Loading...
1/*
2 * mm/percpu.c - percpu memory allocator
3 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated.
17 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
29 *
30 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
32 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be equal to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
43 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
45 *
46 * To use this allocator, arch code should do the followings.
47 *
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be
50 * different from the default
51 *
52 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
54 */
55
56#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
57
58#include <linux/bitmap.h>
59#include <linux/bootmem.h>
60#include <linux/err.h>
61#include <linux/list.h>
62#include <linux/log2.h>
63#include <linux/mm.h>
64#include <linux/module.h>
65#include <linux/mutex.h>
66#include <linux/percpu.h>
67#include <linux/pfn.h>
68#include <linux/slab.h>
69#include <linux/spinlock.h>
70#include <linux/vmalloc.h>
71#include <linux/workqueue.h>
72#include <linux/kmemleak.h>
73
74#include <asm/cacheflush.h>
75#include <asm/sections.h>
76#include <asm/tlbflush.h>
77#include <asm/io.h>
78
79#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
80#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
81#define PCPU_ATOMIC_MAP_MARGIN_LOW 32
82#define PCPU_ATOMIC_MAP_MARGIN_HIGH 64
83#define PCPU_EMPTY_POP_PAGES_LOW 2
84#define PCPU_EMPTY_POP_PAGES_HIGH 4
85
86#ifdef CONFIG_SMP
87/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
88#ifndef __addr_to_pcpu_ptr
89#define __addr_to_pcpu_ptr(addr) \
90 (void __percpu *)((unsigned long)(addr) - \
91 (unsigned long)pcpu_base_addr + \
92 (unsigned long)__per_cpu_start)
93#endif
94#ifndef __pcpu_ptr_to_addr
95#define __pcpu_ptr_to_addr(ptr) \
96 (void __force *)((unsigned long)(ptr) + \
97 (unsigned long)pcpu_base_addr - \
98 (unsigned long)__per_cpu_start)
99#endif
100#else /* CONFIG_SMP */
101/* on UP, it's always identity mapped */
102#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
103#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
104#endif /* CONFIG_SMP */
105
106struct pcpu_chunk {
107 struct list_head list; /* linked to pcpu_slot lists */
108 int free_size; /* free bytes in the chunk */
109 int contig_hint; /* max contiguous size hint */
110 void *base_addr; /* base address of this chunk */
111
112 int map_used; /* # of map entries used before the sentry */
113 int map_alloc; /* # of map entries allocated */
114 int *map; /* allocation map */
115 struct work_struct map_extend_work;/* async ->map[] extension */
116
117 void *data; /* chunk data */
118 int first_free; /* no free below this */
119 bool immutable; /* no [de]population allowed */
120 int nr_populated; /* # of populated pages */
121 unsigned long populated[]; /* populated bitmap */
122};
123
124static int pcpu_unit_pages __read_mostly;
125static int pcpu_unit_size __read_mostly;
126static int pcpu_nr_units __read_mostly;
127static int pcpu_atom_size __read_mostly;
128static int pcpu_nr_slots __read_mostly;
129static size_t pcpu_chunk_struct_size __read_mostly;
130
131/* cpus with the lowest and highest unit addresses */
132static unsigned int pcpu_low_unit_cpu __read_mostly;
133static unsigned int pcpu_high_unit_cpu __read_mostly;
134
135/* the address of the first chunk which starts with the kernel static area */
136void *pcpu_base_addr __read_mostly;
137EXPORT_SYMBOL_GPL(pcpu_base_addr);
138
139static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
140const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
141
142/* group information, used for vm allocation */
143static int pcpu_nr_groups __read_mostly;
144static const unsigned long *pcpu_group_offsets __read_mostly;
145static const size_t *pcpu_group_sizes __read_mostly;
146
147/*
148 * The first chunk which always exists. Note that unlike other
149 * chunks, this one can be allocated and mapped in several different
150 * ways and thus often doesn't live in the vmalloc area.
151 */
152static struct pcpu_chunk *pcpu_first_chunk;
153
154/*
155 * Optional reserved chunk. This chunk reserves part of the first
156 * chunk and serves it for reserved allocations. The amount of
157 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
158 * area doesn't exist, the following variables contain NULL and 0
159 * respectively.
160 */
161static struct pcpu_chunk *pcpu_reserved_chunk;
162static int pcpu_reserved_chunk_limit;
163
164static DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
165static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop */
166
167static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
168
169/*
170 * The number of empty populated pages, protected by pcpu_lock. The
171 * reserved chunk doesn't contribute to the count.
172 */
173static int pcpu_nr_empty_pop_pages;
174
175/*
176 * Balance work is used to populate or destroy chunks asynchronously. We
177 * try to keep the number of populated free pages between
178 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
179 * empty chunk.
180 */
181static void pcpu_balance_workfn(struct work_struct *work);
182static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
183static bool pcpu_async_enabled __read_mostly;
184static bool pcpu_atomic_alloc_failed;
185
186static void pcpu_schedule_balance_work(void)
187{
188 if (pcpu_async_enabled)
189 schedule_work(&pcpu_balance_work);
190}
191
192static bool pcpu_addr_in_first_chunk(void *addr)
193{
194 void *first_start = pcpu_first_chunk->base_addr;
195
196 return addr >= first_start && addr < first_start + pcpu_unit_size;
197}
198
199static bool pcpu_addr_in_reserved_chunk(void *addr)
200{
201 void *first_start = pcpu_first_chunk->base_addr;
202
203 return addr >= first_start &&
204 addr < first_start + pcpu_reserved_chunk_limit;
205}
206
207static int __pcpu_size_to_slot(int size)
208{
209 int highbit = fls(size); /* size is in bytes */
210 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
211}
212
213static int pcpu_size_to_slot(int size)
214{
215 if (size == pcpu_unit_size)
216 return pcpu_nr_slots - 1;
217 return __pcpu_size_to_slot(size);
218}
219
220static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
221{
222 if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
223 return 0;
224
225 return pcpu_size_to_slot(chunk->free_size);
226}
227
228/* set the pointer to a chunk in a page struct */
229static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
230{
231 page->index = (unsigned long)pcpu;
232}
233
234/* obtain pointer to a chunk from a page struct */
235static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
236{
237 return (struct pcpu_chunk *)page->index;
238}
239
240static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
241{
242 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
243}
244
245static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
246 unsigned int cpu, int page_idx)
247{
248 return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
249 (page_idx << PAGE_SHIFT);
250}
251
252static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
253 int *rs, int *re, int end)
254{
255 *rs = find_next_zero_bit(chunk->populated, end, *rs);
256 *re = find_next_bit(chunk->populated, end, *rs + 1);
257}
258
259static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
260 int *rs, int *re, int end)
261{
262 *rs = find_next_bit(chunk->populated, end, *rs);
263 *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
264}
265
266/*
267 * (Un)populated page region iterators. Iterate over (un)populated
268 * page regions between @start and @end in @chunk. @rs and @re should
269 * be integer variables and will be set to start and end page index of
270 * the current region.
271 */
272#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
273 for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
274 (rs) < (re); \
275 (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
276
277#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
278 for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
279 (rs) < (re); \
280 (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
281
282/**
283 * pcpu_mem_zalloc - allocate memory
284 * @size: bytes to allocate
285 *
286 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
287 * kzalloc() is used; otherwise, vzalloc() is used. The returned
288 * memory is always zeroed.
289 *
290 * CONTEXT:
291 * Does GFP_KERNEL allocation.
292 *
293 * RETURNS:
294 * Pointer to the allocated area on success, NULL on failure.
295 */
296static void *pcpu_mem_zalloc(size_t size)
297{
298 if (WARN_ON_ONCE(!slab_is_available()))
299 return NULL;
300
301 if (size <= PAGE_SIZE)
302 return kzalloc(size, GFP_KERNEL);
303 else
304 return vzalloc(size);
305}
306
307/**
308 * pcpu_mem_free - free memory
309 * @ptr: memory to free
310 *
311 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
312 */
313static void pcpu_mem_free(void *ptr)
314{
315 kvfree(ptr);
316}
317
318/**
319 * pcpu_count_occupied_pages - count the number of pages an area occupies
320 * @chunk: chunk of interest
321 * @i: index of the area in question
322 *
323 * Count the number of pages chunk's @i'th area occupies. When the area's
324 * start and/or end address isn't aligned to page boundary, the straddled
325 * page is included in the count iff the rest of the page is free.
326 */
327static int pcpu_count_occupied_pages(struct pcpu_chunk *chunk, int i)
328{
329 int off = chunk->map[i] & ~1;
330 int end = chunk->map[i + 1] & ~1;
331
332 if (!PAGE_ALIGNED(off) && i > 0) {
333 int prev = chunk->map[i - 1];
334
335 if (!(prev & 1) && prev <= round_down(off, PAGE_SIZE))
336 off = round_down(off, PAGE_SIZE);
337 }
338
339 if (!PAGE_ALIGNED(end) && i + 1 < chunk->map_used) {
340 int next = chunk->map[i + 1];
341 int nend = chunk->map[i + 2] & ~1;
342
343 if (!(next & 1) && nend >= round_up(end, PAGE_SIZE))
344 end = round_up(end, PAGE_SIZE);
345 }
346
347 return max_t(int, PFN_DOWN(end) - PFN_UP(off), 0);
348}
349
350/**
351 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
352 * @chunk: chunk of interest
353 * @oslot: the previous slot it was on
354 *
355 * This function is called after an allocation or free changed @chunk.
356 * New slot according to the changed state is determined and @chunk is
357 * moved to the slot. Note that the reserved chunk is never put on
358 * chunk slots.
359 *
360 * CONTEXT:
361 * pcpu_lock.
362 */
363static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
364{
365 int nslot = pcpu_chunk_slot(chunk);
366
367 if (chunk != pcpu_reserved_chunk && oslot != nslot) {
368 if (oslot < nslot)
369 list_move(&chunk->list, &pcpu_slot[nslot]);
370 else
371 list_move_tail(&chunk->list, &pcpu_slot[nslot]);
372 }
373}
374
375/**
376 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
377 * @chunk: chunk of interest
378 * @is_atomic: the allocation context
379 *
380 * Determine whether area map of @chunk needs to be extended. If
381 * @is_atomic, only the amount necessary for a new allocation is
382 * considered; however, async extension is scheduled if the left amount is
383 * low. If !@is_atomic, it aims for more empty space. Combined, this
384 * ensures that the map is likely to have enough available space to
385 * accomodate atomic allocations which can't extend maps directly.
386 *
387 * CONTEXT:
388 * pcpu_lock.
389 *
390 * RETURNS:
391 * New target map allocation length if extension is necessary, 0
392 * otherwise.
393 */
394static int pcpu_need_to_extend(struct pcpu_chunk *chunk, bool is_atomic)
395{
396 int margin, new_alloc;
397
398 if (is_atomic) {
399 margin = 3;
400
401 if (chunk->map_alloc <
402 chunk->map_used + PCPU_ATOMIC_MAP_MARGIN_LOW &&
403 pcpu_async_enabled)
404 schedule_work(&chunk->map_extend_work);
405 } else {
406 margin = PCPU_ATOMIC_MAP_MARGIN_HIGH;
407 }
408
409 if (chunk->map_alloc >= chunk->map_used + margin)
410 return 0;
411
412 new_alloc = PCPU_DFL_MAP_ALLOC;
413 while (new_alloc < chunk->map_used + margin)
414 new_alloc *= 2;
415
416 return new_alloc;
417}
418
419/**
420 * pcpu_extend_area_map - extend area map of a chunk
421 * @chunk: chunk of interest
422 * @new_alloc: new target allocation length of the area map
423 *
424 * Extend area map of @chunk to have @new_alloc entries.
425 *
426 * CONTEXT:
427 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
428 *
429 * RETURNS:
430 * 0 on success, -errno on failure.
431 */
432static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
433{
434 int *old = NULL, *new = NULL;
435 size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
436 unsigned long flags;
437
438 new = pcpu_mem_zalloc(new_size);
439 if (!new)
440 return -ENOMEM;
441
442 /* acquire pcpu_lock and switch to new area map */
443 spin_lock_irqsave(&pcpu_lock, flags);
444
445 if (new_alloc <= chunk->map_alloc)
446 goto out_unlock;
447
448 old_size = chunk->map_alloc * sizeof(chunk->map[0]);
449 old = chunk->map;
450
451 memcpy(new, old, old_size);
452
453 chunk->map_alloc = new_alloc;
454 chunk->map = new;
455 new = NULL;
456
457out_unlock:
458 spin_unlock_irqrestore(&pcpu_lock, flags);
459
460 /*
461 * pcpu_mem_free() might end up calling vfree() which uses
462 * IRQ-unsafe lock and thus can't be called under pcpu_lock.
463 */
464 pcpu_mem_free(old);
465 pcpu_mem_free(new);
466
467 return 0;
468}
469
470static void pcpu_map_extend_workfn(struct work_struct *work)
471{
472 struct pcpu_chunk *chunk = container_of(work, struct pcpu_chunk,
473 map_extend_work);
474 int new_alloc;
475
476 spin_lock_irq(&pcpu_lock);
477 new_alloc = pcpu_need_to_extend(chunk, false);
478 spin_unlock_irq(&pcpu_lock);
479
480 if (new_alloc)
481 pcpu_extend_area_map(chunk, new_alloc);
482}
483
484/**
485 * pcpu_fit_in_area - try to fit the requested allocation in a candidate area
486 * @chunk: chunk the candidate area belongs to
487 * @off: the offset to the start of the candidate area
488 * @this_size: the size of the candidate area
489 * @size: the size of the target allocation
490 * @align: the alignment of the target allocation
491 * @pop_only: only allocate from already populated region
492 *
493 * We're trying to allocate @size bytes aligned at @align. @chunk's area
494 * at @off sized @this_size is a candidate. This function determines
495 * whether the target allocation fits in the candidate area and returns the
496 * number of bytes to pad after @off. If the target area doesn't fit, -1
497 * is returned.
498 *
499 * If @pop_only is %true, this function only considers the already
500 * populated part of the candidate area.
501 */
502static int pcpu_fit_in_area(struct pcpu_chunk *chunk, int off, int this_size,
503 int size, int align, bool pop_only)
504{
505 int cand_off = off;
506
507 while (true) {
508 int head = ALIGN(cand_off, align) - off;
509 int page_start, page_end, rs, re;
510
511 if (this_size < head + size)
512 return -1;
513
514 if (!pop_only)
515 return head;
516
517 /*
518 * If the first unpopulated page is beyond the end of the
519 * allocation, the whole allocation is populated;
520 * otherwise, retry from the end of the unpopulated area.
521 */
522 page_start = PFN_DOWN(head + off);
523 page_end = PFN_UP(head + off + size);
524
525 rs = page_start;
526 pcpu_next_unpop(chunk, &rs, &re, PFN_UP(off + this_size));
527 if (rs >= page_end)
528 return head;
529 cand_off = re * PAGE_SIZE;
530 }
531}
532
533/**
534 * pcpu_alloc_area - allocate area from a pcpu_chunk
535 * @chunk: chunk of interest
536 * @size: wanted size in bytes
537 * @align: wanted align
538 * @pop_only: allocate only from the populated area
539 * @occ_pages_p: out param for the number of pages the area occupies
540 *
541 * Try to allocate @size bytes area aligned at @align from @chunk.
542 * Note that this function only allocates the offset. It doesn't
543 * populate or map the area.
544 *
545 * @chunk->map must have at least two free slots.
546 *
547 * CONTEXT:
548 * pcpu_lock.
549 *
550 * RETURNS:
551 * Allocated offset in @chunk on success, -1 if no matching area is
552 * found.
553 */
554static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align,
555 bool pop_only, int *occ_pages_p)
556{
557 int oslot = pcpu_chunk_slot(chunk);
558 int max_contig = 0;
559 int i, off;
560 bool seen_free = false;
561 int *p;
562
563 for (i = chunk->first_free, p = chunk->map + i; i < chunk->map_used; i++, p++) {
564 int head, tail;
565 int this_size;
566
567 off = *p;
568 if (off & 1)
569 continue;
570
571 this_size = (p[1] & ~1) - off;
572
573 head = pcpu_fit_in_area(chunk, off, this_size, size, align,
574 pop_only);
575 if (head < 0) {
576 if (!seen_free) {
577 chunk->first_free = i;
578 seen_free = true;
579 }
580 max_contig = max(this_size, max_contig);
581 continue;
582 }
583
584 /*
585 * If head is small or the previous block is free,
586 * merge'em. Note that 'small' is defined as smaller
587 * than sizeof(int), which is very small but isn't too
588 * uncommon for percpu allocations.
589 */
590 if (head && (head < sizeof(int) || !(p[-1] & 1))) {
591 *p = off += head;
592 if (p[-1] & 1)
593 chunk->free_size -= head;
594 else
595 max_contig = max(*p - p[-1], max_contig);
596 this_size -= head;
597 head = 0;
598 }
599
600 /* if tail is small, just keep it around */
601 tail = this_size - head - size;
602 if (tail < sizeof(int)) {
603 tail = 0;
604 size = this_size - head;
605 }
606
607 /* split if warranted */
608 if (head || tail) {
609 int nr_extra = !!head + !!tail;
610
611 /* insert new subblocks */
612 memmove(p + nr_extra + 1, p + 1,
613 sizeof(chunk->map[0]) * (chunk->map_used - i));
614 chunk->map_used += nr_extra;
615
616 if (head) {
617 if (!seen_free) {
618 chunk->first_free = i;
619 seen_free = true;
620 }
621 *++p = off += head;
622 ++i;
623 max_contig = max(head, max_contig);
624 }
625 if (tail) {
626 p[1] = off + size;
627 max_contig = max(tail, max_contig);
628 }
629 }
630
631 if (!seen_free)
632 chunk->first_free = i + 1;
633
634 /* update hint and mark allocated */
635 if (i + 1 == chunk->map_used)
636 chunk->contig_hint = max_contig; /* fully scanned */
637 else
638 chunk->contig_hint = max(chunk->contig_hint,
639 max_contig);
640
641 chunk->free_size -= size;
642 *p |= 1;
643
644 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
645 pcpu_chunk_relocate(chunk, oslot);
646 return off;
647 }
648
649 chunk->contig_hint = max_contig; /* fully scanned */
650 pcpu_chunk_relocate(chunk, oslot);
651
652 /* tell the upper layer that this chunk has no matching area */
653 return -1;
654}
655
656/**
657 * pcpu_free_area - free area to a pcpu_chunk
658 * @chunk: chunk of interest
659 * @freeme: offset of area to free
660 * @occ_pages_p: out param for the number of pages the area occupies
661 *
662 * Free area starting from @freeme to @chunk. Note that this function
663 * only modifies the allocation map. It doesn't depopulate or unmap
664 * the area.
665 *
666 * CONTEXT:
667 * pcpu_lock.
668 */
669static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme,
670 int *occ_pages_p)
671{
672 int oslot = pcpu_chunk_slot(chunk);
673 int off = 0;
674 unsigned i, j;
675 int to_free = 0;
676 int *p;
677
678 freeme |= 1; /* we are searching for <given offset, in use> pair */
679
680 i = 0;
681 j = chunk->map_used;
682 while (i != j) {
683 unsigned k = (i + j) / 2;
684 off = chunk->map[k];
685 if (off < freeme)
686 i = k + 1;
687 else if (off > freeme)
688 j = k;
689 else
690 i = j = k;
691 }
692 BUG_ON(off != freeme);
693
694 if (i < chunk->first_free)
695 chunk->first_free = i;
696
697 p = chunk->map + i;
698 *p = off &= ~1;
699 chunk->free_size += (p[1] & ~1) - off;
700
701 *occ_pages_p = pcpu_count_occupied_pages(chunk, i);
702
703 /* merge with next? */
704 if (!(p[1] & 1))
705 to_free++;
706 /* merge with previous? */
707 if (i > 0 && !(p[-1] & 1)) {
708 to_free++;
709 i--;
710 p--;
711 }
712 if (to_free) {
713 chunk->map_used -= to_free;
714 memmove(p + 1, p + 1 + to_free,
715 (chunk->map_used - i) * sizeof(chunk->map[0]));
716 }
717
718 chunk->contig_hint = max(chunk->map[i + 1] - chunk->map[i] - 1, chunk->contig_hint);
719 pcpu_chunk_relocate(chunk, oslot);
720}
721
722static struct pcpu_chunk *pcpu_alloc_chunk(void)
723{
724 struct pcpu_chunk *chunk;
725
726 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size);
727 if (!chunk)
728 return NULL;
729
730 chunk->map = pcpu_mem_zalloc(PCPU_DFL_MAP_ALLOC *
731 sizeof(chunk->map[0]));
732 if (!chunk->map) {
733 pcpu_mem_free(chunk);
734 return NULL;
735 }
736
737 chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
738 chunk->map[0] = 0;
739 chunk->map[1] = pcpu_unit_size | 1;
740 chunk->map_used = 1;
741
742 INIT_LIST_HEAD(&chunk->list);
743 INIT_WORK(&chunk->map_extend_work, pcpu_map_extend_workfn);
744 chunk->free_size = pcpu_unit_size;
745 chunk->contig_hint = pcpu_unit_size;
746
747 return chunk;
748}
749
750static void pcpu_free_chunk(struct pcpu_chunk *chunk)
751{
752 if (!chunk)
753 return;
754 pcpu_mem_free(chunk->map);
755 pcpu_mem_free(chunk);
756}
757
758/**
759 * pcpu_chunk_populated - post-population bookkeeping
760 * @chunk: pcpu_chunk which got populated
761 * @page_start: the start page
762 * @page_end: the end page
763 *
764 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
765 * the bookkeeping information accordingly. Must be called after each
766 * successful population.
767 */
768static void pcpu_chunk_populated(struct pcpu_chunk *chunk,
769 int page_start, int page_end)
770{
771 int nr = page_end - page_start;
772
773 lockdep_assert_held(&pcpu_lock);
774
775 bitmap_set(chunk->populated, page_start, nr);
776 chunk->nr_populated += nr;
777 pcpu_nr_empty_pop_pages += nr;
778}
779
780/**
781 * pcpu_chunk_depopulated - post-depopulation bookkeeping
782 * @chunk: pcpu_chunk which got depopulated
783 * @page_start: the start page
784 * @page_end: the end page
785 *
786 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
787 * Update the bookkeeping information accordingly. Must be called after
788 * each successful depopulation.
789 */
790static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
791 int page_start, int page_end)
792{
793 int nr = page_end - page_start;
794
795 lockdep_assert_held(&pcpu_lock);
796
797 bitmap_clear(chunk->populated, page_start, nr);
798 chunk->nr_populated -= nr;
799 pcpu_nr_empty_pop_pages -= nr;
800}
801
802/*
803 * Chunk management implementation.
804 *
805 * To allow different implementations, chunk alloc/free and
806 * [de]population are implemented in a separate file which is pulled
807 * into this file and compiled together. The following functions
808 * should be implemented.
809 *
810 * pcpu_populate_chunk - populate the specified range of a chunk
811 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
812 * pcpu_create_chunk - create a new chunk
813 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
814 * pcpu_addr_to_page - translate address to physical address
815 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
816 */
817static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
818static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
819static struct pcpu_chunk *pcpu_create_chunk(void);
820static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
821static struct page *pcpu_addr_to_page(void *addr);
822static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
823
824#ifdef CONFIG_NEED_PER_CPU_KM
825#include "percpu-km.c"
826#else
827#include "percpu-vm.c"
828#endif
829
830/**
831 * pcpu_chunk_addr_search - determine chunk containing specified address
832 * @addr: address for which the chunk needs to be determined.
833 *
834 * RETURNS:
835 * The address of the found chunk.
836 */
837static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
838{
839 /* is it in the first chunk? */
840 if (pcpu_addr_in_first_chunk(addr)) {
841 /* is it in the reserved area? */
842 if (pcpu_addr_in_reserved_chunk(addr))
843 return pcpu_reserved_chunk;
844 return pcpu_first_chunk;
845 }
846
847 /*
848 * The address is relative to unit0 which might be unused and
849 * thus unmapped. Offset the address to the unit space of the
850 * current processor before looking it up in the vmalloc
851 * space. Note that any possible cpu id can be used here, so
852 * there's no need to worry about preemption or cpu hotplug.
853 */
854 addr += pcpu_unit_offsets[raw_smp_processor_id()];
855 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
856}
857
858/**
859 * pcpu_alloc - the percpu allocator
860 * @size: size of area to allocate in bytes
861 * @align: alignment of area (max PAGE_SIZE)
862 * @reserved: allocate from the reserved chunk if available
863 * @gfp: allocation flags
864 *
865 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
866 * contain %GFP_KERNEL, the allocation is atomic.
867 *
868 * RETURNS:
869 * Percpu pointer to the allocated area on success, NULL on failure.
870 */
871static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
872 gfp_t gfp)
873{
874 static int warn_limit = 10;
875 struct pcpu_chunk *chunk;
876 const char *err;
877 bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
878 int occ_pages = 0;
879 int slot, off, new_alloc, cpu, ret;
880 unsigned long flags;
881 void __percpu *ptr;
882
883 /*
884 * We want the lowest bit of offset available for in-use/free
885 * indicator, so force >= 16bit alignment and make size even.
886 */
887 if (unlikely(align < 2))
888 align = 2;
889
890 size = ALIGN(size, 2);
891
892 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
893 WARN(true, "illegal size (%zu) or align (%zu) for percpu allocation\n",
894 size, align);
895 return NULL;
896 }
897
898 spin_lock_irqsave(&pcpu_lock, flags);
899
900 /* serve reserved allocations from the reserved chunk if available */
901 if (reserved && pcpu_reserved_chunk) {
902 chunk = pcpu_reserved_chunk;
903
904 if (size > chunk->contig_hint) {
905 err = "alloc from reserved chunk failed";
906 goto fail_unlock;
907 }
908
909 while ((new_alloc = pcpu_need_to_extend(chunk, is_atomic))) {
910 spin_unlock_irqrestore(&pcpu_lock, flags);
911 if (is_atomic ||
912 pcpu_extend_area_map(chunk, new_alloc) < 0) {
913 err = "failed to extend area map of reserved chunk";
914 goto fail;
915 }
916 spin_lock_irqsave(&pcpu_lock, flags);
917 }
918
919 off = pcpu_alloc_area(chunk, size, align, is_atomic,
920 &occ_pages);
921 if (off >= 0)
922 goto area_found;
923
924 err = "alloc from reserved chunk failed";
925 goto fail_unlock;
926 }
927
928restart:
929 /* search through normal chunks */
930 for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
931 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
932 if (size > chunk->contig_hint)
933 continue;
934
935 new_alloc = pcpu_need_to_extend(chunk, is_atomic);
936 if (new_alloc) {
937 if (is_atomic)
938 continue;
939 spin_unlock_irqrestore(&pcpu_lock, flags);
940 if (pcpu_extend_area_map(chunk,
941 new_alloc) < 0) {
942 err = "failed to extend area map";
943 goto fail;
944 }
945 spin_lock_irqsave(&pcpu_lock, flags);
946 /*
947 * pcpu_lock has been dropped, need to
948 * restart cpu_slot list walking.
949 */
950 goto restart;
951 }
952
953 off = pcpu_alloc_area(chunk, size, align, is_atomic,
954 &occ_pages);
955 if (off >= 0)
956 goto area_found;
957 }
958 }
959
960 spin_unlock_irqrestore(&pcpu_lock, flags);
961
962 /*
963 * No space left. Create a new chunk. We don't want multiple
964 * tasks to create chunks simultaneously. Serialize and create iff
965 * there's still no empty chunk after grabbing the mutex.
966 */
967 if (is_atomic)
968 goto fail;
969
970 mutex_lock(&pcpu_alloc_mutex);
971
972 if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
973 chunk = pcpu_create_chunk();
974 if (!chunk) {
975 mutex_unlock(&pcpu_alloc_mutex);
976 err = "failed to allocate new chunk";
977 goto fail;
978 }
979
980 spin_lock_irqsave(&pcpu_lock, flags);
981 pcpu_chunk_relocate(chunk, -1);
982 } else {
983 spin_lock_irqsave(&pcpu_lock, flags);
984 }
985
986 mutex_unlock(&pcpu_alloc_mutex);
987 goto restart;
988
989area_found:
990 spin_unlock_irqrestore(&pcpu_lock, flags);
991
992 /* populate if not all pages are already there */
993 if (!is_atomic) {
994 int page_start, page_end, rs, re;
995
996 mutex_lock(&pcpu_alloc_mutex);
997
998 page_start = PFN_DOWN(off);
999 page_end = PFN_UP(off + size);
1000
1001 pcpu_for_each_unpop_region(chunk, rs, re, page_start, page_end) {
1002 WARN_ON(chunk->immutable);
1003
1004 ret = pcpu_populate_chunk(chunk, rs, re);
1005
1006 spin_lock_irqsave(&pcpu_lock, flags);
1007 if (ret) {
1008 mutex_unlock(&pcpu_alloc_mutex);
1009 pcpu_free_area(chunk, off, &occ_pages);
1010 err = "failed to populate";
1011 goto fail_unlock;
1012 }
1013 pcpu_chunk_populated(chunk, rs, re);
1014 spin_unlock_irqrestore(&pcpu_lock, flags);
1015 }
1016
1017 mutex_unlock(&pcpu_alloc_mutex);
1018 }
1019
1020 if (chunk != pcpu_reserved_chunk)
1021 pcpu_nr_empty_pop_pages -= occ_pages;
1022
1023 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1024 pcpu_schedule_balance_work();
1025
1026 /* clear the areas and return address relative to base address */
1027 for_each_possible_cpu(cpu)
1028 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1029
1030 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1031 kmemleak_alloc_percpu(ptr, size, gfp);
1032 return ptr;
1033
1034fail_unlock:
1035 spin_unlock_irqrestore(&pcpu_lock, flags);
1036fail:
1037 if (!is_atomic && warn_limit) {
1038 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1039 size, align, is_atomic, err);
1040 dump_stack();
1041 if (!--warn_limit)
1042 pr_info("limit reached, disable warning\n");
1043 }
1044 if (is_atomic) {
1045 /* see the flag handling in pcpu_blance_workfn() */
1046 pcpu_atomic_alloc_failed = true;
1047 pcpu_schedule_balance_work();
1048 }
1049 return NULL;
1050}
1051
1052/**
1053 * __alloc_percpu_gfp - allocate dynamic percpu area
1054 * @size: size of area to allocate in bytes
1055 * @align: alignment of area (max PAGE_SIZE)
1056 * @gfp: allocation flags
1057 *
1058 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1059 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1060 * be called from any context but is a lot more likely to fail.
1061 *
1062 * RETURNS:
1063 * Percpu pointer to the allocated area on success, NULL on failure.
1064 */
1065void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1066{
1067 return pcpu_alloc(size, align, false, gfp);
1068}
1069EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1070
1071/**
1072 * __alloc_percpu - allocate dynamic percpu area
1073 * @size: size of area to allocate in bytes
1074 * @align: alignment of area (max PAGE_SIZE)
1075 *
1076 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1077 */
1078void __percpu *__alloc_percpu(size_t size, size_t align)
1079{
1080 return pcpu_alloc(size, align, false, GFP_KERNEL);
1081}
1082EXPORT_SYMBOL_GPL(__alloc_percpu);
1083
1084/**
1085 * __alloc_reserved_percpu - allocate reserved percpu area
1086 * @size: size of area to allocate in bytes
1087 * @align: alignment of area (max PAGE_SIZE)
1088 *
1089 * Allocate zero-filled percpu area of @size bytes aligned at @align
1090 * from reserved percpu area if arch has set it up; otherwise,
1091 * allocation is served from the same dynamic area. Might sleep.
1092 * Might trigger writeouts.
1093 *
1094 * CONTEXT:
1095 * Does GFP_KERNEL allocation.
1096 *
1097 * RETURNS:
1098 * Percpu pointer to the allocated area on success, NULL on failure.
1099 */
1100void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1101{
1102 return pcpu_alloc(size, align, true, GFP_KERNEL);
1103}
1104
1105/**
1106 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1107 * @work: unused
1108 *
1109 * Reclaim all fully free chunks except for the first one.
1110 */
1111static void pcpu_balance_workfn(struct work_struct *work)
1112{
1113 LIST_HEAD(to_free);
1114 struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1115 struct pcpu_chunk *chunk, *next;
1116 int slot, nr_to_pop, ret;
1117
1118 /*
1119 * There's no reason to keep around multiple unused chunks and VM
1120 * areas can be scarce. Destroy all free chunks except for one.
1121 */
1122 mutex_lock(&pcpu_alloc_mutex);
1123 spin_lock_irq(&pcpu_lock);
1124
1125 list_for_each_entry_safe(chunk, next, free_head, list) {
1126 WARN_ON(chunk->immutable);
1127
1128 /* spare the first one */
1129 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1130 continue;
1131
1132 list_move(&chunk->list, &to_free);
1133 }
1134
1135 spin_unlock_irq(&pcpu_lock);
1136
1137 list_for_each_entry_safe(chunk, next, &to_free, list) {
1138 int rs, re;
1139
1140 pcpu_for_each_pop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1141 pcpu_depopulate_chunk(chunk, rs, re);
1142 spin_lock_irq(&pcpu_lock);
1143 pcpu_chunk_depopulated(chunk, rs, re);
1144 spin_unlock_irq(&pcpu_lock);
1145 }
1146 pcpu_destroy_chunk(chunk);
1147 }
1148
1149 /*
1150 * Ensure there are certain number of free populated pages for
1151 * atomic allocs. Fill up from the most packed so that atomic
1152 * allocs don't increase fragmentation. If atomic allocation
1153 * failed previously, always populate the maximum amount. This
1154 * should prevent atomic allocs larger than PAGE_SIZE from keeping
1155 * failing indefinitely; however, large atomic allocs are not
1156 * something we support properly and can be highly unreliable and
1157 * inefficient.
1158 */
1159retry_pop:
1160 if (pcpu_atomic_alloc_failed) {
1161 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1162 /* best effort anyway, don't worry about synchronization */
1163 pcpu_atomic_alloc_failed = false;
1164 } else {
1165 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1166 pcpu_nr_empty_pop_pages,
1167 0, PCPU_EMPTY_POP_PAGES_HIGH);
1168 }
1169
1170 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1171 int nr_unpop = 0, rs, re;
1172
1173 if (!nr_to_pop)
1174 break;
1175
1176 spin_lock_irq(&pcpu_lock);
1177 list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1178 nr_unpop = pcpu_unit_pages - chunk->nr_populated;
1179 if (nr_unpop)
1180 break;
1181 }
1182 spin_unlock_irq(&pcpu_lock);
1183
1184 if (!nr_unpop)
1185 continue;
1186
1187 /* @chunk can't go away while pcpu_alloc_mutex is held */
1188 pcpu_for_each_unpop_region(chunk, rs, re, 0, pcpu_unit_pages) {
1189 int nr = min(re - rs, nr_to_pop);
1190
1191 ret = pcpu_populate_chunk(chunk, rs, rs + nr);
1192 if (!ret) {
1193 nr_to_pop -= nr;
1194 spin_lock_irq(&pcpu_lock);
1195 pcpu_chunk_populated(chunk, rs, rs + nr);
1196 spin_unlock_irq(&pcpu_lock);
1197 } else {
1198 nr_to_pop = 0;
1199 }
1200
1201 if (!nr_to_pop)
1202 break;
1203 }
1204 }
1205
1206 if (nr_to_pop) {
1207 /* ran out of chunks to populate, create a new one and retry */
1208 chunk = pcpu_create_chunk();
1209 if (chunk) {
1210 spin_lock_irq(&pcpu_lock);
1211 pcpu_chunk_relocate(chunk, -1);
1212 spin_unlock_irq(&pcpu_lock);
1213 goto retry_pop;
1214 }
1215 }
1216
1217 mutex_unlock(&pcpu_alloc_mutex);
1218}
1219
1220/**
1221 * free_percpu - free percpu area
1222 * @ptr: pointer to area to free
1223 *
1224 * Free percpu area @ptr.
1225 *
1226 * CONTEXT:
1227 * Can be called from atomic context.
1228 */
1229void free_percpu(void __percpu *ptr)
1230{
1231 void *addr;
1232 struct pcpu_chunk *chunk;
1233 unsigned long flags;
1234 int off, occ_pages;
1235
1236 if (!ptr)
1237 return;
1238
1239 kmemleak_free_percpu(ptr);
1240
1241 addr = __pcpu_ptr_to_addr(ptr);
1242
1243 spin_lock_irqsave(&pcpu_lock, flags);
1244
1245 chunk = pcpu_chunk_addr_search(addr);
1246 off = addr - chunk->base_addr;
1247
1248 pcpu_free_area(chunk, off, &occ_pages);
1249
1250 if (chunk != pcpu_reserved_chunk)
1251 pcpu_nr_empty_pop_pages += occ_pages;
1252
1253 /* if there are more than one fully free chunks, wake up grim reaper */
1254 if (chunk->free_size == pcpu_unit_size) {
1255 struct pcpu_chunk *pos;
1256
1257 list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1258 if (pos != chunk) {
1259 pcpu_schedule_balance_work();
1260 break;
1261 }
1262 }
1263
1264 spin_unlock_irqrestore(&pcpu_lock, flags);
1265}
1266EXPORT_SYMBOL_GPL(free_percpu);
1267
1268/**
1269 * is_kernel_percpu_address - test whether address is from static percpu area
1270 * @addr: address to test
1271 *
1272 * Test whether @addr belongs to in-kernel static percpu area. Module
1273 * static percpu areas are not considered. For those, use
1274 * is_module_percpu_address().
1275 *
1276 * RETURNS:
1277 * %true if @addr is from in-kernel static percpu area, %false otherwise.
1278 */
1279bool is_kernel_percpu_address(unsigned long addr)
1280{
1281#ifdef CONFIG_SMP
1282 const size_t static_size = __per_cpu_end - __per_cpu_start;
1283 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1284 unsigned int cpu;
1285
1286 for_each_possible_cpu(cpu) {
1287 void *start = per_cpu_ptr(base, cpu);
1288
1289 if ((void *)addr >= start && (void *)addr < start + static_size)
1290 return true;
1291 }
1292#endif
1293 /* on UP, can't distinguish from other static vars, always false */
1294 return false;
1295}
1296
1297/**
1298 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
1299 * @addr: the address to be converted to physical address
1300 *
1301 * Given @addr which is dereferenceable address obtained via one of
1302 * percpu access macros, this function translates it into its physical
1303 * address. The caller is responsible for ensuring @addr stays valid
1304 * until this function finishes.
1305 *
1306 * percpu allocator has special setup for the first chunk, which currently
1307 * supports either embedding in linear address space or vmalloc mapping,
1308 * and, from the second one, the backing allocator (currently either vm or
1309 * km) provides translation.
1310 *
1311 * The addr can be translated simply without checking if it falls into the
1312 * first chunk. But the current code reflects better how percpu allocator
1313 * actually works, and the verification can discover both bugs in percpu
1314 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
1315 * code.
1316 *
1317 * RETURNS:
1318 * The physical address for @addr.
1319 */
1320phys_addr_t per_cpu_ptr_to_phys(void *addr)
1321{
1322 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1323 bool in_first_chunk = false;
1324 unsigned long first_low, first_high;
1325 unsigned int cpu;
1326
1327 /*
1328 * The following test on unit_low/high isn't strictly
1329 * necessary but will speed up lookups of addresses which
1330 * aren't in the first chunk.
1331 */
1332 first_low = pcpu_chunk_addr(pcpu_first_chunk, pcpu_low_unit_cpu, 0);
1333 first_high = pcpu_chunk_addr(pcpu_first_chunk, pcpu_high_unit_cpu,
1334 pcpu_unit_pages);
1335 if ((unsigned long)addr >= first_low &&
1336 (unsigned long)addr < first_high) {
1337 for_each_possible_cpu(cpu) {
1338 void *start = per_cpu_ptr(base, cpu);
1339
1340 if (addr >= start && addr < start + pcpu_unit_size) {
1341 in_first_chunk = true;
1342 break;
1343 }
1344 }
1345 }
1346
1347 if (in_first_chunk) {
1348 if (!is_vmalloc_addr(addr))
1349 return __pa(addr);
1350 else
1351 return page_to_phys(vmalloc_to_page(addr)) +
1352 offset_in_page(addr);
1353 } else
1354 return page_to_phys(pcpu_addr_to_page(addr)) +
1355 offset_in_page(addr);
1356}
1357
1358/**
1359 * pcpu_alloc_alloc_info - allocate percpu allocation info
1360 * @nr_groups: the number of groups
1361 * @nr_units: the number of units
1362 *
1363 * Allocate ai which is large enough for @nr_groups groups containing
1364 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1365 * cpu_map array which is long enough for @nr_units and filled with
1366 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1367 * pointer of other groups.
1368 *
1369 * RETURNS:
1370 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1371 * failure.
1372 */
1373struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1374 int nr_units)
1375{
1376 struct pcpu_alloc_info *ai;
1377 size_t base_size, ai_size;
1378 void *ptr;
1379 int unit;
1380
1381 base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1382 __alignof__(ai->groups[0].cpu_map[0]));
1383 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1384
1385 ptr = memblock_virt_alloc_nopanic(PFN_ALIGN(ai_size), 0);
1386 if (!ptr)
1387 return NULL;
1388 ai = ptr;
1389 ptr += base_size;
1390
1391 ai->groups[0].cpu_map = ptr;
1392
1393 for (unit = 0; unit < nr_units; unit++)
1394 ai->groups[0].cpu_map[unit] = NR_CPUS;
1395
1396 ai->nr_groups = nr_groups;
1397 ai->__ai_size = PFN_ALIGN(ai_size);
1398
1399 return ai;
1400}
1401
1402/**
1403 * pcpu_free_alloc_info - free percpu allocation info
1404 * @ai: pcpu_alloc_info to free
1405 *
1406 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1407 */
1408void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1409{
1410 memblock_free_early(__pa(ai), ai->__ai_size);
1411}
1412
1413/**
1414 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1415 * @lvl: loglevel
1416 * @ai: allocation info to dump
1417 *
1418 * Print out information about @ai using loglevel @lvl.
1419 */
1420static void pcpu_dump_alloc_info(const char *lvl,
1421 const struct pcpu_alloc_info *ai)
1422{
1423 int group_width = 1, cpu_width = 1, width;
1424 char empty_str[] = "--------";
1425 int alloc = 0, alloc_end = 0;
1426 int group, v;
1427 int upa, apl; /* units per alloc, allocs per line */
1428
1429 v = ai->nr_groups;
1430 while (v /= 10)
1431 group_width++;
1432
1433 v = num_possible_cpus();
1434 while (v /= 10)
1435 cpu_width++;
1436 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1437
1438 upa = ai->alloc_size / ai->unit_size;
1439 width = upa * (cpu_width + 1) + group_width + 3;
1440 apl = rounddown_pow_of_two(max(60 / width, 1));
1441
1442 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1443 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1444 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1445
1446 for (group = 0; group < ai->nr_groups; group++) {
1447 const struct pcpu_group_info *gi = &ai->groups[group];
1448 int unit = 0, unit_end = 0;
1449
1450 BUG_ON(gi->nr_units % upa);
1451 for (alloc_end += gi->nr_units / upa;
1452 alloc < alloc_end; alloc++) {
1453 if (!(alloc % apl)) {
1454 pr_cont("\n");
1455 printk("%spcpu-alloc: ", lvl);
1456 }
1457 pr_cont("[%0*d] ", group_width, group);
1458
1459 for (unit_end += upa; unit < unit_end; unit++)
1460 if (gi->cpu_map[unit] != NR_CPUS)
1461 pr_cont("%0*d ",
1462 cpu_width, gi->cpu_map[unit]);
1463 else
1464 pr_cont("%s ", empty_str);
1465 }
1466 }
1467 pr_cont("\n");
1468}
1469
1470/**
1471 * pcpu_setup_first_chunk - initialize the first percpu chunk
1472 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1473 * @base_addr: mapped address
1474 *
1475 * Initialize the first percpu chunk which contains the kernel static
1476 * perpcu area. This function is to be called from arch percpu area
1477 * setup path.
1478 *
1479 * @ai contains all information necessary to initialize the first
1480 * chunk and prime the dynamic percpu allocator.
1481 *
1482 * @ai->static_size is the size of static percpu area.
1483 *
1484 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1485 * reserve after the static area in the first chunk. This reserves
1486 * the first chunk such that it's available only through reserved
1487 * percpu allocation. This is primarily used to serve module percpu
1488 * static areas on architectures where the addressing model has
1489 * limited offset range for symbol relocations to guarantee module
1490 * percpu symbols fall inside the relocatable range.
1491 *
1492 * @ai->dyn_size determines the number of bytes available for dynamic
1493 * allocation in the first chunk. The area between @ai->static_size +
1494 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1495 *
1496 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1497 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1498 * @ai->dyn_size.
1499 *
1500 * @ai->atom_size is the allocation atom size and used as alignment
1501 * for vm areas.
1502 *
1503 * @ai->alloc_size is the allocation size and always multiple of
1504 * @ai->atom_size. This is larger than @ai->atom_size if
1505 * @ai->unit_size is larger than @ai->atom_size.
1506 *
1507 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1508 * percpu areas. Units which should be colocated are put into the
1509 * same group. Dynamic VM areas will be allocated according to these
1510 * groupings. If @ai->nr_groups is zero, a single group containing
1511 * all units is assumed.
1512 *
1513 * The caller should have mapped the first chunk at @base_addr and
1514 * copied static data to each unit.
1515 *
1516 * If the first chunk ends up with both reserved and dynamic areas, it
1517 * is served by two chunks - one to serve the core static and reserved
1518 * areas and the other for the dynamic area. They share the same vm
1519 * and page map but uses different area allocation map to stay away
1520 * from each other. The latter chunk is circulated in the chunk slots
1521 * and available for dynamic allocation like any other chunks.
1522 *
1523 * RETURNS:
1524 * 0 on success, -errno on failure.
1525 */
1526int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1527 void *base_addr)
1528{
1529 static int smap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1530 static int dmap[PERCPU_DYNAMIC_EARLY_SLOTS] __initdata;
1531 size_t dyn_size = ai->dyn_size;
1532 size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1533 struct pcpu_chunk *schunk, *dchunk = NULL;
1534 unsigned long *group_offsets;
1535 size_t *group_sizes;
1536 unsigned long *unit_off;
1537 unsigned int cpu;
1538 int *unit_map;
1539 int group, unit, i;
1540
1541#define PCPU_SETUP_BUG_ON(cond) do { \
1542 if (unlikely(cond)) { \
1543 pr_emerg("failed to initialize, %s\n", #cond); \
1544 pr_emerg("cpu_possible_mask=%*pb\n", \
1545 cpumask_pr_args(cpu_possible_mask)); \
1546 pcpu_dump_alloc_info(KERN_EMERG, ai); \
1547 BUG(); \
1548 } \
1549} while (0)
1550
1551 /* sanity checks */
1552 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1553#ifdef CONFIG_SMP
1554 PCPU_SETUP_BUG_ON(!ai->static_size);
1555 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
1556#endif
1557 PCPU_SETUP_BUG_ON(!base_addr);
1558 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
1559 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1560 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
1561 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1562 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
1563 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1564
1565 /* process group information and build config tables accordingly */
1566 group_offsets = memblock_virt_alloc(ai->nr_groups *
1567 sizeof(group_offsets[0]), 0);
1568 group_sizes = memblock_virt_alloc(ai->nr_groups *
1569 sizeof(group_sizes[0]), 0);
1570 unit_map = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_map[0]), 0);
1571 unit_off = memblock_virt_alloc(nr_cpu_ids * sizeof(unit_off[0]), 0);
1572
1573 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1574 unit_map[cpu] = UINT_MAX;
1575
1576 pcpu_low_unit_cpu = NR_CPUS;
1577 pcpu_high_unit_cpu = NR_CPUS;
1578
1579 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1580 const struct pcpu_group_info *gi = &ai->groups[group];
1581
1582 group_offsets[group] = gi->base_offset;
1583 group_sizes[group] = gi->nr_units * ai->unit_size;
1584
1585 for (i = 0; i < gi->nr_units; i++) {
1586 cpu = gi->cpu_map[i];
1587 if (cpu == NR_CPUS)
1588 continue;
1589
1590 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
1591 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1592 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1593
1594 unit_map[cpu] = unit + i;
1595 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1596
1597 /* determine low/high unit_cpu */
1598 if (pcpu_low_unit_cpu == NR_CPUS ||
1599 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
1600 pcpu_low_unit_cpu = cpu;
1601 if (pcpu_high_unit_cpu == NR_CPUS ||
1602 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
1603 pcpu_high_unit_cpu = cpu;
1604 }
1605 }
1606 pcpu_nr_units = unit;
1607
1608 for_each_possible_cpu(cpu)
1609 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1610
1611 /* we're done parsing the input, undefine BUG macro and dump config */
1612#undef PCPU_SETUP_BUG_ON
1613 pcpu_dump_alloc_info(KERN_DEBUG, ai);
1614
1615 pcpu_nr_groups = ai->nr_groups;
1616 pcpu_group_offsets = group_offsets;
1617 pcpu_group_sizes = group_sizes;
1618 pcpu_unit_map = unit_map;
1619 pcpu_unit_offsets = unit_off;
1620
1621 /* determine basic parameters */
1622 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1623 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1624 pcpu_atom_size = ai->atom_size;
1625 pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1626 BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1627
1628 /*
1629 * Allocate chunk slots. The additional last slot is for
1630 * empty chunks.
1631 */
1632 pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1633 pcpu_slot = memblock_virt_alloc(
1634 pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
1635 for (i = 0; i < pcpu_nr_slots; i++)
1636 INIT_LIST_HEAD(&pcpu_slot[i]);
1637
1638 /*
1639 * Initialize static chunk. If reserved_size is zero, the
1640 * static chunk covers static area + dynamic allocation area
1641 * in the first chunk. If reserved_size is not zero, it
1642 * covers static area + reserved area (mostly used for module
1643 * static percpu allocation).
1644 */
1645 schunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1646 INIT_LIST_HEAD(&schunk->list);
1647 INIT_WORK(&schunk->map_extend_work, pcpu_map_extend_workfn);
1648 schunk->base_addr = base_addr;
1649 schunk->map = smap;
1650 schunk->map_alloc = ARRAY_SIZE(smap);
1651 schunk->immutable = true;
1652 bitmap_fill(schunk->populated, pcpu_unit_pages);
1653 schunk->nr_populated = pcpu_unit_pages;
1654
1655 if (ai->reserved_size) {
1656 schunk->free_size = ai->reserved_size;
1657 pcpu_reserved_chunk = schunk;
1658 pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1659 } else {
1660 schunk->free_size = dyn_size;
1661 dyn_size = 0; /* dynamic area covered */
1662 }
1663 schunk->contig_hint = schunk->free_size;
1664
1665 schunk->map[0] = 1;
1666 schunk->map[1] = ai->static_size;
1667 schunk->map_used = 1;
1668 if (schunk->free_size)
1669 schunk->map[++schunk->map_used] = ai->static_size + schunk->free_size;
1670 schunk->map[schunk->map_used] |= 1;
1671
1672 /* init dynamic chunk if necessary */
1673 if (dyn_size) {
1674 dchunk = memblock_virt_alloc(pcpu_chunk_struct_size, 0);
1675 INIT_LIST_HEAD(&dchunk->list);
1676 INIT_WORK(&dchunk->map_extend_work, pcpu_map_extend_workfn);
1677 dchunk->base_addr = base_addr;
1678 dchunk->map = dmap;
1679 dchunk->map_alloc = ARRAY_SIZE(dmap);
1680 dchunk->immutable = true;
1681 bitmap_fill(dchunk->populated, pcpu_unit_pages);
1682 dchunk->nr_populated = pcpu_unit_pages;
1683
1684 dchunk->contig_hint = dchunk->free_size = dyn_size;
1685 dchunk->map[0] = 1;
1686 dchunk->map[1] = pcpu_reserved_chunk_limit;
1687 dchunk->map[2] = (pcpu_reserved_chunk_limit + dchunk->free_size) | 1;
1688 dchunk->map_used = 2;
1689 }
1690
1691 /* link the first chunk in */
1692 pcpu_first_chunk = dchunk ?: schunk;
1693 pcpu_nr_empty_pop_pages +=
1694 pcpu_count_occupied_pages(pcpu_first_chunk, 1);
1695 pcpu_chunk_relocate(pcpu_first_chunk, -1);
1696
1697 /* we're done */
1698 pcpu_base_addr = base_addr;
1699 return 0;
1700}
1701
1702#ifdef CONFIG_SMP
1703
1704const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
1705 [PCPU_FC_AUTO] = "auto",
1706 [PCPU_FC_EMBED] = "embed",
1707 [PCPU_FC_PAGE] = "page",
1708};
1709
1710enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1711
1712static int __init percpu_alloc_setup(char *str)
1713{
1714 if (!str)
1715 return -EINVAL;
1716
1717 if (0)
1718 /* nada */;
1719#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1720 else if (!strcmp(str, "embed"))
1721 pcpu_chosen_fc = PCPU_FC_EMBED;
1722#endif
1723#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1724 else if (!strcmp(str, "page"))
1725 pcpu_chosen_fc = PCPU_FC_PAGE;
1726#endif
1727 else
1728 pr_warn("unknown allocator %s specified\n", str);
1729
1730 return 0;
1731}
1732early_param("percpu_alloc", percpu_alloc_setup);
1733
1734/*
1735 * pcpu_embed_first_chunk() is used by the generic percpu setup.
1736 * Build it if needed by the arch config or the generic setup is going
1737 * to be used.
1738 */
1739#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1740 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1741#define BUILD_EMBED_FIRST_CHUNK
1742#endif
1743
1744/* build pcpu_page_first_chunk() iff needed by the arch config */
1745#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
1746#define BUILD_PAGE_FIRST_CHUNK
1747#endif
1748
1749/* pcpu_build_alloc_info() is used by both embed and page first chunk */
1750#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
1751/**
1752 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1753 * @reserved_size: the size of reserved percpu area in bytes
1754 * @dyn_size: minimum free size for dynamic allocation in bytes
1755 * @atom_size: allocation atom size
1756 * @cpu_distance_fn: callback to determine distance between cpus, optional
1757 *
1758 * This function determines grouping of units, their mappings to cpus
1759 * and other parameters considering needed percpu size, allocation
1760 * atom size and distances between CPUs.
1761 *
1762 * Groups are always multiples of atom size and CPUs which are of
1763 * LOCAL_DISTANCE both ways are grouped together and share space for
1764 * units in the same group. The returned configuration is guaranteed
1765 * to have CPUs on different nodes on different groups and >=75% usage
1766 * of allocated virtual address space.
1767 *
1768 * RETURNS:
1769 * On success, pointer to the new allocation_info is returned. On
1770 * failure, ERR_PTR value is returned.
1771 */
1772static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1773 size_t reserved_size, size_t dyn_size,
1774 size_t atom_size,
1775 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1776{
1777 static int group_map[NR_CPUS] __initdata;
1778 static int group_cnt[NR_CPUS] __initdata;
1779 const size_t static_size = __per_cpu_end - __per_cpu_start;
1780 int nr_groups = 1, nr_units = 0;
1781 size_t size_sum, min_unit_size, alloc_size;
1782 int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1783 int last_allocs, group, unit;
1784 unsigned int cpu, tcpu;
1785 struct pcpu_alloc_info *ai;
1786 unsigned int *cpu_map;
1787
1788 /* this function may be called multiple times */
1789 memset(group_map, 0, sizeof(group_map));
1790 memset(group_cnt, 0, sizeof(group_cnt));
1791
1792 /* calculate size_sum and ensure dyn_size is enough for early alloc */
1793 size_sum = PFN_ALIGN(static_size + reserved_size +
1794 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
1795 dyn_size = size_sum - static_size - reserved_size;
1796
1797 /*
1798 * Determine min_unit_size, alloc_size and max_upa such that
1799 * alloc_size is multiple of atom_size and is the smallest
1800 * which can accommodate 4k aligned segments which are equal to
1801 * or larger than min_unit_size.
1802 */
1803 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1804
1805 alloc_size = roundup(min_unit_size, atom_size);
1806 upa = alloc_size / min_unit_size;
1807 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1808 upa--;
1809 max_upa = upa;
1810
1811 /* group cpus according to their proximity */
1812 for_each_possible_cpu(cpu) {
1813 group = 0;
1814 next_group:
1815 for_each_possible_cpu(tcpu) {
1816 if (cpu == tcpu)
1817 break;
1818 if (group_map[tcpu] == group && cpu_distance_fn &&
1819 (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1820 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1821 group++;
1822 nr_groups = max(nr_groups, group + 1);
1823 goto next_group;
1824 }
1825 }
1826 group_map[cpu] = group;
1827 group_cnt[group]++;
1828 }
1829
1830 /*
1831 * Expand unit size until address space usage goes over 75%
1832 * and then as much as possible without using more address
1833 * space.
1834 */
1835 last_allocs = INT_MAX;
1836 for (upa = max_upa; upa; upa--) {
1837 int allocs = 0, wasted = 0;
1838
1839 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
1840 continue;
1841
1842 for (group = 0; group < nr_groups; group++) {
1843 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1844 allocs += this_allocs;
1845 wasted += this_allocs * upa - group_cnt[group];
1846 }
1847
1848 /*
1849 * Don't accept if wastage is over 1/3. The
1850 * greater-than comparison ensures upa==1 always
1851 * passes the following check.
1852 */
1853 if (wasted > num_possible_cpus() / 3)
1854 continue;
1855
1856 /* and then don't consume more memory */
1857 if (allocs > last_allocs)
1858 break;
1859 last_allocs = allocs;
1860 best_upa = upa;
1861 }
1862 upa = best_upa;
1863
1864 /* allocate and fill alloc_info */
1865 for (group = 0; group < nr_groups; group++)
1866 nr_units += roundup(group_cnt[group], upa);
1867
1868 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1869 if (!ai)
1870 return ERR_PTR(-ENOMEM);
1871 cpu_map = ai->groups[0].cpu_map;
1872
1873 for (group = 0; group < nr_groups; group++) {
1874 ai->groups[group].cpu_map = cpu_map;
1875 cpu_map += roundup(group_cnt[group], upa);
1876 }
1877
1878 ai->static_size = static_size;
1879 ai->reserved_size = reserved_size;
1880 ai->dyn_size = dyn_size;
1881 ai->unit_size = alloc_size / upa;
1882 ai->atom_size = atom_size;
1883 ai->alloc_size = alloc_size;
1884
1885 for (group = 0, unit = 0; group_cnt[group]; group++) {
1886 struct pcpu_group_info *gi = &ai->groups[group];
1887
1888 /*
1889 * Initialize base_offset as if all groups are located
1890 * back-to-back. The caller should update this to
1891 * reflect actual allocation.
1892 */
1893 gi->base_offset = unit * ai->unit_size;
1894
1895 for_each_possible_cpu(cpu)
1896 if (group_map[cpu] == group)
1897 gi->cpu_map[gi->nr_units++] = cpu;
1898 gi->nr_units = roundup(gi->nr_units, upa);
1899 unit += gi->nr_units;
1900 }
1901 BUG_ON(unit != nr_units);
1902
1903 return ai;
1904}
1905#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
1906
1907#if defined(BUILD_EMBED_FIRST_CHUNK)
1908/**
1909 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1910 * @reserved_size: the size of reserved percpu area in bytes
1911 * @dyn_size: minimum free size for dynamic allocation in bytes
1912 * @atom_size: allocation atom size
1913 * @cpu_distance_fn: callback to determine distance between cpus, optional
1914 * @alloc_fn: function to allocate percpu page
1915 * @free_fn: function to free percpu page
1916 *
1917 * This is a helper to ease setting up embedded first percpu chunk and
1918 * can be called where pcpu_setup_first_chunk() is expected.
1919 *
1920 * If this function is used to setup the first chunk, it is allocated
1921 * by calling @alloc_fn and used as-is without being mapped into
1922 * vmalloc area. Allocations are always whole multiples of @atom_size
1923 * aligned to @atom_size.
1924 *
1925 * This enables the first chunk to piggy back on the linear physical
1926 * mapping which often uses larger page size. Please note that this
1927 * can result in very sparse cpu->unit mapping on NUMA machines thus
1928 * requiring large vmalloc address space. Don't use this allocator if
1929 * vmalloc space is not orders of magnitude larger than distances
1930 * between node memory addresses (ie. 32bit NUMA machines).
1931 *
1932 * @dyn_size specifies the minimum dynamic area size.
1933 *
1934 * If the needed size is smaller than the minimum or specified unit
1935 * size, the leftover is returned using @free_fn.
1936 *
1937 * RETURNS:
1938 * 0 on success, -errno on failure.
1939 */
1940int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
1941 size_t atom_size,
1942 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1943 pcpu_fc_alloc_fn_t alloc_fn,
1944 pcpu_fc_free_fn_t free_fn)
1945{
1946 void *base = (void *)ULONG_MAX;
1947 void **areas = NULL;
1948 struct pcpu_alloc_info *ai;
1949 size_t size_sum, areas_size, max_distance;
1950 int group, i, rc;
1951
1952 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1953 cpu_distance_fn);
1954 if (IS_ERR(ai))
1955 return PTR_ERR(ai);
1956
1957 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1958 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1959
1960 areas = memblock_virt_alloc_nopanic(areas_size, 0);
1961 if (!areas) {
1962 rc = -ENOMEM;
1963 goto out_free;
1964 }
1965
1966 /* allocate, copy and determine base address */
1967 for (group = 0; group < ai->nr_groups; group++) {
1968 struct pcpu_group_info *gi = &ai->groups[group];
1969 unsigned int cpu = NR_CPUS;
1970 void *ptr;
1971
1972 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1973 cpu = gi->cpu_map[i];
1974 BUG_ON(cpu == NR_CPUS);
1975
1976 /* allocate space for the whole group */
1977 ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1978 if (!ptr) {
1979 rc = -ENOMEM;
1980 goto out_free_areas;
1981 }
1982 /* kmemleak tracks the percpu allocations separately */
1983 kmemleak_free(ptr);
1984 areas[group] = ptr;
1985
1986 base = min(ptr, base);
1987 }
1988
1989 /*
1990 * Copy data and free unused parts. This should happen after all
1991 * allocations are complete; otherwise, we may end up with
1992 * overlapping groups.
1993 */
1994 for (group = 0; group < ai->nr_groups; group++) {
1995 struct pcpu_group_info *gi = &ai->groups[group];
1996 void *ptr = areas[group];
1997
1998 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1999 if (gi->cpu_map[i] == NR_CPUS) {
2000 /* unused unit, free whole */
2001 free_fn(ptr, ai->unit_size);
2002 continue;
2003 }
2004 /* copy and return the unused part */
2005 memcpy(ptr, __per_cpu_load, ai->static_size);
2006 free_fn(ptr + size_sum, ai->unit_size - size_sum);
2007 }
2008 }
2009
2010 /* base address is now known, determine group base offsets */
2011 max_distance = 0;
2012 for (group = 0; group < ai->nr_groups; group++) {
2013 ai->groups[group].base_offset = areas[group] - base;
2014 max_distance = max_t(size_t, max_distance,
2015 ai->groups[group].base_offset);
2016 }
2017 max_distance += ai->unit_size;
2018
2019 /* warn if maximum distance is further than 75% of vmalloc space */
2020 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
2021 pr_warn("max_distance=0x%zx too large for vmalloc space 0x%lx\n",
2022 max_distance, VMALLOC_TOTAL);
2023#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2024 /* and fail if we have fallback */
2025 rc = -EINVAL;
2026 goto out_free;
2027#endif
2028 }
2029
2030 pr_info("Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
2031 PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
2032 ai->dyn_size, ai->unit_size);
2033
2034 rc = pcpu_setup_first_chunk(ai, base);
2035 goto out_free;
2036
2037out_free_areas:
2038 for (group = 0; group < ai->nr_groups; group++)
2039 if (areas[group])
2040 free_fn(areas[group],
2041 ai->groups[group].nr_units * ai->unit_size);
2042out_free:
2043 pcpu_free_alloc_info(ai);
2044 if (areas)
2045 memblock_free_early(__pa(areas), areas_size);
2046 return rc;
2047}
2048#endif /* BUILD_EMBED_FIRST_CHUNK */
2049
2050#ifdef BUILD_PAGE_FIRST_CHUNK
2051/**
2052 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2053 * @reserved_size: the size of reserved percpu area in bytes
2054 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2055 * @free_fn: function to free percpu page, always called with PAGE_SIZE
2056 * @populate_pte_fn: function to populate pte
2057 *
2058 * This is a helper to ease setting up page-remapped first percpu
2059 * chunk and can be called where pcpu_setup_first_chunk() is expected.
2060 *
2061 * This is the basic allocator. Static percpu area is allocated
2062 * page-by-page into vmalloc area.
2063 *
2064 * RETURNS:
2065 * 0 on success, -errno on failure.
2066 */
2067int __init pcpu_page_first_chunk(size_t reserved_size,
2068 pcpu_fc_alloc_fn_t alloc_fn,
2069 pcpu_fc_free_fn_t free_fn,
2070 pcpu_fc_populate_pte_fn_t populate_pte_fn)
2071{
2072 static struct vm_struct vm;
2073 struct pcpu_alloc_info *ai;
2074 char psize_str[16];
2075 int unit_pages;
2076 size_t pages_size;
2077 struct page **pages;
2078 int unit, i, j, rc;
2079
2080 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2081
2082 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2083 if (IS_ERR(ai))
2084 return PTR_ERR(ai);
2085 BUG_ON(ai->nr_groups != 1);
2086 BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
2087
2088 unit_pages = ai->unit_size >> PAGE_SHIFT;
2089
2090 /* unaligned allocations can't be freed, round up to page size */
2091 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2092 sizeof(pages[0]));
2093 pages = memblock_virt_alloc(pages_size, 0);
2094
2095 /* allocate pages */
2096 j = 0;
2097 for (unit = 0; unit < num_possible_cpus(); unit++)
2098 for (i = 0; i < unit_pages; i++) {
2099 unsigned int cpu = ai->groups[0].cpu_map[unit];
2100 void *ptr;
2101
2102 ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2103 if (!ptr) {
2104 pr_warn("failed to allocate %s page for cpu%u\n",
2105 psize_str, cpu);
2106 goto enomem;
2107 }
2108 /* kmemleak tracks the percpu allocations separately */
2109 kmemleak_free(ptr);
2110 pages[j++] = virt_to_page(ptr);
2111 }
2112
2113 /* allocate vm area, map the pages and copy static data */
2114 vm.flags = VM_ALLOC;
2115 vm.size = num_possible_cpus() * ai->unit_size;
2116 vm_area_register_early(&vm, PAGE_SIZE);
2117
2118 for (unit = 0; unit < num_possible_cpus(); unit++) {
2119 unsigned long unit_addr =
2120 (unsigned long)vm.addr + unit * ai->unit_size;
2121
2122 for (i = 0; i < unit_pages; i++)
2123 populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2124
2125 /* pte already populated, the following shouldn't fail */
2126 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2127 unit_pages);
2128 if (rc < 0)
2129 panic("failed to map percpu area, err=%d\n", rc);
2130
2131 /*
2132 * FIXME: Archs with virtual cache should flush local
2133 * cache for the linear mapping here - something
2134 * equivalent to flush_cache_vmap() on the local cpu.
2135 * flush_cache_vmap() can't be used as most supporting
2136 * data structures are not set up yet.
2137 */
2138
2139 /* copy static data */
2140 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2141 }
2142
2143 /* we're ready, commit */
2144 pr_info("%d %s pages/cpu @%p s%zu r%zu d%zu\n",
2145 unit_pages, psize_str, vm.addr, ai->static_size,
2146 ai->reserved_size, ai->dyn_size);
2147
2148 rc = pcpu_setup_first_chunk(ai, vm.addr);
2149 goto out_free_ar;
2150
2151enomem:
2152 while (--j >= 0)
2153 free_fn(page_address(pages[j]), PAGE_SIZE);
2154 rc = -ENOMEM;
2155out_free_ar:
2156 memblock_free_early(__pa(pages), pages_size);
2157 pcpu_free_alloc_info(ai);
2158 return rc;
2159}
2160#endif /* BUILD_PAGE_FIRST_CHUNK */
2161
2162#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2163/*
2164 * Generic SMP percpu area setup.
2165 *
2166 * The embedding helper is used because its behavior closely resembles
2167 * the original non-dynamic generic percpu area setup. This is
2168 * important because many archs have addressing restrictions and might
2169 * fail if the percpu area is located far away from the previous
2170 * location. As an added bonus, in non-NUMA cases, embedding is
2171 * generally a good idea TLB-wise because percpu area can piggy back
2172 * on the physical linear memory mapping which uses large page
2173 * mappings on applicable archs.
2174 */
2175unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2176EXPORT_SYMBOL(__per_cpu_offset);
2177
2178static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2179 size_t align)
2180{
2181 return memblock_virt_alloc_from_nopanic(
2182 size, align, __pa(MAX_DMA_ADDRESS));
2183}
2184
2185static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2186{
2187 memblock_free_early(__pa(ptr), size);
2188}
2189
2190void __init setup_per_cpu_areas(void)
2191{
2192 unsigned long delta;
2193 unsigned int cpu;
2194 int rc;
2195
2196 /*
2197 * Always reserve area for module percpu variables. That's
2198 * what the legacy allocator did.
2199 */
2200 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2201 PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2202 pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2203 if (rc < 0)
2204 panic("Failed to initialize percpu areas.");
2205
2206 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2207 for_each_possible_cpu(cpu)
2208 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2209}
2210#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2211
2212#else /* CONFIG_SMP */
2213
2214/*
2215 * UP percpu area setup.
2216 *
2217 * UP always uses km-based percpu allocator with identity mapping.
2218 * Static percpu variables are indistinguishable from the usual static
2219 * variables and don't require any special preparation.
2220 */
2221void __init setup_per_cpu_areas(void)
2222{
2223 const size_t unit_size =
2224 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2225 PERCPU_DYNAMIC_RESERVE));
2226 struct pcpu_alloc_info *ai;
2227 void *fc;
2228
2229 ai = pcpu_alloc_alloc_info(1, 1);
2230 fc = memblock_virt_alloc_from_nopanic(unit_size,
2231 PAGE_SIZE,
2232 __pa(MAX_DMA_ADDRESS));
2233 if (!ai || !fc)
2234 panic("Failed to allocate memory for percpu areas.");
2235 /* kmemleak tracks the percpu allocations separately */
2236 kmemleak_free(fc);
2237
2238 ai->dyn_size = unit_size;
2239 ai->unit_size = unit_size;
2240 ai->atom_size = unit_size;
2241 ai->alloc_size = unit_size;
2242 ai->groups[0].nr_units = 1;
2243 ai->groups[0].cpu_map[0] = 0;
2244
2245 if (pcpu_setup_first_chunk(ai, fc) < 0)
2246 panic("Failed to initialize percpu areas.");
2247}
2248
2249#endif /* CONFIG_SMP */
2250
2251/*
2252 * First and reserved chunks are initialized with temporary allocation
2253 * map in initdata so that they can be used before slab is online.
2254 * This function is called after slab is brought up and replaces those
2255 * with properly allocated maps.
2256 */
2257void __init percpu_init_late(void)
2258{
2259 struct pcpu_chunk *target_chunks[] =
2260 { pcpu_first_chunk, pcpu_reserved_chunk, NULL };
2261 struct pcpu_chunk *chunk;
2262 unsigned long flags;
2263 int i;
2264
2265 for (i = 0; (chunk = target_chunks[i]); i++) {
2266 int *map;
2267 const size_t size = PERCPU_DYNAMIC_EARLY_SLOTS * sizeof(map[0]);
2268
2269 BUILD_BUG_ON(size > PAGE_SIZE);
2270
2271 map = pcpu_mem_zalloc(size);
2272 BUG_ON(!map);
2273
2274 spin_lock_irqsave(&pcpu_lock, flags);
2275 memcpy(map, chunk->map, size);
2276 chunk->map = map;
2277 spin_unlock_irqrestore(&pcpu_lock, flags);
2278 }
2279}
2280
2281/*
2282 * Percpu allocator is initialized early during boot when neither slab or
2283 * workqueue is available. Plug async management until everything is up
2284 * and running.
2285 */
2286static int __init percpu_enable_async(void)
2287{
2288 pcpu_async_enabled = true;
2289 return 0;
2290}
2291subsys_initcall(percpu_enable_async);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/percpu.c - percpu memory allocator
4 *
5 * Copyright (C) 2009 SUSE Linux Products GmbH
6 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
7 *
8 * Copyright (C) 2017 Facebook Inc.
9 * Copyright (C) 2017 Dennis Zhou <dennis@kernel.org>
10 *
11 * The percpu allocator handles both static and dynamic areas. Percpu
12 * areas are allocated in chunks which are divided into units. There is
13 * a 1-to-1 mapping for units to possible cpus. These units are grouped
14 * based on NUMA properties of the machine.
15 *
16 * c0 c1 c2
17 * ------------------- ------------------- ------------
18 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19 * ------------------- ...... ------------------- .... ------------
20 *
21 * Allocation is done by offsets into a unit's address space. Ie., an
22 * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23 * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24 * and even sparse. Access is handled by configuring percpu base
25 * registers according to the cpu to unit mappings and offsetting the
26 * base address using pcpu_unit_size.
27 *
28 * There is special consideration for the first chunk which must handle
29 * the static percpu variables in the kernel image as allocation services
30 * are not online yet. In short, the first chunk is structured like so:
31 *
32 * <Static | [Reserved] | Dynamic>
33 *
34 * The static data is copied from the original section managed by the
35 * linker. The reserved section, if non-zero, primarily manages static
36 * percpu variables from kernel modules. Finally, the dynamic section
37 * takes care of normal allocations.
38 *
39 * The allocator organizes chunks into lists according to free size and
40 * memcg-awareness. To make a percpu allocation memcg-aware the __GFP_ACCOUNT
41 * flag should be passed. All memcg-aware allocations are sharing one set
42 * of chunks and all unaccounted allocations and allocations performed
43 * by processes belonging to the root memory cgroup are using the second set.
44 *
45 * The allocator tries to allocate from the fullest chunk first. Each chunk
46 * is managed by a bitmap with metadata blocks. The allocation map is updated
47 * on every allocation and free to reflect the current state while the boundary
48 * map is only updated on allocation. Each metadata block contains
49 * information to help mitigate the need to iterate over large portions
50 * of the bitmap. The reverse mapping from page to chunk is stored in
51 * the page's index. Lastly, units are lazily backed and grow in unison.
52 *
53 * There is a unique conversion that goes on here between bytes and bits.
54 * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
55 * tracks the number of pages it is responsible for in nr_pages. Helper
56 * functions are used to convert from between the bytes, bits, and blocks.
57 * All hints are managed in bits unless explicitly stated.
58 *
59 * To use this allocator, arch code should do the following:
60 *
61 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
62 * regular address to percpu pointer and back if they need to be
63 * different from the default
64 *
65 * - use pcpu_setup_first_chunk() during percpu area initialization to
66 * setup the first chunk containing the kernel static percpu area
67 */
68
69#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
70
71#include <linux/bitmap.h>
72#include <linux/cpumask.h>
73#include <linux/memblock.h>
74#include <linux/err.h>
75#include <linux/list.h>
76#include <linux/log2.h>
77#include <linux/mm.h>
78#include <linux/module.h>
79#include <linux/mutex.h>
80#include <linux/percpu.h>
81#include <linux/pfn.h>
82#include <linux/slab.h>
83#include <linux/spinlock.h>
84#include <linux/vmalloc.h>
85#include <linux/workqueue.h>
86#include <linux/kmemleak.h>
87#include <linux/sched.h>
88#include <linux/sched/mm.h>
89#include <linux/memcontrol.h>
90
91#include <asm/cacheflush.h>
92#include <asm/sections.h>
93#include <asm/tlbflush.h>
94#include <asm/io.h>
95
96#define CREATE_TRACE_POINTS
97#include <trace/events/percpu.h>
98
99#include "percpu-internal.h"
100
101/*
102 * The slots are sorted by the size of the biggest continuous free area.
103 * 1-31 bytes share the same slot.
104 */
105#define PCPU_SLOT_BASE_SHIFT 5
106/* chunks in slots below this are subject to being sidelined on failed alloc */
107#define PCPU_SLOT_FAIL_THRESHOLD 3
108
109#define PCPU_EMPTY_POP_PAGES_LOW 2
110#define PCPU_EMPTY_POP_PAGES_HIGH 4
111
112#ifdef CONFIG_SMP
113/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
114#ifndef __addr_to_pcpu_ptr
115#define __addr_to_pcpu_ptr(addr) \
116 (void __percpu *)((unsigned long)(addr) - \
117 (unsigned long)pcpu_base_addr + \
118 (unsigned long)__per_cpu_start)
119#endif
120#ifndef __pcpu_ptr_to_addr
121#define __pcpu_ptr_to_addr(ptr) \
122 (void __force *)((unsigned long)(ptr) + \
123 (unsigned long)pcpu_base_addr - \
124 (unsigned long)__per_cpu_start)
125#endif
126#else /* CONFIG_SMP */
127/* on UP, it's always identity mapped */
128#define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
129#define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
130#endif /* CONFIG_SMP */
131
132static int pcpu_unit_pages __ro_after_init;
133static int pcpu_unit_size __ro_after_init;
134static int pcpu_nr_units __ro_after_init;
135static int pcpu_atom_size __ro_after_init;
136int pcpu_nr_slots __ro_after_init;
137static int pcpu_free_slot __ro_after_init;
138int pcpu_sidelined_slot __ro_after_init;
139int pcpu_to_depopulate_slot __ro_after_init;
140static size_t pcpu_chunk_struct_size __ro_after_init;
141
142/* cpus with the lowest and highest unit addresses */
143static unsigned int pcpu_low_unit_cpu __ro_after_init;
144static unsigned int pcpu_high_unit_cpu __ro_after_init;
145
146/* the address of the first chunk which starts with the kernel static area */
147void *pcpu_base_addr __ro_after_init;
148
149static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
150const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
151
152/* group information, used for vm allocation */
153static int pcpu_nr_groups __ro_after_init;
154static const unsigned long *pcpu_group_offsets __ro_after_init;
155static const size_t *pcpu_group_sizes __ro_after_init;
156
157/*
158 * The first chunk which always exists. Note that unlike other
159 * chunks, this one can be allocated and mapped in several different
160 * ways and thus often doesn't live in the vmalloc area.
161 */
162struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
163
164/*
165 * Optional reserved chunk. This chunk reserves part of the first
166 * chunk and serves it for reserved allocations. When the reserved
167 * region doesn't exist, the following variable is NULL.
168 */
169struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
170
171DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
172static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
173
174struct list_head *pcpu_chunk_lists __ro_after_init; /* chunk list slots */
175
176/*
177 * The number of empty populated pages, protected by pcpu_lock.
178 * The reserved chunk doesn't contribute to the count.
179 */
180int pcpu_nr_empty_pop_pages;
181
182/*
183 * The number of populated pages in use by the allocator, protected by
184 * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
185 * allocated/deallocated, it is allocated/deallocated in all units of a chunk
186 * and increments/decrements this count by 1).
187 */
188static unsigned long pcpu_nr_populated;
189
190/*
191 * Balance work is used to populate or destroy chunks asynchronously. We
192 * try to keep the number of populated free pages between
193 * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
194 * empty chunk.
195 */
196static void pcpu_balance_workfn(struct work_struct *work);
197static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
198static bool pcpu_async_enabled __read_mostly;
199static bool pcpu_atomic_alloc_failed;
200
201static void pcpu_schedule_balance_work(void)
202{
203 if (pcpu_async_enabled)
204 schedule_work(&pcpu_balance_work);
205}
206
207/**
208 * pcpu_addr_in_chunk - check if the address is served from this chunk
209 * @chunk: chunk of interest
210 * @addr: percpu address
211 *
212 * RETURNS:
213 * True if the address is served from this chunk.
214 */
215static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
216{
217 void *start_addr, *end_addr;
218
219 if (!chunk)
220 return false;
221
222 start_addr = chunk->base_addr + chunk->start_offset;
223 end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
224 chunk->end_offset;
225
226 return addr >= start_addr && addr < end_addr;
227}
228
229static int __pcpu_size_to_slot(int size)
230{
231 int highbit = fls(size); /* size is in bytes */
232 return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
233}
234
235static int pcpu_size_to_slot(int size)
236{
237 if (size == pcpu_unit_size)
238 return pcpu_free_slot;
239 return __pcpu_size_to_slot(size);
240}
241
242static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
243{
244 const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
245
246 if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
247 chunk_md->contig_hint == 0)
248 return 0;
249
250 return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
251}
252
253/* set the pointer to a chunk in a page struct */
254static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
255{
256 page->index = (unsigned long)pcpu;
257}
258
259/* obtain pointer to a chunk from a page struct */
260static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
261{
262 return (struct pcpu_chunk *)page->index;
263}
264
265static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
266{
267 return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
268}
269
270static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
271{
272 return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
273}
274
275static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
276 unsigned int cpu, int page_idx)
277{
278 return (unsigned long)chunk->base_addr +
279 pcpu_unit_page_offset(cpu, page_idx);
280}
281
282/*
283 * The following are helper functions to help access bitmaps and convert
284 * between bitmap offsets to address offsets.
285 */
286static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
287{
288 return chunk->alloc_map +
289 (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
290}
291
292static unsigned long pcpu_off_to_block_index(int off)
293{
294 return off / PCPU_BITMAP_BLOCK_BITS;
295}
296
297static unsigned long pcpu_off_to_block_off(int off)
298{
299 return off & (PCPU_BITMAP_BLOCK_BITS - 1);
300}
301
302static unsigned long pcpu_block_off_to_off(int index, int off)
303{
304 return index * PCPU_BITMAP_BLOCK_BITS + off;
305}
306
307/**
308 * pcpu_check_block_hint - check against the contig hint
309 * @block: block of interest
310 * @bits: size of allocation
311 * @align: alignment of area (max PAGE_SIZE)
312 *
313 * Check to see if the allocation can fit in the block's contig hint.
314 * Note, a chunk uses the same hints as a block so this can also check against
315 * the chunk's contig hint.
316 */
317static bool pcpu_check_block_hint(struct pcpu_block_md *block, int bits,
318 size_t align)
319{
320 int bit_off = ALIGN(block->contig_hint_start, align) -
321 block->contig_hint_start;
322
323 return bit_off + bits <= block->contig_hint;
324}
325
326/*
327 * pcpu_next_hint - determine which hint to use
328 * @block: block of interest
329 * @alloc_bits: size of allocation
330 *
331 * This determines if we should scan based on the scan_hint or first_free.
332 * In general, we want to scan from first_free to fulfill allocations by
333 * first fit. However, if we know a scan_hint at position scan_hint_start
334 * cannot fulfill an allocation, we can begin scanning from there knowing
335 * the contig_hint will be our fallback.
336 */
337static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
338{
339 /*
340 * The three conditions below determine if we can skip past the
341 * scan_hint. First, does the scan hint exist. Second, is the
342 * contig_hint after the scan_hint (possibly not true iff
343 * contig_hint == scan_hint). Third, is the allocation request
344 * larger than the scan_hint.
345 */
346 if (block->scan_hint &&
347 block->contig_hint_start > block->scan_hint_start &&
348 alloc_bits > block->scan_hint)
349 return block->scan_hint_start + block->scan_hint;
350
351 return block->first_free;
352}
353
354/**
355 * pcpu_next_md_free_region - finds the next hint free area
356 * @chunk: chunk of interest
357 * @bit_off: chunk offset
358 * @bits: size of free area
359 *
360 * Helper function for pcpu_for_each_md_free_region. It checks
361 * block->contig_hint and performs aggregation across blocks to find the
362 * next hint. It modifies bit_off and bits in-place to be consumed in the
363 * loop.
364 */
365static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
366 int *bits)
367{
368 int i = pcpu_off_to_block_index(*bit_off);
369 int block_off = pcpu_off_to_block_off(*bit_off);
370 struct pcpu_block_md *block;
371
372 *bits = 0;
373 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
374 block++, i++) {
375 /* handles contig area across blocks */
376 if (*bits) {
377 *bits += block->left_free;
378 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
379 continue;
380 return;
381 }
382
383 /*
384 * This checks three things. First is there a contig_hint to
385 * check. Second, have we checked this hint before by
386 * comparing the block_off. Third, is this the same as the
387 * right contig hint. In the last case, it spills over into
388 * the next block and should be handled by the contig area
389 * across blocks code.
390 */
391 *bits = block->contig_hint;
392 if (*bits && block->contig_hint_start >= block_off &&
393 *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
394 *bit_off = pcpu_block_off_to_off(i,
395 block->contig_hint_start);
396 return;
397 }
398 /* reset to satisfy the second predicate above */
399 block_off = 0;
400
401 *bits = block->right_free;
402 *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
403 }
404}
405
406/**
407 * pcpu_next_fit_region - finds fit areas for a given allocation request
408 * @chunk: chunk of interest
409 * @alloc_bits: size of allocation
410 * @align: alignment of area (max PAGE_SIZE)
411 * @bit_off: chunk offset
412 * @bits: size of free area
413 *
414 * Finds the next free region that is viable for use with a given size and
415 * alignment. This only returns if there is a valid area to be used for this
416 * allocation. block->first_free is returned if the allocation request fits
417 * within the block to see if the request can be fulfilled prior to the contig
418 * hint.
419 */
420static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
421 int align, int *bit_off, int *bits)
422{
423 int i = pcpu_off_to_block_index(*bit_off);
424 int block_off = pcpu_off_to_block_off(*bit_off);
425 struct pcpu_block_md *block;
426
427 *bits = 0;
428 for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
429 block++, i++) {
430 /* handles contig area across blocks */
431 if (*bits) {
432 *bits += block->left_free;
433 if (*bits >= alloc_bits)
434 return;
435 if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
436 continue;
437 }
438
439 /* check block->contig_hint */
440 *bits = ALIGN(block->contig_hint_start, align) -
441 block->contig_hint_start;
442 /*
443 * This uses the block offset to determine if this has been
444 * checked in the prior iteration.
445 */
446 if (block->contig_hint &&
447 block->contig_hint_start >= block_off &&
448 block->contig_hint >= *bits + alloc_bits) {
449 int start = pcpu_next_hint(block, alloc_bits);
450
451 *bits += alloc_bits + block->contig_hint_start -
452 start;
453 *bit_off = pcpu_block_off_to_off(i, start);
454 return;
455 }
456 /* reset to satisfy the second predicate above */
457 block_off = 0;
458
459 *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
460 align);
461 *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
462 *bit_off = pcpu_block_off_to_off(i, *bit_off);
463 if (*bits >= alloc_bits)
464 return;
465 }
466
467 /* no valid offsets were found - fail condition */
468 *bit_off = pcpu_chunk_map_bits(chunk);
469}
470
471/*
472 * Metadata free area iterators. These perform aggregation of free areas
473 * based on the metadata blocks and return the offset @bit_off and size in
474 * bits of the free area @bits. pcpu_for_each_fit_region only returns when
475 * a fit is found for the allocation request.
476 */
477#define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
478 for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
479 (bit_off) < pcpu_chunk_map_bits((chunk)); \
480 (bit_off) += (bits) + 1, \
481 pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
482
483#define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
484 for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
485 &(bits)); \
486 (bit_off) < pcpu_chunk_map_bits((chunk)); \
487 (bit_off) += (bits), \
488 pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
489 &(bits)))
490
491/**
492 * pcpu_mem_zalloc - allocate memory
493 * @size: bytes to allocate
494 * @gfp: allocation flags
495 *
496 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
497 * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
498 * This is to facilitate passing through whitelisted flags. The
499 * returned memory is always zeroed.
500 *
501 * RETURNS:
502 * Pointer to the allocated area on success, NULL on failure.
503 */
504static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
505{
506 if (WARN_ON_ONCE(!slab_is_available()))
507 return NULL;
508
509 if (size <= PAGE_SIZE)
510 return kzalloc(size, gfp);
511 else
512 return __vmalloc(size, gfp | __GFP_ZERO);
513}
514
515/**
516 * pcpu_mem_free - free memory
517 * @ptr: memory to free
518 *
519 * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
520 */
521static void pcpu_mem_free(void *ptr)
522{
523 kvfree(ptr);
524}
525
526static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
527 bool move_front)
528{
529 if (chunk != pcpu_reserved_chunk) {
530 if (move_front)
531 list_move(&chunk->list, &pcpu_chunk_lists[slot]);
532 else
533 list_move_tail(&chunk->list, &pcpu_chunk_lists[slot]);
534 }
535}
536
537static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
538{
539 __pcpu_chunk_move(chunk, slot, true);
540}
541
542/**
543 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
544 * @chunk: chunk of interest
545 * @oslot: the previous slot it was on
546 *
547 * This function is called after an allocation or free changed @chunk.
548 * New slot according to the changed state is determined and @chunk is
549 * moved to the slot. Note that the reserved chunk is never put on
550 * chunk slots.
551 *
552 * CONTEXT:
553 * pcpu_lock.
554 */
555static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
556{
557 int nslot = pcpu_chunk_slot(chunk);
558
559 /* leave isolated chunks in-place */
560 if (chunk->isolated)
561 return;
562
563 if (oslot != nslot)
564 __pcpu_chunk_move(chunk, nslot, oslot < nslot);
565}
566
567static void pcpu_isolate_chunk(struct pcpu_chunk *chunk)
568{
569 lockdep_assert_held(&pcpu_lock);
570
571 if (!chunk->isolated) {
572 chunk->isolated = true;
573 pcpu_nr_empty_pop_pages -= chunk->nr_empty_pop_pages;
574 }
575 list_move(&chunk->list, &pcpu_chunk_lists[pcpu_to_depopulate_slot]);
576}
577
578static void pcpu_reintegrate_chunk(struct pcpu_chunk *chunk)
579{
580 lockdep_assert_held(&pcpu_lock);
581
582 if (chunk->isolated) {
583 chunk->isolated = false;
584 pcpu_nr_empty_pop_pages += chunk->nr_empty_pop_pages;
585 pcpu_chunk_relocate(chunk, -1);
586 }
587}
588
589/*
590 * pcpu_update_empty_pages - update empty page counters
591 * @chunk: chunk of interest
592 * @nr: nr of empty pages
593 *
594 * This is used to keep track of the empty pages now based on the premise
595 * a md_block covers a page. The hint update functions recognize if a block
596 * is made full or broken to calculate deltas for keeping track of free pages.
597 */
598static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
599{
600 chunk->nr_empty_pop_pages += nr;
601 if (chunk != pcpu_reserved_chunk && !chunk->isolated)
602 pcpu_nr_empty_pop_pages += nr;
603}
604
605/*
606 * pcpu_region_overlap - determines if two regions overlap
607 * @a: start of first region, inclusive
608 * @b: end of first region, exclusive
609 * @x: start of second region, inclusive
610 * @y: end of second region, exclusive
611 *
612 * This is used to determine if the hint region [a, b) overlaps with the
613 * allocated region [x, y).
614 */
615static inline bool pcpu_region_overlap(int a, int b, int x, int y)
616{
617 return (a < y) && (x < b);
618}
619
620/**
621 * pcpu_block_update - updates a block given a free area
622 * @block: block of interest
623 * @start: start offset in block
624 * @end: end offset in block
625 *
626 * Updates a block given a known free area. The region [start, end) is
627 * expected to be the entirety of the free area within a block. Chooses
628 * the best starting offset if the contig hints are equal.
629 */
630static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
631{
632 int contig = end - start;
633
634 block->first_free = min(block->first_free, start);
635 if (start == 0)
636 block->left_free = contig;
637
638 if (end == block->nr_bits)
639 block->right_free = contig;
640
641 if (contig > block->contig_hint) {
642 /* promote the old contig_hint to be the new scan_hint */
643 if (start > block->contig_hint_start) {
644 if (block->contig_hint > block->scan_hint) {
645 block->scan_hint_start =
646 block->contig_hint_start;
647 block->scan_hint = block->contig_hint;
648 } else if (start < block->scan_hint_start) {
649 /*
650 * The old contig_hint == scan_hint. But, the
651 * new contig is larger so hold the invariant
652 * scan_hint_start < contig_hint_start.
653 */
654 block->scan_hint = 0;
655 }
656 } else {
657 block->scan_hint = 0;
658 }
659 block->contig_hint_start = start;
660 block->contig_hint = contig;
661 } else if (contig == block->contig_hint) {
662 if (block->contig_hint_start &&
663 (!start ||
664 __ffs(start) > __ffs(block->contig_hint_start))) {
665 /* start has a better alignment so use it */
666 block->contig_hint_start = start;
667 if (start < block->scan_hint_start &&
668 block->contig_hint > block->scan_hint)
669 block->scan_hint = 0;
670 } else if (start > block->scan_hint_start ||
671 block->contig_hint > block->scan_hint) {
672 /*
673 * Knowing contig == contig_hint, update the scan_hint
674 * if it is farther than or larger than the current
675 * scan_hint.
676 */
677 block->scan_hint_start = start;
678 block->scan_hint = contig;
679 }
680 } else {
681 /*
682 * The region is smaller than the contig_hint. So only update
683 * the scan_hint if it is larger than or equal and farther than
684 * the current scan_hint.
685 */
686 if ((start < block->contig_hint_start &&
687 (contig > block->scan_hint ||
688 (contig == block->scan_hint &&
689 start > block->scan_hint_start)))) {
690 block->scan_hint_start = start;
691 block->scan_hint = contig;
692 }
693 }
694}
695
696/*
697 * pcpu_block_update_scan - update a block given a free area from a scan
698 * @chunk: chunk of interest
699 * @bit_off: chunk offset
700 * @bits: size of free area
701 *
702 * Finding the final allocation spot first goes through pcpu_find_block_fit()
703 * to find a block that can hold the allocation and then pcpu_alloc_area()
704 * where a scan is used. When allocations require specific alignments,
705 * we can inadvertently create holes which will not be seen in the alloc
706 * or free paths.
707 *
708 * This takes a given free area hole and updates a block as it may change the
709 * scan_hint. We need to scan backwards to ensure we don't miss free bits
710 * from alignment.
711 */
712static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
713 int bits)
714{
715 int s_off = pcpu_off_to_block_off(bit_off);
716 int e_off = s_off + bits;
717 int s_index, l_bit;
718 struct pcpu_block_md *block;
719
720 if (e_off > PCPU_BITMAP_BLOCK_BITS)
721 return;
722
723 s_index = pcpu_off_to_block_index(bit_off);
724 block = chunk->md_blocks + s_index;
725
726 /* scan backwards in case of alignment skipping free bits */
727 l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
728 s_off = (s_off == l_bit) ? 0 : l_bit + 1;
729
730 pcpu_block_update(block, s_off, e_off);
731}
732
733/**
734 * pcpu_chunk_refresh_hint - updates metadata about a chunk
735 * @chunk: chunk of interest
736 * @full_scan: if we should scan from the beginning
737 *
738 * Iterates over the metadata blocks to find the largest contig area.
739 * A full scan can be avoided on the allocation path as this is triggered
740 * if we broke the contig_hint. In doing so, the scan_hint will be before
741 * the contig_hint or after if the scan_hint == contig_hint. This cannot
742 * be prevented on freeing as we want to find the largest area possibly
743 * spanning blocks.
744 */
745static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
746{
747 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
748 int bit_off, bits;
749
750 /* promote scan_hint to contig_hint */
751 if (!full_scan && chunk_md->scan_hint) {
752 bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
753 chunk_md->contig_hint_start = chunk_md->scan_hint_start;
754 chunk_md->contig_hint = chunk_md->scan_hint;
755 chunk_md->scan_hint = 0;
756 } else {
757 bit_off = chunk_md->first_free;
758 chunk_md->contig_hint = 0;
759 }
760
761 bits = 0;
762 pcpu_for_each_md_free_region(chunk, bit_off, bits)
763 pcpu_block_update(chunk_md, bit_off, bit_off + bits);
764}
765
766/**
767 * pcpu_block_refresh_hint
768 * @chunk: chunk of interest
769 * @index: index of the metadata block
770 *
771 * Scans over the block beginning at first_free and updates the block
772 * metadata accordingly.
773 */
774static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
775{
776 struct pcpu_block_md *block = chunk->md_blocks + index;
777 unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
778 unsigned int start, end; /* region start, region end */
779
780 /* promote scan_hint to contig_hint */
781 if (block->scan_hint) {
782 start = block->scan_hint_start + block->scan_hint;
783 block->contig_hint_start = block->scan_hint_start;
784 block->contig_hint = block->scan_hint;
785 block->scan_hint = 0;
786 } else {
787 start = block->first_free;
788 block->contig_hint = 0;
789 }
790
791 block->right_free = 0;
792
793 /* iterate over free areas and update the contig hints */
794 for_each_clear_bitrange_from(start, end, alloc_map, PCPU_BITMAP_BLOCK_BITS)
795 pcpu_block_update(block, start, end);
796}
797
798/**
799 * pcpu_block_update_hint_alloc - update hint on allocation path
800 * @chunk: chunk of interest
801 * @bit_off: chunk offset
802 * @bits: size of request
803 *
804 * Updates metadata for the allocation path. The metadata only has to be
805 * refreshed by a full scan iff the chunk's contig hint is broken. Block level
806 * scans are required if the block's contig hint is broken.
807 */
808static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
809 int bits)
810{
811 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
812 int nr_empty_pages = 0;
813 struct pcpu_block_md *s_block, *e_block, *block;
814 int s_index, e_index; /* block indexes of the freed allocation */
815 int s_off, e_off; /* block offsets of the freed allocation */
816
817 /*
818 * Calculate per block offsets.
819 * The calculation uses an inclusive range, but the resulting offsets
820 * are [start, end). e_index always points to the last block in the
821 * range.
822 */
823 s_index = pcpu_off_to_block_index(bit_off);
824 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
825 s_off = pcpu_off_to_block_off(bit_off);
826 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
827
828 s_block = chunk->md_blocks + s_index;
829 e_block = chunk->md_blocks + e_index;
830
831 /*
832 * Update s_block.
833 */
834 if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
835 nr_empty_pages++;
836
837 /*
838 * block->first_free must be updated if the allocation takes its place.
839 * If the allocation breaks the contig_hint, a scan is required to
840 * restore this hint.
841 */
842 if (s_off == s_block->first_free)
843 s_block->first_free = find_next_zero_bit(
844 pcpu_index_alloc_map(chunk, s_index),
845 PCPU_BITMAP_BLOCK_BITS,
846 s_off + bits);
847
848 if (pcpu_region_overlap(s_block->scan_hint_start,
849 s_block->scan_hint_start + s_block->scan_hint,
850 s_off,
851 s_off + bits))
852 s_block->scan_hint = 0;
853
854 if (pcpu_region_overlap(s_block->contig_hint_start,
855 s_block->contig_hint_start +
856 s_block->contig_hint,
857 s_off,
858 s_off + bits)) {
859 /* block contig hint is broken - scan to fix it */
860 if (!s_off)
861 s_block->left_free = 0;
862 pcpu_block_refresh_hint(chunk, s_index);
863 } else {
864 /* update left and right contig manually */
865 s_block->left_free = min(s_block->left_free, s_off);
866 if (s_index == e_index)
867 s_block->right_free = min_t(int, s_block->right_free,
868 PCPU_BITMAP_BLOCK_BITS - e_off);
869 else
870 s_block->right_free = 0;
871 }
872
873 /*
874 * Update e_block.
875 */
876 if (s_index != e_index) {
877 if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
878 nr_empty_pages++;
879
880 /*
881 * When the allocation is across blocks, the end is along
882 * the left part of the e_block.
883 */
884 e_block->first_free = find_next_zero_bit(
885 pcpu_index_alloc_map(chunk, e_index),
886 PCPU_BITMAP_BLOCK_BITS, e_off);
887
888 if (e_off == PCPU_BITMAP_BLOCK_BITS) {
889 /* reset the block */
890 e_block++;
891 } else {
892 if (e_off > e_block->scan_hint_start)
893 e_block->scan_hint = 0;
894
895 e_block->left_free = 0;
896 if (e_off > e_block->contig_hint_start) {
897 /* contig hint is broken - scan to fix it */
898 pcpu_block_refresh_hint(chunk, e_index);
899 } else {
900 e_block->right_free =
901 min_t(int, e_block->right_free,
902 PCPU_BITMAP_BLOCK_BITS - e_off);
903 }
904 }
905
906 /* update in-between md_blocks */
907 nr_empty_pages += (e_index - s_index - 1);
908 for (block = s_block + 1; block < e_block; block++) {
909 block->scan_hint = 0;
910 block->contig_hint = 0;
911 block->left_free = 0;
912 block->right_free = 0;
913 }
914 }
915
916 /*
917 * If the allocation is not atomic, some blocks may not be
918 * populated with pages, while we account it here. The number
919 * of pages will be added back with pcpu_chunk_populated()
920 * when populating pages.
921 */
922 if (nr_empty_pages)
923 pcpu_update_empty_pages(chunk, -nr_empty_pages);
924
925 if (pcpu_region_overlap(chunk_md->scan_hint_start,
926 chunk_md->scan_hint_start +
927 chunk_md->scan_hint,
928 bit_off,
929 bit_off + bits))
930 chunk_md->scan_hint = 0;
931
932 /*
933 * The only time a full chunk scan is required is if the chunk
934 * contig hint is broken. Otherwise, it means a smaller space
935 * was used and therefore the chunk contig hint is still correct.
936 */
937 if (pcpu_region_overlap(chunk_md->contig_hint_start,
938 chunk_md->contig_hint_start +
939 chunk_md->contig_hint,
940 bit_off,
941 bit_off + bits))
942 pcpu_chunk_refresh_hint(chunk, false);
943}
944
945/**
946 * pcpu_block_update_hint_free - updates the block hints on the free path
947 * @chunk: chunk of interest
948 * @bit_off: chunk offset
949 * @bits: size of request
950 *
951 * Updates metadata for the allocation path. This avoids a blind block
952 * refresh by making use of the block contig hints. If this fails, it scans
953 * forward and backward to determine the extent of the free area. This is
954 * capped at the boundary of blocks.
955 *
956 * A chunk update is triggered if a page becomes free, a block becomes free,
957 * or the free spans across blocks. This tradeoff is to minimize iterating
958 * over the block metadata to update chunk_md->contig_hint.
959 * chunk_md->contig_hint may be off by up to a page, but it will never be more
960 * than the available space. If the contig hint is contained in one block, it
961 * will be accurate.
962 */
963static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
964 int bits)
965{
966 int nr_empty_pages = 0;
967 struct pcpu_block_md *s_block, *e_block, *block;
968 int s_index, e_index; /* block indexes of the freed allocation */
969 int s_off, e_off; /* block offsets of the freed allocation */
970 int start, end; /* start and end of the whole free area */
971
972 /*
973 * Calculate per block offsets.
974 * The calculation uses an inclusive range, but the resulting offsets
975 * are [start, end). e_index always points to the last block in the
976 * range.
977 */
978 s_index = pcpu_off_to_block_index(bit_off);
979 e_index = pcpu_off_to_block_index(bit_off + bits - 1);
980 s_off = pcpu_off_to_block_off(bit_off);
981 e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
982
983 s_block = chunk->md_blocks + s_index;
984 e_block = chunk->md_blocks + e_index;
985
986 /*
987 * Check if the freed area aligns with the block->contig_hint.
988 * If it does, then the scan to find the beginning/end of the
989 * larger free area can be avoided.
990 *
991 * start and end refer to beginning and end of the free area
992 * within each their respective blocks. This is not necessarily
993 * the entire free area as it may span blocks past the beginning
994 * or end of the block.
995 */
996 start = s_off;
997 if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
998 start = s_block->contig_hint_start;
999 } else {
1000 /*
1001 * Scan backwards to find the extent of the free area.
1002 * find_last_bit returns the starting bit, so if the start bit
1003 * is returned, that means there was no last bit and the
1004 * remainder of the chunk is free.
1005 */
1006 int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
1007 start);
1008 start = (start == l_bit) ? 0 : l_bit + 1;
1009 }
1010
1011 end = e_off;
1012 if (e_off == e_block->contig_hint_start)
1013 end = e_block->contig_hint_start + e_block->contig_hint;
1014 else
1015 end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
1016 PCPU_BITMAP_BLOCK_BITS, end);
1017
1018 /* update s_block */
1019 e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
1020 if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
1021 nr_empty_pages++;
1022 pcpu_block_update(s_block, start, e_off);
1023
1024 /* freeing in the same block */
1025 if (s_index != e_index) {
1026 /* update e_block */
1027 if (end == PCPU_BITMAP_BLOCK_BITS)
1028 nr_empty_pages++;
1029 pcpu_block_update(e_block, 0, end);
1030
1031 /* reset md_blocks in the middle */
1032 nr_empty_pages += (e_index - s_index - 1);
1033 for (block = s_block + 1; block < e_block; block++) {
1034 block->first_free = 0;
1035 block->scan_hint = 0;
1036 block->contig_hint_start = 0;
1037 block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
1038 block->left_free = PCPU_BITMAP_BLOCK_BITS;
1039 block->right_free = PCPU_BITMAP_BLOCK_BITS;
1040 }
1041 }
1042
1043 if (nr_empty_pages)
1044 pcpu_update_empty_pages(chunk, nr_empty_pages);
1045
1046 /*
1047 * Refresh chunk metadata when the free makes a block free or spans
1048 * across blocks. The contig_hint may be off by up to a page, but if
1049 * the contig_hint is contained in a block, it will be accurate with
1050 * the else condition below.
1051 */
1052 if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
1053 pcpu_chunk_refresh_hint(chunk, true);
1054 else
1055 pcpu_block_update(&chunk->chunk_md,
1056 pcpu_block_off_to_off(s_index, start),
1057 end);
1058}
1059
1060/**
1061 * pcpu_is_populated - determines if the region is populated
1062 * @chunk: chunk of interest
1063 * @bit_off: chunk offset
1064 * @bits: size of area
1065 * @next_off: return value for the next offset to start searching
1066 *
1067 * For atomic allocations, check if the backing pages are populated.
1068 *
1069 * RETURNS:
1070 * Bool if the backing pages are populated.
1071 * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1072 */
1073static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1074 int *next_off)
1075{
1076 unsigned int start, end;
1077
1078 start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1079 end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1080
1081 start = find_next_zero_bit(chunk->populated, end, start);
1082 if (start >= end)
1083 return true;
1084
1085 end = find_next_bit(chunk->populated, end, start + 1);
1086
1087 *next_off = end * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1088 return false;
1089}
1090
1091/**
1092 * pcpu_find_block_fit - finds the block index to start searching
1093 * @chunk: chunk of interest
1094 * @alloc_bits: size of request in allocation units
1095 * @align: alignment of area (max PAGE_SIZE bytes)
1096 * @pop_only: use populated regions only
1097 *
1098 * Given a chunk and an allocation spec, find the offset to begin searching
1099 * for a free region. This iterates over the bitmap metadata blocks to
1100 * find an offset that will be guaranteed to fit the requirements. It is
1101 * not quite first fit as if the allocation does not fit in the contig hint
1102 * of a block or chunk, it is skipped. This errs on the side of caution
1103 * to prevent excess iteration. Poor alignment can cause the allocator to
1104 * skip over blocks and chunks that have valid free areas.
1105 *
1106 * RETURNS:
1107 * The offset in the bitmap to begin searching.
1108 * -1 if no offset is found.
1109 */
1110static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1111 size_t align, bool pop_only)
1112{
1113 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1114 int bit_off, bits, next_off;
1115
1116 /*
1117 * This is an optimization to prevent scanning by assuming if the
1118 * allocation cannot fit in the global hint, there is memory pressure
1119 * and creating a new chunk would happen soon.
1120 */
1121 if (!pcpu_check_block_hint(chunk_md, alloc_bits, align))
1122 return -1;
1123
1124 bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1125 bits = 0;
1126 pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1127 if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1128 &next_off))
1129 break;
1130
1131 bit_off = next_off;
1132 bits = 0;
1133 }
1134
1135 if (bit_off == pcpu_chunk_map_bits(chunk))
1136 return -1;
1137
1138 return bit_off;
1139}
1140
1141/*
1142 * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1143 * @map: the address to base the search on
1144 * @size: the bitmap size in bits
1145 * @start: the bitnumber to start searching at
1146 * @nr: the number of zeroed bits we're looking for
1147 * @align_mask: alignment mask for zero area
1148 * @largest_off: offset of the largest area skipped
1149 * @largest_bits: size of the largest area skipped
1150 *
1151 * The @align_mask should be one less than a power of 2.
1152 *
1153 * This is a modified version of bitmap_find_next_zero_area_off() to remember
1154 * the largest area that was skipped. This is imperfect, but in general is
1155 * good enough. The largest remembered region is the largest failed region
1156 * seen. This does not include anything we possibly skipped due to alignment.
1157 * pcpu_block_update_scan() does scan backwards to try and recover what was
1158 * lost to alignment. While this can cause scanning to miss earlier possible
1159 * free areas, smaller allocations will eventually fill those holes.
1160 */
1161static unsigned long pcpu_find_zero_area(unsigned long *map,
1162 unsigned long size,
1163 unsigned long start,
1164 unsigned long nr,
1165 unsigned long align_mask,
1166 unsigned long *largest_off,
1167 unsigned long *largest_bits)
1168{
1169 unsigned long index, end, i, area_off, area_bits;
1170again:
1171 index = find_next_zero_bit(map, size, start);
1172
1173 /* Align allocation */
1174 index = __ALIGN_MASK(index, align_mask);
1175 area_off = index;
1176
1177 end = index + nr;
1178 if (end > size)
1179 return end;
1180 i = find_next_bit(map, end, index);
1181 if (i < end) {
1182 area_bits = i - area_off;
1183 /* remember largest unused area with best alignment */
1184 if (area_bits > *largest_bits ||
1185 (area_bits == *largest_bits && *largest_off &&
1186 (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1187 *largest_off = area_off;
1188 *largest_bits = area_bits;
1189 }
1190
1191 start = i + 1;
1192 goto again;
1193 }
1194 return index;
1195}
1196
1197/**
1198 * pcpu_alloc_area - allocates an area from a pcpu_chunk
1199 * @chunk: chunk of interest
1200 * @alloc_bits: size of request in allocation units
1201 * @align: alignment of area (max PAGE_SIZE)
1202 * @start: bit_off to start searching
1203 *
1204 * This function takes in a @start offset to begin searching to fit an
1205 * allocation of @alloc_bits with alignment @align. It needs to scan
1206 * the allocation map because if it fits within the block's contig hint,
1207 * @start will be block->first_free. This is an attempt to fill the
1208 * allocation prior to breaking the contig hint. The allocation and
1209 * boundary maps are updated accordingly if it confirms a valid
1210 * free area.
1211 *
1212 * RETURNS:
1213 * Allocated addr offset in @chunk on success.
1214 * -1 if no matching area is found.
1215 */
1216static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1217 size_t align, int start)
1218{
1219 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1220 size_t align_mask = (align) ? (align - 1) : 0;
1221 unsigned long area_off = 0, area_bits = 0;
1222 int bit_off, end, oslot;
1223
1224 lockdep_assert_held(&pcpu_lock);
1225
1226 oslot = pcpu_chunk_slot(chunk);
1227
1228 /*
1229 * Search to find a fit.
1230 */
1231 end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1232 pcpu_chunk_map_bits(chunk));
1233 bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1234 align_mask, &area_off, &area_bits);
1235 if (bit_off >= end)
1236 return -1;
1237
1238 if (area_bits)
1239 pcpu_block_update_scan(chunk, area_off, area_bits);
1240
1241 /* update alloc map */
1242 bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1243
1244 /* update boundary map */
1245 set_bit(bit_off, chunk->bound_map);
1246 bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1247 set_bit(bit_off + alloc_bits, chunk->bound_map);
1248
1249 chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1250
1251 /* update first free bit */
1252 if (bit_off == chunk_md->first_free)
1253 chunk_md->first_free = find_next_zero_bit(
1254 chunk->alloc_map,
1255 pcpu_chunk_map_bits(chunk),
1256 bit_off + alloc_bits);
1257
1258 pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1259
1260 pcpu_chunk_relocate(chunk, oslot);
1261
1262 return bit_off * PCPU_MIN_ALLOC_SIZE;
1263}
1264
1265/**
1266 * pcpu_free_area - frees the corresponding offset
1267 * @chunk: chunk of interest
1268 * @off: addr offset into chunk
1269 *
1270 * This function determines the size of an allocation to free using
1271 * the boundary bitmap and clears the allocation map.
1272 *
1273 * RETURNS:
1274 * Number of freed bytes.
1275 */
1276static int pcpu_free_area(struct pcpu_chunk *chunk, int off)
1277{
1278 struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1279 int bit_off, bits, end, oslot, freed;
1280
1281 lockdep_assert_held(&pcpu_lock);
1282 pcpu_stats_area_dealloc(chunk);
1283
1284 oslot = pcpu_chunk_slot(chunk);
1285
1286 bit_off = off / PCPU_MIN_ALLOC_SIZE;
1287
1288 /* find end index */
1289 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1290 bit_off + 1);
1291 bits = end - bit_off;
1292 bitmap_clear(chunk->alloc_map, bit_off, bits);
1293
1294 freed = bits * PCPU_MIN_ALLOC_SIZE;
1295
1296 /* update metadata */
1297 chunk->free_bytes += freed;
1298
1299 /* update first free bit */
1300 chunk_md->first_free = min(chunk_md->first_free, bit_off);
1301
1302 pcpu_block_update_hint_free(chunk, bit_off, bits);
1303
1304 pcpu_chunk_relocate(chunk, oslot);
1305
1306 return freed;
1307}
1308
1309static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1310{
1311 block->scan_hint = 0;
1312 block->contig_hint = nr_bits;
1313 block->left_free = nr_bits;
1314 block->right_free = nr_bits;
1315 block->first_free = 0;
1316 block->nr_bits = nr_bits;
1317}
1318
1319static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1320{
1321 struct pcpu_block_md *md_block;
1322
1323 /* init the chunk's block */
1324 pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1325
1326 for (md_block = chunk->md_blocks;
1327 md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1328 md_block++)
1329 pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1330}
1331
1332/**
1333 * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1334 * @tmp_addr: the start of the region served
1335 * @map_size: size of the region served
1336 *
1337 * This is responsible for creating the chunks that serve the first chunk. The
1338 * base_addr is page aligned down of @tmp_addr while the region end is page
1339 * aligned up. Offsets are kept track of to determine the region served. All
1340 * this is done to appease the bitmap allocator in avoiding partial blocks.
1341 *
1342 * RETURNS:
1343 * Chunk serving the region at @tmp_addr of @map_size.
1344 */
1345static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1346 int map_size)
1347{
1348 struct pcpu_chunk *chunk;
1349 unsigned long aligned_addr;
1350 int start_offset, offset_bits, region_size, region_bits;
1351 size_t alloc_size;
1352
1353 /* region calculations */
1354 aligned_addr = tmp_addr & PAGE_MASK;
1355
1356 start_offset = tmp_addr - aligned_addr;
1357 region_size = ALIGN(start_offset + map_size, PAGE_SIZE);
1358
1359 /* allocate chunk */
1360 alloc_size = struct_size(chunk, populated,
1361 BITS_TO_LONGS(region_size >> PAGE_SHIFT));
1362 chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1363 if (!chunk)
1364 panic("%s: Failed to allocate %zu bytes\n", __func__,
1365 alloc_size);
1366
1367 INIT_LIST_HEAD(&chunk->list);
1368
1369 chunk->base_addr = (void *)aligned_addr;
1370 chunk->start_offset = start_offset;
1371 chunk->end_offset = region_size - chunk->start_offset - map_size;
1372
1373 chunk->nr_pages = region_size >> PAGE_SHIFT;
1374 region_bits = pcpu_chunk_map_bits(chunk);
1375
1376 alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1377 chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1378 if (!chunk->alloc_map)
1379 panic("%s: Failed to allocate %zu bytes\n", __func__,
1380 alloc_size);
1381
1382 alloc_size =
1383 BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1384 chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1385 if (!chunk->bound_map)
1386 panic("%s: Failed to allocate %zu bytes\n", __func__,
1387 alloc_size);
1388
1389 alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1390 chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1391 if (!chunk->md_blocks)
1392 panic("%s: Failed to allocate %zu bytes\n", __func__,
1393 alloc_size);
1394
1395#ifdef CONFIG_MEMCG_KMEM
1396 /* first chunk is free to use */
1397 chunk->obj_cgroups = NULL;
1398#endif
1399 pcpu_init_md_blocks(chunk);
1400
1401 /* manage populated page bitmap */
1402 chunk->immutable = true;
1403 bitmap_fill(chunk->populated, chunk->nr_pages);
1404 chunk->nr_populated = chunk->nr_pages;
1405 chunk->nr_empty_pop_pages = chunk->nr_pages;
1406
1407 chunk->free_bytes = map_size;
1408
1409 if (chunk->start_offset) {
1410 /* hide the beginning of the bitmap */
1411 offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1412 bitmap_set(chunk->alloc_map, 0, offset_bits);
1413 set_bit(0, chunk->bound_map);
1414 set_bit(offset_bits, chunk->bound_map);
1415
1416 chunk->chunk_md.first_free = offset_bits;
1417
1418 pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1419 }
1420
1421 if (chunk->end_offset) {
1422 /* hide the end of the bitmap */
1423 offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1424 bitmap_set(chunk->alloc_map,
1425 pcpu_chunk_map_bits(chunk) - offset_bits,
1426 offset_bits);
1427 set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1428 chunk->bound_map);
1429 set_bit(region_bits, chunk->bound_map);
1430
1431 pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1432 - offset_bits, offset_bits);
1433 }
1434
1435 return chunk;
1436}
1437
1438static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1439{
1440 struct pcpu_chunk *chunk;
1441 int region_bits;
1442
1443 chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1444 if (!chunk)
1445 return NULL;
1446
1447 INIT_LIST_HEAD(&chunk->list);
1448 chunk->nr_pages = pcpu_unit_pages;
1449 region_bits = pcpu_chunk_map_bits(chunk);
1450
1451 chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1452 sizeof(chunk->alloc_map[0]), gfp);
1453 if (!chunk->alloc_map)
1454 goto alloc_map_fail;
1455
1456 chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1457 sizeof(chunk->bound_map[0]), gfp);
1458 if (!chunk->bound_map)
1459 goto bound_map_fail;
1460
1461 chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1462 sizeof(chunk->md_blocks[0]), gfp);
1463 if (!chunk->md_blocks)
1464 goto md_blocks_fail;
1465
1466#ifdef CONFIG_MEMCG_KMEM
1467 if (!mem_cgroup_kmem_disabled()) {
1468 chunk->obj_cgroups =
1469 pcpu_mem_zalloc(pcpu_chunk_map_bits(chunk) *
1470 sizeof(struct obj_cgroup *), gfp);
1471 if (!chunk->obj_cgroups)
1472 goto objcg_fail;
1473 }
1474#endif
1475
1476 pcpu_init_md_blocks(chunk);
1477
1478 /* init metadata */
1479 chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1480
1481 return chunk;
1482
1483#ifdef CONFIG_MEMCG_KMEM
1484objcg_fail:
1485 pcpu_mem_free(chunk->md_blocks);
1486#endif
1487md_blocks_fail:
1488 pcpu_mem_free(chunk->bound_map);
1489bound_map_fail:
1490 pcpu_mem_free(chunk->alloc_map);
1491alloc_map_fail:
1492 pcpu_mem_free(chunk);
1493
1494 return NULL;
1495}
1496
1497static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1498{
1499 if (!chunk)
1500 return;
1501#ifdef CONFIG_MEMCG_KMEM
1502 pcpu_mem_free(chunk->obj_cgroups);
1503#endif
1504 pcpu_mem_free(chunk->md_blocks);
1505 pcpu_mem_free(chunk->bound_map);
1506 pcpu_mem_free(chunk->alloc_map);
1507 pcpu_mem_free(chunk);
1508}
1509
1510/**
1511 * pcpu_chunk_populated - post-population bookkeeping
1512 * @chunk: pcpu_chunk which got populated
1513 * @page_start: the start page
1514 * @page_end: the end page
1515 *
1516 * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1517 * the bookkeeping information accordingly. Must be called after each
1518 * successful population.
1519 */
1520static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1521 int page_end)
1522{
1523 int nr = page_end - page_start;
1524
1525 lockdep_assert_held(&pcpu_lock);
1526
1527 bitmap_set(chunk->populated, page_start, nr);
1528 chunk->nr_populated += nr;
1529 pcpu_nr_populated += nr;
1530
1531 pcpu_update_empty_pages(chunk, nr);
1532}
1533
1534/**
1535 * pcpu_chunk_depopulated - post-depopulation bookkeeping
1536 * @chunk: pcpu_chunk which got depopulated
1537 * @page_start: the start page
1538 * @page_end: the end page
1539 *
1540 * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1541 * Update the bookkeeping information accordingly. Must be called after
1542 * each successful depopulation.
1543 */
1544static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1545 int page_start, int page_end)
1546{
1547 int nr = page_end - page_start;
1548
1549 lockdep_assert_held(&pcpu_lock);
1550
1551 bitmap_clear(chunk->populated, page_start, nr);
1552 chunk->nr_populated -= nr;
1553 pcpu_nr_populated -= nr;
1554
1555 pcpu_update_empty_pages(chunk, -nr);
1556}
1557
1558/*
1559 * Chunk management implementation.
1560 *
1561 * To allow different implementations, chunk alloc/free and
1562 * [de]population are implemented in a separate file which is pulled
1563 * into this file and compiled together. The following functions
1564 * should be implemented.
1565 *
1566 * pcpu_populate_chunk - populate the specified range of a chunk
1567 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1568 * pcpu_post_unmap_tlb_flush - flush tlb for the specified range of a chunk
1569 * pcpu_create_chunk - create a new chunk
1570 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1571 * pcpu_addr_to_page - translate address to physical address
1572 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1573 */
1574static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1575 int page_start, int page_end, gfp_t gfp);
1576static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1577 int page_start, int page_end);
1578static void pcpu_post_unmap_tlb_flush(struct pcpu_chunk *chunk,
1579 int page_start, int page_end);
1580static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1581static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1582static struct page *pcpu_addr_to_page(void *addr);
1583static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1584
1585#ifdef CONFIG_NEED_PER_CPU_KM
1586#include "percpu-km.c"
1587#else
1588#include "percpu-vm.c"
1589#endif
1590
1591/**
1592 * pcpu_chunk_addr_search - determine chunk containing specified address
1593 * @addr: address for which the chunk needs to be determined.
1594 *
1595 * This is an internal function that handles all but static allocations.
1596 * Static percpu address values should never be passed into the allocator.
1597 *
1598 * RETURNS:
1599 * The address of the found chunk.
1600 */
1601static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1602{
1603 /* is it in the dynamic region (first chunk)? */
1604 if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1605 return pcpu_first_chunk;
1606
1607 /* is it in the reserved region? */
1608 if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1609 return pcpu_reserved_chunk;
1610
1611 /*
1612 * The address is relative to unit0 which might be unused and
1613 * thus unmapped. Offset the address to the unit space of the
1614 * current processor before looking it up in the vmalloc
1615 * space. Note that any possible cpu id can be used here, so
1616 * there's no need to worry about preemption or cpu hotplug.
1617 */
1618 addr += pcpu_unit_offsets[raw_smp_processor_id()];
1619 return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1620}
1621
1622#ifdef CONFIG_MEMCG_KMEM
1623static bool pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp,
1624 struct obj_cgroup **objcgp)
1625{
1626 struct obj_cgroup *objcg;
1627
1628 if (!memcg_kmem_online() || !(gfp & __GFP_ACCOUNT))
1629 return true;
1630
1631 objcg = current_obj_cgroup();
1632 if (!objcg)
1633 return true;
1634
1635 if (obj_cgroup_charge(objcg, gfp, pcpu_obj_full_size(size)))
1636 return false;
1637
1638 *objcgp = objcg;
1639 return true;
1640}
1641
1642static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1643 struct pcpu_chunk *chunk, int off,
1644 size_t size)
1645{
1646 if (!objcg)
1647 return;
1648
1649 if (likely(chunk && chunk->obj_cgroups)) {
1650 obj_cgroup_get(objcg);
1651 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = objcg;
1652
1653 rcu_read_lock();
1654 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1655 pcpu_obj_full_size(size));
1656 rcu_read_unlock();
1657 } else {
1658 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1659 }
1660}
1661
1662static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1663{
1664 struct obj_cgroup *objcg;
1665
1666 if (unlikely(!chunk->obj_cgroups))
1667 return;
1668
1669 objcg = chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT];
1670 if (!objcg)
1671 return;
1672 chunk->obj_cgroups[off >> PCPU_MIN_ALLOC_SHIFT] = NULL;
1673
1674 obj_cgroup_uncharge(objcg, pcpu_obj_full_size(size));
1675
1676 rcu_read_lock();
1677 mod_memcg_state(obj_cgroup_memcg(objcg), MEMCG_PERCPU_B,
1678 -pcpu_obj_full_size(size));
1679 rcu_read_unlock();
1680
1681 obj_cgroup_put(objcg);
1682}
1683
1684#else /* CONFIG_MEMCG_KMEM */
1685static bool
1686pcpu_memcg_pre_alloc_hook(size_t size, gfp_t gfp, struct obj_cgroup **objcgp)
1687{
1688 return true;
1689}
1690
1691static void pcpu_memcg_post_alloc_hook(struct obj_cgroup *objcg,
1692 struct pcpu_chunk *chunk, int off,
1693 size_t size)
1694{
1695}
1696
1697static void pcpu_memcg_free_hook(struct pcpu_chunk *chunk, int off, size_t size)
1698{
1699}
1700#endif /* CONFIG_MEMCG_KMEM */
1701
1702/**
1703 * pcpu_alloc - the percpu allocator
1704 * @size: size of area to allocate in bytes
1705 * @align: alignment of area (max PAGE_SIZE)
1706 * @reserved: allocate from the reserved chunk if available
1707 * @gfp: allocation flags
1708 *
1709 * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1710 * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1711 * then no warning will be triggered on invalid or failed allocation
1712 * requests.
1713 *
1714 * RETURNS:
1715 * Percpu pointer to the allocated area on success, NULL on failure.
1716 */
1717static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1718 gfp_t gfp)
1719{
1720 gfp_t pcpu_gfp;
1721 bool is_atomic;
1722 bool do_warn;
1723 struct obj_cgroup *objcg = NULL;
1724 static int warn_limit = 10;
1725 struct pcpu_chunk *chunk, *next;
1726 const char *err;
1727 int slot, off, cpu, ret;
1728 unsigned long flags;
1729 void __percpu *ptr;
1730 size_t bits, bit_align;
1731
1732 gfp = current_gfp_context(gfp);
1733 /* whitelisted flags that can be passed to the backing allocators */
1734 pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1735 is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1736 do_warn = !(gfp & __GFP_NOWARN);
1737
1738 /*
1739 * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1740 * therefore alignment must be a minimum of that many bytes.
1741 * An allocation may have internal fragmentation from rounding up
1742 * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1743 */
1744 if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1745 align = PCPU_MIN_ALLOC_SIZE;
1746
1747 size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1748 bits = size >> PCPU_MIN_ALLOC_SHIFT;
1749 bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1750
1751 if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
1752 !is_power_of_2(align))) {
1753 WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1754 size, align);
1755 return NULL;
1756 }
1757
1758 if (unlikely(!pcpu_memcg_pre_alloc_hook(size, gfp, &objcg)))
1759 return NULL;
1760
1761 if (!is_atomic) {
1762 /*
1763 * pcpu_balance_workfn() allocates memory under this mutex,
1764 * and it may wait for memory reclaim. Allow current task
1765 * to become OOM victim, in case of memory pressure.
1766 */
1767 if (gfp & __GFP_NOFAIL) {
1768 mutex_lock(&pcpu_alloc_mutex);
1769 } else if (mutex_lock_killable(&pcpu_alloc_mutex)) {
1770 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1771 return NULL;
1772 }
1773 }
1774
1775 spin_lock_irqsave(&pcpu_lock, flags);
1776
1777 /* serve reserved allocations from the reserved chunk if available */
1778 if (reserved && pcpu_reserved_chunk) {
1779 chunk = pcpu_reserved_chunk;
1780
1781 off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1782 if (off < 0) {
1783 err = "alloc from reserved chunk failed";
1784 goto fail_unlock;
1785 }
1786
1787 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1788 if (off >= 0)
1789 goto area_found;
1790
1791 err = "alloc from reserved chunk failed";
1792 goto fail_unlock;
1793 }
1794
1795restart:
1796 /* search through normal chunks */
1797 for (slot = pcpu_size_to_slot(size); slot <= pcpu_free_slot; slot++) {
1798 list_for_each_entry_safe(chunk, next, &pcpu_chunk_lists[slot],
1799 list) {
1800 off = pcpu_find_block_fit(chunk, bits, bit_align,
1801 is_atomic);
1802 if (off < 0) {
1803 if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1804 pcpu_chunk_move(chunk, 0);
1805 continue;
1806 }
1807
1808 off = pcpu_alloc_area(chunk, bits, bit_align, off);
1809 if (off >= 0) {
1810 pcpu_reintegrate_chunk(chunk);
1811 goto area_found;
1812 }
1813 }
1814 }
1815
1816 spin_unlock_irqrestore(&pcpu_lock, flags);
1817
1818 if (is_atomic) {
1819 err = "atomic alloc failed, no space left";
1820 goto fail;
1821 }
1822
1823 /* No space left. Create a new chunk. */
1824 if (list_empty(&pcpu_chunk_lists[pcpu_free_slot])) {
1825 chunk = pcpu_create_chunk(pcpu_gfp);
1826 if (!chunk) {
1827 err = "failed to allocate new chunk";
1828 goto fail;
1829 }
1830
1831 spin_lock_irqsave(&pcpu_lock, flags);
1832 pcpu_chunk_relocate(chunk, -1);
1833 } else {
1834 spin_lock_irqsave(&pcpu_lock, flags);
1835 }
1836
1837 goto restart;
1838
1839area_found:
1840 pcpu_stats_area_alloc(chunk, size);
1841 spin_unlock_irqrestore(&pcpu_lock, flags);
1842
1843 /* populate if not all pages are already there */
1844 if (!is_atomic) {
1845 unsigned int page_end, rs, re;
1846
1847 rs = PFN_DOWN(off);
1848 page_end = PFN_UP(off + size);
1849
1850 for_each_clear_bitrange_from(rs, re, chunk->populated, page_end) {
1851 WARN_ON(chunk->immutable);
1852
1853 ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1854
1855 spin_lock_irqsave(&pcpu_lock, flags);
1856 if (ret) {
1857 pcpu_free_area(chunk, off);
1858 err = "failed to populate";
1859 goto fail_unlock;
1860 }
1861 pcpu_chunk_populated(chunk, rs, re);
1862 spin_unlock_irqrestore(&pcpu_lock, flags);
1863 }
1864
1865 mutex_unlock(&pcpu_alloc_mutex);
1866 }
1867
1868 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1869 pcpu_schedule_balance_work();
1870
1871 /* clear the areas and return address relative to base address */
1872 for_each_possible_cpu(cpu)
1873 memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1874
1875 ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1876 kmemleak_alloc_percpu(ptr, size, gfp);
1877
1878 trace_percpu_alloc_percpu(_RET_IP_, reserved, is_atomic, size, align,
1879 chunk->base_addr, off, ptr,
1880 pcpu_obj_full_size(size), gfp);
1881
1882 pcpu_memcg_post_alloc_hook(objcg, chunk, off, size);
1883
1884 return ptr;
1885
1886fail_unlock:
1887 spin_unlock_irqrestore(&pcpu_lock, flags);
1888fail:
1889 trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1890
1891 if (do_warn && warn_limit) {
1892 pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1893 size, align, is_atomic, err);
1894 if (!is_atomic)
1895 dump_stack();
1896 if (!--warn_limit)
1897 pr_info("limit reached, disable warning\n");
1898 }
1899
1900 if (is_atomic) {
1901 /* see the flag handling in pcpu_balance_workfn() */
1902 pcpu_atomic_alloc_failed = true;
1903 pcpu_schedule_balance_work();
1904 } else {
1905 mutex_unlock(&pcpu_alloc_mutex);
1906 }
1907
1908 pcpu_memcg_post_alloc_hook(objcg, NULL, 0, size);
1909
1910 return NULL;
1911}
1912
1913/**
1914 * __alloc_percpu_gfp - allocate dynamic percpu area
1915 * @size: size of area to allocate in bytes
1916 * @align: alignment of area (max PAGE_SIZE)
1917 * @gfp: allocation flags
1918 *
1919 * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1920 * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1921 * be called from any context but is a lot more likely to fail. If @gfp
1922 * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1923 * allocation requests.
1924 *
1925 * RETURNS:
1926 * Percpu pointer to the allocated area on success, NULL on failure.
1927 */
1928void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1929{
1930 return pcpu_alloc(size, align, false, gfp);
1931}
1932EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1933
1934/**
1935 * __alloc_percpu - allocate dynamic percpu area
1936 * @size: size of area to allocate in bytes
1937 * @align: alignment of area (max PAGE_SIZE)
1938 *
1939 * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1940 */
1941void __percpu *__alloc_percpu(size_t size, size_t align)
1942{
1943 return pcpu_alloc(size, align, false, GFP_KERNEL);
1944}
1945EXPORT_SYMBOL_GPL(__alloc_percpu);
1946
1947/**
1948 * __alloc_reserved_percpu - allocate reserved percpu area
1949 * @size: size of area to allocate in bytes
1950 * @align: alignment of area (max PAGE_SIZE)
1951 *
1952 * Allocate zero-filled percpu area of @size bytes aligned at @align
1953 * from reserved percpu area if arch has set it up; otherwise,
1954 * allocation is served from the same dynamic area. Might sleep.
1955 * Might trigger writeouts.
1956 *
1957 * CONTEXT:
1958 * Does GFP_KERNEL allocation.
1959 *
1960 * RETURNS:
1961 * Percpu pointer to the allocated area on success, NULL on failure.
1962 */
1963void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1964{
1965 return pcpu_alloc(size, align, true, GFP_KERNEL);
1966}
1967
1968/**
1969 * pcpu_balance_free - manage the amount of free chunks
1970 * @empty_only: free chunks only if there are no populated pages
1971 *
1972 * If empty_only is %false, reclaim all fully free chunks regardless of the
1973 * number of populated pages. Otherwise, only reclaim chunks that have no
1974 * populated pages.
1975 *
1976 * CONTEXT:
1977 * pcpu_lock (can be dropped temporarily)
1978 */
1979static void pcpu_balance_free(bool empty_only)
1980{
1981 LIST_HEAD(to_free);
1982 struct list_head *free_head = &pcpu_chunk_lists[pcpu_free_slot];
1983 struct pcpu_chunk *chunk, *next;
1984
1985 lockdep_assert_held(&pcpu_lock);
1986
1987 /*
1988 * There's no reason to keep around multiple unused chunks and VM
1989 * areas can be scarce. Destroy all free chunks except for one.
1990 */
1991 list_for_each_entry_safe(chunk, next, free_head, list) {
1992 WARN_ON(chunk->immutable);
1993
1994 /* spare the first one */
1995 if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1996 continue;
1997
1998 if (!empty_only || chunk->nr_empty_pop_pages == 0)
1999 list_move(&chunk->list, &to_free);
2000 }
2001
2002 if (list_empty(&to_free))
2003 return;
2004
2005 spin_unlock_irq(&pcpu_lock);
2006 list_for_each_entry_safe(chunk, next, &to_free, list) {
2007 unsigned int rs, re;
2008
2009 for_each_set_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2010 pcpu_depopulate_chunk(chunk, rs, re);
2011 spin_lock_irq(&pcpu_lock);
2012 pcpu_chunk_depopulated(chunk, rs, re);
2013 spin_unlock_irq(&pcpu_lock);
2014 }
2015 pcpu_destroy_chunk(chunk);
2016 cond_resched();
2017 }
2018 spin_lock_irq(&pcpu_lock);
2019}
2020
2021/**
2022 * pcpu_balance_populated - manage the amount of populated pages
2023 *
2024 * Maintain a certain amount of populated pages to satisfy atomic allocations.
2025 * It is possible that this is called when physical memory is scarce causing
2026 * OOM killer to be triggered. We should avoid doing so until an actual
2027 * allocation causes the failure as it is possible that requests can be
2028 * serviced from already backed regions.
2029 *
2030 * CONTEXT:
2031 * pcpu_lock (can be dropped temporarily)
2032 */
2033static void pcpu_balance_populated(void)
2034{
2035 /* gfp flags passed to underlying allocators */
2036 const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
2037 struct pcpu_chunk *chunk;
2038 int slot, nr_to_pop, ret;
2039
2040 lockdep_assert_held(&pcpu_lock);
2041
2042 /*
2043 * Ensure there are certain number of free populated pages for
2044 * atomic allocs. Fill up from the most packed so that atomic
2045 * allocs don't increase fragmentation. If atomic allocation
2046 * failed previously, always populate the maximum amount. This
2047 * should prevent atomic allocs larger than PAGE_SIZE from keeping
2048 * failing indefinitely; however, large atomic allocs are not
2049 * something we support properly and can be highly unreliable and
2050 * inefficient.
2051 */
2052retry_pop:
2053 if (pcpu_atomic_alloc_failed) {
2054 nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
2055 /* best effort anyway, don't worry about synchronization */
2056 pcpu_atomic_alloc_failed = false;
2057 } else {
2058 nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
2059 pcpu_nr_empty_pop_pages,
2060 0, PCPU_EMPTY_POP_PAGES_HIGH);
2061 }
2062
2063 for (slot = pcpu_size_to_slot(PAGE_SIZE); slot <= pcpu_free_slot; slot++) {
2064 unsigned int nr_unpop = 0, rs, re;
2065
2066 if (!nr_to_pop)
2067 break;
2068
2069 list_for_each_entry(chunk, &pcpu_chunk_lists[slot], list) {
2070 nr_unpop = chunk->nr_pages - chunk->nr_populated;
2071 if (nr_unpop)
2072 break;
2073 }
2074
2075 if (!nr_unpop)
2076 continue;
2077
2078 /* @chunk can't go away while pcpu_alloc_mutex is held */
2079 for_each_clear_bitrange(rs, re, chunk->populated, chunk->nr_pages) {
2080 int nr = min_t(int, re - rs, nr_to_pop);
2081
2082 spin_unlock_irq(&pcpu_lock);
2083 ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
2084 cond_resched();
2085 spin_lock_irq(&pcpu_lock);
2086 if (!ret) {
2087 nr_to_pop -= nr;
2088 pcpu_chunk_populated(chunk, rs, rs + nr);
2089 } else {
2090 nr_to_pop = 0;
2091 }
2092
2093 if (!nr_to_pop)
2094 break;
2095 }
2096 }
2097
2098 if (nr_to_pop) {
2099 /* ran out of chunks to populate, create a new one and retry */
2100 spin_unlock_irq(&pcpu_lock);
2101 chunk = pcpu_create_chunk(gfp);
2102 cond_resched();
2103 spin_lock_irq(&pcpu_lock);
2104 if (chunk) {
2105 pcpu_chunk_relocate(chunk, -1);
2106 goto retry_pop;
2107 }
2108 }
2109}
2110
2111/**
2112 * pcpu_reclaim_populated - scan over to_depopulate chunks and free empty pages
2113 *
2114 * Scan over chunks in the depopulate list and try to release unused populated
2115 * pages back to the system. Depopulated chunks are sidelined to prevent
2116 * repopulating these pages unless required. Fully free chunks are reintegrated
2117 * and freed accordingly (1 is kept around). If we drop below the empty
2118 * populated pages threshold, reintegrate the chunk if it has empty free pages.
2119 * Each chunk is scanned in the reverse order to keep populated pages close to
2120 * the beginning of the chunk.
2121 *
2122 * CONTEXT:
2123 * pcpu_lock (can be dropped temporarily)
2124 *
2125 */
2126static void pcpu_reclaim_populated(void)
2127{
2128 struct pcpu_chunk *chunk;
2129 struct pcpu_block_md *block;
2130 int freed_page_start, freed_page_end;
2131 int i, end;
2132 bool reintegrate;
2133
2134 lockdep_assert_held(&pcpu_lock);
2135
2136 /*
2137 * Once a chunk is isolated to the to_depopulate list, the chunk is no
2138 * longer discoverable to allocations whom may populate pages. The only
2139 * other accessor is the free path which only returns area back to the
2140 * allocator not touching the populated bitmap.
2141 */
2142 while ((chunk = list_first_entry_or_null(
2143 &pcpu_chunk_lists[pcpu_to_depopulate_slot],
2144 struct pcpu_chunk, list))) {
2145 WARN_ON(chunk->immutable);
2146
2147 /*
2148 * Scan chunk's pages in the reverse order to keep populated
2149 * pages close to the beginning of the chunk.
2150 */
2151 freed_page_start = chunk->nr_pages;
2152 freed_page_end = 0;
2153 reintegrate = false;
2154 for (i = chunk->nr_pages - 1, end = -1; i >= 0; i--) {
2155 /* no more work to do */
2156 if (chunk->nr_empty_pop_pages == 0)
2157 break;
2158
2159 /* reintegrate chunk to prevent atomic alloc failures */
2160 if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_HIGH) {
2161 reintegrate = true;
2162 break;
2163 }
2164
2165 /*
2166 * If the page is empty and populated, start or
2167 * extend the (i, end) range. If i == 0, decrease
2168 * i and perform the depopulation to cover the last
2169 * (first) page in the chunk.
2170 */
2171 block = chunk->md_blocks + i;
2172 if (block->contig_hint == PCPU_BITMAP_BLOCK_BITS &&
2173 test_bit(i, chunk->populated)) {
2174 if (end == -1)
2175 end = i;
2176 if (i > 0)
2177 continue;
2178 i--;
2179 }
2180
2181 /* depopulate if there is an active range */
2182 if (end == -1)
2183 continue;
2184
2185 spin_unlock_irq(&pcpu_lock);
2186 pcpu_depopulate_chunk(chunk, i + 1, end + 1);
2187 cond_resched();
2188 spin_lock_irq(&pcpu_lock);
2189
2190 pcpu_chunk_depopulated(chunk, i + 1, end + 1);
2191 freed_page_start = min(freed_page_start, i + 1);
2192 freed_page_end = max(freed_page_end, end + 1);
2193
2194 /* reset the range and continue */
2195 end = -1;
2196 }
2197
2198 /* batch tlb flush per chunk to amortize cost */
2199 if (freed_page_start < freed_page_end) {
2200 spin_unlock_irq(&pcpu_lock);
2201 pcpu_post_unmap_tlb_flush(chunk,
2202 freed_page_start,
2203 freed_page_end);
2204 cond_resched();
2205 spin_lock_irq(&pcpu_lock);
2206 }
2207
2208 if (reintegrate || chunk->free_bytes == pcpu_unit_size)
2209 pcpu_reintegrate_chunk(chunk);
2210 else
2211 list_move_tail(&chunk->list,
2212 &pcpu_chunk_lists[pcpu_sidelined_slot]);
2213 }
2214}
2215
2216/**
2217 * pcpu_balance_workfn - manage the amount of free chunks and populated pages
2218 * @work: unused
2219 *
2220 * For each chunk type, manage the number of fully free chunks and the number of
2221 * populated pages. An important thing to consider is when pages are freed and
2222 * how they contribute to the global counts.
2223 */
2224static void pcpu_balance_workfn(struct work_struct *work)
2225{
2226 /*
2227 * pcpu_balance_free() is called twice because the first time we may
2228 * trim pages in the active pcpu_nr_empty_pop_pages which may cause us
2229 * to grow other chunks. This then gives pcpu_reclaim_populated() time
2230 * to move fully free chunks to the active list to be freed if
2231 * appropriate.
2232 */
2233 mutex_lock(&pcpu_alloc_mutex);
2234 spin_lock_irq(&pcpu_lock);
2235
2236 pcpu_balance_free(false);
2237 pcpu_reclaim_populated();
2238 pcpu_balance_populated();
2239 pcpu_balance_free(true);
2240
2241 spin_unlock_irq(&pcpu_lock);
2242 mutex_unlock(&pcpu_alloc_mutex);
2243}
2244
2245/**
2246 * pcpu_alloc_size - the size of the dynamic percpu area
2247 * @ptr: pointer to the dynamic percpu area
2248 *
2249 * Returns the size of the @ptr allocation. This is undefined for statically
2250 * defined percpu variables as there is no corresponding chunk->bound_map.
2251 *
2252 * RETURNS:
2253 * The size of the dynamic percpu area.
2254 *
2255 * CONTEXT:
2256 * Can be called from atomic context.
2257 */
2258size_t pcpu_alloc_size(void __percpu *ptr)
2259{
2260 struct pcpu_chunk *chunk;
2261 unsigned long bit_off, end;
2262 void *addr;
2263
2264 if (!ptr)
2265 return 0;
2266
2267 addr = __pcpu_ptr_to_addr(ptr);
2268 /* No pcpu_lock here: ptr has not been freed, so chunk is still alive */
2269 chunk = pcpu_chunk_addr_search(addr);
2270 bit_off = (addr - chunk->base_addr) / PCPU_MIN_ALLOC_SIZE;
2271 end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
2272 bit_off + 1);
2273 return (end - bit_off) * PCPU_MIN_ALLOC_SIZE;
2274}
2275
2276/**
2277 * free_percpu - free percpu area
2278 * @ptr: pointer to area to free
2279 *
2280 * Free percpu area @ptr.
2281 *
2282 * CONTEXT:
2283 * Can be called from atomic context.
2284 */
2285void free_percpu(void __percpu *ptr)
2286{
2287 void *addr;
2288 struct pcpu_chunk *chunk;
2289 unsigned long flags;
2290 int size, off;
2291 bool need_balance = false;
2292
2293 if (!ptr)
2294 return;
2295
2296 kmemleak_free_percpu(ptr);
2297
2298 addr = __pcpu_ptr_to_addr(ptr);
2299 chunk = pcpu_chunk_addr_search(addr);
2300 off = addr - chunk->base_addr;
2301
2302 spin_lock_irqsave(&pcpu_lock, flags);
2303 size = pcpu_free_area(chunk, off);
2304
2305 pcpu_memcg_free_hook(chunk, off, size);
2306
2307 /*
2308 * If there are more than one fully free chunks, wake up grim reaper.
2309 * If the chunk is isolated, it may be in the process of being
2310 * reclaimed. Let reclaim manage cleaning up of that chunk.
2311 */
2312 if (!chunk->isolated && chunk->free_bytes == pcpu_unit_size) {
2313 struct pcpu_chunk *pos;
2314
2315 list_for_each_entry(pos, &pcpu_chunk_lists[pcpu_free_slot], list)
2316 if (pos != chunk) {
2317 need_balance = true;
2318 break;
2319 }
2320 } else if (pcpu_should_reclaim_chunk(chunk)) {
2321 pcpu_isolate_chunk(chunk);
2322 need_balance = true;
2323 }
2324
2325 trace_percpu_free_percpu(chunk->base_addr, off, ptr);
2326
2327 spin_unlock_irqrestore(&pcpu_lock, flags);
2328
2329 if (need_balance)
2330 pcpu_schedule_balance_work();
2331}
2332EXPORT_SYMBOL_GPL(free_percpu);
2333
2334bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
2335{
2336#ifdef CONFIG_SMP
2337 const size_t static_size = __per_cpu_end - __per_cpu_start;
2338 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2339 unsigned int cpu;
2340
2341 for_each_possible_cpu(cpu) {
2342 void *start = per_cpu_ptr(base, cpu);
2343 void *va = (void *)addr;
2344
2345 if (va >= start && va < start + static_size) {
2346 if (can_addr) {
2347 *can_addr = (unsigned long) (va - start);
2348 *can_addr += (unsigned long)
2349 per_cpu_ptr(base, get_boot_cpu_id());
2350 }
2351 return true;
2352 }
2353 }
2354#endif
2355 /* on UP, can't distinguish from other static vars, always false */
2356 return false;
2357}
2358
2359/**
2360 * is_kernel_percpu_address - test whether address is from static percpu area
2361 * @addr: address to test
2362 *
2363 * Test whether @addr belongs to in-kernel static percpu area. Module
2364 * static percpu areas are not considered. For those, use
2365 * is_module_percpu_address().
2366 *
2367 * RETURNS:
2368 * %true if @addr is from in-kernel static percpu area, %false otherwise.
2369 */
2370bool is_kernel_percpu_address(unsigned long addr)
2371{
2372 return __is_kernel_percpu_address(addr, NULL);
2373}
2374
2375/**
2376 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2377 * @addr: the address to be converted to physical address
2378 *
2379 * Given @addr which is dereferenceable address obtained via one of
2380 * percpu access macros, this function translates it into its physical
2381 * address. The caller is responsible for ensuring @addr stays valid
2382 * until this function finishes.
2383 *
2384 * percpu allocator has special setup for the first chunk, which currently
2385 * supports either embedding in linear address space or vmalloc mapping,
2386 * and, from the second one, the backing allocator (currently either vm or
2387 * km) provides translation.
2388 *
2389 * The addr can be translated simply without checking if it falls into the
2390 * first chunk. But the current code reflects better how percpu allocator
2391 * actually works, and the verification can discover both bugs in percpu
2392 * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2393 * code.
2394 *
2395 * RETURNS:
2396 * The physical address for @addr.
2397 */
2398phys_addr_t per_cpu_ptr_to_phys(void *addr)
2399{
2400 void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2401 bool in_first_chunk = false;
2402 unsigned long first_low, first_high;
2403 unsigned int cpu;
2404
2405 /*
2406 * The following test on unit_low/high isn't strictly
2407 * necessary but will speed up lookups of addresses which
2408 * aren't in the first chunk.
2409 *
2410 * The address check is against full chunk sizes. pcpu_base_addr
2411 * points to the beginning of the first chunk including the
2412 * static region. Assumes good intent as the first chunk may
2413 * not be full (ie. < pcpu_unit_pages in size).
2414 */
2415 first_low = (unsigned long)pcpu_base_addr +
2416 pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2417 first_high = (unsigned long)pcpu_base_addr +
2418 pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2419 if ((unsigned long)addr >= first_low &&
2420 (unsigned long)addr < first_high) {
2421 for_each_possible_cpu(cpu) {
2422 void *start = per_cpu_ptr(base, cpu);
2423
2424 if (addr >= start && addr < start + pcpu_unit_size) {
2425 in_first_chunk = true;
2426 break;
2427 }
2428 }
2429 }
2430
2431 if (in_first_chunk) {
2432 if (!is_vmalloc_addr(addr))
2433 return __pa(addr);
2434 else
2435 return page_to_phys(vmalloc_to_page(addr)) +
2436 offset_in_page(addr);
2437 } else
2438 return page_to_phys(pcpu_addr_to_page(addr)) +
2439 offset_in_page(addr);
2440}
2441
2442/**
2443 * pcpu_alloc_alloc_info - allocate percpu allocation info
2444 * @nr_groups: the number of groups
2445 * @nr_units: the number of units
2446 *
2447 * Allocate ai which is large enough for @nr_groups groups containing
2448 * @nr_units units. The returned ai's groups[0].cpu_map points to the
2449 * cpu_map array which is long enough for @nr_units and filled with
2450 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2451 * pointer of other groups.
2452 *
2453 * RETURNS:
2454 * Pointer to the allocated pcpu_alloc_info on success, NULL on
2455 * failure.
2456 */
2457struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2458 int nr_units)
2459{
2460 struct pcpu_alloc_info *ai;
2461 size_t base_size, ai_size;
2462 void *ptr;
2463 int unit;
2464
2465 base_size = ALIGN(struct_size(ai, groups, nr_groups),
2466 __alignof__(ai->groups[0].cpu_map[0]));
2467 ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2468
2469 ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2470 if (!ptr)
2471 return NULL;
2472 ai = ptr;
2473 ptr += base_size;
2474
2475 ai->groups[0].cpu_map = ptr;
2476
2477 for (unit = 0; unit < nr_units; unit++)
2478 ai->groups[0].cpu_map[unit] = NR_CPUS;
2479
2480 ai->nr_groups = nr_groups;
2481 ai->__ai_size = PFN_ALIGN(ai_size);
2482
2483 return ai;
2484}
2485
2486/**
2487 * pcpu_free_alloc_info - free percpu allocation info
2488 * @ai: pcpu_alloc_info to free
2489 *
2490 * Free @ai which was allocated by pcpu_alloc_alloc_info().
2491 */
2492void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2493{
2494 memblock_free(ai, ai->__ai_size);
2495}
2496
2497/**
2498 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2499 * @lvl: loglevel
2500 * @ai: allocation info to dump
2501 *
2502 * Print out information about @ai using loglevel @lvl.
2503 */
2504static void pcpu_dump_alloc_info(const char *lvl,
2505 const struct pcpu_alloc_info *ai)
2506{
2507 int group_width = 1, cpu_width = 1, width;
2508 char empty_str[] = "--------";
2509 int alloc = 0, alloc_end = 0;
2510 int group, v;
2511 int upa, apl; /* units per alloc, allocs per line */
2512
2513 v = ai->nr_groups;
2514 while (v /= 10)
2515 group_width++;
2516
2517 v = num_possible_cpus();
2518 while (v /= 10)
2519 cpu_width++;
2520 empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2521
2522 upa = ai->alloc_size / ai->unit_size;
2523 width = upa * (cpu_width + 1) + group_width + 3;
2524 apl = rounddown_pow_of_two(max(60 / width, 1));
2525
2526 printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2527 lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2528 ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2529
2530 for (group = 0; group < ai->nr_groups; group++) {
2531 const struct pcpu_group_info *gi = &ai->groups[group];
2532 int unit = 0, unit_end = 0;
2533
2534 BUG_ON(gi->nr_units % upa);
2535 for (alloc_end += gi->nr_units / upa;
2536 alloc < alloc_end; alloc++) {
2537 if (!(alloc % apl)) {
2538 pr_cont("\n");
2539 printk("%spcpu-alloc: ", lvl);
2540 }
2541 pr_cont("[%0*d] ", group_width, group);
2542
2543 for (unit_end += upa; unit < unit_end; unit++)
2544 if (gi->cpu_map[unit] != NR_CPUS)
2545 pr_cont("%0*d ",
2546 cpu_width, gi->cpu_map[unit]);
2547 else
2548 pr_cont("%s ", empty_str);
2549 }
2550 }
2551 pr_cont("\n");
2552}
2553
2554/**
2555 * pcpu_setup_first_chunk - initialize the first percpu chunk
2556 * @ai: pcpu_alloc_info describing how to percpu area is shaped
2557 * @base_addr: mapped address
2558 *
2559 * Initialize the first percpu chunk which contains the kernel static
2560 * percpu area. This function is to be called from arch percpu area
2561 * setup path.
2562 *
2563 * @ai contains all information necessary to initialize the first
2564 * chunk and prime the dynamic percpu allocator.
2565 *
2566 * @ai->static_size is the size of static percpu area.
2567 *
2568 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2569 * reserve after the static area in the first chunk. This reserves
2570 * the first chunk such that it's available only through reserved
2571 * percpu allocation. This is primarily used to serve module percpu
2572 * static areas on architectures where the addressing model has
2573 * limited offset range for symbol relocations to guarantee module
2574 * percpu symbols fall inside the relocatable range.
2575 *
2576 * @ai->dyn_size determines the number of bytes available for dynamic
2577 * allocation in the first chunk. The area between @ai->static_size +
2578 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2579 *
2580 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2581 * and equal to or larger than @ai->static_size + @ai->reserved_size +
2582 * @ai->dyn_size.
2583 *
2584 * @ai->atom_size is the allocation atom size and used as alignment
2585 * for vm areas.
2586 *
2587 * @ai->alloc_size is the allocation size and always multiple of
2588 * @ai->atom_size. This is larger than @ai->atom_size if
2589 * @ai->unit_size is larger than @ai->atom_size.
2590 *
2591 * @ai->nr_groups and @ai->groups describe virtual memory layout of
2592 * percpu areas. Units which should be colocated are put into the
2593 * same group. Dynamic VM areas will be allocated according to these
2594 * groupings. If @ai->nr_groups is zero, a single group containing
2595 * all units is assumed.
2596 *
2597 * The caller should have mapped the first chunk at @base_addr and
2598 * copied static data to each unit.
2599 *
2600 * The first chunk will always contain a static and a dynamic region.
2601 * However, the static region is not managed by any chunk. If the first
2602 * chunk also contains a reserved region, it is served by two chunks -
2603 * one for the reserved region and one for the dynamic region. They
2604 * share the same vm, but use offset regions in the area allocation map.
2605 * The chunk serving the dynamic region is circulated in the chunk slots
2606 * and available for dynamic allocation like any other chunk.
2607 */
2608void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2609 void *base_addr)
2610{
2611 size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2612 size_t static_size, dyn_size;
2613 unsigned long *group_offsets;
2614 size_t *group_sizes;
2615 unsigned long *unit_off;
2616 unsigned int cpu;
2617 int *unit_map;
2618 int group, unit, i;
2619 unsigned long tmp_addr;
2620 size_t alloc_size;
2621
2622#define PCPU_SETUP_BUG_ON(cond) do { \
2623 if (unlikely(cond)) { \
2624 pr_emerg("failed to initialize, %s\n", #cond); \
2625 pr_emerg("cpu_possible_mask=%*pb\n", \
2626 cpumask_pr_args(cpu_possible_mask)); \
2627 pcpu_dump_alloc_info(KERN_EMERG, ai); \
2628 BUG(); \
2629 } \
2630} while (0)
2631
2632 /* sanity checks */
2633 PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2634#ifdef CONFIG_SMP
2635 PCPU_SETUP_BUG_ON(!ai->static_size);
2636 PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2637#endif
2638 PCPU_SETUP_BUG_ON(!base_addr);
2639 PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2640 PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2641 PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2642 PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2643 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2644 PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2645 PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2646 PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2647 IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2648 PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2649
2650 /* process group information and build config tables accordingly */
2651 alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2652 group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2653 if (!group_offsets)
2654 panic("%s: Failed to allocate %zu bytes\n", __func__,
2655 alloc_size);
2656
2657 alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2658 group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2659 if (!group_sizes)
2660 panic("%s: Failed to allocate %zu bytes\n", __func__,
2661 alloc_size);
2662
2663 alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2664 unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2665 if (!unit_map)
2666 panic("%s: Failed to allocate %zu bytes\n", __func__,
2667 alloc_size);
2668
2669 alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2670 unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2671 if (!unit_off)
2672 panic("%s: Failed to allocate %zu bytes\n", __func__,
2673 alloc_size);
2674
2675 for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2676 unit_map[cpu] = UINT_MAX;
2677
2678 pcpu_low_unit_cpu = NR_CPUS;
2679 pcpu_high_unit_cpu = NR_CPUS;
2680
2681 for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2682 const struct pcpu_group_info *gi = &ai->groups[group];
2683
2684 group_offsets[group] = gi->base_offset;
2685 group_sizes[group] = gi->nr_units * ai->unit_size;
2686
2687 for (i = 0; i < gi->nr_units; i++) {
2688 cpu = gi->cpu_map[i];
2689 if (cpu == NR_CPUS)
2690 continue;
2691
2692 PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2693 PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2694 PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2695
2696 unit_map[cpu] = unit + i;
2697 unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2698
2699 /* determine low/high unit_cpu */
2700 if (pcpu_low_unit_cpu == NR_CPUS ||
2701 unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2702 pcpu_low_unit_cpu = cpu;
2703 if (pcpu_high_unit_cpu == NR_CPUS ||
2704 unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2705 pcpu_high_unit_cpu = cpu;
2706 }
2707 }
2708 pcpu_nr_units = unit;
2709
2710 for_each_possible_cpu(cpu)
2711 PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2712
2713 /* we're done parsing the input, undefine BUG macro and dump config */
2714#undef PCPU_SETUP_BUG_ON
2715 pcpu_dump_alloc_info(KERN_DEBUG, ai);
2716
2717 pcpu_nr_groups = ai->nr_groups;
2718 pcpu_group_offsets = group_offsets;
2719 pcpu_group_sizes = group_sizes;
2720 pcpu_unit_map = unit_map;
2721 pcpu_unit_offsets = unit_off;
2722
2723 /* determine basic parameters */
2724 pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2725 pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2726 pcpu_atom_size = ai->atom_size;
2727 pcpu_chunk_struct_size = struct_size((struct pcpu_chunk *)0, populated,
2728 BITS_TO_LONGS(pcpu_unit_pages));
2729
2730 pcpu_stats_save_ai(ai);
2731
2732 /*
2733 * Allocate chunk slots. The slots after the active slots are:
2734 * sidelined_slot - isolated, depopulated chunks
2735 * free_slot - fully free chunks
2736 * to_depopulate_slot - isolated, chunks to depopulate
2737 */
2738 pcpu_sidelined_slot = __pcpu_size_to_slot(pcpu_unit_size) + 1;
2739 pcpu_free_slot = pcpu_sidelined_slot + 1;
2740 pcpu_to_depopulate_slot = pcpu_free_slot + 1;
2741 pcpu_nr_slots = pcpu_to_depopulate_slot + 1;
2742 pcpu_chunk_lists = memblock_alloc(pcpu_nr_slots *
2743 sizeof(pcpu_chunk_lists[0]),
2744 SMP_CACHE_BYTES);
2745 if (!pcpu_chunk_lists)
2746 panic("%s: Failed to allocate %zu bytes\n", __func__,
2747 pcpu_nr_slots * sizeof(pcpu_chunk_lists[0]));
2748
2749 for (i = 0; i < pcpu_nr_slots; i++)
2750 INIT_LIST_HEAD(&pcpu_chunk_lists[i]);
2751
2752 /*
2753 * The end of the static region needs to be aligned with the
2754 * minimum allocation size as this offsets the reserved and
2755 * dynamic region. The first chunk ends page aligned by
2756 * expanding the dynamic region, therefore the dynamic region
2757 * can be shrunk to compensate while still staying above the
2758 * configured sizes.
2759 */
2760 static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2761 dyn_size = ai->dyn_size - (static_size - ai->static_size);
2762
2763 /*
2764 * Initialize first chunk:
2765 * This chunk is broken up into 3 parts:
2766 * < static | [reserved] | dynamic >
2767 * - static - there is no backing chunk because these allocations can
2768 * never be freed.
2769 * - reserved (pcpu_reserved_chunk) - exists primarily to serve
2770 * allocations from module load.
2771 * - dynamic (pcpu_first_chunk) - serves the dynamic part of the first
2772 * chunk.
2773 */
2774 tmp_addr = (unsigned long)base_addr + static_size;
2775 if (ai->reserved_size)
2776 pcpu_reserved_chunk = pcpu_alloc_first_chunk(tmp_addr,
2777 ai->reserved_size);
2778 tmp_addr = (unsigned long)base_addr + static_size + ai->reserved_size;
2779 pcpu_first_chunk = pcpu_alloc_first_chunk(tmp_addr, dyn_size);
2780
2781 pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2782 pcpu_chunk_relocate(pcpu_first_chunk, -1);
2783
2784 /* include all regions of the first chunk */
2785 pcpu_nr_populated += PFN_DOWN(size_sum);
2786
2787 pcpu_stats_chunk_alloc();
2788 trace_percpu_create_chunk(base_addr);
2789
2790 /* we're done */
2791 pcpu_base_addr = base_addr;
2792}
2793
2794#ifdef CONFIG_SMP
2795
2796const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2797 [PCPU_FC_AUTO] = "auto",
2798 [PCPU_FC_EMBED] = "embed",
2799 [PCPU_FC_PAGE] = "page",
2800};
2801
2802enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2803
2804static int __init percpu_alloc_setup(char *str)
2805{
2806 if (!str)
2807 return -EINVAL;
2808
2809 if (0)
2810 /* nada */;
2811#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2812 else if (!strcmp(str, "embed"))
2813 pcpu_chosen_fc = PCPU_FC_EMBED;
2814#endif
2815#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2816 else if (!strcmp(str, "page"))
2817 pcpu_chosen_fc = PCPU_FC_PAGE;
2818#endif
2819 else
2820 pr_warn("unknown allocator %s specified\n", str);
2821
2822 return 0;
2823}
2824early_param("percpu_alloc", percpu_alloc_setup);
2825
2826/*
2827 * pcpu_embed_first_chunk() is used by the generic percpu setup.
2828 * Build it if needed by the arch config or the generic setup is going
2829 * to be used.
2830 */
2831#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2832 !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2833#define BUILD_EMBED_FIRST_CHUNK
2834#endif
2835
2836/* build pcpu_page_first_chunk() iff needed by the arch config */
2837#if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2838#define BUILD_PAGE_FIRST_CHUNK
2839#endif
2840
2841/* pcpu_build_alloc_info() is used by both embed and page first chunk */
2842#if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2843/**
2844 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2845 * @reserved_size: the size of reserved percpu area in bytes
2846 * @dyn_size: minimum free size for dynamic allocation in bytes
2847 * @atom_size: allocation atom size
2848 * @cpu_distance_fn: callback to determine distance between cpus, optional
2849 *
2850 * This function determines grouping of units, their mappings to cpus
2851 * and other parameters considering needed percpu size, allocation
2852 * atom size and distances between CPUs.
2853 *
2854 * Groups are always multiples of atom size and CPUs which are of
2855 * LOCAL_DISTANCE both ways are grouped together and share space for
2856 * units in the same group. The returned configuration is guaranteed
2857 * to have CPUs on different nodes on different groups and >=75% usage
2858 * of allocated virtual address space.
2859 *
2860 * RETURNS:
2861 * On success, pointer to the new allocation_info is returned. On
2862 * failure, ERR_PTR value is returned.
2863 */
2864static struct pcpu_alloc_info * __init __flatten pcpu_build_alloc_info(
2865 size_t reserved_size, size_t dyn_size,
2866 size_t atom_size,
2867 pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2868{
2869 static int group_map[NR_CPUS] __initdata;
2870 static int group_cnt[NR_CPUS] __initdata;
2871 static struct cpumask mask __initdata;
2872 const size_t static_size = __per_cpu_end - __per_cpu_start;
2873 int nr_groups = 1, nr_units = 0;
2874 size_t size_sum, min_unit_size, alloc_size;
2875 int upa, max_upa, best_upa; /* units_per_alloc */
2876 int last_allocs, group, unit;
2877 unsigned int cpu, tcpu;
2878 struct pcpu_alloc_info *ai;
2879 unsigned int *cpu_map;
2880
2881 /* this function may be called multiple times */
2882 memset(group_map, 0, sizeof(group_map));
2883 memset(group_cnt, 0, sizeof(group_cnt));
2884 cpumask_clear(&mask);
2885
2886 /* calculate size_sum and ensure dyn_size is enough for early alloc */
2887 size_sum = PFN_ALIGN(static_size + reserved_size +
2888 max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2889 dyn_size = size_sum - static_size - reserved_size;
2890
2891 /*
2892 * Determine min_unit_size, alloc_size and max_upa such that
2893 * alloc_size is multiple of atom_size and is the smallest
2894 * which can accommodate 4k aligned segments which are equal to
2895 * or larger than min_unit_size.
2896 */
2897 min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2898
2899 /* determine the maximum # of units that can fit in an allocation */
2900 alloc_size = roundup(min_unit_size, atom_size);
2901 upa = alloc_size / min_unit_size;
2902 while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2903 upa--;
2904 max_upa = upa;
2905
2906 cpumask_copy(&mask, cpu_possible_mask);
2907
2908 /* group cpus according to their proximity */
2909 for (group = 0; !cpumask_empty(&mask); group++) {
2910 /* pop the group's first cpu */
2911 cpu = cpumask_first(&mask);
2912 group_map[cpu] = group;
2913 group_cnt[group]++;
2914 cpumask_clear_cpu(cpu, &mask);
2915
2916 for_each_cpu(tcpu, &mask) {
2917 if (!cpu_distance_fn ||
2918 (cpu_distance_fn(cpu, tcpu) == LOCAL_DISTANCE &&
2919 cpu_distance_fn(tcpu, cpu) == LOCAL_DISTANCE)) {
2920 group_map[tcpu] = group;
2921 group_cnt[group]++;
2922 cpumask_clear_cpu(tcpu, &mask);
2923 }
2924 }
2925 }
2926 nr_groups = group;
2927
2928 /*
2929 * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2930 * Expand the unit_size until we use >= 75% of the units allocated.
2931 * Related to atom_size, which could be much larger than the unit_size.
2932 */
2933 last_allocs = INT_MAX;
2934 best_upa = 0;
2935 for (upa = max_upa; upa; upa--) {
2936 int allocs = 0, wasted = 0;
2937
2938 if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2939 continue;
2940
2941 for (group = 0; group < nr_groups; group++) {
2942 int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2943 allocs += this_allocs;
2944 wasted += this_allocs * upa - group_cnt[group];
2945 }
2946
2947 /*
2948 * Don't accept if wastage is over 1/3. The
2949 * greater-than comparison ensures upa==1 always
2950 * passes the following check.
2951 */
2952 if (wasted > num_possible_cpus() / 3)
2953 continue;
2954
2955 /* and then don't consume more memory */
2956 if (allocs > last_allocs)
2957 break;
2958 last_allocs = allocs;
2959 best_upa = upa;
2960 }
2961 BUG_ON(!best_upa);
2962 upa = best_upa;
2963
2964 /* allocate and fill alloc_info */
2965 for (group = 0; group < nr_groups; group++)
2966 nr_units += roundup(group_cnt[group], upa);
2967
2968 ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2969 if (!ai)
2970 return ERR_PTR(-ENOMEM);
2971 cpu_map = ai->groups[0].cpu_map;
2972
2973 for (group = 0; group < nr_groups; group++) {
2974 ai->groups[group].cpu_map = cpu_map;
2975 cpu_map += roundup(group_cnt[group], upa);
2976 }
2977
2978 ai->static_size = static_size;
2979 ai->reserved_size = reserved_size;
2980 ai->dyn_size = dyn_size;
2981 ai->unit_size = alloc_size / upa;
2982 ai->atom_size = atom_size;
2983 ai->alloc_size = alloc_size;
2984
2985 for (group = 0, unit = 0; group < nr_groups; group++) {
2986 struct pcpu_group_info *gi = &ai->groups[group];
2987
2988 /*
2989 * Initialize base_offset as if all groups are located
2990 * back-to-back. The caller should update this to
2991 * reflect actual allocation.
2992 */
2993 gi->base_offset = unit * ai->unit_size;
2994
2995 for_each_possible_cpu(cpu)
2996 if (group_map[cpu] == group)
2997 gi->cpu_map[gi->nr_units++] = cpu;
2998 gi->nr_units = roundup(gi->nr_units, upa);
2999 unit += gi->nr_units;
3000 }
3001 BUG_ON(unit != nr_units);
3002
3003 return ai;
3004}
3005
3006static void * __init pcpu_fc_alloc(unsigned int cpu, size_t size, size_t align,
3007 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3008{
3009 const unsigned long goal = __pa(MAX_DMA_ADDRESS);
3010#ifdef CONFIG_NUMA
3011 int node = NUMA_NO_NODE;
3012 void *ptr;
3013
3014 if (cpu_to_nd_fn)
3015 node = cpu_to_nd_fn(cpu);
3016
3017 if (node == NUMA_NO_NODE || !node_online(node) || !NODE_DATA(node)) {
3018 ptr = memblock_alloc_from(size, align, goal);
3019 pr_info("cpu %d has no node %d or node-local memory\n",
3020 cpu, node);
3021 pr_debug("per cpu data for cpu%d %zu bytes at 0x%llx\n",
3022 cpu, size, (u64)__pa(ptr));
3023 } else {
3024 ptr = memblock_alloc_try_nid(size, align, goal,
3025 MEMBLOCK_ALLOC_ACCESSIBLE,
3026 node);
3027
3028 pr_debug("per cpu data for cpu%d %zu bytes on node%d at 0x%llx\n",
3029 cpu, size, node, (u64)__pa(ptr));
3030 }
3031 return ptr;
3032#else
3033 return memblock_alloc_from(size, align, goal);
3034#endif
3035}
3036
3037static void __init pcpu_fc_free(void *ptr, size_t size)
3038{
3039 memblock_free(ptr, size);
3040}
3041#endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
3042
3043#if defined(BUILD_EMBED_FIRST_CHUNK)
3044/**
3045 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
3046 * @reserved_size: the size of reserved percpu area in bytes
3047 * @dyn_size: minimum free size for dynamic allocation in bytes
3048 * @atom_size: allocation atom size
3049 * @cpu_distance_fn: callback to determine distance between cpus, optional
3050 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3051 *
3052 * This is a helper to ease setting up embedded first percpu chunk and
3053 * can be called where pcpu_setup_first_chunk() is expected.
3054 *
3055 * If this function is used to setup the first chunk, it is allocated
3056 * by calling pcpu_fc_alloc and used as-is without being mapped into
3057 * vmalloc area. Allocations are always whole multiples of @atom_size
3058 * aligned to @atom_size.
3059 *
3060 * This enables the first chunk to piggy back on the linear physical
3061 * mapping which often uses larger page size. Please note that this
3062 * can result in very sparse cpu->unit mapping on NUMA machines thus
3063 * requiring large vmalloc address space. Don't use this allocator if
3064 * vmalloc space is not orders of magnitude larger than distances
3065 * between node memory addresses (ie. 32bit NUMA machines).
3066 *
3067 * @dyn_size specifies the minimum dynamic area size.
3068 *
3069 * If the needed size is smaller than the minimum or specified unit
3070 * size, the leftover is returned using pcpu_fc_free.
3071 *
3072 * RETURNS:
3073 * 0 on success, -errno on failure.
3074 */
3075int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
3076 size_t atom_size,
3077 pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
3078 pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3079{
3080 void *base = (void *)ULONG_MAX;
3081 void **areas = NULL;
3082 struct pcpu_alloc_info *ai;
3083 size_t size_sum, areas_size;
3084 unsigned long max_distance;
3085 int group, i, highest_group, rc = 0;
3086
3087 ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
3088 cpu_distance_fn);
3089 if (IS_ERR(ai))
3090 return PTR_ERR(ai);
3091
3092 size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
3093 areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
3094
3095 areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
3096 if (!areas) {
3097 rc = -ENOMEM;
3098 goto out_free;
3099 }
3100
3101 /* allocate, copy and determine base address & max_distance */
3102 highest_group = 0;
3103 for (group = 0; group < ai->nr_groups; group++) {
3104 struct pcpu_group_info *gi = &ai->groups[group];
3105 unsigned int cpu = NR_CPUS;
3106 void *ptr;
3107
3108 for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
3109 cpu = gi->cpu_map[i];
3110 BUG_ON(cpu == NR_CPUS);
3111
3112 /* allocate space for the whole group */
3113 ptr = pcpu_fc_alloc(cpu, gi->nr_units * ai->unit_size, atom_size, cpu_to_nd_fn);
3114 if (!ptr) {
3115 rc = -ENOMEM;
3116 goto out_free_areas;
3117 }
3118 /* kmemleak tracks the percpu allocations separately */
3119 kmemleak_ignore_phys(__pa(ptr));
3120 areas[group] = ptr;
3121
3122 base = min(ptr, base);
3123 if (ptr > areas[highest_group])
3124 highest_group = group;
3125 }
3126 max_distance = areas[highest_group] - base;
3127 max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
3128
3129 /* warn if maximum distance is further than 75% of vmalloc space */
3130 if (max_distance > VMALLOC_TOTAL * 3 / 4) {
3131 pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
3132 max_distance, VMALLOC_TOTAL);
3133#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
3134 /* and fail if we have fallback */
3135 rc = -EINVAL;
3136 goto out_free_areas;
3137#endif
3138 }
3139
3140 /*
3141 * Copy data and free unused parts. This should happen after all
3142 * allocations are complete; otherwise, we may end up with
3143 * overlapping groups.
3144 */
3145 for (group = 0; group < ai->nr_groups; group++) {
3146 struct pcpu_group_info *gi = &ai->groups[group];
3147 void *ptr = areas[group];
3148
3149 for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
3150 if (gi->cpu_map[i] == NR_CPUS) {
3151 /* unused unit, free whole */
3152 pcpu_fc_free(ptr, ai->unit_size);
3153 continue;
3154 }
3155 /* copy and return the unused part */
3156 memcpy(ptr, __per_cpu_load, ai->static_size);
3157 pcpu_fc_free(ptr + size_sum, ai->unit_size - size_sum);
3158 }
3159 }
3160
3161 /* base address is now known, determine group base offsets */
3162 for (group = 0; group < ai->nr_groups; group++) {
3163 ai->groups[group].base_offset = areas[group] - base;
3164 }
3165
3166 pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
3167 PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
3168 ai->dyn_size, ai->unit_size);
3169
3170 pcpu_setup_first_chunk(ai, base);
3171 goto out_free;
3172
3173out_free_areas:
3174 for (group = 0; group < ai->nr_groups; group++)
3175 if (areas[group])
3176 pcpu_fc_free(areas[group],
3177 ai->groups[group].nr_units * ai->unit_size);
3178out_free:
3179 pcpu_free_alloc_info(ai);
3180 if (areas)
3181 memblock_free(areas, areas_size);
3182 return rc;
3183}
3184#endif /* BUILD_EMBED_FIRST_CHUNK */
3185
3186#ifdef BUILD_PAGE_FIRST_CHUNK
3187#include <asm/pgalloc.h>
3188
3189#ifndef P4D_TABLE_SIZE
3190#define P4D_TABLE_SIZE PAGE_SIZE
3191#endif
3192
3193#ifndef PUD_TABLE_SIZE
3194#define PUD_TABLE_SIZE PAGE_SIZE
3195#endif
3196
3197#ifndef PMD_TABLE_SIZE
3198#define PMD_TABLE_SIZE PAGE_SIZE
3199#endif
3200
3201#ifndef PTE_TABLE_SIZE
3202#define PTE_TABLE_SIZE PAGE_SIZE
3203#endif
3204void __init __weak pcpu_populate_pte(unsigned long addr)
3205{
3206 pgd_t *pgd = pgd_offset_k(addr);
3207 p4d_t *p4d;
3208 pud_t *pud;
3209 pmd_t *pmd;
3210
3211 if (pgd_none(*pgd)) {
3212 p4d = memblock_alloc(P4D_TABLE_SIZE, P4D_TABLE_SIZE);
3213 if (!p4d)
3214 goto err_alloc;
3215 pgd_populate(&init_mm, pgd, p4d);
3216 }
3217
3218 p4d = p4d_offset(pgd, addr);
3219 if (p4d_none(*p4d)) {
3220 pud = memblock_alloc(PUD_TABLE_SIZE, PUD_TABLE_SIZE);
3221 if (!pud)
3222 goto err_alloc;
3223 p4d_populate(&init_mm, p4d, pud);
3224 }
3225
3226 pud = pud_offset(p4d, addr);
3227 if (pud_none(*pud)) {
3228 pmd = memblock_alloc(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
3229 if (!pmd)
3230 goto err_alloc;
3231 pud_populate(&init_mm, pud, pmd);
3232 }
3233
3234 pmd = pmd_offset(pud, addr);
3235 if (!pmd_present(*pmd)) {
3236 pte_t *new;
3237
3238 new = memblock_alloc(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
3239 if (!new)
3240 goto err_alloc;
3241 pmd_populate_kernel(&init_mm, pmd, new);
3242 }
3243
3244 return;
3245
3246err_alloc:
3247 panic("%s: Failed to allocate memory\n", __func__);
3248}
3249
3250/**
3251 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
3252 * @reserved_size: the size of reserved percpu area in bytes
3253 * @cpu_to_nd_fn: callback to convert cpu to it's node, optional
3254 *
3255 * This is a helper to ease setting up page-remapped first percpu
3256 * chunk and can be called where pcpu_setup_first_chunk() is expected.
3257 *
3258 * This is the basic allocator. Static percpu area is allocated
3259 * page-by-page into vmalloc area.
3260 *
3261 * RETURNS:
3262 * 0 on success, -errno on failure.
3263 */
3264int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
3265{
3266 static struct vm_struct vm;
3267 struct pcpu_alloc_info *ai;
3268 char psize_str[16];
3269 int unit_pages;
3270 size_t pages_size;
3271 struct page **pages;
3272 int unit, i, j, rc = 0;
3273 int upa;
3274 int nr_g0_units;
3275
3276 snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
3277
3278 ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
3279 if (IS_ERR(ai))
3280 return PTR_ERR(ai);
3281 BUG_ON(ai->nr_groups != 1);
3282 upa = ai->alloc_size/ai->unit_size;
3283 nr_g0_units = roundup(num_possible_cpus(), upa);
3284 if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
3285 pcpu_free_alloc_info(ai);
3286 return -EINVAL;
3287 }
3288
3289 unit_pages = ai->unit_size >> PAGE_SHIFT;
3290
3291 /* unaligned allocations can't be freed, round up to page size */
3292 pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
3293 sizeof(pages[0]));
3294 pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
3295 if (!pages)
3296 panic("%s: Failed to allocate %zu bytes\n", __func__,
3297 pages_size);
3298
3299 /* allocate pages */
3300 j = 0;
3301 for (unit = 0; unit < num_possible_cpus(); unit++) {
3302 unsigned int cpu = ai->groups[0].cpu_map[unit];
3303 for (i = 0; i < unit_pages; i++) {
3304 void *ptr;
3305
3306 ptr = pcpu_fc_alloc(cpu, PAGE_SIZE, PAGE_SIZE, cpu_to_nd_fn);
3307 if (!ptr) {
3308 pr_warn("failed to allocate %s page for cpu%u\n",
3309 psize_str, cpu);
3310 goto enomem;
3311 }
3312 /* kmemleak tracks the percpu allocations separately */
3313 kmemleak_ignore_phys(__pa(ptr));
3314 pages[j++] = virt_to_page(ptr);
3315 }
3316 }
3317
3318 /* allocate vm area, map the pages and copy static data */
3319 vm.flags = VM_ALLOC;
3320 vm.size = num_possible_cpus() * ai->unit_size;
3321 vm_area_register_early(&vm, PAGE_SIZE);
3322
3323 for (unit = 0; unit < num_possible_cpus(); unit++) {
3324 unsigned long unit_addr =
3325 (unsigned long)vm.addr + unit * ai->unit_size;
3326
3327 for (i = 0; i < unit_pages; i++)
3328 pcpu_populate_pte(unit_addr + (i << PAGE_SHIFT));
3329
3330 /* pte already populated, the following shouldn't fail */
3331 rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
3332 unit_pages);
3333 if (rc < 0)
3334 panic("failed to map percpu area, err=%d\n", rc);
3335
3336 flush_cache_vmap_early(unit_addr, unit_addr + ai->unit_size);
3337
3338 /* copy static data */
3339 memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
3340 }
3341
3342 /* we're ready, commit */
3343 pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
3344 unit_pages, psize_str, ai->static_size,
3345 ai->reserved_size, ai->dyn_size);
3346
3347 pcpu_setup_first_chunk(ai, vm.addr);
3348 goto out_free_ar;
3349
3350enomem:
3351 while (--j >= 0)
3352 pcpu_fc_free(page_address(pages[j]), PAGE_SIZE);
3353 rc = -ENOMEM;
3354out_free_ar:
3355 memblock_free(pages, pages_size);
3356 pcpu_free_alloc_info(ai);
3357 return rc;
3358}
3359#endif /* BUILD_PAGE_FIRST_CHUNK */
3360
3361#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
3362/*
3363 * Generic SMP percpu area setup.
3364 *
3365 * The embedding helper is used because its behavior closely resembles
3366 * the original non-dynamic generic percpu area setup. This is
3367 * important because many archs have addressing restrictions and might
3368 * fail if the percpu area is located far away from the previous
3369 * location. As an added bonus, in non-NUMA cases, embedding is
3370 * generally a good idea TLB-wise because percpu area can piggy back
3371 * on the physical linear memory mapping which uses large page
3372 * mappings on applicable archs.
3373 */
3374unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
3375EXPORT_SYMBOL(__per_cpu_offset);
3376
3377void __init setup_per_cpu_areas(void)
3378{
3379 unsigned long delta;
3380 unsigned int cpu;
3381 int rc;
3382
3383 /*
3384 * Always reserve area for module percpu variables. That's
3385 * what the legacy allocator did.
3386 */
3387 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE, PERCPU_DYNAMIC_RESERVE,
3388 PAGE_SIZE, NULL, NULL);
3389 if (rc < 0)
3390 panic("Failed to initialize percpu areas.");
3391
3392 delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
3393 for_each_possible_cpu(cpu)
3394 __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
3395}
3396#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
3397
3398#else /* CONFIG_SMP */
3399
3400/*
3401 * UP percpu area setup.
3402 *
3403 * UP always uses km-based percpu allocator with identity mapping.
3404 * Static percpu variables are indistinguishable from the usual static
3405 * variables and don't require any special preparation.
3406 */
3407void __init setup_per_cpu_areas(void)
3408{
3409 const size_t unit_size =
3410 roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
3411 PERCPU_DYNAMIC_RESERVE));
3412 struct pcpu_alloc_info *ai;
3413 void *fc;
3414
3415 ai = pcpu_alloc_alloc_info(1, 1);
3416 fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
3417 if (!ai || !fc)
3418 panic("Failed to allocate memory for percpu areas.");
3419 /* kmemleak tracks the percpu allocations separately */
3420 kmemleak_ignore_phys(__pa(fc));
3421
3422 ai->dyn_size = unit_size;
3423 ai->unit_size = unit_size;
3424 ai->atom_size = unit_size;
3425 ai->alloc_size = unit_size;
3426 ai->groups[0].nr_units = 1;
3427 ai->groups[0].cpu_map[0] = 0;
3428
3429 pcpu_setup_first_chunk(ai, fc);
3430 pcpu_free_alloc_info(ai);
3431}
3432
3433#endif /* CONFIG_SMP */
3434
3435/*
3436 * pcpu_nr_pages - calculate total number of populated backing pages
3437 *
3438 * This reflects the number of pages populated to back chunks. Metadata is
3439 * excluded in the number exposed in meminfo as the number of backing pages
3440 * scales with the number of cpus and can quickly outweigh the memory used for
3441 * metadata. It also keeps this calculation nice and simple.
3442 *
3443 * RETURNS:
3444 * Total number of populated backing pages in use by the allocator.
3445 */
3446unsigned long pcpu_nr_pages(void)
3447{
3448 return pcpu_nr_populated * pcpu_nr_units;
3449}
3450
3451/*
3452 * Percpu allocator is initialized early during boot when neither slab or
3453 * workqueue is available. Plug async management until everything is up
3454 * and running.
3455 */
3456static int __init percpu_enable_async(void)
3457{
3458 pcpu_async_enabled = true;
3459 return 0;
3460}
3461subsys_initcall(percpu_enable_async);