]>
Commit | Line | Data |
---|---|---|
45e898b7 WL |
1 | /* |
2 | * This program is free software; you can redistribute it and/or modify | |
3 | * it under the terms of the GNU General Public License as published by | |
4 | * the Free Software Foundation; either version 2 of the License, or | |
5 | * (at your option) any later version. | |
6 | * | |
7 | * This program is distributed in the hope that it will be useful, | |
8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
10 | * GNU General Public License for more details. | |
11 | * | |
12 | * Authors: Waiman Long <waiman.long@hpe.com> | |
13 | */ | |
14 | ||
15 | /* | |
16 | * When queued spinlock statistical counters are enabled, the following | |
17 | * debugfs files will be created for reporting the counter values: | |
18 | * | |
19 | * <debugfs>/qlockstat/ | |
20 | * pv_hash_hops - average # of hops per hashing operation | |
21 | * pv_kick_unlock - # of vCPU kicks issued at unlock time | |
22 | * pv_kick_wake - # of vCPU kicks used for computing pv_latency_wake | |
23 | * pv_latency_kick - average latency (ns) of vCPU kick operation | |
24 | * pv_latency_wake - average latency (ns) from vCPU kick to wakeup | |
32d62510 | 25 | * pv_lock_slowpath - # of locking operations via the slowpath |
1c4941fd | 26 | * pv_lock_stealing - # of lock stealing operations |
08be8f63 WL |
27 | * pv_spurious_wakeup - # of spurious wakeups in non-head vCPUs |
28 | * pv_wait_again - # of wait's after a queue head vCPU kick | |
cd0272fa | 29 | * pv_wait_early - # of early vCPU wait's |
45e898b7 WL |
30 | * pv_wait_head - # of vCPU wait's at the queue head |
31 | * pv_wait_node - # of vCPU wait's at a non-head queue node | |
32 | * | |
33 | * Writing to the "reset_counters" file will reset all the above counter | |
34 | * values. | |
35 | * | |
36 | * These statistical counters are implemented as per-cpu variables which are | |
37 | * summed and computed whenever the corresponding debugfs files are read. This | |
38 | * minimizes added overhead making the counters usable even in a production | |
39 | * environment. | |
40 | * | |
41 | * There may be slight difference between pv_kick_wake and pv_kick_unlock. | |
42 | */ | |
43 | enum qlock_stats { | |
44 | qstat_pv_hash_hops, | |
45 | qstat_pv_kick_unlock, | |
46 | qstat_pv_kick_wake, | |
47 | qstat_pv_latency_kick, | |
48 | qstat_pv_latency_wake, | |
32d62510 | 49 | qstat_pv_lock_slowpath, |
1c4941fd | 50 | qstat_pv_lock_stealing, |
45e898b7 WL |
51 | qstat_pv_spurious_wakeup, |
52 | qstat_pv_wait_again, | |
cd0272fa | 53 | qstat_pv_wait_early, |
45e898b7 WL |
54 | qstat_pv_wait_head, |
55 | qstat_pv_wait_node, | |
56 | qstat_num, /* Total number of statistical counters */ | |
57 | qstat_reset_cnts = qstat_num, | |
58 | }; | |
59 | ||
60 | #ifdef CONFIG_QUEUED_LOCK_STAT | |
61 | /* | |
62 | * Collect pvqspinlock statistics | |
63 | */ | |
64 | #include <linux/debugfs.h> | |
65 | #include <linux/sched.h> | |
e6017571 | 66 | #include <linux/sched/clock.h> |
45e898b7 WL |
67 | #include <linux/fs.h> |
68 | ||
69 | static const char * const qstat_names[qstat_num + 1] = { | |
70 | [qstat_pv_hash_hops] = "pv_hash_hops", | |
71 | [qstat_pv_kick_unlock] = "pv_kick_unlock", | |
72 | [qstat_pv_kick_wake] = "pv_kick_wake", | |
73 | [qstat_pv_spurious_wakeup] = "pv_spurious_wakeup", | |
74 | [qstat_pv_latency_kick] = "pv_latency_kick", | |
75 | [qstat_pv_latency_wake] = "pv_latency_wake", | |
32d62510 | 76 | [qstat_pv_lock_slowpath] = "pv_lock_slowpath", |
1c4941fd | 77 | [qstat_pv_lock_stealing] = "pv_lock_stealing", |
45e898b7 | 78 | [qstat_pv_wait_again] = "pv_wait_again", |
cd0272fa | 79 | [qstat_pv_wait_early] = "pv_wait_early", |
45e898b7 WL |
80 | [qstat_pv_wait_head] = "pv_wait_head", |
81 | [qstat_pv_wait_node] = "pv_wait_node", | |
82 | [qstat_reset_cnts] = "reset_counters", | |
83 | }; | |
84 | ||
85 | /* | |
86 | * Per-cpu counters | |
87 | */ | |
88 | static DEFINE_PER_CPU(unsigned long, qstats[qstat_num]); | |
89 | static DEFINE_PER_CPU(u64, pv_kick_time); | |
90 | ||
91 | /* | |
92 | * Function to read and return the qlock statistical counter values | |
93 | * | |
94 | * The following counters are handled specially: | |
95 | * 1. qstat_pv_latency_kick | |
96 | * Average kick latency (ns) = pv_latency_kick/pv_kick_unlock | |
97 | * 2. qstat_pv_latency_wake | |
98 | * Average wake latency (ns) = pv_latency_wake/pv_kick_wake | |
99 | * 3. qstat_pv_hash_hops | |
100 | * Average hops/hash = pv_hash_hops/pv_kick_unlock | |
101 | */ | |
102 | static ssize_t qstat_read(struct file *file, char __user *user_buf, | |
103 | size_t count, loff_t *ppos) | |
104 | { | |
105 | char buf[64]; | |
106 | int cpu, counter, len; | |
107 | u64 stat = 0, kicks = 0; | |
108 | ||
109 | /* | |
110 | * Get the counter ID stored in file->f_inode->i_private | |
111 | */ | |
45063097 | 112 | counter = (long)file_inode(file)->i_private; |
45e898b7 WL |
113 | |
114 | if (counter >= qstat_num) | |
115 | return -EBADF; | |
116 | ||
117 | for_each_possible_cpu(cpu) { | |
118 | stat += per_cpu(qstats[counter], cpu); | |
119 | /* | |
120 | * Need to sum additional counter for some of them | |
121 | */ | |
122 | switch (counter) { | |
123 | ||
124 | case qstat_pv_latency_kick: | |
125 | case qstat_pv_hash_hops: | |
126 | kicks += per_cpu(qstats[qstat_pv_kick_unlock], cpu); | |
127 | break; | |
128 | ||
129 | case qstat_pv_latency_wake: | |
130 | kicks += per_cpu(qstats[qstat_pv_kick_wake], cpu); | |
131 | break; | |
132 | } | |
133 | } | |
134 | ||
135 | if (counter == qstat_pv_hash_hops) { | |
66876595 | 136 | u64 frac = 0; |
45e898b7 | 137 | |
66876595 DB |
138 | if (kicks) { |
139 | frac = 100ULL * do_div(stat, kicks); | |
140 | frac = DIV_ROUND_CLOSEST_ULL(frac, kicks); | |
141 | } | |
45e898b7 WL |
142 | |
143 | /* | |
144 | * Return a X.XX decimal number | |
145 | */ | |
146 | len = snprintf(buf, sizeof(buf) - 1, "%llu.%02llu\n", stat, frac); | |
147 | } else { | |
148 | /* | |
149 | * Round to the nearest ns | |
150 | */ | |
151 | if ((counter == qstat_pv_latency_kick) || | |
152 | (counter == qstat_pv_latency_wake)) { | |
45e898b7 WL |
153 | if (kicks) |
154 | stat = DIV_ROUND_CLOSEST_ULL(stat, kicks); | |
155 | } | |
156 | len = snprintf(buf, sizeof(buf) - 1, "%llu\n", stat); | |
157 | } | |
158 | ||
159 | return simple_read_from_buffer(user_buf, count, ppos, buf, len); | |
160 | } | |
161 | ||
162 | /* | |
163 | * Function to handle write request | |
164 | * | |
165 | * When counter = reset_cnts, reset all the counter values. | |
166 | * Since the counter updates aren't atomic, the resetting is done twice | |
167 | * to make sure that the counters are very likely to be all cleared. | |
168 | */ | |
169 | static ssize_t qstat_write(struct file *file, const char __user *user_buf, | |
170 | size_t count, loff_t *ppos) | |
171 | { | |
172 | int cpu; | |
173 | ||
174 | /* | |
175 | * Get the counter ID stored in file->f_inode->i_private | |
176 | */ | |
45063097 | 177 | if ((long)file_inode(file)->i_private != qstat_reset_cnts) |
45e898b7 WL |
178 | return count; |
179 | ||
180 | for_each_possible_cpu(cpu) { | |
181 | int i; | |
182 | unsigned long *ptr = per_cpu_ptr(qstats, cpu); | |
183 | ||
184 | for (i = 0 ; i < qstat_num; i++) | |
185 | WRITE_ONCE(ptr[i], 0); | |
45e898b7 WL |
186 | } |
187 | return count; | |
188 | } | |
189 | ||
190 | /* | |
191 | * Debugfs data structures | |
192 | */ | |
193 | static const struct file_operations fops_qstat = { | |
194 | .read = qstat_read, | |
195 | .write = qstat_write, | |
196 | .llseek = default_llseek, | |
197 | }; | |
198 | ||
199 | /* | |
200 | * Initialize debugfs for the qspinlock statistical counters | |
201 | */ | |
202 | static int __init init_qspinlock_stat(void) | |
203 | { | |
204 | struct dentry *d_qstat = debugfs_create_dir("qlockstat", NULL); | |
205 | int i; | |
206 | ||
b96bbdde DB |
207 | if (!d_qstat) |
208 | goto out; | |
45e898b7 WL |
209 | |
210 | /* | |
211 | * Create the debugfs files | |
212 | * | |
213 | * As reading from and writing to the stat files can be slow, only | |
214 | * root is allowed to do the read/write to limit impact to system | |
215 | * performance. | |
216 | */ | |
217 | for (i = 0; i < qstat_num; i++) | |
b96bbdde DB |
218 | if (!debugfs_create_file(qstat_names[i], 0400, d_qstat, |
219 | (void *)(long)i, &fops_qstat)) | |
220 | goto fail_undo; | |
221 | ||
222 | if (!debugfs_create_file(qstat_names[qstat_reset_cnts], 0200, d_qstat, | |
223 | (void *)(long)qstat_reset_cnts, &fops_qstat)) | |
224 | goto fail_undo; | |
45e898b7 | 225 | |
45e898b7 | 226 | return 0; |
b96bbdde DB |
227 | fail_undo: |
228 | debugfs_remove_recursive(d_qstat); | |
229 | out: | |
230 | pr_warn("Could not create 'qlockstat' debugfs entries\n"); | |
231 | return -ENOMEM; | |
45e898b7 WL |
232 | } |
233 | fs_initcall(init_qspinlock_stat); | |
234 | ||
235 | /* | |
236 | * Increment the PV qspinlock statistical counters | |
237 | */ | |
238 | static inline void qstat_inc(enum qlock_stats stat, bool cond) | |
239 | { | |
240 | if (cond) | |
241 | this_cpu_inc(qstats[stat]); | |
242 | } | |
243 | ||
244 | /* | |
245 | * PV hash hop count | |
246 | */ | |
247 | static inline void qstat_hop(int hopcnt) | |
248 | { | |
249 | this_cpu_add(qstats[qstat_pv_hash_hops], hopcnt); | |
250 | } | |
251 | ||
252 | /* | |
253 | * Replacement function for pv_kick() | |
254 | */ | |
255 | static inline void __pv_kick(int cpu) | |
256 | { | |
257 | u64 start = sched_clock(); | |
258 | ||
259 | per_cpu(pv_kick_time, cpu) = start; | |
260 | pv_kick(cpu); | |
261 | this_cpu_add(qstats[qstat_pv_latency_kick], sched_clock() - start); | |
262 | } | |
263 | ||
264 | /* | |
265 | * Replacement function for pv_wait() | |
266 | */ | |
267 | static inline void __pv_wait(u8 *ptr, u8 val) | |
268 | { | |
269 | u64 *pkick_time = this_cpu_ptr(&pv_kick_time); | |
270 | ||
271 | *pkick_time = 0; | |
272 | pv_wait(ptr, val); | |
273 | if (*pkick_time) { | |
274 | this_cpu_add(qstats[qstat_pv_latency_wake], | |
275 | sched_clock() - *pkick_time); | |
276 | qstat_inc(qstat_pv_kick_wake, true); | |
277 | } | |
278 | } | |
279 | ||
280 | #define pv_kick(c) __pv_kick(c) | |
281 | #define pv_wait(p, v) __pv_wait(p, v) | |
282 | ||
283 | #else /* CONFIG_QUEUED_LOCK_STAT */ | |
284 | ||
285 | static inline void qstat_inc(enum qlock_stats stat, bool cond) { } | |
286 | static inline void qstat_hop(int hopcnt) { } | |
287 | ||
288 | #endif /* CONFIG_QUEUED_LOCK_STAT */ |