GNU Linux-libre 5.10.215-gnu1
[releases.git] / kernel / scftorture.c
1 // SPDX-License-Identifier: GPL-2.0+
2 //
3 // Torture test for smp_call_function() and friends.
4 //
5 // Copyright (C) Facebook, 2020.
6 //
7 // Author: Paul E. McKenney <paulmck@kernel.org>
8
9 #define pr_fmt(fmt) fmt
10
11 #include <linux/atomic.h>
12 #include <linux/bitops.h>
13 #include <linux/completion.h>
14 #include <linux/cpu.h>
15 #include <linux/delay.h>
16 #include <linux/err.h>
17 #include <linux/init.h>
18 #include <linux/interrupt.h>
19 #include <linux/kthread.h>
20 #include <linux/kernel.h>
21 #include <linux/mm.h>
22 #include <linux/module.h>
23 #include <linux/moduleparam.h>
24 #include <linux/notifier.h>
25 #include <linux/percpu.h>
26 #include <linux/rcupdate.h>
27 #include <linux/rcupdate_trace.h>
28 #include <linux/reboot.h>
29 #include <linux/sched.h>
30 #include <linux/spinlock.h>
31 #include <linux/smp.h>
32 #include <linux/stat.h>
33 #include <linux/srcu.h>
34 #include <linux/slab.h>
35 #include <linux/torture.h>
36 #include <linux/types.h>
37
38 #define SCFTORT_STRING "scftorture"
39 #define SCFTORT_FLAG SCFTORT_STRING ": "
40
41 #define SCFTORTOUT(s, x...) \
42         pr_alert(SCFTORT_FLAG s, ## x)
43
44 #define VERBOSE_SCFTORTOUT(s, x...) \
45         do { if (verbose) pr_alert(SCFTORT_FLAG s, ## x); } while (0)
46
47 #define VERBOSE_SCFTORTOUT_ERRSTRING(s, x...) \
48         do { if (verbose) pr_alert(SCFTORT_FLAG "!!! " s, ## x); } while (0)
49
50 MODULE_LICENSE("GPL");
51 MODULE_AUTHOR("Paul E. McKenney <paulmck@kernel.org>");
52
53 // Wait until there are multiple CPUs before starting test.
54 torture_param(int, holdoff, IS_BUILTIN(CONFIG_SCF_TORTURE_TEST) ? 10 : 0,
55               "Holdoff time before test start (s)");
56 torture_param(int, longwait, 0, "Include ridiculously long waits? (seconds)");
57 torture_param(int, nthreads, -1, "# threads, defaults to -1 for all CPUs.");
58 torture_param(int, onoff_holdoff, 0, "Time after boot before CPU hotplugs (s)");
59 torture_param(int, onoff_interval, 0, "Time between CPU hotplugs (s), 0=disable");
60 torture_param(int, shutdown_secs, 0, "Shutdown time (ms), <= zero to disable.");
61 torture_param(int, stat_interval, 60, "Number of seconds between stats printk()s.");
62 torture_param(int, stutter_cpus, 5, "Number of jiffies to change CPUs under test, 0=disable");
63 torture_param(bool, use_cpus_read_lock, 0, "Use cpus_read_lock() to exclude CPU hotplug.");
64 torture_param(int, verbose, 0, "Enable verbose debugging printk()s");
65 torture_param(int, weight_single, -1, "Testing weight for single-CPU no-wait operations.");
66 torture_param(int, weight_single_wait, -1, "Testing weight for single-CPU operations.");
67 torture_param(int, weight_many, -1, "Testing weight for multi-CPU no-wait operations.");
68 torture_param(int, weight_many_wait, -1, "Testing weight for multi-CPU operations.");
69 torture_param(int, weight_all, -1, "Testing weight for all-CPU no-wait operations.");
70 torture_param(int, weight_all_wait, -1, "Testing weight for all-CPU operations.");
71
72 char *torture_type = "";
73
74 #ifdef MODULE
75 # define SCFTORT_SHUTDOWN 0
76 #else
77 # define SCFTORT_SHUTDOWN 1
78 #endif
79
80 torture_param(bool, shutdown, SCFTORT_SHUTDOWN, "Shutdown at end of torture test.");
81
82 struct scf_statistics {
83         struct task_struct *task;
84         int cpu;
85         long long n_single;
86         long long n_single_ofl;
87         long long n_single_wait;
88         long long n_single_wait_ofl;
89         long long n_many;
90         long long n_many_wait;
91         long long n_all;
92         long long n_all_wait;
93 };
94
95 static struct scf_statistics *scf_stats_p;
96 static struct task_struct *scf_torture_stats_task;
97 static DEFINE_PER_CPU(long long, scf_invoked_count);
98
99 // Data for random primitive selection
100 #define SCF_PRIM_SINGLE         0
101 #define SCF_PRIM_MANY           1
102 #define SCF_PRIM_ALL            2
103 #define SCF_NPRIMS              (2 * 3) // Need wait and no-wait versions of each.
104
105 static char *scf_prim_name[] = {
106         "smp_call_function_single",
107         "smp_call_function_many",
108         "smp_call_function",
109 };
110
111 struct scf_selector {
112         unsigned long scfs_weight;
113         int scfs_prim;
114         bool scfs_wait;
115 };
116 static struct scf_selector scf_sel_array[SCF_NPRIMS];
117 static int scf_sel_array_len;
118 static unsigned long scf_sel_totweight;
119
120 // Communicate between caller and handler.
121 struct scf_check {
122         bool scfc_in;
123         bool scfc_out;
124         int scfc_cpu; // -1 for not _single().
125         bool scfc_wait;
126 };
127
128 // Use to wait for all threads to start.
129 static atomic_t n_started;
130 static atomic_t n_errs;
131 static atomic_t n_mb_in_errs;
132 static atomic_t n_mb_out_errs;
133 static atomic_t n_alloc_errs;
134 static bool scfdone;
135 static char *bangstr = "";
136
137 static DEFINE_TORTURE_RANDOM_PERCPU(scf_torture_rand);
138
139 // Print torture statistics.  Caller must ensure serialization.
140 static void scf_torture_stats_print(void)
141 {
142         int cpu;
143         int i;
144         long long invoked_count = 0;
145         bool isdone = READ_ONCE(scfdone);
146         struct scf_statistics scfs = {};
147
148         for_each_possible_cpu(cpu)
149                 invoked_count += data_race(per_cpu(scf_invoked_count, cpu));
150         for (i = 0; i < nthreads; i++) {
151                 scfs.n_single += scf_stats_p[i].n_single;
152                 scfs.n_single_ofl += scf_stats_p[i].n_single_ofl;
153                 scfs.n_single_wait += scf_stats_p[i].n_single_wait;
154                 scfs.n_single_wait_ofl += scf_stats_p[i].n_single_wait_ofl;
155                 scfs.n_many += scf_stats_p[i].n_many;
156                 scfs.n_many_wait += scf_stats_p[i].n_many_wait;
157                 scfs.n_all += scf_stats_p[i].n_all;
158                 scfs.n_all_wait += scf_stats_p[i].n_all_wait;
159         }
160         if (atomic_read(&n_errs) || atomic_read(&n_mb_in_errs) ||
161             atomic_read(&n_mb_out_errs) ||
162             (!IS_ENABLED(CONFIG_KASAN) && atomic_read(&n_alloc_errs)))
163                 bangstr = "!!! ";
164         pr_alert("%s %sscf_invoked_count %s: %lld single: %lld/%lld single_ofl: %lld/%lld many: %lld/%lld all: %lld/%lld ",
165                  SCFTORT_FLAG, bangstr, isdone ? "VER" : "ver", invoked_count,
166                  scfs.n_single, scfs.n_single_wait, scfs.n_single_ofl, scfs.n_single_wait_ofl,
167                  scfs.n_many, scfs.n_many_wait, scfs.n_all, scfs.n_all_wait);
168         torture_onoff_stats();
169         pr_cont("ste: %d stnmie: %d stnmoe: %d staf: %d\n", atomic_read(&n_errs),
170                 atomic_read(&n_mb_in_errs), atomic_read(&n_mb_out_errs),
171                 atomic_read(&n_alloc_errs));
172 }
173
174 // Periodically prints torture statistics, if periodic statistics printing
175 // was specified via the stat_interval module parameter.
176 static int
177 scf_torture_stats(void *arg)
178 {
179         VERBOSE_TOROUT_STRING("scf_torture_stats task started");
180         do {
181                 schedule_timeout_interruptible(stat_interval * HZ);
182                 scf_torture_stats_print();
183                 torture_shutdown_absorb("scf_torture_stats");
184         } while (!torture_must_stop());
185         torture_kthread_stopping("scf_torture_stats");
186         return 0;
187 }
188
189 // Add a primitive to the scf_sel_array[].
190 static void scf_sel_add(unsigned long weight, int prim, bool wait)
191 {
192         struct scf_selector *scfsp = &scf_sel_array[scf_sel_array_len];
193
194         // If no weight, if array would overflow, if computing three-place
195         // percentages would overflow, or if the scf_prim_name[] array would
196         // overflow, don't bother.  In the last three two cases, complain.
197         if (!weight ||
198             WARN_ON_ONCE(scf_sel_array_len >= ARRAY_SIZE(scf_sel_array)) ||
199             WARN_ON_ONCE(0 - 100000 * weight <= 100000 * scf_sel_totweight) ||
200             WARN_ON_ONCE(prim >= ARRAY_SIZE(scf_prim_name)))
201                 return;
202         scf_sel_totweight += weight;
203         scfsp->scfs_weight = scf_sel_totweight;
204         scfsp->scfs_prim = prim;
205         scfsp->scfs_wait = wait;
206         scf_sel_array_len++;
207 }
208
209 // Dump out weighting percentages for scf_prim_name[] array.
210 static void scf_sel_dump(void)
211 {
212         int i;
213         unsigned long oldw = 0;
214         struct scf_selector *scfsp;
215         unsigned long w;
216
217         for (i = 0; i < scf_sel_array_len; i++) {
218                 scfsp = &scf_sel_array[i];
219                 w = (scfsp->scfs_weight - oldw) * 100000 / scf_sel_totweight;
220                 pr_info("%s: %3lu.%03lu %s(%s)\n", __func__, w / 1000, w % 1000,
221                         scf_prim_name[scfsp->scfs_prim],
222                         scfsp->scfs_wait ? "wait" : "nowait");
223                 oldw = scfsp->scfs_weight;
224         }
225 }
226
227 // Randomly pick a primitive and wait/nowait, based on weightings.
228 static struct scf_selector *scf_sel_rand(struct torture_random_state *trsp)
229 {
230         int i;
231         unsigned long w = torture_random(trsp) % (scf_sel_totweight + 1);
232
233         for (i = 0; i < scf_sel_array_len; i++)
234                 if (scf_sel_array[i].scfs_weight >= w)
235                         return &scf_sel_array[i];
236         WARN_ON_ONCE(1);
237         return &scf_sel_array[0];
238 }
239
240 // Update statistics and occasionally burn up mass quantities of CPU time,
241 // if told to do so via scftorture.longwait.  Otherwise, occasionally burn
242 // a little bit.
243 static void scf_handler(void *scfc_in)
244 {
245         int i;
246         int j;
247         unsigned long r = torture_random(this_cpu_ptr(&scf_torture_rand));
248         struct scf_check *scfcp = scfc_in;
249
250         if (likely(scfcp)) {
251                 WRITE_ONCE(scfcp->scfc_out, false); // For multiple receivers.
252                 if (WARN_ON_ONCE(unlikely(!READ_ONCE(scfcp->scfc_in))))
253                         atomic_inc(&n_mb_in_errs);
254         }
255         this_cpu_inc(scf_invoked_count);
256         if (longwait <= 0) {
257                 if (!(r & 0xffc0)) {
258                         udelay(r & 0x3f);
259                         goto out;
260                 }
261         }
262         if (r & 0xfff)
263                 goto out;
264         r = (r >> 12);
265         if (longwait <= 0) {
266                 udelay((r & 0xff) + 1);
267                 goto out;
268         }
269         r = r % longwait + 1;
270         for (i = 0; i < r; i++) {
271                 for (j = 0; j < 1000; j++) {
272                         udelay(1000);
273                         cpu_relax();
274                 }
275         }
276 out:
277         if (unlikely(!scfcp))
278                 return;
279         if (scfcp->scfc_wait)
280                 WRITE_ONCE(scfcp->scfc_out, true);
281         else
282                 kfree(scfcp);
283 }
284
285 // As above, but check for correct CPU.
286 static void scf_handler_1(void *scfc_in)
287 {
288         struct scf_check *scfcp = scfc_in;
289
290         if (likely(scfcp) && WARN_ONCE(smp_processor_id() != scfcp->scfc_cpu, "%s: Wanted CPU %d got CPU %d\n", __func__, scfcp->scfc_cpu, smp_processor_id())) {
291                 atomic_inc(&n_errs);
292         }
293         scf_handler(scfcp);
294 }
295
296 // Randomly do an smp_call_function*() invocation.
297 static void scftorture_invoke_one(struct scf_statistics *scfp, struct torture_random_state *trsp)
298 {
299         uintptr_t cpu;
300         int ret = 0;
301         struct scf_check *scfcp = NULL;
302         struct scf_selector *scfsp = scf_sel_rand(trsp);
303
304         if (use_cpus_read_lock)
305                 cpus_read_lock();
306         else
307                 preempt_disable();
308         if (scfsp->scfs_prim == SCF_PRIM_SINGLE || scfsp->scfs_wait) {
309                 scfcp = kmalloc(sizeof(*scfcp), GFP_ATOMIC);
310                 if (!scfcp) {
311                         WARN_ON_ONCE(!IS_ENABLED(CONFIG_KASAN));
312                         atomic_inc(&n_alloc_errs);
313                 } else {
314                         scfcp->scfc_cpu = -1;
315                         scfcp->scfc_wait = scfsp->scfs_wait;
316                         scfcp->scfc_out = false;
317                 }
318         }
319         switch (scfsp->scfs_prim) {
320         case SCF_PRIM_SINGLE:
321                 cpu = torture_random(trsp) % nr_cpu_ids;
322                 if (scfsp->scfs_wait)
323                         scfp->n_single_wait++;
324                 else
325                         scfp->n_single++;
326                 if (scfcp) {
327                         scfcp->scfc_cpu = cpu;
328                         barrier(); // Prevent race-reduction compiler optimizations.
329                         scfcp->scfc_in = true;
330                 }
331                 ret = smp_call_function_single(cpu, scf_handler_1, (void *)scfcp, scfsp->scfs_wait);
332                 if (ret) {
333                         if (scfsp->scfs_wait)
334                                 scfp->n_single_wait_ofl++;
335                         else
336                                 scfp->n_single_ofl++;
337                         kfree(scfcp);
338                         scfcp = NULL;
339                 }
340                 break;
341         case SCF_PRIM_MANY:
342                 if (scfsp->scfs_wait)
343                         scfp->n_many_wait++;
344                 else
345                         scfp->n_many++;
346                 if (scfcp) {
347                         barrier(); // Prevent race-reduction compiler optimizations.
348                         scfcp->scfc_in = true;
349                 }
350                 smp_call_function_many(cpu_online_mask, scf_handler, scfcp, scfsp->scfs_wait);
351                 break;
352         case SCF_PRIM_ALL:
353                 if (scfsp->scfs_wait)
354                         scfp->n_all_wait++;
355                 else
356                         scfp->n_all++;
357                 if (scfcp) {
358                         barrier(); // Prevent race-reduction compiler optimizations.
359                         scfcp->scfc_in = true;
360                 }
361                 smp_call_function(scf_handler, scfcp, scfsp->scfs_wait);
362                 break;
363         default:
364                 WARN_ON_ONCE(1);
365                 if (scfcp)
366                         scfcp->scfc_out = true;
367         }
368         if (scfcp && scfsp->scfs_wait) {
369                 if (WARN_ON_ONCE((num_online_cpus() > 1 || scfsp->scfs_prim == SCF_PRIM_SINGLE) &&
370                                  !scfcp->scfc_out))
371                         atomic_inc(&n_mb_out_errs); // Leak rather than trash!
372                 else
373                         kfree(scfcp);
374                 barrier(); // Prevent race-reduction compiler optimizations.
375         }
376         if (use_cpus_read_lock)
377                 cpus_read_unlock();
378         else
379                 preempt_enable();
380         if (!(torture_random(trsp) & 0xfff))
381                 schedule_timeout_uninterruptible(1);
382 }
383
384 // SCF test kthread.  Repeatedly does calls to members of the
385 // smp_call_function() family of functions.
386 static int scftorture_invoker(void *arg)
387 {
388         int cpu;
389         DEFINE_TORTURE_RANDOM(rand);
390         struct scf_statistics *scfp = (struct scf_statistics *)arg;
391         bool was_offline = false;
392
393         VERBOSE_SCFTORTOUT("scftorture_invoker %d: task started", scfp->cpu);
394         cpu = scfp->cpu % nr_cpu_ids;
395         set_cpus_allowed_ptr(current, cpumask_of(cpu));
396         set_user_nice(current, MAX_NICE);
397         if (holdoff)
398                 schedule_timeout_interruptible(holdoff * HZ);
399
400         VERBOSE_SCFTORTOUT("scftorture_invoker %d: Waiting for all SCF torturers from cpu %d", scfp->cpu, smp_processor_id());
401
402         // Make sure that the CPU is affinitized appropriately during testing.
403         WARN_ON_ONCE(smp_processor_id() != scfp->cpu);
404
405         if (!atomic_dec_return(&n_started))
406                 while (atomic_read_acquire(&n_started)) {
407                         if (torture_must_stop()) {
408                                 VERBOSE_SCFTORTOUT("scftorture_invoker %d ended before starting", scfp->cpu);
409                                 goto end;
410                         }
411                         schedule_timeout_uninterruptible(1);
412                 }
413
414         VERBOSE_SCFTORTOUT("scftorture_invoker %d started", scfp->cpu);
415
416         do {
417                 scftorture_invoke_one(scfp, &rand);
418                 while (cpu_is_offline(cpu) && !torture_must_stop()) {
419                         schedule_timeout_interruptible(HZ / 5);
420                         was_offline = true;
421                 }
422                 if (was_offline) {
423                         set_cpus_allowed_ptr(current, cpumask_of(cpu));
424                         was_offline = false;
425                 }
426                 cond_resched();
427         } while (!torture_must_stop());
428
429         VERBOSE_SCFTORTOUT("scftorture_invoker %d ended", scfp->cpu);
430 end:
431         torture_kthread_stopping("scftorture_invoker");
432         return 0;
433 }
434
435 static void
436 scftorture_print_module_parms(const char *tag)
437 {
438         pr_alert(SCFTORT_FLAG
439                  "--- %s:  verbose=%d holdoff=%d longwait=%d nthreads=%d onoff_holdoff=%d onoff_interval=%d shutdown_secs=%d stat_interval=%d stutter_cpus=%d use_cpus_read_lock=%d, weight_single=%d, weight_single_wait=%d, weight_many=%d, weight_many_wait=%d, weight_all=%d, weight_all_wait=%d\n", tag,
440                  verbose, holdoff, longwait, nthreads, onoff_holdoff, onoff_interval, shutdown, stat_interval, stutter_cpus, use_cpus_read_lock, weight_single, weight_single_wait, weight_many, weight_many_wait, weight_all, weight_all_wait);
441 }
442
443 static void scf_cleanup_handler(void *unused)
444 {
445 }
446
447 static void scf_torture_cleanup(void)
448 {
449         int i;
450
451         if (torture_cleanup_begin())
452                 return;
453
454         WRITE_ONCE(scfdone, true);
455         if (nthreads)
456                 for (i = 0; i < nthreads; i++)
457                         torture_stop_kthread("scftorture_invoker", scf_stats_p[i].task);
458         else
459                 goto end;
460         smp_call_function(scf_cleanup_handler, NULL, 0);
461         torture_stop_kthread(scf_torture_stats, scf_torture_stats_task);
462         scf_torture_stats_print();  // -After- the stats thread is stopped!
463         kfree(scf_stats_p);  // -After- the last stats print has completed!
464         scf_stats_p = NULL;
465
466         if (atomic_read(&n_errs) || atomic_read(&n_mb_in_errs) || atomic_read(&n_mb_out_errs))
467                 scftorture_print_module_parms("End of test: FAILURE");
468         else if (torture_onoff_failures())
469                 scftorture_print_module_parms("End of test: LOCK_HOTPLUG");
470         else
471                 scftorture_print_module_parms("End of test: SUCCESS");
472
473 end:
474         torture_cleanup_end();
475 }
476
477 static int __init scf_torture_init(void)
478 {
479         long i;
480         int firsterr = 0;
481         unsigned long weight_single1 = weight_single;
482         unsigned long weight_single_wait1 = weight_single_wait;
483         unsigned long weight_many1 = weight_many;
484         unsigned long weight_many_wait1 = weight_many_wait;
485         unsigned long weight_all1 = weight_all;
486         unsigned long weight_all_wait1 = weight_all_wait;
487
488         if (!torture_init_begin(SCFTORT_STRING, verbose))
489                 return -EBUSY;
490
491         scftorture_print_module_parms("Start of test");
492
493         if (weight_single == -1 && weight_single_wait == -1 &&
494             weight_many == -1 && weight_many_wait == -1 &&
495             weight_all == -1 && weight_all_wait == -1) {
496                 weight_single1 = 2 * nr_cpu_ids;
497                 weight_single_wait1 = 2 * nr_cpu_ids;
498                 weight_many1 = 2;
499                 weight_many_wait1 = 2;
500                 weight_all1 = 1;
501                 weight_all_wait1 = 1;
502         } else {
503                 if (weight_single == -1)
504                         weight_single1 = 0;
505                 if (weight_single_wait == -1)
506                         weight_single_wait1 = 0;
507                 if (weight_many == -1)
508                         weight_many1 = 0;
509                 if (weight_many_wait == -1)
510                         weight_many_wait1 = 0;
511                 if (weight_all == -1)
512                         weight_all1 = 0;
513                 if (weight_all_wait == -1)
514                         weight_all_wait1 = 0;
515         }
516         if (weight_single1 == 0 && weight_single_wait1 == 0 &&
517             weight_many1 == 0 && weight_many_wait1 == 0 &&
518             weight_all1 == 0 && weight_all_wait1 == 0) {
519                 VERBOSE_SCFTORTOUT_ERRSTRING("all zero weights makes no sense");
520                 firsterr = -EINVAL;
521                 goto unwind;
522         }
523         scf_sel_add(weight_single1, SCF_PRIM_SINGLE, false);
524         scf_sel_add(weight_single_wait1, SCF_PRIM_SINGLE, true);
525         scf_sel_add(weight_many1, SCF_PRIM_MANY, false);
526         scf_sel_add(weight_many_wait1, SCF_PRIM_MANY, true);
527         scf_sel_add(weight_all1, SCF_PRIM_ALL, false);
528         scf_sel_add(weight_all_wait1, SCF_PRIM_ALL, true);
529         scf_sel_dump();
530
531         if (onoff_interval > 0) {
532                 firsterr = torture_onoff_init(onoff_holdoff * HZ, onoff_interval, NULL);
533                 if (firsterr)
534                         goto unwind;
535         }
536         if (shutdown_secs > 0) {
537                 firsterr = torture_shutdown_init(shutdown_secs, scf_torture_cleanup);
538                 if (firsterr)
539                         goto unwind;
540         }
541
542         // Worker tasks invoking smp_call_function().
543         if (nthreads < 0)
544                 nthreads = num_online_cpus();
545         scf_stats_p = kcalloc(nthreads, sizeof(scf_stats_p[0]), GFP_KERNEL);
546         if (!scf_stats_p) {
547                 VERBOSE_SCFTORTOUT_ERRSTRING("out of memory");
548                 firsterr = -ENOMEM;
549                 goto unwind;
550         }
551
552         VERBOSE_SCFTORTOUT("Starting %d smp_call_function() threads\n", nthreads);
553
554         atomic_set(&n_started, nthreads);
555         for (i = 0; i < nthreads; i++) {
556                 scf_stats_p[i].cpu = i;
557                 firsterr = torture_create_kthread(scftorture_invoker, (void *)&scf_stats_p[i],
558                                                   scf_stats_p[i].task);
559                 if (firsterr)
560                         goto unwind;
561         }
562         if (stat_interval > 0) {
563                 firsterr = torture_create_kthread(scf_torture_stats, NULL, scf_torture_stats_task);
564                 if (firsterr)
565                         goto unwind;
566         }
567
568         torture_init_end();
569         return 0;
570
571 unwind:
572         torture_init_end();
573         scf_torture_cleanup();
574         return firsterr;
575 }
576
577 module_init(scf_torture_init);
578 module_exit(scf_torture_cleanup);