]>
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 | ||
91e914c5 | 256 | static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end) |
ce3141a2 | 257 | { |
91e914c5 DZF |
258 | *rs = find_next_zero_bit(bitmap, end, *rs); |
259 | *re = find_next_bit(bitmap, end, *rs + 1); | |
ce3141a2 TH |
260 | } |
261 | ||
91e914c5 | 262 | static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end) |
ce3141a2 | 263 | { |
91e914c5 DZF |
264 | *rs = find_next_bit(bitmap, end, *rs); |
265 | *re = find_next_zero_bit(bitmap, end, *rs + 1); | |
ce3141a2 TH |
266 | } |
267 | ||
268 | /* | |
91e914c5 DZF |
269 | * Bitmap region iterators. Iterates over the bitmap between |
270 | * [@start, @end) in @chunk. @rs and @re should be integer variables | |
271 | * and will be set to start and end index of the current free region. | |
ce3141a2 | 272 | */ |
91e914c5 DZF |
273 | #define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \ |
274 | for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \ | |
275 | (rs) < (re); \ | |
276 | (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end))) | |
ce3141a2 | 277 | |
91e914c5 DZF |
278 | #define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \ |
279 | for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \ | |
280 | (rs) < (re); \ | |
281 | (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end))) | |
ce3141a2 | 282 | |
fbf59bc9 | 283 | /** |
90459ce0 | 284 | * pcpu_mem_zalloc - allocate memory |
1880d93b | 285 | * @size: bytes to allocate |
fbf59bc9 | 286 | * |
1880d93b | 287 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
90459ce0 | 288 | * kzalloc() is used; otherwise, vzalloc() is used. The returned |
1880d93b | 289 | * memory is always zeroed. |
fbf59bc9 | 290 | * |
ccea34b5 TH |
291 | * CONTEXT: |
292 | * Does GFP_KERNEL allocation. | |
293 | * | |
fbf59bc9 | 294 | * RETURNS: |
1880d93b | 295 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 296 | */ |
90459ce0 | 297 | static void *pcpu_mem_zalloc(size_t size) |
fbf59bc9 | 298 | { |
099a19d9 TH |
299 | if (WARN_ON_ONCE(!slab_is_available())) |
300 | return NULL; | |
301 | ||
1880d93b TH |
302 | if (size <= PAGE_SIZE) |
303 | return kzalloc(size, GFP_KERNEL); | |
7af4c093 JJ |
304 | else |
305 | return vzalloc(size); | |
1880d93b | 306 | } |
fbf59bc9 | 307 | |
1880d93b TH |
308 | /** |
309 | * pcpu_mem_free - free memory | |
310 | * @ptr: memory to free | |
1880d93b | 311 | * |
90459ce0 | 312 | * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc(). |
1880d93b | 313 | */ |
1d5cfdb0 | 314 | static void pcpu_mem_free(void *ptr) |
1880d93b | 315 | { |
1d5cfdb0 | 316 | kvfree(ptr); |
fbf59bc9 TH |
317 | } |
318 | ||
b539b87f TH |
319 | /** |
320 | * pcpu_count_occupied_pages - count the number of pages an area occupies | |
321 | * @chunk: chunk of interest | |
322 | * @i: index of the area in question | |
323 | * | |
324 | * Count the number of pages chunk's @i'th area occupies. When the area's | |
325 | * start and/or end address isn't aligned to page boundary, the straddled | |
326 | * page is included in the count iff the rest of the page is free. | |
327 | */ | |
328 | static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i) | |
329 | { | |
330 | int off = chunk->map[i] & ~1; | |
331 | int end = chunk->map[i + 1] & ~1; | |
332 | ||
333 | if (!PAGE_ALIGNED(off) && i > 0) { | |
334 | int prev = chunk->map[i - 1]; | |
335 | ||
336 | if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE)) | |
337 | off = round_down(off, PAGE_SIZE); | |
338 | } | |
339 | ||
340 | if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) { | |
341 | int next = chunk->map[i + 1]; | |
342 | int nend = chunk->map[i + 2] & ~1; | |
343 | ||
344 | if (!(next & 1) && nend >= round_up(end, PAGE_SIZE)) | |
345 | end = round_up(end, PAGE_SIZE); | |
346 | } | |
347 | ||
348 | return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0); | |
349 | } | |
350 | ||
fbf59bc9 TH |
351 | /** |
352 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
353 | * @chunk: chunk of interest | |
354 | * @oslot: the previous slot it was on | |
355 | * | |
356 | * This function is called after an allocation or free changed @chunk. | |
357 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
358 | * moved to the slot. Note that the reserved chunk is never put on |
359 | * chunk slots. | |
ccea34b5 TH |
360 | * |
361 | * CONTEXT: | |
362 | * pcpu_lock. | |
fbf59bc9 TH |
363 | */ |
364 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
365 | { | |
366 | int nslot = pcpu_chunk_slot(chunk); | |
367 | ||
edcb4639 | 368 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
369 | if (oslot < nslot) |
370 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
371 | else | |
372 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
373 | } | |
374 | } | |
375 | ||
9f7dcf22 | 376 | /** |
833af842 TH |
377 | * pcpu_need_to_extend - determine whether chunk area map needs to be extended |
378 | * @chunk: chunk of interest | |
9c824b6a | 379 | * @is_atomic: the allocation context |
9f7dcf22 | 380 | * |
9c824b6a TH |
381 | * Determine whether area map of @chunk needs to be extended. If |
382 | * @is_atomic, only the amount necessary for a new allocation is | |
383 | * considered; however, async extension is scheduled if the left amount is | |
384 | * low. If !@is_atomic, it aims for more empty space. Combined, this | |
385 | * ensures that the map is likely to have enough available space to | |
386 | * accomodate atomic allocations which can't extend maps directly. | |
9f7dcf22 | 387 | * |
ccea34b5 | 388 | * CONTEXT: |
833af842 | 389 | * pcpu_lock. |
ccea34b5 | 390 | * |
9f7dcf22 | 391 | * RETURNS: |
833af842 TH |
392 | * New target map allocation length if extension is necessary, 0 |
393 | * otherwise. | |
9f7dcf22 | 394 | */ |
9c824b6a | 395 | static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic) |
9f7dcf22 | 396 | { |
9c824b6a TH |
397 | int margin, new_alloc; |
398 | ||
4f996e23 TH |
399 | lockdep_assert_held(&pcpu_lock); |
400 | ||
9c824b6a TH |
401 | if (is_atomic) { |
402 | margin = 3; | |
9f7dcf22 | 403 | |
9c824b6a | 404 | if (chunk->map_alloc < |
4f996e23 TH |
405 | chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW) { |
406 | if (list_empty(&chunk->map_extend_list)) { | |
407 | list_add_tail(&chunk->map_extend_list, | |
408 | &pcpu_map_extend_chunks); | |
409 | pcpu_schedule_balance_work(); | |
410 | } | |
411 | } | |
9c824b6a TH |
412 | } else { |
413 | margin = PCPU_ATOMIC_MAP_MARGIN_HIGH; | |
414 | } | |
415 | ||
416 | if (chunk->map_alloc >= chunk->map_used + margin) | |
9f7dcf22 TH |
417 | return 0; |
418 | ||
419 | new_alloc = PCPU_DFL_MAP_ALLOC; | |
9c824b6a | 420 | while (new_alloc < chunk->map_used + margin) |
9f7dcf22 TH |
421 | new_alloc *= 2; |
422 | ||
833af842 TH |
423 | return new_alloc; |
424 | } | |
425 | ||
426 | /** | |
427 | * pcpu_extend_area_map - extend area map of a chunk | |
428 | * @chunk: chunk of interest | |
429 | * @new_alloc: new target allocation length of the area map | |
430 | * | |
431 | * Extend area map of @chunk to have @new_alloc entries. | |
432 | * | |
433 | * CONTEXT: | |
434 | * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock. | |
435 | * | |
436 | * RETURNS: | |
437 | * 0 on success, -errno on failure. | |
438 | */ | |
439 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc) | |
440 | { | |
441 | int *old = NULL, *new = NULL; | |
442 | size_t old_size = 0, new_size = new_alloc * sizeof(new[0]); | |
443 | unsigned long flags; | |
444 | ||
6710e594 TH |
445 | lockdep_assert_held(&pcpu_alloc_mutex); |
446 | ||
90459ce0 | 447 | new = pcpu_mem_zalloc(new_size); |
833af842 | 448 | if (!new) |
9f7dcf22 | 449 | return -ENOMEM; |
ccea34b5 | 450 | |
833af842 TH |
451 | /* acquire pcpu_lock and switch to new area map */ |
452 | spin_lock_irqsave(&pcpu_lock, flags); | |
453 | ||
454 | if (new_alloc <= chunk->map_alloc) | |
455 | goto out_unlock; | |
9f7dcf22 | 456 | |
833af842 | 457 | old_size = chunk->map_alloc * sizeof(chunk->map[0]); |
a002d148 HS |
458 | old = chunk->map; |
459 | ||
460 | memcpy(new, old, old_size); | |
9f7dcf22 | 461 | |
9f7dcf22 TH |
462 | chunk->map_alloc = new_alloc; |
463 | chunk->map = new; | |
833af842 TH |
464 | new = NULL; |
465 | ||
466 | out_unlock: | |
467 | spin_unlock_irqrestore(&pcpu_lock, flags); | |
468 | ||
469 | /* | |
470 | * pcpu_mem_free() might end up calling vfree() which uses | |
471 | * IRQ-unsafe lock and thus can't be called under pcpu_lock. | |
472 | */ | |
1d5cfdb0 TH |
473 | pcpu_mem_free(old); |
474 | pcpu_mem_free(new); | |
833af842 | 475 | |
9f7dcf22 TH |
476 | return 0; |
477 | } | |
478 | ||
a16037c8 TH |
479 | /** |
480 | * pcpu_fit_in_area - try to fit the requested allocation in a candidate area | |
481 | * @chunk: chunk the candidate area belongs to | |
482 | * @off: the offset to the start of the candidate area | |
483 | * @this_size: the size of the candidate area | |
484 | * @size: the size of the target allocation | |
485 | * @align: the alignment of the target allocation | |
486 | * @pop_only: only allocate from already populated region | |
487 | * | |
488 | * We're trying to allocate @size bytes aligned at @align. @chunk's area | |
489 | * at @off sized @this_size is a candidate. This function determines | |
490 | * whether the target allocation fits in the candidate area and returns the | |
491 | * number of bytes to pad after @off. If the target area doesn't fit, -1 | |
492 | * is returned. | |
493 | * | |
494 | * If @pop_only is %true, this function only considers the already | |
495 | * populated part of the candidate area. | |
496 | */ | |
497 | static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size, | |
498 | int size, int align, bool pop_only) | |
499 | { | |
500 | int cand_off = off; | |
501 | ||
502 | while (true) { | |
503 | int head = ALIGN(cand_off, align) - off; | |
504 | int page_start, page_end, rs, re; | |
505 | ||
506 | if (this_size < head + size) | |
507 | return -1; | |
508 | ||
509 | if (!pop_only) | |
510 | return head; | |
511 | ||
512 | /* | |
513 | * If the first unpopulated page is beyond the end of the | |
514 | * allocation, the whole allocation is populated; | |
515 | * otherwise, retry from the end of the unpopulated area. | |
516 | */ | |
517 | page_start = PFN_DOWN(head + off); | |
518 | page_end = PFN_UP(head + off + size); | |
519 | ||
520 | rs = page_start; | |
91e914c5 DZF |
521 | pcpu_next_unpop(chunk->populated, &rs, &re, |
522 | PFN_UP(off + this_size)); | |
a16037c8 TH |
523 | if (rs >= page_end) |
524 | return head; | |
525 | cand_off = re * PAGE_SIZE; | |
526 | } | |
527 | } | |
528 | ||
fbf59bc9 TH |
529 | /** |
530 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
531 | * @chunk: chunk of interest | |
cae3aeb8 | 532 | * @size: wanted size in bytes |
fbf59bc9 | 533 | * @align: wanted align |
a16037c8 | 534 | * @pop_only: allocate only from the populated area |
b539b87f | 535 | * @occ_pages_p: out param for the number of pages the area occupies |
fbf59bc9 TH |
536 | * |
537 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
538 | * Note that this function only allocates the offset. It doesn't | |
539 | * populate or map the area. | |
540 | * | |
9f7dcf22 TH |
541 | * @chunk->map must have at least two free slots. |
542 | * | |
ccea34b5 TH |
543 | * CONTEXT: |
544 | * pcpu_lock. | |
545 | * | |
fbf59bc9 | 546 | * RETURNS: |
9f7dcf22 TH |
547 | * Allocated offset in @chunk on success, -1 if no matching area is |
548 | * found. | |
fbf59bc9 | 549 | */ |
a16037c8 | 550 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align, |
b539b87f | 551 | bool pop_only, int *occ_pages_p) |
fbf59bc9 TH |
552 | { |
553 | int oslot = pcpu_chunk_slot(chunk); | |
554 | int max_contig = 0; | |
555 | int i, off; | |
3d331ad7 | 556 | bool seen_free = false; |
723ad1d9 | 557 | int *p; |
fbf59bc9 | 558 | |
3d331ad7 | 559 | for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) { |
fbf59bc9 | 560 | int head, tail; |
723ad1d9 AV |
561 | int this_size; |
562 | ||
563 | off = *p; | |
564 | if (off & 1) | |
565 | continue; | |
fbf59bc9 | 566 | |
723ad1d9 | 567 | this_size = (p[1] & ~1) - off; |
a16037c8 TH |
568 | |
569 | head = pcpu_fit_in_area(chunk, off, this_size, size, align, | |
570 | pop_only); | |
571 | if (head < 0) { | |
3d331ad7 AV |
572 | if (!seen_free) { |
573 | chunk->first_free = i; | |
574 | seen_free = true; | |
575 | } | |
723ad1d9 | 576 | max_contig = max(this_size, max_contig); |
fbf59bc9 TH |
577 | continue; |
578 | } | |
579 | ||
580 | /* | |
581 | * If head is small or the previous block is free, | |
582 | * merge'em. Note that 'small' is defined as smaller | |
583 | * than sizeof(int), which is very small but isn't too | |
584 | * uncommon for percpu allocations. | |
585 | */ | |
723ad1d9 | 586 | if (head && (head < sizeof(int) || !(p[-1] & 1))) { |
21ddfd38 | 587 | *p = off += head; |
723ad1d9 | 588 | if (p[-1] & 1) |
fbf59bc9 | 589 | chunk->free_size -= head; |
21ddfd38 JZ |
590 | else |
591 | max_contig = max(*p - p[-1], max_contig); | |
723ad1d9 | 592 | this_size -= head; |
fbf59bc9 TH |
593 | head = 0; |
594 | } | |
595 | ||
596 | /* if tail is small, just keep it around */ | |
723ad1d9 AV |
597 | tail = this_size - head - size; |
598 | if (tail < sizeof(int)) { | |
fbf59bc9 | 599 | tail = 0; |
723ad1d9 AV |
600 | size = this_size - head; |
601 | } | |
fbf59bc9 TH |
602 | |
603 | /* split if warranted */ | |
604 | if (head || tail) { | |
706c16f2 AV |
605 | int nr_extra = !!head + !!tail; |
606 | ||
607 | /* insert new subblocks */ | |
723ad1d9 | 608 | memmove(p + nr_extra + 1, p + 1, |
706c16f2 AV |
609 | sizeof(chunk->map[0]) * (chunk->map_used - i)); |
610 | chunk->map_used += nr_extra; | |
611 | ||
fbf59bc9 | 612 | if (head) { |
3d331ad7 AV |
613 | if (!seen_free) { |
614 | chunk->first_free = i; | |
615 | seen_free = true; | |
616 | } | |
723ad1d9 AV |
617 | *++p = off += head; |
618 | ++i; | |
706c16f2 AV |
619 | max_contig = max(head, max_contig); |
620 | } | |
621 | if (tail) { | |
723ad1d9 | 622 | p[1] = off + size; |
706c16f2 | 623 | max_contig = max(tail, max_contig); |
fbf59bc9 | 624 | } |
fbf59bc9 TH |
625 | } |
626 | ||
3d331ad7 AV |
627 | if (!seen_free) |
628 | chunk->first_free = i + 1; | |
629 | ||
fbf59bc9 | 630 | /* update hint and mark allocated */ |
723ad1d9 | 631 | if (i + 1 == chunk->map_used) |
fbf59bc9 TH |
632 | chunk->contig_hint = max_contig; /* fully scanned */ |
633 | else | |
634 | chunk->contig_hint = max(chunk->contig_hint, | |
635 | max_contig); | |
636 | ||
723ad1d9 AV |
637 | chunk->free_size -= size; |
638 | *p |= 1; | |
fbf59bc9 | 639 | |
b539b87f | 640 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
fbf59bc9 TH |
641 | pcpu_chunk_relocate(chunk, oslot); |
642 | return off; | |
643 | } | |
644 | ||
645 | chunk->contig_hint = max_contig; /* fully scanned */ | |
646 | pcpu_chunk_relocate(chunk, oslot); | |
647 | ||
9f7dcf22 TH |
648 | /* tell the upper layer that this chunk has no matching area */ |
649 | return -1; | |
fbf59bc9 TH |
650 | } |
651 | ||
652 | /** | |
653 | * pcpu_free_area - free area to a pcpu_chunk | |
654 | * @chunk: chunk of interest | |
655 | * @freeme: offset of area to free | |
b539b87f | 656 | * @occ_pages_p: out param for the number of pages the area occupies |
fbf59bc9 TH |
657 | * |
658 | * Free area starting from @freeme to @chunk. Note that this function | |
659 | * only modifies the allocation map. It doesn't depopulate or unmap | |
660 | * the area. | |
ccea34b5 TH |
661 | * |
662 | * CONTEXT: | |
663 | * pcpu_lock. | |
fbf59bc9 | 664 | */ |
b539b87f TH |
665 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme, |
666 | int *occ_pages_p) | |
fbf59bc9 TH |
667 | { |
668 | int oslot = pcpu_chunk_slot(chunk); | |
723ad1d9 AV |
669 | int off = 0; |
670 | unsigned i, j; | |
671 | int to_free = 0; | |
672 | int *p; | |
673 | ||
5ccd30e4 | 674 | lockdep_assert_held(&pcpu_lock); |
30a5b536 | 675 | pcpu_stats_area_dealloc(chunk); |
5ccd30e4 | 676 | |
723ad1d9 AV |
677 | freeme |= 1; /* we are searching for <given offset, in use> pair */ |
678 | ||
679 | i = 0; | |
680 | j = chunk->map_used; | |
681 | while (i != j) { | |
682 | unsigned k = (i + j) / 2; | |
683 | off = chunk->map[k]; | |
684 | if (off < freeme) | |
685 | i = k + 1; | |
686 | else if (off > freeme) | |
687 | j = k; | |
688 | else | |
689 | i = j = k; | |
690 | } | |
fbf59bc9 | 691 | BUG_ON(off != freeme); |
fbf59bc9 | 692 | |
3d331ad7 AV |
693 | if (i < chunk->first_free) |
694 | chunk->first_free = i; | |
695 | ||
723ad1d9 AV |
696 | p = chunk->map + i; |
697 | *p = off &= ~1; | |
698 | chunk->free_size += (p[1] & ~1) - off; | |
fbf59bc9 | 699 | |
b539b87f TH |
700 | *occ_pages_p = pcpu_count_occupied_pages(chunk, i); |
701 | ||
723ad1d9 AV |
702 | /* merge with next? */ |
703 | if (!(p[1] & 1)) | |
704 | to_free++; | |
fbf59bc9 | 705 | /* merge with previous? */ |
723ad1d9 AV |
706 | if (i > 0 && !(p[-1] & 1)) { |
707 | to_free++; | |
fbf59bc9 | 708 | i--; |
723ad1d9 | 709 | p--; |
fbf59bc9 | 710 | } |
723ad1d9 AV |
711 | if (to_free) { |
712 | chunk->map_used -= to_free; | |
713 | memmove(p + 1, p + 1 + to_free, | |
714 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
fbf59bc9 TH |
715 | } |
716 | ||
723ad1d9 | 717 | chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint); |
fbf59bc9 TH |
718 | pcpu_chunk_relocate(chunk, oslot); |
719 | } | |
720 | ||
c0ebfdc3 | 721 | static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr, |
10edf5b0 DZF |
722 | int map_size, |
723 | int *map, | |
724 | int init_map_size) | |
725 | { | |
726 | struct pcpu_chunk *chunk; | |
c0ebfdc3 DZF |
727 | unsigned long aligned_addr; |
728 | int start_offset, region_size; | |
729 | ||
730 | /* region calculations */ | |
731 | aligned_addr = tmp_addr & PAGE_MASK; | |
732 | ||
733 | start_offset = tmp_addr - aligned_addr; | |
6b9d7c8e DZF |
734 | |
735 | region_size = PFN_ALIGN(start_offset + map_size); | |
10edf5b0 | 736 | |
c0ebfdc3 | 737 | /* allocate chunk */ |
8ab16c43 DZF |
738 | chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) + |
739 | BITS_TO_LONGS(region_size >> PAGE_SHIFT), | |
740 | 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 | 748 | |
8ab16c43 | 749 | chunk->nr_pages = region_size >> PAGE_SHIFT; |
c0ebfdc3 | 750 | |
10edf5b0 DZF |
751 | chunk->map = map; |
752 | chunk->map_alloc = init_map_size; | |
753 | ||
754 | /* manage populated page bitmap */ | |
755 | chunk->immutable = true; | |
8ab16c43 DZF |
756 | bitmap_fill(chunk->populated, chunk->nr_pages); |
757 | chunk->nr_populated = chunk->nr_pages; | |
0cecf50c | 758 | chunk->nr_empty_pop_pages = chunk->nr_pages; |
10edf5b0 DZF |
759 | |
760 | chunk->contig_hint = chunk->free_size = map_size; | |
c0ebfdc3 DZF |
761 | |
762 | if (chunk->start_offset) { | |
763 | /* hide the beginning of the bitmap */ | |
0cecf50c DZF |
764 | chunk->nr_empty_pop_pages--; |
765 | ||
c0ebfdc3 DZF |
766 | chunk->map[0] = 1; |
767 | chunk->map[1] = chunk->start_offset; | |
768 | chunk->map_used = 1; | |
769 | } | |
770 | ||
771 | /* set chunk's free region */ | |
772 | chunk->map[++chunk->map_used] = | |
773 | (chunk->start_offset + chunk->free_size) | 1; | |
10edf5b0 | 774 | |
6b9d7c8e DZF |
775 | if (chunk->end_offset) { |
776 | /* hide the end of the bitmap */ | |
0cecf50c DZF |
777 | chunk->nr_empty_pop_pages--; |
778 | ||
6b9d7c8e DZF |
779 | chunk->map[++chunk->map_used] = region_size | 1; |
780 | } | |
781 | ||
10edf5b0 DZF |
782 | return chunk; |
783 | } | |
784 | ||
6081089f TH |
785 | static struct pcpu_chunk *pcpu_alloc_chunk(void) |
786 | { | |
787 | struct pcpu_chunk *chunk; | |
788 | ||
90459ce0 | 789 | chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size); |
6081089f TH |
790 | if (!chunk) |
791 | return NULL; | |
792 | ||
90459ce0 BL |
793 | chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC * |
794 | sizeof(chunk->map[0])); | |
6081089f | 795 | if (!chunk->map) { |
1d5cfdb0 | 796 | pcpu_mem_free(chunk); |
6081089f TH |
797 | return NULL; |
798 | } | |
799 | ||
800 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; | |
723ad1d9 AV |
801 | chunk->map[0] = 0; |
802 | chunk->map[1] = pcpu_unit_size | 1; | |
803 | chunk->map_used = 1; | |
6081089f TH |
804 | |
805 | INIT_LIST_HEAD(&chunk->list); | |
4f996e23 | 806 | INIT_LIST_HEAD(&chunk->map_extend_list); |
6081089f TH |
807 | chunk->free_size = pcpu_unit_size; |
808 | chunk->contig_hint = pcpu_unit_size; | |
809 | ||
c0ebfdc3 DZF |
810 | chunk->nr_pages = pcpu_unit_pages; |
811 | ||
6081089f TH |
812 | return chunk; |
813 | } | |
814 | ||
815 | static void pcpu_free_chunk(struct pcpu_chunk *chunk) | |
816 | { | |
817 | if (!chunk) | |
818 | return; | |
1d5cfdb0 TH |
819 | pcpu_mem_free(chunk->map); |
820 | pcpu_mem_free(chunk); | |
6081089f TH |
821 | } |
822 | ||
b539b87f TH |
823 | /** |
824 | * pcpu_chunk_populated - post-population bookkeeping | |
825 | * @chunk: pcpu_chunk which got populated | |
826 | * @page_start: the start page | |
827 | * @page_end: the end page | |
828 | * | |
829 | * Pages in [@page_start,@page_end) have been populated to @chunk. Update | |
830 | * the bookkeeping information accordingly. Must be called after each | |
831 | * successful population. | |
832 | */ | |
833 | static void pcpu_chunk_populated(struct pcpu_chunk *chunk, | |
834 | int page_start, int page_end) | |
835 | { | |
836 | int nr = page_end - page_start; | |
837 | ||
838 | lockdep_assert_held(&pcpu_lock); | |
839 | ||
840 | bitmap_set(chunk->populated, page_start, nr); | |
841 | chunk->nr_populated += nr; | |
0cecf50c | 842 | chunk->nr_empty_pop_pages += nr; |
b539b87f TH |
843 | pcpu_nr_empty_pop_pages += nr; |
844 | } | |
845 | ||
846 | /** | |
847 | * pcpu_chunk_depopulated - post-depopulation bookkeeping | |
848 | * @chunk: pcpu_chunk which got depopulated | |
849 | * @page_start: the start page | |
850 | * @page_end: the end page | |
851 | * | |
852 | * Pages in [@page_start,@page_end) have been depopulated from @chunk. | |
853 | * Update the bookkeeping information accordingly. Must be called after | |
854 | * each successful depopulation. | |
855 | */ | |
856 | static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk, | |
857 | int page_start, int page_end) | |
858 | { | |
859 | int nr = page_end - page_start; | |
860 | ||
861 | lockdep_assert_held(&pcpu_lock); | |
862 | ||
863 | bitmap_clear(chunk->populated, page_start, nr); | |
864 | chunk->nr_populated -= nr; | |
0cecf50c | 865 | chunk->nr_empty_pop_pages -= nr; |
b539b87f TH |
866 | pcpu_nr_empty_pop_pages -= nr; |
867 | } | |
868 | ||
9f645532 TH |
869 | /* |
870 | * Chunk management implementation. | |
871 | * | |
872 | * To allow different implementations, chunk alloc/free and | |
873 | * [de]population are implemented in a separate file which is pulled | |
874 | * into this file and compiled together. The following functions | |
875 | * should be implemented. | |
876 | * | |
877 | * pcpu_populate_chunk - populate the specified range of a chunk | |
878 | * pcpu_depopulate_chunk - depopulate the specified range of a chunk | |
879 | * pcpu_create_chunk - create a new chunk | |
880 | * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop | |
881 | * pcpu_addr_to_page - translate address to physical address | |
882 | * pcpu_verify_alloc_info - check alloc_info is acceptable during init | |
fbf59bc9 | 883 | */ |
9f645532 TH |
884 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size); |
885 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size); | |
886 | static struct pcpu_chunk *pcpu_create_chunk(void); | |
887 | static void pcpu_destroy_chunk(struct pcpu_chunk *chunk); | |
888 | static struct page *pcpu_addr_to_page(void *addr); | |
889 | static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai); | |
fbf59bc9 | 890 | |
b0c9778b TH |
891 | #ifdef CONFIG_NEED_PER_CPU_KM |
892 | #include "percpu-km.c" | |
893 | #else | |
9f645532 | 894 | #include "percpu-vm.c" |
b0c9778b | 895 | #endif |
fbf59bc9 | 896 | |
88999a89 TH |
897 | /** |
898 | * pcpu_chunk_addr_search - determine chunk containing specified address | |
899 | * @addr: address for which the chunk needs to be determined. | |
900 | * | |
c0ebfdc3 DZF |
901 | * This is an internal function that handles all but static allocations. |
902 | * Static percpu address values should never be passed into the allocator. | |
903 | * | |
88999a89 TH |
904 | * RETURNS: |
905 | * The address of the found chunk. | |
906 | */ | |
907 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
908 | { | |
c0ebfdc3 | 909 | /* is it in the dynamic region (first chunk)? */ |
560f2c23 | 910 | if (pcpu_addr_in_chunk(pcpu_first_chunk, addr)) |
88999a89 | 911 | return pcpu_first_chunk; |
c0ebfdc3 DZF |
912 | |
913 | /* is it in the reserved region? */ | |
560f2c23 | 914 | if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr)) |
c0ebfdc3 | 915 | return pcpu_reserved_chunk; |
88999a89 TH |
916 | |
917 | /* | |
918 | * The address is relative to unit0 which might be unused and | |
919 | * thus unmapped. Offset the address to the unit space of the | |
920 | * current processor before looking it up in the vmalloc | |
921 | * space. Note that any possible cpu id can be used here, so | |
922 | * there's no need to worry about preemption or cpu hotplug. | |
923 | */ | |
924 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; | |
9f645532 | 925 | return pcpu_get_page_chunk(pcpu_addr_to_page(addr)); |
88999a89 TH |
926 | } |
927 | ||
fbf59bc9 | 928 | /** |
edcb4639 | 929 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 930 | * @size: size of area to allocate in bytes |
fbf59bc9 | 931 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 932 | * @reserved: allocate from the reserved chunk if available |
5835d96e | 933 | * @gfp: allocation flags |
fbf59bc9 | 934 | * |
5835d96e TH |
935 | * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't |
936 | * contain %GFP_KERNEL, the allocation is atomic. | |
fbf59bc9 TH |
937 | * |
938 | * RETURNS: | |
939 | * Percpu pointer to the allocated area on success, NULL on failure. | |
940 | */ | |
5835d96e TH |
941 | static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved, |
942 | gfp_t gfp) | |
fbf59bc9 | 943 | { |
f2badb0c | 944 | static int warn_limit = 10; |
fbf59bc9 | 945 | struct pcpu_chunk *chunk; |
f2badb0c | 946 | const char *err; |
6ae833c7 | 947 | bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL; |
b539b87f | 948 | int occ_pages = 0; |
b38d08f3 | 949 | int slot, off, new_alloc, cpu, ret; |
403a91b1 | 950 | unsigned long flags; |
f528f0b8 | 951 | void __percpu *ptr; |
fbf59bc9 | 952 | |
723ad1d9 AV |
953 | /* |
954 | * We want the lowest bit of offset available for in-use/free | |
2f69fa82 | 955 | * indicator, so force >= 16bit alignment and make size even. |
723ad1d9 | 956 | */ |
d2f3c384 DZF |
957 | if (unlikely(align < PCPU_MIN_ALLOC_SIZE)) |
958 | align = PCPU_MIN_ALLOC_SIZE; | |
723ad1d9 | 959 | |
d2f3c384 | 960 | size = ALIGN(size, PCPU_MIN_ALLOC_SIZE); |
2f69fa82 | 961 | |
3ca45a46 | 962 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE || |
963 | !is_power_of_2(align))) { | |
756a025f JP |
964 | WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n", |
965 | size, align); | |
fbf59bc9 TH |
966 | return NULL; |
967 | } | |
968 | ||
6710e594 TH |
969 | if (!is_atomic) |
970 | mutex_lock(&pcpu_alloc_mutex); | |
971 | ||
403a91b1 | 972 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 | 973 | |
edcb4639 TH |
974 | /* serve reserved allocations from the reserved chunk if available */ |
975 | if (reserved && pcpu_reserved_chunk) { | |
976 | chunk = pcpu_reserved_chunk; | |
833af842 TH |
977 | |
978 | if (size > chunk->contig_hint) { | |
979 | err = "alloc from reserved chunk failed"; | |
ccea34b5 | 980 | goto fail_unlock; |
f2badb0c | 981 | } |
833af842 | 982 | |
9c824b6a | 983 | while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) { |
833af842 | 984 | spin_unlock_irqrestore(&pcpu_lock, flags); |
5835d96e TH |
985 | if (is_atomic || |
986 | pcpu_extend_area_map(chunk, new_alloc) < 0) { | |
833af842 | 987 | err = "failed to extend area map of reserved chunk"; |
b38d08f3 | 988 | goto fail; |
833af842 TH |
989 | } |
990 | spin_lock_irqsave(&pcpu_lock, flags); | |
991 | } | |
992 | ||
b539b87f TH |
993 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
994 | &occ_pages); | |
edcb4639 TH |
995 | if (off >= 0) |
996 | goto area_found; | |
833af842 | 997 | |
f2badb0c | 998 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 999 | goto fail_unlock; |
edcb4639 TH |
1000 | } |
1001 | ||
ccea34b5 | 1002 | restart: |
edcb4639 | 1003 | /* search through normal chunks */ |
fbf59bc9 TH |
1004 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
1005 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
1006 | if (size > chunk->contig_hint) | |
1007 | continue; | |
ccea34b5 | 1008 | |
9c824b6a | 1009 | new_alloc = pcpu_need_to_extend(chunk, is_atomic); |
833af842 | 1010 | if (new_alloc) { |
5835d96e TH |
1011 | if (is_atomic) |
1012 | continue; | |
833af842 TH |
1013 | spin_unlock_irqrestore(&pcpu_lock, flags); |
1014 | if (pcpu_extend_area_map(chunk, | |
1015 | new_alloc) < 0) { | |
1016 | err = "failed to extend area map"; | |
b38d08f3 | 1017 | goto fail; |
833af842 TH |
1018 | } |
1019 | spin_lock_irqsave(&pcpu_lock, flags); | |
1020 | /* | |
1021 | * pcpu_lock has been dropped, need to | |
1022 | * restart cpu_slot list walking. | |
1023 | */ | |
1024 | goto restart; | |
ccea34b5 TH |
1025 | } |
1026 | ||
b539b87f TH |
1027 | off = pcpu_alloc_area(chunk, size, align, is_atomic, |
1028 | &occ_pages); | |
fbf59bc9 TH |
1029 | if (off >= 0) |
1030 | goto area_found; | |
fbf59bc9 TH |
1031 | } |
1032 | } | |
1033 | ||
403a91b1 | 1034 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1035 | |
b38d08f3 TH |
1036 | /* |
1037 | * No space left. Create a new chunk. We don't want multiple | |
1038 | * tasks to create chunks simultaneously. Serialize and create iff | |
1039 | * there's still no empty chunk after grabbing the mutex. | |
1040 | */ | |
11df02bf DZ |
1041 | if (is_atomic) { |
1042 | err = "atomic alloc failed, no space left"; | |
5835d96e | 1043 | goto fail; |
11df02bf | 1044 | } |
5835d96e | 1045 | |
b38d08f3 TH |
1046 | if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) { |
1047 | chunk = pcpu_create_chunk(); | |
1048 | if (!chunk) { | |
1049 | err = "failed to allocate new chunk"; | |
1050 | goto fail; | |
1051 | } | |
1052 | ||
1053 | spin_lock_irqsave(&pcpu_lock, flags); | |
1054 | pcpu_chunk_relocate(chunk, -1); | |
1055 | } else { | |
1056 | spin_lock_irqsave(&pcpu_lock, flags); | |
f2badb0c | 1057 | } |
ccea34b5 | 1058 | |
ccea34b5 | 1059 | goto restart; |
fbf59bc9 TH |
1060 | |
1061 | area_found: | |
30a5b536 | 1062 | pcpu_stats_area_alloc(chunk, size); |
403a91b1 | 1063 | spin_unlock_irqrestore(&pcpu_lock, flags); |
ccea34b5 | 1064 | |
dca49645 | 1065 | /* populate if not all pages are already there */ |
5835d96e | 1066 | if (!is_atomic) { |
e04d3208 | 1067 | int page_start, page_end, rs, re; |
dca49645 | 1068 | |
e04d3208 TH |
1069 | page_start = PFN_DOWN(off); |
1070 | page_end = PFN_UP(off + size); | |
b38d08f3 | 1071 | |
91e914c5 DZF |
1072 | pcpu_for_each_unpop_region(chunk->populated, rs, re, |
1073 | page_start, page_end) { | |
e04d3208 TH |
1074 | WARN_ON(chunk->immutable); |
1075 | ||
1076 | ret = pcpu_populate_chunk(chunk, rs, re); | |
1077 | ||
1078 | spin_lock_irqsave(&pcpu_lock, flags); | |
1079 | if (ret) { | |
b539b87f | 1080 | pcpu_free_area(chunk, off, &occ_pages); |
e04d3208 TH |
1081 | err = "failed to populate"; |
1082 | goto fail_unlock; | |
1083 | } | |
b539b87f | 1084 | pcpu_chunk_populated(chunk, rs, re); |
e04d3208 | 1085 | spin_unlock_irqrestore(&pcpu_lock, flags); |
dca49645 | 1086 | } |
fbf59bc9 | 1087 | |
e04d3208 TH |
1088 | mutex_unlock(&pcpu_alloc_mutex); |
1089 | } | |
ccea34b5 | 1090 | |
320661b0 TE |
1091 | if (chunk != pcpu_reserved_chunk) { |
1092 | spin_lock_irqsave(&pcpu_lock, flags); | |
b539b87f | 1093 | pcpu_nr_empty_pop_pages -= occ_pages; |
320661b0 TE |
1094 | spin_unlock_irqrestore(&pcpu_lock, flags); |
1095 | } | |
b539b87f | 1096 | |
1a4d7607 TH |
1097 | if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW) |
1098 | pcpu_schedule_balance_work(); | |
1099 | ||
dca49645 TH |
1100 | /* clear the areas and return address relative to base address */ |
1101 | for_each_possible_cpu(cpu) | |
1102 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); | |
1103 | ||
f528f0b8 | 1104 | ptr = __addr_to_pcpu_ptr(chunk->base_addr + off); |
8a8c35fa | 1105 | kmemleak_alloc_percpu(ptr, size, gfp); |
df95e795 DZ |
1106 | |
1107 | trace_percpu_alloc_percpu(reserved, is_atomic, size, align, | |
1108 | chunk->base_addr, off, ptr); | |
1109 | ||
f528f0b8 | 1110 | return ptr; |
ccea34b5 TH |
1111 | |
1112 | fail_unlock: | |
403a91b1 | 1113 | spin_unlock_irqrestore(&pcpu_lock, flags); |
b38d08f3 | 1114 | fail: |
df95e795 DZ |
1115 | trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align); |
1116 | ||
5835d96e | 1117 | if (!is_atomic && warn_limit) { |
870d4b12 | 1118 | pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n", |
598d8091 | 1119 | size, align, is_atomic, err); |
f2badb0c TH |
1120 | dump_stack(); |
1121 | if (!--warn_limit) | |
870d4b12 | 1122 | pr_info("limit reached, disable warning\n"); |
f2badb0c | 1123 | } |
1a4d7607 TH |
1124 | if (is_atomic) { |
1125 | /* see the flag handling in pcpu_blance_workfn() */ | |
1126 | pcpu_atomic_alloc_failed = true; | |
1127 | pcpu_schedule_balance_work(); | |
6710e594 TH |
1128 | } else { |
1129 | mutex_unlock(&pcpu_alloc_mutex); | |
1a4d7607 | 1130 | } |
ccea34b5 | 1131 | return NULL; |
fbf59bc9 | 1132 | } |
edcb4639 TH |
1133 | |
1134 | /** | |
5835d96e | 1135 | * __alloc_percpu_gfp - allocate dynamic percpu area |
edcb4639 TH |
1136 | * @size: size of area to allocate in bytes |
1137 | * @align: alignment of area (max PAGE_SIZE) | |
5835d96e | 1138 | * @gfp: allocation flags |
edcb4639 | 1139 | * |
5835d96e TH |
1140 | * Allocate zero-filled percpu area of @size bytes aligned at @align. If |
1141 | * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can | |
1142 | * be called from any context but is a lot more likely to fail. | |
ccea34b5 | 1143 | * |
edcb4639 TH |
1144 | * RETURNS: |
1145 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1146 | */ | |
5835d96e TH |
1147 | void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp) |
1148 | { | |
1149 | return pcpu_alloc(size, align, false, gfp); | |
1150 | } | |
1151 | EXPORT_SYMBOL_GPL(__alloc_percpu_gfp); | |
1152 | ||
1153 | /** | |
1154 | * __alloc_percpu - allocate dynamic percpu area | |
1155 | * @size: size of area to allocate in bytes | |
1156 | * @align: alignment of area (max PAGE_SIZE) | |
1157 | * | |
1158 | * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL). | |
1159 | */ | |
43cf38eb | 1160 | void __percpu *__alloc_percpu(size_t size, size_t align) |
edcb4639 | 1161 | { |
5835d96e | 1162 | return pcpu_alloc(size, align, false, GFP_KERNEL); |
edcb4639 | 1163 | } |
fbf59bc9 TH |
1164 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1165 | ||
edcb4639 TH |
1166 | /** |
1167 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1168 | * @size: size of area to allocate in bytes | |
1169 | * @align: alignment of area (max PAGE_SIZE) | |
1170 | * | |
9329ba97 TH |
1171 | * Allocate zero-filled percpu area of @size bytes aligned at @align |
1172 | * from reserved percpu area if arch has set it up; otherwise, | |
1173 | * allocation is served from the same dynamic area. Might sleep. | |
1174 | * Might trigger writeouts. | |
edcb4639 | 1175 | * |
ccea34b5 TH |
1176 | * CONTEXT: |
1177 | * Does GFP_KERNEL allocation. | |
1178 | * | |
edcb4639 TH |
1179 | * RETURNS: |
1180 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1181 | */ | |
43cf38eb | 1182 | void __percpu *__alloc_reserved_percpu(size_t size, size_t align) |
edcb4639 | 1183 | { |
5835d96e | 1184 | return pcpu_alloc(size, align, true, GFP_KERNEL); |
edcb4639 TH |
1185 | } |
1186 | ||
a56dbddf | 1187 | /** |
1a4d7607 | 1188 | * pcpu_balance_workfn - manage the amount of free chunks and populated pages |
a56dbddf TH |
1189 | * @work: unused |
1190 | * | |
1191 | * Reclaim all fully free chunks except for the first one. | |
1192 | */ | |
fe6bd8c3 | 1193 | static void pcpu_balance_workfn(struct work_struct *work) |
fbf59bc9 | 1194 | { |
fe6bd8c3 TH |
1195 | LIST_HEAD(to_free); |
1196 | struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1]; | |
a56dbddf | 1197 | struct pcpu_chunk *chunk, *next; |
1a4d7607 | 1198 | int slot, nr_to_pop, ret; |
a56dbddf | 1199 | |
1a4d7607 TH |
1200 | /* |
1201 | * There's no reason to keep around multiple unused chunks and VM | |
1202 | * areas can be scarce. Destroy all free chunks except for one. | |
1203 | */ | |
ccea34b5 TH |
1204 | mutex_lock(&pcpu_alloc_mutex); |
1205 | spin_lock_irq(&pcpu_lock); | |
a56dbddf | 1206 | |
fe6bd8c3 | 1207 | list_for_each_entry_safe(chunk, next, free_head, list) { |
a56dbddf TH |
1208 | WARN_ON(chunk->immutable); |
1209 | ||
1210 | /* spare the first one */ | |
fe6bd8c3 | 1211 | if (chunk == list_first_entry(free_head, struct pcpu_chunk, list)) |
a56dbddf TH |
1212 | continue; |
1213 | ||
4f996e23 | 1214 | list_del_init(&chunk->map_extend_list); |
fe6bd8c3 | 1215 | list_move(&chunk->list, &to_free); |
a56dbddf TH |
1216 | } |
1217 | ||
ccea34b5 | 1218 | spin_unlock_irq(&pcpu_lock); |
a56dbddf | 1219 | |
fe6bd8c3 | 1220 | list_for_each_entry_safe(chunk, next, &to_free, list) { |
a93ace48 | 1221 | int rs, re; |
dca49645 | 1222 | |
91e914c5 DZF |
1223 | pcpu_for_each_pop_region(chunk->populated, rs, re, 0, |
1224 | chunk->nr_pages) { | |
a93ace48 | 1225 | pcpu_depopulate_chunk(chunk, rs, re); |
b539b87f TH |
1226 | spin_lock_irq(&pcpu_lock); |
1227 | pcpu_chunk_depopulated(chunk, rs, re); | |
1228 | spin_unlock_irq(&pcpu_lock); | |
a93ace48 | 1229 | } |
6081089f | 1230 | pcpu_destroy_chunk(chunk); |
a56dbddf | 1231 | } |
971f3918 | 1232 | |
4f996e23 TH |
1233 | /* service chunks which requested async area map extension */ |
1234 | do { | |
1235 | int new_alloc = 0; | |
1236 | ||
1237 | spin_lock_irq(&pcpu_lock); | |
1238 | ||
1239 | chunk = list_first_entry_or_null(&pcpu_map_extend_chunks, | |
1240 | struct pcpu_chunk, map_extend_list); | |
1241 | if (chunk) { | |
1242 | list_del_init(&chunk->map_extend_list); | |
1243 | new_alloc = pcpu_need_to_extend(chunk, false); | |
1244 | } | |
1245 | ||
1246 | spin_unlock_irq(&pcpu_lock); | |
1247 | ||
1248 | if (new_alloc) | |
1249 | pcpu_extend_area_map(chunk, new_alloc); | |
1250 | } while (chunk); | |
1251 | ||
1a4d7607 TH |
1252 | /* |
1253 | * Ensure there are certain number of free populated pages for | |
1254 | * atomic allocs. Fill up from the most packed so that atomic | |
1255 | * allocs don't increase fragmentation. If atomic allocation | |
1256 | * failed previously, always populate the maximum amount. This | |
1257 | * should prevent atomic allocs larger than PAGE_SIZE from keeping | |
1258 | * failing indefinitely; however, large atomic allocs are not | |
1259 | * something we support properly and can be highly unreliable and | |
1260 | * inefficient. | |
1261 | */ | |
1262 | retry_pop: | |
1263 | if (pcpu_atomic_alloc_failed) { | |
1264 | nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH; | |
1265 | /* best effort anyway, don't worry about synchronization */ | |
1266 | pcpu_atomic_alloc_failed = false; | |
1267 | } else { | |
1268 | nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH - | |
1269 | pcpu_nr_empty_pop_pages, | |
1270 | 0, PCPU_EMPTY_POP_PAGES_HIGH); | |
1271 | } | |
1272 | ||
1273 | for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) { | |
1274 | int nr_unpop = 0, rs, re; | |
1275 | ||
1276 | if (!nr_to_pop) | |
1277 | break; | |
1278 | ||
1279 | spin_lock_irq(&pcpu_lock); | |
1280 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
8ab16c43 | 1281 | nr_unpop = chunk->nr_pages - chunk->nr_populated; |
1a4d7607 TH |
1282 | if (nr_unpop) |
1283 | break; | |
1284 | } | |
1285 | spin_unlock_irq(&pcpu_lock); | |
1286 | ||
1287 | if (!nr_unpop) | |
1288 | continue; | |
1289 | ||
1290 | /* @chunk can't go away while pcpu_alloc_mutex is held */ | |
91e914c5 DZF |
1291 | pcpu_for_each_unpop_region(chunk->populated, rs, re, 0, |
1292 | chunk->nr_pages) { | |
1a4d7607 TH |
1293 | int nr = min(re - rs, nr_to_pop); |
1294 | ||
1295 | ret = pcpu_populate_chunk(chunk, rs, rs + nr); | |
1296 | if (!ret) { | |
1297 | nr_to_pop -= nr; | |
1298 | spin_lock_irq(&pcpu_lock); | |
1299 | pcpu_chunk_populated(chunk, rs, rs + nr); | |
1300 | spin_unlock_irq(&pcpu_lock); | |
1301 | } else { | |
1302 | nr_to_pop = 0; | |
1303 | } | |
1304 | ||
1305 | if (!nr_to_pop) | |
1306 | break; | |
1307 | } | |
1308 | } | |
1309 | ||
1310 | if (nr_to_pop) { | |
1311 | /* ran out of chunks to populate, create a new one and retry */ | |
1312 | chunk = pcpu_create_chunk(); | |
1313 | if (chunk) { | |
1314 | spin_lock_irq(&pcpu_lock); | |
1315 | pcpu_chunk_relocate(chunk, -1); | |
1316 | spin_unlock_irq(&pcpu_lock); | |
1317 | goto retry_pop; | |
1318 | } | |
1319 | } | |
1320 | ||
971f3918 | 1321 | mutex_unlock(&pcpu_alloc_mutex); |
fbf59bc9 TH |
1322 | } |
1323 | ||
1324 | /** | |
1325 | * free_percpu - free percpu area | |
1326 | * @ptr: pointer to area to free | |
1327 | * | |
ccea34b5 TH |
1328 | * Free percpu area @ptr. |
1329 | * | |
1330 | * CONTEXT: | |
1331 | * Can be called from atomic context. | |
fbf59bc9 | 1332 | */ |
43cf38eb | 1333 | void free_percpu(void __percpu *ptr) |
fbf59bc9 | 1334 | { |
129182e5 | 1335 | void *addr; |
fbf59bc9 | 1336 | struct pcpu_chunk *chunk; |
ccea34b5 | 1337 | unsigned long flags; |
b539b87f | 1338 | int off, occ_pages; |
fbf59bc9 TH |
1339 | |
1340 | if (!ptr) | |
1341 | return; | |
1342 | ||
f528f0b8 CM |
1343 | kmemleak_free_percpu(ptr); |
1344 | ||
129182e5 AM |
1345 | addr = __pcpu_ptr_to_addr(ptr); |
1346 | ||
ccea34b5 | 1347 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1348 | |
1349 | chunk = pcpu_chunk_addr_search(addr); | |
bba174f5 | 1350 | off = addr - chunk->base_addr; |
fbf59bc9 | 1351 | |
b539b87f TH |
1352 | pcpu_free_area(chunk, off, &occ_pages); |
1353 | ||
1354 | if (chunk != pcpu_reserved_chunk) | |
1355 | pcpu_nr_empty_pop_pages += occ_pages; | |
fbf59bc9 | 1356 | |
a56dbddf | 1357 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
1358 | if (chunk->free_size == pcpu_unit_size) { |
1359 | struct pcpu_chunk *pos; | |
1360 | ||
a56dbddf | 1361 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1362 | if (pos != chunk) { |
1a4d7607 | 1363 | pcpu_schedule_balance_work(); |
fbf59bc9 TH |
1364 | break; |
1365 | } | |
1366 | } | |
1367 | ||
df95e795 DZ |
1368 | trace_percpu_free_percpu(chunk->base_addr, off, ptr); |
1369 | ||
ccea34b5 | 1370 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1371 | } |
1372 | EXPORT_SYMBOL_GPL(free_percpu); | |
1373 | ||
383776fa | 1374 | bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr) |
10fad5e4 | 1375 | { |
bbddff05 | 1376 | #ifdef CONFIG_SMP |
10fad5e4 TH |
1377 | const size_t static_size = __per_cpu_end - __per_cpu_start; |
1378 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); | |
1379 | unsigned int cpu; | |
1380 | ||
1381 | for_each_possible_cpu(cpu) { | |
1382 | void *start = per_cpu_ptr(base, cpu); | |
383776fa | 1383 | void *va = (void *)addr; |
10fad5e4 | 1384 | |
383776fa | 1385 | if (va >= start && va < start + static_size) { |
8ce371f9 | 1386 | if (can_addr) { |
383776fa | 1387 | *can_addr = (unsigned long) (va - start); |
8ce371f9 PZ |
1388 | *can_addr += (unsigned long) |
1389 | per_cpu_ptr(base, get_boot_cpu_id()); | |
1390 | } | |
10fad5e4 | 1391 | return true; |
383776fa TG |
1392 | } |
1393 | } | |
bbddff05 TH |
1394 | #endif |
1395 | /* on UP, can't distinguish from other static vars, always false */ | |
10fad5e4 TH |
1396 | return false; |
1397 | } | |
1398 | ||
383776fa TG |
1399 | /** |
1400 | * is_kernel_percpu_address - test whether address is from static percpu area | |
1401 | * @addr: address to test | |
1402 | * | |
1403 | * Test whether @addr belongs to in-kernel static percpu area. Module | |
1404 | * static percpu areas are not considered. For those, use | |
1405 | * is_module_percpu_address(). | |
1406 | * | |
1407 | * RETURNS: | |
1408 | * %true if @addr is from in-kernel static percpu area, %false otherwise. | |
1409 | */ | |
1410 | bool is_kernel_percpu_address(unsigned long addr) | |
1411 | { | |
1412 | return __is_kernel_percpu_address(addr, NULL); | |
1413 | } | |
1414 | ||
3b034b0d VG |
1415 | /** |
1416 | * per_cpu_ptr_to_phys - convert translated percpu address to physical address | |
1417 | * @addr: the address to be converted to physical address | |
1418 | * | |
1419 | * Given @addr which is dereferenceable address obtained via one of | |
1420 | * percpu access macros, this function translates it into its physical | |
1421 | * address. The caller is responsible for ensuring @addr stays valid | |
1422 | * until this function finishes. | |
1423 | * | |
67589c71 DY |
1424 | * percpu allocator has special setup for the first chunk, which currently |
1425 | * supports either embedding in linear address space or vmalloc mapping, | |
1426 | * and, from the second one, the backing allocator (currently either vm or | |
1427 | * km) provides translation. | |
1428 | * | |
bffc4375 | 1429 | * The addr can be translated simply without checking if it falls into the |
67589c71 DY |
1430 | * first chunk. But the current code reflects better how percpu allocator |
1431 | * actually works, and the verification can discover both bugs in percpu | |
1432 | * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current | |
1433 | * code. | |
1434 | * | |
3b034b0d VG |
1435 | * RETURNS: |
1436 | * The physical address for @addr. | |
1437 | */ | |
1438 | phys_addr_t per_cpu_ptr_to_phys(void *addr) | |
1439 | { | |
9983b6f0 TH |
1440 | void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr); |
1441 | bool in_first_chunk = false; | |
a855b84c | 1442 | unsigned long first_low, first_high; |
9983b6f0 TH |
1443 | unsigned int cpu; |
1444 | ||
1445 | /* | |
a855b84c | 1446 | * The following test on unit_low/high isn't strictly |
9983b6f0 TH |
1447 | * necessary but will speed up lookups of addresses which |
1448 | * aren't in the first chunk. | |
c0ebfdc3 DZF |
1449 | * |
1450 | * The address check is against full chunk sizes. pcpu_base_addr | |
1451 | * points to the beginning of the first chunk including the | |
1452 | * static region. Assumes good intent as the first chunk may | |
1453 | * not be full (ie. < pcpu_unit_pages in size). | |
9983b6f0 | 1454 | */ |
c0ebfdc3 DZF |
1455 | first_low = (unsigned long)pcpu_base_addr + |
1456 | pcpu_unit_page_offset(pcpu_low_unit_cpu, 0); | |
1457 | first_high = (unsigned long)pcpu_base_addr + | |
1458 | pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages); | |
a855b84c TH |
1459 | if ((unsigned long)addr >= first_low && |
1460 | (unsigned long)addr < first_high) { | |
9983b6f0 TH |
1461 | for_each_possible_cpu(cpu) { |
1462 | void *start = per_cpu_ptr(base, cpu); | |
1463 | ||
1464 | if (addr >= start && addr < start + pcpu_unit_size) { | |
1465 | in_first_chunk = true; | |
1466 | break; | |
1467 | } | |
1468 | } | |
1469 | } | |
1470 | ||
1471 | if (in_first_chunk) { | |
eac522ef | 1472 | if (!is_vmalloc_addr(addr)) |
020ec653 TH |
1473 | return __pa(addr); |
1474 | else | |
9f57bd4d ES |
1475 | return page_to_phys(vmalloc_to_page(addr)) + |
1476 | offset_in_page(addr); | |
020ec653 | 1477 | } else |
9f57bd4d ES |
1478 | return page_to_phys(pcpu_addr_to_page(addr)) + |
1479 | offset_in_page(addr); | |
3b034b0d VG |
1480 | } |
1481 | ||
fbf59bc9 | 1482 | /** |
fd1e8a1f TH |
1483 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
1484 | * @nr_groups: the number of groups | |
1485 | * @nr_units: the number of units | |
1486 | * | |
1487 | * Allocate ai which is large enough for @nr_groups groups containing | |
1488 | * @nr_units units. The returned ai's groups[0].cpu_map points to the | |
1489 | * cpu_map array which is long enough for @nr_units and filled with | |
1490 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map | |
1491 | * pointer of other groups. | |
1492 | * | |
1493 | * RETURNS: | |
1494 | * Pointer to the allocated pcpu_alloc_info on success, NULL on | |
1495 | * failure. | |
1496 | */ | |
1497 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, | |
1498 | int nr_units) | |
1499 | { | |
1500 | struct pcpu_alloc_info *ai; | |
1501 | size_t base_size, ai_size; | |
1502 | void *ptr; | |
1503 | int unit; | |
1504 | ||
1505 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), | |
1506 | __alignof__(ai->groups[0].cpu_map[0])); | |
1507 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); | |
1508 | ||
999c17e3 | 1509 | ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0); |
fd1e8a1f TH |
1510 | if (!ptr) |
1511 | return NULL; | |
1512 | ai = ptr; | |
1513 | ptr += base_size; | |
1514 | ||
1515 | ai->groups[0].cpu_map = ptr; | |
1516 | ||
1517 | for (unit = 0; unit < nr_units; unit++) | |
1518 | ai->groups[0].cpu_map[unit] = NR_CPUS; | |
1519 | ||
1520 | ai->nr_groups = nr_groups; | |
1521 | ai->__ai_size = PFN_ALIGN(ai_size); | |
1522 | ||
1523 | return ai; | |
1524 | } | |
1525 | ||
1526 | /** | |
1527 | * pcpu_free_alloc_info - free percpu allocation info | |
1528 | * @ai: pcpu_alloc_info to free | |
1529 | * | |
1530 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). | |
1531 | */ | |
1532 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) | |
1533 | { | |
999c17e3 | 1534 | memblock_free_early(__pa(ai), ai->__ai_size); |
fd1e8a1f TH |
1535 | } |
1536 | ||
fd1e8a1f TH |
1537 | /** |
1538 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info | |
1539 | * @lvl: loglevel | |
1540 | * @ai: allocation info to dump | |
1541 | * | |
1542 | * Print out information about @ai using loglevel @lvl. | |
1543 | */ | |
1544 | static void pcpu_dump_alloc_info(const char *lvl, | |
1545 | const struct pcpu_alloc_info *ai) | |
033e48fb | 1546 | { |
fd1e8a1f | 1547 | int group_width = 1, cpu_width = 1, width; |
033e48fb | 1548 | char empty_str[] = "--------"; |
fd1e8a1f TH |
1549 | int alloc = 0, alloc_end = 0; |
1550 | int group, v; | |
1551 | int upa, apl; /* units per alloc, allocs per line */ | |
1552 | ||
1553 | v = ai->nr_groups; | |
1554 | while (v /= 10) | |
1555 | group_width++; | |
033e48fb | 1556 | |
fd1e8a1f | 1557 | v = num_possible_cpus(); |
033e48fb | 1558 | while (v /= 10) |
fd1e8a1f TH |
1559 | cpu_width++; |
1560 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; | |
033e48fb | 1561 | |
fd1e8a1f TH |
1562 | upa = ai->alloc_size / ai->unit_size; |
1563 | width = upa * (cpu_width + 1) + group_width + 3; | |
1564 | apl = rounddown_pow_of_two(max(60 / width, 1)); | |
033e48fb | 1565 | |
fd1e8a1f TH |
1566 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
1567 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, | |
1568 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); | |
033e48fb | 1569 | |
fd1e8a1f TH |
1570 | for (group = 0; group < ai->nr_groups; group++) { |
1571 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
1572 | int unit = 0, unit_end = 0; | |
1573 | ||
1574 | BUG_ON(gi->nr_units % upa); | |
1575 | for (alloc_end += gi->nr_units / upa; | |
1576 | alloc < alloc_end; alloc++) { | |
1577 | if (!(alloc % apl)) { | |
1170532b | 1578 | pr_cont("\n"); |
fd1e8a1f TH |
1579 | printk("%spcpu-alloc: ", lvl); |
1580 | } | |
1170532b | 1581 | pr_cont("[%0*d] ", group_width, group); |
fd1e8a1f TH |
1582 | |
1583 | for (unit_end += upa; unit < unit_end; unit++) | |
1584 | if (gi->cpu_map[unit] != NR_CPUS) | |
1170532b JP |
1585 | pr_cont("%0*d ", |
1586 | cpu_width, gi->cpu_map[unit]); | |
fd1e8a1f | 1587 | else |
1170532b | 1588 | pr_cont("%s ", empty_str); |
033e48fb | 1589 | } |
033e48fb | 1590 | } |
1170532b | 1591 | pr_cont("\n"); |
033e48fb | 1592 | } |
033e48fb | 1593 | |
fbf59bc9 | 1594 | /** |
8d408b4b | 1595 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
fd1e8a1f | 1596 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
38a6be52 | 1597 | * @base_addr: mapped address |
8d408b4b TH |
1598 | * |
1599 | * Initialize the first percpu chunk which contains the kernel static | |
1600 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1601 | * setup path. |
8d408b4b | 1602 | * |
fd1e8a1f TH |
1603 | * @ai contains all information necessary to initialize the first |
1604 | * chunk and prime the dynamic percpu allocator. | |
1605 | * | |
1606 | * @ai->static_size is the size of static percpu area. | |
1607 | * | |
1608 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to | |
edcb4639 TH |
1609 | * reserve after the static area in the first chunk. This reserves |
1610 | * the first chunk such that it's available only through reserved | |
1611 | * percpu allocation. This is primarily used to serve module percpu | |
1612 | * static areas on architectures where the addressing model has | |
1613 | * limited offset range for symbol relocations to guarantee module | |
1614 | * percpu symbols fall inside the relocatable range. | |
1615 | * | |
fd1e8a1f TH |
1616 | * @ai->dyn_size determines the number of bytes available for dynamic |
1617 | * allocation in the first chunk. The area between @ai->static_size + | |
1618 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. | |
6074d5b0 | 1619 | * |
fd1e8a1f TH |
1620 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
1621 | * and equal to or larger than @ai->static_size + @ai->reserved_size + | |
1622 | * @ai->dyn_size. | |
8d408b4b | 1623 | * |
fd1e8a1f TH |
1624 | * @ai->atom_size is the allocation atom size and used as alignment |
1625 | * for vm areas. | |
8d408b4b | 1626 | * |
fd1e8a1f TH |
1627 | * @ai->alloc_size is the allocation size and always multiple of |
1628 | * @ai->atom_size. This is larger than @ai->atom_size if | |
1629 | * @ai->unit_size is larger than @ai->atom_size. | |
1630 | * | |
1631 | * @ai->nr_groups and @ai->groups describe virtual memory layout of | |
1632 | * percpu areas. Units which should be colocated are put into the | |
1633 | * same group. Dynamic VM areas will be allocated according to these | |
1634 | * groupings. If @ai->nr_groups is zero, a single group containing | |
1635 | * all units is assumed. | |
8d408b4b | 1636 | * |
38a6be52 TH |
1637 | * The caller should have mapped the first chunk at @base_addr and |
1638 | * copied static data to each unit. | |
fbf59bc9 | 1639 | * |
c0ebfdc3 DZF |
1640 | * The first chunk will always contain a static and a dynamic region. |
1641 | * However, the static region is not managed by any chunk. If the first | |
1642 | * chunk also contains a reserved region, it is served by two chunks - | |
1643 | * one for the reserved region and one for the dynamic region. They | |
1644 | * share the same vm, but use offset regions in the area allocation map. | |
1645 | * The chunk serving the dynamic region is circulated in the chunk slots | |
1646 | * and available for dynamic allocation like any other chunk. | |
edcb4639 | 1647 | * |
fbf59bc9 | 1648 | * RETURNS: |
fb435d52 | 1649 | * 0 on success, -errno on failure. |
fbf59bc9 | 1650 | */ |
fb435d52 TH |
1651 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
1652 | void *base_addr) | |
fbf59bc9 | 1653 | { |
099a19d9 TH |
1654 | static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; |
1655 | static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata; | |
b9c39442 | 1656 | size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
d2f3c384 | 1657 | size_t static_size, dyn_size; |
0c4169c3 | 1658 | struct pcpu_chunk *chunk; |
6563297c TH |
1659 | unsigned long *group_offsets; |
1660 | size_t *group_sizes; | |
fb435d52 | 1661 | unsigned long *unit_off; |
fbf59bc9 | 1662 | unsigned int cpu; |
fd1e8a1f TH |
1663 | int *unit_map; |
1664 | int group, unit, i; | |
c0ebfdc3 DZF |
1665 | int map_size; |
1666 | unsigned long tmp_addr; | |
fbf59bc9 | 1667 | |
635b75fc TH |
1668 | #define PCPU_SETUP_BUG_ON(cond) do { \ |
1669 | if (unlikely(cond)) { \ | |
870d4b12 JP |
1670 | pr_emerg("failed to initialize, %s\n", #cond); \ |
1671 | pr_emerg("cpu_possible_mask=%*pb\n", \ | |
807de073 | 1672 | cpumask_pr_args(cpu_possible_mask)); \ |
635b75fc TH |
1673 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ |
1674 | BUG(); \ | |
1675 | } \ | |
1676 | } while (0) | |
1677 | ||
2f39e637 | 1678 | /* sanity checks */ |
635b75fc | 1679 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
bbddff05 | 1680 | #ifdef CONFIG_SMP |
635b75fc | 1681 | PCPU_SETUP_BUG_ON(!ai->static_size); |
f09f1243 | 1682 | PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start)); |
bbddff05 | 1683 | #endif |
635b75fc | 1684 | PCPU_SETUP_BUG_ON(!base_addr); |
f09f1243 | 1685 | PCPU_SETUP_BUG_ON(offset_in_page(base_addr)); |
635b75fc | 1686 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); |
f09f1243 | 1687 | PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size)); |
635b75fc | 1688 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); |
099a19d9 | 1689 | PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE); |
fb29a2cc | 1690 | PCPU_SETUP_BUG_ON(!ai->dyn_size); |
d2f3c384 | 1691 | PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE)); |
9f645532 | 1692 | PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0); |
8d408b4b | 1693 | |
6563297c | 1694 | /* process group information and build config tables accordingly */ |
999c17e3 SS |
1695 | group_offsets = memblock_virt_alloc(ai->nr_groups * |
1696 | sizeof(group_offsets[0]), 0); | |
1697 | group_sizes = memblock_virt_alloc(ai->nr_groups * | |
1698 | sizeof(group_sizes[0]), 0); | |
1699 | unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0); | |
1700 | unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0); | |
2f39e637 | 1701 | |
fd1e8a1f | 1702 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
ffe0d5a5 | 1703 | unit_map[cpu] = UINT_MAX; |
a855b84c TH |
1704 | |
1705 | pcpu_low_unit_cpu = NR_CPUS; | |
1706 | pcpu_high_unit_cpu = NR_CPUS; | |
2f39e637 | 1707 | |
fd1e8a1f TH |
1708 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
1709 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
2f39e637 | 1710 | |
6563297c TH |
1711 | group_offsets[group] = gi->base_offset; |
1712 | group_sizes[group] = gi->nr_units * ai->unit_size; | |
1713 | ||
fd1e8a1f TH |
1714 | for (i = 0; i < gi->nr_units; i++) { |
1715 | cpu = gi->cpu_map[i]; | |
1716 | if (cpu == NR_CPUS) | |
1717 | continue; | |
8d408b4b | 1718 | |
9f295664 | 1719 | PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids); |
635b75fc TH |
1720 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); |
1721 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); | |
fbf59bc9 | 1722 | |
fd1e8a1f | 1723 | unit_map[cpu] = unit + i; |
fb435d52 TH |
1724 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
1725 | ||
a855b84c TH |
1726 | /* determine low/high unit_cpu */ |
1727 | if (pcpu_low_unit_cpu == NR_CPUS || | |
1728 | unit_off[cpu] < unit_off[pcpu_low_unit_cpu]) | |
1729 | pcpu_low_unit_cpu = cpu; | |
1730 | if (pcpu_high_unit_cpu == NR_CPUS || | |
1731 | unit_off[cpu] > unit_off[pcpu_high_unit_cpu]) | |
1732 | pcpu_high_unit_cpu = cpu; | |
fd1e8a1f | 1733 | } |
2f39e637 | 1734 | } |
fd1e8a1f TH |
1735 | pcpu_nr_units = unit; |
1736 | ||
1737 | for_each_possible_cpu(cpu) | |
635b75fc TH |
1738 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
1739 | ||
1740 | /* we're done parsing the input, undefine BUG macro and dump config */ | |
1741 | #undef PCPU_SETUP_BUG_ON | |
bcbea798 | 1742 | pcpu_dump_alloc_info(KERN_DEBUG, ai); |
fd1e8a1f | 1743 | |
6563297c TH |
1744 | pcpu_nr_groups = ai->nr_groups; |
1745 | pcpu_group_offsets = group_offsets; | |
1746 | pcpu_group_sizes = group_sizes; | |
fd1e8a1f | 1747 | pcpu_unit_map = unit_map; |
fb435d52 | 1748 | pcpu_unit_offsets = unit_off; |
2f39e637 TH |
1749 | |
1750 | /* determine basic parameters */ | |
fd1e8a1f | 1751 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
d9b55eeb | 1752 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
6563297c | 1753 | pcpu_atom_size = ai->atom_size; |
ce3141a2 TH |
1754 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
1755 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
cafe8816 | 1756 | |
30a5b536 DZ |
1757 | pcpu_stats_save_ai(ai); |
1758 | ||
d9b55eeb TH |
1759 | /* |
1760 | * Allocate chunk slots. The additional last slot is for | |
1761 | * empty chunks. | |
1762 | */ | |
1763 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
999c17e3 SS |
1764 | pcpu_slot = memblock_virt_alloc( |
1765 | pcpu_nr_slots * sizeof(pcpu_slot[0]), 0); | |
fbf59bc9 TH |
1766 | for (i = 0; i < pcpu_nr_slots; i++) |
1767 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1768 | ||
d2f3c384 DZF |
1769 | /* |
1770 | * The end of the static region needs to be aligned with the | |
1771 | * minimum allocation size as this offsets the reserved and | |
1772 | * dynamic region. The first chunk ends page aligned by | |
1773 | * expanding the dynamic region, therefore the dynamic region | |
1774 | * can be shrunk to compensate while still staying above the | |
1775 | * configured sizes. | |
1776 | */ | |
1777 | static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE); | |
1778 | dyn_size = ai->dyn_size - (static_size - ai->static_size); | |
1779 | ||
edcb4639 | 1780 | /* |
c0ebfdc3 DZF |
1781 | * Initialize first chunk. |
1782 | * If the reserved_size is non-zero, this initializes the reserved | |
1783 | * chunk. If the reserved_size is zero, the reserved chunk is NULL | |
1784 | * and the dynamic region is initialized here. The first chunk, | |
1785 | * pcpu_first_chunk, will always point to the chunk that serves | |
1786 | * the dynamic region. | |
edcb4639 | 1787 | */ |
d2f3c384 DZF |
1788 | tmp_addr = (unsigned long)base_addr + static_size; |
1789 | map_size = ai->reserved_size ?: dyn_size; | |
c0ebfdc3 | 1790 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size, smap, |
0c4169c3 | 1791 | ARRAY_SIZE(smap)); |
61ace7fa | 1792 | |
edcb4639 | 1793 | /* init dynamic chunk if necessary */ |
b9c39442 | 1794 | if (ai->reserved_size) { |
0c4169c3 | 1795 | pcpu_reserved_chunk = chunk; |
b9c39442 | 1796 | |
d2f3c384 | 1797 | tmp_addr = (unsigned long)base_addr + static_size + |
c0ebfdc3 | 1798 | ai->reserved_size; |
d2f3c384 | 1799 | map_size = dyn_size; |
c0ebfdc3 | 1800 | chunk = pcpu_alloc_first_chunk(tmp_addr, map_size, dmap, |
0c4169c3 | 1801 | ARRAY_SIZE(dmap)); |
edcb4639 TH |
1802 | } |
1803 | ||
2441d15c | 1804 | /* link the first chunk in */ |
0c4169c3 | 1805 | pcpu_first_chunk = chunk; |
0cecf50c | 1806 | pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages; |
ae9e6bc9 | 1807 | pcpu_chunk_relocate(pcpu_first_chunk, -1); |
fbf59bc9 | 1808 | |
30a5b536 | 1809 | pcpu_stats_chunk_alloc(); |
df95e795 | 1810 | trace_percpu_create_chunk(base_addr); |
30a5b536 | 1811 | |
fbf59bc9 | 1812 | /* we're done */ |
bba174f5 | 1813 | pcpu_base_addr = base_addr; |
fb435d52 | 1814 | return 0; |
fbf59bc9 | 1815 | } |
66c3a757 | 1816 | |
bbddff05 TH |
1817 | #ifdef CONFIG_SMP |
1818 | ||
17f3609c | 1819 | const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = { |
f58dc01b TH |
1820 | [PCPU_FC_AUTO] = "auto", |
1821 | [PCPU_FC_EMBED] = "embed", | |
1822 | [PCPU_FC_PAGE] = "page", | |
f58dc01b | 1823 | }; |
66c3a757 | 1824 | |
f58dc01b | 1825 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
66c3a757 | 1826 | |
f58dc01b TH |
1827 | static int __init percpu_alloc_setup(char *str) |
1828 | { | |
5479c78a CG |
1829 | if (!str) |
1830 | return -EINVAL; | |
1831 | ||
f58dc01b TH |
1832 | if (0) |
1833 | /* nada */; | |
1834 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK | |
1835 | else if (!strcmp(str, "embed")) | |
1836 | pcpu_chosen_fc = PCPU_FC_EMBED; | |
1837 | #endif | |
1838 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
1839 | else if (!strcmp(str, "page")) | |
1840 | pcpu_chosen_fc = PCPU_FC_PAGE; | |
f58dc01b TH |
1841 | #endif |
1842 | else | |
870d4b12 | 1843 | pr_warn("unknown allocator %s specified\n", str); |
66c3a757 | 1844 | |
f58dc01b | 1845 | return 0; |
66c3a757 | 1846 | } |
f58dc01b | 1847 | early_param("percpu_alloc", percpu_alloc_setup); |
66c3a757 | 1848 | |
3c9a024f TH |
1849 | /* |
1850 | * pcpu_embed_first_chunk() is used by the generic percpu setup. | |
1851 | * Build it if needed by the arch config or the generic setup is going | |
1852 | * to be used. | |
1853 | */ | |
08fc4580 TH |
1854 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
1855 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
3c9a024f TH |
1856 | #define BUILD_EMBED_FIRST_CHUNK |
1857 | #endif | |
1858 | ||
1859 | /* build pcpu_page_first_chunk() iff needed by the arch config */ | |
1860 | #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK) | |
1861 | #define BUILD_PAGE_FIRST_CHUNK | |
1862 | #endif | |
1863 | ||
1864 | /* pcpu_build_alloc_info() is used by both embed and page first chunk */ | |
1865 | #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK) | |
1866 | /** | |
1867 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs | |
1868 | * @reserved_size: the size of reserved percpu area in bytes | |
1869 | * @dyn_size: minimum free size for dynamic allocation in bytes | |
1870 | * @atom_size: allocation atom size | |
1871 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
1872 | * | |
1873 | * This function determines grouping of units, their mappings to cpus | |
1874 | * and other parameters considering needed percpu size, allocation | |
1875 | * atom size and distances between CPUs. | |
1876 | * | |
bffc4375 | 1877 | * Groups are always multiples of atom size and CPUs which are of |
3c9a024f TH |
1878 | * LOCAL_DISTANCE both ways are grouped together and share space for |
1879 | * units in the same group. The returned configuration is guaranteed | |
1880 | * to have CPUs on different nodes on different groups and >=75% usage | |
1881 | * of allocated virtual address space. | |
1882 | * | |
1883 | * RETURNS: | |
1884 | * On success, pointer to the new allocation_info is returned. On | |
1885 | * failure, ERR_PTR value is returned. | |
1886 | */ | |
1887 | static struct pcpu_alloc_info * __init pcpu_build_alloc_info( | |
1888 | size_t reserved_size, size_t dyn_size, | |
1889 | size_t atom_size, | |
1890 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
1891 | { | |
1892 | static int group_map[NR_CPUS] __initdata; | |
1893 | static int group_cnt[NR_CPUS] __initdata; | |
1894 | const size_t static_size = __per_cpu_end - __per_cpu_start; | |
1895 | int nr_groups = 1, nr_units = 0; | |
1896 | size_t size_sum, min_unit_size, alloc_size; | |
1897 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
1898 | int last_allocs, group, unit; | |
1899 | unsigned int cpu, tcpu; | |
1900 | struct pcpu_alloc_info *ai; | |
1901 | unsigned int *cpu_map; | |
1902 | ||
1903 | /* this function may be called multiple times */ | |
1904 | memset(group_map, 0, sizeof(group_map)); | |
1905 | memset(group_cnt, 0, sizeof(group_cnt)); | |
1906 | ||
1907 | /* calculate size_sum and ensure dyn_size is enough for early alloc */ | |
1908 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
1909 | max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE)); | |
1910 | dyn_size = size_sum - static_size - reserved_size; | |
1911 | ||
1912 | /* | |
1913 | * Determine min_unit_size, alloc_size and max_upa such that | |
1914 | * alloc_size is multiple of atom_size and is the smallest | |
25985edc | 1915 | * which can accommodate 4k aligned segments which are equal to |
3c9a024f TH |
1916 | * or larger than min_unit_size. |
1917 | */ | |
1918 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
1919 | ||
9c015162 | 1920 | /* determine the maximum # of units that can fit in an allocation */ |
3c9a024f TH |
1921 | alloc_size = roundup(min_unit_size, atom_size); |
1922 | upa = alloc_size / min_unit_size; | |
f09f1243 | 1923 | while (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
1924 | upa--; |
1925 | max_upa = upa; | |
1926 | ||
1927 | /* group cpus according to their proximity */ | |
1928 | for_each_possible_cpu(cpu) { | |
1929 | group = 0; | |
1930 | next_group: | |
1931 | for_each_possible_cpu(tcpu) { | |
1932 | if (cpu == tcpu) | |
1933 | break; | |
1934 | if (group_map[tcpu] == group && cpu_distance_fn && | |
1935 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || | |
1936 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
1937 | group++; | |
1938 | nr_groups = max(nr_groups, group + 1); | |
1939 | goto next_group; | |
1940 | } | |
1941 | } | |
1942 | group_map[cpu] = group; | |
1943 | group_cnt[group]++; | |
1944 | } | |
1945 | ||
1946 | /* | |
9c015162 DZF |
1947 | * Wasted space is caused by a ratio imbalance of upa to group_cnt. |
1948 | * Expand the unit_size until we use >= 75% of the units allocated. | |
1949 | * Related to atom_size, which could be much larger than the unit_size. | |
3c9a024f TH |
1950 | */ |
1951 | last_allocs = INT_MAX; | |
1952 | for (upa = max_upa; upa; upa--) { | |
1953 | int allocs = 0, wasted = 0; | |
1954 | ||
f09f1243 | 1955 | if (alloc_size % upa || (offset_in_page(alloc_size / upa))) |
3c9a024f TH |
1956 | continue; |
1957 | ||
1958 | for (group = 0; group < nr_groups; group++) { | |
1959 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); | |
1960 | allocs += this_allocs; | |
1961 | wasted += this_allocs * upa - group_cnt[group]; | |
1962 | } | |
1963 | ||
1964 | /* | |
1965 | * Don't accept if wastage is over 1/3. The | |
1966 | * greater-than comparison ensures upa==1 always | |
1967 | * passes the following check. | |
1968 | */ | |
1969 | if (wasted > num_possible_cpus() / 3) | |
1970 | continue; | |
1971 | ||
1972 | /* and then don't consume more memory */ | |
1973 | if (allocs > last_allocs) | |
1974 | break; | |
1975 | last_allocs = allocs; | |
1976 | best_upa = upa; | |
1977 | } | |
1978 | upa = best_upa; | |
1979 | ||
1980 | /* allocate and fill alloc_info */ | |
1981 | for (group = 0; group < nr_groups; group++) | |
1982 | nr_units += roundup(group_cnt[group], upa); | |
1983 | ||
1984 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); | |
1985 | if (!ai) | |
1986 | return ERR_PTR(-ENOMEM); | |
1987 | cpu_map = ai->groups[0].cpu_map; | |
1988 | ||
1989 | for (group = 0; group < nr_groups; group++) { | |
1990 | ai->groups[group].cpu_map = cpu_map; | |
1991 | cpu_map += roundup(group_cnt[group], upa); | |
1992 | } | |
1993 | ||
1994 | ai->static_size = static_size; | |
1995 | ai->reserved_size = reserved_size; | |
1996 | ai->dyn_size = dyn_size; | |
1997 | ai->unit_size = alloc_size / upa; | |
1998 | ai->atom_size = atom_size; | |
1999 | ai->alloc_size = alloc_size; | |
2000 | ||
2001 | for (group = 0, unit = 0; group_cnt[group]; group++) { | |
2002 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2003 | ||
2004 | /* | |
2005 | * Initialize base_offset as if all groups are located | |
2006 | * back-to-back. The caller should update this to | |
2007 | * reflect actual allocation. | |
2008 | */ | |
2009 | gi->base_offset = unit * ai->unit_size; | |
2010 | ||
2011 | for_each_possible_cpu(cpu) | |
2012 | if (group_map[cpu] == group) | |
2013 | gi->cpu_map[gi->nr_units++] = cpu; | |
2014 | gi->nr_units = roundup(gi->nr_units, upa); | |
2015 | unit += gi->nr_units; | |
2016 | } | |
2017 | BUG_ON(unit != nr_units); | |
2018 | ||
2019 | return ai; | |
2020 | } | |
2021 | #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */ | |
2022 | ||
2023 | #if defined(BUILD_EMBED_FIRST_CHUNK) | |
66c3a757 TH |
2024 | /** |
2025 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
66c3a757 | 2026 | * @reserved_size: the size of reserved percpu area in bytes |
4ba6ce25 | 2027 | * @dyn_size: minimum free size for dynamic allocation in bytes |
c8826dd5 TH |
2028 | * @atom_size: allocation atom size |
2029 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
2030 | * @alloc_fn: function to allocate percpu page | |
25985edc | 2031 | * @free_fn: function to free percpu page |
66c3a757 TH |
2032 | * |
2033 | * This is a helper to ease setting up embedded first percpu chunk and | |
2034 | * can be called where pcpu_setup_first_chunk() is expected. | |
2035 | * | |
2036 | * If this function is used to setup the first chunk, it is allocated | |
c8826dd5 TH |
2037 | * by calling @alloc_fn and used as-is without being mapped into |
2038 | * vmalloc area. Allocations are always whole multiples of @atom_size | |
2039 | * aligned to @atom_size. | |
2040 | * | |
2041 | * This enables the first chunk to piggy back on the linear physical | |
2042 | * mapping which often uses larger page size. Please note that this | |
2043 | * can result in very sparse cpu->unit mapping on NUMA machines thus | |
2044 | * requiring large vmalloc address space. Don't use this allocator if | |
2045 | * vmalloc space is not orders of magnitude larger than distances | |
2046 | * between node memory addresses (ie. 32bit NUMA machines). | |
66c3a757 | 2047 | * |
4ba6ce25 | 2048 | * @dyn_size specifies the minimum dynamic area size. |
66c3a757 TH |
2049 | * |
2050 | * If the needed size is smaller than the minimum or specified unit | |
c8826dd5 | 2051 | * size, the leftover is returned using @free_fn. |
66c3a757 TH |
2052 | * |
2053 | * RETURNS: | |
fb435d52 | 2054 | * 0 on success, -errno on failure. |
66c3a757 | 2055 | */ |
4ba6ce25 | 2056 | int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size, |
c8826dd5 TH |
2057 | size_t atom_size, |
2058 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, | |
2059 | pcpu_fc_alloc_fn_t alloc_fn, | |
2060 | pcpu_fc_free_fn_t free_fn) | |
66c3a757 | 2061 | { |
c8826dd5 TH |
2062 | void *base = (void *)ULONG_MAX; |
2063 | void **areas = NULL; | |
fd1e8a1f | 2064 | struct pcpu_alloc_info *ai; |
93c76b6b | 2065 | size_t size_sum, areas_size; |
2066 | unsigned long max_distance; | |
9b739662 | 2067 | int group, i, highest_group, rc; |
66c3a757 | 2068 | |
c8826dd5 TH |
2069 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
2070 | cpu_distance_fn); | |
fd1e8a1f TH |
2071 | if (IS_ERR(ai)) |
2072 | return PTR_ERR(ai); | |
66c3a757 | 2073 | |
fd1e8a1f | 2074 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
c8826dd5 | 2075 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
fa8a7094 | 2076 | |
999c17e3 | 2077 | areas = memblock_virt_alloc_nopanic(areas_size, 0); |
c8826dd5 | 2078 | if (!areas) { |
fb435d52 | 2079 | rc = -ENOMEM; |
c8826dd5 | 2080 | goto out_free; |
fa8a7094 | 2081 | } |
66c3a757 | 2082 | |
9b739662 | 2083 | /* allocate, copy and determine base address & max_distance */ |
2084 | highest_group = 0; | |
c8826dd5 TH |
2085 | for (group = 0; group < ai->nr_groups; group++) { |
2086 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2087 | unsigned int cpu = NR_CPUS; | |
2088 | void *ptr; | |
2089 | ||
2090 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) | |
2091 | cpu = gi->cpu_map[i]; | |
2092 | BUG_ON(cpu == NR_CPUS); | |
2093 | ||
2094 | /* allocate space for the whole group */ | |
2095 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); | |
2096 | if (!ptr) { | |
2097 | rc = -ENOMEM; | |
2098 | goto out_free_areas; | |
2099 | } | |
f528f0b8 CM |
2100 | /* kmemleak tracks the percpu allocations separately */ |
2101 | kmemleak_free(ptr); | |
c8826dd5 | 2102 | areas[group] = ptr; |
fd1e8a1f | 2103 | |
c8826dd5 | 2104 | base = min(ptr, base); |
9b739662 | 2105 | if (ptr > areas[highest_group]) |
2106 | highest_group = group; | |
2107 | } | |
2108 | max_distance = areas[highest_group] - base; | |
2109 | max_distance += ai->unit_size * ai->groups[highest_group].nr_units; | |
2110 | ||
2111 | /* warn if maximum distance is further than 75% of vmalloc space */ | |
2112 | if (max_distance > VMALLOC_TOTAL * 3 / 4) { | |
2113 | pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n", | |
2114 | max_distance, VMALLOC_TOTAL); | |
2115 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
2116 | /* and fail if we have fallback */ | |
2117 | rc = -EINVAL; | |
2118 | goto out_free_areas; | |
2119 | #endif | |
42b64281 TH |
2120 | } |
2121 | ||
2122 | /* | |
2123 | * Copy data and free unused parts. This should happen after all | |
2124 | * allocations are complete; otherwise, we may end up with | |
2125 | * overlapping groups. | |
2126 | */ | |
2127 | for (group = 0; group < ai->nr_groups; group++) { | |
2128 | struct pcpu_group_info *gi = &ai->groups[group]; | |
2129 | void *ptr = areas[group]; | |
c8826dd5 TH |
2130 | |
2131 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { | |
2132 | if (gi->cpu_map[i] == NR_CPUS) { | |
2133 | /* unused unit, free whole */ | |
2134 | free_fn(ptr, ai->unit_size); | |
2135 | continue; | |
2136 | } | |
2137 | /* copy and return the unused part */ | |
2138 | memcpy(ptr, __per_cpu_load, ai->static_size); | |
2139 | free_fn(ptr + size_sum, ai->unit_size - size_sum); | |
2140 | } | |
fa8a7094 | 2141 | } |
66c3a757 | 2142 | |
c8826dd5 | 2143 | /* base address is now known, determine group base offsets */ |
6ea529a2 | 2144 | for (group = 0; group < ai->nr_groups; group++) { |
c8826dd5 | 2145 | ai->groups[group].base_offset = areas[group] - base; |
6ea529a2 | 2146 | } |
c8826dd5 | 2147 | |
870d4b12 | 2148 | pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
fd1e8a1f TH |
2149 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
2150 | ai->dyn_size, ai->unit_size); | |
d4b95f80 | 2151 | |
fb435d52 | 2152 | rc = pcpu_setup_first_chunk(ai, base); |
c8826dd5 TH |
2153 | goto out_free; |
2154 | ||
2155 | out_free_areas: | |
2156 | for (group = 0; group < ai->nr_groups; group++) | |
f851c8d8 MH |
2157 | if (areas[group]) |
2158 | free_fn(areas[group], | |
2159 | ai->groups[group].nr_units * ai->unit_size); | |
c8826dd5 | 2160 | out_free: |
fd1e8a1f | 2161 | pcpu_free_alloc_info(ai); |
c8826dd5 | 2162 | if (areas) |
999c17e3 | 2163 | memblock_free_early(__pa(areas), areas_size); |
fb435d52 | 2164 | return rc; |
d4b95f80 | 2165 | } |
3c9a024f | 2166 | #endif /* BUILD_EMBED_FIRST_CHUNK */ |
d4b95f80 | 2167 | |
3c9a024f | 2168 | #ifdef BUILD_PAGE_FIRST_CHUNK |
d4b95f80 | 2169 | /** |
00ae4064 | 2170 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
2171 | * @reserved_size: the size of reserved percpu area in bytes |
2172 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
25985edc | 2173 | * @free_fn: function to free percpu page, always called with PAGE_SIZE |
d4b95f80 TH |
2174 | * @populate_pte_fn: function to populate pte |
2175 | * | |
00ae4064 TH |
2176 | * This is a helper to ease setting up page-remapped first percpu |
2177 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
2178 | * |
2179 | * This is the basic allocator. Static percpu area is allocated | |
2180 | * page-by-page into vmalloc area. | |
2181 | * | |
2182 | * RETURNS: | |
fb435d52 | 2183 | * 0 on success, -errno on failure. |
d4b95f80 | 2184 | */ |
fb435d52 TH |
2185 | int __init pcpu_page_first_chunk(size_t reserved_size, |
2186 | pcpu_fc_alloc_fn_t alloc_fn, | |
2187 | pcpu_fc_free_fn_t free_fn, | |
2188 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 2189 | { |
8f05a6a6 | 2190 | static struct vm_struct vm; |
fd1e8a1f | 2191 | struct pcpu_alloc_info *ai; |
00ae4064 | 2192 | char psize_str[16]; |
ce3141a2 | 2193 | int unit_pages; |
d4b95f80 | 2194 | size_t pages_size; |
ce3141a2 | 2195 | struct page **pages; |
fb435d52 | 2196 | int unit, i, j, rc; |
8f606604 | 2197 | int upa; |
2198 | int nr_g0_units; | |
d4b95f80 | 2199 | |
00ae4064 TH |
2200 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
2201 | ||
4ba6ce25 | 2202 | ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL); |
fd1e8a1f TH |
2203 | if (IS_ERR(ai)) |
2204 | return PTR_ERR(ai); | |
2205 | BUG_ON(ai->nr_groups != 1); | |
8f606604 | 2206 | upa = ai->alloc_size/ai->unit_size; |
2207 | nr_g0_units = roundup(num_possible_cpus(), upa); | |
2208 | if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) { | |
2209 | pcpu_free_alloc_info(ai); | |
2210 | return -EINVAL; | |
2211 | } | |
fd1e8a1f TH |
2212 | |
2213 | unit_pages = ai->unit_size >> PAGE_SHIFT; | |
d4b95f80 TH |
2214 | |
2215 | /* unaligned allocations can't be freed, round up to page size */ | |
fd1e8a1f TH |
2216 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
2217 | sizeof(pages[0])); | |
999c17e3 | 2218 | pages = memblock_virt_alloc(pages_size, 0); |
d4b95f80 | 2219 | |
8f05a6a6 | 2220 | /* allocate pages */ |
d4b95f80 | 2221 | j = 0; |
8f606604 | 2222 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
2223 | unsigned int cpu = ai->groups[0].cpu_map[unit]; | |
ce3141a2 | 2224 | for (i = 0; i < unit_pages; i++) { |
d4b95f80 TH |
2225 | void *ptr; |
2226 | ||
3cbc8565 | 2227 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
d4b95f80 | 2228 | if (!ptr) { |
870d4b12 | 2229 | pr_warn("failed to allocate %s page for cpu%u\n", |
8f606604 | 2230 | psize_str, cpu); |
d4b95f80 TH |
2231 | goto enomem; |
2232 | } | |
f528f0b8 CM |
2233 | /* kmemleak tracks the percpu allocations separately */ |
2234 | kmemleak_free(ptr); | |
ce3141a2 | 2235 | pages[j++] = virt_to_page(ptr); |
d4b95f80 | 2236 | } |
8f606604 | 2237 | } |
d4b95f80 | 2238 | |
8f05a6a6 TH |
2239 | /* allocate vm area, map the pages and copy static data */ |
2240 | vm.flags = VM_ALLOC; | |
fd1e8a1f | 2241 | vm.size = num_possible_cpus() * ai->unit_size; |
8f05a6a6 TH |
2242 | vm_area_register_early(&vm, PAGE_SIZE); |
2243 | ||
fd1e8a1f | 2244 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
1d9d3257 | 2245 | unsigned long unit_addr = |
fd1e8a1f | 2246 | (unsigned long)vm.addr + unit * ai->unit_size; |
8f05a6a6 | 2247 | |
ce3141a2 | 2248 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
2249 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
2250 | ||
2251 | /* pte already populated, the following shouldn't fail */ | |
fb435d52 TH |
2252 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
2253 | unit_pages); | |
2254 | if (rc < 0) | |
2255 | panic("failed to map percpu area, err=%d\n", rc); | |
66c3a757 | 2256 | |
8f05a6a6 TH |
2257 | /* |
2258 | * FIXME: Archs with virtual cache should flush local | |
2259 | * cache for the linear mapping here - something | |
2260 | * equivalent to flush_cache_vmap() on the local cpu. | |
2261 | * flush_cache_vmap() can't be used as most supporting | |
2262 | * data structures are not set up yet. | |
2263 | */ | |
2264 | ||
2265 | /* copy static data */ | |
fd1e8a1f | 2266 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
66c3a757 TH |
2267 | } |
2268 | ||
2269 | /* we're ready, commit */ | |
870d4b12 | 2270 | pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n", |
fd1e8a1f TH |
2271 | unit_pages, psize_str, vm.addr, ai->static_size, |
2272 | ai->reserved_size, ai->dyn_size); | |
d4b95f80 | 2273 | |
fb435d52 | 2274 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
d4b95f80 TH |
2275 | goto out_free_ar; |
2276 | ||
2277 | enomem: | |
2278 | while (--j >= 0) | |
ce3141a2 | 2279 | free_fn(page_address(pages[j]), PAGE_SIZE); |
fb435d52 | 2280 | rc = -ENOMEM; |
d4b95f80 | 2281 | out_free_ar: |
999c17e3 | 2282 | memblock_free_early(__pa(pages), pages_size); |
fd1e8a1f | 2283 | pcpu_free_alloc_info(ai); |
fb435d52 | 2284 | return rc; |
d4b95f80 | 2285 | } |
3c9a024f | 2286 | #endif /* BUILD_PAGE_FIRST_CHUNK */ |
d4b95f80 | 2287 | |
bbddff05 | 2288 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA |
e74e3962 | 2289 | /* |
bbddff05 | 2290 | * Generic SMP percpu area setup. |
e74e3962 TH |
2291 | * |
2292 | * The embedding helper is used because its behavior closely resembles | |
2293 | * the original non-dynamic generic percpu area setup. This is | |
2294 | * important because many archs have addressing restrictions and might | |
2295 | * fail if the percpu area is located far away from the previous | |
2296 | * location. As an added bonus, in non-NUMA cases, embedding is | |
2297 | * generally a good idea TLB-wise because percpu area can piggy back | |
2298 | * on the physical linear memory mapping which uses large page | |
2299 | * mappings on applicable archs. | |
2300 | */ | |
e74e3962 TH |
2301 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; |
2302 | EXPORT_SYMBOL(__per_cpu_offset); | |
2303 | ||
c8826dd5 TH |
2304 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
2305 | size_t align) | |
2306 | { | |
999c17e3 SS |
2307 | return memblock_virt_alloc_from_nopanic( |
2308 | size, align, __pa(MAX_DMA_ADDRESS)); | |
c8826dd5 | 2309 | } |
66c3a757 | 2310 | |
c8826dd5 TH |
2311 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
2312 | { | |
999c17e3 | 2313 | memblock_free_early(__pa(ptr), size); |
c8826dd5 TH |
2314 | } |
2315 | ||
e74e3962 TH |
2316 | void __init setup_per_cpu_areas(void) |
2317 | { | |
e74e3962 TH |
2318 | unsigned long delta; |
2319 | unsigned int cpu; | |
fb435d52 | 2320 | int rc; |
e74e3962 TH |
2321 | |
2322 | /* | |
2323 | * Always reserve area for module percpu variables. That's | |
2324 | * what the legacy allocator did. | |
2325 | */ | |
fb435d52 | 2326 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
c8826dd5 TH |
2327 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
2328 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); | |
fb435d52 | 2329 | if (rc < 0) |
bbddff05 | 2330 | panic("Failed to initialize percpu areas."); |
e74e3962 TH |
2331 | |
2332 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2333 | for_each_possible_cpu(cpu) | |
fb435d52 | 2334 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
66c3a757 | 2335 | } |
bbddff05 TH |
2336 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |
2337 | ||
2338 | #else /* CONFIG_SMP */ | |
2339 | ||
2340 | /* | |
2341 | * UP percpu area setup. | |
2342 | * | |
2343 | * UP always uses km-based percpu allocator with identity mapping. | |
2344 | * Static percpu variables are indistinguishable from the usual static | |
2345 | * variables and don't require any special preparation. | |
2346 | */ | |
2347 | void __init setup_per_cpu_areas(void) | |
2348 | { | |
2349 | const size_t unit_size = | |
2350 | roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE, | |
2351 | PERCPU_DYNAMIC_RESERVE)); | |
2352 | struct pcpu_alloc_info *ai; | |
2353 | void *fc; | |
2354 | ||
2355 | ai = pcpu_alloc_alloc_info(1, 1); | |
999c17e3 SS |
2356 | fc = memblock_virt_alloc_from_nopanic(unit_size, |
2357 | PAGE_SIZE, | |
2358 | __pa(MAX_DMA_ADDRESS)); | |
bbddff05 TH |
2359 | if (!ai || !fc) |
2360 | panic("Failed to allocate memory for percpu areas."); | |
100d13c3 CM |
2361 | /* kmemleak tracks the percpu allocations separately */ |
2362 | kmemleak_free(fc); | |
bbddff05 TH |
2363 | |
2364 | ai->dyn_size = unit_size; | |
2365 | ai->unit_size = unit_size; | |
2366 | ai->atom_size = unit_size; | |
2367 | ai->alloc_size = unit_size; | |
2368 | ai->groups[0].nr_units = 1; | |
2369 | ai->groups[0].cpu_map[0] = 0; | |
2370 | ||
2371 | if (pcpu_setup_first_chunk(ai, fc) < 0) | |
2372 | panic("Failed to initialize percpu areas."); | |
2373 | } | |
2374 | ||
2375 | #endif /* CONFIG_SMP */ | |
099a19d9 TH |
2376 | |
2377 | /* | |
2378 | * First and reserved chunks are initialized with temporary allocation | |
2379 | * map in initdata so that they can be used before slab is online. | |
2380 | * This function is called after slab is brought up and replaces those | |
2381 | * with properly allocated maps. | |
2382 | */ | |
2383 | void __init percpu_init_late(void) | |
2384 | { | |
2385 | struct pcpu_chunk *target_chunks[] = | |
2386 | { pcpu_first_chunk, pcpu_reserved_chunk, NULL }; | |
2387 | struct pcpu_chunk *chunk; | |
2388 | unsigned long flags; | |
2389 | int i; | |
2390 | ||
2391 | for (i = 0; (chunk = target_chunks[i]); i++) { | |
2392 | int *map; | |
2393 | const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]); | |
2394 | ||
2395 | BUILD_BUG_ON(size > PAGE_SIZE); | |
2396 | ||
90459ce0 | 2397 | map = pcpu_mem_zalloc(size); |
099a19d9 TH |
2398 | BUG_ON(!map); |
2399 | ||
2400 | spin_lock_irqsave(&pcpu_lock, flags); | |
2401 | memcpy(map, chunk->map, size); | |
2402 | chunk->map = map; | |
2403 | spin_unlock_irqrestore(&pcpu_lock, flags); | |
2404 | } | |
2405 | } | |
1a4d7607 TH |
2406 | |
2407 | /* | |
2408 | * Percpu allocator is initialized early during boot when neither slab or | |
2409 | * workqueue is available. Plug async management until everything is up | |
2410 | * and running. | |
2411 | */ | |
2412 | static int __init percpu_enable_async(void) | |
2413 | { | |
2414 | pcpu_async_enabled = true; | |
2415 | return 0; | |
2416 | } | |
2417 | subsys_initcall(percpu_enable_async); |