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 | * | |
e74e3962 | 49 | * - drop CONFIG_HAVE_LEGACY_PER_CPU_AREA |
fbf59bc9 TH |
50 | * |
51 | * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate | |
e0100983 TH |
52 | * regular address to percpu pointer and back if they need to be |
53 | * different from the default | |
fbf59bc9 | 54 | * |
8d408b4b TH |
55 | * - use pcpu_setup_first_chunk() during percpu area initialization to |
56 | * setup the first chunk containing the kernel static percpu area | |
fbf59bc9 TH |
57 | */ |
58 | ||
59 | #include <linux/bitmap.h> | |
60 | #include <linux/bootmem.h> | |
61 | #include <linux/list.h> | |
a530b795 | 62 | #include <linux/log2.h> |
fbf59bc9 TH |
63 | #include <linux/mm.h> |
64 | #include <linux/module.h> | |
65 | #include <linux/mutex.h> | |
66 | #include <linux/percpu.h> | |
67 | #include <linux/pfn.h> | |
fbf59bc9 | 68 | #include <linux/slab.h> |
ccea34b5 | 69 | #include <linux/spinlock.h> |
fbf59bc9 | 70 | #include <linux/vmalloc.h> |
a56dbddf | 71 | #include <linux/workqueue.h> |
fbf59bc9 TH |
72 | |
73 | #include <asm/cacheflush.h> | |
e0100983 | 74 | #include <asm/sections.h> |
fbf59bc9 TH |
75 | #include <asm/tlbflush.h> |
76 | ||
fbf59bc9 TH |
77 | #define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */ |
78 | #define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */ | |
79 | ||
e0100983 TH |
80 | /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */ |
81 | #ifndef __addr_to_pcpu_ptr | |
82 | #define __addr_to_pcpu_ptr(addr) \ | |
83 | (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \ | |
84 | + (unsigned long)__per_cpu_start) | |
85 | #endif | |
86 | #ifndef __pcpu_ptr_to_addr | |
87 | #define __pcpu_ptr_to_addr(ptr) \ | |
88 | (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \ | |
89 | - (unsigned long)__per_cpu_start) | |
90 | #endif | |
91 | ||
fbf59bc9 TH |
92 | struct pcpu_chunk { |
93 | struct list_head list; /* linked to pcpu_slot lists */ | |
fbf59bc9 TH |
94 | int free_size; /* free bytes in the chunk */ |
95 | int contig_hint; /* max contiguous size hint */ | |
96 | struct vm_struct *vm; /* mapped vmalloc region */ | |
97 | int map_used; /* # of map entries used */ | |
98 | int map_alloc; /* # of map entries allocated */ | |
99 | int *map; /* allocation map */ | |
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; |
40150d37 TH |
107 | static int pcpu_chunk_size __read_mostly; |
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 | ||
2f39e637 TH |
119 | /* cpu -> unit map */ |
120 | const int *pcpu_unit_map __read_mostly; | |
121 | ||
ae9e6bc9 TH |
122 | /* |
123 | * The first chunk which always exists. Note that unlike other | |
124 | * chunks, this one can be allocated and mapped in several different | |
125 | * ways and thus often doesn't live in the vmalloc area. | |
126 | */ | |
127 | static struct pcpu_chunk *pcpu_first_chunk; | |
128 | ||
129 | /* | |
130 | * Optional reserved chunk. This chunk reserves part of the first | |
131 | * chunk and serves it for reserved allocations. The amount of | |
132 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved | |
133 | * area doesn't exist, the following variables contain NULL and 0 | |
134 | * respectively. | |
135 | */ | |
edcb4639 | 136 | static struct pcpu_chunk *pcpu_reserved_chunk; |
edcb4639 TH |
137 | static int pcpu_reserved_chunk_limit; |
138 | ||
fbf59bc9 | 139 | /* |
ccea34b5 TH |
140 | * Synchronization rules. |
141 | * | |
142 | * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former | |
ce3141a2 TH |
143 | * protects allocation/reclaim paths, chunks, populated bitmap and |
144 | * vmalloc mapping. The latter is a spinlock and protects the index | |
145 | * data structures - chunk slots, chunks and area maps in chunks. | |
ccea34b5 TH |
146 | * |
147 | * During allocation, pcpu_alloc_mutex is kept locked all the time and | |
148 | * pcpu_lock is grabbed and released as necessary. All actual memory | |
149 | * allocations are done using GFP_KERNEL with pcpu_lock released. | |
150 | * | |
151 | * Free path accesses and alters only the index data structures, so it | |
152 | * can be safely called from atomic context. When memory needs to be | |
153 | * returned to the system, free path schedules reclaim_work which | |
154 | * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be | |
155 | * reclaimed, release both locks and frees the chunks. Note that it's | |
156 | * necessary to grab both locks to remove a chunk from circulation as | |
157 | * allocation path might be referencing the chunk with only | |
158 | * pcpu_alloc_mutex locked. | |
fbf59bc9 | 159 | */ |
ccea34b5 TH |
160 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ |
161 | static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ | |
fbf59bc9 | 162 | |
40150d37 | 163 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
fbf59bc9 | 164 | |
a56dbddf TH |
165 | /* reclaim work to release fully free chunks, scheduled from free path */ |
166 | static void pcpu_reclaim(struct work_struct *work); | |
167 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); | |
168 | ||
d9b55eeb | 169 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 170 | { |
cae3aeb8 | 171 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
172 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
173 | } | |
174 | ||
d9b55eeb TH |
175 | static int pcpu_size_to_slot(int size) |
176 | { | |
177 | if (size == pcpu_unit_size) | |
178 | return pcpu_nr_slots - 1; | |
179 | return __pcpu_size_to_slot(size); | |
180 | } | |
181 | ||
fbf59bc9 TH |
182 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
183 | { | |
184 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
185 | return 0; | |
186 | ||
187 | return pcpu_size_to_slot(chunk->free_size); | |
188 | } | |
189 | ||
190 | static int pcpu_page_idx(unsigned int cpu, int page_idx) | |
191 | { | |
2f39e637 | 192 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
193 | } |
194 | ||
fbf59bc9 TH |
195 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
196 | unsigned int cpu, int page_idx) | |
197 | { | |
198 | return (unsigned long)chunk->vm->addr + | |
199 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); | |
200 | } | |
201 | ||
ce3141a2 TH |
202 | static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk, |
203 | unsigned int cpu, int page_idx) | |
c8a51be4 | 204 | { |
ce3141a2 TH |
205 | /* must not be used on pre-mapped chunk */ |
206 | WARN_ON(chunk->immutable); | |
c8a51be4 | 207 | |
ce3141a2 | 208 | return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx)); |
fbf59bc9 TH |
209 | } |
210 | ||
e1b9aa3f CL |
211 | /* set the pointer to a chunk in a page struct */ |
212 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
213 | { | |
214 | page->index = (unsigned long)pcpu; | |
215 | } | |
216 | ||
217 | /* obtain pointer to a chunk from a page struct */ | |
218 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
219 | { | |
220 | return (struct pcpu_chunk *)page->index; | |
221 | } | |
222 | ||
ce3141a2 TH |
223 | static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end) |
224 | { | |
225 | *rs = find_next_zero_bit(chunk->populated, end, *rs); | |
226 | *re = find_next_bit(chunk->populated, end, *rs + 1); | |
227 | } | |
228 | ||
229 | static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end) | |
230 | { | |
231 | *rs = find_next_bit(chunk->populated, end, *rs); | |
232 | *re = find_next_zero_bit(chunk->populated, end, *rs + 1); | |
233 | } | |
234 | ||
235 | /* | |
236 | * (Un)populated page region iterators. Iterate over (un)populated | |
237 | * page regions betwen @start and @end in @chunk. @rs and @re should | |
238 | * be integer variables and will be set to start and end page index of | |
239 | * the current region. | |
240 | */ | |
241 | #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ | |
242 | for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ | |
243 | (rs) < (re); \ | |
244 | (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) | |
245 | ||
246 | #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ | |
247 | for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ | |
248 | (rs) < (re); \ | |
249 | (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) | |
250 | ||
fbf59bc9 | 251 | /** |
1880d93b TH |
252 | * pcpu_mem_alloc - allocate memory |
253 | * @size: bytes to allocate | |
fbf59bc9 | 254 | * |
1880d93b TH |
255 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
256 | * kzalloc() is used; otherwise, vmalloc() is used. The returned | |
257 | * memory is always zeroed. | |
fbf59bc9 | 258 | * |
ccea34b5 TH |
259 | * CONTEXT: |
260 | * Does GFP_KERNEL allocation. | |
261 | * | |
fbf59bc9 | 262 | * RETURNS: |
1880d93b | 263 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 264 | */ |
1880d93b | 265 | static void *pcpu_mem_alloc(size_t size) |
fbf59bc9 | 266 | { |
1880d93b TH |
267 | if (size <= PAGE_SIZE) |
268 | return kzalloc(size, GFP_KERNEL); | |
269 | else { | |
270 | void *ptr = vmalloc(size); | |
271 | if (ptr) | |
272 | memset(ptr, 0, size); | |
273 | return ptr; | |
274 | } | |
275 | } | |
fbf59bc9 | 276 | |
1880d93b TH |
277 | /** |
278 | * pcpu_mem_free - free memory | |
279 | * @ptr: memory to free | |
280 | * @size: size of the area | |
281 | * | |
282 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). | |
283 | */ | |
284 | static void pcpu_mem_free(void *ptr, size_t size) | |
285 | { | |
fbf59bc9 | 286 | if (size <= PAGE_SIZE) |
1880d93b | 287 | kfree(ptr); |
fbf59bc9 | 288 | else |
1880d93b | 289 | vfree(ptr); |
fbf59bc9 TH |
290 | } |
291 | ||
292 | /** | |
293 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
294 | * @chunk: chunk of interest | |
295 | * @oslot: the previous slot it was on | |
296 | * | |
297 | * This function is called after an allocation or free changed @chunk. | |
298 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
299 | * moved to the slot. Note that the reserved chunk is never put on |
300 | * chunk slots. | |
ccea34b5 TH |
301 | * |
302 | * CONTEXT: | |
303 | * pcpu_lock. | |
fbf59bc9 TH |
304 | */ |
305 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
306 | { | |
307 | int nslot = pcpu_chunk_slot(chunk); | |
308 | ||
edcb4639 | 309 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
310 | if (oslot < nslot) |
311 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
312 | else | |
313 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
314 | } | |
315 | } | |
316 | ||
fbf59bc9 | 317 | /** |
e1b9aa3f CL |
318 | * pcpu_chunk_addr_search - determine chunk containing specified address |
319 | * @addr: address for which the chunk needs to be determined. | |
ccea34b5 | 320 | * |
fbf59bc9 TH |
321 | * RETURNS: |
322 | * The address of the found chunk. | |
323 | */ | |
324 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
325 | { | |
ae9e6bc9 | 326 | void *first_start = pcpu_first_chunk->vm->addr; |
fbf59bc9 | 327 | |
ae9e6bc9 | 328 | /* is it in the first chunk? */ |
79ba6ac8 | 329 | if (addr >= first_start && addr < first_start + pcpu_unit_size) { |
ae9e6bc9 TH |
330 | /* is it in the reserved area? */ |
331 | if (addr < first_start + pcpu_reserved_chunk_limit) | |
edcb4639 | 332 | return pcpu_reserved_chunk; |
ae9e6bc9 | 333 | return pcpu_first_chunk; |
edcb4639 TH |
334 | } |
335 | ||
2f39e637 TH |
336 | /* |
337 | * The address is relative to unit0 which might be unused and | |
338 | * thus unmapped. Offset the address to the unit space of the | |
339 | * current processor before looking it up in the vmalloc | |
340 | * space. Note that any possible cpu id can be used here, so | |
341 | * there's no need to worry about preemption or cpu hotplug. | |
342 | */ | |
343 | addr += pcpu_unit_map[smp_processor_id()] * pcpu_unit_size; | |
e1b9aa3f | 344 | return pcpu_get_page_chunk(vmalloc_to_page(addr)); |
fbf59bc9 TH |
345 | } |
346 | ||
9f7dcf22 TH |
347 | /** |
348 | * pcpu_extend_area_map - extend area map for allocation | |
349 | * @chunk: target chunk | |
350 | * | |
351 | * Extend area map of @chunk so that it can accomodate an allocation. | |
352 | * A single allocation can split an area into three areas, so this | |
353 | * function makes sure that @chunk->map has at least two extra slots. | |
354 | * | |
ccea34b5 TH |
355 | * CONTEXT: |
356 | * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired | |
357 | * if area map is extended. | |
358 | * | |
9f7dcf22 TH |
359 | * RETURNS: |
360 | * 0 if noop, 1 if successfully extended, -errno on failure. | |
361 | */ | |
362 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) | |
363 | { | |
364 | int new_alloc; | |
365 | int *new; | |
366 | size_t size; | |
367 | ||
368 | /* has enough? */ | |
369 | if (chunk->map_alloc >= chunk->map_used + 2) | |
370 | return 0; | |
371 | ||
ccea34b5 TH |
372 | spin_unlock_irq(&pcpu_lock); |
373 | ||
9f7dcf22 TH |
374 | new_alloc = PCPU_DFL_MAP_ALLOC; |
375 | while (new_alloc < chunk->map_used + 2) | |
376 | new_alloc *= 2; | |
377 | ||
378 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); | |
ccea34b5 TH |
379 | if (!new) { |
380 | spin_lock_irq(&pcpu_lock); | |
9f7dcf22 | 381 | return -ENOMEM; |
ccea34b5 TH |
382 | } |
383 | ||
384 | /* | |
385 | * Acquire pcpu_lock and switch to new area map. Only free | |
386 | * could have happened inbetween, so map_used couldn't have | |
387 | * grown. | |
388 | */ | |
389 | spin_lock_irq(&pcpu_lock); | |
390 | BUG_ON(new_alloc < chunk->map_used + 2); | |
9f7dcf22 TH |
391 | |
392 | size = chunk->map_alloc * sizeof(chunk->map[0]); | |
393 | memcpy(new, chunk->map, size); | |
394 | ||
395 | /* | |
396 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is | |
397 | * one of the first chunks and still using static map. | |
398 | */ | |
399 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) | |
400 | pcpu_mem_free(chunk->map, size); | |
401 | ||
402 | chunk->map_alloc = new_alloc; | |
403 | chunk->map = new; | |
404 | return 0; | |
405 | } | |
406 | ||
fbf59bc9 TH |
407 | /** |
408 | * pcpu_split_block - split a map block | |
409 | * @chunk: chunk of interest | |
410 | * @i: index of map block to split | |
cae3aeb8 TH |
411 | * @head: head size in bytes (can be 0) |
412 | * @tail: tail size in bytes (can be 0) | |
fbf59bc9 TH |
413 | * |
414 | * Split the @i'th map block into two or three blocks. If @head is | |
415 | * non-zero, @head bytes block is inserted before block @i moving it | |
416 | * to @i+1 and reducing its size by @head bytes. | |
417 | * | |
418 | * If @tail is non-zero, the target block, which can be @i or @i+1 | |
419 | * depending on @head, is reduced by @tail bytes and @tail byte block | |
420 | * is inserted after the target block. | |
421 | * | |
9f7dcf22 | 422 | * @chunk->map must have enough free slots to accomodate the split. |
ccea34b5 TH |
423 | * |
424 | * CONTEXT: | |
425 | * pcpu_lock. | |
fbf59bc9 | 426 | */ |
9f7dcf22 TH |
427 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
428 | int head, int tail) | |
fbf59bc9 TH |
429 | { |
430 | int nr_extra = !!head + !!tail; | |
1880d93b | 431 | |
9f7dcf22 | 432 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
fbf59bc9 | 433 | |
9f7dcf22 | 434 | /* insert new subblocks */ |
fbf59bc9 TH |
435 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
436 | sizeof(chunk->map[0]) * (chunk->map_used - i)); | |
437 | chunk->map_used += nr_extra; | |
438 | ||
439 | if (head) { | |
440 | chunk->map[i + 1] = chunk->map[i] - head; | |
441 | chunk->map[i++] = head; | |
442 | } | |
443 | if (tail) { | |
444 | chunk->map[i++] -= tail; | |
445 | chunk->map[i] = tail; | |
446 | } | |
fbf59bc9 TH |
447 | } |
448 | ||
449 | /** | |
450 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
451 | * @chunk: chunk of interest | |
cae3aeb8 | 452 | * @size: wanted size in bytes |
fbf59bc9 TH |
453 | * @align: wanted align |
454 | * | |
455 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
456 | * Note that this function only allocates the offset. It doesn't | |
457 | * populate or map the area. | |
458 | * | |
9f7dcf22 TH |
459 | * @chunk->map must have at least two free slots. |
460 | * | |
ccea34b5 TH |
461 | * CONTEXT: |
462 | * pcpu_lock. | |
463 | * | |
fbf59bc9 | 464 | * RETURNS: |
9f7dcf22 TH |
465 | * Allocated offset in @chunk on success, -1 if no matching area is |
466 | * found. | |
fbf59bc9 TH |
467 | */ |
468 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) | |
469 | { | |
470 | int oslot = pcpu_chunk_slot(chunk); | |
471 | int max_contig = 0; | |
472 | int i, off; | |
473 | ||
fbf59bc9 TH |
474 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
475 | bool is_last = i + 1 == chunk->map_used; | |
476 | int head, tail; | |
477 | ||
478 | /* extra for alignment requirement */ | |
479 | head = ALIGN(off, align) - off; | |
480 | BUG_ON(i == 0 && head != 0); | |
481 | ||
482 | if (chunk->map[i] < 0) | |
483 | continue; | |
484 | if (chunk->map[i] < head + size) { | |
485 | max_contig = max(chunk->map[i], max_contig); | |
486 | continue; | |
487 | } | |
488 | ||
489 | /* | |
490 | * If head is small or the previous block is free, | |
491 | * merge'em. Note that 'small' is defined as smaller | |
492 | * than sizeof(int), which is very small but isn't too | |
493 | * uncommon for percpu allocations. | |
494 | */ | |
495 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { | |
496 | if (chunk->map[i - 1] > 0) | |
497 | chunk->map[i - 1] += head; | |
498 | else { | |
499 | chunk->map[i - 1] -= head; | |
500 | chunk->free_size -= head; | |
501 | } | |
502 | chunk->map[i] -= head; | |
503 | off += head; | |
504 | head = 0; | |
505 | } | |
506 | ||
507 | /* if tail is small, just keep it around */ | |
508 | tail = chunk->map[i] - head - size; | |
509 | if (tail < sizeof(int)) | |
510 | tail = 0; | |
511 | ||
512 | /* split if warranted */ | |
513 | if (head || tail) { | |
9f7dcf22 | 514 | pcpu_split_block(chunk, i, head, tail); |
fbf59bc9 TH |
515 | if (head) { |
516 | i++; | |
517 | off += head; | |
518 | max_contig = max(chunk->map[i - 1], max_contig); | |
519 | } | |
520 | if (tail) | |
521 | max_contig = max(chunk->map[i + 1], max_contig); | |
522 | } | |
523 | ||
524 | /* update hint and mark allocated */ | |
525 | if (is_last) | |
526 | chunk->contig_hint = max_contig; /* fully scanned */ | |
527 | else | |
528 | chunk->contig_hint = max(chunk->contig_hint, | |
529 | max_contig); | |
530 | ||
531 | chunk->free_size -= chunk->map[i]; | |
532 | chunk->map[i] = -chunk->map[i]; | |
533 | ||
534 | pcpu_chunk_relocate(chunk, oslot); | |
535 | return off; | |
536 | } | |
537 | ||
538 | chunk->contig_hint = max_contig; /* fully scanned */ | |
539 | pcpu_chunk_relocate(chunk, oslot); | |
540 | ||
9f7dcf22 TH |
541 | /* tell the upper layer that this chunk has no matching area */ |
542 | return -1; | |
fbf59bc9 TH |
543 | } |
544 | ||
545 | /** | |
546 | * pcpu_free_area - free area to a pcpu_chunk | |
547 | * @chunk: chunk of interest | |
548 | * @freeme: offset of area to free | |
549 | * | |
550 | * Free area starting from @freeme to @chunk. Note that this function | |
551 | * only modifies the allocation map. It doesn't depopulate or unmap | |
552 | * the area. | |
ccea34b5 TH |
553 | * |
554 | * CONTEXT: | |
555 | * pcpu_lock. | |
fbf59bc9 TH |
556 | */ |
557 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) | |
558 | { | |
559 | int oslot = pcpu_chunk_slot(chunk); | |
560 | int i, off; | |
561 | ||
562 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) | |
563 | if (off == freeme) | |
564 | break; | |
565 | BUG_ON(off != freeme); | |
566 | BUG_ON(chunk->map[i] > 0); | |
567 | ||
568 | chunk->map[i] = -chunk->map[i]; | |
569 | chunk->free_size += chunk->map[i]; | |
570 | ||
571 | /* merge with previous? */ | |
572 | if (i > 0 && chunk->map[i - 1] >= 0) { | |
573 | chunk->map[i - 1] += chunk->map[i]; | |
574 | chunk->map_used--; | |
575 | memmove(&chunk->map[i], &chunk->map[i + 1], | |
576 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
577 | i--; | |
578 | } | |
579 | /* merge with next? */ | |
580 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { | |
581 | chunk->map[i] += chunk->map[i + 1]; | |
582 | chunk->map_used--; | |
583 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], | |
584 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); | |
585 | } | |
586 | ||
587 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); | |
588 | pcpu_chunk_relocate(chunk, oslot); | |
589 | } | |
590 | ||
591 | /** | |
ce3141a2 TH |
592 | * pcpu_get_pages_and_bitmap - get temp pages array and bitmap |
593 | * @chunk: chunk of interest | |
594 | * @bitmapp: output parameter for bitmap | |
595 | * @may_alloc: may allocate the array | |
596 | * | |
597 | * Returns pointer to array of pointers to struct page and bitmap, | |
598 | * both of which can be indexed with pcpu_page_idx(). The returned | |
599 | * array is cleared to zero and *@bitmapp is copied from | |
600 | * @chunk->populated. Note that there is only one array and bitmap | |
601 | * and access exclusion is the caller's responsibility. | |
602 | * | |
603 | * CONTEXT: | |
604 | * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc. | |
605 | * Otherwise, don't care. | |
606 | * | |
607 | * RETURNS: | |
608 | * Pointer to temp pages array on success, NULL on failure. | |
609 | */ | |
610 | static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk, | |
611 | unsigned long **bitmapp, | |
612 | bool may_alloc) | |
613 | { | |
614 | static struct page **pages; | |
615 | static unsigned long *bitmap; | |
2f39e637 | 616 | size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]); |
ce3141a2 TH |
617 | size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) * |
618 | sizeof(unsigned long); | |
619 | ||
620 | if (!pages || !bitmap) { | |
621 | if (may_alloc && !pages) | |
622 | pages = pcpu_mem_alloc(pages_size); | |
623 | if (may_alloc && !bitmap) | |
624 | bitmap = pcpu_mem_alloc(bitmap_size); | |
625 | if (!pages || !bitmap) | |
626 | return NULL; | |
627 | } | |
628 | ||
629 | memset(pages, 0, pages_size); | |
630 | bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages); | |
631 | ||
632 | *bitmapp = bitmap; | |
633 | return pages; | |
634 | } | |
635 | ||
636 | /** | |
637 | * pcpu_free_pages - free pages which were allocated for @chunk | |
638 | * @chunk: chunk pages were allocated for | |
639 | * @pages: array of pages to be freed, indexed by pcpu_page_idx() | |
640 | * @populated: populated bitmap | |
641 | * @page_start: page index of the first page to be freed | |
642 | * @page_end: page index of the last page to be freed + 1 | |
643 | * | |
644 | * Free pages [@page_start and @page_end) in @pages for all units. | |
645 | * The pages were allocated for @chunk. | |
646 | */ | |
647 | static void pcpu_free_pages(struct pcpu_chunk *chunk, | |
648 | struct page **pages, unsigned long *populated, | |
649 | int page_start, int page_end) | |
650 | { | |
651 | unsigned int cpu; | |
652 | int i; | |
653 | ||
654 | for_each_possible_cpu(cpu) { | |
655 | for (i = page_start; i < page_end; i++) { | |
656 | struct page *page = pages[pcpu_page_idx(cpu, i)]; | |
657 | ||
658 | if (page) | |
659 | __free_page(page); | |
660 | } | |
661 | } | |
662 | } | |
663 | ||
664 | /** | |
665 | * pcpu_alloc_pages - allocates pages for @chunk | |
666 | * @chunk: target chunk | |
667 | * @pages: array to put the allocated pages into, indexed by pcpu_page_idx() | |
668 | * @populated: populated bitmap | |
669 | * @page_start: page index of the first page to be allocated | |
670 | * @page_end: page index of the last page to be allocated + 1 | |
671 | * | |
672 | * Allocate pages [@page_start,@page_end) into @pages for all units. | |
673 | * The allocation is for @chunk. Percpu core doesn't care about the | |
674 | * content of @pages and will pass it verbatim to pcpu_map_pages(). | |
675 | */ | |
676 | static int pcpu_alloc_pages(struct pcpu_chunk *chunk, | |
677 | struct page **pages, unsigned long *populated, | |
678 | int page_start, int page_end) | |
679 | { | |
680 | const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; | |
681 | unsigned int cpu; | |
682 | int i; | |
683 | ||
684 | for_each_possible_cpu(cpu) { | |
685 | for (i = page_start; i < page_end; i++) { | |
686 | struct page **pagep = &pages[pcpu_page_idx(cpu, i)]; | |
687 | ||
688 | *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0); | |
689 | if (!*pagep) { | |
690 | pcpu_free_pages(chunk, pages, populated, | |
691 | page_start, page_end); | |
692 | return -ENOMEM; | |
693 | } | |
694 | } | |
695 | } | |
696 | return 0; | |
697 | } | |
698 | ||
699 | /** | |
700 | * pcpu_pre_unmap_flush - flush cache prior to unmapping | |
701 | * @chunk: chunk the regions to be flushed belongs to | |
702 | * @page_start: page index of the first page to be flushed | |
703 | * @page_end: page index of the last page to be flushed + 1 | |
704 | * | |
705 | * Pages in [@page_start,@page_end) of @chunk are about to be | |
706 | * unmapped. Flush cache. As each flushing trial can be very | |
707 | * expensive, issue flush on the whole region at once rather than | |
708 | * doing it for each cpu. This could be an overkill but is more | |
709 | * scalable. | |
710 | */ | |
711 | static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk, | |
712 | int page_start, int page_end) | |
713 | { | |
2f39e637 TH |
714 | flush_cache_vunmap( |
715 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
716 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
ce3141a2 TH |
717 | } |
718 | ||
719 | static void __pcpu_unmap_pages(unsigned long addr, int nr_pages) | |
720 | { | |
721 | unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT); | |
722 | } | |
723 | ||
724 | /** | |
725 | * pcpu_unmap_pages - unmap pages out of a pcpu_chunk | |
fbf59bc9 | 726 | * @chunk: chunk of interest |
ce3141a2 TH |
727 | * @pages: pages array which can be used to pass information to free |
728 | * @populated: populated bitmap | |
fbf59bc9 TH |
729 | * @page_start: page index of the first page to unmap |
730 | * @page_end: page index of the last page to unmap + 1 | |
fbf59bc9 TH |
731 | * |
732 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. | |
ce3141a2 TH |
733 | * Corresponding elements in @pages were cleared by the caller and can |
734 | * be used to carry information to pcpu_free_pages() which will be | |
735 | * called after all unmaps are finished. The caller should call | |
736 | * proper pre/post flush functions. | |
fbf59bc9 | 737 | */ |
ce3141a2 TH |
738 | static void pcpu_unmap_pages(struct pcpu_chunk *chunk, |
739 | struct page **pages, unsigned long *populated, | |
740 | int page_start, int page_end) | |
fbf59bc9 | 741 | { |
fbf59bc9 | 742 | unsigned int cpu; |
ce3141a2 | 743 | int i; |
fbf59bc9 | 744 | |
ce3141a2 TH |
745 | for_each_possible_cpu(cpu) { |
746 | for (i = page_start; i < page_end; i++) { | |
747 | struct page *page; | |
8d408b4b | 748 | |
ce3141a2 TH |
749 | page = pcpu_chunk_page(chunk, cpu, i); |
750 | WARN_ON(!page); | |
751 | pages[pcpu_page_idx(cpu, i)] = page; | |
752 | } | |
753 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start), | |
754 | page_end - page_start); | |
755 | } | |
fbf59bc9 | 756 | |
ce3141a2 TH |
757 | for (i = page_start; i < page_end; i++) |
758 | __clear_bit(i, populated); | |
759 | } | |
760 | ||
761 | /** | |
762 | * pcpu_post_unmap_tlb_flush - flush TLB after unmapping | |
763 | * @chunk: pcpu_chunk the regions to be flushed belong to | |
764 | * @page_start: page index of the first page to be flushed | |
765 | * @page_end: page index of the last page to be flushed + 1 | |
766 | * | |
767 | * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush | |
768 | * TLB for the regions. This can be skipped if the area is to be | |
769 | * returned to vmalloc as vmalloc will handle TLB flushing lazily. | |
770 | * | |
771 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once | |
772 | * for the whole region. | |
773 | */ | |
774 | static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk, | |
775 | int page_start, int page_end) | |
776 | { | |
2f39e637 TH |
777 | flush_tlb_kernel_range( |
778 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
779 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
fbf59bc9 TH |
780 | } |
781 | ||
c8a51be4 TH |
782 | static int __pcpu_map_pages(unsigned long addr, struct page **pages, |
783 | int nr_pages) | |
784 | { | |
785 | return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT, | |
786 | PAGE_KERNEL, pages); | |
787 | } | |
788 | ||
789 | /** | |
ce3141a2 | 790 | * pcpu_map_pages - map pages into a pcpu_chunk |
c8a51be4 | 791 | * @chunk: chunk of interest |
ce3141a2 TH |
792 | * @pages: pages array containing pages to be mapped |
793 | * @populated: populated bitmap | |
c8a51be4 TH |
794 | * @page_start: page index of the first page to map |
795 | * @page_end: page index of the last page to map + 1 | |
796 | * | |
ce3141a2 TH |
797 | * For each cpu, map pages [@page_start,@page_end) into @chunk. The |
798 | * caller is responsible for calling pcpu_post_map_flush() after all | |
799 | * mappings are complete. | |
800 | * | |
801 | * This function is responsible for setting corresponding bits in | |
802 | * @chunk->populated bitmap and whatever is necessary for reverse | |
803 | * lookup (addr -> chunk). | |
c8a51be4 | 804 | */ |
ce3141a2 TH |
805 | static int pcpu_map_pages(struct pcpu_chunk *chunk, |
806 | struct page **pages, unsigned long *populated, | |
807 | int page_start, int page_end) | |
c8a51be4 | 808 | { |
ce3141a2 TH |
809 | unsigned int cpu, tcpu; |
810 | int i, err; | |
c8a51be4 TH |
811 | |
812 | for_each_possible_cpu(cpu) { | |
813 | err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start), | |
ce3141a2 | 814 | &pages[pcpu_page_idx(cpu, page_start)], |
c8a51be4 TH |
815 | page_end - page_start); |
816 | if (err < 0) | |
ce3141a2 | 817 | goto err; |
c8a51be4 TH |
818 | } |
819 | ||
ce3141a2 TH |
820 | /* mapping successful, link chunk and mark populated */ |
821 | for (i = page_start; i < page_end; i++) { | |
822 | for_each_possible_cpu(cpu) | |
823 | pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)], | |
824 | chunk); | |
825 | __set_bit(i, populated); | |
826 | } | |
827 | ||
828 | return 0; | |
829 | ||
830 | err: | |
831 | for_each_possible_cpu(tcpu) { | |
832 | if (tcpu == cpu) | |
833 | break; | |
834 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start), | |
835 | page_end - page_start); | |
836 | } | |
837 | return err; | |
838 | } | |
839 | ||
840 | /** | |
841 | * pcpu_post_map_flush - flush cache after mapping | |
842 | * @chunk: pcpu_chunk the regions to be flushed belong to | |
843 | * @page_start: page index of the first page to be flushed | |
844 | * @page_end: page index of the last page to be flushed + 1 | |
845 | * | |
846 | * Pages [@page_start,@page_end) of @chunk have been mapped. Flush | |
847 | * cache. | |
848 | * | |
849 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once | |
850 | * for the whole region. | |
851 | */ | |
852 | static void pcpu_post_map_flush(struct pcpu_chunk *chunk, | |
853 | int page_start, int page_end) | |
854 | { | |
2f39e637 TH |
855 | flush_cache_vmap( |
856 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
857 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
c8a51be4 TH |
858 | } |
859 | ||
fbf59bc9 TH |
860 | /** |
861 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk | |
862 | * @chunk: chunk to depopulate | |
863 | * @off: offset to the area to depopulate | |
cae3aeb8 | 864 | * @size: size of the area to depopulate in bytes |
fbf59bc9 TH |
865 | * @flush: whether to flush cache and tlb or not |
866 | * | |
867 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) | |
868 | * from @chunk. If @flush is true, vcache is flushed before unmapping | |
869 | * and tlb after. | |
ccea34b5 TH |
870 | * |
871 | * CONTEXT: | |
872 | * pcpu_alloc_mutex. | |
fbf59bc9 | 873 | */ |
ce3141a2 | 874 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size) |
fbf59bc9 TH |
875 | { |
876 | int page_start = PFN_DOWN(off); | |
877 | int page_end = PFN_UP(off + size); | |
ce3141a2 TH |
878 | struct page **pages; |
879 | unsigned long *populated; | |
880 | int rs, re; | |
881 | ||
882 | /* quick path, check whether it's empty already */ | |
883 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { | |
884 | if (rs == page_start && re == page_end) | |
885 | return; | |
886 | break; | |
887 | } | |
fbf59bc9 | 888 | |
ce3141a2 TH |
889 | /* immutable chunks can't be depopulated */ |
890 | WARN_ON(chunk->immutable); | |
fbf59bc9 | 891 | |
ce3141a2 TH |
892 | /* |
893 | * If control reaches here, there must have been at least one | |
894 | * successful population attempt so the temp pages array must | |
895 | * be available now. | |
896 | */ | |
897 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, false); | |
898 | BUG_ON(!pages); | |
fbf59bc9 | 899 | |
ce3141a2 TH |
900 | /* unmap and free */ |
901 | pcpu_pre_unmap_flush(chunk, page_start, page_end); | |
fbf59bc9 | 902 | |
ce3141a2 TH |
903 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) |
904 | pcpu_unmap_pages(chunk, pages, populated, rs, re); | |
fbf59bc9 | 905 | |
ce3141a2 TH |
906 | /* no need to flush tlb, vmalloc will handle it lazily */ |
907 | ||
908 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) | |
909 | pcpu_free_pages(chunk, pages, populated, rs, re); | |
fbf59bc9 | 910 | |
ce3141a2 TH |
911 | /* commit new bitmap */ |
912 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); | |
fbf59bc9 TH |
913 | } |
914 | ||
fbf59bc9 TH |
915 | /** |
916 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk | |
917 | * @chunk: chunk of interest | |
918 | * @off: offset to the area to populate | |
cae3aeb8 | 919 | * @size: size of the area to populate in bytes |
fbf59bc9 TH |
920 | * |
921 | * For each cpu, populate and map pages [@page_start,@page_end) into | |
922 | * @chunk. The area is cleared on return. | |
ccea34b5 TH |
923 | * |
924 | * CONTEXT: | |
925 | * pcpu_alloc_mutex, does GFP_KERNEL allocation. | |
fbf59bc9 TH |
926 | */ |
927 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) | |
928 | { | |
fbf59bc9 TH |
929 | int page_start = PFN_DOWN(off); |
930 | int page_end = PFN_UP(off + size); | |
ce3141a2 TH |
931 | int free_end = page_start, unmap_end = page_start; |
932 | struct page **pages; | |
933 | unsigned long *populated; | |
fbf59bc9 | 934 | unsigned int cpu; |
ce3141a2 | 935 | int rs, re, rc; |
fbf59bc9 | 936 | |
ce3141a2 TH |
937 | /* quick path, check whether all pages are already there */ |
938 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) { | |
939 | if (rs == page_start && re == page_end) | |
940 | goto clear; | |
941 | break; | |
942 | } | |
fbf59bc9 | 943 | |
ce3141a2 TH |
944 | /* need to allocate and map pages, this chunk can't be immutable */ |
945 | WARN_ON(chunk->immutable); | |
fbf59bc9 | 946 | |
ce3141a2 TH |
947 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, true); |
948 | if (!pages) | |
949 | return -ENOMEM; | |
fbf59bc9 | 950 | |
ce3141a2 TH |
951 | /* alloc and map */ |
952 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { | |
953 | rc = pcpu_alloc_pages(chunk, pages, populated, rs, re); | |
954 | if (rc) | |
955 | goto err_free; | |
956 | free_end = re; | |
fbf59bc9 TH |
957 | } |
958 | ||
ce3141a2 TH |
959 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
960 | rc = pcpu_map_pages(chunk, pages, populated, rs, re); | |
961 | if (rc) | |
962 | goto err_unmap; | |
963 | unmap_end = re; | |
964 | } | |
965 | pcpu_post_map_flush(chunk, page_start, page_end); | |
fbf59bc9 | 966 | |
ce3141a2 TH |
967 | /* commit new bitmap */ |
968 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); | |
969 | clear: | |
fbf59bc9 | 970 | for_each_possible_cpu(cpu) |
2f39e637 | 971 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); |
fbf59bc9 | 972 | return 0; |
ce3141a2 TH |
973 | |
974 | err_unmap: | |
975 | pcpu_pre_unmap_flush(chunk, page_start, unmap_end); | |
976 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end) | |
977 | pcpu_unmap_pages(chunk, pages, populated, rs, re); | |
978 | pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end); | |
979 | err_free: | |
980 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end) | |
981 | pcpu_free_pages(chunk, pages, populated, rs, re); | |
982 | return rc; | |
fbf59bc9 TH |
983 | } |
984 | ||
985 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) | |
986 | { | |
987 | if (!chunk) | |
988 | return; | |
989 | if (chunk->vm) | |
990 | free_vm_area(chunk->vm); | |
1880d93b | 991 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
fbf59bc9 TH |
992 | kfree(chunk); |
993 | } | |
994 | ||
995 | static struct pcpu_chunk *alloc_pcpu_chunk(void) | |
996 | { | |
997 | struct pcpu_chunk *chunk; | |
998 | ||
999 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); | |
1000 | if (!chunk) | |
1001 | return NULL; | |
1002 | ||
1880d93b | 1003 | chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
fbf59bc9 TH |
1004 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
1005 | chunk->map[chunk->map_used++] = pcpu_unit_size; | |
1006 | ||
142d44b0 | 1007 | chunk->vm = get_vm_area(pcpu_chunk_size, VM_ALLOC); |
fbf59bc9 TH |
1008 | if (!chunk->vm) { |
1009 | free_pcpu_chunk(chunk); | |
1010 | return NULL; | |
1011 | } | |
1012 | ||
1013 | INIT_LIST_HEAD(&chunk->list); | |
1014 | chunk->free_size = pcpu_unit_size; | |
1015 | chunk->contig_hint = pcpu_unit_size; | |
1016 | ||
1017 | return chunk; | |
1018 | } | |
1019 | ||
1020 | /** | |
edcb4639 | 1021 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 1022 | * @size: size of area to allocate in bytes |
fbf59bc9 | 1023 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 1024 | * @reserved: allocate from the reserved chunk if available |
fbf59bc9 | 1025 | * |
ccea34b5 TH |
1026 | * Allocate percpu area of @size bytes aligned at @align. |
1027 | * | |
1028 | * CONTEXT: | |
1029 | * Does GFP_KERNEL allocation. | |
fbf59bc9 TH |
1030 | * |
1031 | * RETURNS: | |
1032 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1033 | */ | |
edcb4639 | 1034 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
fbf59bc9 | 1035 | { |
fbf59bc9 TH |
1036 | struct pcpu_chunk *chunk; |
1037 | int slot, off; | |
1038 | ||
8d408b4b | 1039 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
fbf59bc9 TH |
1040 | WARN(true, "illegal size (%zu) or align (%zu) for " |
1041 | "percpu allocation\n", size, align); | |
1042 | return NULL; | |
1043 | } | |
1044 | ||
ccea34b5 TH |
1045 | mutex_lock(&pcpu_alloc_mutex); |
1046 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 1047 | |
edcb4639 TH |
1048 | /* serve reserved allocations from the reserved chunk if available */ |
1049 | if (reserved && pcpu_reserved_chunk) { | |
1050 | chunk = pcpu_reserved_chunk; | |
9f7dcf22 TH |
1051 | if (size > chunk->contig_hint || |
1052 | pcpu_extend_area_map(chunk) < 0) | |
ccea34b5 | 1053 | goto fail_unlock; |
edcb4639 TH |
1054 | off = pcpu_alloc_area(chunk, size, align); |
1055 | if (off >= 0) | |
1056 | goto area_found; | |
ccea34b5 | 1057 | goto fail_unlock; |
edcb4639 TH |
1058 | } |
1059 | ||
ccea34b5 | 1060 | restart: |
edcb4639 | 1061 | /* search through normal chunks */ |
fbf59bc9 TH |
1062 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
1063 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
1064 | if (size > chunk->contig_hint) | |
1065 | continue; | |
ccea34b5 TH |
1066 | |
1067 | switch (pcpu_extend_area_map(chunk)) { | |
1068 | case 0: | |
1069 | break; | |
1070 | case 1: | |
1071 | goto restart; /* pcpu_lock dropped, restart */ | |
1072 | default: | |
1073 | goto fail_unlock; | |
1074 | } | |
1075 | ||
fbf59bc9 TH |
1076 | off = pcpu_alloc_area(chunk, size, align); |
1077 | if (off >= 0) | |
1078 | goto area_found; | |
fbf59bc9 TH |
1079 | } |
1080 | } | |
1081 | ||
1082 | /* hmmm... no space left, create a new chunk */ | |
ccea34b5 TH |
1083 | spin_unlock_irq(&pcpu_lock); |
1084 | ||
fbf59bc9 TH |
1085 | chunk = alloc_pcpu_chunk(); |
1086 | if (!chunk) | |
ccea34b5 TH |
1087 | goto fail_unlock_mutex; |
1088 | ||
1089 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 1090 | pcpu_chunk_relocate(chunk, -1); |
ccea34b5 | 1091 | goto restart; |
fbf59bc9 TH |
1092 | |
1093 | area_found: | |
ccea34b5 TH |
1094 | spin_unlock_irq(&pcpu_lock); |
1095 | ||
fbf59bc9 TH |
1096 | /* populate, map and clear the area */ |
1097 | if (pcpu_populate_chunk(chunk, off, size)) { | |
ccea34b5 | 1098 | spin_lock_irq(&pcpu_lock); |
fbf59bc9 | 1099 | pcpu_free_area(chunk, off); |
ccea34b5 | 1100 | goto fail_unlock; |
fbf59bc9 TH |
1101 | } |
1102 | ||
ccea34b5 TH |
1103 | mutex_unlock(&pcpu_alloc_mutex); |
1104 | ||
2f39e637 | 1105 | /* return address relative to unit0 */ |
ccea34b5 TH |
1106 | return __addr_to_pcpu_ptr(chunk->vm->addr + off); |
1107 | ||
1108 | fail_unlock: | |
1109 | spin_unlock_irq(&pcpu_lock); | |
1110 | fail_unlock_mutex: | |
1111 | mutex_unlock(&pcpu_alloc_mutex); | |
1112 | return NULL; | |
fbf59bc9 | 1113 | } |
edcb4639 TH |
1114 | |
1115 | /** | |
1116 | * __alloc_percpu - allocate dynamic percpu area | |
1117 | * @size: size of area to allocate in bytes | |
1118 | * @align: alignment of area (max PAGE_SIZE) | |
1119 | * | |
1120 | * Allocate percpu area of @size bytes aligned at @align. Might | |
1121 | * sleep. Might trigger writeouts. | |
1122 | * | |
ccea34b5 TH |
1123 | * CONTEXT: |
1124 | * Does GFP_KERNEL allocation. | |
1125 | * | |
edcb4639 TH |
1126 | * RETURNS: |
1127 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1128 | */ | |
1129 | void *__alloc_percpu(size_t size, size_t align) | |
1130 | { | |
1131 | return pcpu_alloc(size, align, false); | |
1132 | } | |
fbf59bc9 TH |
1133 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1134 | ||
edcb4639 TH |
1135 | /** |
1136 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1137 | * @size: size of area to allocate in bytes | |
1138 | * @align: alignment of area (max PAGE_SIZE) | |
1139 | * | |
1140 | * Allocate percpu area of @size bytes aligned at @align from reserved | |
1141 | * percpu area if arch has set it up; otherwise, allocation is served | |
1142 | * from the same dynamic area. Might sleep. Might trigger writeouts. | |
1143 | * | |
ccea34b5 TH |
1144 | * CONTEXT: |
1145 | * Does GFP_KERNEL allocation. | |
1146 | * | |
edcb4639 TH |
1147 | * RETURNS: |
1148 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1149 | */ | |
1150 | void *__alloc_reserved_percpu(size_t size, size_t align) | |
1151 | { | |
1152 | return pcpu_alloc(size, align, true); | |
1153 | } | |
1154 | ||
a56dbddf TH |
1155 | /** |
1156 | * pcpu_reclaim - reclaim fully free chunks, workqueue function | |
1157 | * @work: unused | |
1158 | * | |
1159 | * Reclaim all fully free chunks except for the first one. | |
ccea34b5 TH |
1160 | * |
1161 | * CONTEXT: | |
1162 | * workqueue context. | |
a56dbddf TH |
1163 | */ |
1164 | static void pcpu_reclaim(struct work_struct *work) | |
fbf59bc9 | 1165 | { |
a56dbddf TH |
1166 | LIST_HEAD(todo); |
1167 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; | |
1168 | struct pcpu_chunk *chunk, *next; | |
1169 | ||
ccea34b5 TH |
1170 | mutex_lock(&pcpu_alloc_mutex); |
1171 | spin_lock_irq(&pcpu_lock); | |
a56dbddf TH |
1172 | |
1173 | list_for_each_entry_safe(chunk, next, head, list) { | |
1174 | WARN_ON(chunk->immutable); | |
1175 | ||
1176 | /* spare the first one */ | |
1177 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) | |
1178 | continue; | |
1179 | ||
a56dbddf TH |
1180 | list_move(&chunk->list, &todo); |
1181 | } | |
1182 | ||
ccea34b5 | 1183 | spin_unlock_irq(&pcpu_lock); |
a56dbddf TH |
1184 | |
1185 | list_for_each_entry_safe(chunk, next, &todo, list) { | |
ce3141a2 | 1186 | pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size); |
a56dbddf TH |
1187 | free_pcpu_chunk(chunk); |
1188 | } | |
971f3918 TH |
1189 | |
1190 | mutex_unlock(&pcpu_alloc_mutex); | |
fbf59bc9 TH |
1191 | } |
1192 | ||
1193 | /** | |
1194 | * free_percpu - free percpu area | |
1195 | * @ptr: pointer to area to free | |
1196 | * | |
ccea34b5 TH |
1197 | * Free percpu area @ptr. |
1198 | * | |
1199 | * CONTEXT: | |
1200 | * Can be called from atomic context. | |
fbf59bc9 TH |
1201 | */ |
1202 | void free_percpu(void *ptr) | |
1203 | { | |
1204 | void *addr = __pcpu_ptr_to_addr(ptr); | |
1205 | struct pcpu_chunk *chunk; | |
ccea34b5 | 1206 | unsigned long flags; |
fbf59bc9 TH |
1207 | int off; |
1208 | ||
1209 | if (!ptr) | |
1210 | return; | |
1211 | ||
ccea34b5 | 1212 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1213 | |
1214 | chunk = pcpu_chunk_addr_search(addr); | |
1215 | off = addr - chunk->vm->addr; | |
1216 | ||
1217 | pcpu_free_area(chunk, off); | |
1218 | ||
a56dbddf | 1219 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
1220 | if (chunk->free_size == pcpu_unit_size) { |
1221 | struct pcpu_chunk *pos; | |
1222 | ||
a56dbddf | 1223 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1224 | if (pos != chunk) { |
a56dbddf | 1225 | schedule_work(&pcpu_reclaim_work); |
fbf59bc9 TH |
1226 | break; |
1227 | } | |
1228 | } | |
1229 | ||
ccea34b5 | 1230 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1231 | } |
1232 | EXPORT_SYMBOL_GPL(free_percpu); | |
1233 | ||
1234 | /** | |
8d408b4b | 1235 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
8d408b4b | 1236 | * @static_size: the size of static percpu area in bytes |
38a6be52 | 1237 | * @reserved_size: the size of reserved percpu area in bytes, 0 for none |
cafe8816 | 1238 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
38a6be52 TH |
1239 | * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE |
1240 | * @base_addr: mapped address | |
2f39e637 | 1241 | * @unit_map: cpu -> unit map, NULL for sequential mapping |
8d408b4b TH |
1242 | * |
1243 | * Initialize the first percpu chunk which contains the kernel static | |
1244 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1245 | * setup path. |
8d408b4b | 1246 | * |
edcb4639 TH |
1247 | * @reserved_size, if non-zero, specifies the amount of bytes to |
1248 | * reserve after the static area in the first chunk. This reserves | |
1249 | * the first chunk such that it's available only through reserved | |
1250 | * percpu allocation. This is primarily used to serve module percpu | |
1251 | * static areas on architectures where the addressing model has | |
1252 | * limited offset range for symbol relocations to guarantee module | |
1253 | * percpu symbols fall inside the relocatable range. | |
1254 | * | |
6074d5b0 TH |
1255 | * @dyn_size, if non-negative, determines the number of bytes |
1256 | * available for dynamic allocation in the first chunk. Specifying | |
1257 | * non-negative value makes percpu leave alone the area beyond | |
1258 | * @static_size + @reserved_size + @dyn_size. | |
1259 | * | |
38a6be52 TH |
1260 | * @unit_size specifies unit size and must be aligned to PAGE_SIZE and |
1261 | * equal to or larger than @static_size + @reserved_size + if | |
1262 | * non-negative, @dyn_size. | |
8d408b4b | 1263 | * |
38a6be52 TH |
1264 | * The caller should have mapped the first chunk at @base_addr and |
1265 | * copied static data to each unit. | |
fbf59bc9 | 1266 | * |
edcb4639 TH |
1267 | * If the first chunk ends up with both reserved and dynamic areas, it |
1268 | * is served by two chunks - one to serve the core static and reserved | |
1269 | * areas and the other for the dynamic area. They share the same vm | |
1270 | * and page map but uses different area allocation map to stay away | |
1271 | * from each other. The latter chunk is circulated in the chunk slots | |
1272 | * and available for dynamic allocation like any other chunks. | |
1273 | * | |
fbf59bc9 TH |
1274 | * RETURNS: |
1275 | * The determined pcpu_unit_size which can be used to initialize | |
1276 | * percpu access. | |
1277 | */ | |
ce3141a2 | 1278 | size_t __init pcpu_setup_first_chunk(size_t static_size, size_t reserved_size, |
38a6be52 | 1279 | ssize_t dyn_size, size_t unit_size, |
2f39e637 | 1280 | void *base_addr, const int *unit_map) |
fbf59bc9 | 1281 | { |
2441d15c | 1282 | static struct vm_struct first_vm; |
edcb4639 | 1283 | static int smap[2], dmap[2]; |
6074d5b0 TH |
1284 | size_t size_sum = static_size + reserved_size + |
1285 | (dyn_size >= 0 ? dyn_size : 0); | |
edcb4639 | 1286 | struct pcpu_chunk *schunk, *dchunk = NULL; |
2f39e637 | 1287 | unsigned int cpu, tcpu; |
ce3141a2 | 1288 | int i; |
fbf59bc9 | 1289 | |
2f39e637 | 1290 | /* sanity checks */ |
edcb4639 TH |
1291 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
1292 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); | |
8d408b4b | 1293 | BUG_ON(!static_size); |
38a6be52 TH |
1294 | BUG_ON(!base_addr); |
1295 | BUG_ON(unit_size < size_sum); | |
1296 | BUG_ON(unit_size & ~PAGE_MASK); | |
1297 | BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE); | |
8d408b4b | 1298 | |
2f39e637 TH |
1299 | /* determine number of units and verify and initialize pcpu_unit_map */ |
1300 | if (unit_map) { | |
1301 | int first_unit = INT_MAX, last_unit = INT_MIN; | |
1302 | ||
1303 | for_each_possible_cpu(cpu) { | |
1304 | int unit = unit_map[cpu]; | |
1305 | ||
1306 | BUG_ON(unit < 0); | |
1307 | for_each_possible_cpu(tcpu) { | |
1308 | if (tcpu == cpu) | |
1309 | break; | |
1310 | /* the mapping should be one-to-one */ | |
1311 | BUG_ON(unit_map[tcpu] == unit); | |
1312 | } | |
1313 | ||
1314 | if (unit < first_unit) { | |
1315 | pcpu_first_unit_cpu = cpu; | |
1316 | first_unit = unit; | |
1317 | } | |
1318 | if (unit > last_unit) { | |
1319 | pcpu_last_unit_cpu = cpu; | |
1320 | last_unit = unit; | |
1321 | } | |
1322 | } | |
1323 | pcpu_nr_units = last_unit + 1; | |
1324 | pcpu_unit_map = unit_map; | |
1325 | } else { | |
1326 | int *identity_map; | |
1327 | ||
1328 | /* #units == #cpus, identity mapped */ | |
384be2b1 | 1329 | identity_map = alloc_bootmem(nr_cpu_ids * |
2f39e637 TH |
1330 | sizeof(identity_map[0])); |
1331 | ||
1332 | for_each_possible_cpu(cpu) | |
1333 | identity_map[cpu] = cpu; | |
1334 | ||
1335 | pcpu_first_unit_cpu = 0; | |
1336 | pcpu_last_unit_cpu = pcpu_nr_units - 1; | |
384be2b1 | 1337 | pcpu_nr_units = nr_cpu_ids; |
2f39e637 TH |
1338 | pcpu_unit_map = identity_map; |
1339 | } | |
1340 | ||
1341 | /* determine basic parameters */ | |
38a6be52 | 1342 | pcpu_unit_pages = unit_size >> PAGE_SHIFT; |
d9b55eeb | 1343 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
2f39e637 | 1344 | pcpu_chunk_size = pcpu_nr_units * pcpu_unit_size; |
ce3141a2 TH |
1345 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
1346 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
fbf59bc9 | 1347 | |
cafe8816 | 1348 | if (dyn_size < 0) |
edcb4639 | 1349 | dyn_size = pcpu_unit_size - static_size - reserved_size; |
cafe8816 | 1350 | |
38a6be52 TH |
1351 | first_vm.flags = VM_ALLOC; |
1352 | first_vm.size = pcpu_chunk_size; | |
1353 | first_vm.addr = base_addr; | |
1354 | ||
d9b55eeb TH |
1355 | /* |
1356 | * Allocate chunk slots. The additional last slot is for | |
1357 | * empty chunks. | |
1358 | */ | |
1359 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
fbf59bc9 TH |
1360 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
1361 | for (i = 0; i < pcpu_nr_slots; i++) | |
1362 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1363 | ||
edcb4639 TH |
1364 | /* |
1365 | * Initialize static chunk. If reserved_size is zero, the | |
1366 | * static chunk covers static area + dynamic allocation area | |
1367 | * in the first chunk. If reserved_size is not zero, it | |
1368 | * covers static area + reserved area (mostly used for module | |
1369 | * static percpu allocation). | |
1370 | */ | |
2441d15c TH |
1371 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
1372 | INIT_LIST_HEAD(&schunk->list); | |
1373 | schunk->vm = &first_vm; | |
61ace7fa TH |
1374 | schunk->map = smap; |
1375 | schunk->map_alloc = ARRAY_SIZE(smap); | |
38a6be52 | 1376 | schunk->immutable = true; |
ce3141a2 | 1377 | bitmap_fill(schunk->populated, pcpu_unit_pages); |
edcb4639 TH |
1378 | |
1379 | if (reserved_size) { | |
1380 | schunk->free_size = reserved_size; | |
ae9e6bc9 TH |
1381 | pcpu_reserved_chunk = schunk; |
1382 | pcpu_reserved_chunk_limit = static_size + reserved_size; | |
edcb4639 TH |
1383 | } else { |
1384 | schunk->free_size = dyn_size; | |
1385 | dyn_size = 0; /* dynamic area covered */ | |
1386 | } | |
2441d15c | 1387 | schunk->contig_hint = schunk->free_size; |
fbf59bc9 | 1388 | |
61ace7fa TH |
1389 | schunk->map[schunk->map_used++] = -static_size; |
1390 | if (schunk->free_size) | |
1391 | schunk->map[schunk->map_used++] = schunk->free_size; | |
1392 | ||
edcb4639 TH |
1393 | /* init dynamic chunk if necessary */ |
1394 | if (dyn_size) { | |
ce3141a2 | 1395 | dchunk = alloc_bootmem(pcpu_chunk_struct_size); |
edcb4639 TH |
1396 | INIT_LIST_HEAD(&dchunk->list); |
1397 | dchunk->vm = &first_vm; | |
1398 | dchunk->map = dmap; | |
1399 | dchunk->map_alloc = ARRAY_SIZE(dmap); | |
38a6be52 | 1400 | dchunk->immutable = true; |
ce3141a2 | 1401 | bitmap_fill(dchunk->populated, pcpu_unit_pages); |
edcb4639 TH |
1402 | |
1403 | dchunk->contig_hint = dchunk->free_size = dyn_size; | |
1404 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; | |
1405 | dchunk->map[dchunk->map_used++] = dchunk->free_size; | |
1406 | } | |
1407 | ||
2441d15c | 1408 | /* link the first chunk in */ |
ae9e6bc9 TH |
1409 | pcpu_first_chunk = dchunk ?: schunk; |
1410 | pcpu_chunk_relocate(pcpu_first_chunk, -1); | |
fbf59bc9 TH |
1411 | |
1412 | /* we're done */ | |
2f39e637 | 1413 | pcpu_base_addr = schunk->vm->addr; |
fbf59bc9 TH |
1414 | return pcpu_unit_size; |
1415 | } | |
66c3a757 | 1416 | |
08fc4580 TH |
1417 | static inline size_t pcpu_calc_fc_sizes(size_t static_size, |
1418 | size_t reserved_size, | |
1419 | ssize_t *dyn_sizep) | |
8c4bfc6e TH |
1420 | { |
1421 | size_t size_sum; | |
1422 | ||
1423 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
1424 | (*dyn_sizep >= 0 ? *dyn_sizep : 0)); | |
1425 | if (*dyn_sizep != 0) | |
1426 | *dyn_sizep = size_sum - static_size - reserved_size; | |
1427 | ||
1428 | return size_sum; | |
1429 | } | |
1430 | ||
08fc4580 TH |
1431 | #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \ |
1432 | !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA) | |
66c3a757 TH |
1433 | /** |
1434 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
1435 | * @static_size: the size of static percpu area in bytes | |
1436 | * @reserved_size: the size of reserved percpu area in bytes | |
1437 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto | |
66c3a757 TH |
1438 | * |
1439 | * This is a helper to ease setting up embedded first percpu chunk and | |
1440 | * can be called where pcpu_setup_first_chunk() is expected. | |
1441 | * | |
1442 | * If this function is used to setup the first chunk, it is allocated | |
1443 | * as a contiguous area using bootmem allocator and used as-is without | |
1444 | * being mapped into vmalloc area. This enables the first chunk to | |
1445 | * piggy back on the linear physical mapping which often uses larger | |
1446 | * page size. | |
1447 | * | |
1448 | * When @dyn_size is positive, dynamic area might be larger than | |
788e5abc TH |
1449 | * specified to fill page alignment. When @dyn_size is auto, |
1450 | * @dyn_size is just big enough to fill page alignment after static | |
1451 | * and reserved areas. | |
66c3a757 TH |
1452 | * |
1453 | * If the needed size is smaller than the minimum or specified unit | |
1454 | * size, the leftover is returned to the bootmem allocator. | |
1455 | * | |
1456 | * RETURNS: | |
1457 | * The determined pcpu_unit_size which can be used to initialize | |
1458 | * percpu access on success, -errno on failure. | |
1459 | */ | |
1460 | ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size, | |
788e5abc | 1461 | ssize_t dyn_size) |
66c3a757 | 1462 | { |
ce3141a2 TH |
1463 | size_t size_sum, unit_size, chunk_size; |
1464 | void *base; | |
66c3a757 TH |
1465 | unsigned int cpu; |
1466 | ||
1467 | /* determine parameters and allocate */ | |
ce3141a2 | 1468 | size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); |
66c3a757 | 1469 | |
ce3141a2 | 1470 | unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); |
384be2b1 | 1471 | chunk_size = unit_size * nr_cpu_ids; |
fa8a7094 | 1472 | |
ce3141a2 TH |
1473 | base = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE, |
1474 | __pa(MAX_DMA_ADDRESS)); | |
1475 | if (!base) { | |
fa8a7094 TH |
1476 | pr_warning("PERCPU: failed to allocate %zu bytes for " |
1477 | "embedding\n", chunk_size); | |
66c3a757 | 1478 | return -ENOMEM; |
fa8a7094 | 1479 | } |
66c3a757 TH |
1480 | |
1481 | /* return the leftover and copy */ | |
74d46d6b | 1482 | for (cpu = 0; cpu < nr_cpu_ids; cpu++) { |
ce3141a2 | 1483 | void *ptr = base + cpu * unit_size; |
66c3a757 | 1484 | |
74d46d6b | 1485 | if (cpu_possible(cpu)) { |
384be2b1 TH |
1486 | free_bootmem(__pa(ptr + size_sum), |
1487 | unit_size - size_sum); | |
74d46d6b TH |
1488 | memcpy(ptr, __per_cpu_load, static_size); |
1489 | } else | |
384be2b1 | 1490 | free_bootmem(__pa(ptr), unit_size); |
66c3a757 TH |
1491 | } |
1492 | ||
1493 | /* we're ready, commit */ | |
004018e2 TH |
1494 | pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n", |
1495 | PFN_DOWN(size_sum), base, static_size, reserved_size, dyn_size, | |
1496 | unit_size); | |
d4b95f80 | 1497 | |
ce3141a2 | 1498 | return pcpu_setup_first_chunk(static_size, reserved_size, dyn_size, |
2f39e637 | 1499 | unit_size, base, NULL); |
d4b95f80 | 1500 | } |
08fc4580 TH |
1501 | #endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK || |
1502 | !CONFIG_HAVE_SETUP_PER_CPU_AREA */ | |
d4b95f80 | 1503 | |
08fc4580 | 1504 | #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK |
d4b95f80 | 1505 | /** |
00ae4064 | 1506 | * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages |
d4b95f80 TH |
1507 | * @static_size: the size of static percpu area in bytes |
1508 | * @reserved_size: the size of reserved percpu area in bytes | |
1509 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
1510 | * @free_fn: funtion to free percpu page, always called with PAGE_SIZE | |
1511 | * @populate_pte_fn: function to populate pte | |
1512 | * | |
00ae4064 TH |
1513 | * This is a helper to ease setting up page-remapped first percpu |
1514 | * chunk and can be called where pcpu_setup_first_chunk() is expected. | |
d4b95f80 TH |
1515 | * |
1516 | * This is the basic allocator. Static percpu area is allocated | |
1517 | * page-by-page into vmalloc area. | |
1518 | * | |
1519 | * RETURNS: | |
1520 | * The determined pcpu_unit_size which can be used to initialize | |
1521 | * percpu access on success, -errno on failure. | |
1522 | */ | |
00ae4064 TH |
1523 | ssize_t __init pcpu_page_first_chunk(size_t static_size, size_t reserved_size, |
1524 | pcpu_fc_alloc_fn_t alloc_fn, | |
1525 | pcpu_fc_free_fn_t free_fn, | |
1526 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
d4b95f80 | 1527 | { |
8f05a6a6 | 1528 | static struct vm_struct vm; |
00ae4064 | 1529 | char psize_str[16]; |
ce3141a2 | 1530 | int unit_pages; |
d4b95f80 | 1531 | size_t pages_size; |
ce3141a2 | 1532 | struct page **pages; |
d4b95f80 TH |
1533 | unsigned int cpu; |
1534 | int i, j; | |
1535 | ssize_t ret; | |
1536 | ||
00ae4064 TH |
1537 | snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10); |
1538 | ||
ce3141a2 TH |
1539 | unit_pages = PFN_UP(max_t(size_t, static_size + reserved_size, |
1540 | PCPU_MIN_UNIT_SIZE)); | |
d4b95f80 TH |
1541 | |
1542 | /* unaligned allocations can't be freed, round up to page size */ | |
384be2b1 | 1543 | pages_size = PFN_ALIGN(unit_pages * nr_cpu_ids * sizeof(pages[0])); |
ce3141a2 | 1544 | pages = alloc_bootmem(pages_size); |
d4b95f80 | 1545 | |
8f05a6a6 | 1546 | /* allocate pages */ |
d4b95f80 TH |
1547 | j = 0; |
1548 | for_each_possible_cpu(cpu) | |
ce3141a2 | 1549 | for (i = 0; i < unit_pages; i++) { |
d4b95f80 TH |
1550 | void *ptr; |
1551 | ||
1552 | ptr = alloc_fn(cpu, PAGE_SIZE); | |
1553 | if (!ptr) { | |
00ae4064 TH |
1554 | pr_warning("PERCPU: failed to allocate %s page " |
1555 | "for cpu%u\n", psize_str, cpu); | |
d4b95f80 TH |
1556 | goto enomem; |
1557 | } | |
ce3141a2 | 1558 | pages[j++] = virt_to_page(ptr); |
d4b95f80 TH |
1559 | } |
1560 | ||
8f05a6a6 TH |
1561 | /* allocate vm area, map the pages and copy static data */ |
1562 | vm.flags = VM_ALLOC; | |
384be2b1 | 1563 | vm.size = nr_cpu_ids * unit_pages << PAGE_SHIFT; |
8f05a6a6 TH |
1564 | vm_area_register_early(&vm, PAGE_SIZE); |
1565 | ||
1566 | for_each_possible_cpu(cpu) { | |
1567 | unsigned long unit_addr = (unsigned long)vm.addr + | |
ce3141a2 | 1568 | (cpu * unit_pages << PAGE_SHIFT); |
8f05a6a6 | 1569 | |
ce3141a2 | 1570 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
1571 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
1572 | ||
1573 | /* pte already populated, the following shouldn't fail */ | |
ce3141a2 TH |
1574 | ret = __pcpu_map_pages(unit_addr, &pages[cpu * unit_pages], |
1575 | unit_pages); | |
8f05a6a6 TH |
1576 | if (ret < 0) |
1577 | panic("failed to map percpu area, err=%zd\n", ret); | |
1578 | ||
1579 | /* | |
1580 | * FIXME: Archs with virtual cache should flush local | |
1581 | * cache for the linear mapping here - something | |
1582 | * equivalent to flush_cache_vmap() on the local cpu. | |
1583 | * flush_cache_vmap() can't be used as most supporting | |
1584 | * data structures are not set up yet. | |
1585 | */ | |
1586 | ||
1587 | /* copy static data */ | |
1588 | memcpy((void *)unit_addr, __per_cpu_load, static_size); | |
1589 | } | |
1590 | ||
d4b95f80 | 1591 | /* we're ready, commit */ |
00ae4064 TH |
1592 | pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu\n", |
1593 | unit_pages, psize_str, vm.addr, static_size, reserved_size); | |
d4b95f80 | 1594 | |
ce3141a2 | 1595 | ret = pcpu_setup_first_chunk(static_size, reserved_size, -1, |
2f39e637 | 1596 | unit_pages << PAGE_SHIFT, vm.addr, NULL); |
d4b95f80 TH |
1597 | goto out_free_ar; |
1598 | ||
1599 | enomem: | |
1600 | while (--j >= 0) | |
ce3141a2 | 1601 | free_fn(page_address(pages[j]), PAGE_SIZE); |
d4b95f80 TH |
1602 | ret = -ENOMEM; |
1603 | out_free_ar: | |
ce3141a2 | 1604 | free_bootmem(__pa(pages), pages_size); |
d4b95f80 TH |
1605 | return ret; |
1606 | } | |
08fc4580 | 1607 | #endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */ |
d4b95f80 | 1608 | |
08fc4580 | 1609 | #ifdef CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK |
a530b795 TH |
1610 | /** |
1611 | * pcpu_lpage_build_unit_map - build unit_map for large page remapping | |
1612 | * @static_size: the size of static percpu area in bytes | |
1613 | * @reserved_size: the size of reserved percpu area in bytes | |
1614 | * @dyn_sizep: in/out parameter for dynamic size, -1 for auto | |
1615 | * @unit_sizep: out parameter for unit size | |
1616 | * @unit_map: unit_map to be filled | |
1617 | * @cpu_distance_fn: callback to determine distance between cpus | |
1618 | * | |
1619 | * This function builds cpu -> unit map and determine other parameters | |
1620 | * considering needed percpu size, large page size and distances | |
1621 | * between CPUs in NUMA. | |
1622 | * | |
1623 | * CPUs which are of LOCAL_DISTANCE both ways are grouped together and | |
1624 | * may share units in the same large page. The returned configuration | |
1625 | * is guaranteed to have CPUs on different nodes on different large | |
1626 | * pages and >=75% usage of allocated virtual address space. | |
1627 | * | |
1628 | * RETURNS: | |
1629 | * On success, fills in @unit_map, sets *@dyn_sizep, *@unit_sizep and | |
1630 | * returns the number of units to be allocated. -errno on failure. | |
1631 | */ | |
1632 | int __init pcpu_lpage_build_unit_map(size_t static_size, size_t reserved_size, | |
1633 | ssize_t *dyn_sizep, size_t *unit_sizep, | |
1634 | size_t lpage_size, int *unit_map, | |
1635 | pcpu_fc_cpu_distance_fn_t cpu_distance_fn) | |
1636 | { | |
1637 | static int group_map[NR_CPUS] __initdata; | |
1638 | static int group_cnt[NR_CPUS] __initdata; | |
1639 | int group_cnt_max = 0; | |
1640 | size_t size_sum, min_unit_size, alloc_size; | |
1641 | int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */ | |
1642 | int last_allocs; | |
1643 | unsigned int cpu, tcpu; | |
1644 | int group, unit; | |
1645 | ||
1646 | /* | |
1647 | * Determine min_unit_size, alloc_size and max_upa such that | |
1648 | * alloc_size is multiple of lpage_size and is the smallest | |
1649 | * which can accomodate 4k aligned segments which are equal to | |
1650 | * or larger than min_unit_size. | |
1651 | */ | |
1652 | size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, dyn_sizep); | |
1653 | min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
1654 | ||
1655 | alloc_size = roundup(min_unit_size, lpage_size); | |
1656 | upa = alloc_size / min_unit_size; | |
1657 | while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) | |
1658 | upa--; | |
1659 | max_upa = upa; | |
1660 | ||
1661 | /* group cpus according to their proximity */ | |
1662 | for_each_possible_cpu(cpu) { | |
1663 | group = 0; | |
1664 | next_group: | |
1665 | for_each_possible_cpu(tcpu) { | |
1666 | if (cpu == tcpu) | |
1667 | break; | |
1668 | if (group_map[tcpu] == group && | |
1669 | (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE || | |
1670 | cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) { | |
1671 | group++; | |
1672 | goto next_group; | |
1673 | } | |
1674 | } | |
1675 | group_map[cpu] = group; | |
1676 | group_cnt[group]++; | |
1677 | group_cnt_max = max(group_cnt_max, group_cnt[group]); | |
1678 | } | |
1679 | ||
1680 | /* | |
1681 | * Expand unit size until address space usage goes over 75% | |
1682 | * and then as much as possible without using more address | |
1683 | * space. | |
1684 | */ | |
1685 | last_allocs = INT_MAX; | |
1686 | for (upa = max_upa; upa; upa--) { | |
1687 | int allocs = 0, wasted = 0; | |
1688 | ||
1689 | if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK)) | |
1690 | continue; | |
1691 | ||
1692 | for (group = 0; group_cnt[group]; group++) { | |
1693 | int this_allocs = DIV_ROUND_UP(group_cnt[group], upa); | |
1694 | allocs += this_allocs; | |
1695 | wasted += this_allocs * upa - group_cnt[group]; | |
1696 | } | |
1697 | ||
1698 | /* | |
1699 | * Don't accept if wastage is over 25%. The | |
1700 | * greater-than comparison ensures upa==1 always | |
1701 | * passes the following check. | |
1702 | */ | |
1703 | if (wasted > num_possible_cpus() / 3) | |
1704 | continue; | |
1705 | ||
1706 | /* and then don't consume more memory */ | |
1707 | if (allocs > last_allocs) | |
1708 | break; | |
1709 | last_allocs = allocs; | |
1710 | best_upa = upa; | |
1711 | } | |
1712 | *unit_sizep = alloc_size / best_upa; | |
1713 | ||
1714 | /* assign units to cpus accordingly */ | |
1715 | unit = 0; | |
1716 | for (group = 0; group_cnt[group]; group++) { | |
1717 | for_each_possible_cpu(cpu) | |
1718 | if (group_map[cpu] == group) | |
1719 | unit_map[cpu] = unit++; | |
1720 | unit = roundup(unit, best_upa); | |
1721 | } | |
1722 | ||
1723 | return unit; /* unit contains aligned number of units */ | |
1724 | } | |
1725 | ||
8c4bfc6e | 1726 | struct pcpul_ent { |
8c4bfc6e | 1727 | void *ptr; |
a530b795 | 1728 | void *map_addr; |
8c4bfc6e TH |
1729 | }; |
1730 | ||
1731 | static size_t pcpul_size; | |
a530b795 TH |
1732 | static size_t pcpul_lpage_size; |
1733 | static int pcpul_nr_lpages; | |
8c4bfc6e | 1734 | static struct pcpul_ent *pcpul_map; |
a530b795 TH |
1735 | |
1736 | static bool __init pcpul_unit_to_cpu(int unit, const int *unit_map, | |
1737 | unsigned int *cpup) | |
1738 | { | |
1739 | unsigned int cpu; | |
1740 | ||
1741 | for_each_possible_cpu(cpu) | |
1742 | if (unit_map[cpu] == unit) { | |
1743 | if (cpup) | |
1744 | *cpup = cpu; | |
1745 | return true; | |
1746 | } | |
1747 | ||
1748 | return false; | |
1749 | } | |
1750 | ||
1751 | static void __init pcpul_lpage_dump_cfg(const char *lvl, size_t static_size, | |
1752 | size_t reserved_size, size_t dyn_size, | |
1753 | size_t unit_size, size_t lpage_size, | |
1754 | const int *unit_map, int nr_units) | |
1755 | { | |
1756 | int width = 1, v = nr_units; | |
1757 | char empty_str[] = "--------"; | |
1758 | int upl, lpl; /* units per lpage, lpage per line */ | |
1759 | unsigned int cpu; | |
1760 | int lpage, unit; | |
1761 | ||
1762 | while (v /= 10) | |
1763 | width++; | |
1764 | empty_str[min_t(int, width, sizeof(empty_str) - 1)] = '\0'; | |
1765 | ||
1766 | upl = max_t(int, lpage_size / unit_size, 1); | |
1767 | lpl = rounddown_pow_of_two(max_t(int, 60 / (upl * (width + 1) + 2), 1)); | |
1768 | ||
1769 | printk("%spcpu-lpage: sta/res/dyn=%zu/%zu/%zu unit=%zu lpage=%zu", lvl, | |
1770 | static_size, reserved_size, dyn_size, unit_size, lpage_size); | |
1771 | ||
1772 | for (lpage = 0, unit = 0; unit < nr_units; unit++) { | |
1773 | if (!(unit % upl)) { | |
1774 | if (!(lpage++ % lpl)) { | |
1775 | printk("\n"); | |
1776 | printk("%spcpu-lpage: ", lvl); | |
1777 | } else | |
1778 | printk("| "); | |
1779 | } | |
1780 | if (pcpul_unit_to_cpu(unit, unit_map, &cpu)) | |
1781 | printk("%0*d ", width, cpu); | |
1782 | else | |
1783 | printk("%s ", empty_str); | |
1784 | } | |
1785 | printk("\n"); | |
1786 | } | |
8c4bfc6e | 1787 | |
8c4bfc6e TH |
1788 | /** |
1789 | * pcpu_lpage_first_chunk - remap the first percpu chunk using large page | |
1790 | * @static_size: the size of static percpu area in bytes | |
1791 | * @reserved_size: the size of reserved percpu area in bytes | |
a530b795 TH |
1792 | * @dyn_size: free size for dynamic allocation in bytes |
1793 | * @unit_size: unit size in bytes | |
8c4bfc6e | 1794 | * @lpage_size: the size of a large page |
a530b795 TH |
1795 | * @unit_map: cpu -> unit mapping |
1796 | * @nr_units: the number of units | |
8c4bfc6e TH |
1797 | * @alloc_fn: function to allocate percpu lpage, always called with lpage_size |
1798 | * @free_fn: function to free percpu memory, @size <= lpage_size | |
1799 | * @map_fn: function to map percpu lpage, always called with lpage_size | |
1800 | * | |
a530b795 TH |
1801 | * This allocator uses large page to build and map the first chunk. |
1802 | * Unlike other helpers, the caller should always specify @dyn_size | |
1803 | * and @unit_size. These parameters along with @unit_map and | |
1804 | * @nr_units can be determined using pcpu_lpage_build_unit_map(). | |
1805 | * This two stage initialization is to allow arch code to evaluate the | |
1806 | * parameters before committing to it. | |
1807 | * | |
1808 | * Large pages are allocated as directed by @unit_map and other | |
1809 | * parameters and mapped to vmalloc space. Unused holes are returned | |
1810 | * to the page allocator. Note that these holes end up being actively | |
1811 | * mapped twice - once to the physical mapping and to the vmalloc area | |
1812 | * for the first percpu chunk. Depending on architecture, this might | |
1813 | * cause problem when changing page attributes of the returned area. | |
1814 | * These double mapped areas can be detected using | |
1815 | * pcpu_lpage_remapped(). | |
8c4bfc6e TH |
1816 | * |
1817 | * RETURNS: | |
1818 | * The determined pcpu_unit_size which can be used to initialize | |
1819 | * percpu access on success, -errno on failure. | |
1820 | */ | |
1821 | ssize_t __init pcpu_lpage_first_chunk(size_t static_size, size_t reserved_size, | |
a530b795 TH |
1822 | size_t dyn_size, size_t unit_size, |
1823 | size_t lpage_size, const int *unit_map, | |
1824 | int nr_units, | |
8c4bfc6e TH |
1825 | pcpu_fc_alloc_fn_t alloc_fn, |
1826 | pcpu_fc_free_fn_t free_fn, | |
1827 | pcpu_fc_map_fn_t map_fn) | |
1828 | { | |
a530b795 TH |
1829 | static struct vm_struct vm; |
1830 | size_t chunk_size = unit_size * nr_units; | |
8c4bfc6e TH |
1831 | size_t map_size; |
1832 | unsigned int cpu; | |
8c4bfc6e | 1833 | ssize_t ret; |
a530b795 | 1834 | int i, j, unit; |
8c4bfc6e | 1835 | |
a530b795 TH |
1836 | pcpul_lpage_dump_cfg(KERN_DEBUG, static_size, reserved_size, dyn_size, |
1837 | unit_size, lpage_size, unit_map, nr_units); | |
8c4bfc6e | 1838 | |
a530b795 TH |
1839 | BUG_ON(chunk_size % lpage_size); |
1840 | ||
1841 | pcpul_size = static_size + reserved_size + dyn_size; | |
1842 | pcpul_lpage_size = lpage_size; | |
1843 | pcpul_nr_lpages = chunk_size / lpage_size; | |
8c4bfc6e TH |
1844 | |
1845 | /* allocate pointer array and alloc large pages */ | |
a530b795 | 1846 | map_size = pcpul_nr_lpages * sizeof(pcpul_map[0]); |
8c4bfc6e TH |
1847 | pcpul_map = alloc_bootmem(map_size); |
1848 | ||
a530b795 TH |
1849 | /* allocate all pages */ |
1850 | for (i = 0; i < pcpul_nr_lpages; i++) { | |
1851 | size_t offset = i * lpage_size; | |
1852 | int first_unit = offset / unit_size; | |
1853 | int last_unit = (offset + lpage_size - 1) / unit_size; | |
8c4bfc6e TH |
1854 | void *ptr; |
1855 | ||
a530b795 TH |
1856 | /* find out which cpu is mapped to this unit */ |
1857 | for (unit = first_unit; unit <= last_unit; unit++) | |
1858 | if (pcpul_unit_to_cpu(unit, unit_map, &cpu)) | |
1859 | goto found; | |
1860 | continue; | |
1861 | found: | |
8c4bfc6e TH |
1862 | ptr = alloc_fn(cpu, lpage_size); |
1863 | if (!ptr) { | |
1864 | pr_warning("PERCPU: failed to allocate large page " | |
1865 | "for cpu%u\n", cpu); | |
1866 | goto enomem; | |
1867 | } | |
1868 | ||
a530b795 TH |
1869 | pcpul_map[i].ptr = ptr; |
1870 | } | |
8c4bfc6e | 1871 | |
a530b795 TH |
1872 | /* return unused holes */ |
1873 | for (unit = 0; unit < nr_units; unit++) { | |
1874 | size_t start = unit * unit_size; | |
1875 | size_t end = start + unit_size; | |
1876 | size_t off, next; | |
1877 | ||
1878 | /* don't free used part of occupied unit */ | |
1879 | if (pcpul_unit_to_cpu(unit, unit_map, NULL)) | |
1880 | start += pcpul_size; | |
1881 | ||
1882 | /* unit can span more than one page, punch the holes */ | |
1883 | for (off = start; off < end; off = next) { | |
1884 | void *ptr = pcpul_map[off / lpage_size].ptr; | |
1885 | next = min(roundup(off + 1, lpage_size), end); | |
1886 | if (ptr) | |
1887 | free_fn(ptr + off % lpage_size, next - off); | |
1888 | } | |
8c4bfc6e TH |
1889 | } |
1890 | ||
a530b795 TH |
1891 | /* allocate address, map and copy */ |
1892 | vm.flags = VM_ALLOC; | |
1893 | vm.size = chunk_size; | |
1894 | vm_area_register_early(&vm, unit_size); | |
1895 | ||
1896 | for (i = 0; i < pcpul_nr_lpages; i++) { | |
1897 | if (!pcpul_map[i].ptr) | |
1898 | continue; | |
1899 | pcpul_map[i].map_addr = vm.addr + i * lpage_size; | |
1900 | map_fn(pcpul_map[i].ptr, lpage_size, pcpul_map[i].map_addr); | |
1901 | } | |
8c4bfc6e TH |
1902 | |
1903 | for_each_possible_cpu(cpu) | |
a530b795 TH |
1904 | memcpy(vm.addr + unit_map[cpu] * unit_size, __per_cpu_load, |
1905 | static_size); | |
8c4bfc6e TH |
1906 | |
1907 | /* we're ready, commit */ | |
004018e2 TH |
1908 | pr_info("PERCPU: large pages @%p s%zu r%zu d%zu u%zu\n", |
1909 | vm.addr, static_size, reserved_size, dyn_size, unit_size); | |
8c4bfc6e | 1910 | |
ce3141a2 | 1911 | ret = pcpu_setup_first_chunk(static_size, reserved_size, dyn_size, |
a530b795 TH |
1912 | unit_size, vm.addr, unit_map); |
1913 | ||
1914 | /* | |
1915 | * Sort pcpul_map array for pcpu_lpage_remapped(). Unmapped | |
1916 | * lpages are pushed to the end and trimmed. | |
1917 | */ | |
1918 | for (i = 0; i < pcpul_nr_lpages - 1; i++) | |
1919 | for (j = i + 1; j < pcpul_nr_lpages; j++) { | |
1920 | struct pcpul_ent tmp; | |
1921 | ||
1922 | if (!pcpul_map[j].ptr) | |
1923 | continue; | |
1924 | if (pcpul_map[i].ptr && | |
1925 | pcpul_map[i].ptr < pcpul_map[j].ptr) | |
1926 | continue; | |
1927 | ||
1928 | tmp = pcpul_map[i]; | |
1929 | pcpul_map[i] = pcpul_map[j]; | |
1930 | pcpul_map[j] = tmp; | |
1931 | } | |
1932 | ||
1933 | while (pcpul_nr_lpages && !pcpul_map[pcpul_nr_lpages - 1].ptr) | |
1934 | pcpul_nr_lpages--; | |
8c4bfc6e TH |
1935 | |
1936 | return ret; | |
1937 | ||
1938 | enomem: | |
a530b795 TH |
1939 | for (i = 0; i < pcpul_nr_lpages; i++) |
1940 | if (pcpul_map[i].ptr) | |
1941 | free_fn(pcpul_map[i].ptr, lpage_size); | |
8c4bfc6e TH |
1942 | free_bootmem(__pa(pcpul_map), map_size); |
1943 | return -ENOMEM; | |
1944 | } | |
1945 | ||
1946 | /** | |
1947 | * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area | |
1948 | * @kaddr: the kernel address in question | |
1949 | * | |
1950 | * Determine whether @kaddr falls in the pcpul recycled area. This is | |
1951 | * used by pageattr to detect VM aliases and break up the pcpu large | |
1952 | * page mapping such that the same physical page is not mapped under | |
1953 | * different attributes. | |
1954 | * | |
1955 | * The recycled area is always at the tail of a partially used large | |
1956 | * page. | |
1957 | * | |
1958 | * RETURNS: | |
1959 | * Address of corresponding remapped pcpu address if match is found; | |
1960 | * otherwise, NULL. | |
1961 | */ | |
1962 | void *pcpu_lpage_remapped(void *kaddr) | |
1963 | { | |
a530b795 TH |
1964 | unsigned long lpage_mask = pcpul_lpage_size - 1; |
1965 | void *lpage_addr = (void *)((unsigned long)kaddr & ~lpage_mask); | |
1966 | unsigned long offset = (unsigned long)kaddr & lpage_mask; | |
1967 | int left = 0, right = pcpul_nr_lpages - 1; | |
8c4bfc6e TH |
1968 | int pos; |
1969 | ||
1970 | /* pcpul in use at all? */ | |
1971 | if (!pcpul_map) | |
1972 | return NULL; | |
1973 | ||
1974 | /* okay, perform binary search */ | |
1975 | while (left <= right) { | |
1976 | pos = (left + right) / 2; | |
1977 | ||
1978 | if (pcpul_map[pos].ptr < lpage_addr) | |
1979 | left = pos + 1; | |
1980 | else if (pcpul_map[pos].ptr > lpage_addr) | |
1981 | right = pos - 1; | |
a530b795 TH |
1982 | else |
1983 | return pcpul_map[pos].map_addr + offset; | |
8c4bfc6e TH |
1984 | } |
1985 | ||
1986 | return NULL; | |
1987 | } | |
08fc4580 | 1988 | #endif /* CONFIG_NEED_PER_CPU_LPAGE_FIRST_CHUNK */ |
8c4bfc6e | 1989 | |
e74e3962 TH |
1990 | /* |
1991 | * Generic percpu area setup. | |
1992 | * | |
1993 | * The embedding helper is used because its behavior closely resembles | |
1994 | * the original non-dynamic generic percpu area setup. This is | |
1995 | * important because many archs have addressing restrictions and might | |
1996 | * fail if the percpu area is located far away from the previous | |
1997 | * location. As an added bonus, in non-NUMA cases, embedding is | |
1998 | * generally a good idea TLB-wise because percpu area can piggy back | |
1999 | * on the physical linear memory mapping which uses large page | |
2000 | * mappings on applicable archs. | |
2001 | */ | |
2002 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA | |
2003 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; | |
2004 | EXPORT_SYMBOL(__per_cpu_offset); | |
2005 | ||
2006 | void __init setup_per_cpu_areas(void) | |
2007 | { | |
2008 | size_t static_size = __per_cpu_end - __per_cpu_start; | |
2009 | ssize_t unit_size; | |
2010 | unsigned long delta; | |
2011 | unsigned int cpu; | |
2012 | ||
2013 | /* | |
2014 | * Always reserve area for module percpu variables. That's | |
2015 | * what the legacy allocator did. | |
2016 | */ | |
2017 | unit_size = pcpu_embed_first_chunk(static_size, PERCPU_MODULE_RESERVE, | |
788e5abc | 2018 | PERCPU_DYNAMIC_RESERVE); |
e74e3962 TH |
2019 | if (unit_size < 0) |
2020 | panic("Failed to initialized percpu areas."); | |
2021 | ||
2022 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
2023 | for_each_possible_cpu(cpu) | |
2024 | __per_cpu_offset[cpu] = delta + cpu * unit_size; | |
2025 | } | |
2026 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |