test_vmalloc.c

  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/*
 62 * This is for synchronization of setup phase.
 63 */
 64DEFINE_STATIC_SRCU(prepare_for_test_srcu);
 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 no_block_alloc_test(void)
288{
289	void *ptr;
290	int i;
291
292	for (i = 0; i < test_loop_count; i++) {
293		bool use_atomic = !!(get_random_u8() % 2);
294		gfp_t gfp = use_atomic ? GFP_ATOMIC : GFP_NOWAIT;
295		unsigned long size = (nr_pages > 0 ? nr_pages : 1) * PAGE_SIZE;
296
297		preempt_disable();
298		ptr = __vmalloc(size, gfp);
299		preempt_enable();
300
301		if (!ptr)
302			return -1;
303
304		*((__u8 *)ptr) = 0;
305		vfree(ptr);
306	}
307
308	return 0;
309}
310
311static int
312pcpu_alloc_test(void)
313{
314	int rv = 0;
315#ifndef CONFIG_NEED_PER_CPU_KM
316	void __percpu **pcpu;
317	size_t size, align;
318	int i;
319
320	pcpu = vmalloc(sizeof(void __percpu *) * 35000);
321	if (!pcpu)
322		return -1;
323
324	for (i = 0; i < 35000; i++) {
325		size = get_random_u32_inclusive(1, PAGE_SIZE / 4);
326
327		/*
328		 * Maximum PAGE_SIZE
329		 */
330		align = 1 << get_random_u32_inclusive(1, 11);
331
332		pcpu[i] = __alloc_percpu(size, align);
333		if (!pcpu[i])
334			rv = -1;
335	}
336
337	for (i = 0; i < 35000; i++)
338		free_percpu(pcpu[i]);
339
340	vfree(pcpu);
341#endif
342	return rv;
343}
344
345struct test_kvfree_rcu {
346	struct rcu_head rcu;
347	unsigned char array[20];
348};
349
350static int
351kvfree_rcu_1_arg_vmalloc_test(void)
352{
353	struct test_kvfree_rcu *p;
354	int i;
355
356	for (i = 0; i < test_loop_count; i++) {
357		p = vmalloc(1 * PAGE_SIZE);
358		if (!p)
359			return -1;
360
361		p->array[0] = 'a';
362		kvfree_rcu_mightsleep(p);
363	}
364
365	return 0;
366}
367
368static int
369kvfree_rcu_2_arg_vmalloc_test(void)
370{
371	struct test_kvfree_rcu *p;
372	int i;
373
374	for (i = 0; i < test_loop_count; i++) {
375		p = vmalloc(1 * PAGE_SIZE);
376		if (!p)
377			return -1;
378
379		p->array[0] = 'a';
380		kvfree_rcu(p, rcu);
381	}
382
383	return 0;
384}
385
386static int
387vm_map_ram_test(void)
388{
389	unsigned long nr_allocated;
390	unsigned int map_nr_pages;
391	unsigned char *v_ptr;
392	struct page **pages;
393	int i;
394
395	map_nr_pages = nr_pages > 0 ? nr_pages:1;
396	pages = kcalloc(map_nr_pages, sizeof(struct page *), GFP_KERNEL);
397	if (!pages)
398		return -1;
399
400	nr_allocated = alloc_pages_bulk(GFP_KERNEL, map_nr_pages, pages);
401	if (nr_allocated != map_nr_pages)
402		goto cleanup;
403
404	/* Run the test loop. */
405	for (i = 0; i < test_loop_count; i++) {
406		v_ptr = vm_map_ram(pages, map_nr_pages, NUMA_NO_NODE);
407		*v_ptr = 'a';
408		vm_unmap_ram(v_ptr, map_nr_pages);
409	}
410
411cleanup:
412	for (i = 0; i < nr_allocated; i++)
413		__free_page(pages[i]);
414
415	kfree(pages);
416
417	/* 0 indicates success. */
418	return nr_allocated != map_nr_pages;
419}
420
421struct test_case_desc {
422	const char *test_name;
423	int (*test_func)(void);
424	bool xfail;
425};
426
427static struct test_case_desc test_case_array[] = {
428	{ "fix_size_alloc_test", fix_size_alloc_test, },
429	{ "full_fit_alloc_test", full_fit_alloc_test, },
430	{ "long_busy_list_alloc_test", long_busy_list_alloc_test, },
431	{ "random_size_alloc_test", random_size_alloc_test, },
432	{ "fix_align_alloc_test", fix_align_alloc_test, },
433	{ "random_size_align_alloc_test", random_size_align_alloc_test, },
434	{ "align_shift_alloc_test", align_shift_alloc_test, true },
435	{ "pcpu_alloc_test", pcpu_alloc_test, },
436	{ "kvfree_rcu_1_arg_vmalloc_test", kvfree_rcu_1_arg_vmalloc_test, },
437	{ "kvfree_rcu_2_arg_vmalloc_test", kvfree_rcu_2_arg_vmalloc_test, },
438	{ "vm_map_ram_test", vm_map_ram_test, },
439	{ "no_block_alloc_test", no_block_alloc_test, true },
440	/* Add a new test case here. */
441};
442
443struct test_case_data {
444	int test_failed;
445	int test_xfailed;
446	int test_passed;
447	u64 time;
448};
449
450static struct test_driver {
451	struct task_struct *task;
452	struct test_case_data data[ARRAY_SIZE(test_case_array)];
453
454	unsigned long start;
455	unsigned long stop;
456} *tdriver;
457
458static void shuffle_array(int *arr, int n)
459{
460	int i, j;
461
462	for (i = n - 1; i > 0; i--)  {
463		/* Cut the range. */
464		j = get_random_u32_below(i);
465
466		/* Swap indexes. */
467		swap(arr[i], arr[j]);
468	}
469}
470
471static int test_func(void *private)
472{
473	struct test_driver *t = private;
474	int random_array[ARRAY_SIZE(test_case_array)];
475	int index, i, j, ret;
476	ktime_t kt;
477	u64 delta;
478
479	for (i = 0; i < ARRAY_SIZE(test_case_array); i++)
480		random_array[i] = i;
481
482	if (!sequential_test_order)
483		shuffle_array(random_array, ARRAY_SIZE(test_case_array));
484
485	/*
486	 * Block until initialization is done.
487	 */
488	synchronize_srcu(&prepare_for_test_srcu);
489
490	t->start = get_cycles();
491	for (i = 0; i < ARRAY_SIZE(test_case_array); i++) {
492		index = random_array[i];
493
494		/*
495		 * Skip tests if run_test_mask has been specified.
496		 */
497		if (!((run_test_mask & (1 << index)) >> index))
498			continue;
499		kt = ktime_get();
500		for (j = 0; j < test_repeat_count; j++) {
501			ret = test_case_array[index].test_func();
502
503			if (!ret)
504				t->data[index].test_passed++;
505			else if (ret && test_case_array[index].xfail)
506				t->data[index].test_xfailed++;
507			else
508				t->data[index].test_failed++;
509		}
510
511		/*
512		 * Take an average time that test took.
513		 */
514		delta = (u64) ktime_us_delta(ktime_get(), kt);
515		do_div(delta, (u32) test_repeat_count);
516
517		t->data[index].time = delta;
518	}
519	t->stop = get_cycles();
520	test_report_one_done();
521
522	/*
523	 * Wait for the kthread_stop() call.
524	 */
525	while (!kthread_should_stop())
526		msleep(10);
527
528	return 0;
529}
530
531static int
532init_test_configuration(void)
533{
534	/*
535	 * A maximum number of workers is defined as hard-coded
536	 * value and set to USHRT_MAX. We add such gap just in
537	 * case and for potential heavy stressing.
538	 */
539	nr_threads = clamp(nr_threads, 1, (int) USHRT_MAX);
540
541	/* Allocate the space for test instances. */
542	tdriver = kvcalloc(nr_threads, sizeof(*tdriver), GFP_KERNEL);
543	if (tdriver == NULL)
544		return -1;
545
546	if (test_repeat_count <= 0)
547		test_repeat_count = 1;
548
549	if (test_loop_count <= 0)
550		test_loop_count = 1;
551
552	return 0;
553}
554
555static void do_concurrent_test(void)
556{
557	int i, ret, idx;
558
559	/*
560	 * Set some basic configurations plus sanity check.
561	 */
562	ret = init_test_configuration();
563	if (ret < 0)
564		return;
565
566	/*
567	 * Put on hold all workers.
568	 */
569	idx = srcu_read_lock(&prepare_for_test_srcu);
570
571	for (i = 0; i < nr_threads; i++) {
572		struct test_driver *t = &tdriver[i];
573
574		t->task = kthread_run(test_func, t, "vmalloc_test/%d", i);
575
576		if (!IS_ERR(t->task))
577			/* Success. */
578			atomic_inc(&test_n_undone);
579		else
580			pr_err("Failed to start %d kthread\n", i);
581	}
582
583	/*
584	 * Now let the workers do their job.
585	 */
586	srcu_read_unlock(&prepare_for_test_srcu, idx);
587
588	/*
589	 * Sleep quiet until all workers are done with 1 second
590	 * interval. Since the test can take a lot of time we
591	 * can run into a stack trace of the hung task. That is
592	 * why we go with completion_timeout and HZ value.
593	 */
594	do {
595		ret = wait_for_completion_timeout(&test_all_done_comp, HZ);
596	} while (!ret);
597
598	for (i = 0; i < nr_threads; i++) {
599		struct test_driver *t = &tdriver[i];
600		int j;
601
602		if (!IS_ERR(t->task))
603			kthread_stop(t->task);
604
605		for (j = 0; j < ARRAY_SIZE(test_case_array); j++) {
606			if (!((run_test_mask & (1 << j)) >> j))
607				continue;
608
609			pr_info(
610				"Summary: %s passed: %d failed: %d xfailed: %d repeat: %d loops: %d avg: %llu usec\n",
611				test_case_array[j].test_name,
612				t->data[j].test_passed,
613				t->data[j].test_failed,
614				t->data[j].test_xfailed,
615				test_repeat_count, test_loop_count,
616				t->data[j].time);
617		}
618
619		pr_info("All test took worker%d=%lu cycles\n",
620			i, t->stop - t->start);
621	}
622
623	kvfree(tdriver);
624}
625
626static int __init vmalloc_test_init(void)
627{
628	do_concurrent_test();
629	/* Fail will directly unload the module */
630	return IS_BUILTIN(CONFIG_TEST_VMALLOC) ? 0:-EAGAIN;
631}
632
633#ifdef MODULE
634module_init(vmalloc_test_init)
635#else
636late_initcall(vmalloc_test_init);
637#endif
638
639MODULE_LICENSE("GPL");
640MODULE_AUTHOR("Uladzislau Rezki");
641MODULE_DESCRIPTION("vmalloc test module");