]>
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 | * | |
5e81ee3e DZF |
7 | * Copyright (C) 2017 Facebook Inc. |
8 | * Copyright (C) 2017 Dennis Zhou <dennisszhou@gmail.com> | |
9 | * | |
9c015162 | 10 | * This file is released under the GPLv2 license. |
fbf59bc9 | 11 | * |
9c015162 DZF |
12 | * The percpu allocator handles both static and dynamic areas. Percpu |
13 | * areas are allocated in chunks which are divided into units. There is | |
14 | * a 1-to-1 mapping for units to possible cpus. These units are grouped | |
15 | * based on NUMA properties of the machine. | |
fbf59bc9 TH |
16 | * |
17 | * c0 c1 c2 | |
18 | * ------------------- ------------------- ------------ | |
19 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u | |
20 | * ------------------- ...... ------------------- .... ------------ | |
21 | * | |
9c015162 DZF |
22 | * Allocation is done by offsets into a unit's address space. Ie., an |
23 | * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, | |
24 | * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear | |
25 | * and even sparse. Access is handled by configuring percpu base | |
26 | * registers according to the cpu to unit mappings and offsetting the | |
27 | * base address using pcpu_unit_size. | |
28 | * | |
29 | * There is special consideration for the first chunk which must handle | |
30 | * the static percpu variables in the kernel image as allocation services | |
5e81ee3e | 31 | * are not online yet. In short, the first chunk is structured like so: |
9c015162 DZF |
32 | * |
33 | * <Static | [Reserved] | Dynamic> | |
34 | * | |
35 | * The static data is copied from the original section managed by the | |
36 | * linker. The reserved section, if non-zero, primarily manages static | |
37 | * percpu variables from kernel modules. Finally, the dynamic section | |
38 | * takes care of normal allocations. | |
fbf59bc9 | 39 | * |
5e81ee3e DZF |
40 | * The allocator organizes chunks into lists according to free size and |
41 | * tries to allocate from the fullest chunk first. Each chunk is managed | |
42 | * by a bitmap with metadata blocks. The allocation map is updated on | |
43 | * every allocation and free to reflect the current state while the boundary | |
44 | * map is only updated on allocation. Each metadata block contains | |
45 | * information to help mitigate the need to iterate over large portions | |
46 | * of the bitmap. The reverse mapping from page to chunk is stored in | |
47 | * the page's index. Lastly, units are lazily backed and grow in unison. | |
48 | * | |
49 | * There is a unique conversion that goes on here between bytes and bits. | |
50 | * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk | |
51 | * tracks the number of pages it is responsible for in nr_pages. Helper | |
52 | * functions are used to convert from between the bytes, bits, and blocks. | |
53 | * All hints are managed in bits unless explicitly stated. | |
9c015162 | 54 | * |
4091fb95 | 55 | * To use this allocator, arch code should do the following: |
fbf59bc9 | 56 | * |
fbf59bc9 | 57 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
e0100983 TH |
58 | * regular address to percpu pointer and back if they need to be |
59 | * different from the default | |
fbf59bc9 | 60 | * |
8d408b4b TH |
61 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
62 | * setup the first chunk containing the kernel static percpu area | |
fbf59bc9 TH |
63 | */ |
64 | ||
870d4b12 JP |
65 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
66 | ||
fbf59bc9 TH |
67 | #include <linux/bitmap.h> |
68 | #include <linux/bootmem.h> | |
fd1e8a1f | 69 | #include <linux/err.h> |
ca460b3c | 70 | #include <linux/lcm.h> |
fbf59bc9 | 71 | #include <linux/list.h> |
a530b795 | 72 | #include <linux/log2.h> |
fbf59bc9 TH |
73 | #include <linux/mm.h> |
74 | #include <linux/module.h> | |
75 | #include <linux/mutex.h> | |
76 | #include <linux/percpu.h> | |
77 | #include <linux/pfn.h> | |
fbf59bc9 | 78 | #include <linux/slab.h> |
ccea34b5 | 79 | #include <linux/spinlock.h> |
fbf59bc9 | 80 | #include <linux/vmalloc.h> |
a56dbddf | 81 | #include <linux/workqueue.h> |
f528f0b8 | 82 | #include <linux/kmemleak.h> |
fbf59bc9 TH |
83 | |
84 | #include <asm/cacheflush.h> | |
e0100983 | 85 | #include <asm/sections.h> |
fbf59bc9 | 86 | #include <asm/tlbflush.h> |
3b034b0d | 87 | #include <asm/io.h> |
fbf59bc9 | 88 | |
df95e795 DZ |
89 | #define CREATE_TRACE_POINTS |
90 | #include <trace/events/percpu.h> | |
91 | ||
8fa3ed80 DZ |
92 | #include "percpu-internal.h" |
93 | ||
40064aec DZF |
94 | /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */ |
95 | #define PCPU_SLOT_BASE_SHIFT 5 | |
96 | ||
1a4d7607 TH |
97 | #define PCPU_EMPTY_POP_PAGES_LOW 2 |
98 | #define PCPU_EMPTY_POP_PAGES_HIGH 4 | |
fbf59bc9 | 99 | |
bbddff05 | 100 | #ifdef CONFIG_SMP |
e0100983 TH |
101 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
102 | #ifndef __addr_to_pcpu_ptr | |
103 | #define __addr_to_pcpu_ptr(addr) \ | |
43cf38eb TH |
104 | (void __percpu *)((unsigned long)(addr) - \ |
105 | (unsigned long)pcpu_base_addr + \ | |
106 | (unsigned long)__per_cpu_start) | |
e0100983 TH |
107 | #endif |
108 | #ifndef __pcpu_ptr_to_addr | |
109 | #define __pcpu_ptr_to_addr(ptr) \ | |
43cf38eb TH |
110 | (void __force *)((unsigned long)(ptr) + \ |
111 | (unsigned long)pcpu_base_addr - \ | |
112 | (unsigned long)__per_cpu_start) | |
e0100983 | 113 | #endif |
bbddff05 TH |
114 | #else /* CONFIG_SMP */ |
115 | /* on UP, it's always identity mapped */ | |
116 | #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) | |
117 | #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) | |
118 | #endif /* CONFIG_SMP */ | |
e0100983 | 119 | |
1328710b DM |
120 | static int pcpu_unit_pages __ro_after_init; |
121 | static int pcpu_unit_size __ro_after_init; | |
122 | static int pcpu_nr_units __ro_after_init; | |
123 | static int pcpu_atom_size __ro_after_init; | |
8fa3ed80 | 124 | int pcpu_nr_slots __ro_after_init; |
1328710b | 125 | static size_t pcpu_chunk_struct_size __ro_after_init; |
fbf59bc9 | 126 | |
a855b84c | 127 | /* cpus with the lowest and highest unit addresses */ |
1328710b DM |
128 | static unsigned int pcpu_low_unit_cpu __ro_after_init; |
129 | static unsigned int pcpu_high_unit_cpu __ro_after_init; | |
2f39e637 | 130 | |
fbf59bc9 | 131 | /* the address of the first chunk which starts with the kernel static area */ |
1328710b | 132 | void *pcpu_base_addr __ro_after_init; |
fbf59bc9 TH |
133 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
134 | ||
1328710b DM |
135 | static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ |
136 | const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ | |
2f39e637 | 137 | |
6563297c | 138 | /* group information, used for vm allocation */ |
1328710b DM |
139 | static int pcpu_nr_groups __ro_after_init; |
140 | static const unsigned long *pcpu_group_offsets __ro_after_init; | |
141 | static const size_t *pcpu_group_sizes __ro_after_init; | |
6563297c | 142 | |
ae9e6bc9 TH |
143 | /* |
144 | * The first chunk which always exists. Note that unlike other | |
145 | * chunks, this one can be allocated and mapped in several different | |
146 | * ways and thus often doesn't live in the vmalloc area. | |
147 | */ | |
8fa3ed80 | 148 | struct pcpu_chunk *pcpu_first_chunk __ro_after_init; |
ae9e6bc9 TH |
149 | |
150 | /* | |
151 | * Optional reserved chunk. This chunk reserves part of the first | |
e2266705 DZF |
152 | * chunk and serves it for reserved allocations. When the reserved |
153 | * region doesn't exist, the following variable is NULL. | |
ae9e6bc9 | 154 | */ |
8fa3ed80 | 155 | struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; |
edcb4639 | 156 | |
8fa3ed80 | 157 | DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ |
6710e594 | 158 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ |
fbf59bc9 | 159 | |
8fa3ed80 | 160 | struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */ |
fbf59bc9 | 161 | |
4f996e23 TH |
162 | /* chunks which need their map areas extended, protected by pcpu_lock */ |
163 | static LIST_HEAD(pcpu_map_extend_chunks); | |
164 | ||
b539b87f TH |
165 | /* |
166 | * The number of empty populated pages, protected by pcpu_lock. The | |
167 | * reserved chunk doesn't contribute to the count. | |
168 | */ | |
6b9b6f39 | 169 | int pcpu_nr_empty_pop_pages; |
b539b87f | 170 | |
1a4d7607 TH |
171 | /* |
172 | * Balance work is used to populate or destroy chunks asynchronously. We | |
173 | * try to keep the number of populated free pages between | |
174 | * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one | |
175 | * empty chunk. | |
176 | */ | |
fe6bd8c3 TH |
177 | static void pcpu_balance_workfn(struct work_struct *work); |
178 | static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); | |
1a4d7607 TH |
179 | static bool pcpu_async_enabled __read_mostly; |
180 | static bool pcpu_atomic_alloc_failed; | |
181 | ||
182 | static void pcpu_schedule_balance_work(void) | |
183 | { | |
184 | if (pcpu_async_enabled) | |
185 | schedule_work(&pcpu_balance_work); | |
186 | } | |
a56dbddf | 187 | |
c0ebfdc3 | 188 | /** |
560f2c23 DZF |
189 | * pcpu_addr_in_chunk - check if the address is served from this chunk |
190 | * @chunk: chunk of interest | |
191 | * @addr: percpu address | |
c0ebfdc3 DZF |
192 | * |
193 | * RETURNS: | |
560f2c23 | 194 | * True if the address is served from this chunk. |
c0ebfdc3 | 195 | */ |
560f2c23 | 196 | static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) |
020ec653 | 197 | { |
c0ebfdc3 DZF |
198 | void *start_addr, *end_addr; |
199 | ||
560f2c23 | 200 | if (!chunk) |
c0ebfdc3 | 201 | return false; |
020ec653 | 202 | |
560f2c23 DZF |
203 | start_addr = chunk->base_addr + chunk->start_offset; |
204 | end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - | |
205 | chunk->end_offset; | |
c0ebfdc3 DZF |
206 | |
207 | return addr >= start_addr && addr < end_addr; | |
020ec653 TH |
208 | } |
209 | ||
d9b55eeb | 210 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 211 | { |
cae3aeb8 | 212 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
213 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
214 | } | |
215 | ||
d9b55eeb TH |
216 | static int pcpu_size_to_slot(int size) |
217 | { | |
218 | if (size == pcpu_unit_size) | |
219 | return pcpu_nr_slots - 1; | |
220 | return __pcpu_size_to_slot(size); | |
221 | } | |
222 | ||
fbf59bc9 TH |
223 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
224 | { | |
40064aec | 225 | if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0) |
fbf59bc9 TH |
226 | return 0; |
227 | ||
40064aec | 228 | return pcpu_size_to_slot(chunk->free_bytes); |
fbf59bc9 TH |
229 | } |
230 | ||
88999a89 TH |
231 | /* set the pointer to a chunk in a page struct */ |
232 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
233 | { | |
234 | page->index = (unsigned long)pcpu; | |
235 | } | |
236 | ||
237 | /* obtain pointer to a chunk from a page struct */ | |
238 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
239 | { | |
240 | return (struct pcpu_chunk *)page->index; | |
241 | } | |
242 | ||
243 | static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) | |
fbf59bc9 | 244 | { |
2f39e637 | 245 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
246 | } |
247 | ||
c0ebfdc3 DZF |
248 | static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) |
249 | { | |
250 | return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); | |
251 | } | |
252 | ||
9983b6f0 TH |
253 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
254 | unsigned int cpu, int page_idx) | |
fbf59bc9 | 255 | { |
c0ebfdc3 DZF |
256 | return (unsigned long)chunk->base_addr + |
257 | pcpu_unit_page_offset(cpu, page_idx); | |
fbf59bc9 TH |
258 | } |
259 | ||
91e914c5 | 260 | static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end) |
ce3141a2 | 261 | { |
91e914c5 DZF |
262 | *rs = find_next_zero_bit(bitmap, end, *rs); |
263 | *re = find_next_bit(bitmap, end, *rs + 1); | |
ce3141a2 TH |
264 | } |
265 | ||
91e914c5 | 266 | static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end) |
ce3141a2 | 267 | { |
91e914c5 DZF |
268 | *rs = find_next_bit(bitmap, end, *rs); |
269 | *re = find_next_zero_bit(bitmap, end, *rs + 1); | |
ce3141a2 TH |
270 | } |
271 | ||
272 | /* | |
91e914c5 DZF |
273 | * Bitmap region iterators. Iterates over the bitmap between |
274 | * [@start, @end) in @chunk. @rs and @re should be integer variables | |
275 | * and will be set to start and end index of the current free region. | |
ce3141a2 | 276 | */ |
91e914c5 DZF |
277 | #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \ |
278 | for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \ | |
279 | (rs) < (re); \ | |
280 | (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end))) | |
ce3141a2 | 281 | |
91e914c5 DZF |
282 | #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \ |
283 | for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \ | |
284 | (rs) < (re); \ | |
285 | (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end))) | |
ce3141a2 | 286 | |
ca460b3c DZF |
287 | /* |
288 | * The following are helper functions to help access bitmaps and convert | |
289 | * between bitmap offsets to address offsets. | |
290 | */ | |
291 | static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index) | |
292 | { | |
293 | return chunk->alloc_map + | |
294 | (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG); | |
295 | } | |
296 | ||
297 | static unsigned long pcpu_off_to_block_index(int off) | |
298 | { | |
299 | return off / PCPU_BITMAP_BLOCK_BITS; | |
300 | } | |
301 | ||
302 | static unsigned long pcpu_off_to_block_off(int off) | |
303 | { | |
304 | return off & (PCPU_BITMAP_BLOCK_BITS - 1); | |
305 | } | |
306 | ||
b185cd0d DZF |
307 | static unsigned long pcpu_block_off_to_off(int index, int off) |
308 | { | |
309 | return index * PCPU_BITMAP_BLOCK_BITS + off; | |
310 | } | |
311 | ||
525ca84d DZF |
312 | /** |
313 | * pcpu_next_md_free_region - finds the next hint free area | |
314 | * @chunk: chunk of interest | |
315 | * @bit_off: chunk offset | |
316 | * @bits: size of free area | |
317 | * | |
318 | * Helper function for pcpu_for_each_md_free_region. It checks | |
319 | * block->contig_hint and performs aggregation across blocks to find the | |
320 | * next hint. It modifies bit_off and bits in-place to be consumed in the | |
321 | * loop. | |
322 | */ | |
323 | static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off, | |
324 | int *bits) | |
325 | { | |
326 | int i = pcpu_off_to_block_index(*bit_off); | |
327 | int block_off = pcpu_off_to_block_off(*bit_off); | |
328 | struct pcpu_block_md *block; | |
329 | ||
330 | *bits = 0; | |
331 | for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); | |
332 | block++, i++) { | |
333 | /* handles contig area across blocks */ | |
334 | if (*bits) { | |
335 | *bits += block->left_free; | |
336 | if (block->left_free == PCPU_BITMAP_BLOCK_BITS) | |
337 | continue; | |
338 | return; | |
339 | } | |
340 | ||
341 | /* | |
342 | * This checks three things. First is there a contig_hint to | |
343 | * check. Second, have we checked this hint before by | |
344 | * comparing the block_off. Third, is this the same as the | |
345 | * right contig hint. In the last case, it spills over into | |
346 | * the next block and should be handled by the contig area | |
347 | * across blocks code. | |
348 | */ | |
349 | *bits = block->contig_hint; | |
350 | if (*bits && block->contig_hint_start >= block_off && | |
351 | *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) { | |
352 | *bit_off = pcpu_block_off_to_off(i, | |
353 | block->contig_hint_start); | |
354 | return; | |
355 | } | |
1fa4df3e DZ |
356 | /* reset to satisfy the second predicate above */ |
357 | block_off = 0; | |
525ca84d DZF |
358 | |
359 | *bits = block->right_free; | |
360 | *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free; | |
361 | } | |
362 | } | |
363 | ||
b4c2116c DZF |
364 | /** |
365 | * pcpu_next_fit_region - finds fit areas for a given allocation request | |
366 | * @chunk: chunk of interest | |
367 | * @alloc_bits: size of allocation | |
368 | * @align: alignment of area (max PAGE_SIZE) | |
369 | * @bit_off: chunk offset | |
370 | * @bits: size of free area | |
371 | * | |
372 | * Finds the next free region that is viable for use with a given size and | |
373 | * alignment. This only returns if there is a valid area to be used for this | |
374 | * allocation. block->first_free is returned if the allocation request fits | |
375 | * within the block to see if the request can be fulfilled prior to the contig | |
376 | * hint. | |
377 | */ | |
378 | static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits, | |
379 | int align, int *bit_off, int *bits) | |
380 | { | |
381 | int i = pcpu_off_to_block_index(*bit_off); | |
382 | int block_off = pcpu_off_to_block_off(*bit_off); | |
383 | struct pcpu_block_md *block; | |
384 | ||
385 | *bits = 0; | |
386 | for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk); | |
387 | block++, i++) { | |
388 | /* handles contig area across blocks */ | |
389 | if (*bits) { | |
390 | *bits += block->left_free; | |
391 | if (*bits >= alloc_bits) | |
392 | return; | |
393 | if (block->left_free == PCPU_BITMAP_BLOCK_BITS) | |
394 | continue; | |
395 | } | |
396 | ||
397 | /* check block->contig_hint */ | |
398 | *bits = ALIGN(block->contig_hint_start, align) - | |
399 | block->contig_hint_start; | |
400 | /* | |
401 | * This uses the block offset to determine if this has been | |
402 | * checked in the prior iteration. | |
403 | */ | |
404 | if (block->contig_hint && | |
405 | block->contig_hint_start >= block_off && | |
406 | block->contig_hint >= *bits + alloc_bits) { | |
407 | *bits += alloc_bits + block->contig_hint_start - | |
408 | block->first_free; | |
409 | *bit_off = pcpu_block_off_to_off(i, block->first_free); | |
410 | return; | |
411 | } | |
1fa4df3e DZ |
412 | /* reset to satisfy the second predicate above */ |
413 | block_off = 0; | |
b4c2116c DZF |
414 | |
415 | *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free, | |
416 | align); | |
417 | *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off; | |
418 | *bit_off = pcpu_block_off_to_off(i, *bit_off); | |
419 | if (*bits >= alloc_bits) | |
420 | return; | |
421 | } | |
422 | ||
423 | /* no valid offsets were found - fail condition */ | |
424 | *bit_off = pcpu_chunk_map_bits(chunk); | |
425 | } | |
426 | ||
525ca84d DZF |
427 | /* |
428 | * Metadata free area iterators. These perform aggregation of free areas | |
429 | * based on the metadata blocks and return the offset @bit_off and size in | |
b4c2116c DZF |
430 | * bits of the free area @bits. pcpu_for_each_fit_region only returns when |
431 | * a fit is found for the allocation request. | |
525ca84d DZF |
432 | */ |
433 | #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \ | |
434 | for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \ | |
435 | (bit_off) < pcpu_chunk_map_bits((chunk)); \ | |
436 | (bit_off) += (bits) + 1, \ | |
437 | pcpu_next_md_free_region((chunk), &(bit_off), &(bits))) | |
438 | ||
b4c2116c DZF |
439 | #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \ |
440 | for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ | |
441 | &(bits)); \ | |
442 | (bit_off) < pcpu_chunk_map_bits((chunk)); \ | |
443 | (bit_off) += (bits), \ | |
444 | pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \ | |
445 | &(bits))) | |
446 | ||
fbf59bc9 | 447 | /** |
90459ce0 | 448 | * pcpu_mem_zalloc - allocate memory |
1880d93b | 449 | * @size: bytes to allocate |
47504ee0 | 450 | * @gfp: allocation flags |
fbf59bc9 | 451 | * |
1880d93b | 452 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
47504ee0 DZ |
453 | * kzalloc() is used; otherwise, the equivalent of vzalloc() is used. |
454 | * This is to facilitate passing through whitelisted flags. The | |
455 | * returned memory is always zeroed. | |
fbf59bc9 | 456 | * |
ccea34b5 TH |
457 | * CONTEXT: |
458 | * Does GFP_KERNEL allocation. | |
459 | * | |
fbf59bc9 | 460 | * RETURNS: |
1880d93b | 461 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 462 | */ |
47504ee0 | 463 | static void *pcpu_mem_zalloc(size_t size, gfp_t gfp) |
fbf59bc9 | 464 | { |
099a19d9 TH |
465 | if (WARN_ON_ONCE(!slab_is_available())) |
466 | return NULL; | |
467 | ||
1880d93b | 468 | if (size <= PAGE_SIZE) |
47504ee0 | 469 | return kzalloc(size, gfp | GFP_KERNEL); |
7af4c093 | 470 | else |
47504ee0 DZ |
471 | return __vmalloc(size, gfp | GFP_KERNEL | __GFP_ZERO, |
472 | PAGE_KERNEL); | |
1880d93b | 473 | } |
fbf59bc9 | 474 | |
1880d93b TH |
475 | /** |
476 | * pcpu_mem_free - free memory | |
477 | * @ptr: memory to free | |
1880d93b | 478 | * |
90459ce0 | 479 | * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). |
1880d93b | 480 | */ |
1d5cfdb0 | 481 | static void pcpu_mem_free(void *ptr) |
1880d93b | 482 | { |
1d5cfdb0 | 483 | kvfree(ptr); |
fbf59bc9 TH |
484 | } |
485 | ||
486 | /** | |
487 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
488 | * @chunk: chunk of interest | |
489 | * @oslot: the previous slot it was on | |
490 | * | |
491 | * This function is called after an allocation or free changed @chunk. | |
492 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
493 | * moved to the slot. Note that the reserved chunk is never put on |
494 | * chunk slots. | |
ccea34b5 TH |
495 | * |
496 | * CONTEXT: | |
497 | * pcpu_lock. | |
fbf59bc9 TH |
498 | */ |
499 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
500 | { | |
501 | int nslot = pcpu_chunk_slot(chunk); | |
502 | ||
edcb4639 | 503 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
504 | if (oslot < nslot) |
505 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
506 | else | |
507 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
508 | } | |
509 | } | |
510 | ||
9f7dcf22 | 511 | /** |
40064aec | 512 | * pcpu_cnt_pop_pages- counts populated backing pages in range |
833af842 | 513 | * @chunk: chunk of interest |
40064aec DZF |
514 | * @bit_off: start offset |
515 | * @bits: size of area to check | |
9f7dcf22 | 516 | * |
40064aec DZF |
517 | * Calculates the number of populated pages in the region |
518 | * [page_start, page_end). This keeps track of how many empty populated | |
519 | * pages are available and decide if async work should be scheduled. | |
ccea34b5 | 520 | * |
9f7dcf22 | 521 | * RETURNS: |
40064aec | 522 | * The nr of populated pages. |
9f7dcf22 | 523 | */ |
40064aec DZF |
524 | static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off, |
525 | int bits) | |
9f7dcf22 | 526 | { |
40064aec DZF |
527 | int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE); |
528 | int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); | |
4f996e23 | 529 | |
40064aec | 530 | if (page_start >= page_end) |
9f7dcf22 TH |
531 | return 0; |
532 | ||
40064aec DZF |
533 | /* |
534 | * bitmap_weight counts the number of bits set in a bitmap up to | |
535 | * the specified number of bits. This is counting the populated | |
536 | * pages up to page_end and then subtracting the populated pages | |
537 | * up to page_start to count the populated pages in | |
538 | * [page_start, page_end). | |
539 | */ | |
540 | return bitmap_weight(chunk->populated, page_end) - | |
541 | bitmap_weight(chunk->populated, page_start); | |
833af842 TH |
542 | } |
543 | ||
544 | /** | |
40064aec | 545 | * pcpu_chunk_update - updates the chunk metadata given a free area |
833af842 | 546 | * @chunk: chunk of interest |
40064aec DZF |
547 | * @bit_off: chunk offset |
548 | * @bits: size of free area | |
833af842 | 549 | * |
13f96637 | 550 | * This updates the chunk's contig hint and starting offset given a free area. |
268625a6 | 551 | * Choose the best starting offset if the contig hint is equal. |
40064aec DZF |
552 | */ |
553 | static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits) | |
554 | { | |
13f96637 DZF |
555 | if (bits > chunk->contig_bits) { |
556 | chunk->contig_bits_start = bit_off; | |
40064aec | 557 | chunk->contig_bits = bits; |
268625a6 DZF |
558 | } else if (bits == chunk->contig_bits && chunk->contig_bits_start && |
559 | (!bit_off || | |
560 | __ffs(bit_off) > __ffs(chunk->contig_bits_start))) { | |
561 | /* use the start with the best alignment */ | |
562 | chunk->contig_bits_start = bit_off; | |
13f96637 | 563 | } |
40064aec DZF |
564 | } |
565 | ||
566 | /** | |
567 | * pcpu_chunk_refresh_hint - updates metadata about a chunk | |
568 | * @chunk: chunk of interest | |
833af842 | 569 | * |
525ca84d DZF |
570 | * Iterates over the metadata blocks to find the largest contig area. |
571 | * It also counts the populated pages and uses the delta to update the | |
572 | * global count. | |
833af842 | 573 | * |
40064aec DZF |
574 | * Updates: |
575 | * chunk->contig_bits | |
13f96637 | 576 | * chunk->contig_bits_start |
525ca84d | 577 | * nr_empty_pop_pages (chunk and global) |
833af842 | 578 | */ |
40064aec | 579 | static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk) |
833af842 | 580 | { |
525ca84d | 581 | int bit_off, bits, nr_empty_pop_pages; |
833af842 | 582 | |
40064aec DZF |
583 | /* clear metadata */ |
584 | chunk->contig_bits = 0; | |
6710e594 | 585 | |
525ca84d | 586 | bit_off = chunk->first_bit; |
40064aec | 587 | bits = nr_empty_pop_pages = 0; |
525ca84d DZF |
588 | pcpu_for_each_md_free_region(chunk, bit_off, bits) { |
589 | pcpu_chunk_update(chunk, bit_off, bits); | |
833af842 | 590 | |
525ca84d | 591 | nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, bit_off, bits); |
40064aec | 592 | } |
9f7dcf22 | 593 | |
40064aec DZF |
594 | /* |
595 | * Keep track of nr_empty_pop_pages. | |
596 | * | |
597 | * The chunk maintains the previous number of free pages it held, | |
598 | * so the delta is used to update the global counter. The reserved | |
599 | * chunk is not part of the free page count as they are populated | |
600 | * at init and are special to serving reserved allocations. | |
601 | */ | |
602 | if (chunk != pcpu_reserved_chunk) | |
603 | pcpu_nr_empty_pop_pages += | |
604 | (nr_empty_pop_pages - chunk->nr_empty_pop_pages); | |
a002d148 | 605 | |
40064aec DZF |
606 | chunk->nr_empty_pop_pages = nr_empty_pop_pages; |
607 | } | |
9f7dcf22 | 608 | |
ca460b3c DZF |
609 | /** |
610 | * pcpu_block_update - updates a block given a free area | |
611 | * @block: block of interest | |
612 | * @start: start offset in block | |
613 | * @end: end offset in block | |
614 | * | |
615 | * Updates a block given a known free area. The region [start, end) is | |
268625a6 DZF |
616 | * expected to be the entirety of the free area within a block. Chooses |
617 | * the best starting offset if the contig hints are equal. | |
ca460b3c DZF |
618 | */ |
619 | static void pcpu_block_update(struct pcpu_block_md *block, int start, int end) | |
620 | { | |
621 | int contig = end - start; | |
622 | ||
623 | block->first_free = min(block->first_free, start); | |
624 | if (start == 0) | |
625 | block->left_free = contig; | |
626 | ||
627 | if (end == PCPU_BITMAP_BLOCK_BITS) | |
628 | block->right_free = contig; | |
629 | ||
630 | if (contig > block->contig_hint) { | |
631 | block->contig_hint_start = start; | |
632 | block->contig_hint = contig; | |
268625a6 DZF |
633 | } else if (block->contig_hint_start && contig == block->contig_hint && |
634 | (!start || __ffs(start) > __ffs(block->contig_hint_start))) { | |
635 | /* use the start with the best alignment */ | |
636 | block->contig_hint_start = start; | |
ca460b3c DZF |
637 | } |
638 | } | |
639 | ||
640 | /** | |
641 | * pcpu_block_refresh_hint | |
642 | * @chunk: chunk of interest | |
643 | * @index: index of the metadata block | |
644 | * | |
645 | * Scans over the block beginning at first_free and updates the block | |
646 | * metadata accordingly. | |
647 | */ | |
648 | static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index) | |
649 | { | |
650 | struct pcpu_block_md *block = chunk->md_blocks + index; | |
651 | unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index); | |
652 | int rs, re; /* region start, region end */ | |
653 | ||
654 | /* clear hints */ | |
655 | block->contig_hint = 0; | |
656 | block->left_free = block->right_free = 0; | |
657 | ||
658 | /* iterate over free areas and update the contig hints */ | |
659 | pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free, | |
660 | PCPU_BITMAP_BLOCK_BITS) { | |
661 | pcpu_block_update(block, rs, re); | |
662 | } | |
663 | } | |
664 | ||
665 | /** | |
666 | * pcpu_block_update_hint_alloc - update hint on allocation path | |
667 | * @chunk: chunk of interest | |
668 | * @bit_off: chunk offset | |
669 | * @bits: size of request | |
fc304334 DZF |
670 | * |
671 | * Updates metadata for the allocation path. The metadata only has to be | |
672 | * refreshed by a full scan iff the chunk's contig hint is broken. Block level | |
673 | * scans are required if the block's contig hint is broken. | |
ca460b3c DZF |
674 | */ |
675 | static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off, | |
676 | int bits) | |
677 | { | |
678 | struct pcpu_block_md *s_block, *e_block, *block; | |
679 | int s_index, e_index; /* block indexes of the freed allocation */ | |
680 | int s_off, e_off; /* block offsets of the freed allocation */ | |
681 | ||
682 | /* | |
683 | * Calculate per block offsets. | |
684 | * The calculation uses an inclusive range, but the resulting offsets | |
685 | * are [start, end). e_index always points to the last block in the | |
686 | * range. | |
687 | */ | |
688 | s_index = pcpu_off_to_block_index(bit_off); | |
689 | e_index = pcpu_off_to_block_index(bit_off + bits - 1); | |
690 | s_off = pcpu_off_to_block_off(bit_off); | |
691 | e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; | |
692 | ||
693 | s_block = chunk->md_blocks + s_index; | |
694 | e_block = chunk->md_blocks + e_index; | |
695 | ||
696 | /* | |
697 | * Update s_block. | |
fc304334 DZF |
698 | * block->first_free must be updated if the allocation takes its place. |
699 | * If the allocation breaks the contig_hint, a scan is required to | |
700 | * restore this hint. | |
ca460b3c | 701 | */ |
fc304334 DZF |
702 | if (s_off == s_block->first_free) |
703 | s_block->first_free = find_next_zero_bit( | |
704 | pcpu_index_alloc_map(chunk, s_index), | |
705 | PCPU_BITMAP_BLOCK_BITS, | |
706 | s_off + bits); | |
707 | ||
708 | if (s_off >= s_block->contig_hint_start && | |
709 | s_off < s_block->contig_hint_start + s_block->contig_hint) { | |
710 | /* block contig hint is broken - scan to fix it */ | |
711 | pcpu_block_refresh_hint(chunk, s_index); | |
712 | } else { | |
713 | /* update left and right contig manually */ | |
714 | s_block->left_free = min(s_block->left_free, s_off); | |
715 | if (s_index == e_index) | |
716 | s_block->right_free = min_t(int, s_block->right_free, | |
717 | PCPU_BITMAP_BLOCK_BITS - e_off); | |
718 | else | |
719 | s_block->right_free = 0; | |
720 | } | |
ca460b3c DZF |
721 | |
722 | /* | |
723 | * Update e_block. | |
724 | */ | |
725 | if (s_index != e_index) { | |
fc304334 DZF |
726 | /* |
727 | * When the allocation is across blocks, the end is along | |
728 | * the left part of the e_block. | |
729 | */ | |
730 | e_block->first_free = find_next_zero_bit( | |
731 | pcpu_index_alloc_map(chunk, e_index), | |
732 | PCPU_BITMAP_BLOCK_BITS, e_off); | |
733 | ||
734 | if (e_off == PCPU_BITMAP_BLOCK_BITS) { | |
735 | /* reset the block */ | |
736 | e_block++; | |
737 | } else { | |
738 | if (e_off > e_block->contig_hint_start) { | |
739 | /* contig hint is broken - scan to fix it */ | |
740 | pcpu_block_refresh_hint(chunk, e_index); | |
741 | } else { | |
742 | e_block->left_free = 0; | |
743 | e_block->right_free = | |
744 | min_t(int, e_block->right_free, | |
745 | PCPU_BITMAP_BLOCK_BITS - e_off); | |
746 | } | |
747 | } | |
ca460b3c DZF |
748 | |
749 | /* update in-between md_blocks */ | |
750 | for (block = s_block + 1; block < e_block; block++) { | |
751 | block->contig_hint = 0; | |
752 | block->left_free = 0; | |
753 | block->right_free = 0; | |
754 | } | |
755 | } | |
756 | ||
fc304334 DZF |
757 | /* |
758 | * The only time a full chunk scan is required is if the chunk | |
759 | * contig hint is broken. Otherwise, it means a smaller space | |
760 | * was used and therefore the chunk contig hint is still correct. | |
761 | */ | |
762 | if (bit_off >= chunk->contig_bits_start && | |
763 | bit_off < chunk->contig_bits_start + chunk->contig_bits) | |
764 | pcpu_chunk_refresh_hint(chunk); | |
ca460b3c DZF |
765 | } |
766 | ||
767 | /** | |
768 | * pcpu_block_update_hint_free - updates the block hints on the free path | |
769 | * @chunk: chunk of interest | |
770 | * @bit_off: chunk offset | |
771 | * @bits: size of request | |
b185cd0d DZF |
772 | * |
773 | * Updates metadata for the allocation path. This avoids a blind block | |
774 | * refresh by making use of the block contig hints. If this fails, it scans | |
775 | * forward and backward to determine the extent of the free area. This is | |
776 | * capped at the boundary of blocks. | |
777 | * | |
778 | * A chunk update is triggered if a page becomes free, a block becomes free, | |
779 | * or the free spans across blocks. This tradeoff is to minimize iterating | |
780 | * over the block metadata to update chunk->contig_bits. chunk->contig_bits | |
781 | * may be off by up to a page, but it will never be more than the available | |
782 | * space. If the contig hint is contained in one block, it will be accurate. | |
ca460b3c DZF |
783 | */ |
784 | static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off, | |
785 | int bits) | |
786 | { | |
787 | struct pcpu_block_md *s_block, *e_block, *block; | |
788 | int s_index, e_index; /* block indexes of the freed allocation */ | |
789 | int s_off, e_off; /* block offsets of the freed allocation */ | |
b185cd0d | 790 | int start, end; /* start and end of the whole free area */ |
ca460b3c DZF |
791 | |
792 | /* | |
793 | * Calculate per block offsets. | |
794 | * The calculation uses an inclusive range, but the resulting offsets | |
795 | * are [start, end). e_index always points to the last block in the | |
796 | * range. | |
797 | */ | |
798 | s_index = pcpu_off_to_block_index(bit_off); | |
799 | e_index = pcpu_off_to_block_index(bit_off + bits - 1); | |
800 | s_off = pcpu_off_to_block_off(bit_off); | |
801 | e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1; | |
802 | ||
803 | s_block = chunk->md_blocks + s_index; | |
804 | e_block = chunk->md_blocks + e_index; | |
805 | ||
b185cd0d DZF |
806 | /* |
807 | * Check if the freed area aligns with the block->contig_hint. | |
808 | * If it does, then the scan to find the beginning/end of the | |
809 | * larger free area can be avoided. | |
810 | * | |
811 | * start and end refer to beginning and end of the free area | |
812 | * within each their respective blocks. This is not necessarily | |
813 | * the entire free area as it may span blocks past the beginning | |
814 | * or end of the block. | |
815 | */ | |
816 | start = s_off; | |
817 | if (s_off == s_block->contig_hint + s_block->contig_hint_start) { | |
818 | start = s_block->contig_hint_start; | |
819 | } else { | |
820 | /* | |
821 | * Scan backwards to find the extent of the free area. | |
822 | * find_last_bit returns the starting bit, so if the start bit | |
823 | * is returned, that means there was no last bit and the | |
824 | * remainder of the chunk is free. | |
825 | */ | |
826 | int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), | |
827 | start); | |
828 | start = (start == l_bit) ? 0 : l_bit + 1; | |
829 | } | |
830 | ||
831 | end = e_off; | |
832 | if (e_off == e_block->contig_hint_start) | |
833 | end = e_block->contig_hint_start + e_block->contig_hint; | |
834 | else | |
835 | end = find_next_bit(pcpu_index_alloc_map(chunk, e_index), | |
836 | PCPU_BITMAP_BLOCK_BITS, end); | |
837 | ||
ca460b3c | 838 | /* update s_block */ |
b185cd0d DZF |
839 | e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS; |
840 | pcpu_block_update(s_block, start, e_off); | |
ca460b3c DZF |
841 | |
842 | /* freeing in the same block */ | |
843 | if (s_index != e_index) { | |
844 | /* update e_block */ | |
b185cd0d | 845 | pcpu_block_update(e_block, 0, end); |
ca460b3c DZF |
846 | |
847 | /* reset md_blocks in the middle */ | |
848 | for (block = s_block + 1; block < e_block; block++) { | |
849 | block->first_free = 0; | |
850 | block->contig_hint_start = 0; | |
851 | block->contig_hint = PCPU_BITMAP_BLOCK_BITS; | |
852 | block->left_free = PCPU_BITMAP_BLOCK_BITS; | |
853 | block->right_free = PCPU_BITMAP_BLOCK_BITS; | |
854 | } | |
855 | } | |
856 | ||
b185cd0d DZF |
857 | /* |
858 | * Refresh chunk metadata when the free makes a page free, a block | |
859 | * free, or spans across blocks. The contig hint may be off by up to | |
860 | * a page, but if the hint is contained in a block, it will be accurate | |
861 | * with the else condition below. | |
862 | */ | |
863 | if ((ALIGN_DOWN(end, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS)) > | |
864 | ALIGN(start, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS))) || | |
865 | s_index != e_index) | |
866 | pcpu_chunk_refresh_hint(chunk); | |
867 | else | |
868 | pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start), | |
869 | s_block->contig_hint); | |
ca460b3c DZF |
870 | } |
871 | ||
40064aec DZF |
872 | /** |
873 | * pcpu_is_populated - determines if the region is populated | |
874 | * @chunk: chunk of interest | |
875 | * @bit_off: chunk offset | |
876 | * @bits: size of area | |
877 | * @next_off: return value for the next offset to start searching | |
878 | * | |
879 | * For atomic allocations, check if the backing pages are populated. | |
880 | * | |
881 | * RETURNS: | |
882 | * Bool if the backing pages are populated. | |
883 | * next_index is to skip over unpopulated blocks in pcpu_find_block_fit. | |
884 | */ | |
885 | static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits, | |
886 | int *next_off) | |
887 | { | |
888 | int page_start, page_end, rs, re; | |
833af842 | 889 | |
40064aec DZF |
890 | page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE); |
891 | page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE); | |
833af842 | 892 | |
40064aec DZF |
893 | rs = page_start; |
894 | pcpu_next_unpop(chunk->populated, &rs, &re, page_end); | |
895 | if (rs >= page_end) | |
896 | return true; | |
833af842 | 897 | |
40064aec DZF |
898 | *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE; |
899 | return false; | |
9f7dcf22 TH |
900 | } |
901 | ||
a16037c8 | 902 | /** |
40064aec DZF |
903 | * pcpu_find_block_fit - finds the block index to start searching |
904 | * @chunk: chunk of interest | |
905 | * @alloc_bits: size of request in allocation units | |
906 | * @align: alignment of area (max PAGE_SIZE bytes) | |
907 | * @pop_only: use populated regions only | |
908 | * | |
b4c2116c DZF |
909 | * Given a chunk and an allocation spec, find the offset to begin searching |
910 | * for a free region. This iterates over the bitmap metadata blocks to | |
911 | * find an offset that will be guaranteed to fit the requirements. It is | |
912 | * not quite first fit as if the allocation does not fit in the contig hint | |
913 | * of a block or chunk, it is skipped. This errs on the side of caution | |
914 | * to prevent excess iteration. Poor alignment can cause the allocator to | |
915 | * skip over blocks and chunks that have valid free areas. | |
916 | * | |
40064aec DZF |
917 | * RETURNS: |
918 | * The offset in the bitmap to begin searching. | |
919 | * -1 if no offset is found. | |
a16037c8 | 920 | */ |
40064aec DZF |
921 | static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits, |
922 | size_t align, bool pop_only) | |
a16037c8 | 923 | { |
b4c2116c | 924 | int bit_off, bits, next_off; |
a16037c8 | 925 | |
13f96637 DZF |
926 | /* |
927 | * Check to see if the allocation can fit in the chunk's contig hint. | |
928 | * This is an optimization to prevent scanning by assuming if it | |
929 | * cannot fit in the global hint, there is memory pressure and creating | |
930 | * a new chunk would happen soon. | |
931 | */ | |
932 | bit_off = ALIGN(chunk->contig_bits_start, align) - | |
933 | chunk->contig_bits_start; | |
934 | if (bit_off + alloc_bits > chunk->contig_bits) | |
935 | return -1; | |
936 | ||
b4c2116c DZF |
937 | bit_off = chunk->first_bit; |
938 | bits = 0; | |
939 | pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) { | |
40064aec | 940 | if (!pop_only || pcpu_is_populated(chunk, bit_off, bits, |
b4c2116c | 941 | &next_off)) |
40064aec | 942 | break; |
a16037c8 | 943 | |
b4c2116c | 944 | bit_off = next_off; |
40064aec | 945 | bits = 0; |
a16037c8 | 946 | } |
40064aec DZF |
947 | |
948 | if (bit_off == pcpu_chunk_map_bits(chunk)) | |
949 | return -1; | |
950 | ||
951 | return bit_off; | |
a16037c8 TH |
952 | } |
953 | ||
fbf59bc9 | 954 | /** |
40064aec | 955 | * pcpu_alloc_area - allocates an area from a pcpu_chunk |
fbf59bc9 | 956 | * @chunk: chunk of interest |
40064aec DZF |
957 | * @alloc_bits: size of request in allocation units |
958 | * @align: alignment of area (max PAGE_SIZE) | |
959 | * @start: bit_off to start searching | |
9f7dcf22 | 960 | * |
40064aec | 961 | * This function takes in a @start offset to begin searching to fit an |
b4c2116c DZF |
962 | * allocation of @alloc_bits with alignment @align. It needs to scan |
963 | * the allocation map because if it fits within the block's contig hint, | |
964 | * @start will be block->first_free. This is an attempt to fill the | |
965 | * allocation prior to breaking the contig hint. The allocation and | |
966 | * boundary maps are updated accordingly if it confirms a valid | |
967 | * free area. | |
ccea34b5 | 968 | * |
fbf59bc9 | 969 | * RETURNS: |
40064aec DZF |
970 | * Allocated addr offset in @chunk on success. |
971 | * -1 if no matching area is found. | |
fbf59bc9 | 972 | */ |
40064aec DZF |
973 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits, |
974 | size_t align, int start) | |
fbf59bc9 | 975 | { |
40064aec DZF |
976 | size_t align_mask = (align) ? (align - 1) : 0; |
977 | int bit_off, end, oslot; | |
a16037c8 | 978 | |
40064aec | 979 | lockdep_assert_held(&pcpu_lock); |
fbf59bc9 | 980 | |
40064aec | 981 | oslot = pcpu_chunk_slot(chunk); |
fbf59bc9 | 982 | |
40064aec DZF |
983 | /* |
984 | * Search to find a fit. | |
985 | */ | |
b4c2116c | 986 | end = start + alloc_bits + PCPU_BITMAP_BLOCK_BITS; |
40064aec DZF |
987 | bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start, |
988 | alloc_bits, align_mask); | |
989 | if (bit_off >= end) | |
990 | return -1; | |
fbf59bc9 | 991 | |
40064aec DZF |
992 | /* update alloc map */ |
993 | bitmap_set(chunk->alloc_map, bit_off, alloc_bits); | |
3d331ad7 | 994 | |
40064aec DZF |
995 | /* update boundary map */ |
996 | set_bit(bit_off, chunk->bound_map); | |
997 | bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1); | |
998 | set_bit(bit_off + alloc_bits, chunk->bound_map); | |
fbf59bc9 | 999 | |
40064aec | 1000 | chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE; |
fbf59bc9 | 1001 | |
86b442fb DZF |
1002 | /* update first free bit */ |
1003 | if (bit_off == chunk->first_bit) | |
1004 | chunk->first_bit = find_next_zero_bit( | |
1005 | chunk->alloc_map, | |
1006 | pcpu_chunk_map_bits(chunk), | |
1007 | bit_off + alloc_bits); | |
1008 | ||
ca460b3c | 1009 | pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits); |
fbf59bc9 | 1010 | |
fbf59bc9 TH |
1011 | pcpu_chunk_relocate(chunk, oslot); |
1012 | ||
40064aec | 1013 | return bit_off * PCPU_MIN_ALLOC_SIZE; |
fbf59bc9 TH |
1014 | } |
1015 | ||
1016 | /** | |
40064aec | 1017 | * pcpu_free_area - frees the corresponding offset |
fbf59bc9 | 1018 | * @chunk: chunk of interest |
40064aec | 1019 | * @off: addr offset into chunk |
ccea34b5 | 1020 | * |
40064aec DZF |
1021 | * This function determines the size of an allocation to free using |
1022 | * the boundary bitmap and clears the allocation map. | |
fbf59bc9 | 1023 | */ |
40064aec | 1024 | static void pcpu_free_area(struct pcpu_chunk *chunk, int off) |
fbf59bc9 | 1025 | { |
40064aec | 1026 | int bit_off, bits, end, oslot; |
723ad1d9 | 1027 | |
5ccd30e4 | 1028 | lockdep_assert_held(&pcpu_lock); |
30a5b536 | 1029 | pcpu_stats_area_dealloc(chunk); |
5ccd30e4 | 1030 | |
40064aec | 1031 | oslot = pcpu_chunk_slot(chunk); |
fbf59bc9 | 1032 | |
40064aec | 1033 | bit_off = off / PCPU_MIN_ALLOC_SIZE; |
3d331ad7 | 1034 | |
40064aec DZF |
1035 | /* find end index */ |
1036 | end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk), | |
1037 | bit_off + 1); | |
1038 | bits = end - bit_off; | |
1039 | bitmap_clear(chunk->alloc_map, bit_off, bits); | |
fbf59bc9 | 1040 | |
40064aec DZF |
1041 | /* update metadata */ |
1042 | chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE; | |
b539b87f | 1043 | |
86b442fb DZF |
1044 | /* update first free bit */ |
1045 | chunk->first_bit = min(chunk->first_bit, bit_off); | |
1046 | ||
ca460b3c | 1047 | pcpu_block_update_hint_free(chunk, bit_off, bits); |
fbf59bc9 | 1048 | |
fbf59bc9 TH |
1049 | pcpu_chunk_relocate(chunk, oslot); |
1050 | } | |
1051 | ||
ca460b3c DZF |
1052 | static void pcpu_init_md_blocks(struct pcpu_chunk *chunk) |
1053 | { | |
1054 | struct pcpu_block_md *md_block; | |
1055 | ||
1056 | for (md_block = chunk->md_blocks; | |
1057 | md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk); | |
1058 | md_block++) { | |
1059 | md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS; | |
1060 | md_block->left_free = PCPU_BITMAP_BLOCK_BITS; | |
1061 | md_block->right_free = PCPU_BITMAP_BLOCK_BITS; | |
1062 | } | |
1063 | } | |
1064 | ||
40064aec DZF |
1065 | /** |
1066 | * pcpu_alloc_first_chunk - creates chunks that serve the first chunk | |
1067 | * @tmp_addr: the start of the region served | |
1068 | * @map_size: size of the region served | |
1069 | * | |
1070 | * This is responsible for creating the chunks that serve the first chunk. The | |
1071 | * base_addr is page aligned down of @tmp_addr while the region end is page | |
1072 | * aligned up. Offsets are kept track of to determine the region served. All | |
1073 | * this is done to appease the bitmap allocator in avoiding partial blocks. | |
1074 | * | |
1075 | * RETURNS: | |
1076 | * Chunk serving the region at @tmp_addr of @map_size. | |
1077 | */ | |
c0ebfdc3 | 1078 | static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, |
40064aec | 1079 | int map_size) |
10edf5b0 DZF |
1080 | { |
1081 | struct pcpu_chunk *chunk; | |
ca460b3c | 1082 | unsigned long aligned_addr, lcm_align; |
40064aec | 1083 | int start_offset, offset_bits, region_size, region_bits; |
c0ebfdc3 DZF |
1084 | |
1085 | /* region calculations */ | |
1086 | aligned_addr = tmp_addr & PAGE_MASK; | |
1087 | ||
1088 | start_offset = tmp_addr - aligned_addr; | |
6b9d7c8e | 1089 | |
ca460b3c DZF |
1090 | /* |
1091 | * Align the end of the region with the LCM of PAGE_SIZE and | |
1092 | * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of | |
1093 | * the other. | |
1094 | */ | |
1095 | lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE); | |
1096 | region_size = ALIGN(start_offset + map_size, lcm_align); | |
10edf5b0 | 1097 | |
c0ebfdc3 | 1098 | /* allocate chunk */ |
8ab16c43 DZF |
1099 | chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) + |
1100 | BITS_TO_LONGS(region_size >> PAGE_SHIFT), | |
1101 | 0); | |
c0ebfdc3 | 1102 | |
10edf5b0 | 1103 | INIT_LIST_HEAD(&chunk->list); |
c0ebfdc3 DZF |
1104 | |
1105 | chunk->base_addr = (void *)aligned_addr; | |
10edf5b0 | 1106 | chunk->start_offset = start_offset; |
6b9d7c8e | 1107 | chunk->end_offset = region_size - chunk->start_offset - map_size; |
c0ebfdc3 | 1108 | |
8ab16c43 | 1109 | chunk->nr_pages = region_size >> PAGE_SHIFT; |
40064aec | 1110 | region_bits = pcpu_chunk_map_bits(chunk); |
c0ebfdc3 | 1111 | |
ca460b3c DZF |
1112 | chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) * |
1113 | sizeof(chunk->alloc_map[0]), 0); | |
1114 | chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) * | |
1115 | sizeof(chunk->bound_map[0]), 0); | |
1116 | chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) * | |
1117 | sizeof(chunk->md_blocks[0]), 0); | |
1118 | pcpu_init_md_blocks(chunk); | |
10edf5b0 DZF |
1119 | |
1120 | /* manage populated page bitmap */ | |
1121 | chunk->immutable = true; | |
8ab16c43 DZF |
1122 | bitmap_fill(chunk->populated, chunk->nr_pages); |
1123 | chunk->nr_populated = chunk->nr_pages; | |
40064aec DZF |
1124 | chunk->nr_empty_pop_pages = |
1125 | pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE, | |
1126 | map_size / PCPU_MIN_ALLOC_SIZE); | |
10edf5b0 | 1127 | |
40064aec DZF |
1128 | chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE; |
1129 | chunk->free_bytes = map_size; | |
c0ebfdc3 DZF |
1130 | |
1131 | if (chunk->start_offset) { | |
1132 | /* hide the beginning of the bitmap */ | |
40064aec DZF |
1133 | offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE; |
1134 | bitmap_set(chunk->alloc_map, 0, offset_bits); | |
1135 | set_bit(0, chunk->bound_map); | |
1136 | set_bit(offset_bits, chunk->bound_map); | |
ca460b3c | 1137 | |
86b442fb DZF |
1138 | chunk->first_bit = offset_bits; |
1139 | ||
ca460b3c | 1140 | pcpu_block_update_hint_alloc(chunk, 0, offset_bits); |
c0ebfdc3 DZF |
1141 | } |
1142 | ||
6b9d7c8e DZF |
1143 | if (chunk->end_offset) { |
1144 | /* hide the end of the bitmap */ | |
40064aec DZF |
1145 | offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE; |
1146 | bitmap_set(chunk->alloc_map, | |
1147 | pcpu_chunk_map_bits(chunk) - offset_bits, | |
1148 | offset_bits); | |
1149 | set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE, | |
1150 | chunk->bound_map); | |
1151 | set_bit(region_bits, chunk->bound_map); | |
6b9d7c8e | 1152 | |
ca460b3c DZF |
1153 | pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk) |
1154 | - offset_bits, offset_bits); | |
1155 | } | |
40064aec | 1156 | |
10edf5b0 DZF |
1157 | return chunk; |
1158 | } | |
1159 | ||
47504ee0 | 1160 | static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp) |
6081089f TH |
1161 | { |
1162 | struct pcpu_chunk *chunk; | |
40064aec | 1163 | int region_bits; |
6081089f | 1164 | |
47504ee0 | 1165 | chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp); |
6081089f TH |
1166 | if (!chunk) |
1167 | return NULL; | |
1168 | ||
40064aec DZF |
1169 | INIT_LIST_HEAD(&chunk->list); |
1170 | chunk->nr_pages = pcpu_unit_pages; | |
1171 | region_bits = pcpu_chunk_map_bits(chunk); | |
6081089f | 1172 | |
40064aec | 1173 | chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) * |
47504ee0 | 1174 | sizeof(chunk->alloc_map[0]), gfp); |
40064aec DZF |
1175 | if (!chunk->alloc_map) |
1176 | goto alloc_map_fail; | |
6081089f | 1177 | |
40064aec | 1178 | chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) * |
47504ee0 | 1179 | sizeof(chunk->bound_map[0]), gfp); |
40064aec DZF |
1180 | if (!chunk->bound_map) |
1181 | goto bound_map_fail; | |
6081089f | 1182 | |
ca460b3c | 1183 | chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) * |
47504ee0 | 1184 | sizeof(chunk->md_blocks[0]), gfp); |
ca460b3c DZF |
1185 | if (!chunk->md_blocks) |
1186 | goto md_blocks_fail; | |
1187 | ||
1188 | pcpu_init_md_blocks(chunk); | |
1189 | ||
40064aec DZF |
1190 | /* init metadata */ |
1191 | chunk->contig_bits = region_bits; | |
1192 | chunk->free_bytes = chunk->nr_pages * PAGE_SIZE; | |
c0ebfdc3 | 1193 | |
6081089f | 1194 | return chunk; |
40064aec | 1195 | |
ca460b3c DZF |
1196 | md_blocks_fail: |
1197 | pcpu_mem_free(chunk->bound_map); | |
40064aec DZF |
1198 | bound_map_fail: |
1199 | pcpu_mem_free(chunk->alloc_map); | |
1200 | alloc_map_fail: | |
1201 | pcpu_mem_free(chunk); | |
1202 | ||
1203 | return NULL; | |
6081089f TH |
1204 | } |
1205 | ||
1206 | static void pcpu_free_chunk(struct pcpu_chunk *chunk) | |
1207 | { | |
1208 | if (!chunk) | |
1209 | return; | |
40064aec DZF |
1210 | pcpu_mem_free(chunk->bound_map); |
1211 | pcpu_mem_free(chunk->alloc_map); | |
1d5cfdb0 | 1212 | pcpu_mem_free(chunk); |
6081089f TH |
1213 | } |
1214 | ||
b539b87f TH |
1215 | /** |
1216 | * pcpu_chunk_populated - post-population bookkeeping | |
1217 | * @chunk: pcpu_chunk which got populated | |
1218 | * @page_start: the start page | |
1219 | * @page_end: the end page | |
40064aec | 1220 | * @for_alloc: if this is to populate for allocation |
b539b87f TH |
1221 | * |
1222 | * Pages in [@page_start,@page_end) have been populated to @chunk. Update | |
1223 | * the bookkeeping information accordingly. Must be called after each | |
1224 | * successful population. | |
40064aec DZF |
1225 | * |
1226 | * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it | |
1227 | * is to serve an allocation in that area. | |
b539b87f | 1228 | */ |
40064aec DZF |
1229 | static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start, |
1230 | int page_end, bool for_alloc) | |
b539b87f TH |
1231 | { |
1232 | int nr = page_end - page_start; | |
1233 | ||
1234 | lockdep_assert_held(&pcpu_lock); | |
1235 | ||
1236 | bitmap_set(chunk->populated, page_start, nr); | |
1237 | chunk->nr_populated += nr; | |
40064aec DZF |
1238 | |
1239 | if (!for_alloc) { | |
1240 | chunk->nr_empty_pop_pages += nr; | |
1241 | pcpu_nr_empty_pop_pages += nr; | |
1242 | } | |
b539b87f TH |
1243 | } |
1244 | ||
1245 | /** | |
1246 | * pcpu_chunk_depopulated - post-depopulation bookkeeping | |
1247 | * @chunk: pcpu_chunk which got depopulated | |
1248 | * @page_start: the start page | |
1249 | * @page_end: the end page | |
1250 | * | |
1251 | * Pages in [@page_start,@page_end) have been depopulated from @chunk. | |
1252 | * Update the bookkeeping information accordingly. Must be called after | |
1253 | * each successful depopulation. | |
1254 | */ | |
1255 | static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, | |
1256 | int page_start, int page_end) | |
1257 | { | |
1258 | int nr = page_end - page_start; | |
1259 | ||
1260 | lockdep_assert_held(&pcpu_lock); | |
1261 | ||
1262 | bitmap_clear(chunk->populated, page_start, nr); | |
1263 | chunk->nr_populated -= nr; | |
0cecf50c | 1264 | chunk->nr_empty_pop_pages -= nr; |
b539b87f TH |
1265 | pcpu_nr_empty_pop_pages -= nr; |
1266 | } | |
1267 | ||
9f645532 TH |
1268 | /* |
1269 | * Chunk management implementation. | |
1270 | * | |
1271 | * To allow different implementations, chunk alloc/free and | |
1272 | * [de]population are implemented in a separate file which is pulled | |
1273 | * into this file and compiled together. The following functions | |
1274 | * should be implemented. | |
1275 | * | |
1276 | * pcpu_populate_chunk - populate the specified range of a chunk | |
1277 | * pcpu_depopulate_chunk - depopulate the specified range of a chunk | |
1278 | * pcpu_create_chunk - create a new chunk | |
1279 | * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop | |
1280 | * pcpu_addr_to_page - translate address to physical address | |
1281 | * pcpu_verify_alloc_info - check alloc_info is acceptable during init | |
fbf59bc9 | 1282 | */ |
15d9f3d1 | 1283 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, |
47504ee0 | 1284 | int page_start, int page_end, gfp_t gfp); |
15d9f3d1 DZ |
1285 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, |
1286 | int page_start, int page_end); | |
47504ee0 | 1287 | static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp); |
9f645532 TH |
1288 | static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); |
1289 | static struct page *pcpu_addr_to_page(void *addr); | |
1290 | static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); | |
fbf59bc9 | 1291 | |
b0c9778b TH |
1292 | #ifdef CONFIG_NEED_PER_CPU_KM |
1293 | #include "percpu-km.c" | |
1294 | #else | |
9f645532 | 1295 | #include "percpu-vm.c" |
b0c9778b | 1296 | #endif |
fbf59bc9 | 1297 | |
88999a89 TH |
1298 | /** |
1299 | * pcpu_chunk_addr_search - determine chunk containing specified address | |
1300 | * @addr: address for which the chunk needs to be determined. | |
1301 | * | |
c0ebfdc3 DZF |
1302 | * This is an internal function that handles all but static allocations. |
1303 | * Static percpu address values should never be passed into the allocator. | |
1304 | * | |
88999a89 TH |
1305 | * RETURNS: |
1306 | * The address of the found chunk. | |
1307 | */ | |
1308 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
1309 | { | |
c0ebfdc3 | 1310 | /* is it in the dynamic region (first chunk)? */ |
560f2c23 | 1311 | if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) |
88999a89 | 1312 | return pcpu_first_chunk; |
c0ebfdc3 DZF |
1313 | |
1314 | /* is it in the reserved region? */ | |
560f2c23 | 1315 | if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) |
c0ebfdc3 | 1316 | return pcpu_reserved_chunk; |
88999a89 TH |
1317 | |
1318 | /* | |
1319 | * The address is relative to unit0 which might be unused and | |
1320 | * thus unmapped. Offset the address to the unit space of the | |
1321 | * current processor before looking it up in the vmalloc | |
1322 | * space. Note that any possible cpu id can be used here, so | |
1323 | * there's no need to worry about preemption or cpu hotplug. | |
1324 | */ | |
1325 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; | |
9f645532 | 1326 | return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); |
88999a89 TH |
1327 | } |
1328 | ||
fbf59bc9 | 1329 | /** |
edcb4639 | 1330 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 1331 | * @size: size of area to allocate in bytes |
fbf59bc9 | 1332 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 1333 | * @reserved: allocate from the reserved chunk if available |
5835d96e | 1334 | * @gfp: allocation flags |
fbf59bc9 | 1335 | * |
5835d96e | 1336 | * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't |
0ea7eeec DB |
1337 | * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN |
1338 | * then no warning will be triggered on invalid or failed allocation | |
1339 | * requests. | |
fbf59bc9 TH |
1340 | * |
1341 | * RETURNS: | |
1342 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1343 | */ | |
5835d96e TH |
1344 | static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, |
1345 | gfp_t gfp) | |
fbf59bc9 | 1346 | { |
0ea7eeec DB |
1347 | bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; |
1348 | bool do_warn = !(gfp & __GFP_NOWARN); | |
f2badb0c | 1349 | static int warn_limit = 10; |
fbf59bc9 | 1350 | struct pcpu_chunk *chunk; |
f2badb0c | 1351 | const char *err; |
40064aec | 1352 | int slot, off, cpu, ret; |
403a91b1 | 1353 | unsigned long flags; |
f528f0b8 | 1354 | void __percpu *ptr; |
40064aec | 1355 | size_t bits, bit_align; |
fbf59bc9 | 1356 | |
723ad1d9 | 1357 | /* |
40064aec DZF |
1358 | * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE, |
1359 | * therefore alignment must be a minimum of that many bytes. | |
1360 | * An allocation may have internal fragmentation from rounding up | |
1361 | * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes. | |
723ad1d9 | 1362 | */ |
d2f3c384 DZF |
1363 | if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) |
1364 | align = PCPU_MIN_ALLOC_SIZE; | |
723ad1d9 | 1365 | |
d2f3c384 | 1366 | size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); |
40064aec DZF |
1367 | bits = size >> PCPU_MIN_ALLOC_SHIFT; |
1368 | bit_align = align >> PCPU_MIN_ALLOC_SHIFT; | |
2f69fa82 | 1369 | |
3ca45a46 | 1370 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || |
1371 | !is_power_of_2(align))) { | |
0ea7eeec | 1372 | WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n", |
756a025f | 1373 | size, align); |
fbf59bc9 TH |
1374 | return NULL; |
1375 | } | |
1376 | ||
6710e594 TH |
1377 | if (!is_atomic) |
1378 | mutex_lock(&pcpu_alloc_mutex); | |
1379 | ||
403a91b1 | 1380 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 | 1381 | |
edcb4639 TH |
1382 | /* serve reserved allocations from the reserved chunk if available */ |
1383 | if (reserved && pcpu_reserved_chunk) { | |
1384 | chunk = pcpu_reserved_chunk; | |
833af842 | 1385 | |
40064aec DZF |
1386 | off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic); |
1387 | if (off < 0) { | |
833af842 | 1388 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 1389 | goto fail_unlock; |
f2badb0c | 1390 | } |
833af842 | 1391 | |
40064aec | 1392 | off = pcpu_alloc_area(chunk, bits, bit_align, off); |
edcb4639 TH |
1393 | if (off >= 0) |
1394 | goto area_found; | |
833af842 | 1395 | |
f2badb0c | 1396 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 1397 | goto fail_unlock; |
edcb4639 TH |
1398 | } |
1399 | ||
ccea34b5 | 1400 | restart: |
edcb4639 | 1401 | /* search through normal chunks */ |
fbf59bc9 TH |
1402 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
1403 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
40064aec DZF |
1404 | off = pcpu_find_block_fit(chunk, bits, bit_align, |
1405 | is_atomic); | |
1406 | if (off < 0) | |
fbf59bc9 | 1407 | continue; |
ccea34b5 | 1408 | |
40064aec | 1409 | off = pcpu_alloc_area(chunk, bits, bit_align, off); |
fbf59bc9 TH |
1410 | if (off >= 0) |
1411 | goto area_found; | |
40064aec | 1412 | |
fbf59bc9 TH |
1413 | } |
1414 | } | |
1415 | ||
403a91b1 | 1416 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1417 | |
b38d08f3 TH |
1418 | /* |
1419 | * No space left. Create a new chunk. We don't want multiple | |
1420 | * tasks to create chunks simultaneously. Serialize and create iff | |
1421 | * there's still no empty chunk after grabbing the mutex. | |
1422 | */ | |
11df02bf DZ |
1423 | if (is_atomic) { |
1424 | err = "atomic alloc failed, no space left"; | |
5835d96e | 1425 | goto fail; |
11df02bf | 1426 | } |
5835d96e | 1427 | |
b38d08f3 | 1428 | if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { |
47504ee0 | 1429 | chunk = pcpu_create_chunk(0); |
b38d08f3 TH |
1430 | if (!chunk) { |
1431 | err = "failed to allocate new chunk"; | |
1432 | goto fail; | |
1433 | } | |
1434 | ||
1435 | spin_lock_irqsave(&pcpu_lock, flags); | |
1436 | pcpu_chunk_relocate(chunk, -1); | |
1437 | } else { | |
1438 | spin_lock_irqsave(&pcpu_lock, flags); | |
f2badb0c | 1439 | } |
ccea34b5 | 1440 | |
ccea34b5 | 1441 | goto restart; |
fbf59bc9 TH |
1442 | |
1443 | area_found: | |
30a5b536 | 1444 | pcpu_stats_area_alloc(chunk, size); |
403a91b1 | 1445 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1446 | |
dca49645 | 1447 | /* populate if not all pages are already there */ |
5835d96e | 1448 | if (!is_atomic) { |
e04d3208 | 1449 | int page_start, page_end, rs, re; |
dca49645 | 1450 | |
e04d3208 TH |
1451 | page_start = PFN_DOWN(off); |
1452 | page_end = PFN_UP(off + size); | |
b38d08f3 | 1453 | |
91e914c5 DZF |
1454 | pcpu_for_each_unpop_region(chunk->populated, rs, re, |
1455 | page_start, page_end) { | |
e04d3208 TH |
1456 | WARN_ON(chunk->immutable); |
1457 | ||
47504ee0 | 1458 | ret = pcpu_populate_chunk(chunk, rs, re, 0); |
e04d3208 TH |
1459 | |
1460 | spin_lock_irqsave(&pcpu_lock, flags); | |
1461 | if (ret) { | |
40064aec | 1462 | pcpu_free_area(chunk, off); |
e04d3208 TH |
1463 | err = "failed to populate"; |
1464 | goto fail_unlock; | |
1465 | } | |
40064aec | 1466 | pcpu_chunk_populated(chunk, rs, re, true); |
e04d3208 | 1467 | spin_unlock_irqrestore(&pcpu_lock, flags); |
dca49645 | 1468 | } |
fbf59bc9 | 1469 | |
e04d3208 TH |
1470 | mutex_unlock(&pcpu_alloc_mutex); |
1471 | } | |
ccea34b5 | 1472 | |
1a4d7607 TH |
1473 | if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) |
1474 | pcpu_schedule_balance_work(); | |
1475 | ||
dca49645 TH |
1476 | /* clear the areas and return address relative to base address */ |
1477 | for_each_possible_cpu(cpu) | |
1478 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); | |
1479 | ||
f528f0b8 | 1480 | ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); |
8a8c35fa | 1481 | kmemleak_alloc_percpu(ptr, size, gfp); |
df95e795 DZ |
1482 | |
1483 | trace_percpu_alloc_percpu(reserved, is_atomic, size, align, | |
1484 | chunk->base_addr, off, ptr); | |
1485 | ||
f528f0b8 | 1486 | return ptr; |
ccea34b5 TH |
1487 | |
1488 | fail_unlock: | |
403a91b1 | 1489 | spin_unlock_irqrestore(&pcpu_lock, flags); |
b38d08f3 | 1490 | fail: |
df95e795 DZ |
1491 | trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); |
1492 | ||
0ea7eeec | 1493 | if (!is_atomic && do_warn && warn_limit) { |
870d4b12 | 1494 | pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", |
598d8091 | 1495 | size, align, is_atomic, err); |
f2badb0c TH |
1496 | dump_stack(); |
1497 | if (!--warn_limit) | |
870d4b12 | 1498 | pr_info("limit reached, disable warning\n"); |
f2badb0c | 1499 | } |
1a4d7607 TH |
1500 | if (is_atomic) { |
1501 | /* see the flag handling in pcpu_blance_workfn() */ | |
1502 | pcpu_atomic_alloc_failed = true; | |
1503 | pcpu_schedule_balance_work(); | |
6710e594 TH |
1504 | } else { |
1505 | mutex_unlock(&pcpu_alloc_mutex); | |
1a4d7607 | 1506 | } |
ccea34b5 | 1507 | return NULL; |
fbf59bc9 | 1508 | } |
edcb4639 TH |
1509 | |
1510 | /** | |
5835d96e | 1511 | * __alloc_percpu_gfp - allocate dynamic percpu area |
edcb4639 TH |
1512 | * @size: size of area to allocate in bytes |
1513 | * @align: alignment of area (max PAGE_SIZE) | |
5835d96e | 1514 | * @gfp: allocation flags |
edcb4639 | 1515 | * |
5835d96e TH |
1516 | * Allocate zero-filled percpu area of @size bytes aligned at @align. If |
1517 | * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can | |
0ea7eeec DB |
1518 | * be called from any context but is a lot more likely to fail. If @gfp |
1519 | * has __GFP_NOWARN then no warning will be triggered on invalid or failed | |
1520 | * allocation requests. | |
ccea34b5 | 1521 | * |
edcb4639 TH |
1522 | * RETURNS: |
1523 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1524 | */ | |
5835d96e TH |
1525 | void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) |
1526 | { | |
1527 | return pcpu_alloc(size, align, false, gfp); | |
1528 | } | |
1529 | EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); | |
1530 | ||
1531 | /** | |
1532 | * __alloc_percpu - allocate dynamic percpu area | |
1533 | * @size: size of area to allocate in bytes | |
1534 | * @align: alignment of area (max PAGE_SIZE) | |
1535 | * | |
1536 | * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). | |
1537 | */ | |
43cf38eb | 1538 | void __percpu *__alloc_percpu(size_t size, size_t align) |
edcb4639 | 1539 | { |
5835d96e | 1540 | return pcpu_alloc(size, align, false, GFP_KERNEL); |
edcb4639 | 1541 | } |
fbf59bc9 TH |
1542 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1543 | ||
edcb4639 TH |
1544 | /** |
1545 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1546 | * @size: size of area to allocate in bytes | |
1547 | * @align: alignment of area (max PAGE_SIZE) | |
1548 | * | |
9329ba97 TH |
1549 | * Allocate zero-filled percpu area of @size bytes aligned at @align |
1550 | * from reserved percpu area if arch has set it up; otherwise, | |
1551 | * allocation is served from the same dynamic area. Might sleep. | |
1552 | * Might trigger writeouts. | |
edcb4639 | 1553 | * |
ccea34b5 TH |
1554 | * CONTEXT: |
1555 | * Does GFP_KERNEL allocation. | |
1556 | * | |
edcb4639 TH |
1557 | * RETURNS: |
1558 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1559 | */ | |
43cf38eb | 1560 | void __percpu *__alloc_reserved_percpu(size_t size, size_t align) |
edcb4639 | 1561 | { |
5835d96e | 1562 | return pcpu_alloc(size, align, true, GFP_KERNEL); |
edcb4639 TH |
1563 | } |
1564 | ||
a56dbddf | 1565 | /** |
1a4d7607 | 1566 | * pcpu_balance_workfn - manage the amount of free chunks and populated pages |
a56dbddf TH |
1567 | * @work: unused |
1568 | * | |
47504ee0 DZ |
1569 | * Reclaim all fully free chunks except for the first one. This is also |
1570 | * responsible for maintaining the pool of empty populated pages. However, | |
1571 | * it is possible that this is called when physical memory is scarce causing | |
1572 | * OOM killer to be triggered. We should avoid doing so until an actual | |
1573 | * allocation causes the failure as it is possible that requests can be | |
1574 | * serviced from already backed regions. | |
a56dbddf | 1575 | */ |
fe6bd8c3 | 1576 | static void pcpu_balance_workfn(struct work_struct *work) |
fbf59bc9 | 1577 | { |
47504ee0 DZ |
1578 | /* gfp flags passed to underlying allocators */ |
1579 | const gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN; | |
fe6bd8c3 TH |
1580 | LIST_HEAD(to_free); |
1581 | struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; | |
a56dbddf | 1582 | struct pcpu_chunk *chunk, *next; |
1a4d7607 | 1583 | int slot, nr_to_pop, ret; |
a56dbddf | 1584 | |
1a4d7607 TH |
1585 | /* |
1586 | * There's no reason to keep around multiple unused chunks and VM | |
1587 | * areas can be scarce. Destroy all free chunks except for one. | |
1588 | */ | |
ccea34b5 TH |
1589 | mutex_lock(&pcpu_alloc_mutex); |
1590 | spin_lock_irq(&pcpu_lock); | |
a56dbddf | 1591 | |
fe6bd8c3 | 1592 | list_for_each_entry_safe(chunk, next, free_head, list) { |
a56dbddf TH |
1593 | WARN_ON(chunk->immutable); |
1594 | ||
1595 | /* spare the first one */ | |
fe6bd8c3 | 1596 | if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) |
a56dbddf TH |
1597 | continue; |
1598 | ||
fe6bd8c3 | 1599 | list_move(&chunk->list, &to_free); |
a56dbddf TH |
1600 | } |
1601 | ||
ccea34b5 | 1602 | spin_unlock_irq(&pcpu_lock); |
a56dbddf | 1603 | |
fe6bd8c3 | 1604 | list_for_each_entry_safe(chunk, next, &to_free, list) { |
a93ace48 | 1605 | int rs, re; |
dca49645 | 1606 | |
91e914c5 DZF |
1607 | pcpu_for_each_pop_region(chunk->populated, rs, re, 0, |
1608 | chunk->nr_pages) { | |
a93ace48 | 1609 | pcpu_depopulate_chunk(chunk, rs, re); |
b539b87f TH |
1610 | spin_lock_irq(&pcpu_lock); |
1611 | pcpu_chunk_depopulated(chunk, rs, re); | |
1612 | spin_unlock_irq(&pcpu_lock); | |
a93ace48 | 1613 | } |
6081089f | 1614 | pcpu_destroy_chunk(chunk); |
a56dbddf | 1615 | } |
971f3918 | 1616 | |
1a4d7607 TH |
1617 | /* |
1618 | * Ensure there are certain number of free populated pages for | |
1619 | * atomic allocs. Fill up from the most packed so that atomic | |
1620 | * allocs don't increase fragmentation. If atomic allocation | |
1621 | * failed previously, always populate the maximum amount. This | |
1622 | * should prevent atomic allocs larger than PAGE_SIZE from keeping | |
1623 | * failing indefinitely; however, large atomic allocs are not | |
1624 | * something we support properly and can be highly unreliable and | |
1625 | * inefficient. | |
1626 | */ | |
1627 | retry_pop: | |
1628 | if (pcpu_atomic_alloc_failed) { | |
1629 | nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; | |
1630 | /* best effort anyway, don't worry about synchronization */ | |
1631 | pcpu_atomic_alloc_failed = false; | |
1632 | } else { | |
1633 | nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - | |
1634 | pcpu_nr_empty_pop_pages, | |
1635 | 0, PCPU_EMPTY_POP_PAGES_HIGH); | |
1636 | } | |
1637 | ||
1638 | for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { | |
1639 | int nr_unpop = 0, rs, re; | |
1640 | ||
1641 | if (!nr_to_pop) | |
1642 | break; | |
1643 | ||
1644 | spin_lock_irq(&pcpu_lock); | |
1645 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
8ab16c43 | 1646 | nr_unpop = chunk->nr_pages - chunk->nr_populated; |
1a4d7607 TH |
1647 | if (nr_unpop) |
1648 | break; | |
1649 | } | |
1650 | spin_unlock_irq(&pcpu_lock); | |
1651 | ||
1652 | if (!nr_unpop) | |
1653 | continue; | |
1654 | ||
1655 | /* @chunk can't go away while pcpu_alloc_mutex is held */ | |
91e914c5 DZF |
1656 | pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, |
1657 | chunk->nr_pages) { | |
1a4d7607 TH |
1658 | int nr = min(re - rs, nr_to_pop); |
1659 | ||
47504ee0 | 1660 | ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp); |
1a4d7607 TH |
1661 | if (!ret) { |
1662 | nr_to_pop -= nr; | |
1663 | spin_lock_irq(&pcpu_lock); | |
40064aec | 1664 | pcpu_chunk_populated(chunk, rs, rs + nr, false); |
1a4d7607 TH |
1665 | spin_unlock_irq(&pcpu_lock); |
1666 | } else { | |
1667 | nr_to_pop = 0; | |
1668 | } | |
1669 | ||
1670 | if (!nr_to_pop) | |
1671 | break; | |
1672 | } | |
1673 | } | |
1674 | ||
1675 | if (nr_to_pop) { | |
1676 | /* ran out of chunks to populate, create a new one and retry */ | |
47504ee0 | 1677 | chunk = pcpu_create_chunk(gfp); |
1a4d7607 TH |
1678 | if (chunk) { |
1679 | spin_lock_irq(&pcpu_lock); | |
1680 | pcpu_chunk_relocate(chunk, -1); | |
1681 | spin_unlock_irq(&pcpu_lock); | |
1682 | goto retry_pop; | |
1683 | } | |
1684 | } | |
1685 | ||
971f3918 | 1686 | mutex_unlock(&pcpu_alloc_mutex); |
fbf59bc9 TH |
1687 | } |
1688 | ||
1689 | /** | |
1690 | * free_percpu - free percpu area | |
1691 | * @ptr: pointer to area to free | |
1692 | * | |
ccea34b5 TH |
1693 | * Free percpu area @ptr. |
1694 | * | |
1695 | * CONTEXT: | |
1696 | * Can be called from atomic context. | |
fbf59bc9 | 1697 | */ |
43cf38eb | 1698 | void free_percpu(void __percpu *ptr) |
fbf59bc9 | 1699 | { |
129182e5 | 1700 | void *addr; |
fbf59bc9 | 1701 | struct pcpu_chunk *chunk; |
ccea34b5 | 1702 | unsigned long flags; |
40064aec | 1703 | int off; |
fbf59bc9 TH |
1704 | |
1705 | if (!ptr) | |
1706 | return; | |
1707 | ||
f528f0b8 CM |
1708 | kmemleak_free_percpu(ptr); |
1709 | ||
129182e5 AM |
1710 | addr = __pcpu_ptr_to_addr(ptr); |
1711 | ||
ccea34b5 | 1712 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1713 | |
1714 | chunk = pcpu_chunk_addr_search(addr); | |
bba174f5 | 1715 | off = addr - chunk->base_addr; |
fbf59bc9 | 1716 | |
40064aec | 1717 | pcpu_free_area(chunk, off); |
fbf59bc9 | 1718 | |
a56dbddf | 1719 | /* if there are more than one fully free chunks, wake up grim reaper */ |
40064aec | 1720 | if (chunk->free_bytes == pcpu_unit_size) { |
fbf59bc9 TH |
1721 | struct pcpu_chunk *pos; |
1722 | ||
a56dbddf | 1723 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1724 | if (pos != chunk) { |
1a4d7607 | 1725 | pcpu_schedule_balance_work(); |
fbf59bc9 TH |
1726 | break; |
1727 | } | |
1728 | } | |
1729 | ||
df95e795 DZ |
1730 | trace_percpu_free_percpu(chunk->base_addr, off, ptr); |
1731 | ||
ccea34b5 | 1732 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1733 | } |
1734 | EXPORT_SYMBOL_GPL(free_percpu); | |
1735 | ||
383776fa | 1736 | bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) |
10fad5e4 | 1737 | { |
bbddff05 | 1738 | #ifdef CONFIG_SMP |
10fad5e4 TH |
1739 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
1740 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); | |
1741 | unsigned int cpu; | |
1742 | ||
1743 | for_each_possible_cpu(cpu) { | |
1744 | void *start = per_cpu_ptr(base, cpu); | |
383776fa | 1745 | void *va = (void *)addr; |
10fad5e4 | 1746 | |
383776fa | 1747 | if (va >= start && va < start + static_size) { |
8ce371f9 | 1748 | if (can_addr) { |
383776fa | 1749 | *can_addr = (unsigned long) (va - start); |
8ce371f9 PZ |
1750 | *can_addr += (unsigned long) |
1751 | per_cpu_ptr(base, get_boot_cpu_id()); | |
1752 | } | |
10fad5e4 | 1753 | return true; |
383776fa TG |
1754 | } |
1755 | } | |
bbddff05 TH |
1756 | #endif |
1757 | /* on UP, can't distinguish from other static vars, always false */ | |
10fad5e4 TH |
1758 | return false; |
1759 | } | |
1760 | ||
383776fa TG |
1761 | /** |
1762 | * is_kernel_percpu_address - test whether address is from static percpu area | |
1763 | * @addr: address to test | |
1764 | * | |
1765 | * Test whether @addr belongs to in-kernel static percpu area. Module | |
1766 | * static percpu areas are not considered. For those, use | |
1767 | * is_module_percpu_address(). | |
1768 | * | |
1769 | * RETURNS: | |
1770 | * %true if @addr is from in-kernel static percpu area, %false otherwise. | |
1771 | */ | |
1772 | bool is_kernel_percpu_address(unsigned long addr) | |
1773 | { | |
1774 | return __is_kernel_percpu_address(addr, NULL); | |
1775 | } | |
1776 | ||
3b034b0d VG |
1777 | /** |
1778 | * per_cpu_ptr_to_phys - convert translated percpu address to physical address | |
1779 | * @addr: the address to be converted to physical address | |
1780 | * | |
1781 | * Given @addr which is dereferenceable address obtained via one of | |
1782 | * percpu access macros, this function translates it into its physical | |
1783 | * address. The caller is responsible for ensuring @addr stays valid | |
1784 | * until this function finishes. | |
1785 | * | |
67589c71 DY |
1786 | * percpu allocator has special setup for the first chunk, which currently |
1787 | * supports either embedding in linear address space or vmalloc mapping, | |
1788 | * and, from the second one, the backing allocator (currently either vm or | |
1789 | * km) provides translation. | |
1790 | * | |
bffc4375 | 1791 | * The addr can be translated simply without checking if it falls into the |
67589c71 DY |
1792 | * first chunk. But the current code reflects better how percpu allocator |
1793 | * actually works, and the verification can discover both bugs in percpu | |
1794 | * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current | |
1795 | * code. | |
1796 | * | |
3b034b0d VG |
1797 | * RETURNS: |
1798 | * The physical address for @addr. | |
1799 | */ | |
1800 | phys_addr_t per_cpu_ptr_to_phys(void *addr) | |
1801 | { | |
9983b6f0 TH |
1802 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
1803 | bool in_first_chunk = false; | |
a855b84c | 1804 | unsigned long first_low, first_high; |
9983b6f0 TH |
1805 | unsigned int cpu; |
1806 | ||
1807 | /* | |
a855b84c | 1808 | * The following test on unit_low/high isn't strictly |
9983b6f0 TH |
1809 | * necessary but will speed up lookups of addresses which |
1810 | * aren't in the first chunk. | |
c0ebfdc3 DZF |
1811 | * |
1812 | * The address check is against full chunk sizes. pcpu_base_addr | |
1813 | * points to the beginning of the first chunk including the | |
1814 | * static region. Assumes good intent as the first chunk may | |
1815 | * not be full (ie. < pcpu_unit_pages in size). | |
9983b6f0 | 1816 | */ |
c0ebfdc3 DZF |
1817 | first_low = (unsigned long)pcpu_base_addr + |
1818 | pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); | |
1819 | first_high = (unsigned long)pcpu_base_addr + | |
1820 | pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); | |
a855b84c TH |
1821 | if ((unsigned long)addr >= first_low && |
1822 | (unsigned long)addr < first_high) { | |
9983b6f0 TH |
1823 | for_each_possible_cpu(cpu) { |
1824 | void *start = per_cpu_ptr(base, cpu); | |
1825 | ||
1826 | if (addr >= start && addr < start + pcpu_unit_size) { | |
1827 | in_first_chunk = true; | |
1828 | break; | |
1829 | } | |
1830 | } | |
1831 | } | |
1832 | ||
1833 | if (in_first_chunk) { | |
eac522ef | 1834 | if (!is_vmalloc_addr(addr)) |
020ec653 TH |
1835 | return __pa(addr); |
1836 | else | |
9f57bd4d ES |
1837 | return page_to_phys(vmalloc_to_page(addr)) + |
1838 | offset_in_page(addr); | |
020ec653 | 1839 | } else |
9f57bd4d ES |
1840 | return page_to_phys(pcpu_addr_to_page(addr)) + |
1841 | offset_in_page(addr); | |
3b034b0d VG |
1842 | } |
1843 | ||
fbf59bc9 | 1844 | /** |
fd1e8a1f TH |
1845 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
1846 | * @nr_groups: the number of groups | |
1847 | * @nr_units: the number of units | |
1848 | * | |
1849 | * Allocate ai which is large enough for @nr_groups groups containing | |
1850 | * @nr_units units. The returned ai's groups[0].cpu_map points to the | |
1851 | * cpu_map array which is long enough for @nr_units and filled with | |
1852 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map | |
1853 | * pointer of other groups. | |
1854 | * | |
1855 | * RETURNS: | |
1856 | * Pointer to the allocated pcpu_alloc_info on success, NULL on | |
1857 | * failure. | |
1858 | */ | |
1859 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, | |
1860 | int nr_units) | |
1861 | { | |
1862 | struct pcpu_alloc_info *ai; | |
1863 | size_t base_size, ai_size; | |
1864 | void *ptr; | |
1865 | int unit; | |
1866 | ||
1867 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), | |
1868 | __alignof__(ai->groups[0].cpu_map[0])); | |
1869 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); | |
1870 | ||
438a5061 | 1871 | ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), PAGE_SIZE); |
fd1e8a1f TH |
1872 | if (!ptr) |
1873 | return NULL; | |
1874 | ai = ptr; | |
1875 | ptr += base_size; | |
1876 | ||
1877 | ai->groups[0].cpu_map = ptr; | |
1878 | ||
1879 | for (unit = 0; unit < nr_units; unit++) | |
1880 | ai->groups[0].cpu_map[unit] = NR_CPUS; | |
1881 | ||
1882 | ai->nr_groups = nr_groups; | |
1883 | ai->__ai_size = PFN_ALIGN(ai_size); | |
1884 | ||
1885 | return ai; | |
1886 | } | |
1887 | ||
1888 | /** | |
1889 | * pcpu_free_alloc_info - free percpu allocation info | |
1890 | * @ai: pcpu_alloc_info to free | |
1891 | * | |
1892 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). | |
1893 | */ | |
1894 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) | |
1895 | { | |
999c17e3 | 1896 | memblock_free_early(__pa(ai), ai->__ai_size); |
fd1e8a1f TH |
1897 | } |
1898 | ||
fd1e8a1f TH |
1899 | /** |
1900 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info | |
1901 | * @lvl: loglevel | |
1902 | * @ai: allocation info to dump | |
1903 | * | |
1904 | * Print out information about @ai using loglevel @lvl. | |
1905 | */ | |
1906 | static void pcpu_dump_alloc_info(const char *lvl, | |
1907 | const struct pcpu_alloc_info *ai) | |
033e48fb | 1908 | { |
fd1e8a1f | 1909 | int group_width = 1, cpu_width = 1, width; |
033e48fb | 1910 | char empty_str[] = "--------"; |
fd1e8a1f TH |
1911 | int alloc = 0, alloc_end = 0; |
1912 | int group, v; | |
1913 | int upa, apl; /* units per alloc, allocs per line */ | |
1914 | ||
1915 | v = ai->nr_groups; | |
1916 | while (v /= 10) | |
1917 | group_width++; | |
033e48fb | 1918 | |
fd1e8a1f | 1919 | v = num_possible_cpus(); |
033e48fb | 1920 | while (v /= 10) |
fd1e8a1f TH |
1921 | cpu_width++; |
1922 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; | |
033e48fb | 1923 | |
fd1e8a1f TH |
1924 | upa = ai->alloc_size / ai->unit_size; |
1925 | width = upa * (cpu_width + 1) + group_width + 3; | |
1926 | apl = rounddown_pow_of_two(max(60 / width, 1)); | |
033e48fb | 1927 | |
fd1e8a1f TH |
1928 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
1929 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, | |
1930 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); | |
033e48fb | 1931 | |
fd1e8a1f TH |
1932 | for (group = 0; group < ai->nr_groups; group++) { |
1933 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
1934 | int unit = 0, unit_end = 0; | |
1935 | ||
1936 | BUG_ON(gi->nr_units % upa); | |
1937 | for (alloc_end += gi->nr_units / upa; | |
1938 | alloc < alloc_end; alloc++) { | |
1939 | if (!(alloc % apl)) { | |
1170532b | 1940 | pr_cont("\n"); |
fd1e8a1f TH |
1941 | printk("%spcpu-alloc: ", lvl); |
1942 | } | |
1170532b | 1943 | pr_cont("[%0*d] ", group_width, group); |
fd1e8a1f TH |
1944 | |
1945 | for (unit_end += upa; unit < unit_end; unit++) | |
1946 | if (gi->cpu_map[unit] != NR_CPUS) | |
1170532b JP |
1947 | pr_cont("%0*d ", |
1948 | cpu_width, gi->cpu_map[unit]); | |
fd1e8a1f | 1949 | else |
1170532b | 1950 | pr_cont("%s ", empty_str); |
033e48fb | 1951 | } |
033e48fb | 1952 | } |
1170532b | 1953 | pr_cont("\n"); |
033e48fb | 1954 | } |
033e48fb | 1955 | |
fbf59bc9 | 1956 | /** |
8d408b4b | 1957 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
fd1e8a1f | 1958 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
38a6be52 | 1959 | * @base_addr: mapped address |
8d408b4b TH |
1960 | * |
1961 | * Initialize the first percpu chunk which contains the kernel static | |
1962 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1963 | * setup path. |
8d408b4b | 1964 | * |
fd1e8a1f TH |
1965 | * @ai contains all information necessary to initialize the first |
1966 | * chunk and prime the dynamic percpu allocator. | |
1967 | * | |
1968 | * @ai->static_size is the size of static percpu area. | |
1969 | * | |
1970 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to | |
edcb4639 TH |
1971 | * reserve after the static area in the first chunk. This reserves |
1972 | * the first chunk such that it's available only through reserved | |
1973 | * percpu allocation. This is primarily used to serve module percpu | |
1974 | * static areas on architectures where the addressing model has | |
1975 | * limited offset range for symbol relocations to guarantee module | |
1976 | * percpu symbols fall inside the relocatable range. | |
1977 | * | |
fd1e8a1f TH |
1978 | * @ai->dyn_size determines the number of bytes available for dynamic |
1979 | * allocation in the first chunk. The area between @ai->static_size + | |
1980 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. | |
6074d5b0 | 1981 | * |
fd1e8a1f TH |
1982 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
1983 | * and equal to or larger than @ai->static_size + @ai->reserved_size + | |
1984 | * @ai->dyn_size. | |
8d408b4b | 1985 | * |
fd1e8a1f TH |
1986 | * @ai->atom_size is the allocation atom size and used as alignment |
1987 | * for vm areas. | |
8d408b4b | 1988 | * |
fd1e8a1f TH |
1989 | * @ai->alloc_size is the allocation size and always multiple of |
1990 | * @ai->atom_size. This is larger than @ai->atom_size if | |
1991 | * @ai->unit_size is larger than @ai->atom_size. | |
1992 | * | |
1993 | * @ai->nr_groups and @ai->groups describe virtual memory layout of | |
1994 | * percpu areas. Units which should be colocated are put into the | |
1995 | * same group. Dynamic VM areas will be allocated according to these | |
1996 | * groupings. If @ai->nr_groups is zero, a single group containing | |
1997 | * all units is assumed. | |
8d408b4b | 1998 | * |
38a6be52 TH |
1999 | * The caller should have mapped the first chunk at @base_addr and |
2000 | * copied static data to each unit. | |
fbf59bc9 | 2001 | * |
c0ebfdc3 DZF |
2002 | * The first chunk will always contain a static and a dynamic region. |
2003 | * However, the static region is not managed by any chunk. If the first | |
2004 | * chunk also contains a reserved region, it is served by two chunks - | |
2005 | * one for the reserved region and one for the dynamic region. They | |
2006 | * share the same vm, but use offset regions in the area allocation map. | |
2007 | * The chunk serving the dynamic region is circulated in the chunk slots | |
2008 | * and available for dynamic allocation like any other chunk. | |
edcb4639 | 2009 | * |
fbf59bc9 | 2010 | * RETURNS: |
fb435d52 | 2011 | * 0 on success, -errno on failure. |
fbf59bc9 | 2012 | */ |
fb435d52 TH |
2013 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
2014 | void *base_addr) | |
fbf59bc9 | 2015 | { |
b9c39442 | 2016 | size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
d2f3c384 | 2017 | size_t static_size, dyn_size; |
0c4169c3 | 2018 | struct pcpu_chunk *chunk; |
6563297c TH |
2019 | unsigned long *group_offsets; |
2020 | size_t *group_sizes; | |
fb435d52 | 2021 | unsigned long *unit_off; |
fbf59bc9 | 2022 | unsigned int cpu; |
fd1e8a1f TH |
2023 | int *unit_map; |
2024 | int group, unit, i; | |
c0ebfdc3 DZF |
2025 | int map_size; |
2026 | unsigned long tmp_addr; | |
fbf59bc9 | 2027 | |
635b75fc TH |
2028 | #define PCPU_SETUP_BUG_ON(cond) do { \ |
2029 | if (unlikely(cond)) { \ | |
870d4b12 JP |
2030 | pr_emerg("failed to initialize, %s\n", #cond); \ |
2031 | pr_emerg("cpu_possible_mask=%*pb\n", \ | |
807de073 | 2032 | cpumask_pr_args(cpu_possible_mask)); \ |
635b75fc TH |
2033 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ |
2034 | BUG(); \ | |
2035 | } \ | |
2036 | } while (0) | |
2037 | ||
2f39e637 | 2038 | /* sanity checks */ |
635b75fc | 2039 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
bbddff05 | 2040 | #ifdef CONFIG_SMP |
635b75fc | 2041 | PCPU_SETUP_BUG_ON(!ai->static_size); |
f09f1243 | 2042 | PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); |
bbddff05 | 2043 | #endif |
635b75fc | 2044 | PCPU_SETUP_BUG_ON(!base_addr); |
f09f1243 | 2045 | PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); |
635b75fc | 2046 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
f09f1243 | 2047 | PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); |
635b75fc | 2048 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); |
ca460b3c | 2049 | PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE)); |
099a19d9 | 2050 | PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); |
fb29a2cc | 2051 | PCPU_SETUP_BUG_ON(!ai->dyn_size); |
d2f3c384 | 2052 | PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); |
ca460b3c DZF |
2053 | PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) || |
2054 | IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE))); | |
9f645532 | 2055 | PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); |
8d408b4b | 2056 | |
6563297c | 2057 | /* process group information and build config tables accordingly */ |
999c17e3 SS |
2058 | group_offsets = memblock_virt_alloc(ai->nr_groups * |
2059 | sizeof(group_offsets[0]), 0); | |
2060 | group_sizes = memblock_virt_alloc(ai->nr_groups * | |
2061 | sizeof(group_sizes[0]), 0); | |
2062 | unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); | |
2063 | unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); | |
2f39e637 | 2064 | |
fd1e8a1f | 2065 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
ffe0d5a5 | 2066 | unit_map[cpu] = UINT_MAX; |
a855b84c TH |
2067 | |
2068 | pcpu_low_unit_cpu = NR_CPUS; | |
2069 | pcpu_high_unit_cpu = NR_CPUS; | |
2f39e637 | 2070 | |
fd1e8a1f TH |
2071 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
2072 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
2f39e637 | 2073 | |
6563297c TH |
2074 | group_offsets[group] = gi->base_offset; |
2075 | group_sizes[group] = gi->nr_units * ai->unit_size; | |
2076 | ||
fd1e8a1f TH |
2077 | for (i = 0; i < gi->nr_units; i++) { |
2078 | cpu = gi->cpu_map[i]; | |
2079 | if (cpu == NR_CPUS) | |
2080 | continue; | |
8d408b4b | 2081 | |
9f295664 | 2082 | PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); |
635b75fc TH |
2083 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); |
2084 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); | |
fbf59bc9 | 2085 | |
fd1e8a1f | 2086 | unit_map[cpu] = unit + i; |
fb435d52 TH |
2087 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
2088 | ||
a855b84c TH |
2089 | /* determine low/high unit_cpu */ |
2090 | if (pcpu_low_unit_cpu == NR_CPUS || | |
2091 | unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) | |
2092 | pcpu_low_unit_cpu = cpu; | |
2093 | if (pcpu_high_unit_cpu == NR_CPUS || | |
2094 | unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) | |
2095 | pcpu_high_unit_cpu = cpu; | |
fd1e8a1f | 2096 | } |
2f39e637 | 2097 | } |
fd1e8a1f TH |
2098 | pcpu_nr_units = unit; |
2099 | ||
2100 | for_each_possible_cpu(cpu) | |
635b75fc TH |
2101 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
2102 | ||
2103 | /* we're done parsing the input, undefine BUG macro and dump config */ | |
2104 | #undef PCPU_SETUP_BUG_ON | |
bcbea798 | 2105 | pcpu_dump_alloc_info(KERN_DEBUG, ai); |
fd1e8a1f | 2106 | |
6563297c TH |
2107 | pcpu_nr_groups = ai->nr_groups; |
2108 | pcpu_group_offsets = group_offsets; | |
2109 | pcpu_group_sizes = group_sizes; | |
fd1e8a1f | 2110 | pcpu_unit_map = unit_map; |
fb435d52 | 2111 | pcpu_unit_offsets = unit_off; |
2f39e637 TH |
2112 | |
2113 | /* determine basic parameters */ | |
fd1e8a1f | 2114 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
d9b55eeb | 2115 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
6563297c | 2116 | pcpu_atom_size = ai->atom_size; |
ce3141a2 TH |
2117 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
2118 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
cafe8816 | 2119 | |
30a5b536 DZ |
2120 | pcpu_stats_save_ai(ai); |
2121 | ||
d9b55eeb TH |
2122 | /* |
2123 | * Allocate chunk slots. The additional last slot is for | |
2124 | * empty chunks. | |
2125 | */ | |
2126 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
999c17e3 SS |
2127 | pcpu_slot = memblock_virt_alloc( |
2128 | pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); | |
fbf59bc9 TH |
2129 | for (i = 0; i < pcpu_nr_slots; i++) |
2130 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
2131 | ||
d2f3c384 DZF |
2132 | /* |
2133 | * The end of the static region needs to be aligned with the | |
2134 | * minimum allocation size as this offsets the reserved and | |
2135 | * dynamic region. The first chunk ends page aligned by | |
2136 | * expanding the dynamic region, therefore the dynamic region | |
2137 | * can be shrunk to compensate while still staying above the | |
2138 | * configured sizes. | |
2139 | */ | |
2140 | static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); | |
2141 | dyn_size = ai->dyn_size - (static_size - ai->static_size); | |
2142 | ||
edcb4639 | 2143 | /* |
c0ebfdc3 DZF |
2144 | * Initialize first chunk. |
2145 | * If the reserved_size is non-zero, this initializes the reserved | |
2146 | * chunk. If the reserved_size is zero, the reserved chunk is NULL | |
2147 | * and the dynamic region is initialized here. The first chunk, | |
2148 | * pcpu_first_chunk, will always point to the chunk that serves | |
2149 | * the dynamic region. | |
edcb4639 | 2150 | */ |
d2f3c384 DZF |
2151 | tmp_addr = (unsigned long)base_addr + static_size; |
2152 | map_size = ai->reserved_size ?: dyn_size; | |
40064aec | 2153 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); |
61ace7fa | 2154 | |
edcb4639 | 2155 | /* init dynamic chunk if necessary */ |
b9c39442 | 2156 | if (ai->reserved_size) { |
0c4169c3 | 2157 | pcpu_reserved_chunk = chunk; |
b9c39442 | 2158 | |
d2f3c384 | 2159 | tmp_addr = (unsigned long)base_addr + static_size + |
c0ebfdc3 | 2160 | ai->reserved_size; |
d2f3c384 | 2161 | map_size = dyn_size; |
40064aec | 2162 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size); |
edcb4639 TH |
2163 | } |
2164 | ||
2441d15c | 2165 | /* link the first chunk in */ |
0c4169c3 | 2166 | pcpu_first_chunk = chunk; |
0cecf50c | 2167 | pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; |
ae9e6bc9 | 2168 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
fbf59bc9 | 2169 | |
30a5b536 | 2170 | pcpu_stats_chunk_alloc(); |
df95e795 | 2171 | trace_percpu_create_chunk(base_addr); |
30a5b536 | 2172 | |
fbf59bc9 | 2173 | /* we're done */ |
bba174f5 | 2174 | pcpu_base_addr = base_addr; |
fb435d52 | 2175 | return 0; |
fbf59bc9 | 2176 | } |
66c3a757 | 2177 | |
bbddff05 TH |
2178 | #ifdef CONFIG_SMP |
2179 | ||
17f3609c | 2180 | const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { |
f58dc01b TH |
2181 | [PCPU_FC_AUTO] = "auto", |
2182 | [PCPU_FC_EMBED] = "embed", | |
2183 | [PCPU_FC_PAGE] = "page", | |
f58dc01b | 2184 | }; |
66c3a757 | 2185 | |
f58dc01b | 2186 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
66c3a757 | 2187 | |
f58dc01b TH |
2188 | static int __init percpu_alloc_setup(char *str) |
2189 | { | |
5479c78a CG |
2190 | if (!str) |
2191 | return -EINVAL; | |
2192 | ||
f58dc01b TH |
2193 | if (0) |
2194 | /* nada */; | |
2195 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK | |
2196 | else if (!strcmp(str, "embed")) | |
2197 | pcpu_chosen_fc = PCPU_FC_EMBED; | |
2198 | #endif | |
2199 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
2200 | else if (!strcmp(str, "page")) | |
2201 | pcpu_chosen_fc = PCPU_FC_PAGE; | |
f58dc01b TH |
2202 | #endif |
2203 | else | |
870d4b12 | 2204 | pr_warn("unknown allocator %s specified\n", str); |
66c3a757 | 2205 | |
f58dc01b | 2206 | return 0; |
66c3a757 | 2207 | } |
f58dc01b | 2208 | early_param("percpu_alloc", percpu_alloc_setup); |
66c3a757 | 2209 | |
3c9a024f TH |
2210 | /* |
2211 | * pcpu_embed_first_chunk() is used by the generic percpu setup. | |
2212 | * Build it if needed by the arch config or the generic setup is going | |
2213 | * to be used. | |
2214 | */ | |
08fc4580 TH |
2215 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
2216 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
3c9a024f TH |
2217 | #define BUILD_EMBED_FIRST_CHUNK |
2218 | #endif | |
2219 | ||
2220 | /* build pcpu_page_first_chunk() iff needed by the arch config */ | |
2221 | #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) | |
2222 | #define BUILD_PAGE_FIRST_CHUNK | |
2223 | #endif | |
2224 | ||
2225 | /* pcpu_build_alloc_info() is used by both embed and page first chunk */ | |
2226 | #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) | |
2227 | /** | |
2228 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs | |
2229 | * @reserved_size: the size of reserved percpu area in bytes | |
2230 | * @dyn_size: minimum free size for dynamic allocation in bytes | |
2231 | * @atom_size: allocation atom size | |
2232 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
2233 | * | |
2234 | * This function determines grouping of units, their mappings to cpus | |
2235 | * and other parameters considering needed percpu size, allocation | |
2236 | * atom size and distances between CPUs. | |
2237 | * | |
bffc4375 | 2238 | * Groups are always multiples of atom size and CPUs which are of |
3c9a024f TH |
2239 | * LOCAL_DISTANCE both ways are grouped together and share space for |
2240 | * units in the same group. The returned configuration is guaranteed | |
2241 | * to have CPUs on different nodes on different groups and >=75% usage | |
2242 | * of allocated virtual address space. | |
2243 | * | |
2244 | * RETURNS: | |
2245 | * On success, pointer to the new allocation_info is returned. On | |
2246 | * failure, ERR_PTR value is returned. | |
2247 | */ | |
2248 | static struct pcpu_alloc_info * __init pcpu_build_alloc_info( | |
2249 | size_t reserved_size, size_t dyn_size, | |
2250 | size_t atom_size, | |
2251 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
2252 | { | |
2253 | static int group_map[NR_CPUS] __initdata; | |
2254 | static int group_cnt[NR_CPUS] __initdata; | |
2255 | const size_t static_size = __per_cpu_end - __per_cpu_start; | |
2256 | int nr_groups = 1, nr_units = 0; | |
2257 | size_t size_sum, min_unit_size, alloc_size; | |
2258 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
2259 | int last_allocs, group, unit; | |
2260 | unsigned int cpu, tcpu; | |
2261 | struct pcpu_alloc_info *ai; | |
2262 | unsigned int *cpu_map; | |
2263 | ||
2264 | /* this function may be called multiple times */ | |
2265 | memset(group_map, 0, sizeof(group_map)); | |
2266 | memset(group_cnt, 0, sizeof(group_cnt)); | |
2267 | ||
2268 | /* calculate size_sum and ensure dyn_size is enough for early alloc */ | |
2269 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
2270 | max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); | |
2271 | dyn_size = size_sum - static_size - reserved_size; | |
2272 | ||
2273 | /* | |
2274 | * Determine min_unit_size, alloc_size and max_upa such that | |
2275 | * alloc_size is multiple of atom_size and is the smallest | |
25985edc | 2276 | * which can accommodate 4k aligned segments which are equal to |
3c9a024f TH |
2277 | * or larger than min_unit_size. |
2278 | */ | |
2279 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
2280 | ||
9c015162 | 2281 | /* determine the maximum # of units that can fit in an allocation */ |
3c9a024f TH |
2282 | alloc_size = roundup(min_unit_size, atom_size); |
2283 | upa = alloc_size / min_unit_size; | |
f09f1243 | 2284 | while (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
2285 | upa--; |
2286 | max_upa = upa; | |
2287 | ||
2288 | /* group cpus according to their proximity */ | |
2289 | for_each_possible_cpu(cpu) { | |
2290 | group = 0; | |
2291 | next_group: | |
2292 | for_each_possible_cpu(tcpu) { | |
2293 | if (cpu == tcpu) | |
2294 | break; | |
2295 | if (group_map[tcpu] == group && cpu_distance_fn && | |
2296 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || | |
2297 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
2298 | group++; | |
2299 | nr_groups = max(nr_groups, group + 1); | |
2300 | goto next_group; | |
2301 | } | |
2302 | } | |
2303 | group_map[cpu] = group; | |
2304 | group_cnt[group]++; | |
2305 | } | |
2306 | ||
2307 | /* | |
9c015162 DZF |
2308 | * Wasted space is caused by a ratio imbalance of upa to group_cnt. |
2309 | * Expand the unit_size until we use >= 75% of the units allocated. | |
2310 | * Related to atom_size, which could be much larger than the unit_size. | |
3c9a024f TH |
2311 | */ |
2312 | last_allocs = INT_MAX; | |
2313 | for (upa = max_upa; upa; upa--) { | |
2314 | int allocs = 0, wasted = 0; | |
2315 | ||
f09f1243 | 2316 | if (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
2317 | continue; |
2318 | ||
2319 | for (group = 0; group < nr_groups; group++) { | |
2320 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); | |
2321 | allocs += this_allocs; | |
2322 | wasted += this_allocs * upa - group_cnt[group]; | |
2323 | } | |
2324 | ||
2325 | /* | |
2326 | * Don't accept if wastage is over 1/3. The | |
2327 | * greater-than comparison ensures upa==1 always | |
2328 | * passes the following check. | |
2329 | */ | |
2330 | if (wasted > num_possible_cpus() / 3) | |
2331 | continue; | |
2332 | ||
2333 | /* and then don't consume more memory */ | |
2334 | if (allocs > last_allocs) | |
2335 | break; | |
2336 | last_allocs = allocs; | |
2337 | best_upa = upa; | |
2338 | } | |
2339 | upa = best_upa; | |
2340 | ||
2341 | /* allocate and fill alloc_info */ | |
2342 | for (group = 0; group < nr_groups; group++) | |
2343 | nr_units += roundup(group_cnt[group], upa); | |
2344 | ||
2345 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); | |
2346 | if (!ai) | |
2347 | return ERR_PTR(-ENOMEM); | |
2348 | cpu_map = ai->groups[0].cpu_map; | |
2349 | ||
2350 | for (group = 0; group < nr_groups; group++) { | |
2351 | ai->groups[group].cpu_map = cpu_map; | |
2352 | cpu_map += roundup(group_cnt[group], upa); | |
2353 | } | |
2354 | ||
2355 | ai->static_size = static_size; | |
2356 | ai->reserved_size = reserved_size; | |
2357 | ai->dyn_size = dyn_size; | |
2358 | ai->unit_size = alloc_size / upa; | |
2359 | ai->atom_size = atom_size; | |
2360 | ai->alloc_size = alloc_size; | |
2361 | ||
2362 | for (group = 0, unit = 0; group_cnt[group]; group++) { | |
2363 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2364 | ||
2365 | /* | |
2366 | * Initialize base_offset as if all groups are located | |
2367 | * back-to-back. The caller should update this to | |
2368 | * reflect actual allocation. | |
2369 | */ | |
2370 | gi->base_offset = unit * ai->unit_size; | |
2371 | ||
2372 | for_each_possible_cpu(cpu) | |
2373 | if (group_map[cpu] == group) | |
2374 | gi->cpu_map[gi->nr_units++] = cpu; | |
2375 | gi->nr_units = roundup(gi->nr_units, upa); | |
2376 | unit += gi->nr_units; | |
2377 | } | |
2378 | BUG_ON(unit != nr_units); | |
2379 | ||
2380 | return ai; | |
2381 | } | |
2382 | #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ | |
2383 | ||
2384 | #if defined(BUILD_EMBED_FIRST_CHUNK) | |
66c3a757 TH |
2385 | /** |
2386 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
66c3a757 | 2387 | * @reserved_size: the size of reserved percpu area in bytes |
4ba6ce25 | 2388 | * @dyn_size: minimum free size for dynamic allocation in bytes |
c8826dd5 TH |
2389 | * @atom_size: allocation atom size |
2390 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
2391 | * @alloc_fn: function to allocate percpu page | |
25985edc | 2392 | * @free_fn: function to free percpu page |
66c3a757 TH |
2393 | * |
2394 | * This is a helper to ease setting up embedded first percpu chunk and | |
2395 | * can be called where pcpu_setup_first_chunk() is expected. | |
2396 | * | |
2397 | * If this function is used to setup the first chunk, it is allocated | |
c8826dd5 TH |
2398 | * by calling @alloc_fn and used as-is without being mapped into |
2399 | * vmalloc area. Allocations are always whole multiples of @atom_size | |
2400 | * aligned to @atom_size. | |
2401 | * | |
2402 | * This enables the first chunk to piggy back on the linear physical | |
2403 | * mapping which often uses larger page size. Please note that this | |
2404 | * can result in very sparse cpu->unit mapping on NUMA machines thus | |
2405 | * requiring large vmalloc address space. Don't use this allocator if | |
2406 | * vmalloc space is not orders of magnitude larger than distances | |
2407 | * between node memory addresses (ie. 32bit NUMA machines). | |
66c3a757 | 2408 | * |
4ba6ce25 | 2409 | * @dyn_size specifies the minimum dynamic area size. |
66c3a757 TH |
2410 | * |
2411 | * If the needed size is smaller than the minimum or specified unit | |
c8826dd5 | 2412 | * size, the leftover is returned using @free_fn. |
66c3a757 TH |
2413 | * |
2414 | * RETURNS: | |
fb435d52 | 2415 | * 0 on success, -errno on failure. |
66c3a757 | 2416 | */ |
4ba6ce25 | 2417 | int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, |
c8826dd5 TH |
2418 | size_t atom_size, |
2419 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, | |
2420 | pcpu_fc_alloc_fn_t alloc_fn, | |
2421 | pcpu_fc_free_fn_t free_fn) | |
66c3a757 | 2422 | { |
c8826dd5 TH |
2423 | void *base = (void *)ULONG_MAX; |
2424 | void **areas = NULL; | |
fd1e8a1f | 2425 | struct pcpu_alloc_info *ai; |
93c76b6b | 2426 | size_t size_sum, areas_size; |
2427 | unsigned long max_distance; | |
9b739662 | 2428 | int group, i, highest_group, rc; |
66c3a757 | 2429 | |
c8826dd5 TH |
2430 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
2431 | cpu_distance_fn); | |
fd1e8a1f TH |
2432 | if (IS_ERR(ai)) |
2433 | return PTR_ERR(ai); | |
66c3a757 | 2434 | |
fd1e8a1f | 2435 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
c8826dd5 | 2436 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
fa8a7094 | 2437 | |
999c17e3 | 2438 | areas = memblock_virt_alloc_nopanic(areas_size, 0); |
c8826dd5 | 2439 | if (!areas) { |
fb435d52 | 2440 | rc = -ENOMEM; |
c8826dd5 | 2441 | goto out_free; |
fa8a7094 | 2442 | } |
66c3a757 | 2443 | |
9b739662 | 2444 | /* allocate, copy and determine base address & max_distance */ |
2445 | highest_group = 0; | |
c8826dd5 TH |
2446 | for (group = 0; group < ai->nr_groups; group++) { |
2447 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2448 | unsigned int cpu = NR_CPUS; | |
2449 | void *ptr; | |
2450 | ||
2451 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) | |
2452 | cpu = gi->cpu_map[i]; | |
2453 | BUG_ON(cpu == NR_CPUS); | |
2454 | ||
2455 | /* allocate space for the whole group */ | |
2456 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); | |
2457 | if (!ptr) { | |
2458 | rc = -ENOMEM; | |
2459 | goto out_free_areas; | |
2460 | } | |
f528f0b8 CM |
2461 | /* kmemleak tracks the percpu allocations separately */ |
2462 | kmemleak_free(ptr); | |
c8826dd5 | 2463 | areas[group] = ptr; |
fd1e8a1f | 2464 | |
c8826dd5 | 2465 | base = min(ptr, base); |
9b739662 | 2466 | if (ptr > areas[highest_group]) |
2467 | highest_group = group; | |
2468 | } | |
2469 | max_distance = areas[highest_group] - base; | |
2470 | max_distance += ai->unit_size * ai->groups[highest_group].nr_units; | |
2471 | ||
2472 | /* warn if maximum distance is further than 75% of vmalloc space */ | |
2473 | if (max_distance > VMALLOC_TOTAL * 3 / 4) { | |
2474 | pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", | |
2475 | max_distance, VMALLOC_TOTAL); | |
2476 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
2477 | /* and fail if we have fallback */ | |
2478 | rc = -EINVAL; | |
2479 | goto out_free_areas; | |
2480 | #endif | |
42b64281 TH |
2481 | } |
2482 | ||
2483 | /* | |
2484 | * Copy data and free unused parts. This should happen after all | |
2485 | * allocations are complete; otherwise, we may end up with | |
2486 | * overlapping groups. | |
2487 | */ | |
2488 | for (group = 0; group < ai->nr_groups; group++) { | |
2489 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2490 | void *ptr = areas[group]; | |
c8826dd5 TH |
2491 | |
2492 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { | |
2493 | if (gi->cpu_map[i] == NR_CPUS) { | |
2494 | /* unused unit, free whole */ | |
2495 | free_fn(ptr, ai->unit_size); | |
2496 | continue; | |
2497 | } | |
2498 | /* copy and return the unused part */ | |
2499 | memcpy(ptr, __per_cpu_load, ai->static_size); | |
2500 | free_fn(ptr + size_sum, ai->unit_size - size_sum); | |
2501 | } | |
fa8a7094 | 2502 | } |
66c3a757 | 2503 | |
c8826dd5 | 2504 | /* base address is now known, determine group base offsets */ |
6ea529a2 | 2505 | for (group = 0; group < ai->nr_groups; group++) { |
c8826dd5 | 2506 | ai->groups[group].base_offset = areas[group] - base; |
6ea529a2 | 2507 | } |
c8826dd5 | 2508 | |
870d4b12 | 2509 | pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
fd1e8a1f TH |
2510 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
2511 | ai->dyn_size, ai->unit_size); | |
d4b95f80 | 2512 | |
fb435d52 | 2513 | rc = pcpu_setup_first_chunk(ai, base); |
c8826dd5 TH |
2514 | goto out_free; |
2515 | ||
2516 | out_free_areas: | |
2517 | for (group = 0; group < ai->nr_groups; group++) | |
f851c8d8 MH |
2518 | if (areas[group]) |
2519 | free_fn(areas[group], | |
2520 | ai->groups[group].nr_units * ai->unit_size); | |
c8826dd5 | 2521 | out_free: |
fd1e8a1f | 2522 | pcpu_free_alloc_info(ai); |
c8826dd5 | 2523 | if (areas) |
999c17e3 | 2524 | memblock_free_early(__pa(areas), areas_size); |
fb435d52 | 2525 | return rc; |
d4b95f80 | 2526 | } |
3c9a024f | 2527 | #endif /* BUILD_EMBED_FIRST_CHUNK */ |
d4b95f80 | 2528 | |
3c9a024f | 2529 | #ifdef BUILD_PAGE_FIRST_CHUNK |
d4b95f80 | 2530 | /** |
00ae4064 | 2531 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
2532 | * @reserved_size: the size of reserved percpu area in bytes |
2533 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
25985edc | 2534 | * @free_fn: function to free percpu page, always called with PAGE_SIZE |
d4b95f80 TH |
2535 | * @populate_pte_fn: function to populate pte |
2536 | * | |
00ae4064 TH |
2537 | * This is a helper to ease setting up page-remapped first percpu |
2538 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
2539 | * |
2540 | * This is the basic allocator. Static percpu area is allocated | |
2541 | * page-by-page into vmalloc area. | |
2542 | * | |
2543 | * RETURNS: | |
fb435d52 | 2544 | * 0 on success, -errno on failure. |
d4b95f80 | 2545 | */ |
fb435d52 TH |
2546 | int __init pcpu_page_first_chunk(size_t reserved_size, |
2547 | pcpu_fc_alloc_fn_t alloc_fn, | |
2548 | pcpu_fc_free_fn_t free_fn, | |
2549 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 2550 | { |
8f05a6a6 | 2551 | static struct vm_struct vm; |
fd1e8a1f | 2552 | struct pcpu_alloc_info *ai; |
00ae4064 | 2553 | char psize_str[16]; |
ce3141a2 | 2554 | int unit_pages; |
d4b95f80 | 2555 | size_t pages_size; |
ce3141a2 | 2556 | struct page **pages; |
fb435d52 | 2557 | int unit, i, j, rc; |
8f606604 | 2558 | int upa; |
2559 | int nr_g0_units; | |
d4b95f80 | 2560 | |
00ae4064 TH |
2561 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
2562 | ||
4ba6ce25 | 2563 | ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); |
fd1e8a1f TH |
2564 | if (IS_ERR(ai)) |
2565 | return PTR_ERR(ai); | |
2566 | BUG_ON(ai->nr_groups != 1); | |
8f606604 | 2567 | upa = ai->alloc_size/ai->unit_size; |
2568 | nr_g0_units = roundup(num_possible_cpus(), upa); | |
2569 | if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) { | |
2570 | pcpu_free_alloc_info(ai); | |
2571 | return -EINVAL; | |
2572 | } | |
fd1e8a1f TH |
2573 | |
2574 | unit_pages = ai->unit_size >> PAGE_SHIFT; | |
d4b95f80 TH |
2575 | |
2576 | /* unaligned allocations can't be freed, round up to page size */ | |
fd1e8a1f TH |
2577 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
2578 | sizeof(pages[0])); | |
999c17e3 | 2579 | pages = memblock_virt_alloc(pages_size, 0); |
d4b95f80 | 2580 | |
8f05a6a6 | 2581 | /* allocate pages */ |
d4b95f80 | 2582 | j = 0; |
8f606604 | 2583 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
2584 | unsigned int cpu = ai->groups[0].cpu_map[unit]; | |
ce3141a2 | 2585 | for (i = 0; i < unit_pages; i++) { |
d4b95f80 TH |
2586 | void *ptr; |
2587 | ||
3cbc8565 | 2588 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
d4b95f80 | 2589 | if (!ptr) { |
870d4b12 | 2590 | pr_warn("failed to allocate %s page for cpu%u\n", |
8f606604 | 2591 | psize_str, cpu); |
d4b95f80 TH |
2592 | goto enomem; |
2593 | } | |
f528f0b8 CM |
2594 | /* kmemleak tracks the percpu allocations separately */ |
2595 | kmemleak_free(ptr); | |
ce3141a2 | 2596 | pages[j++] = virt_to_page(ptr); |
d4b95f80 | 2597 | } |
8f606604 | 2598 | } |
d4b95f80 | 2599 | |
8f05a6a6 TH |
2600 | /* allocate vm area, map the pages and copy static data */ |
2601 | vm.flags = VM_ALLOC; | |
fd1e8a1f | 2602 | vm.size = num_possible_cpus() * ai->unit_size; |
8f05a6a6 TH |
2603 | vm_area_register_early(&vm, PAGE_SIZE); |
2604 | ||
fd1e8a1f | 2605 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
1d9d3257 | 2606 | unsigned long unit_addr = |
fd1e8a1f | 2607 | (unsigned long)vm.addr + unit * ai->unit_size; |
8f05a6a6 | 2608 | |
ce3141a2 | 2609 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
2610 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
2611 | ||
2612 | /* pte already populated, the following shouldn't fail */ | |
fb435d52 TH |
2613 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
2614 | unit_pages); | |
2615 | if (rc < 0) | |
2616 | panic("failed to map percpu area, err=%d\n", rc); | |
66c3a757 | 2617 | |
8f05a6a6 TH |
2618 | /* |
2619 | * FIXME: Archs with virtual cache should flush local | |
2620 | * cache for the linear mapping here - something | |
2621 | * equivalent to flush_cache_vmap() on the local cpu. | |
2622 | * flush_cache_vmap() can't be used as most supporting | |
2623 | * data structures are not set up yet. | |
2624 | */ | |
2625 | ||
2626 | /* copy static data */ | |
fd1e8a1f | 2627 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
66c3a757 TH |
2628 | } |
2629 | ||
2630 | /* we're ready, commit */ | |
870d4b12 | 2631 | pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", |
fd1e8a1f TH |
2632 | unit_pages, psize_str, vm.addr, ai->static_size, |
2633 | ai->reserved_size, ai->dyn_size); | |
d4b95f80 | 2634 | |
fb435d52 | 2635 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
d4b95f80 TH |
2636 | goto out_free_ar; |
2637 | ||
2638 | enomem: | |
2639 | while (--j >= 0) | |
ce3141a2 | 2640 | free_fn(page_address(pages[j]), PAGE_SIZE); |
fb435d52 | 2641 | rc = -ENOMEM; |
d4b95f80 | 2642 | out_free_ar: |
999c17e3 | 2643 | memblock_free_early(__pa(pages), pages_size); |
fd1e8a1f | 2644 | pcpu_free_alloc_info(ai); |
fb435d52 | 2645 | return rc; |
d4b95f80 | 2646 | } |
3c9a024f | 2647 | #endif /* BUILD_PAGE_FIRST_CHUNK */ |
d4b95f80 | 2648 | |
bbddff05 | 2649 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
e74e3962 | 2650 | /* |
bbddff05 | 2651 | * Generic SMP percpu area setup. |
e74e3962 TH |
2652 | * |
2653 | * The embedding helper is used because its behavior closely resembles | |
2654 | * the original non-dynamic generic percpu area setup. This is | |
2655 | * important because many archs have addressing restrictions and might | |
2656 | * fail if the percpu area is located far away from the previous | |
2657 | * location. As an added bonus, in non-NUMA cases, embedding is | |
2658 | * generally a good idea TLB-wise because percpu area can piggy back | |
2659 | * on the physical linear memory mapping which uses large page | |
2660 | * mappings on applicable archs. | |
2661 | */ | |
e74e3962 TH |
2662 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
2663 | EXPORT_SYMBOL(__per_cpu_offset); | |
2664 | ||
c8826dd5 TH |
2665 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
2666 | size_t align) | |
2667 | { | |
999c17e3 SS |
2668 | return memblock_virt_alloc_from_nopanic( |
2669 | size, align, __pa(MAX_DMA_ADDRESS)); | |
c8826dd5 | 2670 | } |
66c3a757 | 2671 | |
c8826dd5 TH |
2672 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
2673 | { | |
999c17e3 | 2674 | memblock_free_early(__pa(ptr), size); |
c8826dd5 TH |
2675 | } |
2676 | ||
e74e3962 TH |
2677 | void __init setup_per_cpu_areas(void) |
2678 | { | |
e74e3962 TH |
2679 | unsigned long delta; |
2680 | unsigned int cpu; | |
fb435d52 | 2681 | int rc; |
e74e3962 TH |
2682 | |
2683 | /* | |
2684 | * Always reserve area for module percpu variables. That's | |
2685 | * what the legacy allocator did. | |
2686 | */ | |
fb435d52 | 2687 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
c8826dd5 TH |
2688 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
2689 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); | |
fb435d52 | 2690 | if (rc < 0) |
bbddff05 | 2691 | panic("Failed to initialize percpu areas."); |
e74e3962 TH |
2692 | |
2693 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2694 | for_each_possible_cpu(cpu) | |
fb435d52 | 2695 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
66c3a757 | 2696 | } |
bbddff05 TH |
2697 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
2698 | ||
2699 | #else /* CONFIG_SMP */ | |
2700 | ||
2701 | /* | |
2702 | * UP percpu area setup. | |
2703 | * | |
2704 | * UP always uses km-based percpu allocator with identity mapping. | |
2705 | * Static percpu variables are indistinguishable from the usual static | |
2706 | * variables and don't require any special preparation. | |
2707 | */ | |
2708 | void __init setup_per_cpu_areas(void) | |
2709 | { | |
2710 | const size_t unit_size = | |
2711 | roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, | |
2712 | PERCPU_DYNAMIC_RESERVE)); | |
2713 | struct pcpu_alloc_info *ai; | |
2714 | void *fc; | |
2715 | ||
2716 | ai = pcpu_alloc_alloc_info(1, 1); | |
999c17e3 SS |
2717 | fc = memblock_virt_alloc_from_nopanic(unit_size, |
2718 | PAGE_SIZE, | |
2719 | __pa(MAX_DMA_ADDRESS)); | |
bbddff05 TH |
2720 | if (!ai || !fc) |
2721 | panic("Failed to allocate memory for percpu areas."); | |
100d13c3 CM |
2722 | /* kmemleak tracks the percpu allocations separately */ |
2723 | kmemleak_free(fc); | |
bbddff05 TH |
2724 | |
2725 | ai->dyn_size = unit_size; | |
2726 | ai->unit_size = unit_size; | |
2727 | ai->atom_size = unit_size; | |
2728 | ai->alloc_size = unit_size; | |
2729 | ai->groups[0].nr_units = 1; | |
2730 | ai->groups[0].cpu_map[0] = 0; | |
2731 | ||
2732 | if (pcpu_setup_first_chunk(ai, fc) < 0) | |
2733 | panic("Failed to initialize percpu areas."); | |
abee2105 NP |
2734 | #ifdef CONFIG_CRIS |
2735 | #warning "the CRIS architecture has physical and virtual addresses confused" | |
2736 | #else | |
438a5061 | 2737 | pcpu_free_alloc_info(ai); |
abee2105 | 2738 | #endif |
bbddff05 TH |
2739 | } |
2740 | ||
2741 | #endif /* CONFIG_SMP */ | |
099a19d9 | 2742 | |
1a4d7607 TH |
2743 | /* |
2744 | * Percpu allocator is initialized early during boot when neither slab or | |
2745 | * workqueue is available. Plug async management until everything is up | |
2746 | * and running. | |
2747 | */ | |
2748 | static int __init percpu_enable_async(void) | |
2749 | { | |
2750 | pcpu_async_enabled = true; | |
2751 | return 0; | |
2752 | } | |
2753 | subsys_initcall(percpu_enable_async); |