On Thu, Feb 04 2021 at 11:48 -0800, Alan Stern wrote:
On Wed, Feb 03, 2021 at 04:13:54PM -0800, Asutosh Das wrote:
Thanks Alan.
I understand the issues with the current ufs design.
ufs has a wlun (well-known lun) that handles power management commands,
such as SSUs. Now this wlun (device wlun) is registered as a scsi_device.
It's queue is also set up for runtime-pm. Likewise there're 2
more wluns, BOOT and RPMB.
Currently, the scsi devices independently runtime suspend/resume - request driven.
So to send SSU while suspending wlun (scsi_device) this scsi device should
be the last to be runtime suspended amongst all other ufs luns (scsi devices). The
reason is syncronize_cache() is sent to luns during their suspend and if SSU has
been sent already, it mostly would fail.
The SCSI subsystem assumes that different LUNs operate independently.
Evidently that isn't true here.
Perhaps that's the reason to send SSU during platform-device suspend. I'm not
sure if that's the right thing to do, but that's what it is now and is causing
this deadlock.
Now this wlun is also registered to bsg and some applications interact with rpmb
wlun and the device-wlun using that interface. Registering the corresponding
queues to runtime-pm ensures that the whole path is resumed before the request
is issued.
Because, we see this deadlock, in the RFC patch, I skipped registering the
queues representing the wluns to runtime-pm, thus removing the restrictions to
issue the request until queue is resumed.
But when the requests come-in via bsg, the device has to be resumed. Hence the
get_sync()/put_sync() in bsg driver.
Does the bsg interface send its I/O requests to the LUNs through the
block request queue?
The reason for initiating get_sync()/put_sync() on the parent device was because
the corresponding queue of this wlun was not setup for runtime-pm anymore.
And for ufs resuming the scsi device essentially means sending a SSU to wlun
which the ufs platform device does in its runtime resume now. I'm not sure if
that was a good idea though, hence the RFC on the patches.
And now it looks to me that adding a cb to sd_suspend_runtime may not work.
Because the scsi devices suspend asynchronously and the wlun suspends earlier than the others.
[ 7.846165]scsi 0:0:0:49488: scsi_runtime_idle
[ 7.851547]scsi 0:0:0:49488: device wlun
[ 7.851809]sd 0:0:0:49488: scsi_runtime_idle
[ 7.861536]sd 0:0:0:49488: scsi_runtime_suspend < suspends prior to other luns
[...]
[ 12.861984]sd 0:0:0:1: [sdb] Synchronizing SCSI cache
[ 12.868894]sd 0:0:0:2: [sdc] Synchronizing SCSI cache
[ 13.124331]sd 0:0:0:0: [sda] Synchronizing SCSI cache
[ 13.143961]sd 0:0:0:3: [sdd] Synchronizing SCSI cache
[ 13.163876]sd 0:0:0:6: [sdg] Synchronizing SCSI cache
[ 13.164024]sd 0:0:0:4: [sde] Synchronizing SCSI cache
[ 13.167066]sd 0:0:0:5: [sdf] Synchronizing SCSI cache
[ 17.101285]sd 0:0:0:7: [sdh] Synchronizing SCSI cache
[ 73.889551]sd 0:0:0:4: [sde] Synchronizing SCSI cache
I'm not sure if there's a way to force the wlun to suspend only after all other luns are suspended.
Is there? I hope Bart/others help provide some inputs on this.
I don't know what would work best for you; it depends on how the LUNs
are used. But one possibility is to make sure that whenever the boot
and rpmb wluns are resumed, the device wlun is also resumed. So for
example, the runtime-resume callback routines for the rpmb and boot
wluns could call pm_runtime_get_sync() for the device wlun, and their
runtime-suspend callback routines could call pm_runtime_put() for the
device wlun. And of course there would have to be appropriate
operations when those LUNs are bound to and unbound from their drivers.
Alan Stern
Thanks Alan.
CanG & I had some discussions on it as well the other day.
I'm now looking into creating a device link between the siblings.
e.g. make the device wlun as a supplier for all the other luns & wluns.
So device wlun (supplier) wouldn't suspend (runtime/system) until all of the other
consumers are suspended. After this linking, I can move all the
pm commands that are being sent by host to the dedicated suspend routine of the device
wlun and the host needn't send any cmds during its suspend and layering
violation wouldn't take place.
-asd
