xref: /OK3568_Linux_fs/kernel/mm/percpu-stats.c (revision 4882a59341e53eb6f0b4789bf948001014eff981)
1*4882a593Smuzhiyun // SPDX-License-Identifier: GPL-2.0-only
2*4882a593Smuzhiyun /*
3*4882a593Smuzhiyun  * mm/percpu-debug.c
4*4882a593Smuzhiyun  *
5*4882a593Smuzhiyun  * Copyright (C) 2017		Facebook Inc.
6*4882a593Smuzhiyun  * Copyright (C) 2017		Dennis Zhou <dennis@kernel.org>
7*4882a593Smuzhiyun  *
8*4882a593Smuzhiyun  * Prints statistics about the percpu allocator and backing chunks.
9*4882a593Smuzhiyun  */
10*4882a593Smuzhiyun #include <linux/debugfs.h>
11*4882a593Smuzhiyun #include <linux/list.h>
12*4882a593Smuzhiyun #include <linux/percpu.h>
13*4882a593Smuzhiyun #include <linux/seq_file.h>
14*4882a593Smuzhiyun #include <linux/sort.h>
15*4882a593Smuzhiyun #include <linux/vmalloc.h>
16*4882a593Smuzhiyun 
17*4882a593Smuzhiyun #include "percpu-internal.h"
18*4882a593Smuzhiyun 
19*4882a593Smuzhiyun #define P(X, Y) \
20*4882a593Smuzhiyun 	seq_printf(m, "  %-20s: %12lld\n", X, (long long int)Y)
21*4882a593Smuzhiyun 
22*4882a593Smuzhiyun struct percpu_stats pcpu_stats;
23*4882a593Smuzhiyun struct pcpu_alloc_info pcpu_stats_ai;
24*4882a593Smuzhiyun 
cmpint(const void * a,const void * b)25*4882a593Smuzhiyun static int cmpint(const void *a, const void *b)
26*4882a593Smuzhiyun {
27*4882a593Smuzhiyun 	return *(int *)a - *(int *)b;
28*4882a593Smuzhiyun }
29*4882a593Smuzhiyun 
30*4882a593Smuzhiyun /*
31*4882a593Smuzhiyun  * Iterates over all chunks to find the max nr_alloc entries.
32*4882a593Smuzhiyun  */
find_max_nr_alloc(void)33*4882a593Smuzhiyun static int find_max_nr_alloc(void)
34*4882a593Smuzhiyun {
35*4882a593Smuzhiyun 	struct pcpu_chunk *chunk;
36*4882a593Smuzhiyun 	int slot, max_nr_alloc;
37*4882a593Smuzhiyun 	enum pcpu_chunk_type type;
38*4882a593Smuzhiyun 
39*4882a593Smuzhiyun 	max_nr_alloc = 0;
40*4882a593Smuzhiyun 	for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++)
41*4882a593Smuzhiyun 		for (slot = 0; slot < pcpu_nr_slots; slot++)
42*4882a593Smuzhiyun 			list_for_each_entry(chunk, &pcpu_chunk_list(type)[slot],
43*4882a593Smuzhiyun 					    list)
44*4882a593Smuzhiyun 				max_nr_alloc = max(max_nr_alloc,
45*4882a593Smuzhiyun 						   chunk->nr_alloc);
46*4882a593Smuzhiyun 
47*4882a593Smuzhiyun 	return max_nr_alloc;
48*4882a593Smuzhiyun }
49*4882a593Smuzhiyun 
50*4882a593Smuzhiyun /*
51*4882a593Smuzhiyun  * Prints out chunk state. Fragmentation is considered between
52*4882a593Smuzhiyun  * the beginning of the chunk to the last allocation.
53*4882a593Smuzhiyun  *
54*4882a593Smuzhiyun  * All statistics are in bytes unless stated otherwise.
55*4882a593Smuzhiyun  */
chunk_map_stats(struct seq_file * m,struct pcpu_chunk * chunk,int * buffer)56*4882a593Smuzhiyun static void chunk_map_stats(struct seq_file *m, struct pcpu_chunk *chunk,
57*4882a593Smuzhiyun 			    int *buffer)
58*4882a593Smuzhiyun {
59*4882a593Smuzhiyun 	struct pcpu_block_md *chunk_md = &chunk->chunk_md;
60*4882a593Smuzhiyun 	int i, last_alloc, as_len, start, end;
61*4882a593Smuzhiyun 	int *alloc_sizes, *p;
62*4882a593Smuzhiyun 	/* statistics */
63*4882a593Smuzhiyun 	int sum_frag = 0, max_frag = 0;
64*4882a593Smuzhiyun 	int cur_min_alloc = 0, cur_med_alloc = 0, cur_max_alloc = 0;
65*4882a593Smuzhiyun 
66*4882a593Smuzhiyun 	alloc_sizes = buffer;
67*4882a593Smuzhiyun 
68*4882a593Smuzhiyun 	/*
69*4882a593Smuzhiyun 	 * find_last_bit returns the start value if nothing found.
70*4882a593Smuzhiyun 	 * Therefore, we must determine if it is a failure of find_last_bit
71*4882a593Smuzhiyun 	 * and set the appropriate value.
72*4882a593Smuzhiyun 	 */
73*4882a593Smuzhiyun 	last_alloc = find_last_bit(chunk->alloc_map,
74*4882a593Smuzhiyun 				   pcpu_chunk_map_bits(chunk) -
75*4882a593Smuzhiyun 				   chunk->end_offset / PCPU_MIN_ALLOC_SIZE - 1);
76*4882a593Smuzhiyun 	last_alloc = test_bit(last_alloc, chunk->alloc_map) ?
77*4882a593Smuzhiyun 		     last_alloc + 1 : 0;
78*4882a593Smuzhiyun 
79*4882a593Smuzhiyun 	as_len = 0;
80*4882a593Smuzhiyun 	start = chunk->start_offset / PCPU_MIN_ALLOC_SIZE;
81*4882a593Smuzhiyun 
82*4882a593Smuzhiyun 	/*
83*4882a593Smuzhiyun 	 * If a bit is set in the allocation map, the bound_map identifies
84*4882a593Smuzhiyun 	 * where the allocation ends.  If the allocation is not set, the
85*4882a593Smuzhiyun 	 * bound_map does not identify free areas as it is only kept accurate
86*4882a593Smuzhiyun 	 * on allocation, not free.
87*4882a593Smuzhiyun 	 *
88*4882a593Smuzhiyun 	 * Positive values are allocations and negative values are free
89*4882a593Smuzhiyun 	 * fragments.
90*4882a593Smuzhiyun 	 */
91*4882a593Smuzhiyun 	while (start < last_alloc) {
92*4882a593Smuzhiyun 		if (test_bit(start, chunk->alloc_map)) {
93*4882a593Smuzhiyun 			end = find_next_bit(chunk->bound_map, last_alloc,
94*4882a593Smuzhiyun 					    start + 1);
95*4882a593Smuzhiyun 			alloc_sizes[as_len] = 1;
96*4882a593Smuzhiyun 		} else {
97*4882a593Smuzhiyun 			end = find_next_bit(chunk->alloc_map, last_alloc,
98*4882a593Smuzhiyun 					    start + 1);
99*4882a593Smuzhiyun 			alloc_sizes[as_len] = -1;
100*4882a593Smuzhiyun 		}
101*4882a593Smuzhiyun 
102*4882a593Smuzhiyun 		alloc_sizes[as_len++] *= (end - start) * PCPU_MIN_ALLOC_SIZE;
103*4882a593Smuzhiyun 
104*4882a593Smuzhiyun 		start = end;
105*4882a593Smuzhiyun 	}
106*4882a593Smuzhiyun 
107*4882a593Smuzhiyun 	/*
108*4882a593Smuzhiyun 	 * The negative values are free fragments and thus sorting gives the
109*4882a593Smuzhiyun 	 * free fragments at the beginning in largest first order.
110*4882a593Smuzhiyun 	 */
111*4882a593Smuzhiyun 	if (as_len > 0) {
112*4882a593Smuzhiyun 		sort(alloc_sizes, as_len, sizeof(int), cmpint, NULL);
113*4882a593Smuzhiyun 
114*4882a593Smuzhiyun 		/* iterate through the unallocated fragments */
115*4882a593Smuzhiyun 		for (i = 0, p = alloc_sizes; *p < 0 && i < as_len; i++, p++) {
116*4882a593Smuzhiyun 			sum_frag -= *p;
117*4882a593Smuzhiyun 			max_frag = max(max_frag, -1 * (*p));
118*4882a593Smuzhiyun 		}
119*4882a593Smuzhiyun 
120*4882a593Smuzhiyun 		cur_min_alloc = alloc_sizes[i];
121*4882a593Smuzhiyun 		cur_med_alloc = alloc_sizes[(i + as_len - 1) / 2];
122*4882a593Smuzhiyun 		cur_max_alloc = alloc_sizes[as_len - 1];
123*4882a593Smuzhiyun 	}
124*4882a593Smuzhiyun 
125*4882a593Smuzhiyun 	P("nr_alloc", chunk->nr_alloc);
126*4882a593Smuzhiyun 	P("max_alloc_size", chunk->max_alloc_size);
127*4882a593Smuzhiyun 	P("empty_pop_pages", chunk->nr_empty_pop_pages);
128*4882a593Smuzhiyun 	P("first_bit", chunk_md->first_free);
129*4882a593Smuzhiyun 	P("free_bytes", chunk->free_bytes);
130*4882a593Smuzhiyun 	P("contig_bytes", chunk_md->contig_hint * PCPU_MIN_ALLOC_SIZE);
131*4882a593Smuzhiyun 	P("sum_frag", sum_frag);
132*4882a593Smuzhiyun 	P("max_frag", max_frag);
133*4882a593Smuzhiyun 	P("cur_min_alloc", cur_min_alloc);
134*4882a593Smuzhiyun 	P("cur_med_alloc", cur_med_alloc);
135*4882a593Smuzhiyun 	P("cur_max_alloc", cur_max_alloc);
136*4882a593Smuzhiyun #ifdef CONFIG_MEMCG_KMEM
137*4882a593Smuzhiyun 	P("memcg_aware", pcpu_is_memcg_chunk(pcpu_chunk_type(chunk)));
138*4882a593Smuzhiyun #endif
139*4882a593Smuzhiyun 	seq_putc(m, '\n');
140*4882a593Smuzhiyun }
141*4882a593Smuzhiyun 
percpu_stats_show(struct seq_file * m,void * v)142*4882a593Smuzhiyun static int percpu_stats_show(struct seq_file *m, void *v)
143*4882a593Smuzhiyun {
144*4882a593Smuzhiyun 	struct pcpu_chunk *chunk;
145*4882a593Smuzhiyun 	int slot, max_nr_alloc;
146*4882a593Smuzhiyun 	int *buffer;
147*4882a593Smuzhiyun 	enum pcpu_chunk_type type;
148*4882a593Smuzhiyun 	int nr_empty_pop_pages;
149*4882a593Smuzhiyun 
150*4882a593Smuzhiyun alloc_buffer:
151*4882a593Smuzhiyun 	spin_lock_irq(&pcpu_lock);
152*4882a593Smuzhiyun 	max_nr_alloc = find_max_nr_alloc();
153*4882a593Smuzhiyun 	spin_unlock_irq(&pcpu_lock);
154*4882a593Smuzhiyun 
155*4882a593Smuzhiyun 	/* there can be at most this many free and allocated fragments */
156*4882a593Smuzhiyun 	buffer = vmalloc(array_size(sizeof(int), (2 * max_nr_alloc + 1)));
157*4882a593Smuzhiyun 	if (!buffer)
158*4882a593Smuzhiyun 		return -ENOMEM;
159*4882a593Smuzhiyun 
160*4882a593Smuzhiyun 	spin_lock_irq(&pcpu_lock);
161*4882a593Smuzhiyun 
162*4882a593Smuzhiyun 	/* if the buffer allocated earlier is too small */
163*4882a593Smuzhiyun 	if (max_nr_alloc < find_max_nr_alloc()) {
164*4882a593Smuzhiyun 		spin_unlock_irq(&pcpu_lock);
165*4882a593Smuzhiyun 		vfree(buffer);
166*4882a593Smuzhiyun 		goto alloc_buffer;
167*4882a593Smuzhiyun 	}
168*4882a593Smuzhiyun 
169*4882a593Smuzhiyun 	nr_empty_pop_pages = 0;
170*4882a593Smuzhiyun 	for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++)
171*4882a593Smuzhiyun 		nr_empty_pop_pages += pcpu_nr_empty_pop_pages[type];
172*4882a593Smuzhiyun 
173*4882a593Smuzhiyun #define PL(X)								\
174*4882a593Smuzhiyun 	seq_printf(m, "  %-20s: %12lld\n", #X, (long long int)pcpu_stats_ai.X)
175*4882a593Smuzhiyun 
176*4882a593Smuzhiyun 	seq_printf(m,
177*4882a593Smuzhiyun 			"Percpu Memory Statistics\n"
178*4882a593Smuzhiyun 			"Allocation Info:\n"
179*4882a593Smuzhiyun 			"----------------------------------------\n");
180*4882a593Smuzhiyun 	PL(unit_size);
181*4882a593Smuzhiyun 	PL(static_size);
182*4882a593Smuzhiyun 	PL(reserved_size);
183*4882a593Smuzhiyun 	PL(dyn_size);
184*4882a593Smuzhiyun 	PL(atom_size);
185*4882a593Smuzhiyun 	PL(alloc_size);
186*4882a593Smuzhiyun 	seq_putc(m, '\n');
187*4882a593Smuzhiyun 
188*4882a593Smuzhiyun #undef PL
189*4882a593Smuzhiyun 
190*4882a593Smuzhiyun #define PU(X) \
191*4882a593Smuzhiyun 	seq_printf(m, "  %-20s: %12llu\n", #X, (unsigned long long)pcpu_stats.X)
192*4882a593Smuzhiyun 
193*4882a593Smuzhiyun 	seq_printf(m,
194*4882a593Smuzhiyun 			"Global Stats:\n"
195*4882a593Smuzhiyun 			"----------------------------------------\n");
196*4882a593Smuzhiyun 	PU(nr_alloc);
197*4882a593Smuzhiyun 	PU(nr_dealloc);
198*4882a593Smuzhiyun 	PU(nr_cur_alloc);
199*4882a593Smuzhiyun 	PU(nr_max_alloc);
200*4882a593Smuzhiyun 	PU(nr_chunks);
201*4882a593Smuzhiyun 	PU(nr_max_chunks);
202*4882a593Smuzhiyun 	PU(min_alloc_size);
203*4882a593Smuzhiyun 	PU(max_alloc_size);
204*4882a593Smuzhiyun 	P("empty_pop_pages", nr_empty_pop_pages);
205*4882a593Smuzhiyun 	seq_putc(m, '\n');
206*4882a593Smuzhiyun 
207*4882a593Smuzhiyun #undef PU
208*4882a593Smuzhiyun 
209*4882a593Smuzhiyun 	seq_printf(m,
210*4882a593Smuzhiyun 			"Per Chunk Stats:\n"
211*4882a593Smuzhiyun 			"----------------------------------------\n");
212*4882a593Smuzhiyun 
213*4882a593Smuzhiyun 	if (pcpu_reserved_chunk) {
214*4882a593Smuzhiyun 		seq_puts(m, "Chunk: <- Reserved Chunk\n");
215*4882a593Smuzhiyun 		chunk_map_stats(m, pcpu_reserved_chunk, buffer);
216*4882a593Smuzhiyun 	}
217*4882a593Smuzhiyun 
218*4882a593Smuzhiyun 	for (type = 0; type < PCPU_NR_CHUNK_TYPES; type++) {
219*4882a593Smuzhiyun 		for (slot = 0; slot < pcpu_nr_slots; slot++) {
220*4882a593Smuzhiyun 			list_for_each_entry(chunk, &pcpu_chunk_list(type)[slot],
221*4882a593Smuzhiyun 					    list) {
222*4882a593Smuzhiyun 				if (chunk == pcpu_first_chunk) {
223*4882a593Smuzhiyun 					seq_puts(m, "Chunk: <- First Chunk\n");
224*4882a593Smuzhiyun 					chunk_map_stats(m, chunk, buffer);
225*4882a593Smuzhiyun 				} else {
226*4882a593Smuzhiyun 					seq_puts(m, "Chunk:\n");
227*4882a593Smuzhiyun 					chunk_map_stats(m, chunk, buffer);
228*4882a593Smuzhiyun 				}
229*4882a593Smuzhiyun 			}
230*4882a593Smuzhiyun 		}
231*4882a593Smuzhiyun 	}
232*4882a593Smuzhiyun 
233*4882a593Smuzhiyun 	spin_unlock_irq(&pcpu_lock);
234*4882a593Smuzhiyun 
235*4882a593Smuzhiyun 	vfree(buffer);
236*4882a593Smuzhiyun 
237*4882a593Smuzhiyun 	return 0;
238*4882a593Smuzhiyun }
239*4882a593Smuzhiyun DEFINE_SHOW_ATTRIBUTE(percpu_stats);
240*4882a593Smuzhiyun 
init_percpu_stats_debugfs(void)241*4882a593Smuzhiyun static int __init init_percpu_stats_debugfs(void)
242*4882a593Smuzhiyun {
243*4882a593Smuzhiyun 	debugfs_create_file("percpu_stats", 0444, NULL, NULL,
244*4882a593Smuzhiyun 			&percpu_stats_fops);
245*4882a593Smuzhiyun 
246*4882a593Smuzhiyun 	return 0;
247*4882a593Smuzhiyun }
248*4882a593Smuzhiyun 
249*4882a593Smuzhiyun late_initcall(init_percpu_stats_debugfs);
250