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