Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 1 | /* |
| 2 | * Copyright 2009 Jerome Glisse. |
| 3 | * All Rights Reserved. |
| 4 | * |
| 5 | * Permission is hereby granted, free of charge, to any person obtaining a |
| 6 | * copy of this software and associated documentation files (the |
| 7 | * "Software"), to deal in the Software without restriction, including |
| 8 | * without limitation the rights to use, copy, modify, merge, publish, |
| 9 | * distribute, sub license, and/or sell copies of the Software, and to |
| 10 | * permit persons to whom the Software is furnished to do so, subject to |
| 11 | * the following conditions: |
| 12 | * |
| 13 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 14 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 15 | * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL |
| 16 | * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, |
| 17 | * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR |
| 18 | * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE |
| 19 | * USE OR OTHER DEALINGS IN THE SOFTWARE. |
| 20 | * |
| 21 | * The above copyright notice and this permission notice (including the |
| 22 | * next paragraph) shall be included in all copies or substantial portions |
| 23 | * of the Software. |
| 24 | * |
| 25 | */ |
| 26 | /* |
| 27 | * Authors: |
| 28 | * Jerome Glisse <glisse@freedesktop.org> |
| 29 | * Dave Airlie |
| 30 | */ |
| 31 | #include <linux/seq_file.h> |
| 32 | #include <linux/atomic.h> |
| 33 | #include <linux/wait.h> |
| 34 | #include <linux/kref.h> |
| 35 | #include <linux/slab.h> |
| 36 | #include <linux/firmware.h> |
| 37 | #include <drm/drmP.h> |
| 38 | #include "amdgpu.h" |
| 39 | #include "amdgpu_trace.h" |
| 40 | |
| 41 | /* |
| 42 | * Fences |
| 43 | * Fences mark an event in the GPUs pipeline and are used |
| 44 | * for GPU/CPU synchronization. When the fence is written, |
| 45 | * it is expected that all buffers associated with that fence |
| 46 | * are no longer in use by the associated ring on the GPU and |
| 47 | * that the the relevant GPU caches have been flushed. |
| 48 | */ |
| 49 | |
| 50 | /** |
| 51 | * amdgpu_fence_write - write a fence value |
| 52 | * |
| 53 | * @ring: ring the fence is associated with |
| 54 | * @seq: sequence number to write |
| 55 | * |
| 56 | * Writes a fence value to memory (all asics). |
| 57 | */ |
| 58 | static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq) |
| 59 | { |
| 60 | struct amdgpu_fence_driver *drv = &ring->fence_drv; |
| 61 | |
| 62 | if (drv->cpu_addr) |
| 63 | *drv->cpu_addr = cpu_to_le32(seq); |
| 64 | } |
| 65 | |
| 66 | /** |
| 67 | * amdgpu_fence_read - read a fence value |
| 68 | * |
| 69 | * @ring: ring the fence is associated with |
| 70 | * |
| 71 | * Reads a fence value from memory (all asics). |
| 72 | * Returns the value of the fence read from memory. |
| 73 | */ |
| 74 | static u32 amdgpu_fence_read(struct amdgpu_ring *ring) |
| 75 | { |
| 76 | struct amdgpu_fence_driver *drv = &ring->fence_drv; |
| 77 | u32 seq = 0; |
| 78 | |
| 79 | if (drv->cpu_addr) |
| 80 | seq = le32_to_cpu(*drv->cpu_addr); |
| 81 | else |
| 82 | seq = lower_32_bits(atomic64_read(&drv->last_seq)); |
| 83 | |
| 84 | return seq; |
| 85 | } |
| 86 | |
| 87 | /** |
| 88 | * amdgpu_fence_schedule_check - schedule lockup check |
| 89 | * |
| 90 | * @ring: pointer to struct amdgpu_ring |
| 91 | * |
| 92 | * Queues a delayed work item to check for lockups. |
| 93 | */ |
| 94 | static void amdgpu_fence_schedule_check(struct amdgpu_ring *ring) |
| 95 | { |
| 96 | /* |
| 97 | * Do not reset the timer here with mod_delayed_work, |
| 98 | * this can livelock in an interaction with TTM delayed destroy. |
| 99 | */ |
| 100 | queue_delayed_work(system_power_efficient_wq, |
| 101 | &ring->fence_drv.lockup_work, |
| 102 | AMDGPU_FENCE_JIFFIES_TIMEOUT); |
| 103 | } |
| 104 | |
| 105 | /** |
| 106 | * amdgpu_fence_emit - emit a fence on the requested ring |
| 107 | * |
| 108 | * @ring: ring the fence is associated with |
| 109 | * @owner: creator of the fence |
| 110 | * @fence: amdgpu fence object |
| 111 | * |
| 112 | * Emits a fence command on the requested ring (all asics). |
| 113 | * Returns 0 on success, -ENOMEM on failure. |
| 114 | */ |
| 115 | int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner, |
| 116 | struct amdgpu_fence **fence) |
| 117 | { |
| 118 | struct amdgpu_device *adev = ring->adev; |
| 119 | |
| 120 | /* we are protected by the ring emission mutex */ |
| 121 | *fence = kmalloc(sizeof(struct amdgpu_fence), GFP_KERNEL); |
| 122 | if ((*fence) == NULL) { |
| 123 | return -ENOMEM; |
| 124 | } |
| 125 | (*fence)->seq = ++ring->fence_drv.sync_seq[ring->idx]; |
| 126 | (*fence)->ring = ring; |
| 127 | (*fence)->owner = owner; |
| 128 | fence_init(&(*fence)->base, &amdgpu_fence_ops, |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 129 | &ring->fence_drv.fence_queue.lock, |
| 130 | adev->fence_context + ring->idx, |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 131 | (*fence)->seq); |
Chunming Zhou | 890ee23 | 2015-06-01 14:35:03 +0800 | [diff] [blame] | 132 | amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, |
| 133 | (*fence)->seq, |
| 134 | AMDGPU_FENCE_FLAG_INT); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 135 | trace_amdgpu_fence_emit(ring->adev->ddev, ring->idx, (*fence)->seq); |
| 136 | return 0; |
| 137 | } |
| 138 | |
| 139 | /** |
| 140 | * amdgpu_fence_check_signaled - callback from fence_queue |
| 141 | * |
| 142 | * this function is called with fence_queue lock held, which is also used |
| 143 | * for the fence locking itself, so unlocked variants are used for |
| 144 | * fence_signal, and remove_wait_queue. |
| 145 | */ |
| 146 | static int amdgpu_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key) |
| 147 | { |
| 148 | struct amdgpu_fence *fence; |
| 149 | struct amdgpu_device *adev; |
| 150 | u64 seq; |
| 151 | int ret; |
| 152 | |
| 153 | fence = container_of(wait, struct amdgpu_fence, fence_wake); |
| 154 | adev = fence->ring->adev; |
| 155 | |
| 156 | /* |
| 157 | * We cannot use amdgpu_fence_process here because we're already |
| 158 | * in the waitqueue, in a call from wake_up_all. |
| 159 | */ |
| 160 | seq = atomic64_read(&fence->ring->fence_drv.last_seq); |
| 161 | if (seq >= fence->seq) { |
| 162 | ret = fence_signal_locked(&fence->base); |
| 163 | if (!ret) |
| 164 | FENCE_TRACE(&fence->base, "signaled from irq context\n"); |
| 165 | else |
| 166 | FENCE_TRACE(&fence->base, "was already signaled\n"); |
| 167 | |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 168 | __remove_wait_queue(&fence->ring->fence_drv.fence_queue, &fence->fence_wake); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 169 | fence_put(&fence->base); |
| 170 | } else |
| 171 | FENCE_TRACE(&fence->base, "pending\n"); |
| 172 | return 0; |
| 173 | } |
| 174 | |
| 175 | /** |
| 176 | * amdgpu_fence_activity - check for fence activity |
| 177 | * |
| 178 | * @ring: pointer to struct amdgpu_ring |
| 179 | * |
| 180 | * Checks the current fence value and calculates the last |
| 181 | * signalled fence value. Returns true if activity occured |
| 182 | * on the ring, and the fence_queue should be waken up. |
| 183 | */ |
| 184 | static bool amdgpu_fence_activity(struct amdgpu_ring *ring) |
| 185 | { |
| 186 | uint64_t seq, last_seq, last_emitted; |
| 187 | unsigned count_loop = 0; |
| 188 | bool wake = false; |
| 189 | |
| 190 | /* Note there is a scenario here for an infinite loop but it's |
| 191 | * very unlikely to happen. For it to happen, the current polling |
| 192 | * process need to be interrupted by another process and another |
| 193 | * process needs to update the last_seq btw the atomic read and |
| 194 | * xchg of the current process. |
| 195 | * |
| 196 | * More over for this to go in infinite loop there need to be |
Jammy Zhou | 86c2b79 | 2015-05-13 22:52:42 +0800 | [diff] [blame] | 197 | * continuously new fence signaled ie amdgpu_fence_read needs |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 198 | * to return a different value each time for both the currently |
| 199 | * polling process and the other process that xchg the last_seq |
| 200 | * btw atomic read and xchg of the current process. And the |
| 201 | * value the other process set as last seq must be higher than |
| 202 | * the seq value we just read. Which means that current process |
Jammy Zhou | 86c2b79 | 2015-05-13 22:52:42 +0800 | [diff] [blame] | 203 | * need to be interrupted after amdgpu_fence_read and before |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 204 | * atomic xchg. |
| 205 | * |
| 206 | * To be even more safe we count the number of time we loop and |
| 207 | * we bail after 10 loop just accepting the fact that we might |
| 208 | * have temporarly set the last_seq not to the true real last |
| 209 | * seq but to an older one. |
| 210 | */ |
| 211 | last_seq = atomic64_read(&ring->fence_drv.last_seq); |
| 212 | do { |
| 213 | last_emitted = ring->fence_drv.sync_seq[ring->idx]; |
| 214 | seq = amdgpu_fence_read(ring); |
| 215 | seq |= last_seq & 0xffffffff00000000LL; |
| 216 | if (seq < last_seq) { |
| 217 | seq &= 0xffffffff; |
| 218 | seq |= last_emitted & 0xffffffff00000000LL; |
| 219 | } |
| 220 | |
| 221 | if (seq <= last_seq || seq > last_emitted) { |
| 222 | break; |
| 223 | } |
| 224 | /* If we loop over we don't want to return without |
| 225 | * checking if a fence is signaled as it means that the |
| 226 | * seq we just read is different from the previous on. |
| 227 | */ |
| 228 | wake = true; |
| 229 | last_seq = seq; |
| 230 | if ((count_loop++) > 10) { |
| 231 | /* We looped over too many time leave with the |
| 232 | * fact that we might have set an older fence |
| 233 | * seq then the current real last seq as signaled |
| 234 | * by the hw. |
| 235 | */ |
| 236 | break; |
| 237 | } |
| 238 | } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq); |
| 239 | |
| 240 | if (seq < last_emitted) |
| 241 | amdgpu_fence_schedule_check(ring); |
| 242 | |
| 243 | return wake; |
| 244 | } |
| 245 | |
| 246 | /** |
| 247 | * amdgpu_fence_check_lockup - check for hardware lockup |
| 248 | * |
| 249 | * @work: delayed work item |
| 250 | * |
| 251 | * Checks for fence activity and if there is none probe |
| 252 | * the hardware if a lockup occured. |
| 253 | */ |
| 254 | static void amdgpu_fence_check_lockup(struct work_struct *work) |
| 255 | { |
| 256 | struct amdgpu_fence_driver *fence_drv; |
| 257 | struct amdgpu_ring *ring; |
| 258 | |
| 259 | fence_drv = container_of(work, struct amdgpu_fence_driver, |
| 260 | lockup_work.work); |
| 261 | ring = fence_drv->ring; |
| 262 | |
Christian König | 0c418f1 | 2015-09-01 15:13:53 +0200 | [diff] [blame] | 263 | if (amdgpu_fence_activity(ring)) |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 264 | wake_up_all(&ring->fence_drv.fence_queue); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 265 | } |
| 266 | |
| 267 | /** |
| 268 | * amdgpu_fence_process - process a fence |
| 269 | * |
| 270 | * @adev: amdgpu_device pointer |
| 271 | * @ring: ring index the fence is associated with |
| 272 | * |
| 273 | * Checks the current fence value and wakes the fence queue |
| 274 | * if the sequence number has increased (all asics). |
| 275 | */ |
| 276 | void amdgpu_fence_process(struct amdgpu_ring *ring) |
| 277 | { |
Christian König | 68ed3de | 2015-08-07 15:57:21 +0200 | [diff] [blame] | 278 | if (amdgpu_fence_activity(ring)) |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 279 | wake_up_all(&ring->fence_drv.fence_queue); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 280 | } |
| 281 | |
| 282 | /** |
| 283 | * amdgpu_fence_seq_signaled - check if a fence sequence number has signaled |
| 284 | * |
| 285 | * @ring: ring the fence is associated with |
| 286 | * @seq: sequence number |
| 287 | * |
| 288 | * Check if the last signaled fence sequnce number is >= the requested |
| 289 | * sequence number (all asics). |
| 290 | * Returns true if the fence has signaled (current fence value |
| 291 | * is >= requested value) or false if it has not (current fence |
| 292 | * value is < the requested value. Helper function for |
| 293 | * amdgpu_fence_signaled(). |
| 294 | */ |
| 295 | static bool amdgpu_fence_seq_signaled(struct amdgpu_ring *ring, u64 seq) |
| 296 | { |
| 297 | if (atomic64_read(&ring->fence_drv.last_seq) >= seq) |
| 298 | return true; |
| 299 | |
| 300 | /* poll new last sequence at least once */ |
| 301 | amdgpu_fence_process(ring); |
| 302 | if (atomic64_read(&ring->fence_drv.last_seq) >= seq) |
| 303 | return true; |
| 304 | |
| 305 | return false; |
| 306 | } |
| 307 | |
| 308 | static bool amdgpu_fence_is_signaled(struct fence *f) |
| 309 | { |
| 310 | struct amdgpu_fence *fence = to_amdgpu_fence(f); |
| 311 | struct amdgpu_ring *ring = fence->ring; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 312 | |
| 313 | if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) |
| 314 | return true; |
| 315 | |
Christian König | 0c418f1 | 2015-09-01 15:13:53 +0200 | [diff] [blame] | 316 | amdgpu_fence_process(ring); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 317 | |
Christian König | 0c418f1 | 2015-09-01 15:13:53 +0200 | [diff] [blame] | 318 | if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) |
| 319 | return true; |
| 320 | |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 321 | return false; |
| 322 | } |
| 323 | |
| 324 | /** |
| 325 | * amdgpu_fence_enable_signaling - enable signalling on fence |
| 326 | * @fence: fence |
| 327 | * |
| 328 | * This function is called with fence_queue lock held, and adds a callback |
| 329 | * to fence_queue that checks if this fence is signaled, and if so it |
| 330 | * signals the fence and removes itself. |
| 331 | */ |
| 332 | static bool amdgpu_fence_enable_signaling(struct fence *f) |
| 333 | { |
| 334 | struct amdgpu_fence *fence = to_amdgpu_fence(f); |
| 335 | struct amdgpu_ring *ring = fence->ring; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 336 | |
| 337 | if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) |
| 338 | return false; |
| 339 | |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 340 | fence->fence_wake.flags = 0; |
| 341 | fence->fence_wake.private = NULL; |
| 342 | fence->fence_wake.func = amdgpu_fence_check_signaled; |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 343 | __add_wait_queue(&ring->fence_drv.fence_queue, &fence->fence_wake); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 344 | fence_get(f); |
| 345 | FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx); |
| 346 | return true; |
| 347 | } |
| 348 | |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 349 | /* |
| 350 | * amdgpu_ring_wait_seq_timeout - wait for seq of the specific ring to signal |
| 351 | * @ring: ring to wait on for the seq number |
| 352 | * @seq: seq number wait for |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 353 | * |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 354 | * return value: |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 355 | * 0: seq signaled, and gpu not hang |
| 356 | * -EDEADL: GPU hang detected |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 357 | * -EINVAL: some paramter is not valid |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 358 | */ |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 359 | static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 360 | { |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 361 | bool signaled = false; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 362 | |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 363 | BUG_ON(!ring); |
| 364 | if (seq > ring->fence_drv.sync_seq[ring->idx]) |
| 365 | return -EINVAL; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 366 | |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 367 | if (atomic64_read(&ring->fence_drv.last_seq) >= seq) |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 368 | return 0; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 369 | |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 370 | wait_event(ring->fence_drv.fence_queue, ( |
Christian König | b7e4dad | 2015-09-01 10:50:26 +0200 | [diff] [blame] | 371 | (signaled = amdgpu_fence_seq_signaled(ring, seq)))); |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 372 | |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 373 | if (signaled) |
| 374 | return 0; |
| 375 | else |
| 376 | return -EDEADLK; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 377 | } |
| 378 | |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 379 | /** |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 380 | * amdgpu_fence_wait_next - wait for the next fence to signal |
| 381 | * |
| 382 | * @adev: amdgpu device pointer |
| 383 | * @ring: ring index the fence is associated with |
| 384 | * |
| 385 | * Wait for the next fence on the requested ring to signal (all asics). |
| 386 | * Returns 0 if the next fence has passed, error for all other cases. |
| 387 | * Caller must hold ring lock. |
| 388 | */ |
| 389 | int amdgpu_fence_wait_next(struct amdgpu_ring *ring) |
| 390 | { |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 391 | uint64_t seq = atomic64_read(&ring->fence_drv.last_seq) + 1ULL; |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 392 | |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 393 | if (seq >= ring->fence_drv.sync_seq[ring->idx]) |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 394 | return -ENOENT; |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 395 | |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 396 | return amdgpu_fence_ring_wait_seq(ring, seq); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 397 | } |
| 398 | |
| 399 | /** |
| 400 | * amdgpu_fence_wait_empty - wait for all fences to signal |
| 401 | * |
| 402 | * @adev: amdgpu device pointer |
| 403 | * @ring: ring index the fence is associated with |
| 404 | * |
| 405 | * Wait for all fences on the requested ring to signal (all asics). |
| 406 | * Returns 0 if the fences have passed, error for all other cases. |
| 407 | * Caller must hold ring lock. |
| 408 | */ |
| 409 | int amdgpu_fence_wait_empty(struct amdgpu_ring *ring) |
| 410 | { |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 411 | uint64_t seq = ring->fence_drv.sync_seq[ring->idx]; |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 412 | |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 413 | if (!seq) |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 414 | return 0; |
| 415 | |
Christian König | 00d2a2b | 2015-08-07 16:15:36 +0200 | [diff] [blame] | 416 | return amdgpu_fence_ring_wait_seq(ring, seq); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 417 | } |
| 418 | |
| 419 | /** |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 420 | * amdgpu_fence_count_emitted - get the count of emitted fences |
| 421 | * |
| 422 | * @ring: ring the fence is associated with |
| 423 | * |
| 424 | * Get the number of fences emitted on the requested ring (all asics). |
| 425 | * Returns the number of emitted fences on the ring. Used by the |
| 426 | * dynpm code to ring track activity. |
| 427 | */ |
| 428 | unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) |
| 429 | { |
| 430 | uint64_t emitted; |
| 431 | |
| 432 | /* We are not protected by ring lock when reading the last sequence |
| 433 | * but it's ok to report slightly wrong fence count here. |
| 434 | */ |
| 435 | amdgpu_fence_process(ring); |
| 436 | emitted = ring->fence_drv.sync_seq[ring->idx] |
| 437 | - atomic64_read(&ring->fence_drv.last_seq); |
| 438 | /* to avoid 32bits warp around */ |
| 439 | if (emitted > 0x10000000) |
| 440 | emitted = 0x10000000; |
| 441 | |
| 442 | return (unsigned)emitted; |
| 443 | } |
| 444 | |
| 445 | /** |
| 446 | * amdgpu_fence_need_sync - do we need a semaphore |
| 447 | * |
| 448 | * @fence: amdgpu fence object |
| 449 | * @dst_ring: which ring to check against |
| 450 | * |
| 451 | * Check if the fence needs to be synced against another ring |
| 452 | * (all asics). If so, we need to emit a semaphore. |
| 453 | * Returns true if we need to sync with another ring, false if |
| 454 | * not. |
| 455 | */ |
| 456 | bool amdgpu_fence_need_sync(struct amdgpu_fence *fence, |
| 457 | struct amdgpu_ring *dst_ring) |
| 458 | { |
| 459 | struct amdgpu_fence_driver *fdrv; |
| 460 | |
| 461 | if (!fence) |
| 462 | return false; |
| 463 | |
| 464 | if (fence->ring == dst_ring) |
| 465 | return false; |
| 466 | |
| 467 | /* we are protected by the ring mutex */ |
| 468 | fdrv = &dst_ring->fence_drv; |
| 469 | if (fence->seq <= fdrv->sync_seq[fence->ring->idx]) |
| 470 | return false; |
| 471 | |
| 472 | return true; |
| 473 | } |
| 474 | |
| 475 | /** |
| 476 | * amdgpu_fence_note_sync - record the sync point |
| 477 | * |
| 478 | * @fence: amdgpu fence object |
| 479 | * @dst_ring: which ring to check against |
| 480 | * |
| 481 | * Note the sequence number at which point the fence will |
| 482 | * be synced with the requested ring (all asics). |
| 483 | */ |
| 484 | void amdgpu_fence_note_sync(struct amdgpu_fence *fence, |
| 485 | struct amdgpu_ring *dst_ring) |
| 486 | { |
| 487 | struct amdgpu_fence_driver *dst, *src; |
| 488 | unsigned i; |
| 489 | |
| 490 | if (!fence) |
| 491 | return; |
| 492 | |
| 493 | if (fence->ring == dst_ring) |
| 494 | return; |
| 495 | |
| 496 | /* we are protected by the ring mutex */ |
| 497 | src = &fence->ring->fence_drv; |
| 498 | dst = &dst_ring->fence_drv; |
| 499 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
| 500 | if (i == dst_ring->idx) |
| 501 | continue; |
| 502 | |
| 503 | dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]); |
| 504 | } |
| 505 | } |
| 506 | |
| 507 | /** |
| 508 | * amdgpu_fence_driver_start_ring - make the fence driver |
| 509 | * ready for use on the requested ring. |
| 510 | * |
| 511 | * @ring: ring to start the fence driver on |
| 512 | * @irq_src: interrupt source to use for this ring |
| 513 | * @irq_type: interrupt type to use for this ring |
| 514 | * |
| 515 | * Make the fence driver ready for processing (all asics). |
| 516 | * Not all asics have all rings, so each asic will only |
| 517 | * start the fence driver on the rings it has. |
| 518 | * Returns 0 for success, errors for failure. |
| 519 | */ |
| 520 | int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, |
| 521 | struct amdgpu_irq_src *irq_src, |
| 522 | unsigned irq_type) |
| 523 | { |
| 524 | struct amdgpu_device *adev = ring->adev; |
| 525 | uint64_t index; |
| 526 | |
| 527 | if (ring != &adev->uvd.ring) { |
| 528 | ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs]; |
| 529 | ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4); |
| 530 | } else { |
| 531 | /* put fence directly behind firmware */ |
| 532 | index = ALIGN(adev->uvd.fw->size, 8); |
| 533 | ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index; |
| 534 | ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index; |
| 535 | } |
| 536 | amdgpu_fence_write(ring, atomic64_read(&ring->fence_drv.last_seq)); |
Chunming Zhou | c6a4079 | 2015-06-01 14:14:32 +0800 | [diff] [blame] | 537 | amdgpu_irq_get(adev, irq_src, irq_type); |
| 538 | |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 539 | ring->fence_drv.irq_src = irq_src; |
| 540 | ring->fence_drv.irq_type = irq_type; |
Chunming Zhou | c6a4079 | 2015-06-01 14:14:32 +0800 | [diff] [blame] | 541 | ring->fence_drv.initialized = true; |
| 542 | |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 543 | dev_info(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, " |
| 544 | "cpu addr 0x%p\n", ring->idx, |
| 545 | ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr); |
| 546 | return 0; |
| 547 | } |
| 548 | |
| 549 | /** |
| 550 | * amdgpu_fence_driver_init_ring - init the fence driver |
| 551 | * for the requested ring. |
| 552 | * |
| 553 | * @ring: ring to init the fence driver on |
| 554 | * |
| 555 | * Init the fence driver for the requested ring (all asics). |
| 556 | * Helper function for amdgpu_fence_driver_init(). |
| 557 | */ |
Christian König | 4f839a2 | 2015-09-08 20:22:31 +0200 | [diff] [blame] | 558 | int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 559 | { |
Christian König | 4f839a2 | 2015-09-08 20:22:31 +0200 | [diff] [blame] | 560 | int i, r; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 561 | |
| 562 | ring->fence_drv.cpu_addr = NULL; |
| 563 | ring->fence_drv.gpu_addr = 0; |
| 564 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) |
| 565 | ring->fence_drv.sync_seq[i] = 0; |
| 566 | |
| 567 | atomic64_set(&ring->fence_drv.last_seq, 0); |
| 568 | ring->fence_drv.initialized = false; |
| 569 | |
| 570 | INIT_DELAYED_WORK(&ring->fence_drv.lockup_work, |
| 571 | amdgpu_fence_check_lockup); |
| 572 | ring->fence_drv.ring = ring; |
Alex Deucher | b80d847 | 2015-08-16 22:55:02 -0400 | [diff] [blame] | 573 | |
Christian König | 5ec92a7 | 2015-09-07 18:43:02 +0200 | [diff] [blame] | 574 | init_waitqueue_head(&ring->fence_drv.fence_queue); |
| 575 | |
Alex Deucher | b80d847 | 2015-08-16 22:55:02 -0400 | [diff] [blame] | 576 | if (amdgpu_enable_scheduler) { |
Junwei Zhang | 2440ff2 | 2015-10-10 08:48:42 +0800 | [diff] [blame] | 577 | long timeout = msecs_to_jiffies(amdgpu_lockup_timeout); |
| 578 | if (timeout == 0) { |
| 579 | /* |
| 580 | * FIXME: |
| 581 | * Delayed workqueue cannot use it directly, |
| 582 | * so the scheduler will not use delayed workqueue if |
| 583 | * MAX_SCHEDULE_TIMEOUT is set. |
| 584 | * Currently keep it simple and silly. |
| 585 | */ |
| 586 | timeout = MAX_SCHEDULE_TIMEOUT; |
| 587 | } |
Christian König | 4f839a2 | 2015-09-08 20:22:31 +0200 | [diff] [blame] | 588 | r = amd_sched_init(&ring->sched, &amdgpu_sched_ops, |
Junwei Zhang | 2440ff2 | 2015-10-10 08:48:42 +0800 | [diff] [blame] | 589 | amdgpu_sched_hw_submission, |
| 590 | timeout, ring->name); |
Christian König | 4f839a2 | 2015-09-08 20:22:31 +0200 | [diff] [blame] | 591 | if (r) { |
| 592 | DRM_ERROR("Failed to create scheduler on ring %s.\n", |
| 593 | ring->name); |
| 594 | return r; |
| 595 | } |
Alex Deucher | b80d847 | 2015-08-16 22:55:02 -0400 | [diff] [blame] | 596 | } |
Christian König | 4f839a2 | 2015-09-08 20:22:31 +0200 | [diff] [blame] | 597 | |
| 598 | return 0; |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 599 | } |
| 600 | |
| 601 | /** |
| 602 | * amdgpu_fence_driver_init - init the fence driver |
| 603 | * for all possible rings. |
| 604 | * |
| 605 | * @adev: amdgpu device pointer |
| 606 | * |
| 607 | * Init the fence driver for all possible rings (all asics). |
| 608 | * Not all asics have all rings, so each asic will only |
| 609 | * start the fence driver on the rings it has using |
| 610 | * amdgpu_fence_driver_start_ring(). |
| 611 | * Returns 0 for success. |
| 612 | */ |
| 613 | int amdgpu_fence_driver_init(struct amdgpu_device *adev) |
| 614 | { |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 615 | if (amdgpu_debugfs_fence_init(adev)) |
| 616 | dev_err(adev->dev, "fence debugfs file creation failed\n"); |
| 617 | |
| 618 | return 0; |
| 619 | } |
| 620 | |
| 621 | /** |
| 622 | * amdgpu_fence_driver_fini - tear down the fence driver |
| 623 | * for all possible rings. |
| 624 | * |
| 625 | * @adev: amdgpu device pointer |
| 626 | * |
| 627 | * Tear down the fence driver for all possible rings (all asics). |
| 628 | */ |
| 629 | void amdgpu_fence_driver_fini(struct amdgpu_device *adev) |
| 630 | { |
| 631 | int i, r; |
| 632 | |
| 633 | mutex_lock(&adev->ring_lock); |
| 634 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { |
| 635 | struct amdgpu_ring *ring = adev->rings[i]; |
| 636 | if (!ring || !ring->fence_drv.initialized) |
| 637 | continue; |
| 638 | r = amdgpu_fence_wait_empty(ring); |
| 639 | if (r) { |
| 640 | /* no need to trigger GPU reset as we are unloading */ |
| 641 | amdgpu_fence_driver_force_completion(adev); |
| 642 | } |
monk.liu | 7f06c23 | 2015-07-30 18:28:12 +0800 | [diff] [blame] | 643 | wake_up_all(&ring->fence_drv.fence_queue); |
Chunming Zhou | c6a4079 | 2015-06-01 14:14:32 +0800 | [diff] [blame] | 644 | amdgpu_irq_put(adev, ring->fence_drv.irq_src, |
| 645 | ring->fence_drv.irq_type); |
Christian König | 4f839a2 | 2015-09-08 20:22:31 +0200 | [diff] [blame] | 646 | amd_sched_fini(&ring->sched); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 647 | ring->fence_drv.initialized = false; |
| 648 | } |
| 649 | mutex_unlock(&adev->ring_lock); |
| 650 | } |
| 651 | |
| 652 | /** |
Alex Deucher | 5ceb54c | 2015-08-05 12:41:48 -0400 | [diff] [blame] | 653 | * amdgpu_fence_driver_suspend - suspend the fence driver |
| 654 | * for all possible rings. |
| 655 | * |
| 656 | * @adev: amdgpu device pointer |
| 657 | * |
| 658 | * Suspend the fence driver for all possible rings (all asics). |
| 659 | */ |
| 660 | void amdgpu_fence_driver_suspend(struct amdgpu_device *adev) |
| 661 | { |
| 662 | int i, r; |
| 663 | |
| 664 | mutex_lock(&adev->ring_lock); |
| 665 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { |
| 666 | struct amdgpu_ring *ring = adev->rings[i]; |
| 667 | if (!ring || !ring->fence_drv.initialized) |
| 668 | continue; |
| 669 | |
| 670 | /* wait for gpu to finish processing current batch */ |
| 671 | r = amdgpu_fence_wait_empty(ring); |
| 672 | if (r) { |
| 673 | /* delay GPU reset to resume */ |
| 674 | amdgpu_fence_driver_force_completion(adev); |
| 675 | } |
| 676 | |
| 677 | /* disable the interrupt */ |
| 678 | amdgpu_irq_put(adev, ring->fence_drv.irq_src, |
| 679 | ring->fence_drv.irq_type); |
| 680 | } |
| 681 | mutex_unlock(&adev->ring_lock); |
| 682 | } |
| 683 | |
| 684 | /** |
| 685 | * amdgpu_fence_driver_resume - resume the fence driver |
| 686 | * for all possible rings. |
| 687 | * |
| 688 | * @adev: amdgpu device pointer |
| 689 | * |
| 690 | * Resume the fence driver for all possible rings (all asics). |
| 691 | * Not all asics have all rings, so each asic will only |
| 692 | * start the fence driver on the rings it has using |
| 693 | * amdgpu_fence_driver_start_ring(). |
| 694 | * Returns 0 for success. |
| 695 | */ |
| 696 | void amdgpu_fence_driver_resume(struct amdgpu_device *adev) |
| 697 | { |
| 698 | int i; |
| 699 | |
| 700 | mutex_lock(&adev->ring_lock); |
| 701 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { |
| 702 | struct amdgpu_ring *ring = adev->rings[i]; |
| 703 | if (!ring || !ring->fence_drv.initialized) |
| 704 | continue; |
| 705 | |
| 706 | /* enable the interrupt */ |
| 707 | amdgpu_irq_get(adev, ring->fence_drv.irq_src, |
| 708 | ring->fence_drv.irq_type); |
| 709 | } |
| 710 | mutex_unlock(&adev->ring_lock); |
| 711 | } |
| 712 | |
| 713 | /** |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 714 | * amdgpu_fence_driver_force_completion - force all fence waiter to complete |
| 715 | * |
| 716 | * @adev: amdgpu device pointer |
| 717 | * |
| 718 | * In case of GPU reset failure make sure no process keep waiting on fence |
| 719 | * that will never complete. |
| 720 | */ |
| 721 | void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev) |
| 722 | { |
| 723 | int i; |
| 724 | |
| 725 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { |
| 726 | struct amdgpu_ring *ring = adev->rings[i]; |
| 727 | if (!ring || !ring->fence_drv.initialized) |
| 728 | continue; |
| 729 | |
| 730 | amdgpu_fence_write(ring, ring->fence_drv.sync_seq[i]); |
| 731 | } |
| 732 | } |
| 733 | |
| 734 | |
| 735 | /* |
| 736 | * Fence debugfs |
| 737 | */ |
| 738 | #if defined(CONFIG_DEBUG_FS) |
| 739 | static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) |
| 740 | { |
| 741 | struct drm_info_node *node = (struct drm_info_node *)m->private; |
| 742 | struct drm_device *dev = node->minor->dev; |
| 743 | struct amdgpu_device *adev = dev->dev_private; |
| 744 | int i, j; |
| 745 | |
| 746 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { |
| 747 | struct amdgpu_ring *ring = adev->rings[i]; |
| 748 | if (!ring || !ring->fence_drv.initialized) |
| 749 | continue; |
| 750 | |
| 751 | amdgpu_fence_process(ring); |
| 752 | |
Christian König | 344c19f | 2015-06-02 15:47:16 +0200 | [diff] [blame] | 753 | seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name); |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 754 | seq_printf(m, "Last signaled fence 0x%016llx\n", |
| 755 | (unsigned long long)atomic64_read(&ring->fence_drv.last_seq)); |
| 756 | seq_printf(m, "Last emitted 0x%016llx\n", |
| 757 | ring->fence_drv.sync_seq[i]); |
| 758 | |
| 759 | for (j = 0; j < AMDGPU_MAX_RINGS; ++j) { |
| 760 | struct amdgpu_ring *other = adev->rings[j]; |
Christian König | 344c19f | 2015-06-02 15:47:16 +0200 | [diff] [blame] | 761 | if (i != j && other && other->fence_drv.initialized && |
| 762 | ring->fence_drv.sync_seq[j]) |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 763 | seq_printf(m, "Last sync to ring %d 0x%016llx\n", |
| 764 | j, ring->fence_drv.sync_seq[j]); |
| 765 | } |
| 766 | } |
| 767 | return 0; |
| 768 | } |
| 769 | |
| 770 | static struct drm_info_list amdgpu_debugfs_fence_list[] = { |
| 771 | {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL}, |
| 772 | }; |
| 773 | #endif |
| 774 | |
| 775 | int amdgpu_debugfs_fence_init(struct amdgpu_device *adev) |
| 776 | { |
| 777 | #if defined(CONFIG_DEBUG_FS) |
| 778 | return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 1); |
| 779 | #else |
| 780 | return 0; |
| 781 | #endif |
| 782 | } |
| 783 | |
| 784 | static const char *amdgpu_fence_get_driver_name(struct fence *fence) |
| 785 | { |
| 786 | return "amdgpu"; |
| 787 | } |
| 788 | |
| 789 | static const char *amdgpu_fence_get_timeline_name(struct fence *f) |
| 790 | { |
| 791 | struct amdgpu_fence *fence = to_amdgpu_fence(f); |
| 792 | return (const char *)fence->ring->name; |
| 793 | } |
| 794 | |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 795 | const struct fence_ops amdgpu_fence_ops = { |
| 796 | .get_driver_name = amdgpu_fence_get_driver_name, |
| 797 | .get_timeline_name = amdgpu_fence_get_timeline_name, |
| 798 | .enable_signaling = amdgpu_fence_enable_signaling, |
| 799 | .signaled = amdgpu_fence_is_signaled, |
Christian König | 318cd34 | 2015-10-15 17:58:09 +0200 | [diff] [blame] | 800 | .wait = fence_default_wait, |
Alex Deucher | d38ceaf | 2015-04-20 16:55:21 -0400 | [diff] [blame] | 801 | .release = NULL, |
| 802 | }; |