percpu: remove compile warnings caused by __verify_pcpu_ptr()
[linux-2.6-block.git] / mm / percpu.c
CommitLineData
fbf59bc9
TH
1/*
2 * linux/mm/percpu.c - percpu memory allocator
3 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks in vmalloc area. Each
2f39e637
TH
11 * chunk is consisted of boot-time determined number of units and the
12 * first chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated. ie. in
17 * vmalloc area
fbf59bc9
TH
18 *
19 * c0 c1 c2
20 * ------------------- ------------------- ------------
21 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
22 * ------------------- ...... ------------------- .... ------------
23 *
24 * Allocation is done in offset-size areas of single unit space. Ie,
25 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
2f39e637
TH
26 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
27 * cpus. On NUMA, the mapping can be non-linear and even sparse.
28 * Percpu access can be done by configuring percpu base registers
29 * according to cpu to unit mapping and pcpu_unit_size.
fbf59bc9 30 *
2f39e637
TH
31 * There are usually many small percpu allocations many of them being
32 * as small as 4 bytes. The allocator organizes chunks into lists
fbf59bc9
TH
33 * according to free size and tries to allocate from the fullest one.
34 * Each chunk keeps the maximum contiguous area size hint which is
35 * guaranteed to be eqaul to or larger than the maximum contiguous
36 * area in the chunk. This helps the allocator not to iterate the
37 * chunk maps unnecessarily.
38 *
39 * Allocation state in each chunk is kept using an array of integers
40 * on chunk->map. A positive value in the map represents a free
41 * region and negative allocated. Allocation inside a chunk is done
42 * by scanning this map sequentially and serving the first matching
43 * entry. This is mostly copied from the percpu_modalloc() allocator.
e1b9aa3f
CL
44 * Chunks can be determined from the address using the index field
45 * in the page struct. The index field contains a pointer to the chunk.
fbf59bc9
TH
46 *
47 * To use this allocator, arch code should do the followings.
48 *
fbf59bc9 49 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
e0100983
TH
50 * regular address to percpu pointer and back if they need to be
51 * different from the default
fbf59bc9 52 *
8d408b4b
TH
53 * - use pcpu_setup_first_chunk() during percpu area initialization to
54 * setup the first chunk containing the kernel static percpu area
fbf59bc9
TH
55 */
56
57#include <linux/bitmap.h>
58#include <linux/bootmem.h>
fd1e8a1f 59#include <linux/err.h>
fbf59bc9 60#include <linux/list.h>
a530b795 61#include <linux/log2.h>
fbf59bc9
TH
62#include <linux/mm.h>
63#include <linux/module.h>
64#include <linux/mutex.h>
65#include <linux/percpu.h>
66#include <linux/pfn.h>
fbf59bc9 67#include <linux/slab.h>
ccea34b5 68#include <linux/spinlock.h>
fbf59bc9 69#include <linux/vmalloc.h>
a56dbddf 70#include <linux/workqueue.h>
fbf59bc9
TH
71
72#include <asm/cacheflush.h>
e0100983 73#include <asm/sections.h>
fbf59bc9
TH
74#include <asm/tlbflush.h>
75
fbf59bc9
TH
76#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
77#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
78
e0100983
TH
79/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
80#ifndef __addr_to_pcpu_ptr
81#define __addr_to_pcpu_ptr(addr) \
82 (void *)((unsigned long)(addr) - (unsigned long)pcpu_base_addr \
83 + (unsigned long)__per_cpu_start)
84#endif
85#ifndef __pcpu_ptr_to_addr
86#define __pcpu_ptr_to_addr(ptr) \
87 (void *)((unsigned long)(ptr) + (unsigned long)pcpu_base_addr \
88 - (unsigned long)__per_cpu_start)
89#endif
90
fbf59bc9
TH
91struct pcpu_chunk {
92 struct list_head list; /* linked to pcpu_slot lists */
fbf59bc9
TH
93 int free_size; /* free bytes in the chunk */
94 int contig_hint; /* max contiguous size hint */
bba174f5 95 void *base_addr; /* base address of this chunk */
fbf59bc9
TH
96 int map_used; /* # of map entries used */
97 int map_alloc; /* # of map entries allocated */
98 int *map; /* allocation map */
6563297c 99 struct vm_struct **vms; /* mapped vmalloc regions */
8d408b4b 100 bool immutable; /* no [de]population allowed */
ce3141a2 101 unsigned long populated[]; /* populated bitmap */
fbf59bc9
TH
102};
103
40150d37
TH
104static int pcpu_unit_pages __read_mostly;
105static int pcpu_unit_size __read_mostly;
2f39e637 106static int pcpu_nr_units __read_mostly;
6563297c 107static int pcpu_atom_size __read_mostly;
40150d37
TH
108static int pcpu_nr_slots __read_mostly;
109static size_t pcpu_chunk_struct_size __read_mostly;
fbf59bc9 110
2f39e637
TH
111/* cpus with the lowest and highest unit numbers */
112static unsigned int pcpu_first_unit_cpu __read_mostly;
113static unsigned int pcpu_last_unit_cpu __read_mostly;
114
fbf59bc9 115/* the address of the first chunk which starts with the kernel static area */
40150d37 116void *pcpu_base_addr __read_mostly;
fbf59bc9
TH
117EXPORT_SYMBOL_GPL(pcpu_base_addr);
118
fb435d52
TH
119static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
120const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
2f39e637 121
6563297c
TH
122/* group information, used for vm allocation */
123static int pcpu_nr_groups __read_mostly;
124static const unsigned long *pcpu_group_offsets __read_mostly;
125static const size_t *pcpu_group_sizes __read_mostly;
126
ae9e6bc9
TH
127/*
128 * The first chunk which always exists. Note that unlike other
129 * chunks, this one can be allocated and mapped in several different
130 * ways and thus often doesn't live in the vmalloc area.
131 */
132static struct pcpu_chunk *pcpu_first_chunk;
133
134/*
135 * Optional reserved chunk. This chunk reserves part of the first
136 * chunk and serves it for reserved allocations. The amount of
137 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
138 * area doesn't exist, the following variables contain NULL and 0
139 * respectively.
140 */
edcb4639 141static struct pcpu_chunk *pcpu_reserved_chunk;
edcb4639
TH
142static int pcpu_reserved_chunk_limit;
143
fbf59bc9 144/*
ccea34b5
TH
145 * Synchronization rules.
146 *
147 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
ce3141a2
TH
148 * protects allocation/reclaim paths, chunks, populated bitmap and
149 * vmalloc mapping. The latter is a spinlock and protects the index
150 * data structures - chunk slots, chunks and area maps in chunks.
ccea34b5
TH
151 *
152 * During allocation, pcpu_alloc_mutex is kept locked all the time and
153 * pcpu_lock is grabbed and released as necessary. All actual memory
154 * allocations are done using GFP_KERNEL with pcpu_lock released.
155 *
156 * Free path accesses and alters only the index data structures, so it
157 * can be safely called from atomic context. When memory needs to be
158 * returned to the system, free path schedules reclaim_work which
159 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
160 * reclaimed, release both locks and frees the chunks. Note that it's
161 * necessary to grab both locks to remove a chunk from circulation as
162 * allocation path might be referencing the chunk with only
163 * pcpu_alloc_mutex locked.
fbf59bc9 164 */
ccea34b5
TH
165static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
166static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
fbf59bc9 167
40150d37 168static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
fbf59bc9 169
a56dbddf
TH
170/* reclaim work to release fully free chunks, scheduled from free path */
171static void pcpu_reclaim(struct work_struct *work);
172static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
173
d9b55eeb 174static int __pcpu_size_to_slot(int size)
fbf59bc9 175{
cae3aeb8 176 int highbit = fls(size); /* size is in bytes */
fbf59bc9
TH
177 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
178}
179
d9b55eeb
TH
180static int pcpu_size_to_slot(int size)
181{
182 if (size == pcpu_unit_size)
183 return pcpu_nr_slots - 1;
184 return __pcpu_size_to_slot(size);
185}
186
fbf59bc9
TH
187static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
188{
189 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
190 return 0;
191
192 return pcpu_size_to_slot(chunk->free_size);
193}
194
195static int pcpu_page_idx(unsigned int cpu, int page_idx)
196{
2f39e637 197 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
fbf59bc9
TH
198}
199
200static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
201 unsigned int cpu, int page_idx)
202{
bba174f5 203 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
fb435d52 204 (page_idx << PAGE_SHIFT);
fbf59bc9
TH
205}
206
ce3141a2
TH
207static struct page *pcpu_chunk_page(struct pcpu_chunk *chunk,
208 unsigned int cpu, int page_idx)
fbf59bc9 209{
ce3141a2
TH
210 /* must not be used on pre-mapped chunk */
211 WARN_ON(chunk->immutable);
c8a51be4 212
ce3141a2 213 return vmalloc_to_page((void *)pcpu_chunk_addr(chunk, cpu, page_idx));
fbf59bc9
TH
214}
215
e1b9aa3f
CL
216/* set the pointer to a chunk in a page struct */
217static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
218{
219 page->index = (unsigned long)pcpu;
220}
221
222/* obtain pointer to a chunk from a page struct */
223static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
224{
225 return (struct pcpu_chunk *)page->index;
226}
227
ce3141a2
TH
228static void pcpu_next_unpop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
229{
230 *rs = find_next_zero_bit(chunk->populated, end, *rs);
231 *re = find_next_bit(chunk->populated, end, *rs + 1);
232}
233
234static void pcpu_next_pop(struct pcpu_chunk *chunk, int *rs, int *re, int end)
235{
236 *rs = find_next_bit(chunk->populated, end, *rs);
237 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
238}
239
240/*
241 * (Un)populated page region iterators. Iterate over (un)populated
242 * page regions betwen @start and @end in @chunk. @rs and @re should
243 * be integer variables and will be set to start and end page index of
244 * the current region.
245 */
246#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
247 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
248 (rs) < (re); \
249 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
250
251#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
252 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
253 (rs) < (re); \
254 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
255
fbf59bc9 256/**
1880d93b
TH
257 * pcpu_mem_alloc - allocate memory
258 * @size: bytes to allocate
fbf59bc9 259 *
1880d93b
TH
260 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
261 * kzalloc() is used; otherwise, vmalloc() is used. The returned
262 * memory is always zeroed.
fbf59bc9 263 *
ccea34b5
TH
264 * CONTEXT:
265 * Does GFP_KERNEL allocation.
266 *
fbf59bc9 267 * RETURNS:
1880d93b 268 * Pointer to the allocated area on success, NULL on failure.
fbf59bc9 269 */
1880d93b 270static void *pcpu_mem_alloc(size_t size)
fbf59bc9 271{
1880d93b
TH
272 if (size <= PAGE_SIZE)
273 return kzalloc(size, GFP_KERNEL);
274 else {
275 void *ptr = vmalloc(size);
276 if (ptr)
277 memset(ptr, 0, size);
278 return ptr;
279 }
280}
fbf59bc9 281
1880d93b
TH
282/**
283 * pcpu_mem_free - free memory
284 * @ptr: memory to free
285 * @size: size of the area
286 *
287 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
288 */
289static void pcpu_mem_free(void *ptr, size_t size)
290{
fbf59bc9 291 if (size <= PAGE_SIZE)
1880d93b 292 kfree(ptr);
fbf59bc9 293 else
1880d93b 294 vfree(ptr);
fbf59bc9
TH
295}
296
297/**
298 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
299 * @chunk: chunk of interest
300 * @oslot: the previous slot it was on
301 *
302 * This function is called after an allocation or free changed @chunk.
303 * New slot according to the changed state is determined and @chunk is
edcb4639
TH
304 * moved to the slot. Note that the reserved chunk is never put on
305 * chunk slots.
ccea34b5
TH
306 *
307 * CONTEXT:
308 * pcpu_lock.
fbf59bc9
TH
309 */
310static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
311{
312 int nslot = pcpu_chunk_slot(chunk);
313
edcb4639 314 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
fbf59bc9
TH
315 if (oslot < nslot)
316 list_move(&chunk->list, &pcpu_slot[nslot]);
317 else
318 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
319 }
320}
321
fbf59bc9 322/**
e1b9aa3f
CL
323 * pcpu_chunk_addr_search - determine chunk containing specified address
324 * @addr: address for which the chunk needs to be determined.
ccea34b5 325 *
fbf59bc9
TH
326 * RETURNS:
327 * The address of the found chunk.
328 */
329static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
330{
bba174f5 331 void *first_start = pcpu_first_chunk->base_addr;
fbf59bc9 332
ae9e6bc9 333 /* is it in the first chunk? */
79ba6ac8 334 if (addr >= first_start && addr < first_start + pcpu_unit_size) {
ae9e6bc9
TH
335 /* is it in the reserved area? */
336 if (addr < first_start + pcpu_reserved_chunk_limit)
edcb4639 337 return pcpu_reserved_chunk;
ae9e6bc9 338 return pcpu_first_chunk;
edcb4639
TH
339 }
340
04a13c7c
TH
341 /*
342 * The address is relative to unit0 which might be unused and
343 * thus unmapped. Offset the address to the unit space of the
344 * current processor before looking it up in the vmalloc
345 * space. Note that any possible cpu id can be used here, so
346 * there's no need to worry about preemption or cpu hotplug.
347 */
5579fd7e 348 addr += pcpu_unit_offsets[raw_smp_processor_id()];
e1b9aa3f 349 return pcpu_get_page_chunk(vmalloc_to_page(addr));
fbf59bc9
TH
350}
351
9f7dcf22
TH
352/**
353 * pcpu_extend_area_map - extend area map for allocation
354 * @chunk: target chunk
355 *
356 * Extend area map of @chunk so that it can accomodate an allocation.
357 * A single allocation can split an area into three areas, so this
358 * function makes sure that @chunk->map has at least two extra slots.
359 *
ccea34b5
TH
360 * CONTEXT:
361 * pcpu_alloc_mutex, pcpu_lock. pcpu_lock is released and reacquired
362 * if area map is extended.
363 *
9f7dcf22
TH
364 * RETURNS:
365 * 0 if noop, 1 if successfully extended, -errno on failure.
366 */
367static int pcpu_extend_area_map(struct pcpu_chunk *chunk)
0f5e4816 368 __releases(lock) __acquires(lock)
9f7dcf22
TH
369{
370 int new_alloc;
371 int *new;
372 size_t size;
373
374 /* has enough? */
375 if (chunk->map_alloc >= chunk->map_used + 2)
376 return 0;
377
ccea34b5
TH
378 spin_unlock_irq(&pcpu_lock);
379
9f7dcf22
TH
380 new_alloc = PCPU_DFL_MAP_ALLOC;
381 while (new_alloc < chunk->map_used + 2)
382 new_alloc *= 2;
383
384 new = pcpu_mem_alloc(new_alloc * sizeof(new[0]));
ccea34b5
TH
385 if (!new) {
386 spin_lock_irq(&pcpu_lock);
9f7dcf22 387 return -ENOMEM;
ccea34b5
TH
388 }
389
390 /*
391 * Acquire pcpu_lock and switch to new area map. Only free
392 * could have happened inbetween, so map_used couldn't have
393 * grown.
394 */
395 spin_lock_irq(&pcpu_lock);
396 BUG_ON(new_alloc < chunk->map_used + 2);
9f7dcf22
TH
397
398 size = chunk->map_alloc * sizeof(chunk->map[0]);
399 memcpy(new, chunk->map, size);
400
401 /*
402 * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
403 * one of the first chunks and still using static map.
404 */
405 if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
406 pcpu_mem_free(chunk->map, size);
407
408 chunk->map_alloc = new_alloc;
409 chunk->map = new;
410 return 0;
411}
412
fbf59bc9
TH
413/**
414 * pcpu_split_block - split a map block
415 * @chunk: chunk of interest
416 * @i: index of map block to split
cae3aeb8
TH
417 * @head: head size in bytes (can be 0)
418 * @tail: tail size in bytes (can be 0)
fbf59bc9
TH
419 *
420 * Split the @i'th map block into two or three blocks. If @head is
421 * non-zero, @head bytes block is inserted before block @i moving it
422 * to @i+1 and reducing its size by @head bytes.
423 *
424 * If @tail is non-zero, the target block, which can be @i or @i+1
425 * depending on @head, is reduced by @tail bytes and @tail byte block
426 * is inserted after the target block.
427 *
9f7dcf22 428 * @chunk->map must have enough free slots to accomodate the split.
ccea34b5
TH
429 *
430 * CONTEXT:
431 * pcpu_lock.
fbf59bc9 432 */
9f7dcf22
TH
433static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
434 int head, int tail)
fbf59bc9
TH
435{
436 int nr_extra = !!head + !!tail;
1880d93b 437
9f7dcf22 438 BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
fbf59bc9 439
9f7dcf22 440 /* insert new subblocks */
fbf59bc9
TH
441 memmove(&chunk->map[i + nr_extra], &chunk->map[i],
442 sizeof(chunk->map[0]) * (chunk->map_used - i));
443 chunk->map_used += nr_extra;
444
445 if (head) {
446 chunk->map[i + 1] = chunk->map[i] - head;
447 chunk->map[i++] = head;
448 }
449 if (tail) {
450 chunk->map[i++] -= tail;
451 chunk->map[i] = tail;
452 }
fbf59bc9
TH
453}
454
455/**
456 * pcpu_alloc_area - allocate area from a pcpu_chunk
457 * @chunk: chunk of interest
cae3aeb8 458 * @size: wanted size in bytes
fbf59bc9
TH
459 * @align: wanted align
460 *
461 * Try to allocate @size bytes area aligned at @align from @chunk.
462 * Note that this function only allocates the offset. It doesn't
463 * populate or map the area.
464 *
9f7dcf22
TH
465 * @chunk->map must have at least two free slots.
466 *
ccea34b5
TH
467 * CONTEXT:
468 * pcpu_lock.
469 *
fbf59bc9 470 * RETURNS:
9f7dcf22
TH
471 * Allocated offset in @chunk on success, -1 if no matching area is
472 * found.
fbf59bc9
TH
473 */
474static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
475{
476 int oslot = pcpu_chunk_slot(chunk);
477 int max_contig = 0;
478 int i, off;
479
fbf59bc9
TH
480 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
481 bool is_last = i + 1 == chunk->map_used;
482 int head, tail;
483
484 /* extra for alignment requirement */
485 head = ALIGN(off, align) - off;
486 BUG_ON(i == 0 && head != 0);
487
488 if (chunk->map[i] < 0)
489 continue;
490 if (chunk->map[i] < head + size) {
491 max_contig = max(chunk->map[i], max_contig);
492 continue;
493 }
494
495 /*
496 * If head is small or the previous block is free,
497 * merge'em. Note that 'small' is defined as smaller
498 * than sizeof(int), which is very small but isn't too
499 * uncommon for percpu allocations.
500 */
501 if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
502 if (chunk->map[i - 1] > 0)
503 chunk->map[i - 1] += head;
504 else {
505 chunk->map[i - 1] -= head;
506 chunk->free_size -= head;
507 }
508 chunk->map[i] -= head;
509 off += head;
510 head = 0;
511 }
512
513 /* if tail is small, just keep it around */
514 tail = chunk->map[i] - head - size;
515 if (tail < sizeof(int))
516 tail = 0;
517
518 /* split if warranted */
519 if (head || tail) {
9f7dcf22 520 pcpu_split_block(chunk, i, head, tail);
fbf59bc9
TH
521 if (head) {
522 i++;
523 off += head;
524 max_contig = max(chunk->map[i - 1], max_contig);
525 }
526 if (tail)
527 max_contig = max(chunk->map[i + 1], max_contig);
528 }
529
530 /* update hint and mark allocated */
531 if (is_last)
532 chunk->contig_hint = max_contig; /* fully scanned */
533 else
534 chunk->contig_hint = max(chunk->contig_hint,
535 max_contig);
536
537 chunk->free_size -= chunk->map[i];
538 chunk->map[i] = -chunk->map[i];
539
540 pcpu_chunk_relocate(chunk, oslot);
541 return off;
542 }
543
544 chunk->contig_hint = max_contig; /* fully scanned */
545 pcpu_chunk_relocate(chunk, oslot);
546
9f7dcf22
TH
547 /* tell the upper layer that this chunk has no matching area */
548 return -1;
fbf59bc9
TH
549}
550
551/**
552 * pcpu_free_area - free area to a pcpu_chunk
553 * @chunk: chunk of interest
554 * @freeme: offset of area to free
555 *
556 * Free area starting from @freeme to @chunk. Note that this function
557 * only modifies the allocation map. It doesn't depopulate or unmap
558 * the area.
ccea34b5
TH
559 *
560 * CONTEXT:
561 * pcpu_lock.
fbf59bc9
TH
562 */
563static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
564{
565 int oslot = pcpu_chunk_slot(chunk);
566 int i, off;
567
568 for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
569 if (off == freeme)
570 break;
571 BUG_ON(off != freeme);
572 BUG_ON(chunk->map[i] > 0);
573
574 chunk->map[i] = -chunk->map[i];
575 chunk->free_size += chunk->map[i];
576
577 /* merge with previous? */
578 if (i > 0 && chunk->map[i - 1] >= 0) {
579 chunk->map[i - 1] += chunk->map[i];
580 chunk->map_used--;
581 memmove(&chunk->map[i], &chunk->map[i + 1],
582 (chunk->map_used - i) * sizeof(chunk->map[0]));
583 i--;
584 }
585 /* merge with next? */
586 if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
587 chunk->map[i] += chunk->map[i + 1];
588 chunk->map_used--;
589 memmove(&chunk->map[i + 1], &chunk->map[i + 2],
590 (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
591 }
592
593 chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
594 pcpu_chunk_relocate(chunk, oslot);
595}
596
597/**
ce3141a2 598 * pcpu_get_pages_and_bitmap - get temp pages array and bitmap
fbf59bc9 599 * @chunk: chunk of interest
ce3141a2
TH
600 * @bitmapp: output parameter for bitmap
601 * @may_alloc: may allocate the array
fbf59bc9 602 *
ce3141a2
TH
603 * Returns pointer to array of pointers to struct page and bitmap,
604 * both of which can be indexed with pcpu_page_idx(). The returned
605 * array is cleared to zero and *@bitmapp is copied from
606 * @chunk->populated. Note that there is only one array and bitmap
607 * and access exclusion is the caller's responsibility.
608 *
609 * CONTEXT:
610 * pcpu_alloc_mutex and does GFP_KERNEL allocation if @may_alloc.
611 * Otherwise, don't care.
612 *
613 * RETURNS:
614 * Pointer to temp pages array on success, NULL on failure.
fbf59bc9 615 */
ce3141a2
TH
616static struct page **pcpu_get_pages_and_bitmap(struct pcpu_chunk *chunk,
617 unsigned long **bitmapp,
618 bool may_alloc)
fbf59bc9 619{
ce3141a2
TH
620 static struct page **pages;
621 static unsigned long *bitmap;
2f39e637 622 size_t pages_size = pcpu_nr_units * pcpu_unit_pages * sizeof(pages[0]);
ce3141a2
TH
623 size_t bitmap_size = BITS_TO_LONGS(pcpu_unit_pages) *
624 sizeof(unsigned long);
625
626 if (!pages || !bitmap) {
627 if (may_alloc && !pages)
628 pages = pcpu_mem_alloc(pages_size);
629 if (may_alloc && !bitmap)
630 bitmap = pcpu_mem_alloc(bitmap_size);
631 if (!pages || !bitmap)
632 return NULL;
633 }
fbf59bc9 634
ce3141a2
TH
635 memset(pages, 0, pages_size);
636 bitmap_copy(bitmap, chunk->populated, pcpu_unit_pages);
8d408b4b 637
ce3141a2
TH
638 *bitmapp = bitmap;
639 return pages;
640}
fbf59bc9 641
ce3141a2
TH
642/**
643 * pcpu_free_pages - free pages which were allocated for @chunk
644 * @chunk: chunk pages were allocated for
645 * @pages: array of pages to be freed, indexed by pcpu_page_idx()
646 * @populated: populated bitmap
647 * @page_start: page index of the first page to be freed
648 * @page_end: page index of the last page to be freed + 1
649 *
650 * Free pages [@page_start and @page_end) in @pages for all units.
651 * The pages were allocated for @chunk.
652 */
653static void pcpu_free_pages(struct pcpu_chunk *chunk,
654 struct page **pages, unsigned long *populated,
655 int page_start, int page_end)
656{
657 unsigned int cpu;
658 int i;
659
660 for_each_possible_cpu(cpu) {
661 for (i = page_start; i < page_end; i++) {
662 struct page *page = pages[pcpu_page_idx(cpu, i)];
663
664 if (page)
665 __free_page(page);
666 }
667 }
fbf59bc9
TH
668}
669
670/**
ce3141a2
TH
671 * pcpu_alloc_pages - allocates pages for @chunk
672 * @chunk: target chunk
673 * @pages: array to put the allocated pages into, indexed by pcpu_page_idx()
674 * @populated: populated bitmap
675 * @page_start: page index of the first page to be allocated
676 * @page_end: page index of the last page to be allocated + 1
677 *
678 * Allocate pages [@page_start,@page_end) into @pages for all units.
679 * The allocation is for @chunk. Percpu core doesn't care about the
680 * content of @pages and will pass it verbatim to pcpu_map_pages().
fbf59bc9 681 */
ce3141a2
TH
682static int pcpu_alloc_pages(struct pcpu_chunk *chunk,
683 struct page **pages, unsigned long *populated,
684 int page_start, int page_end)
fbf59bc9 685{
ce3141a2 686 const gfp_t gfp = GFP_KERNEL | __GFP_HIGHMEM | __GFP_COLD;
fbf59bc9
TH
687 unsigned int cpu;
688 int i;
689
ce3141a2
TH
690 for_each_possible_cpu(cpu) {
691 for (i = page_start; i < page_end; i++) {
692 struct page **pagep = &pages[pcpu_page_idx(cpu, i)];
693
694 *pagep = alloc_pages_node(cpu_to_node(cpu), gfp, 0);
695 if (!*pagep) {
696 pcpu_free_pages(chunk, pages, populated,
697 page_start, page_end);
698 return -ENOMEM;
699 }
700 }
701 }
702 return 0;
703}
fbf59bc9 704
ce3141a2
TH
705/**
706 * pcpu_pre_unmap_flush - flush cache prior to unmapping
707 * @chunk: chunk the regions to be flushed belongs to
708 * @page_start: page index of the first page to be flushed
709 * @page_end: page index of the last page to be flushed + 1
710 *
711 * Pages in [@page_start,@page_end) of @chunk are about to be
712 * unmapped. Flush cache. As each flushing trial can be very
713 * expensive, issue flush on the whole region at once rather than
714 * doing it for each cpu. This could be an overkill but is more
715 * scalable.
716 */
717static void pcpu_pre_unmap_flush(struct pcpu_chunk *chunk,
718 int page_start, int page_end)
719{
2f39e637
TH
720 flush_cache_vunmap(
721 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
722 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
ce3141a2
TH
723}
724
725static void __pcpu_unmap_pages(unsigned long addr, int nr_pages)
726{
727 unmap_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT);
728}
fbf59bc9 729
ce3141a2
TH
730/**
731 * pcpu_unmap_pages - unmap pages out of a pcpu_chunk
fbf59bc9 732 * @chunk: chunk of interest
ce3141a2
TH
733 * @pages: pages array which can be used to pass information to free
734 * @populated: populated bitmap
fbf59bc9
TH
735 * @page_start: page index of the first page to unmap
736 * @page_end: page index of the last page to unmap + 1
fbf59bc9
TH
737 *
738 * For each cpu, unmap pages [@page_start,@page_end) out of @chunk.
ce3141a2
TH
739 * Corresponding elements in @pages were cleared by the caller and can
740 * be used to carry information to pcpu_free_pages() which will be
741 * called after all unmaps are finished. The caller should call
742 * proper pre/post flush functions.
fbf59bc9 743 */
ce3141a2
TH
744static void pcpu_unmap_pages(struct pcpu_chunk *chunk,
745 struct page **pages, unsigned long *populated,
746 int page_start, int page_end)
fbf59bc9 747{
fbf59bc9 748 unsigned int cpu;
ce3141a2 749 int i;
fbf59bc9 750
ce3141a2
TH
751 for_each_possible_cpu(cpu) {
752 for (i = page_start; i < page_end; i++) {
753 struct page *page;
fbf59bc9 754
ce3141a2
TH
755 page = pcpu_chunk_page(chunk, cpu, i);
756 WARN_ON(!page);
757 pages[pcpu_page_idx(cpu, i)] = page;
fbf59bc9 758 }
ce3141a2
TH
759 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, cpu, page_start),
760 page_end - page_start);
fbf59bc9
TH
761 }
762
ce3141a2
TH
763 for (i = page_start; i < page_end; i++)
764 __clear_bit(i, populated);
765}
766
767/**
768 * pcpu_post_unmap_tlb_flush - flush TLB after unmapping
769 * @chunk: pcpu_chunk the regions to be flushed belong to
770 * @page_start: page index of the first page to be flushed
771 * @page_end: page index of the last page to be flushed + 1
772 *
773 * Pages [@page_start,@page_end) of @chunk have been unmapped. Flush
774 * TLB for the regions. This can be skipped if the area is to be
775 * returned to vmalloc as vmalloc will handle TLB flushing lazily.
776 *
777 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
778 * for the whole region.
779 */
780static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
781 int page_start, int page_end)
782{
2f39e637
TH
783 flush_tlb_kernel_range(
784 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
785 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
fbf59bc9
TH
786}
787
c8a51be4
TH
788static int __pcpu_map_pages(unsigned long addr, struct page **pages,
789 int nr_pages)
790{
791 return map_kernel_range_noflush(addr, nr_pages << PAGE_SHIFT,
792 PAGE_KERNEL, pages);
fbf59bc9
TH
793}
794
795/**
ce3141a2 796 * pcpu_map_pages - map pages into a pcpu_chunk
fbf59bc9 797 * @chunk: chunk of interest
ce3141a2
TH
798 * @pages: pages array containing pages to be mapped
799 * @populated: populated bitmap
fbf59bc9
TH
800 * @page_start: page index of the first page to map
801 * @page_end: page index of the last page to map + 1
802 *
ce3141a2
TH
803 * For each cpu, map pages [@page_start,@page_end) into @chunk. The
804 * caller is responsible for calling pcpu_post_map_flush() after all
805 * mappings are complete.
806 *
807 * This function is responsible for setting corresponding bits in
808 * @chunk->populated bitmap and whatever is necessary for reverse
809 * lookup (addr -> chunk).
fbf59bc9 810 */
ce3141a2
TH
811static int pcpu_map_pages(struct pcpu_chunk *chunk,
812 struct page **pages, unsigned long *populated,
813 int page_start, int page_end)
fbf59bc9 814{
ce3141a2
TH
815 unsigned int cpu, tcpu;
816 int i, err;
8d408b4b 817
fbf59bc9 818 for_each_possible_cpu(cpu) {
c8a51be4 819 err = __pcpu_map_pages(pcpu_chunk_addr(chunk, cpu, page_start),
ce3141a2 820 &pages[pcpu_page_idx(cpu, page_start)],
c8a51be4 821 page_end - page_start);
fbf59bc9 822 if (err < 0)
ce3141a2 823 goto err;
c8a51be4
TH
824 }
825
ce3141a2
TH
826 /* mapping successful, link chunk and mark populated */
827 for (i = page_start; i < page_end; i++) {
828 for_each_possible_cpu(cpu)
829 pcpu_set_page_chunk(pages[pcpu_page_idx(cpu, i)],
830 chunk);
831 __set_bit(i, populated);
fbf59bc9
TH
832 }
833
fbf59bc9 834 return 0;
ce3141a2
TH
835
836err:
837 for_each_possible_cpu(tcpu) {
838 if (tcpu == cpu)
839 break;
840 __pcpu_unmap_pages(pcpu_chunk_addr(chunk, tcpu, page_start),
841 page_end - page_start);
842 }
843 return err;
844}
845
846/**
847 * pcpu_post_map_flush - flush cache after mapping
848 * @chunk: pcpu_chunk the regions to be flushed belong to
849 * @page_start: page index of the first page to be flushed
850 * @page_end: page index of the last page to be flushed + 1
851 *
852 * Pages [@page_start,@page_end) of @chunk have been mapped. Flush
853 * cache.
854 *
855 * As with pcpu_pre_unmap_flush(), TLB flushing also is done at once
856 * for the whole region.
857 */
858static void pcpu_post_map_flush(struct pcpu_chunk *chunk,
859 int page_start, int page_end)
860{
2f39e637
TH
861 flush_cache_vmap(
862 pcpu_chunk_addr(chunk, pcpu_first_unit_cpu, page_start),
863 pcpu_chunk_addr(chunk, pcpu_last_unit_cpu, page_end));
c8a51be4
TH
864}
865
fbf59bc9
TH
866/**
867 * pcpu_depopulate_chunk - depopulate and unmap an area of a pcpu_chunk
868 * @chunk: chunk to depopulate
869 * @off: offset to the area to depopulate
cae3aeb8 870 * @size: size of the area to depopulate in bytes
fbf59bc9
TH
871 * @flush: whether to flush cache and tlb or not
872 *
873 * For each cpu, depopulate and unmap pages [@page_start,@page_end)
874 * from @chunk. If @flush is true, vcache is flushed before unmapping
875 * and tlb after.
ccea34b5
TH
876 *
877 * CONTEXT:
878 * pcpu_alloc_mutex.
fbf59bc9 879 */
ce3141a2 880static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size)
fbf59bc9
TH
881{
882 int page_start = PFN_DOWN(off);
883 int page_end = PFN_UP(off + size);
ce3141a2
TH
884 struct page **pages;
885 unsigned long *populated;
886 int rs, re;
887
888 /* quick path, check whether it's empty already */
889 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
890 if (rs == page_start && re == page_end)
891 return;
892 break;
893 }
fbf59bc9 894
ce3141a2
TH
895 /* immutable chunks can't be depopulated */
896 WARN_ON(chunk->immutable);
fbf59bc9 897
ce3141a2
TH
898 /*
899 * If control reaches here, there must have been at least one
900 * successful population attempt so the temp pages array must
901 * be available now.
902 */
903 pages = pcpu_get_pages_and_bitmap(chunk, &populated, false);
904 BUG_ON(!pages);
fbf59bc9 905
ce3141a2
TH
906 /* unmap and free */
907 pcpu_pre_unmap_flush(chunk, page_start, page_end);
fbf59bc9 908
ce3141a2
TH
909 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
910 pcpu_unmap_pages(chunk, pages, populated, rs, re);
fbf59bc9 911
ce3141a2
TH
912 /* no need to flush tlb, vmalloc will handle it lazily */
913
914 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end)
915 pcpu_free_pages(chunk, pages, populated, rs, re);
fbf59bc9 916
ce3141a2
TH
917 /* commit new bitmap */
918 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
fbf59bc9
TH
919}
920
921/**
922 * pcpu_populate_chunk - populate and map an area of a pcpu_chunk
923 * @chunk: chunk of interest
924 * @off: offset to the area to populate
cae3aeb8 925 * @size: size of the area to populate in bytes
fbf59bc9
TH
926 *
927 * For each cpu, populate and map pages [@page_start,@page_end) into
928 * @chunk. The area is cleared on return.
ccea34b5
TH
929 *
930 * CONTEXT:
931 * pcpu_alloc_mutex, does GFP_KERNEL allocation.
fbf59bc9
TH
932 */
933static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size)
934{
fbf59bc9
TH
935 int page_start = PFN_DOWN(off);
936 int page_end = PFN_UP(off + size);
ce3141a2
TH
937 int free_end = page_start, unmap_end = page_start;
938 struct page **pages;
939 unsigned long *populated;
fbf59bc9 940 unsigned int cpu;
ce3141a2 941 int rs, re, rc;
fbf59bc9 942
ce3141a2
TH
943 /* quick path, check whether all pages are already there */
944 pcpu_for_each_pop_region(chunk, rs, re, page_start, page_end) {
945 if (rs == page_start && re == page_end)
946 goto clear;
947 break;
948 }
fbf59bc9 949
ce3141a2
TH
950 /* need to allocate and map pages, this chunk can't be immutable */
951 WARN_ON(chunk->immutable);
fbf59bc9 952
ce3141a2
TH
953 pages = pcpu_get_pages_and_bitmap(chunk, &populated, true);
954 if (!pages)
955 return -ENOMEM;
fbf59bc9 956
ce3141a2
TH
957 /* alloc and map */
958 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
959 rc = pcpu_alloc_pages(chunk, pages, populated, rs, re);
960 if (rc)
961 goto err_free;
962 free_end = re;
fbf59bc9
TH
963 }
964
ce3141a2
TH
965 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
966 rc = pcpu_map_pages(chunk, pages, populated, rs, re);
967 if (rc)
968 goto err_unmap;
969 unmap_end = re;
970 }
971 pcpu_post_map_flush(chunk, page_start, page_end);
fbf59bc9 972
ce3141a2
TH
973 /* commit new bitmap */
974 bitmap_copy(chunk->populated, populated, pcpu_unit_pages);
975clear:
fbf59bc9 976 for_each_possible_cpu(cpu)
2f39e637 977 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
fbf59bc9 978 return 0;
ce3141a2
TH
979
980err_unmap:
981 pcpu_pre_unmap_flush(chunk, page_start, unmap_end);
982 pcpu_for_each_unpop_region(chunk, rs, re, page_start, unmap_end)
983 pcpu_unmap_pages(chunk, pages, populated, rs, re);
984 pcpu_post_unmap_tlb_flush(chunk, page_start, unmap_end);
985err_free:
986 pcpu_for_each_unpop_region(chunk, rs, re, page_start, free_end)
987 pcpu_free_pages(chunk, pages, populated, rs, re);
988 return rc;
fbf59bc9
TH
989}
990
991static void free_pcpu_chunk(struct pcpu_chunk *chunk)
992{
993 if (!chunk)
994 return;
6563297c
TH
995 if (chunk->vms)
996 pcpu_free_vm_areas(chunk->vms, pcpu_nr_groups);
1880d93b 997 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
fbf59bc9
TH
998 kfree(chunk);
999}
1000
1001static struct pcpu_chunk *alloc_pcpu_chunk(void)
1002{
1003 struct pcpu_chunk *chunk;
1004
1005 chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
1006 if (!chunk)
1007 return NULL;
1008
1880d93b 1009 chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
fbf59bc9
TH
1010 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
1011 chunk->map[chunk->map_used++] = pcpu_unit_size;
1012
6563297c
TH
1013 chunk->vms = pcpu_get_vm_areas(pcpu_group_offsets, pcpu_group_sizes,
1014 pcpu_nr_groups, pcpu_atom_size,
1015 GFP_KERNEL);
1016 if (!chunk->vms) {
fbf59bc9
TH
1017 free_pcpu_chunk(chunk);
1018 return NULL;
1019 }
1020
1021 INIT_LIST_HEAD(&chunk->list);
1022 chunk->free_size = pcpu_unit_size;
1023 chunk->contig_hint = pcpu_unit_size;
6563297c 1024 chunk->base_addr = chunk->vms[0]->addr - pcpu_group_offsets[0];
fbf59bc9
TH
1025
1026 return chunk;
1027}
1028
1029/**
edcb4639 1030 * pcpu_alloc - the percpu allocator
cae3aeb8 1031 * @size: size of area to allocate in bytes
fbf59bc9 1032 * @align: alignment of area (max PAGE_SIZE)
edcb4639 1033 * @reserved: allocate from the reserved chunk if available
fbf59bc9 1034 *
ccea34b5
TH
1035 * Allocate percpu area of @size bytes aligned at @align.
1036 *
1037 * CONTEXT:
1038 * Does GFP_KERNEL allocation.
fbf59bc9
TH
1039 *
1040 * RETURNS:
1041 * Percpu pointer to the allocated area on success, NULL on failure.
1042 */
edcb4639 1043static void *pcpu_alloc(size_t size, size_t align, bool reserved)
fbf59bc9 1044{
f2badb0c 1045 static int warn_limit = 10;
fbf59bc9 1046 struct pcpu_chunk *chunk;
f2badb0c 1047 const char *err;
fbf59bc9
TH
1048 int slot, off;
1049
8d408b4b 1050 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
fbf59bc9
TH
1051 WARN(true, "illegal size (%zu) or align (%zu) for "
1052 "percpu allocation\n", size, align);
1053 return NULL;
1054 }
1055
ccea34b5
TH
1056 mutex_lock(&pcpu_alloc_mutex);
1057 spin_lock_irq(&pcpu_lock);
fbf59bc9 1058
edcb4639
TH
1059 /* serve reserved allocations from the reserved chunk if available */
1060 if (reserved && pcpu_reserved_chunk) {
1061 chunk = pcpu_reserved_chunk;
9f7dcf22 1062 if (size > chunk->contig_hint ||
f2badb0c
TH
1063 pcpu_extend_area_map(chunk) < 0) {
1064 err = "failed to extend area map of reserved chunk";
ccea34b5 1065 goto fail_unlock;
f2badb0c 1066 }
edcb4639
TH
1067 off = pcpu_alloc_area(chunk, size, align);
1068 if (off >= 0)
1069 goto area_found;
f2badb0c 1070 err = "alloc from reserved chunk failed";
ccea34b5 1071 goto fail_unlock;
edcb4639
TH
1072 }
1073
ccea34b5 1074restart:
edcb4639 1075 /* search through normal chunks */
fbf59bc9
TH
1076 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1077 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1078 if (size > chunk->contig_hint)
1079 continue;
ccea34b5
TH
1080
1081 switch (pcpu_extend_area_map(chunk)) {
1082 case 0:
1083 break;
1084 case 1:
1085 goto restart; /* pcpu_lock dropped, restart */
1086 default:
f2badb0c 1087 err = "failed to extend area map";
ccea34b5
TH
1088 goto fail_unlock;
1089 }
1090
fbf59bc9
TH
1091 off = pcpu_alloc_area(chunk, size, align);
1092 if (off >= 0)
1093 goto area_found;
fbf59bc9
TH
1094 }
1095 }
1096
1097 /* hmmm... no space left, create a new chunk */
ccea34b5
TH
1098 spin_unlock_irq(&pcpu_lock);
1099
fbf59bc9 1100 chunk = alloc_pcpu_chunk();
f2badb0c
TH
1101 if (!chunk) {
1102 err = "failed to allocate new chunk";
ccea34b5 1103 goto fail_unlock_mutex;
f2badb0c 1104 }
ccea34b5
TH
1105
1106 spin_lock_irq(&pcpu_lock);
fbf59bc9 1107 pcpu_chunk_relocate(chunk, -1);
ccea34b5 1108 goto restart;
fbf59bc9
TH
1109
1110area_found:
ccea34b5
TH
1111 spin_unlock_irq(&pcpu_lock);
1112
fbf59bc9
TH
1113 /* populate, map and clear the area */
1114 if (pcpu_populate_chunk(chunk, off, size)) {
ccea34b5 1115 spin_lock_irq(&pcpu_lock);
fbf59bc9 1116 pcpu_free_area(chunk, off);
f2badb0c 1117 err = "failed to populate";
ccea34b5 1118 goto fail_unlock;
fbf59bc9
TH
1119 }
1120
ccea34b5
TH
1121 mutex_unlock(&pcpu_alloc_mutex);
1122
bba174f5
TH
1123 /* return address relative to base address */
1124 return __addr_to_pcpu_ptr(chunk->base_addr + off);
ccea34b5
TH
1125
1126fail_unlock:
1127 spin_unlock_irq(&pcpu_lock);
1128fail_unlock_mutex:
1129 mutex_unlock(&pcpu_alloc_mutex);
f2badb0c
TH
1130 if (warn_limit) {
1131 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
1132 "%s\n", size, align, err);
1133 dump_stack();
1134 if (!--warn_limit)
1135 pr_info("PERCPU: limit reached, disable warning\n");
1136 }
ccea34b5 1137 return NULL;
fbf59bc9 1138}
edcb4639
TH
1139
1140/**
1141 * __alloc_percpu - allocate dynamic percpu area
1142 * @size: size of area to allocate in bytes
1143 * @align: alignment of area (max PAGE_SIZE)
1144 *
1145 * Allocate percpu area of @size bytes aligned at @align. Might
1146 * sleep. Might trigger writeouts.
1147 *
ccea34b5
TH
1148 * CONTEXT:
1149 * Does GFP_KERNEL allocation.
1150 *
edcb4639
TH
1151 * RETURNS:
1152 * Percpu pointer to the allocated area on success, NULL on failure.
1153 */
1154void *__alloc_percpu(size_t size, size_t align)
1155{
1156 return pcpu_alloc(size, align, false);
1157}
fbf59bc9
TH
1158EXPORT_SYMBOL_GPL(__alloc_percpu);
1159
edcb4639
TH
1160/**
1161 * __alloc_reserved_percpu - allocate reserved percpu area
1162 * @size: size of area to allocate in bytes
1163 * @align: alignment of area (max PAGE_SIZE)
1164 *
1165 * Allocate percpu area of @size bytes aligned at @align from reserved
1166 * percpu area if arch has set it up; otherwise, allocation is served
1167 * from the same dynamic area. Might sleep. Might trigger writeouts.
1168 *
ccea34b5
TH
1169 * CONTEXT:
1170 * Does GFP_KERNEL allocation.
1171 *
edcb4639
TH
1172 * RETURNS:
1173 * Percpu pointer to the allocated area on success, NULL on failure.
1174 */
1175void *__alloc_reserved_percpu(size_t size, size_t align)
1176{
1177 return pcpu_alloc(size, align, true);
1178}
1179
a56dbddf
TH
1180/**
1181 * pcpu_reclaim - reclaim fully free chunks, workqueue function
1182 * @work: unused
1183 *
1184 * Reclaim all fully free chunks except for the first one.
ccea34b5
TH
1185 *
1186 * CONTEXT:
1187 * workqueue context.
a56dbddf
TH
1188 */
1189static void pcpu_reclaim(struct work_struct *work)
fbf59bc9 1190{
a56dbddf
TH
1191 LIST_HEAD(todo);
1192 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
1193 struct pcpu_chunk *chunk, *next;
1194
ccea34b5
TH
1195 mutex_lock(&pcpu_alloc_mutex);
1196 spin_lock_irq(&pcpu_lock);
a56dbddf
TH
1197
1198 list_for_each_entry_safe(chunk, next, head, list) {
1199 WARN_ON(chunk->immutable);
1200
1201 /* spare the first one */
1202 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
1203 continue;
1204
a56dbddf
TH
1205 list_move(&chunk->list, &todo);
1206 }
1207
ccea34b5 1208 spin_unlock_irq(&pcpu_lock);
a56dbddf
TH
1209
1210 list_for_each_entry_safe(chunk, next, &todo, list) {
ce3141a2 1211 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
a56dbddf
TH
1212 free_pcpu_chunk(chunk);
1213 }
971f3918
TH
1214
1215 mutex_unlock(&pcpu_alloc_mutex);
fbf59bc9
TH
1216}
1217
1218/**
1219 * free_percpu - free percpu area
1220 * @ptr: pointer to area to free
1221 *
ccea34b5
TH
1222 * Free percpu area @ptr.
1223 *
1224 * CONTEXT:
1225 * Can be called from atomic context.
fbf59bc9
TH
1226 */
1227void free_percpu(void *ptr)
1228{
1229 void *addr = __pcpu_ptr_to_addr(ptr);
1230 struct pcpu_chunk *chunk;
ccea34b5 1231 unsigned long flags;
fbf59bc9
TH
1232 int off;
1233
1234 if (!ptr)
1235 return;
1236
ccea34b5 1237 spin_lock_irqsave(&pcpu_lock, flags);
fbf59bc9
TH
1238
1239 chunk = pcpu_chunk_addr_search(addr);
bba174f5 1240 off = addr - chunk->base_addr;
fbf59bc9
TH
1241
1242 pcpu_free_area(chunk, off);
1243
a56dbddf 1244 /* if there are more than one fully free chunks, wake up grim reaper */
fbf59bc9
TH
1245 if (chunk->free_size == pcpu_unit_size) {
1246 struct pcpu_chunk *pos;
1247
a56dbddf 1248 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
fbf59bc9 1249 if (pos != chunk) {
a56dbddf 1250 schedule_work(&pcpu_reclaim_work);
fbf59bc9
TH
1251 break;
1252 }
1253 }
1254
ccea34b5 1255 spin_unlock_irqrestore(&pcpu_lock, flags);
fbf59bc9
TH
1256}
1257EXPORT_SYMBOL_GPL(free_percpu);
1258
033e48fb
TH
1259static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1260 size_t reserved_size,
1261 ssize_t *dyn_sizep)
1262{
1263 size_t size_sum;
1264
1265 size_sum = PFN_ALIGN(static_size + reserved_size +
1266 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1267 if (*dyn_sizep != 0)
1268 *dyn_sizep = size_sum - static_size - reserved_size;
1269
1270 return size_sum;
1271}
1272
fbf59bc9 1273/**
fd1e8a1f
TH
1274 * pcpu_alloc_alloc_info - allocate percpu allocation info
1275 * @nr_groups: the number of groups
1276 * @nr_units: the number of units
1277 *
1278 * Allocate ai which is large enough for @nr_groups groups containing
1279 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1280 * cpu_map array which is long enough for @nr_units and filled with
1281 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1282 * pointer of other groups.
1283 *
1284 * RETURNS:
1285 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1286 * failure.
1287 */
1288struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1289 int nr_units)
1290{
1291 struct pcpu_alloc_info *ai;
1292 size_t base_size, ai_size;
1293 void *ptr;
1294 int unit;
1295
1296 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1297 __alignof__(ai->groups[0].cpu_map[0]));
1298 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1299
1300 ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1301 if (!ptr)
1302 return NULL;
1303 ai = ptr;
1304 ptr += base_size;
1305
1306 ai->groups[0].cpu_map = ptr;
1307
1308 for (unit = 0; unit < nr_units; unit++)
1309 ai->groups[0].cpu_map[unit] = NR_CPUS;
1310
1311 ai->nr_groups = nr_groups;
1312 ai->__ai_size = PFN_ALIGN(ai_size);
1313
1314 return ai;
1315}
1316
1317/**
1318 * pcpu_free_alloc_info - free percpu allocation info
1319 * @ai: pcpu_alloc_info to free
1320 *
1321 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1322 */
1323void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1324{
1325 free_bootmem(__pa(ai), ai->__ai_size);
1326}
1327
1328/**
1329 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
edcb4639 1330 * @reserved_size: the size of reserved percpu area in bytes
cafe8816 1331 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
fd1e8a1f
TH
1332 * @atom_size: allocation atom size
1333 * @cpu_distance_fn: callback to determine distance between cpus, optional
033e48fb 1334 *
fd1e8a1f
TH
1335 * This function determines grouping of units, their mappings to cpus
1336 * and other parameters considering needed percpu size, allocation
1337 * atom size and distances between CPUs.
033e48fb 1338 *
fd1e8a1f
TH
1339 * Groups are always mutliples of atom size and CPUs which are of
1340 * LOCAL_DISTANCE both ways are grouped together and share space for
1341 * units in the same group. The returned configuration is guaranteed
1342 * to have CPUs on different nodes on different groups and >=75% usage
1343 * of allocated virtual address space.
033e48fb
TH
1344 *
1345 * RETURNS:
fd1e8a1f
TH
1346 * On success, pointer to the new allocation_info is returned. On
1347 * failure, ERR_PTR value is returned.
033e48fb 1348 */
fd1e8a1f
TH
1349struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1350 size_t reserved_size, ssize_t dyn_size,
1351 size_t atom_size,
1352 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
033e48fb
TH
1353{
1354 static int group_map[NR_CPUS] __initdata;
1355 static int group_cnt[NR_CPUS] __initdata;
1356 const size_t static_size = __per_cpu_end - __per_cpu_start;
fd1e8a1f 1357 int group_cnt_max = 0, nr_groups = 1, nr_units = 0;
033e48fb
TH
1358 size_t size_sum, min_unit_size, alloc_size;
1359 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
fd1e8a1f 1360 int last_allocs, group, unit;
033e48fb 1361 unsigned int cpu, tcpu;
fd1e8a1f
TH
1362 struct pcpu_alloc_info *ai;
1363 unsigned int *cpu_map;
033e48fb 1364
fb59e72e
TH
1365 /* this function may be called multiple times */
1366 memset(group_map, 0, sizeof(group_map));
1367 memset(group_cnt, 0, sizeof(group_map));
1368
033e48fb
TH
1369 /*
1370 * Determine min_unit_size, alloc_size and max_upa such that
fd1e8a1f 1371 * alloc_size is multiple of atom_size and is the smallest
033e48fb
TH
1372 * which can accomodate 4k aligned segments which are equal to
1373 * or larger than min_unit_size.
1374 */
fd1e8a1f 1375 size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
033e48fb
TH
1376 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1377
fd1e8a1f 1378 alloc_size = roundup(min_unit_size, atom_size);
033e48fb
TH
1379 upa = alloc_size / min_unit_size;
1380 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1381 upa--;
1382 max_upa = upa;
1383
1384 /* group cpus according to their proximity */
1385 for_each_possible_cpu(cpu) {
1386 group = 0;
1387 next_group:
1388 for_each_possible_cpu(tcpu) {
1389 if (cpu == tcpu)
1390 break;
fd1e8a1f 1391 if (group_map[tcpu] == group && cpu_distance_fn &&
033e48fb
TH
1392 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1393 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1394 group++;
fd1e8a1f 1395 nr_groups = max(nr_groups, group + 1);
033e48fb
TH
1396 goto next_group;
1397 }
1398 }
1399 group_map[cpu] = group;
1400 group_cnt[group]++;
1401 group_cnt_max = max(group_cnt_max, group_cnt[group]);
1402 }
1403
1404 /*
1405 * Expand unit size until address space usage goes over 75%
1406 * and then as much as possible without using more address
1407 * space.
1408 */
1409 last_allocs = INT_MAX;
1410 for (upa = max_upa; upa; upa--) {
1411 int allocs = 0, wasted = 0;
1412
1413 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1414 continue;
1415
fd1e8a1f 1416 for (group = 0; group < nr_groups; group++) {
033e48fb
TH
1417 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1418 allocs += this_allocs;
1419 wasted += this_allocs * upa - group_cnt[group];
1420 }
1421
1422 /*
1423 * Don't accept if wastage is over 25%. The
1424 * greater-than comparison ensures upa==1 always
1425 * passes the following check.
1426 */
1427 if (wasted > num_possible_cpus() / 3)
1428 continue;
1429
1430 /* and then don't consume more memory */
1431 if (allocs > last_allocs)
1432 break;
1433 last_allocs = allocs;
1434 best_upa = upa;
1435 }
fd1e8a1f
TH
1436 upa = best_upa;
1437
1438 /* allocate and fill alloc_info */
1439 for (group = 0; group < nr_groups; group++)
1440 nr_units += roundup(group_cnt[group], upa);
1441
1442 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1443 if (!ai)
1444 return ERR_PTR(-ENOMEM);
1445 cpu_map = ai->groups[0].cpu_map;
1446
1447 for (group = 0; group < nr_groups; group++) {
1448 ai->groups[group].cpu_map = cpu_map;
1449 cpu_map += roundup(group_cnt[group], upa);
1450 }
1451
1452 ai->static_size = static_size;
1453 ai->reserved_size = reserved_size;
1454 ai->dyn_size = dyn_size;
1455 ai->unit_size = alloc_size / upa;
1456 ai->atom_size = atom_size;
1457 ai->alloc_size = alloc_size;
1458
1459 for (group = 0, unit = 0; group_cnt[group]; group++) {
1460 struct pcpu_group_info *gi = &ai->groups[group];
1461
1462 /*
1463 * Initialize base_offset as if all groups are located
1464 * back-to-back. The caller should update this to
1465 * reflect actual allocation.
1466 */
1467 gi->base_offset = unit * ai->unit_size;
033e48fb 1468
033e48fb
TH
1469 for_each_possible_cpu(cpu)
1470 if (group_map[cpu] == group)
fd1e8a1f
TH
1471 gi->cpu_map[gi->nr_units++] = cpu;
1472 gi->nr_units = roundup(gi->nr_units, upa);
1473 unit += gi->nr_units;
033e48fb 1474 }
fd1e8a1f 1475 BUG_ON(unit != nr_units);
033e48fb 1476
fd1e8a1f 1477 return ai;
033e48fb
TH
1478}
1479
fd1e8a1f
TH
1480/**
1481 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1482 * @lvl: loglevel
1483 * @ai: allocation info to dump
1484 *
1485 * Print out information about @ai using loglevel @lvl.
1486 */
1487static void pcpu_dump_alloc_info(const char *lvl,
1488 const struct pcpu_alloc_info *ai)
033e48fb 1489{
fd1e8a1f 1490 int group_width = 1, cpu_width = 1, width;
033e48fb 1491 char empty_str[] = "--------";
fd1e8a1f
TH
1492 int alloc = 0, alloc_end = 0;
1493 int group, v;
1494 int upa, apl; /* units per alloc, allocs per line */
1495
1496 v = ai->nr_groups;
1497 while (v /= 10)
1498 group_width++;
033e48fb 1499
fd1e8a1f 1500 v = num_possible_cpus();
033e48fb 1501 while (v /= 10)
fd1e8a1f
TH
1502 cpu_width++;
1503 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
033e48fb 1504
fd1e8a1f
TH
1505 upa = ai->alloc_size / ai->unit_size;
1506 width = upa * (cpu_width + 1) + group_width + 3;
1507 apl = rounddown_pow_of_two(max(60 / width, 1));
033e48fb 1508
fd1e8a1f
TH
1509 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1510 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1511 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
033e48fb 1512
fd1e8a1f
TH
1513 for (group = 0; group < ai->nr_groups; group++) {
1514 const struct pcpu_group_info *gi = &ai->groups[group];
1515 int unit = 0, unit_end = 0;
1516
1517 BUG_ON(gi->nr_units % upa);
1518 for (alloc_end += gi->nr_units / upa;
1519 alloc < alloc_end; alloc++) {
1520 if (!(alloc % apl)) {
033e48fb 1521 printk("\n");
fd1e8a1f
TH
1522 printk("%spcpu-alloc: ", lvl);
1523 }
1524 printk("[%0*d] ", group_width, group);
1525
1526 for (unit_end += upa; unit < unit_end; unit++)
1527 if (gi->cpu_map[unit] != NR_CPUS)
1528 printk("%0*d ", cpu_width,
1529 gi->cpu_map[unit]);
1530 else
1531 printk("%s ", empty_str);
033e48fb 1532 }
033e48fb
TH
1533 }
1534 printk("\n");
1535}
033e48fb 1536
fbf59bc9 1537/**
8d408b4b 1538 * pcpu_setup_first_chunk - initialize the first percpu chunk
fd1e8a1f 1539 * @ai: pcpu_alloc_info describing how to percpu area is shaped
38a6be52 1540 * @base_addr: mapped address
8d408b4b
TH
1541 *
1542 * Initialize the first percpu chunk which contains the kernel static
1543 * perpcu area. This function is to be called from arch percpu area
38a6be52 1544 * setup path.
8d408b4b 1545 *
fd1e8a1f
TH
1546 * @ai contains all information necessary to initialize the first
1547 * chunk and prime the dynamic percpu allocator.
1548 *
1549 * @ai->static_size is the size of static percpu area.
1550 *
1551 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
edcb4639
TH
1552 * reserve after the static area in the first chunk. This reserves
1553 * the first chunk such that it's available only through reserved
1554 * percpu allocation. This is primarily used to serve module percpu
1555 * static areas on architectures where the addressing model has
1556 * limited offset range for symbol relocations to guarantee module
1557 * percpu symbols fall inside the relocatable range.
1558 *
fd1e8a1f
TH
1559 * @ai->dyn_size determines the number of bytes available for dynamic
1560 * allocation in the first chunk. The area between @ai->static_size +
1561 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
6074d5b0 1562 *
fd1e8a1f
TH
1563 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1564 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1565 * @ai->dyn_size.
8d408b4b 1566 *
fd1e8a1f
TH
1567 * @ai->atom_size is the allocation atom size and used as alignment
1568 * for vm areas.
8d408b4b 1569 *
fd1e8a1f
TH
1570 * @ai->alloc_size is the allocation size and always multiple of
1571 * @ai->atom_size. This is larger than @ai->atom_size if
1572 * @ai->unit_size is larger than @ai->atom_size.
1573 *
1574 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1575 * percpu areas. Units which should be colocated are put into the
1576 * same group. Dynamic VM areas will be allocated according to these
1577 * groupings. If @ai->nr_groups is zero, a single group containing
1578 * all units is assumed.
8d408b4b 1579 *
38a6be52
TH
1580 * The caller should have mapped the first chunk at @base_addr and
1581 * copied static data to each unit.
fbf59bc9 1582 *
edcb4639
TH
1583 * If the first chunk ends up with both reserved and dynamic areas, it
1584 * is served by two chunks - one to serve the core static and reserved
1585 * areas and the other for the dynamic area. They share the same vm
1586 * and page map but uses different area allocation map to stay away
1587 * from each other. The latter chunk is circulated in the chunk slots
1588 * and available for dynamic allocation like any other chunks.
1589 *
fbf59bc9 1590 * RETURNS:
fb435d52 1591 * 0 on success, -errno on failure.
fbf59bc9 1592 */
fb435d52
TH
1593int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1594 void *base_addr)
fbf59bc9 1595{
635b75fc 1596 static char cpus_buf[4096] __initdata;
edcb4639 1597 static int smap[2], dmap[2];
fd1e8a1f
TH
1598 size_t dyn_size = ai->dyn_size;
1599 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
edcb4639 1600 struct pcpu_chunk *schunk, *dchunk = NULL;
6563297c
TH
1601 unsigned long *group_offsets;
1602 size_t *group_sizes;
fb435d52 1603 unsigned long *unit_off;
fbf59bc9 1604 unsigned int cpu;
fd1e8a1f
TH
1605 int *unit_map;
1606 int group, unit, i;
fbf59bc9 1607
635b75fc
TH
1608 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1609
1610#define PCPU_SETUP_BUG_ON(cond) do { \
1611 if (unlikely(cond)) { \
1612 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1613 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1614 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1615 BUG(); \
1616 } \
1617} while (0)
1618
2f39e637 1619 /* sanity checks */
edcb4639
TH
1620 BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1621 ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
635b75fc
TH
1622 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1623 PCPU_SETUP_BUG_ON(!ai->static_size);
1624 PCPU_SETUP_BUG_ON(!base_addr);
1625 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1626 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1627 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
8d408b4b 1628
6563297c
TH
1629 /* process group information and build config tables accordingly */
1630 group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1631 group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
fd1e8a1f 1632 unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
fb435d52 1633 unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
2f39e637 1634
fd1e8a1f 1635 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
ffe0d5a5 1636 unit_map[cpu] = UINT_MAX;
fd1e8a1f 1637 pcpu_first_unit_cpu = NR_CPUS;
2f39e637 1638
fd1e8a1f
TH
1639 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1640 const struct pcpu_group_info *gi = &ai->groups[group];
2f39e637 1641
6563297c
TH
1642 group_offsets[group] = gi->base_offset;
1643 group_sizes[group] = gi->nr_units * ai->unit_size;
1644
fd1e8a1f
TH
1645 for (i = 0; i < gi->nr_units; i++) {
1646 cpu = gi->cpu_map[i];
1647 if (cpu == NR_CPUS)
1648 continue;
8d408b4b 1649
635b75fc
TH
1650 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1651 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1652 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
fbf59bc9 1653
fd1e8a1f 1654 unit_map[cpu] = unit + i;
fb435d52
TH
1655 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1656
fd1e8a1f
TH
1657 if (pcpu_first_unit_cpu == NR_CPUS)
1658 pcpu_first_unit_cpu = cpu;
1659 }
2f39e637 1660 }
fd1e8a1f
TH
1661 pcpu_last_unit_cpu = cpu;
1662 pcpu_nr_units = unit;
1663
1664 for_each_possible_cpu(cpu)
635b75fc
TH
1665 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1666
1667 /* we're done parsing the input, undefine BUG macro and dump config */
1668#undef PCPU_SETUP_BUG_ON
1669 pcpu_dump_alloc_info(KERN_INFO, ai);
fd1e8a1f 1670
6563297c
TH
1671 pcpu_nr_groups = ai->nr_groups;
1672 pcpu_group_offsets = group_offsets;
1673 pcpu_group_sizes = group_sizes;
fd1e8a1f 1674 pcpu_unit_map = unit_map;
fb435d52 1675 pcpu_unit_offsets = unit_off;
2f39e637
TH
1676
1677 /* determine basic parameters */
fd1e8a1f 1678 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
d9b55eeb 1679 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
6563297c 1680 pcpu_atom_size = ai->atom_size;
ce3141a2
TH
1681 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1682 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
cafe8816 1683
d9b55eeb
TH
1684 /*
1685 * Allocate chunk slots. The additional last slot is for
1686 * empty chunks.
1687 */
1688 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
fbf59bc9
TH
1689 pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1690 for (i = 0; i < pcpu_nr_slots; i++)
1691 INIT_LIST_HEAD(&pcpu_slot[i]);
1692
edcb4639
TH
1693 /*
1694 * Initialize static chunk. If reserved_size is zero, the
1695 * static chunk covers static area + dynamic allocation area
1696 * in the first chunk. If reserved_size is not zero, it
1697 * covers static area + reserved area (mostly used for module
1698 * static percpu allocation).
1699 */
2441d15c
TH
1700 schunk = alloc_bootmem(pcpu_chunk_struct_size);
1701 INIT_LIST_HEAD(&schunk->list);
bba174f5 1702 schunk->base_addr = base_addr;
61ace7fa
TH
1703 schunk->map = smap;
1704 schunk->map_alloc = ARRAY_SIZE(smap);
38a6be52 1705 schunk->immutable = true;
ce3141a2 1706 bitmap_fill(schunk->populated, pcpu_unit_pages);
edcb4639 1707
fd1e8a1f
TH
1708 if (ai->reserved_size) {
1709 schunk->free_size = ai->reserved_size;
ae9e6bc9 1710 pcpu_reserved_chunk = schunk;
fd1e8a1f 1711 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
edcb4639
TH
1712 } else {
1713 schunk->free_size = dyn_size;
1714 dyn_size = 0; /* dynamic area covered */
1715 }
2441d15c 1716 schunk->contig_hint = schunk->free_size;
fbf59bc9 1717
fd1e8a1f 1718 schunk->map[schunk->map_used++] = -ai->static_size;
61ace7fa
TH
1719 if (schunk->free_size)
1720 schunk->map[schunk->map_used++] = schunk->free_size;
1721
edcb4639
TH
1722 /* init dynamic chunk if necessary */
1723 if (dyn_size) {
ce3141a2 1724 dchunk = alloc_bootmem(pcpu_chunk_struct_size);
edcb4639 1725 INIT_LIST_HEAD(&dchunk->list);
bba174f5 1726 dchunk->base_addr = base_addr;
edcb4639
TH
1727 dchunk->map = dmap;
1728 dchunk->map_alloc = ARRAY_SIZE(dmap);
38a6be52 1729 dchunk->immutable = true;
ce3141a2 1730 bitmap_fill(dchunk->populated, pcpu_unit_pages);
edcb4639
TH
1731
1732 dchunk->contig_hint = dchunk->free_size = dyn_size;
1733 dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1734 dchunk->map[dchunk->map_used++] = dchunk->free_size;
1735 }
1736
2441d15c 1737 /* link the first chunk in */
ae9e6bc9
TH
1738 pcpu_first_chunk = dchunk ?: schunk;
1739 pcpu_chunk_relocate(pcpu_first_chunk, -1);
fbf59bc9
TH
1740
1741 /* we're done */
bba174f5 1742 pcpu_base_addr = base_addr;
fb435d52 1743 return 0;
fbf59bc9 1744}
66c3a757 1745
f58dc01b
TH
1746const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1747 [PCPU_FC_AUTO] = "auto",
1748 [PCPU_FC_EMBED] = "embed",
1749 [PCPU_FC_PAGE] = "page",
f58dc01b 1750};
66c3a757 1751
f58dc01b 1752enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
66c3a757 1753
f58dc01b
TH
1754static int __init percpu_alloc_setup(char *str)
1755{
1756 if (0)
1757 /* nada */;
1758#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1759 else if (!strcmp(str, "embed"))
1760 pcpu_chosen_fc = PCPU_FC_EMBED;
1761#endif
1762#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1763 else if (!strcmp(str, "page"))
1764 pcpu_chosen_fc = PCPU_FC_PAGE;
f58dc01b
TH
1765#endif
1766 else
1767 pr_warning("PERCPU: unknown allocator %s specified\n", str);
66c3a757 1768
f58dc01b 1769 return 0;
66c3a757 1770}
f58dc01b 1771early_param("percpu_alloc", percpu_alloc_setup);
66c3a757 1772
08fc4580
TH
1773#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1774 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
66c3a757
TH
1775/**
1776 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
66c3a757
TH
1777 * @reserved_size: the size of reserved percpu area in bytes
1778 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
c8826dd5
TH
1779 * @atom_size: allocation atom size
1780 * @cpu_distance_fn: callback to determine distance between cpus, optional
1781 * @alloc_fn: function to allocate percpu page
1782 * @free_fn: funtion to free percpu page
66c3a757
TH
1783 *
1784 * This is a helper to ease setting up embedded first percpu chunk and
1785 * can be called where pcpu_setup_first_chunk() is expected.
1786 *
1787 * If this function is used to setup the first chunk, it is allocated
c8826dd5
TH
1788 * by calling @alloc_fn and used as-is without being mapped into
1789 * vmalloc area. Allocations are always whole multiples of @atom_size
1790 * aligned to @atom_size.
1791 *
1792 * This enables the first chunk to piggy back on the linear physical
1793 * mapping which often uses larger page size. Please note that this
1794 * can result in very sparse cpu->unit mapping on NUMA machines thus
1795 * requiring large vmalloc address space. Don't use this allocator if
1796 * vmalloc space is not orders of magnitude larger than distances
1797 * between node memory addresses (ie. 32bit NUMA machines).
66c3a757
TH
1798 *
1799 * When @dyn_size is positive, dynamic area might be larger than
788e5abc
TH
1800 * specified to fill page alignment. When @dyn_size is auto,
1801 * @dyn_size is just big enough to fill page alignment after static
1802 * and reserved areas.
66c3a757
TH
1803 *
1804 * If the needed size is smaller than the minimum or specified unit
c8826dd5 1805 * size, the leftover is returned using @free_fn.
66c3a757
TH
1806 *
1807 * RETURNS:
fb435d52 1808 * 0 on success, -errno on failure.
66c3a757 1809 */
c8826dd5
TH
1810int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1811 size_t atom_size,
1812 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1813 pcpu_fc_alloc_fn_t alloc_fn,
1814 pcpu_fc_free_fn_t free_fn)
66c3a757 1815{
c8826dd5
TH
1816 void *base = (void *)ULONG_MAX;
1817 void **areas = NULL;
fd1e8a1f 1818 struct pcpu_alloc_info *ai;
6ea529a2 1819 size_t size_sum, areas_size, max_distance;
c8826dd5 1820 int group, i, rc;
66c3a757 1821
c8826dd5
TH
1822 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1823 cpu_distance_fn);
fd1e8a1f
TH
1824 if (IS_ERR(ai))
1825 return PTR_ERR(ai);
66c3a757 1826
fd1e8a1f 1827 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
c8826dd5 1828 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
fa8a7094 1829
c8826dd5
TH
1830 areas = alloc_bootmem_nopanic(areas_size);
1831 if (!areas) {
fb435d52 1832 rc = -ENOMEM;
c8826dd5 1833 goto out_free;
fa8a7094 1834 }
66c3a757 1835
c8826dd5
TH
1836 /* allocate, copy and determine base address */
1837 for (group = 0; group < ai->nr_groups; group++) {
1838 struct pcpu_group_info *gi = &ai->groups[group];
1839 unsigned int cpu = NR_CPUS;
1840 void *ptr;
1841
1842 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1843 cpu = gi->cpu_map[i];
1844 BUG_ON(cpu == NR_CPUS);
1845
1846 /* allocate space for the whole group */
1847 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1848 if (!ptr) {
1849 rc = -ENOMEM;
1850 goto out_free_areas;
1851 }
1852 areas[group] = ptr;
fd1e8a1f 1853
c8826dd5
TH
1854 base = min(ptr, base);
1855
1856 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1857 if (gi->cpu_map[i] == NR_CPUS) {
1858 /* unused unit, free whole */
1859 free_fn(ptr, ai->unit_size);
1860 continue;
1861 }
1862 /* copy and return the unused part */
1863 memcpy(ptr, __per_cpu_load, ai->static_size);
1864 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1865 }
fa8a7094 1866 }
66c3a757 1867
c8826dd5 1868 /* base address is now known, determine group base offsets */
6ea529a2
TH
1869 max_distance = 0;
1870 for (group = 0; group < ai->nr_groups; group++) {
c8826dd5 1871 ai->groups[group].base_offset = areas[group] - base;
1a0c3298
TH
1872 max_distance = max_t(size_t, max_distance,
1873 ai->groups[group].base_offset);
6ea529a2
TH
1874 }
1875 max_distance += ai->unit_size;
1876
1877 /* warn if maximum distance is further than 75% of vmalloc space */
1878 if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
1a0c3298 1879 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
6ea529a2
TH
1880 "space 0x%lx\n",
1881 max_distance, VMALLOC_END - VMALLOC_START);
1882#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1883 /* and fail if we have fallback */
1884 rc = -EINVAL;
1885 goto out_free;
1886#endif
1887 }
c8826dd5 1888
004018e2 1889 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
fd1e8a1f
TH
1890 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1891 ai->dyn_size, ai->unit_size);
d4b95f80 1892
fb435d52 1893 rc = pcpu_setup_first_chunk(ai, base);
c8826dd5
TH
1894 goto out_free;
1895
1896out_free_areas:
1897 for (group = 0; group < ai->nr_groups; group++)
1898 free_fn(areas[group],
1899 ai->groups[group].nr_units * ai->unit_size);
1900out_free:
fd1e8a1f 1901 pcpu_free_alloc_info(ai);
c8826dd5
TH
1902 if (areas)
1903 free_bootmem(__pa(areas), areas_size);
fb435d52 1904 return rc;
d4b95f80 1905}
08fc4580
TH
1906#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1907 !CONFIG_HAVE_SETUP_PER_CPU_AREA */
d4b95f80 1908
08fc4580 1909#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
d4b95f80 1910/**
00ae4064 1911 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
d4b95f80
TH
1912 * @reserved_size: the size of reserved percpu area in bytes
1913 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1914 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1915 * @populate_pte_fn: function to populate pte
1916 *
00ae4064
TH
1917 * This is a helper to ease setting up page-remapped first percpu
1918 * chunk and can be called where pcpu_setup_first_chunk() is expected.
d4b95f80
TH
1919 *
1920 * This is the basic allocator. Static percpu area is allocated
1921 * page-by-page into vmalloc area.
1922 *
1923 * RETURNS:
fb435d52 1924 * 0 on success, -errno on failure.
d4b95f80 1925 */
fb435d52
TH
1926int __init pcpu_page_first_chunk(size_t reserved_size,
1927 pcpu_fc_alloc_fn_t alloc_fn,
1928 pcpu_fc_free_fn_t free_fn,
1929 pcpu_fc_populate_pte_fn_t populate_pte_fn)
d4b95f80 1930{
8f05a6a6 1931 static struct vm_struct vm;
fd1e8a1f 1932 struct pcpu_alloc_info *ai;
00ae4064 1933 char psize_str[16];
ce3141a2 1934 int unit_pages;
d4b95f80 1935 size_t pages_size;
ce3141a2 1936 struct page **pages;
fb435d52 1937 int unit, i, j, rc;
d4b95f80 1938
00ae4064
TH
1939 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1940
fd1e8a1f
TH
1941 ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1942 if (IS_ERR(ai))
1943 return PTR_ERR(ai);
1944 BUG_ON(ai->nr_groups != 1);
1945 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1946
1947 unit_pages = ai->unit_size >> PAGE_SHIFT;
d4b95f80
TH
1948
1949 /* unaligned allocations can't be freed, round up to page size */
fd1e8a1f
TH
1950 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1951 sizeof(pages[0]));
ce3141a2 1952 pages = alloc_bootmem(pages_size);
d4b95f80 1953
8f05a6a6 1954 /* allocate pages */
d4b95f80 1955 j = 0;
fd1e8a1f 1956 for (unit = 0; unit < num_possible_cpus(); unit++)
ce3141a2 1957 for (i = 0; i < unit_pages; i++) {
fd1e8a1f 1958 unsigned int cpu = ai->groups[0].cpu_map[unit];
d4b95f80
TH
1959 void *ptr;
1960
3cbc8565 1961 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
d4b95f80 1962 if (!ptr) {
00ae4064
TH
1963 pr_warning("PERCPU: failed to allocate %s page "
1964 "for cpu%u\n", psize_str, cpu);
d4b95f80
TH
1965 goto enomem;
1966 }
ce3141a2 1967 pages[j++] = virt_to_page(ptr);
d4b95f80
TH
1968 }
1969
8f05a6a6
TH
1970 /* allocate vm area, map the pages and copy static data */
1971 vm.flags = VM_ALLOC;
fd1e8a1f 1972 vm.size = num_possible_cpus() * ai->unit_size;
8f05a6a6
TH
1973 vm_area_register_early(&vm, PAGE_SIZE);
1974
fd1e8a1f 1975 for (unit = 0; unit < num_possible_cpus(); unit++) {
1d9d3257 1976 unsigned long unit_addr =
fd1e8a1f 1977 (unsigned long)vm.addr + unit * ai->unit_size;
8f05a6a6 1978
ce3141a2 1979 for (i = 0; i < unit_pages; i++)
8f05a6a6
TH
1980 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1981
1982 /* pte already populated, the following shouldn't fail */
fb435d52
TH
1983 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1984 unit_pages);
1985 if (rc < 0)
1986 panic("failed to map percpu area, err=%d\n", rc);
66c3a757 1987
8f05a6a6
TH
1988 /*
1989 * FIXME: Archs with virtual cache should flush local
1990 * cache for the linear mapping here - something
1991 * equivalent to flush_cache_vmap() on the local cpu.
1992 * flush_cache_vmap() can't be used as most supporting
1993 * data structures are not set up yet.
1994 */
1995
1996 /* copy static data */
fd1e8a1f 1997 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
66c3a757
TH
1998 }
1999
2000 /* we're ready, commit */
1d9d3257 2001 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
fd1e8a1f
TH
2002 unit_pages, psize_str, vm.addr, ai->static_size,
2003 ai->reserved_size, ai->dyn_size);
d4b95f80 2004
fb435d52 2005 rc = pcpu_setup_first_chunk(ai, vm.addr);
d4b95f80
TH
2006 goto out_free_ar;
2007
2008enomem:
2009 while (--j >= 0)
ce3141a2 2010 free_fn(page_address(pages[j]), PAGE_SIZE);
fb435d52 2011 rc = -ENOMEM;
d4b95f80 2012out_free_ar:
ce3141a2 2013 free_bootmem(__pa(pages), pages_size);
fd1e8a1f 2014 pcpu_free_alloc_info(ai);
fb435d52 2015 return rc;
d4b95f80 2016}
08fc4580 2017#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
d4b95f80 2018
e74e3962
TH
2019/*
2020 * Generic percpu area setup.
2021 *
2022 * The embedding helper is used because its behavior closely resembles
2023 * the original non-dynamic generic percpu area setup. This is
2024 * important because many archs have addressing restrictions and might
2025 * fail if the percpu area is located far away from the previous
2026 * location. As an added bonus, in non-NUMA cases, embedding is
2027 * generally a good idea TLB-wise because percpu area can piggy back
2028 * on the physical linear memory mapping which uses large page
2029 * mappings on applicable archs.
2030 */
2031#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2032unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2033EXPORT_SYMBOL(__per_cpu_offset);
2034
c8826dd5
TH
2035static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2036 size_t align)
2037{
2038 return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
2039}
66c3a757 2040
c8826dd5
TH
2041static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2042{
2043 free_bootmem(__pa(ptr), size);
2044}
2045
e74e3962
TH
2046void __init setup_per_cpu_areas(void)
2047{
e74e3962
TH
2048 unsigned long delta;
2049 unsigned int cpu;
fb435d52 2050 int rc;
e74e3962
TH
2051
2052 /*
2053 * Always reserve area for module percpu variables. That's
2054 * what the legacy allocator did.
2055 */
fb435d52 2056 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
c8826dd5
TH
2057 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2058 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
fb435d52 2059 if (rc < 0)
e74e3962
TH
2060 panic("Failed to initialized percpu areas.");
2061
2062 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2063 for_each_possible_cpu(cpu)
fb435d52 2064 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
66c3a757 2065}
e74e3962 2066#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */