Root/mm/percpu.c

1/*
2 * mm/percpu.c - percpu memory allocator
3 *
4 * Copyright (C) 2009 SUSE Linux Products GmbH
5 * Copyright (C) 2009 Tejun Heo <tj@kernel.org>
6 *
7 * This file is released under the GPLv2.
8 *
9 * This is percpu allocator which can handle both static and dynamic
10 * areas. Percpu areas are allocated in chunks. Each chunk is
11 * consisted of boot-time determined number of units and the first
12 * chunk is used for static percpu variables in the kernel image
13 * (special boot time alloc/init handling necessary as these areas
14 * need to be brought up before allocation services are running).
15 * Unit grows as necessary and all units grow or shrink in unison.
16 * When a chunk is filled up, another chunk is allocated.
17 *
18 * c0 c1 c2
19 * ------------------- ------------------- ------------
20 * | u0 | u1 | u2 | u3 | | u0 | u1 | u2 | u3 | | u0 | u1 | u
21 * ------------------- ...... ------------------- .... ------------
22 *
23 * Allocation is done in offset-size areas of single unit space. Ie,
24 * an area of 512 bytes at 6k in c1 occupies 512 bytes at 6k of c1:u0,
25 * c1:u1, c1:u2 and c1:u3. On UMA, units corresponds directly to
26 * cpus. On NUMA, the mapping can be non-linear and even sparse.
27 * Percpu access can be done by configuring percpu base registers
28 * according to cpu to unit mapping and pcpu_unit_size.
29 *
30 * There are usually many small percpu allocations many of them being
31 * as small as 4 bytes. The allocator organizes chunks into lists
32 * according to free size and tries to allocate from the fullest one.
33 * Each chunk keeps the maximum contiguous area size hint which is
34 * guaranteed to be eqaul to or larger than the maximum contiguous
35 * area in the chunk. This helps the allocator not to iterate the
36 * chunk maps unnecessarily.
37 *
38 * Allocation state in each chunk is kept using an array of integers
39 * on chunk->map. A positive value in the map represents a free
40 * region and negative allocated. Allocation inside a chunk is done
41 * by scanning this map sequentially and serving the first matching
42 * entry. This is mostly copied from the percpu_modalloc() allocator.
43 * Chunks can be determined from the address using the index field
44 * in the page struct. The index field contains a pointer to the chunk.
45 *
46 * To use this allocator, arch code should do the followings.
47 *
48 * - define __addr_to_pcpu_ptr() and __pcpu_ptr_to_addr() to translate
49 * regular address to percpu pointer and back if they need to be
50 * different from the default
51 *
52 * - use pcpu_setup_first_chunk() during percpu area initialization to
53 * setup the first chunk containing the kernel static percpu area
54 */
55
56#include <linux/bitmap.h>
57#include <linux/bootmem.h>
58#include <linux/err.h>
59#include <linux/list.h>
60#include <linux/log2.h>
61#include <linux/mm.h>
62#include <linux/module.h>
63#include <linux/mutex.h>
64#include <linux/percpu.h>
65#include <linux/pfn.h>
66#include <linux/slab.h>
67#include <linux/spinlock.h>
68#include <linux/vmalloc.h>
69#include <linux/workqueue.h>
70
71#include <asm/cacheflush.h>
72#include <asm/sections.h>
73#include <asm/tlbflush.h>
74#include <asm/io.h>
75
76#define PCPU_SLOT_BASE_SHIFT 5 /* 1-31 shares the same slot */
77#define PCPU_DFL_MAP_ALLOC 16 /* start a map with 16 ents */
78
79/* default addr <-> pcpu_ptr mapping, override in asm/percpu.h if necessary */
80#ifndef __addr_to_pcpu_ptr
81#define __addr_to_pcpu_ptr(addr) \
82    (void __percpu *)((unsigned long)(addr) - \
83              (unsigned long)pcpu_base_addr + \
84              (unsigned long)__per_cpu_start)
85#endif
86#ifndef __pcpu_ptr_to_addr
87#define __pcpu_ptr_to_addr(ptr) \
88    (void __force *)((unsigned long)(ptr) + \
89             (unsigned long)pcpu_base_addr - \
90             (unsigned long)__per_cpu_start)
91#endif
92
93struct pcpu_chunk {
94    struct list_head list; /* linked to pcpu_slot lists */
95    int free_size; /* free bytes in the chunk */
96    int contig_hint; /* max contiguous size hint */
97    void *base_addr; /* base address of this chunk */
98    int map_used; /* # of map entries used */
99    int map_alloc; /* # of map entries allocated */
100    int *map; /* allocation map */
101    void *data; /* chunk data */
102    bool immutable; /* no [de]population allowed */
103    unsigned long populated[]; /* populated bitmap */
104};
105
106static int pcpu_unit_pages __read_mostly;
107static int pcpu_unit_size __read_mostly;
108static int pcpu_nr_units __read_mostly;
109static int pcpu_atom_size __read_mostly;
110static int pcpu_nr_slots __read_mostly;
111static size_t pcpu_chunk_struct_size __read_mostly;
112
113/* cpus with the lowest and highest unit numbers */
114static unsigned int pcpu_first_unit_cpu __read_mostly;
115static unsigned int pcpu_last_unit_cpu __read_mostly;
116
117/* the address of the first chunk which starts with the kernel static area */
118void *pcpu_base_addr __read_mostly;
119EXPORT_SYMBOL_GPL(pcpu_base_addr);
120
121static const int *pcpu_unit_map __read_mostly; /* cpu -> unit */
122const unsigned long *pcpu_unit_offsets __read_mostly; /* cpu -> unit offset */
123
124/* group information, used for vm allocation */
125static int pcpu_nr_groups __read_mostly;
126static const unsigned long *pcpu_group_offsets __read_mostly;
127static const size_t *pcpu_group_sizes __read_mostly;
128
129/*
130 * The first chunk which always exists. Note that unlike other
131 * chunks, this one can be allocated and mapped in several different
132 * ways and thus often doesn't live in the vmalloc area.
133 */
134static struct pcpu_chunk *pcpu_first_chunk;
135
136/*
137 * Optional reserved chunk. This chunk reserves part of the first
138 * chunk and serves it for reserved allocations. The amount of
139 * reserved offset is in pcpu_reserved_chunk_limit. When reserved
140 * area doesn't exist, the following variables contain NULL and 0
141 * respectively.
142 */
143static struct pcpu_chunk *pcpu_reserved_chunk;
144static int pcpu_reserved_chunk_limit;
145
146/*
147 * Synchronization rules.
148 *
149 * There are two locks - pcpu_alloc_mutex and pcpu_lock. The former
150 * protects allocation/reclaim paths, chunks, populated bitmap and
151 * vmalloc mapping. The latter is a spinlock and protects the index
152 * data structures - chunk slots, chunks and area maps in chunks.
153 *
154 * During allocation, pcpu_alloc_mutex is kept locked all the time and
155 * pcpu_lock is grabbed and released as necessary. All actual memory
156 * allocations are done using GFP_KERNEL with pcpu_lock released. In
157 * general, percpu memory can't be allocated with irq off but
158 * irqsave/restore are still used in alloc path so that it can be used
159 * from early init path - sched_init() specifically.
160 *
161 * Free path accesses and alters only the index data structures, so it
162 * can be safely called from atomic context. When memory needs to be
163 * returned to the system, free path schedules reclaim_work which
164 * grabs both pcpu_alloc_mutex and pcpu_lock, unlinks chunks to be
165 * reclaimed, release both locks and frees the chunks. Note that it's
166 * necessary to grab both locks to remove a chunk from circulation as
167 * allocation path might be referencing the chunk with only
168 * pcpu_alloc_mutex locked.
169 */
170static DEFINE_MUTEX(pcpu_alloc_mutex); /* protects whole alloc and reclaim */
171static DEFINE_SPINLOCK(pcpu_lock); /* protects index data structures */
172
173static struct list_head *pcpu_slot __read_mostly; /* chunk list slots */
174
175/* reclaim work to release fully free chunks, scheduled from free path */
176static void pcpu_reclaim(struct work_struct *work);
177static DECLARE_WORK(pcpu_reclaim_work, pcpu_reclaim);
178
179static bool pcpu_addr_in_first_chunk(void *addr)
180{
181    void *first_start = pcpu_first_chunk->base_addr;
182
183    return addr >= first_start && addr < first_start + pcpu_unit_size;
184}
185
186static bool pcpu_addr_in_reserved_chunk(void *addr)
187{
188    void *first_start = pcpu_first_chunk->base_addr;
189
190    return addr >= first_start &&
191        addr < first_start + pcpu_reserved_chunk_limit;
192}
193
194static int __pcpu_size_to_slot(int size)
195{
196    int highbit = fls(size); /* size is in bytes */
197    return max(highbit - PCPU_SLOT_BASE_SHIFT + 2, 1);
198}
199
200static int pcpu_size_to_slot(int size)
201{
202    if (size == pcpu_unit_size)
203        return pcpu_nr_slots - 1;
204    return __pcpu_size_to_slot(size);
205}
206
207static int pcpu_chunk_slot(const struct pcpu_chunk *chunk)
208{
209    if (chunk->free_size < sizeof(int) || chunk->contig_hint < sizeof(int))
210        return 0;
211
212    return pcpu_size_to_slot(chunk->free_size);
213}
214
215/* set the pointer to a chunk in a page struct */
216static void pcpu_set_page_chunk(struct page *page, struct pcpu_chunk *pcpu)
217{
218    page->index = (unsigned long)pcpu;
219}
220
221/* obtain pointer to a chunk from a page struct */
222static struct pcpu_chunk *pcpu_get_page_chunk(struct page *page)
223{
224    return (struct pcpu_chunk *)page->index;
225}
226
227static int __maybe_unused pcpu_page_idx(unsigned int cpu, int page_idx)
228{
229    return pcpu_unit_map[cpu] * pcpu_unit_pages + page_idx;
230}
231
232static unsigned long pcpu_chunk_addr(struct pcpu_chunk *chunk,
233                     unsigned int cpu, int page_idx)
234{
235    return (unsigned long)chunk->base_addr + pcpu_unit_offsets[cpu] +
236        (page_idx << PAGE_SHIFT);
237}
238
239static void __maybe_unused pcpu_next_unpop(struct pcpu_chunk *chunk,
240                       int *rs, int *re, int end)
241{
242    *rs = find_next_zero_bit(chunk->populated, end, *rs);
243    *re = find_next_bit(chunk->populated, end, *rs + 1);
244}
245
246static void __maybe_unused pcpu_next_pop(struct pcpu_chunk *chunk,
247                     int *rs, int *re, int end)
248{
249    *rs = find_next_bit(chunk->populated, end, *rs);
250    *re = find_next_zero_bit(chunk->populated, end, *rs + 1);
251}
252
253/*
254 * (Un)populated page region iterators. Iterate over (un)populated
255 * page regions betwen @start and @end in @chunk. @rs and @re should
256 * be integer variables and will be set to start and end page index of
257 * the current region.
258 */
259#define pcpu_for_each_unpop_region(chunk, rs, re, start, end) \
260    for ((rs) = (start), pcpu_next_unpop((chunk), &(rs), &(re), (end)); \
261         (rs) < (re); \
262         (rs) = (re) + 1, pcpu_next_unpop((chunk), &(rs), &(re), (end)))
263
264#define pcpu_for_each_pop_region(chunk, rs, re, start, end) \
265    for ((rs) = (start), pcpu_next_pop((chunk), &(rs), &(re), (end)); \
266         (rs) < (re); \
267         (rs) = (re) + 1, pcpu_next_pop((chunk), &(rs), &(re), (end)))
268
269/**
270 * pcpu_mem_alloc - allocate memory
271 * @size: bytes to allocate
272 *
273 * Allocate @size bytes. If @size is smaller than PAGE_SIZE,
274 * kzalloc() is used; otherwise, vmalloc() is used. The returned
275 * memory is always zeroed.
276 *
277 * CONTEXT:
278 * Does GFP_KERNEL allocation.
279 *
280 * RETURNS:
281 * Pointer to the allocated area on success, NULL on failure.
282 */
283static void *pcpu_mem_alloc(size_t size)
284{
285    if (size <= PAGE_SIZE)
286        return kzalloc(size, GFP_KERNEL);
287    else {
288        void *ptr = vmalloc(size);
289        if (ptr)
290            memset(ptr, 0, size);
291        return ptr;
292    }
293}
294
295/**
296 * pcpu_mem_free - free memory
297 * @ptr: memory to free
298 * @size: size of the area
299 *
300 * Free @ptr. @ptr should have been allocated using pcpu_mem_alloc().
301 */
302static void pcpu_mem_free(void *ptr, size_t size)
303{
304    if (size <= PAGE_SIZE)
305        kfree(ptr);
306    else
307        vfree(ptr);
308}
309
310/**
311 * pcpu_chunk_relocate - put chunk in the appropriate chunk slot
312 * @chunk: chunk of interest
313 * @oslot: the previous slot it was on
314 *
315 * This function is called after an allocation or free changed @chunk.
316 * New slot according to the changed state is determined and @chunk is
317 * moved to the slot. Note that the reserved chunk is never put on
318 * chunk slots.
319 *
320 * CONTEXT:
321 * pcpu_lock.
322 */
323static void pcpu_chunk_relocate(struct pcpu_chunk *chunk, int oslot)
324{
325    int nslot = pcpu_chunk_slot(chunk);
326
327    if (chunk != pcpu_reserved_chunk && oslot != nslot) {
328        if (oslot < nslot)
329            list_move(&chunk->list, &pcpu_slot[nslot]);
330        else
331            list_move_tail(&chunk->list, &pcpu_slot[nslot]);
332    }
333}
334
335/**
336 * pcpu_need_to_extend - determine whether chunk area map needs to be extended
337 * @chunk: chunk of interest
338 *
339 * Determine whether area map of @chunk needs to be extended to
340 * accomodate a new allocation.
341 *
342 * CONTEXT:
343 * pcpu_lock.
344 *
345 * RETURNS:
346 * New target map allocation length if extension is necessary, 0
347 * otherwise.
348 */
349static int pcpu_need_to_extend(struct pcpu_chunk *chunk)
350{
351    int new_alloc;
352
353    if (chunk->map_alloc >= chunk->map_used + 2)
354        return 0;
355
356    new_alloc = PCPU_DFL_MAP_ALLOC;
357    while (new_alloc < chunk->map_used + 2)
358        new_alloc *= 2;
359
360    return new_alloc;
361}
362
363/**
364 * pcpu_extend_area_map - extend area map of a chunk
365 * @chunk: chunk of interest
366 * @new_alloc: new target allocation length of the area map
367 *
368 * Extend area map of @chunk to have @new_alloc entries.
369 *
370 * CONTEXT:
371 * Does GFP_KERNEL allocation. Grabs and releases pcpu_lock.
372 *
373 * RETURNS:
374 * 0 on success, -errno on failure.
375 */
376static int pcpu_extend_area_map(struct pcpu_chunk *chunk, int new_alloc)
377{
378    int *old = NULL, *new = NULL;
379    size_t old_size = 0, new_size = new_alloc * sizeof(new[0]);
380    unsigned long flags;
381
382    new = pcpu_mem_alloc(new_size);
383    if (!new)
384        return -ENOMEM;
385
386    /* acquire pcpu_lock and switch to new area map */
387    spin_lock_irqsave(&pcpu_lock, flags);
388
389    if (new_alloc <= chunk->map_alloc)
390        goto out_unlock;
391
392    old_size = chunk->map_alloc * sizeof(chunk->map[0]);
393    memcpy(new, chunk->map, old_size);
394
395    /*
396     * map_alloc < PCPU_DFL_MAP_ALLOC indicates that the chunk is
397     * one of the first chunks and still using static map.
398     */
399    if (chunk->map_alloc >= PCPU_DFL_MAP_ALLOC)
400        old = chunk->map;
401
402    chunk->map_alloc = new_alloc;
403    chunk->map = new;
404    new = NULL;
405
406out_unlock:
407    spin_unlock_irqrestore(&pcpu_lock, flags);
408
409    /*
410     * pcpu_mem_free() might end up calling vfree() which uses
411     * IRQ-unsafe lock and thus can't be called under pcpu_lock.
412     */
413    pcpu_mem_free(old, old_size);
414    pcpu_mem_free(new, new_size);
415
416    return 0;
417}
418
419/**
420 * pcpu_split_block - split a map block
421 * @chunk: chunk of interest
422 * @i: index of map block to split
423 * @head: head size in bytes (can be 0)
424 * @tail: tail size in bytes (can be 0)
425 *
426 * Split the @i'th map block into two or three blocks. If @head is
427 * non-zero, @head bytes block is inserted before block @i moving it
428 * to @i+1 and reducing its size by @head bytes.
429 *
430 * If @tail is non-zero, the target block, which can be @i or @i+1
431 * depending on @head, is reduced by @tail bytes and @tail byte block
432 * is inserted after the target block.
433 *
434 * @chunk->map must have enough free slots to accomodate the split.
435 *
436 * CONTEXT:
437 * pcpu_lock.
438 */
439static void pcpu_split_block(struct pcpu_chunk *chunk, int i,
440                 int head, int tail)
441{
442    int nr_extra = !!head + !!tail;
443
444    BUG_ON(chunk->map_alloc < chunk->map_used + nr_extra);
445
446    /* insert new subblocks */
447    memmove(&chunk->map[i + nr_extra], &chunk->map[i],
448        sizeof(chunk->map[0]) * (chunk->map_used - i));
449    chunk->map_used += nr_extra;
450
451    if (head) {
452        chunk->map[i + 1] = chunk->map[i] - head;
453        chunk->map[i++] = head;
454    }
455    if (tail) {
456        chunk->map[i++] -= tail;
457        chunk->map[i] = tail;
458    }
459}
460
461/**
462 * pcpu_alloc_area - allocate area from a pcpu_chunk
463 * @chunk: chunk of interest
464 * @size: wanted size in bytes
465 * @align: wanted align
466 *
467 * Try to allocate @size bytes area aligned at @align from @chunk.
468 * Note that this function only allocates the offset. It doesn't
469 * populate or map the area.
470 *
471 * @chunk->map must have at least two free slots.
472 *
473 * CONTEXT:
474 * pcpu_lock.
475 *
476 * RETURNS:
477 * Allocated offset in @chunk on success, -1 if no matching area is
478 * found.
479 */
480static int pcpu_alloc_area(struct pcpu_chunk *chunk, int size, int align)
481{
482    int oslot = pcpu_chunk_slot(chunk);
483    int max_contig = 0;
484    int i, off;
485
486    for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++])) {
487        bool is_last = i + 1 == chunk->map_used;
488        int head, tail;
489
490        /* extra for alignment requirement */
491        head = ALIGN(off, align) - off;
492        BUG_ON(i == 0 && head != 0);
493
494        if (chunk->map[i] < 0)
495            continue;
496        if (chunk->map[i] < head + size) {
497            max_contig = max(chunk->map[i], max_contig);
498            continue;
499        }
500
501        /*
502         * If head is small or the previous block is free,
503         * merge'em. Note that 'small' is defined as smaller
504         * than sizeof(int), which is very small but isn't too
505         * uncommon for percpu allocations.
506         */
507        if (head && (head < sizeof(int) || chunk->map[i - 1] > 0)) {
508            if (chunk->map[i - 1] > 0)
509                chunk->map[i - 1] += head;
510            else {
511                chunk->map[i - 1] -= head;
512                chunk->free_size -= head;
513            }
514            chunk->map[i] -= head;
515            off += head;
516            head = 0;
517        }
518
519        /* if tail is small, just keep it around */
520        tail = chunk->map[i] - head - size;
521        if (tail < sizeof(int))
522            tail = 0;
523
524        /* split if warranted */
525        if (head || tail) {
526            pcpu_split_block(chunk, i, head, tail);
527            if (head) {
528                i++;
529                off += head;
530                max_contig = max(chunk->map[i - 1], max_contig);
531            }
532            if (tail)
533                max_contig = max(chunk->map[i + 1], max_contig);
534        }
535
536        /* update hint and mark allocated */
537        if (is_last)
538            chunk->contig_hint = max_contig; /* fully scanned */
539        else
540            chunk->contig_hint = max(chunk->contig_hint,
541                         max_contig);
542
543        chunk->free_size -= chunk->map[i];
544        chunk->map[i] = -chunk->map[i];
545
546        pcpu_chunk_relocate(chunk, oslot);
547        return off;
548    }
549
550    chunk->contig_hint = max_contig; /* fully scanned */
551    pcpu_chunk_relocate(chunk, oslot);
552
553    /* tell the upper layer that this chunk has no matching area */
554    return -1;
555}
556
557/**
558 * pcpu_free_area - free area to a pcpu_chunk
559 * @chunk: chunk of interest
560 * @freeme: offset of area to free
561 *
562 * Free area starting from @freeme to @chunk. Note that this function
563 * only modifies the allocation map. It doesn't depopulate or unmap
564 * the area.
565 *
566 * CONTEXT:
567 * pcpu_lock.
568 */
569static void pcpu_free_area(struct pcpu_chunk *chunk, int freeme)
570{
571    int oslot = pcpu_chunk_slot(chunk);
572    int i, off;
573
574    for (i = 0, off = 0; i < chunk->map_used; off += abs(chunk->map[i++]))
575        if (off == freeme)
576            break;
577    BUG_ON(off != freeme);
578    BUG_ON(chunk->map[i] > 0);
579
580    chunk->map[i] = -chunk->map[i];
581    chunk->free_size += chunk->map[i];
582
583    /* merge with previous? */
584    if (i > 0 && chunk->map[i - 1] >= 0) {
585        chunk->map[i - 1] += chunk->map[i];
586        chunk->map_used--;
587        memmove(&chunk->map[i], &chunk->map[i + 1],
588            (chunk->map_used - i) * sizeof(chunk->map[0]));
589        i--;
590    }
591    /* merge with next? */
592    if (i + 1 < chunk->map_used && chunk->map[i + 1] >= 0) {
593        chunk->map[i] += chunk->map[i + 1];
594        chunk->map_used--;
595        memmove(&chunk->map[i + 1], &chunk->map[i + 2],
596            (chunk->map_used - (i + 1)) * sizeof(chunk->map[0]));
597    }
598
599    chunk->contig_hint = max(chunk->map[i], chunk->contig_hint);
600    pcpu_chunk_relocate(chunk, oslot);
601}
602
603static struct pcpu_chunk *pcpu_alloc_chunk(void)
604{
605    struct pcpu_chunk *chunk;
606
607    chunk = kzalloc(pcpu_chunk_struct_size, GFP_KERNEL);
608    if (!chunk)
609        return NULL;
610
611    chunk->map = pcpu_mem_alloc(PCPU_DFL_MAP_ALLOC * sizeof(chunk->map[0]));
612    if (!chunk->map) {
613        kfree(chunk);
614        return NULL;
615    }
616
617    chunk->map_alloc = PCPU_DFL_MAP_ALLOC;
618    chunk->map[chunk->map_used++] = pcpu_unit_size;
619
620    INIT_LIST_HEAD(&chunk->list);
621    chunk->free_size = pcpu_unit_size;
622    chunk->contig_hint = pcpu_unit_size;
623
624    return chunk;
625}
626
627static void pcpu_free_chunk(struct pcpu_chunk *chunk)
628{
629    if (!chunk)
630        return;
631    pcpu_mem_free(chunk->map, chunk->map_alloc * sizeof(chunk->map[0]));
632    kfree(chunk);
633}
634
635/*
636 * Chunk management implementation.
637 *
638 * To allow different implementations, chunk alloc/free and
639 * [de]population are implemented in a separate file which is pulled
640 * into this file and compiled together. The following functions
641 * should be implemented.
642 *
643 * pcpu_populate_chunk - populate the specified range of a chunk
644 * pcpu_depopulate_chunk - depopulate the specified range of a chunk
645 * pcpu_create_chunk - create a new chunk
646 * pcpu_destroy_chunk - destroy a chunk, always preceded by full depop
647 * pcpu_addr_to_page - translate address to physical address
648 * pcpu_verify_alloc_info - check alloc_info is acceptable during init
649 */
650static int pcpu_populate_chunk(struct pcpu_chunk *chunk, int off, int size);
651static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk, int off, int size);
652static struct pcpu_chunk *pcpu_create_chunk(void);
653static void pcpu_destroy_chunk(struct pcpu_chunk *chunk);
654static struct page *pcpu_addr_to_page(void *addr);
655static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai);
656
657#ifdef CONFIG_NEED_PER_CPU_KM
658#include "percpu-km.c"
659#else
660#include "percpu-vm.c"
661#endif
662
663/**
664 * pcpu_chunk_addr_search - determine chunk containing specified address
665 * @addr: address for which the chunk needs to be determined.
666 *
667 * RETURNS:
668 * The address of the found chunk.
669 */
670static struct pcpu_chunk *pcpu_chunk_addr_search(void *addr)
671{
672    /* is it in the first chunk? */
673    if (pcpu_addr_in_first_chunk(addr)) {
674        /* is it in the reserved area? */
675        if (pcpu_addr_in_reserved_chunk(addr))
676            return pcpu_reserved_chunk;
677        return pcpu_first_chunk;
678    }
679
680    /*
681     * The address is relative to unit0 which might be unused and
682     * thus unmapped. Offset the address to the unit space of the
683     * current processor before looking it up in the vmalloc
684     * space. Note that any possible cpu id can be used here, so
685     * there's no need to worry about preemption or cpu hotplug.
686     */
687    addr += pcpu_unit_offsets[raw_smp_processor_id()];
688    return pcpu_get_page_chunk(pcpu_addr_to_page(addr));
689}
690
691/**
692 * pcpu_alloc - the percpu allocator
693 * @size: size of area to allocate in bytes
694 * @align: alignment of area (max PAGE_SIZE)
695 * @reserved: allocate from the reserved chunk if available
696 *
697 * Allocate percpu area of @size bytes aligned at @align.
698 *
699 * CONTEXT:
700 * Does GFP_KERNEL allocation.
701 *
702 * RETURNS:
703 * Percpu pointer to the allocated area on success, NULL on failure.
704 */
705static void __percpu *pcpu_alloc(size_t size, size_t align, bool reserved)
706{
707    static int warn_limit = 10;
708    struct pcpu_chunk *chunk;
709    const char *err;
710    int slot, off, new_alloc;
711    unsigned long flags;
712
713    if (unlikely(!size || size > PCPU_MIN_UNIT_SIZE || align > PAGE_SIZE)) {
714        WARN(true, "illegal size (%zu) or align (%zu) for "
715             "percpu allocation\n", size, align);
716        return NULL;
717    }
718
719    mutex_lock(&pcpu_alloc_mutex);
720    spin_lock_irqsave(&pcpu_lock, flags);
721
722    /* serve reserved allocations from the reserved chunk if available */
723    if (reserved && pcpu_reserved_chunk) {
724        chunk = pcpu_reserved_chunk;
725
726        if (size > chunk->contig_hint) {
727            err = "alloc from reserved chunk failed";
728            goto fail_unlock;
729        }
730
731        while ((new_alloc = pcpu_need_to_extend(chunk))) {
732            spin_unlock_irqrestore(&pcpu_lock, flags);
733            if (pcpu_extend_area_map(chunk, new_alloc) < 0) {
734                err = "failed to extend area map of reserved chunk";
735                goto fail_unlock_mutex;
736            }
737            spin_lock_irqsave(&pcpu_lock, flags);
738        }
739
740        off = pcpu_alloc_area(chunk, size, align);
741        if (off >= 0)
742            goto area_found;
743
744        err = "alloc from reserved chunk failed";
745        goto fail_unlock;
746    }
747
748restart:
749    /* search through normal chunks */
750    for (slot = pcpu_size_to_slot(size); slot < pcpu_nr_slots; slot++) {
751        list_for_each_entry(chunk, &pcpu_slot[slot], list) {
752            if (size > chunk->contig_hint)
753                continue;
754
755            new_alloc = pcpu_need_to_extend(chunk);
756            if (new_alloc) {
757                spin_unlock_irqrestore(&pcpu_lock, flags);
758                if (pcpu_extend_area_map(chunk,
759                             new_alloc) < 0) {
760                    err = "failed to extend area map";
761                    goto fail_unlock_mutex;
762                }
763                spin_lock_irqsave(&pcpu_lock, flags);
764                /*
765                 * pcpu_lock has been dropped, need to
766                 * restart cpu_slot list walking.
767                 */
768                goto restart;
769            }
770
771            off = pcpu_alloc_area(chunk, size, align);
772            if (off >= 0)
773                goto area_found;
774        }
775    }
776
777    /* hmmm... no space left, create a new chunk */
778    spin_unlock_irqrestore(&pcpu_lock, flags);
779
780    chunk = pcpu_create_chunk();
781    if (!chunk) {
782        err = "failed to allocate new chunk";
783        goto fail_unlock_mutex;
784    }
785
786    spin_lock_irqsave(&pcpu_lock, flags);
787    pcpu_chunk_relocate(chunk, -1);
788    goto restart;
789
790area_found:
791    spin_unlock_irqrestore(&pcpu_lock, flags);
792
793    /* populate, map and clear the area */
794    if (pcpu_populate_chunk(chunk, off, size)) {
795        spin_lock_irqsave(&pcpu_lock, flags);
796        pcpu_free_area(chunk, off);
797        err = "failed to populate";
798        goto fail_unlock;
799    }
800
801    mutex_unlock(&pcpu_alloc_mutex);
802
803    /* return address relative to base address */
804    return __addr_to_pcpu_ptr(chunk->base_addr + off);
805
806fail_unlock:
807    spin_unlock_irqrestore(&pcpu_lock, flags);
808fail_unlock_mutex:
809    mutex_unlock(&pcpu_alloc_mutex);
810    if (warn_limit) {
811        pr_warning("PERCPU: allocation failed, size=%zu align=%zu, "
812               "%s\n", size, align, err);
813        dump_stack();
814        if (!--warn_limit)
815            pr_info("PERCPU: limit reached, disable warning\n");
816    }
817    return NULL;
818}
819
820/**
821 * __alloc_percpu - allocate dynamic percpu area
822 * @size: size of area to allocate in bytes
823 * @align: alignment of area (max PAGE_SIZE)
824 *
825 * Allocate percpu area of @size bytes aligned at @align. Might
826 * sleep. Might trigger writeouts.
827 *
828 * CONTEXT:
829 * Does GFP_KERNEL allocation.
830 *
831 * RETURNS:
832 * Percpu pointer to the allocated area on success, NULL on failure.
833 */
834void __percpu *__alloc_percpu(size_t size, size_t align)
835{
836    return pcpu_alloc(size, align, false);
837}
838EXPORT_SYMBOL_GPL(__alloc_percpu);
839
840/**
841 * __alloc_reserved_percpu - allocate reserved percpu area
842 * @size: size of area to allocate in bytes
843 * @align: alignment of area (max PAGE_SIZE)
844 *
845 * Allocate percpu area of @size bytes aligned at @align from reserved
846 * percpu area if arch has set it up; otherwise, allocation is served
847 * from the same dynamic area. Might sleep. Might trigger writeouts.
848 *
849 * CONTEXT:
850 * Does GFP_KERNEL allocation.
851 *
852 * RETURNS:
853 * Percpu pointer to the allocated area on success, NULL on failure.
854 */
855void __percpu *__alloc_reserved_percpu(size_t size, size_t align)
856{
857    return pcpu_alloc(size, align, true);
858}
859
860/**
861 * pcpu_reclaim - reclaim fully free chunks, workqueue function
862 * @work: unused
863 *
864 * Reclaim all fully free chunks except for the first one.
865 *
866 * CONTEXT:
867 * workqueue context.
868 */
869static void pcpu_reclaim(struct work_struct *work)
870{
871    LIST_HEAD(todo);
872    struct list_head *head = &pcpu_slot[pcpu_nr_slots - 1];
873    struct pcpu_chunk *chunk, *next;
874
875    mutex_lock(&pcpu_alloc_mutex);
876    spin_lock_irq(&pcpu_lock);
877
878    list_for_each_entry_safe(chunk, next, head, list) {
879        WARN_ON(chunk->immutable);
880
881        /* spare the first one */
882        if (chunk == list_first_entry(head, struct pcpu_chunk, list))
883            continue;
884
885        list_move(&chunk->list, &todo);
886    }
887
888    spin_unlock_irq(&pcpu_lock);
889
890    list_for_each_entry_safe(chunk, next, &todo, list) {
891        pcpu_depopulate_chunk(chunk, 0, pcpu_unit_size);
892        pcpu_destroy_chunk(chunk);
893    }
894
895    mutex_unlock(&pcpu_alloc_mutex);
896}
897
898/**
899 * free_percpu - free percpu area
900 * @ptr: pointer to area to free
901 *
902 * Free percpu area @ptr.
903 *
904 * CONTEXT:
905 * Can be called from atomic context.
906 */
907void free_percpu(void __percpu *ptr)
908{
909    void *addr;
910    struct pcpu_chunk *chunk;
911    unsigned long flags;
912    int off;
913
914    if (!ptr)
915        return;
916
917    addr = __pcpu_ptr_to_addr(ptr);
918
919    spin_lock_irqsave(&pcpu_lock, flags);
920
921    chunk = pcpu_chunk_addr_search(addr);
922    off = addr - chunk->base_addr;
923
924    pcpu_free_area(chunk, off);
925
926    /* if there are more than one fully free chunks, wake up grim reaper */
927    if (chunk->free_size == pcpu_unit_size) {
928        struct pcpu_chunk *pos;
929
930        list_for_each_entry(pos, &pcpu_slot[pcpu_nr_slots - 1], list)
931            if (pos != chunk) {
932                schedule_work(&pcpu_reclaim_work);
933                break;
934            }
935    }
936
937    spin_unlock_irqrestore(&pcpu_lock, flags);
938}
939EXPORT_SYMBOL_GPL(free_percpu);
940
941/**
942 * is_kernel_percpu_address - test whether address is from static percpu area
943 * @addr: address to test
944 *
945 * Test whether @addr belongs to in-kernel static percpu area. Module
946 * static percpu areas are not considered. For those, use
947 * is_module_percpu_address().
948 *
949 * RETURNS:
950 * %true if @addr is from in-kernel static percpu area, %false otherwise.
951 */
952bool is_kernel_percpu_address(unsigned long addr)
953{
954    const size_t static_size = __per_cpu_end - __per_cpu_start;
955    void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
956    unsigned int cpu;
957
958    for_each_possible_cpu(cpu) {
959        void *start = per_cpu_ptr(base, cpu);
960
961        if ((void *)addr >= start && (void *)addr < start + static_size)
962            return true;
963        }
964    return false;
965}
966
967/**
968 * per_cpu_ptr_to_phys - convert translated percpu address to physical address
969 * @addr: the address to be converted to physical address
970 *
971 * Given @addr which is dereferenceable address obtained via one of
972 * percpu access macros, this function translates it into its physical
973 * address. The caller is responsible for ensuring @addr stays valid
974 * until this function finishes.
975 *
976 * RETURNS:
977 * The physical address for @addr.
978 */
979phys_addr_t per_cpu_ptr_to_phys(void *addr)
980{
981    void __percpu *base = __addr_to_pcpu_ptr(pcpu_base_addr);
982    bool in_first_chunk = false;
983    unsigned long first_start, first_end;
984    unsigned int cpu;
985
986    /*
987     * The following test on first_start/end isn't strictly
988     * necessary but will speed up lookups of addresses which
989     * aren't in the first chunk.
990     */
991    first_start = pcpu_chunk_addr(pcpu_first_chunk, pcpu_first_unit_cpu, 0);
992    first_end = pcpu_chunk_addr(pcpu_first_chunk, pcpu_last_unit_cpu,
993                    pcpu_unit_pages);
994    if ((unsigned long)addr >= first_start &&
995        (unsigned long)addr < first_end) {
996        for_each_possible_cpu(cpu) {
997            void *start = per_cpu_ptr(base, cpu);
998
999            if (addr >= start && addr < start + pcpu_unit_size) {
1000                in_first_chunk = true;
1001                break;
1002            }
1003        }
1004    }
1005
1006    if (in_first_chunk) {
1007        if ((unsigned long)addr < VMALLOC_START ||
1008            (unsigned long)addr >= VMALLOC_END)
1009            return __pa(addr);
1010        else
1011            return page_to_phys(vmalloc_to_page(addr));
1012    } else
1013        return page_to_phys(pcpu_addr_to_page(addr));
1014}
1015
1016static inline size_t pcpu_calc_fc_sizes(size_t static_size,
1017                    size_t reserved_size,
1018                    ssize_t *dyn_sizep)
1019{
1020    size_t size_sum;
1021
1022    size_sum = PFN_ALIGN(static_size + reserved_size +
1023                 (*dyn_sizep >= 0 ? *dyn_sizep : 0));
1024    if (*dyn_sizep != 0)
1025        *dyn_sizep = size_sum - static_size - reserved_size;
1026
1027    return size_sum;
1028}
1029
1030/**
1031 * pcpu_alloc_alloc_info - allocate percpu allocation info
1032 * @nr_groups: the number of groups
1033 * @nr_units: the number of units
1034 *
1035 * Allocate ai which is large enough for @nr_groups groups containing
1036 * @nr_units units. The returned ai's groups[0].cpu_map points to the
1037 * cpu_map array which is long enough for @nr_units and filled with
1038 * NR_CPUS. It's the caller's responsibility to initialize cpu_map
1039 * pointer of other groups.
1040 *
1041 * RETURNS:
1042 * Pointer to the allocated pcpu_alloc_info on success, NULL on
1043 * failure.
1044 */
1045struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
1046                              int nr_units)
1047{
1048    struct pcpu_alloc_info *ai;
1049    size_t base_size, ai_size;
1050    void *ptr;
1051    int unit;
1052
1053    base_size = ALIGN(sizeof(*ai) + nr_groups * sizeof(ai->groups[0]),
1054              __alignof__(ai->groups[0].cpu_map[0]));
1055    ai_size = base_size + nr_units * sizeof(ai->groups[0].cpu_map[0]);
1056
1057    ptr = alloc_bootmem_nopanic(PFN_ALIGN(ai_size));
1058    if (!ptr)
1059        return NULL;
1060    ai = ptr;
1061    ptr += base_size;
1062
1063    ai->groups[0].cpu_map = ptr;
1064
1065    for (unit = 0; unit < nr_units; unit++)
1066        ai->groups[0].cpu_map[unit] = NR_CPUS;
1067
1068    ai->nr_groups = nr_groups;
1069    ai->__ai_size = PFN_ALIGN(ai_size);
1070
1071    return ai;
1072}
1073
1074/**
1075 * pcpu_free_alloc_info - free percpu allocation info
1076 * @ai: pcpu_alloc_info to free
1077 *
1078 * Free @ai which was allocated by pcpu_alloc_alloc_info().
1079 */
1080void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai)
1081{
1082    free_bootmem(__pa(ai), ai->__ai_size);
1083}
1084
1085/**
1086 * pcpu_build_alloc_info - build alloc_info considering distances between CPUs
1087 * @reserved_size: the size of reserved percpu area in bytes
1088 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1089 * @atom_size: allocation atom size
1090 * @cpu_distance_fn: callback to determine distance between cpus, optional
1091 *
1092 * This function determines grouping of units, their mappings to cpus
1093 * and other parameters considering needed percpu size, allocation
1094 * atom size and distances between CPUs.
1095 *
1096 * Groups are always mutliples of atom size and CPUs which are of
1097 * LOCAL_DISTANCE both ways are grouped together and share space for
1098 * units in the same group. The returned configuration is guaranteed
1099 * to have CPUs on different nodes on different groups and >=75% usage
1100 * of allocated virtual address space.
1101 *
1102 * RETURNS:
1103 * On success, pointer to the new allocation_info is returned. On
1104 * failure, ERR_PTR value is returned.
1105 */
1106struct pcpu_alloc_info * __init pcpu_build_alloc_info(
1107                size_t reserved_size, ssize_t dyn_size,
1108                size_t atom_size,
1109                pcpu_fc_cpu_distance_fn_t cpu_distance_fn)
1110{
1111    static int group_map[NR_CPUS] __initdata;
1112    static int group_cnt[NR_CPUS] __initdata;
1113    const size_t static_size = __per_cpu_end - __per_cpu_start;
1114    int nr_groups = 1, nr_units = 0;
1115    size_t size_sum, min_unit_size, alloc_size;
1116    int upa, max_upa, uninitialized_var(best_upa); /* units_per_alloc */
1117    int last_allocs, group, unit;
1118    unsigned int cpu, tcpu;
1119    struct pcpu_alloc_info *ai;
1120    unsigned int *cpu_map;
1121
1122    /* this function may be called multiple times */
1123    memset(group_map, 0, sizeof(group_map));
1124    memset(group_cnt, 0, sizeof(group_cnt));
1125
1126    /*
1127     * Determine min_unit_size, alloc_size and max_upa such that
1128     * alloc_size is multiple of atom_size and is the smallest
1129     * which can accomodate 4k aligned segments which are equal to
1130     * or larger than min_unit_size.
1131     */
1132    size_sum = pcpu_calc_fc_sizes(static_size, reserved_size, &dyn_size);
1133    min_unit_size = max_t(size_t, size_sum, PCPU_MIN_UNIT_SIZE);
1134
1135    alloc_size = roundup(min_unit_size, atom_size);
1136    upa = alloc_size / min_unit_size;
1137    while (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1138        upa--;
1139    max_upa = upa;
1140
1141    /* group cpus according to their proximity */
1142    for_each_possible_cpu(cpu) {
1143        group = 0;
1144    next_group:
1145        for_each_possible_cpu(tcpu) {
1146            if (cpu == tcpu)
1147                break;
1148            if (group_map[tcpu] == group && cpu_distance_fn &&
1149                (cpu_distance_fn(cpu, tcpu) > LOCAL_DISTANCE ||
1150                 cpu_distance_fn(tcpu, cpu) > LOCAL_DISTANCE)) {
1151                group++;
1152                nr_groups = max(nr_groups, group + 1);
1153                goto next_group;
1154            }
1155        }
1156        group_map[cpu] = group;
1157        group_cnt[group]++;
1158    }
1159
1160    /*
1161     * Expand unit size until address space usage goes over 75%
1162     * and then as much as possible without using more address
1163     * space.
1164     */
1165    last_allocs = INT_MAX;
1166    for (upa = max_upa; upa; upa--) {
1167        int allocs = 0, wasted = 0;
1168
1169        if (alloc_size % upa || ((alloc_size / upa) & ~PAGE_MASK))
1170            continue;
1171
1172        for (group = 0; group < nr_groups; group++) {
1173            int this_allocs = DIV_ROUND_UP(group_cnt[group], upa);
1174            allocs += this_allocs;
1175            wasted += this_allocs * upa - group_cnt[group];
1176        }
1177
1178        /*
1179         * Don't accept if wastage is over 25%. The
1180         * greater-than comparison ensures upa==1 always
1181         * passes the following check.
1182         */
1183        if (wasted > num_possible_cpus() / 3)
1184            continue;
1185
1186        /* and then don't consume more memory */
1187        if (allocs > last_allocs)
1188            break;
1189        last_allocs = allocs;
1190        best_upa = upa;
1191    }
1192    upa = best_upa;
1193
1194    /* allocate and fill alloc_info */
1195    for (group = 0; group < nr_groups; group++)
1196        nr_units += roundup(group_cnt[group], upa);
1197
1198    ai = pcpu_alloc_alloc_info(nr_groups, nr_units);
1199    if (!ai)
1200        return ERR_PTR(-ENOMEM);
1201    cpu_map = ai->groups[0].cpu_map;
1202
1203    for (group = 0; group < nr_groups; group++) {
1204        ai->groups[group].cpu_map = cpu_map;
1205        cpu_map += roundup(group_cnt[group], upa);
1206    }
1207
1208    ai->static_size = static_size;
1209    ai->reserved_size = reserved_size;
1210    ai->dyn_size = dyn_size;
1211    ai->unit_size = alloc_size / upa;
1212    ai->atom_size = atom_size;
1213    ai->alloc_size = alloc_size;
1214
1215    for (group = 0, unit = 0; group_cnt[group]; group++) {
1216        struct pcpu_group_info *gi = &ai->groups[group];
1217
1218        /*
1219         * Initialize base_offset as if all groups are located
1220         * back-to-back. The caller should update this to
1221         * reflect actual allocation.
1222         */
1223        gi->base_offset = unit * ai->unit_size;
1224
1225        for_each_possible_cpu(cpu)
1226            if (group_map[cpu] == group)
1227                gi->cpu_map[gi->nr_units++] = cpu;
1228        gi->nr_units = roundup(gi->nr_units, upa);
1229        unit += gi->nr_units;
1230    }
1231    BUG_ON(unit != nr_units);
1232
1233    return ai;
1234}
1235
1236/**
1237 * pcpu_dump_alloc_info - print out information about pcpu_alloc_info
1238 * @lvl: loglevel
1239 * @ai: allocation info to dump
1240 *
1241 * Print out information about @ai using loglevel @lvl.
1242 */
1243static void pcpu_dump_alloc_info(const char *lvl,
1244                 const struct pcpu_alloc_info *ai)
1245{
1246    int group_width = 1, cpu_width = 1, width;
1247    char empty_str[] = "--------";
1248    int alloc = 0, alloc_end = 0;
1249    int group, v;
1250    int upa, apl; /* units per alloc, allocs per line */
1251
1252    v = ai->nr_groups;
1253    while (v /= 10)
1254        group_width++;
1255
1256    v = num_possible_cpus();
1257    while (v /= 10)
1258        cpu_width++;
1259    empty_str[min_t(int, cpu_width, sizeof(empty_str) - 1)] = '\0';
1260
1261    upa = ai->alloc_size / ai->unit_size;
1262    width = upa * (cpu_width + 1) + group_width + 3;
1263    apl = rounddown_pow_of_two(max(60 / width, 1));
1264
1265    printk("%spcpu-alloc: s%zu r%zu d%zu u%zu alloc=%zu*%zu",
1266           lvl, ai->static_size, ai->reserved_size, ai->dyn_size,
1267           ai->unit_size, ai->alloc_size / ai->atom_size, ai->atom_size);
1268
1269    for (group = 0; group < ai->nr_groups; group++) {
1270        const struct pcpu_group_info *gi = &ai->groups[group];
1271        int unit = 0, unit_end = 0;
1272
1273        BUG_ON(gi->nr_units % upa);
1274        for (alloc_end += gi->nr_units / upa;
1275             alloc < alloc_end; alloc++) {
1276            if (!(alloc % apl)) {
1277                printk("\n");
1278                printk("%spcpu-alloc: ", lvl);
1279            }
1280            printk("[%0*d] ", group_width, group);
1281
1282            for (unit_end += upa; unit < unit_end; unit++)
1283                if (gi->cpu_map[unit] != NR_CPUS)
1284                    printk("%0*d ", cpu_width,
1285                           gi->cpu_map[unit]);
1286                else
1287                    printk("%s ", empty_str);
1288        }
1289    }
1290    printk("\n");
1291}
1292
1293/**
1294 * pcpu_setup_first_chunk - initialize the first percpu chunk
1295 * @ai: pcpu_alloc_info describing how to percpu area is shaped
1296 * @base_addr: mapped address
1297 *
1298 * Initialize the first percpu chunk which contains the kernel static
1299 * perpcu area. This function is to be called from arch percpu area
1300 * setup path.
1301 *
1302 * @ai contains all information necessary to initialize the first
1303 * chunk and prime the dynamic percpu allocator.
1304 *
1305 * @ai->static_size is the size of static percpu area.
1306 *
1307 * @ai->reserved_size, if non-zero, specifies the amount of bytes to
1308 * reserve after the static area in the first chunk. This reserves
1309 * the first chunk such that it's available only through reserved
1310 * percpu allocation. This is primarily used to serve module percpu
1311 * static areas on architectures where the addressing model has
1312 * limited offset range for symbol relocations to guarantee module
1313 * percpu symbols fall inside the relocatable range.
1314 *
1315 * @ai->dyn_size determines the number of bytes available for dynamic
1316 * allocation in the first chunk. The area between @ai->static_size +
1317 * @ai->reserved_size + @ai->dyn_size and @ai->unit_size is unused.
1318 *
1319 * @ai->unit_size specifies unit size and must be aligned to PAGE_SIZE
1320 * and equal to or larger than @ai->static_size + @ai->reserved_size +
1321 * @ai->dyn_size.
1322 *
1323 * @ai->atom_size is the allocation atom size and used as alignment
1324 * for vm areas.
1325 *
1326 * @ai->alloc_size is the allocation size and always multiple of
1327 * @ai->atom_size. This is larger than @ai->atom_size if
1328 * @ai->unit_size is larger than @ai->atom_size.
1329 *
1330 * @ai->nr_groups and @ai->groups describe virtual memory layout of
1331 * percpu areas. Units which should be colocated are put into the
1332 * same group. Dynamic VM areas will be allocated according to these
1333 * groupings. If @ai->nr_groups is zero, a single group containing
1334 * all units is assumed.
1335 *
1336 * The caller should have mapped the first chunk at @base_addr and
1337 * copied static data to each unit.
1338 *
1339 * If the first chunk ends up with both reserved and dynamic areas, it
1340 * is served by two chunks - one to serve the core static and reserved
1341 * areas and the other for the dynamic area. They share the same vm
1342 * and page map but uses different area allocation map to stay away
1343 * from each other. The latter chunk is circulated in the chunk slots
1344 * and available for dynamic allocation like any other chunks.
1345 *
1346 * RETURNS:
1347 * 0 on success, -errno on failure.
1348 */
1349int __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
1350                  void *base_addr)
1351{
1352    static char cpus_buf[4096] __initdata;
1353    static int smap[2], dmap[2];
1354    size_t dyn_size = ai->dyn_size;
1355    size_t size_sum = ai->static_size + ai->reserved_size + dyn_size;
1356    struct pcpu_chunk *schunk, *dchunk = NULL;
1357    unsigned long *group_offsets;
1358    size_t *group_sizes;
1359    unsigned long *unit_off;
1360    unsigned int cpu;
1361    int *unit_map;
1362    int group, unit, i;
1363
1364    cpumask_scnprintf(cpus_buf, sizeof(cpus_buf), cpu_possible_mask);
1365
1366#define PCPU_SETUP_BUG_ON(cond) do { \
1367    if (unlikely(cond)) { \
1368        pr_emerg("PERCPU: failed to initialize, %s", #cond); \
1369        pr_emerg("PERCPU: cpu_possible_mask=%s\n", cpus_buf); \
1370        pcpu_dump_alloc_info(KERN_EMERG, ai); \
1371        BUG(); \
1372    } \
1373} while (0)
1374
1375    /* sanity checks */
1376    BUILD_BUG_ON(ARRAY_SIZE(smap) >= PCPU_DFL_MAP_ALLOC ||
1377             ARRAY_SIZE(dmap) >= PCPU_DFL_MAP_ALLOC);
1378    PCPU_SETUP_BUG_ON(ai->nr_groups <= 0);
1379    PCPU_SETUP_BUG_ON(!ai->static_size);
1380    PCPU_SETUP_BUG_ON(!base_addr);
1381    PCPU_SETUP_BUG_ON(ai->unit_size < size_sum);
1382    PCPU_SETUP_BUG_ON(ai->unit_size & ~PAGE_MASK);
1383    PCPU_SETUP_BUG_ON(ai->unit_size < PCPU_MIN_UNIT_SIZE);
1384    PCPU_SETUP_BUG_ON(pcpu_verify_alloc_info(ai) < 0);
1385
1386    /* process group information and build config tables accordingly */
1387    group_offsets = alloc_bootmem(ai->nr_groups * sizeof(group_offsets[0]));
1388    group_sizes = alloc_bootmem(ai->nr_groups * sizeof(group_sizes[0]));
1389    unit_map = alloc_bootmem(nr_cpu_ids * sizeof(unit_map[0]));
1390    unit_off = alloc_bootmem(nr_cpu_ids * sizeof(unit_off[0]));
1391
1392    for (cpu = 0; cpu < nr_cpu_ids; cpu++)
1393        unit_map[cpu] = UINT_MAX;
1394    pcpu_first_unit_cpu = NR_CPUS;
1395
1396    for (group = 0, unit = 0; group < ai->nr_groups; group++, unit += i) {
1397        const struct pcpu_group_info *gi = &ai->groups[group];
1398
1399        group_offsets[group] = gi->base_offset;
1400        group_sizes[group] = gi->nr_units * ai->unit_size;
1401
1402        for (i = 0; i < gi->nr_units; i++) {
1403            cpu = gi->cpu_map[i];
1404            if (cpu == NR_CPUS)
1405                continue;
1406
1407            PCPU_SETUP_BUG_ON(cpu > nr_cpu_ids);
1408            PCPU_SETUP_BUG_ON(!cpu_possible(cpu));
1409            PCPU_SETUP_BUG_ON(unit_map[cpu] != UINT_MAX);
1410
1411            unit_map[cpu] = unit + i;
1412            unit_off[cpu] = gi->base_offset + i * ai->unit_size;
1413
1414            if (pcpu_first_unit_cpu == NR_CPUS)
1415                pcpu_first_unit_cpu = cpu;
1416        }
1417    }
1418    pcpu_last_unit_cpu = cpu;
1419    pcpu_nr_units = unit;
1420
1421    for_each_possible_cpu(cpu)
1422        PCPU_SETUP_BUG_ON(unit_map[cpu] == UINT_MAX);
1423
1424    /* we're done parsing the input, undefine BUG macro and dump config */
1425#undef PCPU_SETUP_BUG_ON
1426    pcpu_dump_alloc_info(KERN_INFO, ai);
1427
1428    pcpu_nr_groups = ai->nr_groups;
1429    pcpu_group_offsets = group_offsets;
1430    pcpu_group_sizes = group_sizes;
1431    pcpu_unit_map = unit_map;
1432    pcpu_unit_offsets = unit_off;
1433
1434    /* determine basic parameters */
1435    pcpu_unit_pages = ai->unit_size >> PAGE_SHIFT;
1436    pcpu_unit_size = pcpu_unit_pages << PAGE_SHIFT;
1437    pcpu_atom_size = ai->atom_size;
1438    pcpu_chunk_struct_size = sizeof(struct pcpu_chunk) +
1439        BITS_TO_LONGS(pcpu_unit_pages) * sizeof(unsigned long);
1440
1441    /*
1442     * Allocate chunk slots. The additional last slot is for
1443     * empty chunks.
1444     */
1445    pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
1446    pcpu_slot = alloc_bootmem(pcpu_nr_slots * sizeof(pcpu_slot[0]));
1447    for (i = 0; i < pcpu_nr_slots; i++)
1448        INIT_LIST_HEAD(&pcpu_slot[i]);
1449
1450    /*
1451     * Initialize static chunk. If reserved_size is zero, the
1452     * static chunk covers static area + dynamic allocation area
1453     * in the first chunk. If reserved_size is not zero, it
1454     * covers static area + reserved area (mostly used for module
1455     * static percpu allocation).
1456     */
1457    schunk = alloc_bootmem(pcpu_chunk_struct_size);
1458    INIT_LIST_HEAD(&schunk->list);
1459    schunk->base_addr = base_addr;
1460    schunk->map = smap;
1461    schunk->map_alloc = ARRAY_SIZE(smap);
1462    schunk->immutable = true;
1463    bitmap_fill(schunk->populated, pcpu_unit_pages);
1464
1465    if (ai->reserved_size) {
1466        schunk->free_size = ai->reserved_size;
1467        pcpu_reserved_chunk = schunk;
1468        pcpu_reserved_chunk_limit = ai->static_size + ai->reserved_size;
1469    } else {
1470        schunk->free_size = dyn_size;
1471        dyn_size = 0; /* dynamic area covered */
1472    }
1473    schunk->contig_hint = schunk->free_size;
1474
1475    schunk->map[schunk->map_used++] = -ai->static_size;
1476    if (schunk->free_size)
1477        schunk->map[schunk->map_used++] = schunk->free_size;
1478
1479    /* init dynamic chunk if necessary */
1480    if (dyn_size) {
1481        dchunk = alloc_bootmem(pcpu_chunk_struct_size);
1482        INIT_LIST_HEAD(&dchunk->list);
1483        dchunk->base_addr = base_addr;
1484        dchunk->map = dmap;
1485        dchunk->map_alloc = ARRAY_SIZE(dmap);
1486        dchunk->immutable = true;
1487        bitmap_fill(dchunk->populated, pcpu_unit_pages);
1488
1489        dchunk->contig_hint = dchunk->free_size = dyn_size;
1490        dchunk->map[dchunk->map_used++] = -pcpu_reserved_chunk_limit;
1491        dchunk->map[dchunk->map_used++] = dchunk->free_size;
1492    }
1493
1494    /* link the first chunk in */
1495    pcpu_first_chunk = dchunk ?: schunk;
1496    pcpu_chunk_relocate(pcpu_first_chunk, -1);
1497
1498    /* we're done */
1499    pcpu_base_addr = base_addr;
1500    return 0;
1501}
1502
1503const char *pcpu_fc_names[PCPU_FC_NR] __initdata = {
1504    [PCPU_FC_AUTO] = "auto",
1505    [PCPU_FC_EMBED] = "embed",
1506    [PCPU_FC_PAGE] = "page",
1507};
1508
1509enum pcpu_fc pcpu_chosen_fc __initdata = PCPU_FC_AUTO;
1510
1511static int __init percpu_alloc_setup(char *str)
1512{
1513    if (0)
1514        /* nada */;
1515#ifdef CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK
1516    else if (!strcmp(str, "embed"))
1517        pcpu_chosen_fc = PCPU_FC_EMBED;
1518#endif
1519#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1520    else if (!strcmp(str, "page"))
1521        pcpu_chosen_fc = PCPU_FC_PAGE;
1522#endif
1523    else
1524        pr_warning("PERCPU: unknown allocator %s specified\n", str);
1525
1526    return 0;
1527}
1528early_param("percpu_alloc", percpu_alloc_setup);
1529
1530#if defined(CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK) || \
1531    !defined(CONFIG_HAVE_SETUP_PER_CPU_AREA)
1532/**
1533 * pcpu_embed_first_chunk - embed the first percpu chunk into bootmem
1534 * @reserved_size: the size of reserved percpu area in bytes
1535 * @dyn_size: free size for dynamic allocation in bytes, -1 for auto
1536 * @atom_size: allocation atom size
1537 * @cpu_distance_fn: callback to determine distance between cpus, optional
1538 * @alloc_fn: function to allocate percpu page
1539 * @free_fn: funtion to free percpu page
1540 *
1541 * This is a helper to ease setting up embedded first percpu chunk and
1542 * can be called where pcpu_setup_first_chunk() is expected.
1543 *
1544 * If this function is used to setup the first chunk, it is allocated
1545 * by calling @alloc_fn and used as-is without being mapped into
1546 * vmalloc area. Allocations are always whole multiples of @atom_size
1547 * aligned to @atom_size.
1548 *
1549 * This enables the first chunk to piggy back on the linear physical
1550 * mapping which often uses larger page size. Please note that this
1551 * can result in very sparse cpu->unit mapping on NUMA machines thus
1552 * requiring large vmalloc address space. Don't use this allocator if
1553 * vmalloc space is not orders of magnitude larger than distances
1554 * between node memory addresses (ie. 32bit NUMA machines).
1555 *
1556 * When @dyn_size is positive, dynamic area might be larger than
1557 * specified to fill page alignment. When @dyn_size is auto,
1558 * @dyn_size is just big enough to fill page alignment after static
1559 * and reserved areas.
1560 *
1561 * If the needed size is smaller than the minimum or specified unit
1562 * size, the leftover is returned using @free_fn.
1563 *
1564 * RETURNS:
1565 * 0 on success, -errno on failure.
1566 */
1567int __init pcpu_embed_first_chunk(size_t reserved_size, ssize_t dyn_size,
1568                  size_t atom_size,
1569                  pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
1570                  pcpu_fc_alloc_fn_t alloc_fn,
1571                  pcpu_fc_free_fn_t free_fn)
1572{
1573    void *base = (void *)ULONG_MAX;
1574    void **areas = NULL;
1575    struct pcpu_alloc_info *ai;
1576    size_t size_sum, areas_size, max_distance;
1577    int group, i, rc;
1578
1579    ai = pcpu_build_alloc_info(reserved_size, dyn_size, atom_size,
1580                   cpu_distance_fn);
1581    if (IS_ERR(ai))
1582        return PTR_ERR(ai);
1583
1584    size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
1585    areas_size = PFN_ALIGN(ai->nr_groups * sizeof(void *));
1586
1587    areas = alloc_bootmem_nopanic(areas_size);
1588    if (!areas) {
1589        rc = -ENOMEM;
1590        goto out_free;
1591    }
1592
1593    /* allocate, copy and determine base address */
1594    for (group = 0; group < ai->nr_groups; group++) {
1595        struct pcpu_group_info *gi = &ai->groups[group];
1596        unsigned int cpu = NR_CPUS;
1597        void *ptr;
1598
1599        for (i = 0; i < gi->nr_units && cpu == NR_CPUS; i++)
1600            cpu = gi->cpu_map[i];
1601        BUG_ON(cpu == NR_CPUS);
1602
1603        /* allocate space for the whole group */
1604        ptr = alloc_fn(cpu, gi->nr_units * ai->unit_size, atom_size);
1605        if (!ptr) {
1606            rc = -ENOMEM;
1607            goto out_free_areas;
1608        }
1609        areas[group] = ptr;
1610
1611        base = min(ptr, base);
1612
1613        for (i = 0; i < gi->nr_units; i++, ptr += ai->unit_size) {
1614            if (gi->cpu_map[i] == NR_CPUS) {
1615                /* unused unit, free whole */
1616                free_fn(ptr, ai->unit_size);
1617                continue;
1618            }
1619            /* copy and return the unused part */
1620            memcpy(ptr, __per_cpu_load, ai->static_size);
1621            free_fn(ptr + size_sum, ai->unit_size - size_sum);
1622        }
1623    }
1624
1625    /* base address is now known, determine group base offsets */
1626    max_distance = 0;
1627    for (group = 0; group < ai->nr_groups; group++) {
1628        ai->groups[group].base_offset = areas[group] - base;
1629        max_distance = max_t(size_t, max_distance,
1630                     ai->groups[group].base_offset);
1631    }
1632    max_distance += ai->unit_size;
1633
1634    /* warn if maximum distance is further than 75% of vmalloc space */
1635    if (max_distance > (VMALLOC_END - VMALLOC_START) * 3 / 4) {
1636        pr_warning("PERCPU: max_distance=0x%zx too large for vmalloc "
1637               "space 0x%lx\n",
1638               max_distance, VMALLOC_END - VMALLOC_START);
1639#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1640        /* and fail if we have fallback */
1641        rc = -EINVAL;
1642        goto out_free;
1643#endif
1644    }
1645
1646    pr_info("PERCPU: Embedded %zu pages/cpu @%p s%zu r%zu d%zu u%zu\n",
1647        PFN_DOWN(size_sum), base, ai->static_size, ai->reserved_size,
1648        ai->dyn_size, ai->unit_size);
1649
1650    rc = pcpu_setup_first_chunk(ai, base);
1651    goto out_free;
1652
1653out_free_areas:
1654    for (group = 0; group < ai->nr_groups; group++)
1655        free_fn(areas[group],
1656            ai->groups[group].nr_units * ai->unit_size);
1657out_free:
1658    pcpu_free_alloc_info(ai);
1659    if (areas)
1660        free_bootmem(__pa(areas), areas_size);
1661    return rc;
1662}
1663#endif /* CONFIG_NEED_PER_CPU_EMBED_FIRST_CHUNK ||
1664      !CONFIG_HAVE_SETUP_PER_CPU_AREA */
1665
1666#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
1667/**
1668 * pcpu_page_first_chunk - map the first chunk using PAGE_SIZE pages
1669 * @reserved_size: the size of reserved percpu area in bytes
1670 * @alloc_fn: function to allocate percpu page, always called with PAGE_SIZE
1671 * @free_fn: funtion to free percpu page, always called with PAGE_SIZE
1672 * @populate_pte_fn: function to populate pte
1673 *
1674 * This is a helper to ease setting up page-remapped first percpu
1675 * chunk and can be called where pcpu_setup_first_chunk() is expected.
1676 *
1677 * This is the basic allocator. Static percpu area is allocated
1678 * page-by-page into vmalloc area.
1679 *
1680 * RETURNS:
1681 * 0 on success, -errno on failure.
1682 */
1683int __init pcpu_page_first_chunk(size_t reserved_size,
1684                 pcpu_fc_alloc_fn_t alloc_fn,
1685                 pcpu_fc_free_fn_t free_fn,
1686                 pcpu_fc_populate_pte_fn_t populate_pte_fn)
1687{
1688    static struct vm_struct vm;
1689    struct pcpu_alloc_info *ai;
1690    char psize_str[16];
1691    int unit_pages;
1692    size_t pages_size;
1693    struct page **pages;
1694    int unit, i, j, rc;
1695
1696    snprintf(psize_str, sizeof(psize_str), "%luK", PAGE_SIZE >> 10);
1697
1698    ai = pcpu_build_alloc_info(reserved_size, -1, PAGE_SIZE, NULL);
1699    if (IS_ERR(ai))
1700        return PTR_ERR(ai);
1701    BUG_ON(ai->nr_groups != 1);
1702    BUG_ON(ai->groups[0].nr_units != num_possible_cpus());
1703
1704    unit_pages = ai->unit_size >> PAGE_SHIFT;
1705
1706    /* unaligned allocations can't be freed, round up to page size */
1707    pages_size = PFN_ALIGN(unit_pages * num_possible_cpus() *
1708                   sizeof(pages[0]));
1709    pages = alloc_bootmem(pages_size);
1710
1711    /* allocate pages */
1712    j = 0;
1713    for (unit = 0; unit < num_possible_cpus(); unit++)
1714        for (i = 0; i < unit_pages; i++) {
1715            unsigned int cpu = ai->groups[0].cpu_map[unit];
1716            void *ptr;
1717
1718            ptr = alloc_fn(cpu, PAGE_SIZE, PAGE_SIZE);
1719            if (!ptr) {
1720                pr_warning("PERCPU: failed to allocate %s page "
1721                       "for cpu%u\n", psize_str, cpu);
1722                goto enomem;
1723            }
1724            pages[j++] = virt_to_page(ptr);
1725        }
1726
1727    /* allocate vm area, map the pages and copy static data */
1728    vm.flags = VM_ALLOC;
1729    vm.size = num_possible_cpus() * ai->unit_size;
1730    vm_area_register_early(&vm, PAGE_SIZE);
1731
1732    for (unit = 0; unit < num_possible_cpus(); unit++) {
1733        unsigned long unit_addr =
1734            (unsigned long)vm.addr + unit * ai->unit_size;
1735
1736        for (i = 0; i < unit_pages; i++)
1737            populate_pte_fn(unit_addr + (i << PAGE_SHIFT));
1738
1739        /* pte already populated, the following shouldn't fail */
1740        rc = __pcpu_map_pages(unit_addr, &pages[unit * unit_pages],
1741                      unit_pages);
1742        if (rc < 0)
1743            panic("failed to map percpu area, err=%d\n", rc);
1744
1745        /*
1746         * FIXME: Archs with virtual cache should flush local
1747         * cache for the linear mapping here - something
1748         * equivalent to flush_cache_vmap() on the local cpu.
1749         * flush_cache_vmap() can't be used as most supporting
1750         * data structures are not set up yet.
1751         */
1752
1753        /* copy static data */
1754        memcpy((void *)unit_addr, __per_cpu_load, ai->static_size);
1755    }
1756
1757    /* we're ready, commit */
1758    pr_info("PERCPU: %d %s pages/cpu @%p s%zu r%zu d%zu\n",
1759        unit_pages, psize_str, vm.addr, ai->static_size,
1760        ai->reserved_size, ai->dyn_size);
1761
1762    rc = pcpu_setup_first_chunk(ai, vm.addr);
1763    goto out_free_ar;
1764
1765enomem:
1766    while (--j >= 0)
1767        free_fn(page_address(pages[j]), PAGE_SIZE);
1768    rc = -ENOMEM;
1769out_free_ar:
1770    free_bootmem(__pa(pages), pages_size);
1771    pcpu_free_alloc_info(ai);
1772    return rc;
1773}
1774#endif /* CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK */
1775
1776/*
1777 * Generic percpu area setup.
1778 *
1779 * The embedding helper is used because its behavior closely resembles
1780 * the original non-dynamic generic percpu area setup. This is
1781 * important because many archs have addressing restrictions and might
1782 * fail if the percpu area is located far away from the previous
1783 * location. As an added bonus, in non-NUMA cases, embedding is
1784 * generally a good idea TLB-wise because percpu area can piggy back
1785 * on the physical linear memory mapping which uses large page
1786 * mappings on applicable archs.
1787 */
1788#ifndef CONFIG_HAVE_SETUP_PER_CPU_AREA
1789unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
1790EXPORT_SYMBOL(__per_cpu_offset);
1791
1792static void * __init pcpu_dfl_fc_alloc(unsigned int cpu, size_t size,
1793                       size_t align)
1794{
1795    return __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
1796}
1797
1798static void __init pcpu_dfl_fc_free(void *ptr, size_t size)
1799{
1800    free_bootmem(__pa(ptr), size);
1801}
1802
1803void __init setup_per_cpu_areas(void)
1804{
1805    unsigned long delta;
1806    unsigned int cpu;
1807    int rc;
1808
1809    /*
1810     * Always reserve area for module percpu variables. That's
1811     * what the legacy allocator did.
1812     */
1813    rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1814                    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE, NULL,
1815                    pcpu_dfl_fc_alloc, pcpu_dfl_fc_free);
1816    if (rc < 0)
1817        panic("Failed to initialized percpu areas.");
1818
1819    delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1820    for_each_possible_cpu(cpu)
1821        __per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
1822}
1823#endif /* CONFIG_HAVE_SETUP_PER_CPU_AREA */
1824

Archive Download this file



interactive