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