drivers/gpu/drm/i915/intel_hangcheck.c

   1 /*
   2  * Copyright © 2016 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_drv.h"
  26
  27 static bool instdone_unchanged(u32 current_instdone, u32 *old_instdone)
  28 {
  29         u32 tmp = current_instdone | *old_instdone;
  30         bool unchanged;
  31
  32         unchanged = tmp == *old_instdone;
  33         *old_instdone |= tmp;
  34
  35         return unchanged;
  36 }
  37
  38 static bool subunits_stuck(struct intel_engine_cs *engine)
  39 {
  40         struct drm_i915_private *dev_priv = engine->i915;
  41         struct intel_instdone instdone;
  42         struct intel_instdone *accu_instdone = &engine->hangcheck.instdone;
  43         bool stuck;
  44         int slice;
  45         int subslice;
  46
  47         if (engine->id != RCS)
  48                 return true;
  49
  50         intel_engine_get_instdone(engine, &instdone);
  51
  52         /* There might be unstable subunit states even when
  53          * actual head is not moving. Filter out the unstable ones by
  54          * accumulating the undone -> done transitions and only
  55          * consider those as progress.
  56          */
  57         stuck = instdone_unchanged(instdone.instdone,
  58                                    &accu_instdone->instdone);
  59         stuck &= instdone_unchanged(instdone.slice_common,
  60                                     &accu_instdone->slice_common);
  61
  62         for_each_instdone_slice_subslice(dev_priv, slice, subslice) {
  63                 stuck &= instdone_unchanged(instdone.sampler[slice][subslice],
  64                                             &accu_instdone->sampler[slice][subslice]);
  65                 stuck &= instdone_unchanged(instdone.row[slice][subslice],
  66                                             &accu_instdone->row[slice][subslice]);
  67         }
  68
  69         return stuck;
  70 }
  71
  72 static enum intel_engine_hangcheck_action
  73 head_stuck(struct intel_engine_cs *engine, u64 acthd)
  74 {
  75         if (acthd != engine->hangcheck.acthd) {
  76
  77                 /* Clear subunit states on head movement */
  78                 memset(&engine->hangcheck.instdone, 0,
  79                        sizeof(engine->hangcheck.instdone));
  80
  81                 return ENGINE_ACTIVE_HEAD;
  82         }
  83
  84         if (!subunits_stuck(engine))
  85                 return ENGINE_ACTIVE_SUBUNITS;
  86
  87         return ENGINE_DEAD;
  88 }
  89
  90 static enum intel_engine_hangcheck_action
  91 engine_stuck(struct intel_engine_cs *engine, u64 acthd)
  92 {
  93         struct drm_i915_private *dev_priv = engine->i915;
  94         enum intel_engine_hangcheck_action ha;
  95         u32 tmp;
  96
  97         ha = head_stuck(engine, acthd);
  98         if (ha != ENGINE_DEAD)
  99                 return ha;
 100
 101         if (IS_GEN(dev_priv, 2))
 102                 return ENGINE_DEAD;
 103
 104         /* Is the chip hanging on a WAIT_FOR_EVENT?
 105          * If so we can simply poke the RB_WAIT bit
 106          * and break the hang. This should work on
 107          * all but the second generation chipsets.
 108          */
 109         tmp = I915_READ_CTL(engine);
 110         if (tmp & RING_WAIT) {
 111                 i915_handle_error(dev_priv, BIT(engine->id), 0,
 112                                   "stuck wait on %s", engine->name);
 113                 I915_WRITE_CTL(engine, tmp);
 114                 return ENGINE_WAIT_KICK;
 115         }
 116
 117         return ENGINE_DEAD;
 118 }
 119
 120 static void hangcheck_load_sample(struct intel_engine_cs *engine,
 121                                   struct intel_engine_hangcheck *hc)
 122 {
 123         hc->acthd = intel_engine_get_active_head(engine);
 124         hc->seqno = intel_engine_get_seqno(engine);
 125 }
 126
 127 static void hangcheck_store_sample(struct intel_engine_cs *engine,
 128                                    const struct intel_engine_hangcheck *hc)
 129 {
 130         engine->hangcheck.acthd = hc->acthd;
 131         engine->hangcheck.seqno = hc->seqno;
 132         engine->hangcheck.action = hc->action;
 133         engine->hangcheck.stalled = hc->stalled;
 134         engine->hangcheck.wedged = hc->wedged;
 135 }
 136
 137 static enum intel_engine_hangcheck_action
 138 hangcheck_get_action(struct intel_engine_cs *engine,
 139                      const struct intel_engine_hangcheck *hc)
 140 {
 141         if (engine->hangcheck.seqno != hc->seqno)
 142                 return ENGINE_ACTIVE_SEQNO;
 143
 144         if (intel_engine_is_idle(engine))
 145                 return ENGINE_IDLE;
 146
 147         return engine_stuck(engine, hc->acthd);
 148 }
 149
 150 static void hangcheck_accumulate_sample(struct intel_engine_cs *engine,
 151                                         struct intel_engine_hangcheck *hc)
 152 {
 153         unsigned long timeout = I915_ENGINE_DEAD_TIMEOUT;
 154
 155         hc->action = hangcheck_get_action(engine, hc);
 156
 157         /* We always increment the progress
 158          * if the engine is busy and still processing
 159          * the same request, so that no single request
 160          * can run indefinitely (such as a chain of
 161          * batches). The only time we do not increment
 162          * the hangcheck score on this ring, if this
 163          * engine is in a legitimate wait for another
 164          * engine. In that case the waiting engine is a
 165          * victim and we want to be sure we catch the
 166          * right culprit. Then every time we do kick
 167          * the ring, make it as a progress as the seqno
 168          * advancement might ensure and if not, it
 169          * will catch the hanging engine.
 170          */
 171
 172         switch (hc->action) {
 173         case ENGINE_IDLE:
 174         case ENGINE_ACTIVE_SEQNO:
 175                 /* Clear head and subunit states on seqno movement */
 176                 hc->acthd = 0;
 177
 178                 memset(&engine->hangcheck.instdone, 0,
 179                        sizeof(engine->hangcheck.instdone));
 180
 181                 /* Intentional fall through */
 182         case ENGINE_WAIT_KICK:
 183         case ENGINE_WAIT:
 184                 engine->hangcheck.action_timestamp = jiffies;
 185                 break;
 186
 187         case ENGINE_ACTIVE_HEAD:
 188         case ENGINE_ACTIVE_SUBUNITS:
 189                 /*
 190                  * Seqno stuck with still active engine gets leeway,
 191                  * in hopes that it is just a long shader.
 192                  */
 193                 timeout = I915_SEQNO_DEAD_TIMEOUT;
 194                 break;
 195
 196         case ENGINE_DEAD:
 197                 if (GEM_SHOW_DEBUG()) {
 198                         struct drm_printer p = drm_debug_printer("hangcheck");
 199                         intel_engine_dump(engine, &p, "%s\n", engine->name);
 200                 }
 201                 break;
 202
 203         default:
 204                 MISSING_CASE(hc->action);
 205         }
 206
 207         hc->stalled = time_after(jiffies,
 208                                  engine->hangcheck.action_timestamp + timeout);
 209         hc->wedged = time_after(jiffies,
 210                                  engine->hangcheck.action_timestamp +
 211                                  I915_ENGINE_WEDGED_TIMEOUT);
 212 }
 213
 214 static void hangcheck_declare_hang(struct drm_i915_private *i915,
 215                                    unsigned int hung,
 216                                    unsigned int stuck)
 217 {
 218         struct intel_engine_cs *engine;
 219         char msg[80];
 220         unsigned int tmp;
 221         int len;
 222
 223         /* If some rings hung but others were still busy, only
 224          * blame the hanging rings in the synopsis.
 225          */
 226         if (stuck != hung)
 227                 hung &= ~stuck;
 228         len = scnprintf(msg, sizeof(msg),
 229                         "%s on ", stuck == hung ? "no progress" : "hang");
 230         for_each_engine_masked(engine, i915, hung, tmp)
 231                 len += scnprintf(msg + len, sizeof(msg) - len,
 232                                  "%s, ", engine->name);
 233         msg[len-2] = '\0';
 234
 235         return i915_handle_error(i915, hung, I915_ERROR_CAPTURE, "%s", msg);
 236 }
 237
 238 /*
 239  * This is called when the chip hasn't reported back with completed
 240  * batchbuffers in a long time. We keep track per ring seqno progress and
 241  * if there are no progress, hangcheck score for that ring is increased.
 242  * Further, acthd is inspected to see if the ring is stuck. On stuck case
 243  * we kick the ring. If we see no progress on three subsequent calls
 244  * we assume chip is wedged and try to fix it by resetting the chip.
 245  */
 246 static void i915_hangcheck_elapsed(struct work_struct *work)
 247 {
 248         struct drm_i915_private *dev_priv =
 249                 container_of(work, typeof(*dev_priv),
 250                              gpu_error.hangcheck_work.work);
 251         struct intel_engine_cs *engine;
 252         enum intel_engine_id id;
 253         unsigned int hung = 0, stuck = 0, wedged = 0;
 254
 255         if (!i915_modparams.enable_hangcheck)
 256                 return;
 257
 258         if (!READ_ONCE(dev_priv->gt.awake))
 259                 return;
 260
 261         if (i915_terminally_wedged(&dev_priv->gpu_error))
 262                 return;
 263
 264         /* As enabling the GPU requires fairly extensive mmio access,
 265          * periodically arm the mmio checker to see if we are triggering
 266          * any invalid access.
 267          */
 268         intel_uncore_arm_unclaimed_mmio_detection(dev_priv);
 269
 270         for_each_engine(engine, dev_priv, id) {
 271                 struct intel_engine_hangcheck hc;
 272
 273                 hangcheck_load_sample(engine, &hc);
 274                 hangcheck_accumulate_sample(engine, &hc);
 275                 hangcheck_store_sample(engine, &hc);
 276
 277                 if (engine->hangcheck.stalled) {
 278                         hung |= intel_engine_flag(engine);
 279                         if (hc.action != ENGINE_DEAD)
 280                                 stuck |= intel_engine_flag(engine);
 281                 }
 282
 283                 if (engine->hangcheck.wedged)
 284                         wedged |= intel_engine_flag(engine);
 285         }
 286
 287         if (wedged) {
 288                 dev_err(dev_priv->drm.dev,
 289                         "GPU recovery timed out,"
 290                         " cancelling all in-flight rendering.\n");
 291                 GEM_TRACE_DUMP();
 292                 i915_gem_set_wedged(dev_priv);
 293         }
 294
 295         if (hung)
 296                 hangcheck_declare_hang(dev_priv, hung, stuck);
 297
 298         /* Reset timer in case GPU hangs without another request being added */
 299         i915_queue_hangcheck(dev_priv);
 300 }
 301
 302 void intel_engine_init_hangcheck(struct intel_engine_cs *engine)
 303 {
 304         memset(&engine->hangcheck, 0, sizeof(engine->hangcheck));
 305         engine->hangcheck.action_timestamp = jiffies;
 306 }
 307
 308 void intel_hangcheck_init(struct drm_i915_private *i915)
 309 {
 310         INIT_DELAYED_WORK(&i915->gpu_error.hangcheck_work,
 311                           i915_hangcheck_elapsed);
 312 }
 313
 314 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 315 #include "selftests/intel_hangcheck.c"
 316 #endif