kernel/dma/map_benchmark.c

   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         __u8 expansion[80];     /* For future use */
  42 };
  43
  44 struct map_benchmark_data {
  45         struct map_benchmark bparam;
  46         struct device *dev;
  47         struct dentry  *debugfs;
  48         enum dma_data_direction dir;
  49         atomic64_t sum_map_100ns;
  50         atomic64_t sum_unmap_100ns;
  51         atomic64_t sum_sq_map;
  52         atomic64_t sum_sq_unmap;
  53         atomic64_t loops;
  54 };
  55
  56 static int map_benchmark_thread(void *data)
  57 {
  58         void *buf;
  59         dma_addr_t dma_addr;
  60         struct map_benchmark_data *map = data;
  61         int ret = 0;
  62
  63         buf = (void *)__get_free_page(GFP_KERNEL);
  64         if (!buf)
  65                 return -ENOMEM;
  66
  67         while (!kthread_should_stop())  {
  68                 u64 map_100ns, unmap_100ns, map_sq, unmap_sq;
  69                 ktime_t map_stime, map_etime, unmap_stime, unmap_etime;
  70                 ktime_t map_delta, unmap_delta;
  71
  72                 /*
  73                  * for a non-coherent device, if we don't stain them in the
  74                  * cache, this will give an underestimate of the real-world
  75                  * overhead of BIDIRECTIONAL or TO_DEVICE mappings;
  76                  * 66 means evertything goes well! 66 is lucky.
  77                  */
  78                 if (map->dir != DMA_FROM_DEVICE)
  79                         memset(buf, 0x66, PAGE_SIZE);
  80
  81                 map_stime = ktime_get();
  82                 dma_addr = dma_map_single(map->dev, buf, PAGE_SIZE, map->dir);
  83                 if (unlikely(dma_mapping_error(map->dev, dma_addr))) {
  84                         pr_err("dma_map_single failed on %s\n",
  85                                 dev_name(map->dev));
  86                         ret = -ENOMEM;
  87                         goto out;
  88                 }
  89                 map_etime = ktime_get();
  90                 map_delta = ktime_sub(map_etime, map_stime);
  91
  92                 /* Pretend DMA is transmitting */
  93                 ndelay(map->bparam.dma_trans_ns);
  94
  95                 unmap_stime = ktime_get();
  96                 dma_unmap_single(map->dev, dma_addr, PAGE_SIZE, map->dir);
  97                 unmap_etime = ktime_get();
  98                 unmap_delta = ktime_sub(unmap_etime, unmap_stime);
  99
 100                 /* calculate sum and sum of squares */
 101
 102                 map_100ns = div64_ul(map_delta,  100);
 103                 unmap_100ns = div64_ul(unmap_delta, 100);
 104                 map_sq = map_100ns * map_100ns;
 105                 unmap_sq = unmap_100ns * unmap_100ns;
 106
 107                 atomic64_add(map_100ns, &map->sum_map_100ns);
 108                 atomic64_add(unmap_100ns, &map->sum_unmap_100ns);
 109                 atomic64_add(map_sq, &map->sum_sq_map);
 110                 atomic64_add(unmap_sq, &map->sum_sq_unmap);
 111                 atomic64_inc(&map->loops);
 112         }
 113
 114 out:
 115         free_page((unsigned long)buf);
 116         return ret;
 117 }
 118
 119 static int do_map_benchmark(struct map_benchmark_data *map)
 120 {
 121         struct task_struct **tsk;
 122         int threads = map->bparam.threads;
 123         int node = map->bparam.node;
 124         const cpumask_t *cpu_mask = cpumask_of_node(node);
 125         u64 loops;
 126         int ret = 0;
 127         int i;
 128
 129         tsk = kmalloc_array(threads, sizeof(*tsk), GFP_KERNEL);
 130         if (!tsk)
 131                 return -ENOMEM;
 132
 133         get_device(map->dev);
 134
 135         for (i = 0; i < threads; i++) {
 136                 tsk[i] = kthread_create_on_node(map_benchmark_thread, map,
 137                                 map->bparam.node, "dma-map-benchmark/%d", i);
 138                 if (IS_ERR(tsk[i])) {
 139                         pr_err("create dma_map thread failed\n");
 140                         ret = PTR_ERR(tsk[i]);
 141                         goto out;
 142                 }
 143
 144                 if (node != NUMA_NO_NODE)
 145                         kthread_bind_mask(tsk[i], cpu_mask);
 146         }
 147
 148         /* clear the old value in the previous benchmark */
 149         atomic64_set(&map->sum_map_100ns, 0);
 150         atomic64_set(&map->sum_unmap_100ns, 0);
 151         atomic64_set(&map->sum_sq_map, 0);
 152         atomic64_set(&map->sum_sq_unmap, 0);
 153         atomic64_set(&map->loops, 0);
 154
 155         for (i = 0; i < threads; i++) {
 156                 get_task_struct(tsk[i]);
 157                 wake_up_process(tsk[i]);
 158         }
 159
 160         msleep_interruptible(map->bparam.seconds * 1000);
 161
 162         /* wait for the completion of benchmark threads */
 163         for (i = 0; i < threads; i++) {
 164                 ret = kthread_stop(tsk[i]);
 165                 if (ret)
 166                         goto out;
 167         }
 168
 169         loops = atomic64_read(&map->loops);
 170         if (likely(loops > 0)) {
 171                 u64 map_variance, unmap_variance;
 172                 u64 sum_map = atomic64_read(&map->sum_map_100ns);
 173                 u64 sum_unmap = atomic64_read(&map->sum_unmap_100ns);
 174                 u64 sum_sq_map = atomic64_read(&map->sum_sq_map);
 175                 u64 sum_sq_unmap = atomic64_read(&map->sum_sq_unmap);
 176
 177                 /* average latency */
 178                 map->bparam.avg_map_100ns = div64_u64(sum_map, loops);
 179                 map->bparam.avg_unmap_100ns = div64_u64(sum_unmap, loops);
 180
 181                 /* standard deviation of latency */
 182                 map_variance = div64_u64(sum_sq_map, loops) -
 183                                 map->bparam.avg_map_100ns *
 184                                 map->bparam.avg_map_100ns;
 185                 unmap_variance = div64_u64(sum_sq_unmap, loops) -
 186                                 map->bparam.avg_unmap_100ns *
 187                                 map->bparam.avg_unmap_100ns;
 188                 map->bparam.map_stddev = int_sqrt64(map_variance);
 189                 map->bparam.unmap_stddev = int_sqrt64(unmap_variance);
 190         }
 191
 192 out:
 193         for (i = 0; i < threads; i++)
 194                 put_task_struct(tsk[i]);
 195         put_device(map->dev);
 196         kfree(tsk);
 197         return ret;
 198 }
 199
 200 static long map_benchmark_ioctl(struct file *file, unsigned int cmd,
 201                 unsigned long arg)
 202 {
 203         struct map_benchmark_data *map = file->private_data;
 204         void __user *argp = (void __user *)arg;
 205         u64 old_dma_mask;
 206
 207         int ret;
 208
 209         if (copy_from_user(&map->bparam, argp, sizeof(map->bparam)))
 210                 return -EFAULT;
 211
 212         switch (cmd) {
 213         case DMA_MAP_BENCHMARK:
 214                 if (map->bparam.threads == 0 ||
 215                     map->bparam.threads > DMA_MAP_MAX_THREADS) {
 216                         pr_err("invalid thread number\n");
 217                         return -EINVAL;
 218                 }
 219
 220                 if (map->bparam.seconds == 0 ||
 221                     map->bparam.seconds > DMA_MAP_MAX_SECONDS) {
 222                         pr_err("invalid duration seconds\n");
 223                         return -EINVAL;
 224                 }
 225
 226                 if (map->bparam.dma_trans_ns > DMA_MAP_MAX_TRANS_DELAY) {
 227                         pr_err("invalid transmission delay\n");
 228                         return -EINVAL;
 229                 }
 230
 231                 if (map->bparam.node != NUMA_NO_NODE &&
 232                     !node_possible(map->bparam.node)) {
 233                         pr_err("invalid numa node\n");
 234                         return -EINVAL;
 235                 }
 236
 237                 switch (map->bparam.dma_dir) {
 238                 case DMA_MAP_BIDIRECTIONAL:
 239                         map->dir = DMA_BIDIRECTIONAL;
 240                         break;
 241                 case DMA_MAP_FROM_DEVICE:
 242                         map->dir = DMA_FROM_DEVICE;
 243                         break;
 244                 case DMA_MAP_TO_DEVICE:
 245                         map->dir = DMA_TO_DEVICE;
 246                         break;
 247                 default:
 248                         pr_err("invalid DMA direction\n");
 249                         return -EINVAL;
 250                 }
 251
 252                 old_dma_mask = dma_get_mask(map->dev);
 253
 254                 ret = dma_set_mask(map->dev,
 255                                    DMA_BIT_MASK(map->bparam.dma_bits));
 256                 if (ret) {
 257                         pr_err("failed to set dma_mask on device %s\n",
 258                                 dev_name(map->dev));
 259                         return -EINVAL;
 260                 }
 261
 262                 ret = do_map_benchmark(map);
 263
 264                 /*
 265                  * restore the original dma_mask as many devices' dma_mask are
 266                  * set by architectures, acpi, busses. When we bind them back
 267                  * to their original drivers, those drivers shouldn't see
 268                  * dma_mask changed by benchmark
 269                  */
 270                 dma_set_mask(map->dev, old_dma_mask);
 271                 break;
 272         default:
 273                 return -EINVAL;
 274         }
 275
 276         if (copy_to_user(argp, &map->bparam, sizeof(map->bparam)))
 277                 return -EFAULT;
 278
 279         return ret;
 280 }
 281
 282 static const struct file_operations map_benchmark_fops = {
 283         .open                   = simple_open,
 284         .unlocked_ioctl         = map_benchmark_ioctl,
 285 };
 286
 287 static void map_benchmark_remove_debugfs(void *data)
 288 {
 289         struct map_benchmark_data *map = (struct map_benchmark_data *)data;
 290
 291         debugfs_remove(map->debugfs);
 292 }
 293
 294 static int __map_benchmark_probe(struct device *dev)
 295 {
 296         struct dentry *entry;
 297         struct map_benchmark_data *map;
 298         int ret;
 299
 300         map = devm_kzalloc(dev, sizeof(*map), GFP_KERNEL);
 301         if (!map)
 302                 return -ENOMEM;
 303         map->dev = dev;
 304
 305         ret = devm_add_action(dev, map_benchmark_remove_debugfs, map);
 306         if (ret) {
 307                 pr_err("Can't add debugfs remove action\n");
 308                 return ret;
 309         }
 310
 311         /*
 312          * we only permit a device bound with this driver, 2nd probe
 313          * will fail
 314          */
 315         entry = debugfs_create_file("dma_map_benchmark", 0600, NULL, map,
 316                         &map_benchmark_fops);
 317         if (IS_ERR(entry))
 318                 return PTR_ERR(entry);
 319         map->debugfs = entry;
 320
 321         return 0;
 322 }
 323
 324 static int map_benchmark_platform_probe(struct platform_device *pdev)
 325 {
 326         return __map_benchmark_probe(&pdev->dev);
 327 }
 328
 329 static struct platform_driver map_benchmark_platform_driver = {
 330         .driver         = {
 331                 .name   = "dma_map_benchmark",
 332         },
 333         .probe = map_benchmark_platform_probe,
 334 };
 335
 336 static int
 337 map_benchmark_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 338 {
 339         return __map_benchmark_probe(&pdev->dev);
 340 }
 341
 342 static struct pci_driver map_benchmark_pci_driver = {
 343         .name   = "dma_map_benchmark",
 344         .probe  = map_benchmark_pci_probe,
 345 };
 346
 347 static int __init map_benchmark_init(void)
 348 {
 349         int ret;
 350
 351         ret = pci_register_driver(&map_benchmark_pci_driver);
 352         if (ret)
 353                 return ret;
 354
 355         ret = platform_driver_register(&map_benchmark_platform_driver);
 356         if (ret) {
 357                 pci_unregister_driver(&map_benchmark_pci_driver);
 358                 return ret;
 359         }
 360
 361         return 0;
 362 }
 363
 364 static void __exit map_benchmark_cleanup(void)
 365 {
 366         platform_driver_unregister(&map_benchmark_platform_driver);
 367         pci_unregister_driver(&map_benchmark_pci_driver);
 368 }
 369
 370 module_init(map_benchmark_init);
 371 module_exit(map_benchmark_cleanup);
 372
 373 MODULE_AUTHOR("Barry Song <song.bao.hua@hisilicon.com>");
 374 MODULE_DESCRIPTION("dma_map benchmark driver");
 375 MODULE_LICENSE("GPL");