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