xref: /linux/drivers/gpu/drm/panfrost/panfrost_perfcnt.c (revision ab93e0dd72c37d378dd936f031ffb83ff2bd87ce)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright 2019 Collabora Ltd */
3 
4 #include <linux/completion.h>
5 #include <linux/iopoll.h>
6 #include <linux/iosys-map.h>
7 #include <linux/pm_runtime.h>
8 #include <linux/slab.h>
9 #include <linux/uaccess.h>
10 
11 #include <drm/drm_file.h>
12 #include <drm/drm_gem_shmem_helper.h>
13 #include <drm/panfrost_drm.h>
14 
15 #include "panfrost_device.h"
16 #include "panfrost_features.h"
17 #include "panfrost_gem.h"
18 #include "panfrost_issues.h"
19 #include "panfrost_job.h"
20 #include "panfrost_mmu.h"
21 #include "panfrost_perfcnt.h"
22 #include "panfrost_regs.h"
23 
24 #define COUNTERS_PER_BLOCK		64
25 #define BYTES_PER_COUNTER		4
26 #define BLOCKS_PER_COREGROUP		8
27 #define V4_SHADERS_PER_COREGROUP	4
28 
29 struct panfrost_perfcnt {
30 	struct panfrost_gem_mapping *mapping;
31 	size_t bosize;
32 	void *buf;
33 	struct panfrost_file_priv *user;
34 	struct mutex lock;
35 	struct completion dump_comp;
36 };
37 
panfrost_perfcnt_clean_cache_done(struct panfrost_device * pfdev)38 void panfrost_perfcnt_clean_cache_done(struct panfrost_device *pfdev)
39 {
40 	complete(&pfdev->perfcnt->dump_comp);
41 }
42 
panfrost_perfcnt_sample_done(struct panfrost_device * pfdev)43 void panfrost_perfcnt_sample_done(struct panfrost_device *pfdev)
44 {
45 	gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_CACHES);
46 }
47 
panfrost_perfcnt_dump_locked(struct panfrost_device * pfdev)48 static int panfrost_perfcnt_dump_locked(struct panfrost_device *pfdev)
49 {
50 	u64 gpuva;
51 	int ret;
52 
53 	reinit_completion(&pfdev->perfcnt->dump_comp);
54 	gpuva = pfdev->perfcnt->mapping->mmnode.start << PAGE_SHIFT;
55 	gpu_write(pfdev, GPU_PERFCNT_BASE_LO, lower_32_bits(gpuva));
56 	gpu_write(pfdev, GPU_PERFCNT_BASE_HI, upper_32_bits(gpuva));
57 	gpu_write(pfdev, GPU_INT_CLEAR,
58 		  GPU_IRQ_CLEAN_CACHES_COMPLETED |
59 		  GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
60 	gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_SAMPLE);
61 	ret = wait_for_completion_interruptible_timeout(&pfdev->perfcnt->dump_comp,
62 							msecs_to_jiffies(1000));
63 	if (!ret)
64 		ret = -ETIMEDOUT;
65 	else if (ret > 0)
66 		ret = 0;
67 
68 	return ret;
69 }
70 
panfrost_perfcnt_enable_locked(struct panfrost_device * pfdev,struct drm_file * file_priv,unsigned int counterset)71 static int panfrost_perfcnt_enable_locked(struct panfrost_device *pfdev,
72 					  struct drm_file *file_priv,
73 					  unsigned int counterset)
74 {
75 	struct panfrost_file_priv *user = file_priv->driver_priv;
76 	struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
77 	struct iosys_map map;
78 	struct drm_gem_shmem_object *bo;
79 	u32 cfg, as;
80 	int ret;
81 
82 	if (user == perfcnt->user)
83 		return 0;
84 	else if (perfcnt->user)
85 		return -EBUSY;
86 
87 	ret = pm_runtime_get_sync(pfdev->dev);
88 	if (ret < 0)
89 		goto err_put_pm;
90 
91 	bo = drm_gem_shmem_create(pfdev->ddev, perfcnt->bosize);
92 	if (IS_ERR(bo)) {
93 		ret = PTR_ERR(bo);
94 		goto err_put_pm;
95 	}
96 
97 	/* Map the perfcnt buf in the address space attached to file_priv. */
98 	ret = panfrost_gem_open(&bo->base, file_priv);
99 	if (ret)
100 		goto err_put_bo;
101 
102 	perfcnt->mapping = panfrost_gem_mapping_get(to_panfrost_bo(&bo->base),
103 						    user);
104 	if (!perfcnt->mapping) {
105 		ret = -EINVAL;
106 		goto err_close_bo;
107 	}
108 
109 	ret = drm_gem_vmap(&bo->base, &map);
110 	if (ret)
111 		goto err_put_mapping;
112 	perfcnt->buf = map.vaddr;
113 
114 	panfrost_gem_internal_set_label(&bo->base, "Perfcnt sample buffer");
115 
116 	/*
117 	 * Invalidate the cache and clear the counters to start from a fresh
118 	 * state.
119 	 */
120 	reinit_completion(&pfdev->perfcnt->dump_comp);
121 	gpu_write(pfdev, GPU_INT_CLEAR,
122 		  GPU_IRQ_CLEAN_CACHES_COMPLETED |
123 		  GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
124 	gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_CLEAR);
125 	gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_INV_CACHES);
126 	ret = wait_for_completion_timeout(&pfdev->perfcnt->dump_comp,
127 					  msecs_to_jiffies(1000));
128 	if (!ret) {
129 		ret = -ETIMEDOUT;
130 		goto err_vunmap;
131 	}
132 
133 	perfcnt->user = user;
134 
135 	as = panfrost_mmu_as_get(pfdev, perfcnt->mapping->mmu);
136 	cfg = GPU_PERFCNT_CFG_AS(as) |
137 	      GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_MANUAL);
138 
139 	/*
140 	 * Bifrost GPUs have 2 set of counters, but we're only interested by
141 	 * the first one for now.
142 	 */
143 	if (panfrost_model_is_bifrost(pfdev))
144 		cfg |= GPU_PERFCNT_CFG_SETSEL(counterset);
145 
146 	gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0xffffffff);
147 	gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0xffffffff);
148 	gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0xffffffff);
149 
150 	/*
151 	 * Due to PRLAM-8186 we need to disable the Tiler before we enable HW
152 	 * counters.
153 	 */
154 	if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
155 		gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
156 	else
157 		gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);
158 
159 	gpu_write(pfdev, GPU_PERFCNT_CFG, cfg);
160 
161 	if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
162 		gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);
163 
164 	/* The BO ref is retained by the mapping. */
165 	drm_gem_object_put(&bo->base);
166 
167 	return 0;
168 
169 err_vunmap:
170 	drm_gem_vunmap(&bo->base, &map);
171 err_put_mapping:
172 	panfrost_gem_mapping_put(perfcnt->mapping);
173 err_close_bo:
174 	panfrost_gem_close(&bo->base, file_priv);
175 err_put_bo:
176 	drm_gem_object_put(&bo->base);
177 err_put_pm:
178 	pm_runtime_put(pfdev->dev);
179 	return ret;
180 }
181 
panfrost_perfcnt_disable_locked(struct panfrost_device * pfdev,struct drm_file * file_priv)182 static int panfrost_perfcnt_disable_locked(struct panfrost_device *pfdev,
183 					   struct drm_file *file_priv)
184 {
185 	struct panfrost_file_priv *user = file_priv->driver_priv;
186 	struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
187 	struct iosys_map map = IOSYS_MAP_INIT_VADDR(perfcnt->buf);
188 
189 	if (user != perfcnt->user)
190 		return -EINVAL;
191 
192 	gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0x0);
193 	gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0x0);
194 	gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0x0);
195 	gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
196 	gpu_write(pfdev, GPU_PERFCNT_CFG,
197 		  GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
198 
199 	perfcnt->user = NULL;
200 	drm_gem_vunmap(&perfcnt->mapping->obj->base.base, &map);
201 	perfcnt->buf = NULL;
202 	panfrost_gem_close(&perfcnt->mapping->obj->base.base, file_priv);
203 	panfrost_mmu_as_put(pfdev, perfcnt->mapping->mmu);
204 	panfrost_gem_mapping_put(perfcnt->mapping);
205 	perfcnt->mapping = NULL;
206 	pm_runtime_mark_last_busy(pfdev->dev);
207 	pm_runtime_put_autosuspend(pfdev->dev);
208 
209 	return 0;
210 }
211 
panfrost_ioctl_perfcnt_enable(struct drm_device * dev,void * data,struct drm_file * file_priv)212 int panfrost_ioctl_perfcnt_enable(struct drm_device *dev, void *data,
213 				  struct drm_file *file_priv)
214 {
215 	struct panfrost_device *pfdev = dev->dev_private;
216 	struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
217 	struct drm_panfrost_perfcnt_enable *req = data;
218 	int ret;
219 
220 	ret = panfrost_unstable_ioctl_check();
221 	if (ret)
222 		return ret;
223 
224 	/* Only Bifrost GPUs have 2 set of counters. */
225 	if (req->counterset > (panfrost_model_is_bifrost(pfdev) ? 1 : 0))
226 		return -EINVAL;
227 
228 	mutex_lock(&perfcnt->lock);
229 	if (req->enable)
230 		ret = panfrost_perfcnt_enable_locked(pfdev, file_priv,
231 						     req->counterset);
232 	else
233 		ret = panfrost_perfcnt_disable_locked(pfdev, file_priv);
234 	mutex_unlock(&perfcnt->lock);
235 
236 	return ret;
237 }
238 
panfrost_ioctl_perfcnt_dump(struct drm_device * dev,void * data,struct drm_file * file_priv)239 int panfrost_ioctl_perfcnt_dump(struct drm_device *dev, void *data,
240 				struct drm_file *file_priv)
241 {
242 	struct panfrost_device *pfdev = dev->dev_private;
243 	struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
244 	struct drm_panfrost_perfcnt_dump *req = data;
245 	void __user *user_ptr = (void __user *)(uintptr_t)req->buf_ptr;
246 	int ret;
247 
248 	ret = panfrost_unstable_ioctl_check();
249 	if (ret)
250 		return ret;
251 
252 	mutex_lock(&perfcnt->lock);
253 	if (perfcnt->user != file_priv->driver_priv) {
254 		ret = -EINVAL;
255 		goto out;
256 	}
257 
258 	ret = panfrost_perfcnt_dump_locked(pfdev);
259 	if (ret)
260 		goto out;
261 
262 	if (copy_to_user(user_ptr, perfcnt->buf, perfcnt->bosize))
263 		ret = -EFAULT;
264 
265 out:
266 	mutex_unlock(&perfcnt->lock);
267 
268 	return ret;
269 }
270 
panfrost_perfcnt_close(struct drm_file * file_priv)271 void panfrost_perfcnt_close(struct drm_file *file_priv)
272 {
273 	struct panfrost_file_priv *pfile = file_priv->driver_priv;
274 	struct panfrost_device *pfdev = pfile->pfdev;
275 	struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
276 
277 	pm_runtime_get_sync(pfdev->dev);
278 	mutex_lock(&perfcnt->lock);
279 	if (perfcnt->user == pfile)
280 		panfrost_perfcnt_disable_locked(pfdev, file_priv);
281 	mutex_unlock(&perfcnt->lock);
282 	pm_runtime_mark_last_busy(pfdev->dev);
283 	pm_runtime_put_autosuspend(pfdev->dev);
284 }
285 
panfrost_perfcnt_init(struct panfrost_device * pfdev)286 int panfrost_perfcnt_init(struct panfrost_device *pfdev)
287 {
288 	struct panfrost_perfcnt *perfcnt;
289 	size_t size;
290 
291 	if (panfrost_has_hw_feature(pfdev, HW_FEATURE_V4)) {
292 		unsigned int ncoregroups;
293 
294 		ncoregroups = hweight64(pfdev->features.l2_present);
295 		size = ncoregroups * BLOCKS_PER_COREGROUP *
296 		       COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
297 	} else {
298 		unsigned int nl2c, ncores;
299 
300 		/*
301 		 * TODO: define a macro to extract the number of l2 caches from
302 		 * mem_features.
303 		 */
304 		nl2c = ((pfdev->features.mem_features >> 8) & GENMASK(3, 0)) + 1;
305 
306 		/*
307 		 * shader_present might be sparse, but the counters layout
308 		 * forces to dump unused regions too, hence the fls64() call
309 		 * instead of hweight64().
310 		 */
311 		ncores = fls64(pfdev->features.shader_present);
312 
313 		/*
314 		 * There's always one JM and one Tiler block, hence the '+ 2'
315 		 * here.
316 		 */
317 		size = (nl2c + ncores + 2) *
318 		       COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
319 	}
320 
321 	perfcnt = devm_kzalloc(pfdev->dev, sizeof(*perfcnt), GFP_KERNEL);
322 	if (!perfcnt)
323 		return -ENOMEM;
324 
325 	perfcnt->bosize = size;
326 
327 	/* Start with everything disabled. */
328 	gpu_write(pfdev, GPU_PERFCNT_CFG,
329 		  GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
330 	gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
331 	gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
332 	gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
333 	gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
334 
335 	init_completion(&perfcnt->dump_comp);
336 	mutex_init(&perfcnt->lock);
337 	pfdev->perfcnt = perfcnt;
338 
339 	return 0;
340 }
341 
panfrost_perfcnt_fini(struct panfrost_device * pfdev)342 void panfrost_perfcnt_fini(struct panfrost_device *pfdev)
343 {
344 	/* Disable everything before leaving. */
345 	gpu_write(pfdev, GPU_PERFCNT_CFG,
346 		  GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
347 	gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
348 	gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
349 	gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
350 	gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
351 }
352