1/* 2 * Copyright © 2008 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 * Authors: 24 * Eric Anholt <eric@anholt.net> 25 * 26 */ 27 28/** 29 * \file 30 * \brief Support for GL_ARB_sync and EGL_KHR_fence_sync. 31 * 32 * GL_ARB_sync is implemented by flushing the current batchbuffer and keeping a 33 * reference on it. We can then check for completion or wait for completion 34 * using the normal buffer object mechanisms. This does mean that if an 35 * application is using many sync objects, it will emit small batchbuffers 36 * which may end up being a significant overhead. In other tests of removing 37 * gratuitous batchbuffer syncs in Mesa, it hasn't appeared to be a significant 38 * performance bottleneck, though. 39 */ 40 41#include "main/imports.h" 42 43#include "intel_context.h" 44#include "intel_batchbuffer.h" 45#include "intel_reg.h" 46 47struct intel_fence { 48 struct intel_context *intel; 49 /** The fence waits for completion of this batch. */ 50 drm_intel_bo *batch_bo; 51 52 mtx_t mutex; 53 bool signalled; 54}; 55 56struct intel_gl_sync_object { 57 struct gl_sync_object Base; 58 struct intel_fence fence; 59}; 60 61static void 62intel_fence_finish(struct intel_fence *fence) 63{ 64 if (fence->batch_bo) 65 drm_intel_bo_unreference(fence->batch_bo); 66} 67 68static void 69intel_fence_insert(struct intel_context *intel, struct intel_fence *fence) 70{ 71 assert(!fence->batch_bo); 72 assert(!fence->signalled); 73 74 intel_batchbuffer_emit_mi_flush(intel); 75 fence->batch_bo = intel->batch.bo; 76 drm_intel_bo_reference(fence->batch_bo); 77 intel_batchbuffer_flush(intel); 78} 79 80static bool 81intel_fence_has_completed_locked(struct intel_fence *fence) 82{ 83 if (fence->signalled) 84 return true; 85 86 if (fence->batch_bo && !drm_intel_bo_busy(fence->batch_bo)) { 87 drm_intel_bo_unreference(fence->batch_bo); 88 fence->batch_bo = NULL; 89 fence->signalled = true; 90 return true; 91 } 92 93 return false; 94} 95 96static bool 97intel_fence_has_completed(struct intel_fence *fence) 98{ 99 bool ret; 100 101 mtx_lock(&fence->mutex); 102 ret = intel_fence_has_completed_locked(fence); 103 mtx_unlock(&fence->mutex); 104 105 return ret; 106} 107 108static bool 109intel_fence_client_wait_locked(struct intel_context *intel, struct intel_fence *fence, 110 uint64_t timeout) 111{ 112 if (fence->signalled) 113 return true; 114 115 assert(fence->batch_bo); 116 117 /* DRM_IOCTL_I915_GEM_WAIT uses a signed 64 bit timeout and returns 118 * immediately for timeouts <= 0. The best we can do is to clamp the 119 * timeout to INT64_MAX. This limits the maximum timeout from 584 years to 120 * 292 years - likely not a big deal. 121 */ 122 if (timeout > INT64_MAX) 123 timeout = INT64_MAX; 124 125 if (drm_intel_gem_bo_wait(fence->batch_bo, timeout) != 0) 126 return false; 127 128 fence->signalled = true; 129 drm_intel_bo_unreference(fence->batch_bo); 130 fence->batch_bo = NULL; 131 132 return true; 133} 134 135/** 136 * Return true if the function successfully signals or has already signalled. 137 * (This matches the behavior expected from __DRI2fence::client_wait_sync). 138 */ 139static bool 140intel_fence_client_wait(struct intel_context *intel, struct intel_fence *fence, 141 uint64_t timeout) 142{ 143 bool ret; 144 145 mtx_lock(&fence->mutex); 146 ret = intel_fence_client_wait_locked(intel, fence, timeout); 147 mtx_unlock(&fence->mutex); 148 149 return ret; 150} 151 152static void 153intel_fence_server_wait(struct intel_context *intel, struct intel_fence *fence) 154{ 155 /* We have nothing to do for WaitSync. Our GL command stream is sequential, 156 * so given that the sync object has already flushed the batchbuffer, any 157 * batchbuffers coming after this waitsync will naturally not occur until 158 * the previous one is done. 159 */ 160} 161 162static struct gl_sync_object * 163intel_gl_new_sync_object(struct gl_context *ctx) 164{ 165 struct intel_gl_sync_object *sync; 166 167 sync = calloc(1, sizeof(*sync)); 168 if (!sync) 169 return NULL; 170 171 return &sync->Base; 172} 173 174static void 175intel_gl_delete_sync_object(struct gl_context *ctx, struct gl_sync_object *s) 176{ 177 struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s; 178 179 intel_fence_finish(&sync->fence); 180 free(sync); 181} 182 183static void 184intel_gl_fence_sync(struct gl_context *ctx, struct gl_sync_object *s, 185 GLenum condition, GLbitfield flags) 186{ 187 struct intel_context *intel = intel_context(ctx); 188 struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s; 189 190 intel_fence_insert(intel, &sync->fence); 191} 192 193static void 194intel_gl_client_wait_sync(struct gl_context *ctx, struct gl_sync_object *s, 195 GLbitfield flags, GLuint64 timeout) 196{ 197 struct intel_context *intel = intel_context(ctx); 198 struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s; 199 200 if (intel_fence_client_wait(intel, &sync->fence, timeout)) 201 s->StatusFlag = 1; 202} 203 204static void 205intel_gl_server_wait_sync(struct gl_context *ctx, struct gl_sync_object *s, 206 GLbitfield flags, GLuint64 timeout) 207{ 208 struct intel_context *intel = intel_context(ctx); 209 struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s; 210 211 intel_fence_server_wait(intel, &sync->fence); 212} 213 214static void 215intel_gl_check_sync(struct gl_context *ctx, struct gl_sync_object *s) 216{ 217 struct intel_gl_sync_object *sync = (struct intel_gl_sync_object *)s; 218 219 if (intel_fence_has_completed(&sync->fence)) 220 s->StatusFlag = 1; 221} 222 223void 224intel_init_syncobj_functions(struct dd_function_table *functions) 225{ 226 functions->NewSyncObject = intel_gl_new_sync_object; 227 functions->DeleteSyncObject = intel_gl_delete_sync_object; 228 functions->FenceSync = intel_gl_fence_sync; 229 functions->CheckSync = intel_gl_check_sync; 230 functions->ClientWaitSync = intel_gl_client_wait_sync; 231 functions->ServerWaitSync = intel_gl_server_wait_sync; 232} 233 234static void * 235intel_dri_create_fence(__DRIcontext *ctx) 236{ 237 struct intel_context *intel = ctx->driverPrivate; 238 struct intel_fence *fence; 239 240 fence = calloc(1, sizeof(*fence)); 241 if (!fence) 242 return NULL; 243 244 mtx_init(&fence->mutex, mtx_plain); 245 fence->intel = intel; 246 intel_fence_insert(intel, fence); 247 248 return fence; 249} 250 251static void 252intel_dri_destroy_fence(__DRIscreen *screen, void *driver_fence) 253{ 254 struct intel_fence *fence = driver_fence; 255 256 intel_fence_finish(fence); 257 free(fence); 258} 259 260static GLboolean 261intel_dri_client_wait_sync(__DRIcontext *ctx, void *driver_fence, unsigned flags, 262 uint64_t timeout) 263{ 264 struct intel_fence *fence = driver_fence; 265 266 return intel_fence_client_wait(fence->intel, fence, timeout); 267} 268 269static void 270intel_dri_server_wait_sync(__DRIcontext *ctx, void *driver_fence, unsigned flags) 271{ 272 struct intel_fence *fence = driver_fence; 273 274 /* We might be called here with a NULL fence as a result of WaitSyncKHR 275 * on a EGL_KHR_reusable_sync fence. Nothing to do here in such case. 276 */ 277 if (!fence) 278 return; 279 280 intel_fence_server_wait(fence->intel, fence); 281} 282 283const __DRI2fenceExtension intelFenceExtension = { 284 .base = { __DRI2_FENCE, 1 }, 285 286 .create_fence = intel_dri_create_fence, 287 .destroy_fence = intel_dri_destroy_fence, 288 .client_wait_sync = intel_dri_client_wait_sync, 289 .server_wait_sync = intel_dri_server_wait_sync, 290 .get_fence_from_cl_event = NULL, 291}; 292