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(int, run_test_mask, INT_MAX,
 42	"Set tests specified in the mask.\n\n"
 43		"\t\tid: 1,    name: fix_size_alloc_test\n"
 44		"\t\tid: 2,    name: full_fit_alloc_test\n"
 45		"\t\tid: 4,    name: long_busy_list_alloc_test\n"
 46		"\t\tid: 8,    name: random_size_alloc_test\n"
 47		"\t\tid: 16,   name: fix_align_alloc_test\n"
 48		"\t\tid: 32,   name: random_size_align_alloc_test\n"
 49		"\t\tid: 64,   name: align_shift_alloc_test\n"
 50		"\t\tid: 128,  name: pcpu_alloc_test\n"
 51		"\t\tid: 256,  name: kvfree_rcu_1_arg_vmalloc_test\n"
 52		"\t\tid: 512,  name: kvfree_rcu_2_arg_vmalloc_test\n"
 53		/* Add a new test case description here. */
 54);
 55
 56/*
 57 * Read write semaphore for synchronization of setup
 58 * phase that is done in main thread and workers.
 59 */
 60static DECLARE_RWSEM(prepare_for_test_rwsem);
 61
 62/*
 63 * Completion tracking for worker threads.
 64 */
 65static DECLARE_COMPLETION(test_all_done_comp);
 66static atomic_t test_n_undone = ATOMIC_INIT(0);
 67
 68static inline void
 69test_report_one_done(void)
 70{
 71	if (atomic_dec_and_test(&test_n_undone))
 72		complete(&test_all_done_comp);
 73}
 74
 75static int random_size_align_alloc_test(void)
 76{
 77	unsigned long size, align;
 78	unsigned int rnd;
 79	void *ptr;
 80	int i;
 81
 82	for (i = 0; i < test_loop_count; i++) {
 83		rnd = get_random_u8();
 84
 85		/*
 86		 * Maximum 1024 pages, if PAGE_SIZE is 4096.
 87		 */
 88		align = 1 << (rnd % 23);
 89
 90		/*
 91		 * Maximum 10 pages.
 92		 */
 93		size = ((rnd % 10) + 1) * PAGE_SIZE;
 94
 95		ptr = __vmalloc_node(size, align, GFP_KERNEL | __GFP_ZERO, 0,
 96				__builtin_return_address(0));
 97		if (!ptr)
 98			return -1;
 99
100		vfree(ptr);
101	}
102
103	return 0;
104}
105
106/*
107 * This test case is supposed to be failed.
108 */
109static int align_shift_alloc_test(void)
110{
111	unsigned long align;
112	void *ptr;
113	int i;
114
115	for (i = 0; i < BITS_PER_LONG; i++) {
116		align = ((unsigned long) 1) << i;
117
118		ptr = __vmalloc_node(PAGE_SIZE, align, GFP_KERNEL|__GFP_ZERO, 0,
119				__builtin_return_address(0));
120		if (!ptr)
121			return -1;
122
123		vfree(ptr);
124	}
125
126	return 0;
127}
128
129static int fix_align_alloc_test(void)
130{
131	void *ptr;
132	int i;
133
134	for (i = 0; i < test_loop_count; i++) {
135		ptr = __vmalloc_node(5 * PAGE_SIZE, THREAD_ALIGN << 1,
136				GFP_KERNEL | __GFP_ZERO, 0,
137				__builtin_return_address(0));
138		if (!ptr)
139			return -1;
140
141		vfree(ptr);
142	}
143
144	return 0;
145}
146
147static int random_size_alloc_test(void)
148{
149	unsigned int n;
150	void *p;
151	int i;
152
153	for (i = 0; i < test_loop_count; i++) {
154		n = get_random_u32_inclusive(1, 100);
155		p = vmalloc(n * PAGE_SIZE);
156
157		if (!p)
158			return -1;
159
160		*((__u8 *)p) = 1;
161		vfree(p);
162	}
163
164	return 0;
165}
166
167static int long_busy_list_alloc_test(void)
168{
169	void *ptr_1, *ptr_2;
170	void **ptr;
171	int rv = -1;
172	int i;
173
174	ptr = vmalloc(sizeof(void *) * 15000);
175	if (!ptr)
176		return rv;
177
178	for (i = 0; i < 15000; i++)
179		ptr[i] = vmalloc(1 * PAGE_SIZE);
180
181	for (i = 0; i < test_loop_count; i++) {
182		ptr_1 = vmalloc(100 * PAGE_SIZE);
183		if (!ptr_1)
184			goto leave;
185
186		ptr_2 = vmalloc(1 * PAGE_SIZE);
187		if (!ptr_2) {
188			vfree(ptr_1);
189			goto leave;
190		}
191
192		*((__u8 *)ptr_1) = 0;
193		*((__u8 *)ptr_2) = 1;
194
195		vfree(ptr_1);
196		vfree(ptr_2);
197	}
198
199	/*  Success */
200	rv = 0;
201
202leave:
203	for (i = 0; i < 15000; i++)
204		vfree(ptr[i]);
205
206	vfree(ptr);
207	return rv;
208}
209
210static int full_fit_alloc_test(void)
211{
212	void **ptr, **junk_ptr, *tmp;
213	int junk_length;
214	int rv = -1;
215	int i;
216
217	junk_length = fls(num_online_cpus());
218	junk_length *= (32 * 1024 * 1024 / PAGE_SIZE);
219
220	ptr = vmalloc(sizeof(void *) * junk_length);
221	if (!ptr)
222		return rv;
223
224	junk_ptr = vmalloc(sizeof(void *) * junk_length);
225	if (!junk_ptr) {
226		vfree(ptr);
227		return rv;
228	}
229
230	for (i = 0; i < junk_length; i++) {
231		ptr[i] = vmalloc(1 * PAGE_SIZE);
232		junk_ptr[i] = vmalloc(1 * PAGE_SIZE);
233	}
234
235	for (i = 0; i < junk_length; i++)
236		vfree(junk_ptr[i]);
237
238	for (i = 0; i < test_loop_count; i++) {
239		tmp = vmalloc(1 * PAGE_SIZE);
240
241		if (!tmp)
242			goto error;
243
244		*((__u8 *)tmp) = 1;
245		vfree(tmp);
246	}
247
248	/* Success */
249	rv = 0;
250
251error:
252	for (i = 0; i < junk_length; i++)
253		vfree(ptr[i]);
254
255	vfree(ptr);
256	vfree(junk_ptr);
257
258	return rv;
259}
260
261static int fix_size_alloc_test(void)
262{
263	void *ptr;
264	int i;
265
266	for (i = 0; i < test_loop_count; i++) {
267		ptr = vmalloc((nr_pages > 0 ? nr_pages:1) * PAGE_SIZE);
268
269		if (!ptr)
270			return -1;
271
272		*((__u8 *)ptr) = 0;
273
274		vfree(ptr);
275	}
276
277	return 0;
278}
279
280static int
281pcpu_alloc_test(void)
282{
283	int rv = 0;
284#ifndef CONFIG_NEED_PER_CPU_KM
285	void __percpu **pcpu;
286	size_t size, align;
287	int i;
288
289	pcpu = vmalloc(sizeof(void __percpu *) * 35000);
290	if (!pcpu)
291		return -1;
292
293	for (i = 0; i < 35000; i++) {
294		size = get_random_u32_inclusive(1, PAGE_SIZE / 4);
295
296		/*
297		 * Maximum PAGE_SIZE
298		 */
299		align = 1 << get_random_u32_inclusive(1, 11);
300
301		pcpu[i] = __alloc_percpu(size, align);
302		if (!pcpu[i])
303			rv = -1;
304	}
305
306	for (i = 0; i < 35000; i++)
307		free_percpu(pcpu[i]);
308
309	vfree(pcpu);
310#endif
311	return rv;
312}
313
314struct test_kvfree_rcu {
315	struct rcu_head rcu;
316	unsigned char array[20];
317};
318
319static int
320kvfree_rcu_1_arg_vmalloc_test(void)
321{
322	struct test_kvfree_rcu *p;
323	int i;
324
325	for (i = 0; i < test_loop_count; i++) {
326		p = vmalloc(1 * PAGE_SIZE);
327		if (!p)
328			return -1;
329
330		p->array[0] = 'a';
331		kvfree_rcu(p);
332	}
333
334	return 0;
335}
336
337static int
338kvfree_rcu_2_arg_vmalloc_test(void)
339{
340	struct test_kvfree_rcu *p;
341	int i;
342
343	for (i = 0; i < test_loop_count; i++) {
344		p = vmalloc(1 * PAGE_SIZE);
345		if (!p)
346			return -1;
347
348		p->array[0] = 'a';
349		kvfree_rcu(p, rcu);
350	}
351
352	return 0;
353}
354
355struct test_case_desc {
356	const char *test_name;
357	int (*test_func)(void);
358};
359
360static struct test_case_desc test_case_array[] = {
361	{ "fix_size_alloc_test", fix_size_alloc_test },
362	{ "full_fit_alloc_test", full_fit_alloc_test },
363	{ "long_busy_list_alloc_test", long_busy_list_alloc_test },
364	{ "random_size_alloc_test", random_size_alloc_test },
365	{ "fix_align_alloc_test", fix_align_alloc_test },
366	{ "random_size_align_alloc_test", random_size_align_alloc_test },
367	{ "align_shift_alloc_test", align_shift_alloc_test },
368	{ "pcpu_alloc_test", pcpu_alloc_test },
369	{ "kvfree_rcu_1_arg_vmalloc_test", kvfree_rcu_1_arg_vmalloc_test },
370	{ "kvfree_rcu_2_arg_vmalloc_test", kvfree_rcu_2_arg_vmalloc_test },
371	/* Add a new test case here. */
372};
373
374struct test_case_data {
375	int test_failed;
376	int test_passed;
377	u64 time;
378};
379
380static struct test_driver {
381	struct task_struct *task;
382	struct test_case_data data[ARRAY_SIZE(test_case_array)];
383
384	unsigned long start;
385	unsigned long stop;
386} *tdriver;
387
388static void shuffle_array(int *arr, int n)
389{
390	int i, j;
391
392	for (i = n - 1; i > 0; i--)  {
393		/* Cut the range. */
394		j = get_random_u32_below(i);
395
396		/* Swap indexes. */
397		swap(arr[i], arr[j]);
398	}
399}
400
401static int test_func(void *private)
402{
403	struct test_driver *t = private;
404	int random_array[ARRAY_SIZE(test_case_array)];
405	int index, i, j;
406	ktime_t kt;
407	u64 delta;
408
409	for (i = 0; i < ARRAY_SIZE(test_case_array); i++)
410		random_array[i] = i;
411
412	if (!sequential_test_order)
413		shuffle_array(random_array, ARRAY_SIZE(test_case_array));
414
415	/*
416	 * Block until initialization is done.
417	 */
418	down_read(&prepare_for_test_rwsem);
419
420	t->start = get_cycles();
421	for (i = 0; i < ARRAY_SIZE(test_case_array); i++) {
422		index = random_array[i];
423
424		/*
425		 * Skip tests if run_test_mask has been specified.
426		 */
427		if (!((run_test_mask & (1 << index)) >> index))
428			continue;
429
430		kt = ktime_get();
431		for (j = 0; j < test_repeat_count; j++) {
432			if (!test_case_array[index].test_func())
433				t->data[index].test_passed++;
434			else
435				t->data[index].test_failed++;
436		}
437
438		/*
439		 * Take an average time that test took.
440		 */
441		delta = (u64) ktime_us_delta(ktime_get(), kt);
442		do_div(delta, (u32) test_repeat_count);
443
444		t->data[index].time = delta;
445	}
446	t->stop = get_cycles();
447
448	up_read(&prepare_for_test_rwsem);
449	test_report_one_done();
450
451	/*
452	 * Wait for the kthread_stop() call.
453	 */
454	while (!kthread_should_stop())
455		msleep(10);
456
457	return 0;
458}
459
460static int
461init_test_configurtion(void)
462{
463	/*
464	 * A maximum number of workers is defined as hard-coded
465	 * value and set to USHRT_MAX. We add such gap just in
466	 * case and for potential heavy stressing.
467	 */
468	nr_threads = clamp(nr_threads, 1, (int) USHRT_MAX);
469
470	/* Allocate the space for test instances. */
471	tdriver = kvcalloc(nr_threads, sizeof(*tdriver), GFP_KERNEL);
472	if (tdriver == NULL)
473		return -1;
474
475	if (test_repeat_count <= 0)
476		test_repeat_count = 1;
477
478	if (test_loop_count <= 0)
479		test_loop_count = 1;
480
481	return 0;
482}
483
484static void do_concurrent_test(void)
485{
486	int i, ret;
487
488	/*
489	 * Set some basic configurations plus sanity check.
490	 */
491	ret = init_test_configurtion();
492	if (ret < 0)
493		return;
494
495	/*
496	 * Put on hold all workers.
497	 */
498	down_write(&prepare_for_test_rwsem);
499
500	for (i = 0; i < nr_threads; i++) {
501		struct test_driver *t = &tdriver[i];
502
503		t->task = kthread_run(test_func, t, "vmalloc_test/%d", i);
504
505		if (!IS_ERR(t->task))
506			/* Success. */
507			atomic_inc(&test_n_undone);
508		else
509			pr_err("Failed to start %d kthread\n", i);
510	}
511
512	/*
513	 * Now let the workers do their job.
514	 */
515	up_write(&prepare_for_test_rwsem);
516
517	/*
518	 * Sleep quiet until all workers are done with 1 second
519	 * interval. Since the test can take a lot of time we
520	 * can run into a stack trace of the hung task. That is
521	 * why we go with completion_timeout and HZ value.
522	 */
523	do {
524		ret = wait_for_completion_timeout(&test_all_done_comp, HZ);
525	} while (!ret);
526
527	for (i = 0; i < nr_threads; i++) {
528		struct test_driver *t = &tdriver[i];
529		int j;
530
531		if (!IS_ERR(t->task))
532			kthread_stop(t->task);
533
534		for (j = 0; j < ARRAY_SIZE(test_case_array); j++) {
535			if (!((run_test_mask & (1 << j)) >> j))
536				continue;
537
538			pr_info(
539				"Summary: %s passed: %d failed: %d repeat: %d loops: %d avg: %llu usec\n",
540				test_case_array[j].test_name,
541				t->data[j].test_passed,
542				t->data[j].test_failed,
543				test_repeat_count, test_loop_count,
544				t->data[j].time);
545		}
546
547		pr_info("All test took worker%d=%lu cycles\n",
548			i, t->stop - t->start);
549	}
550
551	kvfree(tdriver);
552}
553
554static int vmalloc_test_init(void)
555{
556	do_concurrent_test();
557	return -EAGAIN; /* Fail will directly unload the module */
558}
559
560static void vmalloc_test_exit(void)
561{
562}
563
564module_init(vmalloc_test_init)
565module_exit(vmalloc_test_exit)
566
567MODULE_LICENSE("GPL");
568MODULE_AUTHOR("Uladzislau Rezki");
569MODULE_DESCRIPTION("vmalloc test module");