Branch data 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 <dennisszhou@gmail.com>
10 : : *
11 : : * The percpu allocator handles both static and dynamic areas. Percpu
12 : : * areas are allocated in chunks which are divided into units. There is
13 : : * a 1-to-1 mapping for units to possible cpus. These units are grouped
14 : : * based on NUMA properties of the machine.
15 : : *
16 : : * c0 c1 c2
17 : : * ------------------- ------------------- ------------
18 : : * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
19 : : * ------------------- ...... ------------------- .... ------------
20 : : *
21 : : * Allocation is done by offsets into a unit's address space. Ie., an
22 : : * area of 512 bytes at 6k in c1 occupies 512 bytes at 6k in c1:u0,
23 : : * c1:u1, c1:u2, etc. On NUMA machines, the mapping may be non-linear
24 : : * and even sparse. Access is handled by configuring percpu base
25 : : * registers according to the cpu to unit mappings and offsetting the
26 : : * base address using pcpu_unit_size.
27 : : *
28 : : * There is special consideration for the first chunk which must handle
29 : : * the static percpu variables in the kernel image as allocation services
30 : : * are not online yet. In short, the first chunk is structured like so:
31 : : *
32 : : * <Static | [Reserved] | Dynamic>
33 : : *
34 : : * The static data is copied from the original section managed by the
35 : : * linker. The reserved section, if non-zero, primarily manages static
36 : : * percpu variables from kernel modules. Finally, the dynamic section
37 : : * takes care of normal allocations.
38 : : *
39 : : * The allocator organizes chunks into lists according to free size and
40 : : * tries to allocate from the fullest chunk first. Each chunk is managed
41 : : * by a bitmap with metadata blocks. The allocation map is updated on
42 : : * every allocation and free to reflect the current state while the boundary
43 : : * map is only updated on allocation. Each metadata block contains
44 : : * information to help mitigate the need to iterate over large portions
45 : : * of the bitmap. The reverse mapping from page to chunk is stored in
46 : : * the page's index. Lastly, units are lazily backed and grow in unison.
47 : : *
48 : : * There is a unique conversion that goes on here between bytes and bits.
49 : : * Each bit represents a fragment of size PCPU_MIN_ALLOC_SIZE. The chunk
50 : : * tracks the number of pages it is responsible for in nr_pages. Helper
51 : : * functions are used to convert from between the bytes, bits, and blocks.
52 : : * All hints are managed in bits unless explicitly stated.
53 : : *
54 : : * To use this allocator, arch code should do the following:
55 : : *
56 : : * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
57 : : * regular address to percpu pointer and back if they need to be
58 : : * different from the default
59 : : *
60 : : * - use pcpu_setup_first_chunk() during percpu area initialization to
61 : : * setup the first chunk containing the kernel static percpu area
62 : : */
63 : :
64 : : #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
65 : :
66 : : #include <linux/bitmap.h>
67 : : #include <linux/memblock.h>
68 : : #include <linux/err.h>
69 : : #include <linux/lcm.h>
70 : : #include <linux/list.h>
71 : : #include <linux/log2.h>
72 : : #include <linux/mm.h>
73 : : #include <linux/module.h>
74 : : #include <linux/mutex.h>
75 : : #include <linux/percpu.h>
76 : : #include <linux/pfn.h>
77 : : #include <linux/slab.h>
78 : : #include <linux/spinlock.h>
79 : : #include <linux/vmalloc.h>
80 : : #include <linux/workqueue.h>
81 : : #include <linux/kmemleak.h>
82 : : #include <linux/sched.h>
83 : :
84 : : #include <asm/cacheflush.h>
85 : : #include <asm/sections.h>
86 : : #include <asm/tlbflush.h>
87 : : #include <asm/io.h>
88 : :
89 : : #define CREATE_TRACE_POINTS
90 : : #include <trace/events/percpu.h>
91 : :
92 : : #include "percpu-internal.h"
93 : :
94 : : /* the slots are sorted by free bytes left, 1-31 bytes share the same slot */
95 : : #define PCPU_SLOT_BASE_SHIFT 5
96 : : /* chunks in slots below this are subject to being sidelined on failed alloc */
97 : : #define PCPU_SLOT_FAIL_THRESHOLD 3
98 : :
99 : : #define PCPU_EMPTY_POP_PAGES_LOW 2
100 : : #define PCPU_EMPTY_POP_PAGES_HIGH 4
101 : :
102 : : #ifdef CONFIG_SMP
103 : : /* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
104 : : #ifndef __addr_to_pcpu_ptr
105 : : #define __addr_to_pcpu_ptr(addr) \
106 : : (void __percpu *)((unsigned long)(addr) - \
107 : : (unsigned long)pcpu_base_addr + \
108 : : (unsigned long)__per_cpu_start)
109 : : #endif
110 : : #ifndef __pcpu_ptr_to_addr
111 : : #define __pcpu_ptr_to_addr(ptr) \
112 : : (void __force *)((unsigned long)(ptr) + \
113 : : (unsigned long)pcpu_base_addr - \
114 : : (unsigned long)__per_cpu_start)
115 : : #endif
116 : : #else /* CONFIG_SMP */
117 : : /* on UP, it's always identity mapped */
118 : : #define __addr_to_pcpu_ptr(addr) (void __percpu *)(addr)
119 : : #define __pcpu_ptr_to_addr(ptr) (void __force *)(ptr)
120 : : #endif /* CONFIG_SMP */
121 : :
122 : : static int pcpu_unit_pages __ro_after_init;
123 : : static int pcpu_unit_size __ro_after_init;
124 : : static int pcpu_nr_units __ro_after_init;
125 : : static int pcpu_atom_size __ro_after_init;
126 : : int pcpu_nr_slots __ro_after_init;
127 : : static size_t pcpu_chunk_struct_size __ro_after_init;
128 : :
129 : : /* cpus with the lowest and highest unit addresses */
130 : : static unsigned int pcpu_low_unit_cpu __ro_after_init;
131 : : static unsigned int pcpu_high_unit_cpu __ro_after_init;
132 : :
133 : : /* the address of the first chunk which starts with the kernel static area */
134 : : void *pcpu_base_addr __ro_after_init;
135 : : EXPORT_SYMBOL_GPL(pcpu_base_addr);
136 : :
137 : : static const int *pcpu_unit_map __ro_after_init; /* cpu -> unit */
138 : : const unsigned long *pcpu_unit_offsets __ro_after_init; /* cpu -> unit offset */
139 : :
140 : : /* group information, used for vm allocation */
141 : : static int pcpu_nr_groups __ro_after_init;
142 : : static const unsigned long *pcpu_group_offsets __ro_after_init;
143 : : static const size_t *pcpu_group_sizes __ro_after_init;
144 : :
145 : : /*
146 : : * The first chunk which always exists. Note that unlike other
147 : : * chunks, this one can be allocated and mapped in several different
148 : : * ways and thus often doesn't live in the vmalloc area.
149 : : */
150 : : struct pcpu_chunk *pcpu_first_chunk __ro_after_init;
151 : :
152 : : /*
153 : : * Optional reserved chunk. This chunk reserves part of the first
154 : : * chunk and serves it for reserved allocations. When the reserved
155 : : * region doesn't exist, the following variable is NULL.
156 : : */
157 : : struct pcpu_chunk *pcpu_reserved_chunk __ro_after_init;
158 : :
159 : : DEFINE_SPINLOCK(pcpu_lock); /* all internal data structures */
160 : : static DEFINE_MUTEX(pcpu_alloc_mutex); /* chunk create/destroy, [de]pop, map ext */
161 : :
162 : : struct list_head *pcpu_slot __ro_after_init; /* chunk list slots */
163 : :
164 : : /* chunks which need their map areas extended, protected by pcpu_lock */
165 : : static LIST_HEAD(pcpu_map_extend_chunks);
166 : :
167 : : /*
168 : : * The number of empty populated pages, protected by pcpu_lock. The
169 : : * reserved chunk doesn't contribute to the count.
170 : : */
171 : : int pcpu_nr_empty_pop_pages;
172 : :
173 : : /*
174 : : * The number of populated pages in use by the allocator, protected by
175 : : * pcpu_lock. This number is kept per a unit per chunk (i.e. when a page gets
176 : : * allocated/deallocated, it is allocated/deallocated in all units of a chunk
177 : : * and increments/decrements this count by 1).
178 : : */
179 : : static unsigned long pcpu_nr_populated;
180 : :
181 : : /*
182 : : * Balance work is used to populate or destroy chunks asynchronously. We
183 : : * try to keep the number of populated free pages between
184 : : * PCPU_EMPTY_POP_PAGES_LOW and HIGH for atomic allocations and at most one
185 : : * empty chunk.
186 : : */
187 : : static void pcpu_balance_workfn(struct work_struct *work);
188 : : static DECLARE_WORK(pcpu_balance_work, pcpu_balance_workfn);
189 : : static bool pcpu_async_enabled __read_mostly;
190 : : static bool pcpu_atomic_alloc_failed;
191 : :
192 : 56 : static void pcpu_schedule_balance_work(void)
193 : : {
194 : 56 : if (pcpu_async_enabled)
195 : 56 : schedule_work(&pcpu_balance_work);
196 : : }
197 : :
198 : : /**
199 : : * pcpu_addr_in_chunk - check if the address is served from this chunk
200 : : * @chunk: chunk of interest
201 : : * @addr: percpu address
202 : : *
203 : : * RETURNS:
204 : : * True if the address is served from this chunk.
205 : : */
206 : 3870 : static bool pcpu_addr_in_chunk(struct pcpu_chunk *chunk, void *addr)
207 : : {
208 : 3870 : void *start_addr, *end_addr;
209 : :
210 : 3870 : if (!chunk)
211 : : return false;
212 : :
213 : 3870 : start_addr = chunk->base_addr + chunk->start_offset;
214 : 3870 : end_addr = chunk->base_addr + chunk->nr_pages * PAGE_SIZE -
215 : 3870 : chunk->end_offset;
216 : :
217 : 3870 : return addr >= start_addr && addr < end_addr;
218 : : }
219 : :
220 : 88122 : static int __pcpu_size_to_slot(int size)
221 : : {
222 : 88122 : int highbit = fls(size); /* size is in bytes */
223 : 88094 : return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
224 : : }
225 : :
226 : 88150 : static int pcpu_size_to_slot(int size)
227 : : {
228 : 88150 : if (size == pcpu_unit_size)
229 : 56 : return pcpu_nr_slots - 1;
230 : 88094 : return __pcpu_size_to_slot(size);
231 : : }
232 : :
233 : 60206 : static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
234 : : {
235 : 60206 : const struct pcpu_block_md *chunk_md = &chunk->chunk_md;
236 : :
237 : 120384 : if (chunk->free_bytes < PCPU_MIN_ALLOC_SIZE ||
238 [ + - + - : 60178 : chunk_md->contig_hint == 0)
+ - ]
239 : : return 0;
240 : :
241 [ - + + + : 60178 : return pcpu_size_to_slot(chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
+ + ]
242 : : }
243 : :
244 : : /* set the pointer to a chunk in a page struct */
245 : 168 : static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
246 : : {
247 : 168 : page->index = (unsigned long)pcpu;
248 : : }
249 : :
250 : : /* obtain pointer to a chunk from a page struct */
251 : 1683 : static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
252 : : {
253 : 1683 : return (struct pcpu_chunk *)page->index;
254 : : }
255 : :
256 : 224 : static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
257 : : {
258 : 224 : return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
259 : : }
260 : :
261 : 28420 : static unsigned long pcpu_unit_page_offset(unsigned int cpu, int page_idx)
262 : : {
263 : 28420 : return pcpu_unit_offsets[cpu] + (page_idx << PAGE_SHIFT);
264 : : }
265 : :
266 : 27944 : static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
267 : : unsigned int cpu, int page_idx)
268 : : {
269 : 56 : return (unsigned long)chunk->base_addr +
270 : : pcpu_unit_page_offset(cpu, page_idx);
271 : : }
272 : :
273 : : /*
274 : : * The following are helper functions to help access bitmaps and convert
275 : : * between bitmap offsets to address offsets.
276 : : */
277 : 41678 : static unsigned long *pcpu_index_alloc_map(struct pcpu_chunk *chunk, int index)
278 : : {
279 : 41678 : return chunk->alloc_map +
280 : 41678 : (index * PCPU_BITMAP_BLOCK_BITS / BITS_PER_LONG);
281 : : }
282 : :
283 : 97170 : static unsigned long pcpu_off_to_block_index(int off)
284 : : {
285 : 97170 : return off / PCPU_BITMAP_BLOCK_BITS;
286 : : }
287 : :
288 : 97170 : static unsigned long pcpu_off_to_block_off(int off)
289 : : {
290 : 97170 : return off & (PCPU_BITMAP_BLOCK_BITS - 1);
291 : : }
292 : :
293 : 33190 : static unsigned long pcpu_block_off_to_off(int index, int off)
294 : : {
295 : 33190 : return index * PCPU_BITMAP_BLOCK_BITS + off;
296 : : }
297 : :
298 : : /*
299 : : * pcpu_next_hint - determine which hint to use
300 : : * @block: block of interest
301 : : * @alloc_bits: size of allocation
302 : : *
303 : : * This determines if we should scan based on the scan_hint or first_free.
304 : : * In general, we want to scan from first_free to fulfill allocations by
305 : : * first fit. However, if we know a scan_hint at position scan_hint_start
306 : : * cannot fulfill an allocation, we can begin scanning from there knowing
307 : : * the contig_hint will be our fallback.
308 : : */
309 : 55692 : static int pcpu_next_hint(struct pcpu_block_md *block, int alloc_bits)
310 : : {
311 : : /*
312 : : * The three conditions below determine if we can skip past the
313 : : * scan_hint. First, does the scan hint exist. Second, is the
314 : : * contig_hint after the scan_hint (possibly not true iff
315 : : * contig_hint == scan_hint). Third, is the allocation request
316 : : * larger than the scan_hint.
317 : : */
318 : 72072 : if (block->scan_hint &&
319 [ + + + + : 16380 : block->contig_hint_start > block->scan_hint_start &&
+ + + + ]
320 : : alloc_bits > block->scan_hint)
321 : 12142 : return block->scan_hint_start + block->scan_hint;
322 : :
323 : 43550 : return block->first_free;
324 : : }
325 : :
326 : : /**
327 : : * pcpu_next_md_free_region - finds the next hint free area
328 : : * @chunk: chunk of interest
329 : : * @bit_off: chunk offset
330 : : * @bits: size of free area
331 : : *
332 : : * Helper function for pcpu_for_each_md_free_region. It checks
333 : : * block->contig_hint and performs aggregation across blocks to find the
334 : : * next hint. It modifies bit_off and bits in-place to be consumed in the
335 : : * loop.
336 : : */
337 : 38669 : static void pcpu_next_md_free_region(struct pcpu_chunk *chunk, int *bit_off,
338 : : int *bits)
339 : : {
340 : 38669 : int i = pcpu_off_to_block_index(*bit_off);
341 : 38669 : int block_off = pcpu_off_to_block_off(*bit_off);
342 : 38669 : struct pcpu_block_md *block;
343 : :
344 : 38669 : *bits = 0;
345 [ + + ]: 4337575 : for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
346 : 4298906 : block++, i++) {
347 : : /* handles contig area across blocks */
348 [ + + ]: 4299384 : if (*bits) {
349 : 4277350 : *bits += block->left_free;
350 [ + + ]: 4277350 : if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
351 : 4277210 : continue;
352 : : return;
353 : : }
354 : :
355 : : /*
356 : : * This checks three things. First is there a contig_hint to
357 : : * check. Second, have we checked this hint before by
358 : : * comparing the block_off. Third, is this the same as the
359 : : * right contig hint. In the last case, it spills over into
360 : : * the next block and should be handled by the contig area
361 : : * across blocks code.
362 : : */
363 : 22034 : *bits = block->contig_hint;
364 [ + - + + ]: 22034 : if (*bits && block->contig_hint_start >= block_off &&
365 [ + + ]: 19539 : *bits + block->contig_hint_start < PCPU_BITMAP_BLOCK_BITS) {
366 : 338 : *bit_off = pcpu_block_off_to_off(i,
367 : : block->contig_hint_start);
368 : 338 : return;
369 : : }
370 : : /* reset to satisfy the second predicate above */
371 : 21696 : block_off = 0;
372 : :
373 : 21696 : *bits = block->right_free;
374 : 21696 : *bit_off = (i + 1) * PCPU_BITMAP_BLOCK_BITS - block->right_free;
375 : : }
376 : : }
377 : :
378 : : /**
379 : : * pcpu_next_fit_region - finds fit areas for a given allocation request
380 : : * @chunk: chunk of interest
381 : : * @alloc_bits: size of allocation
382 : : * @align: alignment of area (max PAGE_SIZE)
383 : : * @bit_off: chunk offset
384 : : * @bits: size of free area
385 : : *
386 : : * Finds the next free region that is viable for use with a given size and
387 : : * alignment. This only returns if there is a valid area to be used for this
388 : : * allocation. block->first_free is returned if the allocation request fits
389 : : * within the block to see if the request can be fulfilled prior to the contig
390 : : * hint.
391 : : */
392 : 27888 : static void pcpu_next_fit_region(struct pcpu_chunk *chunk, int alloc_bits,
393 : : int align, int *bit_off, int *bits)
394 : : {
395 : 27888 : int i = pcpu_off_to_block_index(*bit_off);
396 : 27888 : int block_off = pcpu_off_to_block_off(*bit_off);
397 : 27888 : struct pcpu_block_md *block;
398 : :
399 : 27888 : *bits = 0;
400 [ + - ]: 30749 : for (block = chunk->md_blocks + i; i < pcpu_chunk_nr_blocks(chunk);
401 : 2861 : block++, i++) {
402 : : /* handles contig area across blocks */
403 [ + + ]: 30749 : if (*bits) {
404 : 140 : *bits += block->left_free;
405 [ + + ]: 140 : if (*bits >= alloc_bits)
406 : : return;
407 [ - + ]: 56 : if (block->left_free == PCPU_BITMAP_BLOCK_BITS)
408 : 0 : continue;
409 : : }
410 : :
411 : : /* check block->contig_hint */
412 : 30665 : *bits = ALIGN(block->contig_hint_start, align) -
413 : : block->contig_hint_start;
414 : : /*
415 : : * This uses the block offset to determine if this has been
416 : : * checked in the prior iteration.
417 : : */
418 [ + - ]: 30665 : if (block->contig_hint &&
419 [ + + ]: 30665 : block->contig_hint_start >= block_off &&
420 [ + + ]: 28246 : block->contig_hint >= *bits + alloc_bits) {
421 [ + + ]: 27804 : int start = pcpu_next_hint(block, alloc_bits);
422 : :
423 : 27804 : *bits += alloc_bits + block->contig_hint_start -
424 : : start;
425 : 27804 : *bit_off = pcpu_block_off_to_off(i, start);
426 : 27804 : return;
427 : : }
428 : : /* reset to satisfy the second predicate above */
429 : 2861 : block_off = 0;
430 : :
431 : 2861 : *bit_off = ALIGN(PCPU_BITMAP_BLOCK_BITS - block->right_free,
432 : : align);
433 : 2861 : *bits = PCPU_BITMAP_BLOCK_BITS - *bit_off;
434 : 2861 : *bit_off = pcpu_block_off_to_off(i, *bit_off);
435 [ + - ]: 2861 : if (*bits >= alloc_bits)
436 : : return;
437 : : }
438 : :
439 : : /* no valid offsets were found - fail condition */
440 : 0 : *bit_off = pcpu_chunk_map_bits(chunk);
441 : : }
442 : :
443 : : /*
444 : : * Metadata free area iterators. These perform aggregation of free areas
445 : : * based on the metadata blocks and return the offset @bit_off and size in
446 : : * bits of the free area @bits. pcpu_for_each_fit_region only returns when
447 : : * a fit is found for the allocation request.
448 : : */
449 : : #define pcpu_for_each_md_free_region(chunk, bit_off, bits) \
450 : : for (pcpu_next_md_free_region((chunk), &(bit_off), &(bits)); \
451 : : (bit_off) < pcpu_chunk_map_bits((chunk)); \
452 : : (bit_off) += (bits) + 1, \
453 : : pcpu_next_md_free_region((chunk), &(bit_off), &(bits)))
454 : :
455 : : #define pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) \
456 : : for (pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
457 : : &(bits)); \
458 : : (bit_off) < pcpu_chunk_map_bits((chunk)); \
459 : : (bit_off) += (bits), \
460 : : pcpu_next_fit_region((chunk), (alloc_bits), (align), &(bit_off), \
461 : : &(bits)))
462 : :
463 : : /**
464 : : * pcpu_mem_zalloc - allocate memory
465 : : * @size: bytes to allocate
466 : : * @gfp: allocation flags
467 : : *
468 : : * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
469 : : * kzalloc() is used; otherwise, the equivalent of vzalloc() is used.
470 : : * This is to facilitate passing through whitelisted flags. The
471 : : * returned memory is always zeroed.
472 : : *
473 : : * RETURNS:
474 : : * Pointer to the allocated area on success, NULL on failure.
475 : : */
476 : 140 : static void *pcpu_mem_zalloc(size_t size, gfp_t gfp)
477 : : {
478 [ - + + - ]: 140 : if (WARN_ON_ONCE(!slab_is_available()))
479 : : return NULL;
480 : :
481 [ + + ]: 140 : if (size <= PAGE_SIZE)
482 : 56 : return kzalloc(size, gfp);
483 : : else
484 : 84 : return __vmalloc(size, gfp | __GFP_ZERO, PAGE_KERNEL);
485 : : }
486 : :
487 : : /**
488 : : * pcpu_mem_free - free memory
489 : : * @ptr: memory to free
490 : : *
491 : : * Free @ptr. @ptr should have been allocated using pcpu_mem_zalloc().
492 : : */
493 : 0 : static void pcpu_mem_free(void *ptr)
494 : : {
495 : 0 : kvfree(ptr);
496 : 0 : }
497 : :
498 : 415 : static void __pcpu_chunk_move(struct pcpu_chunk *chunk, int slot,
499 : : bool move_front)
500 : : {
501 [ + - ]: 415 : if (chunk != pcpu_reserved_chunk) {
502 [ + + ]: 415 : if (move_front)
503 : 56 : list_move(&chunk->list, &pcpu_slot[slot]);
504 : : else
505 : 359 : list_move_tail(&chunk->list, &pcpu_slot[slot]);
506 : : }
507 : 415 : }
508 : :
509 : 0 : static void pcpu_chunk_move(struct pcpu_chunk *chunk, int slot)
510 : : {
511 : 0 : __pcpu_chunk_move(chunk, slot, true);
512 : 0 : }
513 : :
514 : : /**
515 : : * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
516 : : * @chunk: chunk of interest
517 : : * @oslot: the previous slot it was on
518 : : *
519 : : * This function is called after an allocation or free changed @chunk.
520 : : * New slot according to the changed state is determined and @chunk is
521 : : * moved to the slot. Note that the reserved chunk is never put on
522 : : * chunk slots.
523 : : *
524 : : * CONTEXT:
525 : : * pcpu_lock.
526 : : */
527 : 30131 : static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
528 : : {
529 [ + + ]: 30131 : int nslot = pcpu_chunk_slot(chunk);
530 : :
531 [ + + ]: 30131 : if (oslot != nslot)
532 : 415 : __pcpu_chunk_move(chunk, nslot, oslot < nslot);
533 : 30131 : }
534 : :
535 : : /*
536 : : * pcpu_update_empty_pages - update empty page counters
537 : : * @chunk: chunk of interest
538 : : * @nr: nr of empty pages
539 : : *
540 : : * This is used to keep track of the empty pages now based on the premise
541 : : * a md_block covers a page. The hint update functions recognize if a block
542 : : * is made full or broken to calculate deltas for keeping track of free pages.
543 : : */
544 : 392 : static inline void pcpu_update_empty_pages(struct pcpu_chunk *chunk, int nr)
545 : : {
546 : 392 : chunk->nr_empty_pop_pages += nr;
547 : 392 : if (chunk != pcpu_reserved_chunk)
548 : 392 : pcpu_nr_empty_pop_pages += nr;
549 : : }
550 : :
551 : : /*
552 : : * pcpu_region_overlap - determines if two regions overlap
553 : : * @a: start of first region, inclusive
554 : : * @b: end of first region, exclusive
555 : : * @x: start of second region, inclusive
556 : : * @y: end of second region, exclusive
557 : : *
558 : : * This is used to determine if the hint region [a, b) overlaps with the
559 : : * allocated region [x, y).
560 : : */
561 : 111888 : static inline bool pcpu_region_overlap(int a, int b, int x, int y)
562 : : {
563 : 111888 : return (a < y) && (x < b);
564 : : }
565 : :
566 : : /**
567 : : * pcpu_block_update - updates a block given a free area
568 : : * @block: block of interest
569 : : * @start: start offset in block
570 : : * @end: end offset in block
571 : : *
572 : : * Updates a block given a known free area. The region [start, end) is
573 : : * expected to be the entirety of the free area within a block. Chooses
574 : : * the best starting offset if the contig hints are equal.
575 : : */
576 : 48683 : static void pcpu_block_update(struct pcpu_block_md *block, int start, int end)
577 : : {
578 : 48683 : int contig = end - start;
579 : :
580 : 48683 : block->first_free = min(block->first_free, start);
581 [ + + ]: 48683 : if (start == 0)
582 : 31 : block->left_free = contig;
583 : :
584 [ + + ]: 48683 : if (end == block->nr_bits)
585 : 39053 : block->right_free = contig;
586 : :
587 [ + + ]: 48683 : if (contig > block->contig_hint) {
588 : : /* promote the old contig_hint to be the new scan_hint */
589 [ + + ]: 42724 : if (start > block->contig_hint_start) {
590 [ + + ]: 40300 : if (block->contig_hint > block->scan_hint) {
591 : 14131 : block->scan_hint_start =
592 : : block->contig_hint_start;
593 : 14131 : block->scan_hint = block->contig_hint;
594 [ + + ]: 26169 : } else if (start < block->scan_hint_start) {
595 : : /*
596 : : * The old contig_hint == scan_hint. But, the
597 : : * new contig is larger so hold the invariant
598 : : * scan_hint_start < contig_hint_start.
599 : : */
600 : 127 : block->scan_hint = 0;
601 : : }
602 : : } else {
603 : 2424 : block->scan_hint = 0;
604 : : }
605 : 42724 : block->contig_hint_start = start;
606 : 42724 : block->contig_hint = contig;
607 [ + + ]: 5959 : } else if (contig == block->contig_hint) {
608 [ + - + - ]: 203 : if (block->contig_hint_start &&
609 [ + + ]: 203 : (!start ||
610 [ + + ]: 203 : __ffs(start) > __ffs(block->contig_hint_start))) {
611 : : /* start has a better alignment so use it */
612 : 44 : block->contig_hint_start = start;
613 [ + + ]: 44 : if (start < block->scan_hint_start &&
614 [ + - ]: 13 : block->contig_hint > block->scan_hint)
615 : 13 : block->scan_hint = 0;
616 [ + + ]: 159 : } else if (start > block->scan_hint_start ||
617 [ + + ]: 44 : block->contig_hint > block->scan_hint) {
618 : : /*
619 : : * Knowing contig == contig_hint, update the scan_hint
620 : : * if it is farther than or larger than the current
621 : : * scan_hint.
622 : : */
623 : 158 : block->scan_hint_start = start;
624 : 158 : block->scan_hint = contig;
625 : : }
626 : : } else {
627 : : /*
628 : : * The region is smaller than the contig_hint. So only update
629 : : * the scan_hint if it is larger than or equal and farther than
630 : : * the current scan_hint.
631 : : */
632 [ + + ]: 5756 : if ((start < block->contig_hint_start &&
633 [ + + + + ]: 4119 : (contig > block->scan_hint ||
634 : 406 : (contig == block->scan_hint &&
635 [ + + ]: 406 : start > block->scan_hint_start)))) {
636 : 2393 : block->scan_hint_start = start;
637 : 2393 : block->scan_hint = contig;
638 : : }
639 : : }
640 : 48683 : }
641 : :
642 : : /*
643 : : * pcpu_block_update_scan - update a block given a free area from a scan
644 : : * @chunk: chunk of interest
645 : : * @bit_off: chunk offset
646 : : * @bits: size of free area
647 : : *
648 : : * Finding the final allocation spot first goes through pcpu_find_block_fit()
649 : : * to find a block that can hold the allocation and then pcpu_alloc_area()
650 : : * where a scan is used. When allocations require specific alignments,
651 : : * we can inadvertently create holes which will not be seen in the alloc
652 : : * or free paths.
653 : : *
654 : : * This takes a given free area hole and updates a block as it may change the
655 : : * scan_hint. We need to scan backwards to ensure we don't miss free bits
656 : : * from alignment.
657 : : */
658 : 454 : static void pcpu_block_update_scan(struct pcpu_chunk *chunk, int bit_off,
659 : : int bits)
660 : : {
661 : 454 : int s_off = pcpu_off_to_block_off(bit_off);
662 : 454 : int e_off = s_off + bits;
663 : 454 : int s_index, l_bit;
664 : 454 : struct pcpu_block_md *block;
665 : :
666 [ + - ]: 454 : if (e_off > PCPU_BITMAP_BLOCK_BITS)
667 : : return;
668 : :
669 : 454 : s_index = pcpu_off_to_block_index(bit_off);
670 : 454 : block = chunk->md_blocks + s_index;
671 : :
672 : : /* scan backwards in case of alignment skipping free bits */
673 : 454 : l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index), s_off);
674 [ + - ]: 454 : s_off = (s_off == l_bit) ? 0 : l_bit + 1;
675 : :
676 : 454 : pcpu_block_update(block, s_off, e_off);
677 : : }
678 : :
679 : : /**
680 : : * pcpu_chunk_refresh_hint - updates metadata about a chunk
681 : : * @chunk: chunk of interest
682 : : * @full_scan: if we should scan from the beginning
683 : : *
684 : : * Iterates over the metadata blocks to find the largest contig area.
685 : : * A full scan can be avoided on the allocation path as this is triggered
686 : : * if we broke the contig_hint. In doing so, the scan_hint will be before
687 : : * the contig_hint or after if the scan_hint == contig_hint. This cannot
688 : : * be prevented on freeing as we want to find the largest area possibly
689 : : * spanning blocks.
690 : : */
691 : 19124 : static void pcpu_chunk_refresh_hint(struct pcpu_chunk *chunk, bool full_scan)
692 : : {
693 : 19124 : struct pcpu_block_md *chunk_md = &chunk->chunk_md;
694 : 19124 : int bit_off, bits;
695 : :
696 : : /* promote scan_hint to contig_hint */
697 [ + - + + ]: 19124 : if (!full_scan && chunk_md->scan_hint) {
698 : 2449 : bit_off = chunk_md->scan_hint_start + chunk_md->scan_hint;
699 : 2449 : chunk_md->contig_hint_start = chunk_md->scan_hint_start;
700 : 2449 : chunk_md->contig_hint = chunk_md->scan_hint;
701 : 2449 : chunk_md->scan_hint = 0;
702 : : } else {
703 : 16675 : bit_off = chunk_md->first_free;
704 : 16675 : chunk_md->contig_hint = 0;
705 : : }
706 : :
707 : 19124 : bits = 0;
708 [ + + ]: 38669 : pcpu_for_each_md_free_region(chunk, bit_off, bits)
709 : 19545 : pcpu_block_update(chunk_md, bit_off, bit_off + bits);
710 : 19124 : }
711 : :
712 : : /**
713 : : * pcpu_block_refresh_hint
714 : : * @chunk: chunk of interest
715 : : * @index: index of the metadata block
716 : : *
717 : : * Scans over the block beginning at first_free and updates the block
718 : : * metadata accordingly.
719 : : */
720 : 21394 : static void pcpu_block_refresh_hint(struct pcpu_chunk *chunk, int index)
721 : : {
722 : 21394 : struct pcpu_block_md *block = chunk->md_blocks + index;
723 : 21394 : unsigned long *alloc_map = pcpu_index_alloc_map(chunk, index);
724 : 21394 : unsigned int rs, re, start; /* region start, region end */
725 : :
726 : : /* promote scan_hint to contig_hint */
727 [ + + ]: 21394 : if (block->scan_hint) {
728 : 9763 : start = block->scan_hint_start + block->scan_hint;
729 : 9763 : block->contig_hint_start = block->scan_hint_start;
730 : 9763 : block->contig_hint = block->scan_hint;
731 : 9763 : block->scan_hint = 0;
732 : : } else {
733 : 11631 : start = block->first_free;
734 : 11631 : block->contig_hint = 0;
735 : : }
736 : :
737 : 21394 : block->right_free = 0;
738 : :
739 : : /* iterate over free areas and update the contig hints */
740 [ + + ]: 45704 : bitmap_for_each_clear_region(alloc_map, rs, re, start,
741 : : PCPU_BITMAP_BLOCK_BITS)
742 : 24310 : pcpu_block_update(block, rs, re);
743 : 21394 : }
744 : :
745 : : /**
746 : : * pcpu_block_update_hint_alloc - update hint on allocation path
747 : : * @chunk: chunk of interest
748 : : * @bit_off: chunk offset
749 : : * @bits: size of request
750 : : *
751 : : * Updates metadata for the allocation path. The metadata only has to be
752 : : * refreshed by a full scan iff the chunk's contig hint is broken. Block level
753 : : * scans are required if the block's contig hint is broken.
754 : : */
755 : 27972 : static void pcpu_block_update_hint_alloc(struct pcpu_chunk *chunk, int bit_off,
756 : : int bits)
757 : : {
758 : 27972 : struct pcpu_block_md *chunk_md = &chunk->chunk_md;
759 : 27972 : int nr_empty_pages = 0;
760 : 27972 : struct pcpu_block_md *s_block, *e_block, *block;
761 : 27972 : int s_index, e_index; /* block indexes of the freed allocation */
762 : 27972 : int s_off, e_off; /* block offsets of the freed allocation */
763 : :
764 : : /*
765 : : * Calculate per block offsets.
766 : : * The calculation uses an inclusive range, but the resulting offsets
767 : : * are [start, end). e_index always points to the last block in the
768 : : * range.
769 : : */
770 : 27972 : s_index = pcpu_off_to_block_index(bit_off);
771 : 27972 : e_index = pcpu_off_to_block_index(bit_off + bits - 1);
772 : 27972 : s_off = pcpu_off_to_block_off(bit_off);
773 : 27972 : e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
774 : :
775 : 27972 : s_block = chunk->md_blocks + s_index;
776 : 27972 : e_block = chunk->md_blocks + e_index;
777 : :
778 : : /*
779 : : * Update s_block.
780 : : * block->first_free must be updated if the allocation takes its place.
781 : : * If the allocation breaks the contig_hint, a scan is required to
782 : : * restore this hint.
783 : : */
784 [ + + ]: 27972 : if (s_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
785 : 252 : nr_empty_pages++;
786 : :
787 [ + + ]: 27972 : if (s_off == s_block->first_free)
788 : 15753 : s_block->first_free = find_next_zero_bit(
789 : : pcpu_index_alloc_map(chunk, s_index),
790 : : PCPU_BITMAP_BLOCK_BITS,
791 : 15753 : s_off + bits);
792 : :
793 : 27972 : if (pcpu_region_overlap(s_block->scan_hint_start,
794 [ + + ]: 27972 : s_block->scan_hint_start + s_block->scan_hint,
795 : : s_off,
796 : : s_off + bits))
797 : 1763 : s_block->scan_hint = 0;
798 : :
799 : 27972 : if (pcpu_region_overlap(s_block->contig_hint_start,
800 : 27972 : s_block->contig_hint_start +
801 [ + + ]: 27972 : s_block->contig_hint,
802 : : s_off,
803 : : s_off + bits)) {
804 : : /* block contig hint is broken - scan to fix it */
805 [ + + ]: 21310 : if (!s_off)
806 : 224 : s_block->left_free = 0;
807 : 21310 : pcpu_block_refresh_hint(chunk, s_index);
808 : : } else {
809 : : /* update left and right contig manually */
810 : 6662 : s_block->left_free = min(s_block->left_free, s_off);
811 [ + - ]: 6662 : if (s_index == e_index)
812 : 6662 : s_block->right_free = min_t(int, s_block->right_free,
813 : : PCPU_BITMAP_BLOCK_BITS - e_off);
814 : : else
815 : 0 : s_block->right_free = 0;
816 : : }
817 : :
818 : : /*
819 : : * Update e_block.
820 : : */
821 [ + + ]: 27972 : if (s_index != e_index) {
822 [ + - ]: 84 : if (e_block->contig_hint == PCPU_BITMAP_BLOCK_BITS)
823 : 84 : nr_empty_pages++;
824 : :
825 : : /*
826 : : * When the allocation is across blocks, the end is along
827 : : * the left part of the e_block.
828 : : */
829 : 84 : e_block->first_free = find_next_zero_bit(
830 : : pcpu_index_alloc_map(chunk, e_index),
831 : : PCPU_BITMAP_BLOCK_BITS, e_off);
832 : :
833 [ - + ]: 84 : if (e_off == PCPU_BITMAP_BLOCK_BITS) {
834 : : /* reset the block */
835 : 0 : e_block++;
836 : : } else {
837 [ + - ]: 84 : if (e_off > e_block->scan_hint_start)
838 : 84 : e_block->scan_hint = 0;
839 : :
840 : 84 : e_block->left_free = 0;
841 [ + - ]: 84 : if (e_off > e_block->contig_hint_start) {
842 : : /* contig hint is broken - scan to fix it */
843 : 84 : pcpu_block_refresh_hint(chunk, e_index);
844 : : } else {
845 : 0 : e_block->right_free =
846 : 0 : min_t(int, e_block->right_free,
847 : : PCPU_BITMAP_BLOCK_BITS - e_off);
848 : : }
849 : : }
850 : :
851 : : /* update in-between md_blocks */
852 : 84 : nr_empty_pages += (e_index - s_index - 1);
853 [ - + ]: 84 : for (block = s_block + 1; block < e_block; block++) {
854 : 0 : block->scan_hint = 0;
855 : 0 : block->contig_hint = 0;
856 : 0 : block->left_free = 0;
857 : 0 : block->right_free = 0;
858 : : }
859 : : }
860 : :
861 [ + + ]: 27972 : if (nr_empty_pages)
862 [ + - ]: 336 : pcpu_update_empty_pages(chunk, -nr_empty_pages);
863 : :
864 : 27972 : if (pcpu_region_overlap(chunk_md->scan_hint_start,
865 : 27972 : chunk_md->scan_hint_start +
866 [ + + ]: 27972 : chunk_md->scan_hint,
867 : : bit_off,
868 : : bit_off + bits))
869 : 662 : chunk_md->scan_hint = 0;
870 : :
871 : : /*
872 : : * The only time a full chunk scan is required is if the chunk
873 : : * contig hint is broken. Otherwise, it means a smaller space
874 : : * was used and therefore the chunk contig hint is still correct.
875 : : */
876 : 27972 : if (pcpu_region_overlap(chunk_md->contig_hint_start,
877 : 27972 : chunk_md->contig_hint_start +
878 [ + + ]: 27972 : chunk_md->contig_hint,
879 : : bit_off,
880 : : bit_off + bits))
881 : 19124 : pcpu_chunk_refresh_hint(chunk, false);
882 : 27972 : }
883 : :
884 : : /**
885 : : * pcpu_block_update_hint_free - updates the block hints on the free path
886 : : * @chunk: chunk of interest
887 : : * @bit_off: chunk offset
888 : : * @bits: size of request
889 : : *
890 : : * Updates metadata for the allocation path. This avoids a blind block
891 : : * refresh by making use of the block contig hints. If this fails, it scans
892 : : * forward and backward to determine the extent of the free area. This is
893 : : * capped at the boundary of blocks.
894 : : *
895 : : * A chunk update is triggered if a page becomes free, a block becomes free,
896 : : * or the free spans across blocks. This tradeoff is to minimize iterating
897 : : * over the block metadata to update chunk_md->contig_hint.
898 : : * chunk_md->contig_hint may be off by up to a page, but it will never be more
899 : : * than the available space. If the contig hint is contained in one block, it
900 : : * will be accurate.
901 : : */
902 : 2187 : static void pcpu_block_update_hint_free(struct pcpu_chunk *chunk, int bit_off,
903 : : int bits)
904 : : {
905 : 2187 : int nr_empty_pages = 0;
906 : 2187 : struct pcpu_block_md *s_block, *e_block, *block;
907 : 2187 : int s_index, e_index; /* block indexes of the freed allocation */
908 : 2187 : int s_off, e_off; /* block offsets of the freed allocation */
909 : 2187 : int start, end; /* start and end of the whole free area */
910 : :
911 : : /*
912 : : * Calculate per block offsets.
913 : : * The calculation uses an inclusive range, but the resulting offsets
914 : : * are [start, end). e_index always points to the last block in the
915 : : * range.
916 : : */
917 : 2187 : s_index = pcpu_off_to_block_index(bit_off);
918 : 2187 : e_index = pcpu_off_to_block_index(bit_off + bits - 1);
919 : 2187 : s_off = pcpu_off_to_block_off(bit_off);
920 : 2187 : e_off = pcpu_off_to_block_off(bit_off + bits - 1) + 1;
921 : :
922 : 2187 : s_block = chunk->md_blocks + s_index;
923 : 2187 : e_block = chunk->md_blocks + e_index;
924 : :
925 : : /*
926 : : * Check if the freed area aligns with the block->contig_hint.
927 : : * If it does, then the scan to find the beginning/end of the
928 : : * larger free area can be avoided.
929 : : *
930 : : * start and end refer to beginning and end of the free area
931 : : * within each their respective blocks. This is not necessarily
932 : : * the entire free area as it may span blocks past the beginning
933 : : * or end of the block.
934 : : */
935 : 2187 : start = s_off;
936 [ + + ]: 2187 : if (s_off == s_block->contig_hint + s_block->contig_hint_start) {
937 : : start = s_block->contig_hint_start;
938 : : } else {
939 : : /*
940 : : * Scan backwards to find the extent of the free area.
941 : : * find_last_bit returns the starting bit, so if the start bit
942 : : * is returned, that means there was no last bit and the
943 : : * remainder of the chunk is free.
944 : : */
945 : 2021 : int l_bit = find_last_bit(pcpu_index_alloc_map(chunk, s_index),
946 : : start);
947 [ + + ]: 2021 : start = (start == l_bit) ? 0 : l_bit + 1;
948 : : }
949 : :
950 : 2187 : end = e_off;
951 [ + + ]: 2187 : if (e_off == e_block->contig_hint_start)
952 : 215 : end = e_block->contig_hint_start + e_block->contig_hint;
953 : : else
954 : 1972 : end = find_next_bit(pcpu_index_alloc_map(chunk, e_index),
955 : : PCPU_BITMAP_BLOCK_BITS, end);
956 : :
957 : : /* update s_block */
958 [ - + ]: 2187 : e_off = (s_index == e_index) ? end : PCPU_BITMAP_BLOCK_BITS;
959 [ - + ]: 2187 : if (!start && e_off == PCPU_BITMAP_BLOCK_BITS)
960 : 0 : nr_empty_pages++;
961 : 2187 : pcpu_block_update(s_block, start, e_off);
962 : :
963 : : /* freeing in the same block */
964 [ - + ]: 2187 : if (s_index != e_index) {
965 : : /* update e_block */
966 [ # # ]: 0 : if (end == PCPU_BITMAP_BLOCK_BITS)
967 : 0 : nr_empty_pages++;
968 : 0 : pcpu_block_update(e_block, 0, end);
969 : :
970 : : /* reset md_blocks in the middle */
971 : 0 : nr_empty_pages += (e_index - s_index - 1);
972 [ # # ]: 0 : for (block = s_block + 1; block < e_block; block++) {
973 : 0 : block->first_free = 0;
974 : 0 : block->scan_hint = 0;
975 : 0 : block->contig_hint_start = 0;
976 : 0 : block->contig_hint = PCPU_BITMAP_BLOCK_BITS;
977 : 0 : block->left_free = PCPU_BITMAP_BLOCK_BITS;
978 : 0 : block->right_free = PCPU_BITMAP_BLOCK_BITS;
979 : : }
980 : : }
981 : :
982 [ - + ]: 2187 : if (nr_empty_pages)
983 [ # # ]: 0 : pcpu_update_empty_pages(chunk, nr_empty_pages);
984 : :
985 : : /*
986 : : * Refresh chunk metadata when the free makes a block free or spans
987 : : * across blocks. The contig_hint may be off by up to a page, but if
988 : : * the contig_hint is contained in a block, it will be accurate with
989 : : * the else condition below.
990 : : */
991 [ + - - + ]: 2187 : if (((end - start) >= PCPU_BITMAP_BLOCK_BITS) || s_index != e_index)
992 : 0 : pcpu_chunk_refresh_hint(chunk, true);
993 : : else
994 : 2187 : pcpu_block_update(&chunk->chunk_md,
995 : : pcpu_block_off_to_off(s_index, start),
996 : : end);
997 : 2187 : }
998 : :
999 : : /**
1000 : : * pcpu_is_populated - determines if the region is populated
1001 : : * @chunk: chunk of interest
1002 : : * @bit_off: chunk offset
1003 : : * @bits: size of area
1004 : : * @next_off: return value for the next offset to start searching
1005 : : *
1006 : : * For atomic allocations, check if the backing pages are populated.
1007 : : *
1008 : : * RETURNS:
1009 : : * Bool if the backing pages are populated.
1010 : : * next_index is to skip over unpopulated blocks in pcpu_find_block_fit.
1011 : : */
1012 : 56 : static bool pcpu_is_populated(struct pcpu_chunk *chunk, int bit_off, int bits,
1013 : : int *next_off)
1014 : : {
1015 : 56 : unsigned int page_start, page_end, rs, re;
1016 : :
1017 : 56 : page_start = PFN_DOWN(bit_off * PCPU_MIN_ALLOC_SIZE);
1018 : 56 : page_end = PFN_UP((bit_off + bits) * PCPU_MIN_ALLOC_SIZE);
1019 : :
1020 : 56 : rs = page_start;
1021 : 56 : bitmap_next_clear_region(chunk->populated, &rs, &re, page_end);
1022 [ - + ]: 56 : if (rs >= page_end)
1023 : : return true;
1024 : :
1025 : 0 : *next_off = re * PAGE_SIZE / PCPU_MIN_ALLOC_SIZE;
1026 : 0 : return false;
1027 : : }
1028 : :
1029 : : /**
1030 : : * pcpu_find_block_fit - finds the block index to start searching
1031 : : * @chunk: chunk of interest
1032 : : * @alloc_bits: size of request in allocation units
1033 : : * @align: alignment of area (max PAGE_SIZE bytes)
1034 : : * @pop_only: use populated regions only
1035 : : *
1036 : : * Given a chunk and an allocation spec, find the offset to begin searching
1037 : : * for a free region. This iterates over the bitmap metadata blocks to
1038 : : * find an offset that will be guaranteed to fit the requirements. It is
1039 : : * not quite first fit as if the allocation does not fit in the contig hint
1040 : : * of a block or chunk, it is skipped. This errs on the side of caution
1041 : : * to prevent excess iteration. Poor alignment can cause the allocator to
1042 : : * skip over blocks and chunks that have valid free areas.
1043 : : *
1044 : : * RETURNS:
1045 : : * The offset in the bitmap to begin searching.
1046 : : * -1 if no offset is found.
1047 : : */
1048 : 27888 : static int pcpu_find_block_fit(struct pcpu_chunk *chunk, int alloc_bits,
1049 : : size_t align, bool pop_only)
1050 : : {
1051 : 27888 : struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1052 : 27888 : int bit_off, bits, next_off;
1053 : :
1054 : : /*
1055 : : * Check to see if the allocation can fit in the chunk's contig hint.
1056 : : * This is an optimization to prevent scanning by assuming if it
1057 : : * cannot fit in the global hint, there is memory pressure and creating
1058 : : * a new chunk would happen soon.
1059 : : */
1060 : 27888 : bit_off = ALIGN(chunk_md->contig_hint_start, align) -
1061 : : chunk_md->contig_hint_start;
1062 [ + - ]: 27888 : if (bit_off + alloc_bits > chunk_md->contig_hint)
1063 : : return -1;
1064 : :
1065 [ + + ]: 27888 : bit_off = pcpu_next_hint(chunk_md, alloc_bits);
1066 : 27888 : bits = 0;
1067 [ + - ]: 27888 : pcpu_for_each_fit_region(chunk, alloc_bits, align, bit_off, bits) {
1068 [ + + - + ]: 27888 : if (!pop_only || pcpu_is_populated(chunk, bit_off, bits,
1069 : : &next_off))
1070 : : break;
1071 : :
1072 : 0 : bit_off = next_off;
1073 : 0 : bits = 0;
1074 : : }
1075 : :
1076 [ - + ]: 27888 : if (bit_off == pcpu_chunk_map_bits(chunk))
1077 : 0 : return -1;
1078 : :
1079 : : return bit_off;
1080 : : }
1081 : :
1082 : : /*
1083 : : * pcpu_find_zero_area - modified from bitmap_find_next_zero_area_off()
1084 : : * @map: the address to base the search on
1085 : : * @size: the bitmap size in bits
1086 : : * @start: the bitnumber to start searching at
1087 : : * @nr: the number of zeroed bits we're looking for
1088 : : * @align_mask: alignment mask for zero area
1089 : : * @largest_off: offset of the largest area skipped
1090 : : * @largest_bits: size of the largest area skipped
1091 : : *
1092 : : * The @align_mask should be one less than a power of 2.
1093 : : *
1094 : : * This is a modified version of bitmap_find_next_zero_area_off() to remember
1095 : : * the largest area that was skipped. This is imperfect, but in general is
1096 : : * good enough. The largest remembered region is the largest failed region
1097 : : * seen. This does not include anything we possibly skipped due to alignment.
1098 : : * pcpu_block_update_scan() does scan backwards to try and recover what was
1099 : : * lost to alignment. While this can cause scanning to miss earlier possible
1100 : : * free areas, smaller allocations will eventually fill those holes.
1101 : : */
1102 : 27888 : static unsigned long pcpu_find_zero_area(unsigned long *map,
1103 : : unsigned long size,
1104 : : unsigned long start,
1105 : : unsigned long nr,
1106 : : unsigned long align_mask,
1107 : : unsigned long *largest_off,
1108 : : unsigned long *largest_bits)
1109 : : {
1110 : 31183 : unsigned long index, end, i, area_off, area_bits;
1111 : 31183 : again:
1112 : 31183 : index = find_next_zero_bit(map, size, start);
1113 : :
1114 : : /* Align allocation */
1115 : 31183 : index = __ALIGN_MASK(index, align_mask);
1116 : 31183 : area_off = index;
1117 : :
1118 : 31183 : end = index + nr;
1119 [ - + ]: 31183 : if (end > size)
1120 : 0 : return end;
1121 : 31183 : i = find_next_bit(map, end, index);
1122 [ + + ]: 31183 : if (i < end) {
1123 : 3295 : area_bits = i - area_off;
1124 : : /* remember largest unused area with best alignment */
1125 [ + + + + ]: 3295 : if (area_bits > *largest_bits ||
1126 [ + + + - ]: 2713 : (area_bits == *largest_bits && *largest_off &&
1127 [ + + ]: 47 : (!area_off || __ffs(area_off) > __ffs(*largest_off)))) {
1128 : 462 : *largest_off = area_off;
1129 : 462 : *largest_bits = area_bits;
1130 : : }
1131 : :
1132 : 3295 : start = i + 1;
1133 : 3295 : goto again;
1134 : : }
1135 : : return index;
1136 : : }
1137 : :
1138 : : /**
1139 : : * pcpu_alloc_area - allocates an area from a pcpu_chunk
1140 : : * @chunk: chunk of interest
1141 : : * @alloc_bits: size of request in allocation units
1142 : : * @align: alignment of area (max PAGE_SIZE)
1143 : : * @start: bit_off to start searching
1144 : : *
1145 : : * This function takes in a @start offset to begin searching to fit an
1146 : : * allocation of @alloc_bits with alignment @align. It needs to scan
1147 : : * the allocation map because if it fits within the block's contig hint,
1148 : : * @start will be block->first_free. This is an attempt to fill the
1149 : : * allocation prior to breaking the contig hint. The allocation and
1150 : : * boundary maps are updated accordingly if it confirms a valid
1151 : : * free area.
1152 : : *
1153 : : * RETURNS:
1154 : : * Allocated addr offset in @chunk on success.
1155 : : * -1 if no matching area is found.
1156 : : */
1157 : 27888 : static int pcpu_alloc_area(struct pcpu_chunk *chunk, int alloc_bits,
1158 : : size_t align, int start)
1159 : : {
1160 : 27888 : struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1161 [ + - ]: 27888 : size_t align_mask = (align) ? (align - 1) : 0;
1162 : 27888 : unsigned long area_off = 0, area_bits = 0;
1163 : 27888 : int bit_off, end, oslot;
1164 : :
1165 : 27888 : lockdep_assert_held(&pcpu_lock);
1166 : :
1167 [ + - ]: 27888 : oslot = pcpu_chunk_slot(chunk);
1168 : :
1169 : : /*
1170 : : * Search to find a fit.
1171 : : */
1172 : 27888 : end = min_t(int, start + alloc_bits + PCPU_BITMAP_BLOCK_BITS,
1173 : : pcpu_chunk_map_bits(chunk));
1174 : 27888 : bit_off = pcpu_find_zero_area(chunk->alloc_map, end, start, alloc_bits,
1175 : : align_mask, &area_off, &area_bits);
1176 [ + - ]: 27888 : if (bit_off >= end)
1177 : : return -1;
1178 : :
1179 [ + + ]: 27888 : if (area_bits)
1180 : 454 : pcpu_block_update_scan(chunk, area_off, area_bits);
1181 : :
1182 : : /* update alloc map */
1183 [ - + ]: 27888 : bitmap_set(chunk->alloc_map, bit_off, alloc_bits);
1184 : :
1185 : : /* update boundary map */
1186 : 27888 : set_bit(bit_off, chunk->bound_map);
1187 [ - + ]: 27888 : bitmap_clear(chunk->bound_map, bit_off + 1, alloc_bits - 1);
1188 : 27888 : set_bit(bit_off + alloc_bits, chunk->bound_map);
1189 : :
1190 : 27888 : chunk->free_bytes -= alloc_bits * PCPU_MIN_ALLOC_SIZE;
1191 : :
1192 : : /* update first free bit */
1193 [ + + ]: 27888 : if (bit_off == chunk_md->first_free)
1194 : 13204 : chunk_md->first_free = find_next_zero_bit(
1195 : 13204 : chunk->alloc_map,
1196 : : pcpu_chunk_map_bits(chunk),
1197 : : bit_off + alloc_bits);
1198 : :
1199 : 27888 : pcpu_block_update_hint_alloc(chunk, bit_off, alloc_bits);
1200 : :
1201 : 27888 : pcpu_chunk_relocate(chunk, oslot);
1202 : :
1203 : 27888 : return bit_off * PCPU_MIN_ALLOC_SIZE;
1204 : : }
1205 : :
1206 : : /**
1207 : : * pcpu_free_area - frees the corresponding offset
1208 : : * @chunk: chunk of interest
1209 : : * @off: addr offset into chunk
1210 : : *
1211 : : * This function determines the size of an allocation to free using
1212 : : * the boundary bitmap and clears the allocation map.
1213 : : */
1214 : 2187 : static void pcpu_free_area(struct pcpu_chunk *chunk, int off)
1215 : : {
1216 : 2187 : struct pcpu_block_md *chunk_md = &chunk->chunk_md;
1217 : 2187 : int bit_off, bits, end, oslot;
1218 : :
1219 : 2187 : lockdep_assert_held(&pcpu_lock);
1220 [ + - ]: 2187 : pcpu_stats_area_dealloc(chunk);
1221 : :
1222 [ + - ]: 2187 : oslot = pcpu_chunk_slot(chunk);
1223 : :
1224 : 2187 : bit_off = off / PCPU_MIN_ALLOC_SIZE;
1225 : :
1226 : : /* find end index */
1227 : 4374 : end = find_next_bit(chunk->bound_map, pcpu_chunk_map_bits(chunk),
1228 : 2187 : bit_off + 1);
1229 : 2187 : bits = end - bit_off;
1230 [ - + ]: 2187 : bitmap_clear(chunk->alloc_map, bit_off, bits);
1231 : :
1232 : : /* update metadata */
1233 : 2187 : chunk->free_bytes += bits * PCPU_MIN_ALLOC_SIZE;
1234 : :
1235 : : /* update first free bit */
1236 : 2187 : chunk_md->first_free = min(chunk_md->first_free, bit_off);
1237 : :
1238 : 2187 : pcpu_block_update_hint_free(chunk, bit_off, bits);
1239 : :
1240 : 2187 : pcpu_chunk_relocate(chunk, oslot);
1241 : 2187 : }
1242 : :
1243 : 14728 : static void pcpu_init_md_block(struct pcpu_block_md *block, int nr_bits)
1244 : : {
1245 : 14728 : block->scan_hint = 0;
1246 : 14728 : block->contig_hint = nr_bits;
1247 : 14728 : block->left_free = nr_bits;
1248 : 14728 : block->right_free = nr_bits;
1249 : 14728 : block->first_free = 0;
1250 : 14728 : block->nr_bits = nr_bits;
1251 : : }
1252 : :
1253 : 84 : static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
1254 : : {
1255 : 84 : struct pcpu_block_md *md_block;
1256 : :
1257 : : /* init the chunk's block */
1258 : 84 : pcpu_init_md_block(&chunk->chunk_md, pcpu_chunk_map_bits(chunk));
1259 : :
1260 [ + + ]: 14728 : for (md_block = chunk->md_blocks;
1261 [ + + ]: 14728 : md_block != chunk->md_blocks + pcpu_chunk_nr_blocks(chunk);
1262 : 14644 : md_block++)
1263 : 14644 : pcpu_init_md_block(md_block, PCPU_BITMAP_BLOCK_BITS);
1264 : 84 : }
1265 : :
1266 : : /**
1267 : : * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
1268 : : * @tmp_addr: the start of the region served
1269 : : * @map_size: size of the region served
1270 : : *
1271 : : * This is responsible for creating the chunks that serve the first chunk. The
1272 : : * base_addr is page aligned down of @tmp_addr while the region end is page
1273 : : * aligned up. Offsets are kept track of to determine the region served. All
1274 : : * this is done to appease the bitmap allocator in avoiding partial blocks.
1275 : : *
1276 : : * RETURNS:
1277 : : * Chunk serving the region at @tmp_addr of @map_size.
1278 : : */
1279 : 56 : static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
1280 : : int map_size)
1281 : : {
1282 : 56 : struct pcpu_chunk *chunk;
1283 : 56 : unsigned long aligned_addr, lcm_align;
1284 : 56 : int start_offset, offset_bits, region_size, region_bits;
1285 : 56 : size_t alloc_size;
1286 : :
1287 : : /* region calculations */
1288 : 56 : aligned_addr = tmp_addr & PAGE_MASK;
1289 : :
1290 : 56 : start_offset = tmp_addr - aligned_addr;
1291 : :
1292 : : /*
1293 : : * Align the end of the region with the LCM of PAGE_SIZE and
1294 : : * PCPU_BITMAP_BLOCK_SIZE. One of these constants is a multiple of
1295 : : * the other.
1296 : : */
1297 : 56 : lcm_align = lcm(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE);
1298 : 56 : region_size = ALIGN(start_offset + map_size, lcm_align);
1299 : :
1300 : : /* allocate chunk */
1301 : 56 : alloc_size = sizeof(struct pcpu_chunk) +
1302 : 56 : BITS_TO_LONGS(region_size >> PAGE_SHIFT);
1303 : 56 : chunk = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1304 [ - + ]: 56 : if (!chunk)
1305 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
1306 : : alloc_size);
1307 : :
1308 : 56 : INIT_LIST_HEAD(&chunk->list);
1309 : :
1310 : 56 : chunk->base_addr = (void *)aligned_addr;
1311 : 56 : chunk->start_offset = start_offset;
1312 : 56 : chunk->end_offset = region_size - chunk->start_offset - map_size;
1313 : :
1314 : 56 : chunk->nr_pages = region_size >> PAGE_SHIFT;
1315 : 56 : region_bits = pcpu_chunk_map_bits(chunk);
1316 : :
1317 : 56 : alloc_size = BITS_TO_LONGS(region_bits) * sizeof(chunk->alloc_map[0]);
1318 : 56 : chunk->alloc_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1319 [ - + ]: 56 : if (!chunk->alloc_map)
1320 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
1321 : : alloc_size);
1322 : :
1323 : 56 : alloc_size =
1324 : 56 : BITS_TO_LONGS(region_bits + 1) * sizeof(chunk->bound_map[0]);
1325 : 56 : chunk->bound_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1326 [ - + ]: 56 : if (!chunk->bound_map)
1327 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
1328 : : alloc_size);
1329 : :
1330 : 56 : alloc_size = pcpu_chunk_nr_blocks(chunk) * sizeof(chunk->md_blocks[0]);
1331 : 56 : chunk->md_blocks = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
1332 [ - + ]: 56 : if (!chunk->md_blocks)
1333 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
1334 : : alloc_size);
1335 : :
1336 : 56 : pcpu_init_md_blocks(chunk);
1337 : :
1338 : : /* manage populated page bitmap */
1339 : 56 : chunk->immutable = true;
1340 [ + - ]: 56 : bitmap_fill(chunk->populated, chunk->nr_pages);
1341 : 56 : chunk->nr_populated = chunk->nr_pages;
1342 : 56 : chunk->nr_empty_pop_pages = chunk->nr_pages;
1343 : :
1344 : 56 : chunk->free_bytes = map_size;
1345 : :
1346 [ + - ]: 56 : if (chunk->start_offset) {
1347 : : /* hide the beginning of the bitmap */
1348 : 56 : offset_bits = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
1349 [ - + ]: 56 : bitmap_set(chunk->alloc_map, 0, offset_bits);
1350 : 56 : set_bit(0, chunk->bound_map);
1351 : 56 : set_bit(offset_bits, chunk->bound_map);
1352 : :
1353 : 56 : chunk->chunk_md.first_free = offset_bits;
1354 : :
1355 : 56 : pcpu_block_update_hint_alloc(chunk, 0, offset_bits);
1356 : : }
1357 : :
1358 [ + + ]: 56 : if (chunk->end_offset) {
1359 : : /* hide the end of the bitmap */
1360 : 28 : offset_bits = chunk->end_offset / PCPU_MIN_ALLOC_SIZE;
1361 [ - + ]: 28 : bitmap_set(chunk->alloc_map,
1362 [ - + ]: 28 : pcpu_chunk_map_bits(chunk) - offset_bits,
1363 : : offset_bits);
1364 : 28 : set_bit((start_offset + map_size) / PCPU_MIN_ALLOC_SIZE,
1365 : 28 : chunk->bound_map);
1366 : 28 : set_bit(region_bits, chunk->bound_map);
1367 : :
1368 : 28 : pcpu_block_update_hint_alloc(chunk, pcpu_chunk_map_bits(chunk)
1369 : : - offset_bits, offset_bits);
1370 : : }
1371 : :
1372 : 56 : return chunk;
1373 : : }
1374 : :
1375 : 28 : static struct pcpu_chunk *pcpu_alloc_chunk(gfp_t gfp)
1376 : : {
1377 : 28 : struct pcpu_chunk *chunk;
1378 : 28 : int region_bits;
1379 : :
1380 : 28 : chunk = pcpu_mem_zalloc(pcpu_chunk_struct_size, gfp);
1381 [ + - ]: 28 : if (!chunk)
1382 : : return NULL;
1383 : :
1384 : 28 : INIT_LIST_HEAD(&chunk->list);
1385 : 28 : chunk->nr_pages = pcpu_unit_pages;
1386 : 28 : region_bits = pcpu_chunk_map_bits(chunk);
1387 : :
1388 : 28 : chunk->alloc_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits) *
1389 : : sizeof(chunk->alloc_map[0]), gfp);
1390 [ - + ]: 28 : if (!chunk->alloc_map)
1391 : 0 : goto alloc_map_fail;
1392 : :
1393 : 28 : chunk->bound_map = pcpu_mem_zalloc(BITS_TO_LONGS(region_bits + 1) *
1394 : : sizeof(chunk->bound_map[0]), gfp);
1395 [ - + ]: 28 : if (!chunk->bound_map)
1396 : 0 : goto bound_map_fail;
1397 : :
1398 : 28 : chunk->md_blocks = pcpu_mem_zalloc(pcpu_chunk_nr_blocks(chunk) *
1399 : : sizeof(chunk->md_blocks[0]), gfp);
1400 [ - + ]: 28 : if (!chunk->md_blocks)
1401 : 0 : goto md_blocks_fail;
1402 : :
1403 : 28 : pcpu_init_md_blocks(chunk);
1404 : :
1405 : : /* init metadata */
1406 : 28 : chunk->free_bytes = chunk->nr_pages * PAGE_SIZE;
1407 : :
1408 : 28 : return chunk;
1409 : :
1410 : : md_blocks_fail:
1411 : 0 : pcpu_mem_free(chunk->bound_map);
1412 : 0 : bound_map_fail:
1413 : 0 : pcpu_mem_free(chunk->alloc_map);
1414 : 0 : alloc_map_fail:
1415 : 0 : pcpu_mem_free(chunk);
1416 : :
1417 : 0 : return NULL;
1418 : : }
1419 : :
1420 : 0 : static void pcpu_free_chunk(struct pcpu_chunk *chunk)
1421 : : {
1422 [ # # ]: 0 : if (!chunk)
1423 : : return;
1424 : 0 : pcpu_mem_free(chunk->md_blocks);
1425 : 0 : pcpu_mem_free(chunk->bound_map);
1426 : 0 : pcpu_mem_free(chunk->alloc_map);
1427 : 0 : pcpu_mem_free(chunk);
1428 : : }
1429 : :
1430 : : /**
1431 : : * pcpu_chunk_populated - post-population bookkeeping
1432 : : * @chunk: pcpu_chunk which got populated
1433 : : * @page_start: the start page
1434 : : * @page_end: the end page
1435 : : *
1436 : : * Pages in [@page_start,@page_end) have been populated to @chunk. Update
1437 : : * the bookkeeping information accordingly. Must be called after each
1438 : : * successful population.
1439 : : *
1440 : : * If this is @for_alloc, do not increment pcpu_nr_empty_pop_pages because it
1441 : : * is to serve an allocation in that area.
1442 : : */
1443 : 56 : static void pcpu_chunk_populated(struct pcpu_chunk *chunk, int page_start,
1444 : : int page_end)
1445 : : {
1446 : 56 : int nr = page_end - page_start;
1447 : :
1448 : 56 : lockdep_assert_held(&pcpu_lock);
1449 : :
1450 [ - + ]: 56 : bitmap_set(chunk->populated, page_start, nr);
1451 : 56 : chunk->nr_populated += nr;
1452 : 56 : pcpu_nr_populated += nr;
1453 : :
1454 [ + - ]: 56 : pcpu_update_empty_pages(chunk, nr);
1455 : 56 : }
1456 : :
1457 : : /**
1458 : : * pcpu_chunk_depopulated - post-depopulation bookkeeping
1459 : : * @chunk: pcpu_chunk which got depopulated
1460 : : * @page_start: the start page
1461 : : * @page_end: the end page
1462 : : *
1463 : : * Pages in [@page_start,@page_end) have been depopulated from @chunk.
1464 : : * Update the bookkeeping information accordingly. Must be called after
1465 : : * each successful depopulation.
1466 : : */
1467 : 0 : static void pcpu_chunk_depopulated(struct pcpu_chunk *chunk,
1468 : : int page_start, int page_end)
1469 : : {
1470 : 0 : int nr = page_end - page_start;
1471 : :
1472 : 0 : lockdep_assert_held(&pcpu_lock);
1473 : :
1474 [ # # ]: 0 : bitmap_clear(chunk->populated, page_start, nr);
1475 : 0 : chunk->nr_populated -= nr;
1476 : 0 : pcpu_nr_populated -= nr;
1477 : :
1478 [ # # ]: 0 : pcpu_update_empty_pages(chunk, -nr);
1479 : 0 : }
1480 : :
1481 : : /*
1482 : : * Chunk management implementation.
1483 : : *
1484 : : * To allow different implementations, chunk alloc/free and
1485 : : * [de]population are implemented in a separate file which is pulled
1486 : : * into this file and compiled together. The following functions
1487 : : * should be implemented.
1488 : : *
1489 : : * pcpu_populate_chunk - populate the specified range of a chunk
1490 : : * pcpu_depopulate_chunk - depopulate the specified range of a chunk
1491 : : * pcpu_create_chunk - create a new chunk
1492 : : * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
1493 : : * pcpu_addr_to_page - translate address to physical address
1494 : : * pcpu_verify_alloc_info - check alloc_info is acceptable during init
1495 : : */
1496 : : static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
1497 : : int page_start, int page_end, gfp_t gfp);
1498 : : static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
1499 : : int page_start, int page_end);
1500 : : static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp);
1501 : : static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
1502 : : static struct page *pcpu_addr_to_page(void *addr);
1503 : : static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
1504 : :
1505 : : #ifdef CONFIG_NEED_PER_CPU_KM
1506 : : #include "percpu-km.c"
1507 : : #else
1508 : : #include "percpu-vm.c"
1509 : : #endif
1510 : :
1511 : : /**
1512 : : * pcpu_chunk_addr_search - determine chunk containing specified address
1513 : : * @addr: address for which the chunk needs to be determined.
1514 : : *
1515 : : * This is an internal function that handles all but static allocations.
1516 : : * Static percpu address values should never be passed into the allocator.
1517 : : *
1518 : : * RETURNS:
1519 : : * The address of the found chunk.
1520 : : */
1521 : 2187 : static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
1522 : : {
1523 : : /* is it in the dynamic region (first chunk)? */
1524 [ + - + + ]: 2187 : if (pcpu_addr_in_chunk(pcpu_first_chunk, addr))
1525 : : return pcpu_first_chunk;
1526 : :
1527 : : /* is it in the reserved region? */
1528 [ + - + - ]: 1683 : if (pcpu_addr_in_chunk(pcpu_reserved_chunk, addr))
1529 : : return pcpu_reserved_chunk;
1530 : :
1531 : : /*
1532 : : * The address is relative to unit0 which might be unused and
1533 : : * thus unmapped. Offset the address to the unit space of the
1534 : : * current processor before looking it up in the vmalloc
1535 : : * space. Note that any possible cpu id can be used here, so
1536 : : * there's no need to worry about preemption or cpu hotplug.
1537 : : */
1538 : 1683 : addr += pcpu_unit_offsets[raw_smp_processor_id()];
1539 : 1683 : return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
1540 : : }
1541 : :
1542 : : /**
1543 : : * pcpu_alloc - the percpu allocator
1544 : : * @size: size of area to allocate in bytes
1545 : : * @align: alignment of area (max PAGE_SIZE)
1546 : : * @reserved: allocate from the reserved chunk if available
1547 : : * @gfp: allocation flags
1548 : : *
1549 : : * Allocate percpu area of @size bytes aligned at @align. If @gfp doesn't
1550 : : * contain %GFP_KERNEL, the allocation is atomic. If @gfp has __GFP_NOWARN
1551 : : * then no warning will be triggered on invalid or failed allocation
1552 : : * requests.
1553 : : *
1554 : : * RETURNS:
1555 : : * Percpu pointer to the allocated area on success, NULL on failure.
1556 : : */
1557 : 27888 : static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved,
1558 : : gfp_t gfp)
1559 : : {
1560 : : /* whitelisted flags that can be passed to the backing allocators */
1561 : 27888 : gfp_t pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
1562 : 27888 : bool is_atomic = (gfp & GFP_KERNEL) != GFP_KERNEL;
1563 : 27888 : bool do_warn = !(gfp & __GFP_NOWARN);
1564 : 27888 : static int warn_limit = 10;
1565 : 27888 : struct pcpu_chunk *chunk, *next;
1566 : 27888 : const char *err;
1567 : 27888 : int slot, off, cpu, ret;
1568 : 27888 : unsigned long flags;
1569 : 27888 : void __percpu *ptr;
1570 : 27888 : size_t bits, bit_align;
1571 : :
1572 : : /*
1573 : : * There is now a minimum allocation size of PCPU_MIN_ALLOC_SIZE,
1574 : : * therefore alignment must be a minimum of that many bytes.
1575 : : * An allocation may have internal fragmentation from rounding up
1576 : : * of up to PCPU_MIN_ALLOC_SIZE - 1 bytes.
1577 : : */
1578 [ + + ]: 27888 : if (unlikely(align < PCPU_MIN_ALLOC_SIZE))
1579 : 28 : align = PCPU_MIN_ALLOC_SIZE;
1580 : :
1581 : 27888 : size = ALIGN(size, PCPU_MIN_ALLOC_SIZE);
1582 : 27888 : bits = size >> PCPU_MIN_ALLOC_SHIFT;
1583 : 27888 : bit_align = align >> PCPU_MIN_ALLOC_SHIFT;
1584 : :
1585 [ + - + - : 55776 : if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE ||
- + ]
1586 : : !is_power_of_2(align))) {
1587 [ # # ]: 0 : WARN(do_warn, "illegal size (%zu) or align (%zu) for percpu allocation\n",
1588 : : size, align);
1589 : 0 : return NULL;
1590 : : }
1591 : :
1592 [ + + ]: 27888 : if (!is_atomic) {
1593 : : /*
1594 : : * pcpu_balance_workfn() allocates memory under this mutex,
1595 : : * and it may wait for memory reclaim. Allow current task
1596 : : * to become OOM victim, in case of memory pressure.
1597 : : */
1598 [ - + ]: 27832 : if (gfp & __GFP_NOFAIL)
1599 : 0 : mutex_lock(&pcpu_alloc_mutex);
1600 [ + - ]: 27832 : else if (mutex_lock_killable(&pcpu_alloc_mutex))
1601 : : return NULL;
1602 : : }
1603 : :
1604 : 27888 : spin_lock_irqsave(&pcpu_lock, flags);
1605 : :
1606 : : /* serve reserved allocations from the reserved chunk if available */
1607 [ + - - - ]: 27888 : if (reserved && pcpu_reserved_chunk) {
1608 : 0 : chunk = pcpu_reserved_chunk;
1609 : :
1610 : 0 : off = pcpu_find_block_fit(chunk, bits, bit_align, is_atomic);
1611 [ # # ]: 0 : if (off < 0) {
1612 : 0 : err = "alloc from reserved chunk failed";
1613 : 0 : goto fail_unlock;
1614 : : }
1615 : :
1616 : 0 : off = pcpu_alloc_area(chunk, bits, bit_align, off);
1617 [ # # ]: 0 : if (off >= 0)
1618 : 0 : goto area_found;
1619 : :
1620 : 0 : err = "alloc from reserved chunk failed";
1621 : 0 : goto fail_unlock;
1622 : : }
1623 : :
1624 : 27888 : restart:
1625 : : /* search through normal chunks */
1626 [ - + + - ]: 385718 : for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
1627 [ + + ]: 357830 : list_for_each_entry_safe(chunk, next, &pcpu_slot[slot], list) {
1628 : 27888 : off = pcpu_find_block_fit(chunk, bits, bit_align,
1629 : : is_atomic);
1630 [ - + ]: 27888 : if (off < 0) {
1631 [ # # ]: 0 : if (slot < PCPU_SLOT_FAIL_THRESHOLD)
1632 : 0 : pcpu_chunk_move(chunk, 0);
1633 : 0 : continue;
1634 : : }
1635 : :
1636 : 27888 : off = pcpu_alloc_area(chunk, bits, bit_align, off);
1637 [ + - ]: 27888 : if (off >= 0)
1638 : 27888 : goto area_found;
1639 : :
1640 : : }
1641 : : }
1642 : :
1643 : 0 : spin_unlock_irqrestore(&pcpu_lock, flags);
1644 : :
1645 : : /*
1646 : : * No space left. Create a new chunk. We don't want multiple
1647 : : * tasks to create chunks simultaneously. Serialize and create iff
1648 : : * there's still no empty chunk after grabbing the mutex.
1649 : : */
1650 [ # # ]: 0 : if (is_atomic) {
1651 : 0 : err = "atomic alloc failed, no space left";
1652 : 0 : goto fail;
1653 : : }
1654 : :
1655 [ # # ]: 0 : if (list_empty(&pcpu_slot[pcpu_nr_slots - 1])) {
1656 : 0 : chunk = pcpu_create_chunk(pcpu_gfp);
1657 [ # # ]: 0 : if (!chunk) {
1658 : 0 : err = "failed to allocate new chunk";
1659 : 0 : goto fail;
1660 : : }
1661 : :
1662 : 0 : spin_lock_irqsave(&pcpu_lock, flags);
1663 : 0 : pcpu_chunk_relocate(chunk, -1);
1664 : : } else {
1665 : 0 : spin_lock_irqsave(&pcpu_lock, flags);
1666 : : }
1667 : :
1668 : 0 : goto restart;
1669 : :
1670 : 27888 : area_found:
1671 : 27888 : pcpu_stats_area_alloc(chunk, size);
1672 : 27888 : spin_unlock_irqrestore(&pcpu_lock, flags);
1673 : :
1674 : : /* populate if not all pages are already there */
1675 [ + + ]: 27888 : if (!is_atomic) {
1676 : 27832 : unsigned int page_start, page_end, rs, re;
1677 : :
1678 : 27832 : page_start = PFN_DOWN(off);
1679 : 27832 : page_end = PFN_UP(off + size);
1680 : :
1681 [ - + ]: 27832 : bitmap_for_each_clear_region(chunk->populated, rs, re,
1682 : : page_start, page_end) {
1683 [ # # ]: 0 : WARN_ON(chunk->immutable);
1684 : :
1685 : 0 : ret = pcpu_populate_chunk(chunk, rs, re, pcpu_gfp);
1686 : :
1687 : 0 : spin_lock_irqsave(&pcpu_lock, flags);
1688 [ # # ]: 0 : if (ret) {
1689 : 0 : pcpu_free_area(chunk, off);
1690 : 0 : err = "failed to populate";
1691 : 0 : goto fail_unlock;
1692 : : }
1693 : 0 : pcpu_chunk_populated(chunk, rs, re);
1694 : 0 : spin_unlock_irqrestore(&pcpu_lock, flags);
1695 : : }
1696 : :
1697 : 27832 : mutex_unlock(&pcpu_alloc_mutex);
1698 : : }
1699 : :
1700 [ + + ]: 27888 : if (pcpu_nr_empty_pop_pages < PCPU_EMPTY_POP_PAGES_LOW)
1701 [ + - ]: 56 : pcpu_schedule_balance_work();
1702 : :
1703 : : /* clear the areas and return address relative to base address */
1704 [ + + ]: 55776 : for_each_possible_cpu(cpu)
1705 : 27888 : memset((void *)pcpu_chunk_addr(chunk, cpu, 0) + off, 0, size);
1706 : :
1707 : 27888 : ptr = __addr_to_pcpu_ptr(chunk->base_addr + off);
1708 : 27888 : kmemleak_alloc_percpu(ptr, size, gfp);
1709 : :
1710 : 27888 : trace_percpu_alloc_percpu(reserved, is_atomic, size, align,
1711 : : chunk->base_addr, off, ptr);
1712 : :
1713 : 27888 : return ptr;
1714 : :
1715 : 0 : fail_unlock:
1716 : 0 : spin_unlock_irqrestore(&pcpu_lock, flags);
1717 : 0 : fail:
1718 : 0 : trace_percpu_alloc_percpu_fail(reserved, is_atomic, size, align);
1719 : :
1720 [ # # # # ]: 0 : if (!is_atomic && do_warn && warn_limit) {
1721 : 0 : pr_warn("allocation failed, size=%zu align=%zu atomic=%d, %s\n",
1722 : : size, align, is_atomic, err);
1723 : 0 : dump_stack();
1724 [ # # ]: 0 : if (!--warn_limit)
1725 : 0 : pr_info("limit reached, disable warning\n");
1726 : : }
1727 [ # # ]: 0 : if (is_atomic) {
1728 : : /* see the flag handling in pcpu_blance_workfn() */
1729 : 0 : pcpu_atomic_alloc_failed = true;
1730 [ # # ]: 0 : pcpu_schedule_balance_work();
1731 : : } else {
1732 : 0 : mutex_unlock(&pcpu_alloc_mutex);
1733 : : }
1734 : : return NULL;
1735 : : }
1736 : :
1737 : : /**
1738 : : * __alloc_percpu_gfp - allocate dynamic percpu area
1739 : : * @size: size of area to allocate in bytes
1740 : : * @align: alignment of area (max PAGE_SIZE)
1741 : : * @gfp: allocation flags
1742 : : *
1743 : : * Allocate zero-filled percpu area of @size bytes aligned at @align. If
1744 : : * @gfp doesn't contain %GFP_KERNEL, the allocation doesn't block and can
1745 : : * be called from any context but is a lot more likely to fail. If @gfp
1746 : : * has __GFP_NOWARN then no warning will be triggered on invalid or failed
1747 : : * allocation requests.
1748 : : *
1749 : : * RETURNS:
1750 : : * Percpu pointer to the allocated area on success, NULL on failure.
1751 : : */
1752 : 8764 : void __percpu *__alloc_percpu_gfp(size_t size, size_t align, gfp_t gfp)
1753 : : {
1754 : 8764 : return pcpu_alloc(size, align, false, gfp);
1755 : : }
1756 : : EXPORT_SYMBOL_GPL(__alloc_percpu_gfp);
1757 : :
1758 : : /**
1759 : : * __alloc_percpu - allocate dynamic percpu area
1760 : : * @size: size of area to allocate in bytes
1761 : : * @align: alignment of area (max PAGE_SIZE)
1762 : : *
1763 : : * Equivalent to __alloc_percpu_gfp(size, align, %GFP_KERNEL).
1764 : : */
1765 : 19124 : void __percpu *__alloc_percpu(size_t size, size_t align)
1766 : : {
1767 : 19124 : return pcpu_alloc(size, align, false, GFP_KERNEL);
1768 : : }
1769 : : EXPORT_SYMBOL_GPL(__alloc_percpu);
1770 : :
1771 : : /**
1772 : : * __alloc_reserved_percpu - allocate reserved percpu area
1773 : : * @size: size of area to allocate in bytes
1774 : : * @align: alignment of area (max PAGE_SIZE)
1775 : : *
1776 : : * Allocate zero-filled percpu area of @size bytes aligned at @align
1777 : : * from reserved percpu area if arch has set it up; otherwise,
1778 : : * allocation is served from the same dynamic area. Might sleep.
1779 : : * Might trigger writeouts.
1780 : : *
1781 : : * CONTEXT:
1782 : : * Does GFP_KERNEL allocation.
1783 : : *
1784 : : * RETURNS:
1785 : : * Percpu pointer to the allocated area on success, NULL on failure.
1786 : : */
1787 : 0 : void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
1788 : : {
1789 : 0 : return pcpu_alloc(size, align, true, GFP_KERNEL);
1790 : : }
1791 : :
1792 : : /**
1793 : : * pcpu_balance_workfn - manage the amount of free chunks and populated pages
1794 : : * @work: unused
1795 : : *
1796 : : * Reclaim all fully free chunks except for the first one. This is also
1797 : : * responsible for maintaining the pool of empty populated pages. However,
1798 : : * it is possible that this is called when physical memory is scarce causing
1799 : : * OOM killer to be triggered. We should avoid doing so until an actual
1800 : : * allocation causes the failure as it is possible that requests can be
1801 : : * serviced from already backed regions.
1802 : : */
1803 : 56 : static void pcpu_balance_workfn(struct work_struct *work)
1804 : : {
1805 : : /* gfp flags passed to underlying allocators */
1806 : 56 : const gfp_t gfp = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
1807 : 56 : LIST_HEAD(to_free);
1808 : 56 : struct list_head *free_head = &pcpu_slot[pcpu_nr_slots - 1];
1809 : 56 : struct pcpu_chunk *chunk, *next;
1810 : 56 : int slot, nr_to_pop, ret;
1811 : :
1812 : : /*
1813 : : * There's no reason to keep around multiple unused chunks and VM
1814 : : * areas can be scarce. Destroy all free chunks except for one.
1815 : : */
1816 : 56 : mutex_lock(&pcpu_alloc_mutex);
1817 : 56 : spin_lock_irq(&pcpu_lock);
1818 : :
1819 [ - + ]: 56 : list_for_each_entry_safe(chunk, next, free_head, list) {
1820 [ # # ]: 0 : WARN_ON(chunk->immutable);
1821 : :
1822 : : /* spare the first one */
1823 [ # # ]: 0 : if (chunk == list_first_entry(free_head, struct pcpu_chunk, list))
1824 : 0 : continue;
1825 : :
1826 : 0 : list_move(&chunk->list, &to_free);
1827 : : }
1828 : :
1829 : 56 : spin_unlock_irq(&pcpu_lock);
1830 : :
1831 [ - + ]: 56 : list_for_each_entry_safe(chunk, next, &to_free, list) {
1832 : 0 : unsigned int rs, re;
1833 : :
1834 [ # # ]: 0 : bitmap_for_each_set_region(chunk->populated, rs, re, 0,
1835 : : chunk->nr_pages) {
1836 : 0 : pcpu_depopulate_chunk(chunk, rs, re);
1837 : 0 : spin_lock_irq(&pcpu_lock);
1838 : 0 : pcpu_chunk_depopulated(chunk, rs, re);
1839 : 0 : spin_unlock_irq(&pcpu_lock);
1840 : : }
1841 : 0 : pcpu_destroy_chunk(chunk);
1842 : 0 : cond_resched();
1843 : : }
1844 : :
1845 : : /*
1846 : : * Ensure there are certain number of free populated pages for
1847 : : * atomic allocs. Fill up from the most packed so that atomic
1848 : : * allocs don't increase fragmentation. If atomic allocation
1849 : : * failed previously, always populate the maximum amount. This
1850 : : * should prevent atomic allocs larger than PAGE_SIZE from keeping
1851 : : * failing indefinitely; however, large atomic allocs are not
1852 : : * something we support properly and can be highly unreliable and
1853 : : * inefficient.
1854 : : */
1855 : 56 : retry_pop:
1856 [ - + ]: 84 : if (pcpu_atomic_alloc_failed) {
1857 : 0 : nr_to_pop = PCPU_EMPTY_POP_PAGES_HIGH;
1858 : : /* best effort anyway, don't worry about synchronization */
1859 : 0 : pcpu_atomic_alloc_failed = false;
1860 : : } else {
1861 : 84 : nr_to_pop = clamp(PCPU_EMPTY_POP_PAGES_HIGH -
1862 : : pcpu_nr_empty_pop_pages,
1863 : : 0, PCPU_EMPTY_POP_PAGES_HIGH);
1864 : : }
1865 : :
1866 [ - + + + ]: 1036 : for (slot = pcpu_size_to_slot(PAGE_SIZE); slot < pcpu_nr_slots; slot++) {
1867 : 896 : unsigned int nr_unpop = 0, rs, re;
1868 : :
1869 [ + + ]: 896 : if (!nr_to_pop)
1870 : : break;
1871 : :
1872 : 868 : spin_lock_irq(&pcpu_lock);
1873 [ + + ]: 924 : list_for_each_entry(chunk, &pcpu_slot[slot], list) {
1874 : 112 : nr_unpop = chunk->nr_pages - chunk->nr_populated;
1875 [ + + ]: 112 : if (nr_unpop)
1876 : : break;
1877 : : }
1878 : 868 : spin_unlock_irq(&pcpu_lock);
1879 : :
1880 [ + + ]: 868 : if (!nr_unpop)
1881 : 812 : continue;
1882 : :
1883 : : /* @chunk can't go away while pcpu_alloc_mutex is held */
1884 [ + - ]: 56 : bitmap_for_each_clear_region(chunk->populated, rs, re, 0,
1885 : : chunk->nr_pages) {
1886 : 56 : int nr = min_t(int, re - rs, nr_to_pop);
1887 : :
1888 : 56 : ret = pcpu_populate_chunk(chunk, rs, rs + nr, gfp);
1889 [ + - ]: 56 : if (!ret) {
1890 : 56 : nr_to_pop -= nr;
1891 : 56 : spin_lock_irq(&pcpu_lock);
1892 : 56 : pcpu_chunk_populated(chunk, rs, rs + nr);
1893 : 56 : spin_unlock_irq(&pcpu_lock);
1894 : : } else {
1895 : : nr_to_pop = 0;
1896 : : }
1897 : :
1898 [ - + ]: 56 : if (!nr_to_pop)
1899 : : break;
1900 : : }
1901 : : }
1902 : :
1903 [ + + ]: 84 : if (nr_to_pop) {
1904 : : /* ran out of chunks to populate, create a new one and retry */
1905 : 28 : chunk = pcpu_create_chunk(gfp);
1906 [ + - ]: 28 : if (chunk) {
1907 : 28 : spin_lock_irq(&pcpu_lock);
1908 : 28 : pcpu_chunk_relocate(chunk, -1);
1909 : 28 : spin_unlock_irq(&pcpu_lock);
1910 : 28 : goto retry_pop;
1911 : : }
1912 : : }
1913 : :
1914 : 56 : mutex_unlock(&pcpu_alloc_mutex);
1915 : 56 : }
1916 : :
1917 : : /**
1918 : : * free_percpu - free percpu area
1919 : : * @ptr: pointer to area to free
1920 : : *
1921 : : * Free percpu area @ptr.
1922 : : *
1923 : : * CONTEXT:
1924 : : * Can be called from atomic context.
1925 : : */
1926 : 2187 : void free_percpu(void __percpu *ptr)
1927 : : {
1928 : 2187 : void *addr;
1929 : 2187 : struct pcpu_chunk *chunk;
1930 : 2187 : unsigned long flags;
1931 : 2187 : int off;
1932 : 2187 : bool need_balance = false;
1933 : :
1934 [ + - ]: 2187 : if (!ptr)
1935 : : return;
1936 : :
1937 : 2187 : kmemleak_free_percpu(ptr);
1938 : :
1939 : 2187 : addr = __pcpu_ptr_to_addr(ptr);
1940 : :
1941 : 2187 : spin_lock_irqsave(&pcpu_lock, flags);
1942 : :
1943 : 2187 : chunk = pcpu_chunk_addr_search(addr);
1944 : 2187 : off = addr - chunk->base_addr;
1945 : :
1946 : 2187 : pcpu_free_area(chunk, off);
1947 : :
1948 : : /* if there are more than one fully free chunks, wake up grim reaper */
1949 [ - + ]: 2187 : if (chunk->free_bytes == pcpu_unit_size) {
1950 : 0 : struct pcpu_chunk *pos;
1951 : :
1952 [ # # ]: 0 : list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
1953 [ # # ]: 0 : if (pos != chunk) {
1954 : : need_balance = true;
1955 : : break;
1956 : : }
1957 : : }
1958 : :
1959 : 2187 : trace_percpu_free_percpu(chunk->base_addr, off, ptr);
1960 : :
1961 : 2187 : spin_unlock_irqrestore(&pcpu_lock, flags);
1962 : :
1963 [ - + ]: 2187 : if (need_balance)
1964 [ # # ]: 0 : pcpu_schedule_balance_work();
1965 : : }
1966 : : EXPORT_SYMBOL_GPL(free_percpu);
1967 : :
1968 : 0 : bool __is_kernel_percpu_address(unsigned long addr, unsigned long *can_addr)
1969 : : {
1970 : : #ifdef CONFIG_SMP
1971 : 0 : const size_t static_size = __per_cpu_end - __per_cpu_start;
1972 : 0 : void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
1973 : 0 : unsigned int cpu;
1974 : :
1975 [ # # ]: 0 : for_each_possible_cpu(cpu) {
1976 : 0 : void *start = per_cpu_ptr(base, cpu);
1977 : 0 : void *va = (void *)addr;
1978 : :
1979 [ # # # # ]: 0 : if (va >= start && va < start + static_size) {
1980 [ # # ]: 0 : if (can_addr) {
1981 : 0 : *can_addr = (unsigned long) (va - start);
1982 : 0 : *can_addr += (unsigned long)
1983 : 0 : per_cpu_ptr(base, get_boot_cpu_id());
1984 : : }
1985 : 0 : return true;
1986 : : }
1987 : : }
1988 : : #endif
1989 : : /* on UP, can't distinguish from other static vars, always false */
1990 : : return false;
1991 : : }
1992 : :
1993 : : /**
1994 : : * is_kernel_percpu_address - test whether address is from static percpu area
1995 : : * @addr: address to test
1996 : : *
1997 : : * Test whether @addr belongs to in-kernel static percpu area. Module
1998 : : * static percpu areas are not considered. For those, use
1999 : : * is_module_percpu_address().
2000 : : *
2001 : : * RETURNS:
2002 : : * %true if @addr is from in-kernel static percpu area, %false otherwise.
2003 : : */
2004 : 0 : bool is_kernel_percpu_address(unsigned long addr)
2005 : : {
2006 : 0 : return __is_kernel_percpu_address(addr, NULL);
2007 : : }
2008 : :
2009 : : /**
2010 : : * per_cpu_ptr_to_phys - convert translated percpu address to physical address
2011 : : * @addr: the address to be converted to physical address
2012 : : *
2013 : : * Given @addr which is dereferenceable address obtained via one of
2014 : : * percpu access macros, this function translates it into its physical
2015 : : * address. The caller is responsible for ensuring @addr stays valid
2016 : : * until this function finishes.
2017 : : *
2018 : : * percpu allocator has special setup for the first chunk, which currently
2019 : : * supports either embedding in linear address space or vmalloc mapping,
2020 : : * and, from the second one, the backing allocator (currently either vm or
2021 : : * km) provides translation.
2022 : : *
2023 : : * The addr can be translated simply without checking if it falls into the
2024 : : * first chunk. But the current code reflects better how percpu allocator
2025 : : * actually works, and the verification can discover both bugs in percpu
2026 : : * allocator itself and per_cpu_ptr_to_phys() callers. So we keep current
2027 : : * code.
2028 : : *
2029 : : * RETURNS:
2030 : : * The physical address for @addr.
2031 : : */
2032 : 476 : phys_addr_t per_cpu_ptr_to_phys(void *addr)
2033 : : {
2034 : 476 : void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
2035 : 476 : bool in_first_chunk = false;
2036 : 476 : unsigned long first_low, first_high;
2037 : 476 : unsigned int cpu;
2038 : :
2039 : : /*
2040 : : * The following test on unit_low/high isn't strictly
2041 : : * necessary but will speed up lookups of addresses which
2042 : : * aren't in the first chunk.
2043 : : *
2044 : : * The address check is against full chunk sizes. pcpu_base_addr
2045 : : * points to the beginning of the first chunk including the
2046 : : * static region. Assumes good intent as the first chunk may
2047 : : * not be full (ie. < pcpu_unit_pages in size).
2048 : : */
2049 : 476 : first_low = (unsigned long)pcpu_base_addr +
2050 : 476 : pcpu_unit_page_offset(pcpu_low_unit_cpu, 0);
2051 : 476 : first_high = (unsigned long)pcpu_base_addr +
2052 : 476 : pcpu_unit_page_offset(pcpu_high_unit_cpu, pcpu_unit_pages);
2053 : 476 : if ((unsigned long)addr >= first_low &&
2054 [ + - ]: 476 : (unsigned long)addr < first_high) {
2055 [ + - ]: 476 : for_each_possible_cpu(cpu) {
2056 : 476 : void *start = per_cpu_ptr(base, cpu);
2057 : :
2058 [ + - - + ]: 476 : if (addr >= start && addr < start + pcpu_unit_size) {
2059 : : in_first_chunk = true;
2060 : : break;
2061 : : }
2062 : : }
2063 : : }
2064 : :
2065 [ + - ]: 476 : if (in_first_chunk) {
2066 [ + - ]: 476 : if (!is_vmalloc_addr(addr))
2067 [ + - ]: 952 : return __pa(addr);
2068 : : else
2069 : 0 : return page_to_phys(vmalloc_to_page(addr)) +
2070 : 0 : offset_in_page(addr);
2071 : : } else
2072 : 0 : return page_to_phys(pcpu_addr_to_page(addr)) +
2073 : 0 : offset_in_page(addr);
2074 : : }
2075 : :
2076 : : /**
2077 : : * pcpu_alloc_alloc_info - allocate percpu allocation info
2078 : : * @nr_groups: the number of groups
2079 : : * @nr_units: the number of units
2080 : : *
2081 : : * Allocate ai which is large enough for @nr_groups groups containing
2082 : : * @nr_units units. The returned ai's groups[0].cpu_map points to the
2083 : : * cpu_map array which is long enough for @nr_units and filled with
2084 : : * NR_CPUS. It's the caller's responsibility to initialize cpu_map
2085 : : * pointer of other groups.
2086 : : *
2087 : : * RETURNS:
2088 : : * Pointer to the allocated pcpu_alloc_info on success, NULL on
2089 : : * failure.
2090 : : */
2091 : 28 : struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
2092 : : int nr_units)
2093 : : {
2094 : 28 : struct pcpu_alloc_info *ai;
2095 : 28 : size_t base_size, ai_size;
2096 : 28 : void *ptr;
2097 : 28 : int unit;
2098 : :
2099 [ + - ]: 28 : base_size = ALIGN(struct_size(ai, groups, nr_groups),
2100 : : __alignof__(ai->groups[0].cpu_map[0]));
2101 : 28 : ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
2102 : :
2103 : 28 : ptr = memblock_alloc(PFN_ALIGN(ai_size), PAGE_SIZE);
2104 [ + - ]: 28 : if (!ptr)
2105 : : return NULL;
2106 : 28 : ai = ptr;
2107 : 28 : ptr += base_size;
2108 : :
2109 : 28 : ai->groups[0].cpu_map = ptr;
2110 : :
2111 [ + + ]: 56 : for (unit = 0; unit < nr_units; unit++)
2112 : 28 : ai->groups[0].cpu_map[unit] = NR_CPUS;
2113 : :
2114 : 28 : ai->nr_groups = nr_groups;
2115 : 28 : ai->__ai_size = PFN_ALIGN(ai_size);
2116 : :
2117 : 28 : return ai;
2118 : : }
2119 : :
2120 : : /**
2121 : : * pcpu_free_alloc_info - free percpu allocation info
2122 : : * @ai: pcpu_alloc_info to free
2123 : : *
2124 : : * Free @ai which was allocated by pcpu_alloc_alloc_info().
2125 : : */
2126 : 28 : void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
2127 : : {
2128 [ + - ]: 28 : memblock_free_early(__pa(ai), ai->__ai_size);
2129 : 28 : }
2130 : :
2131 : : /**
2132 : : * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
2133 : : * @lvl: loglevel
2134 : : * @ai: allocation info to dump
2135 : : *
2136 : : * Print out information about @ai using loglevel @lvl.
2137 : : */
2138 : 28 : static void pcpu_dump_alloc_info(const char *lvl,
2139 : : const struct pcpu_alloc_info *ai)
2140 : : {
2141 : 28 : int group_width = 1, cpu_width = 1, width;
2142 : 28 : char empty_str[] = "--------";
2143 : 28 : int alloc = 0, alloc_end = 0;
2144 : 28 : int group, v;
2145 : 28 : int upa, apl; /* units per alloc, allocs per line */
2146 : :
2147 : 28 : v = ai->nr_groups;
2148 [ - + ]: 28 : while (v /= 10)
2149 : 0 : group_width++;
2150 : :
2151 : 28 : v = num_possible_cpus();
2152 [ - + ]: 28 : while (v /= 10)
2153 : 0 : cpu_width++;
2154 : 28 : empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
2155 : :
2156 : 28 : upa = ai->alloc_size / ai->unit_size;
2157 : 28 : width = upa * (cpu_width + 1) + group_width + 3;
2158 : 28 : apl = rounddown_pow_of_two(max(60 / width, 1));
2159 : :
2160 : 28 : printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
2161 : : lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
2162 : : ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
2163 : :
2164 [ + + ]: 84 : for (group = 0; group < ai->nr_groups; group++) {
2165 : 28 : const struct pcpu_group_info *gi = &ai->groups[group];
2166 : 28 : int unit = 0, unit_end = 0;
2167 : :
2168 [ - + ]: 28 : BUG_ON(gi->nr_units % upa);
2169 : 28 : for (alloc_end += gi->nr_units / upa;
2170 [ + + ]: 56 : alloc < alloc_end; alloc++) {
2171 [ + - ]: 28 : if (!(alloc % apl)) {
2172 : 28 : pr_cont("\n");
2173 : 28 : printk("%spcpu-alloc: ", lvl);
2174 : : }
2175 : 28 : pr_cont("[%0*d] ", group_width, group);
2176 : :
2177 [ + + ]: 56 : for (unit_end += upa; unit < unit_end; unit++)
2178 [ + - ]: 28 : if (gi->cpu_map[unit] != NR_CPUS)
2179 : 28 : pr_cont("%0*d ",
2180 : : cpu_width, gi->cpu_map[unit]);
2181 : : else
2182 : 0 : pr_cont("%s ", empty_str);
2183 : : }
2184 : : }
2185 : 28 : pr_cont("\n");
2186 : 28 : }
2187 : :
2188 : : /**
2189 : : * pcpu_setup_first_chunk - initialize the first percpu chunk
2190 : : * @ai: pcpu_alloc_info describing how to percpu area is shaped
2191 : : * @base_addr: mapped address
2192 : : *
2193 : : * Initialize the first percpu chunk which contains the kernel static
2194 : : * percpu area. This function is to be called from arch percpu area
2195 : : * setup path.
2196 : : *
2197 : : * @ai contains all information necessary to initialize the first
2198 : : * chunk and prime the dynamic percpu allocator.
2199 : : *
2200 : : * @ai->static_size is the size of static percpu area.
2201 : : *
2202 : : * @ai->reserved_size, if non-zero, specifies the amount of bytes to
2203 : : * reserve after the static area in the first chunk. This reserves
2204 : : * the first chunk such that it's available only through reserved
2205 : : * percpu allocation. This is primarily used to serve module percpu
2206 : : * static areas on architectures where the addressing model has
2207 : : * limited offset range for symbol relocations to guarantee module
2208 : : * percpu symbols fall inside the relocatable range.
2209 : : *
2210 : : * @ai->dyn_size determines the number of bytes available for dynamic
2211 : : * allocation in the first chunk. The area between @ai->static_size +
2212 : : * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
2213 : : *
2214 : : * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
2215 : : * and equal to or larger than @ai->static_size + @ai->reserved_size +
2216 : : * @ai->dyn_size.
2217 : : *
2218 : : * @ai->atom_size is the allocation atom size and used as alignment
2219 : : * for vm areas.
2220 : : *
2221 : : * @ai->alloc_size is the allocation size and always multiple of
2222 : : * @ai->atom_size. This is larger than @ai->atom_size if
2223 : : * @ai->unit_size is larger than @ai->atom_size.
2224 : : *
2225 : : * @ai->nr_groups and @ai->groups describe virtual memory layout of
2226 : : * percpu areas. Units which should be colocated are put into the
2227 : : * same group. Dynamic VM areas will be allocated according to these
2228 : : * groupings. If @ai->nr_groups is zero, a single group containing
2229 : : * all units is assumed.
2230 : : *
2231 : : * The caller should have mapped the first chunk at @base_addr and
2232 : : * copied static data to each unit.
2233 : : *
2234 : : * The first chunk will always contain a static and a dynamic region.
2235 : : * However, the static region is not managed by any chunk. If the first
2236 : : * chunk also contains a reserved region, it is served by two chunks -
2237 : : * one for the reserved region and one for the dynamic region. They
2238 : : * share the same vm, but use offset regions in the area allocation map.
2239 : : * The chunk serving the dynamic region is circulated in the chunk slots
2240 : : * and available for dynamic allocation like any other chunk.
2241 : : */
2242 : 28 : void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
2243 : : void *base_addr)
2244 : : {
2245 : 28 : size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2246 : 28 : size_t static_size, dyn_size;
2247 : 28 : struct pcpu_chunk *chunk;
2248 : 28 : unsigned long *group_offsets;
2249 : 28 : size_t *group_sizes;
2250 : 28 : unsigned long *unit_off;
2251 : 28 : unsigned int cpu;
2252 : 28 : int *unit_map;
2253 : 28 : int group, unit, i;
2254 : 28 : int map_size;
2255 : 28 : unsigned long tmp_addr;
2256 : 28 : size_t alloc_size;
2257 : :
2258 : : #define PCPU_SETUP_BUG_ON(cond) do { \
2259 : : if (unlikely(cond)) { \
2260 : : pr_emerg("failed to initialize, %s\n", #cond); \
2261 : : pr_emerg("cpu_possible_mask=%*pb\n", \
2262 : : cpumask_pr_args(cpu_possible_mask)); \
2263 : : pcpu_dump_alloc_info(KERN_EMERG, ai); \
2264 : : BUG(); \
2265 : : } \
2266 : : } while (0)
2267 : :
2268 : : /* sanity checks */
2269 [ - + ]: 28 : PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
2270 : : #ifdef CONFIG_SMP
2271 [ - + ]: 28 : PCPU_SETUP_BUG_ON(!ai->static_size);
2272 [ - + ]: 28 : PCPU_SETUP_BUG_ON(offset_in_page(__per_cpu_start));
2273 : : #endif
2274 [ - + ]: 28 : PCPU_SETUP_BUG_ON(!base_addr);
2275 [ - + ]: 28 : PCPU_SETUP_BUG_ON(offset_in_page(base_addr));
2276 [ - + ]: 28 : PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
2277 [ - + ]: 28 : PCPU_SETUP_BUG_ON(offset_in_page(ai->unit_size));
2278 [ - + ]: 28 : PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
2279 : 28 : PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->unit_size, PCPU_BITMAP_BLOCK_SIZE));
2280 [ - + ]: 28 : PCPU_SETUP_BUG_ON(ai->dyn_size < PERCPU_DYNAMIC_EARLY_SIZE);
2281 [ - + ]: 28 : PCPU_SETUP_BUG_ON(!ai->dyn_size);
2282 [ - + ]: 28 : PCPU_SETUP_BUG_ON(!IS_ALIGNED(ai->reserved_size, PCPU_MIN_ALLOC_SIZE));
2283 : 28 : PCPU_SETUP_BUG_ON(!(IS_ALIGNED(PCPU_BITMAP_BLOCK_SIZE, PAGE_SIZE) ||
2284 : : IS_ALIGNED(PAGE_SIZE, PCPU_BITMAP_BLOCK_SIZE)));
2285 : 28 : PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
2286 : :
2287 : : /* process group information and build config tables accordingly */
2288 : 28 : alloc_size = ai->nr_groups * sizeof(group_offsets[0]);
2289 : 28 : group_offsets = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2290 [ - + ]: 28 : if (!group_offsets)
2291 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
2292 : : alloc_size);
2293 : :
2294 : 28 : alloc_size = ai->nr_groups * sizeof(group_sizes[0]);
2295 : 28 : group_sizes = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2296 [ - + ]: 28 : if (!group_sizes)
2297 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
2298 : : alloc_size);
2299 : :
2300 : 28 : alloc_size = nr_cpu_ids * sizeof(unit_map[0]);
2301 : 28 : unit_map = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2302 [ - + ]: 28 : if (!unit_map)
2303 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
2304 : : alloc_size);
2305 : :
2306 : 28 : alloc_size = nr_cpu_ids * sizeof(unit_off[0]);
2307 : 28 : unit_off = memblock_alloc(alloc_size, SMP_CACHE_BYTES);
2308 [ - + ]: 28 : if (!unit_off)
2309 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
2310 : : alloc_size);
2311 : :
2312 [ + + ]: 56 : for (cpu = 0; cpu < nr_cpu_ids; cpu++)
2313 : 28 : unit_map[cpu] = UINT_MAX;
2314 : :
2315 : 28 : pcpu_low_unit_cpu = NR_CPUS;
2316 : 28 : pcpu_high_unit_cpu = NR_CPUS;
2317 : :
2318 [ + + ]: 56 : for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
2319 : 28 : const struct pcpu_group_info *gi = &ai->groups[group];
2320 : :
2321 : 28 : group_offsets[group] = gi->base_offset;
2322 : 28 : group_sizes[group] = gi->nr_units * ai->unit_size;
2323 : :
2324 [ + + ]: 56 : for (i = 0; i < gi->nr_units; i++) {
2325 : 28 : cpu = gi->cpu_map[i];
2326 [ - + ]: 28 : if (cpu == NR_CPUS)
2327 : 0 : continue;
2328 : :
2329 [ - + ]: 28 : PCPU_SETUP_BUG_ON(cpu >= nr_cpu_ids);
2330 [ - + ]: 28 : PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
2331 [ - + ]: 28 : PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
2332 : :
2333 : 28 : unit_map[cpu] = unit + i;
2334 : 28 : unit_off[cpu] = gi->base_offset + i * ai->unit_size;
2335 : :
2336 : : /* determine low/high unit_cpu */
2337 [ - + ]: 28 : if (pcpu_low_unit_cpu == NR_CPUS ||
2338 [ # # ]: 0 : unit_off[cpu] < unit_off[pcpu_low_unit_cpu])
2339 : 28 : pcpu_low_unit_cpu = cpu;
2340 [ - + ]: 28 : if (pcpu_high_unit_cpu == NR_CPUS ||
2341 [ # # ]: 0 : unit_off[cpu] > unit_off[pcpu_high_unit_cpu])
2342 : 28 : pcpu_high_unit_cpu = cpu;
2343 : : }
2344 : : }
2345 : 28 : pcpu_nr_units = unit;
2346 : :
2347 [ + + ]: 84 : for_each_possible_cpu(cpu)
2348 [ - + ]: 56 : PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
2349 : :
2350 : : /* we're done parsing the input, undefine BUG macro and dump config */
2351 : : #undef PCPU_SETUP_BUG_ON
2352 : 28 : pcpu_dump_alloc_info(KERN_DEBUG, ai);
2353 : :
2354 : 28 : pcpu_nr_groups = ai->nr_groups;
2355 : 28 : pcpu_group_offsets = group_offsets;
2356 : 28 : pcpu_group_sizes = group_sizes;
2357 : 28 : pcpu_unit_map = unit_map;
2358 : 28 : pcpu_unit_offsets = unit_off;
2359 : :
2360 : : /* determine basic parameters */
2361 : 28 : pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
2362 : 28 : pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
2363 : 28 : pcpu_atom_size = ai->atom_size;
2364 : 28 : pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
2365 : 28 : BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
2366 : :
2367 : 28 : pcpu_stats_save_ai(ai);
2368 : :
2369 : : /*
2370 : : * Allocate chunk slots. The additional last slot is for
2371 : : * empty chunks.
2372 : : */
2373 : 28 : pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
2374 : 28 : pcpu_slot = memblock_alloc(pcpu_nr_slots * sizeof(pcpu_slot[0]),
2375 : : SMP_CACHE_BYTES);
2376 [ - + ]: 28 : if (!pcpu_slot)
2377 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
2378 : : pcpu_nr_slots * sizeof(pcpu_slot[0]));
2379 [ + + ]: 616 : for (i = 0; i < pcpu_nr_slots; i++)
2380 : 588 : INIT_LIST_HEAD(&pcpu_slot[i]);
2381 : :
2382 : : /*
2383 : : * The end of the static region needs to be aligned with the
2384 : : * minimum allocation size as this offsets the reserved and
2385 : : * dynamic region. The first chunk ends page aligned by
2386 : : * expanding the dynamic region, therefore the dynamic region
2387 : : * can be shrunk to compensate while still staying above the
2388 : : * configured sizes.
2389 : : */
2390 : 28 : static_size = ALIGN(ai->static_size, PCPU_MIN_ALLOC_SIZE);
2391 : 28 : dyn_size = ai->dyn_size - (static_size - ai->static_size);
2392 : :
2393 : : /*
2394 : : * Initialize first chunk.
2395 : : * If the reserved_size is non-zero, this initializes the reserved
2396 : : * chunk. If the reserved_size is zero, the reserved chunk is NULL
2397 : : * and the dynamic region is initialized here. The first chunk,
2398 : : * pcpu_first_chunk, will always point to the chunk that serves
2399 : : * the dynamic region.
2400 : : */
2401 : 28 : tmp_addr = (unsigned long)base_addr + static_size;
2402 [ + - ]: 28 : map_size = ai->reserved_size ?: dyn_size;
2403 : 28 : chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2404 : :
2405 : : /* init dynamic chunk if necessary */
2406 [ + - ]: 28 : if (ai->reserved_size) {
2407 : 28 : pcpu_reserved_chunk = chunk;
2408 : :
2409 : 28 : tmp_addr = (unsigned long)base_addr + static_size +
2410 : : ai->reserved_size;
2411 : 28 : map_size = dyn_size;
2412 : 28 : chunk = pcpu_alloc_first_chunk(tmp_addr, map_size);
2413 : : }
2414 : :
2415 : : /* link the first chunk in */
2416 : 28 : pcpu_first_chunk = chunk;
2417 : 28 : pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
2418 : 28 : pcpu_chunk_relocate(pcpu_first_chunk, -1);
2419 : :
2420 : : /* include all regions of the first chunk */
2421 : 28 : pcpu_nr_populated += PFN_DOWN(size_sum);
2422 : :
2423 : 28 : pcpu_stats_chunk_alloc();
2424 : 28 : trace_percpu_create_chunk(base_addr);
2425 : :
2426 : : /* we're done */
2427 : 28 : pcpu_base_addr = base_addr;
2428 : 28 : }
2429 : :
2430 : : #ifdef CONFIG_SMP
2431 : :
2432 : : const char * const pcpu_fc_names[PCPU_FC_NR] __initconst = {
2433 : : [PCPU_FC_AUTO] = "auto",
2434 : : [PCPU_FC_EMBED] = "embed",
2435 : : [PCPU_FC_PAGE] = "page",
2436 : : };
2437 : :
2438 : : enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
2439 : :
2440 : 0 : static int __init percpu_alloc_setup(char *str)
2441 : : {
2442 [ # # ]: 0 : if (!str)
2443 : : return -EINVAL;
2444 : :
2445 : 0 : if (0)
2446 : : /* nada */;
2447 : : #ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
2448 [ # # ]: 0 : else if (!strcmp(str, "embed"))
2449 : 0 : pcpu_chosen_fc = PCPU_FC_EMBED;
2450 : : #endif
2451 : : #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2452 [ # # ]: 0 : else if (!strcmp(str, "page"))
2453 : 0 : pcpu_chosen_fc = PCPU_FC_PAGE;
2454 : : #endif
2455 : : else
2456 : 0 : pr_warn("unknown allocator %s specified\n", str);
2457 : :
2458 : : return 0;
2459 : : }
2460 : : early_param("percpu_alloc", percpu_alloc_setup);
2461 : :
2462 : : /*
2463 : : * pcpu_embed_first_chunk() is used by the generic percpu setup.
2464 : : * Build it if needed by the arch config or the generic setup is going
2465 : : * to be used.
2466 : : */
2467 : : #if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
2468 : : !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
2469 : : #define BUILD_EMBED_FIRST_CHUNK
2470 : : #endif
2471 : :
2472 : : /* build pcpu_page_first_chunk() iff needed by the arch config */
2473 : : #if defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
2474 : : #define BUILD_PAGE_FIRST_CHUNK
2475 : : #endif
2476 : :
2477 : : /* pcpu_build_alloc_info() is used by both embed and page first chunk */
2478 : : #if defined(BUILD_EMBED_FIRST_CHUNK) || defined(BUILD_PAGE_FIRST_CHUNK)
2479 : : /**
2480 : : * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
2481 : : * @reserved_size: the size of reserved percpu area in bytes
2482 : : * @dyn_size: minimum free size for dynamic allocation in bytes
2483 : : * @atom_size: allocation atom size
2484 : : * @cpu_distance_fn: callback to determine distance between cpus, optional
2485 : : *
2486 : : * This function determines grouping of units, their mappings to cpus
2487 : : * and other parameters considering needed percpu size, allocation
2488 : : * atom size and distances between CPUs.
2489 : : *
2490 : : * Groups are always multiples of atom size and CPUs which are of
2491 : : * LOCAL_DISTANCE both ways are grouped together and share space for
2492 : : * units in the same group. The returned configuration is guaranteed
2493 : : * to have CPUs on different nodes on different groups and >=75% usage
2494 : : * of allocated virtual address space.
2495 : : *
2496 : : * RETURNS:
2497 : : * On success, pointer to the new allocation_info is returned. On
2498 : : * failure, ERR_PTR value is returned.
2499 : : */
2500 : 28 : static struct pcpu_alloc_info * __init pcpu_build_alloc_info(
2501 : : size_t reserved_size, size_t dyn_size,
2502 : : size_t atom_size,
2503 : : pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
2504 : : {
2505 : 28 : static int group_map[NR_CPUS] __initdata;
2506 : 28 : static int group_cnt[NR_CPUS] __initdata;
2507 : 28 : const size_t static_size = __per_cpu_end - __per_cpu_start;
2508 : 28 : int nr_groups = 1, nr_units = 0;
2509 : 28 : size_t size_sum, min_unit_size, alloc_size;
2510 : 28 : int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
2511 : 28 : int last_allocs, group, unit;
2512 : 28 : unsigned int cpu, tcpu;
2513 : 28 : struct pcpu_alloc_info *ai;
2514 : 28 : unsigned int *cpu_map;
2515 : :
2516 : : /* this function may be called multiple times */
2517 : 28 : memset(group_map, 0, sizeof(group_map));
2518 : 28 : memset(group_cnt, 0, sizeof(group_cnt));
2519 : :
2520 : : /* calculate size_sum and ensure dyn_size is enough for early alloc */
2521 : 28 : size_sum = PFN_ALIGN(static_size + reserved_size +
2522 : : max_t(size_t, dyn_size, PERCPU_DYNAMIC_EARLY_SIZE));
2523 : 28 : dyn_size = size_sum - static_size - reserved_size;
2524 : :
2525 : : /*
2526 : : * Determine min_unit_size, alloc_size and max_upa such that
2527 : : * alloc_size is multiple of atom_size and is the smallest
2528 : : * which can accommodate 4k aligned segments which are equal to
2529 : : * or larger than min_unit_size.
2530 : : */
2531 : 28 : min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
2532 : :
2533 : : /* determine the maximum # of units that can fit in an allocation */
2534 : 28 : alloc_size = roundup(min_unit_size, atom_size);
2535 : 28 : upa = alloc_size / min_unit_size;
2536 [ - + - + ]: 28 : while (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2537 : 0 : upa--;
2538 : : max_upa = upa;
2539 : :
2540 : : /* group cpus according to their proximity */
2541 [ + + ]: 56 : for_each_possible_cpu(cpu) {
2542 : : group = 0;
2543 : 28 : next_group:
2544 [ + - ]: 28 : for_each_possible_cpu(tcpu) {
2545 [ - + ]: 28 : if (cpu == tcpu)
2546 : : break;
2547 [ # # # # : 0 : if (group_map[tcpu] == group && cpu_distance_fn &&
# # ]
2548 [ # # ]: 0 : (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
2549 : 0 : cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
2550 : 0 : group++;
2551 : 0 : nr_groups = max(nr_groups, group + 1);
2552 : 0 : goto next_group;
2553 : : }
2554 : : }
2555 : 28 : group_map[cpu] = group;
2556 : 28 : group_cnt[group]++;
2557 : : }
2558 : :
2559 : : /*
2560 : : * Wasted space is caused by a ratio imbalance of upa to group_cnt.
2561 : : * Expand the unit_size until we use >= 75% of the units allocated.
2562 : : * Related to atom_size, which could be much larger than the unit_size.
2563 : : */
2564 : : last_allocs = INT_MAX;
2565 [ + + ]: 252 : for (upa = max_upa; upa; upa--) {
2566 : 224 : int allocs = 0, wasted = 0;
2567 : :
2568 [ + + - + ]: 224 : if (alloc_size % upa || (offset_in_page(alloc_size / upa)))
2569 : 112 : continue;
2570 : :
2571 [ + + ]: 224 : for (group = 0; group < nr_groups; group++) {
2572 : 112 : int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
2573 : 112 : allocs += this_allocs;
2574 : 112 : wasted += this_allocs * upa - group_cnt[group];
2575 : : }
2576 : :
2577 : : /*
2578 : : * Don't accept if wastage is over 1/3. The
2579 : : * greater-than comparison ensures upa==1 always
2580 : : * passes the following check.
2581 : : */
2582 [ + + ]: 112 : if (wasted > num_possible_cpus() / 3)
2583 : 84 : continue;
2584 : :
2585 : : /* and then don't consume more memory */
2586 [ + - ]: 28 : if (allocs > last_allocs)
2587 : : break;
2588 : : last_allocs = allocs;
2589 : : best_upa = upa;
2590 : : }
2591 : 28 : upa = best_upa;
2592 : :
2593 : : /* allocate and fill alloc_info */
2594 [ + + ]: 56 : for (group = 0; group < nr_groups; group++)
2595 : 28 : nr_units += roundup(group_cnt[group], upa);
2596 : :
2597 : 28 : ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
2598 [ + - ]: 28 : if (!ai)
2599 : : return ERR_PTR(-ENOMEM);
2600 : 28 : cpu_map = ai->groups[0].cpu_map;
2601 : :
2602 [ + + ]: 56 : for (group = 0; group < nr_groups; group++) {
2603 : 28 : ai->groups[group].cpu_map = cpu_map;
2604 : 28 : cpu_map += roundup(group_cnt[group], upa);
2605 : : }
2606 : :
2607 : 28 : ai->static_size = static_size;
2608 : 28 : ai->reserved_size = reserved_size;
2609 : 28 : ai->dyn_size = dyn_size;
2610 : 28 : ai->unit_size = alloc_size / upa;
2611 : 28 : ai->atom_size = atom_size;
2612 : 28 : ai->alloc_size = alloc_size;
2613 : :
2614 [ + + ]: 56 : for (group = 0, unit = 0; group < nr_groups; group++) {
2615 : 28 : struct pcpu_group_info *gi = &ai->groups[group];
2616 : :
2617 : : /*
2618 : : * Initialize base_offset as if all groups are located
2619 : : * back-to-back. The caller should update this to
2620 : : * reflect actual allocation.
2621 : : */
2622 : 28 : gi->base_offset = unit * ai->unit_size;
2623 : :
2624 [ + + ]: 56 : for_each_possible_cpu(cpu)
2625 [ + - ]: 28 : if (group_map[cpu] == group)
2626 : 28 : gi->cpu_map[gi->nr_units++] = cpu;
2627 : 28 : gi->nr_units = roundup(gi->nr_units, upa);
2628 : 28 : unit += gi->nr_units;
2629 : : }
2630 [ - + ]: 28 : BUG_ON(unit != nr_units);
2631 : :
2632 : : return ai;
2633 : : }
2634 : : #endif /* BUILD_EMBED_FIRST_CHUNK || BUILD_PAGE_FIRST_CHUNK */
2635 : :
2636 : : #if defined(BUILD_EMBED_FIRST_CHUNK)
2637 : : /**
2638 : : * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
2639 : : * @reserved_size: the size of reserved percpu area in bytes
2640 : : * @dyn_size: minimum free size for dynamic allocation in bytes
2641 : : * @atom_size: allocation atom size
2642 : : * @cpu_distance_fn: callback to determine distance between cpus, optional
2643 : : * @alloc_fn: function to allocate percpu page
2644 : : * @free_fn: function to free percpu page
2645 : : *
2646 : : * This is a helper to ease setting up embedded first percpu chunk and
2647 : : * can be called where pcpu_setup_first_chunk() is expected.
2648 : : *
2649 : : * If this function is used to setup the first chunk, it is allocated
2650 : : * by calling @alloc_fn and used as-is without being mapped into
2651 : : * vmalloc area. Allocations are always whole multiples of @atom_size
2652 : : * aligned to @atom_size.
2653 : : *
2654 : : * This enables the first chunk to piggy back on the linear physical
2655 : : * mapping which often uses larger page size. Please note that this
2656 : : * can result in very sparse cpu->unit mapping on NUMA machines thus
2657 : : * requiring large vmalloc address space. Don't use this allocator if
2658 : : * vmalloc space is not orders of magnitude larger than distances
2659 : : * between node memory addresses (ie. 32bit NUMA machines).
2660 : : *
2661 : : * @dyn_size specifies the minimum dynamic area size.
2662 : : *
2663 : : * If the needed size is smaller than the minimum or specified unit
2664 : : * size, the leftover is returned using @free_fn.
2665 : : *
2666 : : * RETURNS:
2667 : : * 0 on success, -errno on failure.
2668 : : */
2669 : 28 : int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
2670 : : size_t atom_size,
2671 : : pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
2672 : : pcpu_fc_alloc_fn_t alloc_fn,
2673 : : pcpu_fc_free_fn_t free_fn)
2674 : : {
2675 : 28 : void *base = (void *)ULONG_MAX;
2676 : 28 : void **areas = NULL;
2677 : 28 : struct pcpu_alloc_info *ai;
2678 : 28 : size_t size_sum, areas_size;
2679 : 28 : unsigned long max_distance;
2680 : 28 : int group, i, highest_group, rc = 0;
2681 : :
2682 : 28 : ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
2683 : : cpu_distance_fn);
2684 [ - + ]: 28 : if (IS_ERR(ai))
2685 : 0 : return PTR_ERR(ai);
2686 : :
2687 : 28 : size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
2688 : 28 : areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
2689 : :
2690 : 28 : areas = memblock_alloc(areas_size, SMP_CACHE_BYTES);
2691 [ - + ]: 28 : if (!areas) {
2692 : 0 : rc = -ENOMEM;
2693 : 0 : goto out_free;
2694 : : }
2695 : :
2696 : : /* allocate, copy and determine base address & max_distance */
2697 : : highest_group = 0;
2698 [ + + ]: 56 : for (group = 0; group < ai->nr_groups; group++) {
2699 : : struct pcpu_group_info *gi = &ai->groups[group];
2700 : : unsigned int cpu = NR_CPUS;
2701 : : void *ptr;
2702 : :
2703 [ + + + - ]: 56 : for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
2704 : 28 : cpu = gi->cpu_map[i];
2705 [ - + ]: 28 : BUG_ON(cpu == NR_CPUS);
2706 : :
2707 : : /* allocate space for the whole group */
2708 : 28 : ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
2709 [ - + ]: 28 : if (!ptr) {
2710 : 0 : rc = -ENOMEM;
2711 : 0 : goto out_free_areas;
2712 : : }
2713 : : /* kmemleak tracks the percpu allocations separately */
2714 [ - + ]: 28 : kmemleak_free(ptr);
2715 : 28 : areas[group] = ptr;
2716 : :
2717 : 28 : base = min(ptr, base);
2718 [ - + ]: 28 : if (ptr > areas[highest_group])
2719 : 0 : highest_group = group;
2720 : : }
2721 : 28 : max_distance = areas[highest_group] - base;
2722 : 28 : max_distance += ai->unit_size * ai->groups[highest_group].nr_units;
2723 : :
2724 : : /* warn if maximum distance is further than 75% of vmalloc space */
2725 [ - + - - : 56 : if (max_distance > VMALLOC_TOTAL * 3 / 4) {
+ ]
2726 [ # # # ]: 0 : pr_warn("max_distance=0x%lx too large for vmalloc space 0x%lx\n",
2727 : : max_distance, VMALLOC_TOTAL);
2728 : : #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
2729 : : /* and fail if we have fallback */
2730 : 0 : rc = -EINVAL;
2731 : 0 : goto out_free_areas;
2732 : : #endif
2733 : : }
2734 : :
2735 : : /*
2736 : : * Copy data and free unused parts. This should happen after all
2737 : : * allocations are complete; otherwise, we may end up with
2738 : : * overlapping groups.
2739 : : */
2740 [ + + ]: 56 : for (group = 0; group < ai->nr_groups; group++) {
2741 : 28 : struct pcpu_group_info *gi = &ai->groups[group];
2742 : 28 : void *ptr = areas[group];
2743 : :
2744 [ + + ]: 56 : for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
2745 [ - + ]: 28 : if (gi->cpu_map[i] == NR_CPUS) {
2746 : : /* unused unit, free whole */
2747 : 0 : free_fn(ptr, ai->unit_size);
2748 : 0 : continue;
2749 : : }
2750 : : /* copy and return the unused part */
2751 : 28 : memcpy(ptr, __per_cpu_load, ai->static_size);
2752 : 28 : free_fn(ptr + size_sum, ai->unit_size - size_sum);
2753 : : }
2754 : : }
2755 : :
2756 : : /* base address is now known, determine group base offsets */
2757 [ + + ]: 56 : for (group = 0; group < ai->nr_groups; group++) {
2758 : 28 : ai->groups[group].base_offset = areas[group] - base;
2759 : : }
2760 : :
2761 : 28 : pr_info("Embedded %zu pages/cpu s%zu r%zu d%zu u%zu\n",
2762 : : PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
2763 : : ai->dyn_size, ai->unit_size);
2764 : :
2765 : 28 : pcpu_setup_first_chunk(ai, base);
2766 : 28 : goto out_free;
2767 : :
2768 : 0 : out_free_areas:
2769 [ # # ]: 0 : for (group = 0; group < ai->nr_groups; group++)
2770 [ # # ]: 0 : if (areas[group])
2771 : 0 : free_fn(areas[group],
2772 : 0 : ai->groups[group].nr_units * ai->unit_size);
2773 : 0 : out_free:
2774 : 28 : pcpu_free_alloc_info(ai);
2775 [ + - ]: 28 : if (areas)
2776 [ + - ]: 28 : memblock_free_early(__pa(areas), areas_size);
2777 : : return rc;
2778 : : }
2779 : : #endif /* BUILD_EMBED_FIRST_CHUNK */
2780 : :
2781 : : #ifdef BUILD_PAGE_FIRST_CHUNK
2782 : : /**
2783 : : * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
2784 : : * @reserved_size: the size of reserved percpu area in bytes
2785 : : * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
2786 : : * @free_fn: function to free percpu page, always called with PAGE_SIZE
2787 : : * @populate_pte_fn: function to populate pte
2788 : : *
2789 : : * This is a helper to ease setting up page-remapped first percpu
2790 : : * chunk and can be called where pcpu_setup_first_chunk() is expected.
2791 : : *
2792 : : * This is the basic allocator. Static percpu area is allocated
2793 : : * page-by-page into vmalloc area.
2794 : : *
2795 : : * RETURNS:
2796 : : * 0 on success, -errno on failure.
2797 : : */
2798 : 0 : int __init pcpu_page_first_chunk(size_t reserved_size,
2799 : : pcpu_fc_alloc_fn_t alloc_fn,
2800 : : pcpu_fc_free_fn_t free_fn,
2801 : : pcpu_fc_populate_pte_fn_t populate_pte_fn)
2802 : : {
2803 : 0 : static struct vm_struct vm;
2804 : 0 : struct pcpu_alloc_info *ai;
2805 : 0 : char psize_str[16];
2806 : 0 : int unit_pages;
2807 : 0 : size_t pages_size;
2808 : 0 : struct page **pages;
2809 : 0 : int unit, i, j, rc = 0;
2810 : 0 : int upa;
2811 : 0 : int nr_g0_units;
2812 : :
2813 : 0 : snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
2814 : :
2815 : 0 : ai = pcpu_build_alloc_info(reserved_size, 0, PAGE_SIZE, NULL);
2816 [ # # ]: 0 : if (IS_ERR(ai))
2817 : 0 : return PTR_ERR(ai);
2818 [ # # ]: 0 : BUG_ON(ai->nr_groups != 1);
2819 : 0 : upa = ai->alloc_size/ai->unit_size;
2820 : 0 : nr_g0_units = roundup(num_possible_cpus(), upa);
2821 [ # # # # ]: 0 : if (WARN_ON(ai->groups[0].nr_units != nr_g0_units)) {
2822 : 0 : pcpu_free_alloc_info(ai);
2823 : 0 : return -EINVAL;
2824 : : }
2825 : :
2826 : 0 : unit_pages = ai->unit_size >> PAGE_SHIFT;
2827 : :
2828 : : /* unaligned allocations can't be freed, round up to page size */
2829 : 0 : pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
2830 : : sizeof(pages[0]));
2831 : 0 : pages = memblock_alloc(pages_size, SMP_CACHE_BYTES);
2832 [ # # ]: 0 : if (!pages)
2833 : 0 : panic("%s: Failed to allocate %zu bytes\n", __func__,
2834 : : pages_size);
2835 : :
2836 : : /* allocate pages */
2837 : : j = 0;
2838 [ # # ]: 0 : for (unit = 0; unit < num_possible_cpus(); unit++) {
2839 : 0 : unsigned int cpu = ai->groups[0].cpu_map[unit];
2840 [ # # ]: 0 : for (i = 0; i < unit_pages; i++) {
2841 : 0 : void *ptr;
2842 : :
2843 : 0 : ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
2844 [ # # ]: 0 : if (!ptr) {
2845 : 0 : pr_warn("failed to allocate %s page for cpu%u\n",
2846 : : psize_str, cpu);
2847 : 0 : goto enomem;
2848 : : }
2849 : : /* kmemleak tracks the percpu allocations separately */
2850 [ # # ]: 0 : kmemleak_free(ptr);
2851 [ # # ]: 0 : pages[j++] = virt_to_page(ptr);
2852 : : }
2853 : : }
2854 : :
2855 : : /* allocate vm area, map the pages and copy static data */
2856 : 0 : vm.flags = VM_ALLOC;
2857 : 0 : vm.size = num_possible_cpus() * ai->unit_size;
2858 : 0 : vm_area_register_early(&vm, PAGE_SIZE);
2859 : :
2860 [ # # ]: 0 : for (unit = 0; unit < num_possible_cpus(); unit++) {
2861 : 0 : unsigned long unit_addr =
2862 : 0 : (unsigned long)vm.addr + unit * ai->unit_size;
2863 : :
2864 [ # # ]: 0 : for (i = 0; i < unit_pages; i++)
2865 : 0 : populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
2866 : :
2867 : : /* pte already populated, the following shouldn't fail */
2868 : 0 : rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
2869 : : unit_pages);
2870 [ # # ]: 0 : if (rc < 0)
2871 : 0 : panic("failed to map percpu area, err=%d\n", rc);
2872 : :
2873 : : /*
2874 : : * FIXME: Archs with virtual cache should flush local
2875 : : * cache for the linear mapping here - something
2876 : : * equivalent to flush_cache_vmap() on the local cpu.
2877 : : * flush_cache_vmap() can't be used as most supporting
2878 : : * data structures are not set up yet.
2879 : : */
2880 : :
2881 : : /* copy static data */
2882 : 0 : memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
2883 : : }
2884 : :
2885 : : /* we're ready, commit */
2886 : 0 : pr_info("%d %s pages/cpu s%zu r%zu d%zu\n",
2887 : : unit_pages, psize_str, ai->static_size,
2888 : : ai->reserved_size, ai->dyn_size);
2889 : :
2890 : 0 : pcpu_setup_first_chunk(ai, vm.addr);
2891 : 0 : goto out_free_ar;
2892 : :
2893 : : enomem:
2894 [ # # ]: 0 : while (--j >= 0)
2895 : 0 : free_fn(page_address(pages[j]), PAGE_SIZE);
2896 : : rc = -ENOMEM;
2897 : 0 : out_free_ar:
2898 [ # # ]: 0 : memblock_free_early(__pa(pages), pages_size);
2899 : 0 : pcpu_free_alloc_info(ai);
2900 : 0 : return rc;
2901 : : }
2902 : : #endif /* BUILD_PAGE_FIRST_CHUNK */
2903 : :
2904 : : #ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
2905 : : /*
2906 : : * Generic SMP percpu area setup.
2907 : : *
2908 : : * The embedding helper is used because its behavior closely resembles
2909 : : * the original non-dynamic generic percpu area setup. This is
2910 : : * important because many archs have addressing restrictions and might
2911 : : * fail if the percpu area is located far away from the previous
2912 : : * location. As an added bonus, in non-NUMA cases, embedding is
2913 : : * generally a good idea TLB-wise because percpu area can piggy back
2914 : : * on the physical linear memory mapping which uses large page
2915 : : * mappings on applicable archs.
2916 : : */
2917 : : unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
2918 : : EXPORT_SYMBOL(__per_cpu_offset);
2919 : :
2920 : : static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
2921 : : size_t align)
2922 : : {
2923 : : return memblock_alloc_from(size, align, __pa(MAX_DMA_ADDRESS));
2924 : : }
2925 : :
2926 : : static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
2927 : : {
2928 : : memblock_free_early(__pa(ptr), size);
2929 : : }
2930 : :
2931 : : void __init setup_per_cpu_areas(void)
2932 : : {
2933 : : unsigned long delta;
2934 : : unsigned int cpu;
2935 : : int rc;
2936 : :
2937 : : /*
2938 : : * Always reserve area for module percpu variables. That's
2939 : : * what the legacy allocator did.
2940 : : */
2941 : : rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
2942 : : PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
2943 : : pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
2944 : : if (rc < 0)
2945 : : panic("Failed to initialize percpu areas.");
2946 : :
2947 : : delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
2948 : : for_each_possible_cpu(cpu)
2949 : : __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
2950 : : }
2951 : : #endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
2952 : :
2953 : : #else /* CONFIG_SMP */
2954 : :
2955 : : /*
2956 : : * UP percpu area setup.
2957 : : *
2958 : : * UP always uses km-based percpu allocator with identity mapping.
2959 : : * Static percpu variables are indistinguishable from the usual static
2960 : : * variables and don't require any special preparation.
2961 : : */
2962 : : void __init setup_per_cpu_areas(void)
2963 : : {
2964 : : const size_t unit_size =
2965 : : roundup_pow_of_two(max_t(size_t, PCPU_MIN_UNIT_SIZE,
2966 : : PERCPU_DYNAMIC_RESERVE));
2967 : : struct pcpu_alloc_info *ai;
2968 : : void *fc;
2969 : :
2970 : : ai = pcpu_alloc_alloc_info(1, 1);
2971 : : fc = memblock_alloc_from(unit_size, PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
2972 : : if (!ai || !fc)
2973 : : panic("Failed to allocate memory for percpu areas.");
2974 : : /* kmemleak tracks the percpu allocations separately */
2975 : : kmemleak_free(fc);
2976 : :
2977 : : ai->dyn_size = unit_size;
2978 : : ai->unit_size = unit_size;
2979 : : ai->atom_size = unit_size;
2980 : : ai->alloc_size = unit_size;
2981 : : ai->groups[0].nr_units = 1;
2982 : : ai->groups[0].cpu_map[0] = 0;
2983 : :
2984 : : pcpu_setup_first_chunk(ai, fc);
2985 : : pcpu_free_alloc_info(ai);
2986 : : }
2987 : :
2988 : : #endif /* CONFIG_SMP */
2989 : :
2990 : : /*
2991 : : * pcpu_nr_pages - calculate total number of populated backing pages
2992 : : *
2993 : : * This reflects the number of pages populated to back chunks. Metadata is
2994 : : * excluded in the number exposed in meminfo as the number of backing pages
2995 : : * scales with the number of cpus and can quickly outweigh the memory used for
2996 : : * metadata. It also keeps this calculation nice and simple.
2997 : : *
2998 : : * RETURNS:
2999 : : * Total number of populated backing pages in use by the allocator.
3000 : : */
3001 : 28 : unsigned long pcpu_nr_pages(void)
3002 : : {
3003 : 28 : return pcpu_nr_populated * pcpu_nr_units;
3004 : : }
3005 : :
3006 : : /*
3007 : : * Percpu allocator is initialized early during boot when neither slab or
3008 : : * workqueue is available. Plug async management until everything is up
3009 : : * and running.
3010 : : */
3011 : 28 : static int __init percpu_enable_async(void)
3012 : : {
3013 : 28 : pcpu_async_enabled = true;
3014 : 28 : return 0;
3015 : : }
3016 : : subsys_initcall(percpu_enable_async);
|
