1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright 2019 Collabora Ltd */
4 #include <drm/drm_file.h>
5 #include <drm/drm_gem_shmem_helper.h>
6 #include <drm/panfrost_drm.h>
7 #include <linux/completion.h>
8 #include <linux/iopoll.h>
9 #include <linux/pm_runtime.h>
10 #include <linux/slab.h>
11 #include <linux/uaccess.h>
13 #include "panfrost_device.h"
14 #include "panfrost_features.h"
15 #include "panfrost_gem.h"
16 #include "panfrost_issues.h"
17 #include "panfrost_job.h"
18 #include "panfrost_mmu.h"
19 #include "panfrost_perfcnt.h"
20 #include "panfrost_regs.h"
22 #define COUNTERS_PER_BLOCK 64
23 #define BYTES_PER_COUNTER 4
24 #define BLOCKS_PER_COREGROUP 8
25 #define V4_SHADERS_PER_COREGROUP 4
27 struct panfrost_perfcnt {
28 struct panfrost_gem_object *bo;
31 struct panfrost_file_priv *user;
33 struct completion dump_comp;
36 void panfrost_perfcnt_clean_cache_done(struct panfrost_device *pfdev)
38 complete(&pfdev->perfcnt->dump_comp);
41 void panfrost_perfcnt_sample_done(struct panfrost_device *pfdev)
43 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_CACHES);
46 static int panfrost_perfcnt_dump_locked(struct panfrost_device *pfdev)
51 reinit_completion(&pfdev->perfcnt->dump_comp);
52 gpuva = pfdev->perfcnt->bo->node.start << PAGE_SHIFT;
53 gpu_write(pfdev, GPU_PERFCNT_BASE_LO, gpuva);
54 gpu_write(pfdev, GPU_PERFCNT_BASE_HI, gpuva >> 32);
55 gpu_write(pfdev, GPU_INT_CLEAR,
56 GPU_IRQ_CLEAN_CACHES_COMPLETED |
57 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
58 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_SAMPLE);
59 ret = wait_for_completion_interruptible_timeout(&pfdev->perfcnt->dump_comp,
60 msecs_to_jiffies(1000));
69 static int panfrost_perfcnt_enable_locked(struct panfrost_device *pfdev,
70 struct drm_file *file_priv,
71 unsigned int counterset)
73 struct panfrost_file_priv *user = file_priv->driver_priv;
74 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
75 struct drm_gem_shmem_object *bo;
79 if (user == perfcnt->user)
81 else if (perfcnt->user)
84 ret = pm_runtime_get_sync(pfdev->dev);
88 bo = drm_gem_shmem_create(pfdev->ddev, perfcnt->bosize);
92 perfcnt->bo = to_panfrost_bo(&bo->base);
94 /* Map the perfcnt buf in the address space attached to file_priv. */
95 ret = panfrost_gem_open(&perfcnt->bo->base.base, file_priv);
99 perfcnt->buf = drm_gem_shmem_vmap(&bo->base);
100 if (IS_ERR(perfcnt->buf)) {
101 ret = PTR_ERR(perfcnt->buf);
106 * Invalidate the cache and clear the counters to start from a fresh
109 reinit_completion(&pfdev->perfcnt->dump_comp);
110 gpu_write(pfdev, GPU_INT_CLEAR,
111 GPU_IRQ_CLEAN_CACHES_COMPLETED |
112 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
113 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_CLEAR);
114 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_INV_CACHES);
115 ret = wait_for_completion_timeout(&pfdev->perfcnt->dump_comp,
116 msecs_to_jiffies(1000));
122 perfcnt->user = user;
125 * Always use address space 0 for now.
126 * FIXME: this needs to be updated when we start using different
129 cfg = GPU_PERFCNT_CFG_AS(0) |
130 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_MANUAL);
133 * Bifrost GPUs have 2 set of counters, but we're only interested by
134 * the first one for now.
136 if (panfrost_model_is_bifrost(pfdev))
137 cfg |= GPU_PERFCNT_CFG_SETSEL(counterset);
139 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0xffffffff);
140 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0xffffffff);
141 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0xffffffff);
144 * Due to PRLAM-8186 we need to disable the Tiler before we enable HW
147 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
148 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
150 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);
152 gpu_write(pfdev, GPU_PERFCNT_CFG, cfg);
154 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
155 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);
160 drm_gem_shmem_vunmap(&perfcnt->bo->base.base, perfcnt->buf);
162 panfrost_gem_close(&perfcnt->bo->base.base, file_priv);
164 drm_gem_object_put_unlocked(&bo->base);
168 static int panfrost_perfcnt_disable_locked(struct panfrost_device *pfdev,
169 struct drm_file *file_priv)
171 struct panfrost_file_priv *user = file_priv->driver_priv;
172 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
174 if (user != perfcnt->user)
177 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0x0);
178 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0x0);
179 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0x0);
180 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
181 gpu_write(pfdev, GPU_PERFCNT_CFG,
182 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
184 perfcnt->user = NULL;
185 drm_gem_shmem_vunmap(&perfcnt->bo->base.base, perfcnt->buf);
187 panfrost_gem_close(&perfcnt->bo->base.base, file_priv);
188 drm_gem_object_put_unlocked(&perfcnt->bo->base.base);
190 pm_runtime_mark_last_busy(pfdev->dev);
191 pm_runtime_put_autosuspend(pfdev->dev);
196 int panfrost_ioctl_perfcnt_enable(struct drm_device *dev, void *data,
197 struct drm_file *file_priv)
199 struct panfrost_device *pfdev = dev->dev_private;
200 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
201 struct drm_panfrost_perfcnt_enable *req = data;
204 ret = panfrost_unstable_ioctl_check();
208 /* Only Bifrost GPUs have 2 set of counters. */
209 if (req->counterset > (panfrost_model_is_bifrost(pfdev) ? 1 : 0))
212 mutex_lock(&perfcnt->lock);
214 ret = panfrost_perfcnt_enable_locked(pfdev, file_priv,
217 ret = panfrost_perfcnt_disable_locked(pfdev, file_priv);
218 mutex_unlock(&perfcnt->lock);
223 int panfrost_ioctl_perfcnt_dump(struct drm_device *dev, void *data,
224 struct drm_file *file_priv)
226 struct panfrost_device *pfdev = dev->dev_private;
227 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
228 struct drm_panfrost_perfcnt_dump *req = data;
229 void __user *user_ptr = (void __user *)(uintptr_t)req->buf_ptr;
232 ret = panfrost_unstable_ioctl_check();
236 mutex_lock(&perfcnt->lock);
237 if (perfcnt->user != file_priv->driver_priv) {
242 ret = panfrost_perfcnt_dump_locked(pfdev);
246 if (copy_to_user(user_ptr, perfcnt->buf, perfcnt->bosize))
250 mutex_unlock(&perfcnt->lock);
255 void panfrost_perfcnt_close(struct drm_file *file_priv)
257 struct panfrost_file_priv *pfile = file_priv->driver_priv;
258 struct panfrost_device *pfdev = pfile->pfdev;
259 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
261 pm_runtime_get_sync(pfdev->dev);
262 mutex_lock(&perfcnt->lock);
263 if (perfcnt->user == pfile)
264 panfrost_perfcnt_disable_locked(pfdev, file_priv);
265 mutex_unlock(&perfcnt->lock);
266 pm_runtime_mark_last_busy(pfdev->dev);
267 pm_runtime_put_autosuspend(pfdev->dev);
270 int panfrost_perfcnt_init(struct panfrost_device *pfdev)
272 struct panfrost_perfcnt *perfcnt;
275 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_V4)) {
276 unsigned int ncoregroups;
278 ncoregroups = hweight64(pfdev->features.l2_present);
279 size = ncoregroups * BLOCKS_PER_COREGROUP *
280 COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
282 unsigned int nl2c, ncores;
285 * TODO: define a macro to extract the number of l2 caches from
288 nl2c = ((pfdev->features.mem_features >> 8) & GENMASK(3, 0)) + 1;
291 * shader_present might be sparse, but the counters layout
292 * forces to dump unused regions too, hence the fls64() call
293 * instead of hweight64().
295 ncores = fls64(pfdev->features.shader_present);
298 * There's always one JM and one Tiler block, hence the '+ 2'
301 size = (nl2c + ncores + 2) *
302 COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
305 perfcnt = devm_kzalloc(pfdev->dev, sizeof(*perfcnt), GFP_KERNEL);
309 perfcnt->bosize = size;
311 /* Start with everything disabled. */
312 gpu_write(pfdev, GPU_PERFCNT_CFG,
313 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
314 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
315 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
316 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
317 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
319 init_completion(&perfcnt->dump_comp);
320 mutex_init(&perfcnt->lock);
321 pfdev->perfcnt = perfcnt;
326 void panfrost_perfcnt_fini(struct panfrost_device *pfdev)
328 /* Disable everything before leaving. */
329 gpu_write(pfdev, GPU_PERFCNT_CFG,
330 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
331 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
332 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
333 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
334 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);