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/dma-buf-map.h>
9 #include <linux/iopoll.h>
10 #include <linux/pm_runtime.h>
11 #include <linux/slab.h>
12 #include <linux/uaccess.h>
14 #include "panfrost_device.h"
15 #include "panfrost_features.h"
16 #include "panfrost_gem.h"
17 #include "panfrost_issues.h"
18 #include "panfrost_job.h"
19 #include "panfrost_mmu.h"
20 #include "panfrost_perfcnt.h"
21 #include "panfrost_regs.h"
23 #define COUNTERS_PER_BLOCK 64
24 #define BYTES_PER_COUNTER 4
25 #define BLOCKS_PER_COREGROUP 8
26 #define V4_SHADERS_PER_COREGROUP 4
28 struct panfrost_perfcnt {
29 struct panfrost_gem_mapping *mapping;
32 struct panfrost_file_priv *user;
34 struct completion dump_comp;
37 void panfrost_perfcnt_clean_cache_done(struct panfrost_device *pfdev)
39 complete(&pfdev->perfcnt->dump_comp);
42 void panfrost_perfcnt_sample_done(struct panfrost_device *pfdev)
44 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_CACHES);
47 static int panfrost_perfcnt_dump_locked(struct panfrost_device *pfdev)
52 reinit_completion(&pfdev->perfcnt->dump_comp);
53 gpuva = pfdev->perfcnt->mapping->mmnode.start << PAGE_SHIFT;
54 gpu_write(pfdev, GPU_PERFCNT_BASE_LO, gpuva);
55 gpu_write(pfdev, GPU_PERFCNT_BASE_HI, gpuva >> 32);
56 gpu_write(pfdev, GPU_INT_CLEAR,
57 GPU_IRQ_CLEAN_CACHES_COMPLETED |
58 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
59 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_SAMPLE);
60 ret = wait_for_completion_interruptible_timeout(&pfdev->perfcnt->dump_comp,
61 msecs_to_jiffies(1000));
70 static int panfrost_perfcnt_enable_locked(struct panfrost_device *pfdev,
71 struct drm_file *file_priv,
72 unsigned int counterset)
74 struct panfrost_file_priv *user = file_priv->driver_priv;
75 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
76 struct dma_buf_map map;
77 struct drm_gem_shmem_object *bo;
81 if (user == perfcnt->user)
83 else if (perfcnt->user)
86 ret = pm_runtime_get_sync(pfdev->dev);
90 bo = drm_gem_shmem_create(pfdev->ddev, perfcnt->bosize);
96 /* Map the perfcnt buf in the address space attached to file_priv. */
97 ret = panfrost_gem_open(&bo->base, file_priv);
101 perfcnt->mapping = panfrost_gem_mapping_get(to_panfrost_bo(&bo->base),
103 if (!perfcnt->mapping) {
108 ret = drm_gem_shmem_vmap(&bo->base, &map);
110 goto err_put_mapping;
111 perfcnt->buf = map.vaddr;
114 * Invalidate the cache and clear the counters to start from a fresh
117 reinit_completion(&pfdev->perfcnt->dump_comp);
118 gpu_write(pfdev, GPU_INT_CLEAR,
119 GPU_IRQ_CLEAN_CACHES_COMPLETED |
120 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
121 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_CLEAR);
122 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_INV_CACHES);
123 ret = wait_for_completion_timeout(&pfdev->perfcnt->dump_comp,
124 msecs_to_jiffies(1000));
130 perfcnt->user = user;
132 as = panfrost_mmu_as_get(pfdev, perfcnt->mapping->mmu);
133 cfg = GPU_PERFCNT_CFG_AS(as) |
134 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_MANUAL);
137 * Bifrost GPUs have 2 set of counters, but we're only interested by
138 * the first one for now.
140 if (panfrost_model_is_bifrost(pfdev))
141 cfg |= GPU_PERFCNT_CFG_SETSEL(counterset);
143 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0xffffffff);
144 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0xffffffff);
145 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0xffffffff);
148 * Due to PRLAM-8186 we need to disable the Tiler before we enable HW
151 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
152 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
154 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);
156 gpu_write(pfdev, GPU_PERFCNT_CFG, cfg);
158 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
159 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);
161 /* The BO ref is retained by the mapping. */
162 drm_gem_object_put(&bo->base);
167 drm_gem_shmem_vunmap(&bo->base, &map);
169 panfrost_gem_mapping_put(perfcnt->mapping);
171 panfrost_gem_close(&bo->base, file_priv);
173 drm_gem_object_put(&bo->base);
175 pm_runtime_put(pfdev->dev);
179 static int panfrost_perfcnt_disable_locked(struct panfrost_device *pfdev,
180 struct drm_file *file_priv)
182 struct panfrost_file_priv *user = file_priv->driver_priv;
183 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
184 struct dma_buf_map map = DMA_BUF_MAP_INIT_VADDR(perfcnt->buf);
186 if (user != perfcnt->user)
189 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0x0);
190 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0x0);
191 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0x0);
192 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
193 gpu_write(pfdev, GPU_PERFCNT_CFG,
194 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
196 perfcnt->user = NULL;
197 drm_gem_shmem_vunmap(&perfcnt->mapping->obj->base.base, &map);
199 panfrost_gem_close(&perfcnt->mapping->obj->base.base, file_priv);
200 panfrost_mmu_as_put(pfdev, perfcnt->mapping->mmu);
201 panfrost_gem_mapping_put(perfcnt->mapping);
202 perfcnt->mapping = NULL;
203 pm_runtime_mark_last_busy(pfdev->dev);
204 pm_runtime_put_autosuspend(pfdev->dev);
209 int panfrost_ioctl_perfcnt_enable(struct drm_device *dev, void *data,
210 struct drm_file *file_priv)
212 struct panfrost_device *pfdev = dev->dev_private;
213 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
214 struct drm_panfrost_perfcnt_enable *req = data;
217 ret = panfrost_unstable_ioctl_check();
221 /* Only Bifrost GPUs have 2 set of counters. */
222 if (req->counterset > (panfrost_model_is_bifrost(pfdev) ? 1 : 0))
225 mutex_lock(&perfcnt->lock);
227 ret = panfrost_perfcnt_enable_locked(pfdev, file_priv,
230 ret = panfrost_perfcnt_disable_locked(pfdev, file_priv);
231 mutex_unlock(&perfcnt->lock);
236 int panfrost_ioctl_perfcnt_dump(struct drm_device *dev, void *data,
237 struct drm_file *file_priv)
239 struct panfrost_device *pfdev = dev->dev_private;
240 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
241 struct drm_panfrost_perfcnt_dump *req = data;
242 void __user *user_ptr = (void __user *)(uintptr_t)req->buf_ptr;
245 ret = panfrost_unstable_ioctl_check();
249 mutex_lock(&perfcnt->lock);
250 if (perfcnt->user != file_priv->driver_priv) {
255 ret = panfrost_perfcnt_dump_locked(pfdev);
259 if (copy_to_user(user_ptr, perfcnt->buf, perfcnt->bosize))
263 mutex_unlock(&perfcnt->lock);
268 void panfrost_perfcnt_close(struct drm_file *file_priv)
270 struct panfrost_file_priv *pfile = file_priv->driver_priv;
271 struct panfrost_device *pfdev = pfile->pfdev;
272 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
274 pm_runtime_get_sync(pfdev->dev);
275 mutex_lock(&perfcnt->lock);
276 if (perfcnt->user == pfile)
277 panfrost_perfcnt_disable_locked(pfdev, file_priv);
278 mutex_unlock(&perfcnt->lock);
279 pm_runtime_mark_last_busy(pfdev->dev);
280 pm_runtime_put_autosuspend(pfdev->dev);
283 int panfrost_perfcnt_init(struct panfrost_device *pfdev)
285 struct panfrost_perfcnt *perfcnt;
288 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_V4)) {
289 unsigned int ncoregroups;
291 ncoregroups = hweight64(pfdev->features.l2_present);
292 size = ncoregroups * BLOCKS_PER_COREGROUP *
293 COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
295 unsigned int nl2c, ncores;
298 * TODO: define a macro to extract the number of l2 caches from
301 nl2c = ((pfdev->features.mem_features >> 8) & GENMASK(3, 0)) + 1;
304 * shader_present might be sparse, but the counters layout
305 * forces to dump unused regions too, hence the fls64() call
306 * instead of hweight64().
308 ncores = fls64(pfdev->features.shader_present);
311 * There's always one JM and one Tiler block, hence the '+ 2'
314 size = (nl2c + ncores + 2) *
315 COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
318 perfcnt = devm_kzalloc(pfdev->dev, sizeof(*perfcnt), GFP_KERNEL);
322 perfcnt->bosize = size;
324 /* Start with everything disabled. */
325 gpu_write(pfdev, GPU_PERFCNT_CFG,
326 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
327 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
328 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
329 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
330 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
332 init_completion(&perfcnt->dump_comp);
333 mutex_init(&perfcnt->lock);
334 pfdev->perfcnt = perfcnt;
339 void panfrost_perfcnt_fini(struct panfrost_device *pfdev)
341 /* Disable everything before leaving. */
342 gpu_write(pfdev, GPU_PERFCNT_CFG,
343 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
344 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
345 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
346 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
347 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);