1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  * Copyright 2008 Red Hat Inc.
4  * Copyright 2009 Jerome Glisse.
5  *
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:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
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.
23  *
24  */
25 #include <drm/drmP.h>
26 #include "amdgpu.h"
27 #include "amdgpu_gfx.h"
28 
29 /*
30  * GPU scratch registers helpers function.
31  */
32 /**
33  * amdgpu_gfx_scratch_get - Allocate a scratch register
34  *
35  * @adev: amdgpu_device pointer
36  * @reg: scratch register mmio offset
37  *
38  * Allocate a CP scratch register for use by the driver (all asics).
39  * Returns 0 on success or -EINVAL on failure.
40  */
41 int amdgpu_gfx_scratch_get(struct amdgpu_device *adev, uint32_t *reg)
42 {
43 	int i;
44 
45 	i = ffs(adev->gfx.scratch.free_mask);
46 	if (i != 0 && i <= adev->gfx.scratch.num_reg) {
47 		i--;
48 		adev->gfx.scratch.free_mask &= ~(1u << i);
49 		*reg = adev->gfx.scratch.reg_base + i;
50 		return 0;
51 	}
52 	return -EINVAL;
53 }
54 
55 /**
56  * amdgpu_gfx_scratch_free - Free a scratch register
57  *
58  * @adev: amdgpu_device pointer
59  * @reg: scratch register mmio offset
60  *
61  * Free a CP scratch register allocated for use by the driver (all asics)
62  */
63 void amdgpu_gfx_scratch_free(struct amdgpu_device *adev, uint32_t reg)
64 {
65 	adev->gfx.scratch.free_mask |= 1u << (reg - adev->gfx.scratch.reg_base);
66 }
67 
68 /**
69  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
70  *
71  * @mask: array in which the per-shader array disable masks will be stored
72  * @max_se: number of SEs
73  * @max_sh: number of SHs
74  *
75  * The bitmask of CUs to be disabled in the shader array determined by se and
76  * sh is stored in mask[se * max_sh + sh].
77  */
78 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
79 {
80 	unsigned se, sh, cu;
81 	const char *p;
82 
83 	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
84 
85 	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
86 		return;
87 
88 	p = amdgpu_disable_cu;
89 	for (;;) {
90 		char *next;
91 		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
92 		if (ret < 3) {
93 			DRM_ERROR("amdgpu: could not parse disable_cu\n");
94 			return;
95 		}
96 
97 		if (se < max_se && sh < max_sh && cu < 16) {
98 			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
99 			mask[se * max_sh + sh] |= 1u << cu;
100 		} else {
101 			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
102 				  se, sh, cu);
103 		}
104 
105 		next = strchr(p, ',');
106 		if (!next)
107 			break;
108 		p = next + 1;
109 	}
110 }
111 
112 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
113 {
114 	int i, queue, pipe, mec;
115 
116 	/* policy for amdgpu compute queue ownership */
117 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
118 		queue = i % adev->gfx.mec.num_queue_per_pipe;
119 		pipe = (i / adev->gfx.mec.num_queue_per_pipe)
120 			% adev->gfx.mec.num_pipe_per_mec;
121 		mec = (i / adev->gfx.mec.num_queue_per_pipe)
122 			/ adev->gfx.mec.num_pipe_per_mec;
123 
124 		/* we've run out of HW */
125 		if (mec >= adev->gfx.mec.num_mec)
126 			break;
127 
128 		/* FIXME: spreading the queues across pipes causes perf regressions */
129 		if (0) {
130 			/* policy: amdgpu owns the first two queues of the first MEC */
131 			if (mec == 0 && queue < 2)
132 				set_bit(i, adev->gfx.mec.queue_bitmap);
133 		} else {
134 			/* policy: amdgpu owns all queues in the first pipe */
135 			if (mec == 0 && pipe == 0)
136 				set_bit(i, adev->gfx.mec.queue_bitmap);
137 		}
138 	}
139 
140 	/* update the number of active compute rings */
141 	adev->gfx.num_compute_rings =
142 		bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
143 
144 	/* If you hit this case and edited the policy, you probably just
145 	 * need to increase AMDGPU_MAX_COMPUTE_RINGS */
146 	if (WARN_ON(adev->gfx.num_compute_rings > AMDGPU_MAX_COMPUTE_RINGS))
147 		adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
148 }
149 
150 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
151 				  struct amdgpu_ring *ring)
152 {
153 	int queue_bit;
154 	int mec, pipe, queue;
155 
156 	queue_bit = adev->gfx.mec.num_mec
157 		    * adev->gfx.mec.num_pipe_per_mec
158 		    * adev->gfx.mec.num_queue_per_pipe;
159 
160 	while (queue_bit-- >= 0) {
161 		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
162 			continue;
163 
164 		amdgpu_gfx_bit_to_queue(adev, queue_bit, &mec, &pipe, &queue);
165 
166 		/* Using pipes 2/3 from MEC 2 seems cause problems */
167 		if (mec == 1 && pipe > 1)
168 			continue;
169 
170 		ring->me = mec + 1;
171 		ring->pipe = pipe;
172 		ring->queue = queue;
173 
174 		return 0;
175 	}
176 
177 	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
178 	return -EINVAL;
179 }
180 
181 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
182 			     struct amdgpu_ring *ring,
183 			     struct amdgpu_irq_src *irq)
184 {
185 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
186 	int r = 0;
187 
188 	mutex_init(&kiq->ring_mutex);
189 
190 	r = amdgpu_wb_get(adev, &adev->virt.reg_val_offs);
191 	if (r)
192 		return r;
193 
194 	ring->adev = NULL;
195 	ring->ring_obj = NULL;
196 	ring->use_doorbell = true;
197 	ring->doorbell_index = AMDGPU_DOORBELL_KIQ;
198 
199 	r = amdgpu_gfx_kiq_acquire(adev, ring);
200 	if (r)
201 		return r;
202 
203 	ring->eop_gpu_addr = kiq->eop_gpu_addr;
204 	sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
205 	r = amdgpu_ring_init(adev, ring, 1024,
206 			     irq, AMDGPU_CP_KIQ_IRQ_DRIVER0);
207 	if (r)
208 		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
209 
210 	return r;
211 }
212 
213 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring,
214 			      struct amdgpu_irq_src *irq)
215 {
216 	amdgpu_wb_free(ring->adev, ring->adev->virt.reg_val_offs);
217 	amdgpu_ring_fini(ring);
218 }
219 
220 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
221 {
222 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
223 
224 	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
225 }
226 
227 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
228 			unsigned hpd_size)
229 {
230 	int r;
231 	u32 *hpd;
232 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
233 
234 	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
235 				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
236 				    &kiq->eop_gpu_addr, (void **)&hpd);
237 	if (r) {
238 		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
239 		return r;
240 	}
241 
242 	memset(hpd, 0, hpd_size);
243 
244 	r = amdgpu_bo_reserve(kiq->eop_obj, true);
245 	if (unlikely(r != 0))
246 		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
247 	amdgpu_bo_kunmap(kiq->eop_obj);
248 	amdgpu_bo_unreserve(kiq->eop_obj);
249 
250 	return 0;
251 }
252 
253 /* create MQD for each compute queue */
254 int amdgpu_gfx_compute_mqd_sw_init(struct amdgpu_device *adev,
255 				   unsigned mqd_size)
256 {
257 	struct amdgpu_ring *ring = NULL;
258 	int r, i;
259 
260 	/* create MQD for KIQ */
261 	ring = &adev->gfx.kiq.ring;
262 	if (!ring->mqd_obj) {
263 		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
264 					    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
265 					    &ring->mqd_gpu_addr, &ring->mqd_ptr);
266 		if (r) {
267 			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
268 			return r;
269 		}
270 
271 		/* prepare MQD backup */
272 		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
273 		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
274 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
275 	}
276 
277 	/* create MQD for each KCQ */
278 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
279 		ring = &adev->gfx.compute_ring[i];
280 		if (!ring->mqd_obj) {
281 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
282 						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
283 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
284 			if (r) {
285 				dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
286 				return r;
287 			}
288 
289 			/* prepare MQD backup */
290 			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
291 			if (!adev->gfx.mec.mqd_backup[i])
292 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
293 		}
294 	}
295 
296 	return 0;
297 }
298 
299 void amdgpu_gfx_compute_mqd_sw_fini(struct amdgpu_device *adev)
300 {
301 	struct amdgpu_ring *ring = NULL;
302 	int i;
303 
304 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
305 		ring = &adev->gfx.compute_ring[i];
306 		kfree(adev->gfx.mec.mqd_backup[i]);
307 		amdgpu_bo_free_kernel(&ring->mqd_obj,
308 				      &ring->mqd_gpu_addr,
309 				      &ring->mqd_ptr);
310 	}
311 
312 	ring = &adev->gfx.kiq.ring;
313 	kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
314 	amdgpu_bo_free_kernel(&ring->mqd_obj,
315 			      &ring->mqd_gpu_addr,
316 			      &ring->mqd_ptr);
317 }
318