, e_index); | |
545 | ||
546 | /* update in-between md_blocks */ | |
547 | for (block = s_block + 1; block < e_block; block++) { | |
548 | block->contig_hint = 0; | |
549 | block->left_free = 0; | |
550 | block->right_free = 0; | |
551 | } | |
552 | } | |
553 | ||
554 | pcpu_chunk_refresh_hint(chunk); | |
555 | } | |
556 | ||
557 | /** | |
558 | * pcpu_block_update_hint_free - updates the block hints on the free path | |
559 | * @chunk: chunk of interest | |
560 | * @bit_off: chunk offset | |
561 | * @bits: size of request | |
562 | */ | |
563 | static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, | |
564 | int bits) | |
565 | { | |
566 | struct pcpu_block_md *s_block, *e_block, *block; | |
567 | int s_index, e_index; /* block indexes of the freed allocation */ | |
568 | int s_off, e_off; /* block offsets of the freed allocation */ | |
569 | ||
570 | /* | |
571 | * Calculate per block offsets. | |
572 | * The calculation uses an inclusive range, but the resulting offsets | |
573 | * are [start, end). e_index always points to the last block in the | |
574 | * range. | |
575 | */ | |
576 | s_index = pcpu_off_to_block_index(bit_off); | |
577 | e_index = pcpu_off_to_block_index(bit_off + bits - 1); | |
578 | s_off = pcpu_off_to_block_off(bit_off); | |
579 | e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; | |
580 | ||
581 | s_block = chunk->md_blocks + s_index; | |
582 | e_block = chunk->md_blocks + e_index; | |
583 | ||
584 | /* update s_block */ | |
585 | pcpu_block_refresh_hint(chunk, s_index); | |
586 | ||
587 | /* freeing in the same block */ | |
588 | if (s_index != e_index) { | |
589 | /* update e_block */ | |
590 | pcpu_block_refresh_hint(chunk, e_index); | |
591 | ||
592 | /* reset md_blocks in the middle */ | |
593 | for (block = s_block + 1; block < e_block; block++) { | |
594 | block->first_free = 0; | |
595 | block->contig_hint_start = 0; | |
596 | block->contig_hint = PCPU_BITMAP_BLOCK_BITS; | |
597 | block->left_free = PCPU_BITMAP_BLOCK_BITS; | |
598 | block->right_free = PCPU_BITMAP_BLOCK_BITS; | |
599 | } | |
600 | } | |
601 | ||
602 | pcpu_chunk_refresh_hint(chunk); | |
603 | } | |
604 | ||
40064aec DZF |
605 | /** |
606 | * pcpu_is_populated - determines if the region is populated | |
607 | * @chunk: chunk of interest | |
608 | * @bit_off: chunk offset | |
609 | * @bits: size of area | |
610 | * @next_off: return value for the next offset to start searching | |
611 | * | |
612 | * For atomic allocations, check if the backing pages are populated. | |
613 | * | |
614 | * RETURNS: | |
615 | * Bool if the backing pages are populated. | |
616 | * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. | |
617 | */ | |
618 | static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, | |
619 | int *next_off) | |
620 | { | |
621 | int page_start, page_end, rs, re; | |
833af842 | 622 | |
40064aec DZF |
623 | page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); |
624 | page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); | |
833af842 | 625 | |
40064aec DZF |
626 | rs = page_start; |
627 | pcpu_next_unpop(chunk->populated, &rs, &re, page_end); | |
628 | if (rs >= page_end) | |
629 | return true; | |
833af842 | 630 | |
40064aec DZF |
631 | *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; |
632 | return false; | |
9f7dcf22 TH |
633 | } |
634 | ||
a16037c8 | 635 | /** |
40064aec DZF |
636 | * pcpu_find_block_fit - finds the block index to start searching |
637 | * @chunk: chunk of interest | |
638 | * @alloc_bits: size of request in allocation units | |
639 | * @align: alignment of area (max PAGE_SIZE bytes) | |
640 | * @pop_only: use populated regions only | |
641 | * | |
642 | * RETURNS: | |
643 | * The offset in the bitmap to begin searching. | |
644 | * -1 if no offset is found. | |
a16037c8 | 645 | */ |
40064aec DZF |
646 | static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, |
647 | size_t align, bool pop_only) | |
a16037c8 | 648 | { |
40064aec DZF |
649 | int bit_off, bits; |
650 | int re; /* region end */ | |
a16037c8 | 651 | |
13f96637 DZF |
652 | /* |
653 | * Check to see if the allocation can fit in the chunk's contig hint. | |
654 | * This is an optimization to prevent scanning by assuming if it | |
655 | * cannot fit in the global hint, there is memory pressure and creating | |
656 | * a new chunk would happen soon. | |
657 | */ | |
658 | bit_off = ALIGN(chunk->contig_bits_start, align) - | |
659 | chunk->contig_bits_start; | |
660 | if (bit_off + alloc_bits > chunk->contig_bits) | |
661 | return -1; | |
662 | ||
86b442fb DZF |
663 | pcpu_for_each_unpop_region(chunk->alloc_map, bit_off, re, |
664 | chunk->first_bit, | |
40064aec DZF |
665 | pcpu_chunk_map_bits(chunk)) { |
666 | bits = re - bit_off; | |
a16037c8 | 667 | |
40064aec DZF |
668 | /* check alignment */ |
669 | bits -= ALIGN(bit_off, align) - bit_off; | |
670 | bit_off = ALIGN(bit_off, align); | |
671 | if (bits < alloc_bits) | |
672 | continue; | |
a16037c8 | 673 | |
40064aec DZF |
674 | bits = alloc_bits; |
675 | if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, | |
676 | &bit_off)) | |
677 | break; | |
a16037c8 | 678 | |
40064aec | 679 | bits = 0; |
a16037c8 | 680 | } |
40064aec DZF |
681 | |
682 | if (bit_off == pcpu_chunk_map_bits(chunk)) | |
683 | return -1; | |
684 | ||
685 | return bit_off; | |
a16037c8 TH |
686 | } |
687 | ||
fbf59bc9 | 688 | /** |
40064aec | 689 | * pcpu_alloc_area - allocates an area from a pcpu_chunk |
fbf59bc9 | 690 | * @chunk: chunk of interest |
40064aec DZF |
691 | * @alloc_bits: size of request in allocation units |
692 | * @align: alignment of area (max PAGE_SIZE) | |
693 | * @start: bit_off to start searching | |
9f7dcf22 | 694 | * |
40064aec DZF |
695 | * This function takes in a @start offset to begin searching to fit an |
696 | * allocation of @alloc_bits with alignment @align. If it confirms a | |
697 | * valid free area, it then updates the allocation and boundary maps | |
698 | * accordingly. | |
ccea34b5 | 699 | * |
fbf59bc9 | 700 | * RETURNS: |
40064aec DZF |
701 | * Allocated addr offset in @chunk on success. |
702 | * -1 if no matching area is found. | |
fbf59bc9 | 703 | */ |
40064aec DZF |
704 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, |
705 | size_t align, int start) | |
fbf59bc9 | 706 | { |
40064aec DZF |
707 | size_t align_mask = (align) ? (align - 1) : 0; |
708 | int bit_off, end, oslot; | |
a16037c8 | 709 | |
40064aec | 710 | lockdep_assert_held(&pcpu_lock); |
fbf59bc9 | 711 | |
40064aec | 712 | oslot = pcpu_chunk_slot(chunk); |
fbf59bc9 | 713 | |
40064aec DZF |
714 | /* |
715 | * Search to find a fit. | |
716 | */ | |
717 | end = start + alloc_bits; | |
718 | bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start, | |
719 | alloc_bits, align_mask); | |
720 | if (bit_off >= end) | |
721 | return -1; | |
fbf59bc9 | 722 | |
40064aec DZF |
723 | /* update alloc map */ |
724 | bitmap_set(chunk->alloc_map, bit_off, alloc_bits); | |
3d331ad7 | 725 | |
40064aec DZF |
726 | /* update boundary map */ |
727 | set_bit(bit_off, chunk->bound_map); | |
728 | bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); | |
729 | set_bit(bit_off + alloc_bits, chunk->bound_map); | |
fbf59bc9 | 730 | |
40064aec | 731 | chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; |
fbf59bc9 | 732 | |
86b442fb DZF |
733 | /* update first free bit */ |
734 | if (bit_off == chunk->first_bit) | |
735 | chunk->first_bit = find_next_zero_bit( | |
736 | chunk->alloc_map, | |
737 | pcpu_chunk_map_bits(chunk), | |
738 | bit_off + alloc_bits); | |
739 | ||
ca460b3c | 740 | pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); |
fbf59bc9 | 741 | |
fbf59bc9 TH |
742 | pcpu_chunk_relocate(chunk, oslot); |
743 | ||
40064aec | 744 | return bit_off * PCPU_MIN_ALLOC_SIZE; |
fbf59bc9 TH |
745 | } |
746 | ||
747 | /** | |
40064aec | 748 | * pcpu_free_area - frees the corresponding offset |
fbf59bc9 | 749 | * @chunk: chunk of interest |
40064aec | 750 | * @off: addr offset into chunk |
ccea34b5 | 751 | * |
40064aec DZF |
752 | * This function determines the size of an allocation to free using |
753 | * the boundary bitmap and clears the allocation map. | |
fbf59bc9 | 754 | */ |
40064aec | 755 | static void pcpu_free_area(struct pcpu_chunk *chunk, int off) |
fbf59bc9 | 756 | { |
40064aec | 757 | int bit_off, bits, end, oslot; |
723ad1d9 | 758 | |
5ccd30e4 | 759 | lockdep_assert_held(&pcpu_lock); |
30a5b536 | 760 | pcpu_stats_area_dealloc(chunk); |
5ccd30e4 | 761 | |
40064aec | 762 | oslot = pcpu_chunk_slot(chunk); |
fbf59bc9 | 763 | |
40064aec | 764 | bit_off = off / PCPU_MIN_ALLOC_SIZE; |
3d331ad7 | 765 | |
40064aec DZF |
766 | /* find end index */ |
767 | end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), | |
768 | bit_off + 1); | |
769 | bits = end - bit_off; | |
770 | bitmap_clear(chunk->alloc_map, bit_off, bits); | |
fbf59bc9 | 771 | |
40064aec DZF |
772 | /* update metadata */ |
773 | chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE; | |
b539b87f | 774 | |
86b442fb DZF |
775 | /* update first free bit */ |
776 | chunk->first_bit = min(chunk->first_bit, bit_off); | |
777 | ||
ca460b3c | 778 | pcpu_block_update_hint_free(chunk, bit_off, bits); |
fbf59bc9 | 779 | |
fbf59bc9 TH |
780 | pcpu_chunk_relocate(chunk, oslot); |
781 | } | |
782 | ||
ca460b3c DZF |
783 | static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) |
784 | { | |
785 | struct pcpu_block_md *md_block; | |
786 | ||
787 | for (md_block = chunk->md_blocks; | |
788 | md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); | |
789 | md_block++) { | |
790 | md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS; | |
791 | md_block->left_free = PCPU_BITMAP_BLOCK_BITS; | |
792 | md_block->right_free = PCPU_BITMAP_BLOCK_BITS; | |
793 | } | |
794 | } | |
795 | ||
40064aec DZF |
796 | /** |
797 | * pcpu_alloc_first_chunk - creates chunks that serve the first chunk | |
798 | * @tmp_addr: the start of the region served | |
799 | * @map_size: size of the region served | |
800 | * | |
801 | * This is responsible for creating the chunks that serve the first chunk. The | |
802 | * base_addr is page aligned down of @tmp_addr while the region end is page | |
803 | * aligned up. Offsets are kept track of to determine the region served. All | |
804 | * this is done to appease the bitmap allocator in avoiding partial blocks. | |
805 | * | |
806 | * RETURNS: | |
807 | * Chunk serving the region at @tmp_addr of @map_size. | |
808 | */ | |
c0ebfdc3 | 809 | static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, |
40064aec | 810 | int map_size) |
10edf5b0 DZF |
811 | { |
812 | struct pcpu_chunk *chunk; | |
ca460b3c | 813 | unsigned long aligned_addr, lcm_align; |
40064aec | 814 | int start_offset, offset_bits, region_size, region_bits; |
c0ebfdc3 DZF |
815 | |
816 | /* region calculations */ | |
817 | aligned_addr = tmp_addr & PAGE_MASK; | |
818 | ||
819 | start_offset = tmp_addr - aligned_addr; | |
6b9d7c8e | 820 | |
ca460b3c DZF |
821 | /* |
822 | * Align the end of the region with the LCM of PAGE_SIZE and | |
823 | * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of | |
824 | * the other. | |
825 | */ | |
826 | lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); | |
827 | region_size = ALIGN(start_offset + map_size, lcm_align); | |
10edf5b0 | 828 | |
c0ebfdc3 | 829 | /* allocate chunk */ |
8ab16c43 DZF |
830 | chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) + |
831 | BITS_TO_LONGS(region_size >> PAGE_SHIFT), | |
832 | 0); | |
c0ebfdc3 | 833 | |
10edf5b0 | 834 | INIT_LIST_HEAD(&chunk->list); |
c0ebfdc3 DZF |
835 | |
836 | chunk->base_addr = (void *)aligned_addr; | |
10edf5b0 | 837 | chunk->start_offset = start_offset; |
6b9d7c8e | 838 | chunk->end_offset = region_size - chunk->start_offset - map_size; |
c0ebfdc3 | 839 | |
8ab16c43 | 840 | chunk->nr_pages = region_size >> PAGE_SHIFT; |
40064aec | 841 | region_bits = pcpu_chunk_map_bits(chunk); |
c0ebfdc3 | 842 | |
ca460b3c DZF |
843 | chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) * |
844 | sizeof(chunk->alloc_map[0]), 0); | |
845 | chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) * | |
846 | sizeof(chunk->bound_map[0]), 0); | |
847 | chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) * | |
848 | sizeof(chunk->md_blocks[0]), 0); | |
849 | pcpu_init_md_blocks(chunk); | |
10edf5b0 DZF |
850 | |
851 | /* manage populated page bitmap */ | |
852 | chunk->immutable = true; | |
8ab16c43 DZF |
853 | bitmap_fill(chunk->populated, chunk->nr_pages); |
854 | chunk->nr_populated = chunk->nr_pages; | |
40064aec DZF |
855 | chunk->nr_empty_pop_pages = |
856 | pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE, | |
857 | map_size / PCPU_MIN_ALLOC_SIZE); | |
10edf5b0 | 858 | |
40064aec DZF |
859 | chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE; |
860 | chunk->free_bytes = map_size; | |
c0ebfdc3 DZF |
861 | |
862 | if (chunk->start_offset) { | |
863 | /* hide the beginning of the bitmap */ | |
40064aec DZF |
864 | offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; |
865 | bitmap_set(chunk->alloc_map, 0, offset_bits); | |
866 | set_bit(0, chunk->bound_map); | |
867 | set_bit(offset_bits, chunk->bound_map); | |
ca460b3c | 868 | |
86b442fb DZF |
869 | chunk->first_bit = offset_bits; |
870 | ||
ca460b3c | 871 | pcpu_block_update_hint_alloc(chunk, 0, offset_bits); |
c0ebfdc3 DZF |
872 | } |
873 | ||
6b9d7c8e DZF |
874 | if (chunk->end_offset) { |
875 | /* hide the end of the bitmap */ | |
40064aec DZF |
876 | offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; |
877 | bitmap_set(chunk->alloc_map, | |
878 | pcpu_chunk_map_bits(chunk) - offset_bits, | |
879 | offset_bits); | |
880 | set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, | |
881 | chunk->bound_map); | |
882 | set_bit(region_bits, chunk->bound_map); | |
6b9d7c8e | 883 | |
ca460b3c DZF |
884 | pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) |
885 | - offset_bits, offset_bits); | |
886 | } | |
40064aec | 887 | |
10edf5b0 DZF |
888 | return chunk; |
889 | } | |
890 | ||
6081089f TH |
891 | static struct pcpu_chunk *pcpu_alloc_chunk(void) |
892 | { | |
893 | struct pcpu_chunk *chunk; | |
40064aec | 894 | int region_bits; |
6081089f | 895 | |
90459ce0 | 896 | chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); |
6081089f TH |
897 | if (!chunk) |
898 | return NULL; | |
899 | ||
40064aec DZF |
900 | INIT_LIST_HEAD(&chunk->list); |
901 | chunk->nr_pages = pcpu_unit_pages; | |
902 | region_bits = pcpu_chunk_map_bits(chunk); | |
6081089f | 903 | |
40064aec DZF |
904 | chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * |
905 | sizeof(chunk->alloc_map[0])); | |
906 | if (!chunk->alloc_map) | |
907 | goto alloc_map_fail; | |
6081089f | 908 | |
40064aec DZF |
909 | chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * |
910 | sizeof(chunk->bound_map[0])); | |
911 | if (!chunk->bound_map) | |
912 | goto bound_map_fail; | |
6081089f | 913 | |
ca460b3c DZF |
914 | chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * |
915 | sizeof(chunk->md_blocks[0])); | |
916 | if (!chunk->md_blocks) | |
917 | goto md_blocks_fail; | |
918 | ||
919 | pcpu_init_md_blocks(chunk); | |
920 | ||
40064aec DZF |
921 | /* init metadata */ |
922 | chunk->contig_bits = region_bits; | |
923 | chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; | |
c0ebfdc3 | 924 | |
6081089f | 925 | return chunk; |
40064aec | 926 | |
ca460b3c DZF |
927 | md_blocks_fail: |
928 | pcpu_mem_free(chunk->bound_map); | |
40064aec DZF |
929 | bound_map_fail: |
930 | pcpu_mem_free(chunk->alloc_map); | |
931 | alloc_map_fail: | |
932 | pcpu_mem_free(chunk); | |
933 | ||
934 | return NULL; | |
6081089f TH |
935 | } |
936 | ||
937 | static void pcpu_free_chunk(struct pcpu_chunk *chunk) | |
938 | { | |
939 | if (!chunk) | |
940 | return; | |
40064aec DZF |
941 | pcpu_mem_free(chunk->bound_map); |
942 | pcpu_mem_free(chunk->alloc_map); | |
1d5cfdb0 | 943 | pcpu_mem_free(chunk); |
6081089f TH |
944 | } |
945 | ||
b539b87f TH |
946 | /** |
947 | * pcpu_chunk_populated - post-population bookkeeping | |
948 | * @chunk: pcpu_chunk which got populated | |
949 | * @page_start: the start page | |
950 | * @page_end: the end page | |
40064aec | 951 | * @for_alloc: if this is to populate for allocation |
b539b87f TH |
952 | * |
953 | * Pages in [@page_start,@page_end) have been populated to @chunk. Update | |
954 | * the bookkeeping information accordingly. Must be called after each | |
955 | * successful population. | |
40064aec DZF |
956 | * |
957 | * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it | |
958 | * is to serve an allocation in that area. | |
b539b87f | 959 | */ |
40064aec DZF |
960 | static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, |
961 | int page_end, bool for_alloc) | |
b539b87f TH |
962 | { |
963 | int nr = page_end - page_start; | |
964 | ||
965 | lockdep_assert_held(&pcpu_lock); | |
966 | ||
967 | bitmap_set(chunk->populated, page_start, nr); | |
968 | chunk->nr_populated += nr; | |
40064aec DZF |
969 | |
970 | if (!for_alloc) { | |
971 | chunk->nr_empty_pop_pages += nr; | |
972 | pcpu_nr_empty_pop_pages += nr; | |
973 | } | |
b539b87f TH |
974 | } |
975 | ||
976 | /** | |
977 | * pcpu_chunk_depopulated - post-depopulation bookkeeping | |
978 | * @chunk: pcpu_chunk which got depopulated | |
979 | * @page_start: the start page | |
980 | * @page_end: the end page | |
981 | * | |
982 | * Pages in [@page_start,@page_end) have been depopulated from @chunk. | |
983 | * Update the bookkeeping information accordingly. Must be called after | |
984 | * each successful depopulation. | |
985 | */ | |
986 | static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, | |
987 | int page_start, int page_end) | |
988 | { | |
989 | int nr = page_end - page_start; | |
990 | ||
991 | lockdep_assert_held(&pcpu_lock); | |
992 | ||
993 | bitmap_clear(chunk->populated, page_start, nr); | |
994 | chunk->nr_populated -= nr; | |
0cecf50c | 995 | chunk->nr_empty_pop_pages -= nr; |
b539b87f TH |
996 | pcpu_nr_empty_pop_pages -= nr; |
997 | } | |
998 | ||
9f645532 TH |
999 | /* |
1000 | * Chunk management implementation. | |
1001 | * | |
1002 | * To allow different implementations, chunk alloc/free and | |
1003 | * [de]population are implemented in a separate file which is pulled | |
1004 | * into this file and compiled together. The following functions | |
1005 | * should be implemented. | |
1006 | * | |
1007 | * pcpu_populate_chunk - populate the specified range of a chunk | |
1008 | * pcpu_depopulate_chunk - depopulate the specified range of a chunk | |
1009 | * pcpu_create_chunk - create a new chunk | |
1010 | * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop | |
1011 | * pcpu_addr_to_page - translate address to physical address | |
1012 | * pcpu_verify_alloc_info - check alloc_info is acceptable during init | |
fbf59bc9 | 1013 | */ |
9f645532 TH |
1014 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); |
1015 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); | |
1016 | static struct pcpu_chunk *pcpu_create_chunk(void); | |
1017 | static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); | |
1018 | static struct page *pcpu_addr_to_page(void *addr); | |
1019 | static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); | |
fbf59bc9 | 1020 | |
b0c9778b TH |
1021 | #ifdef CONFIG_NEED_PER_CPU_KM |
1022 | #include "percpu-km.c" | |
1023 | #else | |
9f645532 | 1024 | #include "percpu-vm.c" |
b0c9778b | 1025 | #endif |
fbf59bc9 | 1026 | |
88999a89 TH |
1027 | /** |
1028 | * pcpu_chunk_addr_search - determine chunk containing specified address | |
1029 | * @addr: address for which the chunk needs to be determined. | |
1030 | * | |
c0ebfdc3 DZF |
1031 | * This is an internal function that handles all but static allocations. |
1032 | * Static percpu address values should never be passed into the allocator. | |
1033 | * | |
88999a89 TH |
1034 | * RETURNS: |
1035 | * The address of the found chunk. | |
1036 | */ | |
1037 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
1038 | { | |
c0ebfdc3 | 1039 | /* is it in the dynamic region (first chunk)? */ |
560f2c23 | 1040 | if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) |
88999a89 | 1041 | return pcpu_first_chunk; |
c0ebfdc3 DZF |
1042 | |
1043 | /* is it in the reserved region? */ | |
560f2c23 | 1044 | if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) |
c0ebfdc3 | 1045 | return pcpu_reserved_chunk; |
88999a89 TH |
1046 | |
1047 | /* | |
1048 | * The address is relative to unit0 which might be unused and | |
1049 | * thus unmapped. Offset the address to the unit space of the | |
1050 | * current processor before looking it up in the vmalloc | |
1051 | * space. Note that any possible cpu id can be used here, so | |
1052 | * there's no need to worry about preemption or cpu hotplug. | |
1053 | */ | |
1054 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; | |
9f645532 | 1055 | return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); |
88999a89 TH |
1056 | } |
1057 | ||
fbf59bc9 | 1058 | /** |
edcb4639 | 1059 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 1060 | * @size: size of area to allocate in bytes |
fbf59bc9 | 1061 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 1062 | * @reserved: allocate from the reserved chunk if available |
5835d96e | 1063 | * @gfp: allocation flags |
fbf59bc9 | 1064 | * |
5835d96e TH |
1065 | * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't |
1066 | * contain %GFP_KERNEL, the allocation is atomic. | |
fbf59bc9 TH |
1067 | * |
1068 | * RETURNS: | |
1069 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1070 | */ | |
5835d96e TH |
1071 | static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, |
1072 | gfp_t gfp) | |
fbf59bc9 | 1073 | { |
f2badb0c | 1074 | static int warn_limit = 10; |
fbf59bc9 | 1075 | struct pcpu_chunk *chunk; |
f2badb0c | 1076 | const char *err; |
6ae833c7 | 1077 | bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; |
40064aec | 1078 | int slot, off, cpu, ret; |
403a91b1 | 1079 | unsigned long flags; |
f528f0b8 | 1080 | void __percpu *ptr; |
40064aec | 1081 | size_t bits, bit_align; |
fbf59bc9 | 1082 | |
723ad1d9 | 1083 | /* |
40064aec DZF |
1084 | * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, |
1085 | * therefore alignment must be a minimum of that many bytes. | |
1086 | * An allocation may have internal fragmentation from rounding up | |
1087 | * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. | |
723ad1d9 | 1088 | */ |
d2f3c384 DZF |
1089 | if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) |
1090 | align = PCPU_MIN_ALLOC_SIZE; | |
723ad1d9 | 1091 | |
d2f3c384 | 1092 | size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); |
40064aec DZF |
1093 | bits = size >> PCPU_MIN_ALLOC_SHIFT; |
1094 | bit_align = align >> PCPU_MIN_ALLOC_SHIFT; | |
2f69fa82 | 1095 | |
3ca45a46 | 1096 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || |
1097 | !is_power_of_2(align))) { | |
756a025f JP |
1098 | WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n", |
1099 | size, align); | |
fbf59bc9 TH |
1100 | return NULL; |
1101 | } | |
1102 | ||
6710e594 TH |
1103 | if (!is_atomic) |
1104 | mutex_lock(&pcpu_alloc_mutex); | |
1105 | ||
403a91b1 | 1106 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 | 1107 | |
edcb4639 TH |
1108 | /* serve reserved allocations from the reserved chunk if available */ |
1109 | if (reserved && pcpu_reserved_chunk) { | |
1110 | chunk = pcpu_reserved_chunk; | |
833af842 | 1111 | |
40064aec DZF |
1112 | off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); |
1113 | if (off < 0) { | |
833af842 | 1114 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 1115 | goto fail_unlock; |
f2badb0c | 1116 | } |
833af842 | 1117 | |
40064aec | 1118 | off = pcpu_alloc_area(chunk, bits, bit_align, off); |
edcb4639 TH |
1119 | if (off >= 0) |
1120 | goto area_found; | |
833af842 | 1121 | |
f2badb0c | 1122 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 1123 | goto fail_unlock; |
edcb4639 TH |
1124 | } |
1125 | ||
ccea34b5 | 1126 | restart: |
edcb4639 | 1127 | /* search through normal chunks */ |
fbf59bc9 TH |
1128 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
1129 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
40064aec DZF |
1130 | off = pcpu_find_block_fit(chunk, bits, bit_align, |
1131 | is_atomic); | |
1132 | if (off < 0) | |
fbf59bc9 | 1133 | continue; |
ccea34b5 | 1134 | |
40064aec | 1135 | off = pcpu_alloc_area(chunk, bits, bit_align, off); |
fbf59bc9 TH |
1136 | if (off >= 0) |
1137 | goto area_found; | |
40064aec | 1138 | |
fbf59bc9 TH |
1139 | } |
1140 | } | |
1141 | ||
403a91b1 | 1142 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1143 | |
b38d08f3 TH |
1144 | /* |
1145 | * No space left. Create a new chunk. We don't want multiple | |
1146 | * tasks to create chunks simultaneously. Serialize and create iff | |
1147 | * there's still no empty chunk after grabbing the mutex. | |
1148 | */ | |
11df02bf DZ |
1149 | if (is_atomic) { |
1150 | err = "atomic alloc failed, no space left"; | |
5835d96e | 1151 | goto fail; |
11df02bf | 1152 | } |
5835d96e | 1153 | |
b38d08f3 TH |
1154 | if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { |
1155 | chunk = pcpu_create_chunk(); | |
1156 | if (!chunk) { | |
1157 | err = "failed to allocate new chunk"; | |
1158 | goto fail; | |
1159 | } | |
1160 | ||
1161 | spin_lock_irqsave(&pcpu_lock, flags); | |
1162 | pcpu_chunk_relocate(chunk, -1); | |
1163 | } else { | |
1164 | spin_lock_irqsave(&pcpu_lock, flags); | |
f2badb0c | 1165 | } |
ccea34b5 | 1166 | |
ccea34b5 | 1167 | goto restart; |
fbf59bc9 TH |
1168 | |
1169 | area_found: | |
30a5b536 | 1170 | pcpu_stats_area_alloc(chunk, size); |
403a91b1 | 1171 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1172 | |
dca49645 | 1173 | /* populate if not all pages are already there */ |
5835d96e | 1174 | if (!is_atomic) { |
e04d3208 | 1175 | int page_start, page_end, rs, re; |
dca49645 | 1176 | |
e04d3208 TH |
1177 | page_start = PFN_DOWN(off); |
1178 | page_end = PFN_UP(off + size); | |
b38d08f3 | 1179 | |
91e914c5 DZF |
1180 | pcpu_for_each_unpop_region(chunk->populated, rs, re, |
1181 | page_start, page_end) { | |
e04d3208 TH |
1182 | WARN_ON(chunk->immutable); |
1183 | ||
1184 | ret = pcpu_populate_chunk(chunk, rs, re); | |
1185 | ||
1186 | spin_lock_irqsave(&pcpu_lock, flags); | |
1187 | if (ret) { | |
40064aec | 1188 | pcpu_free_area(chunk, off); |
e04d3208 TH |
1189 | err = "failed to populate"; |
1190 | goto fail_unlock; | |
1191 | } | |
40064aec | 1192 | pcpu_chunk_populated(chunk, rs, re, true); |
e04d3208 | 1193 | spin_unlock_irqrestore(&pcpu_lock, flags); |
dca49645 | 1194 | } |
fbf59bc9 | 1195 | |
e04d3208 TH |
1196 | mutex_unlock(&pcpu_alloc_mutex); |
1197 | } | |
ccea34b5 | 1198 | |
1a4d7607 TH |
1199 | if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) |
1200 | pcpu_schedule_balance_work(); | |
1201 | ||
dca49645 TH |
1202 | /* clear the areas and return address relative to base address */ |
1203 | for_each_possible_cpu(cpu) | |
1204 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); | |
1205 | ||
f528f0b8 | 1206 | ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); |
8a8c35fa | 1207 | kmemleak_alloc_percpu(ptr, size, gfp); |
df95e795 DZ |
1208 | |
1209 | trace_percpu_alloc_percpu(reserved, is_atomic, size, align, | |
1210 | chunk->base_addr, off, ptr); | |
1211 | ||
f528f0b8 | 1212 | return ptr; |
ccea34b5 TH |
1213 | |
1214 | fail_unlock: | |
403a91b1 | 1215 | spin_unlock_irqrestore(&pcpu_lock, flags); |
b38d08f3 | 1216 | fail: |
df95e795 DZ |
1217 | trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); |
1218 | ||
5835d96e | 1219 | if (!is_atomic && warn_limit) { |
870d4b12 | 1220 | pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", |
598d8091 | 1221 | size, align, is_atomic, err); |
f2badb0c TH |
1222 | dump_stack(); |
1223 | if (!--warn_limit) | |
870d4b12 | 1224 | pr_info("limit reached, disable warning\n"); |
f2badb0c | 1225 | } |
1a4d7607 TH |
1226 | if (is_atomic) { |
1227 | /* see the flag handling in pcpu_blance_workfn() */ | |
1228 | pcpu_atomic_alloc_failed = true; | |
1229 | pcpu_schedule_balance_work(); | |
6710e594 TH |
1230 | } else { |
1231 | mutex_unlock(&pcpu_alloc_mutex); | |
1a4d7607 | 1232 | } |
ccea34b5 | 1233 | return NULL; |
fbf59bc9 | 1234 | } |
edcb4639 TH |
1235 | |
1236 | /** | |
5835d96e | 1237 | * __alloc_percpu_gfp - allocate dynamic percpu area |
edcb4639 TH |
1238 | * @size: size of area to allocate in bytes |
1239 | * @align: alignment of area (max PAGE_SIZE) | |
5835d96e | 1240 | * @gfp: allocation flags |
edcb4639 | 1241 | * |
5835d96e TH |
1242 | * Allocate zero-filled percpu area of @size bytes aligned at @align. If |
1243 | * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can | |
1244 | * be called from any context but is a lot more likely to fail. | |
ccea34b5 | 1245 | * |
edcb4639 TH |
1246 | * RETURNS: |
1247 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1248 | */ | |
5835d96e TH |
1249 | void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) |
1250 | { | |
1251 | return pcpu_alloc(size, align, false, gfp); | |
1252 | } | |
1253 | EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); | |
1254 | ||
1255 | /** | |
1256 | * __alloc_percpu - allocate dynamic percpu area | |
1257 | * @size: size of area to allocate in bytes | |
1258 | * @align: alignment of area (max PAGE_SIZE) | |
1259 | * | |
1260 | * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). | |
1261 | */ | |
43cf38eb | 1262 | void __percpu *__alloc_percpu(size_t size, size_t align) |
edcb4639 | 1263 | { |
5835d96e | 1264 | return pcpu_alloc(size, align, false, GFP_KERNEL); |
edcb4639 | 1265 | } |
fbf59bc9 TH |
1266 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1267 | ||
edcb4639 TH |
1268 | /** |
1269 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1270 | * @size: size of area to allocate in bytes | |
1271 | * @align: alignment of area (max PAGE_SIZE) | |
1272 | * | |
9329ba97 TH |
1273 | * Allocate zero-filled percpu area of @size bytes aligned at @align |
1274 | * from reserved percpu area if arch has set it up; otherwise, | |
1275 | * allocation is served from the same dynamic area. Might sleep. | |
1276 | * Might trigger writeouts. | |
edcb4639 | 1277 | * |
ccea34b5 TH |
1278 | * CONTEXT: |
1279 | * Does GFP_KERNEL allocation. | |
1280 | * | |
edcb4639 TH |
1281 | * RETURNS: |
1282 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1283 | */ | |
43cf38eb | 1284 | void __percpu *__alloc_reserved_percpu(size_t size, size_t align) |
edcb4639 | 1285 | { |
5835d96e | 1286 | return pcpu_alloc(size, align, true, GFP_KERNEL); |
edcb4639 TH |
1287 | } |
1288 | ||
a56dbddf | 1289 | /** |
1a4d7607 | 1290 | * pcpu_balance_workfn - manage the amount of free chunks and populated pages |
a56dbddf TH |
1291 | * @work: unused |
1292 | * | |
1293 | * Reclaim all fully free chunks except for the first one. | |
1294 | */ | |
fe6bd8c3 | 1295 | static void pcpu_balance_workfn(struct work_struct *work) |
fbf59bc9 | 1296 | { |
fe6bd8c3 TH |
1297 | LIST_HEAD(to_free); |
1298 | struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; | |
a56dbddf | 1299 | struct pcpu_chunk *chunk, *next; |
1a4d7607 | 1300 | int slot, nr_to_pop, ret; |
a56dbddf | 1301 | |
1a4d7607 TH |
1302 | /* |
1303 | * There's no reason to keep around multiple unused chunks and VM | |
1304 | * areas can be scarce. Destroy all free chunks except for one. | |
1305 | */ | |
ccea34b5 TH |
1306 | mutex_lock(&pcpu_alloc_mutex); |
1307 | spin_lock_irq(&pcpu_lock); | |
a56dbddf | 1308 | |
fe6bd8c3 | 1309 | list_for_each_entry_safe(chunk, next, free_head, list) { |
a56dbddf TH |
1310 | WARN_ON(chunk->immutable); |
1311 | ||
1312 | /* spare the first one */ | |
fe6bd8c3 | 1313 | if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) |
a56dbddf TH |
1314 | continue; |
1315 | ||
fe6bd8c3 | 1316 | list_move(&chunk->list, &to_free); |
a56dbddf TH |
1317 | } |
1318 | ||
ccea34b5 | 1319 | spin_unlock_irq(&pcpu_lock); |
a56dbddf | 1320 | |
fe6bd8c3 | 1321 | list_for_each_entry_safe(chunk, next, &to_free, list) { |
a93ace48 | 1322 | int rs, re; |
dca49645 | 1323 | |
91e914c5 DZF |
1324 | pcpu_for_each_pop_region(chunk->populated, rs, re, 0, |
1325 | chunk->nr_pages) { | |
a93ace48 | 1326 | pcpu_depopulate_chunk(chunk, rs, re); |
b539b87f TH |
1327 | spin_lock_irq(&pcpu_lock); |
1328 | pcpu_chunk_depopulated(chunk, rs, re); | |
1329 | spin_unlock_irq(&pcpu_lock); | |
a93ace48 | 1330 | } |
6081089f | 1331 | pcpu_destroy_chunk(chunk); |
a56dbddf | 1332 | } |
971f3918 | 1333 | |
1a4d7607 TH |
1334 | /* |
1335 | * Ensure there are certain number of free populated pages for | |
1336 | * atomic allocs. Fill up from the most packed so that atomic | |
1337 | * allocs don't increase fragmentation. If atomic allocation | |
1338 | * failed previously, always populate the maximum amount. This | |
1339 | * should prevent atomic allocs larger than PAGE_SIZE from keeping | |
1340 | * failing indefinitely; however, large atomic allocs are not | |
1341 | * something we support properly and can be highly unreliable and | |
1342 | * inefficient. | |
1343 | */ | |
1344 | retry_pop: | |
1345 | if (pcpu_atomic_alloc_failed) { | |
1346 | nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; | |
1347 | /* best effort anyway, don't worry about synchronization */ | |
1348 | pcpu_atomic_alloc_failed = false; | |
1349 | } else { | |
1350 | nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - | |
1351 | pcpu_nr_empty_pop_pages, | |
1352 | 0, PCPU_EMPTY_POP_PAGES_HIGH); | |
1353 | } | |
1354 | ||
1355 | for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { | |
1356 | int nr_unpop = 0, rs, re; | |
1357 | ||
1358 | if (!nr_to_pop) | |
1359 | break; | |
1360 | ||
1361 | spin_lock_irq(&pcpu_lock); | |
1362 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
8ab16c43 | 1363 | nr_unpop = chunk->nr_pages - chunk->nr_populated; |
1a4d7607 TH |
1364 | if (nr_unpop) |
1365 | break; | |
1366 | } | |
1367 | spin_unlock_irq(&pcpu_lock); | |
1368 | ||
1369 | if (!nr_unpop) | |
1370 | continue; | |
1371 | ||
1372 | /* @chunk can't go away while pcpu_alloc_mutex is held */ | |
91e914c5 DZF |
1373 | pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, |
1374 | chunk->nr_pages) { | |
1a4d7607 TH |
1375 | int nr = min(re - rs, nr_to_pop); |
1376 | ||
1377 | ret = pcpu_populate_chunk(chunk, rs, rs + nr); | |
1378 | if (!ret) { | |
1379 | nr_to_pop -= nr; | |
1380 | spin_lock_irq(&pcpu_lock); | |
40064aec | 1381 | pcpu_chunk_populated(chunk, rs, rs + nr, false); |
1a4d7607 TH |
1382 | spin_unlock_irq(&pcpu_lock); |
1383 | } else { | |
1384 | nr_to_pop = 0; | |
1385 | } | |
1386 | ||
1387 | if (!nr_to_pop) | |
1388 | break; | |
1389 | } | |
1390 | } | |
1391 | ||
1392 | if (nr_to_pop) { | |
1393 | /* ran out of chunks to populate, create a new one and retry */ | |
1394 | chunk = pcpu_create_chunk(); | |
1395 | if (chunk) { | |
1396 | spin_lock_irq(&pcpu_lock); | |
1397 | pcpu_chunk_relocate(chunk, -1); | |
1398 | spin_unlock_irq(&pcpu_lock); | |
1399 | goto retry_pop; | |
1400 | } | |
1401 | } | |
1402 | ||
971f3918 | 1403 | mutex_unlock(&pcpu_alloc_mutex); |
fbf59bc9 TH |
1404 | } |
1405 | ||
1406 | /** | |
1407 | * free_percpu - free percpu area | |
1408 | * @ptr: pointer to area to free | |
1409 | * | |
ccea34b5 TH |
1410 | * Free percpu area @ptr. |
1411 | * | |
1412 | * CONTEXT: | |
1413 | * Can be called from atomic context. | |
fbf59bc9 | 1414 | */ |
43cf38eb | 1415 | void free_percpu(void __percpu *ptr) |
fbf59bc9 | 1416 | { |
129182e5 | 1417 | void *addr; |
fbf59bc9 | 1418 | struct pcpu_chunk *chunk; |
ccea34b5 | 1419 | unsigned long flags; |
40064aec | 1420 | int off; |
fbf59bc9 TH |
1421 | |
1422 | if (!ptr) | |
1423 | return; | |
1424 | ||
f528f0b8 CM |
1425 | kmemleak_free_percpu(ptr); |
1426 | ||
129182e5 AM |
1427 | addr = __pcpu_ptr_to_addr(ptr); |
1428 | ||
ccea34b5 | 1429 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1430 | |
1431 | chunk = pcpu_chunk_addr_search(addr); | |
bba174f5 | 1432 | off = addr - chunk->base_addr; |
fbf59bc9 | 1433 | |
40064aec | 1434 | pcpu_free_area(chunk, off); |
fbf59bc9 | 1435 | |
a56dbddf | 1436 | /* if there are more than one fully free chunks, wake up grim reaper */ |
40064aec | 1437 | if (chunk->free_bytes == pcpu_unit_size) { |
fbf59bc9 TH |
1438 | struct pcpu_chunk *pos; |
1439 | ||
a56dbddf | 1440 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1441 | if (pos != chunk) { |
1a4d7607 | 1442 | pcpu_schedule_balance_work(); |
fbf59bc9 TH |
1443 | break; |
1444 | } | |
1445 | } | |
1446 | ||
df95e795 DZ |
1447 | trace_percpu_free_percpu(chunk->base_addr, off, ptr); |
1448 | ||
ccea34b5 | 1449 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1450 | } |
1451 | EXPORT_SYMBOL_GPL(free_percpu); | |
1452 | ||
383776fa | 1453 | bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) |
10fad5e4 | 1454 | { |
bbddff05 | 1455 | #ifdef CONFIG_SMP |
10fad5e4 TH |
1456 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
1457 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); | |
1458 | unsigned int cpu; | |
1459 | ||
1460 | for_each_possible_cpu(cpu) { | |
1461 | void *start = per_cpu_ptr(base, cpu); | |
383776fa | 1462 | void *va = (void *)addr; |
10fad5e4 | 1463 | |
383776fa | 1464 | if (va >= start && va < start + static_size) { |
8ce371f9 | 1465 | if (can_addr) { |
383776fa | 1466 | *can_addr = (unsigned long) (va - start); |
8ce371f9 PZ |
1467 | *can_addr += (unsigned long) |
1468 | per_cpu_ptr(base, get_boot_cpu_id()); | |
1469 | } | |
10fad5e4 | 1470 | return true; |
383776fa TG |
1471 | } |
1472 | } | |
bbddff05 TH |
1473 | #endif |
1474 | /* on UP, can't distinguish from other static vars, always false */ | |
10fad5e4 TH |
1475 | return false; |
1476 | } | |
1477 | ||
383776fa TG |
1478 | /** |
1479 | * is_kernel_percpu_address - test whether address is from static percpu area | |
1480 | * @addr: address to test | |
1481 | * | |
1482 | * Test whether @addr belongs to in-kernel static percpu area. Module | |
1483 | * static percpu areas are not considered. For those, use | |
1484 | * is_module_percpu_address(). | |
1485 | * | |
1486 | * RETURNS: | |
1487 | * %true if @addr is from in-kernel static percpu area, %false otherwise. | |
1488 | */ | |
1489 | bool is_kernel_percpu_address(unsigned long addr) | |
1490 | { | |
1491 | return __is_kernel_percpu_address(addr, NULL); | |
1492 | } | |
1493 | ||
3b034b0d VG |
1494 | /** |
1495 | * per_cpu_ptr_to_phys - convert translated percpu address to physical address | |
1496 | * @addr: the address to be converted to physical address | |
1497 | * | |
1498 | * Given @addr which is dereferenceable address obtained via one of | |
1499 | * percpu access macros, this function translates it into its physical | |
1500 | * address. The caller is responsible for ensuring @addr stays valid | |
1501 | * until this function finishes. | |
1502 | * | |
67589c71 DY |
1503 | * percpu allocator has special setup for the first chunk, which currently |
1504 | * supports either embedding in linear address space or vmalloc mapping, | |
1505 | * and, from the second one, the backing allocator (currently either vm or | |
1506 | * km) provides translation. | |
1507 | * | |
bffc4375 | 1508 | * The addr can be translated simply without checking if it falls into the |
67589c71 DY |
1509 | * first chunk. But the current code reflects better how percpu allocator |
1510 | * actually works, and the verification can discover both bugs in percpu | |
1511 | * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current | |
1512 | * code. | |
1513 | * | |
3b034b0d VG |
1514 | * RETURNS: |
1515 | * The physical address for @addr. | |
1516 | */ | |
1517 | phys_addr_t per_cpu_ptr_to_phys(void *addr) | |
1518 | { | |
9983b6f0 TH |
1519 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
1520 | bool in_first_chunk = false; | |
a855b84c | 1521 | unsigned long first_low, first_high; |
9983b6f0 TH |
1522 | unsigned int cpu; |
1523 | ||
1524 | /* | |
a855b84c | 1525 | * The following test on unit_low/high isn't strictly |
9983b6f0 TH |
1526 | * necessary but will speed up lookups of addresses which |
1527 | * aren't in the first chunk. | |
c0ebfdc3 DZF |
1528 | * |
1529 | * The address check is against full chunk sizes. pcpu_base_addr | |
1530 | * points to the beginning of the first chunk including the | |
1531 | * static region. Assumes good intent as the first chunk may | |
1532 | * not be full (ie. < pcpu_unit_pages in size). | |
9983b6f0 | 1533 | */ |
c0ebfdc3 DZF |
1534 | first_low = (unsigned long)pcpu_base_addr + |
1535 | pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); | |
1536 | first_high = (unsigned long)pcpu_base_addr + | |
1537 | pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); | |
a855b84c TH |
1538 | if ((unsigned long)addr >= first_low && |
1539 | (unsigned long)addr < first_high) { | |
9983b6f0 TH |
1540 | for_each_possible_cpu(cpu) { |
1541 | void *start = per_cpu_ptr(base, cpu); | |
1542 | ||
1543 | if (addr >= start && addr < start + pcpu_unit_size) { | |
1544 | in_first_chunk = true; | |
1545 | break; | |
1546 | } | |
1547 | } | |
1548 | } | |
1549 | ||
1550 | if (in_first_chunk) { | |
eac522ef | 1551 | if (!is_vmalloc_addr(addr)) |
020ec653 TH |
1552 | return __pa(addr); |
1553 | else | |
9f57bd4d ES |
1554 | return page_to_phys(vmalloc_to_page(addr)) + |
1555 | offset_in_page(addr); | |
020ec653 | 1556 | } else |
9f57bd4d ES |
1557 | return page_to_phys(pcpu_addr_to_page(addr)) + |
1558 | offset_in_page(addr); | |
3b034b0d VG |
1559 | } |
1560 | ||
fbf59bc9 | 1561 | /** |
fd1e8a1f TH |
1562 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
1563 | * @nr_groups: the number of groups | |
1564 | * @nr_units: the number of units | |
1565 | * | |
1566 | * Allocate ai which is large enough for @nr_groups groups containing | |
1567 | * @nr_units units. The returned ai's groups[0].cpu_map points to the | |
1568 | * cpu_map array which is long enough for @nr_units and filled with | |
1569 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map | |
1570 | * pointer of other groups. | |
1571 | * | |
1572 | * RETURNS: | |
1573 | * Pointer to the allocated pcpu_alloc_info on success, NULL on | |
1574 | * failure. | |
1575 | */ | |
1576 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, | |
1577 | int nr_units) | |
1578 | { | |
1579 | struct pcpu_alloc_info *ai; | |
1580 | size_t base_size, ai_size; | |
1581 | void *ptr; | |
1582 | int unit; | |
1583 | ||
1584 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), | |
1585 | __alignof__(ai->groups[0].cpu_map[0])); | |
1586 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); | |
1587 | ||
999c17e3 | 1588 | ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); |
fd1e8a1f TH |
1589 | if (!ptr) |
1590 | return NULL; | |
1591 | ai = ptr; | |
1592 | ptr += base_size; | |
1593 | ||
1594 | ai->groups[0].cpu_map = ptr; | |
1595 | ||
1596 | for (unit = 0; unit < nr_units; unit++) | |
1597 | ai->groups[0].cpu_map[unit] = NR_CPUS; | |
1598 | ||
1599 | ai->nr_groups = nr_groups; | |
1600 | ai->__ai_size = PFN_ALIGN(ai_size); | |
1601 | ||
1602 | return ai; | |
1603 | } | |
1604 | ||
1605 | /** | |
1606 | * pcpu_free_alloc_info - free percpu allocation info | |
1607 | * @ai: pcpu_alloc_info to free | |
1608 | * | |
1609 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). | |
1610 | */ | |
1611 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) | |
1612 | { | |
999c17e3 | 1613 | memblock_free_early(__pa(ai), ai->__ai_size); |
fd1e8a1f TH |
1614 | } |
1615 | ||
fd1e8a1f TH |
1616 | /** |
1617 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info | |
1618 | * @lvl: loglevel | |
1619 | * @ai: allocation info to dump | |
1620 | * | |
1621 | * Print out information about @ai using loglevel @lvl. | |
1622 | */ | |
1623 | static void pcpu_dump_alloc_info(const char *lvl, | |
1624 | const struct pcpu_alloc_info *ai) | |
033e48fb | 1625 | { |
fd1e8a1f | 1626 | int group_width = 1, cpu_width = 1, width; |
033e48fb | 1627 | char empty_str[] = "--------"; |
fd1e8a1f TH |
1628 | int alloc = 0, alloc_end = 0; |
1629 | int group, v; | |
1630 | int upa, apl; /* units per alloc, allocs per line */ | |
1631 | ||
1632 | v = ai->nr_groups; | |
1633 | while (v /= 10) | |
1634 | group_width++; | |
033e48fb | 1635 | |
fd1e8a1f | 1636 | v = num_possible_cpus(); |
033e48fb | 1637 | while (v /= 10) |
fd1e8a1f TH |
1638 | cpu_width++; |
1639 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; | |
033e48fb | 1640 | |
fd1e8a1f TH |
1641 | upa = ai->alloc_size / ai->unit_size; |
1642 | width = upa * (cpu_width + 1) + group_width + 3; | |
1643 | apl = rounddown_pow_of_two(max(60 / width, 1)); | |
033e48fb | 1644 | |
fd1e8a1f TH |
1645 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
1646 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, | |
1647 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); | |
033e48fb | 1648 | |
fd1e8a1f TH |
1649 | for (group = 0; group < ai->nr_groups; group++) { |
1650 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
1651 | int unit = 0, unit_end = 0; | |
1652 | ||
1653 | BUG_ON(gi->nr_units % upa); | |
1654 | for (alloc_end += gi->nr_units / upa; | |
1655 | alloc < alloc_end; alloc++) { | |
1656 | if (!(alloc % apl)) { | |
1170532b | 1657 | pr_cont("\n"); |
fd1e8a1f TH |
1658 | printk("%spcpu-alloc: ", lvl); |
1659 | } | |
1170532b | 1660 | pr_cont("[%0*d] ", group_width, group); |
fd1e8a1f TH |
1661 | |
1662 | for (unit_end += upa; unit < unit_end; unit++) | |
1663 | if (gi->cpu_map[unit] != NR_CPUS) | |
1170532b JP |
1664 | pr_cont("%0*d ", |
1665 | cpu_width, gi->cpu_map[unit]); | |
fd1e8a1f | 1666 | else |
1170532b | 1667 | pr_cont("%s ", empty_str); |
033e48fb | 1668 | } |
033e48fb | 1669 | } |
1170532b | 1670 | pr_cont("\n"); |
033e48fb | 1671 | } |
033e48fb | 1672 | |
fbf59bc9 | 1673 | /** |
8d408b4b | 1674 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
fd1e8a1f | 1675 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
38a6be52 | 1676 | * @base_addr: mapped address |
8d408b4b TH |
1677 | * |
1678 | * Initialize the first percpu chunk which contains the kernel static | |
1679 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1680 | * setup path. |
8d408b4b | 1681 | * |
fd1e8a1f TH |
1682 | * @ai contains all information necessary to initialize the first |
1683 | * chunk and prime the dynamic percpu allocator. | |
1684 | * | |
1685 | * @ai->static_size is the size of static percpu area. | |
1686 | * | |
1687 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to | |
edcb4639 TH |
1688 | * reserve after the static area in the first chunk. This reserves |
1689 | * the first chunk such that it's available only through reserved | |
1690 | * percpu allocation. This is primarily used to serve module percpu | |
1691 | * static areas on architectures where the addressing model has | |
1692 | * limited offset range for symbol relocations to guarantee module | |
1693 | * percpu symbols fall inside the relocatable range. | |
1694 | * | |
fd1e8a1f TH |
1695 | * @ai->dyn_size determines the number of bytes available for dynamic |
1696 | * allocation in the first chunk. The area between @ai->static_size + | |
1697 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. | |
6074d5b0 | 1698 | * |
fd1e8a1f TH |
1699 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
1700 | * and equal to or larger than @ai->static_size + @ai->reserved_size + | |
1701 | * @ai->dyn_size. | |
8d408b4b | 1702 | * |
fd1e8a1f TH |
1703 | * @ai->atom_size is the allocation atom size and used as alignment |
1704 | * for vm areas. | |
8d408b4b | 1705 | * |
fd1e8a1f TH |
1706 | * @ai->alloc_size is the allocation size and always multiple of |
1707 | * @ai->atom_size. This is larger than @ai->atom_size if | |
1708 | * @ai->unit_size is larger than @ai->atom_size. | |
1709 | * | |
1710 | * @ai->nr_groups and @ai->groups describe virtual memory layout of | |
1711 | * percpu areas. Units which should be colocated are put into the | |
1712 | * same group. Dynamic VM areas will be allocated according to these | |
1713 | * groupings. If @ai->nr_groups is zero, a single group containing | |
1714 | * all units is assumed. | |
8d408b4b | 1715 | * |
38a6be52 TH |
1716 | * The caller should have mapped the first chunk at @base_addr and |
1717 | * copied static data to each unit. | |
fbf59bc9 | 1718 | * |
c0ebfdc3 DZF |
1719 | * The first chunk will always contain a static and a dynamic region. |
1720 | * However, the static region is not managed by any chunk. If the first | |
1721 | * chunk also contains a reserved region, it is served by two chunks - | |
1722 | * one for the reserved region and one for the dynamic region. They | |
1723 | * share the same vm, but use offset regions in the area allocation map. | |
1724 | * The chunk serving the dynamic region is circulated in the chunk slots | |
1725 | * and available for dynamic allocation like any other chunk. | |
edcb4639 | 1726 | * |
fbf59bc9 | 1727 | * RETURNS: |
fb435d52 | 1728 | * 0 on success, -errno on failure. |
fbf59bc9 | 1729 | */ |
fb435d52 TH |
1730 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
1731 | void *base_addr) | |
fbf59bc9 | 1732 | { |
b9c39442 | 1733 | size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
d2f3c384 | 1734 | size_t static_size, dyn_size; |
0c4169c3 | 1735 | struct pcpu_chunk *chunk; |
6563297c TH |
1736 | unsigned long *group_offsets; |
1737 | size_t *group_sizes; | |
fb435d52 | 1738 | unsigned long *unit_off; |
fbf59bc9 | 1739 | unsigned int cpu; |
fd1e8a1f TH |
1740 | int *unit_map; |
1741 | int group, unit, i; | |
c0ebfdc3 DZF |
1742 | int map_size; |
1743 | unsigned long tmp_addr; | |
fbf59bc9 | 1744 | |
635b75fc TH |
1745 | #define PCPU_SETUP_BUG_ON(cond) do { \ |
1746 | if (unlikely(cond)) { \ | |
870d4b12 JP |
1747 | pr_emerg("failed to initialize, %s\n", #cond); \ |
1748 | pr_emerg("cpu_possible_mask=%*pb\n", \ | |
807de073 | 1749 | cpumask_pr_args(cpu_possible_mask)); \ |
635b75fc TH |
1750 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ |
1751 | BUG(); \ | |
1752 | } \ | |
1753 | } while (0) | |
1754 | ||
2f39e637 | 1755 | /* sanity checks */ |
635b75fc | 1756 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
bbddff05 | 1757 | #ifdef CONFIG_SMP |
635b75fc | 1758 | PCPU_SETUP_BUG_ON(!ai->static_size); |
f09f1243 | 1759 | PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); |
bbddff05 | 1760 | #endif |
635b75fc | 1761 | PCPU_SETUP_BUG_ON(!base_addr); |
f09f1243 | 1762 | PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); |
635b75fc | 1763 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
f09f1243 | 1764 | PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); |
635b75fc | 1765 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); |
ca460b3c | 1766 | PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); |
099a19d9 | 1767 | PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); |
fb29a2cc | 1768 | PCPU_SETUP_BUG_ON(!ai->dyn_size); |
d2f3c384 | 1769 | PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); |
ca460b3c DZF |
1770 | PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || |
1771 | IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); | |
9f645532 | 1772 | PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); |
8d408b4b | 1773 | |
6563297c | 1774 | /* process group information and build config tables accordingly */ |
999c17e3 SS |
1775 | group_offsets = memblock_virt_alloc(ai->nr_groups * |
1776 | sizeof(group_offsets[0]), 0); | |
1777 | group_sizes = memblock_virt_alloc(ai->nr_groups * | |
1778 | sizeof(group_sizes[0]), 0); | |
1779 | unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); | |
1780 | unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); | |
2f39e637 | 1781 | |
fd1e8a1f | 1782 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
ffe0d5a5 | 1783 | unit_map[cpu] = UINT_MAX; |
a855b84c TH |
1784 | |
1785 | pcpu_low_unit_cpu = NR_CPUS; | |
1786 | pcpu_high_unit_cpu = NR_CPUS; | |
2f39e637 | 1787 | |
fd1e8a1f TH |
1788 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
1789 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
2f39e637 | 1790 | |
6563297c TH |
1791 | group_offsets[group] = gi->base_offset; |
1792 | group_sizes[group] = gi->nr_units * ai->unit_size; | |
1793 | ||
fd1e8a1f TH |
1794 | for (i = 0; i < gi->nr_units; i++) { |
1795 | cpu = gi->cpu_map[i]; | |
1796 | if (cpu == NR_CPUS) | |
1797 | continue; | |
8d408b4b | 1798 | |
9f295664 | 1799 | PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); |
635b75fc TH |
1800 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); |
1801 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); | |
fbf59bc9 | 1802 | |
fd1e8a1f | 1803 | unit_map[cpu] = unit + i; |
fb435d52 TH |
1804 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
1805 | ||
a855b84c TH |
1806 | /* determine low/high unit_cpu */ |
1807 | if (pcpu_low_unit_cpu == NR_CPUS || | |
1808 | unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) | |
1809 | pcpu_low_unit_cpu = cpu; | |
1810 | if (pcpu_high_unit_cpu == NR_CPUS || | |
1811 | unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) | |
1812 | pcpu_high_unit_cpu = cpu; | |
fd1e8a1f | 1813 | } |
2f39e637 | 1814 | } |
fd1e8a1f TH |
1815 | pcpu_nr_units = unit; |
1816 | ||
1817 | for_each_possible_cpu(cpu) | |
635b75fc TH |
1818 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
1819 | ||
1820 | /* we're done parsing the input, undefine BUG macro and dump config */ | |
1821 | #undef PCPU_SETUP_BUG_ON | |
bcbea798 | 1822 | pcpu_dump_alloc_info(KERN_DEBUG, ai); |
fd1e8a1f | 1823 | |
6563297c TH |
1824 | pcpu_nr_groups = ai->nr_groups; |
1825 | pcpu_group_offsets = group_offsets; | |
1826 | pcpu_group_sizes = group_sizes; | |
fd1e8a1f | 1827 | pcpu_unit_map = unit_map; |
fb435d52 | 1828 | pcpu_unit_offsets = unit_off; |
2f39e637 TH |
1829 | |
1830 | /* determine basic parameters */ | |
fd1e8a1f | 1831 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
d9b55eeb | 1832 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
6563297c | 1833 | pcpu_atom_size = ai->atom_size; |
ce3141a2 TH |
1834 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
1835 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
cafe8816 | 1836 | |
30a5b536 DZ |
1837 | pcpu_stats_save_ai(ai); |
1838 | ||
d9b55eeb TH |
1839 | /* |
1840 | * Allocate chunk slots. The additional last slot is for | |
1841 | * empty chunks. | |
1842 | */ | |
1843 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
999c17e3 SS |
1844 | pcpu_slot = memblock_virt_alloc( |
1845 | pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); | |
fbf59bc9 TH |
1846 | for (i = 0; i < pcpu_nr_slots; i++) |
1847 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1848 | ||
d2f3c384 DZF |
1849 | /* |
1850 | * The end of the static region needs to be aligned with the | |
1851 | * minimum allocation size as this offsets the reserved and | |
1852 | * dynamic region. The first chunk ends page aligned by | |
1853 | * expanding the dynamic region, therefore the dynamic region | |
1854 | * can be shrunk to compensate while still staying above the | |
1855 | * configured sizes. | |
1856 | */ | |
1857 | static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); | |
1858 | dyn_size = ai->dyn_size - (static_size - ai->static_size); | |
1859 | ||
edcb4639 | 1860 | /* |
c0ebfdc3 DZF |
1861 | * Initialize first chunk. |
1862 | * If the reserved_size is non-zero, this initializes the reserved | |
1863 | * chunk. If the reserved_size is zero, the reserved chunk is NULL | |
1864 | * and the dynamic region is initialized here. The first chunk, | |
1865 | * pcpu_first_chunk, will always point to the chunk that serves | |
1866 | * the dynamic region. | |
edcb4639 | 1867 | */ |
d2f3c384 DZF |
1868 | tmp_addr = (unsigned long)base_addr + static_size; |
1869 | map_size = ai->reserved_size ?: dyn_size; | |
40064aec | 1870 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); |
61ace7fa | 1871 | |
edcb4639 | 1872 | /* init dynamic chunk if necessary */ |
b9c39442 | 1873 | if (ai->reserved_size) { |
0c4169c3 | 1874 | pcpu_reserved_chunk = chunk; |
b9c39442 | 1875 | |
d2f3c384 | 1876 | tmp_addr = (unsigned long)base_addr + static_size + |
c0ebfdc3 | 1877 | ai->reserved_size; |
d2f3c384 | 1878 | map_size = dyn_size; |
40064aec | 1879 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); |
edcb4639 TH |
1880 | } |
1881 | ||
2441d15c | 1882 | /* link the first chunk in */ |
0c4169c3 | 1883 | pcpu_first_chunk = chunk; |
0cecf50c | 1884 | pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; |
ae9e6bc9 | 1885 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
fbf59bc9 | 1886 | |
30a5b536 | 1887 | pcpu_stats_chunk_alloc(); |
df95e795 | 1888 | trace_percpu_create_chunk(base_addr); |
30a5b536 | 1889 | |
fbf59bc9 | 1890 | /* we're done */ |
bba174f5 | 1891 | pcpu_base_addr = base_addr; |
fb435d52 | 1892 | return 0; |
fbf59bc9 | 1893 | } |
66c3a757 | 1894 | |
bbddff05 TH |
1895 | #ifdef CONFIG_SMP |
1896 | ||
17f3609c | 1897 | const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { |
f58dc01b TH |
1898 | [PCPU_FC_AUTO] = "auto", |
1899 | [PCPU_FC_EMBED] = "embed", | |
1900 | [PCPU_FC_PAGE] = "page", | |
f58dc01b | 1901 | }; |
66c3a757 | 1902 | |
f58dc01b | 1903 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
66c3a757 | 1904 | |
f58dc01b TH |
1905 | static int __init percpu_alloc_setup(char *str) |
1906 | { | |
5479c78a CG |
1907 | if (!str) |
1908 | return -EINVAL; | |
1909 | ||
f58dc01b TH |
1910 | if (0) |
1911 | /* nada */; | |
1912 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK | |
1913 | else if (!strcmp(str, "embed")) | |
1914 | pcpu_chosen_fc = PCPU_FC_EMBED; | |
1915 | #endif | |
1916 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
1917 | else if (!strcmp(str, "page")) | |
1918 | pcpu_chosen_fc = PCPU_FC_PAGE; | |
f58dc01b TH |
1919 | #endif |
1920 | else | |
870d4b12 | 1921 | pr_warn("unknown allocator %s specified\n", str); |
66c3a757 | 1922 | |
f58dc01b | 1923 | return 0; |
66c3a757 | 1924 | } |
f58dc01b | 1925 | early_param("percpu_alloc", percpu_alloc_setup); |
66c3a757 | 1926 | |
3c9a024f TH |
1927 | /* |
1928 | * pcpu_embed_first_chunk() is used by the generic percpu setup. | |
1929 | * Build it if needed by the arch config or the generic setup is going | |
1930 | * to be used. | |
1931 | */ | |
08fc4580 TH |
1932 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
1933 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
3c9a024f TH |
1934 | #define BUILD_EMBED_FIRST_CHUNK |
1935 | #endif | |
1936 | ||
1937 | /* build pcpu_page_first_chunk() iff needed by the arch config */ | |
1938 | #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) | |
1939 | #define BUILD_PAGE_FIRST_CHUNK | |
1940 | #endif | |
1941 | ||
1942 | /* pcpu_build_alloc_info() is used by both embed and page first chunk */ | |
1943 | #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) | |
1944 | /** | |
1945 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs | |
1946 | * @reserved_size: the size of reserved percpu area in bytes | |
1947 | * @dyn_size: minimum free size for dynamic allocation in bytes | |
1948 | * @atom_size: allocation atom size | |
1949 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
1950 | * | |
1951 | * This function determines grouping of units, their mappings to cpus | |
1952 | * and other parameters considering needed percpu size, allocation | |
1953 | * atom size and distances between CPUs. | |
1954 | * | |
bffc4375 | 1955 | * Groups are always multiples of atom size and CPUs which are of |
3c9a024f TH |
1956 | * LOCAL_DISTANCE both ways are grouped together and share space for |
1957 | * units in the same group. The returned configuration is guaranteed | |
1958 | * to have CPUs on different nodes on different groups and >=75% usage | |
1959 | * of allocated virtual address space. | |
1960 | * | |
1961 | * RETURNS: | |
1962 | * On success, pointer to the new allocation_info is returned. On | |
1963 | * failure, ERR_PTR value is returned. | |
1964 | */ | |
1965 | static struct pcpu_alloc_info * __init pcpu_build_alloc_info( | |
1966 | size_t reserved_size, size_t dyn_size, | |
1967 | size_t atom_size, | |
1968 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
1969 | { | |
1970 | static int group_map[NR_CPUS] __initdata; | |
1971 | static int group_cnt[NR_CPUS] __initdata; | |
1972 | const size_t static_size = __per_cpu_end - __per_cpu_start; | |
1973 | int nr_groups = 1, nr_units = 0; | |
1974 | size_t size_sum, min_unit_size, alloc_size; | |
1975 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
1976 | int last_allocs, group, unit; | |
1977 | unsigned int cpu, tcpu; | |
1978 | struct pcpu_alloc_info *ai; | |
1979 | unsigned int *cpu_map; | |
1980 | ||
1981 | /* this function may be called multiple times */ | |
1982 | memset(group_map, 0, sizeof(group_map)); | |
1983 | memset(group_cnt, 0, sizeof(group_cnt)); | |
1984 | ||
1985 | /* calculate size_sum and ensure dyn_size is enough for early alloc */ | |
1986 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
1987 | max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); | |
1988 | dyn_size = size_sum - static_size - reserved_size; | |
1989 | ||
1990 | /* | |
1991 | * Determine min_unit_size, alloc_size and max_upa such that | |
1992 | * alloc_size is multiple of atom_size and is the smallest | |
25985edc | 1993 | * which can accommodate 4k aligned segments which are equal to |
3c9a024f TH |
1994 | * or larger than min_unit_size. |
1995 | */ | |
1996 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
1997 | ||
9c015162 | 1998 | /* determine the maximum # of units that can fit in an allocation */ |
3c9a024f TH |
1999 | alloc_size = roundup(min_unit_size, atom_size); |
2000 | upa = alloc_size / min_unit_size; | |
f09f1243 | 2001 | while (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
2002 | upa--; |
2003 | max_upa = upa; | |
2004 | ||
2005 | /* group cpus according to their proximity */ | |
2006 | for_each_possible_cpu(cpu) { | |
2007 | group = 0; | |
2008 | next_group: | |
2009 | for_each_possible_cpu(tcpu) { | |
2010 | if (cpu == tcpu) | |
2011 | break; | |
2012 | if (group_map[tcpu] == group && cpu_distance_fn && | |
2013 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || | |
2014 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
2015 | group++; | |
2016 | nr_groups = max(nr_groups, group + 1); | |
2017 | goto next_group; | |
2018 | } | |
2019 | } | |
2020 | group_map[cpu] = group; | |
2021 | group_cnt[group]++; | |
2022 | } | |
2023 | ||
2024 | /* | |
9c015162 DZF |
2025 | * Wasted space is caused by a ratio imbalance of upa to group_cnt. |
2026 | * Expand the unit_size until we use >= 75% of the units allocated. | |
2027 | * Related to atom_size, which could be much larger than the unit_size. | |
3c9a024f TH |
2028 | */ |
2029 | last_allocs = INT_MAX; | |
2030 | for (upa = max_upa; upa; upa--) { | |
2031 | int allocs = 0, wasted = 0; | |
2032 | ||
f09f1243 | 2033 | if (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
2034 | continue; |
2035 | ||
2036 | for (group = 0; group < nr_groups; group++) { | |
2037 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); | |
2038 | allocs += this_allocs; | |
2039 | wasted += this_allocs * upa - group_cnt[group]; | |
2040 | } | |
2041 | ||
2042 | /* | |
2043 | * Don't accept if wastage is over 1/3. The | |
2044 | * greater-than comparison ensures upa==1 always | |
2045 | * passes the following check. | |
2046 | */ | |
2047 | if (wasted > num_possible_cpus() / 3) | |
2048 | continue; | |
2049 | ||
2050 | /* and then don't consume more memory */ | |
2051 | if (allocs > last_allocs) | |
2052 | break; | |
2053 | last_allocs = allocs; | |
2054 | best_upa = upa; | |
2055 | } | |
2056 | upa = best_upa; | |
2057 | ||
2058 | /* allocate and fill alloc_info */ | |
2059 | for (group = 0; group < nr_groups; group++) | |
2060 | nr_units += roundup(group_cnt[group], upa); | |
2061 | ||
2062 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); | |
2063 | if (!ai) | |
2064 | return ERR_PTR(-ENOMEM); | |
2065 | cpu_map = ai->groups[0].cpu_map; | |
2066 | ||
2067 | for (group = 0; group < nr_groups; group++) { | |
2068 | ai->groups[group].cpu_map = cpu_map; | |
2069 | cpu_map += roundup(group_cnt[group], upa); | |
2070 | } | |
2071 | ||
2072 | ai->static_size = static_size; | |
2073 | ai->reserved_size = reserved_size; | |
2074 | ai->dyn_size = dyn_size; | |
2075 | ai->unit_size = alloc_size / upa; | |
2076 | ai->atom_size = atom_size; | |
2077 | ai->alloc_size = alloc_size; | |
2078 | ||
2079 | for (group = 0, unit = 0; group_cnt[group]; group++) { | |
2080 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2081 | ||
2082 | /* | |
2083 | * Initialize base_offset as if all groups are located | |
2084 | * back-to-back. The caller should update this to | |
2085 | * reflect actual allocation. | |
2086 | */ | |
2087 | gi->base_offset = unit * ai->unit_size; | |
2088 | ||
2089 | for_each_possible_cpu(cpu) | |
2090 | if (group_map[cpu] == group) | |
2091 | gi->cpu_map[gi->nr_units++] = cpu; | |
2092 | gi->nr_units = roundup(gi->nr_units, upa); | |
2093 | unit += gi->nr_units; | |
2094 | } | |
2095 | BUG_ON(unit != nr_units); | |
2096 | ||
2097 | return ai; | |
2098 | } | |
2099 | #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ | |
2100 | ||
2101 | #if defined(BUILD_EMBED_FIRST_CHUNK) | |
66c3a757 TH |
2102 | /** |
2103 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
66c3a757 | 2104 | * @reserved_size: the size of reserved percpu area in bytes |
4ba6ce25 | 2105 | * @dyn_size: minimum free size for dynamic allocation in bytes |
c8826dd5 TH |
2106 | * @atom_size: allocation atom size |
2107 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
2108 | * @alloc_fn: function to allocate percpu page | |
25985edc | 2109 | * @free_fn: function to free percpu page |
66c3a757 TH |
2110 | * |
2111 | * This is a helper to ease setting up embedded first percpu chunk and | |
2112 | * can be called where pcpu_setup_first_chunk() is expected. | |
2113 | * | |
2114 | * If this function is used to setup the first chunk, it is allocated | |
c8826dd5 TH |
2115 | * by calling @alloc_fn and used as-is without being mapped into |
2116 | * vmalloc area. Allocations are always whole multiples of @atom_size | |
2117 | * aligned to @atom_size. | |
2118 | * | |
2119 | * This enables the first chunk to piggy back on the linear physical | |
2120 | * mapping which often uses larger page size. Please note that this | |
2121 | * can result in very sparse cpu->unit mapping on NUMA machines thus | |
2122 | * requiring large vmalloc address space. Don't use this allocator if | |
2123 | * vmalloc space is not orders of magnitude larger than distances | |
2124 | * between node memory addresses (ie. 32bit NUMA machines). | |
66c3a757 | 2125 | * |
4ba6ce25 | 2126 | * @dyn_size specifies the minimum dynamic area size. |
66c3a757 TH |
2127 | * |
2128 | * If the needed size is smaller than the minimum or specified unit | |
c8826dd5 | 2129 | * size, the leftover is returned using @free_fn. |
66c3a757 TH |
2130 | * |
2131 | * RETURNS: | |
fb435d52 | 2132 | * 0 on success, -errno on failure. |
66c3a757 | 2133 | */ |
4ba6ce25 | 2134 | int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, |
c8826dd5 TH |
2135 | size_t atom_size, |
2136 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, | |
2137 | pcpu_fc_alloc_fn_t alloc_fn, | |
2138 | pcpu_fc_free_fn_t free_fn) | |
66c3a757 | 2139 | { |
c8826dd5 TH |
2140 | void *base = (void *)ULONG_MAX; |
2141 | void **areas = NULL; | |
fd1e8a1f | 2142 | struct pcpu_alloc_info *ai; |
93c76b6b | 2143 | size_t size_sum, areas_size; |
2144 | unsigned long max_distance; | |
9b739662 | 2145 | int group, i, highest_group, rc; |
66c3a757 | 2146 | |
c8826dd5 TH |
2147 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
2148 | cpu_distance_fn); | |
fd1e8a1f TH |
2149 | if (IS_ERR(ai)) |
2150 | return PTR_ERR(ai); | |
66c3a757 | 2151 | |
fd1e8a1f | 2152 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
c8826dd5 | 2153 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
fa8a7094 | 2154 | |
999c17e3 | 2155 | areas = memblock_virt_alloc_nopanic(areas_size, 0); |
c8826dd5 | 2156 | if (!areas) { |
fb435d52 | 2157 | rc = -ENOMEM; |
c8826dd5 | 2158 | goto out_free; |
fa8a7094 | 2159 | } |
66c3a757 | 2160 | |
9b739662 | 2161 | /* allocate, copy and determine base address & max_distance */ |
2162 | highest_group = 0; | |
c8826dd5 TH |
2163 | for (group = 0; group < ai->nr_groups; group++) { |
2164 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2165 | unsigned int cpu = NR_CPUS; | |
2166 | void *ptr; | |
2167 | ||
2168 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) | |
2169 | cpu = gi->cpu_map[i]; | |
2170 | BUG_ON(cpu == NR_CPUS); | |
2171 | ||
2172 | /* allocate space for the whole group */ | |
2173 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); | |
2174 | if (!ptr) { | |
2175 | rc = -ENOMEM; | |
2176 | goto out_free_areas; | |
2177 | } | |
f528f0b8 CM |
2178 | /* kmemleak tracks the percpu allocations separately */ |
2179 | kmemleak_free(ptr); | |
c8826dd5 | 2180 | areas[group] = ptr; |
fd1e8a1f | 2181 | |
c8826dd5 | 2182 | base = min(ptr, base); |
9b739662 | 2183 | if (ptr > areas[highest_group]) |
2184 | highest_group = group; | |
2185 | } | |
2186 | max_distance = areas[highest_group] - base; | |
2187 | max_distance += ai->unit_size * ai->groups[highest_group].nr_units; | |
2188 | ||
2189 | /* warn if maximum distance is further than 75% of vmalloc space */ | |
2190 | if (max_distance > VMALLOC_TOTAL * 3 / 4) { | |
2191 | pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", | |
2192 | max_distance, VMALLOC_TOTAL); | |
2193 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
2194 | /* and fail if we have fallback */ | |
2195 | rc = -EINVAL; | |
2196 | goto out_free_areas; | |
2197 | #endif | |
42b64281 TH |
2198 | } |
2199 | ||
2200 | /* | |
2201 | * Copy data and free unused parts. This should happen after all | |
2202 | * allocations are complete; otherwise, we may end up with | |
2203 | * overlapping groups. | |
2204 | */ | |
2205 | for (group = 0; group < ai->nr_groups; group++) { | |
2206 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2207 | void *ptr = areas[group]; | |
c8826dd5 TH |
2208 | |
2209 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { | |
2210 | if (gi->cpu_map[i] == NR_CPUS) { | |
2211 | /* unused unit, free whole */ | |
2212 | free_fn(ptr, ai->unit_size); | |
2213 | continue; | |
2214 | } | |
2215 | /* copy and return the unused part */ | |
2216 | memcpy(ptr, __per_cpu_load, ai->static_size); | |
2217 | free_fn(ptr + size_sum, ai->unit_size - size_sum); | |
2218 | } | |
fa8a7094 | 2219 | } |
66c3a757 | 2220 | |
c8826dd5 | 2221 | /* base address is now known, determine group base offsets */ |
6ea529a2 | 2222 | for (group = 0; group < ai->nr_groups; group++) { |
c8826dd5 | 2223 | ai->groups[group].base_offset = areas[group] - base; |
6ea529a2 | 2224 | } |
c8826dd5 | 2225 | |
870d4b12 | 2226 | pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
fd1e8a1f TH |
2227 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
2228 | ai->dyn_size, ai->unit_size); | |
d4b95f80 | 2229 | |
fb435d52 | 2230 | rc = pcpu_setup_first_chunk(ai, base); |
c8826dd5 TH |
2231 | goto out_free; |
2232 | ||
2233 | out_free_areas: | |
2234 | for (group = 0; group < ai->nr_groups; group++) | |
f851c8d8 MH |
2235 | if (areas[group]) |
2236 | free_fn(areas[group], | |
2237 | ai->groups[group].nr_units * ai->unit_size); | |
c8826dd5 | 2238 | out_free: |
fd1e8a1f | 2239 | pcpu_free_alloc_info(ai); |
c8826dd5 | 2240 | if (areas) |
999c17e3 | 2241 | memblock_free_early(__pa(areas), areas_size); |
fb435d52 | 2242 | return rc; |
d4b95f80 | 2243 | } |
3c9a024f | 2244 | #endif /* BUILD_EMBED_FIRST_CHUNK */ |
d4b95f80 | 2245 | |
3c9a024f | 2246 | #ifdef BUILD_PAGE_FIRST_CHUNK |
d4b95f80 | 2247 | /** |
00ae4064 | 2248 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
2249 | * @reserved_size: the size of reserved percpu area in bytes |
2250 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
25985edc | 2251 | * @free_fn: function to free percpu page, always called with PAGE_SIZE |
d4b95f80 TH |
2252 | * @populate_pte_fn: function to populate pte |
2253 | * | |
00ae4064 TH |
2254 | * This is a helper to ease setting up page-remapped first percpu |
2255 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
2256 | * |
2257 | * This is the basic allocator. Static percpu area is allocated | |
2258 | * page-by-page into vmalloc area. | |
2259 | * | |
2260 | * RETURNS: | |
fb435d52 | 2261 | * 0 on success, -errno on failure. |
d4b95f80 | 2262 | */ |
fb435d52 TH |
2263 | int __init pcpu_page_first_chunk(size_t reserved_size, |
2264 | pcpu_fc_alloc_fn_t alloc_fn, | |
2265 | pcpu_fc_free_fn_t free_fn, | |
2266 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 2267 | { |
8f05a6a6 | 2268 | static struct vm_struct vm; |
fd1e8a1f | 2269 | struct pcpu_alloc_info *ai; |
00ae4064 | 2270 | char psize_str[16]; |
ce3141a2 | 2271 | int unit_pages; |
d4b95f80 | 2272 | size_t pages_size; |
ce3141a2 | 2273 | struct page **pages; |
fb435d52 | 2274 | int unit, i, j, rc; |
8f606604 | 2275 | int upa; |
2276 | int nr_g0_units; | |
d4b95f80 | 2277 | |
00ae4064 TH |
2278 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
2279 | ||
4ba6ce25 | 2280 | ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); |
fd1e8a1f TH |
2281 | if (IS_ERR(ai)) |
2282 | return PTR_ERR(ai); | |
2283 | BUG_ON(ai->nr_groups != 1); | |
8f606604 | 2284 | upa = ai->alloc_size/ai->unit_size; |
2285 | nr_g0_units = roundup(num_possible_cpus(), upa); | |
2286 | if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) { | |
2287 | pcpu_free_alloc_info(ai); | |
2288 | return -EINVAL; | |
2289 | } | |
fd1e8a1f TH |
2290 | |
2291 | unit_pages = ai->unit_size >> PAGE_SHIFT; | |
d4b95f80 TH |
2292 | |
2293 | /* unaligned allocations can't be freed, round up to page size */ | |
fd1e8a1f TH |
2294 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
2295 | sizeof(pages[0])); | |
999c17e3 | 2296 | pages = memblock_virt_alloc(pages_size, 0); |
d4b95f80 | 2297 | |
8f05a6a6 | 2298 | /* allocate pages */ |
d4b95f80 | 2299 | j = 0; |
8f606604 | 2300 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
2301 | unsigned int cpu = ai->groups[0].cpu_map[unit]; | |
ce3141a2 | 2302 | for (i = 0; i < unit_pages; i++) { |
d4b95f80 TH |
2303 | void *ptr; |
2304 | ||
3cbc8565 | 2305 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
d4b95f80 | 2306 | if (!ptr) { |
870d4b12 | 2307 | pr_warn("failed to allocate %s page for cpu%u\n", |
8f606604 | 2308 | psize_str, cpu); |
d4b95f80 TH |
2309 | goto enomem; |
2310 | } | |
f528f0b8 CM |
2311 | /* kmemleak tracks the percpu allocations separately */ |
2312 | kmemleak_free(ptr); | |
ce3141a2 | 2313 | pages[j++] = virt_to_page(ptr); |
d4b95f80 | 2314 | } |
8f606604 | 2315 | } |
d4b95f80 | 2316 | |
8f05a6a6 TH |
2317 | /* allocate vm area, map the pages and copy static data */ |
2318 | vm.flags = VM_ALLOC; | |
fd1e8a1f | 2319 | vm.size = num_possible_cpus() * ai->unit_size; |
8f05a6a6 TH |
2320 | vm_area_register_early(&vm, PAGE_SIZE); |
2321 | ||
fd1e8a1f | 2322 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
1d9d3257 | 2323 | unsigned long unit_addr = |
fd1e8a1f | 2324 | (unsigned long)vm.addr + unit * ai->unit_size; |
8f05a6a6 | 2325 | |
ce3141a2 | 2326 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
2327 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
2328 | ||
2329 | /* pte already populated, the following shouldn't fail */ | |
fb435d52 TH |
2330 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
2331 | unit_pages); | |
2332 | if (rc < 0) | |
2333 | panic("failed to map percpu area, err=%d\n", rc); | |
66c3a757 | 2334 | |
8f05a6a6 TH |
2335 | /* |
2336 | * FIXME: Archs with virtual cache should flush local | |
2337 | * cache for the linear mapping here - something | |
2338 | * equivalent to flush_cache_vmap() on the local cpu. | |
2339 | * flush_cache_vmap() can't be used as most supporting | |
2340 | * data structures are not set up yet. | |
2341 | */ | |
2342 | ||
2343 | /* copy static data */ | |
fd1e8a1f | 2344 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
66c3a757 TH |
2345 | } |
2346 | ||
2347 | /* we're ready, commit */ | |
870d4b12 | 2348 | pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", |
fd1e8a1f TH |
2349 | unit_pages, psize_str, vm.addr, ai->static_size, |
2350 | ai->reserved_size, ai->dyn_size); | |
d4b95f80 | 2351 | |
fb435d52 | 2352 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
d4b95f80 TH |
2353 | goto out_free_ar; |
2354 | ||
2355 | enomem: | |
2356 | while (--j >= 0) | |
ce3141a2 | 2357 | free_fn(page_address(pages[j]), PAGE_SIZE); |
fb435d52 | 2358 | rc = -ENOMEM; |
d4b95f80 | 2359 | out_free_ar: |
999c17e3 | 2360 | memblock_free_early(__pa(pages), pages_size); |
fd1e8a1f | 2361 | pcpu_free_alloc_info(ai); |
fb435d52 | 2362 | return rc; |
d4b95f80 | 2363 | } |
3c9a024f | 2364 | #endif /* BUILD_PAGE_FIRST_CHUNK */ |
d4b95f80 | 2365 | |
bbddff05 | 2366 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
e74e3962 | 2367 | /* |
bbddff05 | 2368 | * Generic SMP percpu area setup. |
e74e3962 TH |
2369 | * |
2370 | * The embedding helper is used because its behavior closely resembles | |
2371 | * the original non-dynamic generic percpu area setup. This is | |
2372 | * important because many archs have addressing restrictions and might | |
2373 | * fail if the percpu area is located far away from the previous | |
2374 | * location. As an added bonus, in non-NUMA cases, embedding is | |
2375 | * generally a good idea TLB-wise because percpu area can piggy back | |
2376 | * on the physical linear memory mapping which uses large page | |
2377 | * mappings on applicable archs. | |
2378 | */ | |
e74e3962 TH |
2379 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
2380 | EXPORT_SYMBOL(__per_cpu_offset); | |
2381 | ||
c8826dd5 TH |
2382 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
2383 | size_t align) | |
2384 | { | |
999c17e3 SS |
2385 | return memblock_virt_alloc_from_nopanic( |
2386 | size, align, __pa(MAX_DMA_ADDRESS)); | |
c8826dd5 | 2387 | } |
66c3a757 | 2388 | |
c8826dd5 TH |
2389 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
2390 | { | |
999c17e3 | 2391 | memblock_free_early(__pa(ptr), size); |
c8826dd5 TH |
2392 | } |
2393 | ||
e74e3962 TH |
2394 | void __init setup_per_cpu_areas(void) |
2395 | { | |
e74e3962 TH |
2396 | unsigned long delta; |
2397 | unsigned int cpu; | |
fb435d52 | 2398 | int rc; |
e74e3962 TH |
2399 | |
2400 | /* | |
2401 | * Always reserve area for module percpu variables. That's | |
2402 | * what the legacy allocator did. | |
2403 | */ | |
fb435d52 | 2404 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
c8826dd5 TH |
2405 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
2406 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); | |
fb435d52 | 2407 | if (rc < 0) |
bbddff05 | 2408 | panic("Failed to initialize percpu areas."); |
e74e3962 TH |
2409 | |
2410 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2411 | for_each_possible_cpu(cpu) | |
fb435d52 | 2412 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
66c3a757 | 2413 | } |
bbddff05 TH |
2414 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
2415 | ||
2416 | #else /* CONFIG_SMP */ | |
2417 | ||
2418 | /* | |
2419 | * UP percpu area setup. | |
2420 | * | |
2421 | * UP always uses km-based percpu allocator with identity mapping. | |
2422 | * Static percpu variables are indistinguishable from the usual static | |
2423 | * variables and don't require any special preparation. | |
2424 | */ | |
2425 | void __init setup_per_cpu_areas(void) | |
2426 | { | |
2427 | const size_t unit_size = | |
2428 | roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, | |
2429 | PERCPU_DYNAMIC_RESERVE)); | |
2430 | struct pcpu_alloc_info *ai; | |
2431 | void *fc; | |
2432 | ||
2433 | ai = pcpu_alloc_alloc_info(1, 1); | |
999c17e3 SS |
2434 | fc = memblock_virt_alloc_from_nopanic(unit_size, |
2435 | PAGE_SIZE, | |
2436 | __pa(MAX_DMA_ADDRESS)); | |
bbddff05 TH |
2437 | if (!ai || !fc) |
2438 | panic("Failed to allocate memory for percpu areas."); | |
100d13c3 CM |
2439 | /* kmemleak tracks the percpu allocations separately */ |
2440 | kmemleak_free(fc); | |
bbddff05 TH |
2441 | |
2442 | ai->dyn_size = unit_size; | |
2443 | ai->unit_size = unit_size; | |
2444 | ai->atom_size = unit_size; | |
2445 | ai->alloc_size = unit_size; | |
2446 | ai->groups[0].nr_units = 1; | |
2447 | ai->groups[0].cpu_map[0] = 0; | |
2448 | ||
2449 | if (pcpu_setup_first_chunk(ai, fc) < 0) | |
2450 | panic("Failed to initialize percpu areas."); | |
2451 | } | |
2452 | ||
2453 | #endif /* CONFIG_SMP */ | |
099a19d9 | 2454 | |
1a4d7607 TH |
2455 | /* |
2456 | * Percpu allocator is initialized early during boot when neither slab or | |
2457 | * workqueue is available. Plug async management until everything is up | |
2458 | * and running. | |
2459 | */ | |
2460 | static int __init percpu_enable_async(void) | |
2461 | { | |
2462 | pcpu_async_enabled = true; | |
2463 | return 0; | |
2464 | } | |
2465 | subsys_initcall(percpu_enable_async); |