Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/percpu.c - percpu memory allocator
4 *
5 * Copyright (C) 2009 SUSE Linux Products GmbH
6 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 *
8 * Copyright (C) 2017 Facebook Inc.
9 * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org>
10 *
11 * The percpu allocator handles both static and dynamic areas. Percpu
12 * areas are allocated in chunks which are divided into units. There is
13 * a 1-to-1 mapping for units to possible cpus. These units are grouped
14 * based on NUMA properties of the machine.
15 *
16 * c0 c1 c2
17 * ------------------- ------------------- ------------
18 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19 * ------------------- ...... ------------------- .... ------------
20 *
21 * Allocation is done by offsets into a unit's address space. Ie., an
22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24 * and even sparse. Access is handled by configuring percpu base
25 * registers according to the cpu to unit mappings and offsetting the
26 * base address using pcpu_unit_size.
27 *
28 * There is special consideration for the first chunk which must handle
29 * the static percpu variables in the kernel image as allocation services
30 * are not online yet. In short, the first chunk is structured like so:
31 *
32 * <Static | [Reserved] | Dynamic>
33 *
34 * The static data is copied from the original section managed by the
35 * linker. The reserved section, if non-zero, primarily manages static
36 * percpu variables from kernel modules. Finally, the dynamic section
37 * takes care of normal allocations.
38 *
39 * The allocator organizes chunks into lists according to free size and
40 * memcg-awareness. To make a percpu allocation memcg-aware the __GFP_ACCOUNT
41 * flag should be passed. All memcg-aware allocations are sharing one set
42 * of chunks and all unaccounted allocations and allocations performed
43 * by processes belonging to the root memory cgroup are using the second set.
44 *
45 * The allocator tries to allocate from the fullest chunk first. Each chunk
46 * is managed by a bitmap with metadata blocks. The allocation map is updated
47 * on every allocation and free to reflect the current state while the boundary
48 * map is only updated on allocation. Each metadata block contains
49 * information to help mitigate the need to iterate over large portions
50 * of the bitmap. The reverse mapping from page to chunk is stored in
51 * the page's index. Lastly, units are lazily backed and grow in unison.
52 *
53 * There is a unique conversion that goes on here between bytes and bits.
54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
55 * tracks the number of pages it is responsible for in nr_pages. Helper
56 * functions are used to convert from between the bytes, bits, and blocks.
57 * All hints are managed in bits unless explicitly stated.
58 *
59 * To use this allocator, arch code should do the following:
60 *
61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
62 * regular address to percpu pointer and back if they need to be
63 * different from the default
64 *
65 * - use pcpu_setup_first_chunk() during percpu area initialization to
66 * setup the first chunk containing the kernel static percpu area
67 */
68
69#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70
71#include <linux/bitmap.h>
72#include <linux/memblock.h>
73#include <linux/err.h>
74#include <linux/lcm.h>
75#include <linux/list.h>
76#include <linux/log2.h>
77#include <linux/mm.h>
78#include <linux/module.h>
79#include <linux/mutex.h>
80#include <linux/percpu.h>
81#include <linux/pfn.h>
82#include <linux/slab.h>
83#include <linux/spinlock.h>
84#include <linux/vmalloc.h>
85#include <linux/workqueue.h>
86#include <linux/kmemleak.h>
87#include <linux/sched.h>
88#include <linux/sched/mm.h>
89#include <linux/memcontrol.h>
90
91#include <asm/cacheflush.h>
92#include <asm/sections.h>
93#include <asm/tlbflush.h>
94#include <asm/io.h>
95
96#define CREATE_TRACE_POINTS
97#include <trace/events/percpu.h>
98
99#include "percpu-internal.h"
100
101/* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
102#define PCPU_SLOT_BASE_SHIFT 5
103/* chunks in slots below this are subject to being sidelined on failed alloc */
104#define PCPU_SLOT_FAIL_THRESHOLD 3
105
106#define PCPU_EMPTY_POP_PAGES_LOW 2
107#define PCPU_EMPTY_POP_PAGES_HIGH 4
108
109#ifdef CONFIG_SMP
110/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
111#ifndef __addr_to_pcpu_ptr
112#define __addr_to_pcpu_ptr(addr) \
113 (void __percpu *)((unsigned long)(addr) - \
114 (unsigned long)pcpu_base_addr + \
115 (unsigned long)__per_cpu_start)
116#endif
117#ifndef __pcpu_ptr_to_addr
118#define __pcpu_ptr_to_addr(ptr) \
119 (void __force *)((unsigned long)(ptr) + \
120 (unsigned long)pcpu_base_addr - \
121 (unsigned long)__per_cpu_start)
122#endif
123#else /* CONFIG_SMP */
124/* on UP, it's always identity mapped */
125#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
126#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
127#endif /* CONFIG_SMP */
128
129static int pcpu_unit_pages __ro_after_init;
130static int pcpu_unit_size __ro_after_init;
131static int pcpu_nr_units __ro_after_init;
132static int pcpu_atom_size __ro_after_init;
133int pcpu_nr_slots __ro_after_init;
134static size_t pcpu_chunk_struct_size __ro_after_init;
135
136/* cpus with the lowest and highest unit addresses */
137static unsigned int pcpu_low_unit_cpu __ro_after_init;
138static unsigned int pcpu_high_unit_cpu __ro_after_init;
139
140/* the address of the first chunk which starts with the kernel static area */
141void *pcpu_base_addr __ro_after_init;
142EXPORT_SYMBOL_GPL(pcpu_base_addr);
143
144static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
145const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
146
147/* group information, used for vm allocation */
148static int pcpu_nr_groups __ro_after_init;
149static const unsigned long *pcpu_group_offsets __ro_after_init;
150static const size_t *pcpu_group_sizes __ro_after_init;
151
152/*
153 * The first chunk which always exists. Note that unlike other
154 * chunks, this one can be allocated and mapped in several different
155 * ways and thus often doesn't live in the vmalloc area.
156 */
157struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
158
159/*
160 * Optional reserved chunk. This chunk reserves part of the first
161 * chunk and serves it for reserved allocations. When the reserved
162 * region doesn't exist, the following variable is NULL.
163 */
164struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
165
166DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
167static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
168
169struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
170
171/* chunks which need their map areas extended, protected by pcpu_lock */
172static LIST_HEAD(pcpu_map_extend_chunks);
173
174/*
175 * The number of empty populated pages, protected by pcpu_lock. The
176 * reserved chunk doesn't contribute to the count.
177 */
178int pcpu_nr_empty_pop_pages;
179
180/*
181 * The number of populated pages in use by the allocator, protected by
182 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
183 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
184 * and increments/decrements this count by 1).
185 */
186static unsigned long pcpu_nr_populated;
187
188/*
189 * Balance work is used to populate or destroy chunks asynchronously. We
190 * try to keep the number of populated free pages between
191 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
192 * empty chunk.
193 */
194static void pcpu_balance_workfn(struct work_struct *work);
195static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
196static bool pcpu_async_enabled __read_mostly;
197static bool pcpu_atomic_alloc_failed;
198
199static void pcpu_schedule_balance_work(void)
200{
201 if (pcpu_async_enabled)
202 schedule_work(&pcpu_balance_work);
203}
204
205/**
206 * pcpu_addr_in_chunk - check if the address is served from this chunk
207 * @chunk: chunk of interest
208 * @addr: percpu address
209 *
210 * RETURNS:
211 * True if the address is served from this chunk.
212 */
213static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
214{
215 void *start_addr, *end_addr;
216
217 if (!chunk)
218 return false;
219
220 start_addr = chunk->base_addr + chunk->start_offset;
221 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
222 chunk->end_offset;
223
224 return addr >= start_addr && addr < end_addr;
225}
226
227static int __pcpu_size_to_slot(int size)
228{
229 int highbit = fls(size); /* size is in bytes */
230 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
231}
232
233static int pcpu_size_to_slot(int size)
234{
235 if (size == pcpu_unit_size)
236 return pcpu_nr_slots - 1;
237 return __pcpu_size_to_slot(size);
238}
239
240static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
241{
242 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
243
244 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
245 chunk_md->contig_hint == 0)
246 return 0;
247
248 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
249}
250
251/* set the pointer to a chunk in a page struct */
252static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
253{
254 page->index = (unsigned long)pcpu;
255}
256
257/* obtain pointer to a chunk from a page struct */
258static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
259{
260 return (struct pcpu_chunk *)page->index;
261}
262
263static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
264{
265 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
266}
267
268static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
269{
270 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
271}
272
273static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
274 unsigned int cpu, int page_idx)
275{
276 return (unsigned long)chunk->base_addr +
277 pcpu_unit_page_offset(cpu, page_idx);
278}
279
280/*
281 * The following are helper functions to help access bitmaps and convert
282 * between bitmap offsets to address offsets.
283 */
284static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
285{
286 return chunk->alloc_map +
287 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
288}
289
290static unsigned long pcpu_off_to_block_index(int off)
291{
292 return off / PCPU_BITMAP_BLOCK_BITS;
293}
294
295static unsigned long pcpu_off_to_block_off(int off)
296{
297 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
298}
299
300static unsigned long pcpu_block_off_to_off(int index, int off)
301{
302 return index * PCPU_BITMAP_BLOCK_BITS + off;
303}
304
305/*
306 * pcpu_next_hint - determine which hint to use
307 * @block: block of interest
308 * @alloc_bits: size of allocation
309 *
310 * This determines if we should scan based on the scan_hint or first_free.
311 * In general, we want to scan from first_free to fulfill allocations by
312 * first fit. However, if we know a scan_hint at position scan_hint_start
313 * cannot fulfill an allocation, we can begin scanning from there knowing
314 * the contig_hint will be our fallback.
315 */
316static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
317{
318 /*
319 * The three conditions below determine if we can skip past the
320 * scan_hint. First, does the scan hint exist. Second, is the
321 * contig_hint after the scan_hint (possibly not true iff
322 * contig_hint == scan_hint). Third, is the allocation request
323 * larger than the scan_hint.
324 */
325 if (block->scan_hint &&
326 block->contig_hint_start > block->scan_hint_start &&
327 alloc_bits > block->scan_hint)
328 return block->scan_hint_start + block->scan_hint;
329
330 return block->first_free;
331}
332
333/**
334 * pcpu_next_md_free_region - finds the next hint free area
335 * @chunk: chunk of interest
336 * @bit_off: chunk offset
337 * @bits: size of free area
338 *
339 * Helper function for pcpu_for_each_md_free_region. It checks
340 * block->contig_hint and performs aggregation across blocks to find the
341 * next hint. It modifies bit_off and bits in-place to be consumed in the
342 * loop.
343 */
344static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
345 int *bits)
346{
347 int i = pcpu_off_to_block_index(*bit_off);
348 int block_off = pcpu_off_to_block_off(*bit_off);
349 struct pcpu_block_md *block;
350
351 *bits = 0;
352 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
353 block++, i++) {
354 /* handles contig area across blocks */
355 if (*bits) {
356 *bits += block->left_free;
357 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
358 continue;
359 return;
360 }
361
362 /*
363 * This checks three things. First is there a contig_hint to
364 * check. Second, have we checked this hint before by
365 * comparing the block_off. Third, is this the same as the
366 * right contig hint. In the last case, it spills over into
367 * the next block and should be handled by the contig area
368 * across blocks code.
369 */
370 *bits = block->contig_hint;
371 if (*bits && block->contig_hint_start >= block_off &&
372 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
373 *bit_off = pcpu_block_off_to_off(i,
374 block->contig_hint_start);
375 return;
376 }
377 /* reset to satisfy the second predicate above */
378 block_off = 0;
379
380 *bits = block->right_free;
381 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
382 }
383}
384
385/**
386 * pcpu_next_fit_region - finds fit areas for a given allocation request
387 * @chunk: chunk of interest
388 * @alloc_bits: size of allocation
389 * @align: alignment of area (max PAGE_SIZE)
390 * @bit_off: chunk offset
391 * @bits: size of free area
392 *
393 * Finds the next free region that is viable for use with a given size and
394 * alignment. This only returns if there is a valid area to be used for this
395 * allocation. block->first_free is returned if the allocation request fits
396 * within the block to see if the request can be fulfilled prior to the contig
397 * hint.
398 */
399static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
400 int align, int *bit_off, int *bits)
401{
402 int i = pcpu_off_to_block_index(*bit_off);
403 int block_off = pcpu_off_to_block_off(*bit_off);
404 struct pcpu_block_md *block;
405
406 *bits = 0;
407 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
408 block++, i++) {
409 /* handles contig area across blocks */
410 if (*bits) {
411 *bits += block->left_free;
412 if (*bits >= alloc_bits)
413 return;
414 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
415 continue;
416 }
417
418 /* check block->contig_hint */
419 *bits = ALIGN(block->contig_hint_start, align) -
420 block->contig_hint_start;
421 /*
422 * This uses the block offset to determine if this has been
423 * checked in the prior iteration.
424 */
425 if (block->contig_hint &&
426 block->contig_hint_start >= block_off &&
427 block->contig_hint >= *bits + alloc_bits) {
428 int start = pcpu_next_hint(block, alloc_bits);
429
430 *bits += alloc_bits + block->contig_hint_start -
431 start;
432 *bit_off = pcpu_block_off_to_off(i, start);
433 return;
434 }
435 /* reset to satisfy the second predicate above */
436 block_off = 0;
437
438 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
439 align);
440 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
441 *bit_off = pcpu_block_off_to_off(i, *bit_off);
442 if (*bits >= alloc_bits)
443 return;
444 }
445
446 /* no valid offsets were found - fail condition */
447 *bit_off = pcpu_chunk_map_bits(chunk);
448}
449
450/*
451 * Metadata free area iterators. These perform aggregation of free areas
452 * based on the metadata blocks and return the offset @bit_off and size in
453 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
454 * a fit is found for the allocation request.
455 */
456#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
457 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
458 (bit_off) < pcpu_chunk_map_bits((chunk)); \
459 (bit_off) += (bits) + 1, \
460 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
461
462#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
463 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
464 &(bits)); \
465 (bit_off) < pcpu_chunk_map_bits((chunk)); \
466 (bit_off) += (bits), \
467 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
468 &(bits)))
469
470/**
471 * pcpu_mem_zalloc - allocate memory
472 * @size: bytes to allocate
473 * @gfp: allocation flags
474 *
475 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
476 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
477 * This is to facilitate passing through whitelisted flags. The
478 * returned memory is always zeroed.
479 *
480 * RETURNS:
481 * Pointer to the allocated area on success, NULL on failure.
482 */
483static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
484{
485 if (WARN_ON_ONCE(!slab_is_available()))
486 return NULL;
487
488 if (size <= PAGE_SIZE)
489 return kzalloc(size, gfp);
490 else
491 return __vmalloc(size, gfp | __GFP_ZERO);
492}
493
494/**
495 * pcpu_mem_free - free memory
496 * @ptr: memory to free
497 *
498 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
499 */
500static void pcpu_mem_free(void *ptr)
501{
502 kvfree(ptr);
503}
504
505static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
506 bool move_front)
507{
508 if (chunk != pcpu_reserved_chunk) {
509 struct list_head *pcpu_slot;
510
511 pcpu_slot = pcpu_chunk_list(pcpu_chunk_type(chunk));
512 if (move_front)
513 list_move(&chunk->list, &pcpu_slot[slot]);
514 else
515 list_move_tail(&chunk->list, &pcpu_slot[slot]);
516 }
517}
518
519static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
520{
521 __pcpu_chunk_move(chunk, slot, true);
522}
523
524/**
525 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
526 * @chunk: chunk of interest
527 * @oslot: the previous slot it was on
528 *
529 * This function is called after an allocation or free changed @chunk.
530 * New slot according to the changed state is determined and @chunk is
531 * moved to the slot. Note that the reserved chunk is never put on
532 * chunk slots.
533 *
534 * CONTEXT:
535 * pcpu_lock.
536 */
537static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
538{
539 int nslot = pcpu_chunk_slot(chunk);
540
541 if (oslot != nslot)
542 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
543}
544
545/*
546 * pcpu_update_empty_pages - update empty page counters
547 * @chunk: chunk of interest
548 * @nr: nr of empty pages
549 *
550 * This is used to keep track of the empty pages now based on the premise
551 * a md_block covers a page. The hint update functions recognize if a block
552 * is made full or broken to calculate deltas for keeping track of free pages.
553 */
554static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
555{
556 chunk->nr_empty_pop_pages += nr;
557 if (chunk != pcpu_reserved_chunk)
558 pcpu_nr_empty_pop_pages += nr;
559}
560
561/*
562 * pcpu_region_overlap - determines if two regions overlap
563 * @a: start of first region, inclusive
564 * @b: end of first region, exclusive
565 * @x: start of second region, inclusive
566 * @y: end of second region, exclusive
567 *
568 * This is used to determine if the hint region [a, b) overlaps with the
569 * allocated region [x, y).
570 */
571static inline bool pcpu_region_overlap(int a, int b, int x, int y)
572{
573 return (a < y) && (x < b);
574}
575
576/**
577 * pcpu_block_update - updates a block given a free area
578 * @block: block of interest
579 * @start: start offset in block
580 * @end: end offset in block
581 *
582 * Updates a block given a known free area. The region [start, end) is
583 * expected to be the entirety of the free area within a block. Chooses
584 * the best starting offset if the contig hints are equal.
585 */
586static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
587{
588 int contig = end - start;
589
590 block->first_free = min(block->first_free, start);
591 if (start == 0)
592 block->left_free = contig;
593
594 if (end == block->nr_bits)
595 block->right_free = contig;
596
597 if (contig > block->contig_hint) {
598 /* promote the old contig_hint to be the new scan_hint */
599 if (start > block->contig_hint_start) {
600 if (block->contig_hint > block->scan_hint) {
601 block->scan_hint_start =
602 block->contig_hint_start;
603 block->scan_hint = block->contig_hint;
604 } else if (start < block->scan_hint_start) {
605 /*
606 * The old contig_hint == scan_hint. But, the
607 * new contig is larger so hold the invariant
608 * scan_hint_start < contig_hint_start.
609 */
610 block->scan_hint = 0;
611 }
612 } else {
613 block->scan_hint = 0;
614 }
615 block->contig_hint_start = start;
616 block->contig_hint = contig;
617 } else if (contig == block->contig_hint) {
618 if (block->contig_hint_start &&
619 (!start ||
620 __ffs(start) > __ffs(block->contig_hint_start))) {
621 /* start has a better alignment so use it */
622 block->contig_hint_start = start;
623 if (start < block->scan_hint_start &&
624 block->contig_hint > block->scan_hint)
625 block->scan_hint = 0;
626 } else if (start > block->scan_hint_start ||
627 block->contig_hint > block->scan_hint) {
628 /*
629 * Knowing contig == contig_hint, update the scan_hint
630 * if it is farther than or larger than the current
631 * scan_hint.
632 */
633 block->scan_hint_start = start;
634 block->scan_hint = contig;
635 }
636 } else {
637 /*
638 * The region is smaller than the contig_hint. So only update
639 * the scan_hint if it is larger than or equal and farther than
640 * the current scan_hint.
641 */
642 if ((start < block->contig_hint_start &&
643 (contig > block->scan_hint ||
644 (contig == block->scan_hint &&
645 start > block->scan_hint_start)))) {
646 block->scan_hint_start = start;
647 block->scan_hint = contig;
648 }
649 }
650}
651
652/*
653 * pcpu_block_update_scan - update a block given a free area from a scan
654 * @chunk: chunk of interest
655 * @bit_off: chunk offset
656 * @bits: size of free area
657 *
658 * Finding the final allocation spot first goes through pcpu_find_block_fit()
659 * to find a block that can hold the allocation and then pcpu_alloc_area()
660 * where a scan is used. When allocations require specific alignments,
661 * we can inadvertently create holes which will not be seen in the alloc
662 * or free paths.
663 *
664 * This takes a given free area hole and updates a block as it may change the
665 * scan_hint. We need to scan backwards to ensure we don't miss free bits
666 * from alignment.
667 */
668static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
669 int bits)
670{
671 int s_off = pcpu_off_to_block_off(bit_off);
672 int e_off = s_off + bits;
673 int s_index, l_bit;
674 struct pcpu_block_md *block;
675
676 if (e_off > PCPU_BITMAP_BLOCK_BITS)
677 return;
678
679 s_index = pcpu_off_to_block_index(bit_off);
680 block = chunk->md_blocks + s_index;
681
682 /* scan backwards in case of alignment skipping free bits */
683 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
684 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
685
686 pcpu_block_update(block, s_off, e_off);
687}
688
689/**
690 * pcpu_chunk_refresh_hint - updates metadata about a chunk
691 * @chunk: chunk of interest
692 * @full_scan: if we should scan from the beginning
693 *
694 * Iterates over the metadata blocks to find the largest contig area.
695 * A full scan can be avoided on the allocation path as this is triggered
696 * if we broke the contig_hint. In doing so, the scan_hint will be before
697 * the contig_hint or after if the scan_hint == contig_hint. This cannot
698 * be prevented on freeing as we want to find the largest area possibly
699 * spanning blocks.
700 */
701static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
702{
703 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
704 int bit_off, bits;
705
706 /* promote scan_hint to contig_hint */
707 if (!full_scan && chunk_md->scan_hint) {
708 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
709 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
710 chunk_md->contig_hint = chunk_md->scan_hint;
711 chunk_md->scan_hint = 0;
712 } else {
713 bit_off = chunk_md->first_free;
714 chunk_md->contig_hint = 0;
715 }
716
717 bits = 0;
718 pcpu_for_each_md_free_region(chunk, bit_off, bits)
719 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
720}
721
722/**
723 * pcpu_block_refresh_hint
724 * @chunk: chunk of interest
725 * @index: index of the metadata block
726 *
727 * Scans over the block beginning at first_free and updates the block
728 * metadata accordingly.
729 */
730static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
731{
732 struct pcpu_block_md *block = chunk->md_blocks + index;
733 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
734 unsigned int rs, re, start; /* region start, region end */
735
736 /* promote scan_hint to contig_hint */
737 if (block->scan_hint) {
738 start = block->scan_hint_start + block->scan_hint;
739 block->contig_hint_start = block->scan_hint_start;
740 block->contig_hint = block->scan_hint;
741 block->scan_hint = 0;
742 } else {
743 start = block->first_free;
744 block->contig_hint = 0;
745 }
746
747 block->right_free = 0;
748
749 /* iterate over free areas and update the contig hints */
750 bitmap_for_each_clear_region(alloc_map, rs, re, start,
751 PCPU_BITMAP_BLOCK_BITS)
752 pcpu_block_update(block, rs, re);
753}
754
755/**
756 * pcpu_block_update_hint_alloc - update hint on allocation path
757 * @chunk: chunk of interest
758 * @bit_off: chunk offset
759 * @bits: size of request
760 *
761 * Updates metadata for the allocation path. The metadata only has to be
762 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
763 * scans are required if the block's contig hint is broken.
764 */
765static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
766 int bits)
767{
768 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
769 int nr_empty_pages = 0;
770 struct pcpu_block_md *s_block, *e_block, *block;
771 int s_index, e_index; /* block indexes of the freed allocation */
772 int s_off, e_off; /* block offsets of the freed allocation */
773
774 /*
775 * Calculate per block offsets.
776 * The calculation uses an inclusive range, but the resulting offsets
777 * are [start, end). e_index always points to the last block in the
778 * range.
779 */
780 s_index = pcpu_off_to_block_index(bit_off);
781 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
782 s_off = pcpu_off_to_block_off(bit_off);
783 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
784
785 s_block = chunk->md_blocks + s_index;
786 e_block = chunk->md_blocks + e_index;
787
788 /*
789 * Update s_block.
790 * block->first_free must be updated if the allocation takes its place.
791 * If the allocation breaks the contig_hint, a scan is required to
792 * restore this hint.
793 */
794 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
795 nr_empty_pages++;
796
797 if (s_off == s_block->first_free)
798 s_block->first_free = find_next_zero_bit(
799 pcpu_index_alloc_map(chunk, s_index),
800 PCPU_BITMAP_BLOCK_BITS,
801 s_off + bits);
802
803 if (pcpu_region_overlap(s_block->scan_hint_start,
804 s_block->scan_hint_start + s_block->scan_hint,
805 s_off,
806 s_off + bits))
807 s_block->scan_hint = 0;
808
809 if (pcpu_region_overlap(s_block->contig_hint_start,
810 s_block->contig_hint_start +
811 s_block->contig_hint,
812 s_off,
813 s_off + bits)) {
814 /* block contig hint is broken - scan to fix it */
815 if (!s_off)
816 s_block->left_free = 0;
817 pcpu_block_refresh_hint(chunk, s_index);
818 } else {
819 /* update left and right contig manually */
820 s_block->left_free = min(s_block->left_free, s_off);
821 if (s_index == e_index)
822 s_block->right_free = min_t(int, s_block->right_free,
823 PCPU_BITMAP_BLOCK_BITS - e_off);
824 else
825 s_block->right_free = 0;
826 }
827
828 /*
829 * Update e_block.
830 */
831 if (s_index != e_index) {
832 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
833 nr_empty_pages++;
834
835 /*
836 * When the allocation is across blocks, the end is along
837 * the left part of the e_block.
838 */
839 e_block->first_free = find_next_zero_bit(
840 pcpu_index_alloc_map(chunk, e_index),
841 PCPU_BITMAP_BLOCK_BITS, e_off);
842
843 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
844 /* reset the block */
845 e_block++;
846 } else {
847 if (e_off > e_block->scan_hint_start)
848 e_block->scan_hint = 0;
849
850 e_block->left_free = 0;
851 if (e_off > e_block->contig_hint_start) {
852 /* contig hint is broken - scan to fix it */
853 pcpu_block_refresh_hint(chunk, e_index);
854 } else {
855 e_block->right_free =
856 min_t(int, e_block->right_free,
857 PCPU_BITMAP_BLOCK_BITS - e_off);
858 }
859 }
860
861 /* update in-between md_blocks */
862 nr_empty_pages += (e_index - s_index - 1);
863 for (block = s_block + 1; block < e_block; block++) {
864 block->scan_hint = 0;
865 block->contig_hint = 0;
866 block->left_free = 0;
867 block->right_free = 0;
868 }
869 }
870
871 if (nr_empty_pages)
872 pcpu_update_empty_pages(chunk, -nr_empty_pages);
873
874 if (pcpu_region_overlap(chunk_md->scan_hint_start,
875 chunk_md->scan_hint_start +
876 chunk_md->scan_hint,
877 bit_off,
878 bit_off + bits))
879 chunk_md->scan_hint = 0;
880
881 /*
882 * The only time a full chunk scan is required is if the chunk
883 * contig hint is broken. Otherwise, it means a smaller space
884 * was used and therefore the chunk contig hint is still correct.
885 */
886 if (pcpu_region_overlap(chunk_md->contig_hint_start,
887 chunk_md->contig_hint_start +
888 chunk_md->contig_hint,
889 bit_off,
890 bit_off + bits))
891 pcpu_chunk_refresh_hint(chunk, false);
892}
893
894/**
895 * pcpu_block_update_hint_free - updates the block hints on the free path
896 * @chunk: chunk of interest
897 * @bit_off: chunk offset
898 * @bits: size of request
899 *
900 * Updates metadata for the allocation path. This avoids a blind block
901 * refresh by making use of the block contig hints. If this fails, it scans
902 * forward and backward to determine the extent of the free area. This is
903 * capped at the boundary of blocks.
904 *
905 * A chunk update is triggered if a page becomes free, a block becomes free,
906 * or the free spans across blocks. This tradeoff is to minimize iterating
907 * over the block metadata to update chunk_md->contig_hint.
908 * chunk_md->contig_hint may be off by up to a page, but it will never be more
909 * than the available space. If the contig hint is contained in one block, it
910 * will be accurate.
911 */
912static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
913 int bits)
914{
915 int nr_empty_pages = 0;
916 struct pcpu_block_md *s_block, *e_block, *block;
917 int s_index, e_index; /* block indexes of the freed allocation */
918 int s_off, e_off; /* block offsets of the freed allocation */
919 int start, end; /* start and end of the whole free area */
920
921 /*
922 * Calculate per block offsets.
923 * The calculation uses an inclusive range, but the resulting offsets
924 * are [start, end). e_index always points to the last block in the
925 * range.
926 */
927 s_index = pcpu_off_to_block_index(bit_off);
928 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
929 s_off = pcpu_off_to_block_off(bit_off);
930 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
931
932 s_block = chunk->md_blocks + s_index;
933 e_block = chunk->md_blocks + e_index;
934
935 /*
936 * Check if the freed area aligns with the block->contig_hint.
937 * If it does, then the scan to find the beginning/end of the
938 * larger free area can be avoided.
939 *
940 * start and end refer to beginning and end of the free area
941 * within each their respective blocks. This is not necessarily
942 * the entire free area as it may span blocks past the beginning
943 * or end of the block.
944 */
945 start = s_off;
946 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
947 start = s_block->contig_hint_start;
948 } else {
949 /*
950 * Scan backwards to find the extent of the free area.
951 * find_last_bit returns the starting bit, so if the start bit
952 * is returned, that means there was no last bit and the
953 * remainder of the chunk is free.
954 */
955 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
956 start);
957 start = (start == l_bit) ? 0 : l_bit + 1;
958 }
959
960 end = e_off;
961 if (e_off == e_block->contig_hint_start)
962 end = e_block->contig_hint_start + e_block->contig_hint;
963 else
964 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
965 PCPU_BITMAP_BLOCK_BITS, end);
966
967 /* update s_block */
968 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
969 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
970 nr_empty_pages++;
971 pcpu_block_update(s_block, start, e_off);
972
973 /* freeing in the same block */
974 if (s_index != e_index) {
975 /* update e_block */
976 if (end == PCPU_BITMAP_BLOCK_BITS)
977 nr_empty_pages++;
978 pcpu_block_update(e_block, 0, end);
979
980 /* reset md_blocks in the middle */
981 nr_empty_pages += (e_index - s_index - 1);
982 for (block = s_block + 1; block < e_block; block++) {
983 block->first_free = 0;
984 block->scan_hint = 0;
985 block->contig_hint_start = 0;
986 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
987 block->left_free = PCPU_BITMAP_BLOCK_BITS;
988 block->right_free = PCPU_BITMAP_BLOCK_BITS;
989 }
990 }
991
992 if (nr_empty_pages)
993 pcpu_update_empty_pages(chunk, nr_empty_pages);
994
995 /*
996 * Refresh chunk metadata when the free makes a block free or spans
997 * across blocks. The contig_hint may be off by up to a page, but if
998 * the contig_hint is contained in a block, it will be accurate with
999 * the else condition below.
1000 */
1001 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1002 pcpu_chunk_refresh_hint(chunk, true);
1003 else
1004 pcpu_block_update(&chunk->chunk_md,
1005 pcpu_block_off_to_off(s_index, start),
1006 end);
1007}
1008
1009/**
1010 * pcpu_is_populated - determines if the region is populated
1011 * @chunk: chunk of interest
1012 * @bit_off: chunk offset
1013 * @bits: size of area
1014 * @next_off: return value for the next offset to start searching
1015 *
1016 * For atomic allocations, check if the backing pages are populated.
1017 *
1018 * RETURNS:
1019 * Bool if the backing pages are populated.
1020 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1021 */
1022static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1023 int *next_off)
1024{
1025 unsigned int page_start, page_end, rs, re;
1026
1027 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1028 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1029
1030 rs = page_start;
1031 bitmap_next_clear_region(chunk->populated, &rs, &re, page_end);
1032 if (rs >= page_end)
1033 return true;
1034
1035 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1036 return false;
1037}
1038
1039/**
1040 * pcpu_find_block_fit - finds the block index to start searching
1041 * @chunk: chunk of interest
1042 * @alloc_bits: size of request in allocation units
1043 * @align: alignment of area (max PAGE_SIZE bytes)
1044 * @pop_only: use populated regions only
1045 *
1046 * Given a chunk and an allocation spec, find the offset to begin searching
1047 * for a free region. This iterates over the bitmap metadata blocks to
1048 * find an offset that will be guaranteed to fit the requirements. It is
1049 * not quite first fit as if the allocation does not fit in the contig hint
1050 * of a block or chunk, it is skipped. This errs on the side of caution
1051 * to prevent excess iteration. Poor alignment can cause the allocator to
1052 * skip over blocks and chunks that have valid free areas.
1053 *
1054 * RETURNS:
1055 * The offset in the bitmap to begin searching.
1056 * -1 if no offset is found.
1057 */
1058static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1059 size_t align, bool pop_only)
1060{
1061 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1062 int bit_off, bits, next_off;
1063
1064 /*
1065 * Check to see if the allocation can fit in the chunk's contig hint.
1066 * This is an optimization to prevent scanning by assuming if it
1067 * cannot fit in the global hint, there is memory pressure and creating
1068 * a new chunk would happen soon.
1069 */
1070 bit_off = ALIGN(chunk_md->contig_hint_start, align) -
1071 chunk_md->contig_hint_start;
1072 if (bit_off + alloc_bits > chunk_md->contig_hint)
1073 return -1;
1074
1075 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1076 bits = 0;
1077 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1078 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1079 &next_off))
1080 break;
1081
1082 bit_off = next_off;
1083 bits = 0;
1084 }
1085
1086 if (bit_off == pcpu_chunk_map_bits(chunk))
1087 return -1;
1088
1089 return bit_off;
1090}
1091
1092/*
1093 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1094 * @map: the address to base the search on
1095 * @size: the bitmap size in bits
1096 * @start: the bitnumber to start searching at
1097 * @nr: the number of zeroed bits we're looking for
1098 * @align_mask: alignment mask for zero area
1099 * @largest_off: offset of the largest area skipped
1100 * @largest_bits: size of the largest area skipped
1101 *
1102 * The @align_mask should be one less than a power of 2.
1103 *
1104 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1105 * the largest area that was skipped. This is imperfect, but in general is
1106 * good enough. The largest remembered region is the largest failed region
1107 * seen. This does not include anything we possibly skipped due to alignment.
1108 * pcpu_block_update_scan() does scan backwards to try and recover what was
1109 * lost to alignment. While this can cause scanning to miss earlier possible
1110 * free areas, smaller allocations will eventually fill those holes.
1111 */
1112static unsigned long pcpu_find_zero_area(unsigned long *map,
1113 unsigned long size,
1114 unsigned long start,
1115 unsigned long nr,
1116 unsigned long align_mask,
1117 unsigned long *largest_off,
1118 unsigned long *largest_bits)
1119{
1120 unsigned long index, end, i, area_off, area_bits;
1121again:
1122 index = find_next_zero_bit(map, size, start);
1123
1124 /* Align allocation */
1125 index = __ALIGN_MASK(index, align_mask);
1126 area_off = index;
1127
1128 end = index + nr;
1129 if (end > size)
1130 return end;
1131 i = find_next_bit(map, end, index);
1132 if (i < end) {
1133 area_bits = i - area_off;
1134 /* remember largest unused area with best alignment */
1135 if (area_bits > *largest_bits ||
1136 (area_bits == *largest_bits && *largest_off &&
1137 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1138 *largest_off = area_off;
1139 *largest_bits = area_bits;
1140 }
1141
1142 start = i + 1;
1143 goto again;
1144 }
1145 return index;
1146}
1147
1148/**
1149 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1150 * @chunk: chunk of interest
1151 * @alloc_bits: size of request in allocation units
1152 * @align: alignment of area (max PAGE_SIZE)
1153 * @start: bit_off to start searching
1154 *
1155 * This function takes in a @start offset to begin searching to fit an
1156 * allocation of @alloc_bits with alignment @align. It needs to scan
1157 * the allocation map because if it fits within the block's contig hint,
1158 * @start will be block->first_free. This is an attempt to fill the
1159 * allocation prior to breaking the contig hint. The allocation and
1160 * boundary maps are updated accordingly if it confirms a valid
1161 * free area.
1162 *
1163 * RETURNS:
1164 * Allocated addr offset in @chunk on success.
1165 * -1 if no matching area is found.
1166 */
1167static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1168 size_t align, int start)
1169{
1170 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1171 size_t align_mask = (align) ? (align - 1) : 0;
1172 unsigned long area_off = 0, area_bits = 0;
1173 int bit_off, end, oslot;
1174
1175 lockdep_assert_held(&pcpu_lock);
1176
1177 oslot = pcpu_chunk_slot(chunk);
1178
1179 /*
1180 * Search to find a fit.
1181 */
1182 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1183 pcpu_chunk_map_bits(chunk));
1184 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1185 align_mask, &area_off, &area_bits);
1186 if (bit_off >= end)
1187 return -1;
1188
1189 if (area_bits)
1190 pcpu_block_update_scan(chunk, area_off, area_bits);
1191
1192 /* update alloc map */
1193 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1194
1195 /* update boundary map */
1196 set_bit(bit_off, chunk->bound_map);
1197 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1198 set_bit(bit_off + alloc_bits, chunk->bound_map);
1199
1200 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1201
1202 /* update first free bit */
1203 if (bit_off == chunk_md->first_free)
1204 chunk_md->first_free = find_next_zero_bit(
1205 chunk->alloc_map,
1206 pcpu_chunk_map_bits(chunk),
1207 bit_off + alloc_bits);
1208
1209 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1210
1211 pcpu_chunk_relocate(chunk, oslot);
1212
1213 return bit_off * PCPU_MIN_ALLOC_SIZE;
1214}
1215
1216/**
1217 * pcpu_free_area - frees the corresponding offset
1218 * @chunk: chunk of interest
1219 * @off: addr offset into chunk
1220 *
1221 * This function determines the size of an allocation to free using
1222 * the boundary bitmap and clears the allocation map.
1223 *
1224 * RETURNS:
1225 * Number of freed bytes.
1226 */
1227static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1228{
1229 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1230 int bit_off, bits, end, oslot, freed;
1231
1232 lockdep_assert_held(&pcpu_lock);
1233 pcpu_stats_area_dealloc(chunk);
1234
1235 oslot = pcpu_chunk_slot(chunk);
1236
1237 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1238
1239 /* find end index */
1240 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1241 bit_off + 1);
1242 bits = end - bit_off;
1243 bitmap_clear(chunk->alloc_map, bit_off, bits);
1244
1245 freed = bits * PCPU_MIN_ALLOC_SIZE;
1246
1247 /* update metadata */
1248 chunk->free_bytes += freed;
1249
1250 /* update first free bit */
1251 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1252
1253 pcpu_block_update_hint_free(chunk, bit_off, bits);
1254
1255 pcpu_chunk_relocate(chunk, oslot);
1256
1257 return freed;
1258}
1259
1260static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1261{
1262 block->scan_hint = 0;
1263 block->contig_hint = nr_bits;
1264 block->left_free = nr_bits;
1265 block->right_free = nr_bits;
1266 block->first_free = 0;
1267 block->nr_bits = nr_bits;
1268}
1269
1270static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1271{
1272 struct pcpu_block_md *md_block;
1273
1274 /* init the chunk's block */
1275 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1276
1277 for (md_block = chunk->md_blocks;
1278 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1279 md_block++)
1280 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1281}
1282
1283/**
1284 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1285 * @tmp_addr: the start of the region served
1286 * @map_size: size of the region served
1287 *
1288 * This is responsible for creating the chunks that serve the first chunk. The
1289 * base_addr is page aligned down of @tmp_addr while the region end is page
1290 * aligned up. Offsets are kept track of to determine the region served. All
1291 * this is done to appease the bitmap allocator in avoiding partial blocks.
1292 *
1293 * RETURNS:
1294 * Chunk serving the region at @tmp_addr of @map_size.
1295 */
1296static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1297 int map_size)
1298{
1299 struct pcpu_chunk *chunk;
1300 unsigned long aligned_addr, lcm_align;
1301 int start_offset, offset_bits, region_size, region_bits;
1302 size_t alloc_size;
1303
1304 /* region calculations */
1305 aligned_addr = tmp_addr & PAGE_MASK;
1306
1307 start_offset = tmp_addr - aligned_addr;
1308
1309 /*
1310 * Align the end of the region with the LCM of PAGE_SIZE and
1311 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1312 * the other.
1313 */
1314 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1315 region_size = ALIGN(start_offset + map_size, lcm_align);
1316
1317 /* allocate chunk */
1318 alloc_size = sizeof(struct pcpu_chunk) +
1319 BITS_TO_LONGS(region_size >> PAGE_SHIFT) * sizeof(unsigned long);
1320 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1321 if (!chunk)
1322 panic("%s: Failed to allocate %zu bytes\n", __func__,
1323 alloc_size);
1324
1325 INIT_LIST_HEAD(&chunk->list);
1326
1327 chunk->base_addr = (void *)aligned_addr;
1328 chunk->start_offset = start_offset;
1329 chunk->end_offset = region_size - chunk->start_offset - map_size;
1330
1331 chunk->nr_pages = region_size >> PAGE_SHIFT;
1332 region_bits = pcpu_chunk_map_bits(chunk);
1333
1334 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1335 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1336 if (!chunk->alloc_map)
1337 panic("%s: Failed to allocate %zu bytes\n", __func__,
1338 alloc_size);
1339
1340 alloc_size =
1341 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1342 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1343 if (!chunk->bound_map)
1344 panic("%s: Failed to allocate %zu bytes\n", __func__,
1345 alloc_size);
1346
1347 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1348 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1349 if (!chunk->md_blocks)
1350 panic("%s: Failed to allocate %zu bytes\n", __func__,
1351 alloc_size);
1352
1353#ifdef CONFIG_MEMCG_KMEM
1354 /* first chunk isn't memcg-aware */
1355 chunk->obj_cgroups = NULL;
1356#endif
1357 pcpu_init_md_blocks(chunk);
1358
1359 /* manage populated page bitmap */
1360 chunk->immutable = true;
1361 bitmap_fill(chunk->populated, chunk->nr_pages);
1362 chunk->nr_populated = chunk->nr_pages;
1363 chunk->nr_empty_pop_pages = chunk->nr_pages;
1364
1365 chunk->free_bytes = map_size;
1366
1367 if (chunk->start_offset) {
1368 /* hide the beginning of the bitmap */
1369 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1370 bitmap_set(chunk->alloc_map, 0, offset_bits);
1371 set_bit(0, chunk->bound_map);
1372 set_bit(offset_bits, chunk->bound_map);
1373
1374 chunk->chunk_md.first_free = offset_bits;
1375
1376 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1377 }
1378
1379 if (chunk->end_offset) {
1380 /* hide the end of the bitmap */
1381 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1382 bitmap_set(chunk->alloc_map,
1383 pcpu_chunk_map_bits(chunk) - offset_bits,
1384 offset_bits);
1385 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1386 chunk->bound_map);
1387 set_bit(region_bits, chunk->bound_map);
1388
1389 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1390 - offset_bits, offset_bits);
1391 }
1392
1393 return chunk;
1394}
1395
1396static struct pcpu_chunk *pcpu_alloc_chunk(enum pcpu_chunk_type type, gfp_t gfp)
1397{
1398 struct pcpu_chunk *chunk;
1399 int region_bits;
1400
1401 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1402 if (!chunk)
1403 return NULL;
1404
1405 INIT_LIST_HEAD(&chunk->list);
1406 chunk->nr_pages = pcpu_unit_pages;
1407 region_bits = pcpu_chunk_map_bits(chunk);
1408
1409 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1410 sizeof(chunk->alloc_map[0]), gfp);
1411 if (!chunk->alloc_map)
1412 goto alloc_map_fail;
1413
1414 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1415 sizeof(chunk->bound_map[0]), gfp);
1416 if (!chunk->bound_map)
1417 goto bound_map_fail;
1418
1419 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1420 sizeof(chunk->md_blocks[0]), gfp);
1421 if (!chunk->md_blocks)
1422 goto md_blocks_fail;
1423
1424#ifdef CONFIG_MEMCG_KMEM
1425 if (pcpu_is_memcg_chunk(type)) {
1426 chunk->obj_cgroups =
1427 pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1428 sizeof(struct obj_cgroup *), gfp);
1429 if (!chunk->obj_cgroups)
1430 goto objcg_fail;
1431 }
1432#endif
1433
1434 pcpu_init_md_blocks(chunk);
1435
1436 /* init metadata */
1437 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1438
1439 return chunk;
1440
1441#ifdef CONFIG_MEMCG_KMEM
1442objcg_fail:
1443 pcpu_mem_free(chunk->md_blocks);
1444#endif
1445md_blocks_fail:
1446 pcpu_mem_free(chunk->bound_map);
1447bound_map_fail:
1448 pcpu_mem_free(chunk->alloc_map);
1449alloc_map_fail:
1450 pcpu_mem_free(chunk);
1451
1452 return NULL;
1453}
1454
1455static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1456{
1457 if (!chunk)
1458 return;
1459#ifdef CONFIG_MEMCG_KMEM
1460 pcpu_mem_free(chunk->obj_cgroups);
1461#endif
1462 pcpu_mem_free(chunk->md_blocks);
1463 pcpu_mem_free(chunk->bound_map);
1464 pcpu_mem_free(chunk->alloc_map);
1465 pcpu_mem_free(chunk);
1466}
1467
1468/**
1469 * pcpu_chunk_populated - post-population bookkeeping
1470 * @chunk: pcpu_chunk which got populated
1471 * @page_start: the start page
1472 * @page_end: the end page
1473 *
1474 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1475 * the bookkeeping information accordingly. Must be called after each
1476 * successful population.
1477 *
1478 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1479 * is to serve an allocation in that area.
1480 */
1481static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1482 int page_end)
1483{
1484 int nr = page_end - page_start;
1485
1486 lockdep_assert_held(&pcpu_lock);
1487
1488 bitmap_set(chunk->populated, page_start, nr);
1489 chunk->nr_populated += nr;
1490 pcpu_nr_populated += nr;
1491
1492 pcpu_update_empty_pages(chunk, nr);
1493}
1494
1495/**
1496 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1497 * @chunk: pcpu_chunk which got depopulated
1498 * @page_start: the start page
1499 * @page_end: the end page
1500 *
1501 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1502 * Update the bookkeeping information accordingly. Must be called after
1503 * each successful depopulation.
1504 */
1505static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1506 int page_start, int page_end)
1507{
1508 int nr = page_end - page_start;
1509
1510 lockdep_assert_held(&pcpu_lock);
1511
1512 bitmap_clear(chunk->populated, page_start, nr);
1513 chunk->nr_populated -= nr;
1514 pcpu_nr_populated -= nr;
1515
1516 pcpu_update_empty_pages(chunk, -nr);
1517}
1518
1519/*
1520 * Chunk management implementation.
1521 *
1522 * To allow different implementations, chunk alloc/free and
1523 * [de]population are implemented in a separate file which is pulled
1524 * into this file and compiled together. The following functions
1525 * should be implemented.
1526 *
1527 * pcpu_populate_chunk - populate the specified range of a chunk
1528 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1529 * pcpu_create_chunk - create a new chunk
1530 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1531 * pcpu_addr_to_page - translate address to physical address
1532 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1533 */
1534static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1535 int page_start, int page_end, gfp_t gfp);
1536static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1537 int page_start, int page_end);
1538static struct pcpu_chunk *pcpu_create_chunk(enum pcpu_chunk_type type,
1539 gfp_t gfp);
1540static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1541static struct page *pcpu_addr_to_page(void *addr);
1542static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1543
1544#ifdef CONFIG_NEED_PER_CPU_KM
1545#include "percpu-km.c"
1546#else
1547#include "percpu-vm.c"
1548#endif
1549
1550/**
1551 * pcpu_chunk_addr_search - determine chunk containing specified address
1552 * @addr: address for which the chunk needs to be determined.
1553 *
1554 * This is an internal function that handles all but static allocations.
1555 * Static percpu address values should never be passed into the allocator.
1556 *
1557 * RETURNS:
1558 * The address of the found chunk.
1559 */
1560static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1561{
1562 /* is it in the dynamic region (first chunk)? */
1563 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1564 return pcpu_first_chunk;
1565
1566 /* is it in the reserved region? */
1567 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1568 return pcpu_reserved_chunk;
1569
1570 /*
1571 * The address is relative to unit0 which might be unused and
1572 * thus unmapped. Offset the address to the unit space of the
1573 * current processor before looking it up in the vmalloc
1574 * space. Note that any possible cpu id can be used here, so
1575 * there's no need to worry about preemption or cpu hotplug.
1576 */
1577 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1578 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1579}
1580
1581#ifdef CONFIG_MEMCG_KMEM
1582static enum pcpu_chunk_type pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1583 struct obj_cgroup **objcgp)
1584{
1585 struct obj_cgroup *objcg;
1586
1587 if (!memcg_kmem_enabled() || !(gfp & __GFP_ACCOUNT) ||
1588 memcg_kmem_bypass())
1589 return PCPU_CHUNK_ROOT;
1590
1591 objcg = get_obj_cgroup_from_current();
1592 if (!objcg)
1593 return PCPU_CHUNK_ROOT;
1594
1595 if (obj_cgroup_charge(objcg, gfp, size * num_possible_cpus())) {
1596 obj_cgroup_put(objcg);
1597 return PCPU_FAIL_ALLOC;
1598 }
1599
1600 *objcgp = objcg;
1601 return PCPU_CHUNK_MEMCG;
1602}
1603
1604static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1605 struct pcpu_chunk *chunk, int off,
1606 size_t size)
1607{
1608 if (!objcg)
1609 return;
1610
1611 if (chunk) {
1612 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1613
1614 rcu_read_lock();
1615 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1616 size * num_possible_cpus());
1617 rcu_read_unlock();
1618 } else {
1619 obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1620 obj_cgroup_put(objcg);
1621 }
1622}
1623
1624static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1625{
1626 struct obj_cgroup *objcg;
1627
1628 if (!pcpu_is_memcg_chunk(pcpu_chunk_type(chunk)))
1629 return;
1630
1631 objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1632 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1633
1634 obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1635
1636 rcu_read_lock();
1637 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1638 -(size * num_possible_cpus()));
1639 rcu_read_unlock();
1640
1641 obj_cgroup_put(objcg);
1642}
1643
1644#else /* CONFIG_MEMCG_KMEM */
1645static enum pcpu_chunk_type
1646pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1647{
1648 return PCPU_CHUNK_ROOT;
1649}
1650
1651static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1652 struct pcpu_chunk *chunk, int off,
1653 size_t size)
1654{
1655}
1656
1657static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1658{
1659}
1660#endif /* CONFIG_MEMCG_KMEM */
1661
1662/**
1663 * pcpu_alloc - the percpu allocator
1664 * @size: size of area to allocate in bytes
1665 * @align: alignment of area (max PAGE_SIZE)
1666 * @reserved: allocate from the reserved chunk if available
1667 * @gfp: allocation flags
1668 *
1669 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1670 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1671 * then no warning will be triggered on invalid or failed allocation
1672 * requests.
1673 *
1674 * RETURNS:
1675 * Percpu pointer to the allocated area on success, NULL on failure.
1676 */
1677static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1678 gfp_t gfp)
1679{
1680 gfp_t pcpu_gfp;
1681 bool is_atomic;
1682 bool do_warn;
1683 enum pcpu_chunk_type type;
1684 struct list_head *pcpu_slot;
1685 struct obj_cgroup *objcg = NULL;
1686 static int warn_limit = 10;
1687 struct pcpu_chunk *chunk, *next;
1688 const char *err;
1689 int slot, off, cpu, ret;
1690 unsigned long flags;
1691 void __percpu *ptr;
1692 size_t bits, bit_align;
1693
1694 gfp = current_gfp_context(gfp);
1695 /* whitelisted flags that can be passed to the backing allocators */
1696 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1697 is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1698 do_warn = !(gfp & __GFP_NOWARN);
1699
1700 /*
1701 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1702 * therefore alignment must be a minimum of that many bytes.
1703 * An allocation may have internal fragmentation from rounding up
1704 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1705 */
1706 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1707 align = PCPU_MIN_ALLOC_SIZE;
1708
1709 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1710 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1711 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1712
1713 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1714 !is_power_of_2(align))) {
1715 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1716 size, align);
1717 return NULL;
1718 }
1719
1720 type = pcpu_memcg_pre_alloc_hook(size, gfp, &objcg);
1721 if (unlikely(type == PCPU_FAIL_ALLOC))
1722 return NULL;
1723 pcpu_slot = pcpu_chunk_list(type);
1724
1725 if (!is_atomic) {
1726 /*
1727 * pcpu_balance_workfn() allocates memory under this mutex,
1728 * and it may wait for memory reclaim. Allow current task
1729 * to become OOM victim, in case of memory pressure.
1730 */
1731 if (gfp & __GFP_NOFAIL) {
1732 mutex_lock(&pcpu_alloc_mutex);
1733 } else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1734 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1735 return NULL;
1736 }
1737 }
1738
1739 spin_lock_irqsave(&pcpu_lock, flags);
1740
1741 /* serve reserved allocations from the reserved chunk if available */
1742 if (reserved && pcpu_reserved_chunk) {
1743 chunk = pcpu_reserved_chunk;
1744
1745 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1746 if (off < 0) {
1747 err = "alloc from reserved chunk failed";
1748 goto fail_unlock;
1749 }
1750
1751 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1752 if (off >= 0)
1753 goto area_found;
1754
1755 err = "alloc from reserved chunk failed";
1756 goto fail_unlock;
1757 }
1758
1759restart:
1760 /* search through normal chunks */
1761 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1762 list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) {
1763 off = pcpu_find_block_fit(chunk, bits, bit_align,
1764 is_atomic);
1765 if (off < 0) {
1766 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1767 pcpu_chunk_move(chunk, 0);
1768 continue;
1769 }
1770
1771 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1772 if (off >= 0)
1773 goto area_found;
1774
1775 }
1776 }
1777
1778 spin_unlock_irqrestore(&pcpu_lock, flags);
1779
1780 /*
1781 * No space left. Create a new chunk. We don't want multiple
1782 * tasks to create chunks simultaneously. Serialize and create iff
1783 * there's still no empty chunk after grabbing the mutex.
1784 */
1785 if (is_atomic) {
1786 err = "atomic alloc failed, no space left";
1787 goto fail;
1788 }
1789
1790 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1791 chunk = pcpu_create_chunk(type, pcpu_gfp);
1792 if (!chunk) {
1793 err = "failed to allocate new chunk";
1794 goto fail;
1795 }
1796
1797 spin_lock_irqsave(&pcpu_lock, flags);
1798 pcpu_chunk_relocate(chunk, -1);
1799 } else {
1800 spin_lock_irqsave(&pcpu_lock, flags);
1801 }
1802
1803 goto restart;
1804
1805area_found:
1806 pcpu_stats_area_alloc(chunk, size);
1807 spin_unlock_irqrestore(&pcpu_lock, flags);
1808
1809 /* populate if not all pages are already there */
1810 if (!is_atomic) {
1811 unsigned int page_start, page_end, rs, re;
1812
1813 page_start = PFN_DOWN(off);
1814 page_end = PFN_UP(off + size);
1815
1816 bitmap_for_each_clear_region(chunk->populated, rs, re,
1817 page_start, page_end) {
1818 WARN_ON(chunk->immutable);
1819
1820 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1821
1822 spin_lock_irqsave(&pcpu_lock, flags);
1823 if (ret) {
1824 pcpu_free_area(chunk, off);
1825 err = "failed to populate";
1826 goto fail_unlock;
1827 }
1828 pcpu_chunk_populated(chunk, rs, re);
1829 spin_unlock_irqrestore(&pcpu_lock, flags);
1830 }
1831
1832 mutex_unlock(&pcpu_alloc_mutex);
1833 }
1834
1835 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1836 pcpu_schedule_balance_work();
1837
1838 /* clear the areas and return address relative to base address */
1839 for_each_possible_cpu(cpu)
1840 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1841
1842 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1843 kmemleak_alloc_percpu(ptr, size, gfp);
1844
1845 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1846 chunk->base_addr, off, ptr);
1847
1848 pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1849
1850 return ptr;
1851
1852fail_unlock:
1853 spin_unlock_irqrestore(&pcpu_lock, flags);
1854fail:
1855 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1856
1857 if (!is_atomic && do_warn && warn_limit) {
1858 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1859 size, align, is_atomic, err);
1860 dump_stack();
1861 if (!--warn_limit)
1862 pr_info("limit reached, disable warning\n");
1863 }
1864 if (is_atomic) {
1865 /* see the flag handling in pcpu_blance_workfn() */
1866 pcpu_atomic_alloc_failed = true;
1867 pcpu_schedule_balance_work();
1868 } else {
1869 mutex_unlock(&pcpu_alloc_mutex);
1870 }
1871
1872 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1873
1874 return NULL;
1875}
1876
1877/**
1878 * __alloc_percpu_gfp - allocate dynamic percpu area
1879 * @size: size of area to allocate in bytes
1880 * @align: alignment of area (max PAGE_SIZE)
1881 * @gfp: allocation flags
1882 *
1883 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1884 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1885 * be called from any context but is a lot more likely to fail. If @gfp
1886 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1887 * allocation requests.
1888 *
1889 * RETURNS:
1890 * Percpu pointer to the allocated area on success, NULL on failure.
1891 */
1892void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1893{
1894 return pcpu_alloc(size, align, false, gfp);
1895}
1896EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1897
1898/**
1899 * __alloc_percpu - allocate dynamic percpu area
1900 * @size: size of area to allocate in bytes
1901 * @align: alignment of area (max PAGE_SIZE)
1902 *
1903 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1904 */
1905void __percpu *__alloc_percpu(size_t size, size_t align)
1906{
1907 return pcpu_alloc(size, align, false, GFP_KERNEL);
1908}
1909EXPORT_SYMBOL_GPL(__alloc_percpu);
1910
1911/**
1912 * __alloc_reserved_percpu - allocate reserved percpu area
1913 * @size: size of area to allocate in bytes
1914 * @align: alignment of area (max PAGE_SIZE)
1915 *
1916 * Allocate zero-filled percpu area of @size bytes aligned at @align
1917 * from reserved percpu area if arch has set it up; otherwise,
1918 * allocation is served from the same dynamic area. Might sleep.
1919 * Might trigger writeouts.
1920 *
1921 * CONTEXT:
1922 * Does GFP_KERNEL allocation.
1923 *
1924 * RETURNS:
1925 * Percpu pointer to the allocated area on success, NULL on failure.
1926 */
1927void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1928{
1929 return pcpu_alloc(size, align, true, GFP_KERNEL);
1930}
1931
1932/**
1933 * __pcpu_balance_workfn - manage the amount of free chunks and populated pages
1934 * @type: chunk type
1935 *
1936 * Reclaim all fully free chunks except for the first one. This is also
1937 * responsible for maintaining the pool of empty populated pages. However,
1938 * it is possible that this is called when physical memory is scarce causing
1939 * OOM killer to be triggered. We should avoid doing so until an actual
1940 * allocation causes the failure as it is possible that requests can be
1941 * serviced from already backed regions.
1942 */
1943static void __pcpu_balance_workfn(enum pcpu_chunk_type type)
1944{
1945 /* gfp flags passed to underlying allocators */
1946 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
1947 LIST_HEAD(to_free);
1948 struct list_head *pcpu_slot = pcpu_chunk_list(type);
1949 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1950 struct pcpu_chunk *chunk, *next;
1951 int slot, nr_to_pop, ret;
1952
1953 /*
1954 * There's no reason to keep around multiple unused chunks and VM
1955 * areas can be scarce. Destroy all free chunks except for one.
1956 */
1957 mutex_lock(&pcpu_alloc_mutex);
1958 spin_lock_irq(&pcpu_lock);
1959
1960 list_for_each_entry_safe(chunk, next, free_head, list) {
1961 WARN_ON(chunk->immutable);
1962
1963 /* spare the first one */
1964 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1965 continue;
1966
1967 list_move(&chunk->list, &to_free);
1968 }
1969
1970 spin_unlock_irq(&pcpu_lock);
1971
1972 list_for_each_entry_safe(chunk, next, &to_free, list) {
1973 unsigned int rs, re;
1974
1975 bitmap_for_each_set_region(chunk->populated, rs, re, 0,
1976 chunk->nr_pages) {
1977 pcpu_depopulate_chunk(chunk, rs, re);
1978 spin_lock_irq(&pcpu_lock);
1979 pcpu_chunk_depopulated(chunk, rs, re);
1980 spin_unlock_irq(&pcpu_lock);
1981 }
1982 pcpu_destroy_chunk(chunk);
1983 cond_resched();
1984 }
1985
1986 /*
1987 * Ensure there are certain number of free populated pages for
1988 * atomic allocs. Fill up from the most packed so that atomic
1989 * allocs don't increase fragmentation. If atomic allocation
1990 * failed previously, always populate the maximum amount. This
1991 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1992 * failing indefinitely; however, large atomic allocs are not
1993 * something we support properly and can be highly unreliable and
1994 * inefficient.
1995 */
1996retry_pop:
1997 if (pcpu_atomic_alloc_failed) {
1998 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1999 /* best effort anyway, don't worry about synchronization */
2000 pcpu_atomic_alloc_failed = false;
2001 } else {
2002 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2003 pcpu_nr_empty_pop_pages,
2004 0, PCPU_EMPTY_POP_PAGES_HIGH);
2005 }
2006
2007 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
2008 unsigned int nr_unpop = 0, rs, re;
2009
2010 if (!nr_to_pop)
2011 break;
2012
2013 spin_lock_irq(&pcpu_lock);
2014 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
2015 nr_unpop = chunk->nr_pages - chunk->nr_populated;
2016 if (nr_unpop)
2017 break;
2018 }
2019 spin_unlock_irq(&pcpu_lock);
2020
2021 if (!nr_unpop)
2022 continue;
2023
2024 /* @chunk can't go away while pcpu_alloc_mutex is held */
2025 bitmap_for_each_clear_region(chunk->populated, rs, re, 0,
2026 chunk->nr_pages) {
2027 int nr = min_t(int, re - rs, nr_to_pop);
2028
2029 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2030 if (!ret) {
2031 nr_to_pop -= nr;
2032 spin_lock_irq(&pcpu_lock);
2033 pcpu_chunk_populated(chunk, rs, rs + nr);
2034 spin_unlock_irq(&pcpu_lock);
2035 } else {
2036 nr_to_pop = 0;
2037 }
2038
2039 if (!nr_to_pop)
2040 break;
2041 }
2042 }
2043
2044 if (nr_to_pop) {
2045 /* ran out of chunks to populate, create a new one and retry */
2046 chunk = pcpu_create_chunk(type, gfp);
2047 if (chunk) {
2048 spin_lock_irq(&pcpu_lock);
2049 pcpu_chunk_relocate(chunk, -1);
2050 spin_unlock_irq(&pcpu_lock);
2051 goto retry_pop;
2052 }
2053 }
2054
2055 mutex_unlock(&pcpu_alloc_mutex);
2056}
2057
2058/**
2059 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2060 * @work: unused
2061 *
2062 * Call __pcpu_balance_workfn() for each chunk type.
2063 */
2064static void pcpu_balance_workfn(struct work_struct *work)
2065{
2066 enum pcpu_chunk_type type;
2067
2068 for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++)
2069 __pcpu_balance_workfn(type);
2070}
2071
2072/**
2073 * free_percpu - free percpu area
2074 * @ptr: pointer to area to free
2075 *
2076 * Free percpu area @ptr.
2077 *
2078 * CONTEXT:
2079 * Can be called from atomic context.
2080 */
2081void free_percpu(void __percpu *ptr)
2082{
2083 void *addr;
2084 struct pcpu_chunk *chunk;
2085 unsigned long flags;
2086 int size, off;
2087 bool need_balance = false;
2088 struct list_head *pcpu_slot;
2089
2090 if (!ptr)
2091 return;
2092
2093 kmemleak_free_percpu(ptr);
2094
2095 addr = __pcpu_ptr_to_addr(ptr);
2096
2097 spin_lock_irqsave(&pcpu_lock, flags);
2098
2099 chunk = pcpu_chunk_addr_search(addr);
2100 off = addr - chunk->base_addr;
2101
2102 size = pcpu_free_area(chunk, off);
2103
2104 pcpu_slot = pcpu_chunk_list(pcpu_chunk_type(chunk));
2105
2106 pcpu_memcg_free_hook(chunk, off, size);
2107
2108 /* if there are more than one fully free chunks, wake up grim reaper */
2109 if (chunk->free_bytes == pcpu_unit_size) {
2110 struct pcpu_chunk *pos;
2111
2112 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
2113 if (pos != chunk) {
2114 need_balance = true;
2115 break;
2116 }
2117 }
2118
2119 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2120
2121 spin_unlock_irqrestore(&pcpu_lock, flags);
2122
2123 if (need_balance)
2124 pcpu_schedule_balance_work();
2125}
2126EXPORT_SYMBOL_GPL(free_percpu);
2127
2128bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2129{
2130#ifdef CONFIG_SMP
2131 const size_t static_size = __per_cpu_end - __per_cpu_start;
2132 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2133 unsigned int cpu;
2134
2135 for_each_possible_cpu(cpu) {
2136 void *start = per_cpu_ptr(base, cpu);
2137 void *va = (void *)addr;
2138
2139 if (va >= start && va < start + static_size) {
2140 if (can_addr) {
2141 *can_addr = (unsigned long) (va - start);
2142 *can_addr += (unsigned long)
2143 per_cpu_ptr(base, get_boot_cpu_id());
2144 }
2145 return true;
2146 }
2147 }
2148#endif
2149 /* on UP, can't distinguish from other static vars, always false */
2150 return false;
2151}
2152
2153/**
2154 * is_kernel_percpu_address - test whether address is from static percpu area
2155 * @addr: address to test
2156 *
2157 * Test whether @addr belongs to in-kernel static percpu area. Module
2158 * static percpu areas are not considered. For those, use
2159 * is_module_percpu_address().
2160 *
2161 * RETURNS:
2162 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2163 */
2164bool is_kernel_percpu_address(unsigned long addr)
2165{
2166 return __is_kernel_percpu_address(addr, NULL);
2167}
2168
2169/**
2170 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2171 * @addr: the address to be converted to physical address
2172 *
2173 * Given @addr which is dereferenceable address obtained via one of
2174 * percpu access macros, this function translates it into its physical
2175 * address. The caller is responsible for ensuring @addr stays valid
2176 * until this function finishes.
2177 *
2178 * percpu allocator has special setup for the first chunk, which currently
2179 * supports either embedding in linear address space or vmalloc mapping,
2180 * and, from the second one, the backing allocator (currently either vm or
2181 * km) provides translation.
2182 *
2183 * The addr can be translated simply without checking if it falls into the
2184 * first chunk. But the current code reflects better how percpu allocator
2185 * actually works, and the verification can discover both bugs in percpu
2186 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2187 * code.
2188 *
2189 * RETURNS:
2190 * The physical address for @addr.
2191 */
2192phys_addr_t per_cpu_ptr_to_phys(void *addr)
2193{
2194 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2195 bool in_first_chunk = false;
2196 unsigned long first_low, first_high;
2197 unsigned int cpu;
2198
2199 /*
2200 * The following test on unit_low/high isn't strictly
2201 * necessary but will speed up lookups of addresses which
2202 * aren't in the first chunk.
2203 *
2204 * The address check is against full chunk sizes. pcpu_base_addr
2205 * points to the beginning of the first chunk including the
2206 * static region. Assumes good intent as the first chunk may
2207 * not be full (ie. < pcpu_unit_pages in size).
2208 */
2209 first_low = (unsigned long)pcpu_base_addr +
2210 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2211 first_high = (unsigned long)pcpu_base_addr +
2212 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2213 if ((unsigned long)addr >= first_low &&
2214 (unsigned long)addr < first_high) {
2215 for_each_possible_cpu(cpu) {
2216 void *start = per_cpu_ptr(base, cpu);
2217
2218 if (addr >= start && addr < start + pcpu_unit_size) {
2219 in_first_chunk = true;
2220 break;
2221 }
2222 }
2223 }
2224
2225 if (in_first_chunk) {
2226 if (!is_vmalloc_addr(addr))
2227 return __pa(addr);
2228 else
2229 return page_to_phys(vmalloc_to_page(addr)) +
2230 offset_in_page(addr);
2231 } else
2232 return page_to_phys(pcpu_addr_to_page(addr)) +
2233 offset_in_page(addr);
2234}
2235
2236/**
2237 * pcpu_alloc_alloc_info - allocate percpu allocation info
2238 * @nr_groups: the number of groups
2239 * @nr_units: the number of units
2240 *
2241 * Allocate ai which is large enough for @nr_groups groups containing
2242 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2243 * cpu_map array which is long enough for @nr_units and filled with
2244 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2245 * pointer of other groups.
2246 *
2247 * RETURNS:
2248 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2249 * failure.
2250 */
2251struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2252 int nr_units)
2253{
2254 struct pcpu_alloc_info *ai;
2255 size_t base_size, ai_size;
2256 void *ptr;
2257 int unit;
2258
2259 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2260 __alignof__(ai->groups[0].cpu_map[0]));
2261 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2262
2263 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2264 if (!ptr)
2265 return NULL;
2266 ai = ptr;
2267 ptr += base_size;
2268
2269 ai->groups[0].cpu_map = ptr;
2270
2271 for (unit = 0; unit < nr_units; unit++)
2272 ai->groups[0].cpu_map[unit] = NR_CPUS;
2273
2274 ai->nr_groups = nr_groups;
2275 ai->__ai_size = PFN_ALIGN(ai_size);
2276
2277 return ai;
2278}
2279
2280/**
2281 * pcpu_free_alloc_info - free percpu allocation info
2282 * @ai: pcpu_alloc_info to free
2283 *
2284 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2285 */
2286void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2287{
2288 memblock_free_early(__pa(ai), ai->__ai_size);
2289}
2290
2291/**
2292 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2293 * @lvl: loglevel
2294 * @ai: allocation info to dump
2295 *
2296 * Print out information about @ai using loglevel @lvl.
2297 */
2298static void pcpu_dump_alloc_info(const char *lvl,
2299 const struct pcpu_alloc_info *ai)
2300{
2301 int group_width = 1, cpu_width = 1, width;
2302 char empty_str[] = "--------";
2303 int alloc = 0, alloc_end = 0;
2304 int group, v;
2305 int upa, apl; /* units per alloc, allocs per line */
2306
2307 v = ai->nr_groups;
2308 while (v /= 10)
2309 group_width++;
2310
2311 v = num_possible_cpus();
2312 while (v /= 10)
2313 cpu_width++;
2314 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2315
2316 upa = ai->alloc_size / ai->unit_size;
2317 width = upa * (cpu_width + 1) + group_width + 3;
2318 apl = rounddown_pow_of_two(max(60 / width, 1));
2319
2320 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2321 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2322 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2323
2324 for (group = 0; group < ai->nr_groups; group++) {
2325 const struct pcpu_group_info *gi = &ai->groups[group];
2326 int unit = 0, unit_end = 0;
2327
2328 BUG_ON(gi->nr_units % upa);
2329 for (alloc_end += gi->nr_units / upa;
2330 alloc < alloc_end; alloc++) {
2331 if (!(alloc % apl)) {
2332 pr_cont("\n");
2333 printk("%spcpu-alloc: ", lvl);
2334 }
2335 pr_cont("[%0*d] ", group_width, group);
2336
2337 for (unit_end += upa; unit < unit_end; unit++)
2338 if (gi->cpu_map[unit] != NR_CPUS)
2339 pr_cont("%0*d ",
2340 cpu_width, gi->cpu_map[unit]);
2341 else
2342 pr_cont("%s ", empty_str);
2343 }
2344 }
2345 pr_cont("\n");
2346}
2347
2348/**
2349 * pcpu_setup_first_chunk - initialize the first percpu chunk
2350 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2351 * @base_addr: mapped address
2352 *
2353 * Initialize the first percpu chunk which contains the kernel static
2354 * percpu area. This function is to be called from arch percpu area
2355 * setup path.
2356 *
2357 * @ai contains all information necessary to initialize the first
2358 * chunk and prime the dynamic percpu allocator.
2359 *
2360 * @ai->static_size is the size of static percpu area.
2361 *
2362 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2363 * reserve after the static area in the first chunk. This reserves
2364 * the first chunk such that it's available only through reserved
2365 * percpu allocation. This is primarily used to serve module percpu
2366 * static areas on architectures where the addressing model has
2367 * limited offset range for symbol relocations to guarantee module
2368 * percpu symbols fall inside the relocatable range.
2369 *
2370 * @ai->dyn_size determines the number of bytes available for dynamic
2371 * allocation in the first chunk. The area between @ai->static_size +
2372 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2373 *
2374 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2375 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2376 * @ai->dyn_size.
2377 *
2378 * @ai->atom_size is the allocation atom size and used as alignment
2379 * for vm areas.
2380 *
2381 * @ai->alloc_size is the allocation size and always multiple of
2382 * @ai->atom_size. This is larger than @ai->atom_size if
2383 * @ai->unit_size is larger than @ai->atom_size.
2384 *
2385 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2386 * percpu areas. Units which should be colocated are put into the
2387 * same group. Dynamic VM areas will be allocated according to these
2388 * groupings. If @ai->nr_groups is zero, a single group containing
2389 * all units is assumed.
2390 *
2391 * The caller should have mapped the first chunk at @base_addr and
2392 * copied static data to each unit.
2393 *
2394 * The first chunk will always contain a static and a dynamic region.
2395 * However, the static region is not managed by any chunk. If the first
2396 * chunk also contains a reserved region, it is served by two chunks -
2397 * one for the reserved region and one for the dynamic region. They
2398 * share the same vm, but use offset regions in the area allocation map.
2399 * The chunk serving the dynamic region is circulated in the chunk slots
2400 * and available for dynamic allocation like any other chunk.
2401 */
2402void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2403 void *base_addr)
2404{
2405 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2406 size_t static_size, dyn_size;
2407 struct pcpu_chunk *chunk;
2408 unsigned long *group_offsets;
2409 size_t *group_sizes;
2410 unsigned long *unit_off;
2411 unsigned int cpu;
2412 int *unit_map;
2413 int group, unit, i;
2414 int map_size;
2415 unsigned long tmp_addr;
2416 size_t alloc_size;
2417 enum pcpu_chunk_type type;
2418
2419#define PCPU_SETUP_BUG_ON(cond) do { \
2420 if (unlikely(cond)) { \
2421 pr_emerg("failed to initialize, %s\n", #cond); \
2422 pr_emerg("cpu_possible_mask=%*pb\n", \
2423 cpumask_pr_args(cpu_possible_mask)); \
2424 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2425 BUG(); \
2426 } \
2427} while (0)
2428
2429 /* sanity checks */
2430 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2431#ifdef CONFIG_SMP
2432 PCPU_SETUP_BUG_ON(!ai->static_size);
2433 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2434#endif
2435 PCPU_SETUP_BUG_ON(!base_addr);
2436 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2437 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2438 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2439 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2440 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2441 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2442 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2443 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2444 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2445 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2446 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2447
2448 /* process group information and build config tables accordingly */
2449 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2450 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2451 if (!group_offsets)
2452 panic("%s: Failed to allocate %zu bytes\n", __func__,
2453 alloc_size);
2454
2455 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2456 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2457 if (!group_sizes)
2458 panic("%s: Failed to allocate %zu bytes\n", __func__,
2459 alloc_size);
2460
2461 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2462 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2463 if (!unit_map)
2464 panic("%s: Failed to allocate %zu bytes\n", __func__,
2465 alloc_size);
2466
2467 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2468 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2469 if (!unit_off)
2470 panic("%s: Failed to allocate %zu bytes\n", __func__,
2471 alloc_size);
2472
2473 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2474 unit_map[cpu] = UINT_MAX;
2475
2476 pcpu_low_unit_cpu = NR_CPUS;
2477 pcpu_high_unit_cpu = NR_CPUS;
2478
2479 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2480 const struct pcpu_group_info *gi = &ai->groups[group];
2481
2482 group_offsets[group] = gi->base_offset;
2483 group_sizes[group] = gi->nr_units * ai->unit_size;
2484
2485 for (i = 0; i < gi->nr_units; i++) {
2486 cpu = gi->cpu_map[i];
2487 if (cpu == NR_CPUS)
2488 continue;
2489
2490 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2491 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2492 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2493
2494 unit_map[cpu] = unit + i;
2495 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2496
2497 /* determine low/high unit_cpu */
2498 if (pcpu_low_unit_cpu == NR_CPUS ||
2499 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2500 pcpu_low_unit_cpu = cpu;
2501 if (pcpu_high_unit_cpu == NR_CPUS ||
2502 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2503 pcpu_high_unit_cpu = cpu;
2504 }
2505 }
2506 pcpu_nr_units = unit;
2507
2508 for_each_possible_cpu(cpu)
2509 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2510
2511 /* we're done parsing the input, undefine BUG macro and dump config */
2512#undef PCPU_SETUP_BUG_ON
2513 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2514
2515 pcpu_nr_groups = ai->nr_groups;
2516 pcpu_group_offsets = group_offsets;
2517 pcpu_group_sizes = group_sizes;
2518 pcpu_unit_map = unit_map;
2519 pcpu_unit_offsets = unit_off;
2520
2521 /* determine basic parameters */
2522 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2523 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2524 pcpu_atom_size = ai->atom_size;
2525 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2526 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2527
2528 pcpu_stats_save_ai(ai);
2529
2530 /*
2531 * Allocate chunk slots. The additional last slot is for
2532 * empty chunks.
2533 */
2534 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2535 pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2536 sizeof(pcpu_chunk_lists[0]) *
2537 PCPU_NR_CHUNK_TYPES,
2538 SMP_CACHE_BYTES);
2539 if (!pcpu_chunk_lists)
2540 panic("%s: Failed to allocate %zu bytes\n", __func__,
2541 pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]) *
2542 PCPU_NR_CHUNK_TYPES);
2543
2544 for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++)
2545 for (i = 0; i < pcpu_nr_slots; i++)
2546 INIT_LIST_HEAD(&pcpu_chunk_list(type)[i]);
2547
2548 /*
2549 * The end of the static region needs to be aligned with the
2550 * minimum allocation size as this offsets the reserved and
2551 * dynamic region. The first chunk ends page aligned by
2552 * expanding the dynamic region, therefore the dynamic region
2553 * can be shrunk to compensate while still staying above the
2554 * configured sizes.
2555 */
2556 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2557 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2558
2559 /*
2560 * Initialize first chunk.
2561 * If the reserved_size is non-zero, this initializes the reserved
2562 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2563 * and the dynamic region is initialized here. The first chunk,
2564 * pcpu_first_chunk, will always point to the chunk that serves
2565 * the dynamic region.
2566 */
2567 tmp_addr = (unsigned long)base_addr + static_size;
2568 map_size = ai->reserved_size ?: dyn_size;
2569 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2570
2571 /* init dynamic chunk if necessary */
2572 if (ai->reserved_size) {
2573 pcpu_reserved_chunk = chunk;
2574
2575 tmp_addr = (unsigned long)base_addr + static_size +
2576 ai->reserved_size;
2577 map_size = dyn_size;
2578 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2579 }
2580
2581 /* link the first chunk in */
2582 pcpu_first_chunk = chunk;
2583 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2584 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2585
2586 /* include all regions of the first chunk */
2587 pcpu_nr_populated += PFN_DOWN(size_sum);
2588
2589 pcpu_stats_chunk_alloc();
2590 trace_percpu_create_chunk(base_addr);
2591
2592 /* we're done */
2593 pcpu_base_addr = base_addr;
2594}
2595
2596#ifdef CONFIG_SMP
2597
2598const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2599 [PCPU_FC_AUTO] = "auto",
2600 [PCPU_FC_EMBED] = "embed",
2601 [PCPU_FC_PAGE] = "page",
2602};
2603
2604enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2605
2606static int __init percpu_alloc_setup(char *str)
2607{
2608 if (!str)
2609 return -EINVAL;
2610
2611 if (0)
2612 /* nada */;
2613#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2614 else if (!strcmp(str, "embed"))
2615 pcpu_chosen_fc = PCPU_FC_EMBED;
2616#endif
2617#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2618 else if (!strcmp(str, "page"))
2619 pcpu_chosen_fc = PCPU_FC_PAGE;
2620#endif
2621 else
2622 pr_warn("unknown allocator %s specified\n", str);
2623
2624 return 0;
2625}
2626early_param("percpu_alloc", percpu_alloc_setup);
2627
2628/*
2629 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2630 * Build it if needed by the arch config or the generic setup is going
2631 * to be used.
2632 */
2633#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2634 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2635#define BUILD_EMBED_FIRST_CHUNK
2636#endif
2637
2638/* build pcpu_page_first_chunk() iff needed by the arch config */
2639#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2640#define BUILD_PAGE_FIRST_CHUNK
2641#endif
2642
2643/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2644#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2645/**
2646 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2647 * @reserved_size: the size of reserved percpu area in bytes
2648 * @dyn_size: minimum free size for dynamic allocation in bytes
2649 * @atom_size: allocation atom size
2650 * @cpu_distance_fn: callback to determine distance between cpus, optional
2651 *
2652 * This function determines grouping of units, their mappings to cpus
2653 * and other parameters considering needed percpu size, allocation
2654 * atom size and distances between CPUs.
2655 *
2656 * Groups are always multiples of atom size and CPUs which are of
2657 * LOCAL_DISTANCE both ways are grouped together and share space for
2658 * units in the same group. The returned configuration is guaranteed
2659 * to have CPUs on different nodes on different groups and >=75% usage
2660 * of allocated virtual address space.
2661 *
2662 * RETURNS:
2663 * On success, pointer to the new allocation_info is returned. On
2664 * failure, ERR_PTR value is returned.
2665 */
2666static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2667 size_t reserved_size, size_t dyn_size,
2668 size_t atom_size,
2669 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2670{
2671 static int group_map[NR_CPUS] __initdata;
2672 static int group_cnt[NR_CPUS] __initdata;
2673 const size_t static_size = __per_cpu_end - __per_cpu_start;
2674 int nr_groups = 1, nr_units = 0;
2675 size_t size_sum, min_unit_size, alloc_size;
2676 int upa, max_upa, best_upa; /* units_per_alloc */
2677 int last_allocs, group, unit;
2678 unsigned int cpu, tcpu;
2679 struct pcpu_alloc_info *ai;
2680 unsigned int *cpu_map;
2681
2682 /* this function may be called multiple times */
2683 memset(group_map, 0, sizeof(group_map));
2684 memset(group_cnt, 0, sizeof(group_cnt));
2685
2686 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2687 size_sum = PFN_ALIGN(static_size + reserved_size +
2688 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2689 dyn_size = size_sum - static_size - reserved_size;
2690
2691 /*
2692 * Determine min_unit_size, alloc_size and max_upa such that
2693 * alloc_size is multiple of atom_size and is the smallest
2694 * which can accommodate 4k aligned segments which are equal to
2695 * or larger than min_unit_size.
2696 */
2697 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2698
2699 /* determine the maximum # of units that can fit in an allocation */
2700 alloc_size = roundup(min_unit_size, atom_size);
2701 upa = alloc_size / min_unit_size;
2702 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2703 upa--;
2704 max_upa = upa;
2705
2706 /* group cpus according to their proximity */
2707 for_each_possible_cpu(cpu) {
2708 group = 0;
2709 next_group:
2710 for_each_possible_cpu(tcpu) {
2711 if (cpu == tcpu)
2712 break;
2713 if (group_map[tcpu] == group && cpu_distance_fn &&
2714 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2715 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2716 group++;
2717 nr_groups = max(nr_groups, group + 1);
2718 goto next_group;
2719 }
2720 }
2721 group_map[cpu] = group;
2722 group_cnt[group]++;
2723 }
2724
2725 /*
2726 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2727 * Expand the unit_size until we use >= 75% of the units allocated.
2728 * Related to atom_size, which could be much larger than the unit_size.
2729 */
2730 last_allocs = INT_MAX;
2731 for (upa = max_upa; upa; upa--) {
2732 int allocs = 0, wasted = 0;
2733
2734 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2735 continue;
2736
2737 for (group = 0; group < nr_groups; group++) {
2738 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2739 allocs += this_allocs;
2740 wasted += this_allocs * upa - group_cnt[group];
2741 }
2742
2743 /*
2744 * Don't accept if wastage is over 1/3. The
2745 * greater-than comparison ensures upa==1 always
2746 * passes the following check.
2747 */
2748 if (wasted > num_possible_cpus() / 3)
2749 continue;
2750
2751 /* and then don't consume more memory */
2752 if (allocs > last_allocs)
2753 break;
2754 last_allocs = allocs;
2755 best_upa = upa;
2756 }
2757 upa = best_upa;
2758
2759 /* allocate and fill alloc_info */
2760 for (group = 0; group < nr_groups; group++)
2761 nr_units += roundup(group_cnt[group], upa);
2762
2763 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2764 if (!ai)
2765 return ERR_PTR(-ENOMEM);
2766 cpu_map = ai->groups[0].cpu_map;
2767
2768 for (group = 0; group < nr_groups; group++) {
2769 ai->groups[group].cpu_map = cpu_map;
2770 cpu_map += roundup(group_cnt[group], upa);
2771 }
2772
2773 ai->static_size = static_size;
2774 ai->reserved_size = reserved_size;
2775 ai->dyn_size = dyn_size;
2776 ai->unit_size = alloc_size / upa;
2777 ai->atom_size = atom_size;
2778 ai->alloc_size = alloc_size;
2779
2780 for (group = 0, unit = 0; group < nr_groups; group++) {
2781 struct pcpu_group_info *gi = &ai->groups[group];
2782
2783 /*
2784 * Initialize base_offset as if all groups are located
2785 * back-to-back. The caller should update this to
2786 * reflect actual allocation.
2787 */
2788 gi->base_offset = unit * ai->unit_size;
2789
2790 for_each_possible_cpu(cpu)
2791 if (group_map[cpu] == group)
2792 gi->cpu_map[gi->nr_units++] = cpu;
2793 gi->nr_units = roundup(gi->nr_units, upa);
2794 unit += gi->nr_units;
2795 }
2796 BUG_ON(unit != nr_units);
2797
2798 return ai;
2799}
2800#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2801
2802#if defined(BUILD_EMBED_FIRST_CHUNK)
2803/**
2804 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2805 * @reserved_size: the size of reserved percpu area in bytes
2806 * @dyn_size: minimum free size for dynamic allocation in bytes
2807 * @atom_size: allocation atom size
2808 * @cpu_distance_fn: callback to determine distance between cpus, optional
2809 * @alloc_fn: function to allocate percpu page
2810 * @free_fn: function to free percpu page
2811 *
2812 * This is a helper to ease setting up embedded first percpu chunk and
2813 * can be called where pcpu_setup_first_chunk() is expected.
2814 *
2815 * If this function is used to setup the first chunk, it is allocated
2816 * by calling @alloc_fn and used as-is without being mapped into
2817 * vmalloc area. Allocations are always whole multiples of @atom_size
2818 * aligned to @atom_size.
2819 *
2820 * This enables the first chunk to piggy back on the linear physical
2821 * mapping which often uses larger page size. Please note that this
2822 * can result in very sparse cpu->unit mapping on NUMA machines thus
2823 * requiring large vmalloc address space. Don't use this allocator if
2824 * vmalloc space is not orders of magnitude larger than distances
2825 * between node memory addresses (ie. 32bit NUMA machines).
2826 *
2827 * @dyn_size specifies the minimum dynamic area size.
2828 *
2829 * If the needed size is smaller than the minimum or specified unit
2830 * size, the leftover is returned using @free_fn.
2831 *
2832 * RETURNS:
2833 * 0 on success, -errno on failure.
2834 */
2835int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2836 size_t atom_size,
2837 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2838 pcpu_fc_alloc_fn_t alloc_fn,
2839 pcpu_fc_free_fn_t free_fn)
2840{
2841 void *base = (void *)ULONG_MAX;
2842 void **areas = NULL;
2843 struct pcpu_alloc_info *ai;
2844 size_t size_sum, areas_size;
2845 unsigned long max_distance;
2846 int group, i, highest_group, rc = 0;
2847
2848 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2849 cpu_distance_fn);
2850 if (IS_ERR(ai))
2851 return PTR_ERR(ai);
2852
2853 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2854 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2855
2856 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
2857 if (!areas) {
2858 rc = -ENOMEM;
2859 goto out_free;
2860 }
2861
2862 /* allocate, copy and determine base address & max_distance */
2863 highest_group = 0;
2864 for (group = 0; group < ai->nr_groups; group++) {
2865 struct pcpu_group_info *gi = &ai->groups[group];
2866 unsigned int cpu = NR_CPUS;
2867 void *ptr;
2868
2869 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2870 cpu = gi->cpu_map[i];
2871 BUG_ON(cpu == NR_CPUS);
2872
2873 /* allocate space for the whole group */
2874 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2875 if (!ptr) {
2876 rc = -ENOMEM;
2877 goto out_free_areas;
2878 }
2879 /* kmemleak tracks the percpu allocations separately */
2880 kmemleak_free(ptr);
2881 areas[group] = ptr;
2882
2883 base = min(ptr, base);
2884 if (ptr > areas[highest_group])
2885 highest_group = group;
2886 }
2887 max_distance = areas[highest_group] - base;
2888 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2889
2890 /* warn if maximum distance is further than 75% of vmalloc space */
2891 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2892 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2893 max_distance, VMALLOC_TOTAL);
2894#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2895 /* and fail if we have fallback */
2896 rc = -EINVAL;
2897 goto out_free_areas;
2898#endif
2899 }
2900
2901 /*
2902 * Copy data and free unused parts. This should happen after all
2903 * allocations are complete; otherwise, we may end up with
2904 * overlapping groups.
2905 */
2906 for (group = 0; group < ai->nr_groups; group++) {
2907 struct pcpu_group_info *gi = &ai->groups[group];
2908 void *ptr = areas[group];
2909
2910 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2911 if (gi->cpu_map[i] == NR_CPUS) {
2912 /* unused unit, free whole */
2913 free_fn(ptr, ai->unit_size);
2914 continue;
2915 }
2916 /* copy and return the unused part */
2917 memcpy(ptr, __per_cpu_load, ai->static_size);
2918 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2919 }
2920 }
2921
2922 /* base address is now known, determine group base offsets */
2923 for (group = 0; group < ai->nr_groups; group++) {
2924 ai->groups[group].base_offset = areas[group] - base;
2925 }
2926
2927 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2928 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
2929 ai->dyn_size, ai->unit_size);
2930
2931 pcpu_setup_first_chunk(ai, base);
2932 goto out_free;
2933
2934out_free_areas:
2935 for (group = 0; group < ai->nr_groups; group++)
2936 if (areas[group])
2937 free_fn(areas[group],
2938 ai->groups[group].nr_units * ai->unit_size);
2939out_free:
2940 pcpu_free_alloc_info(ai);
2941 if (areas)
2942 memblock_free_early(__pa(areas), areas_size);
2943 return rc;
2944}
2945#endif /* BUILD_EMBED_FIRST_CHUNK */
2946
2947#ifdef BUILD_PAGE_FIRST_CHUNK
2948/**
2949 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2950 * @reserved_size: the size of reserved percpu area in bytes
2951 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2952 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2953 * @populate_pte_fn: function to populate pte
2954 *
2955 * This is a helper to ease setting up page-remapped first percpu
2956 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2957 *
2958 * This is the basic allocator. Static percpu area is allocated
2959 * page-by-page into vmalloc area.
2960 *
2961 * RETURNS:
2962 * 0 on success, -errno on failure.
2963 */
2964int __init pcpu_page_first_chunk(size_t reserved_size,
2965 pcpu_fc_alloc_fn_t alloc_fn,
2966 pcpu_fc_free_fn_t free_fn,
2967 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2968{
2969 static struct vm_struct vm;
2970 struct pcpu_alloc_info *ai;
2971 char psize_str[16];
2972 int unit_pages;
2973 size_t pages_size;
2974 struct page **pages;
2975 int unit, i, j, rc = 0;
2976 int upa;
2977 int nr_g0_units;
2978
2979 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2980
2981 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2982 if (IS_ERR(ai))
2983 return PTR_ERR(ai);
2984 BUG_ON(ai->nr_groups != 1);
2985 upa = ai->alloc_size/ai->unit_size;
2986 nr_g0_units = roundup(num_possible_cpus(), upa);
2987 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
2988 pcpu_free_alloc_info(ai);
2989 return -EINVAL;
2990 }
2991
2992 unit_pages = ai->unit_size >> PAGE_SHIFT;
2993
2994 /* unaligned allocations can't be freed, round up to page size */
2995 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2996 sizeof(pages[0]));
2997 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
2998 if (!pages)
2999 panic("%s: Failed to allocate %zu bytes\n", __func__,
3000 pages_size);
3001
3002 /* allocate pages */
3003 j = 0;
3004 for (unit = 0; unit < num_possible_cpus(); unit++) {
3005 unsigned int cpu = ai->groups[0].cpu_map[unit];
3006 for (i = 0; i < unit_pages; i++) {
3007 void *ptr;
3008
3009 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
3010 if (!ptr) {
3011 pr_warn("failed to allocate %s page for cpu%u\n",
3012 psize_str, cpu);
3013 goto enomem;
3014 }
3015 /* kmemleak tracks the percpu allocations separately */
3016 kmemleak_free(ptr);
3017 pages[j++] = virt_to_page(ptr);
3018 }
3019 }
3020
3021 /* allocate vm area, map the pages and copy static data */
3022 vm.flags = VM_ALLOC;
3023 vm.size = num_possible_cpus() * ai->unit_size;
3024 vm_area_register_early(&vm, PAGE_SIZE);
3025
3026 for (unit = 0; unit < num_possible_cpus(); unit++) {
3027 unsigned long unit_addr =
3028 (unsigned long)vm.addr + unit * ai->unit_size;
3029
3030 for (i = 0; i < unit_pages; i++)
3031 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
3032
3033 /* pte already populated, the following shouldn't fail */
3034 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3035 unit_pages);
3036 if (rc < 0)
3037 panic("failed to map percpu area, err=%d\n", rc);
3038
3039 /*
3040 * FIXME: Archs with virtual cache should flush local
3041 * cache for the linear mapping here - something
3042 * equivalent to flush_cache_vmap() on the local cpu.
3043 * flush_cache_vmap() can't be used as most supporting
3044 * data structures are not set up yet.
3045 */
3046
3047 /* copy static data */
3048 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3049 }
3050
3051 /* we're ready, commit */
3052 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3053 unit_pages, psize_str, ai->static_size,
3054 ai->reserved_size, ai->dyn_size);
3055
3056 pcpu_setup_first_chunk(ai, vm.addr);
3057 goto out_free_ar;
3058
3059enomem:
3060 while (--j >= 0)
3061 free_fn(page_address(pages[j]), PAGE_SIZE);
3062 rc = -ENOMEM;
3063out_free_ar:
3064 memblock_free_early(__pa(pages), pages_size);
3065 pcpu_free_alloc_info(ai);
3066 return rc;
3067}
3068#endif /* BUILD_PAGE_FIRST_CHUNK */
3069
3070#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3071/*
3072 * Generic SMP percpu area setup.
3073 *
3074 * The embedding helper is used because its behavior closely resembles
3075 * the original non-dynamic generic percpu area setup. This is
3076 * important because many archs have addressing restrictions and might
3077 * fail if the percpu area is located far away from the previous
3078 * location. As an added bonus, in non-NUMA cases, embedding is
3079 * generally a good idea TLB-wise because percpu area can piggy back
3080 * on the physical linear memory mapping which uses large page
3081 * mappings on applicable archs.
3082 */
3083unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3084EXPORT_SYMBOL(__per_cpu_offset);
3085
3086static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
3087 size_t align)
3088{
3089 return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS));
3090}
3091
3092static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
3093{
3094 memblock_free_early(__pa(ptr), size);
3095}
3096
3097void __init setup_per_cpu_areas(void)
3098{
3099 unsigned long delta;
3100 unsigned int cpu;
3101 int rc;
3102
3103 /*
3104 * Always reserve area for module percpu variables. That's
3105 * what the legacy allocator did.
3106 */
3107 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
3108 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
3109 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
3110 if (rc < 0)
3111 panic("Failed to initialize percpu areas.");
3112
3113 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3114 for_each_possible_cpu(cpu)
3115 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3116}
3117#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3118
3119#else /* CONFIG_SMP */
3120
3121/*
3122 * UP percpu area setup.
3123 *
3124 * UP always uses km-based percpu allocator with identity mapping.
3125 * Static percpu variables are indistinguishable from the usual static
3126 * variables and don't require any special preparation.
3127 */
3128void __init setup_per_cpu_areas(void)
3129{
3130 const size_t unit_size =
3131 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3132 PERCPU_DYNAMIC_RESERVE));
3133 struct pcpu_alloc_info *ai;
3134 void *fc;
3135
3136 ai = pcpu_alloc_alloc_info(1, 1);
3137 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3138 if (!ai || !fc)
3139 panic("Failed to allocate memory for percpu areas.");
3140 /* kmemleak tracks the percpu allocations separately */
3141 kmemleak_free(fc);
3142
3143 ai->dyn_size = unit_size;
3144 ai->unit_size = unit_size;
3145 ai->atom_size = unit_size;
3146 ai->alloc_size = unit_size;
3147 ai->groups[0].nr_units = 1;
3148 ai->groups[0].cpu_map[0] = 0;
3149
3150 pcpu_setup_first_chunk(ai, fc);
3151 pcpu_free_alloc_info(ai);
3152}
3153
3154#endif /* CONFIG_SMP */
3155
3156/*
3157 * pcpu_nr_pages - calculate total number of populated backing pages
3158 *
3159 * This reflects the number of pages populated to back chunks. Metadata is
3160 * excluded in the number exposed in meminfo as the number of backing pages
3161 * scales with the number of cpus and can quickly outweigh the memory used for
3162 * metadata. It also keeps this calculation nice and simple.
3163 *
3164 * RETURNS:
3165 * Total number of populated backing pages in use by the allocator.
3166 */
3167unsigned long pcpu_nr_pages(void)
3168{
3169 return pcpu_nr_populated * pcpu_nr_units;
3170}
3171
3172/*
3173 * Percpu allocator is initialized early during boot when neither slab or
3174 * workqueue is available. Plug async management until everything is up
3175 * and running.
3176 */
3177static int __init percpu_enable_async(void)
3178{
3179 pcpu_async_enabled = true;
3180 return 0;
3181}
3182subsys_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#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
57
58#include <linux/bitmap.h>
59#include <linux/bootmem.h>
60#include <linux/err.h>
61#include <linux/list.h>
62#include <linux/log2.h>
63#include <linux/mm.h>
64#include <linux/module.h>
65#include <linux/mutex.h>
66#include <linux/percpu.h>
67#include <linux/pfn.h>
68#include <linux/slab.h>
69#include <linux/spinlock.h>
70#include <linux/vmalloc.h>
71#include <linux/workqueue.h>
72#include <linux/kmemleak.h>
73
74#include <asm/cacheflush.h>
75#include <asm/sections.h>
76#include <asm/tlbflush.h>
77#include <asm/io.h>
78
79#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
80#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
81#define PCPU_ATOMIC_MAP_MARGIN_LOW 32
82#define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
83#define PCPU_EMPTY_POP_PAGES_LOW 2
84#define PCPU_EMPTY_POP_PAGES_HIGH 4
85
86#ifdef CONFIG_SMP
87/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
88#ifndef __addr_to_pcpu_ptr
89#define __addr_to_pcpu_ptr(addr) \
90 (void __percpu *)((unsigned long)(addr) - \
91 (unsigned long)pcpu_base_addr + \
92 (unsigned long)__per_cpu_start)
93#endif
94#ifndef __pcpu_ptr_to_addr
95#define __pcpu_ptr_to_addr(ptr) \
96 (void __force *)((unsigned long)(ptr) + \
97 (unsigned long)pcpu_base_addr - \
98 (unsigned long)__per_cpu_start)
99#endif
100#else /* CONFIG_SMP */
101/* on UP, it's always identity mapped */
102#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
103#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
104#endif /* CONFIG_SMP */
105
106struct pcpu_chunk {
107 struct list_head list; /* linked to pcpu_slot lists */
108 int free_size; /* free bytes in the chunk */
109 int contig_hint; /* max contiguous size hint */
110 void *base_addr; /* base address of this chunk */
111
112 int map_used; /* # of map entries used before the sentry */
113 int map_alloc; /* # of map entries allocated */
114 int *map; /* allocation map */
115 struct work_struct map_extend_work;/* async ->map[] extension */
116
117 void *data; /* chunk data */
118 int first_free; /* no free below this */
119 bool immutable; /* no [de]population allowed */
120 int nr_populated; /* # of populated pages */
121 unsigned long populated[]; /* populated bitmap */
122};
123
124static int pcpu_unit_pages __read_mostly;
125static int pcpu_unit_size __read_mostly;
126static int pcpu_nr_units __read_mostly;
127static int pcpu_atom_size __read_mostly;
128static int pcpu_nr_slots __read_mostly;
129static size_t pcpu_chunk_struct_size __read_mostly;
130
131/* cpus with the lowest and highest unit addresses */
132static unsigned int pcpu_low_unit_cpu __read_mostly;
133static unsigned int pcpu_high_unit_cpu __read_mostly;
134
135/* the address of the first chunk which starts with the kernel static area */
136void *pcpu_base_addr __read_mostly;
137EXPORT_SYMBOL_GPL(pcpu_base_addr);
138
139static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
140const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
141
142/* group information, used for vm allocation */
143static int pcpu_nr_groups __read_mostly;
144static const unsigned long *pcpu_group_offsets __read_mostly;
145static const size_t *pcpu_group_sizes __read_mostly;
146
147/*
148 * The first chunk which always exists. Note that unlike other
149 * chunks, this one can be allocated and mapped in several different
150 * ways and thus often doesn't live in the vmalloc area.
151 */
152static struct pcpu_chunk *pcpu_first_chunk;
153
154/*
155 * Optional reserved chunk. This chunk reserves part of the first
156 * chunk and serves it for reserved allocations. The amount of
157 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
158 * area doesn't exist, the following variables contain NULL and 0
159 * respectively.
160 */
161static struct pcpu_chunk *pcpu_reserved_chunk;
162static int pcpu_reserved_chunk_limit;
163
164static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
165static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop */
166
167static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
168
169/*
170 * The number of empty populated pages, protected by pcpu_lock. The
171 * reserved chunk doesn't contribute to the count.
172 */
173static int pcpu_nr_empty_pop_pages;
174
175/*
176 * Balance work is used to populate or destroy chunks asynchronously. We
177 * try to keep the number of populated free pages between
178 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
179 * empty chunk.
180 */
181static void pcpu_balance_workfn(struct work_struct *work);
182static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
183static bool pcpu_async_enabled __read_mostly;
184static bool pcpu_atomic_alloc_failed;
185
186static void pcpu_schedule_balance_work(void)
187{
188 if (pcpu_async_enabled)
189 schedule_work(&pcpu_balance_work);
190}
191
192static bool pcpu_addr_in_first_chunk(void *addr)
193{
194 void *first_start = pcpu_first_chunk->base_addr;
195
196 return addr >= first_start && addr < first_start + pcpu_unit_size;
197}
198
199static bool pcpu_addr_in_reserved_chunk(void *addr)
200{
201 void *first_start = pcpu_first_chunk->base_addr;
202
203 return addr >= first_start &&
204 addr < first_start + pcpu_reserved_chunk_limit;
205}
206
207static int __pcpu_size_to_slot(int size)
208{
209 int highbit = fls(size); /* size is in bytes */
210 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
211}
212
213static int pcpu_size_to_slot(int size)
214{
215 if (size == pcpu_unit_size)
216 return pcpu_nr_slots - 1;
217 return __pcpu_size_to_slot(size);
218}
219
220static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
221{
222 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
223 return 0;
224
225 return pcpu_size_to_slot(chunk->free_size);
226}
227
228/* set the pointer to a chunk in a page struct */
229static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
230{
231 page->index = (unsigned long)pcpu;
232}
233
234/* obtain pointer to a chunk from a page struct */
235static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
236{
237 return (struct pcpu_chunk *)page->index;
238}
239
240static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
241{
242 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
243}
244
245static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
246 unsigned int cpu, int page_idx)
247{
248 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
249 (page_idx << PAGE_SHIFT);
250}
251
252static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
253 int *rs, int *re, int end)
254{
255 *rs = find_next_zero_bit(chunk->populated, end, *rs);
256 *re = find_next_bit(chunk->populated, end, *rs + 1);
257}
258
259static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
260 int *rs, int *re, int end)
261{
262 *rs = find_next_bit(chunk->populated, end, *rs);
263 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
264}
265
266/*
267 * (Un)populated page region iterators. Iterate over (un)populated
268 * page regions between @start and @end in @chunk. @rs and @re should
269 * be integer variables and will be set to start and end page index of
270 * the current region.
271 */
272#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
274 (rs) < (re); \
275 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
276
277#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
278 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
279 (rs) < (re); \
280 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
281
282/**
283 * pcpu_mem_zalloc - allocate memory
284 * @size: bytes to allocate
285 *
286 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
287 * kzalloc() is used; otherwise, vzalloc() is used. The returned
288 * memory is always zeroed.
289 *
290 * CONTEXT:
291 * Does GFP_KERNEL allocation.
292 *
293 * RETURNS:
294 * Pointer to the allocated area on success, NULL on failure.
295 */
296static void *pcpu_mem_zalloc(size_t size)
297{
298 if (WARN_ON_ONCE(!slab_is_available()))
299 return NULL;
300
301 if (size <= PAGE_SIZE)
302 return kzalloc(size, GFP_KERNEL);
303 else
304 return vzalloc(size);
305}
306
307/**
308 * pcpu_mem_free - free memory
309 * @ptr: memory to free
310 *
311 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
312 */
313static void pcpu_mem_free(void *ptr)
314{
315 kvfree(ptr);
316}
317
318/**
319 * pcpu_count_occupied_pages - count the number of pages an area occupies
320 * @chunk: chunk of interest
321 * @i: index of the area in question
322 *
323 * Count the number of pages chunk's @i'th area occupies. When the area's
324 * start and/or end address isn't aligned to page boundary, the straddled
325 * page is included in the count iff the rest of the page is free.
326 */
327static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
328{
329 int off = chunk->map[i] & ~1;
330 int end = chunk->map[i + 1] & ~1;
331
332 if (!PAGE_ALIGNED(off) && i > 0) {
333 int prev = chunk->map[i - 1];
334
335 if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
336 off = round_down(off, PAGE_SIZE);
337 }
338
339 if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
340 int next = chunk->map[i + 1];
341 int nend = chunk->map[i + 2] & ~1;
342
343 if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
344 end = round_up(end, PAGE_SIZE);
345 }
346
347 return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
348}
349
350/**
351 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
352 * @chunk: chunk of interest
353 * @oslot: the previous slot it was on
354 *
355 * This function is called after an allocation or free changed @chunk.
356 * New slot according to the changed state is determined and @chunk is
357 * moved to the slot. Note that the reserved chunk is never put on
358 * chunk slots.
359 *
360 * CONTEXT:
361 * pcpu_lock.
362 */
363static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
364{
365 int nslot = pcpu_chunk_slot(chunk);
366
367 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
368 if (oslot < nslot)
369 list_move(&chunk->list, &pcpu_slot[nslot]);
370 else
371 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
372 }
373}
374
375/**
376 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
377 * @chunk: chunk of interest
378 * @is_atomic: the allocation context
379 *
380 * Determine whether area map of @chunk needs to be extended. If
381 * @is_atomic, only the amount necessary for a new allocation is
382 * considered; however, async extension is scheduled if the left amount is
383 * low. If !@is_atomic, it aims for more empty space. Combined, this
384 * ensures that the map is likely to have enough available space to
385 * accomodate atomic allocations which can't extend maps directly.
386 *
387 * CONTEXT:
388 * pcpu_lock.
389 *
390 * RETURNS:
391 * New target map allocation length if extension is necessary, 0
392 * otherwise.
393 */
394static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
395{
396 int margin, new_alloc;
397
398 if (is_atomic) {
399 margin = 3;
400
401 if (chunk->map_alloc <
402 chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW &&
403 pcpu_async_enabled)
404 schedule_work(&chunk->map_extend_work);
405 } else {
406 margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
407 }
408
409 if (chunk->map_alloc >= chunk->map_used + margin)
410 return 0;
411
412 new_alloc = PCPU_DFL_MAP_ALLOC;
413 while (new_alloc < chunk->map_used + margin)
414 new_alloc *= 2;
415
416 return new_alloc;
417}
418
419/**
420 * pcpu_extend_area_map - extend area map of a chunk
421 * @chunk: chunk of interest
422 * @new_alloc: new target allocation length of the area map
423 *
424 * Extend area map of @chunk to have @new_alloc entries.
425 *
426 * CONTEXT:
427 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
428 *
429 * RETURNS:
430 * 0 on success, -errno on failure.
431 */
432static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
433{
434 int *old = NULL, *new = NULL;
435 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
436 unsigned long flags;
437
438 new = pcpu_mem_zalloc(new_size);
439 if (!new)
440 return -ENOMEM;
441
442 /* acquire pcpu_lock and switch to new area map */
443 spin_lock_irqsave(&pcpu_lock, flags);
444
445 if (new_alloc <= chunk->map_alloc)
446 goto out_unlock;
447
448 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
449 old = chunk->map;
450
451 memcpy(new, old, old_size);
452
453 chunk->map_alloc = new_alloc;
454 chunk->map = new;
455 new = NULL;
456
457out_unlock:
458 spin_unlock_irqrestore(&pcpu_lock, flags);
459
460 /*
461 * pcpu_mem_free() might end up calling vfree() which uses
462 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
463 */
464 pcpu_mem_free(old);
465 pcpu_mem_free(new);
466
467 return 0;
468}
469
470static void pcpu_map_extend_workfn(struct work_struct *work)
471{
472 struct pcpu_chunk *chunk = container_of(work, struct pcpu_chunk,
473 map_extend_work);
474 int new_alloc;
475
476 spin_lock_irq(&pcpu_lock);
477 new_alloc = pcpu_need_to_extend(chunk, false);
478 spin_unlock_irq(&pcpu_lock);
479
480 if (new_alloc)
481 pcpu_extend_area_map(chunk, new_alloc);
482}
483
484/**
485 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
486 * @chunk: chunk the candidate area belongs to
487 * @off: the offset to the start of the candidate area
488 * @this_size: the size of the candidate area
489 * @size: the size of the target allocation
490 * @align: the alignment of the target allocation
491 * @pop_only: only allocate from already populated region
492 *
493 * We're trying to allocate @size bytes aligned at @align. @chunk's area
494 * at @off sized @this_size is a candidate. This function determines
495 * whether the target allocation fits in the candidate area and returns the
496 * number of bytes to pad after @off. If the target area doesn't fit, -1
497 * is returned.
498 *
499 * If @pop_only is %true, this function only considers the already
500 * populated part of the candidate area.
501 */
502static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
503 int size, int align, bool pop_only)
504{
505 int cand_off = off;
506
507 while (true) {
508 int head = ALIGN(cand_off, align) - off;
509 int page_start, page_end, rs, re;
510
511 if (this_size < head + size)
512 return -1;
513
514 if (!pop_only)
515 return head;
516
517 /*
518 * If the first unpopulated page is beyond the end of the
519 * allocation, the whole allocation is populated;
520 * otherwise, retry from the end of the unpopulated area.
521 */
522 page_start = PFN_DOWN(head + off);
523 page_end = PFN_UP(head + off + size);
524
525 rs = page_start;
526 pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
527 if (rs >= page_end)
528 return head;
529 cand_off = re * PAGE_SIZE;
530 }
531}
532
533/**
534 * pcpu_alloc_area - allocate area from a pcpu_chunk
535 * @chunk: chunk of interest
536 * @size: wanted size in bytes
537 * @align: wanted align
538 * @pop_only: allocate only from the populated area
539 * @occ_pages_p: out param for the number of pages the area occupies
540 *
541 * Try to allocate @size bytes area aligned at @align from @chunk.
542 * Note that this function only allocates the offset. It doesn't
543 * populate or map the area.
544 *
545 * @chunk->map must have at least two free slots.
546 *
547 * CONTEXT:
548 * pcpu_lock.
549 *
550 * RETURNS:
551 * Allocated offset in @chunk on success, -1 if no matching area is
552 * found.
553 */
554static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
555 bool pop_only, int *occ_pages_p)
556{
557 int oslot = pcpu_chunk_slot(chunk);
558 int max_contig = 0;
559 int i, off;
560 bool seen_free = false;
561 int *p;
562
563 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
564 int head, tail;
565 int this_size;
566
567 off = *p;
568 if (off & 1)
569 continue;
570
571 this_size = (p[1] & ~1) - off;
572
573 head = pcpu_fit_in_area(chunk, off, this_size, size, align,
574 pop_only);
575 if (head < 0) {
576 if (!seen_free) {
577 chunk->first_free = i;
578 seen_free = true;
579 }
580 max_contig = max(this_size, max_contig);
581 continue;
582 }
583
584 /*
585 * If head is small or the previous block is free,
586 * merge'em. Note that 'small' is defined as smaller
587 * than sizeof(int), which is very small but isn't too
588 * uncommon for percpu allocations.
589 */
590 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
591 *p = off += head;
592 if (p[-1] & 1)
593 chunk->free_size -= head;
594 else
595 max_contig = max(*p - p[-1], max_contig);
596 this_size -= head;
597 head = 0;
598 }
599
600 /* if tail is small, just keep it around */
601 tail = this_size - head - size;
602 if (tail < sizeof(int)) {
603 tail = 0;
604 size = this_size - head;
605 }
606
607 /* split if warranted */
608 if (head || tail) {
609 int nr_extra = !!head + !!tail;
610
611 /* insert new subblocks */
612 memmove(p + nr_extra + 1, p + 1,
613 sizeof(chunk->map[0]) * (chunk->map_used - i));
614 chunk->map_used += nr_extra;
615
616 if (head) {
617 if (!seen_free) {
618 chunk->first_free = i;
619 seen_free = true;
620 }
621 *++p = off += head;
622 ++i;
623 max_contig = max(head, max_contig);
624 }
625 if (tail) {
626 p[1] = off + size;
627 max_contig = max(tail, max_contig);
628 }
629 }
630
631 if (!seen_free)
632 chunk->first_free = i + 1;
633
634 /* update hint and mark allocated */
635 if (i + 1 == chunk->map_used)
636 chunk->contig_hint = max_contig; /* fully scanned */
637 else
638 chunk->contig_hint = max(chunk->contig_hint,
639 max_contig);
640
641 chunk->free_size -= size;
642 *p |= 1;
643
644 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
645 pcpu_chunk_relocate(chunk, oslot);
646 return off;
647 }
648
649 chunk->contig_hint = max_contig; /* fully scanned */
650 pcpu_chunk_relocate(chunk, oslot);
651
652 /* tell the upper layer that this chunk has no matching area */
653 return -1;
654}
655
656/**
657 * pcpu_free_area - free area to a pcpu_chunk
658 * @chunk: chunk of interest
659 * @freeme: offset of area to free
660 * @occ_pages_p: out param for the number of pages the area occupies
661 *
662 * Free area starting from @freeme to @chunk. Note that this function
663 * only modifies the allocation map. It doesn't depopulate or unmap
664 * the area.
665 *
666 * CONTEXT:
667 * pcpu_lock.
668 */
669static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
670 int *occ_pages_p)
671{
672 int oslot = pcpu_chunk_slot(chunk);
673 int off = 0;
674 unsigned i, j;
675 int to_free = 0;
676 int *p;
677
678 freeme |= 1; /* we are searching for <given offset, in use> pair */
679
680 i = 0;
681 j = chunk->map_used;
682 while (i != j) {
683 unsigned k = (i + j) / 2;
684 off = chunk->map[k];
685 if (off < freeme)
686 i = k + 1;
687 else if (off > freeme)
688 j = k;
689 else
690 i = j = k;
691 }
692 BUG_ON(off != freeme);
693
694 if (i < chunk->first_free)
695 chunk->first_free = i;
696
697 p = chunk->map + i;
698 *p = off &= ~1;
699 chunk->free_size += (p[1] & ~1) - off;
700
701 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
702
703 /* merge with next? */
704 if (!(p[1] & 1))
705 to_free++;
706 /* merge with previous? */
707 if (i > 0 && !(p[-1] & 1)) {
708 to_free++;
709 i--;
710 p--;
711 }
712 if (to_free) {
713 chunk->map_used -= to_free;
714 memmove(p + 1, p + 1 + to_free,
715 (chunk->map_used - i) * sizeof(chunk->map[0]));
716 }
717
718 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
719 pcpu_chunk_relocate(chunk, oslot);
720}
721
722static struct pcpu_chunk *pcpu_alloc_chunk(void)
723{
724 struct pcpu_chunk *chunk;
725
726 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
727 if (!chunk)
728 return NULL;
729
730 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
731 sizeof(chunk->map[0]));
732 if (!chunk->map) {
733 pcpu_mem_free(chunk);
734 return NULL;
735 }
736
737 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
738 chunk->map[0] = 0;
739 chunk->map[1] = pcpu_unit_size | 1;
740 chunk->map_used = 1;
741
742 INIT_LIST_HEAD(&chunk->list);
743 INIT_WORK(&chunk->map_extend_work, pcpu_map_extend_workfn);
744 chunk->free_size = pcpu_unit_size;
745 chunk->contig_hint = pcpu_unit_size;
746
747 return chunk;
748}
749
750static void pcpu_free_chunk(struct pcpu_chunk *chunk)
751{
752 if (!chunk)
753 return;
754 pcpu_mem_free(chunk->map);
755 pcpu_mem_free(chunk);
756}
757
758/**
759 * pcpu_chunk_populated - post-population bookkeeping
760 * @chunk: pcpu_chunk which got populated
761 * @page_start: the start page
762 * @page_end: the end page
763 *
764 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
765 * the bookkeeping information accordingly. Must be called after each
766 * successful population.
767 */
768static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
769 int page_start, int page_end)
770{
771 int nr = page_end - page_start;
772
773 lockdep_assert_held(&pcpu_lock);
774
775 bitmap_set(chunk->populated, page_start, nr);
776 chunk->nr_populated += nr;
777 pcpu_nr_empty_pop_pages += nr;
778}
779
780/**
781 * pcpu_chunk_depopulated - post-depopulation bookkeeping
782 * @chunk: pcpu_chunk which got depopulated
783 * @page_start: the start page
784 * @page_end: the end page
785 *
786 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
787 * Update the bookkeeping information accordingly. Must be called after
788 * each successful depopulation.
789 */
790static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
791 int page_start, int page_end)
792{
793 int nr = page_end - page_start;
794
795 lockdep_assert_held(&pcpu_lock);
796
797 bitmap_clear(chunk->populated, page_start, nr);
798 chunk->nr_populated -= nr;
799 pcpu_nr_empty_pop_pages -= nr;
800}
801
802/*
803 * Chunk management implementation.
804 *
805 * To allow different implementations, chunk alloc/free and
806 * [de]population are implemented in a separate file which is pulled
807 * into this file and compiled together. The following functions
808 * should be implemented.
809 *
810 * pcpu_populate_chunk - populate the specified range of a chunk
811 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
812 * pcpu_create_chunk - create a new chunk
813 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
814 * pcpu_addr_to_page - translate address to physical address
815 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
816 */
817static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
818static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
819static struct pcpu_chunk *pcpu_create_chunk(void);
820static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
821static struct page *pcpu_addr_to_page(void *addr);
822static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
823
824#ifdef CONFIG_NEED_PER_CPU_KM
825#include "percpu-km.c"
826#else
827#include "percpu-vm.c"
828#endif
829
830/**
831 * pcpu_chunk_addr_search - determine chunk containing specified address
832 * @addr: address for which the chunk needs to be determined.
833 *
834 * RETURNS:
835 * The address of the found chunk.
836 */
837static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
838{
839 /* is it in the first chunk? */
840 if (pcpu_addr_in_first_chunk(addr)) {
841 /* is it in the reserved area? */
842 if (pcpu_addr_in_reserved_chunk(addr))
843 return pcpu_reserved_chunk;
844 return pcpu_first_chunk;
845 }
846
847 /*
848 * The address is relative to unit0 which might be unused and
849 * thus unmapped. Offset the address to the unit space of the
850 * current processor before looking it up in the vmalloc
851 * space. Note that any possible cpu id can be used here, so
852 * there's no need to worry about preemption or cpu hotplug.
853 */
854 addr += pcpu_unit_offsets[raw_smp_processor_id()];
855 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
856}
857
858/**
859 * pcpu_alloc - the percpu allocator
860 * @size: size of area to allocate in bytes
861 * @align: alignment of area (max PAGE_SIZE)
862 * @reserved: allocate from the reserved chunk if available
863 * @gfp: allocation flags
864 *
865 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
866 * contain %GFP_KERNEL, the allocation is atomic.
867 *
868 * RETURNS:
869 * Percpu pointer to the allocated area on success, NULL on failure.
870 */
871static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
872 gfp_t gfp)
873{
874 static int warn_limit = 10;
875 struct pcpu_chunk *chunk;
876 const char *err;
877 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
878 int occ_pages = 0;
879 int slot, off, new_alloc, cpu, ret;
880 unsigned long flags;
881 void __percpu *ptr;
882
883 /*
884 * We want the lowest bit of offset available for in-use/free
885 * indicator, so force >= 16bit alignment and make size even.
886 */
887 if (unlikely(align < 2))
888 align = 2;
889
890 size = ALIGN(size, 2);
891
892 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
893 WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
894 size, align);
895 return NULL;
896 }
897
898 spin_lock_irqsave(&pcpu_lock, flags);
899
900 /* serve reserved allocations from the reserved chunk if available */
901 if (reserved && pcpu_reserved_chunk) {
902 chunk = pcpu_reserved_chunk;
903
904 if (size > chunk->contig_hint) {
905 err = "alloc from reserved chunk failed";
906 goto fail_unlock;
907 }
908
909 while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
910 spin_unlock_irqrestore(&pcpu_lock, flags);
911 if (is_atomic ||
912 pcpu_extend_area_map(chunk, new_alloc) < 0) {
913 err = "failed to extend area map of reserved chunk";
914 goto fail;
915 }
916 spin_lock_irqsave(&pcpu_lock, flags);
917 }
918
919 off = pcpu_alloc_area(chunk, size, align, is_atomic,
920 &occ_pages);
921 if (off >= 0)
922 goto area_found;
923
924 err = "alloc from reserved chunk failed";
925 goto fail_unlock;
926 }
927
928restart:
929 /* search through normal chunks */
930 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
931 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
932 if (size > chunk->contig_hint)
933 continue;
934
935 new_alloc = pcpu_need_to_extend(chunk, is_atomic);
936 if (new_alloc) {
937 if (is_atomic)
938 continue;
939 spin_unlock_irqrestore(&pcpu_lock, flags);
940 if (pcpu_extend_area_map(chunk,
941 new_alloc) < 0) {
942 err = "failed to extend area map";
943 goto fail;
944 }
945 spin_lock_irqsave(&pcpu_lock, flags);
946 /*
947 * pcpu_lock has been dropped, need to
948 * restart cpu_slot list walking.
949 */
950 goto restart;
951 }
952
953 off = pcpu_alloc_area(chunk, size, align, is_atomic,
954 &occ_pages);
955 if (off >= 0)
956 goto area_found;
957 }
958 }
959
960 spin_unlock_irqrestore(&pcpu_lock, flags);
961
962 /*
963 * No space left. Create a new chunk. We don't want multiple
964 * tasks to create chunks simultaneously. Serialize and create iff
965 * there's still no empty chunk after grabbing the mutex.
966 */
967 if (is_atomic)
968 goto fail;
969
970 mutex_lock(&pcpu_alloc_mutex);
971
972 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
973 chunk = pcpu_create_chunk();
974 if (!chunk) {
975 mutex_unlock(&pcpu_alloc_mutex);
976 err = "failed to allocate new chunk";
977 goto fail;
978 }
979
980 spin_lock_irqsave(&pcpu_lock, flags);
981 pcpu_chunk_relocate(chunk, -1);
982 } else {
983 spin_lock_irqsave(&pcpu_lock, flags);
984 }
985
986 mutex_unlock(&pcpu_alloc_mutex);
987 goto restart;
988
989area_found:
990 spin_unlock_irqrestore(&pcpu_lock, flags);
991
992 /* populate if not all pages are already there */
993 if (!is_atomic) {
994 int page_start, page_end, rs, re;
995
996 mutex_lock(&pcpu_alloc_mutex);
997
998 page_start = PFN_DOWN(off);
999 page_end = PFN_UP(off + size);
1000
1001 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
1002 WARN_ON(chunk->immutable);
1003
1004 ret = pcpu_populate_chunk(chunk, rs, re);
1005
1006 spin_lock_irqsave(&pcpu_lock, flags);
1007 if (ret) {
1008 mutex_unlock(&pcpu_alloc_mutex);
1009 pcpu_free_area(chunk, off, &occ_pages);
1010 err = "failed to populate";
1011 goto fail_unlock;
1012 }
1013 pcpu_chunk_populated(chunk, rs, re);
1014 spin_unlock_irqrestore(&pcpu_lock, flags);
1015 }
1016
1017 mutex_unlock(&pcpu_alloc_mutex);
1018 }
1019
1020 if (chunk != pcpu_reserved_chunk)
1021 pcpu_nr_empty_pop_pages -= occ_pages;
1022
1023 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1024 pcpu_schedule_balance_work();
1025
1026 /* clear the areas and return address relative to base address */
1027 for_each_possible_cpu(cpu)
1028 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1029
1030 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1031 kmemleak_alloc_percpu(ptr, size, gfp);
1032 return ptr;
1033
1034fail_unlock:
1035 spin_unlock_irqrestore(&pcpu_lock, flags);
1036fail:
1037 if (!is_atomic && warn_limit) {
1038 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1039 size, align, is_atomic, err);
1040 dump_stack();
1041 if (!--warn_limit)
1042 pr_info("limit reached, disable warning\n");
1043 }
1044 if (is_atomic) {
1045 /* see the flag handling in pcpu_blance_workfn() */
1046 pcpu_atomic_alloc_failed = true;
1047 pcpu_schedule_balance_work();
1048 }
1049 return NULL;
1050}
1051
1052/**
1053 * __alloc_percpu_gfp - allocate dynamic percpu area
1054 * @size: size of area to allocate in bytes
1055 * @align: alignment of area (max PAGE_SIZE)
1056 * @gfp: allocation flags
1057 *
1058 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1059 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1060 * be called from any context but is a lot more likely to fail.
1061 *
1062 * RETURNS:
1063 * Percpu pointer to the allocated area on success, NULL on failure.
1064 */
1065void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1066{
1067 return pcpu_alloc(size, align, false, gfp);
1068}
1069EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1070
1071/**
1072 * __alloc_percpu - allocate dynamic percpu area
1073 * @size: size of area to allocate in bytes
1074 * @align: alignment of area (max PAGE_SIZE)
1075 *
1076 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1077 */
1078void __percpu *__alloc_percpu(size_t size, size_t align)
1079{
1080 return pcpu_alloc(size, align, false, GFP_KERNEL);
1081}
1082EXPORT_SYMBOL_GPL(__alloc_percpu);
1083
1084/**
1085 * __alloc_reserved_percpu - allocate reserved percpu area
1086 * @size: size of area to allocate in bytes
1087 * @align: alignment of area (max PAGE_SIZE)
1088 *
1089 * Allocate zero-filled percpu area of @size bytes aligned at @align
1090 * from reserved percpu area if arch has set it up; otherwise,
1091 * allocation is served from the same dynamic area. Might sleep.
1092 * Might trigger writeouts.
1093 *
1094 * CONTEXT:
1095 * Does GFP_KERNEL allocation.
1096 *
1097 * RETURNS:
1098 * Percpu pointer to the allocated area on success, NULL on failure.
1099 */
1100void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1101{
1102 return pcpu_alloc(size, align, true, GFP_KERNEL);
1103}
1104
1105/**
1106 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1107 * @work: unused
1108 *
1109 * Reclaim all fully free chunks except for the first one.
1110 */
1111static void pcpu_balance_workfn(struct work_struct *work)
1112{
1113 LIST_HEAD(to_free);
1114 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1115 struct pcpu_chunk *chunk, *next;
1116 int slot, nr_to_pop, ret;
1117
1118 /*
1119 * There's no reason to keep around multiple unused chunks and VM
1120 * areas can be scarce. Destroy all free chunks except for one.
1121 */
1122 mutex_lock(&pcpu_alloc_mutex);
1123 spin_lock_irq(&pcpu_lock);
1124
1125 list_for_each_entry_safe(chunk, next, free_head, list) {
1126 WARN_ON(chunk->immutable);
1127
1128 /* spare the first one */
1129 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1130 continue;
1131
1132 list_move(&chunk->list, &to_free);
1133 }
1134
1135 spin_unlock_irq(&pcpu_lock);
1136
1137 list_for_each_entry_safe(chunk, next, &to_free, list) {
1138 int rs, re;
1139
1140 pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1141 pcpu_depopulate_chunk(chunk, rs, re);
1142 spin_lock_irq(&pcpu_lock);
1143 pcpu_chunk_depopulated(chunk, rs, re);
1144 spin_unlock_irq(&pcpu_lock);
1145 }
1146 pcpu_destroy_chunk(chunk);
1147 }
1148
1149 /*
1150 * Ensure there are certain number of free populated pages for
1151 * atomic allocs. Fill up from the most packed so that atomic
1152 * allocs don't increase fragmentation. If atomic allocation
1153 * failed previously, always populate the maximum amount. This
1154 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1155 * failing indefinitely; however, large atomic allocs are not
1156 * something we support properly and can be highly unreliable and
1157 * inefficient.
1158 */
1159retry_pop:
1160 if (pcpu_atomic_alloc_failed) {
1161 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1162 /* best effort anyway, don't worry about synchronization */
1163 pcpu_atomic_alloc_failed = false;
1164 } else {
1165 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1166 pcpu_nr_empty_pop_pages,
1167 0, PCPU_EMPTY_POP_PAGES_HIGH);
1168 }
1169
1170 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1171 int nr_unpop = 0, rs, re;
1172
1173 if (!nr_to_pop)
1174 break;
1175
1176 spin_lock_irq(&pcpu_lock);
1177 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1178 nr_unpop = pcpu_unit_pages - chunk->nr_populated;
1179 if (nr_unpop)
1180 break;
1181 }
1182 spin_unlock_irq(&pcpu_lock);
1183
1184 if (!nr_unpop)
1185 continue;
1186
1187 /* @chunk can't go away while pcpu_alloc_mutex is held */
1188 pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1189 int nr = min(re - rs, nr_to_pop);
1190
1191 ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1192 if (!ret) {
1193 nr_to_pop -= nr;
1194 spin_lock_irq(&pcpu_lock);
1195 pcpu_chunk_populated(chunk, rs, rs + nr);
1196 spin_unlock_irq(&pcpu_lock);
1197 } else {
1198 nr_to_pop = 0;
1199 }
1200
1201 if (!nr_to_pop)
1202 break;
1203 }
1204 }
1205
1206 if (nr_to_pop) {
1207 /* ran out of chunks to populate, create a new one and retry */
1208 chunk = pcpu_create_chunk();
1209 if (chunk) {
1210 spin_lock_irq(&pcpu_lock);
1211 pcpu_chunk_relocate(chunk, -1);
1212 spin_unlock_irq(&pcpu_lock);
1213 goto retry_pop;
1214 }
1215 }
1216
1217 mutex_unlock(&pcpu_alloc_mutex);
1218}
1219
1220/**
1221 * free_percpu - free percpu area
1222 * @ptr: pointer to area to free
1223 *
1224 * Free percpu area @ptr.
1225 *
1226 * CONTEXT:
1227 * Can be called from atomic context.
1228 */
1229void free_percpu(void __percpu *ptr)
1230{
1231 void *addr;
1232 struct pcpu_chunk *chunk;
1233 unsigned long flags;
1234 int off, occ_pages;
1235
1236 if (!ptr)
1237 return;
1238
1239 kmemleak_free_percpu(ptr);
1240
1241 addr = __pcpu_ptr_to_addr(ptr);
1242
1243 spin_lock_irqsave(&pcpu_lock, flags);
1244
1245 chunk = pcpu_chunk_addr_search(addr);
1246 off = addr - chunk->base_addr;
1247
1248 pcpu_free_area(chunk, off, &occ_pages);
1249
1250 if (chunk != pcpu_reserved_chunk)
1251 pcpu_nr_empty_pop_pages += occ_pages;
1252
1253 /* if there are more than one fully free chunks, wake up grim reaper */
1254 if (chunk->free_size == pcpu_unit_size) {
1255 struct pcpu_chunk *pos;
1256
1257 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1258 if (pos != chunk) {
1259 pcpu_schedule_balance_work();
1260 break;
1261 }
1262 }
1263
1264 spin_unlock_irqrestore(&pcpu_lock, flags);
1265}
1266EXPORT_SYMBOL_GPL(free_percpu);
1267
1268/**
1269 * is_kernel_percpu_address - test whether address is from static percpu area
1270 * @addr: address to test
1271 *
1272 * Test whether @addr belongs to in-kernel static percpu area. Module
1273 * static percpu areas are not considered. For those, use
1274 * is_module_percpu_address().
1275 *
1276 * RETURNS:
1277 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1278 */
1279bool is_kernel_percpu_address(unsigned long addr)
1280{
1281#ifdef CONFIG_SMP
1282 const size_t static_size = __per_cpu_end - __per_cpu_start;
1283 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1284 unsigned int cpu;
1285
1286 for_each_possible_cpu(cpu) {
1287 void *start = per_cpu_ptr(base, cpu);
1288
1289 if ((void *)addr >= start && (void *)addr < start + static_size)
1290 return true;
1291 }
1292#endif
1293 /* on UP, can't distinguish from other static vars, always false */
1294 return false;
1295}
1296
1297/**
1298 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1299 * @addr: the address to be converted to physical address
1300 *
1301 * Given @addr which is dereferenceable address obtained via one of
1302 * percpu access macros, this function translates it into its physical
1303 * address. The caller is responsible for ensuring @addr stays valid
1304 * until this function finishes.
1305 *
1306 * percpu allocator has special setup for the first chunk, which currently
1307 * supports either embedding in linear address space or vmalloc mapping,
1308 * and, from the second one, the backing allocator (currently either vm or
1309 * km) provides translation.
1310 *
1311 * The addr can be translated simply without checking if it falls into the
1312 * first chunk. But the current code reflects better how percpu allocator
1313 * actually works, and the verification can discover both bugs in percpu
1314 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1315 * code.
1316 *
1317 * RETURNS:
1318 * The physical address for @addr.
1319 */
1320phys_addr_t per_cpu_ptr_to_phys(void *addr)
1321{
1322 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1323 bool in_first_chunk = false;
1324 unsigned long first_low, first_high;
1325 unsigned int cpu;
1326
1327 /*
1328 * The following test on unit_low/high isn't strictly
1329 * necessary but will speed up lookups of addresses which
1330 * aren't in the first chunk.
1331 */
1332 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1333 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1334 pcpu_unit_pages);
1335 if ((unsigned long)addr >= first_low &&
1336 (unsigned long)addr < first_high) {
1337 for_each_possible_cpu(cpu) {
1338 void *start = per_cpu_ptr(base, cpu);
1339
1340 if (addr >= start && addr < start + pcpu_unit_size) {
1341 in_first_chunk = true;
1342 break;
1343 }
1344 }
1345 }
1346
1347 if (in_first_chunk) {
1348 if (!is_vmalloc_addr(addr))
1349 return __pa(addr);
1350 else
1351 return page_to_phys(vmalloc_to_page(addr)) +
1352 offset_in_page(addr);
1353 } else
1354 return page_to_phys(pcpu_addr_to_page(addr)) +
1355 offset_in_page(addr);
1356}
1357
1358/**
1359 * pcpu_alloc_alloc_info - allocate percpu allocation info
1360 * @nr_groups: the number of groups
1361 * @nr_units: the number of units
1362 *
1363 * Allocate ai which is large enough for @nr_groups groups containing
1364 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1365 * cpu_map array which is long enough for @nr_units and filled with
1366 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1367 * pointer of other groups.
1368 *
1369 * RETURNS:
1370 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1371 * failure.
1372 */
1373struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1374 int nr_units)
1375{
1376 struct pcpu_alloc_info *ai;
1377 size_t base_size, ai_size;
1378 void *ptr;
1379 int unit;
1380
1381 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1382 __alignof__(ai->groups[0].cpu_map[0]));
1383 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1384
1385 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1386 if (!ptr)
1387 return NULL;
1388 ai = ptr;
1389 ptr += base_size;
1390
1391 ai->groups[0].cpu_map = ptr;
1392
1393 for (unit = 0; unit < nr_units; unit++)
1394 ai->groups[0].cpu_map[unit] = NR_CPUS;
1395
1396 ai->nr_groups = nr_groups;
1397 ai->__ai_size = PFN_ALIGN(ai_size);
1398
1399 return ai;
1400}
1401
1402/**
1403 * pcpu_free_alloc_info - free percpu allocation info
1404 * @ai: pcpu_alloc_info to free
1405 *
1406 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1407 */
1408void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1409{
1410 memblock_free_early(__pa(ai), ai->__ai_size);
1411}
1412
1413/**
1414 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1415 * @lvl: loglevel
1416 * @ai: allocation info to dump
1417 *
1418 * Print out information about @ai using loglevel @lvl.
1419 */
1420static void pcpu_dump_alloc_info(const char *lvl,
1421 const struct pcpu_alloc_info *ai)
1422{
1423 int group_width = 1, cpu_width = 1, width;
1424 char empty_str[] = "--------";
1425 int alloc = 0, alloc_end = 0;
1426 int group, v;
1427 int upa, apl; /* units per alloc, allocs per line */
1428
1429 v = ai->nr_groups;
1430 while (v /= 10)
1431 group_width++;
1432
1433 v = num_possible_cpus();
1434 while (v /= 10)
1435 cpu_width++;
1436 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1437
1438 upa = ai->alloc_size / ai->unit_size;
1439 width = upa * (cpu_width + 1) + group_width + 3;
1440 apl = rounddown_pow_of_two(max(60 / width, 1));
1441
1442 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1443 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1444 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1445
1446 for (group = 0; group < ai->nr_groups; group++) {
1447 const struct pcpu_group_info *gi = &ai->groups[group];
1448 int unit = 0, unit_end = 0;
1449
1450 BUG_ON(gi->nr_units % upa);
1451 for (alloc_end += gi->nr_units / upa;
1452 alloc < alloc_end; alloc++) {
1453 if (!(alloc % apl)) {
1454 pr_cont("\n");
1455 printk("%spcpu-alloc: ", lvl);
1456 }
1457 pr_cont("[%0*d] ", group_width, group);
1458
1459 for (unit_end += upa; unit < unit_end; unit++)
1460 if (gi->cpu_map[unit] != NR_CPUS)
1461 pr_cont("%0*d ",
1462 cpu_width, gi->cpu_map[unit]);
1463 else
1464 pr_cont("%s ", empty_str);
1465 }
1466 }
1467 pr_cont("\n");
1468}
1469
1470/**
1471 * pcpu_setup_first_chunk - initialize the first percpu chunk
1472 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1473 * @base_addr: mapped address
1474 *
1475 * Initialize the first percpu chunk which contains the kernel static
1476 * perpcu area. This function is to be called from arch percpu area
1477 * setup path.
1478 *
1479 * @ai contains all information necessary to initialize the first
1480 * chunk and prime the dynamic percpu allocator.
1481 *
1482 * @ai->static_size is the size of static percpu area.
1483 *
1484 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1485 * reserve after the static area in the first chunk. This reserves
1486 * the first chunk such that it's available only through reserved
1487 * percpu allocation. This is primarily used to serve module percpu
1488 * static areas on architectures where the addressing model has
1489 * limited offset range for symbol relocations to guarantee module
1490 * percpu symbols fall inside the relocatable range.
1491 *
1492 * @ai->dyn_size determines the number of bytes available for dynamic
1493 * allocation in the first chunk. The area between @ai->static_size +
1494 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1495 *
1496 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1497 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1498 * @ai->dyn_size.
1499 *
1500 * @ai->atom_size is the allocation atom size and used as alignment
1501 * for vm areas.
1502 *
1503 * @ai->alloc_size is the allocation size and always multiple of
1504 * @ai->atom_size. This is larger than @ai->atom_size if
1505 * @ai->unit_size is larger than @ai->atom_size.
1506 *
1507 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1508 * percpu areas. Units which should be colocated are put into the
1509 * same group. Dynamic VM areas will be allocated according to these
1510 * groupings. If @ai->nr_groups is zero, a single group containing
1511 * all units is assumed.
1512 *
1513 * The caller should have mapped the first chunk at @base_addr and
1514 * copied static data to each unit.
1515 *
1516 * If the first chunk ends up with both reserved and dynamic areas, it
1517 * is served by two chunks - one to serve the core static and reserved
1518 * areas and the other for the dynamic area. They share the same vm
1519 * and page map but uses different area allocation map to stay away
1520 * from each other. The latter chunk is circulated in the chunk slots
1521 * and available for dynamic allocation like any other chunks.
1522 *
1523 * RETURNS:
1524 * 0 on success, -errno on failure.
1525 */
1526int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1527 void *base_addr)
1528{
1529 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1530 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1531 size_t dyn_size = ai->dyn_size;
1532 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1533 struct pcpu_chunk *schunk, *dchunk = NULL;
1534 unsigned long *group_offsets;
1535 size_t *group_sizes;
1536 unsigned long *unit_off;
1537 unsigned int cpu;
1538 int *unit_map;
1539 int group, unit, i;
1540
1541#define PCPU_SETUP_BUG_ON(cond) do { \
1542 if (unlikely(cond)) { \
1543 pr_emerg("failed to initialize, %s\n", #cond); \
1544 pr_emerg("cpu_possible_mask=%*pb\n", \
1545 cpumask_pr_args(cpu_possible_mask)); \
1546 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1547 BUG(); \
1548 } \
1549} while (0)
1550
1551 /* sanity checks */
1552 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1553#ifdef CONFIG_SMP
1554 PCPU_SETUP_BUG_ON(!ai->static_size);
1555 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
1556#endif
1557 PCPU_SETUP_BUG_ON(!base_addr);
1558 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
1559 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1560 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
1561 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1562 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1563 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1564
1565 /* process group information and build config tables accordingly */
1566 group_offsets = memblock_virt_alloc(ai->nr_groups *
1567 sizeof(group_offsets[0]), 0);
1568 group_sizes = memblock_virt_alloc(ai->nr_groups *
1569 sizeof(group_sizes[0]), 0);
1570 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1571 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1572
1573 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1574 unit_map[cpu] = UINT_MAX;
1575
1576 pcpu_low_unit_cpu = NR_CPUS;
1577 pcpu_high_unit_cpu = NR_CPUS;
1578
1579 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1580 const struct pcpu_group_info *gi = &ai->groups[group];
1581
1582 group_offsets[group] = gi->base_offset;
1583 group_sizes[group] = gi->nr_units * ai->unit_size;
1584
1585 for (i = 0; i < gi->nr_units; i++) {
1586 cpu = gi->cpu_map[i];
1587 if (cpu == NR_CPUS)
1588 continue;
1589
1590 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
1591 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1592 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1593
1594 unit_map[cpu] = unit + i;
1595 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1596
1597 /* determine low/high unit_cpu */
1598 if (pcpu_low_unit_cpu == NR_CPUS ||
1599 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1600 pcpu_low_unit_cpu = cpu;
1601 if (pcpu_high_unit_cpu == NR_CPUS ||
1602 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1603 pcpu_high_unit_cpu = cpu;
1604 }
1605 }
1606 pcpu_nr_units = unit;
1607
1608 for_each_possible_cpu(cpu)
1609 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1610
1611 /* we're done parsing the input, undefine BUG macro and dump config */
1612#undef PCPU_SETUP_BUG_ON
1613 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1614
1615 pcpu_nr_groups = ai->nr_groups;
1616 pcpu_group_offsets = group_offsets;
1617 pcpu_group_sizes = group_sizes;
1618 pcpu_unit_map = unit_map;
1619 pcpu_unit_offsets = unit_off;
1620
1621 /* determine basic parameters */
1622 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1623 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1624 pcpu_atom_size = ai->atom_size;
1625 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1626 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1627
1628 /*
1629 * Allocate chunk slots. The additional last slot is for
1630 * empty chunks.
1631 */
1632 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1633 pcpu_slot = memblock_virt_alloc(
1634 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1635 for (i = 0; i < pcpu_nr_slots; i++)
1636 INIT_LIST_HEAD(&pcpu_slot[i]);
1637
1638 /*
1639 * Initialize static chunk. If reserved_size is zero, the
1640 * static chunk covers static area + dynamic allocation area
1641 * in the first chunk. If reserved_size is not zero, it
1642 * covers static area + reserved area (mostly used for module
1643 * static percpu allocation).
1644 */
1645 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1646 INIT_LIST_HEAD(&schunk->list);
1647 INIT_WORK(&schunk->map_extend_work, pcpu_map_extend_workfn);
1648 schunk->base_addr = base_addr;
1649 schunk->map = smap;
1650 schunk->map_alloc = ARRAY_SIZE(smap);
1651 schunk->immutable = true;
1652 bitmap_fill(schunk->populated, pcpu_unit_pages);
1653 schunk->nr_populated = pcpu_unit_pages;
1654
1655 if (ai->reserved_size) {
1656 schunk->free_size = ai->reserved_size;
1657 pcpu_reserved_chunk = schunk;
1658 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1659 } else {
1660 schunk->free_size = dyn_size;
1661 dyn_size = 0; /* dynamic area covered */
1662 }
1663 schunk->contig_hint = schunk->free_size;
1664
1665 schunk->map[0] = 1;
1666 schunk->map[1] = ai->static_size;
1667 schunk->map_used = 1;
1668 if (schunk->free_size)
1669 schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size;
1670 schunk->map[schunk->map_used] |= 1;
1671
1672 /* init dynamic chunk if necessary */
1673 if (dyn_size) {
1674 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1675 INIT_LIST_HEAD(&dchunk->list);
1676 INIT_WORK(&dchunk->map_extend_work, pcpu_map_extend_workfn);
1677 dchunk->base_addr = base_addr;
1678 dchunk->map = dmap;
1679 dchunk->map_alloc = ARRAY_SIZE(dmap);
1680 dchunk->immutable = true;
1681 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1682 dchunk->nr_populated = pcpu_unit_pages;
1683
1684 dchunk->contig_hint = dchunk->free_size = dyn_size;
1685 dchunk->map[0] = 1;
1686 dchunk->map[1] = pcpu_reserved_chunk_limit;
1687 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1688 dchunk->map_used = 2;
1689 }
1690
1691 /* link the first chunk in */
1692 pcpu_first_chunk = dchunk ?: schunk;
1693 pcpu_nr_empty_pop_pages +=
1694 pcpu_count_occupied_pages(pcpu_first_chunk, 1);
1695 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1696
1697 /* we're done */
1698 pcpu_base_addr = base_addr;
1699 return 0;
1700}
1701
1702#ifdef CONFIG_SMP
1703
1704const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1705 [PCPU_FC_AUTO] = "auto",
1706 [PCPU_FC_EMBED] = "embed",
1707 [PCPU_FC_PAGE] = "page",
1708};
1709
1710enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1711
1712static int __init percpu_alloc_setup(char *str)
1713{
1714 if (!str)
1715 return -EINVAL;
1716
1717 if (0)
1718 /* nada */;
1719#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1720 else if (!strcmp(str, "embed"))
1721 pcpu_chosen_fc = PCPU_FC_EMBED;
1722#endif
1723#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1724 else if (!strcmp(str, "page"))
1725 pcpu_chosen_fc = PCPU_FC_PAGE;
1726#endif
1727 else
1728 pr_warn("unknown allocator %s specified\n", str);
1729
1730 return 0;
1731}
1732early_param("percpu_alloc", percpu_alloc_setup);
1733
1734/*
1735 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1736 * Build it if needed by the arch config or the generic setup is going
1737 * to be used.
1738 */
1739#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1740 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1741#define BUILD_EMBED_FIRST_CHUNK
1742#endif
1743
1744/* build pcpu_page_first_chunk() iff needed by the arch config */
1745#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1746#define BUILD_PAGE_FIRST_CHUNK
1747#endif
1748
1749/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1750#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1751/**
1752 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1753 * @reserved_size: the size of reserved percpu area in bytes
1754 * @dyn_size: minimum free size for dynamic allocation in bytes
1755 * @atom_size: allocation atom size
1756 * @cpu_distance_fn: callback to determine distance between cpus, optional
1757 *
1758 * This function determines grouping of units, their mappings to cpus
1759 * and other parameters considering needed percpu size, allocation
1760 * atom size and distances between CPUs.
1761 *
1762 * Groups are always multiples of atom size and CPUs which are of
1763 * LOCAL_DISTANCE both ways are grouped together and share space for
1764 * units in the same group. The returned configuration is guaranteed
1765 * to have CPUs on different nodes on different groups and >=75% usage
1766 * of allocated virtual address space.
1767 *
1768 * RETURNS:
1769 * On success, pointer to the new allocation_info is returned. On
1770 * failure, ERR_PTR value is returned.
1771 */
1772static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1773 size_t reserved_size, size_t dyn_size,
1774 size_t atom_size,
1775 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1776{
1777 static int group_map[NR_CPUS] __initdata;
1778 static int group_cnt[NR_CPUS] __initdata;
1779 const size_t static_size = __per_cpu_end - __per_cpu_start;
1780 int nr_groups = 1, nr_units = 0;
1781 size_t size_sum, min_unit_size, alloc_size;
1782 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1783 int last_allocs, group, unit;
1784 unsigned int cpu, tcpu;
1785 struct pcpu_alloc_info *ai;
1786 unsigned int *cpu_map;
1787
1788 /* this function may be called multiple times */
1789 memset(group_map, 0, sizeof(group_map));
1790 memset(group_cnt, 0, sizeof(group_cnt));
1791
1792 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1793 size_sum = PFN_ALIGN(static_size + reserved_size +
1794 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1795 dyn_size = size_sum - static_size - reserved_size;
1796
1797 /*
1798 * Determine min_unit_size, alloc_size and max_upa such that
1799 * alloc_size is multiple of atom_size and is the smallest
1800 * which can accommodate 4k aligned segments which are equal to
1801 * or larger than min_unit_size.
1802 */
1803 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1804
1805 alloc_size = roundup(min_unit_size, atom_size);
1806 upa = alloc_size / min_unit_size;
1807 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1808 upa--;
1809 max_upa = upa;
1810
1811 /* group cpus according to their proximity */
1812 for_each_possible_cpu(cpu) {
1813 group = 0;
1814 next_group:
1815 for_each_possible_cpu(tcpu) {
1816 if (cpu == tcpu)
1817 break;
1818 if (group_map[tcpu] == group && cpu_distance_fn &&
1819 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1820 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1821 group++;
1822 nr_groups = max(nr_groups, group + 1);
1823 goto next_group;
1824 }
1825 }
1826 group_map[cpu] = group;
1827 group_cnt[group]++;
1828 }
1829
1830 /*
1831 * Expand unit size until address space usage goes over 75%
1832 * and then as much as possible without using more address
1833 * space.
1834 */
1835 last_allocs = INT_MAX;
1836 for (upa = max_upa; upa; upa--) {
1837 int allocs = 0, wasted = 0;
1838
1839 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1840 continue;
1841
1842 for (group = 0; group < nr_groups; group++) {
1843 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1844 allocs += this_allocs;
1845 wasted += this_allocs * upa - group_cnt[group];
1846 }
1847
1848 /*
1849 * Don't accept if wastage is over 1/3. The
1850 * greater-than comparison ensures upa==1 always
1851 * passes the following check.
1852 */
1853 if (wasted > num_possible_cpus() / 3)
1854 continue;
1855
1856 /* and then don't consume more memory */
1857 if (allocs > last_allocs)
1858 break;
1859 last_allocs = allocs;
1860 best_upa = upa;
1861 }
1862 upa = best_upa;
1863
1864 /* allocate and fill alloc_info */
1865 for (group = 0; group < nr_groups; group++)
1866 nr_units += roundup(group_cnt[group], upa);
1867
1868 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1869 if (!ai)
1870 return ERR_PTR(-ENOMEM);
1871 cpu_map = ai->groups[0].cpu_map;
1872
1873 for (group = 0; group < nr_groups; group++) {
1874 ai->groups[group].cpu_map = cpu_map;
1875 cpu_map += roundup(group_cnt[group], upa);
1876 }
1877
1878 ai->static_size = static_size;
1879 ai->reserved_size = reserved_size;
1880 ai->dyn_size = dyn_size;
1881 ai->unit_size = alloc_size / upa;
1882 ai->atom_size = atom_size;
1883 ai->alloc_size = alloc_size;
1884
1885 for (group = 0, unit = 0; group_cnt[group]; group++) {
1886 struct pcpu_group_info *gi = &ai->groups[group];
1887
1888 /*
1889 * Initialize base_offset as if all groups are located
1890 * back-to-back. The caller should update this to
1891 * reflect actual allocation.
1892 */
1893 gi->base_offset = unit * ai->unit_size;
1894
1895 for_each_possible_cpu(cpu)
1896 if (group_map[cpu] == group)
1897 gi->cpu_map[gi->nr_units++] = cpu;
1898 gi->nr_units = roundup(gi->nr_units, upa);
1899 unit += gi->nr_units;
1900 }
1901 BUG_ON(unit != nr_units);
1902
1903 return ai;
1904}
1905#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1906
1907#if defined(BUILD_EMBED_FIRST_CHUNK)
1908/**
1909 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1910 * @reserved_size: the size of reserved percpu area in bytes
1911 * @dyn_size: minimum free size for dynamic allocation in bytes
1912 * @atom_size: allocation atom size
1913 * @cpu_distance_fn: callback to determine distance between cpus, optional
1914 * @alloc_fn: function to allocate percpu page
1915 * @free_fn: function to free percpu page
1916 *
1917 * This is a helper to ease setting up embedded first percpu chunk and
1918 * can be called where pcpu_setup_first_chunk() is expected.
1919 *
1920 * If this function is used to setup the first chunk, it is allocated
1921 * by calling @alloc_fn and used as-is without being mapped into
1922 * vmalloc area. Allocations are always whole multiples of @atom_size
1923 * aligned to @atom_size.
1924 *
1925 * This enables the first chunk to piggy back on the linear physical
1926 * mapping which often uses larger page size. Please note that this
1927 * can result in very sparse cpu->unit mapping on NUMA machines thus
1928 * requiring large vmalloc address space. Don't use this allocator if
1929 * vmalloc space is not orders of magnitude larger than distances
1930 * between node memory addresses (ie. 32bit NUMA machines).
1931 *
1932 * @dyn_size specifies the minimum dynamic area size.
1933 *
1934 * If the needed size is smaller than the minimum or specified unit
1935 * size, the leftover is returned using @free_fn.
1936 *
1937 * RETURNS:
1938 * 0 on success, -errno on failure.
1939 */
1940int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1941 size_t atom_size,
1942 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1943 pcpu_fc_alloc_fn_t alloc_fn,
1944 pcpu_fc_free_fn_t free_fn)
1945{
1946 void *base = (void *)ULONG_MAX;
1947 void **areas = NULL;
1948 struct pcpu_alloc_info *ai;
1949 size_t size_sum, areas_size, max_distance;
1950 int group, i, rc;
1951
1952 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1953 cpu_distance_fn);
1954 if (IS_ERR(ai))
1955 return PTR_ERR(ai);
1956
1957 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1958 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1959
1960 areas = memblock_virt_alloc_nopanic(areas_size, 0);
1961 if (!areas) {
1962 rc = -ENOMEM;
1963 goto out_free;
1964 }
1965
1966 /* allocate, copy and determine base address */
1967 for (group = 0; group < ai->nr_groups; group++) {
1968 struct pcpu_group_info *gi = &ai->groups[group];
1969 unsigned int cpu = NR_CPUS;
1970 void *ptr;
1971
1972 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1973 cpu = gi->cpu_map[i];
1974 BUG_ON(cpu == NR_CPUS);
1975
1976 /* allocate space for the whole group */
1977 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1978 if (!ptr) {
1979 rc = -ENOMEM;
1980 goto out_free_areas;
1981 }
1982 /* kmemleak tracks the percpu allocations separately */
1983 kmemleak_free(ptr);
1984 areas[group] = ptr;
1985
1986 base = min(ptr, base);
1987 }
1988
1989 /*
1990 * Copy data and free unused parts. This should happen after all
1991 * allocations are complete; otherwise, we may end up with
1992 * overlapping groups.
1993 */
1994 for (group = 0; group < ai->nr_groups; group++) {
1995 struct pcpu_group_info *gi = &ai->groups[group];
1996 void *ptr = areas[group];
1997
1998 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1999 if (gi->cpu_map[i] == NR_CPUS) {
2000 /* unused unit, free whole */
2001 free_fn(ptr, ai->unit_size);
2002 continue;
2003 }
2004 /* copy and return the unused part */
2005 memcpy(ptr, __per_cpu_load, ai->static_size);
2006 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2007 }
2008 }
2009
2010 /* base address is now known, determine group base offsets */
2011 max_distance = 0;
2012 for (group = 0; group < ai->nr_groups; group++) {
2013 ai->groups[group].base_offset = areas[group] - base;
2014 max_distance = max_t(size_t, max_distance,
2015 ai->groups[group].base_offset);
2016 }
2017 max_distance += ai->unit_size;
2018
2019 /* warn if maximum distance is further than 75% of vmalloc space */
2020 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2021 pr_warn("max_distance=0x%zx too large for vmalloc space 0x%lx\n",
2022 max_distance, VMALLOC_TOTAL);
2023#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2024 /* and fail if we have fallback */
2025 rc = -EINVAL;
2026 goto out_free;
2027#endif
2028 }
2029
2030 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2031 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2032 ai->dyn_size, ai->unit_size);
2033
2034 rc = pcpu_setup_first_chunk(ai, base);
2035 goto out_free;
2036
2037out_free_areas:
2038 for (group = 0; group < ai->nr_groups; group++)
2039 if (areas[group])
2040 free_fn(areas[group],
2041 ai->groups[group].nr_units * ai->unit_size);
2042out_free:
2043 pcpu_free_alloc_info(ai);
2044 if (areas)
2045 memblock_free_early(__pa(areas), areas_size);
2046 return rc;
2047}
2048#endif /* BUILD_EMBED_FIRST_CHUNK */
2049
2050#ifdef BUILD_PAGE_FIRST_CHUNK
2051/**
2052 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2053 * @reserved_size: the size of reserved percpu area in bytes
2054 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2055 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2056 * @populate_pte_fn: function to populate pte
2057 *
2058 * This is a helper to ease setting up page-remapped first percpu
2059 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2060 *
2061 * This is the basic allocator. Static percpu area is allocated
2062 * page-by-page into vmalloc area.
2063 *
2064 * RETURNS:
2065 * 0 on success, -errno on failure.
2066 */
2067int __init pcpu_page_first_chunk(size_t reserved_size,
2068 pcpu_fc_alloc_fn_t alloc_fn,
2069 pcpu_fc_free_fn_t free_fn,
2070 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2071{
2072 static struct vm_struct vm;
2073 struct pcpu_alloc_info *ai;
2074 char psize_str[16];
2075 int unit_pages;
2076 size_t pages_size;
2077 struct page **pages;
2078 int unit, i, j, rc;
2079
2080 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2081
2082 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2083 if (IS_ERR(ai))
2084 return PTR_ERR(ai);
2085 BUG_ON(ai->nr_groups != 1);
2086 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
2087
2088 unit_pages = ai->unit_size >> PAGE_SHIFT;
2089
2090 /* unaligned allocations can't be freed, round up to page size */
2091 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2092 sizeof(pages[0]));
2093 pages = memblock_virt_alloc(pages_size, 0);
2094
2095 /* allocate pages */
2096 j = 0;
2097 for (unit = 0; unit < num_possible_cpus(); unit++)
2098 for (i = 0; i < unit_pages; i++) {
2099 unsigned int cpu = ai->groups[0].cpu_map[unit];
2100 void *ptr;
2101
2102 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2103 if (!ptr) {
2104 pr_warn("failed to allocate %s page for cpu%u\n",
2105 psize_str, cpu);
2106 goto enomem;
2107 }
2108 /* kmemleak tracks the percpu allocations separately */
2109 kmemleak_free(ptr);
2110 pages[j++] = virt_to_page(ptr);
2111 }
2112
2113 /* allocate vm area, map the pages and copy static data */
2114 vm.flags = VM_ALLOC;
2115 vm.size = num_possible_cpus() * ai->unit_size;
2116 vm_area_register_early(&vm, PAGE_SIZE);
2117
2118 for (unit = 0; unit < num_possible_cpus(); unit++) {
2119 unsigned long unit_addr =
2120 (unsigned long)vm.addr + unit * ai->unit_size;
2121
2122 for (i = 0; i < unit_pages; i++)
2123 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2124
2125 /* pte already populated, the following shouldn't fail */
2126 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2127 unit_pages);
2128 if (rc < 0)
2129 panic("failed to map percpu area, err=%d\n", rc);
2130
2131 /*
2132 * FIXME: Archs with virtual cache should flush local
2133 * cache for the linear mapping here - something
2134 * equivalent to flush_cache_vmap() on the local cpu.
2135 * flush_cache_vmap() can't be used as most supporting
2136 * data structures are not set up yet.
2137 */
2138
2139 /* copy static data */
2140 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2141 }
2142
2143 /* we're ready, commit */
2144 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2145 unit_pages, psize_str, vm.addr, ai->static_size,
2146 ai->reserved_size, ai->dyn_size);
2147
2148 rc = pcpu_setup_first_chunk(ai, vm.addr);
2149 goto out_free_ar;
2150
2151enomem:
2152 while (--j >= 0)
2153 free_fn(page_address(pages[j]), PAGE_SIZE);
2154 rc = -ENOMEM;
2155out_free_ar:
2156 memblock_free_early(__pa(pages), pages_size);
2157 pcpu_free_alloc_info(ai);
2158 return rc;
2159}
2160#endif /* BUILD_PAGE_FIRST_CHUNK */
2161
2162#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2163/*
2164 * Generic SMP percpu area setup.
2165 *
2166 * The embedding helper is used because its behavior closely resembles
2167 * the original non-dynamic generic percpu area setup. This is
2168 * important because many archs have addressing restrictions and might
2169 * fail if the percpu area is located far away from the previous
2170 * location. As an added bonus, in non-NUMA cases, embedding is
2171 * generally a good idea TLB-wise because percpu area can piggy back
2172 * on the physical linear memory mapping which uses large page
2173 * mappings on applicable archs.
2174 */
2175unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2176EXPORT_SYMBOL(__per_cpu_offset);
2177
2178static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2179 size_t align)
2180{
2181 return memblock_virt_alloc_from_nopanic(
2182 size, align, __pa(MAX_DMA_ADDRESS));
2183}
2184
2185static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2186{
2187 memblock_free_early(__pa(ptr), size);
2188}
2189
2190void __init setup_per_cpu_areas(void)
2191{
2192 unsigned long delta;
2193 unsigned int cpu;
2194 int rc;
2195
2196 /*
2197 * Always reserve area for module percpu variables. That's
2198 * what the legacy allocator did.
2199 */
2200 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2201 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2202 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2203 if (rc < 0)
2204 panic("Failed to initialize percpu areas.");
2205
2206 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2207 for_each_possible_cpu(cpu)
2208 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2209}
2210#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2211
2212#else /* CONFIG_SMP */
2213
2214/*
2215 * UP percpu area setup.
2216 *
2217 * UP always uses km-based percpu allocator with identity mapping.
2218 * Static percpu variables are indistinguishable from the usual static
2219 * variables and don't require any special preparation.
2220 */
2221void __init setup_per_cpu_areas(void)
2222{
2223 const size_t unit_size =
2224 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2225 PERCPU_DYNAMIC_RESERVE));
2226 struct pcpu_alloc_info *ai;
2227 void *fc;
2228
2229 ai = pcpu_alloc_alloc_info(1, 1);
2230 fc = memblock_virt_alloc_from_nopanic(unit_size,
2231 PAGE_SIZE,
2232 __pa(MAX_DMA_ADDRESS));
2233 if (!ai || !fc)
2234 panic("Failed to allocate memory for percpu areas.");
2235 /* kmemleak tracks the percpu allocations separately */
2236 kmemleak_free(fc);
2237
2238 ai->dyn_size = unit_size;
2239 ai->unit_size = unit_size;
2240 ai->atom_size = unit_size;
2241 ai->alloc_size = unit_size;
2242 ai->groups[0].nr_units = 1;
2243 ai->groups[0].cpu_map[0] = 0;
2244
2245 if (pcpu_setup_first_chunk(ai, fc) < 0)
2246 panic("Failed to initialize percpu areas.");
2247}
2248
2249#endif /* CONFIG_SMP */
2250
2251/*
2252 * First and reserved chunks are initialized with temporary allocation
2253 * map in initdata so that they can be used before slab is online.
2254 * This function is called after slab is brought up and replaces those
2255 * with properly allocated maps.
2256 */
2257void __init percpu_init_late(void)
2258{
2259 struct pcpu_chunk *target_chunks[] =
2260 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2261 struct pcpu_chunk *chunk;
2262 unsigned long flags;
2263 int i;
2264
2265 for (i = 0; (chunk = target_chunks[i]); i++) {
2266 int *map;
2267 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2268
2269 BUILD_BUG_ON(size > PAGE_SIZE);
2270
2271 map = pcpu_mem_zalloc(size);
2272 BUG_ON(!map);
2273
2274 spin_lock_irqsave(&pcpu_lock, flags);
2275 memcpy(map, chunk->map, size);
2276 chunk->map = map;
2277 spin_unlock_irqrestore(&pcpu_lock, flags);
2278 }
2279}
2280
2281/*
2282 * Percpu allocator is initialized early during boot when neither slab or
2283 * workqueue is available. Plug async management until everything is up
2284 * and running.
2285 */
2286static int __init percpu_enable_async(void)
2287{
2288 pcpu_async_enabled = true;
2289 return 0;
2290}
2291subsys_initcall(percpu_enable_async);