From: Subhash Jadavani subhashj@codeaurora.org
commit 30fc33f1ef475480dc5bea4fe1bda84b003b992c upstream.
UFS devfreq clock scaling work may require clocks to be ON if it need to execute some UFS commands hence it may request for clock hold before issuing the command. But if UFS clock gating work is already running in parallel, ungate work would end up waiting for the clock gating work to finish and as clock gating work would also wait for the clock scaling work to finish, we would enter in deadlock state. Here is the call trace during this deadlock state:
Workqueue: devfreq_wq devfreq_monitor __switch_to __schedule schedule schedule_timeout wait_for_common wait_for_completion flush_work ufshcd_hold ufshcd_send_uic_cmd ufshcd_dme_get_attr ufs_qcom_set_dme_vs_core_clk_ctrl_clear_div ufs_qcom_clk_scale_notify ufshcd_scale_clks ufshcd_devfreq_target update_devfreq devfreq_monitor process_one_work worker_thread kthread ret_from_fork
Workqueue: events ufshcd_gate_work __switch_to __schedule schedule schedule_preempt_disabled __mutex_lock_slowpath mutex_lock devfreq_monitor_suspend devfreq_simple_ondemand_handler devfreq_suspend_device ufshcd_gate_work process_one_work worker_thread kthread ret_from_fork
Workqueue: events ufshcd_ungate_work __switch_to __schedule schedule schedule_timeout wait_for_common wait_for_completion flush_work __cancel_work_timer cancel_delayed_work_sync ufshcd_ungate_work process_one_work worker_thread kthread ret_from_fork
This change fixes this deadlock by doing this in devfreq work (devfreq_wq): Try cancelling clock gating work. If we are able to cancel gating work or it wasn't scheduled, hold the clock reference count until scaling is in progress. If gate work is already running in parallel, let's skip the frequecy scaling at this time and it will be retried once next scaling window expires.
Reviewed-by: Sahitya Tummala stummala@codeaurora.org Signed-off-by: Subhash Jadavani subhashj@codeaurora.org Signed-off-by: Martin K. Petersen martin.petersen@oracle.com Signed-off-by: Amit Pundir amit.pundir@linaro.org --- drivers/scsi/ufs/ufshcd.c | 32 ++++++++++++++++++++++++++++++++ 1 file changed, 32 insertions(+)
diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c index 7b48fb84a900..e8f3033d991b 100644 --- a/drivers/scsi/ufs/ufshcd.c +++ b/drivers/scsi/ufs/ufshcd.c @@ -5406,15 +5406,47 @@ static int ufshcd_devfreq_target(struct device *dev, { int err = 0; struct ufs_hba *hba = dev_get_drvdata(dev); + bool release_clk_hold = false; + unsigned long irq_flags;
if (!ufshcd_is_clkscaling_enabled(hba)) return -EINVAL;
+ spin_lock_irqsave(hba->host->host_lock, irq_flags); + if (ufshcd_eh_in_progress(hba)) { + spin_unlock_irqrestore(hba->host->host_lock, irq_flags); + return 0; + } + + if (ufshcd_is_clkgating_allowed(hba) && + (hba->clk_gating.state != CLKS_ON)) { + if (cancel_delayed_work(&hba->clk_gating.gate_work)) { + /* hold the vote until the scaling work is completed */ + hba->clk_gating.active_reqs++; + release_clk_hold = true; + hba->clk_gating.state = CLKS_ON; + } else { + /* + * Clock gating work seems to be running in parallel + * hence skip scaling work to avoid deadlock between + * current scaling work and gating work. + */ + spin_unlock_irqrestore(hba->host->host_lock, irq_flags); + return 0; + } + } + spin_unlock_irqrestore(hba->host->host_lock, irq_flags); + if (*freq == UINT_MAX) err = ufshcd_scale_clks(hba, true); else if (*freq == 0) err = ufshcd_scale_clks(hba, false);
+ spin_lock_irqsave(hba->host->host_lock, irq_flags); + if (release_clk_hold) + __ufshcd_release(hba); + spin_unlock_irqrestore(hba->host->host_lock, irq_flags); + return err; }