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/rwsem.h>
 17#include <linux/mm.h>
 18#include <linux/rcupdate.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, run_test_mask, INT_MAX,
 39	"Set tests specified in the mask.\n\n"
 40		"\t\tid: 1,    name: fix_size_alloc_test\n"
 41		"\t\tid: 2,    name: full_fit_alloc_test\n"
 42		"\t\tid: 4,    name: long_busy_list_alloc_test\n"
 43		"\t\tid: 8,    name: random_size_alloc_test\n"
 44		"\t\tid: 16,   name: fix_align_alloc_test\n"
 45		"\t\tid: 32,   name: random_size_align_alloc_test\n"
 46		"\t\tid: 64,   name: align_shift_alloc_test\n"
 47		"\t\tid: 128,  name: pcpu_alloc_test\n"
 48		"\t\tid: 256,  name: kvfree_rcu_1_arg_vmalloc_test\n"
 49		"\t\tid: 512,  name: kvfree_rcu_2_arg_vmalloc_test\n"
 50		/* Add a new test case description here. */
 51);
 52
 53/*
 54 * Read write semaphore for synchronization of setup
 55 * phase that is done in main thread and workers.
 56 */
 57static DECLARE_RWSEM(prepare_for_test_rwsem);
 58
 59/*
 60 * Completion tracking for worker threads.
 61 */
 62static DECLARE_COMPLETION(test_all_done_comp);
 63static atomic_t test_n_undone = ATOMIC_INIT(0);
 64
 65static inline void
 66test_report_one_done(void)
 67{
 68	if (atomic_dec_and_test(&test_n_undone))
 69		complete(&test_all_done_comp);
 70}
 71
 72static int random_size_align_alloc_test(void)
 73{
 74	unsigned long size, align, rnd;
 75	void *ptr;
 76	int i;
 77
 78	for (i = 0; i < test_loop_count; i++) {
 79		get_random_bytes(&rnd, sizeof(rnd));
 80
 81		/*
 82		 * Maximum 1024 pages, if PAGE_SIZE is 4096.
 83		 */
 84		align = 1 << (rnd % 23);
 85
 86		/*
 87		 * Maximum 10 pages.
 88		 */
 89		size = ((rnd % 10) + 1) * PAGE_SIZE;
 90
 91		ptr = __vmalloc_node(size, align, GFP_KERNEL | __GFP_ZERO, 0,
 92				__builtin_return_address(0));
 93		if (!ptr)
 94			return -1;
 95
 96		vfree(ptr);
 97	}
 98
 99	return 0;
100}
101
102/*
103 * This test case is supposed to be failed.
104 */
105static int align_shift_alloc_test(void)
106{
107	unsigned long align;
108	void *ptr;
109	int i;
110
111	for (i = 0; i < BITS_PER_LONG; i++) {
112		align = ((unsigned long) 1) << i;
113
114		ptr = __vmalloc_node(PAGE_SIZE, align, GFP_KERNEL|__GFP_ZERO, 0,
115				__builtin_return_address(0));
116		if (!ptr)
117			return -1;
118
119		vfree(ptr);
120	}
121
122	return 0;
123}
124
125static int fix_align_alloc_test(void)
126{
127	void *ptr;
128	int i;
129
130	for (i = 0; i < test_loop_count; i++) {
131		ptr = __vmalloc_node(5 * PAGE_SIZE, THREAD_ALIGN << 1,
132				GFP_KERNEL | __GFP_ZERO, 0,
133				__builtin_return_address(0));
134		if (!ptr)
135			return -1;
136
137		vfree(ptr);
138	}
139
140	return 0;
141}
142
143static int random_size_alloc_test(void)
144{
145	unsigned int n;
146	void *p;
147	int i;
148
149	for (i = 0; i < test_loop_count; i++) {
150		get_random_bytes(&n, sizeof(i));
151		n = (n % 100) + 1;
152
153		p = vmalloc(n * PAGE_SIZE);
154
155		if (!p)
156			return -1;
157
158		*((__u8 *)p) = 1;
159		vfree(p);
160	}
161
162	return 0;
163}
164
165static int long_busy_list_alloc_test(void)
166{
167	void *ptr_1, *ptr_2;
168	void **ptr;
169	int rv = -1;
170	int i;
171
172	ptr = vmalloc(sizeof(void *) * 15000);
173	if (!ptr)
174		return rv;
175
176	for (i = 0; i < 15000; i++)
177		ptr[i] = vmalloc(1 * PAGE_SIZE);
178
179	for (i = 0; i < test_loop_count; i++) {
180		ptr_1 = vmalloc(100 * PAGE_SIZE);
181		if (!ptr_1)
182			goto leave;
183
184		ptr_2 = vmalloc(1 * PAGE_SIZE);
185		if (!ptr_2) {
186			vfree(ptr_1);
187			goto leave;
188		}
189
190		*((__u8 *)ptr_1) = 0;
191		*((__u8 *)ptr_2) = 1;
192
193		vfree(ptr_1);
194		vfree(ptr_2);
195	}
196
197	/*  Success */
198	rv = 0;
199
200leave:
201	for (i = 0; i < 15000; i++)
202		vfree(ptr[i]);
203
204	vfree(ptr);
205	return rv;
206}
207
208static int full_fit_alloc_test(void)
209{
210	void **ptr, **junk_ptr, *tmp;
211	int junk_length;
212	int rv = -1;
213	int i;
214
215	junk_length = fls(num_online_cpus());
216	junk_length *= (32 * 1024 * 1024 / PAGE_SIZE);
217
218	ptr = vmalloc(sizeof(void *) * junk_length);
219	if (!ptr)
220		return rv;
221
222	junk_ptr = vmalloc(sizeof(void *) * junk_length);
223	if (!junk_ptr) {
224		vfree(ptr);
225		return rv;
226	}
227
228	for (i = 0; i < junk_length; i++) {
229		ptr[i] = vmalloc(1 * PAGE_SIZE);
230		junk_ptr[i] = vmalloc(1 * PAGE_SIZE);
231	}
232
233	for (i = 0; i < junk_length; i++)
234		vfree(junk_ptr[i]);
235
236	for (i = 0; i < test_loop_count; i++) {
237		tmp = vmalloc(1 * PAGE_SIZE);
238
239		if (!tmp)
240			goto error;
241
242		*((__u8 *)tmp) = 1;
243		vfree(tmp);
244	}
245
246	/* Success */
247	rv = 0;
248
249error:
250	for (i = 0; i < junk_length; i++)
251		vfree(ptr[i]);
252
253	vfree(ptr);
254	vfree(junk_ptr);
255
256	return rv;
257}
258
259static int fix_size_alloc_test(void)
260{
261	void *ptr;
262	int i;
263
264	for (i = 0; i < test_loop_count; i++) {
265		ptr = vmalloc(3 * PAGE_SIZE);
266
267		if (!ptr)
268			return -1;
269
270		*((__u8 *)ptr) = 0;
271
272		vfree(ptr);
273	}
274
275	return 0;
276}
277
278static int
279pcpu_alloc_test(void)
280{
281	int rv = 0;
282#ifndef CONFIG_NEED_PER_CPU_KM
283	void __percpu **pcpu;
284	size_t size, align;
285	int i;
286
287	pcpu = vmalloc(sizeof(void __percpu *) * 35000);
288	if (!pcpu)
289		return -1;
290
291	for (i = 0; i < 35000; i++) {
292		unsigned int r;
293
294		get_random_bytes(&r, sizeof(i));
295		size = (r % (PAGE_SIZE / 4)) + 1;
296
297		/*
298		 * Maximum PAGE_SIZE
299		 */
300		get_random_bytes(&r, sizeof(i));
301		align = 1 << ((i % 11) + 1);
302
303		pcpu[i] = __alloc_percpu(size, align);
304		if (!pcpu[i])
305			rv = -1;
306	}
307
308	for (i = 0; i < 35000; i++)
309		free_percpu(pcpu[i]);
310
311	vfree(pcpu);
312#endif
313	return rv;
314}
315
316struct test_kvfree_rcu {
317	struct rcu_head rcu;
318	unsigned char array[20];
319};
320
321static int
322kvfree_rcu_1_arg_vmalloc_test(void)
323{
324	struct test_kvfree_rcu *p;
325	int i;
326
327	for (i = 0; i < test_loop_count; i++) {
328		p = vmalloc(1 * PAGE_SIZE);
329		if (!p)
330			return -1;
331
332		p->array[0] = 'a';
333		kvfree_rcu(p);
334	}
335
336	return 0;
337}
338
339static int
340kvfree_rcu_2_arg_vmalloc_test(void)
341{
342	struct test_kvfree_rcu *p;
343	int i;
344
345	for (i = 0; i < test_loop_count; i++) {
346		p = vmalloc(1 * PAGE_SIZE);
347		if (!p)
348			return -1;
349
350		p->array[0] = 'a';
351		kvfree_rcu(p, rcu);
352	}
353
354	return 0;
355}
356
357struct test_case_desc {
358	const char *test_name;
359	int (*test_func)(void);
360};
361
362static struct test_case_desc test_case_array[] = {
363	{ "fix_size_alloc_test", fix_size_alloc_test },
364	{ "full_fit_alloc_test", full_fit_alloc_test },
365	{ "long_busy_list_alloc_test", long_busy_list_alloc_test },
366	{ "random_size_alloc_test", random_size_alloc_test },
367	{ "fix_align_alloc_test", fix_align_alloc_test },
368	{ "random_size_align_alloc_test", random_size_align_alloc_test },
369	{ "align_shift_alloc_test", align_shift_alloc_test },
370	{ "pcpu_alloc_test", pcpu_alloc_test },
371	{ "kvfree_rcu_1_arg_vmalloc_test", kvfree_rcu_1_arg_vmalloc_test },
372	{ "kvfree_rcu_2_arg_vmalloc_test", kvfree_rcu_2_arg_vmalloc_test },
373	/* Add a new test case here. */
374};
375
376struct test_case_data {
377	int test_failed;
378	int test_passed;
379	u64 time;
380};
381
382static struct test_driver {
383	struct task_struct *task;
384	struct test_case_data data[ARRAY_SIZE(test_case_array)];
385
386	unsigned long start;
387	unsigned long stop;
388} *tdriver;
389
390static void shuffle_array(int *arr, int n)
391{
392	unsigned int rnd;
393	int i, j, x;
394
395	for (i = n - 1; i > 0; i--)  {
396		get_random_bytes(&rnd, sizeof(rnd));
397
398		/* Cut the range. */
399		j = rnd % i;
400
401		/* Swap indexes. */
402		x = arr[i];
403		arr[i] = arr[j];
404		arr[j] = x;
405	}
406}
407
408static int test_func(void *private)
409{
410	struct test_driver *t = private;
411	int random_array[ARRAY_SIZE(test_case_array)];
412	int index, i, j;
413	ktime_t kt;
414	u64 delta;
415
416	for (i = 0; i < ARRAY_SIZE(test_case_array); i++)
417		random_array[i] = i;
418
419	if (!sequential_test_order)
420		shuffle_array(random_array, ARRAY_SIZE(test_case_array));
421
422	/*
423	 * Block until initialization is done.
424	 */
425	down_read(&prepare_for_test_rwsem);
426
427	t->start = get_cycles();
428	for (i = 0; i < ARRAY_SIZE(test_case_array); i++) {
429		index = random_array[i];
430
431		/*
432		 * Skip tests if run_test_mask has been specified.
433		 */
434		if (!((run_test_mask & (1 << index)) >> index))
435			continue;
436
437		kt = ktime_get();
438		for (j = 0; j < test_repeat_count; j++) {
439			if (!test_case_array[index].test_func())
440				t->data[index].test_passed++;
441			else
442				t->data[index].test_failed++;
443		}
444
445		/*
446		 * Take an average time that test took.
447		 */
448		delta = (u64) ktime_us_delta(ktime_get(), kt);
449		do_div(delta, (u32) test_repeat_count);
450
451		t->data[index].time = delta;
452	}
453	t->stop = get_cycles();
454
455	up_read(&prepare_for_test_rwsem);
456	test_report_one_done();
457
458	/*
459	 * Wait for the kthread_stop() call.
460	 */
461	while (!kthread_should_stop())
462		msleep(10);
463
464	return 0;
465}
466
467static int
468init_test_configurtion(void)
469{
470	/*
471	 * A maximum number of workers is defined as hard-coded
472	 * value and set to USHRT_MAX. We add such gap just in
473	 * case and for potential heavy stressing.
474	 */
475	nr_threads = clamp(nr_threads, 1, (int) USHRT_MAX);
476
477	/* Allocate the space for test instances. */
478	tdriver = kvcalloc(nr_threads, sizeof(*tdriver), GFP_KERNEL);
479	if (tdriver == NULL)
480		return -1;
481
482	if (test_repeat_count <= 0)
483		test_repeat_count = 1;
484
485	if (test_loop_count <= 0)
486		test_loop_count = 1;
487
488	return 0;
489}
490
491static void do_concurrent_test(void)
492{
493	int i, ret;
494
495	/*
496	 * Set some basic configurations plus sanity check.
497	 */
498	ret = init_test_configurtion();
499	if (ret < 0)
500		return;
501
502	/*
503	 * Put on hold all workers.
504	 */
505	down_write(&prepare_for_test_rwsem);
506
507	for (i = 0; i < nr_threads; i++) {
508		struct test_driver *t = &tdriver[i];
509
510		t->task = kthread_run(test_func, t, "vmalloc_test/%d", i);
511
512		if (!IS_ERR(t->task))
513			/* Success. */
514			atomic_inc(&test_n_undone);
515		else
516			pr_err("Failed to start %d kthread\n", i);
517	}
518
519	/*
520	 * Now let the workers do their job.
521	 */
522	up_write(&prepare_for_test_rwsem);
523
524	/*
525	 * Sleep quiet until all workers are done with 1 second
526	 * interval. Since the test can take a lot of time we
527	 * can run into a stack trace of the hung task. That is
528	 * why we go with completion_timeout and HZ value.
529	 */
530	do {
531		ret = wait_for_completion_timeout(&test_all_done_comp, HZ);
532	} while (!ret);
533
534	for (i = 0; i < nr_threads; i++) {
535		struct test_driver *t = &tdriver[i];
536		int j;
537
538		if (!IS_ERR(t->task))
539			kthread_stop(t->task);
540
541		for (j = 0; j < ARRAY_SIZE(test_case_array); j++) {
542			if (!((run_test_mask & (1 << j)) >> j))
543				continue;
544
545			pr_info(
546				"Summary: %s passed: %d failed: %d repeat: %d loops: %d avg: %llu usec\n",
547				test_case_array[j].test_name,
548				t->data[j].test_passed,
549				t->data[j].test_failed,
550				test_repeat_count, test_loop_count,
551				t->data[j].time);
552		}
553
554		pr_info("All test took worker%d=%lu cycles\n",
555			i, t->stop - t->start);
556	}
557
558	kvfree(tdriver);
559}
560
561static int vmalloc_test_init(void)
562{
563	do_concurrent_test();
564	return -EAGAIN; /* Fail will directly unload the module */
565}
566
567static void vmalloc_test_exit(void)
568{
569}
570
571module_init(vmalloc_test_init)
572module_exit(vmalloc_test_exit)
573
574MODULE_LICENSE("GPL");
575MODULE_AUTHOR("Uladzislau Rezki");
576MODULE_DESCRIPTION("vmalloc test module");