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
25  *   pv_lock_slowpath	- # of locking operations via the slowpath
26  *   pv_lock_stealing	- # of lock stealing operations
27  *   pv_spurious_wakeup	- # of spurious wakeups in non-head vCPUs
28  *   pv_wait_again	- # of wait's after a queue head vCPU kick
29  *   pv_wait_early	- # of early vCPU wait's
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,
49 	qstat_pv_lock_slowpath,
50 	qstat_pv_lock_stealing,
51 	qstat_pv_spurious_wakeup,
52 	qstat_pv_wait_again,
53 	qstat_pv_wait_early,
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",
75 	[qstat_pv_lock_slowpath]   = "pv_lock_slowpath",
76 	[qstat_pv_lock_stealing]   = "pv_lock_stealing",
77 	[qstat_pv_wait_again]      = "pv_wait_again",
78 	[qstat_pv_wait_early]      = "pv_wait_early",
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 	counter = (long)file_inode(file)->i_private;
112 
113 	if (counter >= qstat_num)
114 		return -EBADF;
115 
116 	for_each_possible_cpu(cpu) {
117 		stat += per_cpu(qstats[counter], cpu);
118 		/*
119 		 * Need to sum additional counter for some of them
120 		 */
121 		switch (counter) {
122 
123 		case qstat_pv_latency_kick:
124 		case qstat_pv_hash_hops:
125 			kicks += per_cpu(qstats[qstat_pv_kick_unlock], cpu);
126 			break;
127 
128 		case qstat_pv_latency_wake:
129 			kicks += per_cpu(qstats[qstat_pv_kick_wake], cpu);
130 			break;
131 		}
132 	}
133 
134 	if (counter == qstat_pv_hash_hops) {
135 		u64 frac = 0;
136 
137 		if (kicks) {
138 			frac = 100ULL * do_div(stat, kicks);
139 			frac = DIV_ROUND_CLOSEST_ULL(frac, kicks);
140 		}
141 
142 		/*
143 		 * Return a X.XX decimal number
144 		 */
145 		len = snprintf(buf, sizeof(buf) - 1, "%llu.%02llu\n", stat, frac);
146 	} else {
147 		/*
148 		 * Round to the nearest ns
149 		 */
150 		if ((counter == qstat_pv_latency_kick) ||
151 		    (counter == qstat_pv_latency_wake)) {
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 ((long)file_inode(file)->i_private != qstat_reset_cnts)
177 		return count;
178 
179 	for_each_possible_cpu(cpu) {
180 		int i;
181 		unsigned long *ptr = per_cpu_ptr(qstats, cpu);
182 
183 		for (i = 0 ; i < qstat_num; i++)
184 			WRITE_ONCE(ptr[i], 0);
185 	}
186 	return count;
187 }
188 
189 /*
190  * Debugfs data structures
191  */
192 static const struct file_operations fops_qstat = {
193 	.read = qstat_read,
194 	.write = qstat_write,
195 	.llseek = default_llseek,
196 };
197 
198 /*
199  * Initialize debugfs for the qspinlock statistical counters
200  */
201 static int __init init_qspinlock_stat(void)
202 {
203 	struct dentry *d_qstat = debugfs_create_dir("qlockstat", NULL);
204 	int i;
205 
206 	if (!d_qstat)
207 		goto out;
208 
209 	/*
210 	 * Create the debugfs files
211 	 *
212 	 * As reading from and writing to the stat files can be slow, only
213 	 * root is allowed to do the read/write to limit impact to system
214 	 * performance.
215 	 */
216 	for (i = 0; i < qstat_num; i++)
217 		if (!debugfs_create_file(qstat_names[i], 0400, d_qstat,
218 					 (void *)(long)i, &fops_qstat))
219 			goto fail_undo;
220 
221 	if (!debugfs_create_file(qstat_names[qstat_reset_cnts], 0200, d_qstat,
222 				 (void *)(long)qstat_reset_cnts, &fops_qstat))
223 		goto fail_undo;
224 
225 	return 0;
226 fail_undo:
227 	debugfs_remove_recursive(d_qstat);
228 out:
229 	pr_warn("Could not create 'qlockstat' debugfs entries\n");
230 	return -ENOMEM;
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 */
288