Loading...
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 * Copyright (C) 2017 Facebook Inc.
8 * Copyright (C) 2017 Dennis Zhou <dennisszhou@gmail.com>
9 *
10 * This file is released under the GPLv2 license.
11 *
12 * The percpu allocator handles both static and dynamic areas. Percpu
13 * areas are allocated in chunks which are divided into units. There is
14 * a 1-to-1 mapping for units to possible cpus. These units are grouped
15 * based on NUMA properties of the machine.
16 *
17 * c0 c1 c2
18 * ------------------- ------------------- ------------
19 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
20 * ------------------- ...... ------------------- .... ------------
21 *
22 * Allocation is done by offsets into a unit's address space. Ie., an
23 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
24 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
25 * and even sparse. Access is handled by configuring percpu base
26 * registers according to the cpu to unit mappings and offsetting the
27 * base address using pcpu_unit_size.
28 *
29 * There is special consideration for the first chunk which must handle
30 * the static percpu variables in the kernel image as allocation services
31 * are not online yet. In short, the first chunk is structured like so:
32 *
33 * <Static | [Reserved] | Dynamic>
34 *
35 * The static data is copied from the original section managed by the
36 * linker. The reserved section, if non-zero, primarily manages static
37 * percpu variables from kernel modules. Finally, the dynamic section
38 * takes care of normal allocations.
39 *
40 * The allocator organizes chunks into lists according to free size and
41 * tries to allocate from the fullest chunk first. Each chunk is managed
42 * by a bitmap with metadata blocks. The allocation map is updated on
43 * every allocation and free to reflect the current state while the boundary
44 * map is only updated on allocation. Each metadata block contains
45 * information to help mitigate the need to iterate over large portions
46 * of the bitmap. The reverse mapping from page to chunk is stored in
47 * the page's index. Lastly, units are lazily backed and grow in unison.
48 *
49 * There is a unique conversion that goes on here between bytes and bits.
50 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
51 * tracks the number of pages it is responsible for in nr_pages. Helper
52 * functions are used to convert from between the bytes, bits, and blocks.
53 * All hints are managed in bits unless explicitly stated.
54 *
55 * To use this allocator, arch code should do the following:
56 *
57 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
58 * regular address to percpu pointer and back if they need to be
59 * different from the default
60 *
61 * - use pcpu_setup_first_chunk() during percpu area initialization to
62 * setup the first chunk containing the kernel static percpu area
63 */
64
65#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
66
67#include <linux/bitmap.h>
68#include <linux/bootmem.h>
69#include <linux/err.h>
70#include <linux/lcm.h>
71#include <linux/list.h>
72#include <linux/log2.h>
73#include <linux/mm.h>
74#include <linux/module.h>
75#include <linux/mutex.h>
76#include <linux/percpu.h>
77#include <linux/pfn.h>
78#include <linux/slab.h>
79#include <linux/spinlock.h>
80#include <linux/vmalloc.h>
81#include <linux/workqueue.h>
82#include <linux/kmemleak.h>
83#include <linux/sched.h>
84
85#include <asm/cacheflush.h>
86#include <asm/sections.h>
87#include <asm/tlbflush.h>
88#include <asm/io.h>
89
90#define CREATE_TRACE_POINTS
91#include <trace/events/percpu.h>
92
93#include "percpu-internal.h"
94
95/* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
96#define PCPU_SLOT_BASE_SHIFT 5
97
98#define PCPU_EMPTY_POP_PAGES_LOW 2
99#define PCPU_EMPTY_POP_PAGES_HIGH 4
100
101#ifdef CONFIG_SMP
102/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
103#ifndef __addr_to_pcpu_ptr
104#define __addr_to_pcpu_ptr(addr) \
105 (void __percpu *)((unsigned long)(addr) - \
106 (unsigned long)pcpu_base_addr + \
107 (unsigned long)__per_cpu_start)
108#endif
109#ifndef __pcpu_ptr_to_addr
110#define __pcpu_ptr_to_addr(ptr) \
111 (void __force *)((unsigned long)(ptr) + \
112 (unsigned long)pcpu_base_addr - \
113 (unsigned long)__per_cpu_start)
114#endif
115#else /* CONFIG_SMP */
116/* on UP, it's always identity mapped */
117#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
118#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
119#endif /* CONFIG_SMP */
120
121static int pcpu_unit_pages __ro_after_init;
122static int pcpu_unit_size __ro_after_init;
123static int pcpu_nr_units __ro_after_init;
124static int pcpu_atom_size __ro_after_init;
125int pcpu_nr_slots __ro_after_init;
126static size_t pcpu_chunk_struct_size __ro_after_init;
127
128/* cpus with the lowest and highest unit addresses */
129static unsigned int pcpu_low_unit_cpu __ro_after_init;
130static unsigned int pcpu_high_unit_cpu __ro_after_init;
131
132/* the address of the first chunk which starts with the kernel static area */
133void *pcpu_base_addr __ro_after_init;
134EXPORT_SYMBOL_GPL(pcpu_base_addr);
135
136static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
137const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
138
139/* group information, used for vm allocation */
140static int pcpu_nr_groups __ro_after_init;
141static const unsigned long *pcpu_group_offsets __ro_after_init;
142static const size_t *pcpu_group_sizes __ro_after_init;
143
144/*
145 * The first chunk which always exists. Note that unlike other
146 * chunks, this one can be allocated and mapped in several different
147 * ways and thus often doesn't live in the vmalloc area.
148 */
149struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
150
151/*
152 * Optional reserved chunk. This chunk reserves part of the first
153 * chunk and serves it for reserved allocations. When the reserved
154 * region doesn't exist, the following variable is NULL.
155 */
156struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
157
158DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
159static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
160
161struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
162
163/* chunks which need their map areas extended, protected by pcpu_lock */
164static LIST_HEAD(pcpu_map_extend_chunks);
165
166/*
167 * The number of empty populated pages, protected by pcpu_lock. The
168 * reserved chunk doesn't contribute to the count.
169 */
170int pcpu_nr_empty_pop_pages;
171
172/*
173 * Balance work is used to populate or destroy chunks asynchronously. We
174 * try to keep the number of populated free pages between
175 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
176 * empty chunk.
177 */
178static void pcpu_balance_workfn(struct work_struct *work);
179static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
180static bool pcpu_async_enabled __read_mostly;
181static bool pcpu_atomic_alloc_failed;
182
183static void pcpu_schedule_balance_work(void)
184{
185 if (pcpu_async_enabled)
186 schedule_work(&pcpu_balance_work);
187}
188
189/**
190 * pcpu_addr_in_chunk - check if the address is served from this chunk
191 * @chunk: chunk of interest
192 * @addr: percpu address
193 *
194 * RETURNS:
195 * True if the address is served from this chunk.
196 */
197static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
198{
199 void *start_addr, *end_addr;
200
201 if (!chunk)
202 return false;
203
204 start_addr = chunk->base_addr + chunk->start_offset;
205 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
206 chunk->end_offset;
207
208 return addr >= start_addr && addr < end_addr;
209}
210
211static int __pcpu_size_to_slot(int size)
212{
213 int highbit = fls(size); /* size is in bytes */
214 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
215}
216
217static int pcpu_size_to_slot(int size)
218{
219 if (size == pcpu_unit_size)
220 return pcpu_nr_slots - 1;
221 return __pcpu_size_to_slot(size);
222}
223
224static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
225{
226 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE || chunk->contig_bits == 0)
227 return 0;
228
229 return pcpu_size_to_slot(chunk->free_bytes);
230}
231
232/* set the pointer to a chunk in a page struct */
233static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
234{
235 page->index = (unsigned long)pcpu;
236}
237
238/* obtain pointer to a chunk from a page struct */
239static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
240{
241 return (struct pcpu_chunk *)page->index;
242}
243
244static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
245{
246 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
247}
248
249static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
250{
251 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
252}
253
254static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
255 unsigned int cpu, int page_idx)
256{
257 return (unsigned long)chunk->base_addr +
258 pcpu_unit_page_offset(cpu, page_idx);
259}
260
261static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end)
262{
263 *rs = find_next_zero_bit(bitmap, end, *rs);
264 *re = find_next_bit(bitmap, end, *rs + 1);
265}
266
267static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end)
268{
269 *rs = find_next_bit(bitmap, end, *rs);
270 *re = find_next_zero_bit(bitmap, end, *rs + 1);
271}
272
273/*
274 * Bitmap region iterators. Iterates over the bitmap between
275 * [@start, @end) in @chunk. @rs and @re should be integer variables
276 * and will be set to start and end index of the current free region.
277 */
278#define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \
279 for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
280 (rs) < (re); \
281 (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
282
283#define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \
284 for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \
285 (rs) < (re); \
286 (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
287
288/*
289 * The following are helper functions to help access bitmaps and convert
290 * between bitmap offsets to address offsets.
291 */
292static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
293{
294 return chunk->alloc_map +
295 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
296}
297
298static unsigned long pcpu_off_to_block_index(int off)
299{
300 return off / PCPU_BITMAP_BLOCK_BITS;
301}
302
303static unsigned long pcpu_off_to_block_off(int off)
304{
305 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
306}
307
308static unsigned long pcpu_block_off_to_off(int index, int off)
309{
310 return index * PCPU_BITMAP_BLOCK_BITS + off;
311}
312
313/**
314 * pcpu_next_md_free_region - finds the next hint free area
315 * @chunk: chunk of interest
316 * @bit_off: chunk offset
317 * @bits: size of free area
318 *
319 * Helper function for pcpu_for_each_md_free_region. It checks
320 * block->contig_hint and performs aggregation across blocks to find the
321 * next hint. It modifies bit_off and bits in-place to be consumed in the
322 * loop.
323 */
324static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
325 int *bits)
326{
327 int i = pcpu_off_to_block_index(*bit_off);
328 int block_off = pcpu_off_to_block_off(*bit_off);
329 struct pcpu_block_md *block;
330
331 *bits = 0;
332 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
333 block++, i++) {
334 /* handles contig area across blocks */
335 if (*bits) {
336 *bits += block->left_free;
337 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
338 continue;
339 return;
340 }
341
342 /*
343 * This checks three things. First is there a contig_hint to
344 * check. Second, have we checked this hint before by
345 * comparing the block_off. Third, is this the same as the
346 * right contig hint. In the last case, it spills over into
347 * the next block and should be handled by the contig area
348 * across blocks code.
349 */
350 *bits = block->contig_hint;
351 if (*bits && block->contig_hint_start >= block_off &&
352 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
353 *bit_off = pcpu_block_off_to_off(i,
354 block->contig_hint_start);
355 return;
356 }
357 /* reset to satisfy the second predicate above */
358 block_off = 0;
359
360 *bits = block->right_free;
361 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
362 }
363}
364
365/**
366 * pcpu_next_fit_region - finds fit areas for a given allocation request
367 * @chunk: chunk of interest
368 * @alloc_bits: size of allocation
369 * @align: alignment of area (max PAGE_SIZE)
370 * @bit_off: chunk offset
371 * @bits: size of free area
372 *
373 * Finds the next free region that is viable for use with a given size and
374 * alignment. This only returns if there is a valid area to be used for this
375 * allocation. block->first_free is returned if the allocation request fits
376 * within the block to see if the request can be fulfilled prior to the contig
377 * hint.
378 */
379static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
380 int align, int *bit_off, int *bits)
381{
382 int i = pcpu_off_to_block_index(*bit_off);
383 int block_off = pcpu_off_to_block_off(*bit_off);
384 struct pcpu_block_md *block;
385
386 *bits = 0;
387 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
388 block++, i++) {
389 /* handles contig area across blocks */
390 if (*bits) {
391 *bits += block->left_free;
392 if (*bits >= alloc_bits)
393 return;
394 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
395 continue;
396 }
397
398 /* check block->contig_hint */
399 *bits = ALIGN(block->contig_hint_start, align) -
400 block->contig_hint_start;
401 /*
402 * This uses the block offset to determine if this has been
403 * checked in the prior iteration.
404 */
405 if (block->contig_hint &&
406 block->contig_hint_start >= block_off &&
407 block->contig_hint >= *bits + alloc_bits) {
408 *bits += alloc_bits + block->contig_hint_start -
409 block->first_free;
410 *bit_off = pcpu_block_off_to_off(i, block->first_free);
411 return;
412 }
413 /* reset to satisfy the second predicate above */
414 block_off = 0;
415
416 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
417 align);
418 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
419 *bit_off = pcpu_block_off_to_off(i, *bit_off);
420 if (*bits >= alloc_bits)
421 return;
422 }
423
424 /* no valid offsets were found - fail condition */
425 *bit_off = pcpu_chunk_map_bits(chunk);
426}
427
428/*
429 * Metadata free area iterators. These perform aggregation of free areas
430 * based on the metadata blocks and return the offset @bit_off and size in
431 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
432 * a fit is found for the allocation request.
433 */
434#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
435 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
436 (bit_off) < pcpu_chunk_map_bits((chunk)); \
437 (bit_off) += (bits) + 1, \
438 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
439
440#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
441 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
442 &(bits)); \
443 (bit_off) < pcpu_chunk_map_bits((chunk)); \
444 (bit_off) += (bits), \
445 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
446 &(bits)))
447
448/**
449 * pcpu_mem_zalloc - allocate memory
450 * @size: bytes to allocate
451 * @gfp: allocation flags
452 *
453 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
454 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
455 * This is to facilitate passing through whitelisted flags. The
456 * returned memory is always zeroed.
457 *
458 * RETURNS:
459 * Pointer to the allocated area on success, NULL on failure.
460 */
461static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
462{
463 if (WARN_ON_ONCE(!slab_is_available()))
464 return NULL;
465
466 if (size <= PAGE_SIZE)
467 return kzalloc(size, gfp);
468 else
469 return __vmalloc(size, gfp | __GFP_ZERO, PAGE_KERNEL);
470}
471
472/**
473 * pcpu_mem_free - free memory
474 * @ptr: memory to free
475 *
476 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
477 */
478static void pcpu_mem_free(void *ptr)
479{
480 kvfree(ptr);
481}
482
483/**
484 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
485 * @chunk: chunk of interest
486 * @oslot: the previous slot it was on
487 *
488 * This function is called after an allocation or free changed @chunk.
489 * New slot according to the changed state is determined and @chunk is
490 * moved to the slot. Note that the reserved chunk is never put on
491 * chunk slots.
492 *
493 * CONTEXT:
494 * pcpu_lock.
495 */
496static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
497{
498 int nslot = pcpu_chunk_slot(chunk);
499
500 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
501 if (oslot < nslot)
502 list_move(&chunk->list, &pcpu_slot[nslot]);
503 else
504 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
505 }
506}
507
508/**
509 * pcpu_cnt_pop_pages- counts populated backing pages in range
510 * @chunk: chunk of interest
511 * @bit_off: start offset
512 * @bits: size of area to check
513 *
514 * Calculates the number of populated pages in the region
515 * [page_start, page_end). This keeps track of how many empty populated
516 * pages are available and decide if async work should be scheduled.
517 *
518 * RETURNS:
519 * The nr of populated pages.
520 */
521static inline int pcpu_cnt_pop_pages(struct pcpu_chunk *chunk, int bit_off,
522 int bits)
523{
524 int page_start = PFN_UP(bit_off * PCPU_MIN_ALLOC_SIZE);
525 int page_end = PFN_DOWN((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
526
527 if (page_start >= page_end)
528 return 0;
529
530 /*
531 * bitmap_weight counts the number of bits set in a bitmap up to
532 * the specified number of bits. This is counting the populated
533 * pages up to page_end and then subtracting the populated pages
534 * up to page_start to count the populated pages in
535 * [page_start, page_end).
536 */
537 return bitmap_weight(chunk->populated, page_end) -
538 bitmap_weight(chunk->populated, page_start);
539}
540
541/**
542 * pcpu_chunk_update - updates the chunk metadata given a free area
543 * @chunk: chunk of interest
544 * @bit_off: chunk offset
545 * @bits: size of free area
546 *
547 * This updates the chunk's contig hint and starting offset given a free area.
548 * Choose the best starting offset if the contig hint is equal.
549 */
550static void pcpu_chunk_update(struct pcpu_chunk *chunk, int bit_off, int bits)
551{
552 if (bits > chunk->contig_bits) {
553 chunk->contig_bits_start = bit_off;
554 chunk->contig_bits = bits;
555 } else if (bits == chunk->contig_bits && chunk->contig_bits_start &&
556 (!bit_off ||
557 __ffs(bit_off) > __ffs(chunk->contig_bits_start))) {
558 /* use the start with the best alignment */
559 chunk->contig_bits_start = bit_off;
560 }
561}
562
563/**
564 * pcpu_chunk_refresh_hint - updates metadata about a chunk
565 * @chunk: chunk of interest
566 *
567 * Iterates over the metadata blocks to find the largest contig area.
568 * It also counts the populated pages and uses the delta to update the
569 * global count.
570 *
571 * Updates:
572 * chunk->contig_bits
573 * chunk->contig_bits_start
574 * nr_empty_pop_pages (chunk and global)
575 */
576static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk)
577{
578 int bit_off, bits, nr_empty_pop_pages;
579
580 /* clear metadata */
581 chunk->contig_bits = 0;
582
583 bit_off = chunk->first_bit;
584 bits = nr_empty_pop_pages = 0;
585 pcpu_for_each_md_free_region(chunk, bit_off, bits) {
586 pcpu_chunk_update(chunk, bit_off, bits);
587
588 nr_empty_pop_pages += pcpu_cnt_pop_pages(chunk, bit_off, bits);
589 }
590
591 /*
592 * Keep track of nr_empty_pop_pages.
593 *
594 * The chunk maintains the previous number of free pages it held,
595 * so the delta is used to update the global counter. The reserved
596 * chunk is not part of the free page count as they are populated
597 * at init and are special to serving reserved allocations.
598 */
599 if (chunk != pcpu_reserved_chunk)
600 pcpu_nr_empty_pop_pages +=
601 (nr_empty_pop_pages - chunk->nr_empty_pop_pages);
602
603 chunk->nr_empty_pop_pages = nr_empty_pop_pages;
604}
605
606/**
607 * pcpu_block_update - updates a block given a free area
608 * @block: block of interest
609 * @start: start offset in block
610 * @end: end offset in block
611 *
612 * Updates a block given a known free area. The region [start, end) is
613 * expected to be the entirety of the free area within a block. Chooses
614 * the best starting offset if the contig hints are equal.
615 */
616static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
617{
618 int contig = end - start;
619
620 block->first_free = min(block->first_free, start);
621 if (start == 0)
622 block->left_free = contig;
623
624 if (end == PCPU_BITMAP_BLOCK_BITS)
625 block->right_free = contig;
626
627 if (contig > block->contig_hint) {
628 block->contig_hint_start = start;
629 block->contig_hint = contig;
630 } else if (block->contig_hint_start && contig == block->contig_hint &&
631 (!start || __ffs(start) > __ffs(block->contig_hint_start))) {
632 /* use the start with the best alignment */
633 block->contig_hint_start = start;
634 }
635}
636
637/**
638 * pcpu_block_refresh_hint
639 * @chunk: chunk of interest
640 * @index: index of the metadata block
641 *
642 * Scans over the block beginning at first_free and updates the block
643 * metadata accordingly.
644 */
645static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
646{
647 struct pcpu_block_md *block = chunk->md_blocks + index;
648 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
649 int rs, re; /* region start, region end */
650
651 /* clear hints */
652 block->contig_hint = 0;
653 block->left_free = block->right_free = 0;
654
655 /* iterate over free areas and update the contig hints */
656 pcpu_for_each_unpop_region(alloc_map, rs, re, block->first_free,
657 PCPU_BITMAP_BLOCK_BITS) {
658 pcpu_block_update(block, rs, re);
659 }
660}
661
662/**
663 * pcpu_block_update_hint_alloc - update hint on allocation path
664 * @chunk: chunk of interest
665 * @bit_off: chunk offset
666 * @bits: size of request
667 *
668 * Updates metadata for the allocation path. The metadata only has to be
669 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
670 * scans are required if the block's contig hint is broken.
671 */
672static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
673 int bits)
674{
675 struct pcpu_block_md *s_block, *e_block, *block;
676 int s_index, e_index; /* block indexes of the freed allocation */
677 int s_off, e_off; /* block offsets of the freed allocation */
678
679 /*
680 * Calculate per block offsets.
681 * The calculation uses an inclusive range, but the resulting offsets
682 * are [start, end). e_index always points to the last block in the
683 * range.
684 */
685 s_index = pcpu_off_to_block_index(bit_off);
686 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
687 s_off = pcpu_off_to_block_off(bit_off);
688 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
689
690 s_block = chunk->md_blocks + s_index;
691 e_block = chunk->md_blocks + e_index;
692
693 /*
694 * Update s_block.
695 * block->first_free must be updated if the allocation takes its place.
696 * If the allocation breaks the contig_hint, a scan is required to
697 * restore this hint.
698 */
699 if (s_off == s_block->first_free)
700 s_block->first_free = find_next_zero_bit(
701 pcpu_index_alloc_map(chunk, s_index),
702 PCPU_BITMAP_BLOCK_BITS,
703 s_off + bits);
704
705 if (s_off >= s_block->contig_hint_start &&
706 s_off < s_block->contig_hint_start + s_block->contig_hint) {
707 /* block contig hint is broken - scan to fix it */
708 pcpu_block_refresh_hint(chunk, s_index);
709 } else {
710 /* update left and right contig manually */
711 s_block->left_free = min(s_block->left_free, s_off);
712 if (s_index == e_index)
713 s_block->right_free = min_t(int, s_block->right_free,
714 PCPU_BITMAP_BLOCK_BITS - e_off);
715 else
716 s_block->right_free = 0;
717 }
718
719 /*
720 * Update e_block.
721 */
722 if (s_index != e_index) {
723 /*
724 * When the allocation is across blocks, the end is along
725 * the left part of the e_block.
726 */
727 e_block->first_free = find_next_zero_bit(
728 pcpu_index_alloc_map(chunk, e_index),
729 PCPU_BITMAP_BLOCK_BITS, e_off);
730
731 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
732 /* reset the block */
733 e_block++;
734 } else {
735 if (e_off > e_block->contig_hint_start) {
736 /* contig hint is broken - scan to fix it */
737 pcpu_block_refresh_hint(chunk, e_index);
738 } else {
739 e_block->left_free = 0;
740 e_block->right_free =
741 min_t(int, e_block->right_free,
742 PCPU_BITMAP_BLOCK_BITS - e_off);
743 }
744 }
745
746 /* update in-between md_blocks */
747 for (block = s_block + 1; block < e_block; block++) {
748 block->contig_hint = 0;
749 block->left_free = 0;
750 block->right_free = 0;
751 }
752 }
753
754 /*
755 * The only time a full chunk scan is required is if the chunk
756 * contig hint is broken. Otherwise, it means a smaller space
757 * was used and therefore the chunk contig hint is still correct.
758 */
759 if (bit_off >= chunk->contig_bits_start &&
760 bit_off < chunk->contig_bits_start + chunk->contig_bits)
761 pcpu_chunk_refresh_hint(chunk);
762}
763
764/**
765 * pcpu_block_update_hint_free - updates the block hints on the free path
766 * @chunk: chunk of interest
767 * @bit_off: chunk offset
768 * @bits: size of request
769 *
770 * Updates metadata for the allocation path. This avoids a blind block
771 * refresh by making use of the block contig hints. If this fails, it scans
772 * forward and backward to determine the extent of the free area. This is
773 * capped at the boundary of blocks.
774 *
775 * A chunk update is triggered if a page becomes free, a block becomes free,
776 * or the free spans across blocks. This tradeoff is to minimize iterating
777 * over the block metadata to update chunk->contig_bits. chunk->contig_bits
778 * may be off by up to a page, but it will never be more than the available
779 * space. If the contig hint is contained in one block, it will be accurate.
780 */
781static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
782 int bits)
783{
784 struct pcpu_block_md *s_block, *e_block, *block;
785 int s_index, e_index; /* block indexes of the freed allocation */
786 int s_off, e_off; /* block offsets of the freed allocation */
787 int start, end; /* start and end of the whole free area */
788
789 /*
790 * Calculate per block offsets.
791 * The calculation uses an inclusive range, but the resulting offsets
792 * are [start, end). e_index always points to the last block in the
793 * range.
794 */
795 s_index = pcpu_off_to_block_index(bit_off);
796 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
797 s_off = pcpu_off_to_block_off(bit_off);
798 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
799
800 s_block = chunk->md_blocks + s_index;
801 e_block = chunk->md_blocks + e_index;
802
803 /*
804 * Check if the freed area aligns with the block->contig_hint.
805 * If it does, then the scan to find the beginning/end of the
806 * larger free area can be avoided.
807 *
808 * start and end refer to beginning and end of the free area
809 * within each their respective blocks. This is not necessarily
810 * the entire free area as it may span blocks past the beginning
811 * or end of the block.
812 */
813 start = s_off;
814 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
815 start = s_block->contig_hint_start;
816 } else {
817 /*
818 * Scan backwards to find the extent of the free area.
819 * find_last_bit returns the starting bit, so if the start bit
820 * is returned, that means there was no last bit and the
821 * remainder of the chunk is free.
822 */
823 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
824 start);
825 start = (start == l_bit) ? 0 : l_bit + 1;
826 }
827
828 end = e_off;
829 if (e_off == e_block->contig_hint_start)
830 end = e_block->contig_hint_start + e_block->contig_hint;
831 else
832 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
833 PCPU_BITMAP_BLOCK_BITS, end);
834
835 /* update s_block */
836 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
837 pcpu_block_update(s_block, start, e_off);
838
839 /* freeing in the same block */
840 if (s_index != e_index) {
841 /* update e_block */
842 pcpu_block_update(e_block, 0, end);
843
844 /* reset md_blocks in the middle */
845 for (block = s_block + 1; block < e_block; block++) {
846 block->first_free = 0;
847 block->contig_hint_start = 0;
848 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
849 block->left_free = PCPU_BITMAP_BLOCK_BITS;
850 block->right_free = PCPU_BITMAP_BLOCK_BITS;
851 }
852 }
853
854 /*
855 * Refresh chunk metadata when the free makes a page free, a block
856 * free, or spans across blocks. The contig hint may be off by up to
857 * a page, but if the hint is contained in a block, it will be accurate
858 * with the else condition below.
859 */
860 if ((ALIGN_DOWN(end, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS)) >
861 ALIGN(start, min(PCPU_BITS_PER_PAGE, PCPU_BITMAP_BLOCK_BITS))) ||
862 s_index != e_index)
863 pcpu_chunk_refresh_hint(chunk);
864 else
865 pcpu_chunk_update(chunk, pcpu_block_off_to_off(s_index, start),
866 s_block->contig_hint);
867}
868
869/**
870 * pcpu_is_populated - determines if the region is populated
871 * @chunk: chunk of interest
872 * @bit_off: chunk offset
873 * @bits: size of area
874 * @next_off: return value for the next offset to start searching
875 *
876 * For atomic allocations, check if the backing pages are populated.
877 *
878 * RETURNS:
879 * Bool if the backing pages are populated.
880 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
881 */
882static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
883 int *next_off)
884{
885 int page_start, page_end, rs, re;
886
887 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
888 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
889
890 rs = page_start;
891 pcpu_next_unpop(chunk->populated, &rs, &re, page_end);
892 if (rs >= page_end)
893 return true;
894
895 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
896 return false;
897}
898
899/**
900 * pcpu_find_block_fit - finds the block index to start searching
901 * @chunk: chunk of interest
902 * @alloc_bits: size of request in allocation units
903 * @align: alignment of area (max PAGE_SIZE bytes)
904 * @pop_only: use populated regions only
905 *
906 * Given a chunk and an allocation spec, find the offset to begin searching
907 * for a free region. This iterates over the bitmap metadata blocks to
908 * find an offset that will be guaranteed to fit the requirements. It is
909 * not quite first fit as if the allocation does not fit in the contig hint
910 * of a block or chunk, it is skipped. This errs on the side of caution
911 * to prevent excess iteration. Poor alignment can cause the allocator to
912 * skip over blocks and chunks that have valid free areas.
913 *
914 * RETURNS:
915 * The offset in the bitmap to begin searching.
916 * -1 if no offset is found.
917 */
918static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
919 size_t align, bool pop_only)
920{
921 int bit_off, bits, next_off;
922
923 /*
924 * Check to see if the allocation can fit in the chunk's contig hint.
925 * This is an optimization to prevent scanning by assuming if it
926 * cannot fit in the global hint, there is memory pressure and creating
927 * a new chunk would happen soon.
928 */
929 bit_off = ALIGN(chunk->contig_bits_start, align) -
930 chunk->contig_bits_start;
931 if (bit_off + alloc_bits > chunk->contig_bits)
932 return -1;
933
934 bit_off = chunk->first_bit;
935 bits = 0;
936 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
937 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
938 &next_off))
939 break;
940
941 bit_off = next_off;
942 bits = 0;
943 }
944
945 if (bit_off == pcpu_chunk_map_bits(chunk))
946 return -1;
947
948 return bit_off;
949}
950
951/**
952 * pcpu_alloc_area - allocates an area from a pcpu_chunk
953 * @chunk: chunk of interest
954 * @alloc_bits: size of request in allocation units
955 * @align: alignment of area (max PAGE_SIZE)
956 * @start: bit_off to start searching
957 *
958 * This function takes in a @start offset to begin searching to fit an
959 * allocation of @alloc_bits with alignment @align. It needs to scan
960 * the allocation map because if it fits within the block's contig hint,
961 * @start will be block->first_free. This is an attempt to fill the
962 * allocation prior to breaking the contig hint. The allocation and
963 * boundary maps are updated accordingly if it confirms a valid
964 * free area.
965 *
966 * RETURNS:
967 * Allocated addr offset in @chunk on success.
968 * -1 if no matching area is found.
969 */
970static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
971 size_t align, int start)
972{
973 size_t align_mask = (align) ? (align - 1) : 0;
974 int bit_off, end, oslot;
975
976 lockdep_assert_held(&pcpu_lock);
977
978 oslot = pcpu_chunk_slot(chunk);
979
980 /*
981 * Search to find a fit.
982 */
983 end = start + alloc_bits + PCPU_BITMAP_BLOCK_BITS;
984 bit_off = bitmap_find_next_zero_area(chunk->alloc_map, end, start,
985 alloc_bits, align_mask);
986 if (bit_off >= end)
987 return -1;
988
989 /* update alloc map */
990 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
991
992 /* update boundary map */
993 set_bit(bit_off, chunk->bound_map);
994 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
995 set_bit(bit_off + alloc_bits, chunk->bound_map);
996
997 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
998
999 /* update first free bit */
1000 if (bit_off == chunk->first_bit)
1001 chunk->first_bit = find_next_zero_bit(
1002 chunk->alloc_map,
1003 pcpu_chunk_map_bits(chunk),
1004 bit_off + alloc_bits);
1005
1006 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1007
1008 pcpu_chunk_relocate(chunk, oslot);
1009
1010 return bit_off * PCPU_MIN_ALLOC_SIZE;
1011}
1012
1013/**
1014 * pcpu_free_area - frees the corresponding offset
1015 * @chunk: chunk of interest
1016 * @off: addr offset into chunk
1017 *
1018 * This function determines the size of an allocation to free using
1019 * the boundary bitmap and clears the allocation map.
1020 */
1021static void pcpu_free_area(struct pcpu_chunk *chunk, int off)
1022{
1023 int bit_off, bits, end, oslot;
1024
1025 lockdep_assert_held(&pcpu_lock);
1026 pcpu_stats_area_dealloc(chunk);
1027
1028 oslot = pcpu_chunk_slot(chunk);
1029
1030 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1031
1032 /* find end index */
1033 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1034 bit_off + 1);
1035 bits = end - bit_off;
1036 bitmap_clear(chunk->alloc_map, bit_off, bits);
1037
1038 /* update metadata */
1039 chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE;
1040
1041 /* update first free bit */
1042 chunk->first_bit = min(chunk->first_bit, bit_off);
1043
1044 pcpu_block_update_hint_free(chunk, bit_off, bits);
1045
1046 pcpu_chunk_relocate(chunk, oslot);
1047}
1048
1049static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1050{
1051 struct pcpu_block_md *md_block;
1052
1053 for (md_block = chunk->md_blocks;
1054 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1055 md_block++) {
1056 md_block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1057 md_block->left_free = PCPU_BITMAP_BLOCK_BITS;
1058 md_block->right_free = PCPU_BITMAP_BLOCK_BITS;
1059 }
1060}
1061
1062/**
1063 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1064 * @tmp_addr: the start of the region served
1065 * @map_size: size of the region served
1066 *
1067 * This is responsible for creating the chunks that serve the first chunk. The
1068 * base_addr is page aligned down of @tmp_addr while the region end is page
1069 * aligned up. Offsets are kept track of to determine the region served. All
1070 * this is done to appease the bitmap allocator in avoiding partial blocks.
1071 *
1072 * RETURNS:
1073 * Chunk serving the region at @tmp_addr of @map_size.
1074 */
1075static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1076 int map_size)
1077{
1078 struct pcpu_chunk *chunk;
1079 unsigned long aligned_addr, lcm_align;
1080 int start_offset, offset_bits, region_size, region_bits;
1081
1082 /* region calculations */
1083 aligned_addr = tmp_addr & PAGE_MASK;
1084
1085 start_offset = tmp_addr - aligned_addr;
1086
1087 /*
1088 * Align the end of the region with the LCM of PAGE_SIZE and
1089 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1090 * the other.
1091 */
1092 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1093 region_size = ALIGN(start_offset + map_size, lcm_align);
1094
1095 /* allocate chunk */
1096 chunk = memblock_virt_alloc(sizeof(struct pcpu_chunk) +
1097 BITS_TO_LONGS(region_size >> PAGE_SHIFT),
1098 0);
1099
1100 INIT_LIST_HEAD(&chunk->list);
1101
1102 chunk->base_addr = (void *)aligned_addr;
1103 chunk->start_offset = start_offset;
1104 chunk->end_offset = region_size - chunk->start_offset - map_size;
1105
1106 chunk->nr_pages = region_size >> PAGE_SHIFT;
1107 region_bits = pcpu_chunk_map_bits(chunk);
1108
1109 chunk->alloc_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits) *
1110 sizeof(chunk->alloc_map[0]), 0);
1111 chunk->bound_map = memblock_virt_alloc(BITS_TO_LONGS(region_bits + 1) *
1112 sizeof(chunk->bound_map[0]), 0);
1113 chunk->md_blocks = memblock_virt_alloc(pcpu_chunk_nr_blocks(chunk) *
1114 sizeof(chunk->md_blocks[0]), 0);
1115 pcpu_init_md_blocks(chunk);
1116
1117 /* manage populated page bitmap */
1118 chunk->immutable = true;
1119 bitmap_fill(chunk->populated, chunk->nr_pages);
1120 chunk->nr_populated = chunk->nr_pages;
1121 chunk->nr_empty_pop_pages =
1122 pcpu_cnt_pop_pages(chunk, start_offset / PCPU_MIN_ALLOC_SIZE,
1123 map_size / PCPU_MIN_ALLOC_SIZE);
1124
1125 chunk->contig_bits = map_size / PCPU_MIN_ALLOC_SIZE;
1126 chunk->free_bytes = map_size;
1127
1128 if (chunk->start_offset) {
1129 /* hide the beginning of the bitmap */
1130 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1131 bitmap_set(chunk->alloc_map, 0, offset_bits);
1132 set_bit(0, chunk->bound_map);
1133 set_bit(offset_bits, chunk->bound_map);
1134
1135 chunk->first_bit = offset_bits;
1136
1137 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1138 }
1139
1140 if (chunk->end_offset) {
1141 /* hide the end of the bitmap */
1142 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1143 bitmap_set(chunk->alloc_map,
1144 pcpu_chunk_map_bits(chunk) - offset_bits,
1145 offset_bits);
1146 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1147 chunk->bound_map);
1148 set_bit(region_bits, chunk->bound_map);
1149
1150 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1151 - offset_bits, offset_bits);
1152 }
1153
1154 return chunk;
1155}
1156
1157static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1158{
1159 struct pcpu_chunk *chunk;
1160 int region_bits;
1161
1162 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1163 if (!chunk)
1164 return NULL;
1165
1166 INIT_LIST_HEAD(&chunk->list);
1167 chunk->nr_pages = pcpu_unit_pages;
1168 region_bits = pcpu_chunk_map_bits(chunk);
1169
1170 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1171 sizeof(chunk->alloc_map[0]), gfp);
1172 if (!chunk->alloc_map)
1173 goto alloc_map_fail;
1174
1175 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1176 sizeof(chunk->bound_map[0]), gfp);
1177 if (!chunk->bound_map)
1178 goto bound_map_fail;
1179
1180 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1181 sizeof(chunk->md_blocks[0]), gfp);
1182 if (!chunk->md_blocks)
1183 goto md_blocks_fail;
1184
1185 pcpu_init_md_blocks(chunk);
1186
1187 /* init metadata */
1188 chunk->contig_bits = region_bits;
1189 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1190
1191 return chunk;
1192
1193md_blocks_fail:
1194 pcpu_mem_free(chunk->bound_map);
1195bound_map_fail:
1196 pcpu_mem_free(chunk->alloc_map);
1197alloc_map_fail:
1198 pcpu_mem_free(chunk);
1199
1200 return NULL;
1201}
1202
1203static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1204{
1205 if (!chunk)
1206 return;
1207 pcpu_mem_free(chunk->bound_map);
1208 pcpu_mem_free(chunk->alloc_map);
1209 pcpu_mem_free(chunk);
1210}
1211
1212/**
1213 * pcpu_chunk_populated - post-population bookkeeping
1214 * @chunk: pcpu_chunk which got populated
1215 * @page_start: the start page
1216 * @page_end: the end page
1217 * @for_alloc: if this is to populate for allocation
1218 *
1219 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1220 * the bookkeeping information accordingly. Must be called after each
1221 * successful population.
1222 *
1223 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1224 * is to serve an allocation in that area.
1225 */
1226static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1227 int page_end, bool for_alloc)
1228{
1229 int nr = page_end - page_start;
1230
1231 lockdep_assert_held(&pcpu_lock);
1232
1233 bitmap_set(chunk->populated, page_start, nr);
1234 chunk->nr_populated += nr;
1235
1236 if (!for_alloc) {
1237 chunk->nr_empty_pop_pages += nr;
1238 pcpu_nr_empty_pop_pages += nr;
1239 }
1240}
1241
1242/**
1243 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1244 * @chunk: pcpu_chunk which got depopulated
1245 * @page_start: the start page
1246 * @page_end: the end page
1247 *
1248 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1249 * Update the bookkeeping information accordingly. Must be called after
1250 * each successful depopulation.
1251 */
1252static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1253 int page_start, int page_end)
1254{
1255 int nr = page_end - page_start;
1256
1257 lockdep_assert_held(&pcpu_lock);
1258
1259 bitmap_clear(chunk->populated, page_start, nr);
1260 chunk->nr_populated -= nr;
1261 chunk->nr_empty_pop_pages -= nr;
1262 pcpu_nr_empty_pop_pages -= nr;
1263}
1264
1265/*
1266 * Chunk management implementation.
1267 *
1268 * To allow different implementations, chunk alloc/free and
1269 * [de]population are implemented in a separate file which is pulled
1270 * into this file and compiled together. The following functions
1271 * should be implemented.
1272 *
1273 * pcpu_populate_chunk - populate the specified range of a chunk
1274 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1275 * pcpu_create_chunk - create a new chunk
1276 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1277 * pcpu_addr_to_page - translate address to physical address
1278 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1279 */
1280static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1281 int page_start, int page_end, gfp_t gfp);
1282static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1283 int page_start, int page_end);
1284static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1285static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1286static struct page *pcpu_addr_to_page(void *addr);
1287static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1288
1289#ifdef CONFIG_NEED_PER_CPU_KM
1290#include "percpu-km.c"
1291#else
1292#include "percpu-vm.c"
1293#endif
1294
1295/**
1296 * pcpu_chunk_addr_search - determine chunk containing specified address
1297 * @addr: address for which the chunk needs to be determined.
1298 *
1299 * This is an internal function that handles all but static allocations.
1300 * Static percpu address values should never be passed into the allocator.
1301 *
1302 * RETURNS:
1303 * The address of the found chunk.
1304 */
1305static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1306{
1307 /* is it in the dynamic region (first chunk)? */
1308 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1309 return pcpu_first_chunk;
1310
1311 /* is it in the reserved region? */
1312 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1313 return pcpu_reserved_chunk;
1314
1315 /*
1316 * The address is relative to unit0 which might be unused and
1317 * thus unmapped. Offset the address to the unit space of the
1318 * current processor before looking it up in the vmalloc
1319 * space. Note that any possible cpu id can be used here, so
1320 * there's no need to worry about preemption or cpu hotplug.
1321 */
1322 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1323 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1324}
1325
1326/**
1327 * pcpu_alloc - the percpu allocator
1328 * @size: size of area to allocate in bytes
1329 * @align: alignment of area (max PAGE_SIZE)
1330 * @reserved: allocate from the reserved chunk if available
1331 * @gfp: allocation flags
1332 *
1333 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1334 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1335 * then no warning will be triggered on invalid or failed allocation
1336 * requests.
1337 *
1338 * RETURNS:
1339 * Percpu pointer to the allocated area on success, NULL on failure.
1340 */
1341static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1342 gfp_t gfp)
1343{
1344 /* whitelisted flags that can be passed to the backing allocators */
1345 gfp_t pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1346 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1347 bool do_warn = !(gfp & __GFP_NOWARN);
1348 static int warn_limit = 10;
1349 struct pcpu_chunk *chunk;
1350 const char *err;
1351 int slot, off, cpu, ret;
1352 unsigned long flags;
1353 void __percpu *ptr;
1354 size_t bits, bit_align;
1355
1356 /*
1357 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1358 * therefore alignment must be a minimum of that many bytes.
1359 * An allocation may have internal fragmentation from rounding up
1360 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1361 */
1362 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1363 align = PCPU_MIN_ALLOC_SIZE;
1364
1365 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1366 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1367 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1368
1369 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1370 !is_power_of_2(align))) {
1371 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1372 size, align);
1373 return NULL;
1374 }
1375
1376 if (!is_atomic) {
1377 /*
1378 * pcpu_balance_workfn() allocates memory under this mutex,
1379 * and it may wait for memory reclaim. Allow current task
1380 * to become OOM victim, in case of memory pressure.
1381 */
1382 if (gfp & __GFP_NOFAIL)
1383 mutex_lock(&pcpu_alloc_mutex);
1384 else if (mutex_lock_killable(&pcpu_alloc_mutex))
1385 return NULL;
1386 }
1387
1388 spin_lock_irqsave(&pcpu_lock, flags);
1389
1390 /* serve reserved allocations from the reserved chunk if available */
1391 if (reserved && pcpu_reserved_chunk) {
1392 chunk = pcpu_reserved_chunk;
1393
1394 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1395 if (off < 0) {
1396 err = "alloc from reserved chunk failed";
1397 goto fail_unlock;
1398 }
1399
1400 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1401 if (off >= 0)
1402 goto area_found;
1403
1404 err = "alloc from reserved chunk failed";
1405 goto fail_unlock;
1406 }
1407
1408restart:
1409 /* search through normal chunks */
1410 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1411 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1412 off = pcpu_find_block_fit(chunk, bits, bit_align,
1413 is_atomic);
1414 if (off < 0)
1415 continue;
1416
1417 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1418 if (off >= 0)
1419 goto area_found;
1420
1421 }
1422 }
1423
1424 spin_unlock_irqrestore(&pcpu_lock, flags);
1425
1426 /*
1427 * No space left. Create a new chunk. We don't want multiple
1428 * tasks to create chunks simultaneously. Serialize and create iff
1429 * there's still no empty chunk after grabbing the mutex.
1430 */
1431 if (is_atomic) {
1432 err = "atomic alloc failed, no space left";
1433 goto fail;
1434 }
1435
1436 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1437 chunk = pcpu_create_chunk(pcpu_gfp);
1438 if (!chunk) {
1439 err = "failed to allocate new chunk";
1440 goto fail;
1441 }
1442
1443 spin_lock_irqsave(&pcpu_lock, flags);
1444 pcpu_chunk_relocate(chunk, -1);
1445 } else {
1446 spin_lock_irqsave(&pcpu_lock, flags);
1447 }
1448
1449 goto restart;
1450
1451area_found:
1452 pcpu_stats_area_alloc(chunk, size);
1453 spin_unlock_irqrestore(&pcpu_lock, flags);
1454
1455 /* populate if not all pages are already there */
1456 if (!is_atomic) {
1457 int page_start, page_end, rs, re;
1458
1459 page_start = PFN_DOWN(off);
1460 page_end = PFN_UP(off + size);
1461
1462 pcpu_for_each_unpop_region(chunk->populated, rs, re,
1463 page_start, page_end) {
1464 WARN_ON(chunk->immutable);
1465
1466 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1467
1468 spin_lock_irqsave(&pcpu_lock, flags);
1469 if (ret) {
1470 pcpu_free_area(chunk, off);
1471 err = "failed to populate";
1472 goto fail_unlock;
1473 }
1474 pcpu_chunk_populated(chunk, rs, re, true);
1475 spin_unlock_irqrestore(&pcpu_lock, flags);
1476 }
1477
1478 mutex_unlock(&pcpu_alloc_mutex);
1479 }
1480
1481 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1482 pcpu_schedule_balance_work();
1483
1484 /* clear the areas and return address relative to base address */
1485 for_each_possible_cpu(cpu)
1486 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1487
1488 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1489 kmemleak_alloc_percpu(ptr, size, gfp);
1490
1491 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1492 chunk->base_addr, off, ptr);
1493
1494 return ptr;
1495
1496fail_unlock:
1497 spin_unlock_irqrestore(&pcpu_lock, flags);
1498fail:
1499 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1500
1501 if (!is_atomic && do_warn && warn_limit) {
1502 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1503 size, align, is_atomic, err);
1504 dump_stack();
1505 if (!--warn_limit)
1506 pr_info("limit reached, disable warning\n");
1507 }
1508 if (is_atomic) {
1509 /* see the flag handling in pcpu_blance_workfn() */
1510 pcpu_atomic_alloc_failed = true;
1511 pcpu_schedule_balance_work();
1512 } else {
1513 mutex_unlock(&pcpu_alloc_mutex);
1514 }
1515 return NULL;
1516}
1517
1518/**
1519 * __alloc_percpu_gfp - allocate dynamic percpu area
1520 * @size: size of area to allocate in bytes
1521 * @align: alignment of area (max PAGE_SIZE)
1522 * @gfp: allocation flags
1523 *
1524 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1525 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1526 * be called from any context but is a lot more likely to fail. If @gfp
1527 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1528 * allocation requests.
1529 *
1530 * RETURNS:
1531 * Percpu pointer to the allocated area on success, NULL on failure.
1532 */
1533void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1534{
1535 return pcpu_alloc(size, align, false, gfp);
1536}
1537EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1538
1539/**
1540 * __alloc_percpu - allocate dynamic percpu area
1541 * @size: size of area to allocate in bytes
1542 * @align: alignment of area (max PAGE_SIZE)
1543 *
1544 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1545 */
1546void __percpu *__alloc_percpu(size_t size, size_t align)
1547{
1548 return pcpu_alloc(size, align, false, GFP_KERNEL);
1549}
1550EXPORT_SYMBOL_GPL(__alloc_percpu);
1551
1552/**
1553 * __alloc_reserved_percpu - allocate reserved percpu area
1554 * @size: size of area to allocate in bytes
1555 * @align: alignment of area (max PAGE_SIZE)
1556 *
1557 * Allocate zero-filled percpu area of @size bytes aligned at @align
1558 * from reserved percpu area if arch has set it up; otherwise,
1559 * allocation is served from the same dynamic area. Might sleep.
1560 * Might trigger writeouts.
1561 *
1562 * CONTEXT:
1563 * Does GFP_KERNEL allocation.
1564 *
1565 * RETURNS:
1566 * Percpu pointer to the allocated area on success, NULL on failure.
1567 */
1568void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1569{
1570 return pcpu_alloc(size, align, true, GFP_KERNEL);
1571}
1572
1573/**
1574 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1575 * @work: unused
1576 *
1577 * Reclaim all fully free chunks except for the first one. This is also
1578 * responsible for maintaining the pool of empty populated pages. However,
1579 * it is possible that this is called when physical memory is scarce causing
1580 * OOM killer to be triggered. We should avoid doing so until an actual
1581 * allocation causes the failure as it is possible that requests can be
1582 * serviced from already backed regions.
1583 */
1584static void pcpu_balance_workfn(struct work_struct *work)
1585{
1586 /* gfp flags passed to underlying allocators */
1587 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
1588 LIST_HEAD(to_free);
1589 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1590 struct pcpu_chunk *chunk, *next;
1591 int slot, nr_to_pop, ret;
1592
1593 /*
1594 * There's no reason to keep around multiple unused chunks and VM
1595 * areas can be scarce. Destroy all free chunks except for one.
1596 */
1597 mutex_lock(&pcpu_alloc_mutex);
1598 spin_lock_irq(&pcpu_lock);
1599
1600 list_for_each_entry_safe(chunk, next, free_head, list) {
1601 WARN_ON(chunk->immutable);
1602
1603 /* spare the first one */
1604 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1605 continue;
1606
1607 list_move(&chunk->list, &to_free);
1608 }
1609
1610 spin_unlock_irq(&pcpu_lock);
1611
1612 list_for_each_entry_safe(chunk, next, &to_free, list) {
1613 int rs, re;
1614
1615 pcpu_for_each_pop_region(chunk->populated, rs, re, 0,
1616 chunk->nr_pages) {
1617 pcpu_depopulate_chunk(chunk, rs, re);
1618 spin_lock_irq(&pcpu_lock);
1619 pcpu_chunk_depopulated(chunk, rs, re);
1620 spin_unlock_irq(&pcpu_lock);
1621 }
1622 pcpu_destroy_chunk(chunk);
1623 cond_resched();
1624 }
1625
1626 /*
1627 * Ensure there are certain number of free populated pages for
1628 * atomic allocs. Fill up from the most packed so that atomic
1629 * allocs don't increase fragmentation. If atomic allocation
1630 * failed previously, always populate the maximum amount. This
1631 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1632 * failing indefinitely; however, large atomic allocs are not
1633 * something we support properly and can be highly unreliable and
1634 * inefficient.
1635 */
1636retry_pop:
1637 if (pcpu_atomic_alloc_failed) {
1638 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1639 /* best effort anyway, don't worry about synchronization */
1640 pcpu_atomic_alloc_failed = false;
1641 } else {
1642 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1643 pcpu_nr_empty_pop_pages,
1644 0, PCPU_EMPTY_POP_PAGES_HIGH);
1645 }
1646
1647 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1648 int nr_unpop = 0, rs, re;
1649
1650 if (!nr_to_pop)
1651 break;
1652
1653 spin_lock_irq(&pcpu_lock);
1654 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1655 nr_unpop = chunk->nr_pages - chunk->nr_populated;
1656 if (nr_unpop)
1657 break;
1658 }
1659 spin_unlock_irq(&pcpu_lock);
1660
1661 if (!nr_unpop)
1662 continue;
1663
1664 /* @chunk can't go away while pcpu_alloc_mutex is held */
1665 pcpu_for_each_unpop_region(chunk->populated, rs, re, 0,
1666 chunk->nr_pages) {
1667 int nr = min(re - rs, nr_to_pop);
1668
1669 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
1670 if (!ret) {
1671 nr_to_pop -= nr;
1672 spin_lock_irq(&pcpu_lock);
1673 pcpu_chunk_populated(chunk, rs, rs + nr, false);
1674 spin_unlock_irq(&pcpu_lock);
1675 } else {
1676 nr_to_pop = 0;
1677 }
1678
1679 if (!nr_to_pop)
1680 break;
1681 }
1682 }
1683
1684 if (nr_to_pop) {
1685 /* ran out of chunks to populate, create a new one and retry */
1686 chunk = pcpu_create_chunk(gfp);
1687 if (chunk) {
1688 spin_lock_irq(&pcpu_lock);
1689 pcpu_chunk_relocate(chunk, -1);
1690 spin_unlock_irq(&pcpu_lock);
1691 goto retry_pop;
1692 }
1693 }
1694
1695 mutex_unlock(&pcpu_alloc_mutex);
1696}
1697
1698/**
1699 * free_percpu - free percpu area
1700 * @ptr: pointer to area to free
1701 *
1702 * Free percpu area @ptr.
1703 *
1704 * CONTEXT:
1705 * Can be called from atomic context.
1706 */
1707void free_percpu(void __percpu *ptr)
1708{
1709 void *addr;
1710 struct pcpu_chunk *chunk;
1711 unsigned long flags;
1712 int off;
1713
1714 if (!ptr)
1715 return;
1716
1717 kmemleak_free_percpu(ptr);
1718
1719 addr = __pcpu_ptr_to_addr(ptr);
1720
1721 spin_lock_irqsave(&pcpu_lock, flags);
1722
1723 chunk = pcpu_chunk_addr_search(addr);
1724 off = addr - chunk->base_addr;
1725
1726 pcpu_free_area(chunk, off);
1727
1728 /* if there are more than one fully free chunks, wake up grim reaper */
1729 if (chunk->free_bytes == pcpu_unit_size) {
1730 struct pcpu_chunk *pos;
1731
1732 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1733 if (pos != chunk) {
1734 pcpu_schedule_balance_work();
1735 break;
1736 }
1737 }
1738
1739 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1740
1741 spin_unlock_irqrestore(&pcpu_lock, flags);
1742}
1743EXPORT_SYMBOL_GPL(free_percpu);
1744
1745bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1746{
1747#ifdef CONFIG_SMP
1748 const size_t static_size = __per_cpu_end - __per_cpu_start;
1749 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1750 unsigned int cpu;
1751
1752 for_each_possible_cpu(cpu) {
1753 void *start = per_cpu_ptr(base, cpu);
1754 void *va = (void *)addr;
1755
1756 if (va >= start && va < start + static_size) {
1757 if (can_addr) {
1758 *can_addr = (unsigned long) (va - start);
1759 *can_addr += (unsigned long)
1760 per_cpu_ptr(base, get_boot_cpu_id());
1761 }
1762 return true;
1763 }
1764 }
1765#endif
1766 /* on UP, can't distinguish from other static vars, always false */
1767 return false;
1768}
1769
1770/**
1771 * is_kernel_percpu_address - test whether address is from static percpu area
1772 * @addr: address to test
1773 *
1774 * Test whether @addr belongs to in-kernel static percpu area. Module
1775 * static percpu areas are not considered. For those, use
1776 * is_module_percpu_address().
1777 *
1778 * RETURNS:
1779 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1780 */
1781bool is_kernel_percpu_address(unsigned long addr)
1782{
1783 return __is_kernel_percpu_address(addr, NULL);
1784}
1785
1786/**
1787 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1788 * @addr: the address to be converted to physical address
1789 *
1790 * Given @addr which is dereferenceable address obtained via one of
1791 * percpu access macros, this function translates it into its physical
1792 * address. The caller is responsible for ensuring @addr stays valid
1793 * until this function finishes.
1794 *
1795 * percpu allocator has special setup for the first chunk, which currently
1796 * supports either embedding in linear address space or vmalloc mapping,
1797 * and, from the second one, the backing allocator (currently either vm or
1798 * km) provides translation.
1799 *
1800 * The addr can be translated simply without checking if it falls into the
1801 * first chunk. But the current code reflects better how percpu allocator
1802 * actually works, and the verification can discover both bugs in percpu
1803 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1804 * code.
1805 *
1806 * RETURNS:
1807 * The physical address for @addr.
1808 */
1809phys_addr_t per_cpu_ptr_to_phys(void *addr)
1810{
1811 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1812 bool in_first_chunk = false;
1813 unsigned long first_low, first_high;
1814 unsigned int cpu;
1815
1816 /*
1817 * The following test on unit_low/high isn't strictly
1818 * necessary but will speed up lookups of addresses which
1819 * aren't in the first chunk.
1820 *
1821 * The address check is against full chunk sizes. pcpu_base_addr
1822 * points to the beginning of the first chunk including the
1823 * static region. Assumes good intent as the first chunk may
1824 * not be full (ie. < pcpu_unit_pages in size).
1825 */
1826 first_low = (unsigned long)pcpu_base_addr +
1827 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
1828 first_high = (unsigned long)pcpu_base_addr +
1829 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
1830 if ((unsigned long)addr >= first_low &&
1831 (unsigned long)addr < first_high) {
1832 for_each_possible_cpu(cpu) {
1833 void *start = per_cpu_ptr(base, cpu);
1834
1835 if (addr >= start && addr < start + pcpu_unit_size) {
1836 in_first_chunk = true;
1837 break;
1838 }
1839 }
1840 }
1841
1842 if (in_first_chunk) {
1843 if (!is_vmalloc_addr(addr))
1844 return __pa(addr);
1845 else
1846 return page_to_phys(vmalloc_to_page(addr)) +
1847 offset_in_page(addr);
1848 } else
1849 return page_to_phys(pcpu_addr_to_page(addr)) +
1850 offset_in_page(addr);
1851}
1852
1853/**
1854 * pcpu_alloc_alloc_info - allocate percpu allocation info
1855 * @nr_groups: the number of groups
1856 * @nr_units: the number of units
1857 *
1858 * Allocate ai which is large enough for @nr_groups groups containing
1859 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1860 * cpu_map array which is long enough for @nr_units and filled with
1861 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1862 * pointer of other groups.
1863 *
1864 * RETURNS:
1865 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1866 * failure.
1867 */
1868struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1869 int nr_units)
1870{
1871 struct pcpu_alloc_info *ai;
1872 size_t base_size, ai_size;
1873 void *ptr;
1874 int unit;
1875
1876 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1877 __alignof__(ai->groups[0].cpu_map[0]));
1878 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1879
1880 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), PAGE_SIZE);
1881 if (!ptr)
1882 return NULL;
1883 ai = ptr;
1884 ptr += base_size;
1885
1886 ai->groups[0].cpu_map = ptr;
1887
1888 for (unit = 0; unit < nr_units; unit++)
1889 ai->groups[0].cpu_map[unit] = NR_CPUS;
1890
1891 ai->nr_groups = nr_groups;
1892 ai->__ai_size = PFN_ALIGN(ai_size);
1893
1894 return ai;
1895}
1896
1897/**
1898 * pcpu_free_alloc_info - free percpu allocation info
1899 * @ai: pcpu_alloc_info to free
1900 *
1901 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1902 */
1903void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1904{
1905 memblock_free_early(__pa(ai), ai->__ai_size);
1906}
1907
1908/**
1909 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1910 * @lvl: loglevel
1911 * @ai: allocation info to dump
1912 *
1913 * Print out information about @ai using loglevel @lvl.
1914 */
1915static void pcpu_dump_alloc_info(const char *lvl,
1916 const struct pcpu_alloc_info *ai)
1917{
1918 int group_width = 1, cpu_width = 1, width;
1919 char empty_str[] = "--------";
1920 int alloc = 0, alloc_end = 0;
1921 int group, v;
1922 int upa, apl; /* units per alloc, allocs per line */
1923
1924 v = ai->nr_groups;
1925 while (v /= 10)
1926 group_width++;
1927
1928 v = num_possible_cpus();
1929 while (v /= 10)
1930 cpu_width++;
1931 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1932
1933 upa = ai->alloc_size / ai->unit_size;
1934 width = upa * (cpu_width + 1) + group_width + 3;
1935 apl = rounddown_pow_of_two(max(60 / width, 1));
1936
1937 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1938 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1939 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1940
1941 for (group = 0; group < ai->nr_groups; group++) {
1942 const struct pcpu_group_info *gi = &ai->groups[group];
1943 int unit = 0, unit_end = 0;
1944
1945 BUG_ON(gi->nr_units % upa);
1946 for (alloc_end += gi->nr_units / upa;
1947 alloc < alloc_end; alloc++) {
1948 if (!(alloc % apl)) {
1949 pr_cont("\n");
1950 printk("%spcpu-alloc: ", lvl);
1951 }
1952 pr_cont("[%0*d] ", group_width, group);
1953
1954 for (unit_end += upa; unit < unit_end; unit++)
1955 if (gi->cpu_map[unit] != NR_CPUS)
1956 pr_cont("%0*d ",
1957 cpu_width, gi->cpu_map[unit]);
1958 else
1959 pr_cont("%s ", empty_str);
1960 }
1961 }
1962 pr_cont("\n");
1963}
1964
1965/**
1966 * pcpu_setup_first_chunk - initialize the first percpu chunk
1967 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1968 * @base_addr: mapped address
1969 *
1970 * Initialize the first percpu chunk which contains the kernel static
1971 * perpcu area. This function is to be called from arch percpu area
1972 * setup path.
1973 *
1974 * @ai contains all information necessary to initialize the first
1975 * chunk and prime the dynamic percpu allocator.
1976 *
1977 * @ai->static_size is the size of static percpu area.
1978 *
1979 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1980 * reserve after the static area in the first chunk. This reserves
1981 * the first chunk such that it's available only through reserved
1982 * percpu allocation. This is primarily used to serve module percpu
1983 * static areas on architectures where the addressing model has
1984 * limited offset range for symbol relocations to guarantee module
1985 * percpu symbols fall inside the relocatable range.
1986 *
1987 * @ai->dyn_size determines the number of bytes available for dynamic
1988 * allocation in the first chunk. The area between @ai->static_size +
1989 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1990 *
1991 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1992 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1993 * @ai->dyn_size.
1994 *
1995 * @ai->atom_size is the allocation atom size and used as alignment
1996 * for vm areas.
1997 *
1998 * @ai->alloc_size is the allocation size and always multiple of
1999 * @ai->atom_size. This is larger than @ai->atom_size if
2000 * @ai->unit_size is larger than @ai->atom_size.
2001 *
2002 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2003 * percpu areas. Units which should be colocated are put into the
2004 * same group. Dynamic VM areas will be allocated according to these
2005 * groupings. If @ai->nr_groups is zero, a single group containing
2006 * all units is assumed.
2007 *
2008 * The caller should have mapped the first chunk at @base_addr and
2009 * copied static data to each unit.
2010 *
2011 * The first chunk will always contain a static and a dynamic region.
2012 * However, the static region is not managed by any chunk. If the first
2013 * chunk also contains a reserved region, it is served by two chunks -
2014 * one for the reserved region and one for the dynamic region. They
2015 * share the same vm, but use offset regions in the area allocation map.
2016 * The chunk serving the dynamic region is circulated in the chunk slots
2017 * and available for dynamic allocation like any other chunk.
2018 *
2019 * RETURNS:
2020 * 0 on success, -errno on failure.
2021 */
2022int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2023 void *base_addr)
2024{
2025 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2026 size_t static_size, dyn_size;
2027 struct pcpu_chunk *chunk;
2028 unsigned long *group_offsets;
2029 size_t *group_sizes;
2030 unsigned long *unit_off;
2031 unsigned int cpu;
2032 int *unit_map;
2033 int group, unit, i;
2034 int map_size;
2035 unsigned long tmp_addr;
2036
2037#define PCPU_SETUP_BUG_ON(cond) do { \
2038 if (unlikely(cond)) { \
2039 pr_emerg("failed to initialize, %s\n", #cond); \
2040 pr_emerg("cpu_possible_mask=%*pb\n", \
2041 cpumask_pr_args(cpu_possible_mask)); \
2042 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2043 BUG(); \
2044 } \
2045} while (0)
2046
2047 /* sanity checks */
2048 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2049#ifdef CONFIG_SMP
2050 PCPU_SETUP_BUG_ON(!ai->static_size);
2051 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2052#endif
2053 PCPU_SETUP_BUG_ON(!base_addr);
2054 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2055 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2056 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2057 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2058 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2059 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2060 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2061 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2062 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2063 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2064 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2065
2066 /* process group information and build config tables accordingly */
2067 group_offsets = memblock_virt_alloc(ai->nr_groups *
2068 sizeof(group_offsets[0]), 0);
2069 group_sizes = memblock_virt_alloc(ai->nr_groups *
2070 sizeof(group_sizes[0]), 0);
2071 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
2072 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
2073
2074 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2075 unit_map[cpu] = UINT_MAX;
2076
2077 pcpu_low_unit_cpu = NR_CPUS;
2078 pcpu_high_unit_cpu = NR_CPUS;
2079
2080 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2081 const struct pcpu_group_info *gi = &ai->groups[group];
2082
2083 group_offsets[group] = gi->base_offset;
2084 group_sizes[group] = gi->nr_units * ai->unit_size;
2085
2086 for (i = 0; i < gi->nr_units; i++) {
2087 cpu = gi->cpu_map[i];
2088 if (cpu == NR_CPUS)
2089 continue;
2090
2091 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2092 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2093 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2094
2095 unit_map[cpu] = unit + i;
2096 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2097
2098 /* determine low/high unit_cpu */
2099 if (pcpu_low_unit_cpu == NR_CPUS ||
2100 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2101 pcpu_low_unit_cpu = cpu;
2102 if (pcpu_high_unit_cpu == NR_CPUS ||
2103 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2104 pcpu_high_unit_cpu = cpu;
2105 }
2106 }
2107 pcpu_nr_units = unit;
2108
2109 for_each_possible_cpu(cpu)
2110 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2111
2112 /* we're done parsing the input, undefine BUG macro and dump config */
2113#undef PCPU_SETUP_BUG_ON
2114 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2115
2116 pcpu_nr_groups = ai->nr_groups;
2117 pcpu_group_offsets = group_offsets;
2118 pcpu_group_sizes = group_sizes;
2119 pcpu_unit_map = unit_map;
2120 pcpu_unit_offsets = unit_off;
2121
2122 /* determine basic parameters */
2123 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2124 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2125 pcpu_atom_size = ai->atom_size;
2126 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2127 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2128
2129 pcpu_stats_save_ai(ai);
2130
2131 /*
2132 * Allocate chunk slots. The additional last slot is for
2133 * empty chunks.
2134 */
2135 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2136 pcpu_slot = memblock_virt_alloc(
2137 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
2138 for (i = 0; i < pcpu_nr_slots; i++)
2139 INIT_LIST_HEAD(&pcpu_slot[i]);
2140
2141 /*
2142 * The end of the static region needs to be aligned with the
2143 * minimum allocation size as this offsets the reserved and
2144 * dynamic region. The first chunk ends page aligned by
2145 * expanding the dynamic region, therefore the dynamic region
2146 * can be shrunk to compensate while still staying above the
2147 * configured sizes.
2148 */
2149 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2150 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2151
2152 /*
2153 * Initialize first chunk.
2154 * If the reserved_size is non-zero, this initializes the reserved
2155 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2156 * and the dynamic region is initialized here. The first chunk,
2157 * pcpu_first_chunk, will always point to the chunk that serves
2158 * the dynamic region.
2159 */
2160 tmp_addr = (unsigned long)base_addr + static_size;
2161 map_size = ai->reserved_size ?: dyn_size;
2162 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2163
2164 /* init dynamic chunk if necessary */
2165 if (ai->reserved_size) {
2166 pcpu_reserved_chunk = chunk;
2167
2168 tmp_addr = (unsigned long)base_addr + static_size +
2169 ai->reserved_size;
2170 map_size = dyn_size;
2171 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2172 }
2173
2174 /* link the first chunk in */
2175 pcpu_first_chunk = chunk;
2176 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2177 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2178
2179 pcpu_stats_chunk_alloc();
2180 trace_percpu_create_chunk(base_addr);
2181
2182 /* we're done */
2183 pcpu_base_addr = base_addr;
2184 return 0;
2185}
2186
2187#ifdef CONFIG_SMP
2188
2189const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2190 [PCPU_FC_AUTO] = "auto",
2191 [PCPU_FC_EMBED] = "embed",
2192 [PCPU_FC_PAGE] = "page",
2193};
2194
2195enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2196
2197static int __init percpu_alloc_setup(char *str)
2198{
2199 if (!str)
2200 return -EINVAL;
2201
2202 if (0)
2203 /* nada */;
2204#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2205 else if (!strcmp(str, "embed"))
2206 pcpu_chosen_fc = PCPU_FC_EMBED;
2207#endif
2208#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2209 else if (!strcmp(str, "page"))
2210 pcpu_chosen_fc = PCPU_FC_PAGE;
2211#endif
2212 else
2213 pr_warn("unknown allocator %s specified\n", str);
2214
2215 return 0;
2216}
2217early_param("percpu_alloc", percpu_alloc_setup);
2218
2219/*
2220 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2221 * Build it if needed by the arch config or the generic setup is going
2222 * to be used.
2223 */
2224#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2225 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2226#define BUILD_EMBED_FIRST_CHUNK
2227#endif
2228
2229/* build pcpu_page_first_chunk() iff needed by the arch config */
2230#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2231#define BUILD_PAGE_FIRST_CHUNK
2232#endif
2233
2234/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2235#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2236/**
2237 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2238 * @reserved_size: the size of reserved percpu area in bytes
2239 * @dyn_size: minimum free size for dynamic allocation in bytes
2240 * @atom_size: allocation atom size
2241 * @cpu_distance_fn: callback to determine distance between cpus, optional
2242 *
2243 * This function determines grouping of units, their mappings to cpus
2244 * and other parameters considering needed percpu size, allocation
2245 * atom size and distances between CPUs.
2246 *
2247 * Groups are always multiples of atom size and CPUs which are of
2248 * LOCAL_DISTANCE both ways are grouped together and share space for
2249 * units in the same group. The returned configuration is guaranteed
2250 * to have CPUs on different nodes on different groups and >=75% usage
2251 * of allocated virtual address space.
2252 *
2253 * RETURNS:
2254 * On success, pointer to the new allocation_info is returned. On
2255 * failure, ERR_PTR value is returned.
2256 */
2257static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2258 size_t reserved_size, size_t dyn_size,
2259 size_t atom_size,
2260 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2261{
2262 static int group_map[NR_CPUS] __initdata;
2263 static int group_cnt[NR_CPUS] __initdata;
2264 const size_t static_size = __per_cpu_end - __per_cpu_start;
2265 int nr_groups = 1, nr_units = 0;
2266 size_t size_sum, min_unit_size, alloc_size;
2267 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
2268 int last_allocs, group, unit;
2269 unsigned int cpu, tcpu;
2270 struct pcpu_alloc_info *ai;
2271 unsigned int *cpu_map;
2272
2273 /* this function may be called multiple times */
2274 memset(group_map, 0, sizeof(group_map));
2275 memset(group_cnt, 0, sizeof(group_cnt));
2276
2277 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2278 size_sum = PFN_ALIGN(static_size + reserved_size +
2279 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2280 dyn_size = size_sum - static_size - reserved_size;
2281
2282 /*
2283 * Determine min_unit_size, alloc_size and max_upa such that
2284 * alloc_size is multiple of atom_size and is the smallest
2285 * which can accommodate 4k aligned segments which are equal to
2286 * or larger than min_unit_size.
2287 */
2288 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2289
2290 /* determine the maximum # of units that can fit in an allocation */
2291 alloc_size = roundup(min_unit_size, atom_size);
2292 upa = alloc_size / min_unit_size;
2293 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2294 upa--;
2295 max_upa = upa;
2296
2297 /* group cpus according to their proximity */
2298 for_each_possible_cpu(cpu) {
2299 group = 0;
2300 next_group:
2301 for_each_possible_cpu(tcpu) {
2302 if (cpu == tcpu)
2303 break;
2304 if (group_map[tcpu] == group && cpu_distance_fn &&
2305 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2306 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2307 group++;
2308 nr_groups = max(nr_groups, group + 1);
2309 goto next_group;
2310 }
2311 }
2312 group_map[cpu] = group;
2313 group_cnt[group]++;
2314 }
2315
2316 /*
2317 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2318 * Expand the unit_size until we use >= 75% of the units allocated.
2319 * Related to atom_size, which could be much larger than the unit_size.
2320 */
2321 last_allocs = INT_MAX;
2322 for (upa = max_upa; upa; upa--) {
2323 int allocs = 0, wasted = 0;
2324
2325 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2326 continue;
2327
2328 for (group = 0; group < nr_groups; group++) {
2329 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2330 allocs += this_allocs;
2331 wasted += this_allocs * upa - group_cnt[group];
2332 }
2333
2334 /*
2335 * Don't accept if wastage is over 1/3. The
2336 * greater-than comparison ensures upa==1 always
2337 * passes the following check.
2338 */
2339 if (wasted > num_possible_cpus() / 3)
2340 continue;
2341
2342 /* and then don't consume more memory */
2343 if (allocs > last_allocs)
2344 break;
2345 last_allocs = allocs;
2346 best_upa = upa;
2347 }
2348 upa = best_upa;
2349
2350 /* allocate and fill alloc_info */
2351 for (group = 0; group < nr_groups; group++)
2352 nr_units += roundup(group_cnt[group], upa);
2353
2354 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2355 if (!ai)
2356 return ERR_PTR(-ENOMEM);
2357 cpu_map = ai->groups[0].cpu_map;
2358
2359 for (group = 0; group < nr_groups; group++) {
2360 ai->groups[group].cpu_map = cpu_map;
2361 cpu_map += roundup(group_cnt[group], upa);
2362 }
2363
2364 ai->static_size = static_size;
2365 ai->reserved_size = reserved_size;
2366 ai->dyn_size = dyn_size;
2367 ai->unit_size = alloc_size / upa;
2368 ai->atom_size = atom_size;
2369 ai->alloc_size = alloc_size;
2370
2371 for (group = 0, unit = 0; group_cnt[group]; group++) {
2372 struct pcpu_group_info *gi = &ai->groups[group];
2373
2374 /*
2375 * Initialize base_offset as if all groups are located
2376 * back-to-back. The caller should update this to
2377 * reflect actual allocation.
2378 */
2379 gi->base_offset = unit * ai->unit_size;
2380
2381 for_each_possible_cpu(cpu)
2382 if (group_map[cpu] == group)
2383 gi->cpu_map[gi->nr_units++] = cpu;
2384 gi->nr_units = roundup(gi->nr_units, upa);
2385 unit += gi->nr_units;
2386 }
2387 BUG_ON(unit != nr_units);
2388
2389 return ai;
2390}
2391#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2392
2393#if defined(BUILD_EMBED_FIRST_CHUNK)
2394/**
2395 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2396 * @reserved_size: the size of reserved percpu area in bytes
2397 * @dyn_size: minimum free size for dynamic allocation in bytes
2398 * @atom_size: allocation atom size
2399 * @cpu_distance_fn: callback to determine distance between cpus, optional
2400 * @alloc_fn: function to allocate percpu page
2401 * @free_fn: function to free percpu page
2402 *
2403 * This is a helper to ease setting up embedded first percpu chunk and
2404 * can be called where pcpu_setup_first_chunk() is expected.
2405 *
2406 * If this function is used to setup the first chunk, it is allocated
2407 * by calling @alloc_fn and used as-is without being mapped into
2408 * vmalloc area. Allocations are always whole multiples of @atom_size
2409 * aligned to @atom_size.
2410 *
2411 * This enables the first chunk to piggy back on the linear physical
2412 * mapping which often uses larger page size. Please note that this
2413 * can result in very sparse cpu->unit mapping on NUMA machines thus
2414 * requiring large vmalloc address space. Don't use this allocator if
2415 * vmalloc space is not orders of magnitude larger than distances
2416 * between node memory addresses (ie. 32bit NUMA machines).
2417 *
2418 * @dyn_size specifies the minimum dynamic area size.
2419 *
2420 * If the needed size is smaller than the minimum or specified unit
2421 * size, the leftover is returned using @free_fn.
2422 *
2423 * RETURNS:
2424 * 0 on success, -errno on failure.
2425 */
2426int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2427 size_t atom_size,
2428 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2429 pcpu_fc_alloc_fn_t alloc_fn,
2430 pcpu_fc_free_fn_t free_fn)
2431{
2432 void *base = (void *)ULONG_MAX;
2433 void **areas = NULL;
2434 struct pcpu_alloc_info *ai;
2435 size_t size_sum, areas_size;
2436 unsigned long max_distance;
2437 int group, i, highest_group, rc;
2438
2439 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2440 cpu_distance_fn);
2441 if (IS_ERR(ai))
2442 return PTR_ERR(ai);
2443
2444 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2445 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2446
2447 areas = memblock_virt_alloc_nopanic(areas_size, 0);
2448 if (!areas) {
2449 rc = -ENOMEM;
2450 goto out_free;
2451 }
2452
2453 /* allocate, copy and determine base address & max_distance */
2454 highest_group = 0;
2455 for (group = 0; group < ai->nr_groups; group++) {
2456 struct pcpu_group_info *gi = &ai->groups[group];
2457 unsigned int cpu = NR_CPUS;
2458 void *ptr;
2459
2460 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2461 cpu = gi->cpu_map[i];
2462 BUG_ON(cpu == NR_CPUS);
2463
2464 /* allocate space for the whole group */
2465 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2466 if (!ptr) {
2467 rc = -ENOMEM;
2468 goto out_free_areas;
2469 }
2470 /* kmemleak tracks the percpu allocations separately */
2471 kmemleak_free(ptr);
2472 areas[group] = ptr;
2473
2474 base = min(ptr, base);
2475 if (ptr > areas[highest_group])
2476 highest_group = group;
2477 }
2478 max_distance = areas[highest_group] - base;
2479 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2480
2481 /* warn if maximum distance is further than 75% of vmalloc space */
2482 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2483 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2484 max_distance, VMALLOC_TOTAL);
2485#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2486 /* and fail if we have fallback */
2487 rc = -EINVAL;
2488 goto out_free_areas;
2489#endif
2490 }
2491
2492 /*
2493 * Copy data and free unused parts. This should happen after all
2494 * allocations are complete; otherwise, we may end up with
2495 * overlapping groups.
2496 */
2497 for (group = 0; group < ai->nr_groups; group++) {
2498 struct pcpu_group_info *gi = &ai->groups[group];
2499 void *ptr = areas[group];
2500
2501 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2502 if (gi->cpu_map[i] == NR_CPUS) {
2503 /* unused unit, free whole */
2504 free_fn(ptr, ai->unit_size);
2505 continue;
2506 }
2507 /* copy and return the unused part */
2508 memcpy(ptr, __per_cpu_load, ai->static_size);
2509 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2510 }
2511 }
2512
2513 /* base address is now known, determine group base offsets */
2514 for (group = 0; group < ai->nr_groups; group++) {
2515 ai->groups[group].base_offset = areas[group] - base;
2516 }
2517
2518 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2519 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2520 ai->dyn_size, ai->unit_size);
2521
2522 rc = pcpu_setup_first_chunk(ai, base);
2523 goto out_free;
2524
2525out_free_areas:
2526 for (group = 0; group < ai->nr_groups; group++)
2527 if (areas[group])
2528 free_fn(areas[group],
2529 ai->groups[group].nr_units * ai->unit_size);
2530out_free:
2531 pcpu_free_alloc_info(ai);
2532 if (areas)
2533 memblock_free_early(__pa(areas), areas_size);
2534 return rc;
2535}
2536#endif /* BUILD_EMBED_FIRST_CHUNK */
2537
2538#ifdef BUILD_PAGE_FIRST_CHUNK
2539/**
2540 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2541 * @reserved_size: the size of reserved percpu area in bytes
2542 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2543 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2544 * @populate_pte_fn: function to populate pte
2545 *
2546 * This is a helper to ease setting up page-remapped first percpu
2547 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2548 *
2549 * This is the basic allocator. Static percpu area is allocated
2550 * page-by-page into vmalloc area.
2551 *
2552 * RETURNS:
2553 * 0 on success, -errno on failure.
2554 */
2555int __init pcpu_page_first_chunk(size_t reserved_size,
2556 pcpu_fc_alloc_fn_t alloc_fn,
2557 pcpu_fc_free_fn_t free_fn,
2558 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2559{
2560 static struct vm_struct vm;
2561 struct pcpu_alloc_info *ai;
2562 char psize_str[16];
2563 int unit_pages;
2564 size_t pages_size;
2565 struct page **pages;
2566 int unit, i, j, rc;
2567 int upa;
2568 int nr_g0_units;
2569
2570 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2571
2572 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2573 if (IS_ERR(ai))
2574 return PTR_ERR(ai);
2575 BUG_ON(ai->nr_groups != 1);
2576 upa = ai->alloc_size/ai->unit_size;
2577 nr_g0_units = roundup(num_possible_cpus(), upa);
2578 if (unlikely(WARN_ON(ai->groups[0].nr_units != nr_g0_units))) {
2579 pcpu_free_alloc_info(ai);
2580 return -EINVAL;
2581 }
2582
2583 unit_pages = ai->unit_size >> PAGE_SHIFT;
2584
2585 /* unaligned allocations can't be freed, round up to page size */
2586 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2587 sizeof(pages[0]));
2588 pages = memblock_virt_alloc(pages_size, 0);
2589
2590 /* allocate pages */
2591 j = 0;
2592 for (unit = 0; unit < num_possible_cpus(); unit++) {
2593 unsigned int cpu = ai->groups[0].cpu_map[unit];
2594 for (i = 0; i < unit_pages; i++) {
2595 void *ptr;
2596
2597 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2598 if (!ptr) {
2599 pr_warn("failed to allocate %s page for cpu%u\n",
2600 psize_str, cpu);
2601 goto enomem;
2602 }
2603 /* kmemleak tracks the percpu allocations separately */
2604 kmemleak_free(ptr);
2605 pages[j++] = virt_to_page(ptr);
2606 }
2607 }
2608
2609 /* allocate vm area, map the pages and copy static data */
2610 vm.flags = VM_ALLOC;
2611 vm.size = num_possible_cpus() * ai->unit_size;
2612 vm_area_register_early(&vm, PAGE_SIZE);
2613
2614 for (unit = 0; unit < num_possible_cpus(); unit++) {
2615 unsigned long unit_addr =
2616 (unsigned long)vm.addr + unit * ai->unit_size;
2617
2618 for (i = 0; i < unit_pages; i++)
2619 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2620
2621 /* pte already populated, the following shouldn't fail */
2622 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2623 unit_pages);
2624 if (rc < 0)
2625 panic("failed to map percpu area, err=%d\n", rc);
2626
2627 /*
2628 * FIXME: Archs with virtual cache should flush local
2629 * cache for the linear mapping here - something
2630 * equivalent to flush_cache_vmap() on the local cpu.
2631 * flush_cache_vmap() can't be used as most supporting
2632 * data structures are not set up yet.
2633 */
2634
2635 /* copy static data */
2636 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2637 }
2638
2639 /* we're ready, commit */
2640 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2641 unit_pages, psize_str, vm.addr, ai->static_size,
2642 ai->reserved_size, ai->dyn_size);
2643
2644 rc = pcpu_setup_first_chunk(ai, vm.addr);
2645 goto out_free_ar;
2646
2647enomem:
2648 while (--j >= 0)
2649 free_fn(page_address(pages[j]), PAGE_SIZE);
2650 rc = -ENOMEM;
2651out_free_ar:
2652 memblock_free_early(__pa(pages), pages_size);
2653 pcpu_free_alloc_info(ai);
2654 return rc;
2655}
2656#endif /* BUILD_PAGE_FIRST_CHUNK */
2657
2658#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2659/*
2660 * Generic SMP percpu area setup.
2661 *
2662 * The embedding helper is used because its behavior closely resembles
2663 * the original non-dynamic generic percpu area setup. This is
2664 * important because many archs have addressing restrictions and might
2665 * fail if the percpu area is located far away from the previous
2666 * location. As an added bonus, in non-NUMA cases, embedding is
2667 * generally a good idea TLB-wise because percpu area can piggy back
2668 * on the physical linear memory mapping which uses large page
2669 * mappings on applicable archs.
2670 */
2671unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2672EXPORT_SYMBOL(__per_cpu_offset);
2673
2674static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2675 size_t align)
2676{
2677 return memblock_virt_alloc_from_nopanic(
2678 size, align, __pa(MAX_DMA_ADDRESS));
2679}
2680
2681static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2682{
2683 memblock_free_early(__pa(ptr), size);
2684}
2685
2686void __init setup_per_cpu_areas(void)
2687{
2688 unsigned long delta;
2689 unsigned int cpu;
2690 int rc;
2691
2692 /*
2693 * Always reserve area for module percpu variables. That's
2694 * what the legacy allocator did.
2695 */
2696 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2697 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2698 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2699 if (rc < 0)
2700 panic("Failed to initialize percpu areas.");
2701
2702 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2703 for_each_possible_cpu(cpu)
2704 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2705}
2706#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2707
2708#else /* CONFIG_SMP */
2709
2710/*
2711 * UP percpu area setup.
2712 *
2713 * UP always uses km-based percpu allocator with identity mapping.
2714 * Static percpu variables are indistinguishable from the usual static
2715 * variables and don't require any special preparation.
2716 */
2717void __init setup_per_cpu_areas(void)
2718{
2719 const size_t unit_size =
2720 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2721 PERCPU_DYNAMIC_RESERVE));
2722 struct pcpu_alloc_info *ai;
2723 void *fc;
2724
2725 ai = pcpu_alloc_alloc_info(1, 1);
2726 fc = memblock_virt_alloc_from_nopanic(unit_size,
2727 PAGE_SIZE,
2728 __pa(MAX_DMA_ADDRESS));
2729 if (!ai || !fc)
2730 panic("Failed to allocate memory for percpu areas.");
2731 /* kmemleak tracks the percpu allocations separately */
2732 kmemleak_free(fc);
2733
2734 ai->dyn_size = unit_size;
2735 ai->unit_size = unit_size;
2736 ai->atom_size = unit_size;
2737 ai->alloc_size = unit_size;
2738 ai->groups[0].nr_units = 1;
2739 ai->groups[0].cpu_map[0] = 0;
2740
2741 if (pcpu_setup_first_chunk(ai, fc) < 0)
2742 panic("Failed to initialize percpu areas.");
2743 pcpu_free_alloc_info(ai);
2744}
2745
2746#endif /* CONFIG_SMP */
2747
2748/*
2749 * Percpu allocator is initialized early during boot when neither slab or
2750 * workqueue is available. Plug async management until everything is up
2751 * and running.
2752 */
2753static int __init percpu_enable_async(void)
2754{
2755 pcpu_async_enabled = true;
2756 return 0;
2757}
2758subsys_initcall(percpu_enable_async);
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 followings.
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#include <linux/bitmap.h>
57#include <linux/bootmem.h>
58#include <linux/err.h>
59#include <linux/list.h>
60#include <linux/log2.h>
61#include <linux/mm.h>
62#include <linux/module.h>
63#include <linux/mutex.h>
64#include <linux/percpu.h>
65#include <linux/pfn.h>
66#include <linux/slab.h>
67#include <linux/spinlock.h>
68#include <linux/vmalloc.h>
69#include <linux/workqueue.h>
70#include <linux/kmemleak.h>
71
72#include <asm/cacheflush.h>
73#include <asm/sections.h>
74#include <asm/tlbflush.h>
75#include <asm/io.h>
76
77#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
79
80#ifdef CONFIG_SMP
81/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
82#ifndef __addr_to_pcpu_ptr
83#define __addr_to_pcpu_ptr(addr) \
84 (void __percpu *)((unsigned long)(addr) - \
85 (unsigned long)pcpu_base_addr + \
86 (unsigned long)__per_cpu_start)
87#endif
88#ifndef __pcpu_ptr_to_addr
89#define __pcpu_ptr_to_addr(ptr) \
90 (void __force *)((unsigned long)(ptr) + \
91 (unsigned long)pcpu_base_addr - \
92 (unsigned long)__per_cpu_start)
93#endif
94#else /* CONFIG_SMP */
95/* on UP, it's always identity mapped */
96#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
97#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
98#endif /* CONFIG_SMP */
99
100struct pcpu_chunk {
101 struct list_head list; /* linked to pcpu_slot lists */
102 int free_size; /* free bytes in the chunk */
103 int contig_hint; /* max contiguous size hint */
104 void *base_addr; /* base address of this chunk */
105 int map_used; /* # of map entries used before the sentry */
106 int map_alloc; /* # of map entries allocated */
107 int *map; /* allocation map */
108 void *data; /* chunk data */
109 int first_free; /* no free below this */
110 bool immutable; /* no [de]population allowed */
111 unsigned long populated[]; /* populated bitmap */
112};
113
114static int pcpu_unit_pages __read_mostly;
115static int pcpu_unit_size __read_mostly;
116static int pcpu_nr_units __read_mostly;
117static int pcpu_atom_size __read_mostly;
118static int pcpu_nr_slots __read_mostly;
119static size_t pcpu_chunk_struct_size __read_mostly;
120
121/* cpus with the lowest and highest unit addresses */
122static unsigned int pcpu_low_unit_cpu __read_mostly;
123static unsigned int pcpu_high_unit_cpu __read_mostly;
124
125/* the address of the first chunk which starts with the kernel static area */
126void *pcpu_base_addr __read_mostly;
127EXPORT_SYMBOL_GPL(pcpu_base_addr);
128
129static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
130const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
131
132/* group information, used for vm allocation */
133static int pcpu_nr_groups __read_mostly;
134static const unsigned long *pcpu_group_offsets __read_mostly;
135static const size_t *pcpu_group_sizes __read_mostly;
136
137/*
138 * The first chunk which always exists. Note that unlike other
139 * chunks, this one can be allocated and mapped in several different
140 * ways and thus often doesn't live in the vmalloc area.
141 */
142static struct pcpu_chunk *pcpu_first_chunk;
143
144/*
145 * Optional reserved chunk. This chunk reserves part of the first
146 * chunk and serves it for reserved allocations. The amount of
147 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
148 * area doesn't exist, the following variables contain NULL and 0
149 * respectively.
150 */
151static struct pcpu_chunk *pcpu_reserved_chunk;
152static int pcpu_reserved_chunk_limit;
153
154/*
155 * Synchronization rules.
156 *
157 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
158 * protects allocation/reclaim paths, chunks, populated bitmap and
159 * vmalloc mapping. The latter is a spinlock and protects the index
160 * data structures - chunk slots, chunks and area maps in chunks.
161 *
162 * During allocation, pcpu_alloc_mutex is kept locked all the time and
163 * pcpu_lock is grabbed and released as necessary. All actual memory
164 * allocations are done using GFP_KERNEL with pcpu_lock released. In
165 * general, percpu memory can't be allocated with irq off but
166 * irqsave/restore are still used in alloc path so that it can be used
167 * from early init path - sched_init() specifically.
168 *
169 * Free path accesses and alters only the index data structures, so it
170 * can be safely called from atomic context. When memory needs to be
171 * returned to the system, free path schedules reclaim_work which
172 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
173 * reclaimed, release both locks and frees the chunks. Note that it's
174 * necessary to grab both locks to remove a chunk from circulation as
175 * allocation path might be referencing the chunk with only
176 * pcpu_alloc_mutex locked.
177 */
178static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
179static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
180
181static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
182
183/* reclaim work to release fully free chunks, scheduled from free path */
184static void pcpu_reclaim(struct work_struct *work);
185static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
186
187static bool pcpu_addr_in_first_chunk(void *addr)
188{
189 void *first_start = pcpu_first_chunk->base_addr;
190
191 return addr >= first_start && addr < first_start + pcpu_unit_size;
192}
193
194static bool pcpu_addr_in_reserved_chunk(void *addr)
195{
196 void *first_start = pcpu_first_chunk->base_addr;
197
198 return addr >= first_start &&
199 addr < first_start + pcpu_reserved_chunk_limit;
200}
201
202static int __pcpu_size_to_slot(int size)
203{
204 int highbit = fls(size); /* size is in bytes */
205 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
206}
207
208static int pcpu_size_to_slot(int size)
209{
210 if (size == pcpu_unit_size)
211 return pcpu_nr_slots - 1;
212 return __pcpu_size_to_slot(size);
213}
214
215static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
216{
217 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
218 return 0;
219
220 return pcpu_size_to_slot(chunk->free_size);
221}
222
223/* set the pointer to a chunk in a page struct */
224static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
225{
226 page->index = (unsigned long)pcpu;
227}
228
229/* obtain pointer to a chunk from a page struct */
230static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
231{
232 return (struct pcpu_chunk *)page->index;
233}
234
235static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
236{
237 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
238}
239
240static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
241 unsigned int cpu, int page_idx)
242{
243 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
244 (page_idx << PAGE_SHIFT);
245}
246
247static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
248 int *rs, int *re, int end)
249{
250 *rs = find_next_zero_bit(chunk->populated, end, *rs);
251 *re = find_next_bit(chunk->populated, end, *rs + 1);
252}
253
254static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
255 int *rs, int *re, int end)
256{
257 *rs = find_next_bit(chunk->populated, end, *rs);
258 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
259}
260
261/*
262 * (Un)populated page region iterators. Iterate over (un)populated
263 * page regions between @start and @end in @chunk. @rs and @re should
264 * be integer variables and will be set to start and end page index of
265 * the current region.
266 */
267#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
268 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
269 (rs) < (re); \
270 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
271
272#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
274 (rs) < (re); \
275 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
276
277/**
278 * pcpu_mem_zalloc - allocate memory
279 * @size: bytes to allocate
280 *
281 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
282 * kzalloc() is used; otherwise, vzalloc() is used. The returned
283 * memory is always zeroed.
284 *
285 * CONTEXT:
286 * Does GFP_KERNEL allocation.
287 *
288 * RETURNS:
289 * Pointer to the allocated area on success, NULL on failure.
290 */
291static void *pcpu_mem_zalloc(size_t size)
292{
293 if (WARN_ON_ONCE(!slab_is_available()))
294 return NULL;
295
296 if (size <= PAGE_SIZE)
297 return kzalloc(size, GFP_KERNEL);
298 else
299 return vzalloc(size);
300}
301
302/**
303 * pcpu_mem_free - free memory
304 * @ptr: memory to free
305 * @size: size of the area
306 *
307 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
308 */
309static void pcpu_mem_free(void *ptr, size_t size)
310{
311 if (size <= PAGE_SIZE)
312 kfree(ptr);
313 else
314 vfree(ptr);
315}
316
317/**
318 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
319 * @chunk: chunk of interest
320 * @oslot: the previous slot it was on
321 *
322 * This function is called after an allocation or free changed @chunk.
323 * New slot according to the changed state is determined and @chunk is
324 * moved to the slot. Note that the reserved chunk is never put on
325 * chunk slots.
326 *
327 * CONTEXT:
328 * pcpu_lock.
329 */
330static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
331{
332 int nslot = pcpu_chunk_slot(chunk);
333
334 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
335 if (oslot < nslot)
336 list_move(&chunk->list, &pcpu_slot[nslot]);
337 else
338 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
339 }
340}
341
342/**
343 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
344 * @chunk: chunk of interest
345 *
346 * Determine whether area map of @chunk needs to be extended to
347 * accommodate a new allocation.
348 *
349 * CONTEXT:
350 * pcpu_lock.
351 *
352 * RETURNS:
353 * New target map allocation length if extension is necessary, 0
354 * otherwise.
355 */
356static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
357{
358 int new_alloc;
359
360 if (chunk->map_alloc >= chunk->map_used + 3)
361 return 0;
362
363 new_alloc = PCPU_DFL_MAP_ALLOC;
364 while (new_alloc < chunk->map_used + 3)
365 new_alloc *= 2;
366
367 return new_alloc;
368}
369
370/**
371 * pcpu_extend_area_map - extend area map of a chunk
372 * @chunk: chunk of interest
373 * @new_alloc: new target allocation length of the area map
374 *
375 * Extend area map of @chunk to have @new_alloc entries.
376 *
377 * CONTEXT:
378 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
379 *
380 * RETURNS:
381 * 0 on success, -errno on failure.
382 */
383static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
384{
385 int *old = NULL, *new = NULL;
386 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
387 unsigned long flags;
388
389 new = pcpu_mem_zalloc(new_size);
390 if (!new)
391 return -ENOMEM;
392
393 /* acquire pcpu_lock and switch to new area map */
394 spin_lock_irqsave(&pcpu_lock, flags);
395
396 if (new_alloc <= chunk->map_alloc)
397 goto out_unlock;
398
399 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
400 old = chunk->map;
401
402 memcpy(new, old, old_size);
403
404 chunk->map_alloc = new_alloc;
405 chunk->map = new;
406 new = NULL;
407
408out_unlock:
409 spin_unlock_irqrestore(&pcpu_lock, flags);
410
411 /*
412 * pcpu_mem_free() might end up calling vfree() which uses
413 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
414 */
415 pcpu_mem_free(old, old_size);
416 pcpu_mem_free(new, new_size);
417
418 return 0;
419}
420
421/**
422 * pcpu_alloc_area - allocate area from a pcpu_chunk
423 * @chunk: chunk of interest
424 * @size: wanted size in bytes
425 * @align: wanted align
426 *
427 * Try to allocate @size bytes area aligned at @align from @chunk.
428 * Note that this function only allocates the offset. It doesn't
429 * populate or map the area.
430 *
431 * @chunk->map must have at least two free slots.
432 *
433 * CONTEXT:
434 * pcpu_lock.
435 *
436 * RETURNS:
437 * Allocated offset in @chunk on success, -1 if no matching area is
438 * found.
439 */
440static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
441{
442 int oslot = pcpu_chunk_slot(chunk);
443 int max_contig = 0;
444 int i, off;
445 bool seen_free = false;
446 int *p;
447
448 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
449 int head, tail;
450 int this_size;
451
452 off = *p;
453 if (off & 1)
454 continue;
455
456 /* extra for alignment requirement */
457 head = ALIGN(off, align) - off;
458
459 this_size = (p[1] & ~1) - off;
460 if (this_size < head + size) {
461 if (!seen_free) {
462 chunk->first_free = i;
463 seen_free = true;
464 }
465 max_contig = max(this_size, max_contig);
466 continue;
467 }
468
469 /*
470 * If head is small or the previous block is free,
471 * merge'em. Note that 'small' is defined as smaller
472 * than sizeof(int), which is very small but isn't too
473 * uncommon for percpu allocations.
474 */
475 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
476 *p = off += head;
477 if (p[-1] & 1)
478 chunk->free_size -= head;
479 else
480 max_contig = max(*p - p[-1], max_contig);
481 this_size -= head;
482 head = 0;
483 }
484
485 /* if tail is small, just keep it around */
486 tail = this_size - head - size;
487 if (tail < sizeof(int)) {
488 tail = 0;
489 size = this_size - head;
490 }
491
492 /* split if warranted */
493 if (head || tail) {
494 int nr_extra = !!head + !!tail;
495
496 /* insert new subblocks */
497 memmove(p + nr_extra + 1, p + 1,
498 sizeof(chunk->map[0]) * (chunk->map_used - i));
499 chunk->map_used += nr_extra;
500
501 if (head) {
502 if (!seen_free) {
503 chunk->first_free = i;
504 seen_free = true;
505 }
506 *++p = off += head;
507 ++i;
508 max_contig = max(head, max_contig);
509 }
510 if (tail) {
511 p[1] = off + size;
512 max_contig = max(tail, max_contig);
513 }
514 }
515
516 if (!seen_free)
517 chunk->first_free = i + 1;
518
519 /* update hint and mark allocated */
520 if (i + 1 == chunk->map_used)
521 chunk->contig_hint = max_contig; /* fully scanned */
522 else
523 chunk->contig_hint = max(chunk->contig_hint,
524 max_contig);
525
526 chunk->free_size -= size;
527 *p |= 1;
528
529 pcpu_chunk_relocate(chunk, oslot);
530 return off;
531 }
532
533 chunk->contig_hint = max_contig; /* fully scanned */
534 pcpu_chunk_relocate(chunk, oslot);
535
536 /* tell the upper layer that this chunk has no matching area */
537 return -1;
538}
539
540/**
541 * pcpu_free_area - free area to a pcpu_chunk
542 * @chunk: chunk of interest
543 * @freeme: offset of area to free
544 *
545 * Free area starting from @freeme to @chunk. Note that this function
546 * only modifies the allocation map. It doesn't depopulate or unmap
547 * the area.
548 *
549 * CONTEXT:
550 * pcpu_lock.
551 */
552static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
553{
554 int oslot = pcpu_chunk_slot(chunk);
555 int off = 0;
556 unsigned i, j;
557 int to_free = 0;
558 int *p;
559
560 freeme |= 1; /* we are searching for <given offset, in use> pair */
561
562 i = 0;
563 j = chunk->map_used;
564 while (i != j) {
565 unsigned k = (i + j) / 2;
566 off = chunk->map[k];
567 if (off < freeme)
568 i = k + 1;
569 else if (off > freeme)
570 j = k;
571 else
572 i = j = k;
573 }
574 BUG_ON(off != freeme);
575
576 if (i < chunk->first_free)
577 chunk->first_free = i;
578
579 p = chunk->map + i;
580 *p = off &= ~1;
581 chunk->free_size += (p[1] & ~1) - off;
582
583 /* merge with next? */
584 if (!(p[1] & 1))
585 to_free++;
586 /* merge with previous? */
587 if (i > 0 && !(p[-1] & 1)) {
588 to_free++;
589 i--;
590 p--;
591 }
592 if (to_free) {
593 chunk->map_used -= to_free;
594 memmove(p + 1, p + 1 + to_free,
595 (chunk->map_used - i) * sizeof(chunk->map[0]));
596 }
597
598 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
599 pcpu_chunk_relocate(chunk, oslot);
600}
601
602static struct pcpu_chunk *pcpu_alloc_chunk(void)
603{
604 struct pcpu_chunk *chunk;
605
606 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
607 if (!chunk)
608 return NULL;
609
610 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
611 sizeof(chunk->map[0]));
612 if (!chunk->map) {
613 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
614 return NULL;
615 }
616
617 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
618 chunk->map[0] = 0;
619 chunk->map[1] = pcpu_unit_size | 1;
620 chunk->map_used = 1;
621
622 INIT_LIST_HEAD(&chunk->list);
623 chunk->free_size = pcpu_unit_size;
624 chunk->contig_hint = pcpu_unit_size;
625
626 return chunk;
627}
628
629static void pcpu_free_chunk(struct pcpu_chunk *chunk)
630{
631 if (!chunk)
632 return;
633 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
634 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
635}
636
637/*
638 * Chunk management implementation.
639 *
640 * To allow different implementations, chunk alloc/free and
641 * [de]population are implemented in a separate file which is pulled
642 * into this file and compiled together. The following functions
643 * should be implemented.
644 *
645 * pcpu_populate_chunk - populate the specified range of a chunk
646 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
647 * pcpu_create_chunk - create a new chunk
648 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
649 * pcpu_addr_to_page - translate address to physical address
650 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
651 */
652static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
653static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
654static struct pcpu_chunk *pcpu_create_chunk(void);
655static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
656static struct page *pcpu_addr_to_page(void *addr);
657static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
658
659#ifdef CONFIG_NEED_PER_CPU_KM
660#include "percpu-km.c"
661#else
662#include "percpu-vm.c"
663#endif
664
665/**
666 * pcpu_chunk_addr_search - determine chunk containing specified address
667 * @addr: address for which the chunk needs to be determined.
668 *
669 * RETURNS:
670 * The address of the found chunk.
671 */
672static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
673{
674 /* is it in the first chunk? */
675 if (pcpu_addr_in_first_chunk(addr)) {
676 /* is it in the reserved area? */
677 if (pcpu_addr_in_reserved_chunk(addr))
678 return pcpu_reserved_chunk;
679 return pcpu_first_chunk;
680 }
681
682 /*
683 * The address is relative to unit0 which might be unused and
684 * thus unmapped. Offset the address to the unit space of the
685 * current processor before looking it up in the vmalloc
686 * space. Note that any possible cpu id can be used here, so
687 * there's no need to worry about preemption or cpu hotplug.
688 */
689 addr += pcpu_unit_offsets[raw_smp_processor_id()];
690 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
691}
692
693/**
694 * pcpu_alloc - the percpu allocator
695 * @size: size of area to allocate in bytes
696 * @align: alignment of area (max PAGE_SIZE)
697 * @reserved: allocate from the reserved chunk if available
698 *
699 * Allocate percpu area of @size bytes aligned at @align.
700 *
701 * CONTEXT:
702 * Does GFP_KERNEL allocation.
703 *
704 * RETURNS:
705 * Percpu pointer to the allocated area on success, NULL on failure.
706 */
707static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
708{
709 static int warn_limit = 10;
710 struct pcpu_chunk *chunk;
711 const char *err;
712 int slot, off, new_alloc;
713 unsigned long flags;
714 void __percpu *ptr;
715
716 /*
717 * We want the lowest bit of offset available for in-use/free
718 * indicator, so force >= 16bit alignment and make size even.
719 */
720 if (unlikely(align < 2))
721 align = 2;
722
723 if (unlikely(size & 1))
724 size++;
725
726 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
727 WARN(true, "illegal size (%zu) or align (%zu) for "
728 "percpu allocation\n", size, align);
729 return NULL;
730 }
731
732 mutex_lock(&pcpu_alloc_mutex);
733 spin_lock_irqsave(&pcpu_lock, flags);
734
735 /* serve reserved allocations from the reserved chunk if available */
736 if (reserved && pcpu_reserved_chunk) {
737 chunk = pcpu_reserved_chunk;
738
739 if (size > chunk->contig_hint) {
740 err = "alloc from reserved chunk failed";
741 goto fail_unlock;
742 }
743
744 while ((new_alloc = pcpu_need_to_extend(chunk))) {
745 spin_unlock_irqrestore(&pcpu_lock, flags);
746 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
747 err = "failed to extend area map of reserved chunk";
748 goto fail_unlock_mutex;
749 }
750 spin_lock_irqsave(&pcpu_lock, flags);
751 }
752
753 off = pcpu_alloc_area(chunk, size, align);
754 if (off >= 0)
755 goto area_found;
756
757 err = "alloc from reserved chunk failed";
758 goto fail_unlock;
759 }
760
761restart:
762 /* search through normal chunks */
763 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
764 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
765 if (size > chunk->contig_hint)
766 continue;
767
768 new_alloc = pcpu_need_to_extend(chunk);
769 if (new_alloc) {
770 spin_unlock_irqrestore(&pcpu_lock, flags);
771 if (pcpu_extend_area_map(chunk,
772 new_alloc) < 0) {
773 err = "failed to extend area map";
774 goto fail_unlock_mutex;
775 }
776 spin_lock_irqsave(&pcpu_lock, flags);
777 /*
778 * pcpu_lock has been dropped, need to
779 * restart cpu_slot list walking.
780 */
781 goto restart;
782 }
783
784 off = pcpu_alloc_area(chunk, size, align);
785 if (off >= 0)
786 goto area_found;
787 }
788 }
789
790 /* hmmm... no space left, create a new chunk */
791 spin_unlock_irqrestore(&pcpu_lock, flags);
792
793 chunk = pcpu_create_chunk();
794 if (!chunk) {
795 err = "failed to allocate new chunk";
796 goto fail_unlock_mutex;
797 }
798
799 spin_lock_irqsave(&pcpu_lock, flags);
800 pcpu_chunk_relocate(chunk, -1);
801 goto restart;
802
803area_found:
804 spin_unlock_irqrestore(&pcpu_lock, flags);
805
806 /* populate, map and clear the area */
807 if (pcpu_populate_chunk(chunk, off, size)) {
808 spin_lock_irqsave(&pcpu_lock, flags);
809 pcpu_free_area(chunk, off);
810 err = "failed to populate";
811 goto fail_unlock;
812 }
813
814 mutex_unlock(&pcpu_alloc_mutex);
815
816 /* return address relative to base address */
817 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
818 kmemleak_alloc_percpu(ptr, size);
819 return ptr;
820
821fail_unlock:
822 spin_unlock_irqrestore(&pcpu_lock, flags);
823fail_unlock_mutex:
824 mutex_unlock(&pcpu_alloc_mutex);
825 if (warn_limit) {
826 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
827 "%s\n", size, align, err);
828 dump_stack();
829 if (!--warn_limit)
830 pr_info("PERCPU: limit reached, disable warning\n");
831 }
832 return NULL;
833}
834
835/**
836 * __alloc_percpu - allocate dynamic percpu area
837 * @size: size of area to allocate in bytes
838 * @align: alignment of area (max PAGE_SIZE)
839 *
840 * Allocate zero-filled percpu area of @size bytes aligned at @align.
841 * Might sleep. Might trigger writeouts.
842 *
843 * CONTEXT:
844 * Does GFP_KERNEL allocation.
845 *
846 * RETURNS:
847 * Percpu pointer to the allocated area on success, NULL on failure.
848 */
849void __percpu *__alloc_percpu(size_t size, size_t align)
850{
851 return pcpu_alloc(size, align, false);
852}
853EXPORT_SYMBOL_GPL(__alloc_percpu);
854
855/**
856 * __alloc_reserved_percpu - allocate reserved percpu area
857 * @size: size of area to allocate in bytes
858 * @align: alignment of area (max PAGE_SIZE)
859 *
860 * Allocate zero-filled percpu area of @size bytes aligned at @align
861 * from reserved percpu area if arch has set it up; otherwise,
862 * allocation is served from the same dynamic area. Might sleep.
863 * Might trigger writeouts.
864 *
865 * CONTEXT:
866 * Does GFP_KERNEL allocation.
867 *
868 * RETURNS:
869 * Percpu pointer to the allocated area on success, NULL on failure.
870 */
871void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
872{
873 return pcpu_alloc(size, align, true);
874}
875
876/**
877 * pcpu_reclaim - reclaim fully free chunks, workqueue function
878 * @work: unused
879 *
880 * Reclaim all fully free chunks except for the first one.
881 *
882 * CONTEXT:
883 * workqueue context.
884 */
885static void pcpu_reclaim(struct work_struct *work)
886{
887 LIST_HEAD(todo);
888 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
889 struct pcpu_chunk *chunk, *next;
890
891 mutex_lock(&pcpu_alloc_mutex);
892 spin_lock_irq(&pcpu_lock);
893
894 list_for_each_entry_safe(chunk, next, head, list) {
895 WARN_ON(chunk->immutable);
896
897 /* spare the first one */
898 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
899 continue;
900
901 list_move(&chunk->list, &todo);
902 }
903
904 spin_unlock_irq(&pcpu_lock);
905
906 list_for_each_entry_safe(chunk, next, &todo, list) {
907 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
908 pcpu_destroy_chunk(chunk);
909 }
910
911 mutex_unlock(&pcpu_alloc_mutex);
912}
913
914/**
915 * free_percpu - free percpu area
916 * @ptr: pointer to area to free
917 *
918 * Free percpu area @ptr.
919 *
920 * CONTEXT:
921 * Can be called from atomic context.
922 */
923void free_percpu(void __percpu *ptr)
924{
925 void *addr;
926 struct pcpu_chunk *chunk;
927 unsigned long flags;
928 int off;
929
930 if (!ptr)
931 return;
932
933 kmemleak_free_percpu(ptr);
934
935 addr = __pcpu_ptr_to_addr(ptr);
936
937 spin_lock_irqsave(&pcpu_lock, flags);
938
939 chunk = pcpu_chunk_addr_search(addr);
940 off = addr - chunk->base_addr;
941
942 pcpu_free_area(chunk, off);
943
944 /* if there are more than one fully free chunks, wake up grim reaper */
945 if (chunk->free_size == pcpu_unit_size) {
946 struct pcpu_chunk *pos;
947
948 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
949 if (pos != chunk) {
950 schedule_work(&pcpu_reclaim_work);
951 break;
952 }
953 }
954
955 spin_unlock_irqrestore(&pcpu_lock, flags);
956}
957EXPORT_SYMBOL_GPL(free_percpu);
958
959/**
960 * is_kernel_percpu_address - test whether address is from static percpu area
961 * @addr: address to test
962 *
963 * Test whether @addr belongs to in-kernel static percpu area. Module
964 * static percpu areas are not considered. For those, use
965 * is_module_percpu_address().
966 *
967 * RETURNS:
968 * %true if @addr is from in-kernel static percpu area, %false otherwise.
969 */
970bool is_kernel_percpu_address(unsigned long addr)
971{
972#ifdef CONFIG_SMP
973 const size_t static_size = __per_cpu_end - __per_cpu_start;
974 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
975 unsigned int cpu;
976
977 for_each_possible_cpu(cpu) {
978 void *start = per_cpu_ptr(base, cpu);
979
980 if ((void *)addr >= start && (void *)addr < start + static_size)
981 return true;
982 }
983#endif
984 /* on UP, can't distinguish from other static vars, always false */
985 return false;
986}
987
988/**
989 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
990 * @addr: the address to be converted to physical address
991 *
992 * Given @addr which is dereferenceable address obtained via one of
993 * percpu access macros, this function translates it into its physical
994 * address. The caller is responsible for ensuring @addr stays valid
995 * until this function finishes.
996 *
997 * percpu allocator has special setup for the first chunk, which currently
998 * supports either embedding in linear address space or vmalloc mapping,
999 * and, from the second one, the backing allocator (currently either vm or
1000 * km) provides translation.
1001 *
1002 * The addr can be tranlated simply without checking if it falls into the
1003 * first chunk. But the current code reflects better how percpu allocator
1004 * actually works, and the verification can discover both bugs in percpu
1005 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1006 * code.
1007 *
1008 * RETURNS:
1009 * The physical address for @addr.
1010 */
1011phys_addr_t per_cpu_ptr_to_phys(void *addr)
1012{
1013 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1014 bool in_first_chunk = false;
1015 unsigned long first_low, first_high;
1016 unsigned int cpu;
1017
1018 /*
1019 * The following test on unit_low/high isn't strictly
1020 * necessary but will speed up lookups of addresses which
1021 * aren't in the first chunk.
1022 */
1023 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1024 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1025 pcpu_unit_pages);
1026 if ((unsigned long)addr >= first_low &&
1027 (unsigned long)addr < first_high) {
1028 for_each_possible_cpu(cpu) {
1029 void *start = per_cpu_ptr(base, cpu);
1030
1031 if (addr >= start && addr < start + pcpu_unit_size) {
1032 in_first_chunk = true;
1033 break;
1034 }
1035 }
1036 }
1037
1038 if (in_first_chunk) {
1039 if (!is_vmalloc_addr(addr))
1040 return __pa(addr);
1041 else
1042 return page_to_phys(vmalloc_to_page(addr)) +
1043 offset_in_page(addr);
1044 } else
1045 return page_to_phys(pcpu_addr_to_page(addr)) +
1046 offset_in_page(addr);
1047}
1048
1049/**
1050 * pcpu_alloc_alloc_info - allocate percpu allocation info
1051 * @nr_groups: the number of groups
1052 * @nr_units: the number of units
1053 *
1054 * Allocate ai which is large enough for @nr_groups groups containing
1055 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1056 * cpu_map array which is long enough for @nr_units and filled with
1057 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1058 * pointer of other groups.
1059 *
1060 * RETURNS:
1061 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1062 * failure.
1063 */
1064struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1065 int nr_units)
1066{
1067 struct pcpu_alloc_info *ai;
1068 size_t base_size, ai_size;
1069 void *ptr;
1070 int unit;
1071
1072 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1073 __alignof__(ai->groups[0].cpu_map[0]));
1074 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1075
1076 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1077 if (!ptr)
1078 return NULL;
1079 ai = ptr;
1080 ptr += base_size;
1081
1082 ai->groups[0].cpu_map = ptr;
1083
1084 for (unit = 0; unit < nr_units; unit++)
1085 ai->groups[0].cpu_map[unit] = NR_CPUS;
1086
1087 ai->nr_groups = nr_groups;
1088 ai->__ai_size = PFN_ALIGN(ai_size);
1089
1090 return ai;
1091}
1092
1093/**
1094 * pcpu_free_alloc_info - free percpu allocation info
1095 * @ai: pcpu_alloc_info to free
1096 *
1097 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1098 */
1099void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1100{
1101 memblock_free_early(__pa(ai), ai->__ai_size);
1102}
1103
1104/**
1105 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1106 * @lvl: loglevel
1107 * @ai: allocation info to dump
1108 *
1109 * Print out information about @ai using loglevel @lvl.
1110 */
1111static void pcpu_dump_alloc_info(const char *lvl,
1112 const struct pcpu_alloc_info *ai)
1113{
1114 int group_width = 1, cpu_width = 1, width;
1115 char empty_str[] = "--------";
1116 int alloc = 0, alloc_end = 0;
1117 int group, v;
1118 int upa, apl; /* units per alloc, allocs per line */
1119
1120 v = ai->nr_groups;
1121 while (v /= 10)
1122 group_width++;
1123
1124 v = num_possible_cpus();
1125 while (v /= 10)
1126 cpu_width++;
1127 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1128
1129 upa = ai->alloc_size / ai->unit_size;
1130 width = upa * (cpu_width + 1) + group_width + 3;
1131 apl = rounddown_pow_of_two(max(60 / width, 1));
1132
1133 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1134 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1135 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1136
1137 for (group = 0; group < ai->nr_groups; group++) {
1138 const struct pcpu_group_info *gi = &ai->groups[group];
1139 int unit = 0, unit_end = 0;
1140
1141 BUG_ON(gi->nr_units % upa);
1142 for (alloc_end += gi->nr_units / upa;
1143 alloc < alloc_end; alloc++) {
1144 if (!(alloc % apl)) {
1145 printk(KERN_CONT "\n");
1146 printk("%spcpu-alloc: ", lvl);
1147 }
1148 printk(KERN_CONT "[%0*d] ", group_width, group);
1149
1150 for (unit_end += upa; unit < unit_end; unit++)
1151 if (gi->cpu_map[unit] != NR_CPUS)
1152 printk(KERN_CONT "%0*d ", cpu_width,
1153 gi->cpu_map[unit]);
1154 else
1155 printk(KERN_CONT "%s ", empty_str);
1156 }
1157 }
1158 printk(KERN_CONT "\n");
1159}
1160
1161/**
1162 * pcpu_setup_first_chunk - initialize the first percpu chunk
1163 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1164 * @base_addr: mapped address
1165 *
1166 * Initialize the first percpu chunk which contains the kernel static
1167 * perpcu area. This function is to be called from arch percpu area
1168 * setup path.
1169 *
1170 * @ai contains all information necessary to initialize the first
1171 * chunk and prime the dynamic percpu allocator.
1172 *
1173 * @ai->static_size is the size of static percpu area.
1174 *
1175 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1176 * reserve after the static area in the first chunk. This reserves
1177 * the first chunk such that it's available only through reserved
1178 * percpu allocation. This is primarily used to serve module percpu
1179 * static areas on architectures where the addressing model has
1180 * limited offset range for symbol relocations to guarantee module
1181 * percpu symbols fall inside the relocatable range.
1182 *
1183 * @ai->dyn_size determines the number of bytes available for dynamic
1184 * allocation in the first chunk. The area between @ai->static_size +
1185 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1186 *
1187 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1188 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1189 * @ai->dyn_size.
1190 *
1191 * @ai->atom_size is the allocation atom size and used as alignment
1192 * for vm areas.
1193 *
1194 * @ai->alloc_size is the allocation size and always multiple of
1195 * @ai->atom_size. This is larger than @ai->atom_size if
1196 * @ai->unit_size is larger than @ai->atom_size.
1197 *
1198 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1199 * percpu areas. Units which should be colocated are put into the
1200 * same group. Dynamic VM areas will be allocated according to these
1201 * groupings. If @ai->nr_groups is zero, a single group containing
1202 * all units is assumed.
1203 *
1204 * The caller should have mapped the first chunk at @base_addr and
1205 * copied static data to each unit.
1206 *
1207 * If the first chunk ends up with both reserved and dynamic areas, it
1208 * is served by two chunks - one to serve the core static and reserved
1209 * areas and the other for the dynamic area. They share the same vm
1210 * and page map but uses different area allocation map to stay away
1211 * from each other. The latter chunk is circulated in the chunk slots
1212 * and available for dynamic allocation like any other chunks.
1213 *
1214 * RETURNS:
1215 * 0 on success, -errno on failure.
1216 */
1217int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1218 void *base_addr)
1219{
1220 static char cpus_buf[4096] __initdata;
1221 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1222 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1223 size_t dyn_size = ai->dyn_size;
1224 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1225 struct pcpu_chunk *schunk, *dchunk = NULL;
1226 unsigned long *group_offsets;
1227 size_t *group_sizes;
1228 unsigned long *unit_off;
1229 unsigned int cpu;
1230 int *unit_map;
1231 int group, unit, i;
1232
1233 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1234
1235#define PCPU_SETUP_BUG_ON(cond) do { \
1236 if (unlikely(cond)) { \
1237 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1238 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1239 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1240 BUG(); \
1241 } \
1242} while (0)
1243
1244 /* sanity checks */
1245 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1246#ifdef CONFIG_SMP
1247 PCPU_SETUP_BUG_ON(!ai->static_size);
1248 PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
1249#endif
1250 PCPU_SETUP_BUG_ON(!base_addr);
1251 PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
1252 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1253 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1254 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1255 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1256 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1257
1258 /* process group information and build config tables accordingly */
1259 group_offsets = memblock_virt_alloc(ai->nr_groups *
1260 sizeof(group_offsets[0]), 0);
1261 group_sizes = memblock_virt_alloc(ai->nr_groups *
1262 sizeof(group_sizes[0]), 0);
1263 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1264 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1265
1266 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1267 unit_map[cpu] = UINT_MAX;
1268
1269 pcpu_low_unit_cpu = NR_CPUS;
1270 pcpu_high_unit_cpu = NR_CPUS;
1271
1272 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1273 const struct pcpu_group_info *gi = &ai->groups[group];
1274
1275 group_offsets[group] = gi->base_offset;
1276 group_sizes[group] = gi->nr_units * ai->unit_size;
1277
1278 for (i = 0; i < gi->nr_units; i++) {
1279 cpu = gi->cpu_map[i];
1280 if (cpu == NR_CPUS)
1281 continue;
1282
1283 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1284 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1285 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1286
1287 unit_map[cpu] = unit + i;
1288 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1289
1290 /* determine low/high unit_cpu */
1291 if (pcpu_low_unit_cpu == NR_CPUS ||
1292 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1293 pcpu_low_unit_cpu = cpu;
1294 if (pcpu_high_unit_cpu == NR_CPUS ||
1295 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1296 pcpu_high_unit_cpu = cpu;
1297 }
1298 }
1299 pcpu_nr_units = unit;
1300
1301 for_each_possible_cpu(cpu)
1302 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1303
1304 /* we're done parsing the input, undefine BUG macro and dump config */
1305#undef PCPU_SETUP_BUG_ON
1306 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1307
1308 pcpu_nr_groups = ai->nr_groups;
1309 pcpu_group_offsets = group_offsets;
1310 pcpu_group_sizes = group_sizes;
1311 pcpu_unit_map = unit_map;
1312 pcpu_unit_offsets = unit_off;
1313
1314 /* determine basic parameters */
1315 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1316 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1317 pcpu_atom_size = ai->atom_size;
1318 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1319 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1320
1321 /*
1322 * Allocate chunk slots. The additional last slot is for
1323 * empty chunks.
1324 */
1325 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1326 pcpu_slot = memblock_virt_alloc(
1327 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1328 for (i = 0; i < pcpu_nr_slots; i++)
1329 INIT_LIST_HEAD(&pcpu_slot[i]);
1330
1331 /*
1332 * Initialize static chunk. If reserved_size is zero, the
1333 * static chunk covers static area + dynamic allocation area
1334 * in the first chunk. If reserved_size is not zero, it
1335 * covers static area + reserved area (mostly used for module
1336 * static percpu allocation).
1337 */
1338 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1339 INIT_LIST_HEAD(&schunk->list);
1340 schunk->base_addr = base_addr;
1341 schunk->map = smap;
1342 schunk->map_alloc = ARRAY_SIZE(smap);
1343 schunk->immutable = true;
1344 bitmap_fill(schunk->populated, pcpu_unit_pages);
1345
1346 if (ai->reserved_size) {
1347 schunk->free_size = ai->reserved_size;
1348 pcpu_reserved_chunk = schunk;
1349 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1350 } else {
1351 schunk->free_size = dyn_size;
1352 dyn_size = 0; /* dynamic area covered */
1353 }
1354 schunk->contig_hint = schunk->free_size;
1355
1356 schunk->map[0] = 1;
1357 schunk->map[1] = ai->static_size;
1358 schunk->map_used = 1;
1359 if (schunk->free_size)
1360 schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size);
1361 else
1362 schunk->map[1] |= 1;
1363
1364 /* init dynamic chunk if necessary */
1365 if (dyn_size) {
1366 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1367 INIT_LIST_HEAD(&dchunk->list);
1368 dchunk->base_addr = base_addr;
1369 dchunk->map = dmap;
1370 dchunk->map_alloc = ARRAY_SIZE(dmap);
1371 dchunk->immutable = true;
1372 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1373
1374 dchunk->contig_hint = dchunk->free_size = dyn_size;
1375 dchunk->map[0] = 1;
1376 dchunk->map[1] = pcpu_reserved_chunk_limit;
1377 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1378 dchunk->map_used = 2;
1379 }
1380
1381 /* link the first chunk in */
1382 pcpu_first_chunk = dchunk ?: schunk;
1383 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1384
1385 /* we're done */
1386 pcpu_base_addr = base_addr;
1387 return 0;
1388}
1389
1390#ifdef CONFIG_SMP
1391
1392const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1393 [PCPU_FC_AUTO] = "auto",
1394 [PCPU_FC_EMBED] = "embed",
1395 [PCPU_FC_PAGE] = "page",
1396};
1397
1398enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1399
1400static int __init percpu_alloc_setup(char *str)
1401{
1402 if (!str)
1403 return -EINVAL;
1404
1405 if (0)
1406 /* nada */;
1407#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1408 else if (!strcmp(str, "embed"))
1409 pcpu_chosen_fc = PCPU_FC_EMBED;
1410#endif
1411#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1412 else if (!strcmp(str, "page"))
1413 pcpu_chosen_fc = PCPU_FC_PAGE;
1414#endif
1415 else
1416 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1417
1418 return 0;
1419}
1420early_param("percpu_alloc", percpu_alloc_setup);
1421
1422/*
1423 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1424 * Build it if needed by the arch config or the generic setup is going
1425 * to be used.
1426 */
1427#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1428 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1429#define BUILD_EMBED_FIRST_CHUNK
1430#endif
1431
1432/* build pcpu_page_first_chunk() iff needed by the arch config */
1433#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1434#define BUILD_PAGE_FIRST_CHUNK
1435#endif
1436
1437/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1438#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1439/**
1440 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1441 * @reserved_size: the size of reserved percpu area in bytes
1442 * @dyn_size: minimum free size for dynamic allocation in bytes
1443 * @atom_size: allocation atom size
1444 * @cpu_distance_fn: callback to determine distance between cpus, optional
1445 *
1446 * This function determines grouping of units, their mappings to cpus
1447 * and other parameters considering needed percpu size, allocation
1448 * atom size and distances between CPUs.
1449 *
1450 * Groups are always mutliples of atom size and CPUs which are of
1451 * LOCAL_DISTANCE both ways are grouped together and share space for
1452 * units in the same group. The returned configuration is guaranteed
1453 * to have CPUs on different nodes on different groups and >=75% usage
1454 * of allocated virtual address space.
1455 *
1456 * RETURNS:
1457 * On success, pointer to the new allocation_info is returned. On
1458 * failure, ERR_PTR value is returned.
1459 */
1460static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1461 size_t reserved_size, size_t dyn_size,
1462 size_t atom_size,
1463 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1464{
1465 static int group_map[NR_CPUS] __initdata;
1466 static int group_cnt[NR_CPUS] __initdata;
1467 const size_t static_size = __per_cpu_end - __per_cpu_start;
1468 int nr_groups = 1, nr_units = 0;
1469 size_t size_sum, min_unit_size, alloc_size;
1470 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1471 int last_allocs, group, unit;
1472 unsigned int cpu, tcpu;
1473 struct pcpu_alloc_info *ai;
1474 unsigned int *cpu_map;
1475
1476 /* this function may be called multiple times */
1477 memset(group_map, 0, sizeof(group_map));
1478 memset(group_cnt, 0, sizeof(group_cnt));
1479
1480 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1481 size_sum = PFN_ALIGN(static_size + reserved_size +
1482 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1483 dyn_size = size_sum - static_size - reserved_size;
1484
1485 /*
1486 * Determine min_unit_size, alloc_size and max_upa such that
1487 * alloc_size is multiple of atom_size and is the smallest
1488 * which can accommodate 4k aligned segments which are equal to
1489 * or larger than min_unit_size.
1490 */
1491 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1492
1493 alloc_size = roundup(min_unit_size, atom_size);
1494 upa = alloc_size / min_unit_size;
1495 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1496 upa--;
1497 max_upa = upa;
1498
1499 /* group cpus according to their proximity */
1500 for_each_possible_cpu(cpu) {
1501 group = 0;
1502 next_group:
1503 for_each_possible_cpu(tcpu) {
1504 if (cpu == tcpu)
1505 break;
1506 if (group_map[tcpu] == group && cpu_distance_fn &&
1507 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1508 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1509 group++;
1510 nr_groups = max(nr_groups, group + 1);
1511 goto next_group;
1512 }
1513 }
1514 group_map[cpu] = group;
1515 group_cnt[group]++;
1516 }
1517
1518 /*
1519 * Expand unit size until address space usage goes over 75%
1520 * and then as much as possible without using more address
1521 * space.
1522 */
1523 last_allocs = INT_MAX;
1524 for (upa = max_upa; upa; upa--) {
1525 int allocs = 0, wasted = 0;
1526
1527 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1528 continue;
1529
1530 for (group = 0; group < nr_groups; group++) {
1531 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1532 allocs += this_allocs;
1533 wasted += this_allocs * upa - group_cnt[group];
1534 }
1535
1536 /*
1537 * Don't accept if wastage is over 1/3. The
1538 * greater-than comparison ensures upa==1 always
1539 * passes the following check.
1540 */
1541 if (wasted > num_possible_cpus() / 3)
1542 continue;
1543
1544 /* and then don't consume more memory */
1545 if (allocs > last_allocs)
1546 break;
1547 last_allocs = allocs;
1548 best_upa = upa;
1549 }
1550 upa = best_upa;
1551
1552 /* allocate and fill alloc_info */
1553 for (group = 0; group < nr_groups; group++)
1554 nr_units += roundup(group_cnt[group], upa);
1555
1556 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1557 if (!ai)
1558 return ERR_PTR(-ENOMEM);
1559 cpu_map = ai->groups[0].cpu_map;
1560
1561 for (group = 0; group < nr_groups; group++) {
1562 ai->groups[group].cpu_map = cpu_map;
1563 cpu_map += roundup(group_cnt[group], upa);
1564 }
1565
1566 ai->static_size = static_size;
1567 ai->reserved_size = reserved_size;
1568 ai->dyn_size = dyn_size;
1569 ai->unit_size = alloc_size / upa;
1570 ai->atom_size = atom_size;
1571 ai->alloc_size = alloc_size;
1572
1573 for (group = 0, unit = 0; group_cnt[group]; group++) {
1574 struct pcpu_group_info *gi = &ai->groups[group];
1575
1576 /*
1577 * Initialize base_offset as if all groups are located
1578 * back-to-back. The caller should update this to
1579 * reflect actual allocation.
1580 */
1581 gi->base_offset = unit * ai->unit_size;
1582
1583 for_each_possible_cpu(cpu)
1584 if (group_map[cpu] == group)
1585 gi->cpu_map[gi->nr_units++] = cpu;
1586 gi->nr_units = roundup(gi->nr_units, upa);
1587 unit += gi->nr_units;
1588 }
1589 BUG_ON(unit != nr_units);
1590
1591 return ai;
1592}
1593#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1594
1595#if defined(BUILD_EMBED_FIRST_CHUNK)
1596/**
1597 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1598 * @reserved_size: the size of reserved percpu area in bytes
1599 * @dyn_size: minimum free size for dynamic allocation in bytes
1600 * @atom_size: allocation atom size
1601 * @cpu_distance_fn: callback to determine distance between cpus, optional
1602 * @alloc_fn: function to allocate percpu page
1603 * @free_fn: function to free percpu page
1604 *
1605 * This is a helper to ease setting up embedded first percpu chunk and
1606 * can be called where pcpu_setup_first_chunk() is expected.
1607 *
1608 * If this function is used to setup the first chunk, it is allocated
1609 * by calling @alloc_fn and used as-is without being mapped into
1610 * vmalloc area. Allocations are always whole multiples of @atom_size
1611 * aligned to @atom_size.
1612 *
1613 * This enables the first chunk to piggy back on the linear physical
1614 * mapping which often uses larger page size. Please note that this
1615 * can result in very sparse cpu->unit mapping on NUMA machines thus
1616 * requiring large vmalloc address space. Don't use this allocator if
1617 * vmalloc space is not orders of magnitude larger than distances
1618 * between node memory addresses (ie. 32bit NUMA machines).
1619 *
1620 * @dyn_size specifies the minimum dynamic area size.
1621 *
1622 * If the needed size is smaller than the minimum or specified unit
1623 * size, the leftover is returned using @free_fn.
1624 *
1625 * RETURNS:
1626 * 0 on success, -errno on failure.
1627 */
1628int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1629 size_t atom_size,
1630 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1631 pcpu_fc_alloc_fn_t alloc_fn,
1632 pcpu_fc_free_fn_t free_fn)
1633{
1634 void *base = (void *)ULONG_MAX;
1635 void **areas = NULL;
1636 struct pcpu_alloc_info *ai;
1637 size_t size_sum, areas_size, max_distance;
1638 int group, i, rc;
1639
1640 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1641 cpu_distance_fn);
1642 if (IS_ERR(ai))
1643 return PTR_ERR(ai);
1644
1645 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1646 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1647
1648 areas = memblock_virt_alloc_nopanic(areas_size, 0);
1649 if (!areas) {
1650 rc = -ENOMEM;
1651 goto out_free;
1652 }
1653
1654 /* allocate, copy and determine base address */
1655 for (group = 0; group < ai->nr_groups; group++) {
1656 struct pcpu_group_info *gi = &ai->groups[group];
1657 unsigned int cpu = NR_CPUS;
1658 void *ptr;
1659
1660 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1661 cpu = gi->cpu_map[i];
1662 BUG_ON(cpu == NR_CPUS);
1663
1664 /* allocate space for the whole group */
1665 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1666 if (!ptr) {
1667 rc = -ENOMEM;
1668 goto out_free_areas;
1669 }
1670 /* kmemleak tracks the percpu allocations separately */
1671 kmemleak_free(ptr);
1672 areas[group] = ptr;
1673
1674 base = min(ptr, base);
1675 }
1676
1677 /*
1678 * Copy data and free unused parts. This should happen after all
1679 * allocations are complete; otherwise, we may end up with
1680 * overlapping groups.
1681 */
1682 for (group = 0; group < ai->nr_groups; group++) {
1683 struct pcpu_group_info *gi = &ai->groups[group];
1684 void *ptr = areas[group];
1685
1686 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1687 if (gi->cpu_map[i] == NR_CPUS) {
1688 /* unused unit, free whole */
1689 free_fn(ptr, ai->unit_size);
1690 continue;
1691 }
1692 /* copy and return the unused part */
1693 memcpy(ptr, __per_cpu_load, ai->static_size);
1694 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1695 }
1696 }
1697
1698 /* base address is now known, determine group base offsets */
1699 max_distance = 0;
1700 for (group = 0; group < ai->nr_groups; group++) {
1701 ai->groups[group].base_offset = areas[group] - base;
1702 max_distance = max_t(size_t, max_distance,
1703 ai->groups[group].base_offset);
1704 }
1705 max_distance += ai->unit_size;
1706
1707 /* warn if maximum distance is further than 75% of vmalloc space */
1708 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
1709 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1710 "space 0x%lx\n", max_distance,
1711 VMALLOC_TOTAL);
1712#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1713 /* and fail if we have fallback */
1714 rc = -EINVAL;
1715 goto out_free;
1716#endif
1717 }
1718
1719 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1720 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1721 ai->dyn_size, ai->unit_size);
1722
1723 rc = pcpu_setup_first_chunk(ai, base);
1724 goto out_free;
1725
1726out_free_areas:
1727 for (group = 0; group < ai->nr_groups; group++)
1728 if (areas[group])
1729 free_fn(areas[group],
1730 ai->groups[group].nr_units * ai->unit_size);
1731out_free:
1732 pcpu_free_alloc_info(ai);
1733 if (areas)
1734 memblock_free_early(__pa(areas), areas_size);
1735 return rc;
1736}
1737#endif /* BUILD_EMBED_FIRST_CHUNK */
1738
1739#ifdef BUILD_PAGE_FIRST_CHUNK
1740/**
1741 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1742 * @reserved_size: the size of reserved percpu area in bytes
1743 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1744 * @free_fn: function to free percpu page, always called with PAGE_SIZE
1745 * @populate_pte_fn: function to populate pte
1746 *
1747 * This is a helper to ease setting up page-remapped first percpu
1748 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1749 *
1750 * This is the basic allocator. Static percpu area is allocated
1751 * page-by-page into vmalloc area.
1752 *
1753 * RETURNS:
1754 * 0 on success, -errno on failure.
1755 */
1756int __init pcpu_page_first_chunk(size_t reserved_size,
1757 pcpu_fc_alloc_fn_t alloc_fn,
1758 pcpu_fc_free_fn_t free_fn,
1759 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1760{
1761 static struct vm_struct vm;
1762 struct pcpu_alloc_info *ai;
1763 char psize_str[16];
1764 int unit_pages;
1765 size_t pages_size;
1766 struct page **pages;
1767 int unit, i, j, rc;
1768
1769 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1770
1771 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
1772 if (IS_ERR(ai))
1773 return PTR_ERR(ai);
1774 BUG_ON(ai->nr_groups != 1);
1775 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1776
1777 unit_pages = ai->unit_size >> PAGE_SHIFT;
1778
1779 /* unaligned allocations can't be freed, round up to page size */
1780 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1781 sizeof(pages[0]));
1782 pages = memblock_virt_alloc(pages_size, 0);
1783
1784 /* allocate pages */
1785 j = 0;
1786 for (unit = 0; unit < num_possible_cpus(); unit++)
1787 for (i = 0; i < unit_pages; i++) {
1788 unsigned int cpu = ai->groups[0].cpu_map[unit];
1789 void *ptr;
1790
1791 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1792 if (!ptr) {
1793 pr_warning("PERCPU: failed to allocate %s page "
1794 "for cpu%u\n", psize_str, cpu);
1795 goto enomem;
1796 }
1797 /* kmemleak tracks the percpu allocations separately */
1798 kmemleak_free(ptr);
1799 pages[j++] = virt_to_page(ptr);
1800 }
1801
1802 /* allocate vm area, map the pages and copy static data */
1803 vm.flags = VM_ALLOC;
1804 vm.size = num_possible_cpus() * ai->unit_size;
1805 vm_area_register_early(&vm, PAGE_SIZE);
1806
1807 for (unit = 0; unit < num_possible_cpus(); unit++) {
1808 unsigned long unit_addr =
1809 (unsigned long)vm.addr + unit * ai->unit_size;
1810
1811 for (i = 0; i < unit_pages; i++)
1812 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1813
1814 /* pte already populated, the following shouldn't fail */
1815 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1816 unit_pages);
1817 if (rc < 0)
1818 panic("failed to map percpu area, err=%d\n", rc);
1819
1820 /*
1821 * FIXME: Archs with virtual cache should flush local
1822 * cache for the linear mapping here - something
1823 * equivalent to flush_cache_vmap() on the local cpu.
1824 * flush_cache_vmap() can't be used as most supporting
1825 * data structures are not set up yet.
1826 */
1827
1828 /* copy static data */
1829 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1830 }
1831
1832 /* we're ready, commit */
1833 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1834 unit_pages, psize_str, vm.addr, ai->static_size,
1835 ai->reserved_size, ai->dyn_size);
1836
1837 rc = pcpu_setup_first_chunk(ai, vm.addr);
1838 goto out_free_ar;
1839
1840enomem:
1841 while (--j >= 0)
1842 free_fn(page_address(pages[j]), PAGE_SIZE);
1843 rc = -ENOMEM;
1844out_free_ar:
1845 memblock_free_early(__pa(pages), pages_size);
1846 pcpu_free_alloc_info(ai);
1847 return rc;
1848}
1849#endif /* BUILD_PAGE_FIRST_CHUNK */
1850
1851#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1852/*
1853 * Generic SMP percpu area setup.
1854 *
1855 * The embedding helper is used because its behavior closely resembles
1856 * the original non-dynamic generic percpu area setup. This is
1857 * important because many archs have addressing restrictions and might
1858 * fail if the percpu area is located far away from the previous
1859 * location. As an added bonus, in non-NUMA cases, embedding is
1860 * generally a good idea TLB-wise because percpu area can piggy back
1861 * on the physical linear memory mapping which uses large page
1862 * mappings on applicable archs.
1863 */
1864unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1865EXPORT_SYMBOL(__per_cpu_offset);
1866
1867static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1868 size_t align)
1869{
1870 return memblock_virt_alloc_from_nopanic(
1871 size, align, __pa(MAX_DMA_ADDRESS));
1872}
1873
1874static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1875{
1876 memblock_free_early(__pa(ptr), size);
1877}
1878
1879void __init setup_per_cpu_areas(void)
1880{
1881 unsigned long delta;
1882 unsigned int cpu;
1883 int rc;
1884
1885 /*
1886 * Always reserve area for module percpu variables. That's
1887 * what the legacy allocator did.
1888 */
1889 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1890 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1891 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
1892 if (rc < 0)
1893 panic("Failed to initialize percpu areas.");
1894
1895 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1896 for_each_possible_cpu(cpu)
1897 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1898}
1899#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1900
1901#else /* CONFIG_SMP */
1902
1903/*
1904 * UP percpu area setup.
1905 *
1906 * UP always uses km-based percpu allocator with identity mapping.
1907 * Static percpu variables are indistinguishable from the usual static
1908 * variables and don't require any special preparation.
1909 */
1910void __init setup_per_cpu_areas(void)
1911{
1912 const size_t unit_size =
1913 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1914 PERCPU_DYNAMIC_RESERVE));
1915 struct pcpu_alloc_info *ai;
1916 void *fc;
1917
1918 ai = pcpu_alloc_alloc_info(1, 1);
1919 fc = memblock_virt_alloc_from_nopanic(unit_size,
1920 PAGE_SIZE,
1921 __pa(MAX_DMA_ADDRESS));
1922 if (!ai || !fc)
1923 panic("Failed to allocate memory for percpu areas.");
1924 /* kmemleak tracks the percpu allocations separately */
1925 kmemleak_free(fc);
1926
1927 ai->dyn_size = unit_size;
1928 ai->unit_size = unit_size;
1929 ai->atom_size = unit_size;
1930 ai->alloc_size = unit_size;
1931 ai->groups[0].nr_units = 1;
1932 ai->groups[0].cpu_map[0] = 0;
1933
1934 if (pcpu_setup_first_chunk(ai, fc) < 0)
1935 panic("Failed to initialize percpu areas.");
1936}
1937
1938#endif /* CONFIG_SMP */
1939
1940/*
1941 * First and reserved chunks are initialized with temporary allocation
1942 * map in initdata so that they can be used before slab is online.
1943 * This function is called after slab is brought up and replaces those
1944 * with properly allocated maps.
1945 */
1946void __init percpu_init_late(void)
1947{
1948 struct pcpu_chunk *target_chunks[] =
1949 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1950 struct pcpu_chunk *chunk;
1951 unsigned long flags;
1952 int i;
1953
1954 for (i = 0; (chunk = target_chunks[i]); i++) {
1955 int *map;
1956 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1957
1958 BUILD_BUG_ON(size > PAGE_SIZE);
1959
1960 map = pcpu_mem_zalloc(size);
1961 BUG_ON(!map);
1962
1963 spin_lock_irqsave(&pcpu_lock, flags);
1964 memcpy(map, chunk->map, size);
1965 chunk->map = map;
1966 spin_unlock_irqrestore(&pcpu_lock, flags);
1967 }
1968}