2 * Copyright 2008 Advanced Micro Devices, Inc.
3 * Copyright 2008 Red Hat Inc.
4 * Copyright 2009 Jerome Glisse.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22 * OTHER DEALINGS IN THE SOFTWARE.
24 * Authors: Dave Airlie
29 #include <linux/seq_file.h>
30 #include <linux/slab.h>
31 #include <linux/uaccess.h>
32 #include <linux/debugfs.h>
34 #include <drm/amdgpu_drm.h>
40 * Most engines on the GPU are fed via ring buffers. Ring
41 * buffers are areas of GPU accessible memory that the host
42 * writes commands into and the GPU reads commands out of.
43 * There is a rptr (read pointer) that determines where the
44 * GPU is currently reading, and a wptr (write pointer)
45 * which determines where the host has written. When the
46 * pointers are equal, the ring is idle. When the host
47 * writes commands to the ring buffer, it increments the
48 * wptr. The GPU then starts fetching commands and executes
49 * them until the pointers are equal again.
51 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
52 struct amdgpu_ring *ring);
53 static void amdgpu_debugfs_ring_fini(struct amdgpu_ring *ring);
56 * amdgpu_ring_alloc - allocate space on the ring buffer
58 * @adev: amdgpu_device pointer
59 * @ring: amdgpu_ring structure holding ring information
60 * @ndw: number of dwords to allocate in the ring buffer
62 * Allocate @ndw dwords in the ring buffer (all asics).
63 * Returns 0 on success, error on failure.
65 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw)
67 /* Align requested size with padding so unlock_commit can
69 ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
71 /* Make sure we aren't trying to allocate more space
72 * than the maximum for one submission
74 if (WARN_ON_ONCE(ndw > ring->max_dw))
78 ring->wptr_old = ring->wptr;
80 if (ring->funcs->begin_use)
81 ring->funcs->begin_use(ring);
86 /** amdgpu_ring_insert_nop - insert NOP packets
88 * @ring: amdgpu_ring structure holding ring information
89 * @count: the number of NOP packets to insert
91 * This is the generic insert_nop function for rings except SDMA
93 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
97 for (i = 0; i < count; i++)
98 amdgpu_ring_write(ring, ring->funcs->nop);
101 /** amdgpu_ring_generic_pad_ib - pad IB with NOP packets
103 * @ring: amdgpu_ring structure holding ring information
104 * @ib: IB to add NOP packets to
106 * This is the generic pad_ib function for rings except SDMA
108 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
110 while (ib->length_dw & ring->funcs->align_mask)
111 ib->ptr[ib->length_dw++] = ring->funcs->nop;
115 * amdgpu_ring_commit - tell the GPU to execute the new
116 * commands on the ring buffer
118 * @adev: amdgpu_device pointer
119 * @ring: amdgpu_ring structure holding ring information
121 * Update the wptr (write pointer) to tell the GPU to
122 * execute new commands on the ring buffer (all asics).
124 void amdgpu_ring_commit(struct amdgpu_ring *ring)
128 /* We pad to match fetch size */
129 count = ring->funcs->align_mask + 1 -
130 (ring->wptr & ring->funcs->align_mask);
131 count %= ring->funcs->align_mask + 1;
132 ring->funcs->insert_nop(ring, count);
135 amdgpu_ring_set_wptr(ring);
137 if (ring->funcs->end_use)
138 ring->funcs->end_use(ring);
142 * amdgpu_ring_undo - reset the wptr
144 * @ring: amdgpu_ring structure holding ring information
146 * Reset the driver's copy of the wptr (all asics).
148 void amdgpu_ring_undo(struct amdgpu_ring *ring)
150 ring->wptr = ring->wptr_old;
152 if (ring->funcs->end_use)
153 ring->funcs->end_use(ring);
157 * amdgpu_ring_priority_put - restore a ring's priority
159 * @ring: amdgpu_ring structure holding the information
160 * @priority: target priority
162 * Release a request for executing at @priority
164 void amdgpu_ring_priority_put(struct amdgpu_ring *ring,
165 enum drm_sched_priority priority)
169 if (!ring->funcs->set_priority)
172 if (atomic_dec_return(&ring->num_jobs[priority]) > 0)
175 /* no need to restore if the job is already at the lowest priority */
176 if (priority == DRM_SCHED_PRIORITY_NORMAL)
179 mutex_lock(&ring->priority_mutex);
180 /* something higher prio is executing, no need to decay */
181 if (ring->priority > priority)
184 /* decay priority to the next level with a job available */
185 for (i = priority; i >= DRM_SCHED_PRIORITY_MIN; i--) {
186 if (i == DRM_SCHED_PRIORITY_NORMAL
187 || atomic_read(&ring->num_jobs[i])) {
189 ring->funcs->set_priority(ring, i);
195 mutex_unlock(&ring->priority_mutex);
199 * amdgpu_ring_priority_get - change the ring's priority
201 * @ring: amdgpu_ring structure holding the information
202 * @priority: target priority
204 * Request a ring's priority to be raised to @priority (refcounted).
206 void amdgpu_ring_priority_get(struct amdgpu_ring *ring,
207 enum drm_sched_priority priority)
209 if (!ring->funcs->set_priority)
212 if (atomic_inc_return(&ring->num_jobs[priority]) <= 0)
215 mutex_lock(&ring->priority_mutex);
216 if (priority <= ring->priority)
219 ring->priority = priority;
220 ring->funcs->set_priority(ring, priority);
223 mutex_unlock(&ring->priority_mutex);
227 * amdgpu_ring_init - init driver ring struct.
229 * @adev: amdgpu_device pointer
230 * @ring: amdgpu_ring structure holding ring information
231 * @max_ndw: maximum number of dw for ring alloc
232 * @nop: nop packet for this ring
234 * Initialize the driver information for the selected ring (all asics).
235 * Returns 0 on success, error on failure.
237 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
238 unsigned max_dw, struct amdgpu_irq_src *irq_src,
242 int sched_hw_submission = amdgpu_sched_hw_submission;
244 /* Set the hw submission limit higher for KIQ because
245 * it's used for a number of gfx/compute tasks by both
246 * KFD and KGD which may have outstanding fences and
247 * it doesn't really use the gpu scheduler anyway;
248 * KIQ tasks get submitted directly to the ring.
250 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
251 sched_hw_submission = max(sched_hw_submission, 256);
252 else if (ring == &adev->sdma.instance[0].page)
253 sched_hw_submission = 256;
255 if (ring->adev == NULL) {
256 if (adev->num_rings >= AMDGPU_MAX_RINGS)
260 ring->idx = adev->num_rings++;
261 adev->rings[ring->idx] = ring;
262 r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission);
267 r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
269 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
273 r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
275 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
279 r = amdgpu_device_wb_get(adev, &ring->fence_offs);
281 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
285 r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
287 dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
290 ring->cond_exe_gpu_addr = adev->wb.gpu_addr + (ring->cond_exe_offs * 4);
291 ring->cond_exe_cpu_addr = &adev->wb.wb[ring->cond_exe_offs];
292 /* always set cond_exec_polling to CONTINUE */
293 *ring->cond_exe_cpu_addr = 1;
295 r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
297 dev_err(adev->dev, "failed initializing fences (%d).\n", r);
301 ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
303 ring->buf_mask = (ring->ring_size / 4) - 1;
304 ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
305 0xffffffffffffffff : ring->buf_mask;
306 /* Allocate ring buffer */
307 if (ring->ring_obj == NULL) {
308 r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE,
309 AMDGPU_GEM_DOMAIN_GTT,
312 (void **)&ring->ring);
314 dev_err(adev->dev, "(%d) ring create failed\n", r);
317 amdgpu_ring_clear_ring(ring);
320 ring->max_dw = max_dw;
321 ring->priority = DRM_SCHED_PRIORITY_NORMAL;
322 mutex_init(&ring->priority_mutex);
324 for (i = 0; i < DRM_SCHED_PRIORITY_MAX; ++i)
325 atomic_set(&ring->num_jobs[i], 0);
327 if (amdgpu_debugfs_ring_init(adev, ring)) {
328 DRM_ERROR("Failed to register debugfs file for rings !\n");
335 * amdgpu_ring_fini - tear down the driver ring struct.
337 * @adev: amdgpu_device pointer
338 * @ring: amdgpu_ring structure holding ring information
340 * Tear down the driver information for the selected ring (all asics).
342 void amdgpu_ring_fini(struct amdgpu_ring *ring)
344 ring->sched.ready = false;
346 /* Not to finish a ring which is not initialized */
347 if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
350 amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
351 amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
353 amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
354 amdgpu_device_wb_free(ring->adev, ring->fence_offs);
356 amdgpu_bo_free_kernel(&ring->ring_obj,
358 (void **)&ring->ring);
360 amdgpu_debugfs_ring_fini(ring);
362 dma_fence_put(ring->vmid_wait);
363 ring->vmid_wait = NULL;
366 ring->adev->rings[ring->idx] = NULL;
370 * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
372 * @adev: amdgpu_device pointer
373 * @reg0: register to write
374 * @reg1: register to wait on
375 * @ref: reference value to write/wait on
376 * @mask: mask to wait on
378 * Helper for rings that don't support write and wait in a
379 * single oneshot packet.
381 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
382 uint32_t reg0, uint32_t reg1,
383 uint32_t ref, uint32_t mask)
385 amdgpu_ring_emit_wreg(ring, reg0, ref);
386 amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
390 * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
392 * @ring: ring to try the recovery on
393 * @vmid: VMID we try to get going again
394 * @fence: timedout fence
396 * Tries to get a ring proceeding again when it is stuck.
398 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
399 struct dma_fence *fence)
401 ktime_t deadline = ktime_add_us(ktime_get(), 10000);
403 if (!ring->funcs->soft_recovery || !fence)
406 atomic_inc(&ring->adev->gpu_reset_counter);
407 while (!dma_fence_is_signaled(fence) &&
408 ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
409 ring->funcs->soft_recovery(ring, vmid);
411 return dma_fence_is_signaled(fence);
417 #if defined(CONFIG_DEBUG_FS)
419 /* Layout of file is 12 bytes consisting of
422 * - driver's copy of wptr
424 * followed by n-words of ring data
426 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
427 size_t size, loff_t *pos)
429 struct amdgpu_ring *ring = file_inode(f)->i_private;
431 uint32_t value, result, early[3];
433 if (*pos & 3 || size & 3)
439 early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
440 early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
441 early[2] = ring->wptr & ring->buf_mask;
442 for (i = *pos / 4; i < 3 && size; i++) {
443 r = put_user(early[i], (uint32_t *)buf);
454 if (*pos >= (ring->ring_size + 12))
457 value = ring->ring[(*pos - 12)/4];
458 r = put_user(value, (uint32_t*)buf);
470 static const struct file_operations amdgpu_debugfs_ring_fops = {
471 .owner = THIS_MODULE,
472 .read = amdgpu_debugfs_ring_read,
473 .llseek = default_llseek
478 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
479 struct amdgpu_ring *ring)
481 #if defined(CONFIG_DEBUG_FS)
482 struct drm_minor *minor = adev->ddev->primary;
483 struct dentry *ent, *root = minor->debugfs_root;
486 sprintf(name, "amdgpu_ring_%s", ring->name);
488 ent = debugfs_create_file(name,
489 S_IFREG | S_IRUGO, root,
490 ring, &amdgpu_debugfs_ring_fops);
494 i_size_write(ent->d_inode, ring->ring_size + 12);
500 static void amdgpu_debugfs_ring_fini(struct amdgpu_ring *ring)
502 #if defined(CONFIG_DEBUG_FS)
503 debugfs_remove(ring->ent);
508 * amdgpu_ring_test_helper - tests ring and set sched readiness status
510 * @ring: ring to try the recovery on
512 * Tests ring and set sched readiness status
514 * Returns 0 on success, error on failure.
516 int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
518 struct amdgpu_device *adev = ring->adev;
521 r = amdgpu_ring_test_ring(ring);
523 DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
526 DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
529 ring->sched.ready = !r;