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 | |
1c4941fd | 25 | * pv_lock_stealing - # of lock stealing operations |
45e898b7 WL |
26 | * pv_spurious_wakeup - # of spurious wakeups |
27 | * pv_wait_again - # of vCPU wait's that happened after a vCPU kick | |
cd0272fa | 28 | * pv_wait_early - # of early vCPU wait's |
45e898b7 WL |
29 | * pv_wait_head - # of vCPU wait's at the queue head |
30 | * pv_wait_node - # of vCPU wait's at a non-head queue node | |
31 | * | |
32 | * Writing to the "reset_counters" file will reset all the above counter | |
33 | * values. | |
34 | * | |
35 | * These statistical counters are implemented as per-cpu variables which are | |
36 | * summed and computed whenever the corresponding debugfs files are read. This | |
37 | * minimizes added overhead making the counters usable even in a production | |
38 | * environment. | |
39 | * | |
40 | * There may be slight difference between pv_kick_wake and pv_kick_unlock. | |
41 | */ | |
42 | enum qlock_stats { | |
43 | qstat_pv_hash_hops, | |
44 | qstat_pv_kick_unlock, | |
45 | qstat_pv_kick_wake, | |
46 | qstat_pv_latency_kick, | |
47 | qstat_pv_latency_wake, | |
1c4941fd | 48 | qstat_pv_lock_stealing, |
45e898b7 WL |
49 | qstat_pv_spurious_wakeup, |
50 | qstat_pv_wait_again, | |
cd0272fa | 51 | qstat_pv_wait_early, |
45e898b7 WL |
52 | qstat_pv_wait_head, |
53 | qstat_pv_wait_node, | |
54 | qstat_num, /* Total number of statistical counters */ | |
55 | qstat_reset_cnts = qstat_num, | |
56 | }; | |
57 | ||
58 | #ifdef CONFIG_QUEUED_LOCK_STAT | |
59 | /* | |
60 | * Collect pvqspinlock statistics | |
61 | */ | |
62 | #include <linux/debugfs.h> | |
63 | #include <linux/sched.h> | |
64 | #include <linux/fs.h> | |
65 | ||
66 | static const char * const qstat_names[qstat_num + 1] = { | |
67 | [qstat_pv_hash_hops] = "pv_hash_hops", | |
68 | [qstat_pv_kick_unlock] = "pv_kick_unlock", | |
69 | [qstat_pv_kick_wake] = "pv_kick_wake", | |
70 | [qstat_pv_spurious_wakeup] = "pv_spurious_wakeup", | |
71 | [qstat_pv_latency_kick] = "pv_latency_kick", | |
72 | [qstat_pv_latency_wake] = "pv_latency_wake", | |
1c4941fd | 73 | [qstat_pv_lock_stealing] = "pv_lock_stealing", |
45e898b7 | 74 | [qstat_pv_wait_again] = "pv_wait_again", |
cd0272fa | 75 | [qstat_pv_wait_early] = "pv_wait_early", |
45e898b7 WL |
76 | [qstat_pv_wait_head] = "pv_wait_head", |
77 | [qstat_pv_wait_node] = "pv_wait_node", | |
78 | [qstat_reset_cnts] = "reset_counters", | |
79 | }; | |
80 | ||
81 | /* | |
82 | * Per-cpu counters | |
83 | */ | |
84 | static DEFINE_PER_CPU(unsigned long, qstats[qstat_num]); | |
85 | static DEFINE_PER_CPU(u64, pv_kick_time); | |
86 | ||
87 | /* | |
88 | * Function to read and return the qlock statistical counter values | |
89 | * | |
90 | * The following counters are handled specially: | |
91 | * 1. qstat_pv_latency_kick | |
92 | * Average kick latency (ns) = pv_latency_kick/pv_kick_unlock | |
93 | * 2. qstat_pv_latency_wake | |
94 | * Average wake latency (ns) = pv_latency_wake/pv_kick_wake | |
95 | * 3. qstat_pv_hash_hops | |
96 | * Average hops/hash = pv_hash_hops/pv_kick_unlock | |
97 | */ | |
98 | static ssize_t qstat_read(struct file *file, char __user *user_buf, | |
99 | size_t count, loff_t *ppos) | |
100 | { | |
101 | char buf[64]; | |
102 | int cpu, counter, len; | |
103 | u64 stat = 0, kicks = 0; | |
104 | ||
105 | /* | |
106 | * Get the counter ID stored in file->f_inode->i_private | |
107 | */ | |
108 | if (!file->f_inode) { | |
109 | WARN_ON_ONCE(1); | |
110 | return -EBADF; | |
111 | } | |
112 | counter = (long)(file->f_inode->i_private); | |
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) { | |
136 | u64 frac; | |
137 | ||
138 | frac = 100ULL * do_div(stat, kicks); | |
139 | frac = DIV_ROUND_CLOSEST_ULL(frac, kicks); | |
140 | ||
141 | /* | |
142 | * Return a X.XX decimal number | |
143 | */ | |
144 | len = snprintf(buf, sizeof(buf) - 1, "%llu.%02llu\n", stat, frac); | |
145 | } else { | |
146 | /* | |
147 | * Round to the nearest ns | |
148 | */ | |
149 | if ((counter == qstat_pv_latency_kick) || | |
150 | (counter == qstat_pv_latency_wake)) { | |
151 | stat = 0; | |
152 | if (kicks) | |
153 | stat = DIV_ROUND_CLOSEST_ULL(stat, kicks); | |
154 | } | |
155 | len = snprintf(buf, sizeof(buf) - 1, "%llu\n", stat); | |
156 | } | |
157 | ||
158 | return simple_read_from_buffer(user_buf, count, ppos, buf, len); | |
159 | } | |
160 | ||
161 | /* | |
162 | * Function to handle write request | |
163 | * | |
164 | * When counter = reset_cnts, reset all the counter values. | |
165 | * Since the counter updates aren't atomic, the resetting is done twice | |
166 | * to make sure that the counters are very likely to be all cleared. | |
167 | */ | |
168 | static ssize_t qstat_write(struct file *file, const char __user *user_buf, | |
169 | size_t count, loff_t *ppos) | |
170 | { | |
171 | int cpu; | |
172 | ||
173 | /* | |
174 | * Get the counter ID stored in file->f_inode->i_private | |
175 | */ | |
176 | if (!file->f_inode) { | |
177 | WARN_ON_ONCE(1); | |
178 | return -EBADF; | |
179 | } | |
180 | if ((long)(file->f_inode->i_private) != qstat_reset_cnts) | |
181 | return count; | |
182 | ||
183 | for_each_possible_cpu(cpu) { | |
184 | int i; | |
185 | unsigned long *ptr = per_cpu_ptr(qstats, cpu); | |
186 | ||
187 | for (i = 0 ; i < qstat_num; i++) | |
188 | WRITE_ONCE(ptr[i], 0); | |
189 | for (i = 0 ; i < qstat_num; i++) | |
190 | WRITE_ONCE(ptr[i], 0); | |
191 | } | |
192 | return count; | |
193 | } | |
194 | ||
195 | /* | |
196 | * Debugfs data structures | |
197 | */ | |
198 | static const struct file_operations fops_qstat = { | |
199 | .read = qstat_read, | |
200 | .write = qstat_write, | |
201 | .llseek = default_llseek, | |
202 | }; | |
203 | ||
204 | /* | |
205 | * Initialize debugfs for the qspinlock statistical counters | |
206 | */ | |
207 | static int __init init_qspinlock_stat(void) | |
208 | { | |
209 | struct dentry *d_qstat = debugfs_create_dir("qlockstat", NULL); | |
210 | int i; | |
211 | ||
212 | if (!d_qstat) { | |
213 | pr_warn("Could not create 'qlockstat' debugfs directory\n"); | |
214 | return 0; | |
215 | } | |
216 | ||
217 | /* | |
218 | * Create the debugfs files | |
219 | * | |
220 | * As reading from and writing to the stat files can be slow, only | |
221 | * root is allowed to do the read/write to limit impact to system | |
222 | * performance. | |
223 | */ | |
224 | for (i = 0; i < qstat_num; i++) | |
225 | debugfs_create_file(qstat_names[i], 0400, d_qstat, | |
226 | (void *)(long)i, &fops_qstat); | |
227 | ||
228 | debugfs_create_file(qstat_names[qstat_reset_cnts], 0200, d_qstat, | |
229 | (void *)(long)qstat_reset_cnts, &fops_qstat); | |
230 | return 0; | |
231 | } | |
232 | fs_initcall(init_qspinlock_stat); | |
233 | ||
234 | /* | |
235 | * Increment the PV qspinlock statistical counters | |
236 | */ | |
237 | static inline void qstat_inc(enum qlock_stats stat, bool cond) | |
238 | { | |
239 | if (cond) | |
240 | this_cpu_inc(qstats[stat]); | |
241 | } | |
242 | ||
243 | /* | |
244 | * PV hash hop count | |
245 | */ | |
246 | static inline void qstat_hop(int hopcnt) | |
247 | { | |
248 | this_cpu_add(qstats[qstat_pv_hash_hops], hopcnt); | |
249 | } | |
250 | ||
251 | /* | |
252 | * Replacement function for pv_kick() | |
253 | */ | |
254 | static inline void __pv_kick(int cpu) | |
255 | { | |
256 | u64 start = sched_clock(); | |
257 | ||
258 | per_cpu(pv_kick_time, cpu) = start; | |
259 | pv_kick(cpu); | |
260 | this_cpu_add(qstats[qstat_pv_latency_kick], sched_clock() - start); | |
261 | } | |
262 | ||
263 | /* | |
264 | * Replacement function for pv_wait() | |
265 | */ | |
266 | static inline void __pv_wait(u8 *ptr, u8 val) | |
267 | { | |
268 | u64 *pkick_time = this_cpu_ptr(&pv_kick_time); | |
269 | ||
270 | *pkick_time = 0; | |
271 | pv_wait(ptr, val); | |
272 | if (*pkick_time) { | |
273 | this_cpu_add(qstats[qstat_pv_latency_wake], | |
274 | sched_clock() - *pkick_time); | |
275 | qstat_inc(qstat_pv_kick_wake, true); | |
276 | } | |
277 | } | |
278 | ||
279 | #define pv_kick(c) __pv_kick(c) | |
280 | #define pv_wait(p, v) __pv_wait(p, v) | |
281 | ||
282 | #else /* CONFIG_QUEUED_LOCK_STAT */ | |
283 | ||
284 | static inline void qstat_inc(enum qlock_stats stat, bool cond) { } | |
285 | static inline void qstat_hop(int hopcnt) { } | |
286 | ||
287 | #endif /* CONFIG_QUEUED_LOCK_STAT */ |