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