Quelle  panfrost_perfcnt.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/* Copyright 2019 Collabora Ltd */

#include <linux/completion.h>
#include <linux/iopoll.h>
#include <linux/iosys-map.h>
#include <linux/pm_runtime.h>
#include <linux/slab.h>
#include <linux/uaccess.h>

#include <drm/drm_file.h>
#include <drm/drm_gem_shmem_helper.h>
#include <drm/panfrost_drm.h>

#include "panfrost_device.h"
#include "panfrost_features.h"
#include "panfrost_gem.h"
#include "panfrost_issues.h"
#include "panfrost_job.h"
#include "panfrost_mmu.h"
#include "panfrost_perfcnt.h"
#include "panfrost_regs.h"

#define COUNTERS_PER_BLOCK  64
#define BYTES_PER_COUNTER  4
#define BLOCKS_PER_COREGROUP  8
#define V4_SHADERS_PER_COREGROUP 4

struct panfrost_perfcnt {
 struct panfrost_gem_mapping *mapping;
 size_t bosize;
 void *buf;
 struct panfrost_file_priv *user;
 struct mutex lock;
 struct completion dump_comp;
};

void panfrost_perfcnt_clean_cache_done(struct panfrost_device *pfdev)
{
 complete(&pfdev->perfcnt->dump_comp);
}

void panfrost_perfcnt_sample_done(struct panfrost_device *pfdev)
{
 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_CACHES);
}

static int panfrost_perfcnt_dump_locked(struct panfrost_device *pfdev)
{
 u64 gpuva;
 int ret;

 reinit_completion(&pfdev->perfcnt->dump_comp);
 gpuva = pfdev->perfcnt->mapping->mmnode.start << PAGE_SHIFT;
 gpu_write(pfdev, GPU_PERFCNT_BASE_LO, lower_32_bits(gpuva));
 gpu_write(pfdev, GPU_PERFCNT_BASE_HI, upper_32_bits(gpuva));
 gpu_write(pfdev, GPU_INT_CLEAR,
    GPU_IRQ_CLEAN_CACHES_COMPLETED |
    GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_SAMPLE);
 ret = wait_for_completion_interruptible_timeout(&pfdev->perfcnt->dump_comp,
       msecs_to_jiffies(1000));
 if (!ret)
  ret = -ETIMEDOUT;
 else if (ret > 0)
  ret = 0;

 return ret;
}

static int panfrost_perfcnt_enable_locked(struct panfrost_device *pfdev,
       struct drm_file *file_priv,
       unsigned int counterset)
{
 struct panfrost_file_priv *user = file_priv->driver_priv;
 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
 struct iosys_map map;
 struct drm_gem_shmem_object *bo;
 u32 cfg, as;
 int ret;

 if (user == perfcnt->user)
  return 0;
 else if (perfcnt->user)
  return -EBUSY;

 ret = pm_runtime_get_sync(pfdev->dev);
 if (ret < 0)
  goto err_put_pm;

 bo = drm_gem_shmem_create(pfdev->ddev, perfcnt->bosize);
 if (IS_ERR(bo)) {
  ret = PTR_ERR(bo);
  goto err_put_pm;
 }

 /* Map the perfcnt buf in the address space attached to file_priv. */
 ret = panfrost_gem_open(&bo->base, file_priv);
 if (ret)
  goto err_put_bo;

 perfcnt->mapping = panfrost_gem_mapping_get(to_panfrost_bo(&bo->base),
          user);
 if (!perfcnt->mapping) {
  ret = -EINVAL;
  goto err_close_bo;
 }

 ret = drm_gem_vmap(&bo->base, &map);
 if (ret)
  goto err_put_mapping;
 perfcnt->buf = map.vaddr;

 panfrost_gem_internal_set_label(&bo->base, "Perfcnt sample buffer");

 /*
 * Invalidate the cache and clear the counters to start from a fresh
 * state.
 */
 reinit_completion(&pfdev->perfcnt->dump_comp);
 gpu_write(pfdev, GPU_INT_CLEAR,
    GPU_IRQ_CLEAN_CACHES_COMPLETED |
    GPU_IRQ_PERFCNT_SAMPLE_COMPLETED);
 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_CLEAR);
 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_INV_CACHES);
 ret = wait_for_completion_timeout(&pfdev->perfcnt->dump_comp,
       msecs_to_jiffies(1000));
 if (!ret) {
  ret = -ETIMEDOUT;
  goto err_vunmap;
 }

 perfcnt->user = user;

 as = panfrost_mmu_as_get(pfdev, perfcnt->mapping->mmu);
 cfg = GPU_PERFCNT_CFG_AS(as) |
       GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_MANUAL);

 /*
 * Bifrost GPUs have 2 set of counters, but we're only interested by
 * the first one for now.
 */
 if (panfrost_model_is_bifrost(pfdev))
  cfg |= GPU_PERFCNT_CFG_SETSEL(counterset);

 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0xffffffff);
 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0xffffffff);
 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0xffffffff);

 /*
 * Due to PRLAM-8186 we need to disable the Tiler before we enable HW
 * counters.
 */
 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
  gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
 else
  gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);

 gpu_write(pfdev, GPU_PERFCNT_CFG, cfg);

 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186))
  gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff);

 /* The BO ref is retained by the mapping. */
 drm_gem_object_put(&bo->base);

 return 0;

err_vunmap:
 drm_gem_vunmap(&bo->base, &map);
err_put_mapping:
 panfrost_gem_mapping_put(perfcnt->mapping);
err_close_bo:
 panfrost_gem_close(&bo->base, file_priv);
err_put_bo:
 drm_gem_object_put(&bo->base);
err_put_pm:
 pm_runtime_put(pfdev->dev);
 return ret;
}

static int panfrost_perfcnt_disable_locked(struct panfrost_device *pfdev,
        struct drm_file *file_priv)
{
 struct panfrost_file_priv *user = file_priv->driver_priv;
 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
 struct iosys_map map = IOSYS_MAP_INIT_VADDR(perfcnt->buf);

 if (user != perfcnt->user)
  return -EINVAL;

 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0x0);
 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0x0);
 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0x0);
 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
 gpu_write(pfdev, GPU_PERFCNT_CFG,
    GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));

 perfcnt->user = NULL;
 drm_gem_vunmap(&perfcnt->mapping->obj->base.base, &map);
 perfcnt->buf = NULL;
 panfrost_gem_close(&perfcnt->mapping->obj->base.base, file_priv);
 panfrost_mmu_as_put(pfdev, perfcnt->mapping->mmu);
 panfrost_gem_mapping_put(perfcnt->mapping);
 perfcnt->mapping = NULL;
 pm_runtime_mark_last_busy(pfdev->dev);
 pm_runtime_put_autosuspend(pfdev->dev);

 return 0;
}

int panfrost_ioctl_perfcnt_enable(struct drm_device *dev, void *data,
      struct drm_file *file_priv)
{
 struct panfrost_device *pfdev = dev->dev_private;
 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
 struct drm_panfrost_perfcnt_enable *req = data;
 int ret;

 ret = panfrost_unstable_ioctl_check();
 if (ret)
  return ret;

 /* Only Bifrost GPUs have 2 set of counters. */
 if (req->counterset > (panfrost_model_is_bifrost(pfdev) ? 1 : 0))
  return -EINVAL;

 mutex_lock(&perfcnt->lock);
 if (req->enable)
  ret = panfrost_perfcnt_enable_locked(pfdev, file_priv,
           req->counterset);
 else
  ret = panfrost_perfcnt_disable_locked(pfdev, file_priv);
 mutex_unlock(&perfcnt->lock);

 return ret;
}

int panfrost_ioctl_perfcnt_dump(struct drm_device *dev, void *data,
    struct drm_file *file_priv)
{
 struct panfrost_device *pfdev = dev->dev_private;
 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;
 struct drm_panfrost_perfcnt_dump *req = data;
 void __user *user_ptr = (void __user *)(uintptr_t)req->buf_ptr;
 int ret;

 ret = panfrost_unstable_ioctl_check();
 if (ret)
  return ret;

 mutex_lock(&perfcnt->lock);
 if (perfcnt->user != file_priv->driver_priv) {
  ret = -EINVAL;
  goto out;
 }

 ret = panfrost_perfcnt_dump_locked(pfdev);
 if (ret)
  goto out;

 if (copy_to_user(user_ptr, perfcnt->buf, perfcnt->bosize))
  ret = -EFAULT;

out:
 mutex_unlock(&perfcnt->lock);

 return ret;
}

void panfrost_perfcnt_close(struct drm_file *file_priv)
{
 struct panfrost_file_priv *pfile = file_priv->driver_priv;
 struct panfrost_device *pfdev = pfile->pfdev;
 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt;

 pm_runtime_get_sync(pfdev->dev);
 mutex_lock(&perfcnt->lock);
 if (perfcnt->user == pfile)
  panfrost_perfcnt_disable_locked(pfdev, file_priv);
 mutex_unlock(&perfcnt->lock);
 pm_runtime_mark_last_busy(pfdev->dev);
 pm_runtime_put_autosuspend(pfdev->dev);
}

int panfrost_perfcnt_init(struct panfrost_device *pfdev)
{
 struct panfrost_perfcnt *perfcnt;
 size_t size;

 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_V4)) {
  unsigned int ncoregroups;

  ncoregroups = hweight64(pfdev->features.l2_present);
  size = ncoregroups * BLOCKS_PER_COREGROUP *
         COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
 } else {
  unsigned int nl2c, ncores;

  /*
 * TODO: define a macro to extract the number of l2 caches from
 * mem_features.
 */
  nl2c = ((pfdev->features.mem_features >> 8) & GENMASK(3, 0)) + 1;

  /*
 * shader_present might be sparse, but the counters layout
 * forces to dump unused regions too, hence the fls64() call
 * instead of hweight64().
 */
  ncores = fls64(pfdev->features.shader_present);

  /*
 * There's always one JM and one Tiler block, hence the '+ 2'
 * here.
 */
  size = (nl2c + ncores + 2) *
         COUNTERS_PER_BLOCK * BYTES_PER_COUNTER;
 }

 perfcnt = devm_kzalloc(pfdev->dev, sizeof(*perfcnt), GFP_KERNEL);
 if (!perfcnt)
  return -ENOMEM;

 perfcnt->bosize = size;

 /* Start with everything disabled. */
 gpu_write(pfdev, GPU_PERFCNT_CFG,
    GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);

 init_completion(&perfcnt->dump_comp);
 mutex_init(&perfcnt->lock);
 pfdev->perfcnt = perfcnt;

 return 0;
}

void panfrost_perfcnt_fini(struct panfrost_device *pfdev)
{
 /* Disable everything before leaving. */
 gpu_write(pfdev, GPU_PERFCNT_CFG,
    GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF));
 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0);
 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0);
 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0);
 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0);
}