]>
Commit | Line | Data |
---|---|---|
fbf59bc9 | 1 | /* |
88999a89 | 2 | * mm/percpu.c - percpu memory allocator |
fbf59bc9 TH |
3 | * |
4 | * Copyright (C) 2009 SUSE Linux Products GmbH | |
5 | * Copyright (C) 2009 Tejun Heo <tj@kernel.org> | |
6 | * | |
9c015162 | 7 | * This file is released under the GPLv2 license. |
fbf59bc9 | 8 | * |
9c015162 DZF |
9 | * The percpu allocator handles both static and dynamic areas. Percpu |
10 | * areas are allocated in chunks which are divided into units. There is | |
11 | * a 1-to-1 mapping for units to possible cpus. These units are grouped | |
12 | * based on NUMA properties of the machine. | |
fbf59bc9 TH |
13 | * |
14 | * c0 c1 c2 | |
15 | * ------------------- ------------------- ------------ | |
16 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u | |
17 | * ------------------- ...... ------------------- .... ------------ | |
18 | * | |
9c015162 DZF |
19 | * Allocation is done by offsets into a unit's address space. Ie., an |
20 | * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0, | |
21 | * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear | |
22 | * and even sparse. Access is handled by configuring percpu base | |
23 | * registers according to the cpu to unit mappings and offsetting the | |
24 | * base address using pcpu_unit_size. | |
25 | * | |
26 | * There is special consideration for the first chunk which must handle | |
27 | * the static percpu variables in the kernel image as allocation services | |
28 | * are not online yet. In short, the first chunk is structure like so: | |
29 | * | |
30 | * <Static | [Reserved] | Dynamic> | |
31 | * | |
32 | * The static data is copied from the original section managed by the | |
33 | * linker. The reserved section, if non-zero, primarily manages static | |
34 | * percpu variables from kernel modules. Finally, the dynamic section | |
35 | * takes care of normal allocations. | |
fbf59bc9 TH |
36 | * |
37 | * Allocation state in each chunk is kept using an array of integers | |
38 | * on chunk->map. A positive value in the map represents a free | |
39 | * region and negative allocated. Allocation inside a chunk is done | |
40 | * by scanning this map sequentially and serving the first matching | |
41 | * entry. This is mostly copied from the percpu_modalloc() allocator. | |
e1b9aa3f CL |
42 | * Chunks can be determined from the address using the index field |
43 | * in the page struct. The index field contains a pointer to the chunk. | |
fbf59bc9 | 44 | * |
9c015162 DZF |
45 | * These chunks are organized into lists according to free_size and |
46 | * tries to allocate from the fullest chunk first. Each chunk maintains | |
47 | * a maximum contiguous area size hint which is guaranteed to be equal | |
48 | * to or larger than the maximum contiguous area in the chunk. This | |
49 | * helps prevent the allocator from iterating over chunks unnecessarily. | |
50 | * | |
4091fb95 | 51 | * To use this allocator, arch code should do the following: |
fbf59bc9 | 52 | * |
fbf59bc9 | 53 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
e0100983 TH |
54 | * regular address to percpu pointer and back if they need to be |
55 | * different from the default | |
fbf59bc9 | 56 | * |
8d408b4b TH |
57 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
58 | * setup the first chunk containing the kernel static percpu area | |
fbf59bc9 TH |
59 | */ |
60 | ||
870d4b12 JP |
61 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
62 | ||
fbf59bc9 TH |
63 | #include <linux/bitmap.h> |
64 | #include <linux/bootmem.h> | |
fd1e8a1f | 65 | #include <linux/err.h> |
fbf59bc9 | 66 | #include <linux/list.h> |
a530b795 | 67 | #include <linux/log2.h> |
fbf59bc9 TH |
68 | #include <linux/mm.h> |
69 | #include <linux/module.h> | |
70 | #include <linux/mutex.h> | |
71 | #include <linux/percpu.h> | |
72 | #include <linux/pfn.h> | |
fbf59bc9 | 73 | #include <linux/slab.h> |
ccea34b5 | 74 | #include <linux/spinlock.h> |
fbf59bc9 | 75 | #include <linux/vmalloc.h> |
a56dbddf | 76 | #include <linux/workqueue.h> |
f528f0b8 | 77 | #include <linux/kmemleak.h> |
fbf59bc9 TH |
78 | |
79 | #include <asm/cacheflush.h> | |
e0100983 | 80 | #include <asm/sections.h> |
fbf59bc9 | 81 | #include <asm/tlbflush.h> |
3b034b0d | 82 | #include <asm/io.h> |
fbf59bc9 | 83 | |
df95e795 DZ |
84 | #define CREATE_TRACE_POINTS |
85 | #include <trace/events/percpu.h> | |
86 | ||
8fa3ed80 DZ |
87 | #include "percpu-internal.h" |
88 | ||
fbf59bc9 TH |
89 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
90 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ | |
9c824b6a TH |
91 | #define PCPU_ATOMIC_MAP_MARGIN_LOW 32 |
92 | #define PCPU_ATOMIC_MAP_MARGIN_HIGH 64 | |
1a4d7607 TH |
93 | #define PCPU_EMPTY_POP_PAGES_LOW 2 |
94 | #define PCPU_EMPTY_POP_PAGES_HIGH 4 | |
fbf59bc9 | 95 | |
bbddff05 | 96 | #ifdef CONFIG_SMP |
e0100983 TH |
97 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
98 | #ifndef __addr_to_pcpu_ptr | |
99 | #define __addr_to_pcpu_ptr(addr) \ | |
43cf38eb TH |
100 | (void __percpu *)((unsigned long)(addr) - \ |
101 | (unsigned long)pcpu_base_addr + \ | |
102 | (unsigned long)__per_cpu_start) | |
e0100983 TH |
103 | #endif |
104 | #ifndef __pcpu_ptr_to_addr | |
105 | #define __pcpu_ptr_to_addr(ptr) \ | |
43cf38eb TH |
106 | (void __force *)((unsigned long)(ptr) + \ |
107 | (unsigned long)pcpu_base_addr - \ | |
108 | (unsigned long)__per_cpu_start) | |
e0100983 | 109 | #endif |
bbddff05 TH |
110 | #else /* CONFIG_SMP */ |
111 | /* on UP, it's always identity mapped */ | |
112 | #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr) | |
113 | #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr) | |
114 | #endif /* CONFIG_SMP */ | |
e0100983 | 115 | |
1328710b DM |
116 | static int pcpu_unit_pages __ro_after_init; |
117 | static int pcpu_unit_size __ro_after_init; | |
118 | static int pcpu_nr_units __ro_after_init; | |
119 | static int pcpu_atom_size __ro_after_init; | |
8fa3ed80 | 120 | int pcpu_nr_slots __ro_after_init; |
1328710b | 121 | static size_t pcpu_chunk_struct_size __ro_after_init; |
fbf59bc9 | 122 | |
a855b84c | 123 | /* cpus with the lowest and highest unit addresses */ |
1328710b DM |
124 | static unsigned int pcpu_low_unit_cpu __ro_after_init; |
125 | static unsigned int pcpu_high_unit_cpu __ro_after_init; | |
2f39e637 | 126 | |
fbf59bc9 | 127 | /* the address of the first chunk which starts with the kernel static area */ |
1328710b | 128 | void *pcpu_base_addr __ro_after_init; |
fbf59bc9 TH |
129 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
130 | ||
1328710b DM |
131 | static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */ |
132 | const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */ | |
2f39e637 | 133 | |
6563297c | 134 | /* group information, used for vm allocation */ |
1328710b DM |
135 | static int pcpu_nr_groups __ro_after_init; |
136 | static const unsigned long *pcpu_group_offsets __ro_after_init; | |
137 | static const size_t *pcpu_group_sizes __ro_after_init; | |
6563297c | 138 | |
ae9e6bc9 TH |
139 | /* |
140 | * The first chunk which always exists. Note that unlike other | |
141 | * chunks, this one can be allocated and mapped in several different | |
142 | * ways and thus often doesn't live in the vmalloc area. | |
143 | */ | |
8fa3ed80 | 144 | struct pcpu_chunk *pcpu_first_chunk __ro_after_init; |
ae9e6bc9 TH |
145 | |
146 | /* | |
147 | * Optional reserved chunk. This chunk reserves part of the first | |
e2266705 DZF |
148 | * chunk and serves it for reserved allocations. When the reserved |
149 | * region doesn't exist, the following variable is NULL. | |
ae9e6bc9 | 150 | */ |
8fa3ed80 | 151 | struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init; |
edcb4639 | 152 | |
8fa3ed80 | 153 | DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */ |
6710e594 | 154 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */ |
fbf59bc9 | 155 | |
8fa3ed80 | 156 | struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */ |
fbf59bc9 | 157 | |
4f996e23 TH |
158 | /* chunks which need their map areas extended, protected by pcpu_lock */ |
159 | static LIST_HEAD(pcpu_map_extend_chunks); | |
160 | ||
b539b87f TH |
161 | /* |
162 | * The number of empty populated pages, protected by pcpu_lock. The | |
163 | * reserved chunk doesn't contribute to the count. | |
164 | */ | |
6b9b6f39 | 165 | int pcpu_nr_empty_pop_pages; |
b539b87f | 166 | |
1a4d7607 TH |
167 | /* |
168 | * Balance work is used to populate or destroy chunks asynchronously. We | |
169 | * try to keep the number of populated free pages between | |
170 | * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one | |
171 | * empty chunk. | |
172 | */ | |
fe6bd8c3 TH |
173 | static void pcpu_balance_workfn(struct work_struct *work); |
174 | static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn); | |
1a4d7607 TH |
175 | static bool pcpu_async_enabled __read_mostly; |
176 | static bool pcpu_atomic_alloc_failed; | |
177 | ||
178 | static void pcpu_schedule_balance_work(void) | |
179 | { | |
180 | if (pcpu_async_enabled) | |
181 | schedule_work(&pcpu_balance_work); | |
182 | } | |
a56dbddf | 183 | |
c0ebfdc3 | 184 | /** |
560f2c23 DZF |
185 | * pcpu_addr_in_chunk - check if the address is served from this chunk |
186 | * @chunk: chunk of interest | |
187 | * @addr: percpu address | |
c0ebfdc3 DZF |
188 | * |
189 | * RETURNS: | |
560f2c23 | 190 | * True if the address is served from this chunk. |
c0ebfdc3 | 191 | */ |
560f2c23 | 192 | static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr) |
020ec653 | 193 | { |
c0ebfdc3 DZF |
194 | void *start_addr, *end_addr; |
195 | ||
560f2c23 | 196 | if (!chunk) |
c0ebfdc3 | 197 | return false; |
020ec653 | 198 | |
560f2c23 DZF |
199 | start_addr = chunk->base_addr + chunk->start_offset; |
200 | end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE - | |
201 | chunk->end_offset; | |
c0ebfdc3 DZF |
202 | |
203 | return addr >= start_addr && addr < end_addr; | |
020ec653 TH |
204 | } |
205 | ||
d9b55eeb | 206 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 207 | { |
cae3aeb8 | 208 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
209 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
210 | } | |
211 | ||
d9b55eeb TH |
212 | static int pcpu_size_to_slot(int size) |
213 | { | |
214 | if (size == pcpu_unit_size) | |
215 | return pcpu_nr_slots - 1; | |
216 | return __pcpu_size_to_slot(size); | |
217 | } | |
218 | ||
fbf59bc9 TH |
219 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
220 | { | |
221 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
222 | return 0; | |
223 | ||
224 | return pcpu_size_to_slot(chunk->free_size); | |
225 | } | |
226 | ||
88999a89 TH |
227 | /* set the pointer to a chunk in a page struct */ |
228 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
229 | { | |
230 | page->index = (unsigned long)pcpu; | |
231 | } | |
232 | ||
233 | /* obtain pointer to a chunk from a page struct */ | |
234 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
235 | { | |
236 | return (struct pcpu_chunk *)page->index; | |
237 | } | |
238 | ||
239 | static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx) | |
fbf59bc9 | 240 | { |
2f39e637 | 241 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
242 | } |
243 | ||
c0ebfdc3 DZF |
244 | static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx) |
245 | { | |
246 | return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT); | |
247 | } | |
248 | ||
9983b6f0 TH |
249 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
250 | unsigned int cpu, int page_idx) | |
fbf59bc9 | 251 | { |
c0ebfdc3 DZF |
252 | return (unsigned long)chunk->base_addr + |
253 | pcpu_unit_page_offset(cpu, page_idx); | |
fbf59bc9 TH |
254 | } |
255 | ||
88999a89 TH |
256 | static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk, |
257 | int *rs, int *re, int end) | |
ce3141a2 TH |
258 | { |
259 | *rs = find_next_zero_bit(chunk->populated, end, *rs); | |
260 | *re = find_next_bit(chunk->populated, end, *rs + 1); | |
261 | } | |
262 | ||
88999a89 TH |
263 | static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk, |
264 | int *rs, int *re, int end) | |
ce3141a2 TH |
265 | { |
266 | *rs = find_next_bit(chunk->populated, end, *rs); | |
267 | *re = find_next_zero_bit(chunk->populated, end, *rs + 1); | |
268 | } | |
269 | ||
270 | /* | |
271 | * (Un)populated page region iterators. Iterate over (un)populated | |
b595076a | 272 | * page regions between @start and @end in @chunk. @rs and @re should |
ce3141a2 TH |
273 | * be integer variables and will be set to start and end page index of |
274 | * the current region. | |
275 | */ | |
276 | #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ | |
277 | for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ | |
278 | (rs) < (re); \ | |
279 | (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) | |
280 | ||
281 | #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ | |
282 | for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ | |
283 | (rs) < (re); \ | |
284 | (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) | |
285 | ||
fbf59bc9 | 286 | /** |
90459ce0 | 287 | * pcpu_mem_zalloc - allocate memory |
1880d93b | 288 | * @size: bytes to allocate |
fbf59bc9 | 289 | * |
1880d93b | 290 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
90459ce0 | 291 | * kzalloc() is used; otherwise, vzalloc() is used. The returned |
1880d93b | 292 | * memory is always zeroed. |
fbf59bc9 | 293 | * |
ccea34b5 TH |
294 | * CONTEXT: |
295 | * Does GFP_KERNEL allocation. | |
296 | * | |
fbf59bc9 | 297 | * RETURNS: |
1880d93b | 298 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 299 | */ |
90459ce0 | 300 | static void *pcpu_mem_zalloc(size_t size) |
fbf59bc9 | 301 | { |
099a19d9 TH |
302 | if (WARN_ON_ONCE(!slab_is_available())) |
303 | return NULL; | |
304 | ||
1880d93b TH |
305 | if (size <= PAGE_SIZE) |
306 | return kzalloc(size, GFP_KERNEL); | |
7af4c093 JJ |
307 | else |
308 | return vzalloc(size); | |
1880d93b | 309 | } |
fbf59bc9 | 310 | |
1880d93b TH |
311 | /** |
312 | * pcpu_mem_free - free memory | |
313 | * @ptr: memory to free | |
1880d93b | 314 | * |
90459ce0 | 315 | * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). |
1880d93b | 316 | */ |
1d5cfdb0 | 317 | static void pcpu_mem_free(void *ptr) |
1880d93b | 318 | { |
1d5cfdb0 | 319 | kvfree(ptr); |
fbf59bc9 TH |
320 | } |
321 | ||
b539b87f TH |
322 | /** |
323 | * pcpu_count_occupied_pages - count the number of pages an area occupies | |
324 | * @chunk: chunk of interest | |
325 | * @i: index of the area in question | |
326 | * | |
327 | * Count the number of pages chunk's @i'th area occupies. When the area's | |
328 | * start and/or end address isn't aligned to page boundary, the straddled | |
329 | * page is included in the count iff the rest of the page is free. | |
330 | */ | |
331 | static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i) | |
332 | { | |
333 | int off = chunk->map[i] & ~1; | |
334 | int end = chunk->map[i + 1] & ~1; | |
335 | ||
336 | if (!PAGE_ALIGNED(off) && i > 0) { | |
337 | int prev = chunk->map[i - 1]; | |
338 | ||
339 | if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE)) | |
340 | off = round_down(off, PAGE_SIZE); | |
341 | } | |
342 | ||
343 | if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) { | |
344 | int next = chunk->map[i + 1]; | |
345 | int nend = chunk->map[i + 2] & ~1; | |
346 | ||
347 | if (!(next & 1) && nend >= round_up(end, PAGE_SIZE)) | |
348 | end = round_up(end, PAGE_SIZE); | |
349 | } | |
350 | ||
351 | return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0); | |
352 | } | |
353 | ||
fbf59bc9 TH |
354 | /** |
355 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
356 | * @chunk: chunk of interest | |
357 | * @oslot: the previous slot it was on | |
358 | * | |
359 | * This function is called after an allocation or free changed @chunk. | |
360 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
361 | * moved to the slot. Note that the reserved chunk is never put on |
362 | * chunk slots. | |
ccea34b5 TH |
363 | * |
364 | * CONTEXT: | |
365 | * pcpu_lock. | |
fbf59bc9 TH |
366 | */ |
367 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
368 | { | |
369 | int nslot = pcpu_chunk_slot(chunk); | |
370 | ||
edcb4639 | 371 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
372 | if (oslot < nslot) |
373 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
374 | else | |
375 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
376 | } | |
377 | } | |
378 | ||
9f7dcf22 | 379 | /** |
833af842 TH |
380 | * pcpu_need_to_extend - determine whether chunk area map needs to be extended |
381 | * @chunk: chunk of interest | |
9c824b6a | 382 | * @is_atomic: the allocation context |
9f7dcf22 | 383 | * |
9c824b6a TH |
384 | * Determine whether area map of @chunk needs to be extended. If |
385 | * @is_atomic, only the amount necessary for a new allocation is | |
386 | * considered; however, async extension is scheduled if the left amount is | |
387 | * low. If !@is_atomic, it aims for more empty space. Combined, this | |
388 | * ensures that the map is likely to have enough available space to | |
389 | * accomodate atomic allocations which can't extend maps directly. | |
9f7dcf22 | 390 | * |
ccea34b5 | 391 | * CONTEXT: |
833af842 | 392 | * pcpu_lock. |
ccea34b5 | 393 | * |
9f7dcf22 | 394 | * RETURNS: |
833af842 TH |
395 | * New target map allocation length if extension is necessary, 0 |
396 | * otherwise. | |
9f7dcf22 | 397 | */ |
9c824b6a | 398 | static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic) |
9f7dcf22 | 399 | { |
9c824b6a TH |
400 | int margin, new_alloc; |
401 | ||
4f996e23 TH |
402 | lockdep_assert_held(&pcpu_lock); |
403 | ||
9c824b6a TH |
404 | if (is_atomic) { |
405 | margin = 3; | |
9f7dcf22 | 406 | |
9c824b6a | 407 | if (chunk->map_alloc < |
4f996e23 TH |
408 | chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) { |
409 | if (list_empty(&chunk->map_extend_list)) { | |
410 | list_add_tail(&chunk->map_extend_list, | |
411 | &pcpu_map_extend_chunks); | |
412 | pcpu_schedule_balance_work(); | |
413 | } | |
414 | } | |
9c824b6a TH |
415 | } else { |
416 | margin = PCPU_ATOMIC_MAP_MARGIN_HIGH; | |
417 | } | |
418 | ||
419 | if (chunk->map_alloc >= chunk->map_used + margin) | |
9f7dcf22 TH |
420 | return 0; |
421 | ||
422 | new_alloc = PCPU_DFL_MAP_ALLOC; | |
9c824b6a | 423 | while (new_alloc < chunk->map_used + margin) |
9f7dcf22 TH |
424 | new_alloc *= 2; |
425 | ||
833af842 TH |
426 | return new_alloc; |
427 | } | |
428 | ||
429 | /** | |
430 | * pcpu_extend_area_map - extend area map of a chunk | |
431 | * @chunk: chunk of interest | |
432 | * @new_alloc: new target allocation length of the area map | |
433 | * | |
434 | * Extend area map of @chunk to have @new_alloc entries. | |
435 | * | |
436 | * CONTEXT: | |
437 | * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. | |
438 | * | |
439 | * RETURNS: | |
440 | * 0 on success, -errno on failure. | |
441 | */ | |
442 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) | |
443 | { | |
444 | int *old = NULL, *new = NULL; | |
445 | size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); | |
446 | unsigned long flags; | |
447 | ||
6710e594 TH |
448 | lockdep_assert_held(&pcpu_alloc_mutex); |
449 | ||
90459ce0 | 450 | new = pcpu_mem_zalloc(new_size); |
833af842 | 451 | if (!new) |
9f7dcf22 | 452 | return -ENOMEM; |
ccea34b5 | 453 | |
833af842 TH |
454 | /* acquire pcpu_lock and switch to new area map */ |
455 | spin_lock_irqsave(&pcpu_lock, flags); | |
456 | ||
457 | if (new_alloc <= chunk->map_alloc) | |
458 | goto out_unlock; | |
9f7dcf22 | 459 | |
833af842 | 460 | old_size = chunk->map_alloc * sizeof(chunk->map[0]); |
a002d148 HS |
461 | old = chunk->map; |
462 | ||
463 | memcpy(new, old, old_size); | |
9f7dcf22 | 464 | |
9f7dcf22 TH |
465 | chunk->map_alloc = new_alloc; |
466 | chunk->map = new; | |
833af842 TH |
467 | new = NULL; |
468 | ||
469 | out_unlock: | |
470 | spin_unlock_irqrestore(&pcpu_lock, flags); | |
471 | ||
472 | /* | |
473 | * pcpu_mem_free() might end up calling vfree() which uses | |
474 | * IRQ-unsafe lock and thus can't be called under pcpu_lock. | |
475 | */ | |
1d5cfdb0 TH |
476 | pcpu_mem_free(old); |
477 | pcpu_mem_free(new); | |
833af842 | 478 | |
9f7dcf22 TH |
479 | return 0; |
480 | } | |
481 | ||
a16037c8 TH |
482 | /** |
483 | * pcpu_fit_in_area - try to fit the requested allocation in a candidate area | |
484 | * @chunk: chunk the candidate area belongs to | |
485 | * @off: the offset to the start of the candidate area | |
486 | * @this_size: the size of the candidate area | |
487 | * @size: the size of the target allocation | |
488 | * @align: the alignment of the target allocation | |
489 | * @pop_only: only allocate from already populated region | |
490 | * | |
491 | * We're trying to allocate @size bytes aligned at @align. @chunk's area | |
492 | * at @off sized @this_size is a candidate. This function determines | |
493 | * whether the target allocation fits in the candidate area and returns the | |
494 | * number of bytes to pad after @off. If the target area doesn't fit, -1 | |
495 | * is returned. | |
496 | * | |
497 | * If @pop_only is %true, this function only considers the already | |
498 | * populated part of the candidate area. | |
499 | */ | |
500 | static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size, | |
501 | int size, int align, bool pop_only) | |
502 | { | |
503 | int cand_off = off; | |
504 | ||
505 | while (true) { | |
506 | int head = ALIGN(cand_off, align) - off; | |
507 | int page_start, page_end, rs, re; | |
508 | ||
509 | if (this_size < head + size) | |
510 | return -1; | |
511 | ||
512 | if (!pop_only) | |
513 | return head; | |
514 | ||
515 | /* | |
516 | * If the first unpopulated page is beyond the end of the | |
517 | * allocation, the whole allocation is populated; | |
518 | * otherwise, retry from the end of the unpopulated area. | |
519 | */ | |
520 | page_start = PFN_DOWN(head + off); | |
521 | page_end = PFN_UP(head + off + size); | |
522 | ||
523 | rs = page_start; | |
524 | pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size)); | |
525 | if (rs >= page_end) | |
526 | return head; | |
527 | cand_off = re * PAGE_SIZE; | |
528 | } | |
529 | } | |
530 | ||
fbf59bc9 TH |
531 | /** |
532 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
533 | * @chunk: chunk of interest | |
cae3aeb8 | 534 | * @size: wanted size in bytes |
fbf59bc9 | 535 | * @align: wanted align |
a16037c8 | 536 | * @pop_only: allocate only from the populated area |
b539b87f | 537 | * @occ_pages_p: out param for the number of pages the area occupies |
fbf59bc9 TH |
538 | * |
539 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
540 | * Note that this function only allocates the offset. It doesn't | |
541 | * populate or map the area. | |
542 | * | |
9f7dcf22 TH |
543 | * @chunk->map must have at least two free slots. |
544 | * | |
ccea34b5 TH |
545 | * CONTEXT: |
546 | * pcpu_lock. | |
547 | * | |
fbf59bc9 | 548 | * RETURNS: |
9f7dcf22 TH |
549 | * Allocated offset in @chunk on success, -1 if no matching area is |
550 | * found. | |
fbf59bc9 | 551 | */ |
a16037c8 | 552 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align, |
b539b87f | 553 | bool pop_only, int *occ_pages_p) |
fbf59bc9 TH |
554 | { |
555 | int oslot = pcpu_chunk_slot(chunk); | |
556 | int max_contig = 0; | |
557 | int i, off; | |
3d331ad7 | 558 | bool seen_free = false; |
723ad1d9 | 559 | int *p; |
fbf59bc9 | 560 | |
3d331ad7 | 561 | for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) { |
fbf59bc9 | 562 | int head, tail; |
723ad1d9 AV |
563 | int this_size; |
564 | ||
565 | off = *p; | |
566 | if (off & 1) | |
567 | continue; | |
fbf59bc9 | 568 | |
723ad1d9 | 569 | this_size = (p[1] & ~1) - off; |
a16037c8 TH |
570 | |
571 | head = pcpu_fit_in_area(chunk, off, this_size, size, align, | |
572 | pop_only); | |
573 | if (head < 0) { | |
3d331ad7 AV |
574 | if (!seen_free) { |
575 | chunk->first_free = i; | |
576 | seen_free = true; | |
577 | } | |
723ad1d9 | 578 | max_contig = max(this_size, max_contig); |
fbf59bc9 TH |
579 | continue; |
580 | } | |
581 | ||
582 | /* | |
583 | * If head is small or the previous block is free, | |
584 | * merge'em. Note that 'small' is defined as smaller | |
585 | * than sizeof(int), which is very small but isn't too | |
586 | * uncommon for percpu allocations. | |
587 | */ | |
723ad1d9 | 588 | if (head && (head < sizeof(int) || !(p[-1] & 1))) { |
21ddfd38 | 589 | *p = off += head; |
723ad1d9 | 590 | if (p[-1] & 1) |
fbf59bc9 | 591 | chunk->free_size -= head; |
21ddfd38 JZ |
592 | else |
593 | max_contig = max(*p - p[-1], max_contig); | |
723ad1d9 | 594 | this_size -= head; |
fbf59bc9 TH |
595 | head = 0; |
596 | } | |
597 | ||
598 | /* if tail is small, just keep it around */ | |
723ad1d9 AV |
599 | tail = this_size - head - size; |
600 | if (tail < sizeof(int)) { | |
fbf59bc9 | 601 | tail = 0; |
723ad1d9 AV |
602 | size = this_size - head; |
603 | } | |
fbf59bc9 TH |
604 | |
605 | /* split if warranted */ | |
606 | if (head || tail) { | |
706c16f2 AV |
607 | int nr_extra = !!head + !!tail; |
608 | ||
609 | /* insert new subblocks */ | |
723ad1d9 | 610 | memmove(p + nr_extra + 1, p + 1, |
706c16f2 AV |
611 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
612 | chunk->map_used += nr_extra; | |
613 | ||
fbf59bc9 | 614 | if (head) { |
3d331ad7 AV |
615 | if (!seen_free) { |
616 | chunk->first_free = i; | |
617 | seen_free = true; | |
618 | } | |
723ad1d9 AV |
619 | *++p = off += head; |
620 | ++i; | |
706c16f2 AV |
621 | max_contig = max(head, max_contig); |
622 | } | |
623 | if (tail) { | |
723ad1d9 | 624 | p[1] = off + size; |
706c16f2 | 625 | max_contig = max(tail, max_contig); |
fbf59bc9 | 626 | } |
fbf59bc9 TH |
627 | } |
628 | ||
3d331ad7 AV |
629 | if (!seen_free) |
630 | chunk->first_free = i + 1; | |
631 | ||
fbf59bc9 | 632 | /* update hint and mark allocated */ |
723ad1d9 | 633 | if (i + 1 == chunk->map_used) |
fbf59bc9 TH |
634 | chunk->contig_hint = max_contig; /* fully scanned */ |
635 | else | |
636 | chunk->contig_hint = max(chunk->contig_hint, | |
637 | max_contig); | |
638 | ||
723ad1d9 AV |
639 | chunk->free_size -= size; |
640 | *p |= 1; | |
fbf59bc9 | 641 | |
b539b87f | 642 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
fbf59bc9 TH |
643 | pcpu_chunk_relocate(chunk, oslot); |
644 | return off; | |
645 | } | |
646 | ||
647 | chunk->contig_hint = max_contig; /* fully scanned */ | |
648 | pcpu_chunk_relocate(chunk, oslot); | |
649 | ||
9f7dcf22 TH |
650 | /* tell the upper layer that this chunk has no matching area */ |
651 | return -1; | |
fbf59bc9 TH |
652 | } |
653 | ||
654 | /** | |
655 | * pcpu_free_area - free area to a pcpu_chunk | |
656 | * @chunk: chunk of interest | |
657 | * @freeme: offset of area to free | |
b539b87f | 658 | * @occ_pages_p: out param for the number of pages the area occupies |
fbf59bc9 TH |
659 | * |
660 | * Free area starting from @freeme to @chunk. Note that this function | |
661 | * only modifies the allocation map. It doesn't depopulate or unmap | |
662 | * the area. | |
ccea34b5 TH |
663 | * |
664 | * CONTEXT: | |
665 | * pcpu_lock. | |
fbf59bc9 | 666 | */ |
b539b87f TH |
667 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme, |
668 | int *occ_pages_p) | |
fbf59bc9 TH |
669 | { |
670 | int oslot = pcpu_chunk_slot(chunk); | |
723ad1d9 AV |
671 | int off = 0; |
672 | unsigned i, j; | |
673 | int to_free = 0; | |
674 | int *p; | |
675 | ||
5ccd30e4 | 676 | lockdep_assert_held(&pcpu_lock); |
30a5b536 | 677 | pcpu_stats_area_dealloc(chunk); |
5ccd30e4 | 678 | |
723ad1d9 AV |
679 | freeme |= 1; /* we are searching for <given offset, in use> pair */ |
680 | ||
681 | i = 0; | |
682 | j = chunk->map_used; | |
683 | while (i != j) { | |
684 | unsigned k = (i + j) / 2; | |
685 | off = chunk->map[k]; | |
686 | if (off < freeme) | |
687 | i = k + 1; | |
688 | else if (off > freeme) | |
689 | j = k; | |
690 | else | |
691 | i = j = k; | |
692 | } | |
fbf59bc9 | 693 | BUG_ON(off != freeme); |
fbf59bc9 | 694 | |
3d331ad7 AV |
695 | if (i < chunk->first_free) |
696 | chunk->first_free = i; | |
697 | ||
723ad1d9 AV |
698 | p = chunk->map + i; |
699 | *p = off &= ~1; | |
700 | chunk->free_size += (p[1] & ~1) - off; | |
fbf59bc9 | 701 | |
b539b87f TH |
702 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
703 | ||
723ad1d9 AV |
704 | /* merge with next? */ |
705 | if (!(p[1] & 1)) | |
706 | to_free++; | |
fbf59bc9 | 707 | /* merge with previous? */ |
723ad1d9 AV |
708 | if (i > 0 && !(p[-1] & 1)) { |
709 | to_free++; | |
fbf59bc9 | 710 | i--; |
723ad1d9 | 711 | p--; |
fbf59bc9 | 712 | } |
723ad1d9 AV |
713 | if (to_free) { |
714 | chunk->map_used -= to_free; | |
715 | memmove(p + 1, p + 1 + to_free, | |
716 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
fbf59bc9 TH |
717 | } |
718 | ||
723ad1d9 | 719 | chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint); |
fbf59bc9 TH |
720 | pcpu_chunk_relocate(chunk, oslot); |
721 | } | |
722 | ||
c0ebfdc3 | 723 | static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, |
10edf5b0 DZF |
724 | int map_size, |
725 | int *map, | |
726 | int init_map_size) | |
727 | { | |
728 | struct pcpu_chunk *chunk; | |
c0ebfdc3 DZF |
729 | unsigned long aligned_addr; |
730 | int start_offset, region_size; | |
731 | ||
732 | /* region calculations */ | |
733 | aligned_addr = tmp_addr & PAGE_MASK; | |
734 | ||
735 | start_offset = tmp_addr - aligned_addr; | |
6b9d7c8e DZF |
736 | |
737 | region_size = PFN_ALIGN(start_offset + map_size); | |
10edf5b0 | 738 | |
c0ebfdc3 | 739 | /* allocate chunk */ |
10edf5b0 | 740 | chunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0); |
c0ebfdc3 | 741 | |
10edf5b0 DZF |
742 | INIT_LIST_HEAD(&chunk->list); |
743 | INIT_LIST_HEAD(&chunk->map_extend_list); | |
c0ebfdc3 DZF |
744 | |
745 | chunk->base_addr = (void *)aligned_addr; | |
10edf5b0 | 746 | chunk->start_offset = start_offset; |
6b9d7c8e | 747 | chunk->end_offset = region_size - chunk->start_offset - map_size; |
c0ebfdc3 DZF |
748 | |
749 | chunk->nr_pages = pcpu_unit_pages; | |
750 | ||
10edf5b0 DZF |
751 | chunk->map = map; |
752 | chunk->map_alloc = init_map_size; | |
753 | ||
754 | /* manage populated page bitmap */ | |
755 | chunk->immutable = true; | |
756 | bitmap_fill(chunk->populated, pcpu_unit_pages); | |
757 | chunk->nr_populated = pcpu_unit_pages; | |
758 | ||
759 | chunk->contig_hint = chunk->free_size = map_size; | |
c0ebfdc3 DZF |
760 | |
761 | if (chunk->start_offset) { | |
762 | /* hide the beginning of the bitmap */ | |
763 | chunk->map[0] = 1; | |
764 | chunk->map[1] = chunk->start_offset; | |
765 | chunk->map_used = 1; | |
766 | } | |
767 | ||
768 | /* set chunk's free region */ | |
769 | chunk->map[++chunk->map_used] = | |
770 | (chunk->start_offset + chunk->free_size) | 1; | |
10edf5b0 | 771 | |
6b9d7c8e DZF |
772 | if (chunk->end_offset) { |
773 | /* hide the end of the bitmap */ | |
774 | chunk->map[++chunk->map_used] = region_size | 1; | |
775 | } | |
776 | ||
10edf5b0 DZF |
777 | return chunk; |
778 | } | |
779 | ||
6081089f TH |
780 | static struct pcpu_chunk *pcpu_alloc_chunk(void) |
781 | { | |
782 | struct pcpu_chunk *chunk; | |
783 | ||
90459ce0 | 784 | chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); |
6081089f TH |
785 | if (!chunk) |
786 | return NULL; | |
787 | ||
90459ce0 BL |
788 | chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC * |
789 | sizeof(chunk->map[0])); | |
6081089f | 790 | if (!chunk->map) { |
1d5cfdb0 | 791 | pcpu_mem_free(chunk); |
6081089f TH |
792 | return NULL; |
793 | } | |
794 | ||
795 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; | |
723ad1d9 AV |
796 | chunk->map[0] = 0; |
797 | chunk->map[1] = pcpu_unit_size | 1; | |
798 | chunk->map_used = 1; | |
6081089f TH |
799 | |
800 | INIT_LIST_HEAD(&chunk->list); | |
4f996e23 | 801 | INIT_LIST_HEAD(&chunk->map_extend_list); |
6081089f TH |
802 | chunk->free_size = pcpu_unit_size; |
803 | chunk->contig_hint = pcpu_unit_size; | |
804 | ||
c0ebfdc3 DZF |
805 | chunk->nr_pages = pcpu_unit_pages; |
806 | ||
6081089f TH |
807 | return chunk; |
808 | } | |
809 | ||
810 | static void pcpu_free_chunk(struct pcpu_chunk *chunk) | |
811 | { | |
812 | if (!chunk) | |
813 | return; | |
1d5cfdb0 TH |
814 | pcpu_mem_free(chunk->map); |
815 | pcpu_mem_free(chunk); | |
6081089f TH |
816 | } |
817 | ||
b539b87f TH |
818 | /** |
819 | * pcpu_chunk_populated - post-population bookkeeping | |
820 | * @chunk: pcpu_chunk which got populated | |
821 | * @page_start: the start page | |
822 | * @page_end: the end page | |
823 | * | |
824 | * Pages in [@page_start,@page_end) have been populated to @chunk. Update | |
825 | * the bookkeeping information accordingly. Must be called after each | |
826 | * successful population. | |
827 | */ | |
828 | static void pcpu_chunk_populated(struct pcpu_chunk *chunk, | |
829 | int page_start, int page_end) | |
830 | { | |
831 | int nr = page_end - page_start; | |
832 | ||
833 | lockdep_assert_held(&pcpu_lock); | |
834 | ||
835 | bitmap_set(chunk->populated, page_start, nr); | |
836 | chunk->nr_populated += nr; | |
837 | pcpu_nr_empty_pop_pages += nr; | |
838 | } | |
839 | ||
840 | /** | |
841 | * pcpu_chunk_depopulated - post-depopulation bookkeeping | |
842 | * @chunk: pcpu_chunk which got depopulated | |
843 | * @page_start: the start page | |
844 | * @page_end: the end page | |
845 | * | |
846 | * Pages in [@page_start,@page_end) have been depopulated from @chunk. | |
847 | * Update the bookkeeping information accordingly. Must be called after | |
848 | * each successful depopulation. | |
849 | */ | |
850 | static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, | |
851 | int page_start, int page_end) | |
852 | { | |
853 | int nr = page_end - page_start; | |
854 | ||
855 | lockdep_assert_held(&pcpu_lock); | |
856 | ||
857 | bitmap_clear(chunk->populated, page_start, nr); | |
858 | chunk->nr_populated -= nr; | |
859 | pcpu_nr_empty_pop_pages -= nr; | |
860 | } | |
861 | ||
9f645532 TH |
862 | /* |
863 | * Chunk management implementation. | |
864 | * | |
865 | * To allow different implementations, chunk alloc/free and | |
866 | * [de]population are implemented in a separate file which is pulled | |
867 | * into this file and compiled together. The following functions | |
868 | * should be implemented. | |
869 | * | |
870 | * pcpu_populate_chunk - populate the specified range of a chunk | |
871 | * pcpu_depopulate_chunk - depopulate the specified range of a chunk | |
872 | * pcpu_create_chunk - create a new chunk | |
873 | * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop | |
874 | * pcpu_addr_to_page - translate address to physical address | |
875 | * pcpu_verify_alloc_info - check alloc_info is acceptable during init | |
fbf59bc9 | 876 | */ |
9f645532 TH |
877 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); |
878 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); | |
879 | static struct pcpu_chunk *pcpu_create_chunk(void); | |
880 | static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); | |
881 | static struct page *pcpu_addr_to_page(void *addr); | |
882 | static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); | |
fbf59bc9 | 883 | |
b0c9778b TH |
884 | #ifdef CONFIG_NEED_PER_CPU_KM |
885 | #include "percpu-km.c" | |
886 | #else | |
9f645532 | 887 | #include "percpu-vm.c" |
b0c9778b | 888 | #endif |
fbf59bc9 | 889 | |
88999a89 TH |
890 | /** |
891 | * pcpu_chunk_addr_search - determine chunk containing specified address | |
892 | * @addr: address for which the chunk needs to be determined. | |
893 | * | |
c0ebfdc3 DZF |
894 | * This is an internal function that handles all but static allocations. |
895 | * Static percpu address values should never be passed into the allocator. | |
896 | * | |
88999a89 TH |
897 | * RETURNS: |
898 | * The address of the found chunk. | |
899 | */ | |
900 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
901 | { | |
c0ebfdc3 | 902 | /* is it in the dynamic region (first chunk)? */ |
560f2c23 | 903 | if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) |
88999a89 | 904 | return pcpu_first_chunk; |
c0ebfdc3 DZF |
905 | |
906 | /* is it in the reserved region? */ | |
560f2c23 | 907 | if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) |
c0ebfdc3 | 908 | return pcpu_reserved_chunk; |
88999a89 TH |
909 | |
910 | /* | |
911 | * The address is relative to unit0 which might be unused and | |
912 | * thus unmapped. Offset the address to the unit space of the | |
913 | * current processor before looking it up in the vmalloc | |
914 | * space. Note that any possible cpu id can be used here, so | |
915 | * there's no need to worry about preemption or cpu hotplug. | |
916 | */ | |
917 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; | |
9f645532 | 918 | return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); |
88999a89 TH |
919 | } |
920 | ||
fbf59bc9 | 921 | /** |
edcb4639 | 922 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 923 | * @size: size of area to allocate in bytes |
fbf59bc9 | 924 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 925 | * @reserved: allocate from the reserved chunk if available |
5835d96e | 926 | * @gfp: allocation flags |
fbf59bc9 | 927 | * |
5835d96e TH |
928 | * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't |
929 | * contain %GFP_KERNEL, the allocation is atomic. | |
fbf59bc9 TH |
930 | * |
931 | * RETURNS: | |
932 | * Percpu pointer to the allocated area on success, NULL on failure. | |
933 | */ | |
5835d96e TH |
934 | static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, |
935 | gfp_t gfp) | |
fbf59bc9 | 936 | { |
f2badb0c | 937 | static int warn_limit = 10; |
fbf59bc9 | 938 | struct pcpu_chunk *chunk; |
f2badb0c | 939 | const char *err; |
6ae833c7 | 940 | bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; |
b539b87f | 941 | int occ_pages = 0; |
b38d08f3 | 942 | int slot, off, new_alloc, cpu, ret; |
403a91b1 | 943 | unsigned long flags; |
f528f0b8 | 944 | void __percpu *ptr; |
fbf59bc9 | 945 | |
723ad1d9 AV |
946 | /* |
947 | * We want the lowest bit of offset available for in-use/free | |
2f69fa82 | 948 | * indicator, so force >= 16bit alignment and make size even. |
723ad1d9 AV |
949 | */ |
950 | if (unlikely(align < 2)) | |
951 | align = 2; | |
952 | ||
fb009e3a | 953 | size = ALIGN(size, 2); |
2f69fa82 | 954 | |
3ca45a46 | 955 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || |
956 | !is_power_of_2(align))) { | |
756a025f JP |
957 | WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n", |
958 | size, align); | |
fbf59bc9 TH |
959 | return NULL; |
960 | } | |
961 | ||
6710e594 TH |
962 | if (!is_atomic) |
963 | mutex_lock(&pcpu_alloc_mutex); | |
964 | ||
403a91b1 | 965 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 | 966 | |
edcb4639 TH |
967 | /* serve reserved allocations from the reserved chunk if available */ |
968 | if (reserved && pcpu_reserved_chunk) { | |
969 | chunk = pcpu_reserved_chunk; | |
833af842 TH |
970 | |
971 | if (size > chunk->contig_hint) { | |
972 | err = "alloc from reserved chunk failed"; | |
ccea34b5 | 973 | goto fail_unlock; |
f2badb0c | 974 | } |
833af842 | 975 | |
9c824b6a | 976 | while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) { |
833af842 | 977 | spin_unlock_irqrestore(&pcpu_lock, flags); |
5835d96e TH |
978 | if (is_atomic || |
979 | pcpu_extend_area_map(chunk, new_alloc) < 0) { | |
833af842 | 980 | err = "failed to extend area map of reserved chunk"; |
b38d08f3 | 981 | goto fail; |
833af842 TH |
982 | } |
983 | spin_lock_irqsave(&pcpu_lock, flags); | |
984 | } | |
985 | ||
b539b87f TH |
986 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
987 | &occ_pages); | |
edcb4639 TH |
988 | if (off >= 0) |
989 | goto area_found; | |
833af842 | 990 | |
f2badb0c | 991 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 992 | goto fail_unlock; |
edcb4639 TH |
993 | } |
994 | ||
ccea34b5 | 995 | restart: |
edcb4639 | 996 | /* search through normal chunks */ |
fbf59bc9 TH |
997 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
998 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
999 | if (size > chunk->contig_hint) | |
1000 | continue; | |
ccea34b5 | 1001 | |
9c824b6a | 1002 | new_alloc = pcpu_need_to_extend(chunk, is_atomic); |
833af842 | 1003 | if (new_alloc) { |
5835d96e TH |
1004 | if (is_atomic) |
1005 | continue; | |
833af842 TH |
1006 | spin_unlock_irqrestore(&pcpu_lock, flags); |
1007 | if (pcpu_extend_area_map(chunk, | |
1008 | new_alloc) < 0) { | |
1009 | err = "failed to extend area map"; | |
b38d08f3 | 1010 | goto fail; |
833af842 TH |
1011 | } |
1012 | spin_lock_irqsave(&pcpu_lock, flags); | |
1013 | /* | |
1014 | * pcpu_lock has been dropped, need to | |
1015 | * restart cpu_slot list walking. | |
1016 | */ | |
1017 | goto restart; | |
ccea34b5 TH |
1018 | } |
1019 | ||
b539b87f TH |
1020 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
1021 | &occ_pages); | |
fbf59bc9 TH |
1022 | if (off >= 0) |
1023 | goto area_found; | |
fbf59bc9 TH |
1024 | } |
1025 | } | |
1026 | ||
403a91b1 | 1027 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1028 | |
b38d08f3 TH |
1029 | /* |
1030 | * No space left. Create a new chunk. We don't want multiple | |
1031 | * tasks to create chunks simultaneously. Serialize and create iff | |
1032 | * there's still no empty chunk after grabbing the mutex. | |
1033 | */ | |
11df02bf DZ |
1034 | if (is_atomic) { |
1035 | err = "atomic alloc failed, no space left"; | |
5835d96e | 1036 | goto fail; |
11df02bf | 1037 | } |
5835d96e | 1038 | |
b38d08f3 TH |
1039 | if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { |
1040 | chunk = pcpu_create_chunk(); | |
1041 | if (!chunk) { | |
1042 | err = "failed to allocate new chunk"; | |
1043 | goto fail; | |
1044 | } | |
1045 | ||
1046 | spin_lock_irqsave(&pcpu_lock, flags); | |
1047 | pcpu_chunk_relocate(chunk, -1); | |
1048 | } else { | |
1049 | spin_lock_irqsave(&pcpu_lock, flags); | |
f2badb0c | 1050 | } |
ccea34b5 | 1051 | |
ccea34b5 | 1052 | goto restart; |
fbf59bc9 TH |
1053 | |
1054 | area_found: | |
30a5b536 | 1055 | pcpu_stats_area_alloc(chunk, size); |
403a91b1 | 1056 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1057 | |
dca49645 | 1058 | /* populate if not all pages are already there */ |
5835d96e | 1059 | if (!is_atomic) { |
e04d3208 | 1060 | int page_start, page_end, rs, re; |
dca49645 | 1061 | |
e04d3208 TH |
1062 | page_start = PFN_DOWN(off); |
1063 | page_end = PFN_UP(off + size); | |
b38d08f3 | 1064 | |
e04d3208 TH |
1065 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
1066 | WARN_ON(chunk->immutable); | |
1067 | ||
1068 | ret = pcpu_populate_chunk(chunk, rs, re); | |
1069 | ||
1070 | spin_lock_irqsave(&pcpu_lock, flags); | |
1071 | if (ret) { | |
b539b87f | 1072 | pcpu_free_area(chunk, off, &occ_pages); |
e04d3208 TH |
1073 | err = "failed to populate"; |
1074 | goto fail_unlock; | |
1075 | } | |
b539b87f | 1076 | pcpu_chunk_populated(chunk, rs, re); |
e04d3208 | 1077 | spin_unlock_irqrestore(&pcpu_lock, flags); |
dca49645 | 1078 | } |
fbf59bc9 | 1079 | |
e04d3208 TH |
1080 | mutex_unlock(&pcpu_alloc_mutex); |
1081 | } | |
ccea34b5 | 1082 | |
320661b0 TE |
1083 | if (chunk != pcpu_reserved_chunk) { |
1084 | spin_lock_irqsave(&pcpu_lock, flags); | |
b539b87f | 1085 | pcpu_nr_empty_pop_pages -= occ_pages; |
320661b0 TE |
1086 | spin_unlock_irqrestore(&pcpu_lock, flags); |
1087 | } | |
b539b87f | 1088 | |
1a4d7607 TH |
1089 | if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) |
1090 | pcpu_schedule_balance_work(); | |
1091 | ||
dca49645 TH |
1092 | /* clear the areas and return address relative to base address */ |
1093 | for_each_possible_cpu(cpu) | |
1094 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); | |
1095 | ||
f528f0b8 | 1096 | ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); |
8a8c35fa | 1097 | kmemleak_alloc_percpu(ptr, size, gfp); |
df95e795 DZ |
1098 | |
1099 | trace_percpu_alloc_percpu(reserved, is_atomic, size, align, | |
1100 | chunk->base_addr, off, ptr); | |
1101 | ||
f528f0b8 | 1102 | return ptr; |
ccea34b5 TH |
1103 | |
1104 | fail_unlock: | |
403a91b1 | 1105 | spin_unlock_irqrestore(&pcpu_lock, flags); |
b38d08f3 | 1106 | fail: |
df95e795 DZ |
1107 | trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); |
1108 | ||
5835d96e | 1109 | if (!is_atomic && warn_limit) { |
870d4b12 | 1110 | pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", |
598d8091 | 1111 | size, align, is_atomic, err); |
f2badb0c TH |
1112 | dump_stack(); |
1113 | if (!--warn_limit) | |
870d4b12 | 1114 | pr_info("limit reached, disable warning\n"); |
f2badb0c | 1115 | } |
1a4d7607 TH |
1116 | if (is_atomic) { |
1117 | /* see the flag handling in pcpu_blance_workfn() */ | |
1118 | pcpu_atomic_alloc_failed = true; | |
1119 | pcpu_schedule_balance_work(); | |
6710e594 TH |
1120 | } else { |
1121 | mutex_unlock(&pcpu_alloc_mutex); | |
1a4d7607 | 1122 | } |
ccea34b5 | 1123 | return NULL; |
fbf59bc9 | 1124 | } |
edcb4639 TH |
1125 | |
1126 | /** | |
5835d96e | 1127 | * __alloc_percpu_gfp - allocate dynamic percpu area |
edcb4639 TH |
1128 | * @size: size of area to allocate in bytes |
1129 | * @align: alignment of area (max PAGE_SIZE) | |
5835d96e | 1130 | * @gfp: allocation flags |
edcb4639 | 1131 | * |
5835d96e TH |
1132 | * Allocate zero-filled percpu area of @size bytes aligned at @align. If |
1133 | * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can | |
1134 | * be called from any context but is a lot more likely to fail. | |
ccea34b5 | 1135 | * |
edcb4639 TH |
1136 | * RETURNS: |
1137 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1138 | */ | |
5835d96e TH |
1139 | void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) |
1140 | { | |
1141 | return pcpu_alloc(size, align, false, gfp); | |
1142 | } | |
1143 | EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); | |
1144 | ||
1145 | /** | |
1146 | * __alloc_percpu - allocate dynamic percpu area | |
1147 | * @size: size of area to allocate in bytes | |
1148 | * @align: alignment of area (max PAGE_SIZE) | |
1149 | * | |
1150 | * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). | |
1151 | */ | |
43cf38eb | 1152 | void __percpu *__alloc_percpu(size_t size, size_t align) |
edcb4639 | 1153 | { |
5835d96e | 1154 | return pcpu_alloc(size, align, false, GFP_KERNEL); |
edcb4639 | 1155 | } |
fbf59bc9 TH |
1156 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1157 | ||
edcb4639 TH |
1158 | /** |
1159 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1160 | * @size: size of area to allocate in bytes | |
1161 | * @align: alignment of area (max PAGE_SIZE) | |
1162 | * | |
9329ba97 TH |
1163 | * Allocate zero-filled percpu area of @size bytes aligned at @align |
1164 | * from reserved percpu area if arch has set it up; otherwise, | |
1165 | * allocation is served from the same dynamic area. Might sleep. | |
1166 | * Might trigger writeouts. | |
edcb4639 | 1167 | * |
ccea34b5 TH |
1168 | * CONTEXT: |
1169 | * Does GFP_KERNEL allocation. | |
1170 | * | |
edcb4639 TH |
1171 | * RETURNS: |
1172 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1173 | */ | |
43cf38eb | 1174 | void __percpu *__alloc_reserved_percpu(size_t size, size_t align) |
edcb4639 | 1175 | { |
5835d96e | 1176 | return pcpu_alloc(size, align, true, GFP_KERNEL); |
edcb4639 TH |
1177 | } |
1178 | ||
a56dbddf | 1179 | /** |
1a4d7607 | 1180 | * pcpu_balance_workfn - manage the amount of free chunks and populated pages |
a56dbddf TH |
1181 | * @work: unused |
1182 | * | |
1183 | * Reclaim all fully free chunks except for the first one. | |
1184 | */ | |
fe6bd8c3 | 1185 | static void pcpu_balance_workfn(struct work_struct *work) |
fbf59bc9 | 1186 | { |
fe6bd8c3 TH |
1187 | LIST_HEAD(to_free); |
1188 | struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; | |
a56dbddf | 1189 | struct pcpu_chunk *chunk, *next; |
1a4d7607 | 1190 | int slot, nr_to_pop, ret; |
a56dbddf | 1191 | |
1a4d7607 TH |
1192 | /* |
1193 | * There's no reason to keep around multiple unused chunks and VM | |
1194 | * areas can be scarce. Destroy all free chunks except for one. | |
1195 | */ | |
ccea34b5 TH |
1196 | mutex_lock(&pcpu_alloc_mutex); |
1197 | spin_lock_irq(&pcpu_lock); | |
a56dbddf | 1198 | |
fe6bd8c3 | 1199 | list_for_each_entry_safe(chunk, next, free_head, list) { |
a56dbddf TH |
1200 | WARN_ON(chunk->immutable); |
1201 | ||
1202 | /* spare the first one */ | |
fe6bd8c3 | 1203 | if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) |
a56dbddf TH |
1204 | continue; |
1205 | ||
4f996e23 | 1206 | list_del_init(&chunk->map_extend_list); |
fe6bd8c3 | 1207 | list_move(&chunk->list, &to_free); |
a56dbddf TH |
1208 | } |
1209 | ||
ccea34b5 | 1210 | spin_unlock_irq(&pcpu_lock); |
a56dbddf | 1211 | |
fe6bd8c3 | 1212 | list_for_each_entry_safe(chunk, next, &to_free, list) { |
a93ace48 | 1213 | int rs, re; |
dca49645 | 1214 | |
a93ace48 TH |
1215 | pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) { |
1216 | pcpu_depopulate_chunk(chunk, rs, re); | |
b539b87f TH |
1217 | spin_lock_irq(&pcpu_lock); |
1218 | pcpu_chunk_depopulated(chunk, rs, re); | |
1219 | spin_unlock_irq(&pcpu_lock); | |
a93ace48 | 1220 | } |
6081089f | 1221 | pcpu_destroy_chunk(chunk); |
a56dbddf | 1222 | } |
971f3918 | 1223 | |
4f996e23 TH |
1224 | /* service chunks which requested async area map extension */ |
1225 | do { | |
1226 | int new_alloc = 0; | |
1227 | ||
1228 | spin_lock_irq(&pcpu_lock); | |
1229 | ||
1230 | chunk = list_first_entry_or_null(&pcpu_map_extend_chunks, | |
1231 | struct pcpu_chunk, map_extend_list); | |
1232 | if (chunk) { | |
1233 | list_del_init(&chunk->map_extend_list); | |
1234 | new_alloc = pcpu_need_to_extend(chunk, false); | |
1235 | } | |
1236 | ||
1237 | spin_unlock_irq(&pcpu_lock); | |
1238 | ||
1239 | if (new_alloc) | |
1240 | pcpu_extend_area_map(chunk, new_alloc); | |
1241 | } while (chunk); | |
1242 | ||
1a4d7607 TH |
1243 | /* |
1244 | * Ensure there are certain number of free populated pages for | |
1245 | * atomic allocs. Fill up from the most packed so that atomic | |
1246 | * allocs don't increase fragmentation. If atomic allocation | |
1247 | * failed previously, always populate the maximum amount. This | |
1248 | * should prevent atomic allocs larger than PAGE_SIZE from keeping | |
1249 | * failing indefinitely; however, large atomic allocs are not | |
1250 | * something we support properly and can be highly unreliable and | |
1251 | * inefficient. | |
1252 | */ | |
1253 | retry_pop: | |
1254 | if (pcpu_atomic_alloc_failed) { | |
1255 | nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; | |
1256 | /* best effort anyway, don't worry about synchronization */ | |
1257 | pcpu_atomic_alloc_failed = false; | |
1258 | } else { | |
1259 | nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - | |
1260 | pcpu_nr_empty_pop_pages, | |
1261 | 0, PCPU_EMPTY_POP_PAGES_HIGH); | |
1262 | } | |
1263 | ||
1264 | for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { | |
1265 | int nr_unpop = 0, rs, re; | |
1266 | ||
1267 | if (!nr_to_pop) | |
1268 | break; | |
1269 | ||
1270 | spin_lock_irq(&pcpu_lock); | |
1271 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
1272 | nr_unpop = pcpu_unit_pages - chunk->nr_populated; | |
1273 | if (nr_unpop) | |
1274 | break; | |
1275 | } | |
1276 | spin_unlock_irq(&pcpu_lock); | |
1277 | ||
1278 | if (!nr_unpop) | |
1279 | continue; | |
1280 | ||
1281 | /* @chunk can't go away while pcpu_alloc_mutex is held */ | |
1282 | pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) { | |
1283 | int nr = min(re - rs, nr_to_pop); | |
1284 | ||
1285 | ret = pcpu_populate_chunk(chunk, rs, rs + nr); | |
1286 | if (!ret) { | |
1287 | nr_to_pop -= nr; | |
1288 | spin_lock_irq(&pcpu_lock); | |
1289 | pcpu_chunk_populated(chunk, rs, rs + nr); | |
1290 | spin_unlock_irq(&pcpu_lock); | |
1291 | } else { | |
1292 | nr_to_pop = 0; | |
1293 | } | |
1294 | ||
1295 | if (!nr_to_pop) | |
1296 | break; | |
1297 | } | |
1298 | } | |
1299 | ||
1300 | if (nr_to_pop) { | |
1301 | /* ran out of chunks to populate, create a new one and retry */ | |
1302 | chunk = pcpu_create_chunk(); | |
1303 | if (chunk) { | |
1304 | spin_lock_irq(&pcpu_lock); | |
1305 | pcpu_chunk_relocate(chunk, -1); | |
1306 | spin_unlock_irq(&pcpu_lock); | |
1307 | goto retry_pop; | |
1308 | } | |
1309 | } | |
1310 | ||
971f3918 | 1311 | mutex_unlock(&pcpu_alloc_mutex); |
fbf59bc9 TH |
1312 | } |
1313 | ||
1314 | /** | |
1315 | * free_percpu - free percpu area | |
1316 | * @ptr: pointer to area to free | |
1317 | * | |
ccea34b5 TH |
1318 | * Free percpu area @ptr. |
1319 | * | |
1320 | * CONTEXT: | |
1321 | * Can be called from atomic context. | |
fbf59bc9 | 1322 | */ |
43cf38eb | 1323 | void free_percpu(void __percpu *ptr) |
fbf59bc9 | 1324 | { |
129182e5 | 1325 | void *addr; |
fbf59bc9 | 1326 | struct pcpu_chunk *chunk; |
ccea34b5 | 1327 | unsigned long flags; |
b539b87f | 1328 | int off, occ_pages; |
fbf59bc9 TH |
1329 | |
1330 | if (!ptr) | |
1331 | return; | |
1332 | ||
f528f0b8 CM |
1333 | kmemleak_free_percpu(ptr); |
1334 | ||
129182e5 AM |
1335 | addr = __pcpu_ptr_to_addr(ptr); |
1336 | ||
ccea34b5 | 1337 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1338 | |
1339 | chunk = pcpu_chunk_addr_search(addr); | |
bba174f5 | 1340 | off = addr - chunk->base_addr; |
fbf59bc9 | 1341 | |
b539b87f TH |
1342 | pcpu_free_area(chunk, off, &occ_pages); |
1343 | ||
1344 | if (chunk != pcpu_reserved_chunk) | |
1345 | pcpu_nr_empty_pop_pages += occ_pages; | |
fbf59bc9 | 1346 | |
a56dbddf | 1347 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
1348 | if (chunk->free_size == pcpu_unit_size) { |
1349 | struct pcpu_chunk *pos; | |
1350 | ||
a56dbddf | 1351 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1352 | if (pos != chunk) { |
1a4d7607 | 1353 | pcpu_schedule_balance_work(); |
fbf59bc9 TH |
1354 | break; |
1355 | } | |
1356 | } | |
1357 | ||
df95e795 DZ |
1358 | trace_percpu_free_percpu(chunk->base_addr, off, ptr); |
1359 | ||
ccea34b5 | 1360 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1361 | } |
1362 | EXPORT_SYMBOL_GPL(free_percpu); | |
1363 | ||
383776fa | 1364 | bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) |
10fad5e4 | 1365 | { |
bbddff05 | 1366 | #ifdef CONFIG_SMP |
10fad5e4 TH |
1367 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
1368 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); | |
1369 | unsigned int cpu; | |
1370 | ||
1371 | for_each_possible_cpu(cpu) { | |
1372 | void *start = per_cpu_ptr(base, cpu); | |
383776fa | 1373 | void *va = (void *)addr; |
10fad5e4 | 1374 | |
383776fa | 1375 | if (va >= start && va < start + static_size) { |
8ce371f9 | 1376 | if (can_addr) { |
383776fa | 1377 | *can_addr = (unsigned long) (va - start); |
8ce371f9 PZ |
1378 | *can_addr += (unsigned long) |
1379 | per_cpu_ptr(base, get_boot_cpu_id()); | |
1380 | } | |
10fad5e4 | 1381 | return true; |
383776fa TG |
1382 | } |
1383 | } | |
bbddff05 TH |
1384 | #endif |
1385 | /* on UP, can't distinguish from other static vars, always false */ | |
10fad5e4 TH |
1386 | return false; |
1387 | } | |
1388 | ||
383776fa TG |
1389 | /** |
1390 | * is_kernel_percpu_address - test whether address is from static percpu area | |
1391 | * @addr: address to test | |
1392 | * | |
1393 | * Test whether @addr belongs to in-kernel static percpu area. Module | |
1394 | * static percpu areas are not considered. For those, use | |
1395 | * is_module_percpu_address(). | |
1396 | * | |
1397 | * RETURNS: | |
1398 | * %true if @addr is from in-kernel static percpu area, %false otherwise. | |
1399 | */ | |
1400 | bool is_kernel_percpu_address(unsigned long addr) | |
1401 | { | |
1402 | return __is_kernel_percpu_address(addr, NULL); | |
1403 | } | |
1404 | ||
3b034b0d VG |
1405 | /** |
1406 | * per_cpu_ptr_to_phys - convert translated percpu address to physical address | |
1407 | * @addr: the address to be converted to physical address | |
1408 | * | |
1409 | * Given @addr which is dereferenceable address obtained via one of | |
1410 | * percpu access macros, this function translates it into its physical | |
1411 | * address. The caller is responsible for ensuring @addr stays valid | |
1412 | * until this function finishes. | |
1413 | * | |
67589c71 DY |
1414 | * percpu allocator has special setup for the first chunk, which currently |
1415 | * supports either embedding in linear address space or vmalloc mapping, | |
1416 | * and, from the second one, the backing allocator (currently either vm or | |
1417 | * km) provides translation. | |
1418 | * | |
bffc4375 | 1419 | * The addr can be translated simply without checking if it falls into the |
67589c71 DY |
1420 | * first chunk. But the current code reflects better how percpu allocator |
1421 | * actually works, and the verification can discover both bugs in percpu | |
1422 | * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current | |
1423 | * code. | |
1424 | * | |
3b034b0d VG |
1425 | * RETURNS: |
1426 | * The physical address for @addr. | |
1427 | */ | |
1428 | phys_addr_t per_cpu_ptr_to_phys(void *addr) | |
1429 | { | |
9983b6f0 TH |
1430 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
1431 | bool in_first_chunk = false; | |
a855b84c | 1432 | unsigned long first_low, first_high; |
9983b6f0 TH |
1433 | unsigned int cpu; |
1434 | ||
1435 | /* | |
a855b84c | 1436 | * The following test on unit_low/high isn't strictly |
9983b6f0 TH |
1437 | * necessary but will speed up lookups of addresses which |
1438 | * aren't in the first chunk. | |
c0ebfdc3 DZF |
1439 | * |
1440 | * The address check is against full chunk sizes. pcpu_base_addr | |
1441 | * points to the beginning of the first chunk including the | |
1442 | * static region. Assumes good intent as the first chunk may | |
1443 | * not be full (ie. < pcpu_unit_pages in size). | |
9983b6f0 | 1444 | */ |
c0ebfdc3 DZF |
1445 | first_low = (unsigned long)pcpu_base_addr + |
1446 | pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); | |
1447 | first_high = (unsigned long)pcpu_base_addr + | |
1448 | pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); | |
a855b84c TH |
1449 | if ((unsigned long)addr >= first_low && |
1450 | (unsigned long)addr < first_high) { | |
9983b6f0 TH |
1451 | for_each_possible_cpu(cpu) { |
1452 | void *start = per_cpu_ptr(base, cpu); | |
1453 | ||
1454 | if (addr >= start && addr < start + pcpu_unit_size) { | |
1455 | in_first_chunk = true; | |
1456 | break; | |
1457 | } | |
1458 | } | |
1459 | } | |
1460 | ||
1461 | if (in_first_chunk) { | |
eac522ef | 1462 | if (!is_vmalloc_addr(addr)) |
020ec653 TH |
1463 | return __pa(addr); |
1464 | else | |
9f57bd4d ES |
1465 | return page_to_phys(vmalloc_to_page(addr)) + |
1466 | offset_in_page(addr); | |
020ec653 | 1467 | } else |
9f57bd4d ES |
1468 | return page_to_phys(pcpu_addr_to_page(addr)) + |
1469 | offset_in_page(addr); | |
3b034b0d VG |
1470 | } |
1471 | ||
fbf59bc9 | 1472 | /** |
fd1e8a1f TH |
1473 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
1474 | * @nr_groups: the number of groups | |
1475 | * @nr_units: the number of units | |
1476 | * | |
1477 | * Allocate ai which is large enough for @nr_groups groups containing | |
1478 | * @nr_units units. The returned ai's groups[0].cpu_map points to the | |
1479 | * cpu_map array which is long enough for @nr_units and filled with | |
1480 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map | |
1481 | * pointer of other groups. | |
1482 | * | |
1483 | * RETURNS: | |
1484 | * Pointer to the allocated pcpu_alloc_info on success, NULL on | |
1485 | * failure. | |
1486 | */ | |
1487 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, | |
1488 | int nr_units) | |
1489 | { | |
1490 | struct pcpu_alloc_info *ai; | |
1491 | size_t base_size, ai_size; | |
1492 | void *ptr; | |
1493 | int unit; | |
1494 | ||
1495 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), | |
1496 | __alignof__(ai->groups[0].cpu_map[0])); | |
1497 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); | |
1498 | ||
999c17e3 | 1499 | ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); |
fd1e8a1f TH |
1500 | if (!ptr) |
1501 | return NULL; | |
1502 | ai = ptr; | |
1503 | ptr += base_size; | |
1504 | ||
1505 | ai->groups[0].cpu_map = ptr; | |
1506 | ||
1507 | for (unit = 0; unit < nr_units; unit++) | |
1508 | ai->groups[0].cpu_map[unit] = NR_CPUS; | |
1509 | ||
1510 | ai->nr_groups = nr_groups; | |
1511 | ai->__ai_size = PFN_ALIGN(ai_size); | |
1512 | ||
1513 | return ai; | |
1514 | } | |
1515 | ||
1516 | /** | |
1517 | * pcpu_free_alloc_info - free percpu allocation info | |
1518 | * @ai: pcpu_alloc_info to free | |
1519 | * | |
1520 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). | |
1521 | */ | |
1522 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) | |
1523 | { | |
999c17e3 | 1524 | memblock_free_early(__pa(ai), ai->__ai_size); |
fd1e8a1f TH |
1525 | } |
1526 | ||
fd1e8a1f TH |
1527 | /** |
1528 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info | |
1529 | * @lvl: loglevel | |
1530 | * @ai: allocation info to dump | |
1531 | * | |
1532 | * Print out information about @ai using loglevel @lvl. | |
1533 | */ | |
1534 | static void pcpu_dump_alloc_info(const char *lvl, | |
1535 | const struct pcpu_alloc_info *ai) | |
033e48fb | 1536 | { |
fd1e8a1f | 1537 | int group_width = 1, cpu_width = 1, width; |
033e48fb | 1538 | char empty_str[] = "--------"; |
fd1e8a1f TH |
1539 | int alloc = 0, alloc_end = 0; |
1540 | int group, v; | |
1541 | int upa, apl; /* units per alloc, allocs per line */ | |
1542 | ||
1543 | v = ai->nr_groups; | |
1544 | while (v /= 10) | |
1545 | group_width++; | |
033e48fb | 1546 | |
fd1e8a1f | 1547 | v = num_possible_cpus(); |
033e48fb | 1548 | while (v /= 10) |
fd1e8a1f TH |
1549 | cpu_width++; |
1550 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; | |
033e48fb | 1551 | |
fd1e8a1f TH |
1552 | upa = ai->alloc_size / ai->unit_size; |
1553 | width = upa * (cpu_width + 1) + group_width + 3; | |
1554 | apl = rounddown_pow_of_two(max(60 / width, 1)); | |
033e48fb | 1555 | |
fd1e8a1f TH |
1556 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
1557 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, | |
1558 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); | |
033e48fb | 1559 | |
fd1e8a1f TH |
1560 | for (group = 0; group < ai->nr_groups; group++) { |
1561 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
1562 | int unit = 0, unit_end = 0; | |
1563 | ||
1564 | BUG_ON(gi->nr_units % upa); | |
1565 | for (alloc_end += gi->nr_units / upa; | |
1566 | alloc < alloc_end; alloc++) { | |
1567 | if (!(alloc % apl)) { | |
1170532b | 1568 | pr_cont("\n"); |
fd1e8a1f TH |
1569 | printk("%spcpu-alloc: ", lvl); |
1570 | } | |
1170532b | 1571 | pr_cont("[%0*d] ", group_width, group); |
fd1e8a1f TH |
1572 | |
1573 | for (unit_end += upa; unit < unit_end; unit++) | |
1574 | if (gi->cpu_map[unit] != NR_CPUS) | |
1170532b JP |
1575 | pr_cont("%0*d ", |
1576 | cpu_width, gi->cpu_map[unit]); | |
fd1e8a1f | 1577 | else |
1170532b | 1578 | pr_cont("%s ", empty_str); |
033e48fb | 1579 | } |
033e48fb | 1580 | } |
1170532b | 1581 | pr_cont("\n"); |
033e48fb | 1582 | } |
033e48fb | 1583 | |
fbf59bc9 | 1584 | /** |
8d408b4b | 1585 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
fd1e8a1f | 1586 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
38a6be52 | 1587 | * @base_addr: mapped address |
8d408b4b TH |
1588 | * |
1589 | * Initialize the first percpu chunk which contains the kernel static | |
1590 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1591 | * setup path. |
8d408b4b | 1592 | * |
fd1e8a1f TH |
1593 | * @ai contains all information necessary to initialize the first |
1594 | * chunk and prime the dynamic percpu allocator. | |
1595 | * | |
1596 | * @ai->static_size is the size of static percpu area. | |
1597 | * | |
1598 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to | |
edcb4639 TH |
1599 | * reserve after the static area in the first chunk. This reserves |
1600 | * the first chunk such that it's available only through reserved | |
1601 | * percpu allocation. This is primarily used to serve module percpu | |
1602 | * static areas on architectures where the addressing model has | |
1603 | * limited offset range for symbol relocations to guarantee module | |
1604 | * percpu symbols fall inside the relocatable range. | |
1605 | * | |
fd1e8a1f TH |
1606 | * @ai->dyn_size determines the number of bytes available for dynamic |
1607 | * allocation in the first chunk. The area between @ai->static_size + | |
1608 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. | |
6074d5b0 | 1609 | * |
fd1e8a1f TH |
1610 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
1611 | * and equal to or larger than @ai->static_size + @ai->reserved_size + | |
1612 | * @ai->dyn_size. | |
8d408b4b | 1613 | * |
fd1e8a1f TH |
1614 | * @ai->atom_size is the allocation atom size and used as alignment |
1615 | * for vm areas. | |
8d408b4b | 1616 | * |
fd1e8a1f TH |
1617 | * @ai->alloc_size is the allocation size and always multiple of |
1618 | * @ai->atom_size. This is larger than @ai->atom_size if | |
1619 | * @ai->unit_size is larger than @ai->atom_size. | |
1620 | * | |
1621 | * @ai->nr_groups and @ai->groups describe virtual memory layout of | |
1622 | * percpu areas. Units which should be colocated are put into the | |
1623 | * same group. Dynamic VM areas will be allocated according to these | |
1624 | * groupings. If @ai->nr_groups is zero, a single group containing | |
1625 | * all units is assumed. | |
8d408b4b | 1626 | * |
38a6be52 TH |
1627 | * The caller should have mapped the first chunk at @base_addr and |
1628 | * copied static data to each unit. | |
fbf59bc9 | 1629 | * |
c0ebfdc3 DZF |
1630 | * The first chunk will always contain a static and a dynamic region. |
1631 | * However, the static region is not managed by any chunk. If the first | |
1632 | * chunk also contains a reserved region, it is served by two chunks - | |
1633 | * one for the reserved region and one for the dynamic region. They | |
1634 | * share the same vm, but use offset regions in the area allocation map. | |
1635 | * The chunk serving the dynamic region is circulated in the chunk slots | |
1636 | * and available for dynamic allocation like any other chunk. | |
edcb4639 | 1637 | * |
fbf59bc9 | 1638 | * RETURNS: |
fb435d52 | 1639 | * 0 on success, -errno on failure. |
fbf59bc9 | 1640 | */ |
fb435d52 TH |
1641 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
1642 | void *base_addr) | |
fbf59bc9 | 1643 | { |
099a19d9 TH |
1644 | static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; |
1645 | static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; | |
b9c39442 | 1646 | size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
0c4169c3 | 1647 | struct pcpu_chunk *chunk; |
6563297c TH |
1648 | unsigned long *group_offsets; |
1649 | size_t *group_sizes; | |
fb435d52 | 1650 | unsigned long *unit_off; |
fbf59bc9 | 1651 | unsigned int cpu; |
fd1e8a1f TH |
1652 | int *unit_map; |
1653 | int group, unit, i; | |
c0ebfdc3 DZF |
1654 | int map_size; |
1655 | unsigned long tmp_addr; | |
fbf59bc9 | 1656 | |
635b75fc TH |
1657 | #define PCPU_SETUP_BUG_ON(cond) do { \ |
1658 | if (unlikely(cond)) { \ | |
870d4b12 JP |
1659 | pr_emerg("failed to initialize, %s\n", #cond); \ |
1660 | pr_emerg("cpu_possible_mask=%*pb\n", \ | |
807de073 | 1661 | cpumask_pr_args(cpu_possible_mask)); \ |
635b75fc TH |
1662 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ |
1663 | BUG(); \ | |
1664 | } \ | |
1665 | } while (0) | |
1666 | ||
2f39e637 | 1667 | /* sanity checks */ |
635b75fc | 1668 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
bbddff05 | 1669 | #ifdef CONFIG_SMP |
635b75fc | 1670 | PCPU_SETUP_BUG_ON(!ai->static_size); |
f09f1243 | 1671 | PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); |
bbddff05 | 1672 | #endif |
635b75fc | 1673 | PCPU_SETUP_BUG_ON(!base_addr); |
f09f1243 | 1674 | PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); |
635b75fc | 1675 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
f09f1243 | 1676 | PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); |
635b75fc | 1677 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); |
099a19d9 | 1678 | PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); |
fb29a2cc | 1679 | PCPU_SETUP_BUG_ON(!ai->dyn_size); |
9f645532 | 1680 | PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); |
8d408b4b | 1681 | |
6563297c | 1682 | /* process group information and build config tables accordingly */ |
999c17e3 SS |
1683 | group_offsets = memblock_virt_alloc(ai->nr_groups * |
1684 | sizeof(group_offsets[0]), 0); | |
1685 | group_sizes = memblock_virt_alloc(ai->nr_groups * | |
1686 | sizeof(group_sizes[0]), 0); | |
1687 | unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); | |
1688 | unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); | |
2f39e637 | 1689 | |
fd1e8a1f | 1690 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
ffe0d5a5 | 1691 | unit_map[cpu] = UINT_MAX; |
a855b84c TH |
1692 | |
1693 | pcpu_low_unit_cpu = NR_CPUS; | |
1694 | pcpu_high_unit_cpu = NR_CPUS; | |
2f39e637 | 1695 | |
fd1e8a1f TH |
1696 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
1697 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
2f39e637 | 1698 | |
6563297c TH |
1699 | group_offsets[group] = gi->base_offset; |
1700 | group_sizes[group] = gi->nr_units * ai->unit_size; | |
1701 | ||
fd1e8a1f TH |
1702 | for (i = 0; i < gi->nr_units; i++) { |
1703 | cpu = gi->cpu_map[i]; | |
1704 | if (cpu == NR_CPUS) | |
1705 | continue; | |
8d408b4b | 1706 | |
9f295664 | 1707 | PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); |
635b75fc TH |
1708 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); |
1709 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); | |
fbf59bc9 | 1710 | |
fd1e8a1f | 1711 | unit_map[cpu] = unit + i; |
fb435d52 TH |
1712 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
1713 | ||
a855b84c TH |
1714 | /* determine low/high unit_cpu */ |
1715 | if (pcpu_low_unit_cpu == NR_CPUS || | |
1716 | unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) | |
1717 | pcpu_low_unit_cpu = cpu; | |
1718 | if (pcpu_high_unit_cpu == NR_CPUS || | |
1719 | unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) | |
1720 | pcpu_high_unit_cpu = cpu; | |
fd1e8a1f | 1721 | } |
2f39e637 | 1722 | } |
fd1e8a1f TH |
1723 | pcpu_nr_units = unit; |
1724 | ||
1725 | for_each_possible_cpu(cpu) | |
635b75fc TH |
1726 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
1727 | ||
1728 | /* we're done parsing the input, undefine BUG macro and dump config */ | |
1729 | #undef PCPU_SETUP_BUG_ON | |
bcbea798 | 1730 | pcpu_dump_alloc_info(KERN_DEBUG, ai); |
fd1e8a1f | 1731 | |
6563297c TH |
1732 | pcpu_nr_groups = ai->nr_groups; |
1733 | pcpu_group_offsets = group_offsets; | |
1734 | pcpu_group_sizes = group_sizes; | |
fd1e8a1f | 1735 | pcpu_unit_map = unit_map; |
fb435d52 | 1736 | pcpu_unit_offsets = unit_off; |
2f39e637 TH |
1737 | |
1738 | /* determine basic parameters */ | |
fd1e8a1f | 1739 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
d9b55eeb | 1740 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
6563297c | 1741 | pcpu_atom_size = ai->atom_size; |
ce3141a2 TH |
1742 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
1743 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
cafe8816 | 1744 | |
30a5b536 DZ |
1745 | pcpu_stats_save_ai(ai); |
1746 | ||
d9b55eeb TH |
1747 | /* |
1748 | * Allocate chunk slots. The additional last slot is for | |
1749 | * empty chunks. | |
1750 | */ | |
1751 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
999c17e3 SS |
1752 | pcpu_slot = memblock_virt_alloc( |
1753 | pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); | |
fbf59bc9 TH |
1754 | for (i = 0; i < pcpu_nr_slots; i++) |
1755 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1756 | ||
edcb4639 | 1757 | /* |
c0ebfdc3 DZF |
1758 | * Initialize first chunk. |
1759 | * If the reserved_size is non-zero, this initializes the reserved | |
1760 | * chunk. If the reserved_size is zero, the reserved chunk is NULL | |
1761 | * and the dynamic region is initialized here. The first chunk, | |
1762 | * pcpu_first_chunk, will always point to the chunk that serves | |
1763 | * the dynamic region. | |
edcb4639 | 1764 | */ |
c0ebfdc3 | 1765 | tmp_addr = (unsigned long)base_addr + ai->static_size; |
10edf5b0 | 1766 | map_size = ai->reserved_size ?: ai->dyn_size; |
c0ebfdc3 | 1767 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size, smap, |
0c4169c3 | 1768 | ARRAY_SIZE(smap)); |
61ace7fa | 1769 | |
edcb4639 | 1770 | /* init dynamic chunk if necessary */ |
b9c39442 | 1771 | if (ai->reserved_size) { |
0c4169c3 | 1772 | pcpu_reserved_chunk = chunk; |
b9c39442 | 1773 | |
c0ebfdc3 DZF |
1774 | tmp_addr = (unsigned long)base_addr + ai->static_size + |
1775 | ai->reserved_size; | |
10edf5b0 | 1776 | map_size = ai->dyn_size; |
c0ebfdc3 | 1777 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size, dmap, |
0c4169c3 | 1778 | ARRAY_SIZE(dmap)); |
edcb4639 TH |
1779 | } |
1780 | ||
2441d15c | 1781 | /* link the first chunk in */ |
0c4169c3 | 1782 | pcpu_first_chunk = chunk; |
e2266705 | 1783 | i = (pcpu_first_chunk->start_offset) ? 1 : 0; |
b539b87f | 1784 | pcpu_nr_empty_pop_pages += |
e2266705 | 1785 | pcpu_count_occupied_pages(pcpu_first_chunk, i); |
ae9e6bc9 | 1786 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
fbf59bc9 | 1787 | |
30a5b536 | 1788 | pcpu_stats_chunk_alloc(); |
df95e795 | 1789 | trace_percpu_create_chunk(base_addr); |
30a5b536 | 1790 | |
fbf59bc9 | 1791 | /* we're done */ |
bba174f5 | 1792 | pcpu_base_addr = base_addr; |
fb435d52 | 1793 | return 0; |
fbf59bc9 | 1794 | } |
66c3a757 | 1795 | |
bbddff05 TH |
1796 | #ifdef CONFIG_SMP |
1797 | ||
17f3609c | 1798 | const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { |
f58dc01b TH |
1799 | [PCPU_FC_AUTO] = "auto", |
1800 | [PCPU_FC_EMBED] = "embed", | |
1801 | [PCPU_FC_PAGE] = "page", | |
f58dc01b | 1802 | }; |
66c3a757 | 1803 | |
f58dc01b | 1804 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
66c3a757 | 1805 | |
f58dc01b TH |
1806 | static int __init percpu_alloc_setup(char *str) |
1807 | { | |
5479c78a CG |
1808 | if (!str) |
1809 | return -EINVAL; | |
1810 | ||
f58dc01b TH |
1811 | if (0) |
1812 | /* nada */; | |
1813 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK | |
1814 | else if (!strcmp(str, "embed")) | |
1815 | pcpu_chosen_fc = PCPU_FC_EMBED; | |
1816 | #endif | |
1817 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
1818 | else if (!strcmp(str, "page")) | |
1819 | pcpu_chosen_fc = PCPU_FC_PAGE; | |
f58dc01b TH |
1820 | #endif |
1821 | else | |
870d4b12 | 1822 | pr_warn("unknown allocator %s specified\n", str); |
66c3a757 | 1823 | |
f58dc01b | 1824 | return 0; |
66c3a757 | 1825 | } |
f58dc01b | 1826 | early_param("percpu_alloc", percpu_alloc_setup); |
66c3a757 | 1827 | |
3c9a024f TH |
1828 | /* |
1829 | * pcpu_embed_first_chunk() is used by the generic percpu setup. | |
1830 | * Build it if needed by the arch config or the generic setup is going | |
1831 | * to be used. | |
1832 | */ | |
08fc4580 TH |
1833 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
1834 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
3c9a024f TH |
1835 | #define BUILD_EMBED_FIRST_CHUNK |
1836 | #endif | |
1837 | ||
1838 | /* build pcpu_page_first_chunk() iff needed by the arch config */ | |
1839 | #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) | |
1840 | #define BUILD_PAGE_FIRST_CHUNK | |
1841 | #endif | |
1842 | ||
1843 | /* pcpu_build_alloc_info() is used by both embed and page first chunk */ | |
1844 | #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) | |
1845 | /** | |
1846 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs | |
1847 | * @reserved_size: the size of reserved percpu area in bytes | |
1848 | * @dyn_size: minimum free size for dynamic allocation in bytes | |
1849 | * @atom_size: allocation atom size | |
1850 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
1851 | * | |
1852 | * This function determines grouping of units, their mappings to cpus | |
1853 | * and other parameters considering needed percpu size, allocation | |
1854 | * atom size and distances between CPUs. | |
1855 | * | |
bffc4375 | 1856 | * Groups are always multiples of atom size and CPUs which are of |
3c9a024f TH |
1857 | * LOCAL_DISTANCE both ways are grouped together and share space for |
1858 | * units in the same group. The returned configuration is guaranteed | |
1859 | * to have CPUs on different nodes on different groups and >=75% usage | |
1860 | * of allocated virtual address space. | |
1861 | * | |
1862 | * RETURNS: | |
1863 | * On success, pointer to the new allocation_info is returned. On | |
1864 | * failure, ERR_PTR value is returned. | |
1865 | */ | |
1866 | static struct pcpu_alloc_info * __init pcpu_build_alloc_info( | |
1867 | size_t reserved_size, size_t dyn_size, | |
1868 | size_t atom_size, | |
1869 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
1870 | { | |
1871 | static int group_map[NR_CPUS] __initdata; | |
1872 | static int group_cnt[NR_CPUS] __initdata; | |
1873 | const size_t static_size = __per_cpu_end - __per_cpu_start; | |
1874 | int nr_groups = 1, nr_units = 0; | |
1875 | size_t size_sum, min_unit_size, alloc_size; | |
1876 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
1877 | int last_allocs, group, unit; | |
1878 | unsigned int cpu, tcpu; | |
1879 | struct pcpu_alloc_info *ai; | |
1880 | unsigned int *cpu_map; | |
1881 | ||
1882 | /* this function may be called multiple times */ | |
1883 | memset(group_map, 0, sizeof(group_map)); | |
1884 | memset(group_cnt, 0, sizeof(group_cnt)); | |
1885 | ||
1886 | /* calculate size_sum and ensure dyn_size is enough for early alloc */ | |
1887 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
1888 | max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); | |
1889 | dyn_size = size_sum - static_size - reserved_size; | |
1890 | ||
1891 | /* | |
1892 | * Determine min_unit_size, alloc_size and max_upa such that | |
1893 | * alloc_size is multiple of atom_size and is the smallest | |
25985edc | 1894 | * which can accommodate 4k aligned segments which are equal to |
3c9a024f TH |
1895 | * or larger than min_unit_size. |
1896 | */ | |
1897 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
1898 | ||
9c015162 | 1899 | /* determine the maximum # of units that can fit in an allocation */ |
3c9a024f TH |
1900 | alloc_size = roundup(min_unit_size, atom_size); |
1901 | upa = alloc_size / min_unit_size; | |
f09f1243 | 1902 | while (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
1903 | upa--; |
1904 | max_upa = upa; | |
1905 | ||
1906 | /* group cpus according to their proximity */ | |
1907 | for_each_possible_cpu(cpu) { | |
1908 | group = 0; | |
1909 | next_group: | |
1910 | for_each_possible_cpu(tcpu) { | |
1911 | if (cpu == tcpu) | |
1912 | break; | |
1913 | if (group_map[tcpu] == group && cpu_distance_fn && | |
1914 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || | |
1915 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
1916 | group++; | |
1917 | nr_groups = max(nr_groups, group + 1); | |
1918 | goto next_group; | |
1919 | } | |
1920 | } | |
1921 | group_map[cpu] = group; | |
1922 | group_cnt[group]++; | |
1923 | } | |
1924 | ||
1925 | /* | |
9c015162 DZF |
1926 | * Wasted space is caused by a ratio imbalance of upa to group_cnt. |
1927 | * Expand the unit_size until we use >= 75% of the units allocated. | |
1928 | * Related to atom_size, which could be much larger than the unit_size. | |
3c9a024f TH |
1929 | */ |
1930 | last_allocs = INT_MAX; | |
1931 | for (upa = max_upa; upa; upa--) { | |
1932 | int allocs = 0, wasted = 0; | |
1933 | ||
f09f1243 | 1934 | if (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
1935 | continue; |
1936 | ||
1937 | for (group = 0; group < nr_groups; group++) { | |
1938 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); | |
1939 | allocs += this_allocs; | |
1940 | wasted += this_allocs * upa - group_cnt[group]; | |
1941 | } | |
1942 | ||
1943 | /* | |
1944 | * Don't accept if wastage is over 1/3. The | |
1945 | * greater-than comparison ensures upa==1 always | |
1946 | * passes the following check. | |
1947 | */ | |
1948 | if (wasted > num_possible_cpus() / 3) | |
1949 | continue; | |
1950 | ||
1951 | /* and then don't consume more memory */ | |
1952 | if (allocs > last_allocs) | |
1953 | break; | |
1954 | last_allocs = allocs; | |
1955 | best_upa = upa; | |
1956 | } | |
1957 | upa = best_upa; | |
1958 | ||
1959 | /* allocate and fill alloc_info */ | |
1960 | for (group = 0; group < nr_groups; group++) | |
1961 | nr_units += roundup(group_cnt[group], upa); | |
1962 | ||
1963 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); | |
1964 | if (!ai) | |
1965 | return ERR_PTR(-ENOMEM); | |
1966 | cpu_map = ai->groups[0].cpu_map; | |
1967 | ||
1968 | for (group = 0; group < nr_groups; group++) { | |
1969 | ai->groups[group].cpu_map = cpu_map; | |
1970 | cpu_map += roundup(group_cnt[group], upa); | |
1971 | } | |
1972 | ||
1973 | ai->static_size = static_size; | |
1974 | ai->reserved_size = reserved_size; | |
1975 | ai->dyn_size = dyn_size; | |
1976 | ai->unit_size = alloc_size / upa; | |
1977 | ai->atom_size = atom_size; | |
1978 | ai->alloc_size = alloc_size; | |
1979 | ||
1980 | for (group = 0, unit = 0; group_cnt[group]; group++) { | |
1981 | struct pcpu_group_info *gi = &ai->groups[group]; | |
1982 | ||
1983 | /* | |
1984 | * Initialize base_offset as if all groups are located | |
1985 | * back-to-back. The caller should update this to | |
1986 | * reflect actual allocation. | |
1987 | */ | |
1988 | gi->base_offset = unit * ai->unit_size; | |
1989 | ||
1990 | for_each_possible_cpu(cpu) | |
1991 | if (group_map[cpu] == group) | |
1992 | gi->cpu_map[gi->nr_units++] = cpu; | |
1993 | gi->nr_units = roundup(gi->nr_units, upa); | |
1994 | unit += gi->nr_units; | |
1995 | } | |
1996 | BUG_ON(unit != nr_units); | |
1997 | ||
1998 | return ai; | |
1999 | } | |
2000 | #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ | |
2001 | ||
2002 | #if defined(BUILD_EMBED_FIRST_CHUNK) | |
66c3a757 TH |
2003 | /** |
2004 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
66c3a757 | 2005 | * @reserved_size: the size of reserved percpu area in bytes |
4ba6ce25 | 2006 | * @dyn_size: minimum free size for dynamic allocation in bytes |
c8826dd5 TH |
2007 | * @atom_size: allocation atom size |
2008 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
2009 | * @alloc_fn: function to allocate percpu page | |
25985edc | 2010 | * @free_fn: function to free percpu page |
66c3a757 TH |
2011 | * |
2012 | * This is a helper to ease setting up embedded first percpu chunk and | |
2013 | * can be called where pcpu_setup_first_chunk() is expected. | |
2014 | * | |
2015 | * If this function is used to setup the first chunk, it is allocated | |
c8826dd5 TH |
2016 | * by calling @alloc_fn and used as-is without being mapped into |
2017 | * vmalloc area. Allocations are always whole multiples of @atom_size | |
2018 | * aligned to @atom_size. | |
2019 | * | |
2020 | * This enables the first chunk to piggy back on the linear physical | |
2021 | * mapping which often uses larger page size. Please note that this | |
2022 | * can result in very sparse cpu->unit mapping on NUMA machines thus | |
2023 | * requiring large vmalloc address space. Don't use this allocator if | |
2024 | * vmalloc space is not orders of magnitude larger than distances | |
2025 | * between node memory addresses (ie. 32bit NUMA machines). | |
66c3a757 | 2026 | * |
4ba6ce25 | 2027 | * @dyn_size specifies the minimum dynamic area size. |
66c3a757 TH |
2028 | * |
2029 | * If the needed size is smaller than the minimum or specified unit | |
c8826dd5 | 2030 | * size, the leftover is returned using @free_fn. |
66c3a757 TH |
2031 | * |
2032 | * RETURNS: | |
fb435d52 | 2033 | * 0 on success, -errno on failure. |
66c3a757 | 2034 | */ |
4ba6ce25 | 2035 | int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, |
c8826dd5 TH |
2036 | size_t atom_size, |
2037 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, | |
2038 | pcpu_fc_alloc_fn_t alloc_fn, | |
2039 | pcpu_fc_free_fn_t free_fn) | |
66c3a757 | 2040 | { |
c8826dd5 TH |
2041 | void *base = (void *)ULONG_MAX; |
2042 | void **areas = NULL; | |
fd1e8a1f | 2043 | struct pcpu_alloc_info *ai; |
93c76b6b | 2044 | size_t size_sum, areas_size; |
2045 | unsigned long max_distance; | |
9b739662 | 2046 | int group, i, highest_group, rc; |
66c3a757 | 2047 | |
c8826dd5 TH |
2048 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
2049 | cpu_distance_fn); | |
fd1e8a1f TH |
2050 | if (IS_ERR(ai)) |
2051 | return PTR_ERR(ai); | |
66c3a757 | 2052 | |
fd1e8a1f | 2053 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
c8826dd5 | 2054 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
fa8a7094 | 2055 | |
999c17e3 | 2056 | areas = memblock_virt_alloc_nopanic(areas_size, 0); |
c8826dd5 | 2057 | if (!areas) { |
fb435d52 | 2058 | rc = -ENOMEM; |
c8826dd5 | 2059 | goto out_free; |
fa8a7094 | 2060 | } |
66c3a757 | 2061 | |
9b739662 | 2062 | /* allocate, copy and determine base address & max_distance */ |
2063 | highest_group = 0; | |
c8826dd5 TH |
2064 | for (group = 0; group < ai->nr_groups; group++) { |
2065 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2066 | unsigned int cpu = NR_CPUS; | |
2067 | void *ptr; | |
2068 | ||
2069 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) | |
2070 | cpu = gi->cpu_map[i]; | |
2071 | BUG_ON(cpu == NR_CPUS); | |
2072 | ||
2073 | /* allocate space for the whole group */ | |
2074 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); | |
2075 | if (!ptr) { | |
2076 | rc = -ENOMEM; | |
2077 | goto out_free_areas; | |
2078 | } | |
f528f0b8 CM |
2079 | /* kmemleak tracks the percpu allocations separately */ |
2080 | kmemleak_free(ptr); | |
c8826dd5 | 2081 | areas[group] = ptr; |
fd1e8a1f | 2082 | |
c8826dd5 | 2083 | base = min(ptr, base); |
9b739662 | 2084 | if (ptr > areas[highest_group]) |
2085 | highest_group = group; | |
2086 | } | |
2087 | max_distance = areas[highest_group] - base; | |
2088 | max_distance += ai->unit_size * ai->groups[highest_group].nr_units; | |
2089 | ||
2090 | /* warn if maximum distance is further than 75% of vmalloc space */ | |
2091 | if (max_distance > VMALLOC_TOTAL * 3 / 4) { | |
2092 | pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", | |
2093 | max_distance, VMALLOC_TOTAL); | |
2094 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
2095 | /* and fail if we have fallback */ | |
2096 | rc = -EINVAL; | |
2097 | goto out_free_areas; | |
2098 | #endif | |
42b64281 TH |
2099 | } |
2100 | ||
2101 | /* | |
2102 | * Copy data and free unused parts. This should happen after all | |
2103 | * allocations are complete; otherwise, we may end up with | |
2104 | * overlapping groups. | |
2105 | */ | |
2106 | for (group = 0; group < ai->nr_groups; group++) { | |
2107 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2108 | void *ptr = areas[group]; | |
c8826dd5 TH |
2109 | |
2110 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { | |
2111 | if (gi->cpu_map[i] == NR_CPUS) { | |
2112 | /* unused unit, free whole */ | |
2113 | free_fn(ptr, ai->unit_size); | |
2114 | continue; | |
2115 | } | |
2116 | /* copy and return the unused part */ | |
2117 | memcpy(ptr, __per_cpu_load, ai->static_size); | |
2118 | free_fn(ptr + size_sum, ai->unit_size - size_sum); | |
2119 | } | |
fa8a7094 | 2120 | } |
66c3a757 | 2121 | |
c8826dd5 | 2122 | /* base address is now known, determine group base offsets */ |
6ea529a2 | 2123 | for (group = 0; group < ai->nr_groups; group++) { |
c8826dd5 | 2124 | ai->groups[group].base_offset = areas[group] - base; |
6ea529a2 | 2125 | } |
c8826dd5 | 2126 | |
870d4b12 | 2127 | pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
fd1e8a1f TH |
2128 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
2129 | ai->dyn_size, ai->unit_size); | |
d4b95f80 | 2130 | |
fb435d52 | 2131 | rc = pcpu_setup_first_chunk(ai, base); |
c8826dd5 TH |
2132 | goto out_free; |
2133 | ||
2134 | out_free_areas: | |
2135 | for (group = 0; group < ai->nr_groups; group++) | |
f851c8d8 MH |
2136 | if (areas[group]) |
2137 | free_fn(areas[group], | |
2138 | ai->groups[group].nr_units * ai->unit_size); | |
c8826dd5 | 2139 | out_free: |
fd1e8a1f | 2140 | pcpu_free_alloc_info(ai); |
c8826dd5 | 2141 | if (areas) |
999c17e3 | 2142 | memblock_free_early(__pa(areas), areas_size); |
fb435d52 | 2143 | return rc; |
d4b95f80 | 2144 | } |
3c9a024f | 2145 | #endif /* BUILD_EMBED_FIRST_CHUNK */ |
d4b95f80 | 2146 | |
3c9a024f | 2147 | #ifdef BUILD_PAGE_FIRST_CHUNK |
d4b95f80 | 2148 | /** |
00ae4064 | 2149 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
2150 | * @reserved_size: the size of reserved percpu area in bytes |
2151 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
25985edc | 2152 | * @free_fn: function to free percpu page, always called with PAGE_SIZE |
d4b95f80 TH |
2153 | * @populate_pte_fn: function to populate pte |
2154 | * | |
00ae4064 TH |
2155 | * This is a helper to ease setting up page-remapped first percpu |
2156 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
2157 | * |
2158 | * This is the basic allocator. Static percpu area is allocated | |
2159 | * page-by-page into vmalloc area. | |
2160 | * | |
2161 | * RETURNS: | |
fb435d52 | 2162 | * 0 on success, -errno on failure. |
d4b95f80 | 2163 | */ |
fb435d52 TH |
2164 | int __init pcpu_page_first_chunk(size_t reserved_size, |
2165 | pcpu_fc_alloc_fn_t alloc_fn, | |
2166 | pcpu_fc_free_fn_t free_fn, | |
2167 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 2168 | { |
8f05a6a6 | 2169 | static struct vm_struct vm; |
fd1e8a1f | 2170 | struct pcpu_alloc_info *ai; |
00ae4064 | 2171 | char psize_str[16]; |
ce3141a2 | 2172 | int unit_pages; |
d4b95f80 | 2173 | size_t pages_size; |
ce3141a2 | 2174 | struct page **pages; |
fb435d52 | 2175 | int unit, i, j, rc; |
8f606604 | 2176 | int upa; |
2177 | int nr_g0_units; | |
d4b95f80 | 2178 | |
00ae4064 TH |
2179 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
2180 | ||
4ba6ce25 | 2181 | ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); |
fd1e8a1f TH |
2182 | if (IS_ERR(ai)) |
2183 | return PTR_ERR(ai); | |
2184 | BUG_ON(ai->nr_groups != 1); | |
8f606604 | 2185 | upa = ai->alloc_size/ai->unit_size; |
2186 | nr_g0_units = roundup(num_possible_cpus(), upa); | |
2187 | if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) { | |
2188 | pcpu_free_alloc_info(ai); | |
2189 | return -EINVAL; | |
2190 | } | |
fd1e8a1f TH |
2191 | |
2192 | unit_pages = ai->unit_size >> PAGE_SHIFT; | |
d4b95f80 TH |
2193 | |
2194 | /* unaligned allocations can't be freed, round up to page size */ | |
fd1e8a1f TH |
2195 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
2196 | sizeof(pages[0])); | |
999c17e3 | 2197 | pages = memblock_virt_alloc(pages_size, 0); |
d4b95f80 | 2198 | |
8f05a6a6 | 2199 | /* allocate pages */ |
d4b95f80 | 2200 | j = 0; |
8f606604 | 2201 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
2202 | unsigned int cpu = ai->groups[0].cpu_map[unit]; | |
ce3141a2 | 2203 | for (i = 0; i < unit_pages; i++) { |
d4b95f80 TH |
2204 | void *ptr; |
2205 | ||
3cbc8565 | 2206 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
d4b95f80 | 2207 | if (!ptr) { |
870d4b12 | 2208 | pr_warn("failed to allocate %s page for cpu%u\n", |
8f606604 | 2209 | psize_str, cpu); |
d4b95f80 TH |
2210 | goto enomem; |
2211 | } | |
f528f0b8 CM |
2212 | /* kmemleak tracks the percpu allocations separately */ |
2213 | kmemleak_free(ptr); | |
ce3141a2 | 2214 | pages[j++] = virt_to_page(ptr); |
d4b95f80 | 2215 | } |
8f606604 | 2216 | } |
d4b95f80 | 2217 | |
8f05a6a6 TH |
2218 | /* allocate vm area, map the pages and copy static data */ |
2219 | vm.flags = VM_ALLOC; | |
fd1e8a1f | 2220 | vm.size = num_possible_cpus() * ai->unit_size; |
8f05a6a6 TH |
2221 | vm_area_register_early(&vm, PAGE_SIZE); |
2222 | ||
fd1e8a1f | 2223 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
1d9d3257 | 2224 | unsigned long unit_addr = |
fd1e8a1f | 2225 | (unsigned long)vm.addr + unit * ai->unit_size; |
8f05a6a6 | 2226 | |
ce3141a2 | 2227 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
2228 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
2229 | ||
2230 | /* pte already populated, the following shouldn't fail */ | |
fb435d52 TH |
2231 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
2232 | unit_pages); | |
2233 | if (rc < 0) | |
2234 | panic("failed to map percpu area, err=%d\n", rc); | |
66c3a757 | 2235 | |
8f05a6a6 TH |
2236 | /* |
2237 | * FIXME: Archs with virtual cache should flush local | |
2238 | * cache for the linear mapping here - something | |
2239 | * equivalent to flush_cache_vmap() on the local cpu. | |
2240 | * flush_cache_vmap() can't be used as most supporting | |
2241 | * data structures are not set up yet. | |
2242 | */ | |
2243 | ||
2244 | /* copy static data */ | |
fd1e8a1f | 2245 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
66c3a757 TH |
2246 | } |
2247 | ||
2248 | /* we're ready, commit */ | |
870d4b12 | 2249 | pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", |
fd1e8a1f TH |
2250 | unit_pages, psize_str, vm.addr, ai->static_size, |
2251 | ai->reserved_size, ai->dyn_size); | |
d4b95f80 | 2252 | |
fb435d52 | 2253 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
d4b95f80 TH |
2254 | goto out_free_ar; |
2255 | ||
2256 | enomem: | |
2257 | while (--j >= 0) | |
ce3141a2 | 2258 | free_fn(page_address(pages[j]), PAGE_SIZE); |
fb435d52 | 2259 | rc = -ENOMEM; |
d4b95f80 | 2260 | out_free_ar: |
999c17e3 | 2261 | memblock_free_early(__pa(pages), pages_size); |
fd1e8a1f | 2262 | pcpu_free_alloc_info(ai); |
fb435d52 | 2263 | return rc; |
d4b95f80 | 2264 | } |
3c9a024f | 2265 | #endif /* BUILD_PAGE_FIRST_CHUNK */ |
d4b95f80 | 2266 | |
bbddff05 | 2267 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
e74e3962 | 2268 | /* |
bbddff05 | 2269 | * Generic SMP percpu area setup. |
e74e3962 TH |
2270 | * |
2271 | * The embedding helper is used because its behavior closely resembles | |
2272 | * the original non-dynamic generic percpu area setup. This is | |
2273 | * important because many archs have addressing restrictions and might | |
2274 | * fail if the percpu area is located far away from the previous | |
2275 | * location. As an added bonus, in non-NUMA cases, embedding is | |
2276 | * generally a good idea TLB-wise because percpu area can piggy back | |
2277 | * on the physical linear memory mapping which uses large page | |
2278 | * mappings on applicable archs. | |
2279 | */ | |
e74e3962 TH |
2280 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
2281 | EXPORT_SYMBOL(__per_cpu_offset); | |
2282 | ||
c8826dd5 TH |
2283 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
2284 | size_t align) | |
2285 | { | |
999c17e3 SS |
2286 | return memblock_virt_alloc_from_nopanic( |
2287 | size, align, __pa(MAX_DMA_ADDRESS)); | |
c8826dd5 | 2288 | } |
66c3a757 | 2289 | |
c8826dd5 TH |
2290 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
2291 | { | |
999c17e3 | 2292 | memblock_free_early(__pa(ptr), size); |
c8826dd5 TH |
2293 | } |
2294 | ||
e74e3962 TH |
2295 | void __init setup_per_cpu_areas(void) |
2296 | { | |
e74e3962 TH |
2297 | unsigned long delta; |
2298 | unsigned int cpu; | |
fb435d52 | 2299 | int rc; |
e74e3962 TH |
2300 | |
2301 | /* | |
2302 | * Always reserve area for module percpu variables. That's | |
2303 | * what the legacy allocator did. | |
2304 | */ | |
fb435d52 | 2305 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
c8826dd5 TH |
2306 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
2307 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); | |
fb435d52 | 2308 | if (rc < 0) |
bbddff05 | 2309 | panic("Failed to initialize percpu areas."); |
e74e3962 TH |
2310 | |
2311 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2312 | for_each_possible_cpu(cpu) | |
fb435d52 | 2313 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
66c3a757 | 2314 | } |
bbddff05 TH |
2315 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
2316 | ||
2317 | #else /* CONFIG_SMP */ | |
2318 | ||
2319 | /* | |
2320 | * UP percpu area setup. | |
2321 | * | |
2322 | * UP always uses km-based percpu allocator with identity mapping. | |
2323 | * Static percpu variables are indistinguishable from the usual static | |
2324 | * variables and don't require any special preparation. | |
2325 | */ | |
2326 | void __init setup_per_cpu_areas(void) | |
2327 | { | |
2328 | const size_t unit_size = | |
2329 | roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, | |
2330 | PERCPU_DYNAMIC_RESERVE)); | |
2331 | struct pcpu_alloc_info *ai; | |
2332 | void *fc; | |
2333 | ||
2334 | ai = pcpu_alloc_alloc_info(1, 1); | |
999c17e3 SS |
2335 | fc = memblock_virt_alloc_from_nopanic(unit_size, |
2336 | PAGE_SIZE, | |
2337 | __pa(MAX_DMA_ADDRESS)); | |
bbddff05 TH |
2338 | if (!ai || !fc) |
2339 | panic("Failed to allocate memory for percpu areas."); | |
100d13c3 CM |
2340 | /* kmemleak tracks the percpu allocations separately */ |
2341 | kmemleak_free(fc); | |
bbddff05 TH |
2342 | |
2343 | ai->dyn_size = unit_size; | |
2344 | ai->unit_size = unit_size; | |
2345 | ai->atom_size = unit_size; | |
2346 | ai->alloc_size = unit_size; | |
2347 | ai->groups[0].nr_units = 1; | |
2348 | ai->groups[0].cpu_map[0] = 0; | |
2349 | ||
2350 | if (pcpu_setup_first_chunk(ai, fc) < 0) | |
2351 | panic("Failed to initialize percpu areas."); | |
2352 | } | |
2353 | ||
2354 | #endif /* CONFIG_SMP */ | |
099a19d9 TH |
2355 | |
2356 | /* | |
2357 | * First and reserved chunks are initialized with temporary allocation | |
2358 | * map in initdata so that they can be used before slab is online. | |
2359 | * This function is called after slab is brought up and replaces those | |
2360 | * with properly allocated maps. | |
2361 | */ | |
2362 | void __init percpu_init_late(void) | |
2363 | { | |
2364 | struct pcpu_chunk *target_chunks[] = | |
2365 | { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; | |
2366 | struct pcpu_chunk *chunk; | |
2367 | unsigned long flags; | |
2368 | int i; | |
2369 | ||
2370 | for (i = 0; (chunk = target_chunks[i]); i++) { | |
2371 | int *map; | |
2372 | const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); | |
2373 | ||
2374 | BUILD_BUG_ON(size > PAGE_SIZE); | |
2375 | ||
90459ce0 | 2376 | map = pcpu_mem_zalloc(size); |
099a19d9 TH |
2377 | BUG_ON(!map); |
2378 | ||
2379 | spin_lock_irqsave(&pcpu_lock, flags); | |
2380 | memcpy(map, chunk->map, size); | |
2381 | chunk->map = map; | |
2382 | spin_unlock_irqrestore(&pcpu_lock, flags); | |
2383 | } | |
2384 | } | |
1a4d7607 TH |
2385 | |
2386 | /* | |
2387 | * Percpu allocator is initialized early during boot when neither slab or | |
2388 | * workqueue is available. Plug async management until everything is up | |
2389 | * and running. | |
2390 | */ | |
2391 | static int __init percpu_enable_async(void) | |
2392 | { | |
2393 | pcpu_async_enabled = true; | |
2394 | return 0; | |
2395 | } | |
2396 | subsys_initcall(percpu_enable_async); |