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_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 * 34 * Writing to the "reset_counters" file will reset all the above counter 35 * values. 36 * 37 * These statistical counters are implemented as per-cpu variables which are 38 * summed and computed whenever the corresponding debugfs files are read. This 39 * minimizes added overhead making the counters usable even in a production 40 * environment. 41 * 42 * There may be slight difference between pv_kick_wake and pv_kick_unlock. 43 */ 44 enum qlock_stats { 45 qstat_pv_hash_hops, 46 qstat_pv_kick_unlock, 47 qstat_pv_kick_wake, 48 qstat_pv_latency_kick, 49 qstat_pv_latency_wake, 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_lock_pending, 57 qstat_lock_slowpath, 58 qstat_lock_idx1, 59 qstat_lock_idx2, 60 qstat_lock_idx3, 61 qstat_num, /* Total number of statistical counters */ 62 qstat_reset_cnts = qstat_num, 63 }; 64 65 #ifdef CONFIG_QUEUED_LOCK_STAT 66 /* 67 * Collect pvqspinlock statistics 68 */ 69 #include <linux/debugfs.h> 70 #include <linux/sched.h> 71 #include <linux/sched/clock.h> 72 #include <linux/fs.h> 73 74 static const char * const qstat_names[qstat_num + 1] = { 75 [qstat_pv_hash_hops] = "pv_hash_hops", 76 [qstat_pv_kick_unlock] = "pv_kick_unlock", 77 [qstat_pv_kick_wake] = "pv_kick_wake", 78 [qstat_pv_spurious_wakeup] = "pv_spurious_wakeup", 79 [qstat_pv_latency_kick] = "pv_latency_kick", 80 [qstat_pv_latency_wake] = "pv_latency_wake", 81 [qstat_pv_lock_stealing] = "pv_lock_stealing", 82 [qstat_pv_wait_again] = "pv_wait_again", 83 [qstat_pv_wait_early] = "pv_wait_early", 84 [qstat_pv_wait_head] = "pv_wait_head", 85 [qstat_pv_wait_node] = "pv_wait_node", 86 [qstat_lock_pending] = "lock_pending", 87 [qstat_lock_slowpath] = "lock_slowpath", 88 [qstat_lock_idx1] = "lock_index1", 89 [qstat_lock_idx2] = "lock_index2", 90 [qstat_lock_idx3] = "lock_index3", 91 [qstat_reset_cnts] = "reset_counters", 92 }; 93 94 /* 95 * Per-cpu counters 96 */ 97 static DEFINE_PER_CPU(unsigned long, qstats[qstat_num]); 98 static DEFINE_PER_CPU(u64, pv_kick_time); 99 100 /* 101 * Function to read and return the qlock statistical counter values 102 * 103 * The following counters are handled specially: 104 * 1. qstat_pv_latency_kick 105 * Average kick latency (ns) = pv_latency_kick/pv_kick_unlock 106 * 2. qstat_pv_latency_wake 107 * Average wake latency (ns) = pv_latency_wake/pv_kick_wake 108 * 3. qstat_pv_hash_hops 109 * Average hops/hash = pv_hash_hops/pv_kick_unlock 110 */ 111 static ssize_t qstat_read(struct file *file, char __user *user_buf, 112 size_t count, loff_t *ppos) 113 { 114 char buf[64]; 115 int cpu, counter, len; 116 u64 stat = 0, kicks = 0; 117 118 /* 119 * Get the counter ID stored in file->f_inode->i_private 120 */ 121 counter = (long)file_inode(file)->i_private; 122 123 if (counter >= qstat_num) 124 return -EBADF; 125 126 for_each_possible_cpu(cpu) { 127 stat += per_cpu(qstats[counter], cpu); 128 /* 129 * Need to sum additional counter for some of them 130 */ 131 switch (counter) { 132 133 case qstat_pv_latency_kick: 134 case qstat_pv_hash_hops: 135 kicks += per_cpu(qstats[qstat_pv_kick_unlock], cpu); 136 break; 137 138 case qstat_pv_latency_wake: 139 kicks += per_cpu(qstats[qstat_pv_kick_wake], cpu); 140 break; 141 } 142 } 143 144 if (counter == qstat_pv_hash_hops) { 145 u64 frac = 0; 146 147 if (kicks) { 148 frac = 100ULL * do_div(stat, kicks); 149 frac = DIV_ROUND_CLOSEST_ULL(frac, kicks); 150 } 151 152 /* 153 * Return a X.XX decimal number 154 */ 155 len = snprintf(buf, sizeof(buf) - 1, "%llu.%02llu\n", stat, frac); 156 } else { 157 /* 158 * Round to the nearest ns 159 */ 160 if ((counter == qstat_pv_latency_kick) || 161 (counter == qstat_pv_latency_wake)) { 162 if (kicks) 163 stat = DIV_ROUND_CLOSEST_ULL(stat, kicks); 164 } 165 len = snprintf(buf, sizeof(buf) - 1, "%llu\n", stat); 166 } 167 168 return simple_read_from_buffer(user_buf, count, ppos, buf, len); 169 } 170 171 /* 172 * Function to handle write request 173 * 174 * When counter = reset_cnts, reset all the counter values. 175 * Since the counter updates aren't atomic, the resetting is done twice 176 * to make sure that the counters are very likely to be all cleared. 177 */ 178 static ssize_t qstat_write(struct file *file, const char __user *user_buf, 179 size_t count, loff_t *ppos) 180 { 181 int cpu; 182 183 /* 184 * Get the counter ID stored in file->f_inode->i_private 185 */ 186 if ((long)file_inode(file)->i_private != qstat_reset_cnts) 187 return count; 188 189 for_each_possible_cpu(cpu) { 190 int i; 191 unsigned long *ptr = per_cpu_ptr(qstats, cpu); 192 193 for (i = 0 ; i < qstat_num; i++) 194 WRITE_ONCE(ptr[i], 0); 195 } 196 return count; 197 } 198 199 /* 200 * Debugfs data structures 201 */ 202 static const struct file_operations fops_qstat = { 203 .read = qstat_read, 204 .write = qstat_write, 205 .llseek = default_llseek, 206 }; 207 208 /* 209 * Initialize debugfs for the qspinlock statistical counters 210 */ 211 static int __init init_qspinlock_stat(void) 212 { 213 struct dentry *d_qstat = debugfs_create_dir("qlockstat", NULL); 214 int i; 215 216 if (!d_qstat) 217 goto out; 218 219 /* 220 * Create the debugfs files 221 * 222 * As reading from and writing to the stat files can be slow, only 223 * root is allowed to do the read/write to limit impact to system 224 * performance. 225 */ 226 for (i = 0; i < qstat_num; i++) 227 if (!debugfs_create_file(qstat_names[i], 0400, d_qstat, 228 (void *)(long)i, &fops_qstat)) 229 goto fail_undo; 230 231 if (!debugfs_create_file(qstat_names[qstat_reset_cnts], 0200, d_qstat, 232 (void *)(long)qstat_reset_cnts, &fops_qstat)) 233 goto fail_undo; 234 235 return 0; 236 fail_undo: 237 debugfs_remove_recursive(d_qstat); 238 out: 239 pr_warn("Could not create 'qlockstat' debugfs entries\n"); 240 return -ENOMEM; 241 } 242 fs_initcall(init_qspinlock_stat); 243 244 /* 245 * Increment the PV qspinlock statistical counters 246 */ 247 static inline void qstat_inc(enum qlock_stats stat, bool cond) 248 { 249 if (cond) 250 this_cpu_inc(qstats[stat]); 251 } 252 253 /* 254 * PV hash hop count 255 */ 256 static inline void qstat_hop(int hopcnt) 257 { 258 this_cpu_add(qstats[qstat_pv_hash_hops], hopcnt); 259 } 260 261 /* 262 * Replacement function for pv_kick() 263 */ 264 static inline void __pv_kick(int cpu) 265 { 266 u64 start = sched_clock(); 267 268 per_cpu(pv_kick_time, cpu) = start; 269 pv_kick(cpu); 270 this_cpu_add(qstats[qstat_pv_latency_kick], sched_clock() - start); 271 } 272 273 /* 274 * Replacement function for pv_wait() 275 */ 276 static inline void __pv_wait(u8 *ptr, u8 val) 277 { 278 u64 *pkick_time = this_cpu_ptr(&pv_kick_time); 279 280 *pkick_time = 0; 281 pv_wait(ptr, val); 282 if (*pkick_time) { 283 this_cpu_add(qstats[qstat_pv_latency_wake], 284 sched_clock() - *pkick_time); 285 qstat_inc(qstat_pv_kick_wake, true); 286 } 287 } 288 289 #define pv_kick(c) __pv_kick(c) 290 #define pv_wait(p, v) __pv_wait(p, v) 291 292 #else /* CONFIG_QUEUED_LOCK_STAT */ 293 294 static inline void qstat_inc(enum qlock_stats stat, bool cond) { } 295 static inline void qstat_hop(int hopcnt) { } 296 297 #endif /* CONFIG_QUEUED_LOCK_STAT */ 298