1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * Read-Copy Update module-based scalability-test facility
4 *
5 * Copyright (C) IBM Corporation, 2015
6 *
7 * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
8 */
9
10 #define pr_fmt(fmt) fmt
11
12 #include <linux/types.h>
13 #include <linux/kernel.h>
14 #include <linux/init.h>
15 #include <linux/mm.h>
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>
24 #include <uapi/linux/sched/types.h>
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 #include <linux/rcupdate_trace.h>
42 #include <linux/sched/debug.h>
43
44 #include "rcu.h"
45
46 MODULE_DESCRIPTION("Read-Copy Update module-based scalability-test facility");
47 MODULE_LICENSE("GPL");
48 MODULE_AUTHOR("Paul E. McKenney <paulmck@linux.ibm.com>");
49
50 #define SCALE_FLAG "-scale:"
51 #define SCALEOUT_STRING(s) \
52 pr_alert("%s" SCALE_FLAG " %s\n", scale_type, s)
53 #define VERBOSE_SCALEOUT_STRING(s) \
54 do { if (verbose) pr_alert("%s" SCALE_FLAG " %s\n", scale_type, s); } while (0)
55 #define SCALEOUT_ERRSTRING(s) \
56 pr_alert("%s" SCALE_FLAG "!!! %s\n", scale_type, s)
57
58 /*
59 * The intended use cases for the nreaders and nwriters module parameters
60 * are as follows:
61 *
62 * 1. Specify only the nr_cpus kernel boot parameter. This will
63 * set both nreaders and nwriters to the value specified by
64 * nr_cpus for a mixed reader/writer test.
65 *
66 * 2. Specify the nr_cpus kernel boot parameter, but set
67 * rcuscale.nreaders to zero. This will set nwriters to the
68 * value specified by nr_cpus for an update-only test.
69 *
70 * 3. Specify the nr_cpus kernel boot parameter, but set
71 * rcuscale.nwriters to zero. This will set nreaders to the
72 * value specified by nr_cpus for a read-only test.
73 *
74 * Various other use cases may of course be specified.
75 *
76 * Note that this test's readers are intended only as a test load for
77 * the writers. The reader scalability statistics will be overly
78 * pessimistic due to the per-critical-section interrupt disabling,
79 * test-end checks, and the pair of calls through pointers.
80 */
81
82 #ifdef MODULE
83 # define RCUSCALE_SHUTDOWN 0
84 #else
85 # define RCUSCALE_SHUTDOWN 1
86 #endif
87
88 torture_param(bool, gp_async, false, "Use asynchronous GP wait primitives");
89 torture_param(int, gp_async_max, 1000, "Max # outstanding waits per writer");
90 torture_param(bool, gp_exp, false, "Use expedited GP wait primitives");
91 torture_param(int, holdoff, 10, "Holdoff time before test start (s)");
92 torture_param(int, minruntime, 0, "Minimum run time (s)");
93 torture_param(int, nreaders, -1, "Number of RCU reader threads");
94 torture_param(int, nwriters, -1, "Number of RCU updater threads");
95 torture_param(bool, shutdown, RCUSCALE_SHUTDOWN,
96 "Shutdown at end of scalability tests.");
97 torture_param(int, verbose, 1, "Enable verbose debugging printk()s");
98 torture_param(int, writer_holdoff, 0, "Holdoff (us) between GPs, zero to disable");
99 torture_param(int, writer_holdoff_jiffies, 0, "Holdoff (jiffies) between GPs, zero to disable");
100 torture_param(int, kfree_rcu_test, 0, "Do we run a kfree_rcu() scale test?");
101 torture_param(int, kfree_mult, 1, "Multiple of kfree_obj size to allocate.");
102 torture_param(int, kfree_by_call_rcu, 0, "Use call_rcu() to emulate kfree_rcu()?");
103
104 static char *scale_type = "rcu";
105 module_param(scale_type, charp, 0444);
106 MODULE_PARM_DESC(scale_type, "Type of RCU to scalability-test (rcu, srcu, ...)");
107
108 // Structure definitions for custom fixed-per-task allocator.
109 struct writer_mblock {
110 struct rcu_head wmb_rh;
111 struct llist_node wmb_node;
112 struct writer_freelist *wmb_wfl;
113 };
114
115 struct writer_freelist {
116 struct llist_head ws_lhg;
117 atomic_t ws_inflight;
118 struct llist_head ____cacheline_internodealigned_in_smp ws_lhp;
119 struct writer_mblock *ws_mblocks;
120 };
121
122 static int nrealreaders;
123 static int nrealwriters;
124 static struct task_struct **writer_tasks;
125 static struct task_struct **reader_tasks;
126 static struct task_struct *shutdown_task;
127
128 static u64 **writer_durations;
129 static bool *writer_done;
130 static struct writer_freelist *writer_freelists;
131 static int *writer_n_durations;
132 static atomic_t n_rcu_scale_reader_started;
133 static atomic_t n_rcu_scale_writer_started;
134 static atomic_t n_rcu_scale_writer_finished;
135 static wait_queue_head_t shutdown_wq;
136 static u64 t_rcu_scale_writer_started;
137 static u64 t_rcu_scale_writer_finished;
138 static unsigned long b_rcu_gp_test_started;
139 static unsigned long b_rcu_gp_test_finished;
140
141 #define MAX_MEAS 10000
142 #define MIN_MEAS 100
143
144 /*
145 * Operations vector for selecting different types of tests.
146 */
147
148 struct rcu_scale_ops {
149 int ptype;
150 void (*init)(void);
151 void (*cleanup)(void);
152 int (*readlock)(void);
153 void (*readunlock)(int idx);
154 unsigned long (*get_gp_seq)(void);
155 unsigned long (*gp_diff)(unsigned long new, unsigned long old);
156 unsigned long (*exp_completed)(void);
157 void (*async)(struct rcu_head *head, rcu_callback_t func);
158 void (*gp_barrier)(void);
159 void (*sync)(void);
160 void (*exp_sync)(void);
161 struct task_struct *(*rso_gp_kthread)(void);
162 void (*stats)(void);
163 const char *name;
164 };
165
166 static struct rcu_scale_ops *cur_ops;
167
168 /*
169 * Definitions for rcu scalability testing.
170 */
171
rcu_scale_read_lock(void)172 static int rcu_scale_read_lock(void) __acquires(RCU)
173 {
174 rcu_read_lock();
175 return 0;
176 }
177
rcu_scale_read_unlock(int idx)178 static void rcu_scale_read_unlock(int idx) __releases(RCU)
179 {
180 rcu_read_unlock();
181 }
182
rcu_no_completed(void)183 static unsigned long __maybe_unused rcu_no_completed(void)
184 {
185 return 0;
186 }
187
rcu_sync_scale_init(void)188 static void rcu_sync_scale_init(void)
189 {
190 }
191
192 static struct rcu_scale_ops rcu_ops = {
193 .ptype = RCU_FLAVOR,
194 .init = rcu_sync_scale_init,
195 .readlock = rcu_scale_read_lock,
196 .readunlock = rcu_scale_read_unlock,
197 .get_gp_seq = rcu_get_gp_seq,
198 .gp_diff = rcu_seq_diff,
199 .exp_completed = rcu_exp_batches_completed,
200 .async = call_rcu_hurry,
201 .gp_barrier = rcu_barrier,
202 .sync = synchronize_rcu,
203 .exp_sync = synchronize_rcu_expedited,
204 .name = "rcu"
205 };
206
207 /*
208 * Definitions for srcu scalability testing.
209 */
210
211 DEFINE_STATIC_SRCU(srcu_ctl_scale);
212 static struct srcu_struct *srcu_ctlp = &srcu_ctl_scale;
213
srcu_scale_read_lock(void)214 static int srcu_scale_read_lock(void) __acquires(srcu_ctlp)
215 {
216 return srcu_read_lock(srcu_ctlp);
217 }
218
srcu_scale_read_unlock(int idx)219 static void srcu_scale_read_unlock(int idx) __releases(srcu_ctlp)
220 {
221 srcu_read_unlock(srcu_ctlp, idx);
222 }
223
srcu_scale_completed(void)224 static unsigned long srcu_scale_completed(void)
225 {
226 return srcu_batches_completed(srcu_ctlp);
227 }
228
srcu_call_rcu(struct rcu_head * head,rcu_callback_t func)229 static void srcu_call_rcu(struct rcu_head *head, rcu_callback_t func)
230 {
231 call_srcu(srcu_ctlp, head, func);
232 }
233
srcu_rcu_barrier(void)234 static void srcu_rcu_barrier(void)
235 {
236 srcu_barrier(srcu_ctlp);
237 }
238
srcu_scale_synchronize(void)239 static void srcu_scale_synchronize(void)
240 {
241 synchronize_srcu(srcu_ctlp);
242 }
243
srcu_scale_stats(void)244 static void srcu_scale_stats(void)
245 {
246 srcu_torture_stats_print(srcu_ctlp, scale_type, SCALE_FLAG);
247 }
248
srcu_scale_synchronize_expedited(void)249 static void srcu_scale_synchronize_expedited(void)
250 {
251 synchronize_srcu_expedited(srcu_ctlp);
252 }
253
254 static struct rcu_scale_ops srcu_ops = {
255 .ptype = SRCU_FLAVOR,
256 .init = rcu_sync_scale_init,
257 .readlock = srcu_scale_read_lock,
258 .readunlock = srcu_scale_read_unlock,
259 .get_gp_seq = srcu_scale_completed,
260 .gp_diff = rcu_seq_diff,
261 .exp_completed = srcu_scale_completed,
262 .async = srcu_call_rcu,
263 .gp_barrier = srcu_rcu_barrier,
264 .sync = srcu_scale_synchronize,
265 .exp_sync = srcu_scale_synchronize_expedited,
266 .stats = srcu_scale_stats,
267 .name = "srcu"
268 };
269
270 static struct srcu_struct srcud;
271
srcu_sync_scale_init(void)272 static void srcu_sync_scale_init(void)
273 {
274 srcu_ctlp = &srcud;
275 init_srcu_struct(srcu_ctlp);
276 }
277
srcu_sync_scale_cleanup(void)278 static void srcu_sync_scale_cleanup(void)
279 {
280 cleanup_srcu_struct(srcu_ctlp);
281 }
282
283 static struct rcu_scale_ops srcud_ops = {
284 .ptype = SRCU_FLAVOR,
285 .init = srcu_sync_scale_init,
286 .cleanup = srcu_sync_scale_cleanup,
287 .readlock = srcu_scale_read_lock,
288 .readunlock = srcu_scale_read_unlock,
289 .get_gp_seq = srcu_scale_completed,
290 .gp_diff = rcu_seq_diff,
291 .exp_completed = srcu_scale_completed,
292 .async = srcu_call_rcu,
293 .gp_barrier = srcu_rcu_barrier,
294 .sync = srcu_scale_synchronize,
295 .exp_sync = srcu_scale_synchronize_expedited,
296 .stats = srcu_scale_stats,
297 .name = "srcud"
298 };
299
300 #ifdef CONFIG_TASKS_RCU
301
302 /*
303 * Definitions for RCU-tasks scalability testing.
304 */
305
tasks_scale_read_lock(void)306 static int tasks_scale_read_lock(void)
307 {
308 return 0;
309 }
310
tasks_scale_read_unlock(int idx)311 static void tasks_scale_read_unlock(int idx)
312 {
313 }
314
rcu_tasks_scale_stats(void)315 static void rcu_tasks_scale_stats(void)
316 {
317 rcu_tasks_torture_stats_print(scale_type, SCALE_FLAG);
318 }
319
320 static struct rcu_scale_ops tasks_ops = {
321 .ptype = RCU_TASKS_FLAVOR,
322 .init = rcu_sync_scale_init,
323 .readlock = tasks_scale_read_lock,
324 .readunlock = tasks_scale_read_unlock,
325 .get_gp_seq = rcu_no_completed,
326 .gp_diff = rcu_seq_diff,
327 .async = call_rcu_tasks,
328 .gp_barrier = rcu_barrier_tasks,
329 .sync = synchronize_rcu_tasks,
330 .exp_sync = synchronize_rcu_tasks,
331 .rso_gp_kthread = get_rcu_tasks_gp_kthread,
332 .stats = IS_ENABLED(CONFIG_TINY_RCU) ? NULL : rcu_tasks_scale_stats,
333 .name = "tasks"
334 };
335
336 #define TASKS_OPS &tasks_ops,
337
338 #else // #ifdef CONFIG_TASKS_RCU
339
340 #define TASKS_OPS
341
342 #endif // #else // #ifdef CONFIG_TASKS_RCU
343
344 #ifdef CONFIG_TASKS_RUDE_RCU
345
346 /*
347 * Definitions for RCU-tasks-rude scalability testing.
348 */
349
tasks_rude_scale_read_lock(void)350 static int tasks_rude_scale_read_lock(void)
351 {
352 return 0;
353 }
354
tasks_rude_scale_read_unlock(int idx)355 static void tasks_rude_scale_read_unlock(int idx)
356 {
357 }
358
rcu_tasks_rude_scale_stats(void)359 static void rcu_tasks_rude_scale_stats(void)
360 {
361 rcu_tasks_rude_torture_stats_print(scale_type, SCALE_FLAG);
362 }
363
364 static struct rcu_scale_ops tasks_rude_ops = {
365 .ptype = RCU_TASKS_RUDE_FLAVOR,
366 .init = rcu_sync_scale_init,
367 .readlock = tasks_rude_scale_read_lock,
368 .readunlock = tasks_rude_scale_read_unlock,
369 .get_gp_seq = rcu_no_completed,
370 .gp_diff = rcu_seq_diff,
371 .sync = synchronize_rcu_tasks_rude,
372 .exp_sync = synchronize_rcu_tasks_rude,
373 .rso_gp_kthread = get_rcu_tasks_rude_gp_kthread,
374 .stats = IS_ENABLED(CONFIG_TINY_RCU) ? NULL : rcu_tasks_rude_scale_stats,
375 .name = "tasks-rude"
376 };
377
378 #define TASKS_RUDE_OPS &tasks_rude_ops,
379
380 #else // #ifdef CONFIG_TASKS_RUDE_RCU
381
382 #define TASKS_RUDE_OPS
383
384 #endif // #else // #ifdef CONFIG_TASKS_RUDE_RCU
385
386 #ifdef CONFIG_TASKS_TRACE_RCU
387
388 /*
389 * Definitions for RCU-tasks-trace scalability testing.
390 */
391
tasks_trace_scale_read_lock(void)392 static int tasks_trace_scale_read_lock(void)
393 {
394 rcu_read_lock_trace();
395 return 0;
396 }
397
tasks_trace_scale_read_unlock(int idx)398 static void tasks_trace_scale_read_unlock(int idx)
399 {
400 rcu_read_unlock_trace();
401 }
402
rcu_tasks_trace_scale_stats(void)403 static void rcu_tasks_trace_scale_stats(void)
404 {
405 rcu_tasks_trace_torture_stats_print(scale_type, SCALE_FLAG);
406 }
407
408 static struct rcu_scale_ops tasks_tracing_ops = {
409 .ptype = RCU_TASKS_FLAVOR,
410 .init = rcu_sync_scale_init,
411 .readlock = tasks_trace_scale_read_lock,
412 .readunlock = tasks_trace_scale_read_unlock,
413 .get_gp_seq = rcu_no_completed,
414 .gp_diff = rcu_seq_diff,
415 .async = call_rcu_tasks_trace,
416 .gp_barrier = rcu_barrier_tasks_trace,
417 .sync = synchronize_rcu_tasks_trace,
418 .exp_sync = synchronize_rcu_tasks_trace,
419 .rso_gp_kthread = get_rcu_tasks_trace_gp_kthread,
420 .stats = IS_ENABLED(CONFIG_TINY_RCU) ? NULL : rcu_tasks_trace_scale_stats,
421 .name = "tasks-tracing"
422 };
423
424 #define TASKS_TRACING_OPS &tasks_tracing_ops,
425
426 #else // #ifdef CONFIG_TASKS_TRACE_RCU
427
428 #define TASKS_TRACING_OPS
429
430 #endif // #else // #ifdef CONFIG_TASKS_TRACE_RCU
431
rcuscale_seq_diff(unsigned long new,unsigned long old)432 static unsigned long rcuscale_seq_diff(unsigned long new, unsigned long old)
433 {
434 if (!cur_ops->gp_diff)
435 return new - old;
436 return cur_ops->gp_diff(new, old);
437 }
438
439 /*
440 * If scalability tests complete, wait for shutdown to commence.
441 */
rcu_scale_wait_shutdown(void)442 static void rcu_scale_wait_shutdown(void)
443 {
444 cond_resched_tasks_rcu_qs();
445 if (atomic_read(&n_rcu_scale_writer_finished) < nrealwriters)
446 return;
447 while (!torture_must_stop())
448 schedule_timeout_uninterruptible(1);
449 }
450
451 /*
452 * RCU scalability reader kthread. Repeatedly does empty RCU read-side
453 * critical section, minimizing update-side interference. However, the
454 * point of this test is not to evaluate reader scalability, but instead
455 * to serve as a test load for update-side scalability testing.
456 */
457 static int
rcu_scale_reader(void * arg)458 rcu_scale_reader(void *arg)
459 {
460 unsigned long flags;
461 int idx;
462 long me = (long)arg;
463
464 VERBOSE_SCALEOUT_STRING("rcu_scale_reader task started");
465 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
466 set_user_nice(current, MAX_NICE);
467 atomic_inc(&n_rcu_scale_reader_started);
468
469 do {
470 local_irq_save(flags);
471 idx = cur_ops->readlock();
472 cur_ops->readunlock(idx);
473 local_irq_restore(flags);
474 rcu_scale_wait_shutdown();
475 } while (!torture_must_stop());
476 torture_kthread_stopping("rcu_scale_reader");
477 return 0;
478 }
479
480 /*
481 * Allocate a writer_mblock structure for the specified rcu_scale_writer
482 * task.
483 */
rcu_scale_alloc(long me)484 static struct writer_mblock *rcu_scale_alloc(long me)
485 {
486 struct llist_node *llnp;
487 struct writer_freelist *wflp;
488 struct writer_mblock *wmbp;
489
490 if (WARN_ON_ONCE(!writer_freelists))
491 return NULL;
492 wflp = &writer_freelists[me];
493 if (llist_empty(&wflp->ws_lhp)) {
494 // ->ws_lhp is private to its rcu_scale_writer task.
495 wmbp = container_of(llist_del_all(&wflp->ws_lhg), struct writer_mblock, wmb_node);
496 wflp->ws_lhp.first = &wmbp->wmb_node;
497 }
498 llnp = llist_del_first(&wflp->ws_lhp);
499 if (!llnp)
500 return NULL;
501 return container_of(llnp, struct writer_mblock, wmb_node);
502 }
503
504 /*
505 * Free a writer_mblock structure to its rcu_scale_writer task.
506 */
rcu_scale_free(struct writer_mblock * wmbp)507 static void rcu_scale_free(struct writer_mblock *wmbp)
508 {
509 struct writer_freelist *wflp;
510
511 if (!wmbp)
512 return;
513 wflp = wmbp->wmb_wfl;
514 llist_add(&wmbp->wmb_node, &wflp->ws_lhg);
515 }
516
517 /*
518 * Callback function for asynchronous grace periods from rcu_scale_writer().
519 */
rcu_scale_async_cb(struct rcu_head * rhp)520 static void rcu_scale_async_cb(struct rcu_head *rhp)
521 {
522 struct writer_mblock *wmbp = container_of(rhp, struct writer_mblock, wmb_rh);
523 struct writer_freelist *wflp = wmbp->wmb_wfl;
524
525 atomic_dec(&wflp->ws_inflight);
526 rcu_scale_free(wmbp);
527 }
528
529 /*
530 * RCU scale writer kthread. Repeatedly does a grace period.
531 */
532 static int
rcu_scale_writer(void * arg)533 rcu_scale_writer(void *arg)
534 {
535 int i = 0;
536 int i_max;
537 unsigned long jdone;
538 long me = (long)arg;
539 bool selfreport = false;
540 bool started = false, done = false, alldone = false;
541 u64 t;
542 DEFINE_TORTURE_RANDOM(tr);
543 u64 *wdp;
544 u64 *wdpp = writer_durations[me];
545 struct writer_freelist *wflp = &writer_freelists[me];
546 struct writer_mblock *wmbp = NULL;
547
548 VERBOSE_SCALEOUT_STRING("rcu_scale_writer task started");
549 WARN_ON(!wdpp);
550 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
551 current->flags |= PF_NO_SETAFFINITY;
552 sched_set_fifo_low(current);
553
554 if (holdoff)
555 schedule_timeout_idle(holdoff * HZ);
556
557 /*
558 * Wait until rcu_end_inkernel_boot() is called for normal GP tests
559 * so that RCU is not always expedited for normal GP tests.
560 * The system_state test is approximate, but works well in practice.
561 */
562 while (!gp_exp && system_state != SYSTEM_RUNNING)
563 schedule_timeout_uninterruptible(1);
564
565 t = ktime_get_mono_fast_ns();
566 if (atomic_inc_return(&n_rcu_scale_writer_started) >= nrealwriters) {
567 t_rcu_scale_writer_started = t;
568 if (gp_exp) {
569 b_rcu_gp_test_started =
570 cur_ops->exp_completed() / 2;
571 } else {
572 b_rcu_gp_test_started = cur_ops->get_gp_seq();
573 }
574 }
575
576 jdone = jiffies + minruntime * HZ;
577 do {
578 bool gp_succeeded = false;
579
580 if (writer_holdoff)
581 udelay(writer_holdoff);
582 if (writer_holdoff_jiffies)
583 schedule_timeout_idle(torture_random(&tr) % writer_holdoff_jiffies + 1);
584 wdp = &wdpp[i];
585 *wdp = ktime_get_mono_fast_ns();
586 if (gp_async && !WARN_ON_ONCE(!cur_ops->async)) {
587 if (!wmbp)
588 wmbp = rcu_scale_alloc(me);
589 if (wmbp && atomic_read(&wflp->ws_inflight) < gp_async_max) {
590 atomic_inc(&wflp->ws_inflight);
591 cur_ops->async(&wmbp->wmb_rh, rcu_scale_async_cb);
592 wmbp = NULL;
593 gp_succeeded = true;
594 } else if (!kthread_should_stop()) {
595 cur_ops->gp_barrier();
596 } else {
597 rcu_scale_free(wmbp); /* Because we are stopping. */
598 wmbp = NULL;
599 }
600 } else if (gp_exp) {
601 cur_ops->exp_sync();
602 gp_succeeded = true;
603 } else {
604 cur_ops->sync();
605 gp_succeeded = true;
606 }
607 t = ktime_get_mono_fast_ns();
608 *wdp = t - *wdp;
609 i_max = i;
610 if (!started &&
611 atomic_read(&n_rcu_scale_writer_started) >= nrealwriters)
612 started = true;
613 if (!done && i >= MIN_MEAS && time_after(jiffies, jdone)) {
614 done = true;
615 WRITE_ONCE(writer_done[me], true);
616 sched_set_normal(current, 0);
617 pr_alert("%s%s rcu_scale_writer %ld has %d measurements\n",
618 scale_type, SCALE_FLAG, me, MIN_MEAS);
619 if (atomic_inc_return(&n_rcu_scale_writer_finished) >=
620 nrealwriters) {
621 schedule_timeout_interruptible(10);
622 rcu_ftrace_dump(DUMP_ALL);
623 SCALEOUT_STRING("Test complete");
624 t_rcu_scale_writer_finished = t;
625 if (gp_exp) {
626 b_rcu_gp_test_finished =
627 cur_ops->exp_completed() / 2;
628 } else {
629 b_rcu_gp_test_finished =
630 cur_ops->get_gp_seq();
631 }
632 if (shutdown) {
633 smp_mb(); /* Assign before wake. */
634 wake_up(&shutdown_wq);
635 }
636 }
637 }
638 if (done && !alldone &&
639 atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters)
640 alldone = true;
641 if (done && !alldone && time_after(jiffies, jdone + HZ * 60)) {
642 static atomic_t dumped;
643 int i;
644
645 if (!atomic_xchg(&dumped, 1)) {
646 for (i = 0; i < nrealwriters; i++) {
647 if (writer_done[i])
648 continue;
649 pr_info("%s: Task %ld flags writer %d:\n", __func__, me, i);
650 sched_show_task(writer_tasks[i]);
651 }
652 if (cur_ops->stats)
653 cur_ops->stats();
654 }
655 }
656 if (!selfreport && time_after(jiffies, jdone + HZ * (70 + me))) {
657 pr_info("%s: Writer %ld self-report: started %d done %d/%d->%d i %d jdone %lu.\n",
658 __func__, me, started, done, writer_done[me], atomic_read(&n_rcu_scale_writer_finished), i, jiffies - jdone);
659 selfreport = true;
660 }
661 if (gp_succeeded && started && !alldone && i < MAX_MEAS - 1)
662 i++;
663 rcu_scale_wait_shutdown();
664 } while (!torture_must_stop());
665 if (gp_async && cur_ops->async) {
666 rcu_scale_free(wmbp);
667 cur_ops->gp_barrier();
668 }
669 writer_n_durations[me] = i_max + 1;
670 torture_kthread_stopping("rcu_scale_writer");
671 return 0;
672 }
673
674 static void
rcu_scale_print_module_parms(struct rcu_scale_ops * cur_ops,const char * tag)675 rcu_scale_print_module_parms(struct rcu_scale_ops *cur_ops, const char *tag)
676 {
677 pr_alert("%s" SCALE_FLAG
678 "--- %s: gp_async=%d gp_async_max=%d gp_exp=%d holdoff=%d minruntime=%d nreaders=%d nwriters=%d writer_holdoff=%d writer_holdoff_jiffies=%d verbose=%d shutdown=%d\n",
679 scale_type, tag, gp_async, gp_async_max, gp_exp, holdoff, minruntime, nrealreaders, nrealwriters, writer_holdoff, writer_holdoff_jiffies, verbose, shutdown);
680 }
681
682 /*
683 * Return the number if non-negative. If -1, the number of CPUs.
684 * If less than -1, that much less than the number of CPUs, but
685 * at least one.
686 */
compute_real(int n)687 static int compute_real(int n)
688 {
689 int nr;
690
691 if (n >= 0) {
692 nr = n;
693 } else {
694 nr = num_online_cpus() + 1 + n;
695 if (nr <= 0)
696 nr = 1;
697 }
698 return nr;
699 }
700
701 /*
702 * kfree_rcu() scalability tests: Start a kfree_rcu() loop on all CPUs for number
703 * of iterations and measure total time and number of GP for all iterations to complete.
704 */
705
706 torture_param(int, kfree_nthreads, -1, "Number of threads running loops of kfree_rcu().");
707 torture_param(int, kfree_alloc_num, 8000, "Number of allocations and frees done in an iteration.");
708 torture_param(int, kfree_loops, 10, "Number of loops doing kfree_alloc_num allocations and frees.");
709 torture_param(bool, kfree_rcu_test_double, false, "Do we run a kfree_rcu() double-argument scale test?");
710 torture_param(bool, kfree_rcu_test_single, false, "Do we run a kfree_rcu() single-argument scale test?");
711
712 static struct task_struct **kfree_reader_tasks;
713 static int kfree_nrealthreads;
714 static atomic_t n_kfree_scale_thread_started;
715 static atomic_t n_kfree_scale_thread_ended;
716 static struct task_struct *kthread_tp;
717 static u64 kthread_stime;
718
719 struct kfree_obj {
720 char kfree_obj[8];
721 struct rcu_head rh;
722 };
723
724 /* Used if doing RCU-kfree'ing via call_rcu(). */
kfree_call_rcu(struct rcu_head * rh)725 static void kfree_call_rcu(struct rcu_head *rh)
726 {
727 struct kfree_obj *obj = container_of(rh, struct kfree_obj, rh);
728
729 kfree(obj);
730 }
731
732 static int
kfree_scale_thread(void * arg)733 kfree_scale_thread(void *arg)
734 {
735 int i, loop = 0;
736 long me = (long)arg;
737 struct kfree_obj *alloc_ptr;
738 u64 start_time, end_time;
739 long long mem_begin, mem_during = 0;
740 bool kfree_rcu_test_both;
741 DEFINE_TORTURE_RANDOM(tr);
742
743 VERBOSE_SCALEOUT_STRING("kfree_scale_thread task started");
744 set_cpus_allowed_ptr(current, cpumask_of(me % nr_cpu_ids));
745 set_user_nice(current, MAX_NICE);
746 kfree_rcu_test_both = (kfree_rcu_test_single == kfree_rcu_test_double);
747
748 start_time = ktime_get_mono_fast_ns();
749
750 if (atomic_inc_return(&n_kfree_scale_thread_started) >= kfree_nrealthreads) {
751 if (gp_exp)
752 b_rcu_gp_test_started = cur_ops->exp_completed() / 2;
753 else
754 b_rcu_gp_test_started = cur_ops->get_gp_seq();
755 }
756
757 do {
758 if (!mem_during) {
759 mem_during = mem_begin = si_mem_available();
760 } else if (loop % (kfree_loops / 4) == 0) {
761 mem_during = (mem_during + si_mem_available()) / 2;
762 }
763
764 for (i = 0; i < kfree_alloc_num; i++) {
765 alloc_ptr = kmalloc(kfree_mult * sizeof(struct kfree_obj), GFP_KERNEL);
766 if (!alloc_ptr)
767 return -ENOMEM;
768
769 if (kfree_by_call_rcu) {
770 call_rcu(&(alloc_ptr->rh), kfree_call_rcu);
771 continue;
772 }
773
774 // By default kfree_rcu_test_single and kfree_rcu_test_double are
775 // initialized to false. If both have the same value (false or true)
776 // both are randomly tested, otherwise only the one with value true
777 // is tested.
778 if ((kfree_rcu_test_single && !kfree_rcu_test_double) ||
779 (kfree_rcu_test_both && torture_random(&tr) & 0x800))
780 kfree_rcu_mightsleep(alloc_ptr);
781 else
782 kfree_rcu(alloc_ptr, rh);
783 }
784
785 cond_resched();
786 } while (!torture_must_stop() && ++loop < kfree_loops);
787
788 if (atomic_inc_return(&n_kfree_scale_thread_ended) >= kfree_nrealthreads) {
789 end_time = ktime_get_mono_fast_ns();
790
791 if (gp_exp)
792 b_rcu_gp_test_finished = cur_ops->exp_completed() / 2;
793 else
794 b_rcu_gp_test_finished = cur_ops->get_gp_seq();
795
796 pr_alert("Total time taken by all kfree'ers: %llu ns, loops: %d, batches: %ld, memory footprint: %lldMB\n",
797 (unsigned long long)(end_time - start_time), kfree_loops,
798 rcuscale_seq_diff(b_rcu_gp_test_finished, b_rcu_gp_test_started),
799 (mem_begin - mem_during) >> (20 - PAGE_SHIFT));
800
801 if (shutdown) {
802 smp_mb(); /* Assign before wake. */
803 wake_up(&shutdown_wq);
804 }
805 }
806
807 torture_kthread_stopping("kfree_scale_thread");
808 return 0;
809 }
810
811 static void
kfree_scale_cleanup(void)812 kfree_scale_cleanup(void)
813 {
814 int i;
815
816 if (torture_cleanup_begin())
817 return;
818
819 if (kfree_reader_tasks) {
820 for (i = 0; i < kfree_nrealthreads; i++)
821 torture_stop_kthread(kfree_scale_thread,
822 kfree_reader_tasks[i]);
823 kfree(kfree_reader_tasks);
824 kfree_reader_tasks = NULL;
825 }
826
827 torture_cleanup_end();
828 }
829
830 /*
831 * shutdown kthread. Just waits to be awakened, then shuts down system.
832 */
833 static int
kfree_scale_shutdown(void * arg)834 kfree_scale_shutdown(void *arg)
835 {
836 wait_event_idle(shutdown_wq,
837 atomic_read(&n_kfree_scale_thread_ended) >= kfree_nrealthreads);
838
839 smp_mb(); /* Wake before output. */
840
841 kfree_scale_cleanup();
842 kernel_power_off();
843 return -EINVAL;
844 }
845
846 // Used if doing RCU-kfree'ing via call_rcu().
847 static unsigned long jiffies_at_lazy_cb;
848 static struct rcu_head lazy_test1_rh;
849 static int rcu_lazy_test1_cb_called;
call_rcu_lazy_test1(struct rcu_head * rh)850 static void call_rcu_lazy_test1(struct rcu_head *rh)
851 {
852 jiffies_at_lazy_cb = jiffies;
853 WRITE_ONCE(rcu_lazy_test1_cb_called, 1);
854 }
855
856 static int __init
kfree_scale_init(void)857 kfree_scale_init(void)
858 {
859 int firsterr = 0;
860 long i;
861 unsigned long jif_start;
862 unsigned long orig_jif;
863
864 pr_alert("%s" SCALE_FLAG
865 "--- kfree_rcu_test: kfree_mult=%d kfree_by_call_rcu=%d kfree_nthreads=%d kfree_alloc_num=%d kfree_loops=%d kfree_rcu_test_double=%d kfree_rcu_test_single=%d\n",
866 scale_type, kfree_mult, kfree_by_call_rcu, kfree_nthreads, kfree_alloc_num, kfree_loops, kfree_rcu_test_double, kfree_rcu_test_single);
867
868 // Also, do a quick self-test to ensure laziness is as much as
869 // expected.
870 if (kfree_by_call_rcu && !IS_ENABLED(CONFIG_RCU_LAZY)) {
871 pr_alert("CONFIG_RCU_LAZY is disabled, falling back to kfree_rcu() for delayed RCU kfree'ing\n");
872 kfree_by_call_rcu = 0;
873 }
874
875 if (kfree_by_call_rcu) {
876 /* do a test to check the timeout. */
877 orig_jif = rcu_get_jiffies_lazy_flush();
878
879 rcu_set_jiffies_lazy_flush(2 * HZ);
880 rcu_barrier();
881
882 jif_start = jiffies;
883 jiffies_at_lazy_cb = 0;
884 call_rcu(&lazy_test1_rh, call_rcu_lazy_test1);
885
886 smp_cond_load_relaxed(&rcu_lazy_test1_cb_called, VAL == 1);
887
888 rcu_set_jiffies_lazy_flush(orig_jif);
889
890 if (WARN_ON_ONCE(jiffies_at_lazy_cb - jif_start < 2 * HZ)) {
891 pr_alert("ERROR: call_rcu() CBs are not being lazy as expected!\n");
892 firsterr = -1;
893 goto unwind;
894 }
895
896 if (WARN_ON_ONCE(jiffies_at_lazy_cb - jif_start > 3 * HZ)) {
897 pr_alert("ERROR: call_rcu() CBs are being too lazy!\n");
898 firsterr = -1;
899 goto unwind;
900 }
901 }
902
903 kfree_nrealthreads = compute_real(kfree_nthreads);
904 /* Start up the kthreads. */
905 if (shutdown) {
906 init_waitqueue_head(&shutdown_wq);
907 firsterr = torture_create_kthread(kfree_scale_shutdown, NULL,
908 shutdown_task);
909 if (torture_init_error(firsterr))
910 goto unwind;
911 schedule_timeout_uninterruptible(1);
912 }
913
914 pr_alert("kfree object size=%zu, kfree_by_call_rcu=%d\n",
915 kfree_mult * sizeof(struct kfree_obj),
916 kfree_by_call_rcu);
917
918 kfree_reader_tasks = kcalloc(kfree_nrealthreads, sizeof(kfree_reader_tasks[0]),
919 GFP_KERNEL);
920 if (kfree_reader_tasks == NULL) {
921 firsterr = -ENOMEM;
922 goto unwind;
923 }
924
925 for (i = 0; i < kfree_nrealthreads; i++) {
926 firsterr = torture_create_kthread(kfree_scale_thread, (void *)i,
927 kfree_reader_tasks[i]);
928 if (torture_init_error(firsterr))
929 goto unwind;
930 }
931
932 while (atomic_read(&n_kfree_scale_thread_started) < kfree_nrealthreads)
933 schedule_timeout_uninterruptible(1);
934
935 torture_init_end();
936 return 0;
937
938 unwind:
939 torture_init_end();
940 kfree_scale_cleanup();
941 return firsterr;
942 }
943
944 static void
rcu_scale_cleanup(void)945 rcu_scale_cleanup(void)
946 {
947 int i;
948 int j;
949 int ngps = 0;
950 u64 *wdp;
951 u64 *wdpp;
952
953 /*
954 * Would like warning at start, but everything is expedited
955 * during the mid-boot phase, so have to wait till the end.
956 */
957 if (rcu_gp_is_expedited() && !rcu_gp_is_normal() && !gp_exp)
958 SCALEOUT_ERRSTRING("All grace periods expedited, no normal ones to measure!");
959 if (rcu_gp_is_normal() && gp_exp)
960 SCALEOUT_ERRSTRING("All grace periods normal, no expedited ones to measure!");
961 if (gp_exp && gp_async)
962 SCALEOUT_ERRSTRING("No expedited async GPs, so went with async!");
963
964 // If built-in, just report all of the GP kthread's CPU time.
965 if (IS_BUILTIN(CONFIG_RCU_SCALE_TEST) && !kthread_tp && cur_ops->rso_gp_kthread)
966 kthread_tp = cur_ops->rso_gp_kthread();
967 if (kthread_tp) {
968 u32 ns;
969 u64 us;
970
971 kthread_stime = kthread_tp->stime - kthread_stime;
972 us = div_u64_rem(kthread_stime, 1000, &ns);
973 pr_info("rcu_scale: Grace-period kthread CPU time: %llu.%03u us\n", us, ns);
974 show_rcu_gp_kthreads();
975 }
976 if (kfree_rcu_test) {
977 kfree_scale_cleanup();
978 return;
979 }
980
981 if (torture_cleanup_begin())
982 return;
983 if (!cur_ops) {
984 torture_cleanup_end();
985 return;
986 }
987
988 if (reader_tasks) {
989 for (i = 0; i < nrealreaders; i++)
990 torture_stop_kthread(rcu_scale_reader,
991 reader_tasks[i]);
992 kfree(reader_tasks);
993 reader_tasks = NULL;
994 }
995
996 if (writer_tasks) {
997 for (i = 0; i < nrealwriters; i++) {
998 torture_stop_kthread(rcu_scale_writer,
999 writer_tasks[i]);
1000 if (!writer_n_durations)
1001 continue;
1002 j = writer_n_durations[i];
1003 pr_alert("%s%s writer %d gps: %d\n",
1004 scale_type, SCALE_FLAG, i, j);
1005 ngps += j;
1006 }
1007 pr_alert("%s%s start: %llu end: %llu duration: %llu gps: %d batches: %ld\n",
1008 scale_type, SCALE_FLAG,
1009 t_rcu_scale_writer_started, t_rcu_scale_writer_finished,
1010 t_rcu_scale_writer_finished -
1011 t_rcu_scale_writer_started,
1012 ngps,
1013 rcuscale_seq_diff(b_rcu_gp_test_finished,
1014 b_rcu_gp_test_started));
1015 for (i = 0; i < nrealwriters; i++) {
1016 if (!writer_durations)
1017 break;
1018 if (!writer_n_durations)
1019 continue;
1020 wdpp = writer_durations[i];
1021 if (!wdpp)
1022 continue;
1023 for (j = 0; j < writer_n_durations[i]; j++) {
1024 wdp = &wdpp[j];
1025 pr_alert("%s%s %4d writer-duration: %5d %llu\n",
1026 scale_type, SCALE_FLAG,
1027 i, j, *wdp);
1028 if (j % 100 == 0)
1029 schedule_timeout_uninterruptible(1);
1030 }
1031 kfree(writer_durations[i]);
1032 if (writer_freelists) {
1033 int ctr = 0;
1034 struct llist_node *llnp;
1035 struct writer_freelist *wflp = &writer_freelists[i];
1036
1037 if (wflp->ws_mblocks) {
1038 llist_for_each(llnp, wflp->ws_lhg.first)
1039 ctr++;
1040 llist_for_each(llnp, wflp->ws_lhp.first)
1041 ctr++;
1042 WARN_ONCE(ctr != gp_async_max,
1043 "%s: ctr = %d gp_async_max = %d\n",
1044 __func__, ctr, gp_async_max);
1045 kfree(wflp->ws_mblocks);
1046 }
1047 }
1048 }
1049 kfree(writer_tasks);
1050 writer_tasks = NULL;
1051 kfree(writer_durations);
1052 writer_durations = NULL;
1053 kfree(writer_n_durations);
1054 writer_n_durations = NULL;
1055 kfree(writer_done);
1056 writer_done = NULL;
1057 kfree(writer_freelists);
1058 writer_freelists = NULL;
1059 }
1060
1061 /* Do torture-type-specific cleanup operations. */
1062 if (cur_ops->cleanup != NULL)
1063 cur_ops->cleanup();
1064
1065 torture_cleanup_end();
1066 }
1067
1068 /*
1069 * RCU scalability shutdown kthread. Just waits to be awakened, then shuts
1070 * down system.
1071 */
1072 static int
rcu_scale_shutdown(void * arg)1073 rcu_scale_shutdown(void *arg)
1074 {
1075 wait_event_idle(shutdown_wq, atomic_read(&n_rcu_scale_writer_finished) >= nrealwriters);
1076 smp_mb(); /* Wake before output. */
1077 rcu_scale_cleanup();
1078 kernel_power_off();
1079 return -EINVAL;
1080 }
1081
1082 static int __init
rcu_scale_init(void)1083 rcu_scale_init(void)
1084 {
1085 int firsterr = 0;
1086 long i;
1087 long j;
1088 static struct rcu_scale_ops *scale_ops[] = {
1089 &rcu_ops, &srcu_ops, &srcud_ops, TASKS_OPS TASKS_RUDE_OPS TASKS_TRACING_OPS
1090 };
1091
1092 if (!torture_init_begin(scale_type, verbose))
1093 return -EBUSY;
1094
1095 /* Process args and announce that the scalability'er is on the job. */
1096 for (i = 0; i < ARRAY_SIZE(scale_ops); i++) {
1097 cur_ops = scale_ops[i];
1098 if (strcmp(scale_type, cur_ops->name) == 0)
1099 break;
1100 }
1101 if (i == ARRAY_SIZE(scale_ops)) {
1102 pr_alert("rcu-scale: invalid scale type: \"%s\"\n", scale_type);
1103 pr_alert("rcu-scale types:");
1104 for (i = 0; i < ARRAY_SIZE(scale_ops); i++)
1105 pr_cont(" %s", scale_ops[i]->name);
1106 pr_cont("\n");
1107 firsterr = -EINVAL;
1108 cur_ops = NULL;
1109 goto unwind;
1110 }
1111 if (cur_ops->init)
1112 cur_ops->init();
1113
1114 if (cur_ops->rso_gp_kthread) {
1115 kthread_tp = cur_ops->rso_gp_kthread();
1116 if (kthread_tp)
1117 kthread_stime = kthread_tp->stime;
1118 }
1119 if (kfree_rcu_test)
1120 return kfree_scale_init();
1121
1122 nrealwriters = compute_real(nwriters);
1123 nrealreaders = compute_real(nreaders);
1124 atomic_set(&n_rcu_scale_reader_started, 0);
1125 atomic_set(&n_rcu_scale_writer_started, 0);
1126 atomic_set(&n_rcu_scale_writer_finished, 0);
1127 rcu_scale_print_module_parms(cur_ops, "Start of test");
1128
1129 /* Start up the kthreads. */
1130
1131 if (shutdown) {
1132 init_waitqueue_head(&shutdown_wq);
1133 firsterr = torture_create_kthread(rcu_scale_shutdown, NULL,
1134 shutdown_task);
1135 if (torture_init_error(firsterr))
1136 goto unwind;
1137 schedule_timeout_uninterruptible(1);
1138 }
1139 reader_tasks = kcalloc(nrealreaders, sizeof(reader_tasks[0]),
1140 GFP_KERNEL);
1141 if (reader_tasks == NULL) {
1142 SCALEOUT_ERRSTRING("out of memory");
1143 firsterr = -ENOMEM;
1144 goto unwind;
1145 }
1146 for (i = 0; i < nrealreaders; i++) {
1147 firsterr = torture_create_kthread(rcu_scale_reader, (void *)i,
1148 reader_tasks[i]);
1149 if (torture_init_error(firsterr))
1150 goto unwind;
1151 }
1152 while (atomic_read(&n_rcu_scale_reader_started) < nrealreaders)
1153 schedule_timeout_uninterruptible(1);
1154 writer_tasks = kcalloc(nrealwriters, sizeof(writer_tasks[0]), GFP_KERNEL);
1155 writer_durations = kcalloc(nrealwriters, sizeof(*writer_durations), GFP_KERNEL);
1156 writer_n_durations = kcalloc(nrealwriters, sizeof(*writer_n_durations), GFP_KERNEL);
1157 writer_done = kcalloc(nrealwriters, sizeof(writer_done[0]), GFP_KERNEL);
1158 if (gp_async) {
1159 if (gp_async_max <= 0) {
1160 pr_warn("%s: gp_async_max = %d must be greater than zero.\n",
1161 __func__, gp_async_max);
1162 WARN_ON_ONCE(IS_BUILTIN(CONFIG_RCU_TORTURE_TEST));
1163 firsterr = -EINVAL;
1164 goto unwind;
1165 }
1166 writer_freelists = kcalloc(nrealwriters, sizeof(writer_freelists[0]), GFP_KERNEL);
1167 }
1168 if (!writer_tasks || !writer_durations || !writer_n_durations || !writer_done ||
1169 (gp_async && !writer_freelists)) {
1170 SCALEOUT_ERRSTRING("out of memory");
1171 firsterr = -ENOMEM;
1172 goto unwind;
1173 }
1174 for (i = 0; i < nrealwriters; i++) {
1175 writer_durations[i] =
1176 kcalloc(MAX_MEAS, sizeof(*writer_durations[i]),
1177 GFP_KERNEL);
1178 if (!writer_durations[i]) {
1179 firsterr = -ENOMEM;
1180 goto unwind;
1181 }
1182 if (writer_freelists) {
1183 struct writer_freelist *wflp = &writer_freelists[i];
1184
1185 init_llist_head(&wflp->ws_lhg);
1186 init_llist_head(&wflp->ws_lhp);
1187 wflp->ws_mblocks = kcalloc(gp_async_max, sizeof(wflp->ws_mblocks[0]),
1188 GFP_KERNEL);
1189 if (!wflp->ws_mblocks) {
1190 firsterr = -ENOMEM;
1191 goto unwind;
1192 }
1193 for (j = 0; j < gp_async_max; j++) {
1194 struct writer_mblock *wmbp = &wflp->ws_mblocks[j];
1195
1196 wmbp->wmb_wfl = wflp;
1197 llist_add(&wmbp->wmb_node, &wflp->ws_lhp);
1198 }
1199 }
1200 firsterr = torture_create_kthread(rcu_scale_writer, (void *)i,
1201 writer_tasks[i]);
1202 if (torture_init_error(firsterr))
1203 goto unwind;
1204 }
1205 torture_init_end();
1206 return 0;
1207
1208 unwind:
1209 torture_init_end();
1210 rcu_scale_cleanup();
1211 if (shutdown) {
1212 WARN_ON(!IS_MODULE(CONFIG_RCU_SCALE_TEST));
1213 kernel_power_off();
1214 }
1215 return firsterr;
1216 }
1217
1218 module_init(rcu_scale_init);
1219 module_exit(rcu_scale_cleanup);
1220