- Synchronize job submission with reset/recovery using reset_lock - Always print recovery reason and call diagnose_failure() - Prevent immediate autosuspend after reset/recovery
Signed-off-by: Jacek Lawrynowicz <[email protected]> --- drivers/accel/ivpu/ivpu_debugfs.c | 20 ++++++++++--- drivers/accel/ivpu/ivpu_hw_37xx.c | 14 +++------ drivers/accel/ivpu/ivpu_hw_40xx.c | 8 +++--- drivers/accel/ivpu/ivpu_ipc.c | 8 ++---- drivers/accel/ivpu/ivpu_job.c | 2 ++ drivers/accel/ivpu/ivpu_mmu.c | 4 +-- drivers/accel/ivpu/ivpu_pm.c | 48 ++++++++++++++++++++----------- drivers/accel/ivpu/ivpu_pm.h | 6 ++-- 8 files changed, 67 insertions(+), 43 deletions(-) diff --git a/drivers/accel/ivpu/ivpu_debugfs.c b/drivers/accel/ivpu/ivpu_debugfs.c index 19035230563d..7cb962e21453 100644 --- a/drivers/accel/ivpu/ivpu_debugfs.c +++ b/drivers/accel/ivpu/ivpu_debugfs.c @@ -102,7 +102,7 @@ static int reset_pending_show(struct seq_file *s, void *v) { struct ivpu_device *vdev = seq_to_ivpu(s); - seq_printf(s, "%d\n", atomic_read(&vdev->pm->in_reset)); + seq_printf(s, "%d\n", atomic_read(&vdev->pm->reset_pending)); return 0; } @@ -130,7 +130,9 @@ dvfs_mode_fops_write(struct file *file, const char __user *user_buf, size_t size fw->dvfs_mode = dvfs_mode; - ivpu_pm_schedule_recovery(vdev); + ret = pci_try_reset_function(to_pci_dev(vdev->drm.dev)); + if (ret) + return ret; return size; } @@ -190,7 +192,10 @@ fw_profiling_freq_fops_write(struct file *file, const char __user *user_buf, return ret; ivpu_hw_profiling_freq_drive(vdev, enable); - ivpu_pm_schedule_recovery(vdev); + + ret = pci_try_reset_function(to_pci_dev(vdev->drm.dev)); + if (ret) + return ret; return size; } @@ -301,11 +306,18 @@ static ssize_t ivpu_force_recovery_fn(struct file *file, const char __user *user_buf, size_t size, loff_t *pos) { struct ivpu_device *vdev = file->private_data; + int ret; if (!size) return -EINVAL; - ivpu_pm_schedule_recovery(vdev); + ret = ivpu_rpm_get(vdev); + if (ret) + return ret; + + ivpu_pm_trigger_recovery(vdev, "debugfs"); + flush_work(&vdev->pm->recovery_work); + ivpu_rpm_put(vdev); return size; } diff --git a/drivers/accel/ivpu/ivpu_hw_37xx.c b/drivers/accel/ivpu/ivpu_hw_37xx.c index e33dfe3089af..2a67a457ec7b 100644 --- a/drivers/accel/ivpu/ivpu_hw_37xx.c +++ b/drivers/accel/ivpu/ivpu_hw_37xx.c @@ -869,24 +869,18 @@ static void ivpu_hw_37xx_irq_disable(struct ivpu_device *vdev) static void ivpu_hw_37xx_irq_wdt_nce_handler(struct ivpu_device *vdev) { - ivpu_err_ratelimited(vdev, "WDT NCE irq\n"); - - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "WDT NCE IRQ"); } static void ivpu_hw_37xx_irq_wdt_mss_handler(struct ivpu_device *vdev) { - ivpu_err_ratelimited(vdev, "WDT MSS irq\n"); - ivpu_hw_wdt_disable(vdev); - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "WDT MSS IRQ"); } static void ivpu_hw_37xx_irq_noc_firewall_handler(struct ivpu_device *vdev) { - ivpu_err_ratelimited(vdev, "NOC Firewall irq\n"); - - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "NOC Firewall IRQ"); } /* Handler for IRQs from VPU core (irqV) */ @@ -964,7 +958,7 @@ static bool ivpu_hw_37xx_irqb_handler(struct ivpu_device *vdev, int irq) REGB_WR32(VPU_37XX_BUTTRESS_INTERRUPT_STAT, status); if (schedule_recovery) - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "Buttress IRQ"); return true; } diff --git a/drivers/accel/ivpu/ivpu_hw_40xx.c b/drivers/accel/ivpu/ivpu_hw_40xx.c index bbb93cac0190..22c9a8f70d83 100644 --- a/drivers/accel/ivpu/ivpu_hw_40xx.c +++ b/drivers/accel/ivpu/ivpu_hw_40xx.c @@ -1027,18 +1027,18 @@ static void ivpu_hw_40xx_irq_disable(struct ivpu_device *vdev) static void ivpu_hw_40xx_irq_wdt_nce_handler(struct ivpu_device *vdev) { /* TODO: For LNN hang consider engine reset instead of full recovery */ - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "WDT NCE IRQ"); } static void ivpu_hw_40xx_irq_wdt_mss_handler(struct ivpu_device *vdev) { ivpu_hw_wdt_disable(vdev); - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "WDT MSS IRQ"); } static void ivpu_hw_40xx_irq_noc_firewall_handler(struct ivpu_device *vdev) { - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "NOC Firewall IRQ"); } /* Handler for IRQs from VPU core (irqV) */ @@ -1132,7 +1132,7 @@ static bool ivpu_hw_40xx_irqb_handler(struct ivpu_device *vdev, int irq) REGB_WR32(VPU_40XX_BUTTRESS_INTERRUPT_STAT, status); if (schedule_recovery) - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "Buttress IRQ"); return true; } diff --git a/drivers/accel/ivpu/ivpu_ipc.c b/drivers/accel/ivpu/ivpu_ipc.c index f69780248803..1e86ec2d62d2 100644 --- a/drivers/accel/ivpu/ivpu_ipc.c +++ b/drivers/accel/ivpu/ivpu_ipc.c @@ -343,10 +343,8 @@ int ivpu_ipc_send_receive_active(struct ivpu_device *vdev, struct vpu_jsm_msg *r hb_ret = ivpu_ipc_send_receive_internal(vdev, &hb_req, VPU_JSM_MSG_QUERY_ENGINE_HB_DONE, &hb_resp, VPU_IPC_CHAN_ASYNC_CMD, vdev->timeout.jsm); - if (hb_ret == -ETIMEDOUT) { - ivpu_hw_diagnose_failure(vdev); - ivpu_pm_schedule_recovery(vdev); - } + if (hb_ret == -ETIMEDOUT) + ivpu_pm_trigger_recovery(vdev, "IPC timeout"); return ret; } @@ -397,7 +395,7 @@ void ivpu_ipc_irq_handler(struct ivpu_device *vdev, bool *wake_thread) */ while (ivpu_hw_reg_ipc_rx_count_get(vdev)) { if (++msg_count > IPC_MAX_RX_MSG) { - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "IPC msg limit"); return; } diff --git a/drivers/accel/ivpu/ivpu_job.c b/drivers/accel/ivpu/ivpu_job.c index d9b47a04b35f..0440bee3ecaf 100644 --- a/drivers/accel/ivpu/ivpu_job.c +++ b/drivers/accel/ivpu/ivpu_job.c @@ -515,7 +515,9 @@ int ivpu_submit_ioctl(struct drm_device *dev, void *data, struct drm_file *file) goto err_destroy_job; } + down_read(&vdev->pm->reset_lock); ret = ivpu_job_submit(job); + up_read(&vdev->pm->reset_lock); if (ret) goto err_signal_fence; diff --git a/drivers/accel/ivpu/ivpu_mmu.c b/drivers/accel/ivpu/ivpu_mmu.c index c82929b0ae9d..9910755c9951 100644 --- a/drivers/accel/ivpu/ivpu_mmu.c +++ b/drivers/accel/ivpu/ivpu_mmu.c @@ -898,13 +898,13 @@ void ivpu_mmu_irq_evtq_handler(struct ivpu_device *vdev) while ((event = ivpu_mmu_get_event(vdev)) != NULL) { ivpu_mmu_dump_event(vdev, event); if (++event_count > IVPU_MMU_MAX_EVENT_COUNT) { - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "MMU event limit"); return; } ssid = FIELD_GET(IVPU_MMU_EVT_SSID_MASK, event[0]); if (ssid == IVPU_GLOBAL_CONTEXT_MMU_SSID) { - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "MMU event"); return; } diff --git a/drivers/accel/ivpu/ivpu_pm.c b/drivers/accel/ivpu/ivpu_pm.c index 8407f1d8c99c..f501f27ebafd 100644 --- a/drivers/accel/ivpu/ivpu_pm.c +++ b/drivers/accel/ivpu/ivpu_pm.c @@ -112,6 +112,14 @@ static void ivpu_pm_recovery_work(struct work_struct *work) char *evt[2] = {"IVPU_PM_EVENT=IVPU_RECOVER", NULL}; int ret; + ivpu_err(vdev, "Recovering the VPU (reset #%d)\n", atomic_read(&vdev->pm->reset_counter)); + + ret = pm_runtime_resume_and_get(vdev->drm.dev); + if (ret) + ivpu_err(vdev, "Failed to resume VPU: %d\n", ret); + + ivpu_fw_log_dump(vdev); + retry: ret = pci_try_reset_function(to_pci_dev(vdev->drm.dev)); if (ret == -EAGAIN && !drm_dev_is_unplugged(&vdev->drm)) { @@ -123,11 +131,13 @@ static void ivpu_pm_recovery_work(struct work_struct *work) ivpu_err(vdev, "Failed to reset VPU: %d\n", ret); kobject_uevent_env(&vdev->drm.dev->kobj, KOBJ_CHANGE, evt); + pm_runtime_mark_last_busy(vdev->drm.dev); + pm_runtime_put_autosuspend(vdev->drm.dev); } -void ivpu_pm_schedule_recovery(struct ivpu_device *vdev) +void ivpu_pm_trigger_recovery(struct ivpu_device *vdev, const char *reason) { - struct ivpu_pm_info *pm = vdev->pm; + ivpu_err(vdev, "Recovery triggered by %s\n", reason); if (ivpu_disable_recovery) { ivpu_err(vdev, "Recovery not available when disable_recovery param is set\n"); @@ -139,10 +149,11 @@ void ivpu_pm_schedule_recovery(struct ivpu_device *vdev) return; } - /* Schedule recovery if it's not in progress */ - if (atomic_cmpxchg(&pm->in_reset, 0, 1) == 0) { - ivpu_hw_irq_disable(vdev); - queue_work(system_long_wq, &pm->recovery_work); + /* Trigger recovery if it's not in progress */ + if (atomic_cmpxchg(&vdev->pm->reset_pending, 0, 1) == 0) { + ivpu_hw_diagnose_failure(vdev); + ivpu_hw_irq_disable(vdev); /* Disable IRQ early to protect from IRQ storm */ + queue_work(system_long_wq, &vdev->pm->recovery_work); } } @@ -150,12 +161,8 @@ static void ivpu_job_timeout_work(struct work_struct *work) { struct ivpu_pm_info *pm = container_of(work, struct ivpu_pm_info, job_timeout_work.work); struct ivpu_device *vdev = pm->vdev; - unsigned long timeout_ms = ivpu_tdr_timeout_ms ? ivpu_tdr_timeout_ms : vdev->timeout.tdr; - ivpu_err(vdev, "TDR detected, timeout %lu ms", timeout_ms); - ivpu_hw_diagnose_failure(vdev); - - ivpu_pm_schedule_recovery(vdev); + ivpu_pm_trigger_recovery(vdev, "TDR"); } void ivpu_start_job_timeout_detection(struct ivpu_device *vdev) @@ -228,6 +235,9 @@ int ivpu_pm_runtime_suspend_cb(struct device *dev) bool hw_is_idle = true; int ret; + drm_WARN_ON(&vdev->drm, !xa_empty(&vdev->submitted_jobs_xa)); + drm_WARN_ON(&vdev->drm, work_pending(&vdev->pm->recovery_work)); + ivpu_dbg(vdev, PM, "Runtime suspend..\n"); if (!ivpu_hw_is_idle(vdev) && vdev->pm->suspend_reschedule_counter) { @@ -310,11 +320,12 @@ void ivpu_pm_reset_prepare_cb(struct pci_dev *pdev) { struct ivpu_device *vdev = pci_get_drvdata(pdev); - pm_runtime_get_sync(vdev->drm.dev); - ivpu_dbg(vdev, PM, "Pre-reset..\n"); atomic_inc(&vdev->pm->reset_counter); - atomic_set(&vdev->pm->in_reset, 1); + atomic_set(&vdev->pm->reset_pending, 1); + + pm_runtime_get_sync(vdev->drm.dev); + down_write(&vdev->pm->reset_lock); ivpu_prepare_for_reset(vdev); ivpu_hw_reset(vdev); ivpu_pm_prepare_cold_boot(vdev); @@ -331,9 +342,11 @@ void ivpu_pm_reset_done_cb(struct pci_dev *pdev) ret = ivpu_resume(vdev); if (ret) ivpu_err(vdev, "Failed to set RESUME state: %d\n", ret); - atomic_set(&vdev->pm->in_reset, 0); + up_write(&vdev->pm->reset_lock); + atomic_set(&vdev->pm->reset_pending, 0); ivpu_dbg(vdev, PM, "Post-reset done.\n"); + pm_runtime_mark_last_busy(vdev->drm.dev); pm_runtime_put_autosuspend(vdev->drm.dev); } @@ -346,7 +359,10 @@ void ivpu_pm_init(struct ivpu_device *vdev) pm->vdev = vdev; pm->suspend_reschedule_counter = PM_RESCHEDULE_LIMIT; - atomic_set(&pm->in_reset, 0); + init_rwsem(&pm->reset_lock); + atomic_set(&pm->reset_pending, 0); + atomic_set(&pm->reset_counter, 0); + INIT_WORK(&pm->recovery_work, ivpu_pm_recovery_work); INIT_DELAYED_WORK(&pm->job_timeout_work, ivpu_job_timeout_work); diff --git a/drivers/accel/ivpu/ivpu_pm.h b/drivers/accel/ivpu/ivpu_pm.h index 97c6e0b0aa42..ec60fbeefefc 100644 --- a/drivers/accel/ivpu/ivpu_pm.h +++ b/drivers/accel/ivpu/ivpu_pm.h @@ -6,6 +6,7 @@ #ifndef __IVPU_PM_H__ #define __IVPU_PM_H__ +#include <linux/rwsem.h> #include <linux/types.h> struct ivpu_device; @@ -14,8 +15,9 @@ struct ivpu_pm_info { struct ivpu_device *vdev; struct delayed_work job_timeout_work; struct work_struct recovery_work; - atomic_t in_reset; + struct rw_semaphore reset_lock; atomic_t reset_counter; + atomic_t reset_pending; bool is_warmboot; u32 suspend_reschedule_counter; }; @@ -37,7 +39,7 @@ int __must_check ivpu_rpm_get(struct ivpu_device *vdev); int __must_check ivpu_rpm_get_if_active(struct ivpu_device *vdev); void ivpu_rpm_put(struct ivpu_device *vdev); -void ivpu_pm_schedule_recovery(struct ivpu_device *vdev); +void ivpu_pm_trigger_recovery(struct ivpu_device *vdev, const char *reason); void ivpu_start_job_timeout_detection(struct ivpu_device *vdev); void ivpu_stop_job_timeout_detection(struct ivpu_device *vdev); -- 2.43.0
