xref: /linux/mm/percpu-km.c (revision f79e4d5f92a129a1159c973735007d4ddc8541f3)
1 /*
2  * mm/percpu-km.c - kernel memory based chunk allocation
3  *
4  * Copyright (C) 2010		SUSE Linux Products GmbH
5  * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
6  *
7  * This file is released under the GPLv2.
8  *
9  * Chunks are allocated as a contiguous kernel memory using gfp
10  * allocation.  This is to be used on nommu architectures.
11  *
12  * To use percpu-km,
13  *
14  * - define CONFIG_NEED_PER_CPU_KM from the arch Kconfig.
15  *
16  * - CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK must not be defined.  It's
17  *   not compatible with PER_CPU_KM.  EMBED_FIRST_CHUNK should work
18  *   fine.
19  *
20  * - NUMA is not supported.  When setting up the first chunk,
21  *   @cpu_distance_fn should be NULL or report all CPUs to be nearer
22  *   than or at LOCAL_DISTANCE.
23  *
24  * - It's best if the chunk size is power of two multiple of
25  *   PAGE_SIZE.  Because each chunk is allocated as a contiguous
26  *   kernel memory block using alloc_pages(), memory will be wasted if
27  *   chunk size is not aligned.  percpu-km code will whine about it.
28  */
29 
30 #if defined(CONFIG_SMP) && defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
31 #error "contiguous percpu allocation is incompatible with paged first chunk"
32 #endif
33 
34 #include <linux/log2.h>
35 
36 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
37 			       int page_start, int page_end, gfp_t gfp)
38 {
39 	return 0;
40 }
41 
42 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
43 				  int page_start, int page_end)
44 {
45 	/* nada */
46 }
47 
48 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp)
49 {
50 	const int nr_pages = pcpu_group_sizes[0] >> PAGE_SHIFT;
51 	struct pcpu_chunk *chunk;
52 	struct page *pages;
53 	int i;
54 
55 	chunk = pcpu_alloc_chunk(gfp);
56 	if (!chunk)
57 		return NULL;
58 
59 	pages = alloc_pages(gfp, order_base_2(nr_pages));
60 	if (!pages) {
61 		pcpu_free_chunk(chunk);
62 		return NULL;
63 	}
64 
65 	for (i = 0; i < nr_pages; i++)
66 		pcpu_set_page_chunk(nth_page(pages, i), chunk);
67 
68 	chunk->data = pages;
69 	chunk->base_addr = page_address(pages) - pcpu_group_offsets[0];
70 
71 	spin_lock_irq(&pcpu_lock);
72 	pcpu_chunk_populated(chunk, 0, nr_pages, false);
73 	spin_unlock_irq(&pcpu_lock);
74 
75 	pcpu_stats_chunk_alloc();
76 	trace_percpu_create_chunk(chunk->base_addr);
77 
78 	return chunk;
79 }
80 
81 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
82 {
83 	const int nr_pages = pcpu_group_sizes[0] >> PAGE_SHIFT;
84 
85 	if (!chunk)
86 		return;
87 
88 	pcpu_stats_chunk_dealloc();
89 	trace_percpu_destroy_chunk(chunk->base_addr);
90 
91 	if (chunk->data)
92 		__free_pages(chunk->data, order_base_2(nr_pages));
93 	pcpu_free_chunk(chunk);
94 }
95 
96 static struct page *pcpu_addr_to_page(void *addr)
97 {
98 	return virt_to_page(addr);
99 }
100 
101 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
102 {
103 	size_t nr_pages, alloc_pages;
104 
105 	/* all units must be in a single group */
106 	if (ai->nr_groups != 1) {
107 		pr_crit("can't handle more than one group\n");
108 		return -EINVAL;
109 	}
110 
111 	nr_pages = (ai->groups[0].nr_units * ai->unit_size) >> PAGE_SHIFT;
112 	alloc_pages = roundup_pow_of_two(nr_pages);
113 
114 	if (alloc_pages > nr_pages)
115 		pr_warn("wasting %zu pages per chunk\n",
116 			alloc_pages - nr_pages);
117 
118 	return 0;
119 }
120