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.
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.
12 * Authors: Waiman Long <waiman.long@hpe.com>
16 * When queued spinlock statistical counters are enabled, the following
17 * debugfs files will be created for reporting the counter values:
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_stealing - # of lock stealing operations
26 * pv_spurious_wakeup - # of spurious wakeups in non-head vCPUs
27 * pv_wait_again - # of wait's after a queue head vCPU kick
28 * pv_wait_early - # of early vCPU wait's
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 * lock_pending - # of locking operations via pending code
32 * lock_slowpath - # of locking operations via MCS lock queue
33 * lock_use_node2 - # of locking operations that use 2nd per-CPU node
34 * lock_use_node3 - # of locking operations that use 3rd per-CPU node
35 * lock_use_node4 - # of locking operations that use 4th per-CPU node
36 * lock_no_node - # of locking operations without using per-CPU node
38 * Subtracting lock_use_node[234] from lock_slowpath will give you
41 * Writing to the "reset_counters" file will reset all the above counter
44 * These statistical counters are implemented as per-cpu variables which are
45 * summed and computed whenever the corresponding debugfs files are read. This
46 * minimizes added overhead making the counters usable even in a production
49 * There may be slight difference between pv_kick_wake and pv_kick_unlock.
55 qstat_pv_latency_kick
,
56 qstat_pv_latency_wake
,
57 qstat_pv_lock_stealing
,
58 qstat_pv_spurious_wakeup
,
69 qstat_num
, /* Total number of statistical counters */
70 qstat_reset_cnts
= qstat_num
,
73 #ifdef CONFIG_QUEUED_LOCK_STAT
75 * Collect pvqspinlock statistics
77 #include <linux/debugfs.h>
78 #include <linux/sched.h>
79 #include <linux/sched/clock.h>
82 static const char * const qstat_names
[qstat_num
+ 1] = {
83 [qstat_pv_hash_hops
] = "pv_hash_hops",
84 [qstat_pv_kick_unlock
] = "pv_kick_unlock",
85 [qstat_pv_kick_wake
] = "pv_kick_wake",
86 [qstat_pv_spurious_wakeup
] = "pv_spurious_wakeup",
87 [qstat_pv_latency_kick
] = "pv_latency_kick",
88 [qstat_pv_latency_wake
] = "pv_latency_wake",
89 [qstat_pv_lock_stealing
] = "pv_lock_stealing",
90 [qstat_pv_wait_again
] = "pv_wait_again",
91 [qstat_pv_wait_early
] = "pv_wait_early",
92 [qstat_pv_wait_head
] = "pv_wait_head",
93 [qstat_pv_wait_node
] = "pv_wait_node",
94 [qstat_lock_pending
] = "lock_pending",
95 [qstat_lock_slowpath
] = "lock_slowpath",
96 [qstat_lock_use_node2
] = "lock_use_node2",
97 [qstat_lock_use_node3
] = "lock_use_node3",
98 [qstat_lock_use_node4
] = "lock_use_node4",
99 [qstat_lock_no_node
] = "lock_no_node",
100 [qstat_reset_cnts
] = "reset_counters",
106 static DEFINE_PER_CPU(unsigned long, qstats
[qstat_num
]);
107 static DEFINE_PER_CPU(u64
, pv_kick_time
);
110 * Function to read and return the qlock statistical counter values
112 * The following counters are handled specially:
113 * 1. qstat_pv_latency_kick
114 * Average kick latency (ns) = pv_latency_kick/pv_kick_unlock
115 * 2. qstat_pv_latency_wake
116 * Average wake latency (ns) = pv_latency_wake/pv_kick_wake
117 * 3. qstat_pv_hash_hops
118 * Average hops/hash = pv_hash_hops/pv_kick_unlock
120 static ssize_t
qstat_read(struct file
*file
, char __user
*user_buf
,
121 size_t count
, loff_t
*ppos
)
124 int cpu
, counter
, len
;
125 u64 stat
= 0, kicks
= 0;
128 * Get the counter ID stored in file->f_inode->i_private
130 counter
= (long)file_inode(file
)->i_private
;
132 if (counter
>= qstat_num
)
135 for_each_possible_cpu(cpu
) {
136 stat
+= per_cpu(qstats
[counter
], cpu
);
138 * Need to sum additional counter for some of them
142 case qstat_pv_latency_kick
:
143 case qstat_pv_hash_hops
:
144 kicks
+= per_cpu(qstats
[qstat_pv_kick_unlock
], cpu
);
147 case qstat_pv_latency_wake
:
148 kicks
+= per_cpu(qstats
[qstat_pv_kick_wake
], cpu
);
153 if (counter
== qstat_pv_hash_hops
) {
157 frac
= 100ULL * do_div(stat
, kicks
);
158 frac
= DIV_ROUND_CLOSEST_ULL(frac
, kicks
);
162 * Return a X.XX decimal number
164 len
= snprintf(buf
, sizeof(buf
) - 1, "%llu.%02llu\n", stat
, frac
);
167 * Round to the nearest ns
169 if ((counter
== qstat_pv_latency_kick
) ||
170 (counter
== qstat_pv_latency_wake
)) {
172 stat
= DIV_ROUND_CLOSEST_ULL(stat
, kicks
);
174 len
= snprintf(buf
, sizeof(buf
) - 1, "%llu\n", stat
);
177 return simple_read_from_buffer(user_buf
, count
, ppos
, buf
, len
);
181 * Function to handle write request
183 * When counter = reset_cnts, reset all the counter values.
184 * Since the counter updates aren't atomic, the resetting is done twice
185 * to make sure that the counters are very likely to be all cleared.
187 static ssize_t
qstat_write(struct file
*file
, const char __user
*user_buf
,
188 size_t count
, loff_t
*ppos
)
193 * Get the counter ID stored in file->f_inode->i_private
195 if ((long)file_inode(file
)->i_private
!= qstat_reset_cnts
)
198 for_each_possible_cpu(cpu
) {
200 unsigned long *ptr
= per_cpu_ptr(qstats
, cpu
);
202 for (i
= 0 ; i
< qstat_num
; i
++)
203 WRITE_ONCE(ptr
[i
], 0);
209 * Debugfs data structures
211 static const struct file_operations fops_qstat
= {
213 .write
= qstat_write
,
214 .llseek
= default_llseek
,
218 * Initialize debugfs for the qspinlock statistical counters
220 static int __init
init_qspinlock_stat(void)
222 struct dentry
*d_qstat
= debugfs_create_dir("qlockstat", NULL
);
229 * Create the debugfs files
231 * As reading from and writing to the stat files can be slow, only
232 * root is allowed to do the read/write to limit impact to system
235 for (i
= 0; i
< qstat_num
; i
++)
236 if (!debugfs_create_file(qstat_names
[i
], 0400, d_qstat
,
237 (void *)(long)i
, &fops_qstat
))
240 if (!debugfs_create_file(qstat_names
[qstat_reset_cnts
], 0200, d_qstat
,
241 (void *)(long)qstat_reset_cnts
, &fops_qstat
))
246 debugfs_remove_recursive(d_qstat
);
248 pr_warn("Could not create 'qlockstat' debugfs entries\n");
251 fs_initcall(init_qspinlock_stat
);
254 * Increment the PV qspinlock statistical counters
256 static inline void qstat_inc(enum qlock_stats stat
, bool cond
)
259 this_cpu_inc(qstats
[stat
]);
265 static inline void qstat_hop(int hopcnt
)
267 this_cpu_add(qstats
[qstat_pv_hash_hops
], hopcnt
);
271 * Replacement function for pv_kick()
273 static inline void __pv_kick(int cpu
)
275 u64 start
= sched_clock();
277 per_cpu(pv_kick_time
, cpu
) = start
;
279 this_cpu_add(qstats
[qstat_pv_latency_kick
], sched_clock() - start
);
283 * Replacement function for pv_wait()
285 static inline void __pv_wait(u8
*ptr
, u8 val
)
287 u64
*pkick_time
= this_cpu_ptr(&pv_kick_time
);
292 this_cpu_add(qstats
[qstat_pv_latency_wake
],
293 sched_clock() - *pkick_time
);
294 qstat_inc(qstat_pv_kick_wake
, true);
298 #define pv_kick(c) __pv_kick(c)
299 #define pv_wait(p, v) __pv_wait(p, v)
301 #else /* CONFIG_QUEUED_LOCK_STAT */
303 static inline void qstat_inc(enum qlock_stats stat
, bool cond
) { }
304 static inline void qstat_hop(int hopcnt
) { }
306 #endif /* CONFIG_QUEUED_LOCK_STAT */