xref: /openbmc/linux/kernel/dma/map_benchmark.c (revision 2cf1c348)
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/math64.h>
15 #include <linux/module.h>
16 #include <linux/pci.h>
17 #include <linux/platform_device.h>
18 #include <linux/slab.h>
19 #include <linux/timekeeping.h>
20 
21 #define DMA_MAP_BENCHMARK	_IOWR('d', 1, struct map_benchmark)
22 #define DMA_MAP_MAX_THREADS	1024
23 #define DMA_MAP_MAX_SECONDS	300
24 #define DMA_MAP_MAX_TRANS_DELAY	(10 * NSEC_PER_MSEC)
25 
26 #define DMA_MAP_BIDIRECTIONAL	0
27 #define DMA_MAP_TO_DEVICE	1
28 #define DMA_MAP_FROM_DEVICE	2
29 
30 struct map_benchmark {
31 	__u64 avg_map_100ns; /* average map latency in 100ns */
32 	__u64 map_stddev; /* standard deviation of map latency */
33 	__u64 avg_unmap_100ns; /* as above */
34 	__u64 unmap_stddev;
35 	__u32 threads; /* how many threads will do map/unmap in parallel */
36 	__u32 seconds; /* how long the test will last */
37 	__s32 node; /* which numa node this benchmark will run on */
38 	__u32 dma_bits; /* DMA addressing capability */
39 	__u32 dma_dir; /* DMA data direction */
40 	__u32 dma_trans_ns; /* time for DMA transmission in ns */
41 	__u32 granule;	/* how many PAGE_SIZE will do map/unmap once a time */
42 	__u8 expansion[76];	/* For future use */
43 };
44 
45 struct map_benchmark_data {
46 	struct map_benchmark bparam;
47 	struct device *dev;
48 	struct dentry  *debugfs;
49 	enum dma_data_direction dir;
50 	atomic64_t sum_map_100ns;
51 	atomic64_t sum_unmap_100ns;
52 	atomic64_t sum_sq_map;
53 	atomic64_t sum_sq_unmap;
54 	atomic64_t loops;
55 };
56 
57 static int map_benchmark_thread(void *data)
58 {
59 	void *buf;
60 	dma_addr_t dma_addr;
61 	struct map_benchmark_data *map = data;
62 	int npages = map->bparam.granule;
63 	u64 size = npages * PAGE_SIZE;
64 	int ret = 0;
65 
66 	buf = alloc_pages_exact(size, GFP_KERNEL);
67 	if (!buf)
68 		return -ENOMEM;
69 
70 	while (!kthread_should_stop())  {
71 		u64 map_100ns, unmap_100ns, map_sq, unmap_sq;
72 		ktime_t map_stime, map_etime, unmap_stime, unmap_etime;
73 		ktime_t map_delta, unmap_delta;
74 
75 		/*
76 		 * for a non-coherent device, if we don't stain them in the
77 		 * cache, this will give an underestimate of the real-world
78 		 * overhead of BIDIRECTIONAL or TO_DEVICE mappings;
79 		 * 66 means evertything goes well! 66 is lucky.
80 		 */
81 		if (map->dir != DMA_FROM_DEVICE)
82 			memset(buf, 0x66, size);
83 
84 		map_stime = ktime_get();
85 		dma_addr = dma_map_single(map->dev, buf, size, map->dir);
86 		if (unlikely(dma_mapping_error(map->dev, dma_addr))) {
87 			pr_err("dma_map_single failed on %s\n",
88 				dev_name(map->dev));
89 			ret = -ENOMEM;
90 			goto out;
91 		}
92 		map_etime = ktime_get();
93 		map_delta = ktime_sub(map_etime, map_stime);
94 
95 		/* Pretend DMA is transmitting */
96 		ndelay(map->bparam.dma_trans_ns);
97 
98 		unmap_stime = ktime_get();
99 		dma_unmap_single(map->dev, dma_addr, size, map->dir);
100 		unmap_etime = ktime_get();
101 		unmap_delta = ktime_sub(unmap_etime, unmap_stime);
102 
103 		/* calculate sum and sum of squares */
104 
105 		map_100ns = div64_ul(map_delta,  100);
106 		unmap_100ns = div64_ul(unmap_delta, 100);
107 		map_sq = map_100ns * map_100ns;
108 		unmap_sq = unmap_100ns * unmap_100ns;
109 
110 		atomic64_add(map_100ns, &map->sum_map_100ns);
111 		atomic64_add(unmap_100ns, &map->sum_unmap_100ns);
112 		atomic64_add(map_sq, &map->sum_sq_map);
113 		atomic64_add(unmap_sq, &map->sum_sq_unmap);
114 		atomic64_inc(&map->loops);
115 	}
116 
117 out:
118 	free_pages_exact(buf, size);
119 	return ret;
120 }
121 
122 static int do_map_benchmark(struct map_benchmark_data *map)
123 {
124 	struct task_struct **tsk;
125 	int threads = map->bparam.threads;
126 	int node = map->bparam.node;
127 	const cpumask_t *cpu_mask = cpumask_of_node(node);
128 	u64 loops;
129 	int ret = 0;
130 	int i;
131 
132 	tsk = kmalloc_array(threads, sizeof(*tsk), GFP_KERNEL);
133 	if (!tsk)
134 		return -ENOMEM;
135 
136 	get_device(map->dev);
137 
138 	for (i = 0; i < threads; i++) {
139 		tsk[i] = kthread_create_on_node(map_benchmark_thread, map,
140 				map->bparam.node, "dma-map-benchmark/%d", i);
141 		if (IS_ERR(tsk[i])) {
142 			pr_err("create dma_map thread failed\n");
143 			ret = PTR_ERR(tsk[i]);
144 			goto out;
145 		}
146 
147 		if (node != NUMA_NO_NODE)
148 			kthread_bind_mask(tsk[i], cpu_mask);
149 	}
150 
151 	/* clear the old value in the previous benchmark */
152 	atomic64_set(&map->sum_map_100ns, 0);
153 	atomic64_set(&map->sum_unmap_100ns, 0);
154 	atomic64_set(&map->sum_sq_map, 0);
155 	atomic64_set(&map->sum_sq_unmap, 0);
156 	atomic64_set(&map->loops, 0);
157 
158 	for (i = 0; i < threads; i++) {
159 		get_task_struct(tsk[i]);
160 		wake_up_process(tsk[i]);
161 	}
162 
163 	msleep_interruptible(map->bparam.seconds * 1000);
164 
165 	/* wait for the completion of benchmark threads */
166 	for (i = 0; i < threads; i++) {
167 		ret = kthread_stop(tsk[i]);
168 		if (ret)
169 			goto out;
170 	}
171 
172 	loops = atomic64_read(&map->loops);
173 	if (likely(loops > 0)) {
174 		u64 map_variance, unmap_variance;
175 		u64 sum_map = atomic64_read(&map->sum_map_100ns);
176 		u64 sum_unmap = atomic64_read(&map->sum_unmap_100ns);
177 		u64 sum_sq_map = atomic64_read(&map->sum_sq_map);
178 		u64 sum_sq_unmap = atomic64_read(&map->sum_sq_unmap);
179 
180 		/* average latency */
181 		map->bparam.avg_map_100ns = div64_u64(sum_map, loops);
182 		map->bparam.avg_unmap_100ns = div64_u64(sum_unmap, loops);
183 
184 		/* standard deviation of latency */
185 		map_variance = div64_u64(sum_sq_map, loops) -
186 				map->bparam.avg_map_100ns *
187 				map->bparam.avg_map_100ns;
188 		unmap_variance = div64_u64(sum_sq_unmap, loops) -
189 				map->bparam.avg_unmap_100ns *
190 				map->bparam.avg_unmap_100ns;
191 		map->bparam.map_stddev = int_sqrt64(map_variance);
192 		map->bparam.unmap_stddev = int_sqrt64(unmap_variance);
193 	}
194 
195 out:
196 	for (i = 0; i < threads; i++)
197 		put_task_struct(tsk[i]);
198 	put_device(map->dev);
199 	kfree(tsk);
200 	return ret;
201 }
202 
203 static long map_benchmark_ioctl(struct file *file, unsigned int cmd,
204 		unsigned long arg)
205 {
206 	struct map_benchmark_data *map = file->private_data;
207 	void __user *argp = (void __user *)arg;
208 	u64 old_dma_mask;
209 	int ret;
210 
211 	if (copy_from_user(&map->bparam, argp, sizeof(map->bparam)))
212 		return -EFAULT;
213 
214 	switch (cmd) {
215 	case DMA_MAP_BENCHMARK:
216 		if (map->bparam.threads == 0 ||
217 		    map->bparam.threads > DMA_MAP_MAX_THREADS) {
218 			pr_err("invalid thread number\n");
219 			return -EINVAL;
220 		}
221 
222 		if (map->bparam.seconds == 0 ||
223 		    map->bparam.seconds > DMA_MAP_MAX_SECONDS) {
224 			pr_err("invalid duration seconds\n");
225 			return -EINVAL;
226 		}
227 
228 		if (map->bparam.dma_trans_ns > DMA_MAP_MAX_TRANS_DELAY) {
229 			pr_err("invalid transmission delay\n");
230 			return -EINVAL;
231 		}
232 
233 		if (map->bparam.node != NUMA_NO_NODE &&
234 		    !node_possible(map->bparam.node)) {
235 			pr_err("invalid numa node\n");
236 			return -EINVAL;
237 		}
238 
239 		if (map->bparam.granule < 1 || map->bparam.granule > 1024) {
240 			pr_err("invalid granule size\n");
241 			return -EINVAL;
242 		}
243 
244 		switch (map->bparam.dma_dir) {
245 		case DMA_MAP_BIDIRECTIONAL:
246 			map->dir = DMA_BIDIRECTIONAL;
247 			break;
248 		case DMA_MAP_FROM_DEVICE:
249 			map->dir = DMA_FROM_DEVICE;
250 			break;
251 		case DMA_MAP_TO_DEVICE:
252 			map->dir = DMA_TO_DEVICE;
253 			break;
254 		default:
255 			pr_err("invalid DMA direction\n");
256 			return -EINVAL;
257 		}
258 
259 		old_dma_mask = dma_get_mask(map->dev);
260 
261 		ret = dma_set_mask(map->dev,
262 				   DMA_BIT_MASK(map->bparam.dma_bits));
263 		if (ret) {
264 			pr_err("failed to set dma_mask on device %s\n",
265 				dev_name(map->dev));
266 			return -EINVAL;
267 		}
268 
269 		ret = do_map_benchmark(map);
270 
271 		/*
272 		 * restore the original dma_mask as many devices' dma_mask are
273 		 * set by architectures, acpi, busses. When we bind them back
274 		 * to their original drivers, those drivers shouldn't see
275 		 * dma_mask changed by benchmark
276 		 */
277 		dma_set_mask(map->dev, old_dma_mask);
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 MODULE_LICENSE("GPL");
383