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