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> | |
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 */ | |
95 | struct vm_struct *vm; /* mapped vmalloc region */ | |
96 | int map_used; /* # of map entries used */ | |
97 | int map_alloc; /* # of map entries allocated */ | |
98 | int *map; /* allocation map */ | |
8d408b4b | 99 | bool immutable; /* no [de]population allowed */ |
ce3141a2 | 100 | unsigned long populated[]; /* populated bitmap */ |
fbf59bc9 TH |
101 | }; |
102 | ||
40150d37 TH |
103 | static int pcpu_unit_pages __read_mostly; |
104 | static int pcpu_unit_size __read_mostly; | |
2f39e637 | 105 | static int pcpu_nr_units __read_mostly; |
40150d37 TH |
106 | static int pcpu_chunk_size __read_mostly; |
107 | static int pcpu_nr_slots __read_mostly; | |
108 | static size_t pcpu_chunk_struct_size __read_mostly; | |
fbf59bc9 | 109 | |
2f39e637 TH |
110 | /* cpus with the lowest and highest unit numbers */ |
111 | static unsigned int pcpu_first_unit_cpu __read_mostly; | |
112 | static unsigned int pcpu_last_unit_cpu __read_mostly; | |
113 | ||
fbf59bc9 | 114 | /* the address of the first chunk which starts with the kernel static area */ |
40150d37 | 115 | void *pcpu_base_addr __read_mostly; |
fbf59bc9 TH |
116 | EXPORT_SYMBOL_GPL(pcpu_base_addr); |
117 | ||
2f39e637 TH |
118 | /* cpu -> unit map */ |
119 | const int *pcpu_unit_map __read_mostly; | |
120 | ||
ae9e6bc9 TH |
121 | /* |
122 | * The first chunk which always exists. Note that unlike other | |
123 | * chunks, this one can be allocated and mapped in several different | |
124 | * ways and thus often doesn't live in the vmalloc area. | |
125 | */ | |
126 | static struct pcpu_chunk *pcpu_first_chunk; | |
127 | ||
128 | /* | |
129 | * Optional reserved chunk. This chunk reserves part of the first | |
130 | * chunk and serves it for reserved allocations. The amount of | |
131 | * reserved offset is in pcpu_reserved_chunk_limit. When reserved | |
132 | * area doesn't exist, the following variables contain NULL and 0 | |
133 | * respectively. | |
134 | */ | |
edcb4639 | 135 | static struct pcpu_chunk *pcpu_reserved_chunk; |
edcb4639 TH |
136 | static int pcpu_reserved_chunk_limit; |
137 | ||
fbf59bc9 | 138 | /* |
ccea34b5 TH |
139 | * Synchronization rules. |
140 | * | |
141 | * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former | |
ce3141a2 TH |
142 | * protects allocation/reclaim paths, chunks, populated bitmap and |
143 | * vmalloc mapping. The latter is a spinlock and protects the index | |
144 | * data structures - chunk slots, chunks and area maps in chunks. | |
ccea34b5 TH |
145 | * |
146 | * During allocation, pcpu_alloc_mutex is kept locked all the time and | |
147 | * pcpu_lock is grabbed and released as necessary. All actual memory | |
148 | * allocations are done using GFP_KERNEL with pcpu_lock released. | |
149 | * | |
150 | * Free path accesses and alters only the index data structures, so it | |
151 | * can be safely called from atomic context. When memory needs to be | |
152 | * returned to the system, free path schedules reclaim_work which | |
153 | * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be | |
154 | * reclaimed, release both locks and frees the chunks. Note that it's | |
155 | * necessary to grab both locks to remove a chunk from circulation as | |
156 | * allocation path might be referencing the chunk with only | |
157 | * pcpu_alloc_mutex locked. | |
fbf59bc9 | 158 | */ |
ccea34b5 TH |
159 | static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */ |
160 | static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */ | |
fbf59bc9 | 161 | |
40150d37 | 162 | static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */ |
fbf59bc9 | 163 | |
a56dbddf TH |
164 | /* reclaim work to release fully free chunks, scheduled from free path */ |
165 | static void pcpu_reclaim(struct work_struct *work); | |
166 | static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim); | |
167 | ||
d9b55eeb | 168 | static int __pcpu_size_to_slot(int size) |
fbf59bc9 | 169 | { |
cae3aeb8 | 170 | int highbit = fls(size); /* size is in bytes */ |
fbf59bc9 TH |
171 | return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1); |
172 | } | |
173 | ||
d9b55eeb TH |
174 | static int pcpu_size_to_slot(int size) |
175 | { | |
176 | if (size == pcpu_unit_size) | |
177 | return pcpu_nr_slots - 1; | |
178 | return __pcpu_size_to_slot(size); | |
179 | } | |
180 | ||
fbf59bc9 TH |
181 | static int pcpu_chunk_slot(const struct pcpu_chunk *chunk) |
182 | { | |
183 | if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int)) | |
184 | return 0; | |
185 | ||
186 | return pcpu_size_to_slot(chunk->free_size); | |
187 | } | |
188 | ||
189 | static int pcpu_page_idx(unsigned int cpu, int page_idx) | |
190 | { | |
2f39e637 | 191 | return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx; |
fbf59bc9 TH |
192 | } |
193 | ||
fbf59bc9 TH |
194 | static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk, |
195 | unsigned int cpu, int page_idx) | |
196 | { | |
197 | return (unsigned long)chunk->vm->addr + | |
198 | (pcpu_page_idx(cpu, page_idx) << PAGE_SHIFT); | |
199 | } | |
200 | ||
ce3141a2 TH |
201 | static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk, |
202 | unsigned int cpu, int page_idx) | |
c8a51be4 | 203 | { |
ce3141a2 TH |
204 | /* must not be used on pre-mapped chunk */ |
205 | WARN_ON(chunk->immutable); | |
c8a51be4 | 206 | |
ce3141a2 | 207 | return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx)); |
fbf59bc9 TH |
208 | } |
209 | ||
e1b9aa3f CL |
210 | /* set the pointer to a chunk in a page struct */ |
211 | static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu) | |
212 | { | |
213 | page->index = (unsigned long)pcpu; | |
214 | } | |
215 | ||
216 | /* obtain pointer to a chunk from a page struct */ | |
217 | static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page) | |
218 | { | |
219 | return (struct pcpu_chunk *)page->index; | |
220 | } | |
221 | ||
ce3141a2 TH |
222 | static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end) |
223 | { | |
224 | *rs = find_next_zero_bit(chunk->populated, end, *rs); | |
225 | *re = find_next_bit(chunk->populated, end, *rs + 1); | |
226 | } | |
227 | ||
228 | static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end) | |
229 | { | |
230 | *rs = find_next_bit(chunk->populated, end, *rs); | |
231 | *re = find_next_zero_bit(chunk->populated, end, *rs + 1); | |
232 | } | |
233 | ||
234 | /* | |
235 | * (Un)populated page region iterators. Iterate over (un)populated | |
236 | * page regions betwen @start and @end in @chunk. @rs and @re should | |
237 | * be integer variables and will be set to start and end page index of | |
238 | * the current region. | |
239 | */ | |
240 | #define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \ | |
241 | for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \ | |
242 | (rs) < (re); \ | |
243 | (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end))) | |
244 | ||
245 | #define pcpu_for_each_pop_region(chunk, rs, re, start, end) \ | |
246 | for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \ | |
247 | (rs) < (re); \ | |
248 | (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end))) | |
249 | ||
fbf59bc9 | 250 | /** |
1880d93b TH |
251 | * pcpu_mem_alloc - allocate memory |
252 | * @size: bytes to allocate | |
fbf59bc9 | 253 | * |
1880d93b TH |
254 | * Allocate @size bytes. If @size is smaller than PAGE_SIZE, |
255 | * kzalloc() is used; otherwise, vmalloc() is used. The returned | |
256 | * memory is always zeroed. | |
fbf59bc9 | 257 | * |
ccea34b5 TH |
258 | * CONTEXT: |
259 | * Does GFP_KERNEL allocation. | |
260 | * | |
fbf59bc9 | 261 | * RETURNS: |
1880d93b | 262 | * Pointer to the allocated area on success, NULL on failure. |
fbf59bc9 | 263 | */ |
1880d93b | 264 | static void *pcpu_mem_alloc(size_t size) |
fbf59bc9 | 265 | { |
1880d93b TH |
266 | if (size <= PAGE_SIZE) |
267 | return kzalloc(size, GFP_KERNEL); | |
268 | else { | |
269 | void *ptr = vmalloc(size); | |
270 | if (ptr) | |
271 | memset(ptr, 0, size); | |
272 | return ptr; | |
273 | } | |
274 | } | |
fbf59bc9 | 275 | |
1880d93b TH |
276 | /** |
277 | * pcpu_mem_free - free memory | |
278 | * @ptr: memory to free | |
279 | * @size: size of the area | |
280 | * | |
281 | * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc(). | |
282 | */ | |
283 | static void pcpu_mem_free(void *ptr, size_t size) | |
284 | { | |
fbf59bc9 | 285 | if (size <= PAGE_SIZE) |
1880d93b | 286 | kfree(ptr); |
fbf59bc9 | 287 | else |
1880d93b | 288 | vfree(ptr); |
fbf59bc9 TH |
289 | } |
290 | ||
291 | /** | |
292 | * pcpu_chunk_relocate - put chunk in the appropriate chunk slot | |
293 | * @chunk: chunk of interest | |
294 | * @oslot: the previous slot it was on | |
295 | * | |
296 | * This function is called after an allocation or free changed @chunk. | |
297 | * New slot according to the changed state is determined and @chunk is | |
edcb4639 TH |
298 | * moved to the slot. Note that the reserved chunk is never put on |
299 | * chunk slots. | |
ccea34b5 TH |
300 | * |
301 | * CONTEXT: | |
302 | * pcpu_lock. | |
fbf59bc9 TH |
303 | */ |
304 | static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot) | |
305 | { | |
306 | int nslot = pcpu_chunk_slot(chunk); | |
307 | ||
edcb4639 | 308 | if (chunk != pcpu_reserved_chunk && oslot != nslot) { |
fbf59bc9 TH |
309 | if (oslot < nslot) |
310 | list_move(&chunk->list, &pcpu_slot[nslot]); | |
311 | else | |
312 | list_move_tail(&chunk->list, &pcpu_slot[nslot]); | |
313 | } | |
314 | } | |
315 | ||
fbf59bc9 | 316 | /** |
e1b9aa3f CL |
317 | * pcpu_chunk_addr_search - determine chunk containing specified address |
318 | * @addr: address for which the chunk needs to be determined. | |
ccea34b5 | 319 | * |
fbf59bc9 TH |
320 | * RETURNS: |
321 | * The address of the found chunk. | |
322 | */ | |
323 | static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr) | |
324 | { | |
ae9e6bc9 | 325 | void *first_start = pcpu_first_chunk->vm->addr; |
fbf59bc9 | 326 | |
ae9e6bc9 | 327 | /* is it in the first chunk? */ |
79ba6ac8 | 328 | if (addr >= first_start && addr < first_start + pcpu_unit_size) { |
ae9e6bc9 TH |
329 | /* is it in the reserved area? */ |
330 | if (addr < first_start + pcpu_reserved_chunk_limit) | |
edcb4639 | 331 | return pcpu_reserved_chunk; |
ae9e6bc9 | 332 | return pcpu_first_chunk; |
edcb4639 TH |
333 | } |
334 | ||
2f39e637 TH |
335 | /* |
336 | * The address is relative to unit0 which might be unused and | |
337 | * thus unmapped. Offset the address to the unit space of the | |
338 | * current processor before looking it up in the vmalloc | |
339 | * space. Note that any possible cpu id can be used here, so | |
340 | * there's no need to worry about preemption or cpu hotplug. | |
341 | */ | |
342 | addr += pcpu_unit_map[smp_processor_id()] * pcpu_unit_size; | |
e1b9aa3f | 343 | return pcpu_get_page_chunk(vmalloc_to_page(addr)); |
fbf59bc9 TH |
344 | } |
345 | ||
9f7dcf22 TH |
346 | /** |
347 | * pcpu_extend_area_map - extend area map for allocation | |
348 | * @chunk: target chunk | |
349 | * | |
350 | * Extend area map of @chunk so that it can accomodate an allocation. | |
351 | * A single allocation can split an area into three areas, so this | |
352 | * function makes sure that @chunk->map has at least two extra slots. | |
353 | * | |
ccea34b5 TH |
354 | * CONTEXT: |
355 | * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired | |
356 | * if area map is extended. | |
357 | * | |
9f7dcf22 TH |
358 | * RETURNS: |
359 | * 0 if noop, 1 if successfully extended, -errno on failure. | |
360 | */ | |
361 | static int pcpu_extend_area_map(struct pcpu_chunk *chunk) | |
362 | { | |
363 | int new_alloc; | |
364 | int *new; | |
365 | size_t size; | |
366 | ||
367 | /* has enough? */ | |
368 | if (chunk->map_alloc >= chunk->map_used + 2) | |
369 | return 0; | |
370 | ||
ccea34b5 TH |
371 | spin_unlock_irq(&pcpu_lock); |
372 | ||
9f7dcf22 TH |
373 | new_alloc = PCPU_DFL_MAP_ALLOC; |
374 | while (new_alloc < chunk->map_used + 2) | |
375 | new_alloc *= 2; | |
376 | ||
377 | new = pcpu_mem_alloc(new_alloc * sizeof(new[0])); | |
ccea34b5 TH |
378 | if (!new) { |
379 | spin_lock_irq(&pcpu_lock); | |
9f7dcf22 | 380 | return -ENOMEM; |
ccea34b5 TH |
381 | } |
382 | ||
383 | /* | |
384 | * Acquire pcpu_lock and switch to new area map. Only free | |
385 | * could have happened inbetween, so map_used couldn't have | |
386 | * grown. | |
387 | */ | |
388 | spin_lock_irq(&pcpu_lock); | |
389 | BUG_ON(new_alloc < chunk->map_used + 2); | |
9f7dcf22 TH |
390 | |
391 | size = chunk->map_alloc * sizeof(chunk->map[0]); | |
392 | memcpy(new, chunk->map, size); | |
393 | ||
394 | /* | |
395 | * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is | |
396 | * one of the first chunks and still using static map. | |
397 | */ | |
398 | if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC) | |
399 | pcpu_mem_free(chunk->map, size); | |
400 | ||
401 | chunk->map_alloc = new_alloc; | |
402 | chunk->map = new; | |
403 | return 0; | |
404 | } | |
405 | ||
fbf59bc9 TH |
406 | /** |
407 | * pcpu_split_block - split a map block | |
408 | * @chunk: chunk of interest | |
409 | * @i: index of map block to split | |
cae3aeb8 TH |
410 | * @head: head size in bytes (can be 0) |
411 | * @tail: tail size in bytes (can be 0) | |
fbf59bc9 TH |
412 | * |
413 | * Split the @i'th map block into two or three blocks. If @head is | |
414 | * non-zero, @head bytes block is inserted before block @i moving it | |
415 | * to @i+1 and reducing its size by @head bytes. | |
416 | * | |
417 | * If @tail is non-zero, the target block, which can be @i or @i+1 | |
418 | * depending on @head, is reduced by @tail bytes and @tail byte block | |
419 | * is inserted after the target block. | |
420 | * | |
9f7dcf22 | 421 | * @chunk->map must have enough free slots to accomodate the split. |
ccea34b5 TH |
422 | * |
423 | * CONTEXT: | |
424 | * pcpu_lock. | |
fbf59bc9 | 425 | */ |
9f7dcf22 TH |
426 | static void pcpu_split_block(struct pcpu_chunk *chunk, int i, |
427 | int head, int tail) | |
fbf59bc9 TH |
428 | { |
429 | int nr_extra = !!head + !!tail; | |
1880d93b | 430 | |
9f7dcf22 | 431 | BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra); |
fbf59bc9 | 432 | |
9f7dcf22 | 433 | /* insert new subblocks */ |
fbf59bc9 TH |
434 | memmove(&chunk->map[i + nr_extra], &chunk->map[i], |
435 | sizeof(chunk->map[0]) * (chunk->map_used - i)); | |
436 | chunk->map_used += nr_extra; | |
437 | ||
438 | if (head) { | |
439 | chunk->map[i + 1] = chunk->map[i] - head; | |
440 | chunk->map[i++] = head; | |
441 | } | |
442 | if (tail) { | |
443 | chunk->map[i++] -= tail; | |
444 | chunk->map[i] = tail; | |
445 | } | |
fbf59bc9 TH |
446 | } |
447 | ||
448 | /** | |
449 | * pcpu_alloc_area - allocate area from a pcpu_chunk | |
450 | * @chunk: chunk of interest | |
cae3aeb8 | 451 | * @size: wanted size in bytes |
fbf59bc9 TH |
452 | * @align: wanted align |
453 | * | |
454 | * Try to allocate @size bytes area aligned at @align from @chunk. | |
455 | * Note that this function only allocates the offset. It doesn't | |
456 | * populate or map the area. | |
457 | * | |
9f7dcf22 TH |
458 | * @chunk->map must have at least two free slots. |
459 | * | |
ccea34b5 TH |
460 | * CONTEXT: |
461 | * pcpu_lock. | |
462 | * | |
fbf59bc9 | 463 | * RETURNS: |
9f7dcf22 TH |
464 | * Allocated offset in @chunk on success, -1 if no matching area is |
465 | * found. | |
fbf59bc9 TH |
466 | */ |
467 | static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align) | |
468 | { | |
469 | int oslot = pcpu_chunk_slot(chunk); | |
470 | int max_contig = 0; | |
471 | int i, off; | |
472 | ||
fbf59bc9 TH |
473 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) { |
474 | bool is_last = i + 1 == chunk->map_used; | |
475 | int head, tail; | |
476 | ||
477 | /* extra for alignment requirement */ | |
478 | head = ALIGN(off, align) - off; | |
479 | BUG_ON(i == 0 && head != 0); | |
480 | ||
481 | if (chunk->map[i] < 0) | |
482 | continue; | |
483 | if (chunk->map[i] < head + size) { | |
484 | max_contig = max(chunk->map[i], max_contig); | |
485 | continue; | |
486 | } | |
487 | ||
488 | /* | |
489 | * If head is small or the previous block is free, | |
490 | * merge'em. Note that 'small' is defined as smaller | |
491 | * than sizeof(int), which is very small but isn't too | |
492 | * uncommon for percpu allocations. | |
493 | */ | |
494 | if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) { | |
495 | if (chunk->map[i - 1] > 0) | |
496 | chunk->map[i - 1] += head; | |
497 | else { | |
498 | chunk->map[i - 1] -= head; | |
499 | chunk->free_size -= head; | |
500 | } | |
501 | chunk->map[i] -= head; | |
502 | off += head; | |
503 | head = 0; | |
504 | } | |
505 | ||
506 | /* if tail is small, just keep it around */ | |
507 | tail = chunk->map[i] - head - size; | |
508 | if (tail < sizeof(int)) | |
509 | tail = 0; | |
510 | ||
511 | /* split if warranted */ | |
512 | if (head || tail) { | |
9f7dcf22 | 513 | pcpu_split_block(chunk, i, head, tail); |
fbf59bc9 TH |
514 | if (head) { |
515 | i++; | |
516 | off += head; | |
517 | max_contig = max(chunk->map[i - 1], max_contig); | |
518 | } | |
519 | if (tail) | |
520 | max_contig = max(chunk->map[i + 1], max_contig); | |
521 | } | |
522 | ||
523 | /* update hint and mark allocated */ | |
524 | if (is_last) | |
525 | chunk->contig_hint = max_contig; /* fully scanned */ | |
526 | else | |
527 | chunk->contig_hint = max(chunk->contig_hint, | |
528 | max_contig); | |
529 | ||
530 | chunk->free_size -= chunk->map[i]; | |
531 | chunk->map[i] = -chunk->map[i]; | |
532 | ||
533 | pcpu_chunk_relocate(chunk, oslot); | |
534 | return off; | |
535 | } | |
536 | ||
537 | chunk->contig_hint = max_contig; /* fully scanned */ | |
538 | pcpu_chunk_relocate(chunk, oslot); | |
539 | ||
9f7dcf22 TH |
540 | /* tell the upper layer that this chunk has no matching area */ |
541 | return -1; | |
fbf59bc9 TH |
542 | } |
543 | ||
544 | /** | |
545 | * pcpu_free_area - free area to a pcpu_chunk | |
546 | * @chunk: chunk of interest | |
547 | * @freeme: offset of area to free | |
548 | * | |
549 | * Free area starting from @freeme to @chunk. Note that this function | |
550 | * only modifies the allocation map. It doesn't depopulate or unmap | |
551 | * the area. | |
ccea34b5 TH |
552 | * |
553 | * CONTEXT: | |
554 | * pcpu_lock. | |
fbf59bc9 TH |
555 | */ |
556 | static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme) | |
557 | { | |
558 | int oslot = pcpu_chunk_slot(chunk); | |
559 | int i, off; | |
560 | ||
561 | for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) | |
562 | if (off == freeme) | |
563 | break; | |
564 | BUG_ON(off != freeme); | |
565 | BUG_ON(chunk->map[i] > 0); | |
566 | ||
567 | chunk->map[i] = -chunk->map[i]; | |
568 | chunk->free_size += chunk->map[i]; | |
569 | ||
570 | /* merge with previous? */ | |
571 | if (i > 0 && chunk->map[i - 1] >= 0) { | |
572 | chunk->map[i - 1] += chunk->map[i]; | |
573 | chunk->map_used--; | |
574 | memmove(&chunk->map[i], &chunk->map[i + 1], | |
575 | (chunk->map_used - i) * sizeof(chunk->map[0])); | |
576 | i--; | |
577 | } | |
578 | /* merge with next? */ | |
579 | if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) { | |
580 | chunk->map[i] += chunk->map[i + 1]; | |
581 | chunk->map_used--; | |
582 | memmove(&chunk->map[i + 1], &chunk->map[i + 2], | |
583 | (chunk->map_used - (i + 1)) * sizeof(chunk->map[0])); | |
584 | } | |
585 | ||
586 | chunk->contig_hint = max(chunk->map[i], chunk->contig_hint); | |
587 | pcpu_chunk_relocate(chunk, oslot); | |
588 | } | |
589 | ||
590 | /** | |
ce3141a2 TH |
591 | * pcpu_get_pages_and_bitmap - get temp pages array and bitmap |
592 | * @chunk: chunk of interest | |
593 | * @bitmapp: output parameter for bitmap | |
594 | * @may_alloc: may allocate the array | |
595 | * | |
596 | * Returns pointer to array of pointers to struct page and bitmap, | |
597 | * both of which can be indexed with pcpu_page_idx(). The returned | |
598 | * array is cleared to zero and *@bitmapp is copied from | |
599 | * @chunk->populated. Note that there is only one array and bitmap | |
600 | * and access exclusion is the caller's responsibility. | |
601 | * | |
602 | * CONTEXT: | |
603 | * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc. | |
604 | * Otherwise, don't care. | |
605 | * | |
606 | * RETURNS: | |
607 | * Pointer to temp pages array on success, NULL on failure. | |
608 | */ | |
609 | static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk, | |
610 | unsigned long **bitmapp, | |
611 | bool may_alloc) | |
612 | { | |
613 | static struct page **pages; | |
614 | static unsigned long *bitmap; | |
2f39e637 | 615 | size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]); |
ce3141a2 TH |
616 | size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) * |
617 | sizeof(unsigned long); | |
618 | ||
619 | if (!pages || !bitmap) { | |
620 | if (may_alloc && !pages) | |
621 | pages = pcpu_mem_alloc(pages_size); | |
622 | if (may_alloc && !bitmap) | |
623 | bitmap = pcpu_mem_alloc(bitmap_size); | |
624 | if (!pages || !bitmap) | |
625 | return NULL; | |
626 | } | |
627 | ||
628 | memset(pages, 0, pages_size); | |
629 | bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages); | |
630 | ||
631 | *bitmapp = bitmap; | |
632 | return pages; | |
633 | } | |
634 | ||
635 | /** | |
636 | * pcpu_free_pages - free pages which were allocated for @chunk | |
637 | * @chunk: chunk pages were allocated for | |
638 | * @pages: array of pages to be freed, indexed by pcpu_page_idx() | |
639 | * @populated: populated bitmap | |
640 | * @page_start: page index of the first page to be freed | |
641 | * @page_end: page index of the last page to be freed + 1 | |
642 | * | |
643 | * Free pages [@page_start and @page_end) in @pages for all units. | |
644 | * The pages were allocated for @chunk. | |
645 | */ | |
646 | static void pcpu_free_pages(struct pcpu_chunk *chunk, | |
647 | struct page **pages, unsigned long *populated, | |
648 | int page_start, int page_end) | |
649 | { | |
650 | unsigned int cpu; | |
651 | int i; | |
652 | ||
653 | for_each_possible_cpu(cpu) { | |
654 | for (i = page_start; i < page_end; i++) { | |
655 | struct page *page = pages[pcpu_page_idx(cpu, i)]; | |
656 | ||
657 | if (page) | |
658 | __free_page(page); | |
659 | } | |
660 | } | |
661 | } | |
662 | ||
663 | /** | |
664 | * pcpu_alloc_pages - allocates pages for @chunk | |
665 | * @chunk: target chunk | |
666 | * @pages: array to put the allocated pages into, indexed by pcpu_page_idx() | |
667 | * @populated: populated bitmap | |
668 | * @page_start: page index of the first page to be allocated | |
669 | * @page_end: page index of the last page to be allocated + 1 | |
670 | * | |
671 | * Allocate pages [@page_start,@page_end) into @pages for all units. | |
672 | * The allocation is for @chunk. Percpu core doesn't care about the | |
673 | * content of @pages and will pass it verbatim to pcpu_map_pages(). | |
674 | */ | |
675 | static int pcpu_alloc_pages(struct pcpu_chunk *chunk, | |
676 | struct page **pages, unsigned long *populated, | |
677 | int page_start, int page_end) | |
678 | { | |
679 | const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD; | |
680 | unsigned int cpu; | |
681 | int i; | |
682 | ||
683 | for_each_possible_cpu(cpu) { | |
684 | for (i = page_start; i < page_end; i++) { | |
685 | struct page **pagep = &pages[pcpu_page_idx(cpu, i)]; | |
686 | ||
687 | *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0); | |
688 | if (!*pagep) { | |
689 | pcpu_free_pages(chunk, pages, populated, | |
690 | page_start, page_end); | |
691 | return -ENOMEM; | |
692 | } | |
693 | } | |
694 | } | |
695 | return 0; | |
696 | } | |
697 | ||
698 | /** | |
699 | * pcpu_pre_unmap_flush - flush cache prior to unmapping | |
700 | * @chunk: chunk the regions to be flushed belongs to | |
701 | * @page_start: page index of the first page to be flushed | |
702 | * @page_end: page index of the last page to be flushed + 1 | |
703 | * | |
704 | * Pages in [@page_start,@page_end) of @chunk are about to be | |
705 | * unmapped. Flush cache. As each flushing trial can be very | |
706 | * expensive, issue flush on the whole region at once rather than | |
707 | * doing it for each cpu. This could be an overkill but is more | |
708 | * scalable. | |
709 | */ | |
710 | static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk, | |
711 | int page_start, int page_end) | |
712 | { | |
2f39e637 TH |
713 | flush_cache_vunmap( |
714 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
715 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
ce3141a2 TH |
716 | } |
717 | ||
718 | static void __pcpu_unmap_pages(unsigned long addr, int nr_pages) | |
719 | { | |
720 | unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT); | |
721 | } | |
722 | ||
723 | /** | |
724 | * pcpu_unmap_pages - unmap pages out of a pcpu_chunk | |
fbf59bc9 | 725 | * @chunk: chunk of interest |
ce3141a2 TH |
726 | * @pages: pages array which can be used to pass information to free |
727 | * @populated: populated bitmap | |
fbf59bc9 TH |
728 | * @page_start: page index of the first page to unmap |
729 | * @page_end: page index of the last page to unmap + 1 | |
fbf59bc9 TH |
730 | * |
731 | * For each cpu, unmap pages [@page_start,@page_end) out of @chunk. | |
ce3141a2 TH |
732 | * Corresponding elements in @pages were cleared by the caller and can |
733 | * be used to carry information to pcpu_free_pages() which will be | |
734 | * called after all unmaps are finished. The caller should call | |
735 | * proper pre/post flush functions. | |
fbf59bc9 | 736 | */ |
ce3141a2 TH |
737 | static void pcpu_unmap_pages(struct pcpu_chunk *chunk, |
738 | struct page **pages, unsigned long *populated, | |
739 | int page_start, int page_end) | |
fbf59bc9 | 740 | { |
fbf59bc9 | 741 | unsigned int cpu; |
ce3141a2 | 742 | int i; |
fbf59bc9 | 743 | |
ce3141a2 TH |
744 | for_each_possible_cpu(cpu) { |
745 | for (i = page_start; i < page_end; i++) { | |
746 | struct page *page; | |
8d408b4b | 747 | |
ce3141a2 TH |
748 | page = pcpu_chunk_page(chunk, cpu, i); |
749 | WARN_ON(!page); | |
750 | pages[pcpu_page_idx(cpu, i)] = page; | |
751 | } | |
752 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start), | |
753 | page_end - page_start); | |
754 | } | |
fbf59bc9 | 755 | |
ce3141a2 TH |
756 | for (i = page_start; i < page_end; i++) |
757 | __clear_bit(i, populated); | |
758 | } | |
759 | ||
760 | /** | |
761 | * pcpu_post_unmap_tlb_flush - flush TLB after unmapping | |
762 | * @chunk: pcpu_chunk the regions to be flushed belong to | |
763 | * @page_start: page index of the first page to be flushed | |
764 | * @page_end: page index of the last page to be flushed + 1 | |
765 | * | |
766 | * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush | |
767 | * TLB for the regions. This can be skipped if the area is to be | |
768 | * returned to vmalloc as vmalloc will handle TLB flushing lazily. | |
769 | * | |
770 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once | |
771 | * for the whole region. | |
772 | */ | |
773 | static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk, | |
774 | int page_start, int page_end) | |
775 | { | |
2f39e637 TH |
776 | flush_tlb_kernel_range( |
777 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
778 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
fbf59bc9 TH |
779 | } |
780 | ||
c8a51be4 TH |
781 | static int __pcpu_map_pages(unsigned long addr, struct page **pages, |
782 | int nr_pages) | |
783 | { | |
784 | return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT, | |
785 | PAGE_KERNEL, pages); | |
786 | } | |
787 | ||
788 | /** | |
ce3141a2 | 789 | * pcpu_map_pages - map pages into a pcpu_chunk |
c8a51be4 | 790 | * @chunk: chunk of interest |
ce3141a2 TH |
791 | * @pages: pages array containing pages to be mapped |
792 | * @populated: populated bitmap | |
c8a51be4 TH |
793 | * @page_start: page index of the first page to map |
794 | * @page_end: page index of the last page to map + 1 | |
795 | * | |
ce3141a2 TH |
796 | * For each cpu, map pages [@page_start,@page_end) into @chunk. The |
797 | * caller is responsible for calling pcpu_post_map_flush() after all | |
798 | * mappings are complete. | |
799 | * | |
800 | * This function is responsible for setting corresponding bits in | |
801 | * @chunk->populated bitmap and whatever is necessary for reverse | |
802 | * lookup (addr -> chunk). | |
c8a51be4 | 803 | */ |
ce3141a2 TH |
804 | static int pcpu_map_pages(struct pcpu_chunk *chunk, |
805 | struct page **pages, unsigned long *populated, | |
806 | int page_start, int page_end) | |
c8a51be4 | 807 | { |
ce3141a2 TH |
808 | unsigned int cpu, tcpu; |
809 | int i, err; | |
c8a51be4 TH |
810 | |
811 | for_each_possible_cpu(cpu) { | |
812 | err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start), | |
ce3141a2 | 813 | &pages[pcpu_page_idx(cpu, page_start)], |
c8a51be4 TH |
814 | page_end - page_start); |
815 | if (err < 0) | |
ce3141a2 | 816 | goto err; |
c8a51be4 TH |
817 | } |
818 | ||
ce3141a2 TH |
819 | /* mapping successful, link chunk and mark populated */ |
820 | for (i = page_start; i < page_end; i++) { | |
821 | for_each_possible_cpu(cpu) | |
822 | pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)], | |
823 | chunk); | |
824 | __set_bit(i, populated); | |
825 | } | |
826 | ||
827 | return 0; | |
828 | ||
829 | err: | |
830 | for_each_possible_cpu(tcpu) { | |
831 | if (tcpu == cpu) | |
832 | break; | |
833 | __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start), | |
834 | page_end - page_start); | |
835 | } | |
836 | return err; | |
837 | } | |
838 | ||
839 | /** | |
840 | * pcpu_post_map_flush - flush cache after mapping | |
841 | * @chunk: pcpu_chunk the regions to be flushed belong to | |
842 | * @page_start: page index of the first page to be flushed | |
843 | * @page_end: page index of the last page to be flushed + 1 | |
844 | * | |
845 | * Pages [@page_start,@page_end) of @chunk have been mapped. Flush | |
846 | * cache. | |
847 | * | |
848 | * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once | |
849 | * for the whole region. | |
850 | */ | |
851 | static void pcpu_post_map_flush(struct pcpu_chunk *chunk, | |
852 | int page_start, int page_end) | |
853 | { | |
2f39e637 TH |
854 | flush_cache_vmap( |
855 | pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start), | |
856 | pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end)); | |
c8a51be4 TH |
857 | } |
858 | ||
fbf59bc9 TH |
859 | /** |
860 | * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk | |
861 | * @chunk: chunk to depopulate | |
862 | * @off: offset to the area to depopulate | |
cae3aeb8 | 863 | * @size: size of the area to depopulate in bytes |
fbf59bc9 TH |
864 | * @flush: whether to flush cache and tlb or not |
865 | * | |
866 | * For each cpu, depopulate and unmap pages [@page_start,@page_end) | |
867 | * from @chunk. If @flush is true, vcache is flushed before unmapping | |
868 | * and tlb after. | |
ccea34b5 TH |
869 | * |
870 | * CONTEXT: | |
871 | * pcpu_alloc_mutex. | |
fbf59bc9 | 872 | */ |
ce3141a2 | 873 | static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size) |
fbf59bc9 TH |
874 | { |
875 | int page_start = PFN_DOWN(off); | |
876 | int page_end = PFN_UP(off + size); | |
ce3141a2 TH |
877 | struct page **pages; |
878 | unsigned long *populated; | |
879 | int rs, re; | |
880 | ||
881 | /* quick path, check whether it's empty already */ | |
882 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { | |
883 | if (rs == page_start && re == page_end) | |
884 | return; | |
885 | break; | |
886 | } | |
fbf59bc9 | 887 | |
ce3141a2 TH |
888 | /* immutable chunks can't be depopulated */ |
889 | WARN_ON(chunk->immutable); | |
fbf59bc9 | 890 | |
ce3141a2 TH |
891 | /* |
892 | * If control reaches here, there must have been at least one | |
893 | * successful population attempt so the temp pages array must | |
894 | * be available now. | |
895 | */ | |
896 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, false); | |
897 | BUG_ON(!pages); | |
fbf59bc9 | 898 | |
ce3141a2 TH |
899 | /* unmap and free */ |
900 | pcpu_pre_unmap_flush(chunk, page_start, page_end); | |
fbf59bc9 | 901 | |
ce3141a2 TH |
902 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) |
903 | pcpu_unmap_pages(chunk, pages, populated, rs, re); | |
fbf59bc9 | 904 | |
ce3141a2 TH |
905 | /* no need to flush tlb, vmalloc will handle it lazily */ |
906 | ||
907 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) | |
908 | pcpu_free_pages(chunk, pages, populated, rs, re); | |
fbf59bc9 | 909 | |
ce3141a2 TH |
910 | /* commit new bitmap */ |
911 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); | |
fbf59bc9 TH |
912 | } |
913 | ||
fbf59bc9 TH |
914 | /** |
915 | * pcpu_populate_chunk - populate and map an area of a pcpu_chunk | |
916 | * @chunk: chunk of interest | |
917 | * @off: offset to the area to populate | |
cae3aeb8 | 918 | * @size: size of the area to populate in bytes |
fbf59bc9 TH |
919 | * |
920 | * For each cpu, populate and map pages [@page_start,@page_end) into | |
921 | * @chunk. The area is cleared on return. | |
ccea34b5 TH |
922 | * |
923 | * CONTEXT: | |
924 | * pcpu_alloc_mutex, does GFP_KERNEL allocation. | |
fbf59bc9 TH |
925 | */ |
926 | static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size) | |
927 | { | |
fbf59bc9 TH |
928 | int page_start = PFN_DOWN(off); |
929 | int page_end = PFN_UP(off + size); | |
ce3141a2 TH |
930 | int free_end = page_start, unmap_end = page_start; |
931 | struct page **pages; | |
932 | unsigned long *populated; | |
fbf59bc9 | 933 | unsigned int cpu; |
ce3141a2 | 934 | int rs, re, rc; |
fbf59bc9 | 935 | |
ce3141a2 TH |
936 | /* quick path, check whether all pages are already there */ |
937 | pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) { | |
938 | if (rs == page_start && re == page_end) | |
939 | goto clear; | |
940 | break; | |
941 | } | |
fbf59bc9 | 942 | |
ce3141a2 TH |
943 | /* need to allocate and map pages, this chunk can't be immutable */ |
944 | WARN_ON(chunk->immutable); | |
fbf59bc9 | 945 | |
ce3141a2 TH |
946 | pages = pcpu_get_pages_and_bitmap(chunk, &populated, true); |
947 | if (!pages) | |
948 | return -ENOMEM; | |
fbf59bc9 | 949 | |
ce3141a2 TH |
950 | /* alloc and map */ |
951 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { | |
952 | rc = pcpu_alloc_pages(chunk, pages, populated, rs, re); | |
953 | if (rc) | |
954 | goto err_free; | |
955 | free_end = re; | |
fbf59bc9 TH |
956 | } |
957 | ||
ce3141a2 TH |
958 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) { |
959 | rc = pcpu_map_pages(chunk, pages, populated, rs, re); | |
960 | if (rc) | |
961 | goto err_unmap; | |
962 | unmap_end = re; | |
963 | } | |
964 | pcpu_post_map_flush(chunk, page_start, page_end); | |
fbf59bc9 | 965 | |
ce3141a2 TH |
966 | /* commit new bitmap */ |
967 | bitmap_copy(chunk->populated, populated, pcpu_unit_pages); | |
968 | clear: | |
fbf59bc9 | 969 | for_each_possible_cpu(cpu) |
2f39e637 | 970 | memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size); |
fbf59bc9 | 971 | return 0; |
ce3141a2 TH |
972 | |
973 | err_unmap: | |
974 | pcpu_pre_unmap_flush(chunk, page_start, unmap_end); | |
975 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end) | |
976 | pcpu_unmap_pages(chunk, pages, populated, rs, re); | |
977 | pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end); | |
978 | err_free: | |
979 | pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end) | |
980 | pcpu_free_pages(chunk, pages, populated, rs, re); | |
981 | return rc; | |
fbf59bc9 TH |
982 | } |
983 | ||
984 | static void free_pcpu_chunk(struct pcpu_chunk *chunk) | |
985 | { | |
986 | if (!chunk) | |
987 | return; | |
988 | if (chunk->vm) | |
989 | free_vm_area(chunk->vm); | |
1880d93b | 990 | pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0])); |
fbf59bc9 TH |
991 | kfree(chunk); |
992 | } | |
993 | ||
994 | static struct pcpu_chunk *alloc_pcpu_chunk(void) | |
995 | { | |
996 | struct pcpu_chunk *chunk; | |
997 | ||
998 | chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL); | |
999 | if (!chunk) | |
1000 | return NULL; | |
1001 | ||
1880d93b | 1002 | chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0])); |
fbf59bc9 TH |
1003 | chunk->map_alloc = PCPU_DFL_MAP_ALLOC; |
1004 | chunk->map[chunk->map_used++] = pcpu_unit_size; | |
1005 | ||
1006 | chunk->vm = get_vm_area(pcpu_chunk_size, GFP_KERNEL); | |
1007 | if (!chunk->vm) { | |
1008 | free_pcpu_chunk(chunk); | |
1009 | return NULL; | |
1010 | } | |
1011 | ||
1012 | INIT_LIST_HEAD(&chunk->list); | |
1013 | chunk->free_size = pcpu_unit_size; | |
1014 | chunk->contig_hint = pcpu_unit_size; | |
1015 | ||
1016 | return chunk; | |
1017 | } | |
1018 | ||
1019 | /** | |
edcb4639 | 1020 | * pcpu_alloc - the percpu allocator |
cae3aeb8 | 1021 | * @size: size of area to allocate in bytes |
fbf59bc9 | 1022 | * @align: alignment of area (max PAGE_SIZE) |
edcb4639 | 1023 | * @reserved: allocate from the reserved chunk if available |
fbf59bc9 | 1024 | * |
ccea34b5 TH |
1025 | * Allocate percpu area of @size bytes aligned at @align. |
1026 | * | |
1027 | * CONTEXT: | |
1028 | * Does GFP_KERNEL allocation. | |
fbf59bc9 TH |
1029 | * |
1030 | * RETURNS: | |
1031 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1032 | */ | |
edcb4639 | 1033 | static void *pcpu_alloc(size_t size, size_t align, bool reserved) |
fbf59bc9 | 1034 | { |
fbf59bc9 TH |
1035 | struct pcpu_chunk *chunk; |
1036 | int slot, off; | |
1037 | ||
8d408b4b | 1038 | if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) { |
fbf59bc9 TH |
1039 | WARN(true, "illegal size (%zu) or align (%zu) for " |
1040 | "percpu allocation\n", size, align); | |
1041 | return NULL; | |
1042 | } | |
1043 | ||
ccea34b5 TH |
1044 | mutex_lock(&pcpu_alloc_mutex); |
1045 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 1046 | |
edcb4639 TH |
1047 | /* serve reserved allocations from the reserved chunk if available */ |
1048 | if (reserved && pcpu_reserved_chunk) { | |
1049 | chunk = pcpu_reserved_chunk; | |
9f7dcf22 TH |
1050 | if (size > chunk->contig_hint || |
1051 | pcpu_extend_area_map(chunk) < 0) | |
ccea34b5 | 1052 | goto fail_unlock; |
edcb4639 TH |
1053 | off = pcpu_alloc_area(chunk, size, align); |
1054 | if (off >= 0) | |
1055 | goto area_found; | |
ccea34b5 | 1056 | goto fail_unlock; |
edcb4639 TH |
1057 | } |
1058 | ||
ccea34b5 | 1059 | restart: |
edcb4639 | 1060 | /* search through normal chunks */ |
fbf59bc9 TH |
1061 | for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) { |
1062 | list_for_each_entry(chunk, &pcpu_slot[slot], list) { | |
1063 | if (size > chunk->contig_hint) | |
1064 | continue; | |
ccea34b5 TH |
1065 | |
1066 | switch (pcpu_extend_area_map(chunk)) { | |
1067 | case 0: | |
1068 | break; | |
1069 | case 1: | |
1070 | goto restart; /* pcpu_lock dropped, restart */ | |
1071 | default: | |
1072 | goto fail_unlock; | |
1073 | } | |
1074 | ||
fbf59bc9 TH |
1075 | off = pcpu_alloc_area(chunk, size, align); |
1076 | if (off >= 0) | |
1077 | goto area_found; | |
fbf59bc9 TH |
1078 | } |
1079 | } | |
1080 | ||
1081 | /* hmmm... no space left, create a new chunk */ | |
ccea34b5 TH |
1082 | spin_unlock_irq(&pcpu_lock); |
1083 | ||
fbf59bc9 TH |
1084 | chunk = alloc_pcpu_chunk(); |
1085 | if (!chunk) | |
ccea34b5 TH |
1086 | goto fail_unlock_mutex; |
1087 | ||
1088 | spin_lock_irq(&pcpu_lock); | |
fbf59bc9 | 1089 | pcpu_chunk_relocate(chunk, -1); |
ccea34b5 | 1090 | goto restart; |
fbf59bc9 TH |
1091 | |
1092 | area_found: | |
ccea34b5 TH |
1093 | spin_unlock_irq(&pcpu_lock); |
1094 | ||
fbf59bc9 TH |
1095 | /* populate, map and clear the area */ |
1096 | if (pcpu_populate_chunk(chunk, off, size)) { | |
ccea34b5 | 1097 | spin_lock_irq(&pcpu_lock); |
fbf59bc9 | 1098 | pcpu_free_area(chunk, off); |
ccea34b5 | 1099 | goto fail_unlock; |
fbf59bc9 TH |
1100 | } |
1101 | ||
ccea34b5 TH |
1102 | mutex_unlock(&pcpu_alloc_mutex); |
1103 | ||
2f39e637 | 1104 | /* return address relative to unit0 */ |
ccea34b5 TH |
1105 | return __addr_to_pcpu_ptr(chunk->vm->addr + off); |
1106 | ||
1107 | fail_unlock: | |
1108 | spin_unlock_irq(&pcpu_lock); | |
1109 | fail_unlock_mutex: | |
1110 | mutex_unlock(&pcpu_alloc_mutex); | |
1111 | return NULL; | |
fbf59bc9 | 1112 | } |
edcb4639 TH |
1113 | |
1114 | /** | |
1115 | * __alloc_percpu - allocate dynamic percpu area | |
1116 | * @size: size of area to allocate in bytes | |
1117 | * @align: alignment of area (max PAGE_SIZE) | |
1118 | * | |
1119 | * Allocate percpu area of @size bytes aligned at @align. Might | |
1120 | * sleep. Might trigger writeouts. | |
1121 | * | |
ccea34b5 TH |
1122 | * CONTEXT: |
1123 | * Does GFP_KERNEL allocation. | |
1124 | * | |
edcb4639 TH |
1125 | * RETURNS: |
1126 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1127 | */ | |
1128 | void *__alloc_percpu(size_t size, size_t align) | |
1129 | { | |
1130 | return pcpu_alloc(size, align, false); | |
1131 | } | |
fbf59bc9 TH |
1132 | EXPORT_SYMBOL_GPL(__alloc_percpu); |
1133 | ||
edcb4639 TH |
1134 | /** |
1135 | * __alloc_reserved_percpu - allocate reserved percpu area | |
1136 | * @size: size of area to allocate in bytes | |
1137 | * @align: alignment of area (max PAGE_SIZE) | |
1138 | * | |
1139 | * Allocate percpu area of @size bytes aligned at @align from reserved | |
1140 | * percpu area if arch has set it up; otherwise, allocation is served | |
1141 | * from the same dynamic area. Might sleep. Might trigger writeouts. | |
1142 | * | |
ccea34b5 TH |
1143 | * CONTEXT: |
1144 | * Does GFP_KERNEL allocation. | |
1145 | * | |
edcb4639 TH |
1146 | * RETURNS: |
1147 | * Percpu pointer to the allocated area on success, NULL on failure. | |
1148 | */ | |
1149 | void *__alloc_reserved_percpu(size_t size, size_t align) | |
1150 | { | |
1151 | return pcpu_alloc(size, align, true); | |
1152 | } | |
1153 | ||
a56dbddf TH |
1154 | /** |
1155 | * pcpu_reclaim - reclaim fully free chunks, workqueue function | |
1156 | * @work: unused | |
1157 | * | |
1158 | * Reclaim all fully free chunks except for the first one. | |
ccea34b5 TH |
1159 | * |
1160 | * CONTEXT: | |
1161 | * workqueue context. | |
a56dbddf TH |
1162 | */ |
1163 | static void pcpu_reclaim(struct work_struct *work) | |
fbf59bc9 | 1164 | { |
a56dbddf TH |
1165 | LIST_HEAD(todo); |
1166 | struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1]; | |
1167 | struct pcpu_chunk *chunk, *next; | |
1168 | ||
ccea34b5 TH |
1169 | mutex_lock(&pcpu_alloc_mutex); |
1170 | spin_lock_irq(&pcpu_lock); | |
a56dbddf TH |
1171 | |
1172 | list_for_each_entry_safe(chunk, next, head, list) { | |
1173 | WARN_ON(chunk->immutable); | |
1174 | ||
1175 | /* spare the first one */ | |
1176 | if (chunk == list_first_entry(head, struct pcpu_chunk, list)) | |
1177 | continue; | |
1178 | ||
a56dbddf TH |
1179 | list_move(&chunk->list, &todo); |
1180 | } | |
1181 | ||
ccea34b5 TH |
1182 | spin_unlock_irq(&pcpu_lock); |
1183 | mutex_unlock(&pcpu_alloc_mutex); | |
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 | } | |
fbf59bc9 TH |
1189 | } |
1190 | ||
1191 | /** | |
1192 | * free_percpu - free percpu area | |
1193 | * @ptr: pointer to area to free | |
1194 | * | |
ccea34b5 TH |
1195 | * Free percpu area @ptr. |
1196 | * | |
1197 | * CONTEXT: | |
1198 | * Can be called from atomic context. | |
fbf59bc9 TH |
1199 | */ |
1200 | void free_percpu(void *ptr) | |
1201 | { | |
1202 | void *addr = __pcpu_ptr_to_addr(ptr); | |
1203 | struct pcpu_chunk *chunk; | |
ccea34b5 | 1204 | unsigned long flags; |
fbf59bc9 TH |
1205 | int off; |
1206 | ||
1207 | if (!ptr) | |
1208 | return; | |
1209 | ||
ccea34b5 | 1210 | spin_lock_irqsave(&pcpu_lock, flags); |
fbf59bc9 TH |
1211 | |
1212 | chunk = pcpu_chunk_addr_search(addr); | |
1213 | off = addr - chunk->vm->addr; | |
1214 | ||
1215 | pcpu_free_area(chunk, off); | |
1216 | ||
a56dbddf | 1217 | /* if there are more than one fully free chunks, wake up grim reaper */ |
fbf59bc9 TH |
1218 | if (chunk->free_size == pcpu_unit_size) { |
1219 | struct pcpu_chunk *pos; | |
1220 | ||
a56dbddf | 1221 | list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list) |
fbf59bc9 | 1222 | if (pos != chunk) { |
a56dbddf | 1223 | schedule_work(&pcpu_reclaim_work); |
fbf59bc9 TH |
1224 | break; |
1225 | } | |
1226 | } | |
1227 | ||
ccea34b5 | 1228 | spin_unlock_irqrestore(&pcpu_lock, flags); |
fbf59bc9 TH |
1229 | } |
1230 | EXPORT_SYMBOL_GPL(free_percpu); | |
1231 | ||
1232 | /** | |
8d408b4b | 1233 | * pcpu_setup_first_chunk - initialize the first percpu chunk |
8d408b4b | 1234 | * @static_size: the size of static percpu area in bytes |
38a6be52 | 1235 | * @reserved_size: the size of reserved percpu area in bytes, 0 for none |
cafe8816 | 1236 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto |
38a6be52 TH |
1237 | * @unit_size: unit size in bytes, must be multiple of PAGE_SIZE |
1238 | * @base_addr: mapped address | |
2f39e637 | 1239 | * @unit_map: cpu -> unit map, NULL for sequential mapping |
8d408b4b TH |
1240 | * |
1241 | * Initialize the first percpu chunk which contains the kernel static | |
1242 | * perpcu area. This function is to be called from arch percpu area | |
38a6be52 | 1243 | * setup path. |
8d408b4b | 1244 | * |
edcb4639 TH |
1245 | * @reserved_size, if non-zero, specifies the amount of bytes to |
1246 | * reserve after the static area in the first chunk. This reserves | |
1247 | * the first chunk such that it's available only through reserved | |
1248 | * percpu allocation. This is primarily used to serve module percpu | |
1249 | * static areas on architectures where the addressing model has | |
1250 | * limited offset range for symbol relocations to guarantee module | |
1251 | * percpu symbols fall inside the relocatable range. | |
1252 | * | |
6074d5b0 TH |
1253 | * @dyn_size, if non-negative, determines the number of bytes |
1254 | * available for dynamic allocation in the first chunk. Specifying | |
1255 | * non-negative value makes percpu leave alone the area beyond | |
1256 | * @static_size + @reserved_size + @dyn_size. | |
1257 | * | |
38a6be52 TH |
1258 | * @unit_size specifies unit size and must be aligned to PAGE_SIZE and |
1259 | * equal to or larger than @static_size + @reserved_size + if | |
1260 | * non-negative, @dyn_size. | |
8d408b4b | 1261 | * |
38a6be52 TH |
1262 | * The caller should have mapped the first chunk at @base_addr and |
1263 | * copied static data to each unit. | |
fbf59bc9 | 1264 | * |
edcb4639 TH |
1265 | * If the first chunk ends up with both reserved and dynamic areas, it |
1266 | * is served by two chunks - one to serve the core static and reserved | |
1267 | * areas and the other for the dynamic area. They share the same vm | |
1268 | * and page map but uses different area allocation map to stay away | |
1269 | * from each other. The latter chunk is circulated in the chunk slots | |
1270 | * and available for dynamic allocation like any other chunks. | |
1271 | * | |
fbf59bc9 TH |
1272 | * RETURNS: |
1273 | * The determined pcpu_unit_size which can be used to initialize | |
1274 | * percpu access. | |
1275 | */ | |
ce3141a2 | 1276 | size_t __init pcpu_setup_first_chunk(size_t static_size, size_t reserved_size, |
38a6be52 | 1277 | ssize_t dyn_size, size_t unit_size, |
2f39e637 | 1278 | void *base_addr, const int *unit_map) |
fbf59bc9 | 1279 | { |
2441d15c | 1280 | static struct vm_struct first_vm; |
edcb4639 | 1281 | static int smap[2], dmap[2]; |
6074d5b0 TH |
1282 | size_t size_sum = static_size + reserved_size + |
1283 | (dyn_size >= 0 ? dyn_size : 0); | |
edcb4639 | 1284 | struct pcpu_chunk *schunk, *dchunk = NULL; |
2f39e637 | 1285 | unsigned int cpu, tcpu; |
ce3141a2 | 1286 | int i; |
fbf59bc9 | 1287 | |
2f39e637 | 1288 | /* sanity checks */ |
edcb4639 TH |
1289 | BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC || |
1290 | ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC); | |
8d408b4b | 1291 | BUG_ON(!static_size); |
38a6be52 TH |
1292 | BUG_ON(!base_addr); |
1293 | BUG_ON(unit_size < size_sum); | |
1294 | BUG_ON(unit_size & ~PAGE_MASK); | |
1295 | BUG_ON(unit_size < PCPU_MIN_UNIT_SIZE); | |
8d408b4b | 1296 | |
2f39e637 TH |
1297 | /* determine number of units and verify and initialize pcpu_unit_map */ |
1298 | if (unit_map) { | |
1299 | int first_unit = INT_MAX, last_unit = INT_MIN; | |
1300 | ||
1301 | for_each_possible_cpu(cpu) { | |
1302 | int unit = unit_map[cpu]; | |
1303 | ||
1304 | BUG_ON(unit < 0); | |
1305 | for_each_possible_cpu(tcpu) { | |
1306 | if (tcpu == cpu) | |
1307 | break; | |
1308 | /* the mapping should be one-to-one */ | |
1309 | BUG_ON(unit_map[tcpu] == unit); | |
1310 | } | |
1311 | ||
1312 | if (unit < first_unit) { | |
1313 | pcpu_first_unit_cpu = cpu; | |
1314 | first_unit = unit; | |
1315 | } | |
1316 | if (unit > last_unit) { | |
1317 | pcpu_last_unit_cpu = cpu; | |
1318 | last_unit = unit; | |
1319 | } | |
1320 | } | |
1321 | pcpu_nr_units = last_unit + 1; | |
1322 | pcpu_unit_map = unit_map; | |
1323 | } else { | |
1324 | int *identity_map; | |
1325 | ||
1326 | /* #units == #cpus, identity mapped */ | |
1327 | identity_map = alloc_bootmem(num_possible_cpus() * | |
1328 | sizeof(identity_map[0])); | |
1329 | ||
1330 | for_each_possible_cpu(cpu) | |
1331 | identity_map[cpu] = cpu; | |
1332 | ||
1333 | pcpu_first_unit_cpu = 0; | |
1334 | pcpu_last_unit_cpu = pcpu_nr_units - 1; | |
1335 | pcpu_nr_units = num_possible_cpus(); | |
1336 | pcpu_unit_map = identity_map; | |
1337 | } | |
1338 | ||
1339 | /* determine basic parameters */ | |
38a6be52 | 1340 | pcpu_unit_pages = unit_size >> PAGE_SHIFT; |
d9b55eeb | 1341 | pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT; |
2f39e637 | 1342 | pcpu_chunk_size = pcpu_nr_units * pcpu_unit_size; |
ce3141a2 TH |
1343 | pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) + |
1344 | BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long); | |
fbf59bc9 | 1345 | |
cafe8816 | 1346 | if (dyn_size < 0) |
edcb4639 | 1347 | dyn_size = pcpu_unit_size - static_size - reserved_size; |
cafe8816 | 1348 | |
38a6be52 TH |
1349 | first_vm.flags = VM_ALLOC; |
1350 | first_vm.size = pcpu_chunk_size; | |
1351 | first_vm.addr = base_addr; | |
1352 | ||
d9b55eeb TH |
1353 | /* |
1354 | * Allocate chunk slots. The additional last slot is for | |
1355 | * empty chunks. | |
1356 | */ | |
1357 | pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2; | |
fbf59bc9 TH |
1358 | pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0])); |
1359 | for (i = 0; i < pcpu_nr_slots; i++) | |
1360 | INIT_LIST_HEAD(&pcpu_slot[i]); | |
1361 | ||
edcb4639 TH |
1362 | /* |
1363 | * Initialize static chunk. If reserved_size is zero, the | |
1364 | * static chunk covers static area + dynamic allocation area | |
1365 | * in the first chunk. If reserved_size is not zero, it | |
1366 | * covers static area + reserved area (mostly used for module | |
1367 | * static percpu allocation). | |
1368 | */ | |
2441d15c TH |
1369 | schunk = alloc_bootmem(pcpu_chunk_struct_size); |
1370 | INIT_LIST_HEAD(&schunk->list); | |
1371 | schunk->vm = &first_vm; | |
61ace7fa TH |
1372 | schunk->map = smap; |
1373 | schunk->map_alloc = ARRAY_SIZE(smap); | |
38a6be52 | 1374 | schunk->immutable = true; |
ce3141a2 | 1375 | bitmap_fill(schunk->populated, pcpu_unit_pages); |
edcb4639 TH |
1376 | |
1377 | if (reserved_size) { | |
1378 | schunk->free_size = reserved_size; | |
ae9e6bc9 TH |
1379 | pcpu_reserved_chunk = schunk; |
1380 | pcpu_reserved_chunk_limit = static_size + reserved_size; | |
edcb4639 TH |
1381 | } else { |
1382 | schunk->free_size = dyn_size; | |
1383 | dyn_size = 0; /* dynamic area covered */ | |
1384 | } | |
2441d15c | 1385 | schunk->contig_hint = schunk->free_size; |
fbf59bc9 | 1386 | |
61ace7fa TH |
1387 | schunk->map[schunk->map_used++] = -static_size; |
1388 | if (schunk->free_size) | |
1389 | schunk->map[schunk->map_used++] = schunk->free_size; | |
1390 | ||
edcb4639 TH |
1391 | /* init dynamic chunk if necessary */ |
1392 | if (dyn_size) { | |
ce3141a2 | 1393 | dchunk = alloc_bootmem(pcpu_chunk_struct_size); |
edcb4639 TH |
1394 | INIT_LIST_HEAD(&dchunk->list); |
1395 | dchunk->vm = &first_vm; | |
1396 | dchunk->map = dmap; | |
1397 | dchunk->map_alloc = ARRAY_SIZE(dmap); | |
38a6be52 | 1398 | dchunk->immutable = true; |
ce3141a2 | 1399 | bitmap_fill(dchunk->populated, pcpu_unit_pages); |
edcb4639 TH |
1400 | |
1401 | dchunk->contig_hint = dchunk->free_size = dyn_size; | |
1402 | dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit; | |
1403 | dchunk->map[dchunk->map_used++] = dchunk->free_size; | |
1404 | } | |
1405 | ||
2441d15c | 1406 | /* link the first chunk in */ |
ae9e6bc9 TH |
1407 | pcpu_first_chunk = dchunk ?: schunk; |
1408 | pcpu_chunk_relocate(pcpu_first_chunk, -1); | |
fbf59bc9 TH |
1409 | |
1410 | /* we're done */ | |
2f39e637 | 1411 | pcpu_base_addr = schunk->vm->addr; |
fbf59bc9 TH |
1412 | return pcpu_unit_size; |
1413 | } | |
66c3a757 | 1414 | |
8c4bfc6e TH |
1415 | static size_t pcpu_calc_fc_sizes(size_t static_size, size_t reserved_size, |
1416 | ssize_t *dyn_sizep) | |
1417 | { | |
1418 | size_t size_sum; | |
1419 | ||
1420 | size_sum = PFN_ALIGN(static_size + reserved_size + | |
1421 | (*dyn_sizep >= 0 ? *dyn_sizep : 0)); | |
1422 | if (*dyn_sizep != 0) | |
1423 | *dyn_sizep = size_sum - static_size - reserved_size; | |
1424 | ||
1425 | return size_sum; | |
1426 | } | |
1427 | ||
66c3a757 TH |
1428 | /** |
1429 | * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem | |
1430 | * @static_size: the size of static percpu area in bytes | |
1431 | * @reserved_size: the size of reserved percpu area in bytes | |
1432 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto | |
66c3a757 TH |
1433 | * |
1434 | * This is a helper to ease setting up embedded first percpu chunk and | |
1435 | * can be called where pcpu_setup_first_chunk() is expected. | |
1436 | * | |
1437 | * If this function is used to setup the first chunk, it is allocated | |
1438 | * as a contiguous area using bootmem allocator and used as-is without | |
1439 | * being mapped into vmalloc area. This enables the first chunk to | |
1440 | * piggy back on the linear physical mapping which often uses larger | |
1441 | * page size. | |
1442 | * | |
1443 | * When @dyn_size is positive, dynamic area might be larger than | |
788e5abc TH |
1444 | * specified to fill page alignment. When @dyn_size is auto, |
1445 | * @dyn_size is just big enough to fill page alignment after static | |
1446 | * and reserved areas. | |
66c3a757 TH |
1447 | * |
1448 | * If the needed size is smaller than the minimum or specified unit | |
1449 | * size, the leftover is returned to the bootmem allocator. | |
1450 | * | |
1451 | * RETURNS: | |
1452 | * The determined pcpu_unit_size which can be used to initialize | |
1453 | * percpu access on success, -errno on failure. | |
1454 | */ | |
1455 | ssize_t __init pcpu_embed_first_chunk(size_t static_size, size_t reserved_size, | |
788e5abc | 1456 | ssize_t dyn_size) |
66c3a757 | 1457 | { |
ce3141a2 TH |
1458 | size_t size_sum, unit_size, chunk_size; |
1459 | void *base; | |
66c3a757 TH |
1460 | unsigned int cpu; |
1461 | ||
1462 | /* determine parameters and allocate */ | |
ce3141a2 | 1463 | size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); |
66c3a757 | 1464 | |
ce3141a2 TH |
1465 | unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); |
1466 | chunk_size = unit_size * num_possible_cpus(); | |
fa8a7094 | 1467 | |
ce3141a2 TH |
1468 | base = __alloc_bootmem_nopanic(chunk_size, PAGE_SIZE, |
1469 | __pa(MAX_DMA_ADDRESS)); | |
1470 | if (!base) { | |
fa8a7094 TH |
1471 | pr_warning("PERCPU: failed to allocate %zu bytes for " |
1472 | "embedding\n", chunk_size); | |
66c3a757 | 1473 | return -ENOMEM; |
fa8a7094 | 1474 | } |
66c3a757 TH |
1475 | |
1476 | /* return the leftover and copy */ | |
1477 | for_each_possible_cpu(cpu) { | |
ce3141a2 | 1478 | void *ptr = base + cpu * unit_size; |
66c3a757 | 1479 | |
ce3141a2 | 1480 | free_bootmem(__pa(ptr + size_sum), unit_size - size_sum); |
66c3a757 TH |
1481 | memcpy(ptr, __per_cpu_load, static_size); |
1482 | } | |
1483 | ||
1484 | /* we're ready, commit */ | |
1485 | pr_info("PERCPU: Embedded %zu pages at %p, static data %zu bytes\n", | |
ce3141a2 | 1486 | size_sum >> PAGE_SHIFT, base, static_size); |
d4b95f80 | 1487 | |
ce3141a2 | 1488 | return pcpu_setup_first_chunk(static_size, reserved_size, dyn_size, |
2f39e637 | 1489 | unit_size, base, NULL); |
d4b95f80 TH |
1490 | } |
1491 | ||
1492 | /** | |
1493 | * pcpu_4k_first_chunk - map the first chunk using PAGE_SIZE pages | |
1494 | * @static_size: the size of static percpu area in bytes | |
1495 | * @reserved_size: the size of reserved percpu area in bytes | |
1496 | * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE | |
1497 | * @free_fn: funtion to free percpu page, always called with PAGE_SIZE | |
1498 | * @populate_pte_fn: function to populate pte | |
1499 | * | |
1500 | * This is a helper to ease setting up embedded first percpu chunk and | |
1501 | * can be called where pcpu_setup_first_chunk() is expected. | |
1502 | * | |
1503 | * This is the basic allocator. Static percpu area is allocated | |
1504 | * page-by-page into vmalloc area. | |
1505 | * | |
1506 | * RETURNS: | |
1507 | * The determined pcpu_unit_size which can be used to initialize | |
1508 | * percpu access on success, -errno on failure. | |
1509 | */ | |
1510 | ssize_t __init pcpu_4k_first_chunk(size_t static_size, size_t reserved_size, | |
1511 | pcpu_fc_alloc_fn_t alloc_fn, | |
1512 | pcpu_fc_free_fn_t free_fn, | |
1513 | pcpu_fc_populate_pte_fn_t populate_pte_fn) | |
1514 | { | |
8f05a6a6 | 1515 | static struct vm_struct vm; |
ce3141a2 | 1516 | int unit_pages; |
d4b95f80 | 1517 | size_t pages_size; |
ce3141a2 | 1518 | struct page **pages; |
d4b95f80 TH |
1519 | unsigned int cpu; |
1520 | int i, j; | |
1521 | ssize_t ret; | |
1522 | ||
ce3141a2 TH |
1523 | unit_pages = PFN_UP(max_t(size_t, static_size + reserved_size, |
1524 | PCPU_MIN_UNIT_SIZE)); | |
d4b95f80 TH |
1525 | |
1526 | /* unaligned allocations can't be freed, round up to page size */ | |
ce3141a2 TH |
1527 | pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() * |
1528 | sizeof(pages[0])); | |
1529 | pages = alloc_bootmem(pages_size); | |
d4b95f80 | 1530 | |
8f05a6a6 | 1531 | /* allocate pages */ |
d4b95f80 TH |
1532 | j = 0; |
1533 | for_each_possible_cpu(cpu) | |
ce3141a2 | 1534 | for (i = 0; i < unit_pages; i++) { |
d4b95f80 TH |
1535 | void *ptr; |
1536 | ||
1537 | ptr = alloc_fn(cpu, PAGE_SIZE); | |
1538 | if (!ptr) { | |
1539 | pr_warning("PERCPU: failed to allocate " | |
1540 | "4k page for cpu%u\n", cpu); | |
1541 | goto enomem; | |
1542 | } | |
ce3141a2 | 1543 | pages[j++] = virt_to_page(ptr); |
d4b95f80 TH |
1544 | } |
1545 | ||
8f05a6a6 TH |
1546 | /* allocate vm area, map the pages and copy static data */ |
1547 | vm.flags = VM_ALLOC; | |
ce3141a2 | 1548 | vm.size = num_possible_cpus() * unit_pages << PAGE_SHIFT; |
8f05a6a6 TH |
1549 | vm_area_register_early(&vm, PAGE_SIZE); |
1550 | ||
1551 | for_each_possible_cpu(cpu) { | |
1552 | unsigned long unit_addr = (unsigned long)vm.addr + | |
ce3141a2 | 1553 | (cpu * unit_pages << PAGE_SHIFT); |
8f05a6a6 | 1554 | |
ce3141a2 | 1555 | for (i = 0; i < unit_pages; i++) |
8f05a6a6 TH |
1556 | populate_pte_fn(unit_addr + (i << PAGE_SHIFT)); |
1557 | ||
1558 | /* pte already populated, the following shouldn't fail */ | |
ce3141a2 TH |
1559 | ret = __pcpu_map_pages(unit_addr, &pages[cpu * unit_pages], |
1560 | unit_pages); | |
8f05a6a6 TH |
1561 | if (ret < 0) |
1562 | panic("failed to map percpu area, err=%zd\n", ret); | |
1563 | ||
1564 | /* | |
1565 | * FIXME: Archs with virtual cache should flush local | |
1566 | * cache for the linear mapping here - something | |
1567 | * equivalent to flush_cache_vmap() on the local cpu. | |
1568 | * flush_cache_vmap() can't be used as most supporting | |
1569 | * data structures are not set up yet. | |
1570 | */ | |
1571 | ||
1572 | /* copy static data */ | |
1573 | memcpy((void *)unit_addr, __per_cpu_load, static_size); | |
1574 | } | |
1575 | ||
d4b95f80 | 1576 | /* we're ready, commit */ |
8f05a6a6 | 1577 | pr_info("PERCPU: %d 4k pages per cpu, static data %zu bytes\n", |
ce3141a2 | 1578 | unit_pages, static_size); |
d4b95f80 | 1579 | |
ce3141a2 | 1580 | ret = pcpu_setup_first_chunk(static_size, reserved_size, -1, |
2f39e637 | 1581 | unit_pages << PAGE_SHIFT, vm.addr, NULL); |
d4b95f80 TH |
1582 | goto out_free_ar; |
1583 | ||
1584 | enomem: | |
1585 | while (--j >= 0) | |
ce3141a2 | 1586 | free_fn(page_address(pages[j]), PAGE_SIZE); |
d4b95f80 TH |
1587 | ret = -ENOMEM; |
1588 | out_free_ar: | |
ce3141a2 | 1589 | free_bootmem(__pa(pages), pages_size); |
d4b95f80 TH |
1590 | return ret; |
1591 | } | |
1592 | ||
8c4bfc6e TH |
1593 | /* |
1594 | * Large page remapping first chunk setup helper | |
1595 | */ | |
1596 | #ifdef CONFIG_NEED_MULTIPLE_NODES | |
1597 | struct pcpul_ent { | |
1598 | unsigned int cpu; | |
1599 | void *ptr; | |
1600 | }; | |
1601 | ||
1602 | static size_t pcpul_size; | |
1603 | static size_t pcpul_unit_size; | |
1604 | static struct pcpul_ent *pcpul_map; | |
1605 | static struct vm_struct pcpul_vm; | |
1606 | ||
8c4bfc6e TH |
1607 | /** |
1608 | * pcpu_lpage_first_chunk - remap the first percpu chunk using large page | |
1609 | * @static_size: the size of static percpu area in bytes | |
1610 | * @reserved_size: the size of reserved percpu area in bytes | |
1611 | * @dyn_size: free size for dynamic allocation in bytes, -1 for auto | |
1612 | * @lpage_size: the size of a large page | |
1613 | * @alloc_fn: function to allocate percpu lpage, always called with lpage_size | |
1614 | * @free_fn: function to free percpu memory, @size <= lpage_size | |
1615 | * @map_fn: function to map percpu lpage, always called with lpage_size | |
1616 | * | |
1617 | * This allocator uses large page as unit. A large page is allocated | |
1618 | * for each cpu and each is remapped into vmalloc area using large | |
1619 | * page mapping. As large page can be quite large, only part of it is | |
1620 | * used for the first chunk. Unused part is returned to the bootmem | |
1621 | * allocator. | |
1622 | * | |
1623 | * So, the large pages are mapped twice - once to the physical mapping | |
1624 | * and to the vmalloc area for the first percpu chunk. The double | |
1625 | * mapping does add one more large TLB entry pressure but still is | |
1626 | * much better than only using 4k mappings while still being NUMA | |
1627 | * friendly. | |
1628 | * | |
1629 | * RETURNS: | |
1630 | * The determined pcpu_unit_size which can be used to initialize | |
1631 | * percpu access on success, -errno on failure. | |
1632 | */ | |
1633 | ssize_t __init pcpu_lpage_first_chunk(size_t static_size, size_t reserved_size, | |
1634 | ssize_t dyn_size, size_t lpage_size, | |
1635 | pcpu_fc_alloc_fn_t alloc_fn, | |
1636 | pcpu_fc_free_fn_t free_fn, | |
1637 | pcpu_fc_map_fn_t map_fn) | |
1638 | { | |
1639 | size_t size_sum; | |
1640 | size_t map_size; | |
1641 | unsigned int cpu; | |
1642 | int i, j; | |
1643 | ssize_t ret; | |
1644 | ||
1645 | /* | |
1646 | * Currently supports only single page. Supporting multiple | |
1647 | * pages won't be too difficult if it ever becomes necessary. | |
1648 | */ | |
1649 | size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size); | |
1650 | ||
1651 | pcpul_unit_size = lpage_size; | |
1652 | pcpul_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE); | |
1653 | if (pcpul_size > pcpul_unit_size) { | |
1654 | pr_warning("PERCPU: static data is larger than large page, " | |
1655 | "can't use large page\n"); | |
1656 | return -EINVAL; | |
1657 | } | |
1658 | ||
1659 | /* allocate pointer array and alloc large pages */ | |
1660 | map_size = PFN_ALIGN(num_possible_cpus() * sizeof(pcpul_map[0])); | |
1661 | pcpul_map = alloc_bootmem(map_size); | |
1662 | ||
1663 | for_each_possible_cpu(cpu) { | |
1664 | void *ptr; | |
1665 | ||
1666 | ptr = alloc_fn(cpu, lpage_size); | |
1667 | if (!ptr) { | |
1668 | pr_warning("PERCPU: failed to allocate large page " | |
1669 | "for cpu%u\n", cpu); | |
1670 | goto enomem; | |
1671 | } | |
1672 | ||
1673 | /* | |
1674 | * Only use pcpul_size bytes and give back the rest. | |
1675 | * | |
1676 | * Ingo: The lpage_size up-rounding bootmem is needed | |
1677 | * to make sure the partial lpage is still fully RAM - | |
1678 | * it's not well-specified to have a incompatible area | |
1679 | * (unmapped RAM, device memory, etc.) in that hole. | |
1680 | */ | |
1681 | free_fn(ptr + pcpul_size, lpage_size - pcpul_size); | |
1682 | ||
1683 | pcpul_map[cpu].cpu = cpu; | |
1684 | pcpul_map[cpu].ptr = ptr; | |
1685 | ||
1686 | memcpy(ptr, __per_cpu_load, static_size); | |
1687 | } | |
1688 | ||
1689 | /* allocate address and map */ | |
1690 | pcpul_vm.flags = VM_ALLOC; | |
1691 | pcpul_vm.size = num_possible_cpus() * pcpul_unit_size; | |
1692 | vm_area_register_early(&pcpul_vm, pcpul_unit_size); | |
1693 | ||
1694 | for_each_possible_cpu(cpu) | |
1695 | map_fn(pcpul_map[cpu].ptr, pcpul_unit_size, | |
1696 | pcpul_vm.addr + cpu * pcpul_unit_size); | |
1697 | ||
1698 | /* we're ready, commit */ | |
1699 | pr_info("PERCPU: Remapped at %p with large pages, static data " | |
1700 | "%zu bytes\n", pcpul_vm.addr, static_size); | |
1701 | ||
ce3141a2 | 1702 | ret = pcpu_setup_first_chunk(static_size, reserved_size, dyn_size, |
2f39e637 | 1703 | pcpul_unit_size, pcpul_vm.addr, NULL); |
8c4bfc6e TH |
1704 | |
1705 | /* sort pcpul_map array for pcpu_lpage_remapped() */ | |
1706 | for (i = 0; i < num_possible_cpus() - 1; i++) | |
1707 | for (j = i + 1; j < num_possible_cpus(); j++) | |
1708 | if (pcpul_map[i].ptr > pcpul_map[j].ptr) { | |
1709 | struct pcpul_ent tmp = pcpul_map[i]; | |
1710 | pcpul_map[i] = pcpul_map[j]; | |
1711 | pcpul_map[j] = tmp; | |
1712 | } | |
1713 | ||
1714 | return ret; | |
1715 | ||
1716 | enomem: | |
1717 | for_each_possible_cpu(cpu) | |
1718 | if (pcpul_map[cpu].ptr) | |
1719 | free_fn(pcpul_map[cpu].ptr, pcpul_size); | |
1720 | free_bootmem(__pa(pcpul_map), map_size); | |
1721 | return -ENOMEM; | |
1722 | } | |
1723 | ||
1724 | /** | |
1725 | * pcpu_lpage_remapped - determine whether a kaddr is in pcpul recycled area | |
1726 | * @kaddr: the kernel address in question | |
1727 | * | |
1728 | * Determine whether @kaddr falls in the pcpul recycled area. This is | |
1729 | * used by pageattr to detect VM aliases and break up the pcpu large | |
1730 | * page mapping such that the same physical page is not mapped under | |
1731 | * different attributes. | |
1732 | * | |
1733 | * The recycled area is always at the tail of a partially used large | |
1734 | * page. | |
1735 | * | |
1736 | * RETURNS: | |
1737 | * Address of corresponding remapped pcpu address if match is found; | |
1738 | * otherwise, NULL. | |
1739 | */ | |
1740 | void *pcpu_lpage_remapped(void *kaddr) | |
1741 | { | |
1742 | unsigned long unit_mask = pcpul_unit_size - 1; | |
1743 | void *lpage_addr = (void *)((unsigned long)kaddr & ~unit_mask); | |
1744 | unsigned long offset = (unsigned long)kaddr & unit_mask; | |
1745 | int left = 0, right = num_possible_cpus() - 1; | |
1746 | int pos; | |
1747 | ||
1748 | /* pcpul in use at all? */ | |
1749 | if (!pcpul_map) | |
1750 | return NULL; | |
1751 | ||
1752 | /* okay, perform binary search */ | |
1753 | while (left <= right) { | |
1754 | pos = (left + right) / 2; | |
1755 | ||
1756 | if (pcpul_map[pos].ptr < lpage_addr) | |
1757 | left = pos + 1; | |
1758 | else if (pcpul_map[pos].ptr > lpage_addr) | |
1759 | right = pos - 1; | |
1760 | else { | |
1761 | /* it shouldn't be in the area for the first chunk */ | |
1762 | WARN_ON(offset < pcpul_size); | |
1763 | ||
1764 | return pcpul_vm.addr + | |
1765 | pcpul_map[pos].cpu * pcpul_unit_size + offset; | |
1766 | } | |
1767 | } | |
1768 | ||
1769 | return NULL; | |
1770 | } | |
1771 | #endif | |
1772 | ||
e74e3962 TH |
1773 | /* |
1774 | * Generic percpu area setup. | |
1775 | * | |
1776 | * The embedding helper is used because its behavior closely resembles | |
1777 | * the original non-dynamic generic percpu area setup. This is | |
1778 | * important because many archs have addressing restrictions and might | |
1779 | * fail if the percpu area is located far away from the previous | |
1780 | * location. As an added bonus, in non-NUMA cases, embedding is | |
1781 | * generally a good idea TLB-wise because percpu area can piggy back | |
1782 | * on the physical linear memory mapping which uses large page | |
1783 | * mappings on applicable archs. | |
1784 | */ | |
1785 | #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA | |
1786 | unsigned long __per_cpu_offset[NR_CPUS] __read_mostly; | |
1787 | EXPORT_SYMBOL(__per_cpu_offset); | |
1788 | ||
1789 | void __init setup_per_cpu_areas(void) | |
1790 | { | |
1791 | size_t static_size = __per_cpu_end - __per_cpu_start; | |
1792 | ssize_t unit_size; | |
1793 | unsigned long delta; | |
1794 | unsigned int cpu; | |
1795 | ||
1796 | /* | |
1797 | * Always reserve area for module percpu variables. That's | |
1798 | * what the legacy allocator did. | |
1799 | */ | |
1800 | unit_size = pcpu_embed_first_chunk(static_size, PERCPU_MODULE_RESERVE, | |
788e5abc | 1801 | PERCPU_DYNAMIC_RESERVE); |
e74e3962 TH |
1802 | if (unit_size < 0) |
1803 | panic("Failed to initialized percpu areas."); | |
1804 | ||
1805 | delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start; | |
1806 | for_each_possible_cpu(cpu) | |
1807 | __per_cpu_offset[cpu] = delta + cpu * unit_size; | |
1808 | } | |
1809 | #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */ |