1 // SPDX-License-Identifier: GPL-2.0
2
3 /*
4 * Test module for stress and analyze performance of vmalloc allocator.
5 * (C) 2018 Uladzislau Rezki (Sony) <urezki@gmail.com>
6 */
7 #include <linux/init.h>
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 #include <linux/vmalloc.h>
11 #include <linux/random.h>
12 #include <linux/kthread.h>
13 #include <linux/moduleparam.h>
14 #include <linux/completion.h>
15 #include <linux/delay.h>
16 #include <linux/mm.h>
17 #include <linux/rcupdate.h>
18 #include <linux/srcu.h>
19 #include <linux/slab.h>
20
21 #define __param(type, name, init, msg) \
22 static type name = init; \
23 module_param(name, type, 0444); \
24 MODULE_PARM_DESC(name, msg) \
25
26 __param(int, nr_threads, 0,
27 "Number of workers to perform tests(min: 1 max: USHRT_MAX)");
28
29 __param(bool, sequential_test_order, false,
30 "Use sequential stress tests order");
31
32 __param(int, test_repeat_count, 1,
33 "Set test repeat counter");
34
35 __param(int, test_loop_count, 1000000,
36 "Set test loop counter");
37
38 __param(int, nr_pages, 0,
39 "Set number of pages for fix_size_alloc_test(default: 1)");
40
41 __param(bool, use_huge, false,
42 "Use vmalloc_huge in fix_size_alloc_test");
43
44 __param(int, run_test_mask, 7,
45 "Set tests specified in the mask.\n\n"
46 "\t\tid: 1, name: fix_size_alloc_test\n"
47 "\t\tid: 2, name: full_fit_alloc_test\n"
48 "\t\tid: 4, name: long_busy_list_alloc_test\n"
49 "\t\tid: 8, name: random_size_alloc_test\n"
50 "\t\tid: 16, name: fix_align_alloc_test\n"
51 "\t\tid: 32, name: random_size_align_alloc_test\n"
52 "\t\tid: 64, name: align_shift_alloc_test\n"
53 "\t\tid: 128, name: pcpu_alloc_test\n"
54 "\t\tid: 256, name: kvfree_rcu_1_arg_vmalloc_test\n"
55 "\t\tid: 512, name: kvfree_rcu_2_arg_vmalloc_test\n"
56 "\t\tid: 1024, name: vm_map_ram_test\n"
57 "\t\tid: 2048, name: no_block_alloc_test\n"
58 /* Add a new test case description here. */
59 );
60
61 __param(int, nr_pcpu_objects, 35000,
62 "Number of pcpu objects to allocate for pcpu_alloc_test");
63
64 /*
65 * This is for synchronization of setup phase.
66 */
67 DEFINE_STATIC_SRCU(prepare_for_test_srcu);
68
69 /*
70 * Completion tracking for worker threads.
71 */
72 static DECLARE_COMPLETION(test_all_done_comp);
73 static atomic_t test_n_undone = ATOMIC_INIT(0);
74
75 static inline void
test_report_one_done(void)76 test_report_one_done(void)
77 {
78 if (atomic_dec_and_test(&test_n_undone))
79 complete(&test_all_done_comp);
80 }
81
random_size_align_alloc_test(void)82 static int random_size_align_alloc_test(void)
83 {
84 unsigned long size, align;
85 unsigned int rnd;
86 void *ptr;
87 int i;
88
89 for (i = 0; i < test_loop_count; i++) {
90 rnd = get_random_u8();
91
92 /*
93 * Maximum 1024 pages, if PAGE_SIZE is 4096.
94 */
95 align = 1 << (rnd % 23);
96
97 /*
98 * Maximum 10 pages.
99 */
100 size = ((rnd % 10) + 1) * PAGE_SIZE;
101
102 ptr = __vmalloc_node(size, align, GFP_KERNEL | __GFP_ZERO, 0,
103 __builtin_return_address(0));
104 if (!ptr)
105 return -1;
106
107 vfree(ptr);
108 }
109
110 return 0;
111 }
112
113 /*
114 * This test case is supposed to be failed.
115 */
align_shift_alloc_test(void)116 static int align_shift_alloc_test(void)
117 {
118 unsigned long align;
119 void *ptr;
120 int i;
121
122 for (i = 0; i < BITS_PER_LONG; i++) {
123 align = 1UL << i;
124
125 ptr = __vmalloc_node(PAGE_SIZE, align, GFP_KERNEL|__GFP_ZERO, 0,
126 __builtin_return_address(0));
127 if (!ptr)
128 return -1;
129
130 vfree(ptr);
131 }
132
133 return 0;
134 }
135
fix_align_alloc_test(void)136 static int fix_align_alloc_test(void)
137 {
138 void *ptr;
139 int i;
140
141 for (i = 0; i < test_loop_count; i++) {
142 ptr = __vmalloc_node(5 * PAGE_SIZE, THREAD_ALIGN << 1,
143 GFP_KERNEL | __GFP_ZERO, 0,
144 __builtin_return_address(0));
145 if (!ptr)
146 return -1;
147
148 vfree(ptr);
149 }
150
151 return 0;
152 }
153
random_size_alloc_test(void)154 static int random_size_alloc_test(void)
155 {
156 unsigned int n;
157 void *p;
158 int i;
159
160 for (i = 0; i < test_loop_count; i++) {
161 n = get_random_u32_inclusive(1, 100);
162 p = vmalloc(n * PAGE_SIZE);
163
164 if (!p)
165 return -1;
166
167 *((__u8 *)p) = 1;
168 vfree(p);
169 }
170
171 return 0;
172 }
173
long_busy_list_alloc_test(void)174 static int long_busy_list_alloc_test(void)
175 {
176 void *ptr_1, *ptr_2;
177 void **ptr;
178 int rv = -1;
179 int i;
180
181 ptr = vmalloc(sizeof(void *) * 15000);
182 if (!ptr)
183 return rv;
184
185 for (i = 0; i < 15000; i++)
186 ptr[i] = vmalloc(1 * PAGE_SIZE);
187
188 for (i = 0; i < test_loop_count; i++) {
189 ptr_1 = vmalloc(100 * PAGE_SIZE);
190 if (!ptr_1)
191 goto leave;
192
193 ptr_2 = vmalloc(1 * PAGE_SIZE);
194 if (!ptr_2) {
195 vfree(ptr_1);
196 goto leave;
197 }
198
199 *((__u8 *)ptr_1) = 0;
200 *((__u8 *)ptr_2) = 1;
201
202 vfree(ptr_1);
203 vfree(ptr_2);
204 }
205
206 /* Success */
207 rv = 0;
208
209 leave:
210 for (i = 0; i < 15000; i++)
211 vfree(ptr[i]);
212
213 vfree(ptr);
214 return rv;
215 }
216
full_fit_alloc_test(void)217 static int full_fit_alloc_test(void)
218 {
219 void **ptr, **junk_ptr, *tmp;
220 int junk_length;
221 int rv = -1;
222 int i;
223
224 junk_length = fls(num_online_cpus());
225 junk_length *= (32 * 1024 * 1024 / PAGE_SIZE);
226
227 ptr = vmalloc(sizeof(void *) * junk_length);
228 if (!ptr)
229 return rv;
230
231 junk_ptr = vmalloc(sizeof(void *) * junk_length);
232 if (!junk_ptr) {
233 vfree(ptr);
234 return rv;
235 }
236
237 for (i = 0; i < junk_length; i++) {
238 ptr[i] = vmalloc(1 * PAGE_SIZE);
239 junk_ptr[i] = vmalloc(1 * PAGE_SIZE);
240 }
241
242 for (i = 0; i < junk_length; i++)
243 vfree(junk_ptr[i]);
244
245 for (i = 0; i < test_loop_count; i++) {
246 tmp = vmalloc(1 * PAGE_SIZE);
247
248 if (!tmp)
249 goto error;
250
251 *((__u8 *)tmp) = 1;
252 vfree(tmp);
253 }
254
255 /* Success */
256 rv = 0;
257
258 error:
259 for (i = 0; i < junk_length; i++)
260 vfree(ptr[i]);
261
262 vfree(ptr);
263 vfree(junk_ptr);
264
265 return rv;
266 }
267
fix_size_alloc_test(void)268 static int fix_size_alloc_test(void)
269 {
270 void *ptr;
271 int i;
272
273 for (i = 0; i < test_loop_count; i++) {
274 if (use_huge)
275 ptr = vmalloc_huge((nr_pages > 0 ? nr_pages:1) * PAGE_SIZE, GFP_KERNEL);
276 else
277 ptr = vmalloc((nr_pages > 0 ? nr_pages:1) * PAGE_SIZE);
278
279 if (!ptr)
280 return -1;
281
282 *((__u8 *)ptr) = 0;
283
284 vfree(ptr);
285 }
286
287 return 0;
288 }
289
no_block_alloc_test(void)290 static int no_block_alloc_test(void)
291 {
292 void *ptr;
293 int i;
294
295 for (i = 0; i < test_loop_count; i++) {
296 bool use_atomic = !!(get_random_u8() % 2);
297 gfp_t gfp = use_atomic ? GFP_ATOMIC : GFP_NOWAIT;
298 unsigned long size = (nr_pages > 0 ? nr_pages : 1) * PAGE_SIZE;
299
300 preempt_disable();
301 ptr = __vmalloc(size, gfp);
302 preempt_enable();
303
304 if (!ptr)
305 return -1;
306
307 *((__u8 *)ptr) = 0;
308 vfree(ptr);
309 }
310
311 return 0;
312 }
313
314 static int
pcpu_alloc_test(void)315 pcpu_alloc_test(void)
316 {
317 int rv = 0;
318 #ifndef CONFIG_NEED_PER_CPU_KM
319 void __percpu **pcpu;
320 size_t size, align;
321 int i;
322
323 pcpu = vmalloc(sizeof(void __percpu *) * nr_pcpu_objects);
324 if (!pcpu)
325 return -1;
326
327 for (i = 0; i < nr_pcpu_objects; i++) {
328 size = get_random_u32_inclusive(1, PAGE_SIZE / 4);
329
330 /*
331 * Maximum PAGE_SIZE
332 */
333 align = 1 << get_random_u32_inclusive(1, PAGE_SHIFT - 1);
334
335 pcpu[i] = __alloc_percpu(size, align);
336 if (!pcpu[i])
337 rv = -1;
338 }
339
340 for (i = 0; i < nr_pcpu_objects; i++)
341 free_percpu(pcpu[i]);
342
343 vfree(pcpu);
344 #endif
345 return rv;
346 }
347
348 struct test_kvfree_rcu {
349 struct rcu_head rcu;
350 unsigned char array[20];
351 };
352
353 static int
kvfree_rcu_1_arg_vmalloc_test(void)354 kvfree_rcu_1_arg_vmalloc_test(void)
355 {
356 struct test_kvfree_rcu *p;
357 int i;
358
359 for (i = 0; i < test_loop_count; i++) {
360 p = vmalloc(1 * PAGE_SIZE);
361 if (!p)
362 return -1;
363
364 p->array[0] = 'a';
365 kvfree_rcu_mightsleep(p);
366 }
367
368 return 0;
369 }
370
371 static int
kvfree_rcu_2_arg_vmalloc_test(void)372 kvfree_rcu_2_arg_vmalloc_test(void)
373 {
374 struct test_kvfree_rcu *p;
375 int i;
376
377 for (i = 0; i < test_loop_count; i++) {
378 p = vmalloc(1 * PAGE_SIZE);
379 if (!p)
380 return -1;
381
382 p->array[0] = 'a';
383 kvfree_rcu(p, rcu);
384 }
385
386 return 0;
387 }
388
389 static int
vm_map_ram_test(void)390 vm_map_ram_test(void)
391 {
392 unsigned long nr_allocated;
393 unsigned int map_nr_pages;
394 unsigned char *v_ptr;
395 struct page **pages;
396 int i;
397
398 map_nr_pages = nr_pages > 0 ? nr_pages:1;
399 pages = kzalloc_objs(struct page *, map_nr_pages);
400 if (!pages)
401 return -1;
402
403 nr_allocated = alloc_pages_bulk(GFP_KERNEL, map_nr_pages, pages);
404 if (nr_allocated != map_nr_pages)
405 goto cleanup;
406
407 /* Run the test loop. */
408 for (i = 0; i < test_loop_count; i++) {
409 v_ptr = vm_map_ram(pages, map_nr_pages, NUMA_NO_NODE);
410 *v_ptr = 'a';
411 vm_unmap_ram(v_ptr, map_nr_pages);
412 }
413
414 cleanup:
415 for (i = 0; i < nr_allocated; i++)
416 __free_page(pages[i]);
417
418 kfree(pages);
419
420 /* 0 indicates success. */
421 return nr_allocated != map_nr_pages;
422 }
423
424 struct test_case_desc {
425 const char *test_name;
426 int (*test_func)(void);
427 bool xfail;
428 };
429
430 static struct test_case_desc test_case_array[] = {
431 { "fix_size_alloc_test", fix_size_alloc_test, },
432 { "full_fit_alloc_test", full_fit_alloc_test, },
433 { "long_busy_list_alloc_test", long_busy_list_alloc_test, },
434 { "random_size_alloc_test", random_size_alloc_test, },
435 { "fix_align_alloc_test", fix_align_alloc_test, },
436 { "random_size_align_alloc_test", random_size_align_alloc_test, },
437 { "align_shift_alloc_test", align_shift_alloc_test, true },
438 { "pcpu_alloc_test", pcpu_alloc_test, },
439 { "kvfree_rcu_1_arg_vmalloc_test", kvfree_rcu_1_arg_vmalloc_test, },
440 { "kvfree_rcu_2_arg_vmalloc_test", kvfree_rcu_2_arg_vmalloc_test, },
441 { "vm_map_ram_test", vm_map_ram_test, },
442 { "no_block_alloc_test", no_block_alloc_test, true },
443 /* Add a new test case here. */
444 };
445
446 struct test_case_data {
447 int test_failed;
448 int test_xfailed;
449 int test_passed;
450 u64 time;
451 };
452
453 static struct test_driver {
454 struct task_struct *task;
455 struct test_case_data data[ARRAY_SIZE(test_case_array)];
456
457 unsigned long start;
458 unsigned long stop;
459 } *tdriver;
460
shuffle_array(int * arr,int n)461 static void shuffle_array(int *arr, int n)
462 {
463 int i, j;
464
465 for (i = n - 1; i > 0; i--) {
466 /* Cut the range. */
467 j = get_random_u32_below(i);
468
469 /* Swap indexes. */
470 swap(arr[i], arr[j]);
471 }
472 }
473
test_func(void * private)474 static int test_func(void *private)
475 {
476 struct test_driver *t = private;
477 int random_array[ARRAY_SIZE(test_case_array)];
478 int index, i, j, ret;
479 ktime_t kt;
480 u64 delta;
481
482 for (i = 0; i < ARRAY_SIZE(test_case_array); i++)
483 random_array[i] = i;
484
485 if (!sequential_test_order)
486 shuffle_array(random_array, ARRAY_SIZE(test_case_array));
487
488 /*
489 * Block until initialization is done.
490 */
491 synchronize_srcu(&prepare_for_test_srcu);
492
493 t->start = get_cycles();
494 for (i = 0; i < ARRAY_SIZE(test_case_array); i++) {
495 index = random_array[i];
496
497 /*
498 * Skip tests if run_test_mask has been specified.
499 */
500 if (!((run_test_mask & (1 << index)) >> index))
501 continue;
502 kt = ktime_get();
503 for (j = 0; j < test_repeat_count; j++) {
504 ret = test_case_array[index].test_func();
505
506 if (!ret)
507 t->data[index].test_passed++;
508 else if (ret && test_case_array[index].xfail)
509 t->data[index].test_xfailed++;
510 else
511 t->data[index].test_failed++;
512 }
513
514 /*
515 * Take an average time that test took.
516 */
517 delta = (u64) ktime_us_delta(ktime_get(), kt);
518 do_div(delta, (u32) test_repeat_count);
519
520 t->data[index].time = delta;
521 }
522 t->stop = get_cycles();
523 test_report_one_done();
524
525 /*
526 * Wait for the kthread_stop() call.
527 */
528 while (!kthread_should_stop())
529 msleep(10);
530
531 return 0;
532 }
533
534 static int
init_test_configuration(void)535 init_test_configuration(void)
536 {
537 /*
538 * A maximum number of workers is defined as hard-coded
539 * value and set to USHRT_MAX. We add such gap just in
540 * case and for potential heavy stressing.
541 */
542 nr_threads = clamp(nr_threads, 1, (int) USHRT_MAX);
543
544 /* Allocate the space for test instances. */
545 tdriver = kvzalloc_objs(*tdriver, nr_threads);
546 if (tdriver == NULL)
547 return -1;
548
549 if (test_repeat_count <= 0)
550 test_repeat_count = 1;
551
552 if (test_loop_count <= 0)
553 test_loop_count = 1;
554
555 return 0;
556 }
557
do_concurrent_test(void)558 static void do_concurrent_test(void)
559 {
560 int i, ret, idx;
561
562 /*
563 * Set some basic configurations plus sanity check.
564 */
565 ret = init_test_configuration();
566 if (ret < 0)
567 return;
568
569 /*
570 * Put on hold all workers.
571 */
572 idx = srcu_read_lock(&prepare_for_test_srcu);
573
574 for (i = 0; i < nr_threads; i++) {
575 struct test_driver *t = &tdriver[i];
576
577 t->task = kthread_run(test_func, t, "vmalloc_test/%d", i);
578
579 if (!IS_ERR(t->task))
580 /* Success. */
581 atomic_inc(&test_n_undone);
582 else
583 pr_err("Failed to start %d kthread\n", i);
584 }
585
586 /*
587 * Now let the workers do their job.
588 */
589 srcu_read_unlock(&prepare_for_test_srcu, idx);
590
591 /*
592 * Sleep quiet until all workers are done with 1 second
593 * interval. Since the test can take a lot of time we
594 * can run into a stack trace of the hung task. That is
595 * why we go with completion_timeout and HZ value.
596 */
597 do {
598 ret = wait_for_completion_timeout(&test_all_done_comp, HZ);
599 } while (!ret);
600
601 for (i = 0; i < nr_threads; i++) {
602 struct test_driver *t = &tdriver[i];
603 int j;
604
605 if (!IS_ERR(t->task))
606 kthread_stop(t->task);
607
608 for (j = 0; j < ARRAY_SIZE(test_case_array); j++) {
609 if (!((run_test_mask & (1 << j)) >> j))
610 continue;
611
612 pr_info(
613 "Summary: %s passed: %d failed: %d xfailed: %d repeat: %d loops: %d avg: %llu usec\n",
614 test_case_array[j].test_name,
615 t->data[j].test_passed,
616 t->data[j].test_failed,
617 t->data[j].test_xfailed,
618 test_repeat_count, test_loop_count,
619 t->data[j].time);
620 }
621
622 pr_info("All test took worker%d=%lu cycles\n",
623 i, t->stop - t->start);
624 }
625
626 kvfree(tdriver);
627 }
628
vmalloc_test_init(void)629 static int __init vmalloc_test_init(void)
630 {
631 do_concurrent_test();
632 /* Fail will directly unload the module */
633 return IS_BUILTIN(CONFIG_TEST_VMALLOC) ? 0:-EAGAIN;
634 }
635
636 #ifdef MODULE
637 module_init(vmalloc_test_init)
638 #else
639 late_initcall(vmalloc_test_init);
640 #endif
641
642 MODULE_LICENSE("GPL");
643 MODULE_AUTHOR("Uladzislau Rezki");
644 MODULE_DESCRIPTION("vmalloc test module");
645