On 8/31/21 12:24 AM, Adrian Hunter wrote:

How does splitting the host_sem address the potential deadlock
between the error handler and runtime resume?

Hmm ... how do runtime resume and the error handler deadlock? If 
shost->eh_noresume == 0 then scsi_error_handler() will call 
scsi_autopm_get_host() before invoking the eh_strategy_handler callback. 
The definition of scsi_autopm_get_host() is as follows:

int scsi_autopm_get_host(struct Scsi_Host *shost)
{
	int	err;

	err = pm_runtime_get_sync(&shost->shost_gendev);
	if (err < 0 && err !=-EACCES)
		pm_runtime_put_sync(&shost->shost_gendev);
	else
		err = 0;
	return err;
}

The power management operations used for shost_gendev instances are 
defined by scsi_bus_pm_ops (see also scsi_host_alloc()). The following 
function is the runtime resume function referenced by scsi_bus_pm_ops 
and skips shost_gendevs since these are not SCSI devices:

static int scsi_runtime_resume(struct device *dev)
{
	int err = 0;

	dev_dbg(dev, "scsi_runtime_resume\n");
	if (scsi_is_sdev_device(dev))
		err = sdev_runtime_resume(dev);

	/* Insert hooks here for targets, hosts, and transport classes*/

	return err;
}

In addition to the above function the runtime resume callback of the UFS 
platform device is also invoked. I think all these functions call 
ufshcd_runtime_resume(). As far as I can see ufshcd_runtime_resume() 
does not touch host_sem?

Bart.