Commit | Line | Data |
---|---|---|
e1d8aa9f FW |
1 | /* |
2 | * Workqueue statistical tracer. | |
3 | * | |
4 | * Copyright (C) 2008 Frederic Weisbecker <fweisbec@gmail.com> | |
5 | * | |
6 | */ | |
7 | ||
8 | ||
9 | #include <trace/workqueue.h> | |
10 | #include <linux/list.h> | |
11 | #include "trace_stat.h" | |
12 | #include "trace.h" | |
13 | ||
14 | ||
15 | /* A cpu workqueue thread */ | |
16 | struct cpu_workqueue_stats { | |
17 | struct list_head list; | |
18 | /* Useful to know if we print the cpu headers */ | |
19 | bool first_entry; | |
20 | int cpu; | |
21 | pid_t pid; | |
22 | /* Can be inserted from interrupt or user context, need to be atomic */ | |
23 | atomic_t inserted; | |
24 | /* | |
25 | * Don't need to be atomic, works are serialized in a single workqueue thread | |
26 | * on a single CPU. | |
27 | */ | |
28 | unsigned int executed; | |
29 | }; | |
30 | ||
31 | /* List of workqueue threads on one cpu */ | |
32 | struct workqueue_global_stats { | |
33 | struct list_head list; | |
34 | spinlock_t lock; | |
35 | }; | |
36 | ||
37 | /* Don't need a global lock because allocated before the workqueues, and | |
38 | * never freed. | |
39 | */ | |
40 | static struct workqueue_global_stats *all_workqueue_stat; | |
41 | ||
42 | /* Insertion of a work */ | |
43 | static void | |
44 | probe_workqueue_insertion(struct task_struct *wq_thread, | |
45 | struct work_struct *work) | |
46 | { | |
47 | int cpu = cpumask_first(&wq_thread->cpus_allowed); | |
48 | struct cpu_workqueue_stats *node, *next; | |
49 | unsigned long flags; | |
50 | ||
51 | spin_lock_irqsave(&all_workqueue_stat[cpu].lock, flags); | |
52 | list_for_each_entry_safe(node, next, &all_workqueue_stat[cpu].list, | |
53 | list) { | |
54 | if (node->pid == wq_thread->pid) { | |
55 | atomic_inc(&node->inserted); | |
56 | goto found; | |
57 | } | |
58 | } | |
59 | pr_debug("trace_workqueue: entry not found\n"); | |
60 | found: | |
61 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
62 | } | |
63 | ||
64 | /* Execution of a work */ | |
65 | static void | |
66 | probe_workqueue_execution(struct task_struct *wq_thread, | |
67 | struct work_struct *work) | |
68 | { | |
69 | int cpu = cpumask_first(&wq_thread->cpus_allowed); | |
70 | struct cpu_workqueue_stats *node, *next; | |
71 | unsigned long flags; | |
72 | ||
73 | spin_lock_irqsave(&all_workqueue_stat[cpu].lock, flags); | |
74 | list_for_each_entry_safe(node, next, &all_workqueue_stat[cpu].list, | |
75 | list) { | |
76 | if (node->pid == wq_thread->pid) { | |
77 | node->executed++; | |
78 | goto found; | |
79 | } | |
80 | } | |
81 | pr_debug("trace_workqueue: entry not found\n"); | |
82 | found: | |
83 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
84 | } | |
85 | ||
86 | /* Creation of a cpu workqueue thread */ | |
87 | static void probe_workqueue_creation(struct task_struct *wq_thread, int cpu) | |
88 | { | |
89 | struct cpu_workqueue_stats *cws; | |
90 | unsigned long flags; | |
91 | ||
92 | WARN_ON(cpu < 0 || cpu >= num_possible_cpus()); | |
93 | ||
94 | /* Workqueues are sometimes created in atomic context */ | |
95 | cws = kzalloc(sizeof(struct cpu_workqueue_stats), GFP_ATOMIC); | |
96 | if (!cws) { | |
97 | pr_warning("trace_workqueue: not enough memory\n"); | |
98 | return; | |
99 | } | |
100 | tracing_record_cmdline(wq_thread); | |
101 | ||
102 | INIT_LIST_HEAD(&cws->list); | |
103 | cws->cpu = cpu; | |
104 | ||
105 | cws->pid = wq_thread->pid; | |
106 | ||
107 | spin_lock_irqsave(&all_workqueue_stat[cpu].lock, flags); | |
108 | if (list_empty(&all_workqueue_stat[cpu].list)) | |
109 | cws->first_entry = true; | |
110 | list_add_tail(&cws->list, &all_workqueue_stat[cpu].list); | |
111 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
112 | } | |
113 | ||
114 | /* Destruction of a cpu workqueue thread */ | |
115 | static void probe_workqueue_destruction(struct task_struct *wq_thread) | |
116 | { | |
117 | /* Workqueue only execute on one cpu */ | |
118 | int cpu = cpumask_first(&wq_thread->cpus_allowed); | |
119 | struct cpu_workqueue_stats *node, *next; | |
120 | unsigned long flags; | |
121 | ||
122 | spin_lock_irqsave(&all_workqueue_stat[cpu].lock, flags); | |
123 | list_for_each_entry_safe(node, next, &all_workqueue_stat[cpu].list, | |
124 | list) { | |
125 | if (node->pid == wq_thread->pid) { | |
126 | list_del(&node->list); | |
127 | kfree(node); | |
128 | goto found; | |
129 | } | |
130 | } | |
131 | ||
132 | pr_debug("trace_workqueue: don't find workqueue to destroy\n"); | |
133 | found: | |
134 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
135 | ||
136 | } | |
137 | ||
138 | static struct cpu_workqueue_stats *workqueue_stat_start_cpu(int cpu) | |
139 | { | |
140 | unsigned long flags; | |
141 | struct cpu_workqueue_stats *ret = NULL; | |
142 | ||
143 | ||
144 | spin_lock_irqsave(&all_workqueue_stat[cpu].lock, flags); | |
145 | ||
146 | if (!list_empty(&all_workqueue_stat[cpu].list)) | |
147 | ret = list_entry(all_workqueue_stat[cpu].list.next, | |
148 | struct cpu_workqueue_stats, list); | |
149 | ||
150 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
151 | ||
152 | return ret; | |
153 | } | |
154 | ||
155 | static void *workqueue_stat_start(void) | |
156 | { | |
157 | int cpu; | |
158 | void *ret = NULL; | |
159 | ||
160 | for_each_possible_cpu(cpu) { | |
161 | ret = workqueue_stat_start_cpu(cpu); | |
162 | if (ret) | |
163 | return ret; | |
164 | } | |
165 | return NULL; | |
166 | } | |
167 | ||
168 | static void *workqueue_stat_next(void *prev, int idx) | |
169 | { | |
170 | struct cpu_workqueue_stats *prev_cws = prev; | |
171 | int cpu = prev_cws->cpu; | |
172 | unsigned long flags; | |
173 | void *ret = NULL; | |
174 | ||
175 | spin_lock_irqsave(&all_workqueue_stat[cpu].lock, flags); | |
176 | if (list_is_last(&prev_cws->list, &all_workqueue_stat[cpu].list)) { | |
177 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
178 | for (++cpu ; cpu < num_possible_cpus(); cpu++) { | |
179 | ret = workqueue_stat_start_cpu(cpu); | |
180 | if (ret) | |
181 | return ret; | |
182 | } | |
183 | return NULL; | |
184 | } | |
185 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
186 | ||
187 | return list_entry(prev_cws->list.next, struct cpu_workqueue_stats, | |
188 | list); | |
189 | } | |
190 | ||
191 | static int workqueue_stat_show(struct seq_file *s, void *p) | |
192 | { | |
193 | struct cpu_workqueue_stats *cws = p; | |
194 | unsigned long flags; | |
195 | int cpu = cws->cpu; | |
196 | ||
197 | seq_printf(s, "%3d %6d %6u %s\n", cws->cpu, | |
198 | atomic_read(&cws->inserted), | |
199 | cws->executed, | |
200 | trace_find_cmdline(cws->pid)); | |
201 | ||
202 | spin_lock_irqsave(&all_workqueue_stat[cpu].lock, flags); | |
203 | if (&cws->list == all_workqueue_stat[cpu].list.next) | |
204 | seq_printf(s, "\n"); | |
205 | spin_unlock_irqrestore(&all_workqueue_stat[cpu].lock, flags); | |
206 | ||
207 | return 0; | |
208 | } | |
209 | ||
210 | static int workqueue_stat_headers(struct seq_file *s) | |
211 | { | |
212 | seq_printf(s, "# CPU INSERTED EXECUTED NAME\n"); | |
213 | seq_printf(s, "# | | | |\n\n"); | |
214 | return 0; | |
215 | } | |
216 | ||
217 | struct tracer_stat workqueue_stats __read_mostly = { | |
218 | .name = "workqueues", | |
219 | .stat_start = workqueue_stat_start, | |
220 | .stat_next = workqueue_stat_next, | |
221 | .stat_show = workqueue_stat_show, | |
222 | .stat_headers = workqueue_stat_headers | |
223 | }; | |
224 | ||
225 | ||
226 | int __init stat_workqueue_init(void) | |
227 | { | |
228 | if (register_stat_tracer(&workqueue_stats)) { | |
229 | pr_warning("Unable to register workqueue stat tracer\n"); | |
230 | return 1; | |
231 | } | |
232 | ||
233 | return 0; | |
234 | } | |
235 | fs_initcall(stat_workqueue_init); | |
236 | ||
237 | /* | |
238 | * Workqueues are created very early, just after pre-smp initcalls. | |
239 | * So we must register our tracepoints at this stage. | |
240 | */ | |
241 | int __init trace_workqueue_early_init(void) | |
242 | { | |
243 | int ret, cpu; | |
244 | ||
245 | ret = register_trace_workqueue_insertion(probe_workqueue_insertion); | |
246 | if (ret) | |
247 | goto out; | |
248 | ||
249 | ret = register_trace_workqueue_execution(probe_workqueue_execution); | |
250 | if (ret) | |
251 | goto no_insertion; | |
252 | ||
253 | ret = register_trace_workqueue_creation(probe_workqueue_creation); | |
254 | if (ret) | |
255 | goto no_execution; | |
256 | ||
257 | ret = register_trace_workqueue_destruction(probe_workqueue_destruction); | |
258 | if (ret) | |
259 | goto no_creation; | |
260 | ||
261 | all_workqueue_stat = kmalloc(sizeof(struct workqueue_global_stats) | |
262 | * num_possible_cpus(), GFP_KERNEL); | |
263 | ||
264 | if (!all_workqueue_stat) { | |
265 | pr_warning("trace_workqueue: not enough memory\n"); | |
266 | goto no_creation; | |
267 | } | |
268 | ||
269 | for_each_possible_cpu(cpu) { | |
270 | spin_lock_init(&all_workqueue_stat[cpu].lock); | |
271 | INIT_LIST_HEAD(&all_workqueue_stat[cpu].list); | |
272 | } | |
273 | ||
274 | return 0; | |
275 | ||
276 | no_creation: | |
277 | unregister_trace_workqueue_creation(probe_workqueue_creation); | |
278 | no_execution: | |
279 | unregister_trace_workqueue_execution(probe_workqueue_execution); | |
280 | no_insertion: | |
281 | unregister_trace_workqueue_insertion(probe_workqueue_insertion); | |
282 | out: | |
283 | pr_warning("trace_workqueue: unable to trace workqueues\n"); | |
284 | ||
285 | return 1; | |
286 | } | |
287 | early_initcall(trace_workqueue_early_init); |