Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/percpu.c - percpu memory allocator
4 *
5 * Copyright (C) 2009 SUSE Linux Products GmbH
6 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 *
8 * Copyright (C) 2017 Facebook Inc.
9 * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org>
10 *
11 * The percpu allocator handles both static and dynamic areas. Percpu
12 * areas are allocated in chunks which are divided into units. There is
13 * a 1-to-1 mapping for units to possible cpus. These units are grouped
14 * based on NUMA properties of the machine.
15 *
16 * c0 c1 c2
17 * ------------------- ------------------- ------------
18 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19 * ------------------- ...... ------------------- .... ------------
20 *
21 * Allocation is done by offsets into a unit's address space. Ie., an
22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24 * and even sparse. Access is handled by configuring percpu base
25 * registers according to the cpu to unit mappings and offsetting the
26 * base address using pcpu_unit_size.
27 *
28 * There is special consideration for the first chunk which must handle
29 * the static percpu variables in the kernel image as allocation services
30 * are not online yet. In short, the first chunk is structured like so:
31 *
32 * <Static | [Reserved] | Dynamic>
33 *
34 * The static data is copied from the original section managed by the
35 * linker. The reserved section, if non-zero, primarily manages static
36 * percpu variables from kernel modules. Finally, the dynamic section
37 * takes care of normal allocations.
38 *
39 * The allocator organizes chunks into lists according to free size and
40 * memcg-awareness. To make a percpu allocation memcg-aware the __GFP_ACCOUNT
41 * flag should be passed. All memcg-aware allocations are sharing one set
42 * of chunks and all unaccounted allocations and allocations performed
43 * by processes belonging to the root memory cgroup are using the second set.
44 *
45 * The allocator tries to allocate from the fullest chunk first. Each chunk
46 * is managed by a bitmap with metadata blocks. The allocation map is updated
47 * on every allocation and free to reflect the current state while the boundary
48 * map is only updated on allocation. Each metadata block contains
49 * information to help mitigate the need to iterate over large portions
50 * of the bitmap. The reverse mapping from page to chunk is stored in
51 * the page's index. Lastly, units are lazily backed and grow in unison.
52 *
53 * There is a unique conversion that goes on here between bytes and bits.
54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
55 * tracks the number of pages it is responsible for in nr_pages. Helper
56 * functions are used to convert from between the bytes, bits, and blocks.
57 * All hints are managed in bits unless explicitly stated.
58 *
59 * To use this allocator, arch code should do the following:
60 *
61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
62 * regular address to percpu pointer and back if they need to be
63 * different from the default
64 *
65 * - use pcpu_setup_first_chunk() during percpu area initialization to
66 * setup the first chunk containing the kernel static percpu area
67 */
68
69#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70
71#include <linux/bitmap.h>
72#include <linux/memblock.h>
73#include <linux/err.h>
74#include <linux/lcm.h>
75#include <linux/list.h>
76#include <linux/log2.h>
77#include <linux/mm.h>
78#include <linux/module.h>
79#include <linux/mutex.h>
80#include <linux/percpu.h>
81#include <linux/pfn.h>
82#include <linux/slab.h>
83#include <linux/spinlock.h>
84#include <linux/vmalloc.h>
85#include <linux/workqueue.h>
86#include <linux/kmemleak.h>
87#include <linux/sched.h>
88#include <linux/sched/mm.h>
89#include <linux/memcontrol.h>
90
91#include <asm/cacheflush.h>
92#include <asm/sections.h>
93#include <asm/tlbflush.h>
94#include <asm/io.h>
95
96#define CREATE_TRACE_POINTS
97#include <trace/events/percpu.h>
98
99#include "percpu-internal.h"
100
101/* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
102#define PCPU_SLOT_BASE_SHIFT 5
103/* chunks in slots below this are subject to being sidelined on failed alloc */
104#define PCPU_SLOT_FAIL_THRESHOLD 3
105
106#define PCPU_EMPTY_POP_PAGES_LOW 2
107#define PCPU_EMPTY_POP_PAGES_HIGH 4
108
109#ifdef CONFIG_SMP
110/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
111#ifndef __addr_to_pcpu_ptr
112#define __addr_to_pcpu_ptr(addr) \
113 (void __percpu *)((unsigned long)(addr) - \
114 (unsigned long)pcpu_base_addr + \
115 (unsigned long)__per_cpu_start)
116#endif
117#ifndef __pcpu_ptr_to_addr
118#define __pcpu_ptr_to_addr(ptr) \
119 (void __force *)((unsigned long)(ptr) + \
120 (unsigned long)pcpu_base_addr - \
121 (unsigned long)__per_cpu_start)
122#endif
123#else /* CONFIG_SMP */
124/* on UP, it's always identity mapped */
125#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
126#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
127#endif /* CONFIG_SMP */
128
129static int pcpu_unit_pages __ro_after_init;
130static int pcpu_unit_size __ro_after_init;
131static int pcpu_nr_units __ro_after_init;
132static int pcpu_atom_size __ro_after_init;
133int pcpu_nr_slots __ro_after_init;
134static size_t pcpu_chunk_struct_size __ro_after_init;
135
136/* cpus with the lowest and highest unit addresses */
137static unsigned int pcpu_low_unit_cpu __ro_after_init;
138static unsigned int pcpu_high_unit_cpu __ro_after_init;
139
140/* the address of the first chunk which starts with the kernel static area */
141void *pcpu_base_addr __ro_after_init;
142EXPORT_SYMBOL_GPL(pcpu_base_addr);
143
144static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
145const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
146
147/* group information, used for vm allocation */
148static int pcpu_nr_groups __ro_after_init;
149static const unsigned long *pcpu_group_offsets __ro_after_init;
150static const size_t *pcpu_group_sizes __ro_after_init;
151
152/*
153 * The first chunk which always exists. Note that unlike other
154 * chunks, this one can be allocated and mapped in several different
155 * ways and thus often doesn't live in the vmalloc area.
156 */
157struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
158
159/*
160 * Optional reserved chunk. This chunk reserves part of the first
161 * chunk and serves it for reserved allocations. When the reserved
162 * region doesn't exist, the following variable is NULL.
163 */
164struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
165
166DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
167static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
168
169struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
170
171/* chunks which need their map areas extended, protected by pcpu_lock */
172static LIST_HEAD(pcpu_map_extend_chunks);
173
174/*
175 * The number of empty populated pages, protected by pcpu_lock. The
176 * reserved chunk doesn't contribute to the count.
177 */
178int pcpu_nr_empty_pop_pages;
179
180/*
181 * The number of populated pages in use by the allocator, protected by
182 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
183 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
184 * and increments/decrements this count by 1).
185 */
186static unsigned long pcpu_nr_populated;
187
188/*
189 * Balance work is used to populate or destroy chunks asynchronously. We
190 * try to keep the number of populated free pages between
191 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
192 * empty chunk.
193 */
194static void pcpu_balance_workfn(struct work_struct *work);
195static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
196static bool pcpu_async_enabled __read_mostly;
197static bool pcpu_atomic_alloc_failed;
198
199static void pcpu_schedule_balance_work(void)
200{
201 if (pcpu_async_enabled)
202 schedule_work(&pcpu_balance_work);
203}
204
205/**
206 * pcpu_addr_in_chunk - check if the address is served from this chunk
207 * @chunk: chunk of interest
208 * @addr: percpu address
209 *
210 * RETURNS:
211 * True if the address is served from this chunk.
212 */
213static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
214{
215 void *start_addr, *end_addr;
216
217 if (!chunk)
218 return false;
219
220 start_addr = chunk->base_addr + chunk->start_offset;
221 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
222 chunk->end_offset;
223
224 return addr >= start_addr && addr < end_addr;
225}
226
227static int __pcpu_size_to_slot(int size)
228{
229 int highbit = fls(size); /* size is in bytes */
230 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
231}
232
233static int pcpu_size_to_slot(int size)
234{
235 if (size == pcpu_unit_size)
236 return pcpu_nr_slots - 1;
237 return __pcpu_size_to_slot(size);
238}
239
240static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
241{
242 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
243
244 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
245 chunk_md->contig_hint == 0)
246 return 0;
247
248 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
249}
250
251/* set the pointer to a chunk in a page struct */
252static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
253{
254 page->index = (unsigned long)pcpu;
255}
256
257/* obtain pointer to a chunk from a page struct */
258static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
259{
260 return (struct pcpu_chunk *)page->index;
261}
262
263static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
264{
265 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
266}
267
268static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
269{
270 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
271}
272
273static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
274 unsigned int cpu, int page_idx)
275{
276 return (unsigned long)chunk->base_addr +
277 pcpu_unit_page_offset(cpu, page_idx);
278}
279
280/*
281 * The following are helper functions to help access bitmaps and convert
282 * between bitmap offsets to address offsets.
283 */
284static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
285{
286 return chunk->alloc_map +
287 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
288}
289
290static unsigned long pcpu_off_to_block_index(int off)
291{
292 return off / PCPU_BITMAP_BLOCK_BITS;
293}
294
295static unsigned long pcpu_off_to_block_off(int off)
296{
297 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
298}
299
300static unsigned long pcpu_block_off_to_off(int index, int off)
301{
302 return index * PCPU_BITMAP_BLOCK_BITS + off;
303}
304
305/*
306 * pcpu_next_hint - determine which hint to use
307 * @block: block of interest
308 * @alloc_bits: size of allocation
309 *
310 * This determines if we should scan based on the scan_hint or first_free.
311 * In general, we want to scan from first_free to fulfill allocations by
312 * first fit. However, if we know a scan_hint at position scan_hint_start
313 * cannot fulfill an allocation, we can begin scanning from there knowing
314 * the contig_hint will be our fallback.
315 */
316static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
317{
318 /*
319 * The three conditions below determine if we can skip past the
320 * scan_hint. First, does the scan hint exist. Second, is the
321 * contig_hint after the scan_hint (possibly not true iff
322 * contig_hint == scan_hint). Third, is the allocation request
323 * larger than the scan_hint.
324 */
325 if (block->scan_hint &&
326 block->contig_hint_start > block->scan_hint_start &&
327 alloc_bits > block->scan_hint)
328 return block->scan_hint_start + block->scan_hint;
329
330 return block->first_free;
331}
332
333/**
334 * pcpu_next_md_free_region - finds the next hint free area
335 * @chunk: chunk of interest
336 * @bit_off: chunk offset
337 * @bits: size of free area
338 *
339 * Helper function for pcpu_for_each_md_free_region. It checks
340 * block->contig_hint and performs aggregation across blocks to find the
341 * next hint. It modifies bit_off and bits in-place to be consumed in the
342 * loop.
343 */
344static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
345 int *bits)
346{
347 int i = pcpu_off_to_block_index(*bit_off);
348 int block_off = pcpu_off_to_block_off(*bit_off);
349 struct pcpu_block_md *block;
350
351 *bits = 0;
352 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
353 block++, i++) {
354 /* handles contig area across blocks */
355 if (*bits) {
356 *bits += block->left_free;
357 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
358 continue;
359 return;
360 }
361
362 /*
363 * This checks three things. First is there a contig_hint to
364 * check. Second, have we checked this hint before by
365 * comparing the block_off. Third, is this the same as the
366 * right contig hint. In the last case, it spills over into
367 * the next block and should be handled by the contig area
368 * across blocks code.
369 */
370 *bits = block->contig_hint;
371 if (*bits && block->contig_hint_start >= block_off &&
372 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
373 *bit_off = pcpu_block_off_to_off(i,
374 block->contig_hint_start);
375 return;
376 }
377 /* reset to satisfy the second predicate above */
378 block_off = 0;
379
380 *bits = block->right_free;
381 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
382 }
383}
384
385/**
386 * pcpu_next_fit_region - finds fit areas for a given allocation request
387 * @chunk: chunk of interest
388 * @alloc_bits: size of allocation
389 * @align: alignment of area (max PAGE_SIZE)
390 * @bit_off: chunk offset
391 * @bits: size of free area
392 *
393 * Finds the next free region that is viable for use with a given size and
394 * alignment. This only returns if there is a valid area to be used for this
395 * allocation. block->first_free is returned if the allocation request fits
396 * within the block to see if the request can be fulfilled prior to the contig
397 * hint.
398 */
399static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
400 int align, int *bit_off, int *bits)
401{
402 int i = pcpu_off_to_block_index(*bit_off);
403 int block_off = pcpu_off_to_block_off(*bit_off);
404 struct pcpu_block_md *block;
405
406 *bits = 0;
407 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
408 block++, i++) {
409 /* handles contig area across blocks */
410 if (*bits) {
411 *bits += block->left_free;
412 if (*bits >= alloc_bits)
413 return;
414 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
415 continue;
416 }
417
418 /* check block->contig_hint */
419 *bits = ALIGN(block->contig_hint_start, align) -
420 block->contig_hint_start;
421 /*
422 * This uses the block offset to determine if this has been
423 * checked in the prior iteration.
424 */
425 if (block->contig_hint &&
426 block->contig_hint_start >= block_off &&
427 block->contig_hint >= *bits + alloc_bits) {
428 int start = pcpu_next_hint(block, alloc_bits);
429
430 *bits += alloc_bits + block->contig_hint_start -
431 start;
432 *bit_off = pcpu_block_off_to_off(i, start);
433 return;
434 }
435 /* reset to satisfy the second predicate above */
436 block_off = 0;
437
438 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
439 align);
440 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
441 *bit_off = pcpu_block_off_to_off(i, *bit_off);
442 if (*bits >= alloc_bits)
443 return;
444 }
445
446 /* no valid offsets were found - fail condition */
447 *bit_off = pcpu_chunk_map_bits(chunk);
448}
449
450/*
451 * Metadata free area iterators. These perform aggregation of free areas
452 * based on the metadata blocks and return the offset @bit_off and size in
453 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
454 * a fit is found for the allocation request.
455 */
456#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
457 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
458 (bit_off) < pcpu_chunk_map_bits((chunk)); \
459 (bit_off) += (bits) + 1, \
460 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
461
462#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
463 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
464 &(bits)); \
465 (bit_off) < pcpu_chunk_map_bits((chunk)); \
466 (bit_off) += (bits), \
467 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
468 &(bits)))
469
470/**
471 * pcpu_mem_zalloc - allocate memory
472 * @size: bytes to allocate
473 * @gfp: allocation flags
474 *
475 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
476 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
477 * This is to facilitate passing through whitelisted flags. The
478 * returned memory is always zeroed.
479 *
480 * RETURNS:
481 * Pointer to the allocated area on success, NULL on failure.
482 */
483static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
484{
485 if (WARN_ON_ONCE(!slab_is_available()))
486 return NULL;
487
488 if (size <= PAGE_SIZE)
489 return kzalloc(size, gfp);
490 else
491 return __vmalloc(size, gfp | __GFP_ZERO);
492}
493
494/**
495 * pcpu_mem_free - free memory
496 * @ptr: memory to free
497 *
498 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
499 */
500static void pcpu_mem_free(void *ptr)
501{
502 kvfree(ptr);
503}
504
505static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
506 bool move_front)
507{
508 if (chunk != pcpu_reserved_chunk) {
509 struct list_head *pcpu_slot;
510
511 pcpu_slot = pcpu_chunk_list(pcpu_chunk_type(chunk));
512 if (move_front)
513 list_move(&chunk->list, &pcpu_slot[slot]);
514 else
515 list_move_tail(&chunk->list, &pcpu_slot[slot]);
516 }
517}
518
519static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
520{
521 __pcpu_chunk_move(chunk, slot, true);
522}
523
524/**
525 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
526 * @chunk: chunk of interest
527 * @oslot: the previous slot it was on
528 *
529 * This function is called after an allocation or free changed @chunk.
530 * New slot according to the changed state is determined and @chunk is
531 * moved to the slot. Note that the reserved chunk is never put on
532 * chunk slots.
533 *
534 * CONTEXT:
535 * pcpu_lock.
536 */
537static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
538{
539 int nslot = pcpu_chunk_slot(chunk);
540
541 if (oslot != nslot)
542 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
543}
544
545/*
546 * pcpu_update_empty_pages - update empty page counters
547 * @chunk: chunk of interest
548 * @nr: nr of empty pages
549 *
550 * This is used to keep track of the empty pages now based on the premise
551 * a md_block covers a page. The hint update functions recognize if a block
552 * is made full or broken to calculate deltas for keeping track of free pages.
553 */
554static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
555{
556 chunk->nr_empty_pop_pages += nr;
557 if (chunk != pcpu_reserved_chunk)
558 pcpu_nr_empty_pop_pages += nr;
559}
560
561/*
562 * pcpu_region_overlap - determines if two regions overlap
563 * @a: start of first region, inclusive
564 * @b: end of first region, exclusive
565 * @x: start of second region, inclusive
566 * @y: end of second region, exclusive
567 *
568 * This is used to determine if the hint region [a, b) overlaps with the
569 * allocated region [x, y).
570 */
571static inline bool pcpu_region_overlap(int a, int b, int x, int y)
572{
573 return (a < y) && (x < b);
574}
575
576/**
577 * pcpu_block_update - updates a block given a free area
578 * @block: block of interest
579 * @start: start offset in block
580 * @end: end offset in block
581 *
582 * Updates a block given a known free area. The region [start, end) is
583 * expected to be the entirety of the free area within a block. Chooses
584 * the best starting offset if the contig hints are equal.
585 */
586static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
587{
588 int contig = end - start;
589
590 block->first_free = min(block->first_free, start);
591 if (start == 0)
592 block->left_free = contig;
593
594 if (end == block->nr_bits)
595 block->right_free = contig;
596
597 if (contig > block->contig_hint) {
598 /* promote the old contig_hint to be the new scan_hint */
599 if (start > block->contig_hint_start) {
600 if (block->contig_hint > block->scan_hint) {
601 block->scan_hint_start =
602 block->contig_hint_start;
603 block->scan_hint = block->contig_hint;
604 } else if (start < block->scan_hint_start) {
605 /*
606 * The old contig_hint == scan_hint. But, the
607 * new contig is larger so hold the invariant
608 * scan_hint_start < contig_hint_start.
609 */
610 block->scan_hint = 0;
611 }
612 } else {
613 block->scan_hint = 0;
614 }
615 block->contig_hint_start = start;
616 block->contig_hint = contig;
617 } else if (contig == block->contig_hint) {
618 if (block->contig_hint_start &&
619 (!start ||
620 __ffs(start) > __ffs(block->contig_hint_start))) {
621 /* start has a better alignment so use it */
622 block->contig_hint_start = start;
623 if (start < block->scan_hint_start &&
624 block->contig_hint > block->scan_hint)
625 block->scan_hint = 0;
626 } else if (start > block->scan_hint_start ||
627 block->contig_hint > block->scan_hint) {
628 /*
629 * Knowing contig == contig_hint, update the scan_hint
630 * if it is farther than or larger than the current
631 * scan_hint.
632 */
633 block->scan_hint_start = start;
634 block->scan_hint = contig;
635 }
636 } else {
637 /*
638 * The region is smaller than the contig_hint. So only update
639 * the scan_hint if it is larger than or equal and farther than
640 * the current scan_hint.
641 */
642 if ((start < block->contig_hint_start &&
643 (contig > block->scan_hint ||
644 (contig == block->scan_hint &&
645 start > block->scan_hint_start)))) {
646 block->scan_hint_start = start;
647 block->scan_hint = contig;
648 }
649 }
650}
651
652/*
653 * pcpu_block_update_scan - update a block given a free area from a scan
654 * @chunk: chunk of interest
655 * @bit_off: chunk offset
656 * @bits: size of free area
657 *
658 * Finding the final allocation spot first goes through pcpu_find_block_fit()
659 * to find a block that can hold the allocation and then pcpu_alloc_area()
660 * where a scan is used. When allocations require specific alignments,
661 * we can inadvertently create holes which will not be seen in the alloc
662 * or free paths.
663 *
664 * This takes a given free area hole and updates a block as it may change the
665 * scan_hint. We need to scan backwards to ensure we don't miss free bits
666 * from alignment.
667 */
668static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
669 int bits)
670{
671 int s_off = pcpu_off_to_block_off(bit_off);
672 int e_off = s_off + bits;
673 int s_index, l_bit;
674 struct pcpu_block_md *block;
675
676 if (e_off > PCPU_BITMAP_BLOCK_BITS)
677 return;
678
679 s_index = pcpu_off_to_block_index(bit_off);
680 block = chunk->md_blocks + s_index;
681
682 /* scan backwards in case of alignment skipping free bits */
683 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
684 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
685
686 pcpu_block_update(block, s_off, e_off);
687}
688
689/**
690 * pcpu_chunk_refresh_hint - updates metadata about a chunk
691 * @chunk: chunk of interest
692 * @full_scan: if we should scan from the beginning
693 *
694 * Iterates over the metadata blocks to find the largest contig area.
695 * A full scan can be avoided on the allocation path as this is triggered
696 * if we broke the contig_hint. In doing so, the scan_hint will be before
697 * the contig_hint or after if the scan_hint == contig_hint. This cannot
698 * be prevented on freeing as we want to find the largest area possibly
699 * spanning blocks.
700 */
701static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
702{
703 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
704 int bit_off, bits;
705
706 /* promote scan_hint to contig_hint */
707 if (!full_scan && chunk_md->scan_hint) {
708 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
709 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
710 chunk_md->contig_hint = chunk_md->scan_hint;
711 chunk_md->scan_hint = 0;
712 } else {
713 bit_off = chunk_md->first_free;
714 chunk_md->contig_hint = 0;
715 }
716
717 bits = 0;
718 pcpu_for_each_md_free_region(chunk, bit_off, bits)
719 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
720}
721
722/**
723 * pcpu_block_refresh_hint
724 * @chunk: chunk of interest
725 * @index: index of the metadata block
726 *
727 * Scans over the block beginning at first_free and updates the block
728 * metadata accordingly.
729 */
730static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
731{
732 struct pcpu_block_md *block = chunk->md_blocks + index;
733 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
734 unsigned int rs, re, start; /* region start, region end */
735
736 /* promote scan_hint to contig_hint */
737 if (block->scan_hint) {
738 start = block->scan_hint_start + block->scan_hint;
739 block->contig_hint_start = block->scan_hint_start;
740 block->contig_hint = block->scan_hint;
741 block->scan_hint = 0;
742 } else {
743 start = block->first_free;
744 block->contig_hint = 0;
745 }
746
747 block->right_free = 0;
748
749 /* iterate over free areas and update the contig hints */
750 bitmap_for_each_clear_region(alloc_map, rs, re, start,
751 PCPU_BITMAP_BLOCK_BITS)
752 pcpu_block_update(block, rs, re);
753}
754
755/**
756 * pcpu_block_update_hint_alloc - update hint on allocation path
757 * @chunk: chunk of interest
758 * @bit_off: chunk offset
759 * @bits: size of request
760 *
761 * Updates metadata for the allocation path. The metadata only has to be
762 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
763 * scans are required if the block's contig hint is broken.
764 */
765static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
766 int bits)
767{
768 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
769 int nr_empty_pages = 0;
770 struct pcpu_block_md *s_block, *e_block, *block;
771 int s_index, e_index; /* block indexes of the freed allocation */
772 int s_off, e_off; /* block offsets of the freed allocation */
773
774 /*
775 * Calculate per block offsets.
776 * The calculation uses an inclusive range, but the resulting offsets
777 * are [start, end). e_index always points to the last block in the
778 * range.
779 */
780 s_index = pcpu_off_to_block_index(bit_off);
781 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
782 s_off = pcpu_off_to_block_off(bit_off);
783 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
784
785 s_block = chunk->md_blocks + s_index;
786 e_block = chunk->md_blocks + e_index;
787
788 /*
789 * Update s_block.
790 * block->first_free must be updated if the allocation takes its place.
791 * If the allocation breaks the contig_hint, a scan is required to
792 * restore this hint.
793 */
794 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
795 nr_empty_pages++;
796
797 if (s_off == s_block->first_free)
798 s_block->first_free = find_next_zero_bit(
799 pcpu_index_alloc_map(chunk, s_index),
800 PCPU_BITMAP_BLOCK_BITS,
801 s_off + bits);
802
803 if (pcpu_region_overlap(s_block->scan_hint_start,
804 s_block->scan_hint_start + s_block->scan_hint,
805 s_off,
806 s_off + bits))
807 s_block->scan_hint = 0;
808
809 if (pcpu_region_overlap(s_block->contig_hint_start,
810 s_block->contig_hint_start +
811 s_block->contig_hint,
812 s_off,
813 s_off + bits)) {
814 /* block contig hint is broken - scan to fix it */
815 if (!s_off)
816 s_block->left_free = 0;
817 pcpu_block_refresh_hint(chunk, s_index);
818 } else {
819 /* update left and right contig manually */
820 s_block->left_free = min(s_block->left_free, s_off);
821 if (s_index == e_index)
822 s_block->right_free = min_t(int, s_block->right_free,
823 PCPU_BITMAP_BLOCK_BITS - e_off);
824 else
825 s_block->right_free = 0;
826 }
827
828 /*
829 * Update e_block.
830 */
831 if (s_index != e_index) {
832 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
833 nr_empty_pages++;
834
835 /*
836 * When the allocation is across blocks, the end is along
837 * the left part of the e_block.
838 */
839 e_block->first_free = find_next_zero_bit(
840 pcpu_index_alloc_map(chunk, e_index),
841 PCPU_BITMAP_BLOCK_BITS, e_off);
842
843 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
844 /* reset the block */
845 e_block++;
846 } else {
847 if (e_off > e_block->scan_hint_start)
848 e_block->scan_hint = 0;
849
850 e_block->left_free = 0;
851 if (e_off > e_block->contig_hint_start) {
852 /* contig hint is broken - scan to fix it */
853 pcpu_block_refresh_hint(chunk, e_index);
854 } else {
855 e_block->right_free =
856 min_t(int, e_block->right_free,
857 PCPU_BITMAP_BLOCK_BITS - e_off);
858 }
859 }
860
861 /* update in-between md_blocks */
862 nr_empty_pages += (e_index - s_index - 1);
863 for (block = s_block + 1; block < e_block; block++) {
864 block->scan_hint = 0;
865 block->contig_hint = 0;
866 block->left_free = 0;
867 block->right_free = 0;
868 }
869 }
870
871 if (nr_empty_pages)
872 pcpu_update_empty_pages(chunk, -nr_empty_pages);
873
874 if (pcpu_region_overlap(chunk_md->scan_hint_start,
875 chunk_md->scan_hint_start +
876 chunk_md->scan_hint,
877 bit_off,
878 bit_off + bits))
879 chunk_md->scan_hint = 0;
880
881 /*
882 * The only time a full chunk scan is required is if the chunk
883 * contig hint is broken. Otherwise, it means a smaller space
884 * was used and therefore the chunk contig hint is still correct.
885 */
886 if (pcpu_region_overlap(chunk_md->contig_hint_start,
887 chunk_md->contig_hint_start +
888 chunk_md->contig_hint,
889 bit_off,
890 bit_off + bits))
891 pcpu_chunk_refresh_hint(chunk, false);
892}
893
894/**
895 * pcpu_block_update_hint_free - updates the block hints on the free path
896 * @chunk: chunk of interest
897 * @bit_off: chunk offset
898 * @bits: size of request
899 *
900 * Updates metadata for the allocation path. This avoids a blind block
901 * refresh by making use of the block contig hints. If this fails, it scans
902 * forward and backward to determine the extent of the free area. This is
903 * capped at the boundary of blocks.
904 *
905 * A chunk update is triggered if a page becomes free, a block becomes free,
906 * or the free spans across blocks. This tradeoff is to minimize iterating
907 * over the block metadata to update chunk_md->contig_hint.
908 * chunk_md->contig_hint may be off by up to a page, but it will never be more
909 * than the available space. If the contig hint is contained in one block, it
910 * will be accurate.
911 */
912static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
913 int bits)
914{
915 int nr_empty_pages = 0;
916 struct pcpu_block_md *s_block, *e_block, *block;
917 int s_index, e_index; /* block indexes of the freed allocation */
918 int s_off, e_off; /* block offsets of the freed allocation */
919 int start, end; /* start and end of the whole free area */
920
921 /*
922 * Calculate per block offsets.
923 * The calculation uses an inclusive range, but the resulting offsets
924 * are [start, end). e_index always points to the last block in the
925 * range.
926 */
927 s_index = pcpu_off_to_block_index(bit_off);
928 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
929 s_off = pcpu_off_to_block_off(bit_off);
930 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
931
932 s_block = chunk->md_blocks + s_index;
933 e_block = chunk->md_blocks + e_index;
934
935 /*
936 * Check if the freed area aligns with the block->contig_hint.
937 * If it does, then the scan to find the beginning/end of the
938 * larger free area can be avoided.
939 *
940 * start and end refer to beginning and end of the free area
941 * within each their respective blocks. This is not necessarily
942 * the entire free area as it may span blocks past the beginning
943 * or end of the block.
944 */
945 start = s_off;
946 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
947 start = s_block->contig_hint_start;
948 } else {
949 /*
950 * Scan backwards to find the extent of the free area.
951 * find_last_bit returns the starting bit, so if the start bit
952 * is returned, that means there was no last bit and the
953 * remainder of the chunk is free.
954 */
955 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
956 start);
957 start = (start == l_bit) ? 0 : l_bit + 1;
958 }
959
960 end = e_off;
961 if (e_off == e_block->contig_hint_start)
962 end = e_block->contig_hint_start + e_block->contig_hint;
963 else
964 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
965 PCPU_BITMAP_BLOCK_BITS, end);
966
967 /* update s_block */
968 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
969 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
970 nr_empty_pages++;
971 pcpu_block_update(s_block, start, e_off);
972
973 /* freeing in the same block */
974 if (s_index != e_index) {
975 /* update e_block */
976 if (end == PCPU_BITMAP_BLOCK_BITS)
977 nr_empty_pages++;
978 pcpu_block_update(e_block, 0, end);
979
980 /* reset md_blocks in the middle */
981 nr_empty_pages += (e_index - s_index - 1);
982 for (block = s_block + 1; block < e_block; block++) {
983 block->first_free = 0;
984 block->scan_hint = 0;
985 block->contig_hint_start = 0;
986 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
987 block->left_free = PCPU_BITMAP_BLOCK_BITS;
988 block->right_free = PCPU_BITMAP_BLOCK_BITS;
989 }
990 }
991
992 if (nr_empty_pages)
993 pcpu_update_empty_pages(chunk, nr_empty_pages);
994
995 /*
996 * Refresh chunk metadata when the free makes a block free or spans
997 * across blocks. The contig_hint may be off by up to a page, but if
998 * the contig_hint is contained in a block, it will be accurate with
999 * the else condition below.
1000 */
1001 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1002 pcpu_chunk_refresh_hint(chunk, true);
1003 else
1004 pcpu_block_update(&chunk->chunk_md,
1005 pcpu_block_off_to_off(s_index, start),
1006 end);
1007}
1008
1009/**
1010 * pcpu_is_populated - determines if the region is populated
1011 * @chunk: chunk of interest
1012 * @bit_off: chunk offset
1013 * @bits: size of area
1014 * @next_off: return value for the next offset to start searching
1015 *
1016 * For atomic allocations, check if the backing pages are populated.
1017 *
1018 * RETURNS:
1019 * Bool if the backing pages are populated.
1020 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1021 */
1022static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1023 int *next_off)
1024{
1025 unsigned int page_start, page_end, rs, re;
1026
1027 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1028 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1029
1030 rs = page_start;
1031 bitmap_next_clear_region(chunk->populated, &rs, &re, page_end);
1032 if (rs >= page_end)
1033 return true;
1034
1035 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1036 return false;
1037}
1038
1039/**
1040 * pcpu_find_block_fit - finds the block index to start searching
1041 * @chunk: chunk of interest
1042 * @alloc_bits: size of request in allocation units
1043 * @align: alignment of area (max PAGE_SIZE bytes)
1044 * @pop_only: use populated regions only
1045 *
1046 * Given a chunk and an allocation spec, find the offset to begin searching
1047 * for a free region. This iterates over the bitmap metadata blocks to
1048 * find an offset that will be guaranteed to fit the requirements. It is
1049 * not quite first fit as if the allocation does not fit in the contig hint
1050 * of a block or chunk, it is skipped. This errs on the side of caution
1051 * to prevent excess iteration. Poor alignment can cause the allocator to
1052 * skip over blocks and chunks that have valid free areas.
1053 *
1054 * RETURNS:
1055 * The offset in the bitmap to begin searching.
1056 * -1 if no offset is found.
1057 */
1058static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1059 size_t align, bool pop_only)
1060{
1061 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1062 int bit_off, bits, next_off;
1063
1064 /*
1065 * Check to see if the allocation can fit in the chunk's contig hint.
1066 * This is an optimization to prevent scanning by assuming if it
1067 * cannot fit in the global hint, there is memory pressure and creating
1068 * a new chunk would happen soon.
1069 */
1070 bit_off = ALIGN(chunk_md->contig_hint_start, align) -
1071 chunk_md->contig_hint_start;
1072 if (bit_off + alloc_bits > chunk_md->contig_hint)
1073 return -1;
1074
1075 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1076 bits = 0;
1077 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1078 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1079 &next_off))
1080 break;
1081
1082 bit_off = next_off;
1083 bits = 0;
1084 }
1085
1086 if (bit_off == pcpu_chunk_map_bits(chunk))
1087 return -1;
1088
1089 return bit_off;
1090}
1091
1092/*
1093 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1094 * @map: the address to base the search on
1095 * @size: the bitmap size in bits
1096 * @start: the bitnumber to start searching at
1097 * @nr: the number of zeroed bits we're looking for
1098 * @align_mask: alignment mask for zero area
1099 * @largest_off: offset of the largest area skipped
1100 * @largest_bits: size of the largest area skipped
1101 *
1102 * The @align_mask should be one less than a power of 2.
1103 *
1104 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1105 * the largest area that was skipped. This is imperfect, but in general is
1106 * good enough. The largest remembered region is the largest failed region
1107 * seen. This does not include anything we possibly skipped due to alignment.
1108 * pcpu_block_update_scan() does scan backwards to try and recover what was
1109 * lost to alignment. While this can cause scanning to miss earlier possible
1110 * free areas, smaller allocations will eventually fill those holes.
1111 */
1112static unsigned long pcpu_find_zero_area(unsigned long *map,
1113 unsigned long size,
1114 unsigned long start,
1115 unsigned long nr,
1116 unsigned long align_mask,
1117 unsigned long *largest_off,
1118 unsigned long *largest_bits)
1119{
1120 unsigned long index, end, i, area_off, area_bits;
1121again:
1122 index = find_next_zero_bit(map, size, start);
1123
1124 /* Align allocation */
1125 index = __ALIGN_MASK(index, align_mask);
1126 area_off = index;
1127
1128 end = index + nr;
1129 if (end > size)
1130 return end;
1131 i = find_next_bit(map, end, index);
1132 if (i < end) {
1133 area_bits = i - area_off;
1134 /* remember largest unused area with best alignment */
1135 if (area_bits > *largest_bits ||
1136 (area_bits == *largest_bits && *largest_off &&
1137 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1138 *largest_off = area_off;
1139 *largest_bits = area_bits;
1140 }
1141
1142 start = i + 1;
1143 goto again;
1144 }
1145 return index;
1146}
1147
1148/**
1149 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1150 * @chunk: chunk of interest
1151 * @alloc_bits: size of request in allocation units
1152 * @align: alignment of area (max PAGE_SIZE)
1153 * @start: bit_off to start searching
1154 *
1155 * This function takes in a @start offset to begin searching to fit an
1156 * allocation of @alloc_bits with alignment @align. It needs to scan
1157 * the allocation map because if it fits within the block's contig hint,
1158 * @start will be block->first_free. This is an attempt to fill the
1159 * allocation prior to breaking the contig hint. The allocation and
1160 * boundary maps are updated accordingly if it confirms a valid
1161 * free area.
1162 *
1163 * RETURNS:
1164 * Allocated addr offset in @chunk on success.
1165 * -1 if no matching area is found.
1166 */
1167static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1168 size_t align, int start)
1169{
1170 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1171 size_t align_mask = (align) ? (align - 1) : 0;
1172 unsigned long area_off = 0, area_bits = 0;
1173 int bit_off, end, oslot;
1174
1175 lockdep_assert_held(&pcpu_lock);
1176
1177 oslot = pcpu_chunk_slot(chunk);
1178
1179 /*
1180 * Search to find a fit.
1181 */
1182 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1183 pcpu_chunk_map_bits(chunk));
1184 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1185 align_mask, &area_off, &area_bits);
1186 if (bit_off >= end)
1187 return -1;
1188
1189 if (area_bits)
1190 pcpu_block_update_scan(chunk, area_off, area_bits);
1191
1192 /* update alloc map */
1193 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1194
1195 /* update boundary map */
1196 set_bit(bit_off, chunk->bound_map);
1197 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1198 set_bit(bit_off + alloc_bits, chunk->bound_map);
1199
1200 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1201
1202 /* update first free bit */
1203 if (bit_off == chunk_md->first_free)
1204 chunk_md->first_free = find_next_zero_bit(
1205 chunk->alloc_map,
1206 pcpu_chunk_map_bits(chunk),
1207 bit_off + alloc_bits);
1208
1209 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1210
1211 pcpu_chunk_relocate(chunk, oslot);
1212
1213 return bit_off * PCPU_MIN_ALLOC_SIZE;
1214}
1215
1216/**
1217 * pcpu_free_area - frees the corresponding offset
1218 * @chunk: chunk of interest
1219 * @off: addr offset into chunk
1220 *
1221 * This function determines the size of an allocation to free using
1222 * the boundary bitmap and clears the allocation map.
1223 *
1224 * RETURNS:
1225 * Number of freed bytes.
1226 */
1227static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1228{
1229 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1230 int bit_off, bits, end, oslot, freed;
1231
1232 lockdep_assert_held(&pcpu_lock);
1233 pcpu_stats_area_dealloc(chunk);
1234
1235 oslot = pcpu_chunk_slot(chunk);
1236
1237 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1238
1239 /* find end index */
1240 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1241 bit_off + 1);
1242 bits = end - bit_off;
1243 bitmap_clear(chunk->alloc_map, bit_off, bits);
1244
1245 freed = bits * PCPU_MIN_ALLOC_SIZE;
1246
1247 /* update metadata */
1248 chunk->free_bytes += freed;
1249
1250 /* update first free bit */
1251 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1252
1253 pcpu_block_update_hint_free(chunk, bit_off, bits);
1254
1255 pcpu_chunk_relocate(chunk, oslot);
1256
1257 return freed;
1258}
1259
1260static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1261{
1262 block->scan_hint = 0;
1263 block->contig_hint = nr_bits;
1264 block->left_free = nr_bits;
1265 block->right_free = nr_bits;
1266 block->first_free = 0;
1267 block->nr_bits = nr_bits;
1268}
1269
1270static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1271{
1272 struct pcpu_block_md *md_block;
1273
1274 /* init the chunk's block */
1275 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1276
1277 for (md_block = chunk->md_blocks;
1278 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1279 md_block++)
1280 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1281}
1282
1283/**
1284 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1285 * @tmp_addr: the start of the region served
1286 * @map_size: size of the region served
1287 *
1288 * This is responsible for creating the chunks that serve the first chunk. The
1289 * base_addr is page aligned down of @tmp_addr while the region end is page
1290 * aligned up. Offsets are kept track of to determine the region served. All
1291 * this is done to appease the bitmap allocator in avoiding partial blocks.
1292 *
1293 * RETURNS:
1294 * Chunk serving the region at @tmp_addr of @map_size.
1295 */
1296static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1297 int map_size)
1298{
1299 struct pcpu_chunk *chunk;
1300 unsigned long aligned_addr, lcm_align;
1301 int start_offset, offset_bits, region_size, region_bits;
1302 size_t alloc_size;
1303
1304 /* region calculations */
1305 aligned_addr = tmp_addr & PAGE_MASK;
1306
1307 start_offset = tmp_addr - aligned_addr;
1308
1309 /*
1310 * Align the end of the region with the LCM of PAGE_SIZE and
1311 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1312 * the other.
1313 */
1314 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1315 region_size = ALIGN(start_offset + map_size, lcm_align);
1316
1317 /* allocate chunk */
1318 alloc_size = sizeof(struct pcpu_chunk) +
1319 BITS_TO_LONGS(region_size >> PAGE_SHIFT) * sizeof(unsigned long);
1320 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1321 if (!chunk)
1322 panic("%s: Failed to allocate %zu bytes\n", __func__,
1323 alloc_size);
1324
1325 INIT_LIST_HEAD(&chunk->list);
1326
1327 chunk->base_addr = (void *)aligned_addr;
1328 chunk->start_offset = start_offset;
1329 chunk->end_offset = region_size - chunk->start_offset - map_size;
1330
1331 chunk->nr_pages = region_size >> PAGE_SHIFT;
1332 region_bits = pcpu_chunk_map_bits(chunk);
1333
1334 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1335 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1336 if (!chunk->alloc_map)
1337 panic("%s: Failed to allocate %zu bytes\n", __func__,
1338 alloc_size);
1339
1340 alloc_size =
1341 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1342 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1343 if (!chunk->bound_map)
1344 panic("%s: Failed to allocate %zu bytes\n", __func__,
1345 alloc_size);
1346
1347 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1348 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1349 if (!chunk->md_blocks)
1350 panic("%s: Failed to allocate %zu bytes\n", __func__,
1351 alloc_size);
1352
1353#ifdef CONFIG_MEMCG_KMEM
1354 /* first chunk isn't memcg-aware */
1355 chunk->obj_cgroups = NULL;
1356#endif
1357 pcpu_init_md_blocks(chunk);
1358
1359 /* manage populated page bitmap */
1360 chunk->immutable = true;
1361 bitmap_fill(chunk->populated, chunk->nr_pages);
1362 chunk->nr_populated = chunk->nr_pages;
1363 chunk->nr_empty_pop_pages = chunk->nr_pages;
1364
1365 chunk->free_bytes = map_size;
1366
1367 if (chunk->start_offset) {
1368 /* hide the beginning of the bitmap */
1369 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1370 bitmap_set(chunk->alloc_map, 0, offset_bits);
1371 set_bit(0, chunk->bound_map);
1372 set_bit(offset_bits, chunk->bound_map);
1373
1374 chunk->chunk_md.first_free = offset_bits;
1375
1376 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1377 }
1378
1379 if (chunk->end_offset) {
1380 /* hide the end of the bitmap */
1381 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1382 bitmap_set(chunk->alloc_map,
1383 pcpu_chunk_map_bits(chunk) - offset_bits,
1384 offset_bits);
1385 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1386 chunk->bound_map);
1387 set_bit(region_bits, chunk->bound_map);
1388
1389 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1390 - offset_bits, offset_bits);
1391 }
1392
1393 return chunk;
1394}
1395
1396static struct pcpu_chunk *pcpu_alloc_chunk(enum pcpu_chunk_type type, gfp_t gfp)
1397{
1398 struct pcpu_chunk *chunk;
1399 int region_bits;
1400
1401 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1402 if (!chunk)
1403 return NULL;
1404
1405 INIT_LIST_HEAD(&chunk->list);
1406 chunk->nr_pages = pcpu_unit_pages;
1407 region_bits = pcpu_chunk_map_bits(chunk);
1408
1409 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1410 sizeof(chunk->alloc_map[0]), gfp);
1411 if (!chunk->alloc_map)
1412 goto alloc_map_fail;
1413
1414 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1415 sizeof(chunk->bound_map[0]), gfp);
1416 if (!chunk->bound_map)
1417 goto bound_map_fail;
1418
1419 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1420 sizeof(chunk->md_blocks[0]), gfp);
1421 if (!chunk->md_blocks)
1422 goto md_blocks_fail;
1423
1424#ifdef CONFIG_MEMCG_KMEM
1425 if (pcpu_is_memcg_chunk(type)) {
1426 chunk->obj_cgroups =
1427 pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1428 sizeof(struct obj_cgroup *), gfp);
1429 if (!chunk->obj_cgroups)
1430 goto objcg_fail;
1431 }
1432#endif
1433
1434 pcpu_init_md_blocks(chunk);
1435
1436 /* init metadata */
1437 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1438
1439 return chunk;
1440
1441#ifdef CONFIG_MEMCG_KMEM
1442objcg_fail:
1443 pcpu_mem_free(chunk->md_blocks);
1444#endif
1445md_blocks_fail:
1446 pcpu_mem_free(chunk->bound_map);
1447bound_map_fail:
1448 pcpu_mem_free(chunk->alloc_map);
1449alloc_map_fail:
1450 pcpu_mem_free(chunk);
1451
1452 return NULL;
1453}
1454
1455static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1456{
1457 if (!chunk)
1458 return;
1459#ifdef CONFIG_MEMCG_KMEM
1460 pcpu_mem_free(chunk->obj_cgroups);
1461#endif
1462 pcpu_mem_free(chunk->md_blocks);
1463 pcpu_mem_free(chunk->bound_map);
1464 pcpu_mem_free(chunk->alloc_map);
1465 pcpu_mem_free(chunk);
1466}
1467
1468/**
1469 * pcpu_chunk_populated - post-population bookkeeping
1470 * @chunk: pcpu_chunk which got populated
1471 * @page_start: the start page
1472 * @page_end: the end page
1473 *
1474 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1475 * the bookkeeping information accordingly. Must be called after each
1476 * successful population.
1477 *
1478 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1479 * is to serve an allocation in that area.
1480 */
1481static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1482 int page_end)
1483{
1484 int nr = page_end - page_start;
1485
1486 lockdep_assert_held(&pcpu_lock);
1487
1488 bitmap_set(chunk->populated, page_start, nr);
1489 chunk->nr_populated += nr;
1490 pcpu_nr_populated += nr;
1491
1492 pcpu_update_empty_pages(chunk, nr);
1493}
1494
1495/**
1496 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1497 * @chunk: pcpu_chunk which got depopulated
1498 * @page_start: the start page
1499 * @page_end: the end page
1500 *
1501 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1502 * Update the bookkeeping information accordingly. Must be called after
1503 * each successful depopulation.
1504 */
1505static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1506 int page_start, int page_end)
1507{
1508 int nr = page_end - page_start;
1509
1510 lockdep_assert_held(&pcpu_lock);
1511
1512 bitmap_clear(chunk->populated, page_start, nr);
1513 chunk->nr_populated -= nr;
1514 pcpu_nr_populated -= nr;
1515
1516 pcpu_update_empty_pages(chunk, -nr);
1517}
1518
1519/*
1520 * Chunk management implementation.
1521 *
1522 * To allow different implementations, chunk alloc/free and
1523 * [de]population are implemented in a separate file which is pulled
1524 * into this file and compiled together. The following functions
1525 * should be implemented.
1526 *
1527 * pcpu_populate_chunk - populate the specified range of a chunk
1528 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1529 * pcpu_create_chunk - create a new chunk
1530 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1531 * pcpu_addr_to_page - translate address to physical address
1532 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1533 */
1534static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1535 int page_start, int page_end, gfp_t gfp);
1536static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1537 int page_start, int page_end);
1538static struct pcpu_chunk *pcpu_create_chunk(enum pcpu_chunk_type type,
1539 gfp_t gfp);
1540static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1541static struct page *pcpu_addr_to_page(void *addr);
1542static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1543
1544#ifdef CONFIG_NEED_PER_CPU_KM
1545#include "percpu-km.c"
1546#else
1547#include "percpu-vm.c"
1548#endif
1549
1550/**
1551 * pcpu_chunk_addr_search - determine chunk containing specified address
1552 * @addr: address for which the chunk needs to be determined.
1553 *
1554 * This is an internal function that handles all but static allocations.
1555 * Static percpu address values should never be passed into the allocator.
1556 *
1557 * RETURNS:
1558 * The address of the found chunk.
1559 */
1560static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1561{
1562 /* is it in the dynamic region (first chunk)? */
1563 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1564 return pcpu_first_chunk;
1565
1566 /* is it in the reserved region? */
1567 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1568 return pcpu_reserved_chunk;
1569
1570 /*
1571 * The address is relative to unit0 which might be unused and
1572 * thus unmapped. Offset the address to the unit space of the
1573 * current processor before looking it up in the vmalloc
1574 * space. Note that any possible cpu id can be used here, so
1575 * there's no need to worry about preemption or cpu hotplug.
1576 */
1577 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1578 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1579}
1580
1581#ifdef CONFIG_MEMCG_KMEM
1582static enum pcpu_chunk_type pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1583 struct obj_cgroup **objcgp)
1584{
1585 struct obj_cgroup *objcg;
1586
1587 if (!memcg_kmem_enabled() || !(gfp & __GFP_ACCOUNT) ||
1588 memcg_kmem_bypass())
1589 return PCPU_CHUNK_ROOT;
1590
1591 objcg = get_obj_cgroup_from_current();
1592 if (!objcg)
1593 return PCPU_CHUNK_ROOT;
1594
1595 if (obj_cgroup_charge(objcg, gfp, size * num_possible_cpus())) {
1596 obj_cgroup_put(objcg);
1597 return PCPU_FAIL_ALLOC;
1598 }
1599
1600 *objcgp = objcg;
1601 return PCPU_CHUNK_MEMCG;
1602}
1603
1604static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1605 struct pcpu_chunk *chunk, int off,
1606 size_t size)
1607{
1608 if (!objcg)
1609 return;
1610
1611 if (chunk) {
1612 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1613
1614 rcu_read_lock();
1615 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1616 size * num_possible_cpus());
1617 rcu_read_unlock();
1618 } else {
1619 obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1620 obj_cgroup_put(objcg);
1621 }
1622}
1623
1624static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1625{
1626 struct obj_cgroup *objcg;
1627
1628 if (!pcpu_is_memcg_chunk(pcpu_chunk_type(chunk)))
1629 return;
1630
1631 objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1632 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1633
1634 obj_cgroup_uncharge(objcg, size * num_possible_cpus());
1635
1636 rcu_read_lock();
1637 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1638 -(size * num_possible_cpus()));
1639 rcu_read_unlock();
1640
1641 obj_cgroup_put(objcg);
1642}
1643
1644#else /* CONFIG_MEMCG_KMEM */
1645static enum pcpu_chunk_type
1646pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1647{
1648 return PCPU_CHUNK_ROOT;
1649}
1650
1651static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1652 struct pcpu_chunk *chunk, int off,
1653 size_t size)
1654{
1655}
1656
1657static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1658{
1659}
1660#endif /* CONFIG_MEMCG_KMEM */
1661
1662/**
1663 * pcpu_alloc - the percpu allocator
1664 * @size: size of area to allocate in bytes
1665 * @align: alignment of area (max PAGE_SIZE)
1666 * @reserved: allocate from the reserved chunk if available
1667 * @gfp: allocation flags
1668 *
1669 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1670 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1671 * then no warning will be triggered on invalid or failed allocation
1672 * requests.
1673 *
1674 * RETURNS:
1675 * Percpu pointer to the allocated area on success, NULL on failure.
1676 */
1677static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1678 gfp_t gfp)
1679{
1680 gfp_t pcpu_gfp;
1681 bool is_atomic;
1682 bool do_warn;
1683 enum pcpu_chunk_type type;
1684 struct list_head *pcpu_slot;
1685 struct obj_cgroup *objcg = NULL;
1686 static int warn_limit = 10;
1687 struct pcpu_chunk *chunk, *next;
1688 const char *err;
1689 int slot, off, cpu, ret;
1690 unsigned long flags;
1691 void __percpu *ptr;
1692 size_t bits, bit_align;
1693
1694 gfp = current_gfp_context(gfp);
1695 /* whitelisted flags that can be passed to the backing allocators */
1696 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1697 is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1698 do_warn = !(gfp & __GFP_NOWARN);
1699
1700 /*
1701 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1702 * therefore alignment must be a minimum of that many bytes.
1703 * An allocation may have internal fragmentation from rounding up
1704 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1705 */
1706 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1707 align = PCPU_MIN_ALLOC_SIZE;
1708
1709 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1710 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1711 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1712
1713 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1714 !is_power_of_2(align))) {
1715 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1716 size, align);
1717 return NULL;
1718 }
1719
1720 type = pcpu_memcg_pre_alloc_hook(size, gfp, &objcg);
1721 if (unlikely(type == PCPU_FAIL_ALLOC))
1722 return NULL;
1723 pcpu_slot = pcpu_chunk_list(type);
1724
1725 if (!is_atomic) {
1726 /*
1727 * pcpu_balance_workfn() allocates memory under this mutex,
1728 * and it may wait for memory reclaim. Allow current task
1729 * to become OOM victim, in case of memory pressure.
1730 */
1731 if (gfp & __GFP_NOFAIL) {
1732 mutex_lock(&pcpu_alloc_mutex);
1733 } else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1734 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1735 return NULL;
1736 }
1737 }
1738
1739 spin_lock_irqsave(&pcpu_lock, flags);
1740
1741 /* serve reserved allocations from the reserved chunk if available */
1742 if (reserved && pcpu_reserved_chunk) {
1743 chunk = pcpu_reserved_chunk;
1744
1745 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1746 if (off < 0) {
1747 err = "alloc from reserved chunk failed";
1748 goto fail_unlock;
1749 }
1750
1751 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1752 if (off >= 0)
1753 goto area_found;
1754
1755 err = "alloc from reserved chunk failed";
1756 goto fail_unlock;
1757 }
1758
1759restart:
1760 /* search through normal chunks */
1761 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1762 list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) {
1763 off = pcpu_find_block_fit(chunk, bits, bit_align,
1764 is_atomic);
1765 if (off < 0) {
1766 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1767 pcpu_chunk_move(chunk, 0);
1768 continue;
1769 }
1770
1771 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1772 if (off >= 0)
1773 goto area_found;
1774
1775 }
1776 }
1777
1778 spin_unlock_irqrestore(&pcpu_lock, flags);
1779
1780 /*
1781 * No space left. Create a new chunk. We don't want multiple
1782 * tasks to create chunks simultaneously. Serialize and create iff
1783 * there's still no empty chunk after grabbing the mutex.
1784 */
1785 if (is_atomic) {
1786 err = "atomic alloc failed, no space left";
1787 goto fail;
1788 }
1789
1790 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1791 chunk = pcpu_create_chunk(type, pcpu_gfp);
1792 if (!chunk) {
1793 err = "failed to allocate new chunk";
1794 goto fail;
1795 }
1796
1797 spin_lock_irqsave(&pcpu_lock, flags);
1798 pcpu_chunk_relocate(chunk, -1);
1799 } else {
1800 spin_lock_irqsave(&pcpu_lock, flags);
1801 }
1802
1803 goto restart;
1804
1805area_found:
1806 pcpu_stats_area_alloc(chunk, size);
1807 spin_unlock_irqrestore(&pcpu_lock, flags);
1808
1809 /* populate if not all pages are already there */
1810 if (!is_atomic) {
1811 unsigned int page_start, page_end, rs, re;
1812
1813 page_start = PFN_DOWN(off);
1814 page_end = PFN_UP(off + size);
1815
1816 bitmap_for_each_clear_region(chunk->populated, rs, re,
1817 page_start, page_end) {
1818 WARN_ON(chunk->immutable);
1819
1820 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1821
1822 spin_lock_irqsave(&pcpu_lock, flags);
1823 if (ret) {
1824 pcpu_free_area(chunk, off);
1825 err = "failed to populate";
1826 goto fail_unlock;
1827 }
1828 pcpu_chunk_populated(chunk, rs, re);
1829 spin_unlock_irqrestore(&pcpu_lock, flags);
1830 }
1831
1832 mutex_unlock(&pcpu_alloc_mutex);
1833 }
1834
1835 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1836 pcpu_schedule_balance_work();
1837
1838 /* clear the areas and return address relative to base address */
1839 for_each_possible_cpu(cpu)
1840 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1841
1842 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1843 kmemleak_alloc_percpu(ptr, size, gfp);
1844
1845 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1846 chunk->base_addr, off, ptr);
1847
1848 pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1849
1850 return ptr;
1851
1852fail_unlock:
1853 spin_unlock_irqrestore(&pcpu_lock, flags);
1854fail:
1855 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1856
1857 if (!is_atomic && do_warn && warn_limit) {
1858 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1859 size, align, is_atomic, err);
1860 dump_stack();
1861 if (!--warn_limit)
1862 pr_info("limit reached, disable warning\n");
1863 }
1864 if (is_atomic) {
1865 /* see the flag handling in pcpu_blance_workfn() */
1866 pcpu_atomic_alloc_failed = true;
1867 pcpu_schedule_balance_work();
1868 } else {
1869 mutex_unlock(&pcpu_alloc_mutex);
1870 }
1871
1872 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1873
1874 return NULL;
1875}
1876
1877/**
1878 * __alloc_percpu_gfp - allocate dynamic percpu area
1879 * @size: size of area to allocate in bytes
1880 * @align: alignment of area (max PAGE_SIZE)
1881 * @gfp: allocation flags
1882 *
1883 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1884 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1885 * be called from any context but is a lot more likely to fail. If @gfp
1886 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1887 * allocation requests.
1888 *
1889 * RETURNS:
1890 * Percpu pointer to the allocated area on success, NULL on failure.
1891 */
1892void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1893{
1894 return pcpu_alloc(size, align, false, gfp);
1895}
1896EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1897
1898/**
1899 * __alloc_percpu - allocate dynamic percpu area
1900 * @size: size of area to allocate in bytes
1901 * @align: alignment of area (max PAGE_SIZE)
1902 *
1903 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1904 */
1905void __percpu *__alloc_percpu(size_t size, size_t align)
1906{
1907 return pcpu_alloc(size, align, false, GFP_KERNEL);
1908}
1909EXPORT_SYMBOL_GPL(__alloc_percpu);
1910
1911/**
1912 * __alloc_reserved_percpu - allocate reserved percpu area
1913 * @size: size of area to allocate in bytes
1914 * @align: alignment of area (max PAGE_SIZE)
1915 *
1916 * Allocate zero-filled percpu area of @size bytes aligned at @align
1917 * from reserved percpu area if arch has set it up; otherwise,
1918 * allocation is served from the same dynamic area. Might sleep.
1919 * Might trigger writeouts.
1920 *
1921 * CONTEXT:
1922 * Does GFP_KERNEL allocation.
1923 *
1924 * RETURNS:
1925 * Percpu pointer to the allocated area on success, NULL on failure.
1926 */
1927void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1928{
1929 return pcpu_alloc(size, align, true, GFP_KERNEL);
1930}
1931
1932/**
1933 * __pcpu_balance_workfn - manage the amount of free chunks and populated pages
1934 * @type: chunk type
1935 *
1936 * Reclaim all fully free chunks except for the first one. This is also
1937 * responsible for maintaining the pool of empty populated pages. However,
1938 * it is possible that this is called when physical memory is scarce causing
1939 * OOM killer to be triggered. We should avoid doing so until an actual
1940 * allocation causes the failure as it is possible that requests can be
1941 * serviced from already backed regions.
1942 */
1943static void __pcpu_balance_workfn(enum pcpu_chunk_type type)
1944{
1945 /* gfp flags passed to underlying allocators */
1946 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
1947 LIST_HEAD(to_free);
1948 struct list_head *pcpu_slot = pcpu_chunk_list(type);
1949 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1950 struct pcpu_chunk *chunk, *next;
1951 int slot, nr_to_pop, ret;
1952
1953 /*
1954 * There's no reason to keep around multiple unused chunks and VM
1955 * areas can be scarce. Destroy all free chunks except for one.
1956 */
1957 mutex_lock(&pcpu_alloc_mutex);
1958 spin_lock_irq(&pcpu_lock);
1959
1960 list_for_each_entry_safe(chunk, next, free_head, list) {
1961 WARN_ON(chunk->immutable);
1962
1963 /* spare the first one */
1964 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1965 continue;
1966
1967 list_move(&chunk->list, &to_free);
1968 }
1969
1970 spin_unlock_irq(&pcpu_lock);
1971
1972 list_for_each_entry_safe(chunk, next, &to_free, list) {
1973 unsigned int rs, re;
1974
1975 bitmap_for_each_set_region(chunk->populated, rs, re, 0,
1976 chunk->nr_pages) {
1977 pcpu_depopulate_chunk(chunk, rs, re);
1978 spin_lock_irq(&pcpu_lock);
1979 pcpu_chunk_depopulated(chunk, rs, re);
1980 spin_unlock_irq(&pcpu_lock);
1981 }
1982 pcpu_destroy_chunk(chunk);
1983 cond_resched();
1984 }
1985
1986 /*
1987 * Ensure there are certain number of free populated pages for
1988 * atomic allocs. Fill up from the most packed so that atomic
1989 * allocs don't increase fragmentation. If atomic allocation
1990 * failed previously, always populate the maximum amount. This
1991 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1992 * failing indefinitely; however, large atomic allocs are not
1993 * something we support properly and can be highly unreliable and
1994 * inefficient.
1995 */
1996retry_pop:
1997 if (pcpu_atomic_alloc_failed) {
1998 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1999 /* best effort anyway, don't worry about synchronization */
2000 pcpu_atomic_alloc_failed = false;
2001 } else {
2002 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2003 pcpu_nr_empty_pop_pages,
2004 0, PCPU_EMPTY_POP_PAGES_HIGH);
2005 }
2006
2007 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
2008 unsigned int nr_unpop = 0, rs, re;
2009
2010 if (!nr_to_pop)
2011 break;
2012
2013 spin_lock_irq(&pcpu_lock);
2014 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
2015 nr_unpop = chunk->nr_pages - chunk->nr_populated;
2016 if (nr_unpop)
2017 break;
2018 }
2019 spin_unlock_irq(&pcpu_lock);
2020
2021 if (!nr_unpop)
2022 continue;
2023
2024 /* @chunk can't go away while pcpu_alloc_mutex is held */
2025 bitmap_for_each_clear_region(chunk->populated, rs, re, 0,
2026 chunk->nr_pages) {
2027 int nr = min_t(int, re - rs, nr_to_pop);
2028
2029 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2030 if (!ret) {
2031 nr_to_pop -= nr;
2032 spin_lock_irq(&pcpu_lock);
2033 pcpu_chunk_populated(chunk, rs, rs + nr);
2034 spin_unlock_irq(&pcpu_lock);
2035 } else {
2036 nr_to_pop = 0;
2037 }
2038
2039 if (!nr_to_pop)
2040 break;
2041 }
2042 }
2043
2044 if (nr_to_pop) {
2045 /* ran out of chunks to populate, create a new one and retry */
2046 chunk = pcpu_create_chunk(type, gfp);
2047 if (chunk) {
2048 spin_lock_irq(&pcpu_lock);
2049 pcpu_chunk_relocate(chunk, -1);
2050 spin_unlock_irq(&pcpu_lock);
2051 goto retry_pop;
2052 }
2053 }
2054
2055 mutex_unlock(&pcpu_alloc_mutex);
2056}
2057
2058/**
2059 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2060 * @work: unused
2061 *
2062 * Call __pcpu_balance_workfn() for each chunk type.
2063 */
2064static void pcpu_balance_workfn(struct work_struct *work)
2065{
2066 enum pcpu_chunk_type type;
2067
2068 for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++)
2069 __pcpu_balance_workfn(type);
2070}
2071
2072/**
2073 * free_percpu - free percpu area
2074 * @ptr: pointer to area to free
2075 *
2076 * Free percpu area @ptr.
2077 *
2078 * CONTEXT:
2079 * Can be called from atomic context.
2080 */
2081void free_percpu(void __percpu *ptr)
2082{
2083 void *addr;
2084 struct pcpu_chunk *chunk;
2085 unsigned long flags;
2086 int size, off;
2087 bool need_balance = false;
2088 struct list_head *pcpu_slot;
2089
2090 if (!ptr)
2091 return;
2092
2093 kmemleak_free_percpu(ptr);
2094
2095 addr = __pcpu_ptr_to_addr(ptr);
2096
2097 spin_lock_irqsave(&pcpu_lock, flags);
2098
2099 chunk = pcpu_chunk_addr_search(addr);
2100 off = addr - chunk->base_addr;
2101
2102 size = pcpu_free_area(chunk, off);
2103
2104 pcpu_slot = pcpu_chunk_list(pcpu_chunk_type(chunk));
2105
2106 pcpu_memcg_free_hook(chunk, off, size);
2107
2108 /* if there are more than one fully free chunks, wake up grim reaper */
2109 if (chunk->free_bytes == pcpu_unit_size) {
2110 struct pcpu_chunk *pos;
2111
2112 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
2113 if (pos != chunk) {
2114 need_balance = true;
2115 break;
2116 }
2117 }
2118
2119 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2120
2121 spin_unlock_irqrestore(&pcpu_lock, flags);
2122
2123 if (need_balance)
2124 pcpu_schedule_balance_work();
2125}
2126EXPORT_SYMBOL_GPL(free_percpu);
2127
2128bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2129{
2130#ifdef CONFIG_SMP
2131 const size_t static_size = __per_cpu_end - __per_cpu_start;
2132 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2133 unsigned int cpu;
2134
2135 for_each_possible_cpu(cpu) {
2136 void *start = per_cpu_ptr(base, cpu);
2137 void *va = (void *)addr;
2138
2139 if (va >= start && va < start + static_size) {
2140 if (can_addr) {
2141 *can_addr = (unsigned long) (va - start);
2142 *can_addr += (unsigned long)
2143 per_cpu_ptr(base, get_boot_cpu_id());
2144 }
2145 return true;
2146 }
2147 }
2148#endif
2149 /* on UP, can't distinguish from other static vars, always false */
2150 return false;
2151}
2152
2153/**
2154 * is_kernel_percpu_address - test whether address is from static percpu area
2155 * @addr: address to test
2156 *
2157 * Test whether @addr belongs to in-kernel static percpu area. Module
2158 * static percpu areas are not considered. For those, use
2159 * is_module_percpu_address().
2160 *
2161 * RETURNS:
2162 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2163 */
2164bool is_kernel_percpu_address(unsigned long addr)
2165{
2166 return __is_kernel_percpu_address(addr, NULL);
2167}
2168
2169/**
2170 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2171 * @addr: the address to be converted to physical address
2172 *
2173 * Given @addr which is dereferenceable address obtained via one of
2174 * percpu access macros, this function translates it into its physical
2175 * address. The caller is responsible for ensuring @addr stays valid
2176 * until this function finishes.
2177 *
2178 * percpu allocator has special setup for the first chunk, which currently
2179 * supports either embedding in linear address space or vmalloc mapping,
2180 * and, from the second one, the backing allocator (currently either vm or
2181 * km) provides translation.
2182 *
2183 * The addr can be translated simply without checking if it falls into the
2184 * first chunk. But the current code reflects better how percpu allocator
2185 * actually works, and the verification can discover both bugs in percpu
2186 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2187 * code.
2188 *
2189 * RETURNS:
2190 * The physical address for @addr.
2191 */
2192phys_addr_t per_cpu_ptr_to_phys(void *addr)
2193{
2194 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2195 bool in_first_chunk = false;
2196 unsigned long first_low, first_high;
2197 unsigned int cpu;
2198
2199 /*
2200 * The following test on unit_low/high isn't strictly
2201 * necessary but will speed up lookups of addresses which
2202 * aren't in the first chunk.
2203 *
2204 * The address check is against full chunk sizes. pcpu_base_addr
2205 * points to the beginning of the first chunk including the
2206 * static region. Assumes good intent as the first chunk may
2207 * not be full (ie. < pcpu_unit_pages in size).
2208 */
2209 first_low = (unsigned long)pcpu_base_addr +
2210 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2211 first_high = (unsigned long)pcpu_base_addr +
2212 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2213 if ((unsigned long)addr >= first_low &&
2214 (unsigned long)addr < first_high) {
2215 for_each_possible_cpu(cpu) {
2216 void *start = per_cpu_ptr(base, cpu);
2217
2218 if (addr >= start && addr < start + pcpu_unit_size) {
2219 in_first_chunk = true;
2220 break;
2221 }
2222 }
2223 }
2224
2225 if (in_first_chunk) {
2226 if (!is_vmalloc_addr(addr))
2227 return __pa(addr);
2228 else
2229 return page_to_phys(vmalloc_to_page(addr)) +
2230 offset_in_page(addr);
2231 } else
2232 return page_to_phys(pcpu_addr_to_page(addr)) +
2233 offset_in_page(addr);
2234}
2235
2236/**
2237 * pcpu_alloc_alloc_info - allocate percpu allocation info
2238 * @nr_groups: the number of groups
2239 * @nr_units: the number of units
2240 *
2241 * Allocate ai which is large enough for @nr_groups groups containing
2242 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2243 * cpu_map array which is long enough for @nr_units and filled with
2244 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2245 * pointer of other groups.
2246 *
2247 * RETURNS:
2248 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2249 * failure.
2250 */
2251struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2252 int nr_units)
2253{
2254 struct pcpu_alloc_info *ai;
2255 size_t base_size, ai_size;
2256 void *ptr;
2257 int unit;
2258
2259 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2260 __alignof__(ai->groups[0].cpu_map[0]));
2261 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2262
2263 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2264 if (!ptr)
2265 return NULL;
2266 ai = ptr;
2267 ptr += base_size;
2268
2269 ai->groups[0].cpu_map = ptr;
2270
2271 for (unit = 0; unit < nr_units; unit++)
2272 ai->groups[0].cpu_map[unit] = NR_CPUS;
2273
2274 ai->nr_groups = nr_groups;
2275 ai->__ai_size = PFN_ALIGN(ai_size);
2276
2277 return ai;
2278}
2279
2280/**
2281 * pcpu_free_alloc_info - free percpu allocation info
2282 * @ai: pcpu_alloc_info to free
2283 *
2284 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2285 */
2286void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2287{
2288 memblock_free_early(__pa(ai), ai->__ai_size);
2289}
2290
2291/**
2292 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2293 * @lvl: loglevel
2294 * @ai: allocation info to dump
2295 *
2296 * Print out information about @ai using loglevel @lvl.
2297 */
2298static void pcpu_dump_alloc_info(const char *lvl,
2299 const struct pcpu_alloc_info *ai)
2300{
2301 int group_width = 1, cpu_width = 1, width;
2302 char empty_str[] = "--------";
2303 int alloc = 0, alloc_end = 0;
2304 int group, v;
2305 int upa, apl; /* units per alloc, allocs per line */
2306
2307 v = ai->nr_groups;
2308 while (v /= 10)
2309 group_width++;
2310
2311 v = num_possible_cpus();
2312 while (v /= 10)
2313 cpu_width++;
2314 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2315
2316 upa = ai->alloc_size / ai->unit_size;
2317 width = upa * (cpu_width + 1) + group_width + 3;
2318 apl = rounddown_pow_of_two(max(60 / width, 1));
2319
2320 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2321 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2322 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2323
2324 for (group = 0; group < ai->nr_groups; group++) {
2325 const struct pcpu_group_info *gi = &ai->groups[group];
2326 int unit = 0, unit_end = 0;
2327
2328 BUG_ON(gi->nr_units % upa);
2329 for (alloc_end += gi->nr_units / upa;
2330 alloc < alloc_end; alloc++) {
2331 if (!(alloc % apl)) {
2332 pr_cont("\n");
2333 printk("%spcpu-alloc: ", lvl);
2334 }
2335 pr_cont("[%0*d] ", group_width, group);
2336
2337 for (unit_end += upa; unit < unit_end; unit++)
2338 if (gi->cpu_map[unit] != NR_CPUS)
2339 pr_cont("%0*d ",
2340 cpu_width, gi->cpu_map[unit]);
2341 else
2342 pr_cont("%s ", empty_str);
2343 }
2344 }
2345 pr_cont("\n");
2346}
2347
2348/**
2349 * pcpu_setup_first_chunk - initialize the first percpu chunk
2350 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2351 * @base_addr: mapped address
2352 *
2353 * Initialize the first percpu chunk which contains the kernel static
2354 * percpu area. This function is to be called from arch percpu area
2355 * setup path.
2356 *
2357 * @ai contains all information necessary to initialize the first
2358 * chunk and prime the dynamic percpu allocator.
2359 *
2360 * @ai->static_size is the size of static percpu area.
2361 *
2362 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2363 * reserve after the static area in the first chunk. This reserves
2364 * the first chunk such that it's available only through reserved
2365 * percpu allocation. This is primarily used to serve module percpu
2366 * static areas on architectures where the addressing model has
2367 * limited offset range for symbol relocations to guarantee module
2368 * percpu symbols fall inside the relocatable range.
2369 *
2370 * @ai->dyn_size determines the number of bytes available for dynamic
2371 * allocation in the first chunk. The area between @ai->static_size +
2372 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2373 *
2374 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2375 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2376 * @ai->dyn_size.
2377 *
2378 * @ai->atom_size is the allocation atom size and used as alignment
2379 * for vm areas.
2380 *
2381 * @ai->alloc_size is the allocation size and always multiple of
2382 * @ai->atom_size. This is larger than @ai->atom_size if
2383 * @ai->unit_size is larger than @ai->atom_size.
2384 *
2385 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2386 * percpu areas. Units which should be colocated are put into the
2387 * same group. Dynamic VM areas will be allocated according to these
2388 * groupings. If @ai->nr_groups is zero, a single group containing
2389 * all units is assumed.
2390 *
2391 * The caller should have mapped the first chunk at @base_addr and
2392 * copied static data to each unit.
2393 *
2394 * The first chunk will always contain a static and a dynamic region.
2395 * However, the static region is not managed by any chunk. If the first
2396 * chunk also contains a reserved region, it is served by two chunks -
2397 * one for the reserved region and one for the dynamic region. They
2398 * share the same vm, but use offset regions in the area allocation map.
2399 * The chunk serving the dynamic region is circulated in the chunk slots
2400 * and available for dynamic allocation like any other chunk.
2401 */
2402void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2403 void *base_addr)
2404{
2405 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2406 size_t static_size, dyn_size;
2407 struct pcpu_chunk *chunk;
2408 unsigned long *group_offsets;
2409 size_t *group_sizes;
2410 unsigned long *unit_off;
2411 unsigned int cpu;
2412 int *unit_map;
2413 int group, unit, i;
2414 int map_size;
2415 unsigned long tmp_addr;
2416 size_t alloc_size;
2417 enum pcpu_chunk_type type;
2418
2419#define PCPU_SETUP_BUG_ON(cond) do { \
2420 if (unlikely(cond)) { \
2421 pr_emerg("failed to initialize, %s\n", #cond); \
2422 pr_emerg("cpu_possible_mask=%*pb\n", \
2423 cpumask_pr_args(cpu_possible_mask)); \
2424 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2425 BUG(); \
2426 } \
2427} while (0)
2428
2429 /* sanity checks */
2430 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2431#ifdef CONFIG_SMP
2432 PCPU_SETUP_BUG_ON(!ai->static_size);
2433 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2434#endif
2435 PCPU_SETUP_BUG_ON(!base_addr);
2436 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2437 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2438 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2439 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2440 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2441 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2442 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2443 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2444 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2445 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2446 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2447
2448 /* process group information and build config tables accordingly */
2449 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2450 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2451 if (!group_offsets)
2452 panic("%s: Failed to allocate %zu bytes\n", __func__,
2453 alloc_size);
2454
2455 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2456 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2457 if (!group_sizes)
2458 panic("%s: Failed to allocate %zu bytes\n", __func__,
2459 alloc_size);
2460
2461 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2462 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2463 if (!unit_map)
2464 panic("%s: Failed to allocate %zu bytes\n", __func__,
2465 alloc_size);
2466
2467 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2468 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2469 if (!unit_off)
2470 panic("%s: Failed to allocate %zu bytes\n", __func__,
2471 alloc_size);
2472
2473 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2474 unit_map[cpu] = UINT_MAX;
2475
2476 pcpu_low_unit_cpu = NR_CPUS;
2477 pcpu_high_unit_cpu = NR_CPUS;
2478
2479 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2480 const struct pcpu_group_info *gi = &ai->groups[group];
2481
2482 group_offsets[group] = gi->base_offset;
2483 group_sizes[group] = gi->nr_units * ai->unit_size;
2484
2485 for (i = 0; i < gi->nr_units; i++) {
2486 cpu = gi->cpu_map[i];
2487 if (cpu == NR_CPUS)
2488 continue;
2489
2490 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2491 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2492 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2493
2494 unit_map[cpu] = unit + i;
2495 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2496
2497 /* determine low/high unit_cpu */
2498 if (pcpu_low_unit_cpu == NR_CPUS ||
2499 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2500 pcpu_low_unit_cpu = cpu;
2501 if (pcpu_high_unit_cpu == NR_CPUS ||
2502 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2503 pcpu_high_unit_cpu = cpu;
2504 }
2505 }
2506 pcpu_nr_units = unit;
2507
2508 for_each_possible_cpu(cpu)
2509 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2510
2511 /* we're done parsing the input, undefine BUG macro and dump config */
2512#undef PCPU_SETUP_BUG_ON
2513 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2514
2515 pcpu_nr_groups = ai->nr_groups;
2516 pcpu_group_offsets = group_offsets;
2517 pcpu_group_sizes = group_sizes;
2518 pcpu_unit_map = unit_map;
2519 pcpu_unit_offsets = unit_off;
2520
2521 /* determine basic parameters */
2522 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2523 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2524 pcpu_atom_size = ai->atom_size;
2525 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2526 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2527
2528 pcpu_stats_save_ai(ai);
2529
2530 /*
2531 * Allocate chunk slots. The additional last slot is for
2532 * empty chunks.
2533 */
2534 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2535 pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2536 sizeof(pcpu_chunk_lists[0]) *
2537 PCPU_NR_CHUNK_TYPES,
2538 SMP_CACHE_BYTES);
2539 if (!pcpu_chunk_lists)
2540 panic("%s: Failed to allocate %zu bytes\n", __func__,
2541 pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]) *
2542 PCPU_NR_CHUNK_TYPES);
2543
2544 for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++)
2545 for (i = 0; i < pcpu_nr_slots; i++)
2546 INIT_LIST_HEAD(&pcpu_chunk_list(type)[i]);
2547
2548 /*
2549 * The end of the static region needs to be aligned with the
2550 * minimum allocation size as this offsets the reserved and
2551 * dynamic region. The first chunk ends page aligned by
2552 * expanding the dynamic region, therefore the dynamic region
2553 * can be shrunk to compensate while still staying above the
2554 * configured sizes.
2555 */
2556 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2557 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2558
2559 /*
2560 * Initialize first chunk.
2561 * If the reserved_size is non-zero, this initializes the reserved
2562 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2563 * and the dynamic region is initialized here. The first chunk,
2564 * pcpu_first_chunk, will always point to the chunk that serves
2565 * the dynamic region.
2566 */
2567 tmp_addr = (unsigned long)base_addr + static_size;
2568 map_size = ai->reserved_size ?: dyn_size;
2569 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2570
2571 /* init dynamic chunk if necessary */
2572 if (ai->reserved_size) {
2573 pcpu_reserved_chunk = chunk;
2574
2575 tmp_addr = (unsigned long)base_addr + static_size +
2576 ai->reserved_size;
2577 map_size = dyn_size;
2578 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2579 }
2580
2581 /* link the first chunk in */
2582 pcpu_first_chunk = chunk;
2583 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2584 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2585
2586 /* include all regions of the first chunk */
2587 pcpu_nr_populated += PFN_DOWN(size_sum);
2588
2589 pcpu_stats_chunk_alloc();
2590 trace_percpu_create_chunk(base_addr);
2591
2592 /* we're done */
2593 pcpu_base_addr = base_addr;
2594}
2595
2596#ifdef CONFIG_SMP
2597
2598const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2599 [PCPU_FC_AUTO] = "auto",
2600 [PCPU_FC_EMBED] = "embed",
2601 [PCPU_FC_PAGE] = "page",
2602};
2603
2604enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2605
2606static int __init percpu_alloc_setup(char *str)
2607{
2608 if (!str)
2609 return -EINVAL;
2610
2611 if (0)
2612 /* nada */;
2613#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2614 else if (!strcmp(str, "embed"))
2615 pcpu_chosen_fc = PCPU_FC_EMBED;
2616#endif
2617#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2618 else if (!strcmp(str, "page"))
2619 pcpu_chosen_fc = PCPU_FC_PAGE;
2620#endif
2621 else
2622 pr_warn("unknown allocator %s specified\n", str);
2623
2624 return 0;
2625}
2626early_param("percpu_alloc", percpu_alloc_setup);
2627
2628/*
2629 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2630 * Build it if needed by the arch config or the generic setup is going
2631 * to be used.
2632 */
2633#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2634 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2635#define BUILD_EMBED_FIRST_CHUNK
2636#endif
2637
2638/* build pcpu_page_first_chunk() iff needed by the arch config */
2639#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2640#define BUILD_PAGE_FIRST_CHUNK
2641#endif
2642
2643/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2644#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2645/**
2646 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2647 * @reserved_size: the size of reserved percpu area in bytes
2648 * @dyn_size: minimum free size for dynamic allocation in bytes
2649 * @atom_size: allocation atom size
2650 * @cpu_distance_fn: callback to determine distance between cpus, optional
2651 *
2652 * This function determines grouping of units, their mappings to cpus
2653 * and other parameters considering needed percpu size, allocation
2654 * atom size and distances between CPUs.
2655 *
2656 * Groups are always multiples of atom size and CPUs which are of
2657 * LOCAL_DISTANCE both ways are grouped together and share space for
2658 * units in the same group. The returned configuration is guaranteed
2659 * to have CPUs on different nodes on different groups and >=75% usage
2660 * of allocated virtual address space.
2661 *
2662 * RETURNS:
2663 * On success, pointer to the new allocation_info is returned. On
2664 * failure, ERR_PTR value is returned.
2665 */
2666static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2667 size_t reserved_size, size_t dyn_size,
2668 size_t atom_size,
2669 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2670{
2671 static int group_map[NR_CPUS] __initdata;
2672 static int group_cnt[NR_CPUS] __initdata;
2673 const size_t static_size = __per_cpu_end - __per_cpu_start;
2674 int nr_groups = 1, nr_units = 0;
2675 size_t size_sum, min_unit_size, alloc_size;
2676 int upa, max_upa, best_upa; /* units_per_alloc */
2677 int last_allocs, group, unit;
2678 unsigned int cpu, tcpu;
2679 struct pcpu_alloc_info *ai;
2680 unsigned int *cpu_map;
2681
2682 /* this function may be called multiple times */
2683 memset(group_map, 0, sizeof(group_map));
2684 memset(group_cnt, 0, sizeof(group_cnt));
2685
2686 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2687 size_sum = PFN_ALIGN(static_size + reserved_size +
2688 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2689 dyn_size = size_sum - static_size - reserved_size;
2690
2691 /*
2692 * Determine min_unit_size, alloc_size and max_upa such that
2693 * alloc_size is multiple of atom_size and is the smallest
2694 * which can accommodate 4k aligned segments which are equal to
2695 * or larger than min_unit_size.
2696 */
2697 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2698
2699 /* determine the maximum # of units that can fit in an allocation */
2700 alloc_size = roundup(min_unit_size, atom_size);
2701 upa = alloc_size / min_unit_size;
2702 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2703 upa--;
2704 max_upa = upa;
2705
2706 /* group cpus according to their proximity */
2707 for_each_possible_cpu(cpu) {
2708 group = 0;
2709 next_group:
2710 for_each_possible_cpu(tcpu) {
2711 if (cpu == tcpu)
2712 break;
2713 if (group_map[tcpu] == group && cpu_distance_fn &&
2714 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2715 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2716 group++;
2717 nr_groups = max(nr_groups, group + 1);
2718 goto next_group;
2719 }
2720 }
2721 group_map[cpu] = group;
2722 group_cnt[group]++;
2723 }
2724
2725 /*
2726 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2727 * Expand the unit_size until we use >= 75% of the units allocated.
2728 * Related to atom_size, which could be much larger than the unit_size.
2729 */
2730 last_allocs = INT_MAX;
2731 for (upa = max_upa; upa; upa--) {
2732 int allocs = 0, wasted = 0;
2733
2734 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2735 continue;
2736
2737 for (group = 0; group < nr_groups; group++) {
2738 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2739 allocs += this_allocs;
2740 wasted += this_allocs * upa - group_cnt[group];
2741 }
2742
2743 /*
2744 * Don't accept if wastage is over 1/3. The
2745 * greater-than comparison ensures upa==1 always
2746 * passes the following check.
2747 */
2748 if (wasted > num_possible_cpus() / 3)
2749 continue;
2750
2751 /* and then don't consume more memory */
2752 if (allocs > last_allocs)
2753 break;
2754 last_allocs = allocs;
2755 best_upa = upa;
2756 }
2757 upa = best_upa;
2758
2759 /* allocate and fill alloc_info */
2760 for (group = 0; group < nr_groups; group++)
2761 nr_units += roundup(group_cnt[group], upa);
2762
2763 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2764 if (!ai)
2765 return ERR_PTR(-ENOMEM);
2766 cpu_map = ai->groups[0].cpu_map;
2767
2768 for (group = 0; group < nr_groups; group++) {
2769 ai->groups[group].cpu_map = cpu_map;
2770 cpu_map += roundup(group_cnt[group], upa);
2771 }
2772
2773 ai->static_size = static_size;
2774 ai->reserved_size = reserved_size;
2775 ai->dyn_size = dyn_size;
2776 ai->unit_size = alloc_size / upa;
2777 ai->atom_size = atom_size;
2778 ai->alloc_size = alloc_size;
2779
2780 for (group = 0, unit = 0; group < nr_groups; group++) {
2781 struct pcpu_group_info *gi = &ai->groups[group];
2782
2783 /*
2784 * Initialize base_offset as if all groups are located
2785 * back-to-back. The caller should update this to
2786 * reflect actual allocation.
2787 */
2788 gi->base_offset = unit * ai->unit_size;
2789
2790 for_each_possible_cpu(cpu)
2791 if (group_map[cpu] == group)
2792 gi->cpu_map[gi->nr_units++] = cpu;
2793 gi->nr_units = roundup(gi->nr_units, upa);
2794 unit += gi->nr_units;
2795 }
2796 BUG_ON(unit != nr_units);
2797
2798 return ai;
2799}
2800#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2801
2802#if defined(BUILD_EMBED_FIRST_CHUNK)
2803/**
2804 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2805 * @reserved_size: the size of reserved percpu area in bytes
2806 * @dyn_size: minimum free size for dynamic allocation in bytes
2807 * @atom_size: allocation atom size
2808 * @cpu_distance_fn: callback to determine distance between cpus, optional
2809 * @alloc_fn: function to allocate percpu page
2810 * @free_fn: function to free percpu page
2811 *
2812 * This is a helper to ease setting up embedded first percpu chunk and
2813 * can be called where pcpu_setup_first_chunk() is expected.
2814 *
2815 * If this function is used to setup the first chunk, it is allocated
2816 * by calling @alloc_fn and used as-is without being mapped into
2817 * vmalloc area. Allocations are always whole multiples of @atom_size
2818 * aligned to @atom_size.
2819 *
2820 * This enables the first chunk to piggy back on the linear physical
2821 * mapping which often uses larger page size. Please note that this
2822 * can result in very sparse cpu->unit mapping on NUMA machines thus
2823 * requiring large vmalloc address space. Don't use this allocator if
2824 * vmalloc space is not orders of magnitude larger than distances
2825 * between node memory addresses (ie. 32bit NUMA machines).
2826 *
2827 * @dyn_size specifies the minimum dynamic area size.
2828 *
2829 * If the needed size is smaller than the minimum or specified unit
2830 * size, the leftover is returned using @free_fn.
2831 *
2832 * RETURNS:
2833 * 0 on success, -errno on failure.
2834 */
2835int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2836 size_t atom_size,
2837 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2838 pcpu_fc_alloc_fn_t alloc_fn,
2839 pcpu_fc_free_fn_t free_fn)
2840{
2841 void *base = (void *)ULONG_MAX;
2842 void **areas = NULL;
2843 struct pcpu_alloc_info *ai;
2844 size_t size_sum, areas_size;
2845 unsigned long max_distance;
2846 int group, i, highest_group, rc = 0;
2847
2848 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2849 cpu_distance_fn);
2850 if (IS_ERR(ai))
2851 return PTR_ERR(ai);
2852
2853 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2854 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2855
2856 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
2857 if (!areas) {
2858 rc = -ENOMEM;
2859 goto out_free;
2860 }
2861
2862 /* allocate, copy and determine base address & max_distance */
2863 highest_group = 0;
2864 for (group = 0; group < ai->nr_groups; group++) {
2865 struct pcpu_group_info *gi = &ai->groups[group];
2866 unsigned int cpu = NR_CPUS;
2867 void *ptr;
2868
2869 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2870 cpu = gi->cpu_map[i];
2871 BUG_ON(cpu == NR_CPUS);
2872
2873 /* allocate space for the whole group */
2874 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2875 if (!ptr) {
2876 rc = -ENOMEM;
2877 goto out_free_areas;
2878 }
2879 /* kmemleak tracks the percpu allocations separately */
2880 kmemleak_free(ptr);
2881 areas[group] = ptr;
2882
2883 base = min(ptr, base);
2884 if (ptr > areas[highest_group])
2885 highest_group = group;
2886 }
2887 max_distance = areas[highest_group] - base;
2888 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2889
2890 /* warn if maximum distance is further than 75% of vmalloc space */
2891 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2892 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2893 max_distance, VMALLOC_TOTAL);
2894#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2895 /* and fail if we have fallback */
2896 rc = -EINVAL;
2897 goto out_free_areas;
2898#endif
2899 }
2900
2901 /*
2902 * Copy data and free unused parts. This should happen after all
2903 * allocations are complete; otherwise, we may end up with
2904 * overlapping groups.
2905 */
2906 for (group = 0; group < ai->nr_groups; group++) {
2907 struct pcpu_group_info *gi = &ai->groups[group];
2908 void *ptr = areas[group];
2909
2910 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2911 if (gi->cpu_map[i] == NR_CPUS) {
2912 /* unused unit, free whole */
2913 free_fn(ptr, ai->unit_size);
2914 continue;
2915 }
2916 /* copy and return the unused part */
2917 memcpy(ptr, __per_cpu_load, ai->static_size);
2918 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2919 }
2920 }
2921
2922 /* base address is now known, determine group base offsets */
2923 for (group = 0; group < ai->nr_groups; group++) {
2924 ai->groups[group].base_offset = areas[group] - base;
2925 }
2926
2927 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2928 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
2929 ai->dyn_size, ai->unit_size);
2930
2931 pcpu_setup_first_chunk(ai, base);
2932 goto out_free;
2933
2934out_free_areas:
2935 for (group = 0; group < ai->nr_groups; group++)
2936 if (areas[group])
2937 free_fn(areas[group],
2938 ai->groups[group].nr_units * ai->unit_size);
2939out_free:
2940 pcpu_free_alloc_info(ai);
2941 if (areas)
2942 memblock_free_early(__pa(areas), areas_size);
2943 return rc;
2944}
2945#endif /* BUILD_EMBED_FIRST_CHUNK */
2946
2947#ifdef BUILD_PAGE_FIRST_CHUNK
2948/**
2949 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2950 * @reserved_size: the size of reserved percpu area in bytes
2951 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2952 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2953 * @populate_pte_fn: function to populate pte
2954 *
2955 * This is a helper to ease setting up page-remapped first percpu
2956 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2957 *
2958 * This is the basic allocator. Static percpu area is allocated
2959 * page-by-page into vmalloc area.
2960 *
2961 * RETURNS:
2962 * 0 on success, -errno on failure.
2963 */
2964int __init pcpu_page_first_chunk(size_t reserved_size,
2965 pcpu_fc_alloc_fn_t alloc_fn,
2966 pcpu_fc_free_fn_t free_fn,
2967 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2968{
2969 static struct vm_struct vm;
2970 struct pcpu_alloc_info *ai;
2971 char psize_str[16];
2972 int unit_pages;
2973 size_t pages_size;
2974 struct page **pages;
2975 int unit, i, j, rc = 0;
2976 int upa;
2977 int nr_g0_units;
2978
2979 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2980
2981 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2982 if (IS_ERR(ai))
2983 return PTR_ERR(ai);
2984 BUG_ON(ai->nr_groups != 1);
2985 upa = ai->alloc_size/ai->unit_size;
2986 nr_g0_units = roundup(num_possible_cpus(), upa);
2987 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
2988 pcpu_free_alloc_info(ai);
2989 return -EINVAL;
2990 }
2991
2992 unit_pages = ai->unit_size >> PAGE_SHIFT;
2993
2994 /* unaligned allocations can't be freed, round up to page size */
2995 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2996 sizeof(pages[0]));
2997 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
2998 if (!pages)
2999 panic("%s: Failed to allocate %zu bytes\n", __func__,
3000 pages_size);
3001
3002 /* allocate pages */
3003 j = 0;
3004 for (unit = 0; unit < num_possible_cpus(); unit++) {
3005 unsigned int cpu = ai->groups[0].cpu_map[unit];
3006 for (i = 0; i < unit_pages; i++) {
3007 void *ptr;
3008
3009 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
3010 if (!ptr) {
3011 pr_warn("failed to allocate %s page for cpu%u\n",
3012 psize_str, cpu);
3013 goto enomem;
3014 }
3015 /* kmemleak tracks the percpu allocations separately */
3016 kmemleak_free(ptr);
3017 pages[j++] = virt_to_page(ptr);
3018 }
3019 }
3020
3021 /* allocate vm area, map the pages and copy static data */
3022 vm.flags = VM_ALLOC;
3023 vm.size = num_possible_cpus() * ai->unit_size;
3024 vm_area_register_early(&vm, PAGE_SIZE);
3025
3026 for (unit = 0; unit < num_possible_cpus(); unit++) {
3027 unsigned long unit_addr =
3028 (unsigned long)vm.addr + unit * ai->unit_size;
3029
3030 for (i = 0; i < unit_pages; i++)
3031 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
3032
3033 /* pte already populated, the following shouldn't fail */
3034 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3035 unit_pages);
3036 if (rc < 0)
3037 panic("failed to map percpu area, err=%d\n", rc);
3038
3039 /*
3040 * FIXME: Archs with virtual cache should flush local
3041 * cache for the linear mapping here - something
3042 * equivalent to flush_cache_vmap() on the local cpu.
3043 * flush_cache_vmap() can't be used as most supporting
3044 * data structures are not set up yet.
3045 */
3046
3047 /* copy static data */
3048 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3049 }
3050
3051 /* we're ready, commit */
3052 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3053 unit_pages, psize_str, ai->static_size,
3054 ai->reserved_size, ai->dyn_size);
3055
3056 pcpu_setup_first_chunk(ai, vm.addr);
3057 goto out_free_ar;
3058
3059enomem:
3060 while (--j >= 0)
3061 free_fn(page_address(pages[j]), PAGE_SIZE);
3062 rc = -ENOMEM;
3063out_free_ar:
3064 memblock_free_early(__pa(pages), pages_size);
3065 pcpu_free_alloc_info(ai);
3066 return rc;
3067}
3068#endif /* BUILD_PAGE_FIRST_CHUNK */
3069
3070#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3071/*
3072 * Generic SMP percpu area setup.
3073 *
3074 * The embedding helper is used because its behavior closely resembles
3075 * the original non-dynamic generic percpu area setup. This is
3076 * important because many archs have addressing restrictions and might
3077 * fail if the percpu area is located far away from the previous
3078 * location. As an added bonus, in non-NUMA cases, embedding is
3079 * generally a good idea TLB-wise because percpu area can piggy back
3080 * on the physical linear memory mapping which uses large page
3081 * mappings on applicable archs.
3082 */
3083unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3084EXPORT_SYMBOL(__per_cpu_offset);
3085
3086static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
3087 size_t align)
3088{
3089 return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS));
3090}
3091
3092static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
3093{
3094 memblock_free_early(__pa(ptr), size);
3095}
3096
3097void __init setup_per_cpu_areas(void)
3098{
3099 unsigned long delta;
3100 unsigned int cpu;
3101 int rc;
3102
3103 /*
3104 * Always reserve area for module percpu variables. That's
3105 * what the legacy allocator did.
3106 */
3107 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
3108 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
3109 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
3110 if (rc < 0)
3111 panic("Failed to initialize percpu areas.");
3112
3113 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3114 for_each_possible_cpu(cpu)
3115 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3116}
3117#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3118
3119#else /* CONFIG_SMP */
3120
3121/*
3122 * UP percpu area setup.
3123 *
3124 * UP always uses km-based percpu allocator with identity mapping.
3125 * Static percpu variables are indistinguishable from the usual static
3126 * variables and don't require any special preparation.
3127 */
3128void __init setup_per_cpu_areas(void)
3129{
3130 const size_t unit_size =
3131 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3132 PERCPU_DYNAMIC_RESERVE));
3133 struct pcpu_alloc_info *ai;
3134 void *fc;
3135
3136 ai = pcpu_alloc_alloc_info(1, 1);
3137 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3138 if (!ai || !fc)
3139 panic("Failed to allocate memory for percpu areas.");
3140 /* kmemleak tracks the percpu allocations separately */
3141 kmemleak_free(fc);
3142
3143 ai->dyn_size = unit_size;
3144 ai->unit_size = unit_size;
3145 ai->atom_size = unit_size;
3146 ai->alloc_size = unit_size;
3147 ai->groups[0].nr_units = 1;
3148 ai->groups[0].cpu_map[0] = 0;
3149
3150 pcpu_setup_first_chunk(ai, fc);
3151 pcpu_free_alloc_info(ai);
3152}
3153
3154#endif /* CONFIG_SMP */
3155
3156/*
3157 * pcpu_nr_pages - calculate total number of populated backing pages
3158 *
3159 * This reflects the number of pages populated to back chunks. Metadata is
3160 * excluded in the number exposed in meminfo as the number of backing pages
3161 * scales with the number of cpus and can quickly outweigh the memory used for
3162 * metadata. It also keeps this calculation nice and simple.
3163 *
3164 * RETURNS:
3165 * Total number of populated backing pages in use by the allocator.
3166 */
3167unsigned long pcpu_nr_pages(void)
3168{
3169 return pcpu_nr_populated * pcpu_nr_units;
3170}
3171
3172/*
3173 * Percpu allocator is initialized early during boot when neither slab or
3174 * workqueue is available. Plug async management until everything is up
3175 * and running.
3176 */
3177static int __init percpu_enable_async(void)
3178{
3179 pcpu_async_enabled = true;
3180 return 0;
3181}
3182subsys_initcall(percpu_enable_async);
1/*
2 * mm/percpu.c - percpu memory allocator
3 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated.
17 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
29 *
30 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
32 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be equal to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
43 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
45 *
46 * To use this allocator, arch code should do the followings.
47 *
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be
50 * different from the default
51 *
52 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
54 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
58#include <linux/err.h>
59#include <linux/list.h>
60#include <linux/log2.h>
61#include <linux/mm.h>
62#include <linux/module.h>
63#include <linux/mutex.h>
64#include <linux/percpu.h>
65#include <linux/pfn.h>
66#include <linux/slab.h>
67#include <linux/spinlock.h>
68#include <linux/vmalloc.h>
69#include <linux/workqueue.h>
70#include <linux/kmemleak.h>
71
72#include <asm/cacheflush.h>
73#include <asm/sections.h>
74#include <asm/tlbflush.h>
75#include <asm/io.h>
76
77#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
78#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
79
80#ifdef CONFIG_SMP
81/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
82#ifndef __addr_to_pcpu_ptr
83#define __addr_to_pcpu_ptr(addr) \
84 (void __percpu *)((unsigned long)(addr) - \
85 (unsigned long)pcpu_base_addr + \
86 (unsigned long)__per_cpu_start)
87#endif
88#ifndef __pcpu_ptr_to_addr
89#define __pcpu_ptr_to_addr(ptr) \
90 (void __force *)((unsigned long)(ptr) + \
91 (unsigned long)pcpu_base_addr - \
92 (unsigned long)__per_cpu_start)
93#endif
94#else /* CONFIG_SMP */
95/* on UP, it's always identity mapped */
96#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
97#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
98#endif /* CONFIG_SMP */
99
100struct pcpu_chunk {
101 struct list_head list; /* linked to pcpu_slot lists */
102 int free_size; /* free bytes in the chunk */
103 int contig_hint; /* max contiguous size hint */
104 void *base_addr; /* base address of this chunk */
105 int map_used; /* # of map entries used before the sentry */
106 int map_alloc; /* # of map entries allocated */
107 int *map; /* allocation map */
108 void *data; /* chunk data */
109 int first_free; /* no free below this */
110 bool immutable; /* no [de]population allowed */
111 unsigned long populated[]; /* populated bitmap */
112};
113
114static int pcpu_unit_pages __read_mostly;
115static int pcpu_unit_size __read_mostly;
116static int pcpu_nr_units __read_mostly;
117static int pcpu_atom_size __read_mostly;
118static int pcpu_nr_slots __read_mostly;
119static size_t pcpu_chunk_struct_size __read_mostly;
120
121/* cpus with the lowest and highest unit addresses */
122static unsigned int pcpu_low_unit_cpu __read_mostly;
123static unsigned int pcpu_high_unit_cpu __read_mostly;
124
125/* the address of the first chunk which starts with the kernel static area */
126void *pcpu_base_addr __read_mostly;
127EXPORT_SYMBOL_GPL(pcpu_base_addr);
128
129static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
130const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
131
132/* group information, used for vm allocation */
133static int pcpu_nr_groups __read_mostly;
134static const unsigned long *pcpu_group_offsets __read_mostly;
135static const size_t *pcpu_group_sizes __read_mostly;
136
137/*
138 * The first chunk which always exists. Note that unlike other
139 * chunks, this one can be allocated and mapped in several different
140 * ways and thus often doesn't live in the vmalloc area.
141 */
142static struct pcpu_chunk *pcpu_first_chunk;
143
144/*
145 * Optional reserved chunk. This chunk reserves part of the first
146 * chunk and serves it for reserved allocations. The amount of
147 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
148 * area doesn't exist, the following variables contain NULL and 0
149 * respectively.
150 */
151static struct pcpu_chunk *pcpu_reserved_chunk;
152static int pcpu_reserved_chunk_limit;
153
154/*
155 * Synchronization rules.
156 *
157 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
158 * protects allocation/reclaim paths, chunks, populated bitmap and
159 * vmalloc mapping. The latter is a spinlock and protects the index
160 * data structures - chunk slots, chunks and area maps in chunks.
161 *
162 * During allocation, pcpu_alloc_mutex is kept locked all the time and
163 * pcpu_lock is grabbed and released as necessary. All actual memory
164 * allocations are done using GFP_KERNEL with pcpu_lock released. In
165 * general, percpu memory can't be allocated with irq off but
166 * irqsave/restore are still used in alloc path so that it can be used
167 * from early init path - sched_init() specifically.
168 *
169 * Free path accesses and alters only the index data structures, so it
170 * can be safely called from atomic context. When memory needs to be
171 * returned to the system, free path schedules reclaim_work which
172 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
173 * reclaimed, release both locks and frees the chunks. Note that it's
174 * necessary to grab both locks to remove a chunk from circulation as
175 * allocation path might be referencing the chunk with only
176 * pcpu_alloc_mutex locked.
177 */
178static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
179static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
180
181static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
182
183/* reclaim work to release fully free chunks, scheduled from free path */
184static void pcpu_reclaim(struct work_struct *work);
185static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
186
187static bool pcpu_addr_in_first_chunk(void *addr)
188{
189 void *first_start = pcpu_first_chunk->base_addr;
190
191 return addr >= first_start && addr < first_start + pcpu_unit_size;
192}
193
194static bool pcpu_addr_in_reserved_chunk(void *addr)
195{
196 void *first_start = pcpu_first_chunk->base_addr;
197
198 return addr >= first_start &&
199 addr < first_start + pcpu_reserved_chunk_limit;
200}
201
202static int __pcpu_size_to_slot(int size)
203{
204 int highbit = fls(size); /* size is in bytes */
205 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
206}
207
208static int pcpu_size_to_slot(int size)
209{
210 if (size == pcpu_unit_size)
211 return pcpu_nr_slots - 1;
212 return __pcpu_size_to_slot(size);
213}
214
215static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
216{
217 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
218 return 0;
219
220 return pcpu_size_to_slot(chunk->free_size);
221}
222
223/* set the pointer to a chunk in a page struct */
224static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
225{
226 page->index = (unsigned long)pcpu;
227}
228
229/* obtain pointer to a chunk from a page struct */
230static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
231{
232 return (struct pcpu_chunk *)page->index;
233}
234
235static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
236{
237 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
238}
239
240static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
241 unsigned int cpu, int page_idx)
242{
243 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
244 (page_idx << PAGE_SHIFT);
245}
246
247static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
248 int *rs, int *re, int end)
249{
250 *rs = find_next_zero_bit(chunk->populated, end, *rs);
251 *re = find_next_bit(chunk->populated, end, *rs + 1);
252}
253
254static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
255 int *rs, int *re, int end)
256{
257 *rs = find_next_bit(chunk->populated, end, *rs);
258 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
259}
260
261/*
262 * (Un)populated page region iterators. Iterate over (un)populated
263 * page regions between @start and @end in @chunk. @rs and @re should
264 * be integer variables and will be set to start and end page index of
265 * the current region.
266 */
267#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
268 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
269 (rs) < (re); \
270 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
271
272#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
274 (rs) < (re); \
275 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
276
277/**
278 * pcpu_mem_zalloc - allocate memory
279 * @size: bytes to allocate
280 *
281 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
282 * kzalloc() is used; otherwise, vzalloc() is used. The returned
283 * memory is always zeroed.
284 *
285 * CONTEXT:
286 * Does GFP_KERNEL allocation.
287 *
288 * RETURNS:
289 * Pointer to the allocated area on success, NULL on failure.
290 */
291static void *pcpu_mem_zalloc(size_t size)
292{
293 if (WARN_ON_ONCE(!slab_is_available()))
294 return NULL;
295
296 if (size <= PAGE_SIZE)
297 return kzalloc(size, GFP_KERNEL);
298 else
299 return vzalloc(size);
300}
301
302/**
303 * pcpu_mem_free - free memory
304 * @ptr: memory to free
305 * @size: size of the area
306 *
307 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
308 */
309static void pcpu_mem_free(void *ptr, size_t size)
310{
311 if (size <= PAGE_SIZE)
312 kfree(ptr);
313 else
314 vfree(ptr);
315}
316
317/**
318 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
319 * @chunk: chunk of interest
320 * @oslot: the previous slot it was on
321 *
322 * This function is called after an allocation or free changed @chunk.
323 * New slot according to the changed state is determined and @chunk is
324 * moved to the slot. Note that the reserved chunk is never put on
325 * chunk slots.
326 *
327 * CONTEXT:
328 * pcpu_lock.
329 */
330static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
331{
332 int nslot = pcpu_chunk_slot(chunk);
333
334 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
335 if (oslot < nslot)
336 list_move(&chunk->list, &pcpu_slot[nslot]);
337 else
338 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
339 }
340}
341
342/**
343 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
344 * @chunk: chunk of interest
345 *
346 * Determine whether area map of @chunk needs to be extended to
347 * accommodate a new allocation.
348 *
349 * CONTEXT:
350 * pcpu_lock.
351 *
352 * RETURNS:
353 * New target map allocation length if extension is necessary, 0
354 * otherwise.
355 */
356static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
357{
358 int new_alloc;
359
360 if (chunk->map_alloc >= chunk->map_used + 3)
361 return 0;
362
363 new_alloc = PCPU_DFL_MAP_ALLOC;
364 while (new_alloc < chunk->map_used + 3)
365 new_alloc *= 2;
366
367 return new_alloc;
368}
369
370/**
371 * pcpu_extend_area_map - extend area map of a chunk
372 * @chunk: chunk of interest
373 * @new_alloc: new target allocation length of the area map
374 *
375 * Extend area map of @chunk to have @new_alloc entries.
376 *
377 * CONTEXT:
378 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
379 *
380 * RETURNS:
381 * 0 on success, -errno on failure.
382 */
383static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
384{
385 int *old = NULL, *new = NULL;
386 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
387 unsigned long flags;
388
389 new = pcpu_mem_zalloc(new_size);
390 if (!new)
391 return -ENOMEM;
392
393 /* acquire pcpu_lock and switch to new area map */
394 spin_lock_irqsave(&pcpu_lock, flags);
395
396 if (new_alloc <= chunk->map_alloc)
397 goto out_unlock;
398
399 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
400 old = chunk->map;
401
402 memcpy(new, old, old_size);
403
404 chunk->map_alloc = new_alloc;
405 chunk->map = new;
406 new = NULL;
407
408out_unlock:
409 spin_unlock_irqrestore(&pcpu_lock, flags);
410
411 /*
412 * pcpu_mem_free() might end up calling vfree() which uses
413 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
414 */
415 pcpu_mem_free(old, old_size);
416 pcpu_mem_free(new, new_size);
417
418 return 0;
419}
420
421/**
422 * pcpu_alloc_area - allocate area from a pcpu_chunk
423 * @chunk: chunk of interest
424 * @size: wanted size in bytes
425 * @align: wanted align
426 *
427 * Try to allocate @size bytes area aligned at @align from @chunk.
428 * Note that this function only allocates the offset. It doesn't
429 * populate or map the area.
430 *
431 * @chunk->map must have at least two free slots.
432 *
433 * CONTEXT:
434 * pcpu_lock.
435 *
436 * RETURNS:
437 * Allocated offset in @chunk on success, -1 if no matching area is
438 * found.
439 */
440static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
441{
442 int oslot = pcpu_chunk_slot(chunk);
443 int max_contig = 0;
444 int i, off;
445 bool seen_free = false;
446 int *p;
447
448 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
449 int head, tail;
450 int this_size;
451
452 off = *p;
453 if (off & 1)
454 continue;
455
456 /* extra for alignment requirement */
457 head = ALIGN(off, align) - off;
458
459 this_size = (p[1] & ~1) - off;
460 if (this_size < head + size) {
461 if (!seen_free) {
462 chunk->first_free = i;
463 seen_free = true;
464 }
465 max_contig = max(this_size, max_contig);
466 continue;
467 }
468
469 /*
470 * If head is small or the previous block is free,
471 * merge'em. Note that 'small' is defined as smaller
472 * than sizeof(int), which is very small but isn't too
473 * uncommon for percpu allocations.
474 */
475 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
476 *p = off += head;
477 if (p[-1] & 1)
478 chunk->free_size -= head;
479 else
480 max_contig = max(*p - p[-1], max_contig);
481 this_size -= head;
482 head = 0;
483 }
484
485 /* if tail is small, just keep it around */
486 tail = this_size - head - size;
487 if (tail < sizeof(int)) {
488 tail = 0;
489 size = this_size - head;
490 }
491
492 /* split if warranted */
493 if (head || tail) {
494 int nr_extra = !!head + !!tail;
495
496 /* insert new subblocks */
497 memmove(p + nr_extra + 1, p + 1,
498 sizeof(chunk->map[0]) * (chunk->map_used - i));
499 chunk->map_used += nr_extra;
500
501 if (head) {
502 if (!seen_free) {
503 chunk->first_free = i;
504 seen_free = true;
505 }
506 *++p = off += head;
507 ++i;
508 max_contig = max(head, max_contig);
509 }
510 if (tail) {
511 p[1] = off + size;
512 max_contig = max(tail, max_contig);
513 }
514 }
515
516 if (!seen_free)
517 chunk->first_free = i + 1;
518
519 /* update hint and mark allocated */
520 if (i + 1 == chunk->map_used)
521 chunk->contig_hint = max_contig; /* fully scanned */
522 else
523 chunk->contig_hint = max(chunk->contig_hint,
524 max_contig);
525
526 chunk->free_size -= size;
527 *p |= 1;
528
529 pcpu_chunk_relocate(chunk, oslot);
530 return off;
531 }
532
533 chunk->contig_hint = max_contig; /* fully scanned */
534 pcpu_chunk_relocate(chunk, oslot);
535
536 /* tell the upper layer that this chunk has no matching area */
537 return -1;
538}
539
540/**
541 * pcpu_free_area - free area to a pcpu_chunk
542 * @chunk: chunk of interest
543 * @freeme: offset of area to free
544 *
545 * Free area starting from @freeme to @chunk. Note that this function
546 * only modifies the allocation map. It doesn't depopulate or unmap
547 * the area.
548 *
549 * CONTEXT:
550 * pcpu_lock.
551 */
552static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
553{
554 int oslot = pcpu_chunk_slot(chunk);
555 int off = 0;
556 unsigned i, j;
557 int to_free = 0;
558 int *p;
559
560 freeme |= 1; /* we are searching for <given offset, in use> pair */
561
562 i = 0;
563 j = chunk->map_used;
564 while (i != j) {
565 unsigned k = (i + j) / 2;
566 off = chunk->map[k];
567 if (off < freeme)
568 i = k + 1;
569 else if (off > freeme)
570 j = k;
571 else
572 i = j = k;
573 }
574 BUG_ON(off != freeme);
575
576 if (i < chunk->first_free)
577 chunk->first_free = i;
578
579 p = chunk->map + i;
580 *p = off &= ~1;
581 chunk->free_size += (p[1] & ~1) - off;
582
583 /* merge with next? */
584 if (!(p[1] & 1))
585 to_free++;
586 /* merge with previous? */
587 if (i > 0 && !(p[-1] & 1)) {
588 to_free++;
589 i--;
590 p--;
591 }
592 if (to_free) {
593 chunk->map_used -= to_free;
594 memmove(p + 1, p + 1 + to_free,
595 (chunk->map_used - i) * sizeof(chunk->map[0]));
596 }
597
598 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
599 pcpu_chunk_relocate(chunk, oslot);
600}
601
602static struct pcpu_chunk *pcpu_alloc_chunk(void)
603{
604 struct pcpu_chunk *chunk;
605
606 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
607 if (!chunk)
608 return NULL;
609
610 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
611 sizeof(chunk->map[0]));
612 if (!chunk->map) {
613 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
614 return NULL;
615 }
616
617 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
618 chunk->map[0] = 0;
619 chunk->map[1] = pcpu_unit_size | 1;
620 chunk->map_used = 1;
621
622 INIT_LIST_HEAD(&chunk->list);
623 chunk->free_size = pcpu_unit_size;
624 chunk->contig_hint = pcpu_unit_size;
625
626 return chunk;
627}
628
629static void pcpu_free_chunk(struct pcpu_chunk *chunk)
630{
631 if (!chunk)
632 return;
633 pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
634 pcpu_mem_free(chunk, pcpu_chunk_struct_size);
635}
636
637/*
638 * Chunk management implementation.
639 *
640 * To allow different implementations, chunk alloc/free and
641 * [de]population are implemented in a separate file which is pulled
642 * into this file and compiled together. The following functions
643 * should be implemented.
644 *
645 * pcpu_populate_chunk - populate the specified range of a chunk
646 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
647 * pcpu_create_chunk - create a new chunk
648 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
649 * pcpu_addr_to_page - translate address to physical address
650 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
651 */
652static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
653static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
654static struct pcpu_chunk *pcpu_create_chunk(void);
655static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
656static struct page *pcpu_addr_to_page(void *addr);
657static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
658
659#ifdef CONFIG_NEED_PER_CPU_KM
660#include "percpu-km.c"
661#else
662#include "percpu-vm.c"
663#endif
664
665/**
666 * pcpu_chunk_addr_search - determine chunk containing specified address
667 * @addr: address for which the chunk needs to be determined.
668 *
669 * RETURNS:
670 * The address of the found chunk.
671 */
672static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
673{
674 /* is it in the first chunk? */
675 if (pcpu_addr_in_first_chunk(addr)) {
676 /* is it in the reserved area? */
677 if (pcpu_addr_in_reserved_chunk(addr))
678 return pcpu_reserved_chunk;
679 return pcpu_first_chunk;
680 }
681
682 /*
683 * The address is relative to unit0 which might be unused and
684 * thus unmapped. Offset the address to the unit space of the
685 * current processor before looking it up in the vmalloc
686 * space. Note that any possible cpu id can be used here, so
687 * there's no need to worry about preemption or cpu hotplug.
688 */
689 addr += pcpu_unit_offsets[raw_smp_processor_id()];
690 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
691}
692
693/**
694 * pcpu_alloc - the percpu allocator
695 * @size: size of area to allocate in bytes
696 * @align: alignment of area (max PAGE_SIZE)
697 * @reserved: allocate from the reserved chunk if available
698 *
699 * Allocate percpu area of @size bytes aligned at @align.
700 *
701 * CONTEXT:
702 * Does GFP_KERNEL allocation.
703 *
704 * RETURNS:
705 * Percpu pointer to the allocated area on success, NULL on failure.
706 */
707static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
708{
709 static int warn_limit = 10;
710 struct pcpu_chunk *chunk;
711 const char *err;
712 int slot, off, new_alloc;
713 unsigned long flags;
714 void __percpu *ptr;
715
716 /*
717 * We want the lowest bit of offset available for in-use/free
718 * indicator, so force >= 16bit alignment and make size even.
719 */
720 if (unlikely(align < 2))
721 align = 2;
722
723 if (unlikely(size & 1))
724 size++;
725
726 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
727 WARN(true, "illegal size (%zu) or align (%zu) for "
728 "percpu allocation\n", size, align);
729 return NULL;
730 }
731
732 mutex_lock(&pcpu_alloc_mutex);
733 spin_lock_irqsave(&pcpu_lock, flags);
734
735 /* serve reserved allocations from the reserved chunk if available */
736 if (reserved && pcpu_reserved_chunk) {
737 chunk = pcpu_reserved_chunk;
738
739 if (size > chunk->contig_hint) {
740 err = "alloc from reserved chunk failed";
741 goto fail_unlock;
742 }
743
744 while ((new_alloc = pcpu_need_to_extend(chunk))) {
745 spin_unlock_irqrestore(&pcpu_lock, flags);
746 if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
747 err = "failed to extend area map of reserved chunk";
748 goto fail_unlock_mutex;
749 }
750 spin_lock_irqsave(&pcpu_lock, flags);
751 }
752
753 off = pcpu_alloc_area(chunk, size, align);
754 if (off >= 0)
755 goto area_found;
756
757 err = "alloc from reserved chunk failed";
758 goto fail_unlock;
759 }
760
761restart:
762 /* search through normal chunks */
763 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
764 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
765 if (size > chunk->contig_hint)
766 continue;
767
768 new_alloc = pcpu_need_to_extend(chunk);
769 if (new_alloc) {
770 spin_unlock_irqrestore(&pcpu_lock, flags);
771 if (pcpu_extend_area_map(chunk,
772 new_alloc) < 0) {
773 err = "failed to extend area map";
774 goto fail_unlock_mutex;
775 }
776 spin_lock_irqsave(&pcpu_lock, flags);
777 /*
778 * pcpu_lock has been dropped, need to
779 * restart cpu_slot list walking.
780 */
781 goto restart;
782 }
783
784 off = pcpu_alloc_area(chunk, size, align);
785 if (off >= 0)
786 goto area_found;
787 }
788 }
789
790 /* hmmm... no space left, create a new chunk */
791 spin_unlock_irqrestore(&pcpu_lock, flags);
792
793 chunk = pcpu_create_chunk();
794 if (!chunk) {
795 err = "failed to allocate new chunk";
796 goto fail_unlock_mutex;
797 }
798
799 spin_lock_irqsave(&pcpu_lock, flags);
800 pcpu_chunk_relocate(chunk, -1);
801 goto restart;
802
803area_found:
804 spin_unlock_irqrestore(&pcpu_lock, flags);
805
806 /* populate, map and clear the area */
807 if (pcpu_populate_chunk(chunk, off, size)) {
808 spin_lock_irqsave(&pcpu_lock, flags);
809 pcpu_free_area(chunk, off);
810 err = "failed to populate";
811 goto fail_unlock;
812 }
813
814 mutex_unlock(&pcpu_alloc_mutex);
815
816 /* return address relative to base address */
817 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
818 kmemleak_alloc_percpu(ptr, size);
819 return ptr;
820
821fail_unlock:
822 spin_unlock_irqrestore(&pcpu_lock, flags);
823fail_unlock_mutex:
824 mutex_unlock(&pcpu_alloc_mutex);
825 if (warn_limit) {
826 pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
827 "%s\n", size, align, err);
828 dump_stack();
829 if (!--warn_limit)
830 pr_info("PERCPU: limit reached, disable warning\n");
831 }
832 return NULL;
833}
834
835/**
836 * __alloc_percpu - allocate dynamic percpu area
837 * @size: size of area to allocate in bytes
838 * @align: alignment of area (max PAGE_SIZE)
839 *
840 * Allocate zero-filled percpu area of @size bytes aligned at @align.
841 * Might sleep. Might trigger writeouts.
842 *
843 * CONTEXT:
844 * Does GFP_KERNEL allocation.
845 *
846 * RETURNS:
847 * Percpu pointer to the allocated area on success, NULL on failure.
848 */
849void __percpu *__alloc_percpu(size_t size, size_t align)
850{
851 return pcpu_alloc(size, align, false);
852}
853EXPORT_SYMBOL_GPL(__alloc_percpu);
854
855/**
856 * __alloc_reserved_percpu - allocate reserved percpu area
857 * @size: size of area to allocate in bytes
858 * @align: alignment of area (max PAGE_SIZE)
859 *
860 * Allocate zero-filled percpu area of @size bytes aligned at @align
861 * from reserved percpu area if arch has set it up; otherwise,
862 * allocation is served from the same dynamic area. Might sleep.
863 * Might trigger writeouts.
864 *
865 * CONTEXT:
866 * Does GFP_KERNEL allocation.
867 *
868 * RETURNS:
869 * Percpu pointer to the allocated area on success, NULL on failure.
870 */
871void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
872{
873 return pcpu_alloc(size, align, true);
874}
875
876/**
877 * pcpu_reclaim - reclaim fully free chunks, workqueue function
878 * @work: unused
879 *
880 * Reclaim all fully free chunks except for the first one.
881 *
882 * CONTEXT:
883 * workqueue context.
884 */
885static void pcpu_reclaim(struct work_struct *work)
886{
887 LIST_HEAD(todo);
888 struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
889 struct pcpu_chunk *chunk, *next;
890
891 mutex_lock(&pcpu_alloc_mutex);
892 spin_lock_irq(&pcpu_lock);
893
894 list_for_each_entry_safe(chunk, next, head, list) {
895 WARN_ON(chunk->immutable);
896
897 /* spare the first one */
898 if (chunk == list_first_entry(head, struct pcpu_chunk, list))
899 continue;
900
901 list_move(&chunk->list, &todo);
902 }
903
904 spin_unlock_irq(&pcpu_lock);
905
906 list_for_each_entry_safe(chunk, next, &todo, list) {
907 pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
908 pcpu_destroy_chunk(chunk);
909 }
910
911 mutex_unlock(&pcpu_alloc_mutex);
912}
913
914/**
915 * free_percpu - free percpu area
916 * @ptr: pointer to area to free
917 *
918 * Free percpu area @ptr.
919 *
920 * CONTEXT:
921 * Can be called from atomic context.
922 */
923void free_percpu(void __percpu *ptr)
924{
925 void *addr;
926 struct pcpu_chunk *chunk;
927 unsigned long flags;
928 int off;
929
930 if (!ptr)
931 return;
932
933 kmemleak_free_percpu(ptr);
934
935 addr = __pcpu_ptr_to_addr(ptr);
936
937 spin_lock_irqsave(&pcpu_lock, flags);
938
939 chunk = pcpu_chunk_addr_search(addr);
940 off = addr - chunk->base_addr;
941
942 pcpu_free_area(chunk, off);
943
944 /* if there are more than one fully free chunks, wake up grim reaper */
945 if (chunk->free_size == pcpu_unit_size) {
946 struct pcpu_chunk *pos;
947
948 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
949 if (pos != chunk) {
950 schedule_work(&pcpu_reclaim_work);
951 break;
952 }
953 }
954
955 spin_unlock_irqrestore(&pcpu_lock, flags);
956}
957EXPORT_SYMBOL_GPL(free_percpu);
958
959/**
960 * is_kernel_percpu_address - test whether address is from static percpu area
961 * @addr: address to test
962 *
963 * Test whether @addr belongs to in-kernel static percpu area. Module
964 * static percpu areas are not considered. For those, use
965 * is_module_percpu_address().
966 *
967 * RETURNS:
968 * %true if @addr is from in-kernel static percpu area, %false otherwise.
969 */
970bool is_kernel_percpu_address(unsigned long addr)
971{
972#ifdef CONFIG_SMP
973 const size_t static_size = __per_cpu_end - __per_cpu_start;
974 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
975 unsigned int cpu;
976
977 for_each_possible_cpu(cpu) {
978 void *start = per_cpu_ptr(base, cpu);
979
980 if ((void *)addr >= start && (void *)addr < start + static_size)
981 return true;
982 }
983#endif
984 /* on UP, can't distinguish from other static vars, always false */
985 return false;
986}
987
988/**
989 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
990 * @addr: the address to be converted to physical address
991 *
992 * Given @addr which is dereferenceable address obtained via one of
993 * percpu access macros, this function translates it into its physical
994 * address. The caller is responsible for ensuring @addr stays valid
995 * until this function finishes.
996 *
997 * percpu allocator has special setup for the first chunk, which currently
998 * supports either embedding in linear address space or vmalloc mapping,
999 * and, from the second one, the backing allocator (currently either vm or
1000 * km) provides translation.
1001 *
1002 * The addr can be tranlated simply without checking if it falls into the
1003 * first chunk. But the current code reflects better how percpu allocator
1004 * actually works, and the verification can discover both bugs in percpu
1005 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1006 * code.
1007 *
1008 * RETURNS:
1009 * The physical address for @addr.
1010 */
1011phys_addr_t per_cpu_ptr_to_phys(void *addr)
1012{
1013 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1014 bool in_first_chunk = false;
1015 unsigned long first_low, first_high;
1016 unsigned int cpu;
1017
1018 /*
1019 * The following test on unit_low/high isn't strictly
1020 * necessary but will speed up lookups of addresses which
1021 * aren't in the first chunk.
1022 */
1023 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1024 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1025 pcpu_unit_pages);
1026 if ((unsigned long)addr >= first_low &&
1027 (unsigned long)addr < first_high) {
1028 for_each_possible_cpu(cpu) {
1029 void *start = per_cpu_ptr(base, cpu);
1030
1031 if (addr >= start && addr < start + pcpu_unit_size) {
1032 in_first_chunk = true;
1033 break;
1034 }
1035 }
1036 }
1037
1038 if (in_first_chunk) {
1039 if (!is_vmalloc_addr(addr))
1040 return __pa(addr);
1041 else
1042 return page_to_phys(vmalloc_to_page(addr)) +
1043 offset_in_page(addr);
1044 } else
1045 return page_to_phys(pcpu_addr_to_page(addr)) +
1046 offset_in_page(addr);
1047}
1048
1049/**
1050 * pcpu_alloc_alloc_info - allocate percpu allocation info
1051 * @nr_groups: the number of groups
1052 * @nr_units: the number of units
1053 *
1054 * Allocate ai which is large enough for @nr_groups groups containing
1055 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1056 * cpu_map array which is long enough for @nr_units and filled with
1057 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1058 * pointer of other groups.
1059 *
1060 * RETURNS:
1061 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1062 * failure.
1063 */
1064struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1065 int nr_units)
1066{
1067 struct pcpu_alloc_info *ai;
1068 size_t base_size, ai_size;
1069 void *ptr;
1070 int unit;
1071
1072 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1073 __alignof__(ai->groups[0].cpu_map[0]));
1074 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1075
1076 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1077 if (!ptr)
1078 return NULL;
1079 ai = ptr;
1080 ptr += base_size;
1081
1082 ai->groups[0].cpu_map = ptr;
1083
1084 for (unit = 0; unit < nr_units; unit++)
1085 ai->groups[0].cpu_map[unit] = NR_CPUS;
1086
1087 ai->nr_groups = nr_groups;
1088 ai->__ai_size = PFN_ALIGN(ai_size);
1089
1090 return ai;
1091}
1092
1093/**
1094 * pcpu_free_alloc_info - free percpu allocation info
1095 * @ai: pcpu_alloc_info to free
1096 *
1097 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1098 */
1099void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1100{
1101 memblock_free_early(__pa(ai), ai->__ai_size);
1102}
1103
1104/**
1105 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1106 * @lvl: loglevel
1107 * @ai: allocation info to dump
1108 *
1109 * Print out information about @ai using loglevel @lvl.
1110 */
1111static void pcpu_dump_alloc_info(const char *lvl,
1112 const struct pcpu_alloc_info *ai)
1113{
1114 int group_width = 1, cpu_width = 1, width;
1115 char empty_str[] = "--------";
1116 int alloc = 0, alloc_end = 0;
1117 int group, v;
1118 int upa, apl; /* units per alloc, allocs per line */
1119
1120 v = ai->nr_groups;
1121 while (v /= 10)
1122 group_width++;
1123
1124 v = num_possible_cpus();
1125 while (v /= 10)
1126 cpu_width++;
1127 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1128
1129 upa = ai->alloc_size / ai->unit_size;
1130 width = upa * (cpu_width + 1) + group_width + 3;
1131 apl = rounddown_pow_of_two(max(60 / width, 1));
1132
1133 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1134 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1135 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1136
1137 for (group = 0; group < ai->nr_groups; group++) {
1138 const struct pcpu_group_info *gi = &ai->groups[group];
1139 int unit = 0, unit_end = 0;
1140
1141 BUG_ON(gi->nr_units % upa);
1142 for (alloc_end += gi->nr_units / upa;
1143 alloc < alloc_end; alloc++) {
1144 if (!(alloc % apl)) {
1145 printk(KERN_CONT "\n");
1146 printk("%spcpu-alloc: ", lvl);
1147 }
1148 printk(KERN_CONT "[%0*d] ", group_width, group);
1149
1150 for (unit_end += upa; unit < unit_end; unit++)
1151 if (gi->cpu_map[unit] != NR_CPUS)
1152 printk(KERN_CONT "%0*d ", cpu_width,
1153 gi->cpu_map[unit]);
1154 else
1155 printk(KERN_CONT "%s ", empty_str);
1156 }
1157 }
1158 printk(KERN_CONT "\n");
1159}
1160
1161/**
1162 * pcpu_setup_first_chunk - initialize the first percpu chunk
1163 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1164 * @base_addr: mapped address
1165 *
1166 * Initialize the first percpu chunk which contains the kernel static
1167 * perpcu area. This function is to be called from arch percpu area
1168 * setup path.
1169 *
1170 * @ai contains all information necessary to initialize the first
1171 * chunk and prime the dynamic percpu allocator.
1172 *
1173 * @ai->static_size is the size of static percpu area.
1174 *
1175 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1176 * reserve after the static area in the first chunk. This reserves
1177 * the first chunk such that it's available only through reserved
1178 * percpu allocation. This is primarily used to serve module percpu
1179 * static areas on architectures where the addressing model has
1180 * limited offset range for symbol relocations to guarantee module
1181 * percpu symbols fall inside the relocatable range.
1182 *
1183 * @ai->dyn_size determines the number of bytes available for dynamic
1184 * allocation in the first chunk. The area between @ai->static_size +
1185 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1186 *
1187 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1188 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1189 * @ai->dyn_size.
1190 *
1191 * @ai->atom_size is the allocation atom size and used as alignment
1192 * for vm areas.
1193 *
1194 * @ai->alloc_size is the allocation size and always multiple of
1195 * @ai->atom_size. This is larger than @ai->atom_size if
1196 * @ai->unit_size is larger than @ai->atom_size.
1197 *
1198 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1199 * percpu areas. Units which should be colocated are put into the
1200 * same group. Dynamic VM areas will be allocated according to these
1201 * groupings. If @ai->nr_groups is zero, a single group containing
1202 * all units is assumed.
1203 *
1204 * The caller should have mapped the first chunk at @base_addr and
1205 * copied static data to each unit.
1206 *
1207 * If the first chunk ends up with both reserved and dynamic areas, it
1208 * is served by two chunks - one to serve the core static and reserved
1209 * areas and the other for the dynamic area. They share the same vm
1210 * and page map but uses different area allocation map to stay away
1211 * from each other. The latter chunk is circulated in the chunk slots
1212 * and available for dynamic allocation like any other chunks.
1213 *
1214 * RETURNS:
1215 * 0 on success, -errno on failure.
1216 */
1217int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1218 void *base_addr)
1219{
1220 static char cpus_buf[4096] __initdata;
1221 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1222 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1223 size_t dyn_size = ai->dyn_size;
1224 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1225 struct pcpu_chunk *schunk, *dchunk = NULL;
1226 unsigned long *group_offsets;
1227 size_t *group_sizes;
1228 unsigned long *unit_off;
1229 unsigned int cpu;
1230 int *unit_map;
1231 int group, unit, i;
1232
1233 cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1234
1235#define PCPU_SETUP_BUG_ON(cond) do { \
1236 if (unlikely(cond)) { \
1237 pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1238 pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1239 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1240 BUG(); \
1241 } \
1242} while (0)
1243
1244 /* sanity checks */
1245 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1246#ifdef CONFIG_SMP
1247 PCPU_SETUP_BUG_ON(!ai->static_size);
1248 PCPU_SETUP_BUG_ON((unsigned long)__per_cpu_start & ~PAGE_MASK);
1249#endif
1250 PCPU_SETUP_BUG_ON(!base_addr);
1251 PCPU_SETUP_BUG_ON((unsigned long)base_addr & ~PAGE_MASK);
1252 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1253 PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1254 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1255 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1256 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1257
1258 /* process group information and build config tables accordingly */
1259 group_offsets = memblock_virt_alloc(ai->nr_groups *
1260 sizeof(group_offsets[0]), 0);
1261 group_sizes = memblock_virt_alloc(ai->nr_groups *
1262 sizeof(group_sizes[0]), 0);
1263 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1264 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1265
1266 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1267 unit_map[cpu] = UINT_MAX;
1268
1269 pcpu_low_unit_cpu = NR_CPUS;
1270 pcpu_high_unit_cpu = NR_CPUS;
1271
1272 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1273 const struct pcpu_group_info *gi = &ai->groups[group];
1274
1275 group_offsets[group] = gi->base_offset;
1276 group_sizes[group] = gi->nr_units * ai->unit_size;
1277
1278 for (i = 0; i < gi->nr_units; i++) {
1279 cpu = gi->cpu_map[i];
1280 if (cpu == NR_CPUS)
1281 continue;
1282
1283 PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1284 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1285 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1286
1287 unit_map[cpu] = unit + i;
1288 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1289
1290 /* determine low/high unit_cpu */
1291 if (pcpu_low_unit_cpu == NR_CPUS ||
1292 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1293 pcpu_low_unit_cpu = cpu;
1294 if (pcpu_high_unit_cpu == NR_CPUS ||
1295 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1296 pcpu_high_unit_cpu = cpu;
1297 }
1298 }
1299 pcpu_nr_units = unit;
1300
1301 for_each_possible_cpu(cpu)
1302 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1303
1304 /* we're done parsing the input, undefine BUG macro and dump config */
1305#undef PCPU_SETUP_BUG_ON
1306 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1307
1308 pcpu_nr_groups = ai->nr_groups;
1309 pcpu_group_offsets = group_offsets;
1310 pcpu_group_sizes = group_sizes;
1311 pcpu_unit_map = unit_map;
1312 pcpu_unit_offsets = unit_off;
1313
1314 /* determine basic parameters */
1315 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1316 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1317 pcpu_atom_size = ai->atom_size;
1318 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1319 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1320
1321 /*
1322 * Allocate chunk slots. The additional last slot is for
1323 * empty chunks.
1324 */
1325 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1326 pcpu_slot = memblock_virt_alloc(
1327 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1328 for (i = 0; i < pcpu_nr_slots; i++)
1329 INIT_LIST_HEAD(&pcpu_slot[i]);
1330
1331 /*
1332 * Initialize static chunk. If reserved_size is zero, the
1333 * static chunk covers static area + dynamic allocation area
1334 * in the first chunk. If reserved_size is not zero, it
1335 * covers static area + reserved area (mostly used for module
1336 * static percpu allocation).
1337 */
1338 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1339 INIT_LIST_HEAD(&schunk->list);
1340 schunk->base_addr = base_addr;
1341 schunk->map = smap;
1342 schunk->map_alloc = ARRAY_SIZE(smap);
1343 schunk->immutable = true;
1344 bitmap_fill(schunk->populated, pcpu_unit_pages);
1345
1346 if (ai->reserved_size) {
1347 schunk->free_size = ai->reserved_size;
1348 pcpu_reserved_chunk = schunk;
1349 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1350 } else {
1351 schunk->free_size = dyn_size;
1352 dyn_size = 0; /* dynamic area covered */
1353 }
1354 schunk->contig_hint = schunk->free_size;
1355
1356 schunk->map[0] = 1;
1357 schunk->map[1] = ai->static_size;
1358 schunk->map_used = 1;
1359 if (schunk->free_size)
1360 schunk->map[++schunk->map_used] = 1 | (ai->static_size + schunk->free_size);
1361 else
1362 schunk->map[1] |= 1;
1363
1364 /* init dynamic chunk if necessary */
1365 if (dyn_size) {
1366 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1367 INIT_LIST_HEAD(&dchunk->list);
1368 dchunk->base_addr = base_addr;
1369 dchunk->map = dmap;
1370 dchunk->map_alloc = ARRAY_SIZE(dmap);
1371 dchunk->immutable = true;
1372 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1373
1374 dchunk->contig_hint = dchunk->free_size = dyn_size;
1375 dchunk->map[0] = 1;
1376 dchunk->map[1] = pcpu_reserved_chunk_limit;
1377 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1378 dchunk->map_used = 2;
1379 }
1380
1381 /* link the first chunk in */
1382 pcpu_first_chunk = dchunk ?: schunk;
1383 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1384
1385 /* we're done */
1386 pcpu_base_addr = base_addr;
1387 return 0;
1388}
1389
1390#ifdef CONFIG_SMP
1391
1392const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1393 [PCPU_FC_AUTO] = "auto",
1394 [PCPU_FC_EMBED] = "embed",
1395 [PCPU_FC_PAGE] = "page",
1396};
1397
1398enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1399
1400static int __init percpu_alloc_setup(char *str)
1401{
1402 if (!str)
1403 return -EINVAL;
1404
1405 if (0)
1406 /* nada */;
1407#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1408 else if (!strcmp(str, "embed"))
1409 pcpu_chosen_fc = PCPU_FC_EMBED;
1410#endif
1411#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1412 else if (!strcmp(str, "page"))
1413 pcpu_chosen_fc = PCPU_FC_PAGE;
1414#endif
1415 else
1416 pr_warning("PERCPU: unknown allocator %s specified\n", str);
1417
1418 return 0;
1419}
1420early_param("percpu_alloc", percpu_alloc_setup);
1421
1422/*
1423 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1424 * Build it if needed by the arch config or the generic setup is going
1425 * to be used.
1426 */
1427#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1428 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1429#define BUILD_EMBED_FIRST_CHUNK
1430#endif
1431
1432/* build pcpu_page_first_chunk() iff needed by the arch config */
1433#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1434#define BUILD_PAGE_FIRST_CHUNK
1435#endif
1436
1437/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1438#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1439/**
1440 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1441 * @reserved_size: the size of reserved percpu area in bytes
1442 * @dyn_size: minimum free size for dynamic allocation in bytes
1443 * @atom_size: allocation atom size
1444 * @cpu_distance_fn: callback to determine distance between cpus, optional
1445 *
1446 * This function determines grouping of units, their mappings to cpus
1447 * and other parameters considering needed percpu size, allocation
1448 * atom size and distances between CPUs.
1449 *
1450 * Groups are always mutliples of atom size and CPUs which are of
1451 * LOCAL_DISTANCE both ways are grouped together and share space for
1452 * units in the same group. The returned configuration is guaranteed
1453 * to have CPUs on different nodes on different groups and >=75% usage
1454 * of allocated virtual address space.
1455 *
1456 * RETURNS:
1457 * On success, pointer to the new allocation_info is returned. On
1458 * failure, ERR_PTR value is returned.
1459 */
1460static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1461 size_t reserved_size, size_t dyn_size,
1462 size_t atom_size,
1463 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1464{
1465 static int group_map[NR_CPUS] __initdata;
1466 static int group_cnt[NR_CPUS] __initdata;
1467 const size_t static_size = __per_cpu_end - __per_cpu_start;
1468 int nr_groups = 1, nr_units = 0;
1469 size_t size_sum, min_unit_size, alloc_size;
1470 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1471 int last_allocs, group, unit;
1472 unsigned int cpu, tcpu;
1473 struct pcpu_alloc_info *ai;
1474 unsigned int *cpu_map;
1475
1476 /* this function may be called multiple times */
1477 memset(group_map, 0, sizeof(group_map));
1478 memset(group_cnt, 0, sizeof(group_cnt));
1479
1480 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1481 size_sum = PFN_ALIGN(static_size + reserved_size +
1482 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1483 dyn_size = size_sum - static_size - reserved_size;
1484
1485 /*
1486 * Determine min_unit_size, alloc_size and max_upa such that
1487 * alloc_size is multiple of atom_size and is the smallest
1488 * which can accommodate 4k aligned segments which are equal to
1489 * or larger than min_unit_size.
1490 */
1491 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1492
1493 alloc_size = roundup(min_unit_size, atom_size);
1494 upa = alloc_size / min_unit_size;
1495 while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1496 upa--;
1497 max_upa = upa;
1498
1499 /* group cpus according to their proximity */
1500 for_each_possible_cpu(cpu) {
1501 group = 0;
1502 next_group:
1503 for_each_possible_cpu(tcpu) {
1504 if (cpu == tcpu)
1505 break;
1506 if (group_map[tcpu] == group && cpu_distance_fn &&
1507 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1508 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1509 group++;
1510 nr_groups = max(nr_groups, group + 1);
1511 goto next_group;
1512 }
1513 }
1514 group_map[cpu] = group;
1515 group_cnt[group]++;
1516 }
1517
1518 /*
1519 * Expand unit size until address space usage goes over 75%
1520 * and then as much as possible without using more address
1521 * space.
1522 */
1523 last_allocs = INT_MAX;
1524 for (upa = max_upa; upa; upa--) {
1525 int allocs = 0, wasted = 0;
1526
1527 if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1528 continue;
1529
1530 for (group = 0; group < nr_groups; group++) {
1531 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1532 allocs += this_allocs;
1533 wasted += this_allocs * upa - group_cnt[group];
1534 }
1535
1536 /*
1537 * Don't accept if wastage is over 1/3. The
1538 * greater-than comparison ensures upa==1 always
1539 * passes the following check.
1540 */
1541 if (wasted > num_possible_cpus() / 3)
1542 continue;
1543
1544 /* and then don't consume more memory */
1545 if (allocs > last_allocs)
1546 break;
1547 last_allocs = allocs;
1548 best_upa = upa;
1549 }
1550 upa = best_upa;
1551
1552 /* allocate and fill alloc_info */
1553 for (group = 0; group < nr_groups; group++)
1554 nr_units += roundup(group_cnt[group], upa);
1555
1556 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1557 if (!ai)
1558 return ERR_PTR(-ENOMEM);
1559 cpu_map = ai->groups[0].cpu_map;
1560
1561 for (group = 0; group < nr_groups; group++) {
1562 ai->groups[group].cpu_map = cpu_map;
1563 cpu_map += roundup(group_cnt[group], upa);
1564 }
1565
1566 ai->static_size = static_size;
1567 ai->reserved_size = reserved_size;
1568 ai->dyn_size = dyn_size;
1569 ai->unit_size = alloc_size / upa;
1570 ai->atom_size = atom_size;
1571 ai->alloc_size = alloc_size;
1572
1573 for (group = 0, unit = 0; group_cnt[group]; group++) {
1574 struct pcpu_group_info *gi = &ai->groups[group];
1575
1576 /*
1577 * Initialize base_offset as if all groups are located
1578 * back-to-back. The caller should update this to
1579 * reflect actual allocation.
1580 */
1581 gi->base_offset = unit * ai->unit_size;
1582
1583 for_each_possible_cpu(cpu)
1584 if (group_map[cpu] == group)
1585 gi->cpu_map[gi->nr_units++] = cpu;
1586 gi->nr_units = roundup(gi->nr_units, upa);
1587 unit += gi->nr_units;
1588 }
1589 BUG_ON(unit != nr_units);
1590
1591 return ai;
1592}
1593#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1594
1595#if defined(BUILD_EMBED_FIRST_CHUNK)
1596/**
1597 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1598 * @reserved_size: the size of reserved percpu area in bytes
1599 * @dyn_size: minimum free size for dynamic allocation in bytes
1600 * @atom_size: allocation atom size
1601 * @cpu_distance_fn: callback to determine distance between cpus, optional
1602 * @alloc_fn: function to allocate percpu page
1603 * @free_fn: function to free percpu page
1604 *
1605 * This is a helper to ease setting up embedded first percpu chunk and
1606 * can be called where pcpu_setup_first_chunk() is expected.
1607 *
1608 * If this function is used to setup the first chunk, it is allocated
1609 * by calling @alloc_fn and used as-is without being mapped into
1610 * vmalloc area. Allocations are always whole multiples of @atom_size
1611 * aligned to @atom_size.
1612 *
1613 * This enables the first chunk to piggy back on the linear physical
1614 * mapping which often uses larger page size. Please note that this
1615 * can result in very sparse cpu->unit mapping on NUMA machines thus
1616 * requiring large vmalloc address space. Don't use this allocator if
1617 * vmalloc space is not orders of magnitude larger than distances
1618 * between node memory addresses (ie. 32bit NUMA machines).
1619 *
1620 * @dyn_size specifies the minimum dynamic area size.
1621 *
1622 * If the needed size is smaller than the minimum or specified unit
1623 * size, the leftover is returned using @free_fn.
1624 *
1625 * RETURNS:
1626 * 0 on success, -errno on failure.
1627 */
1628int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1629 size_t atom_size,
1630 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1631 pcpu_fc_alloc_fn_t alloc_fn,
1632 pcpu_fc_free_fn_t free_fn)
1633{
1634 void *base = (void *)ULONG_MAX;
1635 void **areas = NULL;
1636 struct pcpu_alloc_info *ai;
1637 size_t size_sum, areas_size, max_distance;
1638 int group, i, rc;
1639
1640 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1641 cpu_distance_fn);
1642 if (IS_ERR(ai))
1643 return PTR_ERR(ai);
1644
1645 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1646 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1647
1648 areas = memblock_virt_alloc_nopanic(areas_size, 0);
1649 if (!areas) {
1650 rc = -ENOMEM;
1651 goto out_free;
1652 }
1653
1654 /* allocate, copy and determine base address */
1655 for (group = 0; group < ai->nr_groups; group++) {
1656 struct pcpu_group_info *gi = &ai->groups[group];
1657 unsigned int cpu = NR_CPUS;
1658 void *ptr;
1659
1660 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1661 cpu = gi->cpu_map[i];
1662 BUG_ON(cpu == NR_CPUS);
1663
1664 /* allocate space for the whole group */
1665 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1666 if (!ptr) {
1667 rc = -ENOMEM;
1668 goto out_free_areas;
1669 }
1670 /* kmemleak tracks the percpu allocations separately */
1671 kmemleak_free(ptr);
1672 areas[group] = ptr;
1673
1674 base = min(ptr, base);
1675 }
1676
1677 /*
1678 * Copy data and free unused parts. This should happen after all
1679 * allocations are complete; otherwise, we may end up with
1680 * overlapping groups.
1681 */
1682 for (group = 0; group < ai->nr_groups; group++) {
1683 struct pcpu_group_info *gi = &ai->groups[group];
1684 void *ptr = areas[group];
1685
1686 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1687 if (gi->cpu_map[i] == NR_CPUS) {
1688 /* unused unit, free whole */
1689 free_fn(ptr, ai->unit_size);
1690 continue;
1691 }
1692 /* copy and return the unused part */
1693 memcpy(ptr, __per_cpu_load, ai->static_size);
1694 free_fn(ptr + size_sum, ai->unit_size - size_sum);
1695 }
1696 }
1697
1698 /* base address is now known, determine group base offsets */
1699 max_distance = 0;
1700 for (group = 0; group < ai->nr_groups; group++) {
1701 ai->groups[group].base_offset = areas[group] - base;
1702 max_distance = max_t(size_t, max_distance,
1703 ai->groups[group].base_offset);
1704 }
1705 max_distance += ai->unit_size;
1706
1707 /* warn if maximum distance is further than 75% of vmalloc space */
1708 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
1709 pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1710 "space 0x%lx\n", max_distance,
1711 VMALLOC_TOTAL);
1712#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1713 /* and fail if we have fallback */
1714 rc = -EINVAL;
1715 goto out_free;
1716#endif
1717 }
1718
1719 pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1720 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1721 ai->dyn_size, ai->unit_size);
1722
1723 rc = pcpu_setup_first_chunk(ai, base);
1724 goto out_free;
1725
1726out_free_areas:
1727 for (group = 0; group < ai->nr_groups; group++)
1728 if (areas[group])
1729 free_fn(areas[group],
1730 ai->groups[group].nr_units * ai->unit_size);
1731out_free:
1732 pcpu_free_alloc_info(ai);
1733 if (areas)
1734 memblock_free_early(__pa(areas), areas_size);
1735 return rc;
1736}
1737#endif /* BUILD_EMBED_FIRST_CHUNK */
1738
1739#ifdef BUILD_PAGE_FIRST_CHUNK
1740/**
1741 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1742 * @reserved_size: the size of reserved percpu area in bytes
1743 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1744 * @free_fn: function to free percpu page, always called with PAGE_SIZE
1745 * @populate_pte_fn: function to populate pte
1746 *
1747 * This is a helper to ease setting up page-remapped first percpu
1748 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1749 *
1750 * This is the basic allocator. Static percpu area is allocated
1751 * page-by-page into vmalloc area.
1752 *
1753 * RETURNS:
1754 * 0 on success, -errno on failure.
1755 */
1756int __init pcpu_page_first_chunk(size_t reserved_size,
1757 pcpu_fc_alloc_fn_t alloc_fn,
1758 pcpu_fc_free_fn_t free_fn,
1759 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1760{
1761 static struct vm_struct vm;
1762 struct pcpu_alloc_info *ai;
1763 char psize_str[16];
1764 int unit_pages;
1765 size_t pages_size;
1766 struct page **pages;
1767 int unit, i, j, rc;
1768
1769 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1770
1771 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
1772 if (IS_ERR(ai))
1773 return PTR_ERR(ai);
1774 BUG_ON(ai->nr_groups != 1);
1775 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1776
1777 unit_pages = ai->unit_size >> PAGE_SHIFT;
1778
1779 /* unaligned allocations can't be freed, round up to page size */
1780 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1781 sizeof(pages[0]));
1782 pages = memblock_virt_alloc(pages_size, 0);
1783
1784 /* allocate pages */
1785 j = 0;
1786 for (unit = 0; unit < num_possible_cpus(); unit++)
1787 for (i = 0; i < unit_pages; i++) {
1788 unsigned int cpu = ai->groups[0].cpu_map[unit];
1789 void *ptr;
1790
1791 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1792 if (!ptr) {
1793 pr_warning("PERCPU: failed to allocate %s page "
1794 "for cpu%u\n", psize_str, cpu);
1795 goto enomem;
1796 }
1797 /* kmemleak tracks the percpu allocations separately */
1798 kmemleak_free(ptr);
1799 pages[j++] = virt_to_page(ptr);
1800 }
1801
1802 /* allocate vm area, map the pages and copy static data */
1803 vm.flags = VM_ALLOC;
1804 vm.size = num_possible_cpus() * ai->unit_size;
1805 vm_area_register_early(&vm, PAGE_SIZE);
1806
1807 for (unit = 0; unit < num_possible_cpus(); unit++) {
1808 unsigned long unit_addr =
1809 (unsigned long)vm.addr + unit * ai->unit_size;
1810
1811 for (i = 0; i < unit_pages; i++)
1812 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1813
1814 /* pte already populated, the following shouldn't fail */
1815 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1816 unit_pages);
1817 if (rc < 0)
1818 panic("failed to map percpu area, err=%d\n", rc);
1819
1820 /*
1821 * FIXME: Archs with virtual cache should flush local
1822 * cache for the linear mapping here - something
1823 * equivalent to flush_cache_vmap() on the local cpu.
1824 * flush_cache_vmap() can't be used as most supporting
1825 * data structures are not set up yet.
1826 */
1827
1828 /* copy static data */
1829 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1830 }
1831
1832 /* we're ready, commit */
1833 pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1834 unit_pages, psize_str, vm.addr, ai->static_size,
1835 ai->reserved_size, ai->dyn_size);
1836
1837 rc = pcpu_setup_first_chunk(ai, vm.addr);
1838 goto out_free_ar;
1839
1840enomem:
1841 while (--j >= 0)
1842 free_fn(page_address(pages[j]), PAGE_SIZE);
1843 rc = -ENOMEM;
1844out_free_ar:
1845 memblock_free_early(__pa(pages), pages_size);
1846 pcpu_free_alloc_info(ai);
1847 return rc;
1848}
1849#endif /* BUILD_PAGE_FIRST_CHUNK */
1850
1851#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1852/*
1853 * Generic SMP percpu area setup.
1854 *
1855 * The embedding helper is used because its behavior closely resembles
1856 * the original non-dynamic generic percpu area setup. This is
1857 * important because many archs have addressing restrictions and might
1858 * fail if the percpu area is located far away from the previous
1859 * location. As an added bonus, in non-NUMA cases, embedding is
1860 * generally a good idea TLB-wise because percpu area can piggy back
1861 * on the physical linear memory mapping which uses large page
1862 * mappings on applicable archs.
1863 */
1864unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1865EXPORT_SYMBOL(__per_cpu_offset);
1866
1867static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1868 size_t align)
1869{
1870 return memblock_virt_alloc_from_nopanic(
1871 size, align, __pa(MAX_DMA_ADDRESS));
1872}
1873
1874static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1875{
1876 memblock_free_early(__pa(ptr), size);
1877}
1878
1879void __init setup_per_cpu_areas(void)
1880{
1881 unsigned long delta;
1882 unsigned int cpu;
1883 int rc;
1884
1885 /*
1886 * Always reserve area for module percpu variables. That's
1887 * what the legacy allocator did.
1888 */
1889 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1890 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1891 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
1892 if (rc < 0)
1893 panic("Failed to initialize percpu areas.");
1894
1895 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1896 for_each_possible_cpu(cpu)
1897 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1898}
1899#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1900
1901#else /* CONFIG_SMP */
1902
1903/*
1904 * UP percpu area setup.
1905 *
1906 * UP always uses km-based percpu allocator with identity mapping.
1907 * Static percpu variables are indistinguishable from the usual static
1908 * variables and don't require any special preparation.
1909 */
1910void __init setup_per_cpu_areas(void)
1911{
1912 const size_t unit_size =
1913 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
1914 PERCPU_DYNAMIC_RESERVE));
1915 struct pcpu_alloc_info *ai;
1916 void *fc;
1917
1918 ai = pcpu_alloc_alloc_info(1, 1);
1919 fc = memblock_virt_alloc_from_nopanic(unit_size,
1920 PAGE_SIZE,
1921 __pa(MAX_DMA_ADDRESS));
1922 if (!ai || !fc)
1923 panic("Failed to allocate memory for percpu areas.");
1924 /* kmemleak tracks the percpu allocations separately */
1925 kmemleak_free(fc);
1926
1927 ai->dyn_size = unit_size;
1928 ai->unit_size = unit_size;
1929 ai->atom_size = unit_size;
1930 ai->alloc_size = unit_size;
1931 ai->groups[0].nr_units = 1;
1932 ai->groups[0].cpu_map[0] = 0;
1933
1934 if (pcpu_setup_first_chunk(ai, fc) < 0)
1935 panic("Failed to initialize percpu areas.");
1936}
1937
1938#endif /* CONFIG_SMP */
1939
1940/*
1941 * First and reserved chunks are initialized with temporary allocation
1942 * map in initdata so that they can be used before slab is online.
1943 * This function is called after slab is brought up and replaces those
1944 * with properly allocated maps.
1945 */
1946void __init percpu_init_late(void)
1947{
1948 struct pcpu_chunk *target_chunks[] =
1949 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
1950 struct pcpu_chunk *chunk;
1951 unsigned long flags;
1952 int i;
1953
1954 for (i = 0; (chunk = target_chunks[i]); i++) {
1955 int *map;
1956 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
1957
1958 BUILD_BUG_ON(size > PAGE_SIZE);
1959
1960 map = pcpu_mem_zalloc(size);
1961 BUG_ON(!map);
1962
1963 spin_lock_irqsave(&pcpu_lock, flags);
1964 memcpy(map, chunk->map, size);
1965 chunk->map = map;
1966 spin_unlock_irqrestore(&pcpu_lock, flags);
1967 }
1968}