]>
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 | ||
b185cd0d DZF |
303 | static unsigned long pcpu_block_off_to_off(int index, int off) |
304 | { | |
305 | return index * PCPU_BITMAP_BLOCK_BITS + off; | |
306 | } | |
307 | ||
525ca84d DZF |
308 | /** |
309 | * pcpu_next_md_free_region - finds the next hint free area | |
310 | * @chunk: chunk of interest | |
311 | * @bit_off: chunk offset | |
312 | * @bits: size of free area | |
313 | * | |
314 | * Helper function for pcpu_for_each_md_free_region. It checks | |
315 | * block->contig_hint and performs aggregation across blocks to find the | |
316 | * next hint. It modifies bit_off and bits in-place to be consumed in the | |
317 | * loop. | |
318 | */ | |
319 | static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off, | |
320 | int *bits) | |
321 | { | |
322 | int i = pcpu_off_to_block_index(*bit_off); | |
323 | int block_off = pcpu_off_to_block_off(*bit_off); | |
324 | struct pcpu_block_md *block; | |
325 | ||
326 | *bits = 0; | |
327 | for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); | |
328 | block++, i++) { | |
329 | /* handles contig area across blocks */ | |
330 | if (*bits) { | |
331 | *bits += block->left_free; | |
332 | if (block->left_free == PCPU_BITMAP_BLOCK_BITS) | |
333 | continue; | |
334 | return; | |
335 | } | |
336 | ||
337 | /* | |
338 | * This checks three things. First is there a contig_hint to | |
339 | * check. Second, have we checked this hint before by | |
340 | * comparing the block_off. Third, is this the same as the | |
341 | * right contig hint. In the last case, it spills over into | |
342 | * the next block and should be handled by the contig area | |
343 | * across blocks code. | |
344 | */ | |
345 | *bits = block->contig_hint; | |
346 | if (*bits && block->contig_hint_start >= block_off && | |
347 | *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) { | |
348 | *bit_off = pcpu_block_off_to_off(i, | |
349 | block->contig_hint_start); | |
350 | return; | |
351 | } | |
352 | ||
353 | *bits = block->right_free; | |
354 | *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free; | |
355 | } | |
356 | } | |
357 | ||
358 | /* | |
359 | * Metadata free area iterators. These perform aggregation of free areas | |
360 | * based on the metadata blocks and return the offset @bit_off and size in | |
361 | * bits of the free area @bits. | |
362 | */ | |
363 | #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \ | |
364 | for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \ | |
365 | (bit_off) < pcpu_chunk_map_bits((chunk)); \ | |
366 | (bit_off) += (bits) + 1, \ | |
367 | pcpu_next_md_free_region((chunk), &(bit_off), &(bits))) | |
368 | ||
fbf59bc9 | 369 | /** |
90459ce0 | 370 | * pcpu_mem_zalloc - allocate memory |
1880d93b | 371 | * @size: bytes to allocate |
fbf59bc9 | 372 | * |
1880d93b | 373 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
90459ce0 | 374 | * kzalloc() is used; otherwise, vzalloc() is used. The returned |
1880d93b | 375 | * memory is always zeroed. |
fbf59bc9 | 376 | * |
ccea34b5 TH |
377 | * CONTEXT: |
378 | * Does GFP_KERNEL allocation. | |
379 | * | |
fbf59bc9 | 380 | * RETURNS: |
1880d93b | 381 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 382 | */ |
90459ce0 | 383 | static void *pcpu_mem_zalloc(size_t size) |
fbf59bc9 | 384 | { |
099a19d9 TH |
385 | if (WARN_ON_ONCE(!slab_is_available())) |
386 | return NULL; | |
387 | ||
1880d93b TH |
388 | if (size <= PAGE_SIZE) |
389 | return kzalloc(size, GFP_KERNEL); | |
7af4c093 JJ |
390 | else |
391 | return vzalloc(size); | |
1880d93b | 392 | } |
fbf59bc9 | 393 | |
1880d93b TH |
394 | /** |
395 | * pcpu_mem_free - free memory | |
396 | * @ptr: memory to free | |
1880d93b | 397 | * |
90459ce0 | 398 | * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). |
1880d93b | 399 | */ |
1d5cfdb0 | 400 | static void pcpu_mem_free(void *ptr) |
1880d93b | 401 | { |
1d5cfdb0 | 402 | kvfree(ptr); |
fbf59bc9 TH |
403 | } |
404 | ||
405 | /** | |
406 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
407 | * @chunk: chunk of interest | |
408 | * @oslot: the previous slot it was on | |
409 | * | |
410 | * This function is called after an allocation or free changed @chunk. | |
411 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
412 | * moved to the slot. Note that the reserved chunk is never put on |
413 | * chunk slots. | |
ccea34b5 TH |
414 | * |
415 | * CONTEXT: | |
416 | * pcpu_lock. | |
fbf59bc9 TH |
417 | */ |
418 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
419 | { | |
420 | int nslot = pcpu_chunk_slot(chunk); | |
421 | ||
edcb4639 | 422 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
423 | if (oslot < nslot) |
424 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
425 | else | |
426 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
427 | } | |
428 | } | |
429 | ||
9f7dcf22 | 430 | /** |
40064aec | 431 | * pcpu_cnt_pop_pages- counts populated backing pages in range |
833af842 | 432 | * @chunk: chunk of interest |
40064aec DZF |
433 | * @bit_off: start offset |
434 | * @bits: size of area to check | |
9f7dcf22 | 435 | * |
40064aec DZF |
436 | * Calculates the number of populated pages in the region |
437 | * [page_start, page_end). This keeps track of how many empty populated | |
438 | * pages are available and decide if async work should be scheduled. | |
ccea34b5 | 439 | * |
9f7dcf22 | 440 | * RETURNS: |
40064aec | 441 | * The nr of populated pages. |
9f7dcf22 | 442 | */ |
40064aec DZF |
443 | static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off, |
444 | int bits) | |
9f7dcf22 | 445 | { |
40064aec DZF |
446 | int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE); |
447 | int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); | |
4f996e23 | 448 | |
40064aec | 449 | if (page_start >= page_end) |
9f7dcf22 TH |
450 | return 0; |
451 | ||
40064aec DZF |
452 | /* |
453 | * bitmap_weight counts the number of bits set in a bitmap up to | |
454 | * the specified number of bits. This is counting the populated | |
455 | * pages up to page_end and then subtracting the populated pages | |
456 | * up to page_start to count the populated pages in | |
457 | * [page_start, page_end). | |
458 | */ | |
459 | return bitmap_weight(chunk->populated, page_end) - | |
460 | bitmap_weight(chunk->populated, page_start); | |
833af842 TH |
461 | } |
462 | ||
463 | /** | |
40064aec | 464 | * pcpu_chunk_update - updates the chunk metadata given a free area |
833af842 | 465 | * @chunk: chunk of interest |
40064aec DZF |
466 | * @bit_off: chunk offset |
467 | * @bits: size of free area | |
833af842 | 468 | * |
13f96637 | 469 | * This updates the chunk's contig hint and starting offset given a free area. |
268625a6 | 470 | * Choose the best starting offset if the contig hint is equal. |
40064aec DZF |
471 | */ |
472 | static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits) | |
473 | { | |
13f96637 DZF |
474 | if (bits > chunk->contig_bits) { |
475 | chunk->contig_bits_start = bit_off; | |
40064aec | 476 | chunk->contig_bits = bits; |
268625a6 DZF |
477 | } else if (bits == chunk->contig_bits && chunk->contig_bits_start && |
478 | (!bit_off || | |
479 | __ffs(bit_off) > __ffs(chunk->contig_bits_start))) { | |
480 | /* use the start with the best alignment */ | |
481 | chunk->contig_bits_start = bit_off; | |
13f96637 | 482 | } |
40064aec DZF |
483 | } |
484 | ||
485 | /** | |
486 | * pcpu_chunk_refresh_hint - updates metadata about a chunk | |
487 | * @chunk: chunk of interest | |
833af842 | 488 | * |
525ca84d DZF |
489 | * Iterates over the metadata blocks to find the largest contig area. |
490 | * It also counts the populated pages and uses the delta to update the | |
491 | * global count. | |
833af842 | 492 | * |
40064aec DZF |
493 | * Updates: |
494 | * chunk->contig_bits | |
13f96637 | 495 | * chunk->contig_bits_start |
525ca84d | 496 | * nr_empty_pop_pages (chunk and global) |
833af842 | 497 | */ |
40064aec | 498 | static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk) |
833af842 | 499 | { |
525ca84d | 500 | int bit_off, bits, nr_empty_pop_pages; |
833af842 | 501 | |
40064aec DZF |
502 | /* clear metadata */ |
503 | chunk->contig_bits = 0; | |
6710e594 | 504 | |
525ca84d | 505 | bit_off = chunk->first_bit; |
40064aec | 506 | bits = nr_empty_pop_pages = 0; |
525ca84d DZF |
507 | pcpu_for_each_md_free_region(chunk, bit_off, bits) { |
508 | pcpu_chunk_update(chunk, bit_off, bits); | |
833af842 | 509 | |
525ca84d | 510 | nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, bit_off, bits); |
40064aec | 511 | } |
9f7dcf22 | 512 | |
40064aec DZF |
513 | /* |
514 | * Keep track of nr_empty_pop_pages. | |
515 | * | |
516 | * The chunk maintains the previous number of free pages it held, | |
517 | * so the delta is used to update the global counter. The reserved | |
518 | * chunk is not part of the free page count as they are populated | |
519 | * at init and are special to serving reserved allocations. | |
520 | */ | |
521 | if (chunk != pcpu_reserved_chunk) | |
522 | pcpu_nr_empty_pop_pages += | |
523 | (nr_empty_pop_pages - chunk->nr_empty_pop_pages); | |
a002d148 | 524 | |
40064aec DZF |
525 | chunk->nr_empty_pop_pages = nr_empty_pop_pages; |
526 | } | |
9f7dcf22 | 527 | |
ca460b3c DZF |
528 | /** |
529 | * pcpu_block_update - updates a block given a free area | |
530 | * @block: block of interest | |
531 | * @start: start offset in block | |
532 | * @end: end offset in block | |
533 | * | |
534 | * Updates a block given a known free area. The region [start, end) is | |
268625a6 DZF |
535 | * expected to be the entirety of the free area within a block. Chooses |
536 | * the best starting offset if the contig hints are equal. | |
ca460b3c DZF |
537 | */ |
538 | static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) | |
539 | { | |
540 | int contig = end - start; | |
541 | ||
542 | block->first_free = min(block->first_free, start); | |
543 | if (start == 0) | |
544 | block->left_free = contig; | |
545 | ||
546 | if (end == PCPU_BITMAP_BLOCK_BITS) | |
547 | block->right_free = contig; | |
548 | ||
549 | if (contig > block->contig_hint) { | |
550 | block->contig_hint_start = start; | |
551 | block->contig_hint = contig; | |
268625a6 DZF |
552 | } else if (block->contig_hint_start && contig == block->contig_hint && |
553 | (!start || __ffs(start) > __ffs(block->contig_hint_start))) { | |
554 | /* use the start with the best alignment */ | |
555 | block->contig_hint_start = start; | |
ca460b3c DZF |
556 | } |
557 | } | |
558 | ||
559 | /** | |
560 | * pcpu_block_refresh_hint | |
561 | * @chunk: chunk of interest | |
562 | * @index: index of the metadata block | |
563 | * | |
564 | * Scans over the block beginning at first_free and updates the block | |
565 | * metadata accordingly. | |
566 | */ | |
567 | static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) | |
568 | { | |
569 | struct pcpu_block_md *block = chunk->md_blocks + index; | |
570 | unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); | |
571 | int rs, re; /* region start, region end */ | |
572 | ||
573 | /* clear hints */ | |
574 | block->contig_hint = 0; | |
575 | block->left_free = block->right_free = 0; | |
576 | ||
577 | /* iterate over free areas and update the contig hints */ | |
578 | pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free, | |
579 | PCPU_BITMAP_BLOCK_BITS) { | |
580 | pcpu_block_update(block, rs, re); | |
581 | } | |
582 | } | |
583 | ||
584 | /** | |
585 | * pcpu_block_update_hint_alloc - update hint on allocation path | |
586 | * @chunk: chunk of interest | |
587 | * @bit_off: chunk offset | |
588 | * @bits: size of request | |
fc304334 DZF |
589 | * |
590 | * Updates metadata for the allocation path. The metadata only has to be | |
591 | * refreshed by a full scan iff the chunk's contig hint is broken. Block level | |
592 | * scans are required if the block's contig hint is broken. | |
ca460b3c DZF |
593 | */ |
594 | static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, | |
595 | int bits) | |
596 | { | |
597 | struct pcpu_block_md *s_block, *e_block, *block; | |
598 | int s_index, e_index; /* block indexes of the freed allocation */ | |
599 | int s_off, e_off; /* block offsets of the freed allocation */ | |
600 | ||
601 | /* | |
602 | * Calculate per block offsets. | |
603 | * The calculation uses an inclusive range, but the resulting offsets | |
604 | * are [start, end). e_index always points to the last block in the | |
605 | * range. | |
606 | */ | |
607 | s_index = pcpu_off_to_block_index(bit_off); | |
608 | e_index = pcpu_off_to_block_index(bit_off + bits - 1); | |
609 | s_off = pcpu_off_to_block_off(bit_off); | |
610 | e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; | |
611 | ||
612 | s_block = chunk->md_blocks + s_index; | |
613 | e_block = chunk->md_blocks + e_index; | |
614 | ||
615 | /* | |
616 | * Update s_block. | |
fc304334 DZF |
617 | * block->first_free must be updated if the allocation takes its place. |
618 | * If the allocation breaks the contig_hint, a scan is required to | |
619 | * restore this hint. | |
ca460b3c | 620 | */ |
fc304334 DZF |
621 | if (s_off == s_block->first_free) |
622 | s_block->first_free = find_next_zero_bit( | |
623 | pcpu_index_alloc_map(chunk, s_index), | |
624 | PCPU_BITMAP_BLOCK_BITS, | |
625 | s_off + bits); | |
626 | ||
627 | if (s_off >= s_block->contig_hint_start && | |
628 | s_off < s_block->contig_hint_start + s_block->contig_hint) { | |
629 | /* block contig hint is broken - scan to fix it */ | |
630 | pcpu_block_refresh_hint(chunk, s_index); | |
631 | } else { | |
632 | /* update left and right contig manually */ | |
633 | s_block->left_free = min(s_block->left_free, s_off); | |
634 | if (s_index == e_index) | |
635 | s_block->right_free = min_t(int, s_block->right_free, | |
636 | PCPU_BITMAP_BLOCK_BITS - e_off); | |
637 | else | |
638 | s_block->right_free = 0; | |
639 | } | |
ca460b3c DZF |
640 | |
641 | /* | |
642 | * Update e_block. | |
643 | */ | |
644 | if (s_index != e_index) { | |
fc304334 DZF |
645 | /* |
646 | * When the allocation is across blocks, the end is along | |
647 | * the left part of the e_block. | |
648 | */ | |
649 | e_block->first_free = find_next_zero_bit( | |
650 | pcpu_index_alloc_map(chunk, e_index), | |
651 | PCPU_BITMAP_BLOCK_BITS, e_off); | |
652 | ||
653 | if (e_off == PCPU_BITMAP_BLOCK_BITS) { | |
654 | /* reset the block */ | |
655 | e_block++; | |
656 | } else { | |
657 | if (e_off > e_block->contig_hint_start) { | |
658 | /* contig hint is broken - scan to fix it */ | |
659 | pcpu_block_refresh_hint(chunk, e_index); | |
660 | } else { | |
661 | e_block->left_free = 0; | |
662 | e_block->right_free = | |
663 | min_t(int, e_block->right_free, | |
664 | PCPU_BITMAP_BLOCK_BITS - e_off); | |
665 | } | |
666 | } | |
ca460b3c DZF |
667 | |
668 | /* update in-between md_blocks */ | |
669 | for (block = s_block + 1; block < e_block; block++) { | |
670 | block->contig_hint = 0; | |
671 | block->left_free = 0; | |
672 | block->right_free = 0; | |
673 | } | |
674 | } | |
675 | ||
fc304334 DZF |
676 | /* |
677 | * The only time a full chunk scan is required is if the chunk | |
678 | * contig hint is broken. Otherwise, it means a smaller space | |
679 | * was used and therefore the chunk contig hint is still correct. | |
680 | */ | |
681 | if (bit_off >= chunk->contig_bits_start && | |
682 | bit_off < chunk->contig_bits_start + chunk->contig_bits) | |
683 | pcpu_chunk_refresh_hint(chunk); | |
ca460b3c DZF |
684 | } |
685 | ||
686 | /** | |
687 | * pcpu_block_update_hint_free - updates the block hints on the free path | |
688 | * @chunk: chunk of interest | |
689 | * @bit_off: chunk offset | |
690 | * @bits: size of request | |
b185cd0d DZF |
691 | * |
692 | * Updates metadata for the allocation path. This avoids a blind block | |
693 | * refresh by making use of the block contig hints. If this fails, it scans | |
694 | * forward and backward to determine the extent of the free area. This is | |
695 | * capped at the boundary of blocks. | |
696 | * | |
697 | * A chunk update is triggered if a page becomes free, a block becomes free, | |
698 | * or the free spans across blocks. This tradeoff is to minimize iterating | |
699 | * over the block metadata to update chunk->contig_bits. chunk->contig_bits | |
700 | * may be off by up to a page, but it will never be more than the available | |
701 | * space. If the contig hint is contained in one block, it will be accurate. | |
ca460b3c DZF |
702 | */ |
703 | static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, | |
704 | int bits) | |
705 | { | |
706 | struct pcpu_block_md *s_block, *e_block, *block; | |
707 | int s_index, e_index; /* block indexes of the freed allocation */ | |
708 | int s_off, e_off; /* block offsets of the freed allocation */ | |
b185cd0d | 709 | int start, end; /* start and end of the whole free area */ |
ca460b3c DZF |
710 | |
711 | /* | |
712 | * Calculate per block offsets. | |
713 | * The calculation uses an inclusive range, but the resulting offsets | |
714 | * are [start, end). e_index always points to the last block in the | |
715 | * range. | |
716 | */ | |
717 | s_index = pcpu_off_to_block_index(bit_off); | |
718 | e_index = pcpu_off_to_block_index(bit_off + bits - 1); | |
719 | s_off = pcpu_off_to_block_off(bit_off); | |
720 | e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; | |
721 | ||
722 | s_block = chunk->md_blocks + s_index; | |
723 | e_block = chunk->md_blocks + e_index; | |
724 | ||
b185cd0d DZF |
725 | /* |
726 | * Check if the freed area aligns with the block->contig_hint. | |
727 | * If it does, then the scan to find the beginning/end of the | |
728 | * larger free area can be avoided. | |
729 | * | |
730 | * start and end refer to beginning and end of the free area | |
731 | * within each their respective blocks. This is not necessarily | |
732 | * the entire free area as it may span blocks past the beginning | |
733 | * or end of the block. | |
734 | */ | |
735 | start = s_off; | |
736 | if (s_off == s_block->contig_hint + s_block->contig_hint_start) { | |
737 | start = s_block->contig_hint_start; | |
738 | } else { | |
739 | /* | |
740 | * Scan backwards to find the extent of the free area. | |
741 | * find_last_bit returns the starting bit, so if the start bit | |
742 | * is returned, that means there was no last bit and the | |
743 | * remainder of the chunk is free. | |
744 | */ | |
745 | int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), | |
746 | start); | |
747 | start = (start == l_bit) ? 0 : l_bit + 1; | |
748 | } | |
749 | ||
750 | end = e_off; | |
751 | if (e_off == e_block->contig_hint_start) | |
752 | end = e_block->contig_hint_start + e_block->contig_hint; | |
753 | else | |
754 | end = find_next_bit(pcpu_index_alloc_map(chunk, e_index), | |
755 | PCPU_BITMAP_BLOCK_BITS, end); | |
756 | ||
ca460b3c | 757 | /* update s_block */ |
b185cd0d DZF |
758 | e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS; |
759 | pcpu_block_update(s_block, start, e_off); | |
ca460b3c DZF |
760 | |
761 | /* freeing in the same block */ | |
762 | if (s_index != e_index) { | |
763 | /* update e_block */ | |
b185cd0d | 764 | pcpu_block_update(e_block, 0, end); |
ca460b3c DZF |
765 | |
766 | /* reset md_blocks in the middle */ | |
767 | for (block = s_block + 1; block < e_block; block++) { | |
768 | block->first_free = 0; | |
769 | block->contig_hint_start = 0; | |
770 | block->contig_hint = PCPU_BITMAP_BLOCK_BITS; | |
771 | block->left_free = PCPU_BITMAP_BLOCK_BITS; | |
772 | block->right_free = PCPU_BITMAP_BLOCK_BITS; | |
773 | } | |
774 | } | |
775 | ||
b185cd0d DZF |
776 | /* |
777 | * Refresh chunk metadata when the free makes a page free, a block | |
778 | * free, or spans across blocks. The contig hint may be off by up to | |
779 | * a page, but if the hint is contained in a block, it will be accurate | |
780 | * with the else condition below. | |
781 | */ | |
782 | if ((ALIGN_DOWN(end, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS)) > | |
783 | ALIGN(start, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS))) || | |
784 | s_index != e_index) | |
785 | pcpu_chunk_refresh_hint(chunk); | |
786 | else | |
787 | pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start), | |
788 | s_block->contig_hint); | |
ca460b3c DZF |
789 | } |
790 | ||
40064aec DZF |
791 | /** |
792 | * pcpu_is_populated - determines if the region is populated | |
793 | * @chunk: chunk of interest | |
794 | * @bit_off: chunk offset | |
795 | * @bits: size of area | |
796 | * @next_off: return value for the next offset to start searching | |
797 | * | |
798 | * For atomic allocations, check if the backing pages are populated. | |
799 | * | |
800 | * RETURNS: | |
801 | * Bool if the backing pages are populated. | |
802 | * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. | |
803 | */ | |
804 | static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, | |
805 | int *next_off) | |
806 | { | |
807 | int page_start, page_end, rs, re; | |
833af842 | 808 | |
40064aec DZF |
809 | page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); |
810 | page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); | |
833af842 | 811 | |
40064aec DZF |
812 | rs = page_start; |
813 | pcpu_next_unpop(chunk->populated, &rs, &re, page_end); | |
814 | if (rs >= page_end) | |
815 | return true; | |
833af842 | 816 | |
40064aec DZF |
817 | *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; |
818 | return false; | |
9f7dcf22 TH |
819 | } |
820 | ||
a16037c8 | 821 | /** |
40064aec DZF |
822 | * pcpu_find_block_fit - finds the block index to start searching |
823 | * @chunk: chunk of interest | |
824 | * @alloc_bits: size of request in allocation units | |
825 | * @align: alignment of area (max PAGE_SIZE bytes) | |
826 | * @pop_only: use populated regions only | |
827 | * | |
828 | * RETURNS: | |
829 | * The offset in the bitmap to begin searching. | |
830 | * -1 if no offset is found. | |
a16037c8 | 831 | */ |
40064aec DZF |
832 | static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, |
833 | size_t align, bool pop_only) | |
a16037c8 | 834 | { |
40064aec DZF |
835 | int bit_off, bits; |
836 | int re; /* region end */ | |
a16037c8 | 837 | |
13f96637 DZF |
838 | /* |
839 | * Check to see if the allocation can fit in the chunk's contig hint. | |
840 | * This is an optimization to prevent scanning by assuming if it | |
841 | * cannot fit in the global hint, there is memory pressure and creating | |
842 | * a new chunk would happen soon. | |
843 | */ | |
844 | bit_off = ALIGN(chunk->contig_bits_start, align) - | |
845 | chunk->contig_bits_start; | |
846 | if (bit_off + alloc_bits > chunk->contig_bits) | |
847 | return -1; | |
848 | ||
86b442fb DZF |
849 | pcpu_for_each_unpop_region(chunk->alloc_map, bit_off, re, |
850 | chunk->first_bit, | |
40064aec DZF |
851 | pcpu_chunk_map_bits(chunk)) { |
852 | bits = re - bit_off; | |
a16037c8 | 853 | |
40064aec DZF |
854 | /* check alignment */ |
855 | bits -= ALIGN(bit_off, align) - bit_off; | |
856 | bit_off = ALIGN(bit_off, align); | |
857 | if (bits < alloc_bits) | |
858 | continue; | |
a16037c8 | 859 | |
40064aec DZF |
860 | bits = alloc_bits; |
861 | if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, | |
862 | &bit_off)) | |
863 | break; | |
a16037c8 | 864 | |
40064aec | 865 | bits = 0; |
a16037c8 | 866 | } |
40064aec DZF |
867 | |
868 | if (bit_off == pcpu_chunk_map_bits(chunk)) | |
869 | return -1; | |
870 | ||
871 | return bit_off; | |
a16037c8 TH |
872 | } |
873 | ||
fbf59bc9 | 874 | /** |
40064aec | 875 | * pcpu_alloc_area - allocates an area from a pcpu_chunk |
fbf59bc9 | 876 | * @chunk: chunk of interest |
40064aec DZF |
877 | * @alloc_bits: size of request in allocation units |
878 | * @align: alignment of area (max PAGE_SIZE) | |
879 | * @start: bit_off to start searching | |
9f7dcf22 | 880 | * |
40064aec DZF |
881 | * This function takes in a @start offset to begin searching to fit an |
882 | * allocation of @alloc_bits with alignment @align. If it confirms a | |
883 | * valid free area, it then updates the allocation and boundary maps | |
884 | * accordingly. | |
ccea34b5 | 885 | * |
fbf59bc9 | 886 | * RETURNS: |
40064aec DZF |
887 | * Allocated addr offset in @chunk on success. |
888 | * -1 if no matching area is found. | |
fbf59bc9 | 889 | */ |
40064aec DZF |
890 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, |
891 | size_t align, int start) | |
fbf59bc9 | 892 | { |
40064aec DZF |
893 | size_t align_mask = (align) ? (align - 1) : 0; |
894 | int bit_off, end, oslot; | |
a16037c8 | 895 | |
40064aec | 896 | lockdep_assert_held(&pcpu_lock); |
fbf59bc9 | 897 | |
40064aec | 898 | oslot = pcpu_chunk_slot(chunk); |
fbf59bc9 | 899 | |
40064aec DZF |
900 | /* |
901 | * Search to find a fit. | |
902 | */ | |
903 | end = start + alloc_bits; | |
904 | bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start, | |
905 | alloc_bits, align_mask); | |
906 | if (bit_off >= end) | |
907 | return -1; | |
fbf59bc9 | 908 | |
40064aec DZF |
909 | /* update alloc map */ |
910 | bitmap_set(chunk->alloc_map, bit_off, alloc_bits); | |
3d331ad7 | 911 | |
40064aec DZF |
912 | /* update boundary map */ |
913 | set_bit(bit_off, chunk->bound_map); | |
914 | bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); | |
915 | set_bit(bit_off + alloc_bits, chunk->bound_map); | |
fbf59bc9 | 916 | |
40064aec | 917 | chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; |
fbf59bc9 | 918 | |
86b442fb DZF |
919 | /* update first free bit */ |
920 | if (bit_off == chunk->first_bit) | |
921 | chunk->first_bit = find_next_zero_bit( | |
922 | chunk->alloc_map, | |
923 | pcpu_chunk_map_bits(chunk), | |
924 | bit_off + alloc_bits); | |
925 | ||
ca460b3c | 926 | pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); |
fbf59bc9 | 927 | |
fbf59bc9 TH |
928 | pcpu_chunk_relocate(chunk, oslot); |
929 | ||
40064aec | 930 | return bit_off * PCPU_MIN_ALLOC_SIZE; |
fbf59bc9 TH |
931 | } |
932 | ||
933 | /** | |
40064aec | 934 | * pcpu_free_area - frees the corresponding offset |
fbf59bc9 | 935 | * @chunk: chunk of interest |
40064aec | 936 | * @off: addr offset into chunk |
ccea34b5 | 937 | * |
40064aec DZF |
938 | * This function determines the size of an allocation to free using |
939 | * the boundary bitmap and clears the allocation map. | |
fbf59bc9 | 940 | */ |
40064aec | 941 | static void pcpu_free_area(struct pcpu_chunk *chunk, int off) |
fbf59bc9 | 942 | { |
40064aec | 943 | int bit_off, bits, end, oslot; |
723ad1d9 | 944 | |
5ccd30e4 | 945 | lockdep_assert_held(&pcpu_lock); |
30a5b536 | 946 | pcpu_stats_area_dealloc(chunk); |
5ccd30e4 | 947 | |
40064aec | 948 | oslot = pcpu_chunk_slot(chunk); |
fbf59bc9 | 949 | |
40064aec | 950 | bit_off = off / PCPU_MIN_ALLOC_SIZE; |
3d331ad7 | 951 | |
40064aec DZF |
952 | /* find end index */ |
953 | end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), | |
954 | bit_off + 1); | |
955 | bits = end - bit_off; | |
956 | bitmap_clear(chunk->alloc_map, bit_off, bits); | |
fbf59bc9 | 957 | |
40064aec DZF |
958 | /* update metadata */ |
959 | chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE; | |
b539b87f | 960 | |
86b442fb DZF |
961 | /* update first free bit */ |
962 | chunk->first_bit = min(chunk->first_bit, bit_off); | |
963 | ||
ca460b3c | 964 | pcpu_block_update_hint_free(chunk, bit_off, bits); |
fbf59bc9 | 965 | |
fbf59bc9 TH |
966 | pcpu_chunk_relocate(chunk, oslot); |
967 | } | |
968 | ||
ca460b3c DZF |
969 | static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) |
970 | { | |
971 | struct pcpu_block_md *md_block; | |
972 | ||
973 | for (md_block = chunk->md_blocks; | |
974 | md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); | |
975 | md_block++) { | |
976 | md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS; | |
977 | md_block->left_free = PCPU_BITMAP_BLOCK_BITS; | |
978 | md_block->right_free = PCPU_BITMAP_BLOCK_BITS; | |
979 | } | |
980 | } | |
981 | ||
40064aec DZF |
982 | /** |
983 | * pcpu_alloc_first_chunk - creates chunks that serve the first chunk | |
984 | * @tmp_addr: the start of the region served | |
985 | * @map_size: size of the region served | |
986 | * | |
987 | * This is responsible for creating the chunks that serve the first chunk. The | |
988 | * base_addr is page aligned down of @tmp_addr while the region end is page | |
989 | * aligned up. Offsets are kept track of to determine the region served. All | |
990 | * this is done to appease the bitmap allocator in avoiding partial blocks. | |
991 | * | |
992 | * RETURNS: | |
993 | * Chunk serving the region at @tmp_addr of @map_size. | |
994 | */ | |
c0ebfdc3 | 995 | static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, |
40064aec | 996 | int map_size) |
10edf5b0 DZF |
997 | { |
998 | struct pcpu_chunk *chunk; | |
ca460b3c | 999 | unsigned long aligned_addr, lcm_align; |
40064aec | 1000 | int start_offset, offset_bits, region_size, region_bits; |
c0ebfdc3 DZF |
1001 | |
1002 | /* region calculations */ | |
1003 | aligned_addr = tmp_addr & PAGE_MASK; | |
1004 | ||
1005 | start_offset = tmp_addr - aligned_addr; | |
6b9d7c8e | 1006 | |
ca460b3c DZF |
1007 | /* |
1008 | * Align the end of the region with the LCM of PAGE_SIZE and | |
1009 | * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of | |
1010 | * the other. | |
1011 | */ | |
1012 | lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); | |
1013 | region_size = ALIGN(start_offset + map_size, lcm_align); | |
10edf5b0 | 1014 | |
c0ebfdc3 | 1015 | /* allocate chunk */ |
8ab16c43 DZF |
1016 | chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) + |
1017 | BITS_TO_LONGS(region_size >> PAGE_SHIFT), | |
1018 | 0); | |
c0ebfdc3 | 1019 | |
10edf5b0 | 1020 | INIT_LIST_HEAD(&chunk->list); |
c0ebfdc3 DZF |
1021 | |
1022 | chunk->base_addr = (void *)aligned_addr; | |
10edf5b0 | 1023 | chunk->start_offset = start_offset; |
6b9d7c8e | 1024 | chunk->end_offset = region_size - chunk->start_offset - map_size; |
c0ebfdc3 | 1025 | |
8ab16c43 | 1026 | chunk->nr_pages = region_size >> PAGE_SHIFT; |
40064aec | 1027 | region_bits = pcpu_chunk_map_bits(chunk); |
c0ebfdc3 | 1028 | |
ca460b3c DZF |
1029 | chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) * |
1030 | sizeof(chunk->alloc_map[0]), 0); | |
1031 | chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) * | |
1032 | sizeof(chunk->bound_map[0]), 0); | |
1033 | chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) * | |
1034 | sizeof(chunk->md_blocks[0]), 0); | |
1035 | pcpu_init_md_blocks(chunk); | |
10edf5b0 DZF |
1036 | |
1037 | /* manage populated page bitmap */ | |
1038 | chunk->immutable = true; | |
8ab16c43 DZF |
1039 | bitmap_fill(chunk->populated, chunk->nr_pages); |
1040 | chunk->nr_populated = chunk->nr_pages; | |
40064aec DZF |
1041 | chunk->nr_empty_pop_pages = |
1042 | pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE, | |
1043 | map_size / PCPU_MIN_ALLOC_SIZE); | |
10edf5b0 | 1044 | |
40064aec DZF |
1045 | chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE; |
1046 | chunk->free_bytes = map_size; | |
c0ebfdc3 DZF |
1047 | |
1048 | if (chunk->start_offset) { | |
1049 | /* hide the beginning of the bitmap */ | |
40064aec DZF |
1050 | offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; |
1051 | bitmap_set(chunk->alloc_map, 0, offset_bits); | |
1052 | set_bit(0, chunk->bound_map); | |
1053 | set_bit(offset_bits, chunk->bound_map); | |
ca460b3c | 1054 | |
86b442fb DZF |
1055 | chunk->first_bit = offset_bits; |
1056 | ||
ca460b3c | 1057 | pcpu_block_update_hint_alloc(chunk, 0, offset_bits); |
c0ebfdc3 DZF |
1058 | } |
1059 | ||
6b9d7c8e DZF |
1060 | if (chunk->end_offset) { |
1061 | /* hide the end of the bitmap */ | |
40064aec DZF |
1062 | offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; |
1063 | bitmap_set(chunk->alloc_map, | |
1064 | pcpu_chunk_map_bits(chunk) - offset_bits, | |
1065 | offset_bits); | |
1066 | set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, | |
1067 | chunk->bound_map); | |
1068 | set_bit(region_bits, chunk->bound_map); | |
6b9d7c8e | 1069 | |
ca460b3c DZF |
1070 | pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) |
1071 | - offset_bits, offset_bits); | |
1072 | } | |
40064aec | 1073 | |
10edf5b0 DZF |
1074 | return chunk; |
1075 | } | |
1076 | ||
6081089f TH |
1077 | static struct pcpu_chunk *pcpu_alloc_chunk(void) |
1078 | { | |
1079 | struct pcpu_chunk *chunk; | |
40064aec | 1080 | int region_bits; |
6081089f | 1081 | |
90459ce0 | 1082 | chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); |
6081089f TH |
1083 | if (!chunk) |
1084 | return NULL; | |
1085 | ||
40064aec DZF |
1086 | INIT_LIST_HEAD(&chunk->list); |
1087 | chunk->nr_pages = pcpu_unit_pages; | |
1088 | region_bits = pcpu_chunk_map_bits(chunk); | |
6081089f | 1089 | |
40064aec DZF |
1090 | chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * |
1091 | sizeof(chunk->alloc_map[0])); | |
1092 | if (!chunk->alloc_map) | |
1093 | goto alloc_map_fail; | |
6081089f | 1094 | |
40064aec DZF |
1095 | chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * |
1096 | sizeof(chunk->bound_map[0])); | |
1097 | if (!chunk->bound_map) | |
1098 | goto bound_map_fail; | |
6081089f | 1099 | |
ca460b3c DZF |
1100 | chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * |
1101 | sizeof(chunk->md_blocks[0])); | |
1102 | if (!chunk->md_blocks) | |
1103 | goto md_blocks_fail; | |
1104 | ||
1105 | pcpu_init_md_blocks(chunk); | |
1106 | ||
40064aec DZF |
1107 | /* init metadata */ |
1108 | chunk->contig_bits = region_bits; | |
1109 | chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; | |
c0ebfdc3 | 1110 | |
6081089f | 1111 | return chunk; |
40064aec | 1112 | |
ca460b3c DZF |
1113 | md_blocks_fail: |
1114 | pcpu_mem_free(chunk->bound_map); | |
40064aec DZF |
1115 | bound_map_fail: |
1116 | pcpu_mem_free(chunk->alloc_map); | |
1117 | alloc_map_fail: | |
1118 | pcpu_mem_free(chunk); | |
1119 | ||
1120 | return NULL; | |
6081089f TH |
1121 | } |
1122 | ||
1123 | static void pcpu_free_chunk(struct pcpu_chunk *chunk) | |
1124 | { | |
1125 | if (!chunk) | |
1126 | return; | |
40064aec DZF |
1127 | pcpu_mem_free(chunk->bound_map); |
1128 | pcpu_mem_free(chunk->alloc_map); | |
1d5cfdb0 | 1129 | pcpu_mem_free(chunk); |
6081089f TH |
1130 | } |
1131 | ||
b539b87f TH |
1132 | /** |
1133 | * pcpu_chunk_populated - post-population bookkeeping | |
1134 | * @chunk: pcpu_chunk which got populated | |
1135 | * @page_start: the start page | |
1136 | * @page_end: the end page | |
40064aec | 1137 | * @for_alloc: if this is to populate for allocation |
b539b87f TH |
1138 | * |
1139 | * Pages in [@page_start,@page_end) have been populated to @chunk. Update | |
1140 | * the bookkeeping information accordingly. Must be called after each | |
1141 | * successful population. | |
40064aec DZF |
1142 | * |
1143 | * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it | |
1144 | * is to serve an allocation in that area. | |
b539b87f | 1145 | */ |
40064aec DZF |
1146 | static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, |
1147 | int page_end, bool for_alloc) | |
b539b87f TH |
1148 | { |
1149 | int nr = page_end - page_start; | |
1150 | ||
1151 | lockdep_assert_held(&pcpu_lock); | |
1152 | ||
1153 | bitmap_set(chunk->populated, page_start, nr); | |
1154 | chunk->nr_populated += nr; | |
40064aec DZF |
1155 | |
1156 | if (!for_alloc) { | |
1157 | chunk->nr_empty_pop_pages += nr; | |
1158 | pcpu_nr_empty_pop_pages += nr; | |
1159 | } | |
b539b87f TH |
1160 | } |
1161 | ||
1162 | /** | |
1163 | * pcpu_chunk_depopulated - post-depopulation bookkeeping | |
1164 | * @chunk: pcpu_chunk which got depopulated | |
1165 | * @page_start: the start page | |
1166 | * @page_end: the end page | |
1167 | * | |
1168 | * Pages in [@page_start,@page_end) have been depopulated from @chunk. | |
1169 | * Update the bookkeeping information accordingly. Must be called after | |
1170 | * each successful depopulation. | |
1171 | */ | |
1172 | static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, | |
1173 | int page_start, int page_end) | |
1174 | { | |
1175 | int nr = page_end - page_start; | |
1176 | ||
1177 | lockdep_assert_held(&pcpu_lock); | |
1178 | ||
1179 | bitmap_clear(chunk->populated, page_start, nr); | |
1180 | chunk->nr_populated -= nr; | |
0cecf50c | 1181 | chunk->nr_empty_pop_pages -= nr; |
b539b87f TH |
1182 | pcpu_nr_empty_pop_pages -= nr; |
1183 | } | |
1184 | ||
9f645532 TH |
1185 | /* |
1186 | * Chunk management implementation. | |
1187 | * | |
1188 | * To allow different implementations, chunk alloc/free and | |
1189 | * [de]population are implemented in a separate file which is pulled | |
1190 | * into this file and compiled together. The following functions | |
1191 | * should be implemented. | |
1192 | * | |
1193 | * pcpu_populate_chunk - populate the specified range of a chunk | |
1194 | * pcpu_depopulate_chunk - depopulate the specified range of a chunk | |
1195 | * pcpu_create_chunk - create a new chunk | |
1196 | * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop | |
1197 | * pcpu_addr_to_page - translate address to physical address | |
1198 | * pcpu_verify_alloc_info - check alloc_info is acceptable during init | |
fbf59bc9 | 1199 | */ |
9f645532 TH |
1200 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); |
1201 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); | |
1202 | static struct pcpu_chunk *pcpu_create_chunk(void); | |
1203 | static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); | |
1204 | static struct page *pcpu_addr_to_page(void *addr); | |
1205 | static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); | |
fbf59bc9 | 1206 | |
b0c9778b TH |
1207 | #ifdef CONFIG_NEED_PER_CPU_KM |
1208 | #include "percpu-km.c" | |
1209 | #else | |
9f645532 | 1210 | #include "percpu-vm.c" |
b0c9778b | 1211 | #endif |
fbf59bc9 | 1212 | |
88999a89 TH |
1213 | /** |
1214 | * pcpu_chunk_addr_search - determine chunk containing specified address | |
1215 | * @addr: address for which the chunk needs to be determined. | |
1216 | * | |
c0ebfdc3 DZF |
1217 | * This is an internal function that handles all but static allocations. |
1218 | * Static percpu address values should never be passed into the allocator. | |
1219 | * | |
88999a89 TH |
1220 | * RETURNS: |
1221 | * The address of the found chunk. | |
1222 | */ | |
1223 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
1224 | { | |
c0ebfdc3 | 1225 | /* is it in the dynamic region (first chunk)? */ |
560f2c23 | 1226 | if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) |
88999a89 | 1227 | return pcpu_first_chunk; |
c0ebfdc3 DZF |
1228 | |
1229 | /* is it in the reserved region? */ | |
560f2c23 | 1230 | if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) |
c0ebfdc3 | 1231 | return pcpu_reserved_chunk; |
88999a89 TH |
1232 | |
1233 | /* | |
1234 | * The address is relative to unit0 which might be unused and | |
1235 | * thus unmapped. Offset the address to the unit space of the | |
1236 | * current processor before looking it up in the vmalloc | |
1237 | * space. Note that any possible cpu id can be used here, so | |
1238 | * there's no need to worry about preemption or cpu hotplug. | |
1239 | */ | |
1240 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; | |
9f645532 | 1241 | return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); |
88999a89 TH |
1242 | } |
1243 | ||
fbf59bc9 | 1244 | /** |
edcb4639 | 1245 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 1246 | * @size: size of area to allocate in bytes |
fbf59bc9 | 1247 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 1248 | * @reserved: allocate from the reserved chunk if available |
5835d96e | 1249 | * @gfp: allocation flags |
fbf59bc9 | 1250 | * |
5835d96e TH |
1251 | * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't |
1252 | * contain %GFP_KERNEL, the allocation is atomic. | |
fbf59bc9 TH |
1253 | * |
1254 | * RETURNS: | |
1255 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1256 | */ | |
5835d96e TH |
1257 | static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, |
1258 | gfp_t gfp) | |
fbf59bc9 | 1259 | { |
f2badb0c | 1260 | static int warn_limit = 10; |
fbf59bc9 | 1261 | struct pcpu_chunk *chunk; |
f2badb0c | 1262 | const char *err; |
6ae833c7 | 1263 | bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; |
40064aec | 1264 | int slot, off, cpu, ret; |
403a91b1 | 1265 | unsigned long flags; |
f528f0b8 | 1266 | void __percpu *ptr; |
40064aec | 1267 | size_t bits, bit_align; |
fbf59bc9 | 1268 | |
723ad1d9 | 1269 | /* |
40064aec DZF |
1270 | * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, |
1271 | * therefore alignment must be a minimum of that many bytes. | |
1272 | * An allocation may have internal fragmentation from rounding up | |
1273 | * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. | |
723ad1d9 | 1274 | */ |
d2f3c384 DZF |
1275 | if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) |
1276 | align = PCPU_MIN_ALLOC_SIZE; | |
723ad1d9 | 1277 | |
d2f3c384 | 1278 | size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); |
40064aec DZF |
1279 | bits = size >> PCPU_MIN_ALLOC_SHIFT; |
1280 | bit_align = align >> PCPU_MIN_ALLOC_SHIFT; | |
2f69fa82 | 1281 | |
3ca45a46 | 1282 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || |
1283 | !is_power_of_2(align))) { | |
756a025f JP |
1284 | WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n", |
1285 | size, align); | |
fbf59bc9 TH |
1286 | return NULL; |
1287 | } | |
1288 | ||
6710e594 TH |
1289 | if (!is_atomic) |
1290 | mutex_lock(&pcpu_alloc_mutex); | |
1291 | ||
403a91b1 | 1292 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 | 1293 | |
edcb4639 TH |
1294 | /* serve reserved allocations from the reserved chunk if available */ |
1295 | if (reserved && pcpu_reserved_chunk) { | |
1296 | chunk = pcpu_reserved_chunk; | |
833af842 | 1297 | |
40064aec DZF |
1298 | off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); |
1299 | if (off < 0) { | |
833af842 | 1300 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 1301 | goto fail_unlock; |
f2badb0c | 1302 | } |
833af842 | 1303 | |
40064aec | 1304 | off = pcpu_alloc_area(chunk, bits, bit_align, off); |
edcb4639 TH |
1305 | if (off >= 0) |
1306 | goto area_found; | |
833af842 | 1307 | |
f2badb0c | 1308 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 1309 | goto fail_unlock; |
edcb4639 TH |
1310 | } |
1311 | ||
ccea34b5 | 1312 | restart: |
edcb4639 | 1313 | /* search through normal chunks */ |
fbf59bc9 TH |
1314 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
1315 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
40064aec DZF |
1316 | off = pcpu_find_block_fit(chunk, bits, bit_align, |
1317 | is_atomic); | |
1318 | if (off < 0) | |
fbf59bc9 | 1319 | continue; |
ccea34b5 | 1320 | |
40064aec | 1321 | off = pcpu_alloc_area(chunk, bits, bit_align, off); |
fbf59bc9 TH |
1322 | if (off >= 0) |
1323 | goto area_found; | |
40064aec | 1324 | |
fbf59bc9 TH |
1325 | } |
1326 | } | |
1327 | ||
403a91b1 | 1328 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1329 | |
b38d08f3 TH |
1330 | /* |
1331 | * No space left. Create a new chunk. We don't want multiple | |
1332 | * tasks to create chunks simultaneously. Serialize and create iff | |
1333 | * there's still no empty chunk after grabbing the mutex. | |
1334 | */ | |
11df02bf DZ |
1335 | if (is_atomic) { |
1336 | err = "atomic alloc failed, no space left"; | |
5835d96e | 1337 | goto fail; |
11df02bf | 1338 | } |
5835d96e | 1339 | |
b38d08f3 TH |
1340 | if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { |
1341 | chunk = pcpu_create_chunk(); | |
1342 | if (!chunk) { | |
1343 | err = "failed to allocate new chunk"; | |
1344 | goto fail; | |
1345 | } | |
1346 | ||
1347 | spin_lock_irqsave(&pcpu_lock, flags); | |
1348 | pcpu_chunk_relocate(chunk, -1); | |
1349 | } else { | |
1350 | spin_lock_irqsave(&pcpu_lock, flags); | |
f2badb0c | 1351 | } |
ccea34b5 | 1352 | |
ccea34b5 | 1353 | goto restart; |
fbf59bc9 TH |
1354 | |
1355 | area_found: | |
30a5b536 | 1356 | pcpu_stats_area_alloc(chunk, size); |
403a91b1 | 1357 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1358 | |
dca49645 | 1359 | /* populate if not all pages are already there */ |
5835d96e | 1360 | if (!is_atomic) { |
e04d3208 | 1361 | int page_start, page_end, rs, re; |
dca49645 | 1362 | |
e04d3208 TH |
1363 | page_start = PFN_DOWN(off); |
1364 | page_end = PFN_UP(off + size); | |
b38d08f3 | 1365 | |
91e914c5 DZF |
1366 | pcpu_for_each_unpop_region(chunk->populated, rs, re, |
1367 | page_start, page_end) { | |
e04d3208 TH |
1368 | WARN_ON(chunk->immutable); |
1369 | ||
1370 | ret = pcpu_populate_chunk(chunk, rs, re); | |
1371 | ||
1372 | spin_lock_irqsave(&pcpu_lock, flags); | |
1373 | if (ret) { | |
40064aec | 1374 | pcpu_free_area(chunk, off); |
e04d3208 TH |
1375 | err = "failed to populate"; |
1376 | goto fail_unlock; | |
1377 | } | |
40064aec | 1378 | pcpu_chunk_populated(chunk, rs, re, true); |
e04d3208 | 1379 | spin_unlock_irqrestore(&pcpu_lock, flags); |
dca49645 | 1380 | } |
fbf59bc9 | 1381 | |
e04d3208 TH |
1382 | mutex_unlock(&pcpu_alloc_mutex); |
1383 | } | |
ccea34b5 | 1384 | |
1a4d7607 TH |
1385 | if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) |
1386 | pcpu_schedule_balance_work(); | |
1387 | ||
dca49645 TH |
1388 | /* clear the areas and return address relative to base address */ |
1389 | for_each_possible_cpu(cpu) | |
1390 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); | |
1391 | ||
f528f0b8 | 1392 | ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); |
8a8c35fa | 1393 | kmemleak_alloc_percpu(ptr, size, gfp); |
df95e795 DZ |
1394 | |
1395 | trace_percpu_alloc_percpu(reserved, is_atomic, size, align, | |
1396 | chunk->base_addr, off, ptr); | |
1397 | ||
f528f0b8 | 1398 | return ptr; |
ccea34b5 TH |
1399 | |
1400 | fail_unlock: | |
403a91b1 | 1401 | spin_unlock_irqrestore(&pcpu_lock, flags); |
b38d08f3 | 1402 | fail: |
df95e795 DZ |
1403 | trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); |
1404 | ||
5835d96e | 1405 | if (!is_atomic && warn_limit) { |
870d4b12 | 1406 | pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", |
598d8091 | 1407 | size, align, is_atomic, err); |
f2badb0c TH |
1408 | dump_stack(); |
1409 | if (!--warn_limit) | |
870d4b12 | 1410 | pr_info("limit reached, disable warning\n"); |
f2badb0c | 1411 | } |
1a4d7607 TH |
1412 | if (is_atomic) { |
1413 | /* see the flag handling in pcpu_blance_workfn() */ | |
1414 | pcpu_atomic_alloc_failed = true; | |
1415 | pcpu_schedule_balance_work(); | |
6710e594 TH |
1416 | } else { |
1417 | mutex_unlock(&pcpu_alloc_mutex); | |
1a4d7607 | 1418 | } |
ccea34b5 | 1419 | return NULL; |
fbf59bc9 | 1420 | } |
edcb4639 TH |
1421 | |
1422 | /** | |
5835d96e | 1423 | * __alloc_percpu_gfp - allocate dynamic percpu area |
edcb4639 TH |
1424 | * @size: size of area to allocate in bytes |
1425 | * @align: alignment of area (max PAGE_SIZE) | |
5835d96e | 1426 | * @gfp: allocation flags |
edcb4639 | 1427 | * |
5835d96e TH |
1428 | * Allocate zero-filled percpu area of @size bytes aligned at @align. If |
1429 | * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can | |
1430 | * be called from any context but is a lot more likely to fail. | |
ccea34b5 | 1431 | * |
edcb4639 TH |
1432 | * RETURNS: |
1433 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1434 | */ | |
5835d96e TH |
1435 | void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) |
1436 | { | |
1437 | return pcpu_alloc(size, align, false, gfp); | |
1438 | } | |
1439 | EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); | |
1440 | ||
1441 | /** | |
1442 | * __alloc_percpu - allocate dynamic percpu area | |
1443 | * @size: size of area to allocate in bytes | |
1444 | * @align: alignment of area (max PAGE_SIZE) | |
1445 | * | |
1446 | * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). | |
1447 | */ | |
43cf38eb | 1448 | void __percpu *__alloc_percpu(size_t size, size_t align) |
edcb4639 | 1449 | { |
5835d96e | 1450 | return pcpu_alloc(size, align, false, GFP_KERNEL); |
edcb4639 | 1451 | } |
fbf59bc9 TH |
1452 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1453 | ||
edcb4639 TH |
1454 | /** |
1455 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1456 | * @size: size of area to allocate in bytes | |
1457 | * @align: alignment of area (max PAGE_SIZE) | |
1458 | * | |
9329ba97 TH |
1459 | * Allocate zero-filled percpu area of @size bytes aligned at @align |
1460 | * from reserved percpu area if arch has set it up; otherwise, | |
1461 | * allocation is served from the same dynamic area. Might sleep. | |
1462 | * Might trigger writeouts. | |
edcb4639 | 1463 | * |
ccea34b5 TH |
1464 | * CONTEXT: |
1465 | * Does GFP_KERNEL allocation. | |
1466 | * | |
edcb4639 TH |
1467 | * RETURNS: |
1468 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1469 | */ | |
43cf38eb | 1470 | void __percpu *__alloc_reserved_percpu(size_t size, size_t align) |
edcb4639 | 1471 | { |
5835d96e | 1472 | return pcpu_alloc(size, align, true, GFP_KERNEL); |
edcb4639 TH |
1473 | } |
1474 | ||
a56dbddf | 1475 | /** |
1a4d7607 | 1476 | * pcpu_balance_workfn - manage the amount of free chunks and populated pages |
a56dbddf TH |
1477 | * @work: unused |
1478 | * | |
1479 | * Reclaim all fully free chunks except for the first one. | |
1480 | */ | |
fe6bd8c3 | 1481 | static void pcpu_balance_workfn(struct work_struct *work) |
fbf59bc9 | 1482 | { |
fe6bd8c3 TH |
1483 | LIST_HEAD(to_free); |
1484 | struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; | |
a56dbddf | 1485 | struct pcpu_chunk *chunk, *next; |
1a4d7607 | 1486 | int slot, nr_to_pop, ret; |
a56dbddf | 1487 | |
1a4d7607 TH |
1488 | /* |
1489 | * There's no reason to keep around multiple unused chunks and VM | |
1490 | * areas can be scarce. Destroy all free chunks except for one. | |
1491 | */ | |
ccea34b5 TH |
1492 | mutex_lock(&pcpu_alloc_mutex); |
1493 | spin_lock_irq(&pcpu_lock); | |
a56dbddf | 1494 | |
fe6bd8c3 | 1495 | list_for_each_entry_safe(chunk, next, free_head, list) { |
a56dbddf TH |
1496 | WARN_ON(chunk->immutable); |
1497 | ||
1498 | /* spare the first one */ | |
fe6bd8c3 | 1499 | if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) |
a56dbddf TH |
1500 | continue; |
1501 | ||
fe6bd8c3 | 1502 | list_move(&chunk->list, &to_free); |
a56dbddf TH |
1503 | } |
1504 | ||
ccea34b5 | 1505 | spin_unlock_irq(&pcpu_lock); |
a56dbddf | 1506 | |
fe6bd8c3 | 1507 | list_for_each_entry_safe(chunk, next, &to_free, list) { |
a93ace48 | 1508 | int rs, re; |
dca49645 | 1509 | |
91e914c5 DZF |
1510 | pcpu_for_each_pop_region(chunk->populated, rs, re, 0, |
1511 | chunk->nr_pages) { | |
a93ace48 | 1512 | pcpu_depopulate_chunk(chunk, rs, re); |
b539b87f TH |
1513 | spin_lock_irq(&pcpu_lock); |
1514 | pcpu_chunk_depopulated(chunk, rs, re); | |
1515 | spin_unlock_irq(&pcpu_lock); | |
a93ace48 | 1516 | } |
6081089f | 1517 | pcpu_destroy_chunk(chunk); |
a56dbddf | 1518 | } |
971f3918 | 1519 | |
1a4d7607 TH |
1520 | /* |
1521 | * Ensure there are certain number of free populated pages for | |
1522 | * atomic allocs. Fill up from the most packed so that atomic | |
1523 | * allocs don't increase fragmentation. If atomic allocation | |
1524 | * failed previously, always populate the maximum amount. This | |
1525 | * should prevent atomic allocs larger than PAGE_SIZE from keeping | |
1526 | * failing indefinitely; however, large atomic allocs are not | |
1527 | * something we support properly and can be highly unreliable and | |
1528 | * inefficient. | |
1529 | */ | |
1530 | retry_pop: | |
1531 | if (pcpu_atomic_alloc_failed) { | |
1532 | nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; | |
1533 | /* best effort anyway, don't worry about synchronization */ | |
1534 | pcpu_atomic_alloc_failed = false; | |
1535 | } else { | |
1536 | nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - | |
1537 | pcpu_nr_empty_pop_pages, | |
1538 | 0, PCPU_EMPTY_POP_PAGES_HIGH); | |
1539 | } | |
1540 | ||
1541 | for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { | |
1542 | int nr_unpop = 0, rs, re; | |
1543 | ||
1544 | if (!nr_to_pop) | |
1545 | break; | |
1546 | ||
1547 | spin_lock_irq(&pcpu_lock); | |
1548 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
8ab16c43 | 1549 | nr_unpop = chunk->nr_pages - chunk->nr_populated; |
1a4d7607 TH |
1550 | if (nr_unpop) |
1551 | break; | |
1552 | } | |
1553 | spin_unlock_irq(&pcpu_lock); | |
1554 | ||
1555 | if (!nr_unpop) | |
1556 | continue; | |
1557 | ||
1558 | /* @chunk can't go away while pcpu_alloc_mutex is held */ | |
91e914c5 DZF |
1559 | pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, |
1560 | chunk->nr_pages) { | |
1a4d7607 TH |
1561 | int nr = min(re - rs, nr_to_pop); |
1562 | ||
1563 | ret = pcpu_populate_chunk(chunk, rs, rs + nr); | |
1564 | if (!ret) { | |
1565 | nr_to_pop -= nr; | |
1566 | spin_lock_irq(&pcpu_lock); | |
40064aec | 1567 | pcpu_chunk_populated(chunk, rs, rs + nr, false); |
1a4d7607 TH |
1568 | spin_unlock_irq(&pcpu_lock); |
1569 | } else { | |
1570 | nr_to_pop = 0; | |
1571 | } | |
1572 | ||
1573 | if (!nr_to_pop) | |
1574 | break; | |
1575 | } | |
1576 | } | |
1577 | ||
1578 | if (nr_to_pop) { | |
1579 | /* ran out of chunks to populate, create a new one and retry */ | |
1580 | chunk = pcpu_create_chunk(); | |
1581 | if (chunk) { | |
1582 | spin_lock_irq(&pcpu_lock); | |
1583 | pcpu_chunk_relocate(chunk, -1); | |
1584 | spin_unlock_irq(&pcpu_lock); | |
1585 | goto retry_pop; | |
1586 | } | |
1587 | } | |
1588 | ||
971f3918 | 1589 | mutex_unlock(&pcpu_alloc_mutex); |
fbf59bc9 TH |
1590 | } |
1591 | ||
1592 | /** | |
1593 | * free_percpu - free percpu area | |
1594 | * @ptr: pointer to area to free | |
1595 | * | |
ccea34b5 TH |
1596 | * Free percpu area @ptr. |
1597 | * | |
1598 | * CONTEXT: | |
1599 | * Can be called from atomic context. | |
fbf59bc9 | 1600 | */ |
43cf38eb | 1601 | void free_percpu(void __percpu *ptr) |
fbf59bc9 | 1602 | { |
129182e5 | 1603 | void *addr; |
fbf59bc9 | 1604 | struct pcpu_chunk *chunk; |
ccea34b5 | 1605 | unsigned long flags; |
40064aec | 1606 | int off; |
fbf59bc9 TH |
1607 | |
1608 | if (!ptr) | |
1609 | return; | |
1610 | ||
f528f0b8 CM |
1611 | kmemleak_free_percpu(ptr); |
1612 | ||
129182e5 AM |
1613 | addr = __pcpu_ptr_to_addr(ptr); |
1614 | ||
ccea34b5 | 1615 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1616 | |
1617 | chunk = pcpu_chunk_addr_search(addr); | |
bba174f5 | 1618 | off = addr - chunk->base_addr; |
fbf59bc9 | 1619 | |
40064aec | 1620 | pcpu_free_area(chunk, off); |
fbf59bc9 | 1621 | |
a56dbddf | 1622 | /* if there are more than one fully free chunks, wake up grim reaper */ |
40064aec | 1623 | if (chunk->free_bytes == pcpu_unit_size) { |
fbf59bc9 TH |
1624 | struct pcpu_chunk *pos; |
1625 | ||
a56dbddf | 1626 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1627 | if (pos != chunk) { |
1a4d7607 | 1628 | pcpu_schedule_balance_work(); |
fbf59bc9 TH |
1629 | break; |
1630 | } | |
1631 | } | |
1632 | ||
df95e795 DZ |
1633 | trace_percpu_free_percpu(chunk->base_addr, off, ptr); |
1634 | ||
ccea34b5 | 1635 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1636 | } |
1637 | EXPORT_SYMBOL_GPL(free_percpu); | |
1638 | ||
383776fa | 1639 | bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) |
10fad5e4 | 1640 | { |
bbddff05 | 1641 | #ifdef CONFIG_SMP |
10fad5e4 TH |
1642 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
1643 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); | |
1644 | unsigned int cpu; | |
1645 | ||
1646 | for_each_possible_cpu(cpu) { | |
1647 | void *start = per_cpu_ptr(base, cpu); | |
383776fa | 1648 | void *va = (void *)addr; |
10fad5e4 | 1649 | |
383776fa | 1650 | if (va >= start && va < start + static_size) { |
8ce371f9 | 1651 | if (can_addr) { |
383776fa | 1652 | *can_addr = (unsigned long) (va - start); |
8ce371f9 PZ |
1653 | *can_addr += (unsigned long) |
1654 | per_cpu_ptr(base, get_boot_cpu_id()); | |
1655 | } | |
10fad5e4 | 1656 | return true; |
383776fa TG |
1657 | } |
1658 | } | |
bbddff05 TH |
1659 | #endif |
1660 | /* on UP, can't distinguish from other static vars, always false */ | |
10fad5e4 TH |
1661 | return false; |
1662 | } | |
1663 | ||
383776fa TG |
1664 | /** |
1665 | * is_kernel_percpu_address - test whether address is from static percpu area | |
1666 | * @addr: address to test | |
1667 | * | |
1668 | * Test whether @addr belongs to in-kernel static percpu area. Module | |
1669 | * static percpu areas are not considered. For those, use | |
1670 | * is_module_percpu_address(). | |
1671 | * | |
1672 | * RETURNS: | |
1673 | * %true if @addr is from in-kernel static percpu area, %false otherwise. | |
1674 | */ | |
1675 | bool is_kernel_percpu_address(unsigned long addr) | |
1676 | { | |
1677 | return __is_kernel_percpu_address(addr, NULL); | |
1678 | } | |
1679 | ||
3b034b0d VG |
1680 | /** |
1681 | * per_cpu_ptr_to_phys - convert translated percpu address to physical address | |
1682 | * @addr: the address to be converted to physical address | |
1683 | * | |
1684 | * Given @addr which is dereferenceable address obtained via one of | |
1685 | * percpu access macros, this function translates it into its physical | |
1686 | * address. The caller is responsible for ensuring @addr stays valid | |
1687 | * until this function finishes. | |
1688 | * | |
67589c71 DY |
1689 | * percpu allocator has special setup for the first chunk, which currently |
1690 | * supports either embedding in linear address space or vmalloc mapping, | |
1691 | * and, from the second one, the backing allocator (currently either vm or | |
1692 | * km) provides translation. | |
1693 | * | |
bffc4375 | 1694 | * The addr can be translated simply without checking if it falls into the |
67589c71 DY |
1695 | * first chunk. But the current code reflects better how percpu allocator |
1696 | * actually works, and the verification can discover both bugs in percpu | |
1697 | * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current | |
1698 | * code. | |
1699 | * | |
3b034b0d VG |
1700 | * RETURNS: |
1701 | * The physical address for @addr. | |
1702 | */ | |
1703 | phys_addr_t per_cpu_ptr_to_phys(void *addr) | |
1704 | { | |
9983b6f0 TH |
1705 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
1706 | bool in_first_chunk = false; | |
a855b84c | 1707 | unsigned long first_low, first_high; |
9983b6f0 TH |
1708 | unsigned int cpu; |
1709 | ||
1710 | /* | |
a855b84c | 1711 | * The following test on unit_low/high isn't strictly |
9983b6f0 TH |
1712 | * necessary but will speed up lookups of addresses which |
1713 | * aren't in the first chunk. | |
c0ebfdc3 DZF |
1714 | * |
1715 | * The address check is against full chunk sizes. pcpu_base_addr | |
1716 | * points to the beginning of the first chunk including the | |
1717 | * static region. Assumes good intent as the first chunk may | |
1718 | * not be full (ie. < pcpu_unit_pages in size). | |
9983b6f0 | 1719 | */ |
c0ebfdc3 DZF |
1720 | first_low = (unsigned long)pcpu_base_addr + |
1721 | pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); | |
1722 | first_high = (unsigned long)pcpu_base_addr + | |
1723 | pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); | |
a855b84c TH |
1724 | if ((unsigned long)addr >= first_low && |
1725 | (unsigned long)addr < first_high) { | |
9983b6f0 TH |
1726 | for_each_possible_cpu(cpu) { |
1727 | void *start = per_cpu_ptr(base, cpu); | |
1728 | ||
1729 | if (addr >= start && addr < start + pcpu_unit_size) { | |
1730 | in_first_chunk = true; | |
1731 | break; | |
1732 | } | |
1733 | } | |
1734 | } | |
1735 | ||
1736 | if (in_first_chunk) { | |
eac522ef | 1737 | if (!is_vmalloc_addr(addr)) |
020ec653 TH |
1738 | return __pa(addr); |
1739 | else | |
9f57bd4d ES |
1740 | return page_to_phys(vmalloc_to_page(addr)) + |
1741 | offset_in_page(addr); | |
020ec653 | 1742 | } else |
9f57bd4d ES |
1743 | return page_to_phys(pcpu_addr_to_page(addr)) + |
1744 | offset_in_page(addr); | |
3b034b0d VG |
1745 | } |
1746 | ||
fbf59bc9 | 1747 | /** |
fd1e8a1f TH |
1748 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
1749 | * @nr_groups: the number of groups | |
1750 | * @nr_units: the number of units | |
1751 | * | |
1752 | * Allocate ai which is large enough for @nr_groups groups containing | |
1753 | * @nr_units units. The returned ai's groups[0].cpu_map points to the | |
1754 | * cpu_map array which is long enough for @nr_units and filled with | |
1755 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map | |
1756 | * pointer of other groups. | |
1757 | * | |
1758 | * RETURNS: | |
1759 | * Pointer to the allocated pcpu_alloc_info on success, NULL on | |
1760 | * failure. | |
1761 | */ | |
1762 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, | |
1763 | int nr_units) | |
1764 | { | |
1765 | struct pcpu_alloc_info *ai; | |
1766 | size_t base_size, ai_size; | |
1767 | void *ptr; | |
1768 | int unit; | |
1769 | ||
1770 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), | |
1771 | __alignof__(ai->groups[0].cpu_map[0])); | |
1772 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); | |
1773 | ||
999c17e3 | 1774 | ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); |
fd1e8a1f TH |
1775 | if (!ptr) |
1776 | return NULL; | |
1777 | ai = ptr; | |
1778 | ptr += base_size; | |
1779 | ||
1780 | ai->groups[0].cpu_map = ptr; | |
1781 | ||
1782 | for (unit = 0; unit < nr_units; unit++) | |
1783 | ai->groups[0].cpu_map[unit] = NR_CPUS; | |
1784 | ||
1785 | ai->nr_groups = nr_groups; | |
1786 | ai->__ai_size = PFN_ALIGN(ai_size); | |
1787 | ||
1788 | return ai; | |
1789 | } | |
1790 | ||
1791 | /** | |
1792 | * pcpu_free_alloc_info - free percpu allocation info | |
1793 | * @ai: pcpu_alloc_info to free | |
1794 | * | |
1795 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). | |
1796 | */ | |
1797 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) | |
1798 | { | |
999c17e3 | 1799 | memblock_free_early(__pa(ai), ai->__ai_size); |
fd1e8a1f TH |
1800 | } |
1801 | ||
fd1e8a1f TH |
1802 | /** |
1803 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info | |
1804 | * @lvl: loglevel | |
1805 | * @ai: allocation info to dump | |
1806 | * | |
1807 | * Print out information about @ai using loglevel @lvl. | |
1808 | */ | |
1809 | static void pcpu_dump_alloc_info(const char *lvl, | |
1810 | const struct pcpu_alloc_info *ai) | |
033e48fb | 1811 | { |
fd1e8a1f | 1812 | int group_width = 1, cpu_width = 1, width; |
033e48fb | 1813 | char empty_str[] = "--------"; |
fd1e8a1f TH |
1814 | int alloc = 0, alloc_end = 0; |
1815 | int group, v; | |
1816 | int upa, apl; /* units per alloc, allocs per line */ | |
1817 | ||
1818 | v = ai->nr_groups; | |
1819 | while (v /= 10) | |
1820 | group_width++; | |
033e48fb | 1821 | |
fd1e8a1f | 1822 | v = num_possible_cpus(); |
033e48fb | 1823 | while (v /= 10) |
fd1e8a1f TH |
1824 | cpu_width++; |
1825 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; | |
033e48fb | 1826 | |
fd1e8a1f TH |
1827 | upa = ai->alloc_size / ai->unit_size; |
1828 | width = upa * (cpu_width + 1) + group_width + 3; | |
1829 | apl = rounddown_pow_of_two(max(60 / width, 1)); | |
033e48fb | 1830 | |
fd1e8a1f TH |
1831 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
1832 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, | |
1833 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); | |
033e48fb | 1834 | |
fd1e8a1f TH |
1835 | for (group = 0; group < ai->nr_groups; group++) { |
1836 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
1837 | int unit = 0, unit_end = 0; | |
1838 | ||
1839 | BUG_ON(gi->nr_units % upa); | |
1840 | for (alloc_end += gi->nr_units / upa; | |
1841 | alloc < alloc_end; alloc++) { | |
1842 | if (!(alloc % apl)) { | |
1170532b | 1843 | pr_cont("\n"); |
fd1e8a1f TH |
1844 | printk("%spcpu-alloc: ", lvl); |
1845 | } | |
1170532b | 1846 | pr_cont("[%0*d] ", group_width, group); |
fd1e8a1f TH |
1847 | |
1848 | for (unit_end += upa; unit < unit_end; unit++) | |
1849 | if (gi->cpu_map[unit] != NR_CPUS) | |
1170532b JP |
1850 | pr_cont("%0*d ", |
1851 | cpu_width, gi->cpu_map[unit]); | |
fd1e8a1f | 1852 | else |
1170532b | 1853 | pr_cont("%s ", empty_str); |
033e48fb | 1854 | } |
033e48fb | 1855 | } |
1170532b | 1856 | pr_cont("\n"); |
033e48fb | 1857 | } |
033e48fb | 1858 | |
fbf59bc9 | 1859 | /** |
8d408b4b | 1860 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
fd1e8a1f | 1861 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
38a6be52 | 1862 | * @base_addr: mapped address |
8d408b4b TH |
1863 | * |
1864 | * Initialize the first percpu chunk which contains the kernel static | |
1865 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1866 | * setup path. |
8d408b4b | 1867 | * |
fd1e8a1f TH |
1868 | * @ai contains all information necessary to initialize the first |
1869 | * chunk and prime the dynamic percpu allocator. | |
1870 | * | |
1871 | * @ai->static_size is the size of static percpu area. | |
1872 | * | |
1873 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to | |
edcb4639 TH |
1874 | * reserve after the static area in the first chunk. This reserves |
1875 | * the first chunk such that it's available only through reserved | |
1876 | * percpu allocation. This is primarily used to serve module percpu | |
1877 | * static areas on architectures where the addressing model has | |
1878 | * limited offset range for symbol relocations to guarantee module | |
1879 | * percpu symbols fall inside the relocatable range. | |
1880 | * | |
fd1e8a1f TH |
1881 | * @ai->dyn_size determines the number of bytes available for dynamic |
1882 | * allocation in the first chunk. The area between @ai->static_size + | |
1883 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. | |
6074d5b0 | 1884 | * |
fd1e8a1f TH |
1885 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
1886 | * and equal to or larger than @ai->static_size + @ai->reserved_size + | |
1887 | * @ai->dyn_size. | |
8d408b4b | 1888 | * |
fd1e8a1f TH |
1889 | * @ai->atom_size is the allocation atom size and used as alignment |
1890 | * for vm areas. | |
8d408b4b | 1891 | * |
fd1e8a1f TH |
1892 | * @ai->alloc_size is the allocation size and always multiple of |
1893 | * @ai->atom_size. This is larger than @ai->atom_size if | |
1894 | * @ai->unit_size is larger than @ai->atom_size. | |
1895 | * | |
1896 | * @ai->nr_groups and @ai->groups describe virtual memory layout of | |
1897 | * percpu areas. Units which should be colocated are put into the | |
1898 | * same group. Dynamic VM areas will be allocated according to these | |
1899 | * groupings. If @ai->nr_groups is zero, a single group containing | |
1900 | * all units is assumed. | |
8d408b4b | 1901 | * |
38a6be52 TH |
1902 | * The caller should have mapped the first chunk at @base_addr and |
1903 | * copied static data to each unit. | |
fbf59bc9 | 1904 | * |
c0ebfdc3 DZF |
1905 | * The first chunk will always contain a static and a dynamic region. |
1906 | * However, the static region is not managed by any chunk. If the first | |
1907 | * chunk also contains a reserved region, it is served by two chunks - | |
1908 | * one for the reserved region and one for the dynamic region. They | |
1909 | * share the same vm, but use offset regions in the area allocation map. | |
1910 | * The chunk serving the dynamic region is circulated in the chunk slots | |
1911 | * and available for dynamic allocation like any other chunk. | |
edcb4639 | 1912 | * |
fbf59bc9 | 1913 | * RETURNS: |
fb435d52 | 1914 | * 0 on success, -errno on failure. |
fbf59bc9 | 1915 | */ |
fb435d52 TH |
1916 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
1917 | void *base_addr) | |
fbf59bc9 | 1918 | { |
b9c39442 | 1919 | size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
d2f3c384 | 1920 | size_t static_size, dyn_size; |
0c4169c3 | 1921 | struct pcpu_chunk *chunk; |
6563297c TH |
1922 | unsigned long *group_offsets; |
1923 | size_t *group_sizes; | |
fb435d52 | 1924 | unsigned long *unit_off; |
fbf59bc9 | 1925 | unsigned int cpu; |
fd1e8a1f TH |
1926 | int *unit_map; |
1927 | int group, unit, i; | |
c0ebfdc3 DZF |
1928 | int map_size; |
1929 | unsigned long tmp_addr; | |
fbf59bc9 | 1930 | |
635b75fc TH |
1931 | #define PCPU_SETUP_BUG_ON(cond) do { \ |
1932 | if (unlikely(cond)) { \ | |
870d4b12 JP |
1933 | pr_emerg("failed to initialize, %s\n", #cond); \ |
1934 | pr_emerg("cpu_possible_mask=%*pb\n", \ | |
807de073 | 1935 | cpumask_pr_args(cpu_possible_mask)); \ |
635b75fc TH |
1936 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ |
1937 | BUG(); \ | |
1938 | } \ | |
1939 | } while (0) | |
1940 | ||
2f39e637 | 1941 | /* sanity checks */ |
635b75fc | 1942 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
bbddff05 | 1943 | #ifdef CONFIG_SMP |
635b75fc | 1944 | PCPU_SETUP_BUG_ON(!ai->static_size); |
f09f1243 | 1945 | PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); |
bbddff05 | 1946 | #endif |
635b75fc | 1947 | PCPU_SETUP_BUG_ON(!base_addr); |
f09f1243 | 1948 | PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); |
635b75fc | 1949 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
f09f1243 | 1950 | PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); |
635b75fc | 1951 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); |
ca460b3c | 1952 | PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); |
099a19d9 | 1953 | PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); |
fb29a2cc | 1954 | PCPU_SETUP_BUG_ON(!ai->dyn_size); |
d2f3c384 | 1955 | PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); |
ca460b3c DZF |
1956 | PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || |
1957 | IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); | |
9f645532 | 1958 | PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); |
8d408b4b | 1959 | |
6563297c | 1960 | /* process group information and build config tables accordingly */ |
999c17e3 SS |
1961 | group_offsets = memblock_virt_alloc(ai->nr_groups * |
1962 | sizeof(group_offsets[0]), 0); | |
1963 | group_sizes = memblock_virt_alloc(ai->nr_groups * | |
1964 | sizeof(group_sizes[0]), 0); | |
1965 | unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); | |
1966 | unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); | |
2f39e637 | 1967 | |
fd1e8a1f | 1968 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
ffe0d5a5 | 1969 | unit_map[cpu] = UINT_MAX; |
a855b84c TH |
1970 | |
1971 | pcpu_low_unit_cpu = NR_CPUS; | |
1972 | pcpu_high_unit_cpu = NR_CPUS; | |
2f39e637 | 1973 | |
fd1e8a1f TH |
1974 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
1975 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
2f39e637 | 1976 | |
6563297c TH |
1977 | group_offsets[group] = gi->base_offset; |
1978 | group_sizes[group] = gi->nr_units * ai->unit_size; | |
1979 | ||
fd1e8a1f TH |
1980 | for (i = 0; i < gi->nr_units; i++) { |
1981 | cpu = gi->cpu_map[i]; | |
1982 | if (cpu == NR_CPUS) | |
1983 | continue; | |
8d408b4b | 1984 | |
9f295664 | 1985 | PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); |
635b75fc TH |
1986 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); |
1987 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); | |
fbf59bc9 | 1988 | |
fd1e8a1f | 1989 | unit_map[cpu] = unit + i; |
fb435d52 TH |
1990 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
1991 | ||
a855b84c TH |
1992 | /* determine low/high unit_cpu */ |
1993 | if (pcpu_low_unit_cpu == NR_CPUS || | |
1994 | unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) | |
1995 | pcpu_low_unit_cpu = cpu; | |
1996 | if (pcpu_high_unit_cpu == NR_CPUS || | |
1997 | unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) | |
1998 | pcpu_high_unit_cpu = cpu; | |
fd1e8a1f | 1999 | } |
2f39e637 | 2000 | } |
fd1e8a1f TH |
2001 | pcpu_nr_units = unit; |
2002 | ||
2003 | for_each_possible_cpu(cpu) | |
635b75fc TH |
2004 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
2005 | ||
2006 | /* we're done parsing the input, undefine BUG macro and dump config */ | |
2007 | #undef PCPU_SETUP_BUG_ON | |
bcbea798 | 2008 | pcpu_dump_alloc_info(KERN_DEBUG, ai); |
fd1e8a1f | 2009 | |
6563297c TH |
2010 | pcpu_nr_groups = ai->nr_groups; |
2011 | pcpu_group_offsets = group_offsets; | |
2012 | pcpu_group_sizes = group_sizes; | |
fd1e8a1f | 2013 | pcpu_unit_map = unit_map; |
fb435d52 | 2014 | pcpu_unit_offsets = unit_off; |
2f39e637 TH |
2015 | |
2016 | /* determine basic parameters */ | |
fd1e8a1f | 2017 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
d9b55eeb | 2018 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
6563297c | 2019 | pcpu_atom_size = ai->atom_size; |
ce3141a2 TH |
2020 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
2021 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
cafe8816 | 2022 | |
30a5b536 DZ |
2023 | pcpu_stats_save_ai(ai); |
2024 | ||
d9b55eeb TH |
2025 | /* |
2026 | * Allocate chunk slots. The additional last slot is for | |
2027 | * empty chunks. | |
2028 | */ | |
2029 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
999c17e3 SS |
2030 | pcpu_slot = memblock_virt_alloc( |
2031 | pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); | |
fbf59bc9 TH |
2032 | for (i = 0; i < pcpu_nr_slots; i++) |
2033 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
2034 | ||
d2f3c384 DZF |
2035 | /* |
2036 | * The end of the static region needs to be aligned with the | |
2037 | * minimum allocation size as this offsets the reserved and | |
2038 | * dynamic region. The first chunk ends page aligned by | |
2039 | * expanding the dynamic region, therefore the dynamic region | |
2040 | * can be shrunk to compensate while still staying above the | |
2041 | * configured sizes. | |
2042 | */ | |
2043 | static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); | |
2044 | dyn_size = ai->dyn_size - (static_size - ai->static_size); | |
2045 | ||
edcb4639 | 2046 | /* |
c0ebfdc3 DZF |
2047 | * Initialize first chunk. |
2048 | * If the reserved_size is non-zero, this initializes the reserved | |
2049 | * chunk. If the reserved_size is zero, the reserved chunk is NULL | |
2050 | * and the dynamic region is initialized here. The first chunk, | |
2051 | * pcpu_first_chunk, will always point to the chunk that serves | |
2052 | * the dynamic region. | |
edcb4639 | 2053 | */ |
d2f3c384 DZF |
2054 | tmp_addr = (unsigned long)base_addr + static_size; |
2055 | map_size = ai->reserved_size ?: dyn_size; | |
40064aec | 2056 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); |
61ace7fa | 2057 | |
edcb4639 | 2058 | /* init dynamic chunk if necessary */ |
b9c39442 | 2059 | if (ai->reserved_size) { |
0c4169c3 | 2060 | pcpu_reserved_chunk = chunk; |
b9c39442 | 2061 | |
d2f3c384 | 2062 | tmp_addr = (unsigned long)base_addr + static_size + |
c0ebfdc3 | 2063 | ai->reserved_size; |
d2f3c384 | 2064 | map_size = dyn_size; |
40064aec | 2065 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); |
edcb4639 TH |
2066 | } |
2067 | ||
2441d15c | 2068 | /* link the first chunk in */ |
0c4169c3 | 2069 | pcpu_first_chunk = chunk; |
0cecf50c | 2070 | pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; |
ae9e6bc9 | 2071 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
fbf59bc9 | 2072 | |
30a5b536 | 2073 | pcpu_stats_chunk_alloc(); |
df95e795 | 2074 | trace_percpu_create_chunk(base_addr); |
30a5b536 | 2075 | |
fbf59bc9 | 2076 | /* we're done */ |
bba174f5 | 2077 | pcpu_base_addr = base_addr; |
fb435d52 | 2078 | return 0; |
fbf59bc9 | 2079 | } |
66c3a757 | 2080 | |
bbddff05 TH |
2081 | #ifdef CONFIG_SMP |
2082 | ||
17f3609c | 2083 | const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { |
f58dc01b TH |
2084 | [PCPU_FC_AUTO] = "auto", |
2085 | [PCPU_FC_EMBED] = "embed", | |
2086 | [PCPU_FC_PAGE] = "page", | |
f58dc01b | 2087 | }; |
66c3a757 | 2088 | |
f58dc01b | 2089 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
66c3a757 | 2090 | |
f58dc01b TH |
2091 | static int __init percpu_alloc_setup(char *str) |
2092 | { | |
5479c78a CG |
2093 | if (!str) |
2094 | return -EINVAL; | |
2095 | ||
f58dc01b TH |
2096 | if (0) |
2097 | /* nada */; | |
2098 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK | |
2099 | else if (!strcmp(str, "embed")) | |
2100 | pcpu_chosen_fc = PCPU_FC_EMBED; | |
2101 | #endif | |
2102 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
2103 | else if (!strcmp(str, "page")) | |
2104 | pcpu_chosen_fc = PCPU_FC_PAGE; | |
f58dc01b TH |
2105 | #endif |
2106 | else | |
870d4b12 | 2107 | pr_warn("unknown allocator %s specified\n", str); |
66c3a757 | 2108 | |
f58dc01b | 2109 | return 0; |
66c3a757 | 2110 | } |
f58dc01b | 2111 | early_param("percpu_alloc", percpu_alloc_setup); |
66c3a757 | 2112 | |
3c9a024f TH |
2113 | /* |
2114 | * pcpu_embed_first_chunk() is used by the generic percpu setup. | |
2115 | * Build it if needed by the arch config or the generic setup is going | |
2116 | * to be used. | |
2117 | */ | |
08fc4580 TH |
2118 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
2119 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
3c9a024f TH |
2120 | #define BUILD_EMBED_FIRST_CHUNK |
2121 | #endif | |
2122 | ||
2123 | /* build pcpu_page_first_chunk() iff needed by the arch config */ | |
2124 | #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) | |
2125 | #define BUILD_PAGE_FIRST_CHUNK | |
2126 | #endif | |
2127 | ||
2128 | /* pcpu_build_alloc_info() is used by both embed and page first chunk */ | |
2129 | #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) | |
2130 | /** | |
2131 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs | |
2132 | * @reserved_size: the size of reserved percpu area in bytes | |
2133 | * @dyn_size: minimum free size for dynamic allocation in bytes | |
2134 | * @atom_size: allocation atom size | |
2135 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
2136 | * | |
2137 | * This function determines grouping of units, their mappings to cpus | |
2138 | * and other parameters considering needed percpu size, allocation | |
2139 | * atom size and distances between CPUs. | |
2140 | * | |
bffc4375 | 2141 | * Groups are always multiples of atom size and CPUs which are of |
3c9a024f TH |
2142 | * LOCAL_DISTANCE both ways are grouped together and share space for |
2143 | * units in the same group. The returned configuration is guaranteed | |
2144 | * to have CPUs on different nodes on different groups and >=75% usage | |
2145 | * of allocated virtual address space. | |
2146 | * | |
2147 | * RETURNS: | |
2148 | * On success, pointer to the new allocation_info is returned. On | |
2149 | * failure, ERR_PTR value is returned. | |
2150 | */ | |
2151 | static struct pcpu_alloc_info * __init pcpu_build_alloc_info( | |
2152 | size_t reserved_size, size_t dyn_size, | |
2153 | size_t atom_size, | |
2154 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
2155 | { | |
2156 | static int group_map[NR_CPUS] __initdata; | |
2157 | static int group_cnt[NR_CPUS] __initdata; | |
2158 | const size_t static_size = __per_cpu_end - __per_cpu_start; | |
2159 | int nr_groups = 1, nr_units = 0; | |
2160 | size_t size_sum, min_unit_size, alloc_size; | |
2161 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
2162 | int last_allocs, group, unit; | |
2163 | unsigned int cpu, tcpu; | |
2164 | struct pcpu_alloc_info *ai; | |
2165 | unsigned int *cpu_map; | |
2166 | ||
2167 | /* this function may be called multiple times */ | |
2168 | memset(group_map, 0, sizeof(group_map)); | |
2169 | memset(group_cnt, 0, sizeof(group_cnt)); | |
2170 | ||
2171 | /* calculate size_sum and ensure dyn_size is enough for early alloc */ | |
2172 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
2173 | max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); | |
2174 | dyn_size = size_sum - static_size - reserved_size; | |
2175 | ||
2176 | /* | |
2177 | * Determine min_unit_size, alloc_size and max_upa such that | |
2178 | * alloc_size is multiple of atom_size and is the smallest | |
25985edc | 2179 | * which can accommodate 4k aligned segments which are equal to |
3c9a024f TH |
2180 | * or larger than min_unit_size. |
2181 | */ | |
2182 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
2183 | ||
9c015162 | 2184 | /* determine the maximum # of units that can fit in an allocation */ |
3c9a024f TH |
2185 | alloc_size = roundup(min_unit_size, atom_size); |
2186 | upa = alloc_size / min_unit_size; | |
f09f1243 | 2187 | while (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
2188 | upa--; |
2189 | max_upa = upa; | |
2190 | ||
2191 | /* group cpus according to their proximity */ | |
2192 | for_each_possible_cpu(cpu) { | |
2193 | group = 0; | |
2194 | next_group: | |
2195 | for_each_possible_cpu(tcpu) { | |
2196 | if (cpu == tcpu) | |
2197 | break; | |
2198 | if (group_map[tcpu] == group && cpu_distance_fn && | |
2199 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || | |
2200 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
2201 | group++; | |
2202 | nr_groups = max(nr_groups, group + 1); | |
2203 | goto next_group; | |
2204 | } | |
2205 | } | |
2206 | group_map[cpu] = group; | |
2207 | group_cnt[group]++; | |
2208 | } | |
2209 | ||
2210 | /* | |
9c015162 DZF |
2211 | * Wasted space is caused by a ratio imbalance of upa to group_cnt. |
2212 | * Expand the unit_size until we use >= 75% of the units allocated. | |
2213 | * Related to atom_size, which could be much larger than the unit_size. | |
3c9a024f TH |
2214 | */ |
2215 | last_allocs = INT_MAX; | |
2216 | for (upa = max_upa; upa; upa--) { | |
2217 | int allocs = 0, wasted = 0; | |
2218 | ||
f09f1243 | 2219 | if (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
2220 | continue; |
2221 | ||
2222 | for (group = 0; group < nr_groups; group++) { | |
2223 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); | |
2224 | allocs += this_allocs; | |
2225 | wasted += this_allocs * upa - group_cnt[group]; | |
2226 | } | |
2227 | ||
2228 | /* | |
2229 | * Don't accept if wastage is over 1/3. The | |
2230 | * greater-than comparison ensures upa==1 always | |
2231 | * passes the following check. | |
2232 | */ | |
2233 | if (wasted > num_possible_cpus() / 3) | |
2234 | continue; | |
2235 | ||
2236 | /* and then don't consume more memory */ | |
2237 | if (allocs > last_allocs) | |
2238 | break; | |
2239 | last_allocs = allocs; | |
2240 | best_upa = upa; | |
2241 | } | |
2242 | upa = best_upa; | |
2243 | ||
2244 | /* allocate and fill alloc_info */ | |
2245 | for (group = 0; group < nr_groups; group++) | |
2246 | nr_units += roundup(group_cnt[group], upa); | |
2247 | ||
2248 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); | |
2249 | if (!ai) | |
2250 | return ERR_PTR(-ENOMEM); | |
2251 | cpu_map = ai->groups[0].cpu_map; | |
2252 | ||
2253 | for (group = 0; group < nr_groups; group++) { | |
2254 | ai->groups[group].cpu_map = cpu_map; | |
2255 | cpu_map += roundup(group_cnt[group], upa); | |
2256 | } | |
2257 | ||
2258 | ai->static_size = static_size; | |
2259 | ai->reserved_size = reserved_size; | |
2260 | ai->dyn_size = dyn_size; | |
2261 | ai->unit_size = alloc_size / upa; | |
2262 | ai->atom_size = atom_size; | |
2263 | ai->alloc_size = alloc_size; | |
2264 | ||
2265 | for (group = 0, unit = 0; group_cnt[group]; group++) { | |
2266 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2267 | ||
2268 | /* | |
2269 | * Initialize base_offset as if all groups are located | |
2270 | * back-to-back. The caller should update this to | |
2271 | * reflect actual allocation. | |
2272 | */ | |
2273 | gi->base_offset = unit * ai->unit_size; | |
2274 | ||
2275 | for_each_possible_cpu(cpu) | |
2276 | if (group_map[cpu] == group) | |
2277 | gi->cpu_map[gi->nr_units++] = cpu; | |
2278 | gi->nr_units = roundup(gi->nr_units, upa); | |
2279 | unit += gi->nr_units; | |
2280 | } | |
2281 | BUG_ON(unit != nr_units); | |
2282 | ||
2283 | return ai; | |
2284 | } | |
2285 | #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ | |
2286 | ||
2287 | #if defined(BUILD_EMBED_FIRST_CHUNK) | |
66c3a757 TH |
2288 | /** |
2289 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
66c3a757 | 2290 | * @reserved_size: the size of reserved percpu area in bytes |
4ba6ce25 | 2291 | * @dyn_size: minimum free size for dynamic allocation in bytes |
c8826dd5 TH |
2292 | * @atom_size: allocation atom size |
2293 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
2294 | * @alloc_fn: function to allocate percpu page | |
25985edc | 2295 | * @free_fn: function to free percpu page |
66c3a757 TH |
2296 | * |
2297 | * This is a helper to ease setting up embedded first percpu chunk and | |
2298 | * can be called where pcpu_setup_first_chunk() is expected. | |
2299 | * | |
2300 | * If this function is used to setup the first chunk, it is allocated | |
c8826dd5 TH |
2301 | * by calling @alloc_fn and used as-is without being mapped into |
2302 | * vmalloc area. Allocations are always whole multiples of @atom_size | |
2303 | * aligned to @atom_size. | |
2304 | * | |
2305 | * This enables the first chunk to piggy back on the linear physical | |
2306 | * mapping which often uses larger page size. Please note that this | |
2307 | * can result in very sparse cpu->unit mapping on NUMA machines thus | |
2308 | * requiring large vmalloc address space. Don't use this allocator if | |
2309 | * vmalloc space is not orders of magnitude larger than distances | |
2310 | * between node memory addresses (ie. 32bit NUMA machines). | |
66c3a757 | 2311 | * |
4ba6ce25 | 2312 | * @dyn_size specifies the minimum dynamic area size. |
66c3a757 TH |
2313 | * |
2314 | * If the needed size is smaller than the minimum or specified unit | |
c8826dd5 | 2315 | * size, the leftover is returned using @free_fn. |
66c3a757 TH |
2316 | * |
2317 | * RETURNS: | |
fb435d52 | 2318 | * 0 on success, -errno on failure. |
66c3a757 | 2319 | */ |
4ba6ce25 | 2320 | int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, |
c8826dd5 TH |
2321 | size_t atom_size, |
2322 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, | |
2323 | pcpu_fc_alloc_fn_t alloc_fn, | |
2324 | pcpu_fc_free_fn_t free_fn) | |
66c3a757 | 2325 | { |
c8826dd5 TH |
2326 | void *base = (void *)ULONG_MAX; |
2327 | void **areas = NULL; | |
fd1e8a1f | 2328 | struct pcpu_alloc_info *ai; |
93c76b6b | 2329 | size_t size_sum, areas_size; |
2330 | unsigned long max_distance; | |
9b739662 | 2331 | int group, i, highest_group, rc; |
66c3a757 | 2332 | |
c8826dd5 TH |
2333 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
2334 | cpu_distance_fn); | |
fd1e8a1f TH |
2335 | if (IS_ERR(ai)) |
2336 | return PTR_ERR(ai); | |
66c3a757 | 2337 | |
fd1e8a1f | 2338 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
c8826dd5 | 2339 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
fa8a7094 | 2340 | |
999c17e3 | 2341 | areas = memblock_virt_alloc_nopanic(areas_size, 0); |
c8826dd5 | 2342 | if (!areas) { |
fb435d52 | 2343 | rc = -ENOMEM; |
c8826dd5 | 2344 | goto out_free; |
fa8a7094 | 2345 | } |
66c3a757 | 2346 | |
9b739662 | 2347 | /* allocate, copy and determine base address & max_distance */ |
2348 | highest_group = 0; | |
c8826dd5 TH |
2349 | for (group = 0; group < ai->nr_groups; group++) { |
2350 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2351 | unsigned int cpu = NR_CPUS; | |
2352 | void *ptr; | |
2353 | ||
2354 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) | |
2355 | cpu = gi->cpu_map[i]; | |
2356 | BUG_ON(cpu == NR_CPUS); | |
2357 | ||
2358 | /* allocate space for the whole group */ | |
2359 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); | |
2360 | if (!ptr) { | |
2361 | rc = -ENOMEM; | |
2362 | goto out_free_areas; | |
2363 | } | |
f528f0b8 CM |
2364 | /* kmemleak tracks the percpu allocations separately */ |
2365 | kmemleak_free(ptr); | |
c8826dd5 | 2366 | areas[group] = ptr; |
fd1e8a1f | 2367 | |
c8826dd5 | 2368 | base = min(ptr, base); |
9b739662 | 2369 | if (ptr > areas[highest_group]) |
2370 | highest_group = group; | |
2371 | } | |
2372 | max_distance = areas[highest_group] - base; | |
2373 | max_distance += ai->unit_size * ai->groups[highest_group].nr_units; | |
2374 | ||
2375 | /* warn if maximum distance is further than 75% of vmalloc space */ | |
2376 | if (max_distance > VMALLOC_TOTAL * 3 / 4) { | |
2377 | pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", | |
2378 | max_distance, VMALLOC_TOTAL); | |
2379 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
2380 | /* and fail if we have fallback */ | |
2381 | rc = -EINVAL; | |
2382 | goto out_free_areas; | |
2383 | #endif | |
42b64281 TH |
2384 | } |
2385 | ||
2386 | /* | |
2387 | * Copy data and free unused parts. This should happen after all | |
2388 | * allocations are complete; otherwise, we may end up with | |
2389 | * overlapping groups. | |
2390 | */ | |
2391 | for (group = 0; group < ai->nr_groups; group++) { | |
2392 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2393 | void *ptr = areas[group]; | |
c8826dd5 TH |
2394 | |
2395 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { | |
2396 | if (gi->cpu_map[i] == NR_CPUS) { | |
2397 | /* unused unit, free whole */ | |
2398 | free_fn(ptr, ai->unit_size); | |
2399 | continue; | |
2400 | } | |
2401 | /* copy and return the unused part */ | |
2402 | memcpy(ptr, __per_cpu_load, ai->static_size); | |
2403 | free_fn(ptr + size_sum, ai->unit_size - size_sum); | |
2404 | } | |
fa8a7094 | 2405 | } |
66c3a757 | 2406 | |
c8826dd5 | 2407 | /* base address is now known, determine group base offsets */ |
6ea529a2 | 2408 | for (group = 0; group < ai->nr_groups; group++) { |
c8826dd5 | 2409 | ai->groups[group].base_offset = areas[group] - base; |
6ea529a2 | 2410 | } |
c8826dd5 | 2411 | |
870d4b12 | 2412 | pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
fd1e8a1f TH |
2413 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
2414 | ai->dyn_size, ai->unit_size); | |
d4b95f80 | 2415 | |
fb435d52 | 2416 | rc = pcpu_setup_first_chunk(ai, base); |
c8826dd5 TH |
2417 | goto out_free; |
2418 | ||
2419 | out_free_areas: | |
2420 | for (group = 0; group < ai->nr_groups; group++) | |
f851c8d8 MH |
2421 | if (areas[group]) |
2422 | free_fn(areas[group], | |
2423 | ai->groups[group].nr_units * ai->unit_size); | |
c8826dd5 | 2424 | out_free: |
fd1e8a1f | 2425 | pcpu_free_alloc_info(ai); |
c8826dd5 | 2426 | if (areas) |
999c17e3 | 2427 | memblock_free_early(__pa(areas), areas_size); |
fb435d52 | 2428 | return rc; |
d4b95f80 | 2429 | } |
3c9a024f | 2430 | #endif /* BUILD_EMBED_FIRST_CHUNK */ |
d4b95f80 | 2431 | |
3c9a024f | 2432 | #ifdef BUILD_PAGE_FIRST_CHUNK |
d4b95f80 | 2433 | /** |
00ae4064 | 2434 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
2435 | * @reserved_size: the size of reserved percpu area in bytes |
2436 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
25985edc | 2437 | * @free_fn: function to free percpu page, always called with PAGE_SIZE |
d4b95f80 TH |
2438 | * @populate_pte_fn: function to populate pte |
2439 | * | |
00ae4064 TH |
2440 | * This is a helper to ease setting up page-remapped first percpu |
2441 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
2442 | * |
2443 | * This is the basic allocator. Static percpu area is allocated | |
2444 | * page-by-page into vmalloc area. | |
2445 | * | |
2446 | * RETURNS: | |
fb435d52 | 2447 | * 0 on success, -errno on failure. |
d4b95f80 | 2448 | */ |
fb435d52 TH |
2449 | int __init pcpu_page_first_chunk(size_t reserved_size, |
2450 | pcpu_fc_alloc_fn_t alloc_fn, | |
2451 | pcpu_fc_free_fn_t free_fn, | |
2452 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 2453 | { |
8f05a6a6 | 2454 | static struct vm_struct vm; |
fd1e8a1f | 2455 | struct pcpu_alloc_info *ai; |
00ae4064 | 2456 | char psize_str[16]; |
ce3141a2 | 2457 | int unit_pages; |
d4b95f80 | 2458 | size_t pages_size; |
ce3141a2 | 2459 | struct page **pages; |
fb435d52 | 2460 | int unit, i, j, rc; |
8f606604 | 2461 | int upa; |
2462 | int nr_g0_units; | |
d4b95f80 | 2463 | |
00ae4064 TH |
2464 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
2465 | ||
4ba6ce25 | 2466 | ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); |
fd1e8a1f TH |
2467 | if (IS_ERR(ai)) |
2468 | return PTR_ERR(ai); | |
2469 | BUG_ON(ai->nr_groups != 1); | |
8f606604 | 2470 | upa = ai->alloc_size/ai->unit_size; |
2471 | nr_g0_units = roundup(num_possible_cpus(), upa); | |
2472 | if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) { | |
2473 | pcpu_free_alloc_info(ai); | |
2474 | return -EINVAL; | |
2475 | } | |
fd1e8a1f TH |
2476 | |
2477 | unit_pages = ai->unit_size >> PAGE_SHIFT; | |
d4b95f80 TH |
2478 | |
2479 | /* unaligned allocations can't be freed, round up to page size */ | |
fd1e8a1f TH |
2480 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
2481 | sizeof(pages[0])); | |
999c17e3 | 2482 | pages = memblock_virt_alloc(pages_size, 0); |
d4b95f80 | 2483 | |
8f05a6a6 | 2484 | /* allocate pages */ |
d4b95f80 | 2485 | j = 0; |
8f606604 | 2486 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
2487 | unsigned int cpu = ai->groups[0].cpu_map[unit]; | |
ce3141a2 | 2488 | for (i = 0; i < unit_pages; i++) { |
d4b95f80 TH |
2489 | void *ptr; |
2490 | ||
3cbc8565 | 2491 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
d4b95f80 | 2492 | if (!ptr) { |
870d4b12 | 2493 | pr_warn("failed to allocate %s page for cpu%u\n", |
8f606604 | 2494 | psize_str, cpu); |
d4b95f80 TH |
2495 | goto enomem; |
2496 | } | |
f528f0b8 CM |
2497 | /* kmemleak tracks the percpu allocations separately */ |
2498 | kmemleak_free(ptr); | |
ce3141a2 | 2499 | pages[j++] = virt_to_page(ptr); |
d4b95f80 | 2500 | } |
8f606604 | 2501 | } |
d4b95f80 | 2502 | |
8f05a6a6 TH |
2503 | /* allocate vm area, map the pages and copy static data */ |
2504 | vm.flags = VM_ALLOC; | |
fd1e8a1f | 2505 | vm.size = num_possible_cpus() * ai->unit_size; |
8f05a6a6 TH |
2506 | vm_area_register_early(&vm, PAGE_SIZE); |
2507 | ||
fd1e8a1f | 2508 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
1d9d3257 | 2509 | unsigned long unit_addr = |
fd1e8a1f | 2510 | (unsigned long)vm.addr + unit * ai->unit_size; |
8f05a6a6 | 2511 | |
ce3141a2 | 2512 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
2513 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
2514 | ||
2515 | /* pte already populated, the following shouldn't fail */ | |
fb435d52 TH |
2516 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
2517 | unit_pages); | |
2518 | if (rc < 0) | |
2519 | panic("failed to map percpu area, err=%d\n", rc); | |
66c3a757 | 2520 | |
8f05a6a6 TH |
2521 | /* |
2522 | * FIXME: Archs with virtual cache should flush local | |
2523 | * cache for the linear mapping here - something | |
2524 | * equivalent to flush_cache_vmap() on the local cpu. | |
2525 | * flush_cache_vmap() can't be used as most supporting | |
2526 | * data structures are not set up yet. | |
2527 | */ | |
2528 | ||
2529 | /* copy static data */ | |
fd1e8a1f | 2530 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
66c3a757 TH |
2531 | } |
2532 | ||
2533 | /* we're ready, commit */ | |
870d4b12 | 2534 | pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", |
fd1e8a1f TH |
2535 | unit_pages, psize_str, vm.addr, ai->static_size, |
2536 | ai->reserved_size, ai->dyn_size); | |
d4b95f80 | 2537 | |
fb435d52 | 2538 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
d4b95f80 TH |
2539 | goto out_free_ar; |
2540 | ||
2541 | enomem: | |
2542 | while (--j >= 0) | |
ce3141a2 | 2543 | free_fn(page_address(pages[j]), PAGE_SIZE); |
fb435d52 | 2544 | rc = -ENOMEM; |
d4b95f80 | 2545 | out_free_ar: |
999c17e3 | 2546 | memblock_free_early(__pa(pages), pages_size); |
fd1e8a1f | 2547 | pcpu_free_alloc_info(ai); |
fb435d52 | 2548 | return rc; |
d4b95f80 | 2549 | } |
3c9a024f | 2550 | #endif /* BUILD_PAGE_FIRST_CHUNK */ |
d4b95f80 | 2551 | |
bbddff05 | 2552 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
e74e3962 | 2553 | /* |
bbddff05 | 2554 | * Generic SMP percpu area setup. |
e74e3962 TH |
2555 | * |
2556 | * The embedding helper is used because its behavior closely resembles | |
2557 | * the original non-dynamic generic percpu area setup. This is | |
2558 | * important because many archs have addressing restrictions and might | |
2559 | * fail if the percpu area is located far away from the previous | |
2560 | * location. As an added bonus, in non-NUMA cases, embedding is | |
2561 | * generally a good idea TLB-wise because percpu area can piggy back | |
2562 | * on the physical linear memory mapping which uses large page | |
2563 | * mappings on applicable archs. | |
2564 | */ | |
e74e3962 TH |
2565 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
2566 | EXPORT_SYMBOL(__per_cpu_offset); | |
2567 | ||
c8826dd5 TH |
2568 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
2569 | size_t align) | |
2570 | { | |
999c17e3 SS |
2571 | return memblock_virt_alloc_from_nopanic( |
2572 | size, align, __pa(MAX_DMA_ADDRESS)); | |
c8826dd5 | 2573 | } |
66c3a757 | 2574 | |
c8826dd5 TH |
2575 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
2576 | { | |
999c17e3 | 2577 | memblock_free_early(__pa(ptr), size); |
c8826dd5 TH |
2578 | } |
2579 | ||
e74e3962 TH |
2580 | void __init setup_per_cpu_areas(void) |
2581 | { | |
e74e3962 TH |
2582 | unsigned long delta; |
2583 | unsigned int cpu; | |
fb435d52 | 2584 | int rc; |
e74e3962 TH |
2585 | |
2586 | /* | |
2587 | * Always reserve area for module percpu variables. That's | |
2588 | * what the legacy allocator did. | |
2589 | */ | |
fb435d52 | 2590 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
c8826dd5 TH |
2591 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
2592 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); | |
fb435d52 | 2593 | if (rc < 0) |
bbddff05 | 2594 | panic("Failed to initialize percpu areas."); |
e74e3962 TH |
2595 | |
2596 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2597 | for_each_possible_cpu(cpu) | |
fb435d52 | 2598 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
66c3a757 | 2599 | } |
bbddff05 TH |
2600 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
2601 | ||
2602 | #else /* CONFIG_SMP */ | |
2603 | ||
2604 | /* | |
2605 | * UP percpu area setup. | |
2606 | * | |
2607 | * UP always uses km-based percpu allocator with identity mapping. | |
2608 | * Static percpu variables are indistinguishable from the usual static | |
2609 | * variables and don't require any special preparation. | |
2610 | */ | |
2611 | void __init setup_per_cpu_areas(void) | |
2612 | { | |
2613 | const size_t unit_size = | |
2614 | roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, | |
2615 | PERCPU_DYNAMIC_RESERVE)); | |
2616 | struct pcpu_alloc_info *ai; | |
2617 | void *fc; | |
2618 | ||
2619 | ai = pcpu_alloc_alloc_info(1, 1); | |
999c17e3 SS |
2620 | fc = memblock_virt_alloc_from_nopanic(unit_size, |
2621 | PAGE_SIZE, | |
2622 | __pa(MAX_DMA_ADDRESS)); | |
bbddff05 TH |
2623 | if (!ai || !fc) |
2624 | panic("Failed to allocate memory for percpu areas."); | |
100d13c3 CM |
2625 | /* kmemleak tracks the percpu allocations separately */ |
2626 | kmemleak_free(fc); | |
bbddff05 TH |
2627 | |
2628 | ai->dyn_size = unit_size; | |
2629 | ai->unit_size = unit_size; | |
2630 | ai->atom_size = unit_size; | |
2631 | ai->alloc_size = unit_size; | |
2632 | ai->groups[0].nr_units = 1; | |
2633 | ai->groups[0].cpu_map[0] = 0; | |
2634 | ||
2635 | if (pcpu_setup_first_chunk(ai, fc) < 0) | |
2636 | panic("Failed to initialize percpu areas."); | |
2637 | } | |
2638 | ||
2639 | #endif /* CONFIG_SMP */ | |
099a19d9 | 2640 | |
1a4d7607 TH |
2641 | /* |
2642 | * Percpu allocator is initialized early during boot when neither slab or | |
2643 | * workqueue is available. Plug async management until everything is up | |
2644 | * and running. | |
2645 | */ | |
2646 | static int __init percpu_enable_async(void) | |
2647 | { | |
2648 | pcpu_async_enabled = true; | |
2649 | return 0; | |
2650 | } | |
2651 | subsys_initcall(percpu_enable_async); |