Re: [PATCH 01/12] usbnet: introduce usbnet 3 command helpers

On Thursday 11 October 2012 10:36:22 Alan Stern wrote:

It's worse than you may realize.  When a SCSI disk is suspended, all of
its ancestor devices may be suspended too.  Pages can't be read in from
the drive until all those ancestors are resumed.  This means that all
runtime resume code paths for all drivers that could be bound to an
ancestor of a block device must avoid GFP_KERNEL.  In practice it's
probably easiest for the runtime PM core to use tsk_set_allowd_gfp()
before calling any runtime_resume method.

Or at least, this will be true when sd supports nontrivial autosuspend.

Up to now, I've found three driver for which tsk_set_allowd_gfp() wouldn't
do the job. They boil down into two types of errors. That is surprisingly good.

First we have workqueues. bas-gigaset is a good example.
The driver kills a scheduled work in pre_reset(). If this is done synchronously
the driver may need to wait for a memory allocation inside the work.
In principle we could provide a workqueue limited to GFP_NOIO. Is that worth
it, or do we just check?

Second there is a problem just like priority inversion with realtime tasks.
usb-skeleton and ati_remote2
They take mutexes which are also taken in other code paths. So the error
handler may need to wait for a mutex to be dropped which can only happen
if a memory allocation succeeds, which is waiting for the error handler.

usb-skeleton is even worse, as it does copy_to_user(). I guess copy_to/from_user
must simply not be done under such a mutex.

I am afraid there is no generic solution in the last two cases. What do you think?

	Regards
		Oliver

--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Fri, Oct 12, 2012 at 9:51 PM, Oliver Neukum [off-list ref] wrote:

On Thursday 11 October 2012 10:36:22 Alan Stern wrote:

quoted

It's worse than you may realize.  When a SCSI disk is suspended, all of
its ancestor devices may be suspended too.  Pages can't be read in from
the drive until all those ancestors are resumed.  This means that all
runtime resume code paths for all drivers that could be bound to an
ancestor of a block device must avoid GFP_KERNEL.  In practice it's
probably easiest for the runtime PM core to use tsk_set_allowd_gfp()
before calling any runtime_resume method.

Or at least, this will be true when sd supports nontrivial autosuspend.

Up to now, I've found three driver for which tsk_set_allowd_gfp() wouldn't
do the job. They boil down into two types of errors. That is surprisingly good.

Looks all are very good examples, :-)

First we have workqueues. bas-gigaset is a good example.
The driver kills a scheduled work in pre_reset(). If this is done synchronously
the driver may need to wait for a memory allocation inside the work.
In principle we could provide a workqueue limited to GFP_NOIO. Is that worth
it, or do we just check?

The easiest way is to always call tsk_set_allowd_gfp(~GFP_IOFS) in the
start of work function under the situation, and restore the flag in the end
of the work function.

Second there is a problem just like priority inversion with realtime tasks.
usb-skeleton and ati_remote2
They take mutexes which are also taken in other code paths. So the error
handler may need to wait for a mutex to be dropped which can only happen
if a memory allocation succeeds, which is waiting for the error handler.

Suppose mutex_lock(A) is called in pre_reset(), one solution is that
always calling tsk_set_allowd_gfp(~GFP_IOFS) before each mutex_lock(A).
We can do it only for devices with storage interface in current
configuration.


Thanks，
--
Ming Lei