]>
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 |
45e898b7 WL |
27 | * pv_spurious_wakeup - # of spurious wakeups |
28 | * pv_wait_again - # of vCPU wait's that happened after a 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> | |
66 | #include <linux/fs.h> | |
67 | ||
68 | static const char * const qstat_names[qstat_num + 1] = { | |
69 | [qstat_pv_hash_hops] = "pv_hash_hops", | |
70 | [qstat_pv_kick_unlock] = "pv_kick_unlock", | |
71 | [qstat_pv_kick_wake] = "pv_kick_wake", | |
72 | [qstat_pv_spurious_wakeup] = "pv_spurious_wakeup", | |
73 | [qstat_pv_latency_kick] = "pv_latency_kick", | |
74 | [qstat_pv_latency_wake] = "pv_latency_wake", | |
32d62510 | 75 | [qstat_pv_lock_slowpath] = "pv_lock_slowpath", |
1c4941fd | 76 | [qstat_pv_lock_stealing] = "pv_lock_stealing", |
45e898b7 | 77 | [qstat_pv_wait_again] = "pv_wait_again", |
cd0272fa | 78 | [qstat_pv_wait_early] = "pv_wait_early", |
45e898b7 WL |
79 | [qstat_pv_wait_head] = "pv_wait_head", |
80 | [qstat_pv_wait_node] = "pv_wait_node", | |
81 | [qstat_reset_cnts] = "reset_counters", | |
82 | }; | |
83 | ||
84 | /* | |
85 | * Per-cpu counters | |
86 | */ | |
87 | static DEFINE_PER_CPU(unsigned long, qstats[qstat_num]); | |
88 | static DEFINE_PER_CPU(u64, pv_kick_time); | |
89 | ||
90 | /* | |
91 | * Function to read and return the qlock statistical counter values | |
92 | * | |
93 | * The following counters are handled specially: | |
94 | * 1. qstat_pv_latency_kick | |
95 | * Average kick latency (ns) = pv_latency_kick/pv_kick_unlock | |
96 | * 2. qstat_pv_latency_wake | |
97 | * Average wake latency (ns) = pv_latency_wake/pv_kick_wake | |
98 | * 3. qstat_pv_hash_hops | |
99 | * Average hops/hash = pv_hash_hops/pv_kick_unlock | |
100 | */ | |
101 | static ssize_t qstat_read(struct file *file, char __user *user_buf, | |
102 | size_t count, loff_t *ppos) | |
103 | { | |
104 | char buf[64]; | |
105 | int cpu, counter, len; | |
106 | u64 stat = 0, kicks = 0; | |
107 | ||
108 | /* | |
109 | * Get the counter ID stored in file->f_inode->i_private | |
110 | */ | |
111 | if (!file->f_inode) { | |
112 | WARN_ON_ONCE(1); | |
113 | return -EBADF; | |
114 | } | |
115 | counter = (long)(file->f_inode->i_private); | |
116 | ||
117 | if (counter >= qstat_num) | |
118 | return -EBADF; | |
119 | ||
120 | for_each_possible_cpu(cpu) { | |
121 | stat += per_cpu(qstats[counter], cpu); | |
122 | /* | |
123 | * Need to sum additional counter for some of them | |
124 | */ | |
125 | switch (counter) { | |
126 | ||
127 | case qstat_pv_latency_kick: | |
128 | case qstat_pv_hash_hops: | |
129 | kicks += per_cpu(qstats[qstat_pv_kick_unlock], cpu); | |
130 | break; | |
131 | ||
132 | case qstat_pv_latency_wake: | |
133 | kicks += per_cpu(qstats[qstat_pv_kick_wake], cpu); | |
134 | break; | |
135 | } | |
136 | } | |
137 | ||
138 | if (counter == qstat_pv_hash_hops) { | |
66876595 | 139 | u64 frac = 0; |
45e898b7 | 140 | |
66876595 DB |
141 | if (kicks) { |
142 | frac = 100ULL * do_div(stat, kicks); | |
143 | frac = DIV_ROUND_CLOSEST_ULL(frac, kicks); | |
144 | } | |
45e898b7 WL |
145 | |
146 | /* | |
147 | * Return a X.XX decimal number | |
148 | */ | |
149 | len = snprintf(buf, sizeof(buf) - 1, "%llu.%02llu\n", stat, frac); | |
150 | } else { | |
151 | /* | |
152 | * Round to the nearest ns | |
153 | */ | |
154 | if ((counter == qstat_pv_latency_kick) || | |
155 | (counter == qstat_pv_latency_wake)) { | |
45e898b7 WL |
156 | if (kicks) |
157 | stat = DIV_ROUND_CLOSEST_ULL(stat, kicks); | |
158 | } | |
159 | len = snprintf(buf, sizeof(buf) - 1, "%llu\n", stat); | |
160 | } | |
161 | ||
162 | return simple_read_from_buffer(user_buf, count, ppos, buf, len); | |
163 | } | |
164 | ||
165 | /* | |
166 | * Function to handle write request | |
167 | * | |
168 | * When counter = reset_cnts, reset all the counter values. | |
169 | * Since the counter updates aren't atomic, the resetting is done twice | |
170 | * to make sure that the counters are very likely to be all cleared. | |
171 | */ | |
172 | static ssize_t qstat_write(struct file *file, const char __user *user_buf, | |
173 | size_t count, loff_t *ppos) | |
174 | { | |
175 | int cpu; | |
176 | ||
177 | /* | |
178 | * Get the counter ID stored in file->f_inode->i_private | |
179 | */ | |
180 | if (!file->f_inode) { | |
181 | WARN_ON_ONCE(1); | |
182 | return -EBADF; | |
183 | } | |
184 | if ((long)(file->f_inode->i_private) != qstat_reset_cnts) | |
185 | return count; | |
186 | ||
187 | for_each_possible_cpu(cpu) { | |
188 | int i; | |
189 | unsigned long *ptr = per_cpu_ptr(qstats, cpu); | |
190 | ||
191 | for (i = 0 ; i < qstat_num; i++) | |
192 | WRITE_ONCE(ptr[i], 0); | |
45e898b7 WL |
193 | } |
194 | return count; | |
195 | } | |
196 | ||
197 | /* | |
198 | * Debugfs data structures | |
199 | */ | |
200 | static const struct file_operations fops_qstat = { | |
201 | .read = qstat_read, | |
202 | .write = qstat_write, | |
203 | .llseek = default_llseek, | |
204 | }; | |
205 | ||
206 | /* | |
207 | * Initialize debugfs for the qspinlock statistical counters | |
208 | */ | |
209 | static int __init init_qspinlock_stat(void) | |
210 | { | |
211 | struct dentry *d_qstat = debugfs_create_dir("qlockstat", NULL); | |
212 | int i; | |
213 | ||
b96bbdde DB |
214 | if (!d_qstat) |
215 | goto out; | |
45e898b7 WL |
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++) | |
b96bbdde DB |
225 | if (!debugfs_create_file(qstat_names[i], 0400, d_qstat, |
226 | (void *)(long)i, &fops_qstat)) | |
227 | goto fail_undo; | |
228 | ||
229 | if (!debugfs_create_file(qstat_names[qstat_reset_cnts], 0200, d_qstat, | |
230 | (void *)(long)qstat_reset_cnts, &fops_qstat)) | |
231 | goto fail_undo; | |
45e898b7 | 232 | |
45e898b7 | 233 | return 0; |
b96bbdde DB |
234 | fail_undo: |
235 | debugfs_remove_recursive(d_qstat); | |
236 | out: | |
237 | pr_warn("Could not create 'qlockstat' debugfs entries\n"); | |
238 | return -ENOMEM; | |
45e898b7 WL |
239 | } |
240 | fs_initcall(init_qspinlock_stat); | |
241 | ||
242 | /* | |
243 | * Increment the PV qspinlock statistical counters | |
244 | */ | |
245 | static inline void qstat_inc(enum qlock_stats stat, bool cond) | |
246 | { | |
247 | if (cond) | |
248 | this_cpu_inc(qstats[stat]); | |
249 | } | |
250 | ||
251 | /* | |
252 | * PV hash hop count | |
253 | */ | |
254 | static inline void qstat_hop(int hopcnt) | |
255 | { | |
256 | this_cpu_add(qstats[qstat_pv_hash_hops], hopcnt); | |
257 | } | |
258 | ||
259 | /* | |
260 | * Replacement function for pv_kick() | |
261 | */ | |
262 | static inline void __pv_kick(int cpu) | |
263 | { | |
264 | u64 start = sched_clock(); | |
265 | ||
266 | per_cpu(pv_kick_time, cpu) = start; | |
267 | pv_kick(cpu); | |
268 | this_cpu_add(qstats[qstat_pv_latency_kick], sched_clock() - start); | |
269 | } | |
270 | ||
271 | /* | |
272 | * Replacement function for pv_wait() | |
273 | */ | |
274 | static inline void __pv_wait(u8 *ptr, u8 val) | |
275 | { | |
276 | u64 *pkick_time = this_cpu_ptr(&pv_kick_time); | |
277 | ||
278 | *pkick_time = 0; | |
279 | pv_wait(ptr, val); | |
280 | if (*pkick_time) { | |
281 | this_cpu_add(qstats[qstat_pv_latency_wake], | |
282 | sched_clock() - *pkick_time); | |
283 | qstat_inc(qstat_pv_kick_wake, true); | |
284 | } | |
285 | } | |
286 | ||
287 | #define pv_kick(c) __pv_kick(c) | |
288 | #define pv_wait(p, v) __pv_wait(p, v) | |
289 | ||
290 | #else /* CONFIG_QUEUED_LOCK_STAT */ | |
291 | ||
292 | static inline void qstat_inc(enum qlock_stats stat, bool cond) { } | |
293 | static inline void qstat_hop(int hopcnt) { } | |
294 | ||
295 | #endif /* CONFIG_QUEUED_LOCK_STAT */ |