src/gallium/drivers/iris/iris_fence.c

   1 /*
   2  * Copyright © 2018 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 shall be included
  12  * in all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  15  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  18  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  19  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  20  * DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 /**
  24  * @file iris_fence.c
  25  *
  26  * Fences for driver and IPC serialisation, scheduling and synchronisation.
  27  */
  28
  29 #include "drm-uapi/sync_file.h"
  30 #include "util/u_debug.h"
  31 #include "util/u_inlines.h"
  32 #include "intel/common/gen_gem.h"
  33
  34 #include "iris_batch.h"
  35 #include "iris_bufmgr.h"
  36 #include "iris_context.h"
  37 #include "iris_fence.h"
  38 #include "iris_screen.h"
  39
  40 static uint32_t
  41 gem_syncobj_create(int fd, uint32_t flags)
  42 {
  43    struct drm_syncobj_create args = {
  44       .flags = flags,
  45    };
  46
  47    gen_ioctl(fd, DRM_IOCTL_SYNCOBJ_CREATE, &args);
  48
  49    return args.handle;
  50 }
  51
  52 static void
  53 gem_syncobj_destroy(int fd, uint32_t handle)
  54 {
  55    struct drm_syncobj_destroy args = {
  56       .handle = handle,
  57    };
  58
  59    gen_ioctl(fd, DRM_IOCTL_SYNCOBJ_DESTROY, &args);
  60 }
  61
  62 /**
  63  * Make a new sync-point.
  64  */
  65 struct iris_syncobj *
  66 iris_create_syncobj(struct iris_screen *screen)
  67 {
  68    struct iris_syncobj *syncobj = malloc(sizeof(*syncobj));
  69
  70    if (!syncobj)
  71       return NULL;
  72
  73    syncobj->handle = gem_syncobj_create(screen->fd, 0);
  74    assert(syncobj->handle);
  75
  76    pipe_reference_init(&syncobj->ref, 1);
  77
  78    return syncobj;
  79 }
  80
  81 void
  82 iris_syncobj_destroy(struct iris_screen *screen, struct iris_syncobj *syncobj)
  83 {
  84    gem_syncobj_destroy(screen->fd, syncobj->handle);
  85    free(syncobj);
  86 }
  87
  88 /**
  89  * Add a sync-point to the batch, with the given flags.
  90  *
  91  * \p flags   One of I915_EXEC_FENCE_WAIT or I915_EXEC_FENCE_SIGNAL.
  92  */
  93 void
  94 iris_batch_add_syncobj(struct iris_batch *batch,
  95                        struct iris_syncobj *syncobj,
  96                        unsigned flags)
  97 {
  98    struct drm_i915_gem_exec_fence *fence =
  99       util_dynarray_grow(&batch->exec_fences, struct drm_i915_gem_exec_fence, 1);
 100
 101    *fence = (struct drm_i915_gem_exec_fence) {
 102       .handle = syncobj->handle,
 103       .flags = flags,
 104    };
 105
 106    struct iris_syncobj **store =
 107       util_dynarray_grow(&batch->syncobjs, struct iris_syncobj *, 1);
 108
 109    *store = NULL;
 110    iris_syncobj_reference(batch->screen, store, syncobj);
 111 }
 112
 113 /* ------------------------------------------------------------------- */
 114
 115 struct pipe_fence_handle {
 116    struct pipe_reference ref;
 117
 118    struct pipe_context *unflushed_ctx;
 119
 120    struct iris_fine_fence *fine[IRIS_BATCH_COUNT];
 121 };
 122
 123 static void
 124 iris_fence_destroy(struct pipe_screen *p_screen,
 125                    struct pipe_fence_handle *fence)
 126 {
 127    struct iris_screen *screen = (struct iris_screen *)p_screen;
 128
 129    for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++)
 130       iris_fine_fence_reference(screen, &fence->fine[i], NULL);
 131
 132    free(fence);
 133 }
 134
 135 static void
 136 iris_fence_reference(struct pipe_screen *p_screen,
 137                      struct pipe_fence_handle **dst,
 138                      struct pipe_fence_handle *src)
 139 {
 140    if (pipe_reference(*dst ? &(*dst)->ref : NULL,
 141                       src ? &src->ref : NULL))
 142       iris_fence_destroy(p_screen, *dst);
 143
 144    *dst = src;
 145 }
 146
 147 bool
 148 iris_wait_syncobj(struct pipe_screen *p_screen,
 149                   struct iris_syncobj *syncobj,
 150                   int64_t timeout_nsec)
 151 {
 152    if (!syncobj)
 153       return false;
 154
 155    struct iris_screen *screen = (struct iris_screen *)p_screen;
 156    struct drm_syncobj_wait args = {
 157       .handles = (uintptr_t)&syncobj->handle,
 158       .count_handles = 1,
 159       .timeout_nsec = timeout_nsec,
 160    };
 161    return gen_ioctl(screen->fd, DRM_IOCTL_SYNCOBJ_WAIT, &args);
 162 }
 163
 164 #define CSI "\e["
 165 #define BLUE_HEADER  CSI "0;97;44m"
 166 #define NORMAL       CSI "0m"
 167
 168 static void
 169 iris_fence_flush(struct pipe_context *ctx,
 170                  struct pipe_fence_handle **out_fence,
 171                  unsigned flags)
 172 {
 173    struct iris_screen *screen = (void *) ctx->screen;
 174    struct iris_context *ice = (struct iris_context *)ctx;
 175
 176    /* We require DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT (kernel 5.2+) for
 177     * deferred flushes.  Just ignore the request to defer on older kernels.
 178     */
 179    if (!(screen->kernel_features & KERNEL_HAS_WAIT_FOR_SUBMIT))
 180       flags &= ~PIPE_FLUSH_DEFERRED;
 181
 182    const bool deferred = flags & PIPE_FLUSH_DEFERRED;
 183
 184    if (flags & PIPE_FLUSH_END_OF_FRAME) {
 185       ice->frame++;
 186
 187       if (INTEL_DEBUG & DEBUG_SUBMIT) {
 188          fprintf(stderr, "%s ::: FRAME %-10u (ctx %p)%-35c%s\n",
 189                  (INTEL_DEBUG & DEBUG_COLOR) ? BLUE_HEADER : "",
 190                  ice->frame, ctx, ' ',
 191                  (INTEL_DEBUG & DEBUG_COLOR) ? NORMAL : "");
 192       }
 193    }
 194
 195    if (!deferred) {
 196       for (unsigned i = 0; i < IRIS_BATCH_COUNT; i++)
 197          iris_batch_flush(&ice->batches[i]);
 198    }
 199
 200    if (!out_fence)
 201       return;
 202
 203    struct pipe_fence_handle *fence = calloc(1, sizeof(*fence));
 204    if (!fence)
 205       return;
 206
 207    pipe_reference_init(&fence->ref, 1);
 208
 209    if (deferred)
 210       fence->unflushed_ctx = ctx;
 211
 212    for (unsigned b = 0; b < IRIS_BATCH_COUNT; b++) {
 213       struct iris_batch *batch = &ice->batches[b];
 214
 215       if (deferred && iris_batch_bytes_used(batch) > 0) {
 216          struct iris_fine_fence *fine =
 217             iris_fine_fence_new(batch, IRIS_FENCE_BOTTOM_OF_PIPE);
 218          iris_fine_fence_reference(screen, &fence->fine[b], fine);
 219          iris_fine_fence_reference(screen, &fine, NULL);
 220       } else {
 221          /* This batch has no commands queued up (perhaps we just flushed,
 222           * or all the commands are on the other batch).  Wait for the last
 223           * syncobj on this engine - unless it's already finished by now.
 224           */
 225          if (iris_fine_fence_signaled(batch->last_fence))
 226             continue;
 227
 228          iris_fine_fence_reference(screen, &fence->fine[b], batch->last_fence);
 229       }
 230    }
 231
 232    iris_fence_reference(ctx->screen, out_fence, NULL);
 233    *out_fence = fence;
 234 }
 235
 236 static void
 237 iris_fence_await(struct pipe_context *ctx,
 238                  struct pipe_fence_handle *fence)
 239 {
 240    struct iris_context *ice = (struct iris_context *)ctx;
 241
 242    /* Unflushed fences from the same context are no-ops. */
 243    if (ctx && ctx == fence->unflushed_ctx)
 244       return;
 245
 246    /* XXX: We can't safely flush the other context, because it might be
 247     *      bound to another thread, and poking at its internals wouldn't
 248     *      be safe.  In the future we should use MI_SEMAPHORE_WAIT and
 249     *      block until the other job has been submitted, relying on
 250     *      kernel timeslicing to preempt us until the other job is
 251     *      actually flushed and the seqno finally passes.
 252     */
 253    if (fence->unflushed_ctx) {
 254       pipe_debug_message(&ice->dbg, CONFORMANCE, "%s",
 255                          "glWaitSync on unflushed fence from another context "
 256                          "is unlikely to work without kernel 5.8+\n");
 257    }
 258
 259    /* Flush any current work in our context as it doesn't need to wait
 260     * for this fence.  Any future work in our context must wait.
 261     */
 262    for (unsigned b = 0; b < IRIS_BATCH_COUNT; b++) {
 263       struct iris_batch *batch = &ice->batches[b];
 264
 265       for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++) {
 266          struct iris_fine_fence *fine = fence->fine[i];
 267
 268          if (iris_fine_fence_signaled(fine))
 269             continue;
 270
 271          iris_batch_flush(batch);
 272          iris_batch_add_syncobj(batch, fine->syncobj, I915_EXEC_FENCE_WAIT);
 273       }
 274    }
 275 }
 276
 277 #define NSEC_PER_SEC (1000 * USEC_PER_SEC)
 278 #define USEC_PER_SEC (1000 * MSEC_PER_SEC)
 279 #define MSEC_PER_SEC (1000)
 280
 281 static uint64_t
 282 gettime_ns(void)
 283 {
 284    struct timespec current;
 285    clock_gettime(CLOCK_MONOTONIC, &current);
 286    return (uint64_t)current.tv_sec * NSEC_PER_SEC + current.tv_nsec;
 287 }
 288
 289 static uint64_t
 290 rel2abs(uint64_t timeout)
 291 {
 292    if (timeout == 0)
 293       return 0;
 294
 295    uint64_t current_time = gettime_ns();
 296    uint64_t max_timeout = (uint64_t) INT64_MAX - current_time;
 297
 298    timeout = MIN2(max_timeout, timeout);
 299
 300    return current_time + timeout;
 301 }
 302
 303 static bool
 304 iris_fence_finish(struct pipe_screen *p_screen,
 305                   struct pipe_context *ctx,
 306                   struct pipe_fence_handle *fence,
 307                   uint64_t timeout)
 308 {
 309    struct iris_context *ice = (struct iris_context *)ctx;
 310    struct iris_screen *screen = (struct iris_screen *)p_screen;
 311
 312    /* If we created the fence with PIPE_FLUSH_DEFERRED, we may not have
 313     * flushed yet.  Check if our syncobj is the current batch's signalling
 314     * syncobj - if so, we haven't flushed and need to now.
 315     *
 316     * The Gallium docs mention that a flush will occur if \p ctx matches
 317     * the context the fence was created with.  It may be NULL, so we check
 318     * that it matches first.
 319     */
 320    if (ctx && ctx == fence->unflushed_ctx) {
 321       for (unsigned i = 0; i < IRIS_BATCH_COUNT; i++) {
 322          struct iris_fine_fence *fine = fence->fine[i];
 323
 324          if (iris_fine_fence_signaled(fine))
 325             continue;
 326
 327          if (fine->syncobj == iris_batch_get_signal_syncobj(&ice->batches[i]))
 328             iris_batch_flush(&ice->batches[i]);
 329       }
 330
 331       /* The fence is no longer deferred. */
 332       fence->unflushed_ctx = NULL;
 333    }
 334
 335    unsigned int handle_count = 0;
 336    uint32_t handles[ARRAY_SIZE(fence->fine)];
 337    for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++) {
 338       struct iris_fine_fence *fine = fence->fine[i];
 339
 340       if (iris_fine_fence_signaled(fine))
 341          continue;
 342
 343       handles[handle_count++] = fine->syncobj->handle;
 344    }
 345
 346    if (handle_count == 0)
 347       return true;
 348
 349    struct drm_syncobj_wait args = {
 350       .handles = (uintptr_t)handles,
 351       .count_handles = handle_count,
 352       .timeout_nsec = rel2abs(timeout),
 353       .flags = DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL
 354    };
 355
 356    if (fence->unflushed_ctx) {
 357       /* This fence had a deferred flush from another context.  We can't
 358        * safely flush it here, because the context might be bound to a
 359        * different thread, and poking at its internals wouldn't be safe.
 360        *
 361        * Instead, use the WAIT_FOR_SUBMIT flag to block and hope that
 362        * another thread submits the work.
 363        */
 364       args.flags |= DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT;
 365    }
 366
 367    return gen_ioctl(screen->fd, DRM_IOCTL_SYNCOBJ_WAIT, &args) == 0;
 368 }
 369
 370 static int
 371 sync_merge_fd(int sync_fd, int new_fd)
 372 {
 373    if (sync_fd == -1)
 374       return new_fd;
 375
 376    if (new_fd == -1)
 377       return sync_fd;
 378
 379    struct sync_merge_data args = {
 380       .name = "iris fence",
 381       .fd2 = new_fd,
 382       .fence = -1,
 383    };
 384
 385    gen_ioctl(sync_fd, SYNC_IOC_MERGE, &args);
 386    close(new_fd);
 387    close(sync_fd);
 388
 389    return args.fence;
 390 }
 391
 392 static int
 393 iris_fence_get_fd(struct pipe_screen *p_screen,
 394                   struct pipe_fence_handle *fence)
 395 {
 396    struct iris_screen *screen = (struct iris_screen *)p_screen;
 397    int fd = -1;
 398
 399    /* Deferred fences aren't supported. */
 400    if (fence->unflushed_ctx)
 401       return -1;
 402
 403    for (unsigned i = 0; i < ARRAY_SIZE(fence->fine); i++) {
 404       struct iris_fine_fence *fine = fence->fine[i];
 405
 406       if (iris_fine_fence_signaled(fine))
 407          continue;
 408
 409       struct drm_syncobj_handle args = {
 410          .handle = fine->syncobj->handle,
 411          .flags = DRM_SYNCOBJ_HANDLE_TO_FD_FLAGS_EXPORT_SYNC_FILE,
 412          .fd = -1,
 413       };
 414
 415       gen_ioctl(screen->fd, DRM_IOCTL_SYNCOBJ_HANDLE_TO_FD, &args);
 416       fd = sync_merge_fd(fd, args.fd);
 417    }
 418
 419    if (fd == -1) {
 420       /* Our fence has no syncobj's recorded.  This means that all of the
 421        * batches had already completed, their syncobj's had been signalled,
 422        * and so we didn't bother to record them.  But we're being asked to
 423        * export such a fence.  So export a dummy already-signalled syncobj.
 424        */
 425       struct drm_syncobj_handle args = {
 426          .flags = DRM_SYNCOBJ_HANDLE_TO_FD_FLAGS_EXPORT_SYNC_FILE, .fd = -1,
 427       };
 428
 429       args.handle = gem_syncobj_create(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED);
 430       gen_ioctl(screen->fd, DRM_IOCTL_SYNCOBJ_HANDLE_TO_FD, &args);
 431       gem_syncobj_destroy(screen->fd, args.handle);
 432       return args.fd;
 433    }
 434
 435    return fd;
 436 }
 437
 438 static void
 439 iris_fence_create_fd(struct pipe_context *ctx,
 440                      struct pipe_fence_handle **out,
 441                      int fd,
 442                      enum pipe_fd_type type)
 443 {
 444    assert(type == PIPE_FD_TYPE_NATIVE_SYNC);
 445
 446    struct iris_screen *screen = (struct iris_screen *)ctx->screen;
 447    struct drm_syncobj_handle args = {
 448       .handle = gem_syncobj_create(screen->fd, DRM_SYNCOBJ_CREATE_SIGNALED),
 449       .flags = DRM_SYNCOBJ_FD_TO_HANDLE_FLAGS_IMPORT_SYNC_FILE,
 450       .fd = fd,
 451    };
 452    if (gen_ioctl(screen->fd, DRM_IOCTL_SYNCOBJ_FD_TO_HANDLE, &args) == -1) {
 453       fprintf(stderr, "DRM_IOCTL_SYNCOBJ_FD_TO_HANDLE failed: %s\n",
 454               strerror(errno));
 455       gem_syncobj_destroy(screen->fd, args.handle);
 456       *out = NULL;
 457       return;
 458    }
 459
 460    struct iris_syncobj *syncobj = malloc(sizeof(*syncobj));
 461    if (!syncobj) {
 462       *out = NULL;
 463       return;
 464    }
 465    syncobj->handle = args.handle;
 466    pipe_reference_init(&syncobj->ref, 1);
 467
 468    struct iris_fine_fence *fine = calloc(1, sizeof(*fine));
 469    if (!fine) {
 470       free(syncobj);
 471       *out = NULL;
 472       return;
 473    }
 474
 475    static const uint32_t zero = 0;
 476
 477    /* Fences work in terms of iris_fine_fence, but we don't actually have a
 478     * seqno for an imported fence.  So, create a fake one which always
 479     * returns as 'not signaled' so we fall back to using the sync object.
 480     */
 481    fine->seqno = UINT32_MAX;
 482    fine->map = &zero;
 483    fine->syncobj = syncobj;
 484    fine->flags = IRIS_FENCE_END;
 485    pipe_reference_init(&fine->reference, 1);
 486
 487    struct pipe_fence_handle *fence = calloc(1, sizeof(*fence));
 488    if (!fence) {
 489       free(fine);
 490       free(syncobj);
 491       *out = NULL;
 492       return;
 493    }
 494    pipe_reference_init(&fence->ref, 1);
 495    fence->fine[0] = fine;
 496
 497    *out = fence;
 498 }
 499
 500 void
 501 iris_init_screen_fence_functions(struct pipe_screen *screen)
 502 {
 503    screen->fence_reference = iris_fence_reference;
 504    screen->fence_finish = iris_fence_finish;
 505    screen->fence_get_fd = iris_fence_get_fd;
 506 }
 507
 508 void
 509 iris_init_context_fence_functions(struct pipe_context *ctx)
 510 {
 511    ctx->flush = iris_fence_flush;
 512    ctx->create_fence_fd = iris_fence_create_fd;
 513    ctx->fence_server_sync = iris_fence_await;
 514 }