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 /* delay 0.1 second to enable gfx off feature */
30 #define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
31 
32 /*
33  * GPU GFX IP block helpers function.
34  */
35 
36 int amdgpu_gfx_queue_to_bit(struct amdgpu_device *adev, int mec,
37 			    int pipe, int queue)
38 {
39 	int bit = 0;
40 
41 	bit += mec * adev->gfx.mec.num_pipe_per_mec
42 		* adev->gfx.mec.num_queue_per_pipe;
43 	bit += pipe * adev->gfx.mec.num_queue_per_pipe;
44 	bit += queue;
45 
46 	return bit;
47 }
48 
49 void amdgpu_gfx_bit_to_queue(struct amdgpu_device *adev, int bit,
50 			     int *mec, int *pipe, int *queue)
51 {
52 	*queue = bit % adev->gfx.mec.num_queue_per_pipe;
53 	*pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
54 		% adev->gfx.mec.num_pipe_per_mec;
55 	*mec = (bit / adev->gfx.mec.num_queue_per_pipe)
56 	       / adev->gfx.mec.num_pipe_per_mec;
57 
58 }
59 
60 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
61 				     int mec, int pipe, int queue)
62 {
63 	return test_bit(amdgpu_gfx_queue_to_bit(adev, mec, pipe, queue),
64 			adev->gfx.mec.queue_bitmap);
65 }
66 
67 /**
68  * amdgpu_gfx_scratch_get - Allocate a scratch register
69  *
70  * @adev: amdgpu_device pointer
71  * @reg: scratch register mmio offset
72  *
73  * Allocate a CP scratch register for use by the driver (all asics).
74  * Returns 0 on success or -EINVAL on failure.
75  */
76 int amdgpu_gfx_scratch_get(struct amdgpu_device *adev, uint32_t *reg)
77 {
78 	int i;
79 
80 	i = ffs(adev->gfx.scratch.free_mask);
81 	if (i != 0 && i <= adev->gfx.scratch.num_reg) {
82 		i--;
83 		adev->gfx.scratch.free_mask &= ~(1u << i);
84 		*reg = adev->gfx.scratch.reg_base + i;
85 		return 0;
86 	}
87 	return -EINVAL;
88 }
89 
90 /**
91  * amdgpu_gfx_scratch_free - Free a scratch register
92  *
93  * @adev: amdgpu_device pointer
94  * @reg: scratch register mmio offset
95  *
96  * Free a CP scratch register allocated for use by the driver (all asics)
97  */
98 void amdgpu_gfx_scratch_free(struct amdgpu_device *adev, uint32_t reg)
99 {
100 	adev->gfx.scratch.free_mask |= 1u << (reg - adev->gfx.scratch.reg_base);
101 }
102 
103 /**
104  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
105  *
106  * @mask: array in which the per-shader array disable masks will be stored
107  * @max_se: number of SEs
108  * @max_sh: number of SHs
109  *
110  * The bitmask of CUs to be disabled in the shader array determined by se and
111  * sh is stored in mask[se * max_sh + sh].
112  */
113 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
114 {
115 	unsigned se, sh, cu;
116 	const char *p;
117 
118 	memset(mask, 0, sizeof(*mask) * max_se * max_sh);
119 
120 	if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
121 		return;
122 
123 	p = amdgpu_disable_cu;
124 	for (;;) {
125 		char *next;
126 		int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
127 		if (ret < 3) {
128 			DRM_ERROR("amdgpu: could not parse disable_cu\n");
129 			return;
130 		}
131 
132 		if (se < max_se && sh < max_sh && cu < 16) {
133 			DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
134 			mask[se * max_sh + sh] |= 1u << cu;
135 		} else {
136 			DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
137 				  se, sh, cu);
138 		}
139 
140 		next = strchr(p, ',');
141 		if (!next)
142 			break;
143 		p = next + 1;
144 	}
145 }
146 
147 static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
148 {
149 	if (amdgpu_compute_multipipe != -1) {
150 		DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
151 			 amdgpu_compute_multipipe);
152 		return amdgpu_compute_multipipe == 1;
153 	}
154 
155 	/* FIXME: spreading the queues across pipes causes perf regressions
156 	 * on POLARIS11 compute workloads */
157 	if (adev->asic_type == CHIP_POLARIS11)
158 		return false;
159 
160 	return adev->gfx.mec.num_mec > 1;
161 }
162 
163 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
164 {
165 	int i, queue, pipe, mec;
166 	bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
167 
168 	/* policy for amdgpu compute queue ownership */
169 	for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
170 		queue = i % adev->gfx.mec.num_queue_per_pipe;
171 		pipe = (i / adev->gfx.mec.num_queue_per_pipe)
172 			% adev->gfx.mec.num_pipe_per_mec;
173 		mec = (i / adev->gfx.mec.num_queue_per_pipe)
174 			/ adev->gfx.mec.num_pipe_per_mec;
175 
176 		/* we've run out of HW */
177 		if (mec >= adev->gfx.mec.num_mec)
178 			break;
179 
180 		if (multipipe_policy) {
181 			/* policy: amdgpu owns the first two queues of the first MEC */
182 			if (mec == 0 && queue < 2)
183 				set_bit(i, adev->gfx.mec.queue_bitmap);
184 		} else {
185 			/* policy: amdgpu owns all queues in the first pipe */
186 			if (mec == 0 && pipe == 0)
187 				set_bit(i, adev->gfx.mec.queue_bitmap);
188 		}
189 	}
190 
191 	/* update the number of active compute rings */
192 	adev->gfx.num_compute_rings =
193 		bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
194 
195 	/* If you hit this case and edited the policy, you probably just
196 	 * need to increase AMDGPU_MAX_COMPUTE_RINGS */
197 	if (WARN_ON(adev->gfx.num_compute_rings > AMDGPU_MAX_COMPUTE_RINGS))
198 		adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
199 }
200 
201 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
202 				  struct amdgpu_ring *ring)
203 {
204 	int queue_bit;
205 	int mec, pipe, queue;
206 
207 	queue_bit = adev->gfx.mec.num_mec
208 		    * adev->gfx.mec.num_pipe_per_mec
209 		    * adev->gfx.mec.num_queue_per_pipe;
210 
211 	while (queue_bit-- >= 0) {
212 		if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
213 			continue;
214 
215 		amdgpu_gfx_bit_to_queue(adev, queue_bit, &mec, &pipe, &queue);
216 
217 		/*
218 		 * 1. Using pipes 2/3 from MEC 2 seems cause problems.
219 		 * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
220 		 * only can be issued on queue 0.
221 		 */
222 		if ((mec == 1 && pipe > 1) || queue != 0)
223 			continue;
224 
225 		ring->me = mec + 1;
226 		ring->pipe = pipe;
227 		ring->queue = queue;
228 
229 		return 0;
230 	}
231 
232 	dev_err(adev->dev, "Failed to find a queue for KIQ\n");
233 	return -EINVAL;
234 }
235 
236 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
237 			     struct amdgpu_ring *ring,
238 			     struct amdgpu_irq_src *irq)
239 {
240 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
241 	int r = 0;
242 
243 	spin_lock_init(&kiq->ring_lock);
244 
245 	r = amdgpu_device_wb_get(adev, &adev->virt.reg_val_offs);
246 	if (r)
247 		return r;
248 
249 	ring->adev = NULL;
250 	ring->ring_obj = NULL;
251 	ring->use_doorbell = true;
252 	ring->doorbell_index = AMDGPU_DOORBELL_KIQ;
253 
254 	r = amdgpu_gfx_kiq_acquire(adev, ring);
255 	if (r)
256 		return r;
257 
258 	ring->eop_gpu_addr = kiq->eop_gpu_addr;
259 	sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
260 	r = amdgpu_ring_init(adev, ring, 1024,
261 			     irq, AMDGPU_CP_KIQ_IRQ_DRIVER0);
262 	if (r)
263 		dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
264 
265 	return r;
266 }
267 
268 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring,
269 			      struct amdgpu_irq_src *irq)
270 {
271 	amdgpu_device_wb_free(ring->adev, ring->adev->virt.reg_val_offs);
272 	amdgpu_ring_fini(ring);
273 }
274 
275 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
276 {
277 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
278 
279 	amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
280 }
281 
282 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
283 			unsigned hpd_size)
284 {
285 	int r;
286 	u32 *hpd;
287 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
288 
289 	r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
290 				    AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
291 				    &kiq->eop_gpu_addr, (void **)&hpd);
292 	if (r) {
293 		dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
294 		return r;
295 	}
296 
297 	memset(hpd, 0, hpd_size);
298 
299 	r = amdgpu_bo_reserve(kiq->eop_obj, true);
300 	if (unlikely(r != 0))
301 		dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
302 	amdgpu_bo_kunmap(kiq->eop_obj);
303 	amdgpu_bo_unreserve(kiq->eop_obj);
304 
305 	return 0;
306 }
307 
308 /* create MQD for each compute queue */
309 int amdgpu_gfx_compute_mqd_sw_init(struct amdgpu_device *adev,
310 				   unsigned mqd_size)
311 {
312 	struct amdgpu_ring *ring = NULL;
313 	int r, i;
314 
315 	/* create MQD for KIQ */
316 	ring = &adev->gfx.kiq.ring;
317 	if (!ring->mqd_obj) {
318 		/* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
319 		 * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
320 		 * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
321 		 * KIQ MQD no matter SRIOV or Bare-metal
322 		 */
323 		r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
324 					    AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
325 					    &ring->mqd_gpu_addr, &ring->mqd_ptr);
326 		if (r) {
327 			dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
328 			return r;
329 		}
330 
331 		/* prepare MQD backup */
332 		adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
333 		if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
334 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
335 	}
336 
337 	/* create MQD for each KCQ */
338 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
339 		ring = &adev->gfx.compute_ring[i];
340 		if (!ring->mqd_obj) {
341 			r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
342 						    AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
343 						    &ring->mqd_gpu_addr, &ring->mqd_ptr);
344 			if (r) {
345 				dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
346 				return r;
347 			}
348 
349 			/* prepare MQD backup */
350 			adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
351 			if (!adev->gfx.mec.mqd_backup[i])
352 				dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
353 		}
354 	}
355 
356 	return 0;
357 }
358 
359 void amdgpu_gfx_compute_mqd_sw_fini(struct amdgpu_device *adev)
360 {
361 	struct amdgpu_ring *ring = NULL;
362 	int i;
363 
364 	for (i = 0; i < adev->gfx.num_compute_rings; i++) {
365 		ring = &adev->gfx.compute_ring[i];
366 		kfree(adev->gfx.mec.mqd_backup[i]);
367 		amdgpu_bo_free_kernel(&ring->mqd_obj,
368 				      &ring->mqd_gpu_addr,
369 				      &ring->mqd_ptr);
370 	}
371 
372 	ring = &adev->gfx.kiq.ring;
373 	kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
374 	amdgpu_bo_free_kernel(&ring->mqd_obj,
375 			      &ring->mqd_gpu_addr,
376 			      &ring->mqd_ptr);
377 }
378 
379 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
380  *
381  * @adev: amdgpu_device pointer
382  * @bool enable true: enable gfx off feature, false: disable gfx off feature
383  *
384  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
385  * 2. other client can send request to disable gfx off feature, the request should be honored.
386  * 3. other client can cancel their request of disable gfx off feature
387  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
388  */
389 
390 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
391 {
392 	if (!(adev->powerplay.pp_feature & PP_GFXOFF_MASK))
393 		return;
394 
395 	if (!adev->powerplay.pp_funcs->set_powergating_by_smu)
396 		return;
397 
398 
399 	mutex_lock(&adev->gfx.gfx_off_mutex);
400 
401 	if (!enable)
402 		adev->gfx.gfx_off_req_count++;
403 	else if (adev->gfx.gfx_off_req_count > 0)
404 		adev->gfx.gfx_off_req_count--;
405 
406 	if (enable && !adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
407 		schedule_delayed_work(&adev->gfx.gfx_off_delay_work, GFX_OFF_DELAY_ENABLE);
408 	} else if (!enable && adev->gfx.gfx_off_state) {
409 		if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false))
410 			adev->gfx.gfx_off_state = false;
411 	}
412 
413 	mutex_unlock(&adev->gfx.gfx_off_mutex);
414 }
415