Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/percpu.c - percpu memory allocator
4 *
5 * Copyright (C) 2009 SUSE Linux Products GmbH
6 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 *
8 * Copyright (C) 2017 Facebook Inc.
9 * Copyright (C) 2017 Dennis Zhou <dennisszhou@gmail.com>
10 *
11 * The percpu allocator handles both static and dynamic areas. Percpu
12 * areas are allocated in chunks which are divided into units. There is
13 * a 1-to-1 mapping for units to possible cpus. These units are grouped
14 * based on NUMA properties of the machine.
15 *
16 * c0 c1 c2
17 * ------------------- ------------------- ------------
18 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19 * ------------------- ...... ------------------- .... ------------
20 *
21 * Allocation is done by offsets into a unit's address space. Ie., an
22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24 * and even sparse. Access is handled by configuring percpu base
25 * registers according to the cpu to unit mappings and offsetting the
26 * base address using pcpu_unit_size.
27 *
28 * There is special consideration for the first chunk which must handle
29 * the static percpu variables in the kernel image as allocation services
30 * are not online yet. In short, the first chunk is structured like so:
31 *
32 * <Static | [Reserved] | Dynamic>
33 *
34 * The static data is copied from the original section managed by the
35 * linker. The reserved section, if non-zero, primarily manages static
36 * percpu variables from kernel modules. Finally, the dynamic section
37 * takes care of normal allocations.
38 *
39 * The allocator organizes chunks into lists according to free size and
40 * tries to allocate from the fullest chunk first. Each chunk is managed
41 * by a bitmap with metadata blocks. The allocation map is updated on
42 * every allocation and free to reflect the current state while the boundary
43 * map is only updated on allocation. Each metadata block contains
44 * information to help mitigate the need to iterate over large portions
45 * of the bitmap. The reverse mapping from page to chunk is stored in
46 * the page's index. Lastly, units are lazily backed and grow in unison.
47 *
48 * There is a unique conversion that goes on here between bytes and bits.
49 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
50 * tracks the number of pages it is responsible for in nr_pages. Helper
51 * functions are used to convert from between the bytes, bits, and blocks.
52 * All hints are managed in bits unless explicitly stated.
53 *
54 * To use this allocator, arch code should do the following:
55 *
56 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
57 * regular address to percpu pointer and back if they need to be
58 * different from the default
59 *
60 * - use pcpu_setup_first_chunk() during percpu area initialization to
61 * setup the first chunk containing the kernel static percpu area
62 */
63
64#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
65
66#include <linux/bitmap.h>
67#include <linux/memblock.h>
68#include <linux/err.h>
69#include <linux/lcm.h>
70#include <linux/list.h>
71#include <linux/log2.h>
72#include <linux/mm.h>
73#include <linux/module.h>
74#include <linux/mutex.h>
75#include <linux/percpu.h>
76#include <linux/pfn.h>
77#include <linux/slab.h>
78#include <linux/spinlock.h>
79#include <linux/vmalloc.h>
80#include <linux/workqueue.h>
81#include <linux/kmemleak.h>
82#include <linux/sched.h>
83
84#include <asm/cacheflush.h>
85#include <asm/sections.h>
86#include <asm/tlbflush.h>
87#include <asm/io.h>
88
89#define CREATE_TRACE_POINTS
90#include <trace/events/percpu.h>
91
92#include "percpu-internal.h"
93
94/* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
95#define PCPU_SLOT_BASE_SHIFT 5
96/* chunks in slots below this are subject to being sidelined on failed alloc */
97#define PCPU_SLOT_FAIL_THRESHOLD 3
98
99#define PCPU_EMPTY_POP_PAGES_LOW 2
100#define PCPU_EMPTY_POP_PAGES_HIGH 4
101
102#ifdef CONFIG_SMP
103/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
104#ifndef __addr_to_pcpu_ptr
105#define __addr_to_pcpu_ptr(addr) \
106 (void __percpu *)((unsigned long)(addr) - \
107 (unsigned long)pcpu_base_addr + \
108 (unsigned long)__per_cpu_start)
109#endif
110#ifndef __pcpu_ptr_to_addr
111#define __pcpu_ptr_to_addr(ptr) \
112 (void __force *)((unsigned long)(ptr) + \
113 (unsigned long)pcpu_base_addr - \
114 (unsigned long)__per_cpu_start)
115#endif
116#else /* CONFIG_SMP */
117/* on UP, it's always identity mapped */
118#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
119#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
120#endif /* CONFIG_SMP */
121
122static int pcpu_unit_pages __ro_after_init;
123static int pcpu_unit_size __ro_after_init;
124static int pcpu_nr_units __ro_after_init;
125static int pcpu_atom_size __ro_after_init;
126int pcpu_nr_slots __ro_after_init;
127static size_t pcpu_chunk_struct_size __ro_after_init;
128
129/* cpus with the lowest and highest unit addresses */
130static unsigned int pcpu_low_unit_cpu __ro_after_init;
131static unsigned int pcpu_high_unit_cpu __ro_after_init;
132
133/* the address of the first chunk which starts with the kernel static area */
134void *pcpu_base_addr __ro_after_init;
135EXPORT_SYMBOL_GPL(pcpu_base_addr);
136
137static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
138const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
139
140/* group information, used for vm allocation */
141static int pcpu_nr_groups __ro_after_init;
142static const unsigned long *pcpu_group_offsets __ro_after_init;
143static const size_t *pcpu_group_sizes __ro_after_init;
144
145/*
146 * The first chunk which always exists. Note that unlike other
147 * chunks, this one can be allocated and mapped in several different
148 * ways and thus often doesn't live in the vmalloc area.
149 */
150struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
151
152/*
153 * Optional reserved chunk. This chunk reserves part of the first
154 * chunk and serves it for reserved allocations. When the reserved
155 * region doesn't exist, the following variable is NULL.
156 */
157struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
158
159DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
160static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
161
162struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
163
164/* chunks which need their map areas extended, protected by pcpu_lock */
165static LIST_HEAD(pcpu_map_extend_chunks);
166
167/*
168 * The number of empty populated pages, protected by pcpu_lock. The
169 * reserved chunk doesn't contribute to the count.
170 */
171int pcpu_nr_empty_pop_pages;
172
173/*
174 * The number of populated pages in use by the allocator, protected by
175 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
176 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
177 * and increments/decrements this count by 1).
178 */
179static unsigned long pcpu_nr_populated;
180
181/*
182 * Balance work is used to populate or destroy chunks asynchronously. We
183 * try to keep the number of populated free pages between
184 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
185 * empty chunk.
186 */
187static void pcpu_balance_workfn(struct work_struct *work);
188static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
189static bool pcpu_async_enabled __read_mostly;
190static bool pcpu_atomic_alloc_failed;
191
192static void pcpu_schedule_balance_work(void)
193{
194 if (pcpu_async_enabled)
195 schedule_work(&pcpu_balance_work);
196}
197
198/**
199 * pcpu_addr_in_chunk - check if the address is served from this chunk
200 * @chunk: chunk of interest
201 * @addr: percpu address
202 *
203 * RETURNS:
204 * True if the address is served from this chunk.
205 */
206static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
207{
208 void *start_addr, *end_addr;
209
210 if (!chunk)
211 return false;
212
213 start_addr = chunk->base_addr + chunk->start_offset;
214 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
215 chunk->end_offset;
216
217 return addr >= start_addr && addr < end_addr;
218}
219
220static int __pcpu_size_to_slot(int size)
221{
222 int highbit = fls(size); /* size is in bytes */
223 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
224}
225
226static int pcpu_size_to_slot(int size)
227{
228 if (size == pcpu_unit_size)
229 return pcpu_nr_slots - 1;
230 return __pcpu_size_to_slot(size);
231}
232
233static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
234{
235 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
236
237 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
238 chunk_md->contig_hint == 0)
239 return 0;
240
241 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
242}
243
244/* set the pointer to a chunk in a page struct */
245static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
246{
247 page->index = (unsigned long)pcpu;
248}
249
250/* obtain pointer to a chunk from a page struct */
251static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
252{
253 return (struct pcpu_chunk *)page->index;
254}
255
256static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
257{
258 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
259}
260
261static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
262{
263 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
264}
265
266static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
267 unsigned int cpu, int page_idx)
268{
269 return (unsigned long)chunk->base_addr +
270 pcpu_unit_page_offset(cpu, page_idx);
271}
272
273static void pcpu_next_unpop(unsigned long *bitmap, int *rs, int *re, int end)
274{
275 *rs = find_next_zero_bit(bitmap, end, *rs);
276 *re = find_next_bit(bitmap, end, *rs + 1);
277}
278
279static void pcpu_next_pop(unsigned long *bitmap, int *rs, int *re, int end)
280{
281 *rs = find_next_bit(bitmap, end, *rs);
282 *re = find_next_zero_bit(bitmap, end, *rs + 1);
283}
284
285/*
286 * Bitmap region iterators. Iterates over the bitmap between
287 * [@start, @end) in @chunk. @rs and @re should be integer variables
288 * and will be set to start and end index of the current free region.
289 */
290#define pcpu_for_each_unpop_region(bitmap, rs, re, start, end) \
291 for ((rs) = (start), pcpu_next_unpop((bitmap), &(rs), &(re), (end)); \
292 (rs) < (re); \
293 (rs) = (re) + 1, pcpu_next_unpop((bitmap), &(rs), &(re), (end)))
294
295#define pcpu_for_each_pop_region(bitmap, rs, re, start, end) \
296 for ((rs) = (start), pcpu_next_pop((bitmap), &(rs), &(re), (end)); \
297 (rs) < (re); \
298 (rs) = (re) + 1, pcpu_next_pop((bitmap), &(rs), &(re), (end)))
299
300/*
301 * The following are helper functions to help access bitmaps and convert
302 * between bitmap offsets to address offsets.
303 */
304static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
305{
306 return chunk->alloc_map +
307 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
308}
309
310static unsigned long pcpu_off_to_block_index(int off)
311{
312 return off / PCPU_BITMAP_BLOCK_BITS;
313}
314
315static unsigned long pcpu_off_to_block_off(int off)
316{
317 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
318}
319
320static unsigned long pcpu_block_off_to_off(int index, int off)
321{
322 return index * PCPU_BITMAP_BLOCK_BITS + off;
323}
324
325/*
326 * pcpu_next_hint - determine which hint to use
327 * @block: block of interest
328 * @alloc_bits: size of allocation
329 *
330 * This determines if we should scan based on the scan_hint or first_free.
331 * In general, we want to scan from first_free to fulfill allocations by
332 * first fit. However, if we know a scan_hint at position scan_hint_start
333 * cannot fulfill an allocation, we can begin scanning from there knowing
334 * the contig_hint will be our fallback.
335 */
336static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
337{
338 /*
339 * The three conditions below determine if we can skip past the
340 * scan_hint. First, does the scan hint exist. Second, is the
341 * contig_hint after the scan_hint (possibly not true iff
342 * contig_hint == scan_hint). Third, is the allocation request
343 * larger than the scan_hint.
344 */
345 if (block->scan_hint &&
346 block->contig_hint_start > block->scan_hint_start &&
347 alloc_bits > block->scan_hint)
348 return block->scan_hint_start + block->scan_hint;
349
350 return block->first_free;
351}
352
353/**
354 * pcpu_next_md_free_region - finds the next hint free area
355 * @chunk: chunk of interest
356 * @bit_off: chunk offset
357 * @bits: size of free area
358 *
359 * Helper function for pcpu_for_each_md_free_region. It checks
360 * block->contig_hint and performs aggregation across blocks to find the
361 * next hint. It modifies bit_off and bits in-place to be consumed in the
362 * loop.
363 */
364static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
365 int *bits)
366{
367 int i = pcpu_off_to_block_index(*bit_off);
368 int block_off = pcpu_off_to_block_off(*bit_off);
369 struct pcpu_block_md *block;
370
371 *bits = 0;
372 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
373 block++, i++) {
374 /* handles contig area across blocks */
375 if (*bits) {
376 *bits += block->left_free;
377 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
378 continue;
379 return;
380 }
381
382 /*
383 * This checks three things. First is there a contig_hint to
384 * check. Second, have we checked this hint before by
385 * comparing the block_off. Third, is this the same as the
386 * right contig hint. In the last case, it spills over into
387 * the next block and should be handled by the contig area
388 * across blocks code.
389 */
390 *bits = block->contig_hint;
391 if (*bits && block->contig_hint_start >= block_off &&
392 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
393 *bit_off = pcpu_block_off_to_off(i,
394 block->contig_hint_start);
395 return;
396 }
397 /* reset to satisfy the second predicate above */
398 block_off = 0;
399
400 *bits = block->right_free;
401 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
402 }
403}
404
405/**
406 * pcpu_next_fit_region - finds fit areas for a given allocation request
407 * @chunk: chunk of interest
408 * @alloc_bits: size of allocation
409 * @align: alignment of area (max PAGE_SIZE)
410 * @bit_off: chunk offset
411 * @bits: size of free area
412 *
413 * Finds the next free region that is viable for use with a given size and
414 * alignment. This only returns if there is a valid area to be used for this
415 * allocation. block->first_free is returned if the allocation request fits
416 * within the block to see if the request can be fulfilled prior to the contig
417 * hint.
418 */
419static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
420 int align, int *bit_off, int *bits)
421{
422 int i = pcpu_off_to_block_index(*bit_off);
423 int block_off = pcpu_off_to_block_off(*bit_off);
424 struct pcpu_block_md *block;
425
426 *bits = 0;
427 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
428 block++, i++) {
429 /* handles contig area across blocks */
430 if (*bits) {
431 *bits += block->left_free;
432 if (*bits >= alloc_bits)
433 return;
434 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
435 continue;
436 }
437
438 /* check block->contig_hint */
439 *bits = ALIGN(block->contig_hint_start, align) -
440 block->contig_hint_start;
441 /*
442 * This uses the block offset to determine if this has been
443 * checked in the prior iteration.
444 */
445 if (block->contig_hint &&
446 block->contig_hint_start >= block_off &&
447 block->contig_hint >= *bits + alloc_bits) {
448 int start = pcpu_next_hint(block, alloc_bits);
449
450 *bits += alloc_bits + block->contig_hint_start -
451 start;
452 *bit_off = pcpu_block_off_to_off(i, start);
453 return;
454 }
455 /* reset to satisfy the second predicate above */
456 block_off = 0;
457
458 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
459 align);
460 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
461 *bit_off = pcpu_block_off_to_off(i, *bit_off);
462 if (*bits >= alloc_bits)
463 return;
464 }
465
466 /* no valid offsets were found - fail condition */
467 *bit_off = pcpu_chunk_map_bits(chunk);
468}
469
470/*
471 * Metadata free area iterators. These perform aggregation of free areas
472 * based on the metadata blocks and return the offset @bit_off and size in
473 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
474 * a fit is found for the allocation request.
475 */
476#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
477 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
478 (bit_off) < pcpu_chunk_map_bits((chunk)); \
479 (bit_off) += (bits) + 1, \
480 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
481
482#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
483 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
484 &(bits)); \
485 (bit_off) < pcpu_chunk_map_bits((chunk)); \
486 (bit_off) += (bits), \
487 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
488 &(bits)))
489
490/**
491 * pcpu_mem_zalloc - allocate memory
492 * @size: bytes to allocate
493 * @gfp: allocation flags
494 *
495 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
496 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
497 * This is to facilitate passing through whitelisted flags. The
498 * returned memory is always zeroed.
499 *
500 * RETURNS:
501 * Pointer to the allocated area on success, NULL on failure.
502 */
503static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
504{
505 if (WARN_ON_ONCE(!slab_is_available()))
506 return NULL;
507
508 if (size <= PAGE_SIZE)
509 return kzalloc(size, gfp);
510 else
511 return __vmalloc(size, gfp | __GFP_ZERO, PAGE_KERNEL);
512}
513
514/**
515 * pcpu_mem_free - free memory
516 * @ptr: memory to free
517 *
518 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
519 */
520static void pcpu_mem_free(void *ptr)
521{
522 kvfree(ptr);
523}
524
525static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
526 bool move_front)
527{
528 if (chunk != pcpu_reserved_chunk) {
529 if (move_front)
530 list_move(&chunk->list, &pcpu_slot[slot]);
531 else
532 list_move_tail(&chunk->list, &pcpu_slot[slot]);
533 }
534}
535
536static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
537{
538 __pcpu_chunk_move(chunk, slot, true);
539}
540
541/**
542 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
543 * @chunk: chunk of interest
544 * @oslot: the previous slot it was on
545 *
546 * This function is called after an allocation or free changed @chunk.
547 * New slot according to the changed state is determined and @chunk is
548 * moved to the slot. Note that the reserved chunk is never put on
549 * chunk slots.
550 *
551 * CONTEXT:
552 * pcpu_lock.
553 */
554static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
555{
556 int nslot = pcpu_chunk_slot(chunk);
557
558 if (oslot != nslot)
559 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
560}
561
562/*
563 * pcpu_update_empty_pages - update empty page counters
564 * @chunk: chunk of interest
565 * @nr: nr of empty pages
566 *
567 * This is used to keep track of the empty pages now based on the premise
568 * a md_block covers a page. The hint update functions recognize if a block
569 * is made full or broken to calculate deltas for keeping track of free pages.
570 */
571static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
572{
573 chunk->nr_empty_pop_pages += nr;
574 if (chunk != pcpu_reserved_chunk)
575 pcpu_nr_empty_pop_pages += nr;
576}
577
578/*
579 * pcpu_region_overlap - determines if two regions overlap
580 * @a: start of first region, inclusive
581 * @b: end of first region, exclusive
582 * @x: start of second region, inclusive
583 * @y: end of second region, exclusive
584 *
585 * This is used to determine if the hint region [a, b) overlaps with the
586 * allocated region [x, y).
587 */
588static inline bool pcpu_region_overlap(int a, int b, int x, int y)
589{
590 return (a < y) && (x < b);
591}
592
593/**
594 * pcpu_block_update - updates a block given a free area
595 * @block: block of interest
596 * @start: start offset in block
597 * @end: end offset in block
598 *
599 * Updates a block given a known free area. The region [start, end) is
600 * expected to be the entirety of the free area within a block. Chooses
601 * the best starting offset if the contig hints are equal.
602 */
603static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
604{
605 int contig = end - start;
606
607 block->first_free = min(block->first_free, start);
608 if (start == 0)
609 block->left_free = contig;
610
611 if (end == block->nr_bits)
612 block->right_free = contig;
613
614 if (contig > block->contig_hint) {
615 /* promote the old contig_hint to be the new scan_hint */
616 if (start > block->contig_hint_start) {
617 if (block->contig_hint > block->scan_hint) {
618 block->scan_hint_start =
619 block->contig_hint_start;
620 block->scan_hint = block->contig_hint;
621 } else if (start < block->scan_hint_start) {
622 /*
623 * The old contig_hint == scan_hint. But, the
624 * new contig is larger so hold the invariant
625 * scan_hint_start < contig_hint_start.
626 */
627 block->scan_hint = 0;
628 }
629 } else {
630 block->scan_hint = 0;
631 }
632 block->contig_hint_start = start;
633 block->contig_hint = contig;
634 } else if (contig == block->contig_hint) {
635 if (block->contig_hint_start &&
636 (!start ||
637 __ffs(start) > __ffs(block->contig_hint_start))) {
638 /* start has a better alignment so use it */
639 block->contig_hint_start = start;
640 if (start < block->scan_hint_start &&
641 block->contig_hint > block->scan_hint)
642 block->scan_hint = 0;
643 } else if (start > block->scan_hint_start ||
644 block->contig_hint > block->scan_hint) {
645 /*
646 * Knowing contig == contig_hint, update the scan_hint
647 * if it is farther than or larger than the current
648 * scan_hint.
649 */
650 block->scan_hint_start = start;
651 block->scan_hint = contig;
652 }
653 } else {
654 /*
655 * The region is smaller than the contig_hint. So only update
656 * the scan_hint if it is larger than or equal and farther than
657 * the current scan_hint.
658 */
659 if ((start < block->contig_hint_start &&
660 (contig > block->scan_hint ||
661 (contig == block->scan_hint &&
662 start > block->scan_hint_start)))) {
663 block->scan_hint_start = start;
664 block->scan_hint = contig;
665 }
666 }
667}
668
669/*
670 * pcpu_block_update_scan - update a block given a free area from a scan
671 * @chunk: chunk of interest
672 * @bit_off: chunk offset
673 * @bits: size of free area
674 *
675 * Finding the final allocation spot first goes through pcpu_find_block_fit()
676 * to find a block that can hold the allocation and then pcpu_alloc_area()
677 * where a scan is used. When allocations require specific alignments,
678 * we can inadvertently create holes which will not be seen in the alloc
679 * or free paths.
680 *
681 * This takes a given free area hole and updates a block as it may change the
682 * scan_hint. We need to scan backwards to ensure we don't miss free bits
683 * from alignment.
684 */
685static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
686 int bits)
687{
688 int s_off = pcpu_off_to_block_off(bit_off);
689 int e_off = s_off + bits;
690 int s_index, l_bit;
691 struct pcpu_block_md *block;
692
693 if (e_off > PCPU_BITMAP_BLOCK_BITS)
694 return;
695
696 s_index = pcpu_off_to_block_index(bit_off);
697 block = chunk->md_blocks + s_index;
698
699 /* scan backwards in case of alignment skipping free bits */
700 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
701 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
702
703 pcpu_block_update(block, s_off, e_off);
704}
705
706/**
707 * pcpu_chunk_refresh_hint - updates metadata about a chunk
708 * @chunk: chunk of interest
709 * @full_scan: if we should scan from the beginning
710 *
711 * Iterates over the metadata blocks to find the largest contig area.
712 * A full scan can be avoided on the allocation path as this is triggered
713 * if we broke the contig_hint. In doing so, the scan_hint will be before
714 * the contig_hint or after if the scan_hint == contig_hint. This cannot
715 * be prevented on freeing as we want to find the largest area possibly
716 * spanning blocks.
717 */
718static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
719{
720 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
721 int bit_off, bits;
722
723 /* promote scan_hint to contig_hint */
724 if (!full_scan && chunk_md->scan_hint) {
725 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
726 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
727 chunk_md->contig_hint = chunk_md->scan_hint;
728 chunk_md->scan_hint = 0;
729 } else {
730 bit_off = chunk_md->first_free;
731 chunk_md->contig_hint = 0;
732 }
733
734 bits = 0;
735 pcpu_for_each_md_free_region(chunk, bit_off, bits) {
736 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
737 }
738}
739
740/**
741 * pcpu_block_refresh_hint
742 * @chunk: chunk of interest
743 * @index: index of the metadata block
744 *
745 * Scans over the block beginning at first_free and updates the block
746 * metadata accordingly.
747 */
748static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
749{
750 struct pcpu_block_md *block = chunk->md_blocks + index;
751 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
752 int rs, re, start; /* region start, region end */
753
754 /* promote scan_hint to contig_hint */
755 if (block->scan_hint) {
756 start = block->scan_hint_start + block->scan_hint;
757 block->contig_hint_start = block->scan_hint_start;
758 block->contig_hint = block->scan_hint;
759 block->scan_hint = 0;
760 } else {
761 start = block->first_free;
762 block->contig_hint = 0;
763 }
764
765 block->right_free = 0;
766
767 /* iterate over free areas and update the contig hints */
768 pcpu_for_each_unpop_region(alloc_map, rs, re, start,
769 PCPU_BITMAP_BLOCK_BITS) {
770 pcpu_block_update(block, rs, re);
771 }
772}
773
774/**
775 * pcpu_block_update_hint_alloc - update hint on allocation path
776 * @chunk: chunk of interest
777 * @bit_off: chunk offset
778 * @bits: size of request
779 *
780 * Updates metadata for the allocation path. The metadata only has to be
781 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
782 * scans are required if the block's contig hint is broken.
783 */
784static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
785 int bits)
786{
787 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
788 int nr_empty_pages = 0;
789 struct pcpu_block_md *s_block, *e_block, *block;
790 int s_index, e_index; /* block indexes of the freed allocation */
791 int s_off, e_off; /* block offsets of the freed allocation */
792
793 /*
794 * Calculate per block offsets.
795 * The calculation uses an inclusive range, but the resulting offsets
796 * are [start, end). e_index always points to the last block in the
797 * range.
798 */
799 s_index = pcpu_off_to_block_index(bit_off);
800 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
801 s_off = pcpu_off_to_block_off(bit_off);
802 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
803
804 s_block = chunk->md_blocks + s_index;
805 e_block = chunk->md_blocks + e_index;
806
807 /*
808 * Update s_block.
809 * block->first_free must be updated if the allocation takes its place.
810 * If the allocation breaks the contig_hint, a scan is required to
811 * restore this hint.
812 */
813 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
814 nr_empty_pages++;
815
816 if (s_off == s_block->first_free)
817 s_block->first_free = find_next_zero_bit(
818 pcpu_index_alloc_map(chunk, s_index),
819 PCPU_BITMAP_BLOCK_BITS,
820 s_off + bits);
821
822 if (pcpu_region_overlap(s_block->scan_hint_start,
823 s_block->scan_hint_start + s_block->scan_hint,
824 s_off,
825 s_off + bits))
826 s_block->scan_hint = 0;
827
828 if (pcpu_region_overlap(s_block->contig_hint_start,
829 s_block->contig_hint_start +
830 s_block->contig_hint,
831 s_off,
832 s_off + bits)) {
833 /* block contig hint is broken - scan to fix it */
834 if (!s_off)
835 s_block->left_free = 0;
836 pcpu_block_refresh_hint(chunk, s_index);
837 } else {
838 /* update left and right contig manually */
839 s_block->left_free = min(s_block->left_free, s_off);
840 if (s_index == e_index)
841 s_block->right_free = min_t(int, s_block->right_free,
842 PCPU_BITMAP_BLOCK_BITS - e_off);
843 else
844 s_block->right_free = 0;
845 }
846
847 /*
848 * Update e_block.
849 */
850 if (s_index != e_index) {
851 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
852 nr_empty_pages++;
853
854 /*
855 * When the allocation is across blocks, the end is along
856 * the left part of the e_block.
857 */
858 e_block->first_free = find_next_zero_bit(
859 pcpu_index_alloc_map(chunk, e_index),
860 PCPU_BITMAP_BLOCK_BITS, e_off);
861
862 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
863 /* reset the block */
864 e_block++;
865 } else {
866 if (e_off > e_block->scan_hint_start)
867 e_block->scan_hint = 0;
868
869 e_block->left_free = 0;
870 if (e_off > e_block->contig_hint_start) {
871 /* contig hint is broken - scan to fix it */
872 pcpu_block_refresh_hint(chunk, e_index);
873 } else {
874 e_block->right_free =
875 min_t(int, e_block->right_free,
876 PCPU_BITMAP_BLOCK_BITS - e_off);
877 }
878 }
879
880 /* update in-between md_blocks */
881 nr_empty_pages += (e_index - s_index - 1);
882 for (block = s_block + 1; block < e_block; block++) {
883 block->scan_hint = 0;
884 block->contig_hint = 0;
885 block->left_free = 0;
886 block->right_free = 0;
887 }
888 }
889
890 if (nr_empty_pages)
891 pcpu_update_empty_pages(chunk, -nr_empty_pages);
892
893 if (pcpu_region_overlap(chunk_md->scan_hint_start,
894 chunk_md->scan_hint_start +
895 chunk_md->scan_hint,
896 bit_off,
897 bit_off + bits))
898 chunk_md->scan_hint = 0;
899
900 /*
901 * The only time a full chunk scan is required is if the chunk
902 * contig hint is broken. Otherwise, it means a smaller space
903 * was used and therefore the chunk contig hint is still correct.
904 */
905 if (pcpu_region_overlap(chunk_md->contig_hint_start,
906 chunk_md->contig_hint_start +
907 chunk_md->contig_hint,
908 bit_off,
909 bit_off + bits))
910 pcpu_chunk_refresh_hint(chunk, false);
911}
912
913/**
914 * pcpu_block_update_hint_free - updates the block hints on the free path
915 * @chunk: chunk of interest
916 * @bit_off: chunk offset
917 * @bits: size of request
918 *
919 * Updates metadata for the allocation path. This avoids a blind block
920 * refresh by making use of the block contig hints. If this fails, it scans
921 * forward and backward to determine the extent of the free area. This is
922 * capped at the boundary of blocks.
923 *
924 * A chunk update is triggered if a page becomes free, a block becomes free,
925 * or the free spans across blocks. This tradeoff is to minimize iterating
926 * over the block metadata to update chunk_md->contig_hint.
927 * chunk_md->contig_hint may be off by up to a page, but it will never be more
928 * than the available space. If the contig hint is contained in one block, it
929 * will be accurate.
930 */
931static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
932 int bits)
933{
934 int nr_empty_pages = 0;
935 struct pcpu_block_md *s_block, *e_block, *block;
936 int s_index, e_index; /* block indexes of the freed allocation */
937 int s_off, e_off; /* block offsets of the freed allocation */
938 int start, end; /* start and end of the whole free area */
939
940 /*
941 * Calculate per block offsets.
942 * The calculation uses an inclusive range, but the resulting offsets
943 * are [start, end). e_index always points to the last block in the
944 * range.
945 */
946 s_index = pcpu_off_to_block_index(bit_off);
947 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
948 s_off = pcpu_off_to_block_off(bit_off);
949 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
950
951 s_block = chunk->md_blocks + s_index;
952 e_block = chunk->md_blocks + e_index;
953
954 /*
955 * Check if the freed area aligns with the block->contig_hint.
956 * If it does, then the scan to find the beginning/end of the
957 * larger free area can be avoided.
958 *
959 * start and end refer to beginning and end of the free area
960 * within each their respective blocks. This is not necessarily
961 * the entire free area as it may span blocks past the beginning
962 * or end of the block.
963 */
964 start = s_off;
965 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
966 start = s_block->contig_hint_start;
967 } else {
968 /*
969 * Scan backwards to find the extent of the free area.
970 * find_last_bit returns the starting bit, so if the start bit
971 * is returned, that means there was no last bit and the
972 * remainder of the chunk is free.
973 */
974 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
975 start);
976 start = (start == l_bit) ? 0 : l_bit + 1;
977 }
978
979 end = e_off;
980 if (e_off == e_block->contig_hint_start)
981 end = e_block->contig_hint_start + e_block->contig_hint;
982 else
983 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
984 PCPU_BITMAP_BLOCK_BITS, end);
985
986 /* update s_block */
987 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
988 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
989 nr_empty_pages++;
990 pcpu_block_update(s_block, start, e_off);
991
992 /* freeing in the same block */
993 if (s_index != e_index) {
994 /* update e_block */
995 if (end == PCPU_BITMAP_BLOCK_BITS)
996 nr_empty_pages++;
997 pcpu_block_update(e_block, 0, end);
998
999 /* reset md_blocks in the middle */
1000 nr_empty_pages += (e_index - s_index - 1);
1001 for (block = s_block + 1; block < e_block; block++) {
1002 block->first_free = 0;
1003 block->scan_hint = 0;
1004 block->contig_hint_start = 0;
1005 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1006 block->left_free = PCPU_BITMAP_BLOCK_BITS;
1007 block->right_free = PCPU_BITMAP_BLOCK_BITS;
1008 }
1009 }
1010
1011 if (nr_empty_pages)
1012 pcpu_update_empty_pages(chunk, nr_empty_pages);
1013
1014 /*
1015 * Refresh chunk metadata when the free makes a block free or spans
1016 * across blocks. The contig_hint may be off by up to a page, but if
1017 * the contig_hint is contained in a block, it will be accurate with
1018 * the else condition below.
1019 */
1020 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1021 pcpu_chunk_refresh_hint(chunk, true);
1022 else
1023 pcpu_block_update(&chunk->chunk_md,
1024 pcpu_block_off_to_off(s_index, start),
1025 end);
1026}
1027
1028/**
1029 * pcpu_is_populated - determines if the region is populated
1030 * @chunk: chunk of interest
1031 * @bit_off: chunk offset
1032 * @bits: size of area
1033 * @next_off: return value for the next offset to start searching
1034 *
1035 * For atomic allocations, check if the backing pages are populated.
1036 *
1037 * RETURNS:
1038 * Bool if the backing pages are populated.
1039 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1040 */
1041static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1042 int *next_off)
1043{
1044 int page_start, page_end, rs, re;
1045
1046 page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1047 page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1048
1049 rs = page_start;
1050 pcpu_next_unpop(chunk->populated, &rs, &re, page_end);
1051 if (rs >= page_end)
1052 return true;
1053
1054 *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1055 return false;
1056}
1057
1058/**
1059 * pcpu_find_block_fit - finds the block index to start searching
1060 * @chunk: chunk of interest
1061 * @alloc_bits: size of request in allocation units
1062 * @align: alignment of area (max PAGE_SIZE bytes)
1063 * @pop_only: use populated regions only
1064 *
1065 * Given a chunk and an allocation spec, find the offset to begin searching
1066 * for a free region. This iterates over the bitmap metadata blocks to
1067 * find an offset that will be guaranteed to fit the requirements. It is
1068 * not quite first fit as if the allocation does not fit in the contig hint
1069 * of a block or chunk, it is skipped. This errs on the side of caution
1070 * to prevent excess iteration. Poor alignment can cause the allocator to
1071 * skip over blocks and chunks that have valid free areas.
1072 *
1073 * RETURNS:
1074 * The offset in the bitmap to begin searching.
1075 * -1 if no offset is found.
1076 */
1077static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1078 size_t align, bool pop_only)
1079{
1080 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1081 int bit_off, bits, next_off;
1082
1083 /*
1084 * Check to see if the allocation can fit in the chunk's contig hint.
1085 * This is an optimization to prevent scanning by assuming if it
1086 * cannot fit in the global hint, there is memory pressure and creating
1087 * a new chunk would happen soon.
1088 */
1089 bit_off = ALIGN(chunk_md->contig_hint_start, align) -
1090 chunk_md->contig_hint_start;
1091 if (bit_off + alloc_bits > chunk_md->contig_hint)
1092 return -1;
1093
1094 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1095 bits = 0;
1096 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1097 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1098 &next_off))
1099 break;
1100
1101 bit_off = next_off;
1102 bits = 0;
1103 }
1104
1105 if (bit_off == pcpu_chunk_map_bits(chunk))
1106 return -1;
1107
1108 return bit_off;
1109}
1110
1111/*
1112 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1113 * @map: the address to base the search on
1114 * @size: the bitmap size in bits
1115 * @start: the bitnumber to start searching at
1116 * @nr: the number of zeroed bits we're looking for
1117 * @align_mask: alignment mask for zero area
1118 * @largest_off: offset of the largest area skipped
1119 * @largest_bits: size of the largest area skipped
1120 *
1121 * The @align_mask should be one less than a power of 2.
1122 *
1123 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1124 * the largest area that was skipped. This is imperfect, but in general is
1125 * good enough. The largest remembered region is the largest failed region
1126 * seen. This does not include anything we possibly skipped due to alignment.
1127 * pcpu_block_update_scan() does scan backwards to try and recover what was
1128 * lost to alignment. While this can cause scanning to miss earlier possible
1129 * free areas, smaller allocations will eventually fill those holes.
1130 */
1131static unsigned long pcpu_find_zero_area(unsigned long *map,
1132 unsigned long size,
1133 unsigned long start,
1134 unsigned long nr,
1135 unsigned long align_mask,
1136 unsigned long *largest_off,
1137 unsigned long *largest_bits)
1138{
1139 unsigned long index, end, i, area_off, area_bits;
1140again:
1141 index = find_next_zero_bit(map, size, start);
1142
1143 /* Align allocation */
1144 index = __ALIGN_MASK(index, align_mask);
1145 area_off = index;
1146
1147 end = index + nr;
1148 if (end > size)
1149 return end;
1150 i = find_next_bit(map, end, index);
1151 if (i < end) {
1152 area_bits = i - area_off;
1153 /* remember largest unused area with best alignment */
1154 if (area_bits > *largest_bits ||
1155 (area_bits == *largest_bits && *largest_off &&
1156 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1157 *largest_off = area_off;
1158 *largest_bits = area_bits;
1159 }
1160
1161 start = i + 1;
1162 goto again;
1163 }
1164 return index;
1165}
1166
1167/**
1168 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1169 * @chunk: chunk of interest
1170 * @alloc_bits: size of request in allocation units
1171 * @align: alignment of area (max PAGE_SIZE)
1172 * @start: bit_off to start searching
1173 *
1174 * This function takes in a @start offset to begin searching to fit an
1175 * allocation of @alloc_bits with alignment @align. It needs to scan
1176 * the allocation map because if it fits within the block's contig hint,
1177 * @start will be block->first_free. This is an attempt to fill the
1178 * allocation prior to breaking the contig hint. The allocation and
1179 * boundary maps are updated accordingly if it confirms a valid
1180 * free area.
1181 *
1182 * RETURNS:
1183 * Allocated addr offset in @chunk on success.
1184 * -1 if no matching area is found.
1185 */
1186static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1187 size_t align, int start)
1188{
1189 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1190 size_t align_mask = (align) ? (align - 1) : 0;
1191 unsigned long area_off = 0, area_bits = 0;
1192 int bit_off, end, oslot;
1193
1194 lockdep_assert_held(&pcpu_lock);
1195
1196 oslot = pcpu_chunk_slot(chunk);
1197
1198 /*
1199 * Search to find a fit.
1200 */
1201 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1202 pcpu_chunk_map_bits(chunk));
1203 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1204 align_mask, &area_off, &area_bits);
1205 if (bit_off >= end)
1206 return -1;
1207
1208 if (area_bits)
1209 pcpu_block_update_scan(chunk, area_off, area_bits);
1210
1211 /* update alloc map */
1212 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1213
1214 /* update boundary map */
1215 set_bit(bit_off, chunk->bound_map);
1216 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1217 set_bit(bit_off + alloc_bits, chunk->bound_map);
1218
1219 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1220
1221 /* update first free bit */
1222 if (bit_off == chunk_md->first_free)
1223 chunk_md->first_free = find_next_zero_bit(
1224 chunk->alloc_map,
1225 pcpu_chunk_map_bits(chunk),
1226 bit_off + alloc_bits);
1227
1228 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1229
1230 pcpu_chunk_relocate(chunk, oslot);
1231
1232 return bit_off * PCPU_MIN_ALLOC_SIZE;
1233}
1234
1235/**
1236 * pcpu_free_area - frees the corresponding offset
1237 * @chunk: chunk of interest
1238 * @off: addr offset into chunk
1239 *
1240 * This function determines the size of an allocation to free using
1241 * the boundary bitmap and clears the allocation map.
1242 */
1243static void pcpu_free_area(struct pcpu_chunk *chunk, int off)
1244{
1245 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1246 int bit_off, bits, end, oslot;
1247
1248 lockdep_assert_held(&pcpu_lock);
1249 pcpu_stats_area_dealloc(chunk);
1250
1251 oslot = pcpu_chunk_slot(chunk);
1252
1253 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1254
1255 /* find end index */
1256 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1257 bit_off + 1);
1258 bits = end - bit_off;
1259 bitmap_clear(chunk->alloc_map, bit_off, bits);
1260
1261 /* update metadata */
1262 chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE;
1263
1264 /* update first free bit */
1265 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1266
1267 pcpu_block_update_hint_free(chunk, bit_off, bits);
1268
1269 pcpu_chunk_relocate(chunk, oslot);
1270}
1271
1272static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1273{
1274 block->scan_hint = 0;
1275 block->contig_hint = nr_bits;
1276 block->left_free = nr_bits;
1277 block->right_free = nr_bits;
1278 block->first_free = 0;
1279 block->nr_bits = nr_bits;
1280}
1281
1282static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1283{
1284 struct pcpu_block_md *md_block;
1285
1286 /* init the chunk's block */
1287 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1288
1289 for (md_block = chunk->md_blocks;
1290 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1291 md_block++)
1292 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1293}
1294
1295/**
1296 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1297 * @tmp_addr: the start of the region served
1298 * @map_size: size of the region served
1299 *
1300 * This is responsible for creating the chunks that serve the first chunk. The
1301 * base_addr is page aligned down of @tmp_addr while the region end is page
1302 * aligned up. Offsets are kept track of to determine the region served. All
1303 * this is done to appease the bitmap allocator in avoiding partial blocks.
1304 *
1305 * RETURNS:
1306 * Chunk serving the region at @tmp_addr of @map_size.
1307 */
1308static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1309 int map_size)
1310{
1311 struct pcpu_chunk *chunk;
1312 unsigned long aligned_addr, lcm_align;
1313 int start_offset, offset_bits, region_size, region_bits;
1314 size_t alloc_size;
1315
1316 /* region calculations */
1317 aligned_addr = tmp_addr & PAGE_MASK;
1318
1319 start_offset = tmp_addr - aligned_addr;
1320
1321 /*
1322 * Align the end of the region with the LCM of PAGE_SIZE and
1323 * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1324 * the other.
1325 */
1326 lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1327 region_size = ALIGN(start_offset + map_size, lcm_align);
1328
1329 /* allocate chunk */
1330 alloc_size = sizeof(struct pcpu_chunk) +
1331 BITS_TO_LONGS(region_size >> PAGE_SHIFT);
1332 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1333 if (!chunk)
1334 panic("%s: Failed to allocate %zu bytes\n", __func__,
1335 alloc_size);
1336
1337 INIT_LIST_HEAD(&chunk->list);
1338
1339 chunk->base_addr = (void *)aligned_addr;
1340 chunk->start_offset = start_offset;
1341 chunk->end_offset = region_size - chunk->start_offset - map_size;
1342
1343 chunk->nr_pages = region_size >> PAGE_SHIFT;
1344 region_bits = pcpu_chunk_map_bits(chunk);
1345
1346 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1347 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1348 if (!chunk->alloc_map)
1349 panic("%s: Failed to allocate %zu bytes\n", __func__,
1350 alloc_size);
1351
1352 alloc_size =
1353 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1354 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1355 if (!chunk->bound_map)
1356 panic("%s: Failed to allocate %zu bytes\n", __func__,
1357 alloc_size);
1358
1359 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1360 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1361 if (!chunk->md_blocks)
1362 panic("%s: Failed to allocate %zu bytes\n", __func__,
1363 alloc_size);
1364
1365 pcpu_init_md_blocks(chunk);
1366
1367 /* manage populated page bitmap */
1368 chunk->immutable = true;
1369 bitmap_fill(chunk->populated, chunk->nr_pages);
1370 chunk->nr_populated = chunk->nr_pages;
1371 chunk->nr_empty_pop_pages = chunk->nr_pages;
1372
1373 chunk->free_bytes = map_size;
1374
1375 if (chunk->start_offset) {
1376 /* hide the beginning of the bitmap */
1377 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1378 bitmap_set(chunk->alloc_map, 0, offset_bits);
1379 set_bit(0, chunk->bound_map);
1380 set_bit(offset_bits, chunk->bound_map);
1381
1382 chunk->chunk_md.first_free = offset_bits;
1383
1384 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1385 }
1386
1387 if (chunk->end_offset) {
1388 /* hide the end of the bitmap */
1389 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1390 bitmap_set(chunk->alloc_map,
1391 pcpu_chunk_map_bits(chunk) - offset_bits,
1392 offset_bits);
1393 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1394 chunk->bound_map);
1395 set_bit(region_bits, chunk->bound_map);
1396
1397 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1398 - offset_bits, offset_bits);
1399 }
1400
1401 return chunk;
1402}
1403
1404static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1405{
1406 struct pcpu_chunk *chunk;
1407 int region_bits;
1408
1409 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1410 if (!chunk)
1411 return NULL;
1412
1413 INIT_LIST_HEAD(&chunk->list);
1414 chunk->nr_pages = pcpu_unit_pages;
1415 region_bits = pcpu_chunk_map_bits(chunk);
1416
1417 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1418 sizeof(chunk->alloc_map[0]), gfp);
1419 if (!chunk->alloc_map)
1420 goto alloc_map_fail;
1421
1422 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1423 sizeof(chunk->bound_map[0]), gfp);
1424 if (!chunk->bound_map)
1425 goto bound_map_fail;
1426
1427 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1428 sizeof(chunk->md_blocks[0]), gfp);
1429 if (!chunk->md_blocks)
1430 goto md_blocks_fail;
1431
1432 pcpu_init_md_blocks(chunk);
1433
1434 /* init metadata */
1435 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1436
1437 return chunk;
1438
1439md_blocks_fail:
1440 pcpu_mem_free(chunk->bound_map);
1441bound_map_fail:
1442 pcpu_mem_free(chunk->alloc_map);
1443alloc_map_fail:
1444 pcpu_mem_free(chunk);
1445
1446 return NULL;
1447}
1448
1449static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1450{
1451 if (!chunk)
1452 return;
1453 pcpu_mem_free(chunk->md_blocks);
1454 pcpu_mem_free(chunk->bound_map);
1455 pcpu_mem_free(chunk->alloc_map);
1456 pcpu_mem_free(chunk);
1457}
1458
1459/**
1460 * pcpu_chunk_populated - post-population bookkeeping
1461 * @chunk: pcpu_chunk which got populated
1462 * @page_start: the start page
1463 * @page_end: the end page
1464 *
1465 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1466 * the bookkeeping information accordingly. Must be called after each
1467 * successful population.
1468 *
1469 * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1470 * is to serve an allocation in that area.
1471 */
1472static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1473 int page_end)
1474{
1475 int nr = page_end - page_start;
1476
1477 lockdep_assert_held(&pcpu_lock);
1478
1479 bitmap_set(chunk->populated, page_start, nr);
1480 chunk->nr_populated += nr;
1481 pcpu_nr_populated += nr;
1482
1483 pcpu_update_empty_pages(chunk, nr);
1484}
1485
1486/**
1487 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1488 * @chunk: pcpu_chunk which got depopulated
1489 * @page_start: the start page
1490 * @page_end: the end page
1491 *
1492 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1493 * Update the bookkeeping information accordingly. Must be called after
1494 * each successful depopulation.
1495 */
1496static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1497 int page_start, int page_end)
1498{
1499 int nr = page_end - page_start;
1500
1501 lockdep_assert_held(&pcpu_lock);
1502
1503 bitmap_clear(chunk->populated, page_start, nr);
1504 chunk->nr_populated -= nr;
1505 pcpu_nr_populated -= nr;
1506
1507 pcpu_update_empty_pages(chunk, -nr);
1508}
1509
1510/*
1511 * Chunk management implementation.
1512 *
1513 * To allow different implementations, chunk alloc/free and
1514 * [de]population are implemented in a separate file which is pulled
1515 * into this file and compiled together. The following functions
1516 * should be implemented.
1517 *
1518 * pcpu_populate_chunk - populate the specified range of a chunk
1519 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1520 * pcpu_create_chunk - create a new chunk
1521 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1522 * pcpu_addr_to_page - translate address to physical address
1523 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1524 */
1525static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1526 int page_start, int page_end, gfp_t gfp);
1527static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1528 int page_start, int page_end);
1529static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1530static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1531static struct page *pcpu_addr_to_page(void *addr);
1532static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1533
1534#ifdef CONFIG_NEED_PER_CPU_KM
1535#include "percpu-km.c"
1536#else
1537#include "percpu-vm.c"
1538#endif
1539
1540/**
1541 * pcpu_chunk_addr_search - determine chunk containing specified address
1542 * @addr: address for which the chunk needs to be determined.
1543 *
1544 * This is an internal function that handles all but static allocations.
1545 * Static percpu address values should never be passed into the allocator.
1546 *
1547 * RETURNS:
1548 * The address of the found chunk.
1549 */
1550static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1551{
1552 /* is it in the dynamic region (first chunk)? */
1553 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1554 return pcpu_first_chunk;
1555
1556 /* is it in the reserved region? */
1557 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1558 return pcpu_reserved_chunk;
1559
1560 /*
1561 * The address is relative to unit0 which might be unused and
1562 * thus unmapped. Offset the address to the unit space of the
1563 * current processor before looking it up in the vmalloc
1564 * space. Note that any possible cpu id can be used here, so
1565 * there's no need to worry about preemption or cpu hotplug.
1566 */
1567 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1568 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1569}
1570
1571/**
1572 * pcpu_alloc - the percpu allocator
1573 * @size: size of area to allocate in bytes
1574 * @align: alignment of area (max PAGE_SIZE)
1575 * @reserved: allocate from the reserved chunk if available
1576 * @gfp: allocation flags
1577 *
1578 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1579 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1580 * then no warning will be triggered on invalid or failed allocation
1581 * requests.
1582 *
1583 * RETURNS:
1584 * Percpu pointer to the allocated area on success, NULL on failure.
1585 */
1586static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1587 gfp_t gfp)
1588{
1589 /* whitelisted flags that can be passed to the backing allocators */
1590 gfp_t pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1591 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1592 bool do_warn = !(gfp & __GFP_NOWARN);
1593 static int warn_limit = 10;
1594 struct pcpu_chunk *chunk, *next;
1595 const char *err;
1596 int slot, off, cpu, ret;
1597 unsigned long flags;
1598 void __percpu *ptr;
1599 size_t bits, bit_align;
1600
1601 /*
1602 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1603 * therefore alignment must be a minimum of that many bytes.
1604 * An allocation may have internal fragmentation from rounding up
1605 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1606 */
1607 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1608 align = PCPU_MIN_ALLOC_SIZE;
1609
1610 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1611 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1612 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1613
1614 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1615 !is_power_of_2(align))) {
1616 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1617 size, align);
1618 return NULL;
1619 }
1620
1621 if (!is_atomic) {
1622 /*
1623 * pcpu_balance_workfn() allocates memory under this mutex,
1624 * and it may wait for memory reclaim. Allow current task
1625 * to become OOM victim, in case of memory pressure.
1626 */
1627 if (gfp & __GFP_NOFAIL)
1628 mutex_lock(&pcpu_alloc_mutex);
1629 else if (mutex_lock_killable(&pcpu_alloc_mutex))
1630 return NULL;
1631 }
1632
1633 spin_lock_irqsave(&pcpu_lock, flags);
1634
1635 /* serve reserved allocations from the reserved chunk if available */
1636 if (reserved && pcpu_reserved_chunk) {
1637 chunk = pcpu_reserved_chunk;
1638
1639 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1640 if (off < 0) {
1641 err = "alloc from reserved chunk failed";
1642 goto fail_unlock;
1643 }
1644
1645 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1646 if (off >= 0)
1647 goto area_found;
1648
1649 err = "alloc from reserved chunk failed";
1650 goto fail_unlock;
1651 }
1652
1653restart:
1654 /* search through normal chunks */
1655 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1656 list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) {
1657 off = pcpu_find_block_fit(chunk, bits, bit_align,
1658 is_atomic);
1659 if (off < 0) {
1660 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1661 pcpu_chunk_move(chunk, 0);
1662 continue;
1663 }
1664
1665 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1666 if (off >= 0)
1667 goto area_found;
1668
1669 }
1670 }
1671
1672 spin_unlock_irqrestore(&pcpu_lock, flags);
1673
1674 /*
1675 * No space left. Create a new chunk. We don't want multiple
1676 * tasks to create chunks simultaneously. Serialize and create iff
1677 * there's still no empty chunk after grabbing the mutex.
1678 */
1679 if (is_atomic) {
1680 err = "atomic alloc failed, no space left";
1681 goto fail;
1682 }
1683
1684 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1685 chunk = pcpu_create_chunk(pcpu_gfp);
1686 if (!chunk) {
1687 err = "failed to allocate new chunk";
1688 goto fail;
1689 }
1690
1691 spin_lock_irqsave(&pcpu_lock, flags);
1692 pcpu_chunk_relocate(chunk, -1);
1693 } else {
1694 spin_lock_irqsave(&pcpu_lock, flags);
1695 }
1696
1697 goto restart;
1698
1699area_found:
1700 pcpu_stats_area_alloc(chunk, size);
1701 spin_unlock_irqrestore(&pcpu_lock, flags);
1702
1703 /* populate if not all pages are already there */
1704 if (!is_atomic) {
1705 int page_start, page_end, rs, re;
1706
1707 page_start = PFN_DOWN(off);
1708 page_end = PFN_UP(off + size);
1709
1710 pcpu_for_each_unpop_region(chunk->populated, rs, re,
1711 page_start, page_end) {
1712 WARN_ON(chunk->immutable);
1713
1714 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1715
1716 spin_lock_irqsave(&pcpu_lock, flags);
1717 if (ret) {
1718 pcpu_free_area(chunk, off);
1719 err = "failed to populate";
1720 goto fail_unlock;
1721 }
1722 pcpu_chunk_populated(chunk, rs, re);
1723 spin_unlock_irqrestore(&pcpu_lock, flags);
1724 }
1725
1726 mutex_unlock(&pcpu_alloc_mutex);
1727 }
1728
1729 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1730 pcpu_schedule_balance_work();
1731
1732 /* clear the areas and return address relative to base address */
1733 for_each_possible_cpu(cpu)
1734 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1735
1736 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1737 kmemleak_alloc_percpu(ptr, size, gfp);
1738
1739 trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1740 chunk->base_addr, off, ptr);
1741
1742 return ptr;
1743
1744fail_unlock:
1745 spin_unlock_irqrestore(&pcpu_lock, flags);
1746fail:
1747 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1748
1749 if (!is_atomic && do_warn && warn_limit) {
1750 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1751 size, align, is_atomic, err);
1752 dump_stack();
1753 if (!--warn_limit)
1754 pr_info("limit reached, disable warning\n");
1755 }
1756 if (is_atomic) {
1757 /* see the flag handling in pcpu_blance_workfn() */
1758 pcpu_atomic_alloc_failed = true;
1759 pcpu_schedule_balance_work();
1760 } else {
1761 mutex_unlock(&pcpu_alloc_mutex);
1762 }
1763 return NULL;
1764}
1765
1766/**
1767 * __alloc_percpu_gfp - allocate dynamic percpu area
1768 * @size: size of area to allocate in bytes
1769 * @align: alignment of area (max PAGE_SIZE)
1770 * @gfp: allocation flags
1771 *
1772 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1773 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1774 * be called from any context but is a lot more likely to fail. If @gfp
1775 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1776 * allocation requests.
1777 *
1778 * RETURNS:
1779 * Percpu pointer to the allocated area on success, NULL on failure.
1780 */
1781void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1782{
1783 return pcpu_alloc(size, align, false, gfp);
1784}
1785EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1786
1787/**
1788 * __alloc_percpu - allocate dynamic percpu area
1789 * @size: size of area to allocate in bytes
1790 * @align: alignment of area (max PAGE_SIZE)
1791 *
1792 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1793 */
1794void __percpu *__alloc_percpu(size_t size, size_t align)
1795{
1796 return pcpu_alloc(size, align, false, GFP_KERNEL);
1797}
1798EXPORT_SYMBOL_GPL(__alloc_percpu);
1799
1800/**
1801 * __alloc_reserved_percpu - allocate reserved percpu area
1802 * @size: size of area to allocate in bytes
1803 * @align: alignment of area (max PAGE_SIZE)
1804 *
1805 * Allocate zero-filled percpu area of @size bytes aligned at @align
1806 * from reserved percpu area if arch has set it up; otherwise,
1807 * allocation is served from the same dynamic area. Might sleep.
1808 * Might trigger writeouts.
1809 *
1810 * CONTEXT:
1811 * Does GFP_KERNEL allocation.
1812 *
1813 * RETURNS:
1814 * Percpu pointer to the allocated area on success, NULL on failure.
1815 */
1816void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1817{
1818 return pcpu_alloc(size, align, true, GFP_KERNEL);
1819}
1820
1821/**
1822 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1823 * @work: unused
1824 *
1825 * Reclaim all fully free chunks except for the first one. This is also
1826 * responsible for maintaining the pool of empty populated pages. However,
1827 * it is possible that this is called when physical memory is scarce causing
1828 * OOM killer to be triggered. We should avoid doing so until an actual
1829 * allocation causes the failure as it is possible that requests can be
1830 * serviced from already backed regions.
1831 */
1832static void pcpu_balance_workfn(struct work_struct *work)
1833{
1834 /* gfp flags passed to underlying allocators */
1835 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
1836 LIST_HEAD(to_free);
1837 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1838 struct pcpu_chunk *chunk, *next;
1839 int slot, nr_to_pop, ret;
1840
1841 /*
1842 * There's no reason to keep around multiple unused chunks and VM
1843 * areas can be scarce. Destroy all free chunks except for one.
1844 */
1845 mutex_lock(&pcpu_alloc_mutex);
1846 spin_lock_irq(&pcpu_lock);
1847
1848 list_for_each_entry_safe(chunk, next, free_head, list) {
1849 WARN_ON(chunk->immutable);
1850
1851 /* spare the first one */
1852 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1853 continue;
1854
1855 list_move(&chunk->list, &to_free);
1856 }
1857
1858 spin_unlock_irq(&pcpu_lock);
1859
1860 list_for_each_entry_safe(chunk, next, &to_free, list) {
1861 int rs, re;
1862
1863 pcpu_for_each_pop_region(chunk->populated, rs, re, 0,
1864 chunk->nr_pages) {
1865 pcpu_depopulate_chunk(chunk, rs, re);
1866 spin_lock_irq(&pcpu_lock);
1867 pcpu_chunk_depopulated(chunk, rs, re);
1868 spin_unlock_irq(&pcpu_lock);
1869 }
1870 pcpu_destroy_chunk(chunk);
1871 cond_resched();
1872 }
1873
1874 /*
1875 * Ensure there are certain number of free populated pages for
1876 * atomic allocs. Fill up from the most packed so that atomic
1877 * allocs don't increase fragmentation. If atomic allocation
1878 * failed previously, always populate the maximum amount. This
1879 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1880 * failing indefinitely; however, large atomic allocs are not
1881 * something we support properly and can be highly unreliable and
1882 * inefficient.
1883 */
1884retry_pop:
1885 if (pcpu_atomic_alloc_failed) {
1886 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1887 /* best effort anyway, don't worry about synchronization */
1888 pcpu_atomic_alloc_failed = false;
1889 } else {
1890 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1891 pcpu_nr_empty_pop_pages,
1892 0, PCPU_EMPTY_POP_PAGES_HIGH);
1893 }
1894
1895 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1896 int nr_unpop = 0, rs, re;
1897
1898 if (!nr_to_pop)
1899 break;
1900
1901 spin_lock_irq(&pcpu_lock);
1902 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1903 nr_unpop = chunk->nr_pages - chunk->nr_populated;
1904 if (nr_unpop)
1905 break;
1906 }
1907 spin_unlock_irq(&pcpu_lock);
1908
1909 if (!nr_unpop)
1910 continue;
1911
1912 /* @chunk can't go away while pcpu_alloc_mutex is held */
1913 pcpu_for_each_unpop_region(chunk->populated, rs, re, 0,
1914 chunk->nr_pages) {
1915 int nr = min(re - rs, nr_to_pop);
1916
1917 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
1918 if (!ret) {
1919 nr_to_pop -= nr;
1920 spin_lock_irq(&pcpu_lock);
1921 pcpu_chunk_populated(chunk, rs, rs + nr);
1922 spin_unlock_irq(&pcpu_lock);
1923 } else {
1924 nr_to_pop = 0;
1925 }
1926
1927 if (!nr_to_pop)
1928 break;
1929 }
1930 }
1931
1932 if (nr_to_pop) {
1933 /* ran out of chunks to populate, create a new one and retry */
1934 chunk = pcpu_create_chunk(gfp);
1935 if (chunk) {
1936 spin_lock_irq(&pcpu_lock);
1937 pcpu_chunk_relocate(chunk, -1);
1938 spin_unlock_irq(&pcpu_lock);
1939 goto retry_pop;
1940 }
1941 }
1942
1943 mutex_unlock(&pcpu_alloc_mutex);
1944}
1945
1946/**
1947 * free_percpu - free percpu area
1948 * @ptr: pointer to area to free
1949 *
1950 * Free percpu area @ptr.
1951 *
1952 * CONTEXT:
1953 * Can be called from atomic context.
1954 */
1955void free_percpu(void __percpu *ptr)
1956{
1957 void *addr;
1958 struct pcpu_chunk *chunk;
1959 unsigned long flags;
1960 int off;
1961 bool need_balance = false;
1962
1963 if (!ptr)
1964 return;
1965
1966 kmemleak_free_percpu(ptr);
1967
1968 addr = __pcpu_ptr_to_addr(ptr);
1969
1970 spin_lock_irqsave(&pcpu_lock, flags);
1971
1972 chunk = pcpu_chunk_addr_search(addr);
1973 off = addr - chunk->base_addr;
1974
1975 pcpu_free_area(chunk, off);
1976
1977 /* if there are more than one fully free chunks, wake up grim reaper */
1978 if (chunk->free_bytes == pcpu_unit_size) {
1979 struct pcpu_chunk *pos;
1980
1981 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1982 if (pos != chunk) {
1983 need_balance = true;
1984 break;
1985 }
1986 }
1987
1988 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1989
1990 spin_unlock_irqrestore(&pcpu_lock, flags);
1991
1992 if (need_balance)
1993 pcpu_schedule_balance_work();
1994}
1995EXPORT_SYMBOL_GPL(free_percpu);
1996
1997bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1998{
1999#ifdef CONFIG_SMP
2000 const size_t static_size = __per_cpu_end - __per_cpu_start;
2001 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2002 unsigned int cpu;
2003
2004 for_each_possible_cpu(cpu) {
2005 void *start = per_cpu_ptr(base, cpu);
2006 void *va = (void *)addr;
2007
2008 if (va >= start && va < start + static_size) {
2009 if (can_addr) {
2010 *can_addr = (unsigned long) (va - start);
2011 *can_addr += (unsigned long)
2012 per_cpu_ptr(base, get_boot_cpu_id());
2013 }
2014 return true;
2015 }
2016 }
2017#endif
2018 /* on UP, can't distinguish from other static vars, always false */
2019 return false;
2020}
2021
2022/**
2023 * is_kernel_percpu_address - test whether address is from static percpu area
2024 * @addr: address to test
2025 *
2026 * Test whether @addr belongs to in-kernel static percpu area. Module
2027 * static percpu areas are not considered. For those, use
2028 * is_module_percpu_address().
2029 *
2030 * RETURNS:
2031 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2032 */
2033bool is_kernel_percpu_address(unsigned long addr)
2034{
2035 return __is_kernel_percpu_address(addr, NULL);
2036}
2037
2038/**
2039 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2040 * @addr: the address to be converted to physical address
2041 *
2042 * Given @addr which is dereferenceable address obtained via one of
2043 * percpu access macros, this function translates it into its physical
2044 * address. The caller is responsible for ensuring @addr stays valid
2045 * until this function finishes.
2046 *
2047 * percpu allocator has special setup for the first chunk, which currently
2048 * supports either embedding in linear address space or vmalloc mapping,
2049 * and, from the second one, the backing allocator (currently either vm or
2050 * km) provides translation.
2051 *
2052 * The addr can be translated simply without checking if it falls into the
2053 * first chunk. But the current code reflects better how percpu allocator
2054 * actually works, and the verification can discover both bugs in percpu
2055 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2056 * code.
2057 *
2058 * RETURNS:
2059 * The physical address for @addr.
2060 */
2061phys_addr_t per_cpu_ptr_to_phys(void *addr)
2062{
2063 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2064 bool in_first_chunk = false;
2065 unsigned long first_low, first_high;
2066 unsigned int cpu;
2067
2068 /*
2069 * The following test on unit_low/high isn't strictly
2070 * necessary but will speed up lookups of addresses which
2071 * aren't in the first chunk.
2072 *
2073 * The address check is against full chunk sizes. pcpu_base_addr
2074 * points to the beginning of the first chunk including the
2075 * static region. Assumes good intent as the first chunk may
2076 * not be full (ie. < pcpu_unit_pages in size).
2077 */
2078 first_low = (unsigned long)pcpu_base_addr +
2079 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2080 first_high = (unsigned long)pcpu_base_addr +
2081 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2082 if ((unsigned long)addr >= first_low &&
2083 (unsigned long)addr < first_high) {
2084 for_each_possible_cpu(cpu) {
2085 void *start = per_cpu_ptr(base, cpu);
2086
2087 if (addr >= start && addr < start + pcpu_unit_size) {
2088 in_first_chunk = true;
2089 break;
2090 }
2091 }
2092 }
2093
2094 if (in_first_chunk) {
2095 if (!is_vmalloc_addr(addr))
2096 return __pa(addr);
2097 else
2098 return page_to_phys(vmalloc_to_page(addr)) +
2099 offset_in_page(addr);
2100 } else
2101 return page_to_phys(pcpu_addr_to_page(addr)) +
2102 offset_in_page(addr);
2103}
2104
2105/**
2106 * pcpu_alloc_alloc_info - allocate percpu allocation info
2107 * @nr_groups: the number of groups
2108 * @nr_units: the number of units
2109 *
2110 * Allocate ai which is large enough for @nr_groups groups containing
2111 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2112 * cpu_map array which is long enough for @nr_units and filled with
2113 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2114 * pointer of other groups.
2115 *
2116 * RETURNS:
2117 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2118 * failure.
2119 */
2120struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2121 int nr_units)
2122{
2123 struct pcpu_alloc_info *ai;
2124 size_t base_size, ai_size;
2125 void *ptr;
2126 int unit;
2127
2128 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2129 __alignof__(ai->groups[0].cpu_map[0]));
2130 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2131
2132 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2133 if (!ptr)
2134 return NULL;
2135 ai = ptr;
2136 ptr += base_size;
2137
2138 ai->groups[0].cpu_map = ptr;
2139
2140 for (unit = 0; unit < nr_units; unit++)
2141 ai->groups[0].cpu_map[unit] = NR_CPUS;
2142
2143 ai->nr_groups = nr_groups;
2144 ai->__ai_size = PFN_ALIGN(ai_size);
2145
2146 return ai;
2147}
2148
2149/**
2150 * pcpu_free_alloc_info - free percpu allocation info
2151 * @ai: pcpu_alloc_info to free
2152 *
2153 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2154 */
2155void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2156{
2157 memblock_free_early(__pa(ai), ai->__ai_size);
2158}
2159
2160/**
2161 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2162 * @lvl: loglevel
2163 * @ai: allocation info to dump
2164 *
2165 * Print out information about @ai using loglevel @lvl.
2166 */
2167static void pcpu_dump_alloc_info(const char *lvl,
2168 const struct pcpu_alloc_info *ai)
2169{
2170 int group_width = 1, cpu_width = 1, width;
2171 char empty_str[] = "--------";
2172 int alloc = 0, alloc_end = 0;
2173 int group, v;
2174 int upa, apl; /* units per alloc, allocs per line */
2175
2176 v = ai->nr_groups;
2177 while (v /= 10)
2178 group_width++;
2179
2180 v = num_possible_cpus();
2181 while (v /= 10)
2182 cpu_width++;
2183 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2184
2185 upa = ai->alloc_size / ai->unit_size;
2186 width = upa * (cpu_width + 1) + group_width + 3;
2187 apl = rounddown_pow_of_two(max(60 / width, 1));
2188
2189 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2190 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2191 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2192
2193 for (group = 0; group < ai->nr_groups; group++) {
2194 const struct pcpu_group_info *gi = &ai->groups[group];
2195 int unit = 0, unit_end = 0;
2196
2197 BUG_ON(gi->nr_units % upa);
2198 for (alloc_end += gi->nr_units / upa;
2199 alloc < alloc_end; alloc++) {
2200 if (!(alloc % apl)) {
2201 pr_cont("\n");
2202 printk("%spcpu-alloc: ", lvl);
2203 }
2204 pr_cont("[%0*d] ", group_width, group);
2205
2206 for (unit_end += upa; unit < unit_end; unit++)
2207 if (gi->cpu_map[unit] != NR_CPUS)
2208 pr_cont("%0*d ",
2209 cpu_width, gi->cpu_map[unit]);
2210 else
2211 pr_cont("%s ", empty_str);
2212 }
2213 }
2214 pr_cont("\n");
2215}
2216
2217/**
2218 * pcpu_setup_first_chunk - initialize the first percpu chunk
2219 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2220 * @base_addr: mapped address
2221 *
2222 * Initialize the first percpu chunk which contains the kernel static
2223 * percpu area. This function is to be called from arch percpu area
2224 * setup path.
2225 *
2226 * @ai contains all information necessary to initialize the first
2227 * chunk and prime the dynamic percpu allocator.
2228 *
2229 * @ai->static_size is the size of static percpu area.
2230 *
2231 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2232 * reserve after the static area in the first chunk. This reserves
2233 * the first chunk such that it's available only through reserved
2234 * percpu allocation. This is primarily used to serve module percpu
2235 * static areas on architectures where the addressing model has
2236 * limited offset range for symbol relocations to guarantee module
2237 * percpu symbols fall inside the relocatable range.
2238 *
2239 * @ai->dyn_size determines the number of bytes available for dynamic
2240 * allocation in the first chunk. The area between @ai->static_size +
2241 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2242 *
2243 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2244 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2245 * @ai->dyn_size.
2246 *
2247 * @ai->atom_size is the allocation atom size and used as alignment
2248 * for vm areas.
2249 *
2250 * @ai->alloc_size is the allocation size and always multiple of
2251 * @ai->atom_size. This is larger than @ai->atom_size if
2252 * @ai->unit_size is larger than @ai->atom_size.
2253 *
2254 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2255 * percpu areas. Units which should be colocated are put into the
2256 * same group. Dynamic VM areas will be allocated according to these
2257 * groupings. If @ai->nr_groups is zero, a single group containing
2258 * all units is assumed.
2259 *
2260 * The caller should have mapped the first chunk at @base_addr and
2261 * copied static data to each unit.
2262 *
2263 * The first chunk will always contain a static and a dynamic region.
2264 * However, the static region is not managed by any chunk. If the first
2265 * chunk also contains a reserved region, it is served by two chunks -
2266 * one for the reserved region and one for the dynamic region. They
2267 * share the same vm, but use offset regions in the area allocation map.
2268 * The chunk serving the dynamic region is circulated in the chunk slots
2269 * and available for dynamic allocation like any other chunk.
2270 */
2271void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2272 void *base_addr)
2273{
2274 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2275 size_t static_size, dyn_size;
2276 struct pcpu_chunk *chunk;
2277 unsigned long *group_offsets;
2278 size_t *group_sizes;
2279 unsigned long *unit_off;
2280 unsigned int cpu;
2281 int *unit_map;
2282 int group, unit, i;
2283 int map_size;
2284 unsigned long tmp_addr;
2285 size_t alloc_size;
2286
2287#define PCPU_SETUP_BUG_ON(cond) do { \
2288 if (unlikely(cond)) { \
2289 pr_emerg("failed to initialize, %s\n", #cond); \
2290 pr_emerg("cpu_possible_mask=%*pb\n", \
2291 cpumask_pr_args(cpu_possible_mask)); \
2292 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2293 BUG(); \
2294 } \
2295} while (0)
2296
2297 /* sanity checks */
2298 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2299#ifdef CONFIG_SMP
2300 PCPU_SETUP_BUG_ON(!ai->static_size);
2301 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2302#endif
2303 PCPU_SETUP_BUG_ON(!base_addr);
2304 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2305 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2306 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2307 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2308 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2309 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2310 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2311 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2312 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2313 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2314 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2315
2316 /* process group information and build config tables accordingly */
2317 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2318 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2319 if (!group_offsets)
2320 panic("%s: Failed to allocate %zu bytes\n", __func__,
2321 alloc_size);
2322
2323 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2324 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2325 if (!group_sizes)
2326 panic("%s: Failed to allocate %zu bytes\n", __func__,
2327 alloc_size);
2328
2329 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2330 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2331 if (!unit_map)
2332 panic("%s: Failed to allocate %zu bytes\n", __func__,
2333 alloc_size);
2334
2335 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2336 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2337 if (!unit_off)
2338 panic("%s: Failed to allocate %zu bytes\n", __func__,
2339 alloc_size);
2340
2341 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2342 unit_map[cpu] = UINT_MAX;
2343
2344 pcpu_low_unit_cpu = NR_CPUS;
2345 pcpu_high_unit_cpu = NR_CPUS;
2346
2347 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2348 const struct pcpu_group_info *gi = &ai->groups[group];
2349
2350 group_offsets[group] = gi->base_offset;
2351 group_sizes[group] = gi->nr_units * ai->unit_size;
2352
2353 for (i = 0; i < gi->nr_units; i++) {
2354 cpu = gi->cpu_map[i];
2355 if (cpu == NR_CPUS)
2356 continue;
2357
2358 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2359 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2360 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2361
2362 unit_map[cpu] = unit + i;
2363 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2364
2365 /* determine low/high unit_cpu */
2366 if (pcpu_low_unit_cpu == NR_CPUS ||
2367 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2368 pcpu_low_unit_cpu = cpu;
2369 if (pcpu_high_unit_cpu == NR_CPUS ||
2370 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2371 pcpu_high_unit_cpu = cpu;
2372 }
2373 }
2374 pcpu_nr_units = unit;
2375
2376 for_each_possible_cpu(cpu)
2377 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2378
2379 /* we're done parsing the input, undefine BUG macro and dump config */
2380#undef PCPU_SETUP_BUG_ON
2381 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2382
2383 pcpu_nr_groups = ai->nr_groups;
2384 pcpu_group_offsets = group_offsets;
2385 pcpu_group_sizes = group_sizes;
2386 pcpu_unit_map = unit_map;
2387 pcpu_unit_offsets = unit_off;
2388
2389 /* determine basic parameters */
2390 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2391 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2392 pcpu_atom_size = ai->atom_size;
2393 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2394 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2395
2396 pcpu_stats_save_ai(ai);
2397
2398 /*
2399 * Allocate chunk slots. The additional last slot is for
2400 * empty chunks.
2401 */
2402 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2403 pcpu_slot = memblock_alloc(pcpu_nr_slots * sizeof(pcpu_slot[0]),
2404 SMP_CACHE_BYTES);
2405 if (!pcpu_slot)
2406 panic("%s: Failed to allocate %zu bytes\n", __func__,
2407 pcpu_nr_slots * sizeof(pcpu_slot[0]));
2408 for (i = 0; i < pcpu_nr_slots; i++)
2409 INIT_LIST_HEAD(&pcpu_slot[i]);
2410
2411 /*
2412 * The end of the static region needs to be aligned with the
2413 * minimum allocation size as this offsets the reserved and
2414 * dynamic region. The first chunk ends page aligned by
2415 * expanding the dynamic region, therefore the dynamic region
2416 * can be shrunk to compensate while still staying above the
2417 * configured sizes.
2418 */
2419 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2420 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2421
2422 /*
2423 * Initialize first chunk.
2424 * If the reserved_size is non-zero, this initializes the reserved
2425 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2426 * and the dynamic region is initialized here. The first chunk,
2427 * pcpu_first_chunk, will always point to the chunk that serves
2428 * the dynamic region.
2429 */
2430 tmp_addr = (unsigned long)base_addr + static_size;
2431 map_size = ai->reserved_size ?: dyn_size;
2432 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2433
2434 /* init dynamic chunk if necessary */
2435 if (ai->reserved_size) {
2436 pcpu_reserved_chunk = chunk;
2437
2438 tmp_addr = (unsigned long)base_addr + static_size +
2439 ai->reserved_size;
2440 map_size = dyn_size;
2441 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2442 }
2443
2444 /* link the first chunk in */
2445 pcpu_first_chunk = chunk;
2446 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2447 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2448
2449 /* include all regions of the first chunk */
2450 pcpu_nr_populated += PFN_DOWN(size_sum);
2451
2452 pcpu_stats_chunk_alloc();
2453 trace_percpu_create_chunk(base_addr);
2454
2455 /* we're done */
2456 pcpu_base_addr = base_addr;
2457}
2458
2459#ifdef CONFIG_SMP
2460
2461const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2462 [PCPU_FC_AUTO] = "auto",
2463 [PCPU_FC_EMBED] = "embed",
2464 [PCPU_FC_PAGE] = "page",
2465};
2466
2467enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2468
2469static int __init percpu_alloc_setup(char *str)
2470{
2471 if (!str)
2472 return -EINVAL;
2473
2474 if (0)
2475 /* nada */;
2476#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2477 else if (!strcmp(str, "embed"))
2478 pcpu_chosen_fc = PCPU_FC_EMBED;
2479#endif
2480#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2481 else if (!strcmp(str, "page"))
2482 pcpu_chosen_fc = PCPU_FC_PAGE;
2483#endif
2484 else
2485 pr_warn("unknown allocator %s specified\n", str);
2486
2487 return 0;
2488}
2489early_param("percpu_alloc", percpu_alloc_setup);
2490
2491/*
2492 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2493 * Build it if needed by the arch config or the generic setup is going
2494 * to be used.
2495 */
2496#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2497 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2498#define BUILD_EMBED_FIRST_CHUNK
2499#endif
2500
2501/* build pcpu_page_first_chunk() iff needed by the arch config */
2502#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2503#define BUILD_PAGE_FIRST_CHUNK
2504#endif
2505
2506/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2507#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2508/**
2509 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2510 * @reserved_size: the size of reserved percpu area in bytes
2511 * @dyn_size: minimum free size for dynamic allocation in bytes
2512 * @atom_size: allocation atom size
2513 * @cpu_distance_fn: callback to determine distance between cpus, optional
2514 *
2515 * This function determines grouping of units, their mappings to cpus
2516 * and other parameters considering needed percpu size, allocation
2517 * atom size and distances between CPUs.
2518 *
2519 * Groups are always multiples of atom size and CPUs which are of
2520 * LOCAL_DISTANCE both ways are grouped together and share space for
2521 * units in the same group. The returned configuration is guaranteed
2522 * to have CPUs on different nodes on different groups and >=75% usage
2523 * of allocated virtual address space.
2524 *
2525 * RETURNS:
2526 * On success, pointer to the new allocation_info is returned. On
2527 * failure, ERR_PTR value is returned.
2528 */
2529static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2530 size_t reserved_size, size_t dyn_size,
2531 size_t atom_size,
2532 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2533{
2534 static int group_map[NR_CPUS] __initdata;
2535 static int group_cnt[NR_CPUS] __initdata;
2536 const size_t static_size = __per_cpu_end - __per_cpu_start;
2537 int nr_groups = 1, nr_units = 0;
2538 size_t size_sum, min_unit_size, alloc_size;
2539 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
2540 int last_allocs, group, unit;
2541 unsigned int cpu, tcpu;
2542 struct pcpu_alloc_info *ai;
2543 unsigned int *cpu_map;
2544
2545 /* this function may be called multiple times */
2546 memset(group_map, 0, sizeof(group_map));
2547 memset(group_cnt, 0, sizeof(group_cnt));
2548
2549 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2550 size_sum = PFN_ALIGN(static_size + reserved_size +
2551 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2552 dyn_size = size_sum - static_size - reserved_size;
2553
2554 /*
2555 * Determine min_unit_size, alloc_size and max_upa such that
2556 * alloc_size is multiple of atom_size and is the smallest
2557 * which can accommodate 4k aligned segments which are equal to
2558 * or larger than min_unit_size.
2559 */
2560 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2561
2562 /* determine the maximum # of units that can fit in an allocation */
2563 alloc_size = roundup(min_unit_size, atom_size);
2564 upa = alloc_size / min_unit_size;
2565 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2566 upa--;
2567 max_upa = upa;
2568
2569 /* group cpus according to their proximity */
2570 for_each_possible_cpu(cpu) {
2571 group = 0;
2572 next_group:
2573 for_each_possible_cpu(tcpu) {
2574 if (cpu == tcpu)
2575 break;
2576 if (group_map[tcpu] == group && cpu_distance_fn &&
2577 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2578 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2579 group++;
2580 nr_groups = max(nr_groups, group + 1);
2581 goto next_group;
2582 }
2583 }
2584 group_map[cpu] = group;
2585 group_cnt[group]++;
2586 }
2587
2588 /*
2589 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2590 * Expand the unit_size until we use >= 75% of the units allocated.
2591 * Related to atom_size, which could be much larger than the unit_size.
2592 */
2593 last_allocs = INT_MAX;
2594 for (upa = max_upa; upa; upa--) {
2595 int allocs = 0, wasted = 0;
2596
2597 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2598 continue;
2599
2600 for (group = 0; group < nr_groups; group++) {
2601 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2602 allocs += this_allocs;
2603 wasted += this_allocs * upa - group_cnt[group];
2604 }
2605
2606 /*
2607 * Don't accept if wastage is over 1/3. The
2608 * greater-than comparison ensures upa==1 always
2609 * passes the following check.
2610 */
2611 if (wasted > num_possible_cpus() / 3)
2612 continue;
2613
2614 /* and then don't consume more memory */
2615 if (allocs > last_allocs)
2616 break;
2617 last_allocs = allocs;
2618 best_upa = upa;
2619 }
2620 upa = best_upa;
2621
2622 /* allocate and fill alloc_info */
2623 for (group = 0; group < nr_groups; group++)
2624 nr_units += roundup(group_cnt[group], upa);
2625
2626 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2627 if (!ai)
2628 return ERR_PTR(-ENOMEM);
2629 cpu_map = ai->groups[0].cpu_map;
2630
2631 for (group = 0; group < nr_groups; group++) {
2632 ai->groups[group].cpu_map = cpu_map;
2633 cpu_map += roundup(group_cnt[group], upa);
2634 }
2635
2636 ai->static_size = static_size;
2637 ai->reserved_size = reserved_size;
2638 ai->dyn_size = dyn_size;
2639 ai->unit_size = alloc_size / upa;
2640 ai->atom_size = atom_size;
2641 ai->alloc_size = alloc_size;
2642
2643 for (group = 0, unit = 0; group < nr_groups; group++) {
2644 struct pcpu_group_info *gi = &ai->groups[group];
2645
2646 /*
2647 * Initialize base_offset as if all groups are located
2648 * back-to-back. The caller should update this to
2649 * reflect actual allocation.
2650 */
2651 gi->base_offset = unit * ai->unit_size;
2652
2653 for_each_possible_cpu(cpu)
2654 if (group_map[cpu] == group)
2655 gi->cpu_map[gi->nr_units++] = cpu;
2656 gi->nr_units = roundup(gi->nr_units, upa);
2657 unit += gi->nr_units;
2658 }
2659 BUG_ON(unit != nr_units);
2660
2661 return ai;
2662}
2663#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2664
2665#if defined(BUILD_EMBED_FIRST_CHUNK)
2666/**
2667 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2668 * @reserved_size: the size of reserved percpu area in bytes
2669 * @dyn_size: minimum free size for dynamic allocation in bytes
2670 * @atom_size: allocation atom size
2671 * @cpu_distance_fn: callback to determine distance between cpus, optional
2672 * @alloc_fn: function to allocate percpu page
2673 * @free_fn: function to free percpu page
2674 *
2675 * This is a helper to ease setting up embedded first percpu chunk and
2676 * can be called where pcpu_setup_first_chunk() is expected.
2677 *
2678 * If this function is used to setup the first chunk, it is allocated
2679 * by calling @alloc_fn and used as-is without being mapped into
2680 * vmalloc area. Allocations are always whole multiples of @atom_size
2681 * aligned to @atom_size.
2682 *
2683 * This enables the first chunk to piggy back on the linear physical
2684 * mapping which often uses larger page size. Please note that this
2685 * can result in very sparse cpu->unit mapping on NUMA machines thus
2686 * requiring large vmalloc address space. Don't use this allocator if
2687 * vmalloc space is not orders of magnitude larger than distances
2688 * between node memory addresses (ie. 32bit NUMA machines).
2689 *
2690 * @dyn_size specifies the minimum dynamic area size.
2691 *
2692 * If the needed size is smaller than the minimum or specified unit
2693 * size, the leftover is returned using @free_fn.
2694 *
2695 * RETURNS:
2696 * 0 on success, -errno on failure.
2697 */
2698int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2699 size_t atom_size,
2700 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2701 pcpu_fc_alloc_fn_t alloc_fn,
2702 pcpu_fc_free_fn_t free_fn)
2703{
2704 void *base = (void *)ULONG_MAX;
2705 void **areas = NULL;
2706 struct pcpu_alloc_info *ai;
2707 size_t size_sum, areas_size;
2708 unsigned long max_distance;
2709 int group, i, highest_group, rc = 0;
2710
2711 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2712 cpu_distance_fn);
2713 if (IS_ERR(ai))
2714 return PTR_ERR(ai);
2715
2716 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2717 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2718
2719 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
2720 if (!areas) {
2721 rc = -ENOMEM;
2722 goto out_free;
2723 }
2724
2725 /* allocate, copy and determine base address & max_distance */
2726 highest_group = 0;
2727 for (group = 0; group < ai->nr_groups; group++) {
2728 struct pcpu_group_info *gi = &ai->groups[group];
2729 unsigned int cpu = NR_CPUS;
2730 void *ptr;
2731
2732 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2733 cpu = gi->cpu_map[i];
2734 BUG_ON(cpu == NR_CPUS);
2735
2736 /* allocate space for the whole group */
2737 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2738 if (!ptr) {
2739 rc = -ENOMEM;
2740 goto out_free_areas;
2741 }
2742 /* kmemleak tracks the percpu allocations separately */
2743 kmemleak_free(ptr);
2744 areas[group] = ptr;
2745
2746 base = min(ptr, base);
2747 if (ptr > areas[highest_group])
2748 highest_group = group;
2749 }
2750 max_distance = areas[highest_group] - base;
2751 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2752
2753 /* warn if maximum distance is further than 75% of vmalloc space */
2754 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2755 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2756 max_distance, VMALLOC_TOTAL);
2757#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2758 /* and fail if we have fallback */
2759 rc = -EINVAL;
2760 goto out_free_areas;
2761#endif
2762 }
2763
2764 /*
2765 * Copy data and free unused parts. This should happen after all
2766 * allocations are complete; otherwise, we may end up with
2767 * overlapping groups.
2768 */
2769 for (group = 0; group < ai->nr_groups; group++) {
2770 struct pcpu_group_info *gi = &ai->groups[group];
2771 void *ptr = areas[group];
2772
2773 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2774 if (gi->cpu_map[i] == NR_CPUS) {
2775 /* unused unit, free whole */
2776 free_fn(ptr, ai->unit_size);
2777 continue;
2778 }
2779 /* copy and return the unused part */
2780 memcpy(ptr, __per_cpu_load, ai->static_size);
2781 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2782 }
2783 }
2784
2785 /* base address is now known, determine group base offsets */
2786 for (group = 0; group < ai->nr_groups; group++) {
2787 ai->groups[group].base_offset = areas[group] - base;
2788 }
2789
2790 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2791 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
2792 ai->dyn_size, ai->unit_size);
2793
2794 pcpu_setup_first_chunk(ai, base);
2795 goto out_free;
2796
2797out_free_areas:
2798 for (group = 0; group < ai->nr_groups; group++)
2799 if (areas[group])
2800 free_fn(areas[group],
2801 ai->groups[group].nr_units * ai->unit_size);
2802out_free:
2803 pcpu_free_alloc_info(ai);
2804 if (areas)
2805 memblock_free_early(__pa(areas), areas_size);
2806 return rc;
2807}
2808#endif /* BUILD_EMBED_FIRST_CHUNK */
2809
2810#ifdef BUILD_PAGE_FIRST_CHUNK
2811/**
2812 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2813 * @reserved_size: the size of reserved percpu area in bytes
2814 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2815 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2816 * @populate_pte_fn: function to populate pte
2817 *
2818 * This is a helper to ease setting up page-remapped first percpu
2819 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2820 *
2821 * This is the basic allocator. Static percpu area is allocated
2822 * page-by-page into vmalloc area.
2823 *
2824 * RETURNS:
2825 * 0 on success, -errno on failure.
2826 */
2827int __init pcpu_page_first_chunk(size_t reserved_size,
2828 pcpu_fc_alloc_fn_t alloc_fn,
2829 pcpu_fc_free_fn_t free_fn,
2830 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2831{
2832 static struct vm_struct vm;
2833 struct pcpu_alloc_info *ai;
2834 char psize_str[16];
2835 int unit_pages;
2836 size_t pages_size;
2837 struct page **pages;
2838 int unit, i, j, rc = 0;
2839 int upa;
2840 int nr_g0_units;
2841
2842 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2843
2844 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2845 if (IS_ERR(ai))
2846 return PTR_ERR(ai);
2847 BUG_ON(ai->nr_groups != 1);
2848 upa = ai->alloc_size/ai->unit_size;
2849 nr_g0_units = roundup(num_possible_cpus(), upa);
2850 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
2851 pcpu_free_alloc_info(ai);
2852 return -EINVAL;
2853 }
2854
2855 unit_pages = ai->unit_size >> PAGE_SHIFT;
2856
2857 /* unaligned allocations can't be freed, round up to page size */
2858 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2859 sizeof(pages[0]));
2860 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
2861 if (!pages)
2862 panic("%s: Failed to allocate %zu bytes\n", __func__,
2863 pages_size);
2864
2865 /* allocate pages */
2866 j = 0;
2867 for (unit = 0; unit < num_possible_cpus(); unit++) {
2868 unsigned int cpu = ai->groups[0].cpu_map[unit];
2869 for (i = 0; i < unit_pages; i++) {
2870 void *ptr;
2871
2872 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2873 if (!ptr) {
2874 pr_warn("failed to allocate %s page for cpu%u\n",
2875 psize_str, cpu);
2876 goto enomem;
2877 }
2878 /* kmemleak tracks the percpu allocations separately */
2879 kmemleak_free(ptr);
2880 pages[j++] = virt_to_page(ptr);
2881 }
2882 }
2883
2884 /* allocate vm area, map the pages and copy static data */
2885 vm.flags = VM_ALLOC;
2886 vm.size = num_possible_cpus() * ai->unit_size;
2887 vm_area_register_early(&vm, PAGE_SIZE);
2888
2889 for (unit = 0; unit < num_possible_cpus(); unit++) {
2890 unsigned long unit_addr =
2891 (unsigned long)vm.addr + unit * ai->unit_size;
2892
2893 for (i = 0; i < unit_pages; i++)
2894 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2895
2896 /* pte already populated, the following shouldn't fail */
2897 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2898 unit_pages);
2899 if (rc < 0)
2900 panic("failed to map percpu area, err=%d\n", rc);
2901
2902 /*
2903 * FIXME: Archs with virtual cache should flush local
2904 * cache for the linear mapping here - something
2905 * equivalent to flush_cache_vmap() on the local cpu.
2906 * flush_cache_vmap() can't be used as most supporting
2907 * data structures are not set up yet.
2908 */
2909
2910 /* copy static data */
2911 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2912 }
2913
2914 /* we're ready, commit */
2915 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
2916 unit_pages, psize_str, ai->static_size,
2917 ai->reserved_size, ai->dyn_size);
2918
2919 pcpu_setup_first_chunk(ai, vm.addr);
2920 goto out_free_ar;
2921
2922enomem:
2923 while (--j >= 0)
2924 free_fn(page_address(pages[j]), PAGE_SIZE);
2925 rc = -ENOMEM;
2926out_free_ar:
2927 memblock_free_early(__pa(pages), pages_size);
2928 pcpu_free_alloc_info(ai);
2929 return rc;
2930}
2931#endif /* BUILD_PAGE_FIRST_CHUNK */
2932
2933#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2934/*
2935 * Generic SMP percpu area setup.
2936 *
2937 * The embedding helper is used because its behavior closely resembles
2938 * the original non-dynamic generic percpu area setup. This is
2939 * important because many archs have addressing restrictions and might
2940 * fail if the percpu area is located far away from the previous
2941 * location. As an added bonus, in non-NUMA cases, embedding is
2942 * generally a good idea TLB-wise because percpu area can piggy back
2943 * on the physical linear memory mapping which uses large page
2944 * mappings on applicable archs.
2945 */
2946unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2947EXPORT_SYMBOL(__per_cpu_offset);
2948
2949static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2950 size_t align)
2951{
2952 return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS));
2953}
2954
2955static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2956{
2957 memblock_free_early(__pa(ptr), size);
2958}
2959
2960void __init setup_per_cpu_areas(void)
2961{
2962 unsigned long delta;
2963 unsigned int cpu;
2964 int rc;
2965
2966 /*
2967 * Always reserve area for module percpu variables. That's
2968 * what the legacy allocator did.
2969 */
2970 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2971 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2972 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2973 if (rc < 0)
2974 panic("Failed to initialize percpu areas.");
2975
2976 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2977 for_each_possible_cpu(cpu)
2978 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2979}
2980#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2981
2982#else /* CONFIG_SMP */
2983
2984/*
2985 * UP percpu area setup.
2986 *
2987 * UP always uses km-based percpu allocator with identity mapping.
2988 * Static percpu variables are indistinguishable from the usual static
2989 * variables and don't require any special preparation.
2990 */
2991void __init setup_per_cpu_areas(void)
2992{
2993 const size_t unit_size =
2994 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2995 PERCPU_DYNAMIC_RESERVE));
2996 struct pcpu_alloc_info *ai;
2997 void *fc;
2998
2999 ai = pcpu_alloc_alloc_info(1, 1);
3000 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3001 if (!ai || !fc)
3002 panic("Failed to allocate memory for percpu areas.");
3003 /* kmemleak tracks the percpu allocations separately */
3004 kmemleak_free(fc);
3005
3006 ai->dyn_size = unit_size;
3007 ai->unit_size = unit_size;
3008 ai->atom_size = unit_size;
3009 ai->alloc_size = unit_size;
3010 ai->groups[0].nr_units = 1;
3011 ai->groups[0].cpu_map[0] = 0;
3012
3013 pcpu_setup_first_chunk(ai, fc);
3014 pcpu_free_alloc_info(ai);
3015}
3016
3017#endif /* CONFIG_SMP */
3018
3019/*
3020 * pcpu_nr_pages - calculate total number of populated backing pages
3021 *
3022 * This reflects the number of pages populated to back chunks. Metadata is
3023 * excluded in the number exposed in meminfo as the number of backing pages
3024 * scales with the number of cpus and can quickly outweigh the memory used for
3025 * metadata. It also keeps this calculation nice and simple.
3026 *
3027 * RETURNS:
3028 * Total number of populated backing pages in use by the allocator.
3029 */
3030unsigned long pcpu_nr_pages(void)
3031{
3032 return pcpu_nr_populated * pcpu_nr_units;
3033}
3034
3035/*
3036 * Percpu allocator is initialized early during boot when neither slab or
3037 * workqueue is available. Plug async management until everything is up
3038 * and running.
3039 */
3040static int __init percpu_enable_async(void)
3041{
3042 pcpu_async_enabled = true;
3043 return 0;
3044}
3045subsys_initcall(percpu_enable_async);
1// 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/cpumask.h>
73#include <linux/memblock.h>
74#include <linux/err.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/*
102 * The slots are sorted by the size of the biggest continuous free area.
103 * 1-31 bytes share the same slot.
104 */
105#define PCPU_SLOT_BASE_SHIFT 5
106/* chunks in slots below this are subject to being sidelined on failed alloc */
107#define PCPU_SLOT_FAIL_THRESHOLD 3
108
109#define PCPU_EMPTY_POP_PAGES_LOW 2
110#define PCPU_EMPTY_POP_PAGES_HIGH 4
111
112#ifdef CONFIG_SMP
113/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
114#ifndef __addr_to_pcpu_ptr
115#define __addr_to_pcpu_ptr(addr) \
116 (void __percpu *)((unsigned long)(addr) - \
117 (unsigned long)pcpu_base_addr + \
118 (unsigned long)__per_cpu_start)
119#endif
120#ifndef __pcpu_ptr_to_addr
121#define __pcpu_ptr_to_addr(ptr) \
122 (void __force *)((unsigned long)(ptr) + \
123 (unsigned long)pcpu_base_addr - \
124 (unsigned long)__per_cpu_start)
125#endif
126#else /* CONFIG_SMP */
127/* on UP, it's always identity mapped */
128#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
129#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
130#endif /* CONFIG_SMP */
131
132static int pcpu_unit_pages __ro_after_init;
133static int pcpu_unit_size __ro_after_init;
134static int pcpu_nr_units __ro_after_init;
135static int pcpu_atom_size __ro_after_init;
136int pcpu_nr_slots __ro_after_init;
137static int pcpu_free_slot __ro_after_init;
138int pcpu_sidelined_slot __ro_after_init;
139int pcpu_to_depopulate_slot __ro_after_init;
140static size_t pcpu_chunk_struct_size __ro_after_init;
141
142/* cpus with the lowest and highest unit addresses */
143static unsigned int pcpu_low_unit_cpu __ro_after_init;
144static unsigned int pcpu_high_unit_cpu __ro_after_init;
145
146/* the address of the first chunk which starts with the kernel static area */
147void *pcpu_base_addr __ro_after_init;
148
149static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
150const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
151
152/* group information, used for vm allocation */
153static int pcpu_nr_groups __ro_after_init;
154static const unsigned long *pcpu_group_offsets __ro_after_init;
155static const size_t *pcpu_group_sizes __ro_after_init;
156
157/*
158 * The first chunk which always exists. Note that unlike other
159 * chunks, this one can be allocated and mapped in several different
160 * ways and thus often doesn't live in the vmalloc area.
161 */
162struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
163
164/*
165 * Optional reserved chunk. This chunk reserves part of the first
166 * chunk and serves it for reserved allocations. When the reserved
167 * region doesn't exist, the following variable is NULL.
168 */
169struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
170
171DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
172static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
173
174struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
175
176/*
177 * The number of empty populated pages, protected by pcpu_lock.
178 * The reserved chunk doesn't contribute to the count.
179 */
180int pcpu_nr_empty_pop_pages;
181
182/*
183 * The number of populated pages in use by the allocator, protected by
184 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
185 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
186 * and increments/decrements this count by 1).
187 */
188static unsigned long pcpu_nr_populated;
189
190/*
191 * Balance work is used to populate or destroy chunks asynchronously. We
192 * try to keep the number of populated free pages between
193 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
194 * empty chunk.
195 */
196static void pcpu_balance_workfn(struct work_struct *work);
197static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
198static bool pcpu_async_enabled __read_mostly;
199static bool pcpu_atomic_alloc_failed;
200
201static void pcpu_schedule_balance_work(void)
202{
203 if (pcpu_async_enabled)
204 schedule_work(&pcpu_balance_work);
205}
206
207/**
208 * pcpu_addr_in_chunk - check if the address is served from this chunk
209 * @chunk: chunk of interest
210 * @addr: percpu address
211 *
212 * RETURNS:
213 * True if the address is served from this chunk.
214 */
215static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
216{
217 void *start_addr, *end_addr;
218
219 if (!chunk)
220 return false;
221
222 start_addr = chunk->base_addr + chunk->start_offset;
223 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
224 chunk->end_offset;
225
226 return addr >= start_addr && addr < end_addr;
227}
228
229static int __pcpu_size_to_slot(int size)
230{
231 int highbit = fls(size); /* size is in bytes */
232 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
233}
234
235static int pcpu_size_to_slot(int size)
236{
237 if (size == pcpu_unit_size)
238 return pcpu_free_slot;
239 return __pcpu_size_to_slot(size);
240}
241
242static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
243{
244 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
245
246 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
247 chunk_md->contig_hint == 0)
248 return 0;
249
250 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
251}
252
253/* set the pointer to a chunk in a page struct */
254static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
255{
256 page->index = (unsigned long)pcpu;
257}
258
259/* obtain pointer to a chunk from a page struct */
260static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
261{
262 return (struct pcpu_chunk *)page->index;
263}
264
265static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
266{
267 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
268}
269
270static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
271{
272 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
273}
274
275static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
276 unsigned int cpu, int page_idx)
277{
278 return (unsigned long)chunk->base_addr +
279 pcpu_unit_page_offset(cpu, page_idx);
280}
281
282/*
283 * The following are helper functions to help access bitmaps and convert
284 * between bitmap offsets to address offsets.
285 */
286static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
287{
288 return chunk->alloc_map +
289 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
290}
291
292static unsigned long pcpu_off_to_block_index(int off)
293{
294 return off / PCPU_BITMAP_BLOCK_BITS;
295}
296
297static unsigned long pcpu_off_to_block_off(int off)
298{
299 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
300}
301
302static unsigned long pcpu_block_off_to_off(int index, int off)
303{
304 return index * PCPU_BITMAP_BLOCK_BITS + off;
305}
306
307/**
308 * pcpu_check_block_hint - check against the contig hint
309 * @block: block of interest
310 * @bits: size of allocation
311 * @align: alignment of area (max PAGE_SIZE)
312 *
313 * Check to see if the allocation can fit in the block's contig hint.
314 * Note, a chunk uses the same hints as a block so this can also check against
315 * the chunk's contig hint.
316 */
317static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
318 size_t align)
319{
320 int bit_off = ALIGN(block->contig_hint_start, align) -
321 block->contig_hint_start;
322
323 return bit_off + bits <= block->contig_hint;
324}
325
326/*
327 * pcpu_next_hint - determine which hint to use
328 * @block: block of interest
329 * @alloc_bits: size of allocation
330 *
331 * This determines if we should scan based on the scan_hint or first_free.
332 * In general, we want to scan from first_free to fulfill allocations by
333 * first fit. However, if we know a scan_hint at position scan_hint_start
334 * cannot fulfill an allocation, we can begin scanning from there knowing
335 * the contig_hint will be our fallback.
336 */
337static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
338{
339 /*
340 * The three conditions below determine if we can skip past the
341 * scan_hint. First, does the scan hint exist. Second, is the
342 * contig_hint after the scan_hint (possibly not true iff
343 * contig_hint == scan_hint). Third, is the allocation request
344 * larger than the scan_hint.
345 */
346 if (block->scan_hint &&
347 block->contig_hint_start > block->scan_hint_start &&
348 alloc_bits > block->scan_hint)
349 return block->scan_hint_start + block->scan_hint;
350
351 return block->first_free;
352}
353
354/**
355 * pcpu_next_md_free_region - finds the next hint free area
356 * @chunk: chunk of interest
357 * @bit_off: chunk offset
358 * @bits: size of free area
359 *
360 * Helper function for pcpu_for_each_md_free_region. It checks
361 * block->contig_hint and performs aggregation across blocks to find the
362 * next hint. It modifies bit_off and bits in-place to be consumed in the
363 * loop.
364 */
365static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
366 int *bits)
367{
368 int i = pcpu_off_to_block_index(*bit_off);
369 int block_off = pcpu_off_to_block_off(*bit_off);
370 struct pcpu_block_md *block;
371
372 *bits = 0;
373 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
374 block++, i++) {
375 /* handles contig area across blocks */
376 if (*bits) {
377 *bits += block->left_free;
378 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
379 continue;
380 return;
381 }
382
383 /*
384 * This checks three things. First is there a contig_hint to
385 * check. Second, have we checked this hint before by
386 * comparing the block_off. Third, is this the same as the
387 * right contig hint. In the last case, it spills over into
388 * the next block and should be handled by the contig area
389 * across blocks code.
390 */
391 *bits = block->contig_hint;
392 if (*bits && block->contig_hint_start >= block_off &&
393 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
394 *bit_off = pcpu_block_off_to_off(i,
395 block->contig_hint_start);
396 return;
397 }
398 /* reset to satisfy the second predicate above */
399 block_off = 0;
400
401 *bits = block->right_free;
402 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
403 }
404}
405
406/**
407 * pcpu_next_fit_region - finds fit areas for a given allocation request
408 * @chunk: chunk of interest
409 * @alloc_bits: size of allocation
410 * @align: alignment of area (max PAGE_SIZE)
411 * @bit_off: chunk offset
412 * @bits: size of free area
413 *
414 * Finds the next free region that is viable for use with a given size and
415 * alignment. This only returns if there is a valid area to be used for this
416 * allocation. block->first_free is returned if the allocation request fits
417 * within the block to see if the request can be fulfilled prior to the contig
418 * hint.
419 */
420static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
421 int align, int *bit_off, int *bits)
422{
423 int i = pcpu_off_to_block_index(*bit_off);
424 int block_off = pcpu_off_to_block_off(*bit_off);
425 struct pcpu_block_md *block;
426
427 *bits = 0;
428 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
429 block++, i++) {
430 /* handles contig area across blocks */
431 if (*bits) {
432 *bits += block->left_free;
433 if (*bits >= alloc_bits)
434 return;
435 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
436 continue;
437 }
438
439 /* check block->contig_hint */
440 *bits = ALIGN(block->contig_hint_start, align) -
441 block->contig_hint_start;
442 /*
443 * This uses the block offset to determine if this has been
444 * checked in the prior iteration.
445 */
446 if (block->contig_hint &&
447 block->contig_hint_start >= block_off &&
448 block->contig_hint >= *bits + alloc_bits) {
449 int start = pcpu_next_hint(block, alloc_bits);
450
451 *bits += alloc_bits + block->contig_hint_start -
452 start;
453 *bit_off = pcpu_block_off_to_off(i, start);
454 return;
455 }
456 /* reset to satisfy the second predicate above */
457 block_off = 0;
458
459 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
460 align);
461 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
462 *bit_off = pcpu_block_off_to_off(i, *bit_off);
463 if (*bits >= alloc_bits)
464 return;
465 }
466
467 /* no valid offsets were found - fail condition */
468 *bit_off = pcpu_chunk_map_bits(chunk);
469}
470
471/*
472 * Metadata free area iterators. These perform aggregation of free areas
473 * based on the metadata blocks and return the offset @bit_off and size in
474 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
475 * a fit is found for the allocation request.
476 */
477#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
478 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
479 (bit_off) < pcpu_chunk_map_bits((chunk)); \
480 (bit_off) += (bits) + 1, \
481 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
482
483#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
484 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
485 &(bits)); \
486 (bit_off) < pcpu_chunk_map_bits((chunk)); \
487 (bit_off) += (bits), \
488 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
489 &(bits)))
490
491/**
492 * pcpu_mem_zalloc - allocate memory
493 * @size: bytes to allocate
494 * @gfp: allocation flags
495 *
496 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
497 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
498 * This is to facilitate passing through whitelisted flags. The
499 * returned memory is always zeroed.
500 *
501 * RETURNS:
502 * Pointer to the allocated area on success, NULL on failure.
503 */
504static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
505{
506 if (WARN_ON_ONCE(!slab_is_available()))
507 return NULL;
508
509 if (size <= PAGE_SIZE)
510 return kzalloc(size, gfp);
511 else
512 return __vmalloc(size, gfp | __GFP_ZERO);
513}
514
515/**
516 * pcpu_mem_free - free memory
517 * @ptr: memory to free
518 *
519 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
520 */
521static void pcpu_mem_free(void *ptr)
522{
523 kvfree(ptr);
524}
525
526static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
527 bool move_front)
528{
529 if (chunk != pcpu_reserved_chunk) {
530 if (move_front)
531 list_move(&chunk->list, &pcpu_chunk_lists[slot]);
532 else
533 list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]);
534 }
535}
536
537static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
538{
539 __pcpu_chunk_move(chunk, slot, true);
540}
541
542/**
543 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
544 * @chunk: chunk of interest
545 * @oslot: the previous slot it was on
546 *
547 * This function is called after an allocation or free changed @chunk.
548 * New slot according to the changed state is determined and @chunk is
549 * moved to the slot. Note that the reserved chunk is never put on
550 * chunk slots.
551 *
552 * CONTEXT:
553 * pcpu_lock.
554 */
555static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
556{
557 int nslot = pcpu_chunk_slot(chunk);
558
559 /* leave isolated chunks in-place */
560 if (chunk->isolated)
561 return;
562
563 if (oslot != nslot)
564 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
565}
566
567static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
568{
569 lockdep_assert_held(&pcpu_lock);
570
571 if (!chunk->isolated) {
572 chunk->isolated = true;
573 pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
574 }
575 list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
576}
577
578static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
579{
580 lockdep_assert_held(&pcpu_lock);
581
582 if (chunk->isolated) {
583 chunk->isolated = false;
584 pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
585 pcpu_chunk_relocate(chunk, -1);
586 }
587}
588
589/*
590 * pcpu_update_empty_pages - update empty page counters
591 * @chunk: chunk of interest
592 * @nr: nr of empty pages
593 *
594 * This is used to keep track of the empty pages now based on the premise
595 * a md_block covers a page. The hint update functions recognize if a block
596 * is made full or broken to calculate deltas for keeping track of free pages.
597 */
598static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
599{
600 chunk->nr_empty_pop_pages += nr;
601 if (chunk != pcpu_reserved_chunk && !chunk->isolated)
602 pcpu_nr_empty_pop_pages += nr;
603}
604
605/*
606 * pcpu_region_overlap - determines if two regions overlap
607 * @a: start of first region, inclusive
608 * @b: end of first region, exclusive
609 * @x: start of second region, inclusive
610 * @y: end of second region, exclusive
611 *
612 * This is used to determine if the hint region [a, b) overlaps with the
613 * allocated region [x, y).
614 */
615static inline bool pcpu_region_overlap(int a, int b, int x, int y)
616{
617 return (a < y) && (x < b);
618}
619
620/**
621 * pcpu_block_update - updates a block given a free area
622 * @block: block of interest
623 * @start: start offset in block
624 * @end: end offset in block
625 *
626 * Updates a block given a known free area. The region [start, end) is
627 * expected to be the entirety of the free area within a block. Chooses
628 * the best starting offset if the contig hints are equal.
629 */
630static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
631{
632 int contig = end - start;
633
634 block->first_free = min(block->first_free, start);
635 if (start == 0)
636 block->left_free = contig;
637
638 if (end == block->nr_bits)
639 block->right_free = contig;
640
641 if (contig > block->contig_hint) {
642 /* promote the old contig_hint to be the new scan_hint */
643 if (start > block->contig_hint_start) {
644 if (block->contig_hint > block->scan_hint) {
645 block->scan_hint_start =
646 block->contig_hint_start;
647 block->scan_hint = block->contig_hint;
648 } else if (start < block->scan_hint_start) {
649 /*
650 * The old contig_hint == scan_hint. But, the
651 * new contig is larger so hold the invariant
652 * scan_hint_start < contig_hint_start.
653 */
654 block->scan_hint = 0;
655 }
656 } else {
657 block->scan_hint = 0;
658 }
659 block->contig_hint_start = start;
660 block->contig_hint = contig;
661 } else if (contig == block->contig_hint) {
662 if (block->contig_hint_start &&
663 (!start ||
664 __ffs(start) > __ffs(block->contig_hint_start))) {
665 /* start has a better alignment so use it */
666 block->contig_hint_start = start;
667 if (start < block->scan_hint_start &&
668 block->contig_hint > block->scan_hint)
669 block->scan_hint = 0;
670 } else if (start > block->scan_hint_start ||
671 block->contig_hint > block->scan_hint) {
672 /*
673 * Knowing contig == contig_hint, update the scan_hint
674 * if it is farther than or larger than the current
675 * scan_hint.
676 */
677 block->scan_hint_start = start;
678 block->scan_hint = contig;
679 }
680 } else {
681 /*
682 * The region is smaller than the contig_hint. So only update
683 * the scan_hint if it is larger than or equal and farther than
684 * the current scan_hint.
685 */
686 if ((start < block->contig_hint_start &&
687 (contig > block->scan_hint ||
688 (contig == block->scan_hint &&
689 start > block->scan_hint_start)))) {
690 block->scan_hint_start = start;
691 block->scan_hint = contig;
692 }
693 }
694}
695
696/*
697 * pcpu_block_update_scan - update a block given a free area from a scan
698 * @chunk: chunk of interest
699 * @bit_off: chunk offset
700 * @bits: size of free area
701 *
702 * Finding the final allocation spot first goes through pcpu_find_block_fit()
703 * to find a block that can hold the allocation and then pcpu_alloc_area()
704 * where a scan is used. When allocations require specific alignments,
705 * we can inadvertently create holes which will not be seen in the alloc
706 * or free paths.
707 *
708 * This takes a given free area hole and updates a block as it may change the
709 * scan_hint. We need to scan backwards to ensure we don't miss free bits
710 * from alignment.
711 */
712static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
713 int bits)
714{
715 int s_off = pcpu_off_to_block_off(bit_off);
716 int e_off = s_off + bits;
717 int s_index, l_bit;
718 struct pcpu_block_md *block;
719
720 if (e_off > PCPU_BITMAP_BLOCK_BITS)
721 return;
722
723 s_index = pcpu_off_to_block_index(bit_off);
724 block = chunk->md_blocks + s_index;
725
726 /* scan backwards in case of alignment skipping free bits */
727 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
728 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
729
730 pcpu_block_update(block, s_off, e_off);
731}
732
733/**
734 * pcpu_chunk_refresh_hint - updates metadata about a chunk
735 * @chunk: chunk of interest
736 * @full_scan: if we should scan from the beginning
737 *
738 * Iterates over the metadata blocks to find the largest contig area.
739 * A full scan can be avoided on the allocation path as this is triggered
740 * if we broke the contig_hint. In doing so, the scan_hint will be before
741 * the contig_hint or after if the scan_hint == contig_hint. This cannot
742 * be prevented on freeing as we want to find the largest area possibly
743 * spanning blocks.
744 */
745static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
746{
747 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
748 int bit_off, bits;
749
750 /* promote scan_hint to contig_hint */
751 if (!full_scan && chunk_md->scan_hint) {
752 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
753 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
754 chunk_md->contig_hint = chunk_md->scan_hint;
755 chunk_md->scan_hint = 0;
756 } else {
757 bit_off = chunk_md->first_free;
758 chunk_md->contig_hint = 0;
759 }
760
761 bits = 0;
762 pcpu_for_each_md_free_region(chunk, bit_off, bits)
763 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
764}
765
766/**
767 * pcpu_block_refresh_hint
768 * @chunk: chunk of interest
769 * @index: index of the metadata block
770 *
771 * Scans over the block beginning at first_free and updates the block
772 * metadata accordingly.
773 */
774static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
775{
776 struct pcpu_block_md *block = chunk->md_blocks + index;
777 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
778 unsigned int start, end; /* region start, region end */
779
780 /* promote scan_hint to contig_hint */
781 if (block->scan_hint) {
782 start = block->scan_hint_start + block->scan_hint;
783 block->contig_hint_start = block->scan_hint_start;
784 block->contig_hint = block->scan_hint;
785 block->scan_hint = 0;
786 } else {
787 start = block->first_free;
788 block->contig_hint = 0;
789 }
790
791 block->right_free = 0;
792
793 /* iterate over free areas and update the contig hints */
794 for_each_clear_bitrange_from(start, end, alloc_map, PCPU_BITMAP_BLOCK_BITS)
795 pcpu_block_update(block, start, end);
796}
797
798/**
799 * pcpu_block_update_hint_alloc - update hint on allocation path
800 * @chunk: chunk of interest
801 * @bit_off: chunk offset
802 * @bits: size of request
803 *
804 * Updates metadata for the allocation path. The metadata only has to be
805 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
806 * scans are required if the block's contig hint is broken.
807 */
808static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
809 int bits)
810{
811 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
812 int nr_empty_pages = 0;
813 struct pcpu_block_md *s_block, *e_block, *block;
814 int s_index, e_index; /* block indexes of the freed allocation */
815 int s_off, e_off; /* block offsets of the freed allocation */
816
817 /*
818 * Calculate per block offsets.
819 * The calculation uses an inclusive range, but the resulting offsets
820 * are [start, end). e_index always points to the last block in the
821 * range.
822 */
823 s_index = pcpu_off_to_block_index(bit_off);
824 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
825 s_off = pcpu_off_to_block_off(bit_off);
826 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
827
828 s_block = chunk->md_blocks + s_index;
829 e_block = chunk->md_blocks + e_index;
830
831 /*
832 * Update s_block.
833 */
834 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
835 nr_empty_pages++;
836
837 /*
838 * block->first_free must be updated if the allocation takes its place.
839 * If the allocation breaks the contig_hint, a scan is required to
840 * restore this hint.
841 */
842 if (s_off == s_block->first_free)
843 s_block->first_free = find_next_zero_bit(
844 pcpu_index_alloc_map(chunk, s_index),
845 PCPU_BITMAP_BLOCK_BITS,
846 s_off + bits);
847
848 if (pcpu_region_overlap(s_block->scan_hint_start,
849 s_block->scan_hint_start + s_block->scan_hint,
850 s_off,
851 s_off + bits))
852 s_block->scan_hint = 0;
853
854 if (pcpu_region_overlap(s_block->contig_hint_start,
855 s_block->contig_hint_start +
856 s_block->contig_hint,
857 s_off,
858 s_off + bits)) {
859 /* block contig hint is broken - scan to fix it */
860 if (!s_off)
861 s_block->left_free = 0;
862 pcpu_block_refresh_hint(chunk, s_index);
863 } else {
864 /* update left and right contig manually */
865 s_block->left_free = min(s_block->left_free, s_off);
866 if (s_index == e_index)
867 s_block->right_free = min_t(int, s_block->right_free,
868 PCPU_BITMAP_BLOCK_BITS - e_off);
869 else
870 s_block->right_free = 0;
871 }
872
873 /*
874 * Update e_block.
875 */
876 if (s_index != e_index) {
877 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
878 nr_empty_pages++;
879
880 /*
881 * When the allocation is across blocks, the end is along
882 * the left part of the e_block.
883 */
884 e_block->first_free = find_next_zero_bit(
885 pcpu_index_alloc_map(chunk, e_index),
886 PCPU_BITMAP_BLOCK_BITS, e_off);
887
888 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
889 /* reset the block */
890 e_block++;
891 } else {
892 if (e_off > e_block->scan_hint_start)
893 e_block->scan_hint = 0;
894
895 e_block->left_free = 0;
896 if (e_off > e_block->contig_hint_start) {
897 /* contig hint is broken - scan to fix it */
898 pcpu_block_refresh_hint(chunk, e_index);
899 } else {
900 e_block->right_free =
901 min_t(int, e_block->right_free,
902 PCPU_BITMAP_BLOCK_BITS - e_off);
903 }
904 }
905
906 /* update in-between md_blocks */
907 nr_empty_pages += (e_index - s_index - 1);
908 for (block = s_block + 1; block < e_block; block++) {
909 block->scan_hint = 0;
910 block->contig_hint = 0;
911 block->left_free = 0;
912 block->right_free = 0;
913 }
914 }
915
916 /*
917 * If the allocation is not atomic, some blocks may not be
918 * populated with pages, while we account it here. The number
919 * of pages will be added back with pcpu_chunk_populated()
920 * when populating pages.
921 */
922 if (nr_empty_pages)
923 pcpu_update_empty_pages(chunk, -nr_empty_pages);
924
925 if (pcpu_region_overlap(chunk_md->scan_hint_start,
926 chunk_md->scan_hint_start +
927 chunk_md->scan_hint,
928 bit_off,
929 bit_off + bits))
930 chunk_md->scan_hint = 0;
931
932 /*
933 * The only time a full chunk scan is required is if the chunk
934 * contig hint is broken. Otherwise, it means a smaller space
935 * was used and therefore the chunk contig hint is still correct.
936 */
937 if (pcpu_region_overlap(chunk_md->contig_hint_start,
938 chunk_md->contig_hint_start +
939 chunk_md->contig_hint,
940 bit_off,
941 bit_off + bits))
942 pcpu_chunk_refresh_hint(chunk, false);
943}
944
945/**
946 * pcpu_block_update_hint_free - updates the block hints on the free path
947 * @chunk: chunk of interest
948 * @bit_off: chunk offset
949 * @bits: size of request
950 *
951 * Updates metadata for the allocation path. This avoids a blind block
952 * refresh by making use of the block contig hints. If this fails, it scans
953 * forward and backward to determine the extent of the free area. This is
954 * capped at the boundary of blocks.
955 *
956 * A chunk update is triggered if a page becomes free, a block becomes free,
957 * or the free spans across blocks. This tradeoff is to minimize iterating
958 * over the block metadata to update chunk_md->contig_hint.
959 * chunk_md->contig_hint may be off by up to a page, but it will never be more
960 * than the available space. If the contig hint is contained in one block, it
961 * will be accurate.
962 */
963static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
964 int bits)
965{
966 int nr_empty_pages = 0;
967 struct pcpu_block_md *s_block, *e_block, *block;
968 int s_index, e_index; /* block indexes of the freed allocation */
969 int s_off, e_off; /* block offsets of the freed allocation */
970 int start, end; /* start and end of the whole free area */
971
972 /*
973 * Calculate per block offsets.
974 * The calculation uses an inclusive range, but the resulting offsets
975 * are [start, end). e_index always points to the last block in the
976 * range.
977 */
978 s_index = pcpu_off_to_block_index(bit_off);
979 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
980 s_off = pcpu_off_to_block_off(bit_off);
981 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
982
983 s_block = chunk->md_blocks + s_index;
984 e_block = chunk->md_blocks + e_index;
985
986 /*
987 * Check if the freed area aligns with the block->contig_hint.
988 * If it does, then the scan to find the beginning/end of the
989 * larger free area can be avoided.
990 *
991 * start and end refer to beginning and end of the free area
992 * within each their respective blocks. This is not necessarily
993 * the entire free area as it may span blocks past the beginning
994 * or end of the block.
995 */
996 start = s_off;
997 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
998 start = s_block->contig_hint_start;
999 } else {
1000 /*
1001 * Scan backwards to find the extent of the free area.
1002 * find_last_bit returns the starting bit, so if the start bit
1003 * is returned, that means there was no last bit and the
1004 * remainder of the chunk is free.
1005 */
1006 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
1007 start);
1008 start = (start == l_bit) ? 0 : l_bit + 1;
1009 }
1010
1011 end = e_off;
1012 if (e_off == e_block->contig_hint_start)
1013 end = e_block->contig_hint_start + e_block->contig_hint;
1014 else
1015 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
1016 PCPU_BITMAP_BLOCK_BITS, end);
1017
1018 /* update s_block */
1019 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1020 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1021 nr_empty_pages++;
1022 pcpu_block_update(s_block, start, e_off);
1023
1024 /* freeing in the same block */
1025 if (s_index != e_index) {
1026 /* update e_block */
1027 if (end == PCPU_BITMAP_BLOCK_BITS)
1028 nr_empty_pages++;
1029 pcpu_block_update(e_block, 0, end);
1030
1031 /* reset md_blocks in the middle */
1032 nr_empty_pages += (e_index - s_index - 1);
1033 for (block = s_block + 1; block < e_block; block++) {
1034 block->first_free = 0;
1035 block->scan_hint = 0;
1036 block->contig_hint_start = 0;
1037 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1038 block->left_free = PCPU_BITMAP_BLOCK_BITS;
1039 block->right_free = PCPU_BITMAP_BLOCK_BITS;
1040 }
1041 }
1042
1043 if (nr_empty_pages)
1044 pcpu_update_empty_pages(chunk, nr_empty_pages);
1045
1046 /*
1047 * Refresh chunk metadata when the free makes a block free or spans
1048 * across blocks. The contig_hint may be off by up to a page, but if
1049 * the contig_hint is contained in a block, it will be accurate with
1050 * the else condition below.
1051 */
1052 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1053 pcpu_chunk_refresh_hint(chunk, true);
1054 else
1055 pcpu_block_update(&chunk->chunk_md,
1056 pcpu_block_off_to_off(s_index, start),
1057 end);
1058}
1059
1060/**
1061 * pcpu_is_populated - determines if the region is populated
1062 * @chunk: chunk of interest
1063 * @bit_off: chunk offset
1064 * @bits: size of area
1065 * @next_off: return value for the next offset to start searching
1066 *
1067 * For atomic allocations, check if the backing pages are populated.
1068 *
1069 * RETURNS:
1070 * Bool if the backing pages are populated.
1071 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1072 */
1073static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1074 int *next_off)
1075{
1076 unsigned int start, end;
1077
1078 start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1079 end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1080
1081 start = find_next_zero_bit(chunk->populated, end, start);
1082 if (start >= end)
1083 return true;
1084
1085 end = find_next_bit(chunk->populated, end, start + 1);
1086
1087 *next_off = end * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1088 return false;
1089}
1090
1091/**
1092 * pcpu_find_block_fit - finds the block index to start searching
1093 * @chunk: chunk of interest
1094 * @alloc_bits: size of request in allocation units
1095 * @align: alignment of area (max PAGE_SIZE bytes)
1096 * @pop_only: use populated regions only
1097 *
1098 * Given a chunk and an allocation spec, find the offset to begin searching
1099 * for a free region. This iterates over the bitmap metadata blocks to
1100 * find an offset that will be guaranteed to fit the requirements. It is
1101 * not quite first fit as if the allocation does not fit in the contig hint
1102 * of a block or chunk, it is skipped. This errs on the side of caution
1103 * to prevent excess iteration. Poor alignment can cause the allocator to
1104 * skip over blocks and chunks that have valid free areas.
1105 *
1106 * RETURNS:
1107 * The offset in the bitmap to begin searching.
1108 * -1 if no offset is found.
1109 */
1110static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1111 size_t align, bool pop_only)
1112{
1113 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1114 int bit_off, bits, next_off;
1115
1116 /*
1117 * This is an optimization to prevent scanning by assuming if the
1118 * allocation cannot fit in the global hint, there is memory pressure
1119 * and creating a new chunk would happen soon.
1120 */
1121 if (!pcpu_check_block_hint(chunk_md, alloc_bits, align))
1122 return -1;
1123
1124 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1125 bits = 0;
1126 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1127 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1128 &next_off))
1129 break;
1130
1131 bit_off = next_off;
1132 bits = 0;
1133 }
1134
1135 if (bit_off == pcpu_chunk_map_bits(chunk))
1136 return -1;
1137
1138 return bit_off;
1139}
1140
1141/*
1142 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1143 * @map: the address to base the search on
1144 * @size: the bitmap size in bits
1145 * @start: the bitnumber to start searching at
1146 * @nr: the number of zeroed bits we're looking for
1147 * @align_mask: alignment mask for zero area
1148 * @largest_off: offset of the largest area skipped
1149 * @largest_bits: size of the largest area skipped
1150 *
1151 * The @align_mask should be one less than a power of 2.
1152 *
1153 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1154 * the largest area that was skipped. This is imperfect, but in general is
1155 * good enough. The largest remembered region is the largest failed region
1156 * seen. This does not include anything we possibly skipped due to alignment.
1157 * pcpu_block_update_scan() does scan backwards to try and recover what was
1158 * lost to alignment. While this can cause scanning to miss earlier possible
1159 * free areas, smaller allocations will eventually fill those holes.
1160 */
1161static unsigned long pcpu_find_zero_area(unsigned long *map,
1162 unsigned long size,
1163 unsigned long start,
1164 unsigned long nr,
1165 unsigned long align_mask,
1166 unsigned long *largest_off,
1167 unsigned long *largest_bits)
1168{
1169 unsigned long index, end, i, area_off, area_bits;
1170again:
1171 index = find_next_zero_bit(map, size, start);
1172
1173 /* Align allocation */
1174 index = __ALIGN_MASK(index, align_mask);
1175 area_off = index;
1176
1177 end = index + nr;
1178 if (end > size)
1179 return end;
1180 i = find_next_bit(map, end, index);
1181 if (i < end) {
1182 area_bits = i - area_off;
1183 /* remember largest unused area with best alignment */
1184 if (area_bits > *largest_bits ||
1185 (area_bits == *largest_bits && *largest_off &&
1186 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1187 *largest_off = area_off;
1188 *largest_bits = area_bits;
1189 }
1190
1191 start = i + 1;
1192 goto again;
1193 }
1194 return index;
1195}
1196
1197/**
1198 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1199 * @chunk: chunk of interest
1200 * @alloc_bits: size of request in allocation units
1201 * @align: alignment of area (max PAGE_SIZE)
1202 * @start: bit_off to start searching
1203 *
1204 * This function takes in a @start offset to begin searching to fit an
1205 * allocation of @alloc_bits with alignment @align. It needs to scan
1206 * the allocation map because if it fits within the block's contig hint,
1207 * @start will be block->first_free. This is an attempt to fill the
1208 * allocation prior to breaking the contig hint. The allocation and
1209 * boundary maps are updated accordingly if it confirms a valid
1210 * free area.
1211 *
1212 * RETURNS:
1213 * Allocated addr offset in @chunk on success.
1214 * -1 if no matching area is found.
1215 */
1216static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1217 size_t align, int start)
1218{
1219 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1220 size_t align_mask = (align) ? (align - 1) : 0;
1221 unsigned long area_off = 0, area_bits = 0;
1222 int bit_off, end, oslot;
1223
1224 lockdep_assert_held(&pcpu_lock);
1225
1226 oslot = pcpu_chunk_slot(chunk);
1227
1228 /*
1229 * Search to find a fit.
1230 */
1231 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1232 pcpu_chunk_map_bits(chunk));
1233 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1234 align_mask, &area_off, &area_bits);
1235 if (bit_off >= end)
1236 return -1;
1237
1238 if (area_bits)
1239 pcpu_block_update_scan(chunk, area_off, area_bits);
1240
1241 /* update alloc map */
1242 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1243
1244 /* update boundary map */
1245 set_bit(bit_off, chunk->bound_map);
1246 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1247 set_bit(bit_off + alloc_bits, chunk->bound_map);
1248
1249 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1250
1251 /* update first free bit */
1252 if (bit_off == chunk_md->first_free)
1253 chunk_md->first_free = find_next_zero_bit(
1254 chunk->alloc_map,
1255 pcpu_chunk_map_bits(chunk),
1256 bit_off + alloc_bits);
1257
1258 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1259
1260 pcpu_chunk_relocate(chunk, oslot);
1261
1262 return bit_off * PCPU_MIN_ALLOC_SIZE;
1263}
1264
1265/**
1266 * pcpu_free_area - frees the corresponding offset
1267 * @chunk: chunk of interest
1268 * @off: addr offset into chunk
1269 *
1270 * This function determines the size of an allocation to free using
1271 * the boundary bitmap and clears the allocation map.
1272 *
1273 * RETURNS:
1274 * Number of freed bytes.
1275 */
1276static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1277{
1278 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1279 int bit_off, bits, end, oslot, freed;
1280
1281 lockdep_assert_held(&pcpu_lock);
1282 pcpu_stats_area_dealloc(chunk);
1283
1284 oslot = pcpu_chunk_slot(chunk);
1285
1286 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1287
1288 /* find end index */
1289 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1290 bit_off + 1);
1291 bits = end - bit_off;
1292 bitmap_clear(chunk->alloc_map, bit_off, bits);
1293
1294 freed = bits * PCPU_MIN_ALLOC_SIZE;
1295
1296 /* update metadata */
1297 chunk->free_bytes += freed;
1298
1299 /* update first free bit */
1300 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1301
1302 pcpu_block_update_hint_free(chunk, bit_off, bits);
1303
1304 pcpu_chunk_relocate(chunk, oslot);
1305
1306 return freed;
1307}
1308
1309static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1310{
1311 block->scan_hint = 0;
1312 block->contig_hint = nr_bits;
1313 block->left_free = nr_bits;
1314 block->right_free = nr_bits;
1315 block->first_free = 0;
1316 block->nr_bits = nr_bits;
1317}
1318
1319static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1320{
1321 struct pcpu_block_md *md_block;
1322
1323 /* init the chunk's block */
1324 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1325
1326 for (md_block = chunk->md_blocks;
1327 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1328 md_block++)
1329 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1330}
1331
1332/**
1333 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1334 * @tmp_addr: the start of the region served
1335 * @map_size: size of the region served
1336 *
1337 * This is responsible for creating the chunks that serve the first chunk. The
1338 * base_addr is page aligned down of @tmp_addr while the region end is page
1339 * aligned up. Offsets are kept track of to determine the region served. All
1340 * this is done to appease the bitmap allocator in avoiding partial blocks.
1341 *
1342 * RETURNS:
1343 * Chunk serving the region at @tmp_addr of @map_size.
1344 */
1345static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1346 int map_size)
1347{
1348 struct pcpu_chunk *chunk;
1349 unsigned long aligned_addr;
1350 int start_offset, offset_bits, region_size, region_bits;
1351 size_t alloc_size;
1352
1353 /* region calculations */
1354 aligned_addr = tmp_addr & PAGE_MASK;
1355
1356 start_offset = tmp_addr - aligned_addr;
1357 region_size = ALIGN(start_offset + map_size, PAGE_SIZE);
1358
1359 /* allocate chunk */
1360 alloc_size = struct_size(chunk, populated,
1361 BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1362 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1363 if (!chunk)
1364 panic("%s: Failed to allocate %zu bytes\n", __func__,
1365 alloc_size);
1366
1367 INIT_LIST_HEAD(&chunk->list);
1368
1369 chunk->base_addr = (void *)aligned_addr;
1370 chunk->start_offset = start_offset;
1371 chunk->end_offset = region_size - chunk->start_offset - map_size;
1372
1373 chunk->nr_pages = region_size >> PAGE_SHIFT;
1374 region_bits = pcpu_chunk_map_bits(chunk);
1375
1376 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1377 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1378 if (!chunk->alloc_map)
1379 panic("%s: Failed to allocate %zu bytes\n", __func__,
1380 alloc_size);
1381
1382 alloc_size =
1383 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1384 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1385 if (!chunk->bound_map)
1386 panic("%s: Failed to allocate %zu bytes\n", __func__,
1387 alloc_size);
1388
1389 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1390 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1391 if (!chunk->md_blocks)
1392 panic("%s: Failed to allocate %zu bytes\n", __func__,
1393 alloc_size);
1394
1395#ifdef CONFIG_MEMCG_KMEM
1396 /* first chunk is free to use */
1397 chunk->obj_cgroups = NULL;
1398#endif
1399 pcpu_init_md_blocks(chunk);
1400
1401 /* manage populated page bitmap */
1402 chunk->immutable = true;
1403 bitmap_fill(chunk->populated, chunk->nr_pages);
1404 chunk->nr_populated = chunk->nr_pages;
1405 chunk->nr_empty_pop_pages = chunk->nr_pages;
1406
1407 chunk->free_bytes = map_size;
1408
1409 if (chunk->start_offset) {
1410 /* hide the beginning of the bitmap */
1411 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1412 bitmap_set(chunk->alloc_map, 0, offset_bits);
1413 set_bit(0, chunk->bound_map);
1414 set_bit(offset_bits, chunk->bound_map);
1415
1416 chunk->chunk_md.first_free = offset_bits;
1417
1418 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1419 }
1420
1421 if (chunk->end_offset) {
1422 /* hide the end of the bitmap */
1423 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1424 bitmap_set(chunk->alloc_map,
1425 pcpu_chunk_map_bits(chunk) - offset_bits,
1426 offset_bits);
1427 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1428 chunk->bound_map);
1429 set_bit(region_bits, chunk->bound_map);
1430
1431 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1432 - offset_bits, offset_bits);
1433 }
1434
1435 return chunk;
1436}
1437
1438static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1439{
1440 struct pcpu_chunk *chunk;
1441 int region_bits;
1442
1443 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1444 if (!chunk)
1445 return NULL;
1446
1447 INIT_LIST_HEAD(&chunk->list);
1448 chunk->nr_pages = pcpu_unit_pages;
1449 region_bits = pcpu_chunk_map_bits(chunk);
1450
1451 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1452 sizeof(chunk->alloc_map[0]), gfp);
1453 if (!chunk->alloc_map)
1454 goto alloc_map_fail;
1455
1456 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1457 sizeof(chunk->bound_map[0]), gfp);
1458 if (!chunk->bound_map)
1459 goto bound_map_fail;
1460
1461 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1462 sizeof(chunk->md_blocks[0]), gfp);
1463 if (!chunk->md_blocks)
1464 goto md_blocks_fail;
1465
1466#ifdef CONFIG_MEMCG_KMEM
1467 if (!mem_cgroup_kmem_disabled()) {
1468 chunk->obj_cgroups =
1469 pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1470 sizeof(struct obj_cgroup *), gfp);
1471 if (!chunk->obj_cgroups)
1472 goto objcg_fail;
1473 }
1474#endif
1475
1476 pcpu_init_md_blocks(chunk);
1477
1478 /* init metadata */
1479 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1480
1481 return chunk;
1482
1483#ifdef CONFIG_MEMCG_KMEM
1484objcg_fail:
1485 pcpu_mem_free(chunk->md_blocks);
1486#endif
1487md_blocks_fail:
1488 pcpu_mem_free(chunk->bound_map);
1489bound_map_fail:
1490 pcpu_mem_free(chunk->alloc_map);
1491alloc_map_fail:
1492 pcpu_mem_free(chunk);
1493
1494 return NULL;
1495}
1496
1497static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1498{
1499 if (!chunk)
1500 return;
1501#ifdef CONFIG_MEMCG_KMEM
1502 pcpu_mem_free(chunk->obj_cgroups);
1503#endif
1504 pcpu_mem_free(chunk->md_blocks);
1505 pcpu_mem_free(chunk->bound_map);
1506 pcpu_mem_free(chunk->alloc_map);
1507 pcpu_mem_free(chunk);
1508}
1509
1510/**
1511 * pcpu_chunk_populated - post-population bookkeeping
1512 * @chunk: pcpu_chunk which got populated
1513 * @page_start: the start page
1514 * @page_end: the end page
1515 *
1516 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1517 * the bookkeeping information accordingly. Must be called after each
1518 * successful population.
1519 */
1520static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1521 int page_end)
1522{
1523 int nr = page_end - page_start;
1524
1525 lockdep_assert_held(&pcpu_lock);
1526
1527 bitmap_set(chunk->populated, page_start, nr);
1528 chunk->nr_populated += nr;
1529 pcpu_nr_populated += nr;
1530
1531 pcpu_update_empty_pages(chunk, nr);
1532}
1533
1534/**
1535 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1536 * @chunk: pcpu_chunk which got depopulated
1537 * @page_start: the start page
1538 * @page_end: the end page
1539 *
1540 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1541 * Update the bookkeeping information accordingly. Must be called after
1542 * each successful depopulation.
1543 */
1544static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1545 int page_start, int page_end)
1546{
1547 int nr = page_end - page_start;
1548
1549 lockdep_assert_held(&pcpu_lock);
1550
1551 bitmap_clear(chunk->populated, page_start, nr);
1552 chunk->nr_populated -= nr;
1553 pcpu_nr_populated -= nr;
1554
1555 pcpu_update_empty_pages(chunk, -nr);
1556}
1557
1558/*
1559 * Chunk management implementation.
1560 *
1561 * To allow different implementations, chunk alloc/free and
1562 * [de]population are implemented in a separate file which is pulled
1563 * into this file and compiled together. The following functions
1564 * should be implemented.
1565 *
1566 * pcpu_populate_chunk - populate the specified range of a chunk
1567 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1568 * pcpu_post_unmap_tlb_flush - flush tlb for the specified range of a chunk
1569 * pcpu_create_chunk - create a new chunk
1570 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1571 * pcpu_addr_to_page - translate address to physical address
1572 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1573 */
1574static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1575 int page_start, int page_end, gfp_t gfp);
1576static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1577 int page_start, int page_end);
1578static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1579 int page_start, int page_end);
1580static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1581static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1582static struct page *pcpu_addr_to_page(void *addr);
1583static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1584
1585#ifdef CONFIG_NEED_PER_CPU_KM
1586#include "percpu-km.c"
1587#else
1588#include "percpu-vm.c"
1589#endif
1590
1591/**
1592 * pcpu_chunk_addr_search - determine chunk containing specified address
1593 * @addr: address for which the chunk needs to be determined.
1594 *
1595 * This is an internal function that handles all but static allocations.
1596 * Static percpu address values should never be passed into the allocator.
1597 *
1598 * RETURNS:
1599 * The address of the found chunk.
1600 */
1601static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1602{
1603 /* is it in the dynamic region (first chunk)? */
1604 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1605 return pcpu_first_chunk;
1606
1607 /* is it in the reserved region? */
1608 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1609 return pcpu_reserved_chunk;
1610
1611 /*
1612 * The address is relative to unit0 which might be unused and
1613 * thus unmapped. Offset the address to the unit space of the
1614 * current processor before looking it up in the vmalloc
1615 * space. Note that any possible cpu id can be used here, so
1616 * there's no need to worry about preemption or cpu hotplug.
1617 */
1618 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1619 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1620}
1621
1622#ifdef CONFIG_MEMCG_KMEM
1623static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1624 struct obj_cgroup **objcgp)
1625{
1626 struct obj_cgroup *objcg;
1627
1628 if (!memcg_kmem_enabled() || !(gfp & __GFP_ACCOUNT))
1629 return true;
1630
1631 objcg = get_obj_cgroup_from_current();
1632 if (!objcg)
1633 return true;
1634
1635 if (obj_cgroup_charge(objcg, gfp, pcpu_obj_full_size(size))) {
1636 obj_cgroup_put(objcg);
1637 return false;
1638 }
1639
1640 *objcgp = objcg;
1641 return true;
1642}
1643
1644static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1645 struct pcpu_chunk *chunk, int off,
1646 size_t size)
1647{
1648 if (!objcg)
1649 return;
1650
1651 if (likely(chunk && chunk->obj_cgroups)) {
1652 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1653
1654 rcu_read_lock();
1655 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1656 pcpu_obj_full_size(size));
1657 rcu_read_unlock();
1658 } else {
1659 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1660 obj_cgroup_put(objcg);
1661 }
1662}
1663
1664static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1665{
1666 struct obj_cgroup *objcg;
1667
1668 if (unlikely(!chunk->obj_cgroups))
1669 return;
1670
1671 objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1672 if (!objcg)
1673 return;
1674 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1675
1676 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1677
1678 rcu_read_lock();
1679 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1680 -pcpu_obj_full_size(size));
1681 rcu_read_unlock();
1682
1683 obj_cgroup_put(objcg);
1684}
1685
1686#else /* CONFIG_MEMCG_KMEM */
1687static bool
1688pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1689{
1690 return true;
1691}
1692
1693static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1694 struct pcpu_chunk *chunk, int off,
1695 size_t size)
1696{
1697}
1698
1699static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1700{
1701}
1702#endif /* CONFIG_MEMCG_KMEM */
1703
1704/**
1705 * pcpu_alloc - the percpu allocator
1706 * @size: size of area to allocate in bytes
1707 * @align: alignment of area (max PAGE_SIZE)
1708 * @reserved: allocate from the reserved chunk if available
1709 * @gfp: allocation flags
1710 *
1711 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1712 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1713 * then no warning will be triggered on invalid or failed allocation
1714 * requests.
1715 *
1716 * RETURNS:
1717 * Percpu pointer to the allocated area on success, NULL on failure.
1718 */
1719static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1720 gfp_t gfp)
1721{
1722 gfp_t pcpu_gfp;
1723 bool is_atomic;
1724 bool do_warn;
1725 struct obj_cgroup *objcg = NULL;
1726 static int warn_limit = 10;
1727 struct pcpu_chunk *chunk, *next;
1728 const char *err;
1729 int slot, off, cpu, ret;
1730 unsigned long flags;
1731 void __percpu *ptr;
1732 size_t bits, bit_align;
1733
1734 gfp = current_gfp_context(gfp);
1735 /* whitelisted flags that can be passed to the backing allocators */
1736 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1737 is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1738 do_warn = !(gfp & __GFP_NOWARN);
1739
1740 /*
1741 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1742 * therefore alignment must be a minimum of that many bytes.
1743 * An allocation may have internal fragmentation from rounding up
1744 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1745 */
1746 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1747 align = PCPU_MIN_ALLOC_SIZE;
1748
1749 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1750 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1751 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1752
1753 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1754 !is_power_of_2(align))) {
1755 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1756 size, align);
1757 return NULL;
1758 }
1759
1760 if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1761 return NULL;
1762
1763 if (!is_atomic) {
1764 /*
1765 * pcpu_balance_workfn() allocates memory under this mutex,
1766 * and it may wait for memory reclaim. Allow current task
1767 * to become OOM victim, in case of memory pressure.
1768 */
1769 if (gfp & __GFP_NOFAIL) {
1770 mutex_lock(&pcpu_alloc_mutex);
1771 } else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1772 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1773 return NULL;
1774 }
1775 }
1776
1777 spin_lock_irqsave(&pcpu_lock, flags);
1778
1779 /* serve reserved allocations from the reserved chunk if available */
1780 if (reserved && pcpu_reserved_chunk) {
1781 chunk = pcpu_reserved_chunk;
1782
1783 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1784 if (off < 0) {
1785 err = "alloc from reserved chunk failed";
1786 goto fail_unlock;
1787 }
1788
1789 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1790 if (off >= 0)
1791 goto area_found;
1792
1793 err = "alloc from reserved chunk failed";
1794 goto fail_unlock;
1795 }
1796
1797restart:
1798 /* search through normal chunks */
1799 for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1800 list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1801 list) {
1802 off = pcpu_find_block_fit(chunk, bits, bit_align,
1803 is_atomic);
1804 if (off < 0) {
1805 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1806 pcpu_chunk_move(chunk, 0);
1807 continue;
1808 }
1809
1810 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1811 if (off >= 0) {
1812 pcpu_reintegrate_chunk(chunk);
1813 goto area_found;
1814 }
1815 }
1816 }
1817
1818 spin_unlock_irqrestore(&pcpu_lock, flags);
1819
1820 if (is_atomic) {
1821 err = "atomic alloc failed, no space left";
1822 goto fail;
1823 }
1824
1825 /* No space left. Create a new chunk. */
1826 if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) {
1827 chunk = pcpu_create_chunk(pcpu_gfp);
1828 if (!chunk) {
1829 err = "failed to allocate new chunk";
1830 goto fail;
1831 }
1832
1833 spin_lock_irqsave(&pcpu_lock, flags);
1834 pcpu_chunk_relocate(chunk, -1);
1835 } else {
1836 spin_lock_irqsave(&pcpu_lock, flags);
1837 }
1838
1839 goto restart;
1840
1841area_found:
1842 pcpu_stats_area_alloc(chunk, size);
1843 spin_unlock_irqrestore(&pcpu_lock, flags);
1844
1845 /* populate if not all pages are already there */
1846 if (!is_atomic) {
1847 unsigned int page_end, rs, re;
1848
1849 rs = PFN_DOWN(off);
1850 page_end = PFN_UP(off + size);
1851
1852 for_each_clear_bitrange_from(rs, re, chunk->populated, page_end) {
1853 WARN_ON(chunk->immutable);
1854
1855 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1856
1857 spin_lock_irqsave(&pcpu_lock, flags);
1858 if (ret) {
1859 pcpu_free_area(chunk, off);
1860 err = "failed to populate";
1861 goto fail_unlock;
1862 }
1863 pcpu_chunk_populated(chunk, rs, re);
1864 spin_unlock_irqrestore(&pcpu_lock, flags);
1865 }
1866
1867 mutex_unlock(&pcpu_alloc_mutex);
1868 }
1869
1870 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1871 pcpu_schedule_balance_work();
1872
1873 /* clear the areas and return address relative to base address */
1874 for_each_possible_cpu(cpu)
1875 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1876
1877 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1878 kmemleak_alloc_percpu(ptr, size, gfp);
1879
1880 trace_percpu_alloc_percpu(_RET_IP_, reserved, is_atomic, size, align,
1881 chunk->base_addr, off, ptr,
1882 pcpu_obj_full_size(size), gfp);
1883
1884 pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1885
1886 return ptr;
1887
1888fail_unlock:
1889 spin_unlock_irqrestore(&pcpu_lock, flags);
1890fail:
1891 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1892
1893 if (!is_atomic && do_warn && warn_limit) {
1894 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1895 size, align, is_atomic, err);
1896 dump_stack();
1897 if (!--warn_limit)
1898 pr_info("limit reached, disable warning\n");
1899 }
1900 if (is_atomic) {
1901 /* see the flag handling in pcpu_balance_workfn() */
1902 pcpu_atomic_alloc_failed = true;
1903 pcpu_schedule_balance_work();
1904 } else {
1905 mutex_unlock(&pcpu_alloc_mutex);
1906 }
1907
1908 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1909
1910 return NULL;
1911}
1912
1913/**
1914 * __alloc_percpu_gfp - allocate dynamic percpu area
1915 * @size: size of area to allocate in bytes
1916 * @align: alignment of area (max PAGE_SIZE)
1917 * @gfp: allocation flags
1918 *
1919 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1920 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1921 * be called from any context but is a lot more likely to fail. If @gfp
1922 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1923 * allocation requests.
1924 *
1925 * RETURNS:
1926 * Percpu pointer to the allocated area on success, NULL on failure.
1927 */
1928void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1929{
1930 return pcpu_alloc(size, align, false, gfp);
1931}
1932EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1933
1934/**
1935 * __alloc_percpu - allocate dynamic percpu area
1936 * @size: size of area to allocate in bytes
1937 * @align: alignment of area (max PAGE_SIZE)
1938 *
1939 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1940 */
1941void __percpu *__alloc_percpu(size_t size, size_t align)
1942{
1943 return pcpu_alloc(size, align, false, GFP_KERNEL);
1944}
1945EXPORT_SYMBOL_GPL(__alloc_percpu);
1946
1947/**
1948 * __alloc_reserved_percpu - allocate reserved percpu area
1949 * @size: size of area to allocate in bytes
1950 * @align: alignment of area (max PAGE_SIZE)
1951 *
1952 * Allocate zero-filled percpu area of @size bytes aligned at @align
1953 * from reserved percpu area if arch has set it up; otherwise,
1954 * allocation is served from the same dynamic area. Might sleep.
1955 * Might trigger writeouts.
1956 *
1957 * CONTEXT:
1958 * Does GFP_KERNEL allocation.
1959 *
1960 * RETURNS:
1961 * Percpu pointer to the allocated area on success, NULL on failure.
1962 */
1963void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1964{
1965 return pcpu_alloc(size, align, true, GFP_KERNEL);
1966}
1967
1968/**
1969 * pcpu_balance_free - manage the amount of free chunks
1970 * @empty_only: free chunks only if there are no populated pages
1971 *
1972 * If empty_only is %false, reclaim all fully free chunks regardless of the
1973 * number of populated pages. Otherwise, only reclaim chunks that have no
1974 * populated pages.
1975 *
1976 * CONTEXT:
1977 * pcpu_lock (can be dropped temporarily)
1978 */
1979static void pcpu_balance_free(bool empty_only)
1980{
1981 LIST_HEAD(to_free);
1982 struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1983 struct pcpu_chunk *chunk, *next;
1984
1985 lockdep_assert_held(&pcpu_lock);
1986
1987 /*
1988 * There's no reason to keep around multiple unused chunks and VM
1989 * areas can be scarce. Destroy all free chunks except for one.
1990 */
1991 list_for_each_entry_safe(chunk, next, free_head, list) {
1992 WARN_ON(chunk->immutable);
1993
1994 /* spare the first one */
1995 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1996 continue;
1997
1998 if (!empty_only || chunk->nr_empty_pop_pages == 0)
1999 list_move(&chunk->list, &to_free);
2000 }
2001
2002 if (list_empty(&to_free))
2003 return;
2004
2005 spin_unlock_irq(&pcpu_lock);
2006 list_for_each_entry_safe(chunk, next, &to_free, list) {
2007 unsigned int rs, re;
2008
2009 for_each_set_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2010 pcpu_depopulate_chunk(chunk, rs, re);
2011 spin_lock_irq(&pcpu_lock);
2012 pcpu_chunk_depopulated(chunk, rs, re);
2013 spin_unlock_irq(&pcpu_lock);
2014 }
2015 pcpu_destroy_chunk(chunk);
2016 cond_resched();
2017 }
2018 spin_lock_irq(&pcpu_lock);
2019}
2020
2021/**
2022 * pcpu_balance_populated - manage the amount of populated pages
2023 *
2024 * Maintain a certain amount of populated pages to satisfy atomic allocations.
2025 * It is possible that this is called when physical memory is scarce causing
2026 * OOM killer to be triggered. We should avoid doing so until an actual
2027 * allocation causes the failure as it is possible that requests can be
2028 * serviced from already backed regions.
2029 *
2030 * CONTEXT:
2031 * pcpu_lock (can be dropped temporarily)
2032 */
2033static void pcpu_balance_populated(void)
2034{
2035 /* gfp flags passed to underlying allocators */
2036 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2037 struct pcpu_chunk *chunk;
2038 int slot, nr_to_pop, ret;
2039
2040 lockdep_assert_held(&pcpu_lock);
2041
2042 /*
2043 * Ensure there are certain number of free populated pages for
2044 * atomic allocs. Fill up from the most packed so that atomic
2045 * allocs don't increase fragmentation. If atomic allocation
2046 * failed previously, always populate the maximum amount. This
2047 * should prevent atomic allocs larger than PAGE_SIZE from keeping
2048 * failing indefinitely; however, large atomic allocs are not
2049 * something we support properly and can be highly unreliable and
2050 * inefficient.
2051 */
2052retry_pop:
2053 if (pcpu_atomic_alloc_failed) {
2054 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2055 /* best effort anyway, don't worry about synchronization */
2056 pcpu_atomic_alloc_failed = false;
2057 } else {
2058 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2059 pcpu_nr_empty_pop_pages,
2060 0, PCPU_EMPTY_POP_PAGES_HIGH);
2061 }
2062
2063 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2064 unsigned int nr_unpop = 0, rs, re;
2065
2066 if (!nr_to_pop)
2067 break;
2068
2069 list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
2070 nr_unpop = chunk->nr_pages - chunk->nr_populated;
2071 if (nr_unpop)
2072 break;
2073 }
2074
2075 if (!nr_unpop)
2076 continue;
2077
2078 /* @chunk can't go away while pcpu_alloc_mutex is held */
2079 for_each_clear_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2080 int nr = min_t(int, re - rs, nr_to_pop);
2081
2082 spin_unlock_irq(&pcpu_lock);
2083 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2084 cond_resched();
2085 spin_lock_irq(&pcpu_lock);
2086 if (!ret) {
2087 nr_to_pop -= nr;
2088 pcpu_chunk_populated(chunk, rs, rs + nr);
2089 } else {
2090 nr_to_pop = 0;
2091 }
2092
2093 if (!nr_to_pop)
2094 break;
2095 }
2096 }
2097
2098 if (nr_to_pop) {
2099 /* ran out of chunks to populate, create a new one and retry */
2100 spin_unlock_irq(&pcpu_lock);
2101 chunk = pcpu_create_chunk(gfp);
2102 cond_resched();
2103 spin_lock_irq(&pcpu_lock);
2104 if (chunk) {
2105 pcpu_chunk_relocate(chunk, -1);
2106 goto retry_pop;
2107 }
2108 }
2109}
2110
2111/**
2112 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2113 *
2114 * Scan over chunks in the depopulate list and try to release unused populated
2115 * pages back to the system. Depopulated chunks are sidelined to prevent
2116 * repopulating these pages unless required. Fully free chunks are reintegrated
2117 * and freed accordingly (1 is kept around). If we drop below the empty
2118 * populated pages threshold, reintegrate the chunk if it has empty free pages.
2119 * Each chunk is scanned in the reverse order to keep populated pages close to
2120 * the beginning of the chunk.
2121 *
2122 * CONTEXT:
2123 * pcpu_lock (can be dropped temporarily)
2124 *
2125 */
2126static void pcpu_reclaim_populated(void)
2127{
2128 struct pcpu_chunk *chunk;
2129 struct pcpu_block_md *block;
2130 int freed_page_start, freed_page_end;
2131 int i, end;
2132 bool reintegrate;
2133
2134 lockdep_assert_held(&pcpu_lock);
2135
2136 /*
2137 * Once a chunk is isolated to the to_depopulate list, the chunk is no
2138 * longer discoverable to allocations whom may populate pages. The only
2139 * other accessor is the free path which only returns area back to the
2140 * allocator not touching the populated bitmap.
2141 */
2142 while ((chunk = list_first_entry_or_null(
2143 &pcpu_chunk_lists[pcpu_to_depopulate_slot],
2144 struct pcpu_chunk, list))) {
2145 WARN_ON(chunk->immutable);
2146
2147 /*
2148 * Scan chunk's pages in the reverse order to keep populated
2149 * pages close to the beginning of the chunk.
2150 */
2151 freed_page_start = chunk->nr_pages;
2152 freed_page_end = 0;
2153 reintegrate = false;
2154 for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2155 /* no more work to do */
2156 if (chunk->nr_empty_pop_pages == 0)
2157 break;
2158
2159 /* reintegrate chunk to prevent atomic alloc failures */
2160 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2161 reintegrate = true;
2162 break;
2163 }
2164
2165 /*
2166 * If the page is empty and populated, start or
2167 * extend the (i, end) range. If i == 0, decrease
2168 * i and perform the depopulation to cover the last
2169 * (first) page in the chunk.
2170 */
2171 block = chunk->md_blocks + i;
2172 if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2173 test_bit(i, chunk->populated)) {
2174 if (end == -1)
2175 end = i;
2176 if (i > 0)
2177 continue;
2178 i--;
2179 }
2180
2181 /* depopulate if there is an active range */
2182 if (end == -1)
2183 continue;
2184
2185 spin_unlock_irq(&pcpu_lock);
2186 pcpu_depopulate_chunk(chunk, i + 1, end + 1);
2187 cond_resched();
2188 spin_lock_irq(&pcpu_lock);
2189
2190 pcpu_chunk_depopulated(chunk, i + 1, end + 1);
2191 freed_page_start = min(freed_page_start, i + 1);
2192 freed_page_end = max(freed_page_end, end + 1);
2193
2194 /* reset the range and continue */
2195 end = -1;
2196 }
2197
2198 /* batch tlb flush per chunk to amortize cost */
2199 if (freed_page_start < freed_page_end) {
2200 spin_unlock_irq(&pcpu_lock);
2201 pcpu_post_unmap_tlb_flush(chunk,
2202 freed_page_start,
2203 freed_page_end);
2204 cond_resched();
2205 spin_lock_irq(&pcpu_lock);
2206 }
2207
2208 if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2209 pcpu_reintegrate_chunk(chunk);
2210 else
2211 list_move_tail(&chunk->list,
2212 &pcpu_chunk_lists[pcpu_sidelined_slot]);
2213 }
2214}
2215
2216/**
2217 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2218 * @work: unused
2219 *
2220 * For each chunk type, manage the number of fully free chunks and the number of
2221 * populated pages. An important thing to consider is when pages are freed and
2222 * how they contribute to the global counts.
2223 */
2224static void pcpu_balance_workfn(struct work_struct *work)
2225{
2226 /*
2227 * pcpu_balance_free() is called twice because the first time we may
2228 * trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2229 * to grow other chunks. This then gives pcpu_reclaim_populated() time
2230 * to move fully free chunks to the active list to be freed if
2231 * appropriate.
2232 */
2233 mutex_lock(&pcpu_alloc_mutex);
2234 spin_lock_irq(&pcpu_lock);
2235
2236 pcpu_balance_free(false);
2237 pcpu_reclaim_populated();
2238 pcpu_balance_populated();
2239 pcpu_balance_free(true);
2240
2241 spin_unlock_irq(&pcpu_lock);
2242 mutex_unlock(&pcpu_alloc_mutex);
2243}
2244
2245/**
2246 * free_percpu - free percpu area
2247 * @ptr: pointer to area to free
2248 *
2249 * Free percpu area @ptr.
2250 *
2251 * CONTEXT:
2252 * Can be called from atomic context.
2253 */
2254void free_percpu(void __percpu *ptr)
2255{
2256 void *addr;
2257 struct pcpu_chunk *chunk;
2258 unsigned long flags;
2259 int size, off;
2260 bool need_balance = false;
2261
2262 if (!ptr)
2263 return;
2264
2265 kmemleak_free_percpu(ptr);
2266
2267 addr = __pcpu_ptr_to_addr(ptr);
2268
2269 spin_lock_irqsave(&pcpu_lock, flags);
2270
2271 chunk = pcpu_chunk_addr_search(addr);
2272 off = addr - chunk->base_addr;
2273
2274 size = pcpu_free_area(chunk, off);
2275
2276 pcpu_memcg_free_hook(chunk, off, size);
2277
2278 /*
2279 * If there are more than one fully free chunks, wake up grim reaper.
2280 * If the chunk is isolated, it may be in the process of being
2281 * reclaimed. Let reclaim manage cleaning up of that chunk.
2282 */
2283 if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2284 struct pcpu_chunk *pos;
2285
2286 list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2287 if (pos != chunk) {
2288 need_balance = true;
2289 break;
2290 }
2291 } else if (pcpu_should_reclaim_chunk(chunk)) {
2292 pcpu_isolate_chunk(chunk);
2293 need_balance = true;
2294 }
2295
2296 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2297
2298 spin_unlock_irqrestore(&pcpu_lock, flags);
2299
2300 if (need_balance)
2301 pcpu_schedule_balance_work();
2302}
2303EXPORT_SYMBOL_GPL(free_percpu);
2304
2305bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2306{
2307#ifdef CONFIG_SMP
2308 const size_t static_size = __per_cpu_end - __per_cpu_start;
2309 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2310 unsigned int cpu;
2311
2312 for_each_possible_cpu(cpu) {
2313 void *start = per_cpu_ptr(base, cpu);
2314 void *va = (void *)addr;
2315
2316 if (va >= start && va < start + static_size) {
2317 if (can_addr) {
2318 *can_addr = (unsigned long) (va - start);
2319 *can_addr += (unsigned long)
2320 per_cpu_ptr(base, get_boot_cpu_id());
2321 }
2322 return true;
2323 }
2324 }
2325#endif
2326 /* on UP, can't distinguish from other static vars, always false */
2327 return false;
2328}
2329
2330/**
2331 * is_kernel_percpu_address - test whether address is from static percpu area
2332 * @addr: address to test
2333 *
2334 * Test whether @addr belongs to in-kernel static percpu area. Module
2335 * static percpu areas are not considered. For those, use
2336 * is_module_percpu_address().
2337 *
2338 * RETURNS:
2339 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2340 */
2341bool is_kernel_percpu_address(unsigned long addr)
2342{
2343 return __is_kernel_percpu_address(addr, NULL);
2344}
2345
2346/**
2347 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2348 * @addr: the address to be converted to physical address
2349 *
2350 * Given @addr which is dereferenceable address obtained via one of
2351 * percpu access macros, this function translates it into its physical
2352 * address. The caller is responsible for ensuring @addr stays valid
2353 * until this function finishes.
2354 *
2355 * percpu allocator has special setup for the first chunk, which currently
2356 * supports either embedding in linear address space or vmalloc mapping,
2357 * and, from the second one, the backing allocator (currently either vm or
2358 * km) provides translation.
2359 *
2360 * The addr can be translated simply without checking if it falls into the
2361 * first chunk. But the current code reflects better how percpu allocator
2362 * actually works, and the verification can discover both bugs in percpu
2363 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2364 * code.
2365 *
2366 * RETURNS:
2367 * The physical address for @addr.
2368 */
2369phys_addr_t per_cpu_ptr_to_phys(void *addr)
2370{
2371 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2372 bool in_first_chunk = false;
2373 unsigned long first_low, first_high;
2374 unsigned int cpu;
2375
2376 /*
2377 * The following test on unit_low/high isn't strictly
2378 * necessary but will speed up lookups of addresses which
2379 * aren't in the first chunk.
2380 *
2381 * The address check is against full chunk sizes. pcpu_base_addr
2382 * points to the beginning of the first chunk including the
2383 * static region. Assumes good intent as the first chunk may
2384 * not be full (ie. < pcpu_unit_pages in size).
2385 */
2386 first_low = (unsigned long)pcpu_base_addr +
2387 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2388 first_high = (unsigned long)pcpu_base_addr +
2389 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2390 if ((unsigned long)addr >= first_low &&
2391 (unsigned long)addr < first_high) {
2392 for_each_possible_cpu(cpu) {
2393 void *start = per_cpu_ptr(base, cpu);
2394
2395 if (addr >= start && addr < start + pcpu_unit_size) {
2396 in_first_chunk = true;
2397 break;
2398 }
2399 }
2400 }
2401
2402 if (in_first_chunk) {
2403 if (!is_vmalloc_addr(addr))
2404 return __pa(addr);
2405 else
2406 return page_to_phys(vmalloc_to_page(addr)) +
2407 offset_in_page(addr);
2408 } else
2409 return page_to_phys(pcpu_addr_to_page(addr)) +
2410 offset_in_page(addr);
2411}
2412
2413/**
2414 * pcpu_alloc_alloc_info - allocate percpu allocation info
2415 * @nr_groups: the number of groups
2416 * @nr_units: the number of units
2417 *
2418 * Allocate ai which is large enough for @nr_groups groups containing
2419 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2420 * cpu_map array which is long enough for @nr_units and filled with
2421 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2422 * pointer of other groups.
2423 *
2424 * RETURNS:
2425 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2426 * failure.
2427 */
2428struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2429 int nr_units)
2430{
2431 struct pcpu_alloc_info *ai;
2432 size_t base_size, ai_size;
2433 void *ptr;
2434 int unit;
2435
2436 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2437 __alignof__(ai->groups[0].cpu_map[0]));
2438 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2439
2440 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2441 if (!ptr)
2442 return NULL;
2443 ai = ptr;
2444 ptr += base_size;
2445
2446 ai->groups[0].cpu_map = ptr;
2447
2448 for (unit = 0; unit < nr_units; unit++)
2449 ai->groups[0].cpu_map[unit] = NR_CPUS;
2450
2451 ai->nr_groups = nr_groups;
2452 ai->__ai_size = PFN_ALIGN(ai_size);
2453
2454 return ai;
2455}
2456
2457/**
2458 * pcpu_free_alloc_info - free percpu allocation info
2459 * @ai: pcpu_alloc_info to free
2460 *
2461 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2462 */
2463void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2464{
2465 memblock_free(ai, ai->__ai_size);
2466}
2467
2468/**
2469 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2470 * @lvl: loglevel
2471 * @ai: allocation info to dump
2472 *
2473 * Print out information about @ai using loglevel @lvl.
2474 */
2475static void pcpu_dump_alloc_info(const char *lvl,
2476 const struct pcpu_alloc_info *ai)
2477{
2478 int group_width = 1, cpu_width = 1, width;
2479 char empty_str[] = "--------";
2480 int alloc = 0, alloc_end = 0;
2481 int group, v;
2482 int upa, apl; /* units per alloc, allocs per line */
2483
2484 v = ai->nr_groups;
2485 while (v /= 10)
2486 group_width++;
2487
2488 v = num_possible_cpus();
2489 while (v /= 10)
2490 cpu_width++;
2491 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2492
2493 upa = ai->alloc_size / ai->unit_size;
2494 width = upa * (cpu_width + 1) + group_width + 3;
2495 apl = rounddown_pow_of_two(max(60 / width, 1));
2496
2497 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2498 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2499 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2500
2501 for (group = 0; group < ai->nr_groups; group++) {
2502 const struct pcpu_group_info *gi = &ai->groups[group];
2503 int unit = 0, unit_end = 0;
2504
2505 BUG_ON(gi->nr_units % upa);
2506 for (alloc_end += gi->nr_units / upa;
2507 alloc < alloc_end; alloc++) {
2508 if (!(alloc % apl)) {
2509 pr_cont("\n");
2510 printk("%spcpu-alloc: ", lvl);
2511 }
2512 pr_cont("[%0*d] ", group_width, group);
2513
2514 for (unit_end += upa; unit < unit_end; unit++)
2515 if (gi->cpu_map[unit] != NR_CPUS)
2516 pr_cont("%0*d ",
2517 cpu_width, gi->cpu_map[unit]);
2518 else
2519 pr_cont("%s ", empty_str);
2520 }
2521 }
2522 pr_cont("\n");
2523}
2524
2525/**
2526 * pcpu_setup_first_chunk - initialize the first percpu chunk
2527 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2528 * @base_addr: mapped address
2529 *
2530 * Initialize the first percpu chunk which contains the kernel static
2531 * percpu area. This function is to be called from arch percpu area
2532 * setup path.
2533 *
2534 * @ai contains all information necessary to initialize the first
2535 * chunk and prime the dynamic percpu allocator.
2536 *
2537 * @ai->static_size is the size of static percpu area.
2538 *
2539 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2540 * reserve after the static area in the first chunk. This reserves
2541 * the first chunk such that it's available only through reserved
2542 * percpu allocation. This is primarily used to serve module percpu
2543 * static areas on architectures where the addressing model has
2544 * limited offset range for symbol relocations to guarantee module
2545 * percpu symbols fall inside the relocatable range.
2546 *
2547 * @ai->dyn_size determines the number of bytes available for dynamic
2548 * allocation in the first chunk. The area between @ai->static_size +
2549 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2550 *
2551 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2552 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2553 * @ai->dyn_size.
2554 *
2555 * @ai->atom_size is the allocation atom size and used as alignment
2556 * for vm areas.
2557 *
2558 * @ai->alloc_size is the allocation size and always multiple of
2559 * @ai->atom_size. This is larger than @ai->atom_size if
2560 * @ai->unit_size is larger than @ai->atom_size.
2561 *
2562 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2563 * percpu areas. Units which should be colocated are put into the
2564 * same group. Dynamic VM areas will be allocated according to these
2565 * groupings. If @ai->nr_groups is zero, a single group containing
2566 * all units is assumed.
2567 *
2568 * The caller should have mapped the first chunk at @base_addr and
2569 * copied static data to each unit.
2570 *
2571 * The first chunk will always contain a static and a dynamic region.
2572 * However, the static region is not managed by any chunk. If the first
2573 * chunk also contains a reserved region, it is served by two chunks -
2574 * one for the reserved region and one for the dynamic region. They
2575 * share the same vm, but use offset regions in the area allocation map.
2576 * The chunk serving the dynamic region is circulated in the chunk slots
2577 * and available for dynamic allocation like any other chunk.
2578 */
2579void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2580 void *base_addr)
2581{
2582 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2583 size_t static_size, dyn_size;
2584 struct pcpu_chunk *chunk;
2585 unsigned long *group_offsets;
2586 size_t *group_sizes;
2587 unsigned long *unit_off;
2588 unsigned int cpu;
2589 int *unit_map;
2590 int group, unit, i;
2591 int map_size;
2592 unsigned long tmp_addr;
2593 size_t alloc_size;
2594
2595#define PCPU_SETUP_BUG_ON(cond) do { \
2596 if (unlikely(cond)) { \
2597 pr_emerg("failed to initialize, %s\n", #cond); \
2598 pr_emerg("cpu_possible_mask=%*pb\n", \
2599 cpumask_pr_args(cpu_possible_mask)); \
2600 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2601 BUG(); \
2602 } \
2603} while (0)
2604
2605 /* sanity checks */
2606 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2607#ifdef CONFIG_SMP
2608 PCPU_SETUP_BUG_ON(!ai->static_size);
2609 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2610#endif
2611 PCPU_SETUP_BUG_ON(!base_addr);
2612 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2613 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2614 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2615 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2616 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2617 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2618 PCPU_SETUP_BUG_ON(!ai->dyn_size);
2619 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2620 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2621 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2622 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2623
2624 /* process group information and build config tables accordingly */
2625 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2626 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2627 if (!group_offsets)
2628 panic("%s: Failed to allocate %zu bytes\n", __func__,
2629 alloc_size);
2630
2631 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2632 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2633 if (!group_sizes)
2634 panic("%s: Failed to allocate %zu bytes\n", __func__,
2635 alloc_size);
2636
2637 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2638 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2639 if (!unit_map)
2640 panic("%s: Failed to allocate %zu bytes\n", __func__,
2641 alloc_size);
2642
2643 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2644 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2645 if (!unit_off)
2646 panic("%s: Failed to allocate %zu bytes\n", __func__,
2647 alloc_size);
2648
2649 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2650 unit_map[cpu] = UINT_MAX;
2651
2652 pcpu_low_unit_cpu = NR_CPUS;
2653 pcpu_high_unit_cpu = NR_CPUS;
2654
2655 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2656 const struct pcpu_group_info *gi = &ai->groups[group];
2657
2658 group_offsets[group] = gi->base_offset;
2659 group_sizes[group] = gi->nr_units * ai->unit_size;
2660
2661 for (i = 0; i < gi->nr_units; i++) {
2662 cpu = gi->cpu_map[i];
2663 if (cpu == NR_CPUS)
2664 continue;
2665
2666 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2667 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2668 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2669
2670 unit_map[cpu] = unit + i;
2671 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2672
2673 /* determine low/high unit_cpu */
2674 if (pcpu_low_unit_cpu == NR_CPUS ||
2675 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2676 pcpu_low_unit_cpu = cpu;
2677 if (pcpu_high_unit_cpu == NR_CPUS ||
2678 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2679 pcpu_high_unit_cpu = cpu;
2680 }
2681 }
2682 pcpu_nr_units = unit;
2683
2684 for_each_possible_cpu(cpu)
2685 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2686
2687 /* we're done parsing the input, undefine BUG macro and dump config */
2688#undef PCPU_SETUP_BUG_ON
2689 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2690
2691 pcpu_nr_groups = ai->nr_groups;
2692 pcpu_group_offsets = group_offsets;
2693 pcpu_group_sizes = group_sizes;
2694 pcpu_unit_map = unit_map;
2695 pcpu_unit_offsets = unit_off;
2696
2697 /* determine basic parameters */
2698 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2699 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2700 pcpu_atom_size = ai->atom_size;
2701 pcpu_chunk_struct_size = struct_size(chunk, populated,
2702 BITS_TO_LONGS(pcpu_unit_pages));
2703
2704 pcpu_stats_save_ai(ai);
2705
2706 /*
2707 * Allocate chunk slots. The slots after the active slots are:
2708 * sidelined_slot - isolated, depopulated chunks
2709 * free_slot - fully free chunks
2710 * to_depopulate_slot - isolated, chunks to depopulate
2711 */
2712 pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1;
2713 pcpu_free_slot = pcpu_sidelined_slot + 1;
2714 pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2715 pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2716 pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2717 sizeof(pcpu_chunk_lists[0]),
2718 SMP_CACHE_BYTES);
2719 if (!pcpu_chunk_lists)
2720 panic("%s: Failed to allocate %zu bytes\n", __func__,
2721 pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2722
2723 for (i = 0; i < pcpu_nr_slots; i++)
2724 INIT_LIST_HEAD(&pcpu_chunk_lists[i]);
2725
2726 /*
2727 * The end of the static region needs to be aligned with the
2728 * minimum allocation size as this offsets the reserved and
2729 * dynamic region. The first chunk ends page aligned by
2730 * expanding the dynamic region, therefore the dynamic region
2731 * can be shrunk to compensate while still staying above the
2732 * configured sizes.
2733 */
2734 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2735 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2736
2737 /*
2738 * Initialize first chunk.
2739 * If the reserved_size is non-zero, this initializes the reserved
2740 * chunk. If the reserved_size is zero, the reserved chunk is NULL
2741 * and the dynamic region is initialized here. The first chunk,
2742 * pcpu_first_chunk, will always point to the chunk that serves
2743 * the dynamic region.
2744 */
2745 tmp_addr = (unsigned long)base_addr + static_size;
2746 map_size = ai->reserved_size ?: dyn_size;
2747 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2748
2749 /* init dynamic chunk if necessary */
2750 if (ai->reserved_size) {
2751 pcpu_reserved_chunk = chunk;
2752
2753 tmp_addr = (unsigned long)base_addr + static_size +
2754 ai->reserved_size;
2755 map_size = dyn_size;
2756 chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2757 }
2758
2759 /* link the first chunk in */
2760 pcpu_first_chunk = chunk;
2761 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2762 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2763
2764 /* include all regions of the first chunk */
2765 pcpu_nr_populated += PFN_DOWN(size_sum);
2766
2767 pcpu_stats_chunk_alloc();
2768 trace_percpu_create_chunk(base_addr);
2769
2770 /* we're done */
2771 pcpu_base_addr = base_addr;
2772}
2773
2774#ifdef CONFIG_SMP
2775
2776const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2777 [PCPU_FC_AUTO] = "auto",
2778 [PCPU_FC_EMBED] = "embed",
2779 [PCPU_FC_PAGE] = "page",
2780};
2781
2782enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2783
2784static int __init percpu_alloc_setup(char *str)
2785{
2786 if (!str)
2787 return -EINVAL;
2788
2789 if (0)
2790 /* nada */;
2791#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2792 else if (!strcmp(str, "embed"))
2793 pcpu_chosen_fc = PCPU_FC_EMBED;
2794#endif
2795#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2796 else if (!strcmp(str, "page"))
2797 pcpu_chosen_fc = PCPU_FC_PAGE;
2798#endif
2799 else
2800 pr_warn("unknown allocator %s specified\n", str);
2801
2802 return 0;
2803}
2804early_param("percpu_alloc", percpu_alloc_setup);
2805
2806/*
2807 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2808 * Build it if needed by the arch config or the generic setup is going
2809 * to be used.
2810 */
2811#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2812 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2813#define BUILD_EMBED_FIRST_CHUNK
2814#endif
2815
2816/* build pcpu_page_first_chunk() iff needed by the arch config */
2817#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2818#define BUILD_PAGE_FIRST_CHUNK
2819#endif
2820
2821/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2822#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2823/**
2824 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2825 * @reserved_size: the size of reserved percpu area in bytes
2826 * @dyn_size: minimum free size for dynamic allocation in bytes
2827 * @atom_size: allocation atom size
2828 * @cpu_distance_fn: callback to determine distance between cpus, optional
2829 *
2830 * This function determines grouping of units, their mappings to cpus
2831 * and other parameters considering needed percpu size, allocation
2832 * atom size and distances between CPUs.
2833 *
2834 * Groups are always multiples of atom size and CPUs which are of
2835 * LOCAL_DISTANCE both ways are grouped together and share space for
2836 * units in the same group. The returned configuration is guaranteed
2837 * to have CPUs on different nodes on different groups and >=75% usage
2838 * of allocated virtual address space.
2839 *
2840 * RETURNS:
2841 * On success, pointer to the new allocation_info is returned. On
2842 * failure, ERR_PTR value is returned.
2843 */
2844static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2845 size_t reserved_size, size_t dyn_size,
2846 size_t atom_size,
2847 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2848{
2849 static int group_map[NR_CPUS] __initdata;
2850 static int group_cnt[NR_CPUS] __initdata;
2851 static struct cpumask mask __initdata;
2852 const size_t static_size = __per_cpu_end - __per_cpu_start;
2853 int nr_groups = 1, nr_units = 0;
2854 size_t size_sum, min_unit_size, alloc_size;
2855 int upa, max_upa, best_upa; /* units_per_alloc */
2856 int last_allocs, group, unit;
2857 unsigned int cpu, tcpu;
2858 struct pcpu_alloc_info *ai;
2859 unsigned int *cpu_map;
2860
2861 /* this function may be called multiple times */
2862 memset(group_map, 0, sizeof(group_map));
2863 memset(group_cnt, 0, sizeof(group_cnt));
2864 cpumask_clear(&mask);
2865
2866 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2867 size_sum = PFN_ALIGN(static_size + reserved_size +
2868 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2869 dyn_size = size_sum - static_size - reserved_size;
2870
2871 /*
2872 * Determine min_unit_size, alloc_size and max_upa such that
2873 * alloc_size is multiple of atom_size and is the smallest
2874 * which can accommodate 4k aligned segments which are equal to
2875 * or larger than min_unit_size.
2876 */
2877 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2878
2879 /* determine the maximum # of units that can fit in an allocation */
2880 alloc_size = roundup(min_unit_size, atom_size);
2881 upa = alloc_size / min_unit_size;
2882 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2883 upa--;
2884 max_upa = upa;
2885
2886 cpumask_copy(&mask, cpu_possible_mask);
2887
2888 /* group cpus according to their proximity */
2889 for (group = 0; !cpumask_empty(&mask); group++) {
2890 /* pop the group's first cpu */
2891 cpu = cpumask_first(&mask);
2892 group_map[cpu] = group;
2893 group_cnt[group]++;
2894 cpumask_clear_cpu(cpu, &mask);
2895
2896 for_each_cpu(tcpu, &mask) {
2897 if (!cpu_distance_fn ||
2898 (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2899 cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2900 group_map[tcpu] = group;
2901 group_cnt[group]++;
2902 cpumask_clear_cpu(tcpu, &mask);
2903 }
2904 }
2905 }
2906 nr_groups = group;
2907
2908 /*
2909 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2910 * Expand the unit_size until we use >= 75% of the units allocated.
2911 * Related to atom_size, which could be much larger than the unit_size.
2912 */
2913 last_allocs = INT_MAX;
2914 best_upa = 0;
2915 for (upa = max_upa; upa; upa--) {
2916 int allocs = 0, wasted = 0;
2917
2918 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2919 continue;
2920
2921 for (group = 0; group < nr_groups; group++) {
2922 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2923 allocs += this_allocs;
2924 wasted += this_allocs * upa - group_cnt[group];
2925 }
2926
2927 /*
2928 * Don't accept if wastage is over 1/3. The
2929 * greater-than comparison ensures upa==1 always
2930 * passes the following check.
2931 */
2932 if (wasted > num_possible_cpus() / 3)
2933 continue;
2934
2935 /* and then don't consume more memory */
2936 if (allocs > last_allocs)
2937 break;
2938 last_allocs = allocs;
2939 best_upa = upa;
2940 }
2941 BUG_ON(!best_upa);
2942 upa = best_upa;
2943
2944 /* allocate and fill alloc_info */
2945 for (group = 0; group < nr_groups; group++)
2946 nr_units += roundup(group_cnt[group], upa);
2947
2948 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2949 if (!ai)
2950 return ERR_PTR(-ENOMEM);
2951 cpu_map = ai->groups[0].cpu_map;
2952
2953 for (group = 0; group < nr_groups; group++) {
2954 ai->groups[group].cpu_map = cpu_map;
2955 cpu_map += roundup(group_cnt[group], upa);
2956 }
2957
2958 ai->static_size = static_size;
2959 ai->reserved_size = reserved_size;
2960 ai->dyn_size = dyn_size;
2961 ai->unit_size = alloc_size / upa;
2962 ai->atom_size = atom_size;
2963 ai->alloc_size = alloc_size;
2964
2965 for (group = 0, unit = 0; group < nr_groups; group++) {
2966 struct pcpu_group_info *gi = &ai->groups[group];
2967
2968 /*
2969 * Initialize base_offset as if all groups are located
2970 * back-to-back. The caller should update this to
2971 * reflect actual allocation.
2972 */
2973 gi->base_offset = unit * ai->unit_size;
2974
2975 for_each_possible_cpu(cpu)
2976 if (group_map[cpu] == group)
2977 gi->cpu_map[gi->nr_units++] = cpu;
2978 gi->nr_units = roundup(gi->nr_units, upa);
2979 unit += gi->nr_units;
2980 }
2981 BUG_ON(unit != nr_units);
2982
2983 return ai;
2984}
2985
2986static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align,
2987 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
2988{
2989 const unsigned long goal = __pa(MAX_DMA_ADDRESS);
2990#ifdef CONFIG_NUMA
2991 int node = NUMA_NO_NODE;
2992 void *ptr;
2993
2994 if (cpu_to_nd_fn)
2995 node = cpu_to_nd_fn(cpu);
2996
2997 if (node == NUMA_NO_NODE || !node_online(node) || !NODE_DATA(node)) {
2998 ptr = memblock_alloc_from(size, align, goal);
2999 pr_info("cpu %d has no node %d or node-local memory\n",
3000 cpu, node);
3001 pr_debug("per cpu data for cpu%d %zu bytes at 0x%llx\n",
3002 cpu, size, (u64)__pa(ptr));
3003 } else {
3004 ptr = memblock_alloc_try_nid(size, align, goal,
3005 MEMBLOCK_ALLOC_ACCESSIBLE,
3006 node);
3007
3008 pr_debug("per cpu data for cpu%d %zu bytes on node%d at 0x%llx\n",
3009 cpu, size, node, (u64)__pa(ptr));
3010 }
3011 return ptr;
3012#else
3013 return memblock_alloc_from(size, align, goal);
3014#endif
3015}
3016
3017static void __init pcpu_fc_free(void *ptr, size_t size)
3018{
3019 memblock_free(ptr, size);
3020}
3021#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
3022
3023#if defined(BUILD_EMBED_FIRST_CHUNK)
3024/**
3025 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
3026 * @reserved_size: the size of reserved percpu area in bytes
3027 * @dyn_size: minimum free size for dynamic allocation in bytes
3028 * @atom_size: allocation atom size
3029 * @cpu_distance_fn: callback to determine distance between cpus, optional
3030 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3031 *
3032 * This is a helper to ease setting up embedded first percpu chunk and
3033 * can be called where pcpu_setup_first_chunk() is expected.
3034 *
3035 * If this function is used to setup the first chunk, it is allocated
3036 * by calling pcpu_fc_alloc and used as-is without being mapped into
3037 * vmalloc area. Allocations are always whole multiples of @atom_size
3038 * aligned to @atom_size.
3039 *
3040 * This enables the first chunk to piggy back on the linear physical
3041 * mapping which often uses larger page size. Please note that this
3042 * can result in very sparse cpu->unit mapping on NUMA machines thus
3043 * requiring large vmalloc address space. Don't use this allocator if
3044 * vmalloc space is not orders of magnitude larger than distances
3045 * between node memory addresses (ie. 32bit NUMA machines).
3046 *
3047 * @dyn_size specifies the minimum dynamic area size.
3048 *
3049 * If the needed size is smaller than the minimum or specified unit
3050 * size, the leftover is returned using pcpu_fc_free.
3051 *
3052 * RETURNS:
3053 * 0 on success, -errno on failure.
3054 */
3055int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3056 size_t atom_size,
3057 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3058 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3059{
3060 void *base = (void *)ULONG_MAX;
3061 void **areas = NULL;
3062 struct pcpu_alloc_info *ai;
3063 size_t size_sum, areas_size;
3064 unsigned long max_distance;
3065 int group, i, highest_group, rc = 0;
3066
3067 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3068 cpu_distance_fn);
3069 if (IS_ERR(ai))
3070 return PTR_ERR(ai);
3071
3072 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3073 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3074
3075 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
3076 if (!areas) {
3077 rc = -ENOMEM;
3078 goto out_free;
3079 }
3080
3081 /* allocate, copy and determine base address & max_distance */
3082 highest_group = 0;
3083 for (group = 0; group < ai->nr_groups; group++) {
3084 struct pcpu_group_info *gi = &ai->groups[group];
3085 unsigned int cpu = NR_CPUS;
3086 void *ptr;
3087
3088 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3089 cpu = gi->cpu_map[i];
3090 BUG_ON(cpu == NR_CPUS);
3091
3092 /* allocate space for the whole group */
3093 ptr = pcpu_fc_alloc(cpu, gi->nr_units * ai->unit_size, atom_size, cpu_to_nd_fn);
3094 if (!ptr) {
3095 rc = -ENOMEM;
3096 goto out_free_areas;
3097 }
3098 /* kmemleak tracks the percpu allocations separately */
3099 kmemleak_ignore_phys(__pa(ptr));
3100 areas[group] = ptr;
3101
3102 base = min(ptr, base);
3103 if (ptr > areas[highest_group])
3104 highest_group = group;
3105 }
3106 max_distance = areas[highest_group] - base;
3107 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3108
3109 /* warn if maximum distance is further than 75% of vmalloc space */
3110 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3111 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3112 max_distance, VMALLOC_TOTAL);
3113#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3114 /* and fail if we have fallback */
3115 rc = -EINVAL;
3116 goto out_free_areas;
3117#endif
3118 }
3119
3120 /*
3121 * Copy data and free unused parts. This should happen after all
3122 * allocations are complete; otherwise, we may end up with
3123 * overlapping groups.
3124 */
3125 for (group = 0; group < ai->nr_groups; group++) {
3126 struct pcpu_group_info *gi = &ai->groups[group];
3127 void *ptr = areas[group];
3128
3129 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3130 if (gi->cpu_map[i] == NR_CPUS) {
3131 /* unused unit, free whole */
3132 pcpu_fc_free(ptr, ai->unit_size);
3133 continue;
3134 }
3135 /* copy and return the unused part */
3136 memcpy(ptr, __per_cpu_load, ai->static_size);
3137 pcpu_fc_free(ptr + size_sum, ai->unit_size - size_sum);
3138 }
3139 }
3140
3141 /* base address is now known, determine group base offsets */
3142 for (group = 0; group < ai->nr_groups; group++) {
3143 ai->groups[group].base_offset = areas[group] - base;
3144 }
3145
3146 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3147 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3148 ai->dyn_size, ai->unit_size);
3149
3150 pcpu_setup_first_chunk(ai, base);
3151 goto out_free;
3152
3153out_free_areas:
3154 for (group = 0; group < ai->nr_groups; group++)
3155 if (areas[group])
3156 pcpu_fc_free(areas[group],
3157 ai->groups[group].nr_units * ai->unit_size);
3158out_free:
3159 pcpu_free_alloc_info(ai);
3160 if (areas)
3161 memblock_free(areas, areas_size);
3162 return rc;
3163}
3164#endif /* BUILD_EMBED_FIRST_CHUNK */
3165
3166#ifdef BUILD_PAGE_FIRST_CHUNK
3167#include <asm/pgalloc.h>
3168
3169#ifndef P4D_TABLE_SIZE
3170#define P4D_TABLE_SIZE PAGE_SIZE
3171#endif
3172
3173#ifndef PUD_TABLE_SIZE
3174#define PUD_TABLE_SIZE PAGE_SIZE
3175#endif
3176
3177#ifndef PMD_TABLE_SIZE
3178#define PMD_TABLE_SIZE PAGE_SIZE
3179#endif
3180
3181#ifndef PTE_TABLE_SIZE
3182#define PTE_TABLE_SIZE PAGE_SIZE
3183#endif
3184void __init __weak pcpu_populate_pte(unsigned long addr)
3185{
3186 pgd_t *pgd = pgd_offset_k(addr);
3187 p4d_t *p4d;
3188 pud_t *pud;
3189 pmd_t *pmd;
3190
3191 if (pgd_none(*pgd)) {
3192 p4d_t *new;
3193
3194 new = memblock_alloc(P4D_TABLE_SIZE, P4D_TABLE_SIZE);
3195 if (!new)
3196 goto err_alloc;
3197 pgd_populate(&init_mm, pgd, new);
3198 }
3199
3200 p4d = p4d_offset(pgd, addr);
3201 if (p4d_none(*p4d)) {
3202 pud_t *new;
3203
3204 new = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
3205 if (!new)
3206 goto err_alloc;
3207 p4d_populate(&init_mm, p4d, new);
3208 }
3209
3210 pud = pud_offset(p4d, addr);
3211 if (pud_none(*pud)) {
3212 pmd_t *new;
3213
3214 new = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
3215 if (!new)
3216 goto err_alloc;
3217 pud_populate(&init_mm, pud, new);
3218 }
3219
3220 pmd = pmd_offset(pud, addr);
3221 if (!pmd_present(*pmd)) {
3222 pte_t *new;
3223
3224 new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
3225 if (!new)
3226 goto err_alloc;
3227 pmd_populate_kernel(&init_mm, pmd, new);
3228 }
3229
3230 return;
3231
3232err_alloc:
3233 panic("%s: Failed to allocate memory\n", __func__);
3234}
3235
3236/**
3237 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3238 * @reserved_size: the size of reserved percpu area in bytes
3239 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3240 *
3241 * This is a helper to ease setting up page-remapped first percpu
3242 * chunk and can be called where pcpu_setup_first_chunk() is expected.
3243 *
3244 * This is the basic allocator. Static percpu area is allocated
3245 * page-by-page into vmalloc area.
3246 *
3247 * RETURNS:
3248 * 0 on success, -errno on failure.
3249 */
3250int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3251{
3252 static struct vm_struct vm;
3253 struct pcpu_alloc_info *ai;
3254 char psize_str[16];
3255 int unit_pages;
3256 size_t pages_size;
3257 struct page **pages;
3258 int unit, i, j, rc = 0;
3259 int upa;
3260 int nr_g0_units;
3261
3262 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
3263
3264 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
3265 if (IS_ERR(ai))
3266 return PTR_ERR(ai);
3267 BUG_ON(ai->nr_groups != 1);
3268 upa = ai->alloc_size/ai->unit_size;
3269 nr_g0_units = roundup(num_possible_cpus(), upa);
3270 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3271 pcpu_free_alloc_info(ai);
3272 return -EINVAL;
3273 }
3274
3275 unit_pages = ai->unit_size >> PAGE_SHIFT;
3276
3277 /* unaligned allocations can't be freed, round up to page size */
3278 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3279 sizeof(pages[0]));
3280 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
3281 if (!pages)
3282 panic("%s: Failed to allocate %zu bytes\n", __func__,
3283 pages_size);
3284
3285 /* allocate pages */
3286 j = 0;
3287 for (unit = 0; unit < num_possible_cpus(); unit++) {
3288 unsigned int cpu = ai->groups[0].cpu_map[unit];
3289 for (i = 0; i < unit_pages; i++) {
3290 void *ptr;
3291
3292 ptr = pcpu_fc_alloc(cpu, PAGE_SIZE, PAGE_SIZE, cpu_to_nd_fn);
3293 if (!ptr) {
3294 pr_warn("failed to allocate %s page for cpu%u\n",
3295 psize_str, cpu);
3296 goto enomem;
3297 }
3298 /* kmemleak tracks the percpu allocations separately */
3299 kmemleak_ignore_phys(__pa(ptr));
3300 pages[j++] = virt_to_page(ptr);
3301 }
3302 }
3303
3304 /* allocate vm area, map the pages and copy static data */
3305 vm.flags = VM_ALLOC;
3306 vm.size = num_possible_cpus() * ai->unit_size;
3307 vm_area_register_early(&vm, PAGE_SIZE);
3308
3309 for (unit = 0; unit < num_possible_cpus(); unit++) {
3310 unsigned long unit_addr =
3311 (unsigned long)vm.addr + unit * ai->unit_size;
3312
3313 for (i = 0; i < unit_pages; i++)
3314 pcpu_populate_pte(unit_addr + (i << PAGE_SHIFT));
3315
3316 /* pte already populated, the following shouldn't fail */
3317 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3318 unit_pages);
3319 if (rc < 0)
3320 panic("failed to map percpu area, err=%d\n", rc);
3321
3322 /*
3323 * FIXME: Archs with virtual cache should flush local
3324 * cache for the linear mapping here - something
3325 * equivalent to flush_cache_vmap() on the local cpu.
3326 * flush_cache_vmap() can't be used as most supporting
3327 * data structures are not set up yet.
3328 */
3329
3330 /* copy static data */
3331 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3332 }
3333
3334 /* we're ready, commit */
3335 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3336 unit_pages, psize_str, ai->static_size,
3337 ai->reserved_size, ai->dyn_size);
3338
3339 pcpu_setup_first_chunk(ai, vm.addr);
3340 goto out_free_ar;
3341
3342enomem:
3343 while (--j >= 0)
3344 pcpu_fc_free(page_address(pages[j]), PAGE_SIZE);
3345 rc = -ENOMEM;
3346out_free_ar:
3347 memblock_free(pages, pages_size);
3348 pcpu_free_alloc_info(ai);
3349 return rc;
3350}
3351#endif /* BUILD_PAGE_FIRST_CHUNK */
3352
3353#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3354/*
3355 * Generic SMP percpu area setup.
3356 *
3357 * The embedding helper is used because its behavior closely resembles
3358 * the original non-dynamic generic percpu area setup. This is
3359 * important because many archs have addressing restrictions and might
3360 * fail if the percpu area is located far away from the previous
3361 * location. As an added bonus, in non-NUMA cases, embedding is
3362 * generally a good idea TLB-wise because percpu area can piggy back
3363 * on the physical linear memory mapping which uses large page
3364 * mappings on applicable archs.
3365 */
3366unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3367EXPORT_SYMBOL(__per_cpu_offset);
3368
3369void __init setup_per_cpu_areas(void)
3370{
3371 unsigned long delta;
3372 unsigned int cpu;
3373 int rc;
3374
3375 /*
3376 * Always reserve area for module percpu variables. That's
3377 * what the legacy allocator did.
3378 */
3379 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, PERCPU_DYNAMIC_RESERVE,
3380 PAGE_SIZE, NULL, NULL);
3381 if (rc < 0)
3382 panic("Failed to initialize percpu areas.");
3383
3384 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3385 for_each_possible_cpu(cpu)
3386 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3387}
3388#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3389
3390#else /* CONFIG_SMP */
3391
3392/*
3393 * UP percpu area setup.
3394 *
3395 * UP always uses km-based percpu allocator with identity mapping.
3396 * Static percpu variables are indistinguishable from the usual static
3397 * variables and don't require any special preparation.
3398 */
3399void __init setup_per_cpu_areas(void)
3400{
3401 const size_t unit_size =
3402 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3403 PERCPU_DYNAMIC_RESERVE));
3404 struct pcpu_alloc_info *ai;
3405 void *fc;
3406
3407 ai = pcpu_alloc_alloc_info(1, 1);
3408 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3409 if (!ai || !fc)
3410 panic("Failed to allocate memory for percpu areas.");
3411 /* kmemleak tracks the percpu allocations separately */
3412 kmemleak_ignore_phys(__pa(fc));
3413
3414 ai->dyn_size = unit_size;
3415 ai->unit_size = unit_size;
3416 ai->atom_size = unit_size;
3417 ai->alloc_size = unit_size;
3418 ai->groups[0].nr_units = 1;
3419 ai->groups[0].cpu_map[0] = 0;
3420
3421 pcpu_setup_first_chunk(ai, fc);
3422 pcpu_free_alloc_info(ai);
3423}
3424
3425#endif /* CONFIG_SMP */
3426
3427/*
3428 * pcpu_nr_pages - calculate total number of populated backing pages
3429 *
3430 * This reflects the number of pages populated to back chunks. Metadata is
3431 * excluded in the number exposed in meminfo as the number of backing pages
3432 * scales with the number of cpus and can quickly outweigh the memory used for
3433 * metadata. It also keeps this calculation nice and simple.
3434 *
3435 * RETURNS:
3436 * Total number of populated backing pages in use by the allocator.
3437 */
3438unsigned long pcpu_nr_pages(void)
3439{
3440 return pcpu_nr_populated * pcpu_nr_units;
3441}
3442
3443/*
3444 * Percpu allocator is initialized early during boot when neither slab or
3445 * workqueue is available. Plug async management until everything is up
3446 * and running.
3447 */
3448static int __init percpu_enable_async(void)
3449{
3450 pcpu_async_enabled = true;
3451 return 0;
3452}
3453subsys_initcall(percpu_enable_async);