]>
Commit | Line | Data |
---|---|---|
8bf05ed3 | 1 | // SPDX-License-Identifier: GPL-2.0+ |
8704baab PM |
2 | /* |
3 | * Read-Copy Update module-based performance-test facility | |
4 | * | |
8704baab PM |
5 | * Copyright (C) IBM Corporation, 2015 |
6 | * | |
8bf05ed3 | 7 | * Authors: Paul E. McKenney <paulmck@linux.ibm.com> |
8704baab | 8 | */ |
60500037 PM |
9 | |
10 | #define pr_fmt(fmt) fmt | |
11 | ||
8704baab PM |
12 | #include <linux/types.h> |
13 | #include <linux/kernel.h> | |
14 | #include <linux/init.h> | |
12af6603 | 15 | #include <linux/mm.h> |
8704baab PM |
16 | #include <linux/module.h> |
17 | #include <linux/kthread.h> | |
18 | #include <linux/err.h> | |
19 | #include <linux/spinlock.h> | |
20 | #include <linux/smp.h> | |
21 | #include <linux/rcupdate.h> | |
22 | #include <linux/interrupt.h> | |
23 | #include <linux/sched.h> | |
ae7e81c0 | 24 | #include <uapi/linux/sched/types.h> |
8704baab PM |
25 | #include <linux/atomic.h> |
26 | #include <linux/bitops.h> | |
27 | #include <linux/completion.h> | |
28 | #include <linux/moduleparam.h> | |
29 | #include <linux/percpu.h> | |
30 | #include <linux/notifier.h> | |
31 | #include <linux/reboot.h> | |
32 | #include <linux/freezer.h> | |
33 | #include <linux/cpu.h> | |
34 | #include <linux/delay.h> | |
35 | #include <linux/stat.h> | |
36 | #include <linux/srcu.h> | |
37 | #include <linux/slab.h> | |
38 | #include <asm/byteorder.h> | |
39 | #include <linux/torture.h> | |
40 | #include <linux/vmalloc.h> | |
41 | ||
25c36329 PM |
42 | #include "rcu.h" |
43 | ||
8704baab | 44 | MODULE_LICENSE("GPL"); |
8bf05ed3 | 45 | MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>"); |
8704baab PM |
46 | |
47 | #define PERF_FLAG "-perf:" | |
48 | #define PERFOUT_STRING(s) \ | |
a56fefa2 | 49 | pr_alert("%s" PERF_FLAG " %s\n", perf_type, s) |
8704baab PM |
50 | #define VERBOSE_PERFOUT_STRING(s) \ |
51 | do { if (verbose) pr_alert("%s" PERF_FLAG " %s\n", perf_type, s); } while (0) | |
52 | #define VERBOSE_PERFOUT_ERRSTRING(s) \ | |
53 | do { if (verbose) pr_alert("%s" PERF_FLAG "!!! %s\n", perf_type, s); } while (0) | |
54 | ||
85ba6bfe PM |
55 | /* |
56 | * The intended use cases for the nreaders and nwriters module parameters | |
57 | * are as follows: | |
58 | * | |
59 | * 1. Specify only the nr_cpus kernel boot parameter. This will | |
60 | * set both nreaders and nwriters to the value specified by | |
61 | * nr_cpus for a mixed reader/writer test. | |
62 | * | |
63 | * 2. Specify the nr_cpus kernel boot parameter, but set | |
64 | * rcuperf.nreaders to zero. This will set nwriters to the | |
65 | * value specified by nr_cpus for an update-only test. | |
66 | * | |
67 | * 3. Specify the nr_cpus kernel boot parameter, but set | |
68 | * rcuperf.nwriters to zero. This will set nreaders to the | |
69 | * value specified by nr_cpus for a read-only test. | |
70 | * | |
71 | * Various other use cases may of course be specified. | |
72 | */ | |
73 | ||
e838a7d6 PM |
74 | #ifdef MODULE |
75 | # define RCUPERF_SHUTDOWN 0 | |
76 | #else | |
77 | # define RCUPERF_SHUTDOWN 1 | |
78 | #endif | |
79 | ||
881ed593 PM |
80 | torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives"); |
81 | torture_param(int, gp_async_max, 1000, "Max # outstanding waits per reader"); | |
af06d4f7 | 82 | torture_param(bool, gp_exp, false, "Use expedited GP wait primitives"); |
df37e66b | 83 | torture_param(int, holdoff, 10, "Holdoff time before test start (s)"); |
85ba6bfe | 84 | torture_param(int, nreaders, -1, "Number of RCU reader threads"); |
8704baab | 85 | torture_param(int, nwriters, -1, "Number of RCU updater threads"); |
e838a7d6 | 86 | torture_param(bool, shutdown, RCUPERF_SHUTDOWN, |
492b95e5 | 87 | "Shutdown at end of performance tests."); |
90127d60 | 88 | torture_param(int, verbose, 1, "Enable verbose debugging printk()s"); |
820687a7 | 89 | torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable"); |
e6e78b00 | 90 | torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() perf test?"); |
f87dc808 | 91 | torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate."); |
8704baab PM |
92 | |
93 | static char *perf_type = "rcu"; | |
94 | module_param(perf_type, charp, 0444); | |
b3f3886c | 95 | MODULE_PARM_DESC(perf_type, "Type of RCU to performance-test (rcu, srcu, ...)"); |
8704baab PM |
96 | |
97 | static int nrealreaders; | |
98 | static int nrealwriters; | |
99 | static struct task_struct **writer_tasks; | |
100 | static struct task_struct **reader_tasks; | |
101 | static struct task_struct *shutdown_task; | |
102 | ||
103 | static u64 **writer_durations; | |
104 | static int *writer_n_durations; | |
105 | static atomic_t n_rcu_perf_reader_started; | |
106 | static atomic_t n_rcu_perf_writer_started; | |
107 | static atomic_t n_rcu_perf_writer_finished; | |
108 | static wait_queue_head_t shutdown_wq; | |
109 | static u64 t_rcu_perf_writer_started; | |
110 | static u64 t_rcu_perf_writer_finished; | |
e6e78b00 JFG |
111 | static unsigned long b_rcu_gp_test_started; |
112 | static unsigned long b_rcu_gp_test_finished; | |
881ed593 | 113 | static DEFINE_PER_CPU(atomic_t, n_async_inflight); |
8704baab | 114 | |
8704baab PM |
115 | #define MAX_MEAS 10000 |
116 | #define MIN_MEAS 100 | |
117 | ||
8704baab PM |
118 | /* |
119 | * Operations vector for selecting different types of tests. | |
120 | */ | |
121 | ||
122 | struct rcu_perf_ops { | |
123 | int ptype; | |
124 | void (*init)(void); | |
125 | void (*cleanup)(void); | |
126 | int (*readlock)(void); | |
127 | void (*readunlock)(int idx); | |
17ef2fe9 | 128 | unsigned long (*get_gp_seq)(void); |
d7219312 | 129 | unsigned long (*gp_diff)(unsigned long new, unsigned long old); |
8704baab | 130 | unsigned long (*exp_completed)(void); |
881ed593 PM |
131 | void (*async)(struct rcu_head *head, rcu_callback_t func); |
132 | void (*gp_barrier)(void); | |
8704baab PM |
133 | void (*sync)(void); |
134 | void (*exp_sync)(void); | |
135 | const char *name; | |
136 | }; | |
137 | ||
138 | static struct rcu_perf_ops *cur_ops; | |
139 | ||
140 | /* | |
141 | * Definitions for rcu perf testing. | |
142 | */ | |
143 | ||
144 | static int rcu_perf_read_lock(void) __acquires(RCU) | |
145 | { | |
146 | rcu_read_lock(); | |
147 | return 0; | |
148 | } | |
149 | ||
150 | static void rcu_perf_read_unlock(int idx) __releases(RCU) | |
151 | { | |
152 | rcu_read_unlock(); | |
153 | } | |
154 | ||
155 | static unsigned long __maybe_unused rcu_no_completed(void) | |
156 | { | |
157 | return 0; | |
158 | } | |
159 | ||
160 | static void rcu_sync_perf_init(void) | |
161 | { | |
162 | } | |
163 | ||
164 | static struct rcu_perf_ops rcu_ops = { | |
165 | .ptype = RCU_FLAVOR, | |
166 | .init = rcu_sync_perf_init, | |
167 | .readlock = rcu_perf_read_lock, | |
168 | .readunlock = rcu_perf_read_unlock, | |
17ef2fe9 | 169 | .get_gp_seq = rcu_get_gp_seq, |
d7219312 | 170 | .gp_diff = rcu_seq_diff, |
8704baab | 171 | .exp_completed = rcu_exp_batches_completed, |
881ed593 PM |
172 | .async = call_rcu, |
173 | .gp_barrier = rcu_barrier, | |
8704baab PM |
174 | .sync = synchronize_rcu, |
175 | .exp_sync = synchronize_rcu_expedited, | |
176 | .name = "rcu" | |
177 | }; | |
178 | ||
8704baab PM |
179 | /* |
180 | * Definitions for srcu perf testing. | |
181 | */ | |
182 | ||
183 | DEFINE_STATIC_SRCU(srcu_ctl_perf); | |
184 | static struct srcu_struct *srcu_ctlp = &srcu_ctl_perf; | |
185 | ||
186 | static int srcu_perf_read_lock(void) __acquires(srcu_ctlp) | |
187 | { | |
188 | return srcu_read_lock(srcu_ctlp); | |
189 | } | |
190 | ||
191 | static void srcu_perf_read_unlock(int idx) __releases(srcu_ctlp) | |
192 | { | |
193 | srcu_read_unlock(srcu_ctlp, idx); | |
194 | } | |
195 | ||
196 | static unsigned long srcu_perf_completed(void) | |
197 | { | |
198 | return srcu_batches_completed(srcu_ctlp); | |
199 | } | |
200 | ||
881ed593 PM |
201 | static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func) |
202 | { | |
203 | call_srcu(srcu_ctlp, head, func); | |
204 | } | |
205 | ||
206 | static void srcu_rcu_barrier(void) | |
207 | { | |
208 | srcu_barrier(srcu_ctlp); | |
209 | } | |
210 | ||
8704baab PM |
211 | static void srcu_perf_synchronize(void) |
212 | { | |
213 | synchronize_srcu(srcu_ctlp); | |
214 | } | |
215 | ||
216 | static void srcu_perf_synchronize_expedited(void) | |
217 | { | |
218 | synchronize_srcu_expedited(srcu_ctlp); | |
219 | } | |
220 | ||
221 | static struct rcu_perf_ops srcu_ops = { | |
222 | .ptype = SRCU_FLAVOR, | |
223 | .init = rcu_sync_perf_init, | |
224 | .readlock = srcu_perf_read_lock, | |
225 | .readunlock = srcu_perf_read_unlock, | |
17ef2fe9 | 226 | .get_gp_seq = srcu_perf_completed, |
d7219312 | 227 | .gp_diff = rcu_seq_diff, |
8704baab | 228 | .exp_completed = srcu_perf_completed, |
881ed593 PM |
229 | .async = srcu_call_rcu, |
230 | .gp_barrier = srcu_rcu_barrier, | |
8704baab PM |
231 | .sync = srcu_perf_synchronize, |
232 | .exp_sync = srcu_perf_synchronize_expedited, | |
233 | .name = "srcu" | |
234 | }; | |
235 | ||
f60cb4d4 PM |
236 | static struct srcu_struct srcud; |
237 | ||
238 | static void srcu_sync_perf_init(void) | |
239 | { | |
240 | srcu_ctlp = &srcud; | |
241 | init_srcu_struct(srcu_ctlp); | |
242 | } | |
243 | ||
244 | static void srcu_sync_perf_cleanup(void) | |
245 | { | |
246 | cleanup_srcu_struct(srcu_ctlp); | |
247 | } | |
248 | ||
249 | static struct rcu_perf_ops srcud_ops = { | |
250 | .ptype = SRCU_FLAVOR, | |
251 | .init = srcu_sync_perf_init, | |
252 | .cleanup = srcu_sync_perf_cleanup, | |
253 | .readlock = srcu_perf_read_lock, | |
254 | .readunlock = srcu_perf_read_unlock, | |
17ef2fe9 | 255 | .get_gp_seq = srcu_perf_completed, |
d7219312 | 256 | .gp_diff = rcu_seq_diff, |
f60cb4d4 PM |
257 | .exp_completed = srcu_perf_completed, |
258 | .async = srcu_call_rcu, | |
259 | .gp_barrier = srcu_rcu_barrier, | |
260 | .sync = srcu_perf_synchronize, | |
261 | .exp_sync = srcu_perf_synchronize_expedited, | |
262 | .name = "srcud" | |
263 | }; | |
264 | ||
8704baab PM |
265 | /* |
266 | * Definitions for RCU-tasks perf testing. | |
267 | */ | |
268 | ||
269 | static int tasks_perf_read_lock(void) | |
270 | { | |
271 | return 0; | |
272 | } | |
273 | ||
274 | static void tasks_perf_read_unlock(int idx) | |
275 | { | |
276 | } | |
277 | ||
278 | static struct rcu_perf_ops tasks_ops = { | |
279 | .ptype = RCU_TASKS_FLAVOR, | |
280 | .init = rcu_sync_perf_init, | |
281 | .readlock = tasks_perf_read_lock, | |
282 | .readunlock = tasks_perf_read_unlock, | |
17ef2fe9 | 283 | .get_gp_seq = rcu_no_completed, |
d7219312 | 284 | .gp_diff = rcu_seq_diff, |
881ed593 PM |
285 | .async = call_rcu_tasks, |
286 | .gp_barrier = rcu_barrier_tasks, | |
8704baab PM |
287 | .sync = synchronize_rcu_tasks, |
288 | .exp_sync = synchronize_rcu_tasks, | |
289 | .name = "tasks" | |
290 | }; | |
291 | ||
d7219312 PM |
292 | static unsigned long rcuperf_seq_diff(unsigned long new, unsigned long old) |
293 | { | |
294 | if (!cur_ops->gp_diff) | |
295 | return new - old; | |
296 | return cur_ops->gp_diff(new, old); | |
297 | } | |
298 | ||
8704baab PM |
299 | /* |
300 | * If performance tests complete, wait for shutdown to commence. | |
301 | */ | |
302 | static void rcu_perf_wait_shutdown(void) | |
303 | { | |
cee43939 | 304 | cond_resched_tasks_rcu_qs(); |
8704baab PM |
305 | if (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters) |
306 | return; | |
307 | while (!torture_must_stop()) | |
308 | schedule_timeout_uninterruptible(1); | |
309 | } | |
310 | ||
311 | /* | |
312 | * RCU perf reader kthread. Repeatedly does empty RCU read-side | |
313 | * critical section, minimizing update-side interference. | |
314 | */ | |
315 | static int | |
316 | rcu_perf_reader(void *arg) | |
317 | { | |
318 | unsigned long flags; | |
319 | int idx; | |
6b558c4c | 320 | long me = (long)arg; |
8704baab PM |
321 | |
322 | VERBOSE_PERFOUT_STRING("rcu_perf_reader task started"); | |
6b558c4c | 323 | set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids)); |
8704baab PM |
324 | set_user_nice(current, MAX_NICE); |
325 | atomic_inc(&n_rcu_perf_reader_started); | |
326 | ||
327 | do { | |
328 | local_irq_save(flags); | |
329 | idx = cur_ops->readlock(); | |
330 | cur_ops->readunlock(idx); | |
331 | local_irq_restore(flags); | |
332 | rcu_perf_wait_shutdown(); | |
333 | } while (!torture_must_stop()); | |
334 | torture_kthread_stopping("rcu_perf_reader"); | |
335 | return 0; | |
336 | } | |
337 | ||
881ed593 PM |
338 | /* |
339 | * Callback function for asynchronous grace periods from rcu_perf_writer(). | |
340 | */ | |
341 | static void rcu_perf_async_cb(struct rcu_head *rhp) | |
342 | { | |
343 | atomic_dec(this_cpu_ptr(&n_async_inflight)); | |
344 | kfree(rhp); | |
345 | } | |
346 | ||
8704baab PM |
347 | /* |
348 | * RCU perf writer kthread. Repeatedly does a grace period. | |
349 | */ | |
350 | static int | |
351 | rcu_perf_writer(void *arg) | |
352 | { | |
353 | int i = 0; | |
354 | int i_max; | |
355 | long me = (long)arg; | |
881ed593 | 356 | struct rcu_head *rhp = NULL; |
2094c995 | 357 | struct sched_param sp; |
8704baab PM |
358 | bool started = false, done = false, alldone = false; |
359 | u64 t; | |
360 | u64 *wdp; | |
361 | u64 *wdpp = writer_durations[me]; | |
362 | ||
363 | VERBOSE_PERFOUT_STRING("rcu_perf_writer task started"); | |
8704baab | 364 | WARN_ON(!wdpp); |
6b558c4c | 365 | set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids)); |
2094c995 PM |
366 | sp.sched_priority = 1; |
367 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); | |
df37e66b PM |
368 | |
369 | if (holdoff) | |
370 | schedule_timeout_uninterruptible(holdoff * HZ); | |
371 | ||
77e9752c JFG |
372 | /* |
373 | * Wait until rcu_end_inkernel_boot() is called for normal GP tests | |
374 | * so that RCU is not always expedited for normal GP tests. | |
375 | * The system_state test is approximate, but works well in practice. | |
376 | */ | |
377 | while (!gp_exp && system_state != SYSTEM_RUNNING) | |
378 | schedule_timeout_uninterruptible(1); | |
379 | ||
8704baab PM |
380 | t = ktime_get_mono_fast_ns(); |
381 | if (atomic_inc_return(&n_rcu_perf_writer_started) >= nrealwriters) { | |
382 | t_rcu_perf_writer_started = t; | |
383 | if (gp_exp) { | |
e6e78b00 | 384 | b_rcu_gp_test_started = |
8704baab PM |
385 | cur_ops->exp_completed() / 2; |
386 | } else { | |
e6e78b00 | 387 | b_rcu_gp_test_started = cur_ops->get_gp_seq(); |
8704baab PM |
388 | } |
389 | } | |
390 | ||
391 | do { | |
820687a7 PM |
392 | if (writer_holdoff) |
393 | udelay(writer_holdoff); | |
8704baab PM |
394 | wdp = &wdpp[i]; |
395 | *wdp = ktime_get_mono_fast_ns(); | |
881ed593 PM |
396 | if (gp_async) { |
397 | retry: | |
398 | if (!rhp) | |
399 | rhp = kmalloc(sizeof(*rhp), GFP_KERNEL); | |
400 | if (rhp && atomic_read(this_cpu_ptr(&n_async_inflight)) < gp_async_max) { | |
881ed593 PM |
401 | atomic_inc(this_cpu_ptr(&n_async_inflight)); |
402 | cur_ops->async(rhp, rcu_perf_async_cb); | |
403 | rhp = NULL; | |
404 | } else if (!kthread_should_stop()) { | |
881ed593 PM |
405 | cur_ops->gp_barrier(); |
406 | goto retry; | |
407 | } else { | |
408 | kfree(rhp); /* Because we are stopping. */ | |
409 | } | |
410 | } else if (gp_exp) { | |
8704baab PM |
411 | cur_ops->exp_sync(); |
412 | } else { | |
8704baab PM |
413 | cur_ops->sync(); |
414 | } | |
8704baab PM |
415 | t = ktime_get_mono_fast_ns(); |
416 | *wdp = t - *wdp; | |
417 | i_max = i; | |
418 | if (!started && | |
419 | atomic_read(&n_rcu_perf_writer_started) >= nrealwriters) | |
420 | started = true; | |
421 | if (!done && i >= MIN_MEAS) { | |
422 | done = true; | |
620316e5 PM |
423 | sp.sched_priority = 0; |
424 | sched_setscheduler_nocheck(current, | |
425 | SCHED_NORMAL, &sp); | |
a56fefa2 SP |
426 | pr_alert("%s%s rcu_perf_writer %ld has %d measurements\n", |
427 | perf_type, PERF_FLAG, me, MIN_MEAS); | |
8704baab PM |
428 | if (atomic_inc_return(&n_rcu_perf_writer_finished) >= |
429 | nrealwriters) { | |
620316e5 | 430 | schedule_timeout_interruptible(10); |
ac2bb275 | 431 | rcu_ftrace_dump(DUMP_ALL); |
8704baab PM |
432 | PERFOUT_STRING("Test complete"); |
433 | t_rcu_perf_writer_finished = t; | |
434 | if (gp_exp) { | |
e6e78b00 | 435 | b_rcu_gp_test_finished = |
8704baab PM |
436 | cur_ops->exp_completed() / 2; |
437 | } else { | |
e6e78b00 | 438 | b_rcu_gp_test_finished = |
17ef2fe9 | 439 | cur_ops->get_gp_seq(); |
8704baab | 440 | } |
e6fb1fc1 AS |
441 | if (shutdown) { |
442 | smp_mb(); /* Assign before wake. */ | |
443 | wake_up(&shutdown_wq); | |
444 | } | |
8704baab PM |
445 | } |
446 | } | |
447 | if (done && !alldone && | |
448 | atomic_read(&n_rcu_perf_writer_finished) >= nrealwriters) | |
449 | alldone = true; | |
450 | if (started && !alldone && i < MAX_MEAS - 1) | |
451 | i++; | |
452 | rcu_perf_wait_shutdown(); | |
453 | } while (!torture_must_stop()); | |
881ed593 | 454 | if (gp_async) { |
881ed593 PM |
455 | cur_ops->gp_barrier(); |
456 | } | |
8704baab PM |
457 | writer_n_durations[me] = i_max; |
458 | torture_kthread_stopping("rcu_perf_writer"); | |
459 | return 0; | |
460 | } | |
461 | ||
96221795 | 462 | static void |
8704baab PM |
463 | rcu_perf_print_module_parms(struct rcu_perf_ops *cur_ops, const char *tag) |
464 | { | |
465 | pr_alert("%s" PERF_FLAG | |
466 | "--- %s: nreaders=%d nwriters=%d verbose=%d shutdown=%d\n", | |
467 | perf_type, tag, nrealreaders, nrealwriters, verbose, shutdown); | |
468 | } | |
469 | ||
470 | static void | |
471 | rcu_perf_cleanup(void) | |
472 | { | |
473 | int i; | |
474 | int j; | |
475 | int ngps = 0; | |
476 | u64 *wdp; | |
477 | u64 *wdpp; | |
478 | ||
9683937d PM |
479 | /* |
480 | * Would like warning at start, but everything is expedited | |
481 | * during the mid-boot phase, so have to wait till the end. | |
482 | */ | |
483 | if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp) | |
484 | VERBOSE_PERFOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!"); | |
485 | if (rcu_gp_is_normal() && gp_exp) | |
486 | VERBOSE_PERFOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!"); | |
881ed593 PM |
487 | if (gp_exp && gp_async) |
488 | VERBOSE_PERFOUT_ERRSTRING("No expedited async GPs, so went with async!"); | |
9683937d | 489 | |
8704baab PM |
490 | if (torture_cleanup_begin()) |
491 | return; | |
ad092c02 PM |
492 | if (!cur_ops) { |
493 | torture_cleanup_end(); | |
494 | return; | |
495 | } | |
8704baab PM |
496 | |
497 | if (reader_tasks) { | |
498 | for (i = 0; i < nrealreaders; i++) | |
499 | torture_stop_kthread(rcu_perf_reader, | |
500 | reader_tasks[i]); | |
501 | kfree(reader_tasks); | |
502 | } | |
503 | ||
504 | if (writer_tasks) { | |
505 | for (i = 0; i < nrealwriters; i++) { | |
506 | torture_stop_kthread(rcu_perf_writer, | |
507 | writer_tasks[i]); | |
508 | if (!writer_n_durations) | |
509 | continue; | |
510 | j = writer_n_durations[i]; | |
511 | pr_alert("%s%s writer %d gps: %d\n", | |
512 | perf_type, PERF_FLAG, i, j); | |
513 | ngps += j; | |
514 | } | |
515 | pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n", | |
516 | perf_type, PERF_FLAG, | |
517 | t_rcu_perf_writer_started, t_rcu_perf_writer_finished, | |
518 | t_rcu_perf_writer_finished - | |
519 | t_rcu_perf_writer_started, | |
520 | ngps, | |
e6e78b00 JFG |
521 | rcuperf_seq_diff(b_rcu_gp_test_finished, |
522 | b_rcu_gp_test_started)); | |
8704baab PM |
523 | for (i = 0; i < nrealwriters; i++) { |
524 | if (!writer_durations) | |
525 | break; | |
526 | if (!writer_n_durations) | |
527 | continue; | |
528 | wdpp = writer_durations[i]; | |
529 | if (!wdpp) | |
530 | continue; | |
531 | for (j = 0; j <= writer_n_durations[i]; j++) { | |
532 | wdp = &wdpp[j]; | |
533 | pr_alert("%s%s %4d writer-duration: %5d %llu\n", | |
534 | perf_type, PERF_FLAG, | |
535 | i, j, *wdp); | |
536 | if (j % 100 == 0) | |
537 | schedule_timeout_uninterruptible(1); | |
538 | } | |
539 | kfree(writer_durations[i]); | |
540 | } | |
541 | kfree(writer_tasks); | |
542 | kfree(writer_durations); | |
543 | kfree(writer_n_durations); | |
544 | } | |
545 | ||
620d2460 | 546 | /* Do torture-type-specific cleanup operations. */ |
8704baab PM |
547 | if (cur_ops->cleanup != NULL) |
548 | cur_ops->cleanup(); | |
549 | ||
550 | torture_cleanup_end(); | |
551 | } | |
552 | ||
553 | /* | |
554 | * Return the number if non-negative. If -1, the number of CPUs. | |
555 | * If less than -1, that much less than the number of CPUs, but | |
556 | * at least one. | |
557 | */ | |
558 | static int compute_real(int n) | |
559 | { | |
560 | int nr; | |
561 | ||
562 | if (n >= 0) { | |
563 | nr = n; | |
564 | } else { | |
565 | nr = num_online_cpus() + 1 + n; | |
566 | if (nr <= 0) | |
567 | nr = 1; | |
568 | } | |
569 | return nr; | |
570 | } | |
571 | ||
572 | /* | |
573 | * RCU perf shutdown kthread. Just waits to be awakened, then shuts | |
574 | * down system. | |
575 | */ | |
576 | static int | |
577 | rcu_perf_shutdown(void *arg) | |
578 | { | |
579 | do { | |
580 | wait_event(shutdown_wq, | |
581 | atomic_read(&n_rcu_perf_writer_finished) >= | |
582 | nrealwriters); | |
583 | } while (atomic_read(&n_rcu_perf_writer_finished) < nrealwriters); | |
584 | smp_mb(); /* Wake before output. */ | |
585 | rcu_perf_cleanup(); | |
586 | kernel_power_off(); | |
587 | return -EINVAL; | |
588 | } | |
589 | ||
e6e78b00 JFG |
590 | /* |
591 | * kfree_rcu() performance tests: Start a kfree_rcu() loop on all CPUs for number | |
592 | * of iterations and measure total time and number of GP for all iterations to complete. | |
593 | */ | |
594 | ||
595 | torture_param(int, kfree_nthreads, -1, "Number of threads running loops of kfree_rcu()."); | |
596 | torture_param(int, kfree_alloc_num, 8000, "Number of allocations and frees done in an iteration."); | |
597 | torture_param(int, kfree_loops, 10, "Number of loops doing kfree_alloc_num allocations and frees."); | |
e6e78b00 JFG |
598 | |
599 | static struct task_struct **kfree_reader_tasks; | |
600 | static int kfree_nrealthreads; | |
601 | static atomic_t n_kfree_perf_thread_started; | |
602 | static atomic_t n_kfree_perf_thread_ended; | |
603 | ||
604 | struct kfree_obj { | |
605 | char kfree_obj[8]; | |
606 | struct rcu_head rh; | |
607 | }; | |
608 | ||
609 | static int | |
610 | kfree_perf_thread(void *arg) | |
611 | { | |
612 | int i, loop = 0; | |
613 | long me = (long)arg; | |
614 | struct kfree_obj *alloc_ptr; | |
615 | u64 start_time, end_time; | |
12af6603 | 616 | long long mem_begin, mem_during = 0; |
e6e78b00 JFG |
617 | |
618 | VERBOSE_PERFOUT_STRING("kfree_perf_thread task started"); | |
619 | set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids)); | |
620 | set_user_nice(current, MAX_NICE); | |
621 | ||
622 | start_time = ktime_get_mono_fast_ns(); | |
623 | ||
624 | if (atomic_inc_return(&n_kfree_perf_thread_started) >= kfree_nrealthreads) { | |
625 | if (gp_exp) | |
626 | b_rcu_gp_test_started = cur_ops->exp_completed() / 2; | |
627 | else | |
628 | b_rcu_gp_test_started = cur_ops->get_gp_seq(); | |
629 | } | |
630 | ||
631 | do { | |
12af6603 JFG |
632 | if (!mem_during) { |
633 | mem_during = mem_begin = si_mem_available(); | |
634 | } else if (loop % (kfree_loops / 4) == 0) { | |
635 | mem_during = (mem_during + si_mem_available()) / 2; | |
636 | } | |
637 | ||
e6e78b00 | 638 | for (i = 0; i < kfree_alloc_num; i++) { |
f87dc808 | 639 | alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL); |
e6e78b00 JFG |
640 | if (!alloc_ptr) |
641 | return -ENOMEM; | |
642 | ||
189a6883 | 643 | kfree_rcu(alloc_ptr, rh); |
e6e78b00 JFG |
644 | } |
645 | ||
646 | cond_resched(); | |
647 | } while (!torture_must_stop() && ++loop < kfree_loops); | |
648 | ||
649 | if (atomic_inc_return(&n_kfree_perf_thread_ended) >= kfree_nrealthreads) { | |
650 | end_time = ktime_get_mono_fast_ns(); | |
651 | ||
652 | if (gp_exp) | |
653 | b_rcu_gp_test_finished = cur_ops->exp_completed() / 2; | |
654 | else | |
655 | b_rcu_gp_test_finished = cur_ops->get_gp_seq(); | |
656 | ||
12af6603 | 657 | pr_alert("Total time taken by all kfree'ers: %llu ns, loops: %d, batches: %ld, memory footprint: %lldMB\n", |
e6e78b00 | 658 | (unsigned long long)(end_time - start_time), kfree_loops, |
12af6603 JFG |
659 | rcuperf_seq_diff(b_rcu_gp_test_finished, b_rcu_gp_test_started), |
660 | (mem_begin - mem_during) >> (20 - PAGE_SHIFT)); | |
661 | ||
e6e78b00 JFG |
662 | if (shutdown) { |
663 | smp_mb(); /* Assign before wake. */ | |
664 | wake_up(&shutdown_wq); | |
665 | } | |
666 | } | |
667 | ||
668 | torture_kthread_stopping("kfree_perf_thread"); | |
669 | return 0; | |
670 | } | |
671 | ||
672 | static void | |
673 | kfree_perf_cleanup(void) | |
674 | { | |
675 | int i; | |
676 | ||
677 | if (torture_cleanup_begin()) | |
678 | return; | |
679 | ||
680 | if (kfree_reader_tasks) { | |
681 | for (i = 0; i < kfree_nrealthreads; i++) | |
682 | torture_stop_kthread(kfree_perf_thread, | |
683 | kfree_reader_tasks[i]); | |
684 | kfree(kfree_reader_tasks); | |
685 | } | |
686 | ||
687 | torture_cleanup_end(); | |
688 | } | |
689 | ||
690 | /* | |
691 | * shutdown kthread. Just waits to be awakened, then shuts down system. | |
692 | */ | |
693 | static int | |
694 | kfree_perf_shutdown(void *arg) | |
695 | { | |
696 | do { | |
697 | wait_event(shutdown_wq, | |
698 | atomic_read(&n_kfree_perf_thread_ended) >= | |
699 | kfree_nrealthreads); | |
700 | } while (atomic_read(&n_kfree_perf_thread_ended) < kfree_nrealthreads); | |
701 | ||
702 | smp_mb(); /* Wake before output. */ | |
703 | ||
704 | kfree_perf_cleanup(); | |
705 | kernel_power_off(); | |
706 | return -EINVAL; | |
707 | } | |
708 | ||
709 | static int __init | |
710 | kfree_perf_init(void) | |
711 | { | |
712 | long i; | |
713 | int firsterr = 0; | |
714 | ||
715 | kfree_nrealthreads = compute_real(kfree_nthreads); | |
716 | /* Start up the kthreads. */ | |
717 | if (shutdown) { | |
718 | init_waitqueue_head(&shutdown_wq); | |
719 | firsterr = torture_create_kthread(kfree_perf_shutdown, NULL, | |
720 | shutdown_task); | |
721 | if (firsterr) | |
722 | goto unwind; | |
723 | schedule_timeout_uninterruptible(1); | |
724 | } | |
725 | ||
f87dc808 JFG |
726 | pr_alert("kfree object size=%lu\n", kfree_mult * sizeof(struct kfree_obj)); |
727 | ||
e6e78b00 JFG |
728 | kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]), |
729 | GFP_KERNEL); | |
730 | if (kfree_reader_tasks == NULL) { | |
731 | firsterr = -ENOMEM; | |
732 | goto unwind; | |
733 | } | |
734 | ||
735 | for (i = 0; i < kfree_nrealthreads; i++) { | |
736 | firsterr = torture_create_kthread(kfree_perf_thread, (void *)i, | |
737 | kfree_reader_tasks[i]); | |
738 | if (firsterr) | |
739 | goto unwind; | |
740 | } | |
741 | ||
742 | while (atomic_read(&n_kfree_perf_thread_started) < kfree_nrealthreads) | |
743 | schedule_timeout_uninterruptible(1); | |
744 | ||
745 | torture_init_end(); | |
746 | return 0; | |
747 | ||
748 | unwind: | |
749 | torture_init_end(); | |
750 | kfree_perf_cleanup(); | |
751 | return firsterr; | |
752 | } | |
753 | ||
8704baab PM |
754 | static int __init |
755 | rcu_perf_init(void) | |
756 | { | |
757 | long i; | |
758 | int firsterr = 0; | |
759 | static struct rcu_perf_ops *perf_ops[] = { | |
620d2460 | 760 | &rcu_ops, &srcu_ops, &srcud_ops, &tasks_ops, |
8704baab PM |
761 | }; |
762 | ||
a2f2577d | 763 | if (!torture_init_begin(perf_type, verbose)) |
8704baab PM |
764 | return -EBUSY; |
765 | ||
766 | /* Process args and tell the world that the perf'er is on the job. */ | |
767 | for (i = 0; i < ARRAY_SIZE(perf_ops); i++) { | |
768 | cur_ops = perf_ops[i]; | |
769 | if (strcmp(perf_type, cur_ops->name) == 0) | |
770 | break; | |
771 | } | |
772 | if (i == ARRAY_SIZE(perf_ops)) { | |
a7538352 | 773 | pr_alert("rcu-perf: invalid perf type: \"%s\"\n", perf_type); |
8704baab PM |
774 | pr_alert("rcu-perf types:"); |
775 | for (i = 0; i < ARRAY_SIZE(perf_ops); i++) | |
a7538352 JP |
776 | pr_cont(" %s", perf_ops[i]->name); |
777 | pr_cont("\n"); | |
f0288064 | 778 | WARN_ON(!IS_MODULE(CONFIG_RCU_PERF_TEST)); |
8704baab | 779 | firsterr = -EINVAL; |
ad092c02 | 780 | cur_ops = NULL; |
8704baab PM |
781 | goto unwind; |
782 | } | |
783 | if (cur_ops->init) | |
784 | cur_ops->init(); | |
785 | ||
e6e78b00 JFG |
786 | if (kfree_rcu_test) |
787 | return kfree_perf_init(); | |
788 | ||
8704baab PM |
789 | nrealwriters = compute_real(nwriters); |
790 | nrealreaders = compute_real(nreaders); | |
791 | atomic_set(&n_rcu_perf_reader_started, 0); | |
792 | atomic_set(&n_rcu_perf_writer_started, 0); | |
793 | atomic_set(&n_rcu_perf_writer_finished, 0); | |
794 | rcu_perf_print_module_parms(cur_ops, "Start of test"); | |
795 | ||
796 | /* Start up the kthreads. */ | |
797 | ||
798 | if (shutdown) { | |
799 | init_waitqueue_head(&shutdown_wq); | |
800 | firsterr = torture_create_kthread(rcu_perf_shutdown, NULL, | |
801 | shutdown_task); | |
802 | if (firsterr) | |
803 | goto unwind; | |
804 | schedule_timeout_uninterruptible(1); | |
805 | } | |
806 | reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]), | |
807 | GFP_KERNEL); | |
808 | if (reader_tasks == NULL) { | |
809 | VERBOSE_PERFOUT_ERRSTRING("out of memory"); | |
810 | firsterr = -ENOMEM; | |
811 | goto unwind; | |
812 | } | |
813 | for (i = 0; i < nrealreaders; i++) { | |
6b558c4c | 814 | firsterr = torture_create_kthread(rcu_perf_reader, (void *)i, |
8704baab PM |
815 | reader_tasks[i]); |
816 | if (firsterr) | |
817 | goto unwind; | |
818 | } | |
819 | while (atomic_read(&n_rcu_perf_reader_started) < nrealreaders) | |
820 | schedule_timeout_uninterruptible(1); | |
821 | writer_tasks = kcalloc(nrealwriters, sizeof(reader_tasks[0]), | |
822 | GFP_KERNEL); | |
823 | writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations), | |
824 | GFP_KERNEL); | |
825 | writer_n_durations = | |
826 | kcalloc(nrealwriters, sizeof(*writer_n_durations), | |
827 | GFP_KERNEL); | |
828 | if (!writer_tasks || !writer_durations || !writer_n_durations) { | |
829 | VERBOSE_PERFOUT_ERRSTRING("out of memory"); | |
830 | firsterr = -ENOMEM; | |
831 | goto unwind; | |
832 | } | |
833 | for (i = 0; i < nrealwriters; i++) { | |
834 | writer_durations[i] = | |
835 | kcalloc(MAX_MEAS, sizeof(*writer_durations[i]), | |
836 | GFP_KERNEL); | |
05dbbfe7 WY |
837 | if (!writer_durations[i]) { |
838 | firsterr = -ENOMEM; | |
8704baab | 839 | goto unwind; |
05dbbfe7 | 840 | } |
8704baab PM |
841 | firsterr = torture_create_kthread(rcu_perf_writer, (void *)i, |
842 | writer_tasks[i]); | |
843 | if (firsterr) | |
844 | goto unwind; | |
845 | } | |
846 | torture_init_end(); | |
847 | return 0; | |
848 | ||
849 | unwind: | |
850 | torture_init_end(); | |
851 | rcu_perf_cleanup(); | |
852 | return firsterr; | |
853 | } | |
854 | ||
855 | module_init(rcu_perf_init); | |
856 | module_exit(rcu_perf_cleanup); |