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