int kbase_instr_hwcnt_enable(struct kbase_context *kctx,
struct kbase_uk_hwcnt_setup *setup)
{
struct kbase_device *kbdev;
int err;
kbdev = kctx->kbdev;
/* Mark the context as active so the GPU is kept turned on */
/* A suspend won't happen here, because we're in a syscall from a
* userspace thread. */
kbase_pm_context_active(kbdev);
/* Schedule the context in */
kbasep_js_schedule_privileged_ctx(kbdev, kctx);
err = kbase_instr_hwcnt_enable_internal(kbdev, kctx, setup);
if (err) {
/* Release the context. This had its own Power Manager Active
* reference */
kbasep_js_release_privileged_ctx(kbdev, kctx);
/* Also release our Power Manager Active reference */
kbase_pm_context_idle(kbdev);
}
return err;
}
int kbase_instr_hwcnt_enable(struct kbase_context *kctx,
struct kbase_uk_hwcnt_setup *setup)
{
struct kbase_device *kbdev;
bool access_allowed;
int err;
kbdev = kctx->kbdev;
/* Determine if the calling task has access to this capability */
access_allowed = kbase_security_has_capability(kctx,
KBASE_SEC_INSTR_HW_COUNTERS_COLLECT,
KBASE_SEC_FLAG_NOAUDIT);
if (!access_allowed)
return -EINVAL;
/* Mark the context as active so the GPU is kept turned on */
/* A suspend won't happen here, because we're in a syscall from a
* userspace thread. */
kbase_pm_context_active(kbdev);
/* Schedule the context in */
kbasep_js_schedule_privileged_ctx(kbdev, kctx);
err = kbase_instr_hwcnt_enable_internal(kbdev, kctx, setup);
if (err) {
/* Release the context. This had its own Power Manager Active
* reference */
kbasep_js_release_privileged_ctx(kbdev, kctx);
/* Also release our Power Manager Active reference */
kbase_pm_context_idle(kbdev);
}
return err;
}
STATIC mali_error kbase_instr_hwcnt_enable_internal(kbase_device *kbdev, kbase_context *kctx, kbase_uk_hwcnt_setup *setup)
{
unsigned long flags, pm_flags;
mali_error err = MALI_ERROR_FUNCTION_FAILED;
kbasep_js_device_data *js_devdata;
u32 irq_mask;
int ret;
u64 shader_cores_needed;
KBASE_DEBUG_ASSERT(NULL != kctx);
KBASE_DEBUG_ASSERT(NULL != kbdev);
KBASE_DEBUG_ASSERT(NULL != setup);
KBASE_DEBUG_ASSERT(NULL == kbdev->hwcnt.suspended_kctx);
shader_cores_needed = kbase_pm_get_present_cores(kbdev, KBASE_PM_CORE_SHADER);
js_devdata = &kbdev->js_data;
/* alignment failure */
if ((setup->dump_buffer == 0ULL) || (setup->dump_buffer & (2048 - 1)))
goto out_err;
/* Override core availability policy to ensure all cores are available */
kbase_pm_ca_instr_enable(kbdev);
/* Mark the context as active so the GPU is kept turned on */
/* A suspend won't happen here, because we're in a syscall from a userspace
* thread. */
kbase_pm_context_active(kbdev);
/* Request the cores early on synchronously - we'll release them on any errors
* (e.g. instrumentation already active) */
kbase_pm_request_cores_sync(kbdev, MALI_TRUE, shader_cores_needed);
spin_lock_irqsave(&kbdev->hwcnt.lock, flags);
if (kbdev->hwcnt.state == KBASE_INSTR_STATE_RESETTING) {
/* GPU is being reset */
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
wait_event(kbdev->hwcnt.wait, kbdev->hwcnt.triggered != 0);
spin_lock_irqsave(&kbdev->hwcnt.lock, flags);
}
if (kbdev->hwcnt.state != KBASE_INSTR_STATE_DISABLED) {
/* Instrumentation is already enabled */
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
goto out_unrequest_cores;
}
/* Enable interrupt */
spin_lock_irqsave(&kbdev->pm.power_change_lock, pm_flags);
irq_mask = kbase_reg_read(kbdev, GPU_CONTROL_REG(GPU_IRQ_MASK), NULL);
kbase_reg_write(kbdev, GPU_CONTROL_REG(GPU_IRQ_MASK), irq_mask | PRFCNT_SAMPLE_COMPLETED, NULL);
spin_unlock_irqrestore(&kbdev->pm.power_change_lock, pm_flags);
/* In use, this context is the owner */
kbdev->hwcnt.kctx = kctx;
/* Remember the dump address so we can reprogram it later */
kbdev->hwcnt.addr = setup->dump_buffer;
/* Remember all the settings for suspend/resume */
if (&kbdev->hwcnt.suspended_state != setup)
memcpy(&kbdev->hwcnt.suspended_state, setup, sizeof(kbdev->hwcnt.suspended_state));
/* Request the clean */
kbdev->hwcnt.state = KBASE_INSTR_STATE_REQUEST_CLEAN;
kbdev->hwcnt.triggered = 0;
/* Clean&invalidate the caches so we're sure the mmu tables for the dump buffer is valid */
ret = queue_work(kbdev->hwcnt.cache_clean_wq, &kbdev->hwcnt.cache_clean_work);
KBASE_DEBUG_ASSERT(ret);
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
/* Wait for cacheclean to complete */
wait_event(kbdev->hwcnt.wait, kbdev->hwcnt.triggered != 0);
KBASE_DEBUG_ASSERT(kbdev->hwcnt.state == KBASE_INSTR_STATE_IDLE);
/* Schedule the context in */
kbasep_js_schedule_privileged_ctx(kbdev, kctx);
/* Configure */
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_CONFIG), (kctx->as_nr << PRFCNT_CONFIG_AS_SHIFT) | PRFCNT_CONFIG_MODE_OFF, kctx);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_BASE_LO), setup->dump_buffer & 0xFFFFFFFF, kctx);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_BASE_HI), setup->dump_buffer >> 32, kctx);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_JM_EN), setup->jm_bm, kctx);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_SHADER_EN), setup->shader_bm, kctx);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_L3_CACHE_EN), setup->l3_cache_bm, kctx);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_MMU_L2_EN), setup->mmu_l2_bm, kctx);
/* Due to PRLAM-8186 we need to disable the Tiler before we enable the HW counter dump. */
if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8186))
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_TILER_EN), 0, kctx);
else
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_TILER_EN), setup->tiler_bm, kctx);
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_CONFIG), (kctx->as_nr << PRFCNT_CONFIG_AS_SHIFT) | PRFCNT_CONFIG_MODE_MANUAL, kctx);
/* If HW has PRLAM-8186 we can now re-enable the tiler HW counters dump */
if (kbase_hw_has_issue(kbdev, BASE_HW_ISSUE_8186))
kbase_reg_write(kbdev, GPU_CONTROL_REG(PRFCNT_TILER_EN), setup->tiler_bm, kctx);
spin_lock_irqsave(&kbdev->hwcnt.lock, flags);
if (kbdev->hwcnt.state == KBASE_INSTR_STATE_RESETTING) {
/* GPU is being reset */
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
wait_event(kbdev->hwcnt.wait, kbdev->hwcnt.triggered != 0);
spin_lock_irqsave(&kbdev->hwcnt.lock, flags);
}
kbdev->hwcnt.state = KBASE_INSTR_STATE_IDLE;
kbdev->hwcnt.triggered = 1;
wake_up(&kbdev->hwcnt.wait);
spin_unlock_irqrestore(&kbdev->hwcnt.lock, flags);
err = MALI_ERROR_NONE;
dev_dbg(kbdev->dev, "HW counters dumping set-up for context %p", kctx);
return err;
out_unrequest_cores:
kbase_pm_unrequest_cores(kbdev, MALI_TRUE, shader_cores_needed);
kbase_pm_context_idle(kbdev);
out_err:
return err;
}
