Commit | Line | Data |
---|---|---|
fbf59bc9 TH |
1 | /* |
2 | * linux/mm/percpu.c - percpu memory allocator | |
3 | * | |
4 | * Copyright (C) 2009 SUSE Linux Products GmbH | |
5 | * Copyright (C) 2009 Tejun Heo <tj@kernel.org> | |
6 | * | |
7 | * This file is released under the GPLv2. | |
8 | * | |
9 | * This is percpu allocator which can handle both static and dynamic | |
10 | * areas. Percpu areas are allocated in chunks in vmalloc area. Each | |
2f39e637 TH |
11 | * chunk is consisted of boot-time determined number of units and the |
12 | * first chunk is used for static percpu variables in the kernel image | |
13 | * (special boot time alloc/init handling necessary as these areas | |
14 | * need to be brought up before allocation services are running). | |
15 | * Unit grows as necessary and all units grow or shrink in unison. | |
16 | * When a chunk is filled up, another chunk is allocated. ie. in | |
17 | * vmalloc area | |
fbf59bc9 TH |
18 | * |
19 | * c0 c1 c2 | |
20 | * ------------------- ------------------- ------------ | |
21 | * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u | |
22 | * ------------------- ...... ------------------- .... ------------ | |
23 | * | |
24 | * Allocation is done in offset-size areas of single unit space. Ie, | |
25 | * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0, | |
2f39e637 TH |
26 | * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to |
27 | * cpus. On NUMA, the mapping can be non-linear and even sparse. | |
28 | * Percpu access can be done by configuring percpu base registers | |
29 | * according to cpu to unit mapping and pcpu_unit_size. | |
fbf59bc9 | 30 | * |
2f39e637 TH |
31 | * There are usually many small percpu allocations many of them being |
32 | * as small as 4 bytes. The allocator organizes chunks into lists | |
fbf59bc9 TH |
33 | * according to free size and tries to allocate from the fullest one. |
34 | * Each chunk keeps the maximum contiguous area size hint which is | |
35 | * guaranteed to be eqaul to or larger than the maximum contiguous | |
36 | * area in the chunk. This helps the allocator not to iterate the | |
37 | * chunk maps unnecessarily. | |
38 | * | |
39 | * Allocation state in each chunk is kept using an array of integers | |
40 | * on chunk->map. A positive value in the map represents a free | |
41 | * region and negative allocated. Allocation inside a chunk is done | |
42 | * by scanning this map sequentially and serving the first matching | |
43 | * entry. This is mostly copied from the percpu_modalloc() allocator. | |
e1b9aa3f CL |
44 | * Chunks can be determined from the address using the index field |
45 | * in the page struct. The index field contains a pointer to the chunk. | |
fbf59bc9 TH |
46 | * |
47 | * To use this allocator, arch code should do the followings. | |
48 | * | |
fbf59bc9 | 49 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate |
e0100983 TH |
50 | * regular address to percpu pointer and back if they need to be |
51 | * different from the default | |
fbf59bc9 | 52 | * |
8d408b4b TH |
53 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
54 | * setup the first chunk containing the kernel static percpu area | |
fbf59bc9 TH |
55 | */ |
56 | ||
57 | #include <linux/bitmap.h> | |
58 | #include <linux/bootmem.h> | |
fd1e8a1f | 59 | #include <linux/err.h> |
fbf59bc9 | 60 | #include <linux/list.h> |
a530b795 | 61 | #include <linux/log2.h> |
fbf59bc9 TH |
62 | #include <linux/mm.h> |
63 | #include <linux/module.h> | |
64 | #include <linux/mutex.h> | |
65 | #include <linux/percpu.h> | |
66 | #include <linux/pfn.h> | |
fbf59bc9 | 67 | #include <linux/slab.h> |
ccea34b5 | 68 | #include <linux/spinlock.h> |
fbf59bc9 | 69 | #include <linux/vmalloc.h> |
a56dbddf | 70 | #include <linux/workqueue.h> |
fbf59bc9 TH |
71 | |
72 | #include <asm/cacheflush.h> | |
e0100983 | 73 | #include <asm/sections.h> |
fbf59bc9 TH |
74 | #include <asm/tlbflush.h> |
75 | ||
fbf59bc9 TH |
76 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
77 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ | |
78 | ||
e0100983 TH |
79 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
80 | #ifndef __addr_to_pcpu_ptr | |
81 | #define __addr_to_pcpu_ptr(addr) \ | |
82 | (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \ | |
83 | + (unsigned long)__per_cpu_start) | |
84 | #endif | |
85 | #ifndef __pcpu_ptr_to_addr | |
86 | #define __pcpu_ptr_to_addr(ptr) \ | |
87 | (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \ | |
88 | - (unsigned long)__per_cpu_start) | |
89 | #endif | |
90 | ||
fbf59bc9 TH |
91 | struct pcpu_chunk { |
92 | struct list_head list; /* linked to pcpu_slot lists */ | |
fbf59bc9 TH |
93 | int free_size; /* free bytes in the chunk */ |
94 | int contig_hint; /* max contiguous size hint */ | |
bba174f5 | 95 | void *base_addr; /* base address of this chunk */ |
fbf59bc9 TH |
96 | int map_used; /* # of map entries used */ |
97 | int map_alloc; /* # of map entries allocated */ | |
98 | int *map; /* allocation map */ | |
6563297c | 99 | struct vm_struct **vms; /* mapped vmalloc regions */ |
8d408b4b | 100 | bool immutable; /* no [de]population allowed */ |
ce3141a2 | 101 | unsigned long populated[]; /* populated bitmap */ |
fbf59bc9 TH |
102 | }; |
103 | ||
40150d37 TH |
104 | static int pcpu_unit_pages __read_mostly; |
105 | static int pcpu_unit_size __read_mostly; | |
2f39e637 | 106 | static int pcpu_nr_units __read_mostly; |
6563297c | 107 | static int pcpu_atom_size __read_mostly; |
40150d37 TH |
108 | static int pcpu_nr_slots __read_mostly; |
109 | static size_t pcpu_chunk_struct_size __read_mostly; | |
fbf59bc9 | 110 | |
2f39e637 TH |
111 | /* cpus with the lowest and highest unit numbers */ |
112 | static unsigned int pcpu_first_unit_cpu __read_mostly; | |
113 | static unsigned int pcpu_last_unit_cpu __read_mostly; | |
114 | ||
fbf59bc9 | 115 | /* the address of the first chunk which starts with the kernel static area */ |
40150d37 | 116 | void *pcpu_base_addr __read_mostly; |
fbf59bc9 TH |
117 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
118 | ||
fb435d52 TH |
119 | static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */ |
120 | const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */ | |
2f39e637 | 121 | |
6563297c TH |
122 | /* group information, used for vm allocation */ |
123 | static int pcpu_nr_groups __read_mostly; | |
124 | static const unsigned long *pcpu_group_offsets __read_mostly; | |
125 | static const size_t *pcpu_group_sizes __read_mostly; | |
126 | ||
ae9e6bc9 TH |
127 | /* |
128 | * The first chunk which always exists. Note that unlike other | |
129 | * chunks, this one can be allocated and mapped in several different | |
130 | * ways and thus often doesn't live in the vmalloc area. | |
131 | */ | |
132 | static struct pcpu_chunk *pcpu_first_chunk; | |
133 | ||
134 | /* | |
135 | * Optional reserved chunk. This chunk reserves part of the first | |
136 | * chunk and serves it for reserved allocations. The amount of | |
137 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved | |
138 | * area doesn't exist, the following variables contain NULL and 0 | |
139 | * respectively. | |
140 | */ | |
edcb4639 | 141 | static struct pcpu_chunk *pcpu_reserved_chunk; |
edcb4639 TH |
142 | static int pcpu_reserved_chunk_limit; |
143 | ||
fbf59bc9 | 144 | /* |
ccea34b5 TH |
145 | * Synchronization rules. |
146 | * | |
147 | * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former | |
ce3141a2 TH |
148 | * protects allocation/reclaim paths, chunks, populated bitmap and |
149 | * vmalloc mapping. The latter is a spinlock and protects the index | |
150 | * data structures - chunk slots, chunks and area maps in chunks. | |
ccea34b5 TH |
151 | * |
152 | * During allocation, pcpu_alloc_mutex is kept locked all the time and | |
153 | * pcpu_lock is grabbed and released as necessary. All actual memory | |
154 | * allocations are done using GFP_KERNEL with pcpu_lock released. | |
155 | * | |
156 | * Free path accesses and alters only the index data structures, so it | |
157 | * can be safely called from atomic context. When memory needs to be | |
158 | * returned to the system, free path schedules reclaim_work which | |
159 | * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be | |
160 | * reclaimed, release both locks and frees the chunks. Note that it's | |
161 | * necessary to grab both locks to remove a chunk from circulation as | |
162 | * allocation path might be referencing the chunk with only | |
163 | * pcpu_alloc_mutex locked. | |
fbf59bc9 | 164 | */ |
ccea34b5 TH |
165 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ |
166 | static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ | |
fbf59bc9 | 167 | |
40150d37 | 168 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
fbf59bc9 | 169 | |
a56dbddf TH |
170 | /* reclaim work to release fully free chunks, scheduled from free path */ |
171 | static void pcpu_reclaim(struct work_struct *work); | |
172 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); | |
173 | ||
d9b55eeb | 174 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 175 | { |
cae3aeb8 | 176 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
177 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
178 | } | |
179 | ||
d9b55eeb TH |
180 | static int pcpu_size_to_slot(int size) |
181 | { | |
182 | if (size == pcpu_unit_size) | |
183 | return pcpu_nr_slots - 1; | |
184 | return __pcpu_size_to_slot(size); | |
185 | } | |
186 | ||
fbf59bc9 TH |
187 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
188 | { | |
189 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
190 | return 0; | |
191 | ||
192 | return pcpu_size_to_slot(chunk->free_size); | |
193 | } | |
194 | ||
195 | static int pcpu_page_idx(unsigned int cpu, int page_idx) | |
196 | { | |
2f39e637 | 197 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
198 | } |
199 | ||
200 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, | |
201 | unsigned int cpu, int page_idx) | |
202 | { | |
bba174f5 | 203 | return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] + |
fb435d52 | 204 | (page_idx << PAGE_SHIFT); |
fbf59bc9 TH |
205 | } |
206 | ||
ce3141a2 TH |
207 | static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk, |
208 | unsigned int cpu, int page_idx) | |
fbf59bc9 | 209 | { |
ce3141a2 TH |
210 | /* must not be used on pre-mapped chunk */ |
211 | WARN_ON(chunk->immutable); | |
c8a51be4 | 212 | |
ce3141a2 | 213 | return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx)); |
fbf59bc9 TH |
214 | } |
215 | ||
e1b9aa3f CL |
216 | /* set the pointer to a chunk in a page struct */ |
217 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
218 | { | |
219 | page->index = (unsigned long)pcpu; | |
220 | } | |
221 | ||
222 | /* obtain pointer to a chunk from a page struct */ | |
223 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
224 | { | |
225 | return (struct pcpu_chunk *)page->index; | |
226 | } | |
227 | ||
ce3141a2 TH |
228 | static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end) |
229 | { | |
230 | *rs = find_next_zero_bit(chunk->populated, end, *rs); | |
231 | *re = find_next_bit(chunk->populated, end, *rs + 1); | |
232 | } | |
233 | ||
234 | static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end) | |
235 | { | |
236 | *rs = find_next_bit(chunk->populated, end, *rs); | |
237 | *re = find_next_zero_bit(chunk->populated, end, *rs + 1); | |
238 | } | |
239 | ||
240 | /* | |
241 | * (Un)populated page region iterators. Iterate over (un)populated | |
242 | * page regions betwen @start and @end in @chunk. @rs and @re should | |
243 | * be integer variables and will be set to start and end page index of | |
244 | * the current region. | |
245 | */ | |
246 | #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ | |
247 | for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ | |
248 | (rs) < (re); \ | |
249 | (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) | |
250 | ||
251 | #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ | |
252 | for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ | |
253 | (rs) < (re); \ | |
254 | (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) | |
255 | ||
fbf59bc9 | 256 | /** |
1880d93b TH |
257 | * pcpu_mem_alloc - allocate memory |
258 | * @size: bytes to allocate | |
fbf59bc9 | 259 | * |
1880d93b TH |
260 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
261 | * kzalloc() is used; otherwise, vmalloc() is used. The returned | |
262 | * memory is always zeroed. | |
fbf59bc9 | 263 | * |
ccea34b5 TH |
264 | * CONTEXT: |
265 | * Does GFP_KERNEL allocation. | |
266 | * | |
fbf59bc9 | 267 | * RETURNS: |
1880d93b | 268 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 269 | */ |
1880d93b | 270 | static void *pcpu_mem_alloc(size_t size) |
fbf59bc9 | 271 | { |
1880d93b TH |
272 | if (size <= PAGE_SIZE) |
273 | return kzalloc(size, GFP_KERNEL); | |
274 | else { | |
275 | void *ptr = vmalloc(size); | |
276 | if (ptr) | |
277 | memset(ptr, 0, size); | |
278 | return ptr; | |
279 | } | |
280 | } | |
fbf59bc9 | 281 | |
1880d93b TH |
282 | /** |
283 | * pcpu_mem_free - free memory | |
284 | * @ptr: memory to free | |
285 | * @size: size of the area | |
286 | * | |
287 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). | |
288 | */ | |
289 | static void pcpu_mem_free(void *ptr, size_t size) | |
290 | { | |
fbf59bc9 | 291 | if (size <= PAGE_SIZE) |
1880d93b | 292 | kfree(ptr); |
fbf59bc9 | 293 | else |
1880d93b | 294 | vfree(ptr); |
fbf59bc9 TH |
295 | } |
296 | ||
297 | /** | |
298 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
299 | * @chunk: chunk of interest | |
300 | * @oslot: the previous slot it was on | |
301 | * | |
302 | * This function is called after an allocation or free changed @chunk. | |
303 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
304 | * moved to the slot. Note that the reserved chunk is never put on |
305 | * chunk slots. | |
ccea34b5 TH |
306 | * |
307 | * CONTEXT: | |
308 | * pcpu_lock. | |
fbf59bc9 TH |
309 | */ |
310 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
311 | { | |
312 | int nslot = pcpu_chunk_slot(chunk); | |
313 | ||
edcb4639 | 314 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
315 | if (oslot < nslot) |
316 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
317 | else | |
318 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
319 | } | |
320 | } | |
321 | ||
fbf59bc9 | 322 | /** |
e1b9aa3f CL |
323 | * pcpu_chunk_addr_search - determine chunk containing specified address |
324 | * @addr: address for which the chunk needs to be determined. | |
ccea34b5 | 325 | * |
fbf59bc9 TH |
326 | * RETURNS: |
327 | * The address of the found chunk. | |
328 | */ | |
329 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
330 | { | |
bba174f5 | 331 | void *first_start = pcpu_first_chunk->base_addr; |
fbf59bc9 | 332 | |
ae9e6bc9 | 333 | /* is it in the first chunk? */ |
79ba6ac8 | 334 | if (addr >= first_start && addr < first_start + pcpu_unit_size) { |
ae9e6bc9 TH |
335 | /* is it in the reserved area? */ |
336 | if (addr < first_start + pcpu_reserved_chunk_limit) | |
edcb4639 | 337 | return pcpu_reserved_chunk; |
ae9e6bc9 | 338 | return pcpu_first_chunk; |
edcb4639 TH |
339 | } |
340 | ||
04a13c7c TH |
341 | /* |
342 | * The address is relative to unit0 which might be unused and | |
343 | * thus unmapped. Offset the address to the unit space of the | |
344 | * current processor before looking it up in the vmalloc | |
345 | * space. Note that any possible cpu id can be used here, so | |
346 | * there's no need to worry about preemption or cpu hotplug. | |
347 | */ | |
5579fd7e | 348 | addr += pcpu_unit_offsets[raw_smp_processor_id()]; |
e1b9aa3f | 349 | return pcpu_get_page_chunk(vmalloc_to_page(addr)); |
fbf59bc9 TH |
350 | } |
351 | ||
9f7dcf22 TH |
352 | /** |
353 | * pcpu_extend_area_map - extend area map for allocation | |
354 | * @chunk: target chunk | |
355 | * | |
356 | * Extend area map of @chunk so that it can accomodate an allocation. | |
357 | * A single allocation can split an area into three areas, so this | |
358 | * function makes sure that @chunk->map has at least two extra slots. | |
359 | * | |
ccea34b5 TH |
360 | * CONTEXT: |
361 | * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired | |
362 | * if area map is extended. | |
363 | * | |
9f7dcf22 TH |
364 | * RETURNS: |
365 | * 0 if noop, 1 if successfully extended, -errno on failure. | |
366 | */ | |
367 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) | |
368 | { | |
369 | int new_alloc; | |
370 | int *new; | |
371 | size_t size; | |
372 | ||
373 | /* has enough? */ | |
374 | if (chunk->map_alloc >= chunk->map_used + 2) | |
375 | return 0; | |
376 | ||
ccea34b5 TH |
377 | spin_unlock_irq(&pcpu_lock); |
378 | ||
9f7dcf22 TH |
379 | new_alloc = PCPU_DFL_MAP_ALLOC; |
380 | while (new_alloc < chunk->map_used + 2) | |
381 | new_alloc *= 2; | |
382 | ||
383 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); | |
ccea34b5 TH |
384 | if (!new) { |
385 | spin_lock_irq(&pcpu_lock); | |
9f7dcf22 | 386 | return -ENOMEM; |
ccea34b5 TH |
387 | } |
388 | ||
389 | /* | |
390 | * Acquire pcpu_lock and switch to new area map. Only free | |
391 | * could have happened inbetween, so map_used couldn't have | |
392 | * grown. | |
393 | */ | |
394 | spin_lock_irq(&pcpu_lock); | |
395 | BUG_ON(new_alloc < chunk->map_used + 2); | |
9f7dcf22 TH |
396 | |
397 | size = chunk->map_alloc * sizeof(chunk->map[0]); | |
398 | memcpy(new, chunk->map, size); | |
399 | ||
400 | /* | |
401 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is | |
402 | * one of the first chunks and still using static map. | |
403 | */ | |
404 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) | |
405 | pcpu_mem_free(chunk->map, size); | |
406 | ||
407 | chunk->map_alloc = new_alloc; | |
408 | chunk->map = new; | |
409 | return 0; | |
410 | } | |
411 | ||
fbf59bc9 TH |
412 | /** |
413 | * pcpu_split_block - split a map block | |
414 | * @chunk: chunk of interest | |
415 | * @i: index of map block to split | |
cae3aeb8 TH |
416 | * @head: head size in bytes (can be 0) |
417 | * @tail: tail size in bytes (can be 0) | |
fbf59bc9 TH |
418 | * |
419 | * Split the @i'th map block into two or three blocks. If @head is | |
420 | * non-zero, @head bytes block is inserted before block @i moving it | |
421 | * to @i+1 and reducing its size by @head bytes. | |
422 | * | |
423 | * If @tail is non-zero, the target block, which can be @i or @i+1 | |
424 | * depending on @head, is reduced by @tail bytes and @tail byte block | |
425 | * is inserted after the target block. | |
426 | * | |
9f7dcf22 | 427 | * @chunk->map must have enough free slots to accomodate the split. |
ccea34b5 TH |
428 | * |
429 | * CONTEXT: | |
430 | * pcpu_lock. | |
fbf59bc9 | 431 | */ |
9f7dcf22 TH |
432 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
433 | int head, int tail) | |
fbf59bc9 TH |
434 | { |
435 | int nr_extra = !!head + !!tail; | |
1880d93b | 436 | |
9f7dcf22 | 437 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
fbf59bc9 | 438 | |
9f7dcf22 | 439 | /* insert new subblocks */ |
fbf59bc9 TH |
440 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
441 | sizeof(chunk->map[0]) * (chunk->map_used - i)); | |
442 | chunk->map_used += nr_extra; | |
443 | ||
444 | if (head) { | |
445 | chunk->map[i + 1] = chunk->map[i] - head; | |
446 | chunk->map[i++] = head; | |
447 | } | |
448 | if (tail) { | |
449 | chunk->map[i++] -= tail; | |
450 | chunk->map[i] = tail; | |
451 | } | |
fbf59bc9 TH |
452 | } |
453 | ||
454 | /** | |
455 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
456 | * @chunk: chunk of interest | |
cae3aeb8 | 457 | * @size: wanted size in bytes |
fbf59bc9 TH |
458 | * @align: wanted align |
459 | * | |
460 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
461 | * Note that this function only allocates the offset. It doesn't | |
462 | * populate or map the area. | |
463 | * | |
9f7dcf22 TH |
464 | * @chunk->map must have at least two free slots. |
465 | * | |
ccea34b5 TH |
466 | * CONTEXT: |
467 | * pcpu_lock. | |
468 | * | |
fbf59bc9 | 469 | * RETURNS: |
9f7dcf22 TH |
470 | * Allocated offset in @chunk on success, -1 if no matching area is |
471 | * found. | |
fbf59bc9 TH |
472 | */ |
473 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) | |
474 | { | |
475 | int oslot = pcpu_chunk_slot(chunk); | |
476 | int max_contig = 0; | |
477 | int i, off; | |
478 | ||
fbf59bc9 TH |
479 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
480 | bool is_last = i + 1 == chunk->map_used; | |
481 | int head, tail; | |
482 | ||
483 | /* extra for alignment requirement */ | |
484 | head = ALIGN(off, align) - off; | |
485 | BUG_ON(i == 0 && head != 0); | |
486 | ||
487 | if (chunk->map[i] < 0) | |
488 | continue; | |
489 | if (chunk->map[i] < head + size) { | |
490 | max_contig = max(chunk->map[i], max_contig); | |
491 | continue; | |
492 | } | |
493 | ||
494 | /* | |
495 | * If head is small or the previous block is free, | |
496 | * merge'em. Note that 'small' is defined as smaller | |
497 | * than sizeof(int), which is very small but isn't too | |
498 | * uncommon for percpu allocations. | |
499 | */ | |
500 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { | |
501 | if (chunk->map[i - 1] > 0) | |
502 | chunk->map[i - 1] += head; | |
503 | else { | |
504 | chunk->map[i - 1] -= head; | |
505 | chunk->free_size -= head; | |
506 | } | |
507 | chunk->map[i] -= head; | |
508 | off += head; | |
509 | head = 0; | |
510 | } | |
511 | ||
512 | /* if tail is small, just keep it around */ | |
513 | tail = chunk->map[i] - head - size; | |
514 | if (tail < sizeof(int)) | |
515 | tail = 0; | |
516 | ||
517 | /* split if warranted */ | |
518 | if (head || tail) { | |
9f7dcf22 | 519 | pcpu_split_block(chunk, i, head, tail); |
fbf59bc9 TH |
520 | if (head) { |
521 | i++; | |
522 | off += head; | |
523 | max_contig = max(chunk->map[i - 1], max_contig); | |
524 | } | |
525 | if (tail) | |
526 | max_contig = max(chunk->map[i + 1], max_contig); | |
527 | } | |
528 | ||
529 | /* update hint and mark allocated */ | |
530 | if (is_last) | |
531 | chunk->contig_hint = max_contig; /* fully scanned */ | |
532 | else | |
533 | chunk->contig_hint = max(chunk->contig_hint, | |
534 | max_contig); | |
535 | ||
536 | chunk->free_size -= chunk->map[i]; | |
537 | chunk->map[i] = -chunk->map[i]; | |
538 | ||
539 | pcpu_chunk_relocate(chunk, oslot); | |
540 | return off; | |
541 | } | |
542 | ||
543 | chunk->contig_hint = max_contig; /* fully scanned */ | |
544 | pcpu_chunk_relocate(chunk, oslot); | |
545 | ||
9f7dcf22 TH |
546 | /* tell the upper layer that this chunk has no matching area */ |
547 | return -1; | |
fbf59bc9 TH |
548 | } |
549 | ||
550 | /** | |
551 | * pcpu_free_area - free area to a pcpu_chunk | |
552 | * @chunk: chunk of interest | |
553 | * @freeme: offset of area to free | |
554 | * | |
555 | * Free area starting from @freeme to @chunk. Note that this function | |
556 | * only modifies the allocation map. It doesn't depopulate or unmap | |
557 | * the area. | |
ccea34b5 TH |
558 | * |
559 | * CONTEXT: | |
560 | * pcpu_lock. | |
fbf59bc9 TH |
561 | */ |
562 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) | |
563 | { | |
564 | int oslot = pcpu_chunk_slot(chunk); | |
565 | int i, off; | |
566 | ||
567 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) | |
568 | if (off == freeme) | |
569 | break; | |
570 | BUG_ON(off != freeme); | |
571 | BUG_ON(chunk->map[i] > 0); | |
572 | ||
573 | chunk->map[i] = -chunk->map[i]; | |
574 | chunk->free_size += chunk->map[i]; | |
575 | ||
576 | /* merge with previous? */ | |
577 | if (i > 0 && chunk->map[i - 1] >= 0) { | |
578 | chunk->map[i - 1] += chunk->map[i]; | |
579 | chunk->map_used--; | |
580 | memmove(&chunk->map[i], &chunk->map[i + 1], | |
581 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
582 | i--; | |
583 | } | |
584 | /* merge with next? */ | |
585 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { | |
586 | chunk->map[i] += chunk->map[i + 1]; | |
587 | chunk->map_used--; | |
588 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], | |
589 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); | |
590 | } | |
591 | ||
592 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); | |
593 | pcpu_chunk_relocate(chunk, oslot); | |
594 | } | |
595 | ||
596 | /** | |
ce3141a2 | 597 | * pcpu_get_pages_and_bitmap - get temp pages array and bitmap |
fbf59bc9 | 598 | * @chunk: chunk of interest |
ce3141a2 TH |
599 | * @bitmapp: output parameter for bitmap |
600 | * @may_alloc: may allocate the array | |
fbf59bc9 | 601 | * |
ce3141a2 TH |
602 | * Returns pointer to array of pointers to struct page and bitmap, |
603 | * both of which can be indexed with pcpu_page_idx(). The returned | |
604 | * array is cleared to zero and *@bitmapp is copied from | |
605 | * @chunk->populated. Note that there is only one array and bitmap | |
606 | * and access exclusion is the caller's responsibility. | |
607 | * | |
608 | * CONTEXT: | |
609 | * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc. | |
610 | * Otherwise, don't care. | |
611 | * | |
612 | * RETURNS: | |
613 | * Pointer to temp pages array on success, NULL on failure. | |
fbf59bc9 | 614 | */ |
ce3141a2 TH |
615 | static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk, |
616 | unsigned long **bitmapp, | |
617 | bool may_alloc) | |
fbf59bc9 | 618 | { |
ce3141a2 TH |
619 | static struct page **pages; |
620 | static unsigned long *bitmap; | |
2f39e637 | 621 | size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]); |
ce3141a2 TH |
622 | size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) * |
623 | sizeof(unsigned long); | |
624 | ||
625 | if (!pages || !bitmap) { | |
626 | if (may_alloc && !pages) | |
627 | pages = pcpu_mem_alloc(pages_size); | |
628 | if (may_alloc && !bitmap) | |
629 | bitmap = pcpu_mem_alloc(bitmap_size); | |
630 | if (!pages || !bitmap) | |
631 | return NULL; | |
632 | } | |
fbf59bc9 | 633 | |
ce3141a2 TH |
634 | memset(pages, 0, pages_size); |
635 | bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages); | |
8d408b4b | 636 | |
ce3141a2 TH |
637 | *bitmapp = bitmap; |
638 | return pages; | |
639 | } | |
fbf59bc9 | 640 | |
ce3141a2 TH |
641 | /** |
642 | * pcpu_free_pages - free pages which were allocated for @chunk | |
643 | * @chunk: chunk pages were allocated for | |
644 | * @pages: array of pages to be freed, indexed by pcpu_page_idx() | |
645 | * @populated: populated bitmap | |
646 | * @page_start: page index of the first page to be freed | |
647 | * @page_end: page index of the last page to be freed + 1 | |
648 | * | |
649 | * Free pages [@page_start and @page_end) in @pages for all units. | |
650 | * The pages were allocated for @chunk. | |
651 | */ | |
652 | static void pcpu_free_pages(struct pcpu_chunk *chunk, | |
653 | struct page **pages, unsigned long *populated, | |
654 | int page_start, int page_end) | |
655 | { | |
656 | unsigned int cpu; | |
657 | int i; | |
658 | ||
659 | for_each_possible_cpu(cpu) { | |
660 | for (i = page_start; i < page_end; i++) { | |
661 | struct page *page = pages[pcpu_page_idx(cpu, i)]; | |
662 | ||
663 | if (page) | |
664 | __free_page(page); | |
665 | } | |
666 | } | |
fbf59bc9 TH |
667 | } |
668 | ||
669 | /** | |
ce3141a2 TH |
670 | * pcpu_alloc_pages - allocates pages for @chunk |
671 | * @chunk: target chunk | |
672 | * @pages: array to put the allocated pages into, indexed by pcpu_page_idx() | |
673 | * @populated: populated bitmap | |
674 | * @page_start: page index of the first page to be allocated | |
675 | * @page_end: page index of the last page to be allocated + 1 | |
676 | * | |
677 | * Allocate pages [@page_start,@page_end) into @pages for all units. | |
678 | * The allocation is for @chunk. Percpu core doesn't care about the | |
679 | * content of @pages and will pass it verbatim to pcpu_map_pages(). | |
fbf59bc9 | 680 | */ |
ce3141a2 TH |
681 | static int pcpu_alloc_pages(struct pcpu_chunk *chunk, |
682 | struct page **pages, unsigned long *populated, | |
683 | int page_start, int page_end) | |
fbf59bc9 | 684 | { |
ce3141a2 | 685 | const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; |
fbf59bc9 TH |
686 | unsigned int cpu; |
687 | int i; | |
688 | ||
ce3141a2 TH |
689 | for_each_possible_cpu(cpu) { |
690 | for (i = page_start; i < page_end; i++) { | |
691 | struct page **pagep = &pages[pcpu_page_idx(cpu, i)]; | |
692 | ||
693 | *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0); | |
694 | if (!*pagep) { | |
695 | pcpu_free_pages(chunk, pages, populated, | |
696 | page_start, page_end); | |
697 | return -ENOMEM; | |
698 | } | |
699 | } | |
700 | } | |
701 | return 0; | |
702 | } | |
fbf59bc9 | 703 | |
ce3141a2 TH |
704 | /** |
705 | * pcpu_pre_unmap_flush - flush cache prior to unmapping | |
706 | * @chunk: chunk the regions to be flushed belongs to | |
707 | * @page_start: page index of the first page to be flushed | |
708 | * @page_end: page index of the last page to be flushed + 1 | |
709 | * | |
710 | * Pages in [@page_start,@page_end) of @chunk are about to be | |
711 | * unmapped. Flush cache. As each flushing trial can be very | |
712 | * expensive, issue flush on the whole region at once rather than | |
713 | * doing it for each cpu. This could be an overkill but is more | |
714 | * scalable. | |
715 | */ | |
716 | static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk, | |
717 | int page_start, int page_end) | |
718 | { | |
2f39e637 TH |
719 | flush_cache_vunmap( |
720 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
721 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
ce3141a2 TH |
722 | } |
723 | ||
724 | static void __pcpu_unmap_pages(unsigned long addr, int nr_pages) | |
725 | { | |
726 | unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT); | |
727 | } | |
fbf59bc9 | 728 | |
ce3141a2 TH |
729 | /** |
730 | * pcpu_unmap_pages - unmap pages out of a pcpu_chunk | |
fbf59bc9 | 731 | * @chunk: chunk of interest |
ce3141a2 TH |
732 | * @pages: pages array which can be used to pass information to free |
733 | * @populated: populated bitmap | |
fbf59bc9 TH |
734 | * @page_start: page index of the first page to unmap |
735 | * @page_end: page index of the last page to unmap + 1 | |
fbf59bc9 TH |
736 | * |
737 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. | |
ce3141a2 TH |
738 | * Corresponding elements in @pages were cleared by the caller and can |
739 | * be used to carry information to pcpu_free_pages() which will be | |
740 | * called after all unmaps are finished. The caller should call | |
741 | * proper pre/post flush functions. | |
fbf59bc9 | 742 | */ |
ce3141a2 TH |
743 | static void pcpu_unmap_pages(struct pcpu_chunk *chunk, |
744 | struct page **pages, unsigned long *populated, | |
745 | int page_start, int page_end) | |
fbf59bc9 | 746 | { |
fbf59bc9 | 747 | unsigned int cpu; |
ce3141a2 | 748 | int i; |
fbf59bc9 | 749 | |
ce3141a2 TH |
750 | for_each_possible_cpu(cpu) { |
751 | for (i = page_start; i < page_end; i++) { | |
752 | struct page *page; | |
fbf59bc9 | 753 | |
ce3141a2 TH |
754 | page = pcpu_chunk_page(chunk, cpu, i); |
755 | WARN_ON(!page); | |
756 | pages[pcpu_page_idx(cpu, i)] = page; | |
fbf59bc9 | 757 | } |
ce3141a2 TH |
758 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start), |
759 | page_end - page_start); | |
fbf59bc9 TH |
760 | } |
761 | ||
ce3141a2 TH |
762 | for (i = page_start; i < page_end; i++) |
763 | __clear_bit(i, populated); | |
764 | } | |
765 | ||
766 | /** | |
767 | * pcpu_post_unmap_tlb_flush - flush TLB after unmapping | |
768 | * @chunk: pcpu_chunk the regions to be flushed belong to | |
769 | * @page_start: page index of the first page to be flushed | |
770 | * @page_end: page index of the last page to be flushed + 1 | |
771 | * | |
772 | * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush | |
773 | * TLB for the regions. This can be skipped if the area is to be | |
774 | * returned to vmalloc as vmalloc will handle TLB flushing lazily. | |
775 | * | |
776 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once | |
777 | * for the whole region. | |
778 | */ | |
779 | static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk, | |
780 | int page_start, int page_end) | |
781 | { | |
2f39e637 TH |
782 | flush_tlb_kernel_range( |
783 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
784 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
fbf59bc9 TH |
785 | } |
786 | ||
c8a51be4 TH |
787 | static int __pcpu_map_pages(unsigned long addr, struct page **pages, |
788 | int nr_pages) | |
789 | { | |
790 | return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT, | |
791 | PAGE_KERNEL, pages); | |
fbf59bc9 TH |
792 | } |
793 | ||
794 | /** | |
ce3141a2 | 795 | * pcpu_map_pages - map pages into a pcpu_chunk |
fbf59bc9 | 796 | * @chunk: chunk of interest |
ce3141a2 TH |
797 | * @pages: pages array containing pages to be mapped |
798 | * @populated: populated bitmap | |
fbf59bc9 TH |
799 | * @page_start: page index of the first page to map |
800 | * @page_end: page index of the last page to map + 1 | |
801 | * | |
ce3141a2 TH |
802 | * For each cpu, map pages [@page_start,@page_end) into @chunk. The |
803 | * caller is responsible for calling pcpu_post_map_flush() after all | |
804 | * mappings are complete. | |
805 | * | |
806 | * This function is responsible for setting corresponding bits in | |
807 | * @chunk->populated bitmap and whatever is necessary for reverse | |
808 | * lookup (addr -> chunk). | |
fbf59bc9 | 809 | */ |
ce3141a2 TH |
810 | static int pcpu_map_pages(struct pcpu_chunk *chunk, |
811 | struct page **pages, unsigned long *populated, | |
812 | int page_start, int page_end) | |
fbf59bc9 | 813 | { |
ce3141a2 TH |
814 | unsigned int cpu, tcpu; |
815 | int i, err; | |
8d408b4b | 816 | |
fbf59bc9 | 817 | for_each_possible_cpu(cpu) { |
c8a51be4 | 818 | err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start), |
ce3141a2 | 819 | &pages[pcpu_page_idx(cpu, page_start)], |
c8a51be4 | 820 | page_end - page_start); |
fbf59bc9 | 821 | if (err < 0) |
ce3141a2 | 822 | goto err; |
c8a51be4 TH |
823 | } |
824 | ||
ce3141a2 TH |
825 | /* mapping successful, link chunk and mark populated */ |
826 | for (i = page_start; i < page_end; i++) { | |
827 | for_each_possible_cpu(cpu) | |
828 | pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)], | |
829 | chunk); | |
830 | __set_bit(i, populated); | |
fbf59bc9 TH |
831 | } |
832 | ||
fbf59bc9 | 833 | return 0; |
ce3141a2 TH |
834 | |
835 | err: | |
836 | for_each_possible_cpu(tcpu) { | |
837 | if (tcpu == cpu) | |
838 | break; | |
839 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start), | |
840 | page_end - page_start); | |
841 | } | |
842 | return err; | |
843 | } | |
844 | ||
845 | /** | |
846 | * pcpu_post_map_flush - flush cache after mapping | |
847 | * @chunk: pcpu_chunk the regions to be flushed belong to | |
848 | * @page_start: page index of the first page to be flushed | |
849 | * @page_end: page index of the last page to be flushed + 1 | |
850 | * | |
851 | * Pages [@page_start,@page_end) of @chunk have been mapped. Flush | |
852 | * cache. | |
853 | * | |
854 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once | |
855 | * for the whole region. | |
856 | */ | |
857 | static void pcpu_post_map_flush(struct pcpu_chunk *chunk, | |
858 | int page_start, int page_end) | |
859 | { | |
2f39e637 TH |
860 | flush_cache_vmap( |
861 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
862 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
c8a51be4 TH |
863 | } |
864 | ||
fbf59bc9 TH |
865 | /** |
866 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk | |
867 | * @chunk: chunk to depopulate | |
868 | * @off: offset to the area to depopulate | |
cae3aeb8 | 869 | * @size: size of the area to depopulate in bytes |
fbf59bc9 TH |
870 | * @flush: whether to flush cache and tlb or not |
871 | * | |
872 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) | |
873 | * from @chunk. If @flush is true, vcache is flushed before unmapping | |
874 | * and tlb after. | |
ccea34b5 TH |
875 | * |
876 | * CONTEXT: | |
877 | * pcpu_alloc_mutex. | |
fbf59bc9 | 878 | */ |
ce3141a2 | 879 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size) |
fbf59bc9 TH |
880 | { |
881 | int page_start = PFN_DOWN(off); | |
882 | int page_end = PFN_UP(off + size); | |
ce3141a2 TH |
883 | struct page **pages; |
884 | unsigned long *populated; | |
885 | int rs, re; | |
886 | ||
887 | /* quick path, check whether it's empty already */ | |
888 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { | |
889 | if (rs == page_start && re == page_end) | |
890 | return; | |
891 | break; | |
892 | } | |
fbf59bc9 | 893 | |
ce3141a2 TH |
894 | /* immutable chunks can't be depopulated */ |
895 | WARN_ON(chunk->immutable); | |
fbf59bc9 | 896 | |
ce3141a2 TH |
897 | /* |
898 | * If control reaches here, there must have been at least one | |
899 | * successful population attempt so the temp pages array must | |
900 | * be available now. | |
901 | */ | |
902 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, false); | |
903 | BUG_ON(!pages); | |
fbf59bc9 | 904 | |
ce3141a2 TH |
905 | /* unmap and free */ |
906 | pcpu_pre_unmap_flush(chunk, page_start, page_end); | |
fbf59bc9 | 907 | |
ce3141a2 TH |
908 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) |
909 | pcpu_unmap_pages(chunk, pages, populated, rs, re); | |
fbf59bc9 | 910 | |
ce3141a2 TH |
911 | /* no need to flush tlb, vmalloc will handle it lazily */ |
912 | ||
913 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) | |
914 | pcpu_free_pages(chunk, pages, populated, rs, re); | |
fbf59bc9 | 915 | |
ce3141a2 TH |
916 | /* commit new bitmap */ |
917 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); | |
fbf59bc9 TH |
918 | } |
919 | ||
920 | /** | |
921 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk | |
922 | * @chunk: chunk of interest | |
923 | * @off: offset to the area to populate | |
cae3aeb8 | 924 | * @size: size of the area to populate in bytes |
fbf59bc9 TH |
925 | * |
926 | * For each cpu, populate and map pages [@page_start,@page_end) into | |
927 | * @chunk. The area is cleared on return. | |
ccea34b5 TH |
928 | * |
929 | * CONTEXT: | |
930 | * pcpu_alloc_mutex, does GFP_KERNEL allocation. | |
fbf59bc9 TH |
931 | */ |
932 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) | |
933 | { | |
fbf59bc9 TH |
934 | int page_start = PFN_DOWN(off); |
935 | int page_end = PFN_UP(off + size); | |
ce3141a2 TH |
936 | int free_end = page_start, unmap_end = page_start; |
937 | struct page **pages; | |
938 | unsigned long *populated; | |
fbf59bc9 | 939 | unsigned int cpu; |
ce3141a2 | 940 | int rs, re, rc; |
fbf59bc9 | 941 | |
ce3141a2 TH |
942 | /* quick path, check whether all pages are already there */ |
943 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) { | |
944 | if (rs == page_start && re == page_end) | |
945 | goto clear; | |
946 | break; | |
947 | } | |
fbf59bc9 | 948 | |
ce3141a2 TH |
949 | /* need to allocate and map pages, this chunk can't be immutable */ |
950 | WARN_ON(chunk->immutable); | |
fbf59bc9 | 951 | |
ce3141a2 TH |
952 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, true); |
953 | if (!pages) | |
954 | return -ENOMEM; | |
fbf59bc9 | 955 | |
ce3141a2 TH |
956 | /* alloc and map */ |
957 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { | |
958 | rc = pcpu_alloc_pages(chunk, pages, populated, rs, re); | |
959 | if (rc) | |
960 | goto err_free; | |
961 | free_end = re; | |
fbf59bc9 TH |
962 | } |
963 | ||
ce3141a2 TH |
964 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
965 | rc = pcpu_map_pages(chunk, pages, populated, rs, re); | |
966 | if (rc) | |
967 | goto err_unmap; | |
968 | unmap_end = re; | |
969 | } | |
970 | pcpu_post_map_flush(chunk, page_start, page_end); | |
fbf59bc9 | 971 | |
ce3141a2 TH |
972 | /* commit new bitmap */ |
973 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); | |
974 | clear: | |
fbf59bc9 | 975 | for_each_possible_cpu(cpu) |
2f39e637 | 976 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); |
fbf59bc9 | 977 | return 0; |
ce3141a2 TH |
978 | |
979 | err_unmap: | |
980 | pcpu_pre_unmap_flush(chunk, page_start, unmap_end); | |
981 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end) | |
982 | pcpu_unmap_pages(chunk, pages, populated, rs, re); | |
983 | pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end); | |
984 | err_free: | |
985 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end) | |
986 | pcpu_free_pages(chunk, pages, populated, rs, re); | |
987 | return rc; | |
fbf59bc9 TH |
988 | } |
989 | ||
990 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) | |
991 | { | |
992 | if (!chunk) | |
993 | return; | |
6563297c TH |
994 | if (chunk->vms) |
995 | pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups); | |
1880d93b | 996 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
fbf59bc9 TH |
997 | kfree(chunk); |
998 | } | |
999 | ||
1000 | static struct pcpu_chunk *alloc_pcpu_chunk(void) | |
1001 | { | |
1002 | struct pcpu_chunk *chunk; | |
1003 | ||
1004 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); | |
1005 | if (!chunk) | |
1006 | return NULL; | |
1007 | ||
1880d93b | 1008 | chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
fbf59bc9 TH |
1009 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
1010 | chunk->map[chunk->map_used++] = pcpu_unit_size; | |
1011 | ||
6563297c TH |
1012 | chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes, |
1013 | pcpu_nr_groups, pcpu_atom_size, | |
1014 | GFP_KERNEL); | |
1015 | if (!chunk->vms) { | |
fbf59bc9 TH |
1016 | free_pcpu_chunk(chunk); |
1017 | return NULL; | |
1018 | } | |
1019 | ||
1020 | INIT_LIST_HEAD(&chunk->list); | |
1021 | chunk->free_size = pcpu_unit_size; | |
1022 | chunk->contig_hint = pcpu_unit_size; | |
6563297c | 1023 | chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0]; |
fbf59bc9 TH |
1024 | |
1025 | return chunk; | |
1026 | } | |
1027 | ||
1028 | /** | |
edcb4639 | 1029 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 1030 | * @size: size of area to allocate in bytes |
fbf59bc9 | 1031 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 1032 | * @reserved: allocate from the reserved chunk if available |
fbf59bc9 | 1033 | * |
ccea34b5 TH |
1034 | * Allocate percpu area of @size bytes aligned at @align. |
1035 | * | |
1036 | * CONTEXT: | |
1037 | * Does GFP_KERNEL allocation. | |
fbf59bc9 TH |
1038 | * |
1039 | * RETURNS: | |
1040 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1041 | */ | |
edcb4639 | 1042 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
fbf59bc9 | 1043 | { |
f2badb0c | 1044 | static int warn_limit = 10; |
fbf59bc9 | 1045 | struct pcpu_chunk *chunk; |
f2badb0c | 1046 | const char *err; |
fbf59bc9 TH |
1047 | int slot, off; |
1048 | ||
8d408b4b | 1049 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
fbf59bc9 TH |
1050 | WARN(true, "illegal size (%zu) or align (%zu) for " |
1051 | "percpu allocation\n", size, align); | |
1052 | return NULL; | |
1053 | } | |
1054 | ||
ccea34b5 TH |
1055 | mutex_lock(&pcpu_alloc_mutex); |
1056 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 1057 | |
edcb4639 TH |
1058 | /* serve reserved allocations from the reserved chunk if available */ |
1059 | if (reserved && pcpu_reserved_chunk) { | |
1060 | chunk = pcpu_reserved_chunk; | |
9f7dcf22 | 1061 | if (size > chunk->contig_hint || |
f2badb0c TH |
1062 | pcpu_extend_area_map(chunk) < 0) { |
1063 | err = "failed to extend area map of reserved chunk"; | |
ccea34b5 | 1064 | goto fail_unlock; |
f2badb0c | 1065 | } |
edcb4639 TH |
1066 | off = pcpu_alloc_area(chunk, size, align); |
1067 | if (off >= 0) | |
1068 | goto area_found; | |
f2badb0c | 1069 | err = "alloc from reserved chunk failed"; |
ccea34b5 | 1070 | goto fail_unlock; |
edcb4639 TH |
1071 | } |
1072 | ||
ccea34b5 | 1073 | restart: |
edcb4639 | 1074 | /* search through normal chunks */ |
fbf59bc9 TH |
1075 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
1076 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
1077 | if (size > chunk->contig_hint) | |
1078 | continue; | |
ccea34b5 TH |
1079 | |
1080 | switch (pcpu_extend_area_map(chunk)) { | |
1081 | case 0: | |
1082 | break; | |
1083 | case 1: | |
1084 | goto restart; /* pcpu_lock dropped, restart */ | |
1085 | default: | |
f2badb0c | 1086 | err = "failed to extend area map"; |
ccea34b5 TH |
1087 | goto fail_unlock; |
1088 | } | |
1089 | ||
fbf59bc9 TH |
1090 | off = pcpu_alloc_area(chunk, size, align); |
1091 | if (off >= 0) | |
1092 | goto area_found; | |
fbf59bc9 TH |
1093 | } |
1094 | } | |
1095 | ||
1096 | /* hmmm... no space left, create a new chunk */ | |
ccea34b5 TH |
1097 | spin_unlock_irq(&pcpu_lock); |
1098 | ||
fbf59bc9 | 1099 | chunk = alloc_pcpu_chunk(); |
f2badb0c TH |
1100 | if (!chunk) { |
1101 | err = "failed to allocate new chunk"; | |
ccea34b5 | 1102 | goto fail_unlock_mutex; |
f2badb0c | 1103 | } |
ccea34b5 TH |
1104 | |
1105 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 1106 | pcpu_chunk_relocate(chunk, -1); |
ccea34b5 | 1107 | goto restart; |
fbf59bc9 TH |
1108 | |
1109 | area_found: | |
ccea34b5 TH |
1110 | spin_unlock_irq(&pcpu_lock); |
1111 | ||
fbf59bc9 TH |
1112 | /* populate, map and clear the area */ |
1113 | if (pcpu_populate_chunk(chunk, off, size)) { | |
ccea34b5 | 1114 | spin_lock_irq(&pcpu_lock); |
fbf59bc9 | 1115 | pcpu_free_area(chunk, off); |
f2badb0c | 1116 | err = "failed to populate"; |
ccea34b5 | 1117 | goto fail_unlock; |
fbf59bc9 TH |
1118 | } |
1119 | ||
ccea34b5 TH |
1120 | mutex_unlock(&pcpu_alloc_mutex); |
1121 | ||
bba174f5 TH |
1122 | /* return address relative to base address */ |
1123 | return __addr_to_pcpu_ptr(chunk->base_addr + off); | |
ccea34b5 TH |
1124 | |
1125 | fail_unlock: | |
1126 | spin_unlock_irq(&pcpu_lock); | |
1127 | fail_unlock_mutex: | |
1128 | mutex_unlock(&pcpu_alloc_mutex); | |
f2badb0c TH |
1129 | if (warn_limit) { |
1130 | pr_warning("PERCPU: allocation failed, size=%zu align=%zu, " | |
1131 | "%s\n", size, align, err); | |
1132 | dump_stack(); | |
1133 | if (!--warn_limit) | |
1134 | pr_info("PERCPU: limit reached, disable warning\n"); | |
1135 | } | |
ccea34b5 | 1136 | return NULL; |
fbf59bc9 | 1137 | } |
edcb4639 TH |
1138 | |
1139 | /** | |
1140 | * __alloc_percpu - allocate dynamic percpu area | |
1141 | * @size: size of area to allocate in bytes | |
1142 | * @align: alignment of area (max PAGE_SIZE) | |
1143 | * | |
1144 | * Allocate percpu area of @size bytes aligned at @align. Might | |
1145 | * sleep. Might trigger writeouts. | |
1146 | * | |
ccea34b5 TH |
1147 | * CONTEXT: |
1148 | * Does GFP_KERNEL allocation. | |
1149 | * | |
edcb4639 TH |
1150 | * RETURNS: |
1151 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1152 | */ | |
1153 | void *__alloc_percpu(size_t size, size_t align) | |
1154 | { | |
1155 | return pcpu_alloc(size, align, false); | |
1156 | } | |
fbf59bc9 TH |
1157 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1158 | ||
edcb4639 TH |
1159 | /** |
1160 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1161 | * @size: size of area to allocate in bytes | |
1162 | * @align: alignment of area (max PAGE_SIZE) | |
1163 | * | |
1164 | * Allocate percpu area of @size bytes aligned at @align from reserved | |
1165 | * percpu area if arch has set it up; otherwise, allocation is served | |
1166 | * from the same dynamic area. Might sleep. Might trigger writeouts. | |
1167 | * | |
ccea34b5 TH |
1168 | * CONTEXT: |
1169 | * Does GFP_KERNEL allocation. | |
1170 | * | |
edcb4639 TH |
1171 | * RETURNS: |
1172 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1173 | */ | |
1174 | void *__alloc_reserved_percpu(size_t size, size_t align) | |
1175 | { | |
1176 | return pcpu_alloc(size, align, true); | |
1177 | } | |
1178 | ||
a56dbddf TH |
1179 | /** |
1180 | * pcpu_reclaim - reclaim fully free chunks, workqueue function | |
1181 | * @work: unused | |
1182 | * | |
1183 | * Reclaim all fully free chunks except for the first one. | |
ccea34b5 TH |
1184 | * |
1185 | * CONTEXT: | |
1186 | * workqueue context. | |
a56dbddf TH |
1187 | */ |
1188 | static void pcpu_reclaim(struct work_struct *work) | |
fbf59bc9 | 1189 | { |
a56dbddf TH |
1190 | LIST_HEAD(todo); |
1191 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; | |
1192 | struct pcpu_chunk *chunk, *next; | |
1193 | ||
ccea34b5 TH |
1194 | mutex_lock(&pcpu_alloc_mutex); |
1195 | spin_lock_irq(&pcpu_lock); | |
a56dbddf TH |
1196 | |
1197 | list_for_each_entry_safe(chunk, next, head, list) { | |
1198 | WARN_ON(chunk->immutable); | |
1199 | ||
1200 | /* spare the first one */ | |
1201 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) | |
1202 | continue; | |
1203 | ||
a56dbddf TH |
1204 | list_move(&chunk->list, &todo); |
1205 | } | |
1206 | ||
ccea34b5 | 1207 | spin_unlock_irq(&pcpu_lock); |
a56dbddf TH |
1208 | |
1209 | list_for_each_entry_safe(chunk, next, &todo, list) { | |
ce3141a2 | 1210 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); |
a56dbddf TH |
1211 | free_pcpu_chunk(chunk); |
1212 | } | |
971f3918 TH |
1213 | |
1214 | mutex_unlock(&pcpu_alloc_mutex); | |
fbf59bc9 TH |
1215 | } |
1216 | ||
1217 | /** | |
1218 | * free_percpu - free percpu area | |
1219 | * @ptr: pointer to area to free | |
1220 | * | |
ccea34b5 TH |
1221 | * Free percpu area @ptr. |
1222 | * | |
1223 | * CONTEXT: | |
1224 | * Can be called from atomic context. | |
fbf59bc9 TH |
1225 | */ |
1226 | void free_percpu(void *ptr) | |
1227 | { | |
1228 | void *addr = __pcpu_ptr_to_addr(ptr); | |
1229 | struct pcpu_chunk *chunk; | |
ccea34b5 | 1230 | unsigned long flags; |
fbf59bc9 TH |
1231 | int off; |
1232 | ||
1233 | if (!ptr) | |
1234 | return; | |
1235 | ||
ccea34b5 | 1236 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1237 | |
1238 | chunk = pcpu_chunk_addr_search(addr); | |
bba174f5 | 1239 | off = addr - chunk->base_addr; |
fbf59bc9 TH |
1240 | |
1241 | pcpu_free_area(chunk, off); | |
1242 | ||
a56dbddf | 1243 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
1244 | if (chunk->free_size == pcpu_unit_size) { |
1245 | struct pcpu_chunk *pos; | |
1246 | ||
a56dbddf | 1247 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1248 | if (pos != chunk) { |
a56dbddf | 1249 | schedule_work(&pcpu_reclaim_work); |
fbf59bc9 TH |
1250 | break; |
1251 | } | |
1252 | } | |
1253 | ||
ccea34b5 | 1254 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1255 | } |
1256 | EXPORT_SYMBOL_GPL(free_percpu); | |
1257 | ||
033e48fb TH |
1258 | static inline size_t pcpu_calc_fc_sizes(size_t static_size, |
1259 | size_t reserved_size, | |
1260 | ssize_t *dyn_sizep) | |
1261 | { | |
1262 | size_t size_sum; | |
1263 | ||
1264 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
1265 | (*dyn_sizep >= 0 ? *dyn_sizep : 0)); | |
1266 | if (*dyn_sizep != 0) | |
1267 | *dyn_sizep = size_sum - static_size - reserved_size; | |
1268 | ||
1269 | return size_sum; | |
1270 | } | |
1271 | ||
fbf59bc9 | 1272 | /** |
fd1e8a1f TH |
1273 | * pcpu_alloc_alloc_info - allocate percpu allocation info |
1274 | * @nr_groups: the number of groups | |
1275 | * @nr_units: the number of units | |
1276 | * | |
1277 | * Allocate ai which is large enough for @nr_groups groups containing | |
1278 | * @nr_units units. The returned ai's groups[0].cpu_map points to the | |
1279 | * cpu_map array which is long enough for @nr_units and filled with | |
1280 | * NR_CPUS. It's the caller's responsibility to initialize cpu_map | |
1281 | * pointer of other groups. | |
1282 | * | |
1283 | * RETURNS: | |
1284 | * Pointer to the allocated pcpu_alloc_info on success, NULL on | |
1285 | * failure. | |
1286 | */ | |
1287 | struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups, | |
1288 | int nr_units) | |
1289 | { | |
1290 | struct pcpu_alloc_info *ai; | |
1291 | size_t base_size, ai_size; | |
1292 | void *ptr; | |
1293 | int unit; | |
1294 | ||
1295 | base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]), | |
1296 | __alignof__(ai->groups[0].cpu_map[0])); | |
1297 | ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]); | |
1298 | ||
1299 | ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size)); | |
1300 | if (!ptr) | |
1301 | return NULL; | |
1302 | ai = ptr; | |
1303 | ptr += base_size; | |
1304 | ||
1305 | ai->groups[0].cpu_map = ptr; | |
1306 | ||
1307 | for (unit = 0; unit < nr_units; unit++) | |
1308 | ai->groups[0].cpu_map[unit] = NR_CPUS; | |
1309 | ||
1310 | ai->nr_groups = nr_groups; | |
1311 | ai->__ai_size = PFN_ALIGN(ai_size); | |
1312 | ||
1313 | return ai; | |
1314 | } | |
1315 | ||
1316 | /** | |
1317 | * pcpu_free_alloc_info - free percpu allocation info | |
1318 | * @ai: pcpu_alloc_info to free | |
1319 | * | |
1320 | * Free @ai which was allocated by pcpu_alloc_alloc_info(). | |
1321 | */ | |
1322 | void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai) | |
1323 | { | |
1324 | free_bootmem(__pa(ai), ai->__ai_size); | |
1325 | } | |
1326 | ||
1327 | /** | |
1328 | * pcpu_build_alloc_info - build alloc_info considering distances between CPUs | |
edcb4639 | 1329 | * @reserved_size: the size of reserved percpu area in bytes |
cafe8816 | 1330 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
fd1e8a1f TH |
1331 | * @atom_size: allocation atom size |
1332 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
033e48fb | 1333 | * |
fd1e8a1f TH |
1334 | * This function determines grouping of units, their mappings to cpus |
1335 | * and other parameters considering needed percpu size, allocation | |
1336 | * atom size and distances between CPUs. | |
033e48fb | 1337 | * |
fd1e8a1f TH |
1338 | * Groups are always mutliples of atom size and CPUs which are of |
1339 | * LOCAL_DISTANCE both ways are grouped together and share space for | |
1340 | * units in the same group. The returned configuration is guaranteed | |
1341 | * to have CPUs on different nodes on different groups and >=75% usage | |
1342 | * of allocated virtual address space. | |
033e48fb TH |
1343 | * |
1344 | * RETURNS: | |
fd1e8a1f TH |
1345 | * On success, pointer to the new allocation_info is returned. On |
1346 | * failure, ERR_PTR value is returned. | |
033e48fb | 1347 | */ |
fd1e8a1f TH |
1348 | struct pcpu_alloc_info * __init pcpu_build_alloc_info( |
1349 | size_t reserved_size, ssize_t dyn_size, | |
1350 | size_t atom_size, | |
1351 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
033e48fb TH |
1352 | { |
1353 | static int group_map[NR_CPUS] __initdata; | |
1354 | static int group_cnt[NR_CPUS] __initdata; | |
1355 | const size_t static_size = __per_cpu_end - __per_cpu_start; | |
fd1e8a1f | 1356 | int group_cnt_max = 0, nr_groups = 1, nr_units = 0; |
033e48fb TH |
1357 | size_t size_sum, min_unit_size, alloc_size; |
1358 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
fd1e8a1f | 1359 | int last_allocs, group, unit; |
033e48fb | 1360 | unsigned int cpu, tcpu; |
fd1e8a1f TH |
1361 | struct pcpu_alloc_info *ai; |
1362 | unsigned int *cpu_map; | |
033e48fb | 1363 | |
fb59e72e TH |
1364 | /* this function may be called multiple times */ |
1365 | memset(group_map, 0, sizeof(group_map)); | |
1366 | memset(group_cnt, 0, sizeof(group_map)); | |
1367 | ||
033e48fb TH |
1368 | /* |
1369 | * Determine min_unit_size, alloc_size and max_upa such that | |
fd1e8a1f | 1370 | * alloc_size is multiple of atom_size and is the smallest |
033e48fb TH |
1371 | * which can accomodate 4k aligned segments which are equal to |
1372 | * or larger than min_unit_size. | |
1373 | */ | |
fd1e8a1f | 1374 | size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); |
033e48fb TH |
1375 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); |
1376 | ||
fd1e8a1f | 1377 | alloc_size = roundup(min_unit_size, atom_size); |
033e48fb TH |
1378 | upa = alloc_size / min_unit_size; |
1379 | while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) | |
1380 | upa--; | |
1381 | max_upa = upa; | |
1382 | ||
1383 | /* group cpus according to their proximity */ | |
1384 | for_each_possible_cpu(cpu) { | |
1385 | group = 0; | |
1386 | next_group: | |
1387 | for_each_possible_cpu(tcpu) { | |
1388 | if (cpu == tcpu) | |
1389 | break; | |
fd1e8a1f | 1390 | if (group_map[tcpu] == group && cpu_distance_fn && |
033e48fb TH |
1391 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || |
1392 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
1393 | group++; | |
fd1e8a1f | 1394 | nr_groups = max(nr_groups, group + 1); |
033e48fb TH |
1395 | goto next_group; |
1396 | } | |
1397 | } | |
1398 | group_map[cpu] = group; | |
1399 | group_cnt[group]++; | |
1400 | group_cnt_max = max(group_cnt_max, group_cnt[group]); | |
1401 | } | |
1402 | ||
1403 | /* | |
1404 | * Expand unit size until address space usage goes over 75% | |
1405 | * and then as much as possible without using more address | |
1406 | * space. | |
1407 | */ | |
1408 | last_allocs = INT_MAX; | |
1409 | for (upa = max_upa; upa; upa--) { | |
1410 | int allocs = 0, wasted = 0; | |
1411 | ||
1412 | if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) | |
1413 | continue; | |
1414 | ||
fd1e8a1f | 1415 | for (group = 0; group < nr_groups; group++) { |
033e48fb TH |
1416 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); |
1417 | allocs += this_allocs; | |
1418 | wasted += this_allocs * upa - group_cnt[group]; | |
1419 | } | |
1420 | ||
1421 | /* | |
1422 | * Don't accept if wastage is over 25%. The | |
1423 | * greater-than comparison ensures upa==1 always | |
1424 | * passes the following check. | |
1425 | */ | |
1426 | if (wasted > num_possible_cpus() / 3) | |
1427 | continue; | |
1428 | ||
1429 | /* and then don't consume more memory */ | |
1430 | if (allocs > last_allocs) | |
1431 | break; | |
1432 | last_allocs = allocs; | |
1433 | best_upa = upa; | |
1434 | } | |
fd1e8a1f TH |
1435 | upa = best_upa; |
1436 | ||
1437 | /* allocate and fill alloc_info */ | |
1438 | for (group = 0; group < nr_groups; group++) | |
1439 | nr_units += roundup(group_cnt[group], upa); | |
1440 | ||
1441 | ai = pcpu_alloc_alloc_info(nr_groups, nr_units); | |
1442 | if (!ai) | |
1443 | return ERR_PTR(-ENOMEM); | |
1444 | cpu_map = ai->groups[0].cpu_map; | |
1445 | ||
1446 | for (group = 0; group < nr_groups; group++) { | |
1447 | ai->groups[group].cpu_map = cpu_map; | |
1448 | cpu_map += roundup(group_cnt[group], upa); | |
1449 | } | |
1450 | ||
1451 | ai->static_size = static_size; | |
1452 | ai->reserved_size = reserved_size; | |
1453 | ai->dyn_size = dyn_size; | |
1454 | ai->unit_size = alloc_size / upa; | |
1455 | ai->atom_size = atom_size; | |
1456 | ai->alloc_size = alloc_size; | |
1457 | ||
1458 | for (group = 0, unit = 0; group_cnt[group]; group++) { | |
1459 | struct pcpu_group_info *gi = &ai->groups[group]; | |
1460 | ||
1461 | /* | |
1462 | * Initialize base_offset as if all groups are located | |
1463 | * back-to-back. The caller should update this to | |
1464 | * reflect actual allocation. | |
1465 | */ | |
1466 | gi->base_offset = unit * ai->unit_size; | |
033e48fb | 1467 | |
033e48fb TH |
1468 | for_each_possible_cpu(cpu) |
1469 | if (group_map[cpu] == group) | |
fd1e8a1f TH |
1470 | gi->cpu_map[gi->nr_units++] = cpu; |
1471 | gi->nr_units = roundup(gi->nr_units, upa); | |
1472 | unit += gi->nr_units; | |
033e48fb | 1473 | } |
fd1e8a1f | 1474 | BUG_ON(unit != nr_units); |
033e48fb | 1475 | |
fd1e8a1f | 1476 | return ai; |
033e48fb TH |
1477 | } |
1478 | ||
fd1e8a1f TH |
1479 | /** |
1480 | * pcpu_dump_alloc_info - print out information about pcpu_alloc_info | |
1481 | * @lvl: loglevel | |
1482 | * @ai: allocation info to dump | |
1483 | * | |
1484 | * Print out information about @ai using loglevel @lvl. | |
1485 | */ | |
1486 | static void pcpu_dump_alloc_info(const char *lvl, | |
1487 | const struct pcpu_alloc_info *ai) | |
033e48fb | 1488 | { |
fd1e8a1f | 1489 | int group_width = 1, cpu_width = 1, width; |
033e48fb | 1490 | char empty_str[] = "--------"; |
fd1e8a1f TH |
1491 | int alloc = 0, alloc_end = 0; |
1492 | int group, v; | |
1493 | int upa, apl; /* units per alloc, allocs per line */ | |
1494 | ||
1495 | v = ai->nr_groups; | |
1496 | while (v /= 10) | |
1497 | group_width++; | |
033e48fb | 1498 | |
fd1e8a1f | 1499 | v = num_possible_cpus(); |
033e48fb | 1500 | while (v /= 10) |
fd1e8a1f TH |
1501 | cpu_width++; |
1502 | empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0'; | |
033e48fb | 1503 | |
fd1e8a1f TH |
1504 | upa = ai->alloc_size / ai->unit_size; |
1505 | width = upa * (cpu_width + 1) + group_width + 3; | |
1506 | apl = rounddown_pow_of_two(max(60 / width, 1)); | |
033e48fb | 1507 | |
fd1e8a1f TH |
1508 | printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu", |
1509 | lvl, ai->static_size, ai->reserved_size, ai->dyn_size, | |
1510 | ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size); | |
033e48fb | 1511 | |
fd1e8a1f TH |
1512 | for (group = 0; group < ai->nr_groups; group++) { |
1513 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
1514 | int unit = 0, unit_end = 0; | |
1515 | ||
1516 | BUG_ON(gi->nr_units % upa); | |
1517 | for (alloc_end += gi->nr_units / upa; | |
1518 | alloc < alloc_end; alloc++) { | |
1519 | if (!(alloc % apl)) { | |
033e48fb | 1520 | printk("\n"); |
fd1e8a1f TH |
1521 | printk("%spcpu-alloc: ", lvl); |
1522 | } | |
1523 | printk("[%0*d] ", group_width, group); | |
1524 | ||
1525 | for (unit_end += upa; unit < unit_end; unit++) | |
1526 | if (gi->cpu_map[unit] != NR_CPUS) | |
1527 | printk("%0*d ", cpu_width, | |
1528 | gi->cpu_map[unit]); | |
1529 | else | |
1530 | printk("%s ", empty_str); | |
033e48fb | 1531 | } |
033e48fb TH |
1532 | } |
1533 | printk("\n"); | |
1534 | } | |
033e48fb | 1535 | |
fbf59bc9 | 1536 | /** |
8d408b4b | 1537 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
fd1e8a1f | 1538 | * @ai: pcpu_alloc_info describing how to percpu area is shaped |
38a6be52 | 1539 | * @base_addr: mapped address |
8d408b4b TH |
1540 | * |
1541 | * Initialize the first percpu chunk which contains the kernel static | |
1542 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1543 | * setup path. |
8d408b4b | 1544 | * |
fd1e8a1f TH |
1545 | * @ai contains all information necessary to initialize the first |
1546 | * chunk and prime the dynamic percpu allocator. | |
1547 | * | |
1548 | * @ai->static_size is the size of static percpu area. | |
1549 | * | |
1550 | * @ai->reserved_size, if non-zero, specifies the amount of bytes to | |
edcb4639 TH |
1551 | * reserve after the static area in the first chunk. This reserves |
1552 | * the first chunk such that it's available only through reserved | |
1553 | * percpu allocation. This is primarily used to serve module percpu | |
1554 | * static areas on architectures where the addressing model has | |
1555 | * limited offset range for symbol relocations to guarantee module | |
1556 | * percpu symbols fall inside the relocatable range. | |
1557 | * | |
fd1e8a1f TH |
1558 | * @ai->dyn_size determines the number of bytes available for dynamic |
1559 | * allocation in the first chunk. The area between @ai->static_size + | |
1560 | * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused. | |
6074d5b0 | 1561 | * |
fd1e8a1f TH |
1562 | * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE |
1563 | * and equal to or larger than @ai->static_size + @ai->reserved_size + | |
1564 | * @ai->dyn_size. | |
8d408b4b | 1565 | * |
fd1e8a1f TH |
1566 | * @ai->atom_size is the allocation atom size and used as alignment |
1567 | * for vm areas. | |
8d408b4b | 1568 | * |
fd1e8a1f TH |
1569 | * @ai->alloc_size is the allocation size and always multiple of |
1570 | * @ai->atom_size. This is larger than @ai->atom_size if | |
1571 | * @ai->unit_size is larger than @ai->atom_size. | |
1572 | * | |
1573 | * @ai->nr_groups and @ai->groups describe virtual memory layout of | |
1574 | * percpu areas. Units which should be colocated are put into the | |
1575 | * same group. Dynamic VM areas will be allocated according to these | |
1576 | * groupings. If @ai->nr_groups is zero, a single group containing | |
1577 | * all units is assumed. | |
8d408b4b | 1578 | * |
38a6be52 TH |
1579 | * The caller should have mapped the first chunk at @base_addr and |
1580 | * copied static data to each unit. | |
fbf59bc9 | 1581 | * |
edcb4639 TH |
1582 | * If the first chunk ends up with both reserved and dynamic areas, it |
1583 | * is served by two chunks - one to serve the core static and reserved | |
1584 | * areas and the other for the dynamic area. They share the same vm | |
1585 | * and page map but uses different area allocation map to stay away | |
1586 | * from each other. The latter chunk is circulated in the chunk slots | |
1587 | * and available for dynamic allocation like any other chunks. | |
1588 | * | |
fbf59bc9 | 1589 | * RETURNS: |
fb435d52 | 1590 | * 0 on success, -errno on failure. |
fbf59bc9 | 1591 | */ |
fb435d52 TH |
1592 | int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai, |
1593 | void *base_addr) | |
fbf59bc9 | 1594 | { |
635b75fc | 1595 | static char cpus_buf[4096] __initdata; |
edcb4639 | 1596 | static int smap[2], dmap[2]; |
fd1e8a1f TH |
1597 | size_t dyn_size = ai->dyn_size; |
1598 | size_t size_sum = ai->static_size + ai->reserved_size + dyn_size; | |
edcb4639 | 1599 | struct pcpu_chunk *schunk, *dchunk = NULL; |
6563297c TH |
1600 | unsigned long *group_offsets; |
1601 | size_t *group_sizes; | |
fb435d52 | 1602 | unsigned long *unit_off; |
fbf59bc9 | 1603 | unsigned int cpu; |
fd1e8a1f TH |
1604 | int *unit_map; |
1605 | int group, unit, i; | |
fbf59bc9 | 1606 | |
635b75fc TH |
1607 | cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask); |
1608 | ||
1609 | #define PCPU_SETUP_BUG_ON(cond) do { \ | |
1610 | if (unlikely(cond)) { \ | |
1611 | pr_emerg("PERCPU: failed to initialize, %s", #cond); \ | |
1612 | pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \ | |
1613 | pcpu_dump_alloc_info(KERN_EMERG, ai); \ | |
1614 | BUG(); \ | |
1615 | } \ | |
1616 | } while (0) | |
1617 | ||
2f39e637 | 1618 | /* sanity checks */ |
edcb4639 TH |
1619 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
1620 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); | |
635b75fc TH |
1621 | PCPU_SETUP_BUG_ON(ai->nr_groups <= 0); |
1622 | PCPU_SETUP_BUG_ON(!ai->static_size); | |
1623 | PCPU_SETUP_BUG_ON(!base_addr); | |
1624 | PCPU_SETUP_BUG_ON(ai->unit_size < size_sum); | |
1625 | PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK); | |
1626 | PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE); | |
8d408b4b | 1627 | |
6563297c TH |
1628 | /* process group information and build config tables accordingly */ |
1629 | group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0])); | |
1630 | group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0])); | |
fd1e8a1f | 1631 | unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0])); |
fb435d52 | 1632 | unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0])); |
2f39e637 | 1633 | |
fd1e8a1f | 1634 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) |
ffe0d5a5 | 1635 | unit_map[cpu] = UINT_MAX; |
fd1e8a1f | 1636 | pcpu_first_unit_cpu = NR_CPUS; |
2f39e637 | 1637 | |
fd1e8a1f TH |
1638 | for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) { |
1639 | const struct pcpu_group_info *gi = &ai->groups[group]; | |
2f39e637 | 1640 | |
6563297c TH |
1641 | group_offsets[group] = gi->base_offset; |
1642 | group_sizes[group] = gi->nr_units * ai->unit_size; | |
1643 | ||
fd1e8a1f TH |
1644 | for (i = 0; i < gi->nr_units; i++) { |
1645 | cpu = gi->cpu_map[i]; | |
1646 | if (cpu == NR_CPUS) | |
1647 | continue; | |
8d408b4b | 1648 | |
635b75fc TH |
1649 | PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids); |
1650 | PCPU_SETUP_BUG_ON(!cpu_possible(cpu)); | |
1651 | PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX); | |
fbf59bc9 | 1652 | |
fd1e8a1f | 1653 | unit_map[cpu] = unit + i; |
fb435d52 TH |
1654 | unit_off[cpu] = gi->base_offset + i * ai->unit_size; |
1655 | ||
fd1e8a1f TH |
1656 | if (pcpu_first_unit_cpu == NR_CPUS) |
1657 | pcpu_first_unit_cpu = cpu; | |
1658 | } | |
2f39e637 | 1659 | } |
fd1e8a1f TH |
1660 | pcpu_last_unit_cpu = cpu; |
1661 | pcpu_nr_units = unit; | |
1662 | ||
1663 | for_each_possible_cpu(cpu) | |
635b75fc TH |
1664 | PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX); |
1665 | ||
1666 | /* we're done parsing the input, undefine BUG macro and dump config */ | |
1667 | #undef PCPU_SETUP_BUG_ON | |
1668 | pcpu_dump_alloc_info(KERN_INFO, ai); | |
fd1e8a1f | 1669 | |
6563297c TH |
1670 | pcpu_nr_groups = ai->nr_groups; |
1671 | pcpu_group_offsets = group_offsets; | |
1672 | pcpu_group_sizes = group_sizes; | |
fd1e8a1f | 1673 | pcpu_unit_map = unit_map; |
fb435d52 | 1674 | pcpu_unit_offsets = unit_off; |
2f39e637 TH |
1675 | |
1676 | /* determine basic parameters */ | |
fd1e8a1f | 1677 | pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT; |
d9b55eeb | 1678 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
6563297c | 1679 | pcpu_atom_size = ai->atom_size; |
ce3141a2 TH |
1680 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
1681 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
cafe8816 | 1682 | |
d9b55eeb TH |
1683 | /* |
1684 | * Allocate chunk slots. The additional last slot is for | |
1685 | * empty chunks. | |
1686 | */ | |
1687 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
fbf59bc9 TH |
1688 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
1689 | for (i = 0; i < pcpu_nr_slots; i++) | |
1690 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1691 | ||
edcb4639 TH |
1692 | /* |
1693 | * Initialize static chunk. If reserved_size is zero, the | |
1694 | * static chunk covers static area + dynamic allocation area | |
1695 | * in the first chunk. If reserved_size is not zero, it | |
1696 | * covers static area + reserved area (mostly used for module | |
1697 | * static percpu allocation). | |
1698 | */ | |
2441d15c TH |
1699 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
1700 | INIT_LIST_HEAD(&schunk->list); | |
bba174f5 | 1701 | schunk->base_addr = base_addr; |
61ace7fa TH |
1702 | schunk->map = smap; |
1703 | schunk->map_alloc = ARRAY_SIZE(smap); | |
38a6be52 | 1704 | schunk->immutable = true; |
ce3141a2 | 1705 | bitmap_fill(schunk->populated, pcpu_unit_pages); |
edcb4639 | 1706 | |
fd1e8a1f TH |
1707 | if (ai->reserved_size) { |
1708 | schunk->free_size = ai->reserved_size; | |
ae9e6bc9 | 1709 | pcpu_reserved_chunk = schunk; |
fd1e8a1f | 1710 | pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size; |
edcb4639 TH |
1711 | } else { |
1712 | schunk->free_size = dyn_size; | |
1713 | dyn_size = 0; /* dynamic area covered */ | |
1714 | } | |
2441d15c | 1715 | schunk->contig_hint = schunk->free_size; |
fbf59bc9 | 1716 | |
fd1e8a1f | 1717 | schunk->map[schunk->map_used++] = -ai->static_size; |
61ace7fa TH |
1718 | if (schunk->free_size) |
1719 | schunk->map[schunk->map_used++] = schunk->free_size; | |
1720 | ||
edcb4639 TH |
1721 | /* init dynamic chunk if necessary */ |
1722 | if (dyn_size) { | |
ce3141a2 | 1723 | dchunk = alloc_bootmem(pcpu_chunk_struct_size); |
edcb4639 | 1724 | INIT_LIST_HEAD(&dchunk->list); |
bba174f5 | 1725 | dchunk->base_addr = base_addr; |
edcb4639 TH |
1726 | dchunk->map = dmap; |
1727 | dchunk->map_alloc = ARRAY_SIZE(dmap); | |
38a6be52 | 1728 | dchunk->immutable = true; |
ce3141a2 | 1729 | bitmap_fill(dchunk->populated, pcpu_unit_pages); |
edcb4639 TH |
1730 | |
1731 | dchunk->contig_hint = dchunk->free_size = dyn_size; | |
1732 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; | |
1733 | dchunk->map[dchunk->map_used++] = dchunk->free_size; | |
1734 | } | |
1735 | ||
2441d15c | 1736 | /* link the first chunk in */ |
ae9e6bc9 TH |
1737 | pcpu_first_chunk = dchunk ?: schunk; |
1738 | pcpu_chunk_relocate(pcpu_first_chunk, -1); | |
fbf59bc9 TH |
1739 | |
1740 | /* we're done */ | |
bba174f5 | 1741 | pcpu_base_addr = base_addr; |
fb435d52 | 1742 | return 0; |
fbf59bc9 | 1743 | } |
66c3a757 | 1744 | |
f58dc01b TH |
1745 | const char *pcpu_fc_names[PCPU_FC_NR] __initdata = { |
1746 | [PCPU_FC_AUTO] = "auto", | |
1747 | [PCPU_FC_EMBED] = "embed", | |
1748 | [PCPU_FC_PAGE] = "page", | |
f58dc01b | 1749 | }; |
66c3a757 | 1750 | |
f58dc01b | 1751 | enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO; |
66c3a757 | 1752 | |
f58dc01b TH |
1753 | static int __init percpu_alloc_setup(char *str) |
1754 | { | |
1755 | if (0) | |
1756 | /* nada */; | |
1757 | #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK | |
1758 | else if (!strcmp(str, "embed")) | |
1759 | pcpu_chosen_fc = PCPU_FC_EMBED; | |
1760 | #endif | |
1761 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
1762 | else if (!strcmp(str, "page")) | |
1763 | pcpu_chosen_fc = PCPU_FC_PAGE; | |
f58dc01b TH |
1764 | #endif |
1765 | else | |
1766 | pr_warning("PERCPU: unknown allocator %s specified\n", str); | |
66c3a757 | 1767 | |
f58dc01b | 1768 | return 0; |
66c3a757 | 1769 | } |
f58dc01b | 1770 | early_param("percpu_alloc", percpu_alloc_setup); |
66c3a757 | 1771 | |
08fc4580 TH |
1772 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
1773 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
66c3a757 TH |
1774 | /** |
1775 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
66c3a757 TH |
1776 | * @reserved_size: the size of reserved percpu area in bytes |
1777 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto | |
c8826dd5 TH |
1778 | * @atom_size: allocation atom size |
1779 | * @cpu_distance_fn: callback to determine distance between cpus, optional | |
1780 | * @alloc_fn: function to allocate percpu page | |
1781 | * @free_fn: funtion to free percpu page | |
66c3a757 TH |
1782 | * |
1783 | * This is a helper to ease setting up embedded first percpu chunk and | |
1784 | * can be called where pcpu_setup_first_chunk() is expected. | |
1785 | * | |
1786 | * If this function is used to setup the first chunk, it is allocated | |
c8826dd5 TH |
1787 | * by calling @alloc_fn and used as-is without being mapped into |
1788 | * vmalloc area. Allocations are always whole multiples of @atom_size | |
1789 | * aligned to @atom_size. | |
1790 | * | |
1791 | * This enables the first chunk to piggy back on the linear physical | |
1792 | * mapping which often uses larger page size. Please note that this | |
1793 | * can result in very sparse cpu->unit mapping on NUMA machines thus | |
1794 | * requiring large vmalloc address space. Don't use this allocator if | |
1795 | * vmalloc space is not orders of magnitude larger than distances | |
1796 | * between node memory addresses (ie. 32bit NUMA machines). | |
66c3a757 TH |
1797 | * |
1798 | * When @dyn_size is positive, dynamic area might be larger than | |
788e5abc TH |
1799 | * specified to fill page alignment. When @dyn_size is auto, |
1800 | * @dyn_size is just big enough to fill page alignment after static | |
1801 | * and reserved areas. | |
66c3a757 TH |
1802 | * |
1803 | * If the needed size is smaller than the minimum or specified unit | |
c8826dd5 | 1804 | * size, the leftover is returned using @free_fn. |
66c3a757 TH |
1805 | * |
1806 | * RETURNS: | |
fb435d52 | 1807 | * 0 on success, -errno on failure. |
66c3a757 | 1808 | */ |
c8826dd5 TH |
1809 | int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size, |
1810 | size_t atom_size, | |
1811 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn, | |
1812 | pcpu_fc_alloc_fn_t alloc_fn, | |
1813 | pcpu_fc_free_fn_t free_fn) | |
66c3a757 | 1814 | { |
c8826dd5 TH |
1815 | void *base = (void *)ULONG_MAX; |
1816 | void **areas = NULL; | |
fd1e8a1f | 1817 | struct pcpu_alloc_info *ai; |
6ea529a2 | 1818 | size_t size_sum, areas_size, max_distance; |
c8826dd5 | 1819 | int group, i, rc; |
66c3a757 | 1820 | |
c8826dd5 TH |
1821 | ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size, |
1822 | cpu_distance_fn); | |
fd1e8a1f TH |
1823 | if (IS_ERR(ai)) |
1824 | return PTR_ERR(ai); | |
66c3a757 | 1825 | |
fd1e8a1f | 1826 | size_sum = ai->static_size + ai->reserved_size + ai->dyn_size; |
c8826dd5 | 1827 | areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *)); |
fa8a7094 | 1828 | |
c8826dd5 TH |
1829 | areas = alloc_bootmem_nopanic(areas_size); |
1830 | if (!areas) { | |
fb435d52 | 1831 | rc = -ENOMEM; |
c8826dd5 | 1832 | goto out_free; |
fa8a7094 | 1833 | } |
66c3a757 | 1834 | |
c8826dd5 TH |
1835 | /* allocate, copy and determine base address */ |
1836 | for (group = 0; group < ai->nr_groups; group++) { | |
1837 | struct pcpu_group_info *gi = &ai->groups[group]; | |
1838 | unsigned int cpu = NR_CPUS; | |
1839 | void *ptr; | |
1840 | ||
1841 | for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++) | |
1842 | cpu = gi->cpu_map[i]; | |
1843 | BUG_ON(cpu == NR_CPUS); | |
1844 | ||
1845 | /* allocate space for the whole group */ | |
1846 | ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size); | |
1847 | if (!ptr) { | |
1848 | rc = -ENOMEM; | |
1849 | goto out_free_areas; | |
1850 | } | |
1851 | areas[group] = ptr; | |
fd1e8a1f | 1852 | |
c8826dd5 TH |
1853 | base = min(ptr, base); |
1854 | ||
1855 | for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) { | |
1856 | if (gi->cpu_map[i] == NR_CPUS) { | |
1857 | /* unused unit, free whole */ | |
1858 | free_fn(ptr, ai->unit_size); | |
1859 | continue; | |
1860 | } | |
1861 | /* copy and return the unused part */ | |
1862 | memcpy(ptr, __per_cpu_load, ai->static_size); | |
1863 | free_fn(ptr + size_sum, ai->unit_size - size_sum); | |
1864 | } | |
fa8a7094 | 1865 | } |
66c3a757 | 1866 | |
c8826dd5 | 1867 | /* base address is now known, determine group base offsets */ |
6ea529a2 TH |
1868 | max_distance = 0; |
1869 | for (group = 0; group < ai->nr_groups; group++) { | |
c8826dd5 | 1870 | ai->groups[group].base_offset = areas[group] - base; |
6ea529a2 TH |
1871 | max_distance = max(max_distance, ai->groups[group].base_offset); |
1872 | } | |
1873 | max_distance += ai->unit_size; | |
1874 | ||
1875 | /* warn if maximum distance is further than 75% of vmalloc space */ | |
1876 | if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) { | |
1877 | pr_warning("PERCPU: max_distance=0x%lx too large for vmalloc " | |
1878 | "space 0x%lx\n", | |
1879 | max_distance, VMALLOC_END - VMALLOC_START); | |
1880 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK | |
1881 | /* and fail if we have fallback */ | |
1882 | rc = -EINVAL; | |
1883 | goto out_free; | |
1884 | #endif | |
1885 | } | |
c8826dd5 | 1886 | |
004018e2 | 1887 | pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
fd1e8a1f TH |
1888 | PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size, |
1889 | ai->dyn_size, ai->unit_size); | |
d4b95f80 | 1890 | |
fb435d52 | 1891 | rc = pcpu_setup_first_chunk(ai, base); |
c8826dd5 TH |
1892 | goto out_free; |
1893 | ||
1894 | out_free_areas: | |
1895 | for (group = 0; group < ai->nr_groups; group++) | |
1896 | free_fn(areas[group], | |
1897 | ai->groups[group].nr_units * ai->unit_size); | |
1898 | out_free: | |
fd1e8a1f | 1899 | pcpu_free_alloc_info(ai); |
c8826dd5 TH |
1900 | if (areas) |
1901 | free_bootmem(__pa(areas), areas_size); | |
fb435d52 | 1902 | return rc; |
d4b95f80 | 1903 | } |
08fc4580 TH |
1904 | #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || |
1905 | !CONFIG_HAVE_SETUP_PER_CPU_AREA */ | |
d4b95f80 | 1906 | |
08fc4580 | 1907 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
d4b95f80 | 1908 | /** |
00ae4064 | 1909 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
1910 | * @reserved_size: the size of reserved percpu area in bytes |
1911 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
1912 | * @free_fn: funtion to free percpu page, always called with PAGE_SIZE | |
1913 | * @populate_pte_fn: function to populate pte | |
1914 | * | |
00ae4064 TH |
1915 | * This is a helper to ease setting up page-remapped first percpu |
1916 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
1917 | * |
1918 | * This is the basic allocator. Static percpu area is allocated | |
1919 | * page-by-page into vmalloc area. | |
1920 | * | |
1921 | * RETURNS: | |
fb435d52 | 1922 | * 0 on success, -errno on failure. |
d4b95f80 | 1923 | */ |
fb435d52 TH |
1924 | int __init pcpu_page_first_chunk(size_t reserved_size, |
1925 | pcpu_fc_alloc_fn_t alloc_fn, | |
1926 | pcpu_fc_free_fn_t free_fn, | |
1927 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 1928 | { |
8f05a6a6 | 1929 | static struct vm_struct vm; |
fd1e8a1f | 1930 | struct pcpu_alloc_info *ai; |
00ae4064 | 1931 | char psize_str[16]; |
ce3141a2 | 1932 | int unit_pages; |
d4b95f80 | 1933 | size_t pages_size; |
ce3141a2 | 1934 | struct page **pages; |
fb435d52 | 1935 | int unit, i, j, rc; |
d4b95f80 | 1936 | |
00ae4064 TH |
1937 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
1938 | ||
fd1e8a1f TH |
1939 | ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL); |
1940 | if (IS_ERR(ai)) | |
1941 | return PTR_ERR(ai); | |
1942 | BUG_ON(ai->nr_groups != 1); | |
1943 | BUG_ON(ai->groups[0].nr_units != num_possible_cpus()); | |
1944 | ||
1945 | unit_pages = ai->unit_size >> PAGE_SHIFT; | |
d4b95f80 TH |
1946 | |
1947 | /* unaligned allocations can't be freed, round up to page size */ | |
fd1e8a1f TH |
1948 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
1949 | sizeof(pages[0])); | |
ce3141a2 | 1950 | pages = alloc_bootmem(pages_size); |
d4b95f80 | 1951 | |
8f05a6a6 | 1952 | /* allocate pages */ |
d4b95f80 | 1953 | j = 0; |
fd1e8a1f | 1954 | for (unit = 0; unit < num_possible_cpus(); unit++) |
ce3141a2 | 1955 | for (i = 0; i < unit_pages; i++) { |
fd1e8a1f | 1956 | unsigned int cpu = ai->groups[0].cpu_map[unit]; |
d4b95f80 TH |
1957 | void *ptr; |
1958 | ||
3cbc8565 | 1959 | ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE); |
d4b95f80 | 1960 | if (!ptr) { |
00ae4064 TH |
1961 | pr_warning("PERCPU: failed to allocate %s page " |
1962 | "for cpu%u\n", psize_str, cpu); | |
d4b95f80 TH |
1963 | goto enomem; |
1964 | } | |
ce3141a2 | 1965 | pages[j++] = virt_to_page(ptr); |
d4b95f80 TH |
1966 | } |
1967 | ||
8f05a6a6 TH |
1968 | /* allocate vm area, map the pages and copy static data */ |
1969 | vm.flags = VM_ALLOC; | |
fd1e8a1f | 1970 | vm.size = num_possible_cpus() * ai->unit_size; |
8f05a6a6 TH |
1971 | vm_area_register_early(&vm, PAGE_SIZE); |
1972 | ||
fd1e8a1f | 1973 | for (unit = 0; unit < num_possible_cpus(); unit++) { |
1d9d3257 | 1974 | unsigned long unit_addr = |
fd1e8a1f | 1975 | (unsigned long)vm.addr + unit * ai->unit_size; |
8f05a6a6 | 1976 | |
ce3141a2 | 1977 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
1978 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
1979 | ||
1980 | /* pte already populated, the following shouldn't fail */ | |
fb435d52 TH |
1981 | rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages], |
1982 | unit_pages); | |
1983 | if (rc < 0) | |
1984 | panic("failed to map percpu area, err=%d\n", rc); | |
66c3a757 | 1985 | |
8f05a6a6 TH |
1986 | /* |
1987 | * FIXME: Archs with virtual cache should flush local | |
1988 | * cache for the linear mapping here - something | |
1989 | * equivalent to flush_cache_vmap() on the local cpu. | |
1990 | * flush_cache_vmap() can't be used as most supporting | |
1991 | * data structures are not set up yet. | |
1992 | */ | |
1993 | ||
1994 | /* copy static data */ | |
fd1e8a1f | 1995 | memcpy((void *)unit_addr, __per_cpu_load, ai->static_size); |
66c3a757 TH |
1996 | } |
1997 | ||
1998 | /* we're ready, commit */ | |
1d9d3257 | 1999 | pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n", |
fd1e8a1f TH |
2000 | unit_pages, psize_str, vm.addr, ai->static_size, |
2001 | ai->reserved_size, ai->dyn_size); | |
d4b95f80 | 2002 | |
fb435d52 | 2003 | rc = pcpu_setup_first_chunk(ai, vm.addr); |
d4b95f80 TH |
2004 | goto out_free_ar; |
2005 | ||
2006 | enomem: | |
2007 | while (--j >= 0) | |
ce3141a2 | 2008 | free_fn(page_address(pages[j]), PAGE_SIZE); |
fb435d52 | 2009 | rc = -ENOMEM; |
d4b95f80 | 2010 | out_free_ar: |
ce3141a2 | 2011 | free_bootmem(__pa(pages), pages_size); |
fd1e8a1f | 2012 | pcpu_free_alloc_info(ai); |
fb435d52 | 2013 | return rc; |
d4b95f80 | 2014 | } |
08fc4580 | 2015 | #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ |
d4b95f80 | 2016 | |
e74e3962 TH |
2017 | /* |
2018 | * Generic percpu area setup. | |
2019 | * | |
2020 | * The embedding helper is used because its behavior closely resembles | |
2021 | * the original non-dynamic generic percpu area setup. This is | |
2022 | * important because many archs have addressing restrictions and might | |
2023 | * fail if the percpu area is located far away from the previous | |
2024 | * location. As an added bonus, in non-NUMA cases, embedding is | |
2025 | * generally a good idea TLB-wise because percpu area can piggy back | |
2026 | * on the physical linear memory mapping which uses large page | |
2027 | * mappings on applicable archs. | |
2028 | */ | |
2029 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA | |
2030 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; | |
2031 | EXPORT_SYMBOL(__per_cpu_offset); | |
2032 | ||
c8826dd5 TH |
2033 | static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size, |
2034 | size_t align) | |
2035 | { | |
2036 | return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); | |
2037 | } | |
66c3a757 | 2038 | |
c8826dd5 TH |
2039 | static void __init pcpu_dfl_fc_free(void *ptr, size_t size) |
2040 | { | |
2041 | free_bootmem(__pa(ptr), size); | |
2042 | } | |
2043 | ||
e74e3962 TH |
2044 | void __init setup_per_cpu_areas(void) |
2045 | { | |
e74e3962 TH |
2046 | unsigned long delta; |
2047 | unsigned int cpu; | |
fb435d52 | 2048 | int rc; |
e74e3962 TH |
2049 | |
2050 | /* | |
2051 | * Always reserve area for module percpu variables. That's | |
2052 | * what the legacy allocator did. | |
2053 | */ | |
fb435d52 | 2054 | rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, |
c8826dd5 TH |
2055 | PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL, |
2056 | pcpu_dfl_fc_alloc, pcpu_dfl_fc_free); | |
fb435d52 | 2057 | if (rc < 0) |
e74e3962 TH |
2058 | panic("Failed to initialized percpu areas."); |
2059 | ||
2060 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2061 | for_each_possible_cpu(cpu) | |
fb435d52 | 2062 | __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu]; |
66c3a757 | 2063 | } |
e74e3962 | 2064 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |