diff options
Diffstat (limited to 'mali_kbase/csf/mali_kbase_csf_reset_gpu.c')
| -rw-r--r-- | mali_kbase/csf/mali_kbase_csf_reset_gpu.c | 140 |
1 files changed, 64 insertions, 76 deletions
diff --git a/mali_kbase/csf/mali_kbase_csf_reset_gpu.c b/mali_kbase/csf/mali_kbase_csf_reset_gpu.c index 952209f..240397e 100644 --- a/mali_kbase/csf/mali_kbase_csf_reset_gpu.c +++ b/mali_kbase/csf/mali_kbase_csf_reset_gpu.c @@ -38,21 +38,18 @@ enum kbasep_soft_reset_status { MCU_REINIT_FAILED }; -static inline bool -kbase_csf_reset_state_is_silent(enum kbase_csf_reset_gpu_state state) +static inline bool kbase_csf_reset_state_is_silent(enum kbase_csf_reset_gpu_state state) { return (state == KBASE_CSF_RESET_GPU_COMMITTED_SILENT); } -static inline bool -kbase_csf_reset_state_is_committed(enum kbase_csf_reset_gpu_state state) +static inline bool kbase_csf_reset_state_is_committed(enum kbase_csf_reset_gpu_state state) { return (state == KBASE_CSF_RESET_GPU_COMMITTED || state == KBASE_CSF_RESET_GPU_COMMITTED_SILENT); } -static inline bool -kbase_csf_reset_state_is_active(enum kbase_csf_reset_gpu_state state) +static inline bool kbase_csf_reset_state_is_active(enum kbase_csf_reset_gpu_state state) { return (state == KBASE_CSF_RESET_GPU_HAPPENING); } @@ -100,8 +97,7 @@ int kbase_reset_gpu_prevent_and_wait(struct kbase_device *kbdev) { down_read(&kbdev->csf.reset.sem); - if (atomic_read(&kbdev->csf.reset.state) == - KBASE_CSF_RESET_GPU_FAILED) { + if (atomic_read(&kbdev->csf.reset.state) == KBASE_CSF_RESET_GPU_FAILED) { up_read(&kbdev->csf.reset.sem); return -ENOMEM; } @@ -120,8 +116,7 @@ int kbase_reset_gpu_try_prevent(struct kbase_device *kbdev) if (!down_read_trylock(&kbdev->csf.reset.sem)) return -EAGAIN; - if (atomic_read(&kbdev->csf.reset.state) == - KBASE_CSF_RESET_GPU_FAILED) { + if (atomic_read(&kbdev->csf.reset.state) == KBASE_CSF_RESET_GPU_FAILED) { up_read(&kbdev->csf.reset.sem); return -ENOMEM; } @@ -166,9 +161,8 @@ void kbase_reset_gpu_assert_failed_or_prevented(struct kbase_device *kbdev) /* Mark the reset as now happening, and synchronize with other threads that * might be trying to access the GPU */ -static void kbase_csf_reset_begin_hw_access_sync( - struct kbase_device *kbdev, - enum kbase_csf_reset_gpu_state initial_reset_state) +static void kbase_csf_reset_begin_hw_access_sync(struct kbase_device *kbdev, + enum kbase_csf_reset_gpu_state initial_reset_state) { unsigned long hwaccess_lock_flags; unsigned long scheduler_spin_lock_flags; @@ -193,15 +187,13 @@ static void kbase_csf_reset_begin_hw_access_sync( /* Mark the reset as finished and allow others threads to once more access the * GPU */ -static void kbase_csf_reset_end_hw_access(struct kbase_device *kbdev, - int err_during_reset, +static void kbase_csf_reset_end_hw_access(struct kbase_device *kbdev, int err_during_reset, bool firmware_inited) { unsigned long hwaccess_lock_flags; unsigned long scheduler_spin_lock_flags; - WARN_ON(!kbase_csf_reset_state_is_active( - atomic_read(&kbdev->csf.reset.state))); + WARN_ON(!kbase_csf_reset_state_is_active(atomic_read(&kbdev->csf.reset.state))); /* Once again, we synchronize with atomic context threads accessing the * HW, as otherwise any actions they defer could get lost @@ -210,8 +202,7 @@ static void kbase_csf_reset_end_hw_access(struct kbase_device *kbdev, kbase_csf_scheduler_spin_lock(kbdev, &scheduler_spin_lock_flags); if (!err_during_reset) { - atomic_set(&kbdev->csf.reset.state, - KBASE_CSF_RESET_GPU_NOT_PENDING); + atomic_set(&kbdev->csf.reset.state, KBASE_CSF_RESET_GPU_NOT_PENDING); } else { dev_err(kbdev->dev, "Reset failed to complete"); atomic_set(&kbdev->csf.reset.state, KBASE_CSF_RESET_GPU_FAILED); @@ -236,26 +227,30 @@ static void kbase_csf_debug_dump_registers(struct kbase_device *kbdev) kbase_io_history_dump(kbdev); dev_err(kbdev->dev, "Register state:"); - dev_err(kbdev->dev, " GPU_IRQ_RAWSTAT=0x%08x GPU_STATUS=0x%08x MCU_STATUS=0x%08x", - kbase_reg_read(kbdev, GPU_CONTROL_REG(GPU_IRQ_RAWSTAT)), - kbase_reg_read(kbdev, GPU_CONTROL_REG(GPU_STATUS)), - kbase_reg_read(kbdev, GPU_CONTROL_REG(MCU_STATUS))); + dev_err(kbdev->dev, " GPU_IRQ_RAWSTAT=0x%08x GPU_STATUS=0x%08x MCU_STATUS=0x%08x", + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(GPU_IRQ_RAWSTAT)), + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(GPU_STATUS)), + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(MCU_STATUS))); dev_err(kbdev->dev, - " JOB_IRQ_RAWSTAT=0x%08x MMU_IRQ_RAWSTAT=0x%08x GPU_FAULTSTATUS=0x%08x", - kbase_reg_read(kbdev, JOB_CONTROL_REG(JOB_IRQ_RAWSTAT)), - kbase_reg_read(kbdev, MMU_CONTROL_REG(MMU_IRQ_RAWSTAT)), - kbase_reg_read(kbdev, GPU_CONTROL_REG(GPU_FAULTSTATUS))); - dev_err(kbdev->dev, " GPU_IRQ_MASK=0x%08x JOB_IRQ_MASK=0x%08x MMU_IRQ_MASK=0x%08x", - kbase_reg_read(kbdev, GPU_CONTROL_REG(GPU_IRQ_MASK)), - kbase_reg_read(kbdev, JOB_CONTROL_REG(JOB_IRQ_MASK)), - kbase_reg_read(kbdev, MMU_CONTROL_REG(MMU_IRQ_MASK))); - dev_err(kbdev->dev, " PWR_OVERRIDE0=0x%08x PWR_OVERRIDE1=0x%08x", - kbase_reg_read(kbdev, GPU_CONTROL_REG(PWR_OVERRIDE0)), - kbase_reg_read(kbdev, GPU_CONTROL_REG(PWR_OVERRIDE1))); - dev_err(kbdev->dev, " SHADER_CONFIG=0x%08x L2_MMU_CONFIG=0x%08x TILER_CONFIG=0x%08x", - kbase_reg_read(kbdev, GPU_CONTROL_REG(SHADER_CONFIG)), - kbase_reg_read(kbdev, GPU_CONTROL_REG(L2_MMU_CONFIG)), - kbase_reg_read(kbdev, GPU_CONTROL_REG(TILER_CONFIG))); + " JOB_IRQ_RAWSTAT=0x%08x MMU_IRQ_RAWSTAT=0x%08x GPU_FAULTSTATUS=0x%08x", + kbase_reg_read32(kbdev, JOB_CONTROL_ENUM(JOB_IRQ_RAWSTAT)), + kbase_reg_read32(kbdev, MMU_CONTROL_ENUM(IRQ_RAWSTAT)), + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(GPU_FAULTSTATUS))); + dev_err(kbdev->dev, " GPU_IRQ_MASK=0x%08x JOB_IRQ_MASK=0x%08x MMU_IRQ_MASK=0x%08x", + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(GPU_IRQ_MASK)), + kbase_reg_read32(kbdev, JOB_CONTROL_ENUM(JOB_IRQ_MASK)), + kbase_reg_read32(kbdev, MMU_CONTROL_ENUM(IRQ_MASK))); + if (kbdev->gpu_props.gpu_id.arch_id < GPU_ID_ARCH_MAKE(14, 10, 0)) { + dev_err(kbdev->dev, " PWR_OVERRIDE0=0x%08x PWR_OVERRIDE1=0x%08x", + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(PWR_OVERRIDE0)), + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(PWR_OVERRIDE1))); + dev_err(kbdev->dev, + " SHADER_CONFIG=0x%08x L2_MMU_CONFIG=0x%08x TILER_CONFIG=0x%08x", + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(SHADER_CONFIG)), + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(L2_MMU_CONFIG)), + kbase_reg_read32(kbdev, GPU_CONTROL_ENUM(TILER_CONFIG))); + } + } /** @@ -294,8 +289,7 @@ static enum kbasep_soft_reset_status kbase_csf_reset_gpu_once(struct kbase_devic spin_lock(&kbdev->mmu_mask_change); kbase_pm_reset_start_locked(kbdev); - dev_dbg(kbdev->dev, - "We're about to flush out the IRQs and their bottom halves\n"); + dev_dbg(kbdev->dev, "We're about to flush out the IRQs and their bottom halves\n"); kbdev->irq_reset_flush = true; /* Disable IRQ to avoid IRQ handlers to kick in after releasing the @@ -313,13 +307,11 @@ static enum kbasep_soft_reset_status kbase_csf_reset_gpu_once(struct kbase_devic dev_dbg(kbdev->dev, "Flush out any in-flight work items\n"); kbase_flush_mmu_wqs(kbdev); - dev_dbg(kbdev->dev, - "The flush has completed so reset the active indicator\n"); + dev_dbg(kbdev->dev, "The flush has completed so reset the active indicator\n"); kbdev->irq_reset_flush = false; if (!silent) - dev_err(kbdev->dev, "Resetting GPU (allowing up to %d ms)", - RESET_TIMEOUT); + dev_err(kbdev->dev, "Resetting GPU (allowing up to %d ms)", RESET_TIMEOUT); /* Output the state of some interesting registers to help in the * debugging of GPU resets, and dump the firmware trace buffer @@ -330,9 +322,11 @@ static enum kbasep_soft_reset_status kbase_csf_reset_gpu_once(struct kbase_devic kbase_csf_firmware_log_dump_buffer(kbdev); } - spin_lock_irqsave(&kbdev->hwaccess_lock, flags); - kbase_ipa_control_handle_gpu_reset_pre(kbdev); - spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags); + { + spin_lock_irqsave(&kbdev->hwaccess_lock, flags); + kbase_ipa_control_handle_gpu_reset_pre(kbdev); + spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags); + } /* Tell hardware counters a reset is about to occur. * If the backend is in an unrecoverable error state (e.g. due to @@ -353,7 +347,9 @@ static enum kbasep_soft_reset_status kbase_csf_reset_gpu_once(struct kbase_devic mutex_lock(&kbdev->mmu_hw_mutex); spin_lock_irqsave(&kbdev->hwaccess_lock, flags); kbase_ctx_sched_restore_all_as(kbdev); - kbase_ipa_control_handle_gpu_reset_post(kbdev); + { + kbase_ipa_control_handle_gpu_reset_post(kbdev); + } spin_unlock_irqrestore(&kbdev->hwaccess_lock, flags); mutex_unlock(&kbdev->mmu_hw_mutex); @@ -443,16 +439,14 @@ err: static void kbase_csf_reset_gpu_worker(struct work_struct *data) { - struct kbase_device *kbdev = container_of(data, struct kbase_device, - csf.reset.work); + struct kbase_device *kbdev = container_of(data, struct kbase_device, csf.reset.work); bool gpu_sleep_mode_active = false; bool firmware_inited; unsigned long flags; int err = 0; const enum kbase_csf_reset_gpu_state initial_reset_state = atomic_read(&kbdev->csf.reset.state); - const bool silent = - kbase_csf_reset_state_is_silent(initial_reset_state); + const bool silent = kbase_csf_reset_state_is_silent(initial_reset_state); /* Ensure any threads (e.g. executing the CSF scheduler) have finished * using the HW @@ -482,8 +476,8 @@ static void kbase_csf_reset_gpu_worker(struct work_struct *data) err = kbase_csf_reset_gpu_now(kbdev, firmware_inited, silent); #endif - } else if (!kbase_pm_context_active_handle_suspend(kbdev, - KBASE_PM_SUSPEND_HANDLER_DONT_REACTIVATE)) { + } else if (!kbase_pm_context_active_handle_suspend( + kbdev, KBASE_PM_SUSPEND_HANDLER_DONT_REACTIVATE)) { err = kbase_csf_reset_gpu_now(kbdev, firmware_inited, silent); kbase_pm_context_idle(kbdev); } @@ -497,22 +491,22 @@ static void kbase_csf_reset_gpu_worker(struct work_struct *data) bool kbase_prepare_to_reset_gpu(struct kbase_device *kbdev, unsigned int flags) { if (flags & RESET_FLAGS_HWC_UNRECOVERABLE_ERROR) - kbase_hwcnt_backend_csf_on_unrecoverable_error( - &kbdev->hwcnt_gpu_iface); + kbase_hwcnt_backend_csf_on_unrecoverable_error(&kbdev->hwcnt_gpu_iface); - if (atomic_cmpxchg(&kbdev->csf.reset.state, - KBASE_CSF_RESET_GPU_NOT_PENDING, - KBASE_CSF_RESET_GPU_PREPARED) != - KBASE_CSF_RESET_GPU_NOT_PENDING) + if (atomic_cmpxchg(&kbdev->csf.reset.state, KBASE_CSF_RESET_GPU_NOT_PENDING, + KBASE_CSF_RESET_GPU_PREPARED) != KBASE_CSF_RESET_GPU_NOT_PENDING) /* Some other thread is already resetting the GPU */ return false; + /* Issue the wake up of threads waiting for PM state transition. + * They might want to exit the wait since GPU reset has been triggered. + */ + wake_up(&kbdev->pm.backend.gpu_in_desired_state_wait); return true; } KBASE_EXPORT_TEST_API(kbase_prepare_to_reset_gpu); -bool kbase_prepare_to_reset_gpu_locked(struct kbase_device *kbdev, - unsigned int flags) +bool kbase_prepare_to_reset_gpu_locked(struct kbase_device *kbdev, unsigned int flags) { lockdep_assert_held(&kbdev->hwaccess_lock); @@ -524,8 +518,7 @@ void kbase_reset_gpu(struct kbase_device *kbdev) /* Note this is a WARN/atomic_set because it is a software issue for * a race to be occurring here */ - if (WARN_ON(atomic_read(&kbdev->csf.reset.state) != - KBASE_RESET_GPU_PREPARED)) + if (WARN_ON(atomic_read(&kbdev->csf.reset.state) != KBASE_RESET_GPU_PREPARED)) return; atomic_set(&kbdev->csf.reset.state, KBASE_CSF_RESET_GPU_COMMITTED); @@ -546,10 +539,9 @@ void kbase_reset_gpu_locked(struct kbase_device *kbdev) int kbase_reset_gpu_silent(struct kbase_device *kbdev) { - if (atomic_cmpxchg(&kbdev->csf.reset.state, - KBASE_CSF_RESET_GPU_NOT_PENDING, - KBASE_CSF_RESET_GPU_COMMITTED_SILENT) != - KBASE_CSF_RESET_GPU_NOT_PENDING) { + if (atomic_cmpxchg(&kbdev->csf.reset.state, KBASE_CSF_RESET_GPU_NOT_PENDING, + KBASE_CSF_RESET_GPU_COMMITTED_SILENT) != + KBASE_CSF_RESET_GPU_NOT_PENDING) { /* Some other thread is already resetting the GPU */ return -EAGAIN; } @@ -564,8 +556,7 @@ KBASE_EXPORT_TEST_API(kbase_reset_gpu_silent); bool kbase_reset_gpu_is_active(struct kbase_device *kbdev) { - enum kbase_csf_reset_gpu_state reset_state = - atomic_read(&kbdev->csf.reset.state); + enum kbase_csf_reset_gpu_state reset_state = atomic_read(&kbdev->csf.reset.state); /* For CSF, the reset is considered active only when the reset worker * is actually executing and other threads would have to wait for it to @@ -600,10 +591,8 @@ int kbase_reset_gpu_wait(struct kbase_device *kbdev) remaining = wait_event_timeout( kbdev->csf.reset.wait, - (atomic_read(&kbdev->csf.reset.state) == - KBASE_CSF_RESET_GPU_NOT_PENDING) || - (atomic_read(&kbdev->csf.reset.state) == - KBASE_CSF_RESET_GPU_FAILED), + (atomic_read(&kbdev->csf.reset.state) == KBASE_CSF_RESET_GPU_NOT_PENDING) || + (atomic_read(&kbdev->csf.reset.state) == KBASE_CSF_RESET_GPU_FAILED), wait_timeout); if (!remaining) { @@ -611,8 +600,7 @@ int kbase_reset_gpu_wait(struct kbase_device *kbdev) return -ETIMEDOUT; - } else if (atomic_read(&kbdev->csf.reset.state) == - KBASE_CSF_RESET_GPU_FAILED) { + } else if (atomic_read(&kbdev->csf.reset.state) == KBASE_CSF_RESET_GPU_FAILED) { return -ENOMEM; } |