drivers/gpu/drm/i915/selftests/i915_gem_timeline.c

   1 /*
   2  * Copyright © 2017 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include "../i915_selftest.h"
  26 #include "i915_random.h"
  27
  28 #include "mock_gem_device.h"
  29 #include "mock_timeline.h"
  30
  31 struct __igt_sync {
  32         const char *name;
  33         u32 seqno;
  34         bool expected;
  35         bool set;
  36 };
  37
  38 static int __igt_sync(struct intel_timeline *tl,
  39                       u64 ctx,
  40                       const struct __igt_sync *p,
  41                       const char *name)
  42 {
  43         int ret;
  44
  45         if (__intel_timeline_sync_is_later(tl, ctx, p->seqno) != p->expected) {
  46                 pr_err("%s: %s(ctx=%llu, seqno=%u) expected passed %s but failed\n",
  47                        name, p->name, ctx, p->seqno, yesno(p->expected));
  48                 return -EINVAL;
  49         }
  50
  51         if (p->set) {
  52                 ret = __intel_timeline_sync_set(tl, ctx, p->seqno);
  53                 if (ret)
  54                         return ret;
  55         }
  56
  57         return 0;
  58 }
  59
  60 static int igt_sync(void *arg)
  61 {
  62         const struct __igt_sync pass[] = {
  63                 { "unset", 0, false, false },
  64                 { "new", 0, false, true },
  65                 { "0a", 0, true, true },
  66                 { "1a", 1, false, true },
  67                 { "1b", 1, true, true },
  68                 { "0b", 0, true, false },
  69                 { "2a", 2, false, true },
  70                 { "4", 4, false, true },
  71                 { "INT_MAX", INT_MAX, false, true },
  72                 { "INT_MAX-1", INT_MAX-1, true, false },
  73                 { "INT_MAX+1", (u32)INT_MAX+1, false, true },
  74                 { "INT_MAX", INT_MAX, true, false },
  75                 { "UINT_MAX", UINT_MAX, false, true },
  76                 { "wrap", 0, false, true },
  77                 { "unwrap", UINT_MAX, true, false },
  78                 {},
  79         }, *p;
  80         struct intel_timeline *tl;
  81         int order, offset;
  82         int ret = -ENODEV;
  83
  84         tl = mock_timeline(0);
  85         if (!tl)
  86                 return -ENOMEM;
  87
  88         for (p = pass; p->name; p++) {
  89                 for (order = 1; order < 64; order++) {
  90                         for (offset = -1; offset <= (order > 1); offset++) {
  91                                 u64 ctx = BIT_ULL(order) + offset;
  92
  93                                 ret = __igt_sync(tl, ctx, p, "1");
  94                                 if (ret)
  95                                         goto out;
  96                         }
  97                 }
  98         }
  99         mock_timeline_destroy(tl);
 100
 101         tl = mock_timeline(0);
 102         if (!tl)
 103                 return -ENOMEM;
 104
 105         for (order = 1; order < 64; order++) {
 106                 for (offset = -1; offset <= (order > 1); offset++) {
 107                         u64 ctx = BIT_ULL(order) + offset;
 108
 109                         for (p = pass; p->name; p++) {
 110                                 ret = __igt_sync(tl, ctx, p, "2");
 111                                 if (ret)
 112                                         goto out;
 113                         }
 114                 }
 115         }
 116
 117 out:
 118         mock_timeline_destroy(tl);
 119         return ret;
 120 }
 121
 122 static unsigned int random_engine(struct rnd_state *rnd)
 123 {
 124         return i915_prandom_u32_max_state(I915_NUM_ENGINES, rnd);
 125 }
 126
 127 static int bench_sync(void *arg)
 128 {
 129         struct rnd_state prng;
 130         struct intel_timeline *tl;
 131         unsigned long end_time, count;
 132         u64 prng32_1M;
 133         ktime_t kt;
 134         int order, last_order;
 135
 136         tl = mock_timeline(0);
 137         if (!tl)
 138                 return -ENOMEM;
 139
 140         /* Lookups from cache are very fast and so the random number generation
 141          * and the loop itself becomes a significant factor in the per-iteration
 142          * timings. We try to compensate the results by measuring the overhead
 143          * of the prng and subtract it from the reported results.
 144          */
 145         prandom_seed_state(&prng, i915_selftest.random_seed);
 146         count = 0;
 147         kt = ktime_get();
 148         end_time = jiffies + HZ/10;
 149         do {
 150                 u32 x;
 151
 152                 /* Make sure the compiler doesn't optimise away the prng call */
 153                 WRITE_ONCE(x, prandom_u32_state(&prng));
 154
 155                 count++;
 156         } while (!time_after(jiffies, end_time));
 157         kt = ktime_sub(ktime_get(), kt);
 158         pr_debug("%s: %lu random evaluations, %lluns/prng\n",
 159                  __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
 160         prng32_1M = div64_ul(ktime_to_ns(kt) << 20, count);
 161
 162         /* Benchmark (only) setting random context ids */
 163         prandom_seed_state(&prng, i915_selftest.random_seed);
 164         count = 0;
 165         kt = ktime_get();
 166         end_time = jiffies + HZ/10;
 167         do {
 168                 u64 id = i915_prandom_u64_state(&prng);
 169
 170                 __intel_timeline_sync_set(tl, id, 0);
 171                 count++;
 172         } while (!time_after(jiffies, end_time));
 173         kt = ktime_sub(ktime_get(), kt);
 174         kt = ktime_sub_ns(kt, (count * prng32_1M * 2) >> 20);
 175         pr_info("%s: %lu random insertions, %lluns/insert\n",
 176                 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
 177
 178         /* Benchmark looking up the exact same context ids as we just set */
 179         prandom_seed_state(&prng, i915_selftest.random_seed);
 180         end_time = count;
 181         kt = ktime_get();
 182         while (end_time--) {
 183                 u64 id = i915_prandom_u64_state(&prng);
 184
 185                 if (!__intel_timeline_sync_is_later(tl, id, 0)) {
 186                         mock_timeline_destroy(tl);
 187                         pr_err("Lookup of %llu failed\n", id);
 188                         return -EINVAL;
 189                 }
 190         }
 191         kt = ktime_sub(ktime_get(), kt);
 192         kt = ktime_sub_ns(kt, (count * prng32_1M * 2) >> 20);
 193         pr_info("%s: %lu random lookups, %lluns/lookup\n",
 194                 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
 195
 196         mock_timeline_destroy(tl);
 197         cond_resched();
 198
 199         tl = mock_timeline(0);
 200         if (!tl)
 201                 return -ENOMEM;
 202
 203         /* Benchmark setting the first N (in order) contexts */
 204         count = 0;
 205         kt = ktime_get();
 206         end_time = jiffies + HZ/10;
 207         do {
 208                 __intel_timeline_sync_set(tl, count++, 0);
 209         } while (!time_after(jiffies, end_time));
 210         kt = ktime_sub(ktime_get(), kt);
 211         pr_info("%s: %lu in-order insertions, %lluns/insert\n",
 212                 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
 213
 214         /* Benchmark looking up the exact same context ids as we just set */
 215         end_time = count;
 216         kt = ktime_get();
 217         while (end_time--) {
 218                 if (!__intel_timeline_sync_is_later(tl, end_time, 0)) {
 219                         pr_err("Lookup of %lu failed\n", end_time);
 220                         mock_timeline_destroy(tl);
 221                         return -EINVAL;
 222                 }
 223         }
 224         kt = ktime_sub(ktime_get(), kt);
 225         pr_info("%s: %lu in-order lookups, %lluns/lookup\n",
 226                 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
 227
 228         mock_timeline_destroy(tl);
 229         cond_resched();
 230
 231         tl = mock_timeline(0);
 232         if (!tl)
 233                 return -ENOMEM;
 234
 235         /* Benchmark searching for a random context id and maybe changing it */
 236         prandom_seed_state(&prng, i915_selftest.random_seed);
 237         count = 0;
 238         kt = ktime_get();
 239         end_time = jiffies + HZ/10;
 240         do {
 241                 u32 id = random_engine(&prng);
 242                 u32 seqno = prandom_u32_state(&prng);
 243
 244                 if (!__intel_timeline_sync_is_later(tl, id, seqno))
 245                         __intel_timeline_sync_set(tl, id, seqno);
 246
 247                 count++;
 248         } while (!time_after(jiffies, end_time));
 249         kt = ktime_sub(ktime_get(), kt);
 250         kt = ktime_sub_ns(kt, (count * prng32_1M * 2) >> 20);
 251         pr_info("%s: %lu repeated insert/lookups, %lluns/op\n",
 252                 __func__, count, (long long)div64_ul(ktime_to_ns(kt), count));
 253         mock_timeline_destroy(tl);
 254         cond_resched();
 255
 256         /* Benchmark searching for a known context id and changing the seqno */
 257         for (last_order = 1, order = 1; order < 32;
 258              ({ int tmp = last_order; last_order = order; order += tmp; })) {
 259                 unsigned int mask = BIT(order) - 1;
 260
 261                 tl = mock_timeline(0);
 262                 if (!tl)
 263                         return -ENOMEM;
 264
 265                 count = 0;
 266                 kt = ktime_get();
 267                 end_time = jiffies + HZ/10;
 268                 do {
 269                         /* Without assuming too many details of the underlying
 270                          * implementation, try to identify its phase-changes
 271                          * (if any)!
 272                          */
 273                         u64 id = (u64)(count & mask) << order;
 274
 275                         __intel_timeline_sync_is_later(tl, id, 0);
 276                         __intel_timeline_sync_set(tl, id, 0);
 277
 278                         count++;
 279                 } while (!time_after(jiffies, end_time));
 280                 kt = ktime_sub(ktime_get(), kt);
 281                 pr_info("%s: %lu cyclic/%d insert/lookups, %lluns/op\n",
 282                         __func__, count, order,
 283                         (long long)div64_ul(ktime_to_ns(kt), count));
 284                 mock_timeline_destroy(tl);
 285                 cond_resched();
 286         }
 287
 288         return 0;
 289 }
 290
 291 int i915_gem_timeline_mock_selftests(void)
 292 {
 293         static const struct i915_subtest tests[] = {
 294                 SUBTEST(igt_sync),
 295                 SUBTEST(bench_sync),
 296         };
 297
 298         return i915_subtests(tests, NULL);
 299 }