freedreno/ir3: fix load_interpolated_input slot
[mesa.git] / src / freedreno / vulkan / tu_fence.c
1 /*
2 * Copyright © 2019 Google LLC
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
21 * DEALINGS IN THE SOFTWARE.
22 */
23
24 #include "tu_private.h"
25
26 #include <fcntl.h>
27 #include <libsync.h>
28 #include <unistd.h>
29
30 #include "util/os_time.h"
31
32 /**
33 * Internally, a fence can be in one of these states.
34 */
35 enum tu_fence_state
36 {
37 TU_FENCE_STATE_RESET,
38 TU_FENCE_STATE_PENDING,
39 TU_FENCE_STATE_SIGNALED,
40 };
41
42 static enum tu_fence_state
43 tu_fence_get_state(const struct tu_fence *fence)
44 {
45 if (fence->signaled)
46 assert(fence->fd < 0);
47
48 if (fence->signaled)
49 return TU_FENCE_STATE_SIGNALED;
50 else if (fence->fd >= 0)
51 return TU_FENCE_STATE_PENDING;
52 else
53 return TU_FENCE_STATE_RESET;
54 }
55
56 static void
57 tu_fence_set_state(struct tu_fence *fence, enum tu_fence_state state, int fd)
58 {
59 if (fence->fd >= 0)
60 close(fence->fd);
61
62 switch (state) {
63 case TU_FENCE_STATE_RESET:
64 assert(fd < 0);
65 fence->signaled = false;
66 fence->fd = -1;
67 break;
68 case TU_FENCE_STATE_PENDING:
69 assert(fd >= 0);
70 fence->signaled = false;
71 fence->fd = fd;
72 break;
73 case TU_FENCE_STATE_SIGNALED:
74 assert(fd < 0);
75 fence->signaled = true;
76 fence->fd = -1;
77 break;
78 default:
79 unreachable("unknown fence state");
80 break;
81 }
82 }
83
84 void
85 tu_fence_init(struct tu_fence *fence, bool signaled)
86 {
87 fence->signaled = signaled;
88 fence->fd = -1;
89 }
90
91 void
92 tu_fence_finish(struct tu_fence *fence)
93 {
94 if (fence->fd >= 0)
95 close(fence->fd);
96 }
97
98 /**
99 * Update the associated fd of a fence. Ownership of \a fd is transferred to
100 * \a fence.
101 *
102 * This function does not block. \a fence can also be in any state when this
103 * function is called. To be able to do that, the caller must make sure that,
104 * when both the currently associated fd and the new fd are valid, they are on
105 * the same timeline with the new fd being later on the timeline.
106 */
107 void
108 tu_fence_update_fd(struct tu_fence *fence, int fd)
109 {
110 const enum tu_fence_state state =
111 fd >= 0 ? TU_FENCE_STATE_PENDING : TU_FENCE_STATE_SIGNALED;
112 tu_fence_set_state(fence, state, fd);
113 }
114
115 /**
116 * Make a fence a copy of another fence. \a fence must be in the reset state.
117 */
118 void
119 tu_fence_copy(struct tu_fence *fence, const struct tu_fence *src)
120 {
121 assert(tu_fence_get_state(fence) == TU_FENCE_STATE_RESET);
122
123 /* dup src->fd */
124 int fd = -1;
125 if (src->fd >= 0) {
126 fd = fcntl(src->fd, F_DUPFD_CLOEXEC, 0);
127 if (fd < 0) {
128 tu_loge("failed to dup fd %d for fence", src->fd);
129 sync_wait(src->fd, -1);
130 }
131 }
132
133 tu_fence_update_fd(fence, fd);
134 }
135
136 /**
137 * Signal a fence. \a fence must be in the reset state.
138 */
139 void
140 tu_fence_signal(struct tu_fence *fence)
141 {
142 assert(tu_fence_get_state(fence) == TU_FENCE_STATE_RESET);
143 tu_fence_set_state(fence, TU_FENCE_STATE_SIGNALED, -1);
144 }
145
146 /**
147 * Wait until a fence is idle (i.e., not pending).
148 */
149 void
150 tu_fence_wait_idle(struct tu_fence *fence)
151 {
152 if (fence->fd >= 0) {
153 if (sync_wait(fence->fd, -1))
154 tu_loge("sync_wait on fence fd %d failed", fence->fd);
155
156 tu_fence_set_state(fence, TU_FENCE_STATE_SIGNALED, -1);
157 }
158 }
159
160 VkResult
161 tu_CreateFence(VkDevice _device,
162 const VkFenceCreateInfo *pCreateInfo,
163 const VkAllocationCallbacks *pAllocator,
164 VkFence *pFence)
165 {
166 TU_FROM_HANDLE(tu_device, device, _device);
167
168 struct tu_fence *fence =
169 vk_alloc2(&device->alloc, pAllocator, sizeof(*fence), 8,
170 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
171
172 if (!fence)
173 return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
174
175 tu_fence_init(fence, pCreateInfo->flags & VK_FENCE_CREATE_SIGNALED_BIT);
176
177 *pFence = tu_fence_to_handle(fence);
178
179 return VK_SUCCESS;
180 }
181
182 void
183 tu_DestroyFence(VkDevice _device,
184 VkFence _fence,
185 const VkAllocationCallbacks *pAllocator)
186 {
187 TU_FROM_HANDLE(tu_device, device, _device);
188 TU_FROM_HANDLE(tu_fence, fence, _fence);
189
190 if (!fence)
191 return;
192
193 tu_fence_finish(fence);
194
195 vk_free2(&device->alloc, pAllocator, fence);
196 }
197
198 /**
199 * Initialize a pollfd array from fences.
200 */
201 static nfds_t
202 tu_fence_init_poll_fds(uint32_t fence_count,
203 const VkFence *fences,
204 bool wait_all,
205 struct pollfd *fds)
206 {
207 nfds_t nfds = 0;
208 for (uint32_t i = 0; i < fence_count; i++) {
209 TU_FROM_HANDLE(tu_fence, fence, fences[i]);
210
211 if (fence->signaled) {
212 if (wait_all) {
213 /* skip signaled fences */
214 continue;
215 } else {
216 /* no need to poll any fd */
217 nfds = 0;
218 break;
219 }
220 }
221
222 /* negative fds are never ready, which is the desired behavior */
223 fds[nfds].fd = fence->fd;
224 fds[nfds].events = POLLIN;
225 fds[nfds].revents = 0;
226 nfds++;
227 }
228
229 return nfds;
230 }
231
232 /**
233 * Translate timeout from nanoseconds to milliseconds for poll().
234 */
235 static int
236 tu_fence_get_poll_timeout(uint64_t timeout_ns)
237 {
238 const uint64_t ns_per_ms = 1000 * 1000;
239 uint64_t timeout_ms = timeout_ns / ns_per_ms;
240
241 /* round up if needed */
242 if (timeout_ns - timeout_ms * ns_per_ms >= ns_per_ms / 2)
243 timeout_ms++;
244
245 return timeout_ms < INT_MAX ? timeout_ms : INT_MAX;
246 }
247
248 /**
249 * Poll a pollfd array.
250 */
251 static VkResult
252 tu_fence_poll_fds(struct pollfd *fds, nfds_t nfds, uint64_t *timeout_ns)
253 {
254 while (true) {
255 /* poll */
256 uint64_t duration = os_time_get_nano();
257 int ret = poll(fds, nfds, tu_fence_get_poll_timeout(*timeout_ns));
258 duration = os_time_get_nano() - duration;
259
260 /* update timeout_ns */
261 if (*timeout_ns > duration)
262 *timeout_ns -= duration;
263 else
264 *timeout_ns = 0;
265
266 if (ret > 0) {
267 return VK_SUCCESS;
268 } else if (ret == 0) {
269 if (!*timeout_ns)
270 return VK_TIMEOUT;
271 } else if (errno != EINTR && errno != EAGAIN) {
272 return VK_ERROR_OUT_OF_HOST_MEMORY;
273 }
274 }
275 }
276
277 /**
278 * Update a pollfd array and the fence states. This should be called after a
279 * successful call to tu_fence_poll_fds.
280 */
281 static nfds_t
282 tu_fence_update_fences_and_poll_fds(uint32_t fence_count,
283 const VkFence *fences,
284 bool wait_all,
285 struct pollfd *fds)
286 {
287 uint32_t nfds = 0;
288 uint32_t fds_idx = 0;
289 for (uint32_t i = 0; i < fence_count; i++) {
290 TU_FROM_HANDLE(tu_fence, fence, fences[i]);
291
292 /* no signaled fence in fds */
293 if (fence->signaled)
294 continue;
295
296 /* fds[fds_idx] corresponds to fences[i] */
297 assert(fence->fd == fds[fds_idx].fd);
298
299 assert(nfds <= fds_idx && fds_idx <= i);
300
301 /* fd is ready (errors are treated as ready) */
302 if (fds[fds_idx].revents) {
303 tu_fence_set_state(fence, TU_FENCE_STATE_SIGNALED, -1);
304 } else if (wait_all) {
305 /* add to fds again for another poll */
306 fds[nfds].fd = fence->fd;
307 fds[nfds].events = POLLIN;
308 fds[nfds].revents = 0;
309 nfds++;
310 }
311
312 fds_idx++;
313 }
314
315 return nfds;
316 }
317
318 VkResult
319 tu_WaitForFences(VkDevice _device,
320 uint32_t fenceCount,
321 const VkFence *pFences,
322 VkBool32 waitAll,
323 uint64_t timeout)
324 {
325 TU_FROM_HANDLE(tu_device, device, _device);
326
327 /* add a simpler path for when fenceCount == 1? */
328
329 struct pollfd stack_fds[8];
330 struct pollfd *fds = stack_fds;
331 if (fenceCount > ARRAY_SIZE(stack_fds)) {
332 fds = vk_alloc(&device->alloc, sizeof(*fds) * fenceCount, 8,
333 VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
334 if (!fds)
335 return VK_ERROR_OUT_OF_HOST_MEMORY;
336 }
337
338 /* set up pollfd array and start polling */
339 nfds_t nfds = tu_fence_init_poll_fds(fenceCount, pFences, waitAll, fds);
340 VkResult result = VK_SUCCESS;
341 while (nfds) {
342 result = tu_fence_poll_fds(fds, nfds, &timeout);
343 if (result != VK_SUCCESS)
344 break;
345 nfds = tu_fence_update_fences_and_poll_fds(fenceCount, pFences, waitAll,
346 fds);
347 }
348
349 if (fds != stack_fds)
350 vk_free(&device->alloc, fds);
351
352 return result;
353 }
354
355 VkResult
356 tu_ResetFences(VkDevice _device, uint32_t fenceCount, const VkFence *pFences)
357 {
358 for (unsigned i = 0; i < fenceCount; ++i) {
359 TU_FROM_HANDLE(tu_fence, fence, pFences[i]);
360 assert(tu_fence_get_state(fence) != TU_FENCE_STATE_PENDING);
361 tu_fence_set_state(fence, TU_FENCE_STATE_RESET, -1);
362 }
363
364 return VK_SUCCESS;
365 }
366
367 VkResult
368 tu_GetFenceStatus(VkDevice _device, VkFence _fence)
369 {
370 TU_FROM_HANDLE(tu_fence, fence, _fence);
371
372 if (fence->fd >= 0) {
373 int err = sync_wait(fence->fd, 0);
374 if (!err)
375 tu_fence_set_state(fence, TU_FENCE_STATE_SIGNALED, -1);
376 else if (err && errno != ETIME)
377 return VK_ERROR_OUT_OF_HOST_MEMORY;
378 }
379
380 return fence->signaled ? VK_SUCCESS : VK_NOT_READY;
381 }