xref: /openbmc/linux/kernel/dma/map_benchmark.c (revision 193b1fc1)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2020 HiSilicon Limited.
4  */
5 
6 #define pr_fmt(fmt)	KBUILD_MODNAME ": " fmt
7 
8 #include <linux/debugfs.h>
9 #include <linux/delay.h>
10 #include <linux/device.h>
11 #include <linux/dma-mapping.h>
12 #include <linux/kernel.h>
13 #include <linux/kthread.h>
14 #include <linux/map_benchmark.h>
15 #include <linux/math64.h>
16 #include <linux/module.h>
17 #include <linux/pci.h>
18 #include <linux/platform_device.h>
19 #include <linux/slab.h>
20 #include <linux/timekeeping.h>
21 
22 struct map_benchmark_data {
23 	struct map_benchmark bparam;
24 	struct device *dev;
25 	struct dentry  *debugfs;
26 	enum dma_data_direction dir;
27 	atomic64_t sum_map_100ns;
28 	atomic64_t sum_unmap_100ns;
29 	atomic64_t sum_sq_map;
30 	atomic64_t sum_sq_unmap;
31 	atomic64_t loops;
32 };
33 
34 static int map_benchmark_thread(void *data)
35 {
36 	void *buf;
37 	dma_addr_t dma_addr;
38 	struct map_benchmark_data *map = data;
39 	int npages = map->bparam.granule;
40 	u64 size = npages * PAGE_SIZE;
41 	int ret = 0;
42 
43 	buf = alloc_pages_exact(size, GFP_KERNEL);
44 	if (!buf)
45 		return -ENOMEM;
46 
47 	while (!kthread_should_stop())  {
48 		u64 map_100ns, unmap_100ns, map_sq, unmap_sq;
49 		ktime_t map_stime, map_etime, unmap_stime, unmap_etime;
50 		ktime_t map_delta, unmap_delta;
51 
52 		/*
53 		 * for a non-coherent device, if we don't stain them in the
54 		 * cache, this will give an underestimate of the real-world
55 		 * overhead of BIDIRECTIONAL or TO_DEVICE mappings;
56 		 * 66 means evertything goes well! 66 is lucky.
57 		 */
58 		if (map->dir != DMA_FROM_DEVICE)
59 			memset(buf, 0x66, size);
60 
61 		map_stime = ktime_get();
62 		dma_addr = dma_map_single(map->dev, buf, size, map->dir);
63 		if (unlikely(dma_mapping_error(map->dev, dma_addr))) {
64 			pr_err("dma_map_single failed on %s\n",
65 				dev_name(map->dev));
66 			ret = -ENOMEM;
67 			goto out;
68 		}
69 		map_etime = ktime_get();
70 		map_delta = ktime_sub(map_etime, map_stime);
71 
72 		/* Pretend DMA is transmitting */
73 		ndelay(map->bparam.dma_trans_ns);
74 
75 		unmap_stime = ktime_get();
76 		dma_unmap_single(map->dev, dma_addr, size, map->dir);
77 		unmap_etime = ktime_get();
78 		unmap_delta = ktime_sub(unmap_etime, unmap_stime);
79 
80 		/* calculate sum and sum of squares */
81 
82 		map_100ns = div64_ul(map_delta,  100);
83 		unmap_100ns = div64_ul(unmap_delta, 100);
84 		map_sq = map_100ns * map_100ns;
85 		unmap_sq = unmap_100ns * unmap_100ns;
86 
87 		atomic64_add(map_100ns, &map->sum_map_100ns);
88 		atomic64_add(unmap_100ns, &map->sum_unmap_100ns);
89 		atomic64_add(map_sq, &map->sum_sq_map);
90 		atomic64_add(unmap_sq, &map->sum_sq_unmap);
91 		atomic64_inc(&map->loops);
92 
93 		/*
94 		 * We may test for a long time so periodically check whether
95 		 * we need to schedule to avoid starving the others. Otherwise
96 		 * we may hangup the kernel in a non-preemptible kernel when
97 		 * the test kthreads number >= CPU number, the test kthreads
98 		 * will run endless on every CPU since the thread resposible
99 		 * for notifying the kthread stop (in do_map_benchmark())
100 		 * could not be scheduled.
101 		 *
102 		 * Note this may degrade the test concurrency since the test
103 		 * threads may need to share the CPU time with other load
104 		 * in the system. So it's recommended to run this benchmark
105 		 * on an idle system.
106 		 */
107 		cond_resched();
108 	}
109 
110 out:
111 	free_pages_exact(buf, size);
112 	return ret;
113 }
114 
115 static int do_map_benchmark(struct map_benchmark_data *map)
116 {
117 	struct task_struct **tsk;
118 	int threads = map->bparam.threads;
119 	int node = map->bparam.node;
120 	u64 loops;
121 	int ret = 0;
122 	int i;
123 
124 	tsk = kmalloc_array(threads, sizeof(*tsk), GFP_KERNEL);
125 	if (!tsk)
126 		return -ENOMEM;
127 
128 	get_device(map->dev);
129 
130 	for (i = 0; i < threads; i++) {
131 		tsk[i] = kthread_create_on_node(map_benchmark_thread, map,
132 				map->bparam.node, "dma-map-benchmark/%d", i);
133 		if (IS_ERR(tsk[i])) {
134 			pr_err("create dma_map thread failed\n");
135 			ret = PTR_ERR(tsk[i]);
136 			while (--i >= 0)
137 				kthread_stop(tsk[i]);
138 			goto out;
139 		}
140 
141 		if (node != NUMA_NO_NODE)
142 			kthread_bind_mask(tsk[i], cpumask_of_node(node));
143 	}
144 
145 	/* clear the old value in the previous benchmark */
146 	atomic64_set(&map->sum_map_100ns, 0);
147 	atomic64_set(&map->sum_unmap_100ns, 0);
148 	atomic64_set(&map->sum_sq_map, 0);
149 	atomic64_set(&map->sum_sq_unmap, 0);
150 	atomic64_set(&map->loops, 0);
151 
152 	for (i = 0; i < threads; i++) {
153 		get_task_struct(tsk[i]);
154 		wake_up_process(tsk[i]);
155 	}
156 
157 	msleep_interruptible(map->bparam.seconds * 1000);
158 
159 	/* wait for the completion of all started benchmark threads */
160 	for (i = 0; i < threads; i++) {
161 		int kthread_ret = kthread_stop_put(tsk[i]);
162 
163 		if (kthread_ret)
164 			ret = kthread_ret;
165 	}
166 
167 	if (ret)
168 		goto out;
169 
170 	loops = atomic64_read(&map->loops);
171 	if (likely(loops > 0)) {
172 		u64 map_variance, unmap_variance;
173 		u64 sum_map = atomic64_read(&map->sum_map_100ns);
174 		u64 sum_unmap = atomic64_read(&map->sum_unmap_100ns);
175 		u64 sum_sq_map = atomic64_read(&map->sum_sq_map);
176 		u64 sum_sq_unmap = atomic64_read(&map->sum_sq_unmap);
177 
178 		/* average latency */
179 		map->bparam.avg_map_100ns = div64_u64(sum_map, loops);
180 		map->bparam.avg_unmap_100ns = div64_u64(sum_unmap, loops);
181 
182 		/* standard deviation of latency */
183 		map_variance = div64_u64(sum_sq_map, loops) -
184 				map->bparam.avg_map_100ns *
185 				map->bparam.avg_map_100ns;
186 		unmap_variance = div64_u64(sum_sq_unmap, loops) -
187 				map->bparam.avg_unmap_100ns *
188 				map->bparam.avg_unmap_100ns;
189 		map->bparam.map_stddev = int_sqrt64(map_variance);
190 		map->bparam.unmap_stddev = int_sqrt64(unmap_variance);
191 	}
192 
193 out:
194 	put_device(map->dev);
195 	kfree(tsk);
196 	return ret;
197 }
198 
199 static long map_benchmark_ioctl(struct file *file, unsigned int cmd,
200 		unsigned long arg)
201 {
202 	struct map_benchmark_data *map = file->private_data;
203 	void __user *argp = (void __user *)arg;
204 	u64 old_dma_mask;
205 	int ret;
206 
207 	if (copy_from_user(&map->bparam, argp, sizeof(map->bparam)))
208 		return -EFAULT;
209 
210 	switch (cmd) {
211 	case DMA_MAP_BENCHMARK:
212 		if (map->bparam.threads == 0 ||
213 		    map->bparam.threads > DMA_MAP_MAX_THREADS) {
214 			pr_err("invalid thread number\n");
215 			return -EINVAL;
216 		}
217 
218 		if (map->bparam.seconds == 0 ||
219 		    map->bparam.seconds > DMA_MAP_MAX_SECONDS) {
220 			pr_err("invalid duration seconds\n");
221 			return -EINVAL;
222 		}
223 
224 		if (map->bparam.dma_trans_ns > DMA_MAP_MAX_TRANS_DELAY) {
225 			pr_err("invalid transmission delay\n");
226 			return -EINVAL;
227 		}
228 
229 		if (map->bparam.node != NUMA_NO_NODE &&
230 		    (map->bparam.node < 0 || map->bparam.node >= MAX_NUMNODES ||
231 		     !node_possible(map->bparam.node))) {
232 			pr_err("invalid numa node\n");
233 			return -EINVAL;
234 		}
235 
236 		if (map->bparam.granule < 1 || map->bparam.granule > 1024) {
237 			pr_err("invalid granule size\n");
238 			return -EINVAL;
239 		}
240 
241 		switch (map->bparam.dma_dir) {
242 		case DMA_MAP_BIDIRECTIONAL:
243 			map->dir = DMA_BIDIRECTIONAL;
244 			break;
245 		case DMA_MAP_FROM_DEVICE:
246 			map->dir = DMA_FROM_DEVICE;
247 			break;
248 		case DMA_MAP_TO_DEVICE:
249 			map->dir = DMA_TO_DEVICE;
250 			break;
251 		default:
252 			pr_err("invalid DMA direction\n");
253 			return -EINVAL;
254 		}
255 
256 		old_dma_mask = dma_get_mask(map->dev);
257 
258 		ret = dma_set_mask(map->dev,
259 				   DMA_BIT_MASK(map->bparam.dma_bits));
260 		if (ret) {
261 			pr_err("failed to set dma_mask on device %s\n",
262 				dev_name(map->dev));
263 			return -EINVAL;
264 		}
265 
266 		ret = do_map_benchmark(map);
267 
268 		/*
269 		 * restore the original dma_mask as many devices' dma_mask are
270 		 * set by architectures, acpi, busses. When we bind them back
271 		 * to their original drivers, those drivers shouldn't see
272 		 * dma_mask changed by benchmark
273 		 */
274 		dma_set_mask(map->dev, old_dma_mask);
275 
276 		if (ret)
277 			return ret;
278 		break;
279 	default:
280 		return -EINVAL;
281 	}
282 
283 	if (copy_to_user(argp, &map->bparam, sizeof(map->bparam)))
284 		return -EFAULT;
285 
286 	return ret;
287 }
288 
289 static const struct file_operations map_benchmark_fops = {
290 	.open			= simple_open,
291 	.unlocked_ioctl		= map_benchmark_ioctl,
292 };
293 
294 static void map_benchmark_remove_debugfs(void *data)
295 {
296 	struct map_benchmark_data *map = (struct map_benchmark_data *)data;
297 
298 	debugfs_remove(map->debugfs);
299 }
300 
301 static int __map_benchmark_probe(struct device *dev)
302 {
303 	struct dentry *entry;
304 	struct map_benchmark_data *map;
305 	int ret;
306 
307 	map = devm_kzalloc(dev, sizeof(*map), GFP_KERNEL);
308 	if (!map)
309 		return -ENOMEM;
310 	map->dev = dev;
311 
312 	ret = devm_add_action(dev, map_benchmark_remove_debugfs, map);
313 	if (ret) {
314 		pr_err("Can't add debugfs remove action\n");
315 		return ret;
316 	}
317 
318 	/*
319 	 * we only permit a device bound with this driver, 2nd probe
320 	 * will fail
321 	 */
322 	entry = debugfs_create_file("dma_map_benchmark", 0600, NULL, map,
323 			&map_benchmark_fops);
324 	if (IS_ERR(entry))
325 		return PTR_ERR(entry);
326 	map->debugfs = entry;
327 
328 	return 0;
329 }
330 
331 static int map_benchmark_platform_probe(struct platform_device *pdev)
332 {
333 	return __map_benchmark_probe(&pdev->dev);
334 }
335 
336 static struct platform_driver map_benchmark_platform_driver = {
337 	.driver		= {
338 		.name	= "dma_map_benchmark",
339 	},
340 	.probe = map_benchmark_platform_probe,
341 };
342 
343 static int
344 map_benchmark_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
345 {
346 	return __map_benchmark_probe(&pdev->dev);
347 }
348 
349 static struct pci_driver map_benchmark_pci_driver = {
350 	.name	= "dma_map_benchmark",
351 	.probe	= map_benchmark_pci_probe,
352 };
353 
354 static int __init map_benchmark_init(void)
355 {
356 	int ret;
357 
358 	ret = pci_register_driver(&map_benchmark_pci_driver);
359 	if (ret)
360 		return ret;
361 
362 	ret = platform_driver_register(&map_benchmark_platform_driver);
363 	if (ret) {
364 		pci_unregister_driver(&map_benchmark_pci_driver);
365 		return ret;
366 	}
367 
368 	return 0;
369 }
370 
371 static void __exit map_benchmark_cleanup(void)
372 {
373 	platform_driver_unregister(&map_benchmark_platform_driver);
374 	pci_unregister_driver(&map_benchmark_pci_driver);
375 }
376 
377 module_init(map_benchmark_init);
378 module_exit(map_benchmark_cleanup);
379 
380 MODULE_AUTHOR("Barry Song <song.bao.hua@hisilicon.com>");
381 MODULE_DESCRIPTION("dma_map benchmark driver");
382