On Fri, Sep 12, 2025 at 04:40:27PM +0200, Greg KH wrote: > On Fri, Sep 12, 2025 at 05:26:46PM +0300, Laurent Pinchart wrote: > > (CC'ing Dan Williams) > > > > On Fri, Sep 12, 2025 at 04:19:53PM +0200, Greg KH wrote: > > > On Fri, Sep 12, 2025 at 04:59:16PM +0300, Laurent Pinchart wrote: > > > > On Fri, Sep 12, 2025 at 03:46:27PM +0200, Bartosz Golaszewski wrote: > > > > > On Fri, Sep 12, 2025 at 3:39 PM Greg Kroah-Hartman wrote: > > > > > > > > > > > > I have no objection moving this to the cdev api, BUT given that > > > > > > 'struct > > > > > > cdev' is embedded everywhere, I don't think it's going to be a > > > > > > simple > > > > > > task, but rather have to be done one-driver-at-a-time like the > > > > > > patch in > > > > > > this series does it. > > > > > > > > > > I don't think cdev is the right place for this as user-space keeping a > > > > > reference to a file-descriptor whose "backend" disappeared is not the > > > > > only possible problem. We can easily create a use-case of a USB I2C > > > > > expander being used by some in-kernel consumer and then unplugged. > > > > > This has nothing to do with the character device. I believe the > > > > > sub-system level is the right place for this and every driver > > > > > subsystem would have to integrate it separately, taking its various > > > > > quirks into account. > > > > > > > > That's why I mentioned in-kernel users previously. Drivers routinely > > > > acquire resources provided by other drivers, and having a way to revoke > > > > those is needed. > > > > > > > > It is a different but related problem compared to userspace racing with > > > > .remove(). Could we solve both using the same backend concepts ? > > > > Perhaps, time will tell, and if that works nicely, great. But we still > > > > have lots of drivers exposing character devices to userspace (usually > > > > through a subsystem-specific API, drivers that create a cdev manually > > > > are the minority). That problem is in my opinion more urgent than > > > > handling the removal of in-kernel resources, because it's more common, > > > > and is easily triggerable by userspace. The good news is that it should > > > > also be simpler to solve, we should be able to address the enter/exit > > > > part entirely in cdev, and limit the changes to drivers in .remove() to > > > > the strict minimum. > > > > > > > > What I'd like to see is if the proposed implementation of revocable > > > > resources can be used as a building block to fix the cdev issue. If it > > > > ca, great, let's solve it then. If it can't, that's still fine, it will > > > > still be useful for in-kernel resources, even if we need a different > > > > implementation for cdev. > > > > > > Patch 5/5 in this series does just this for a specific use of a cdev in > > > the driver. Is that what you are looking for? > > > > Not quite, I would like to see the enter/exit (aka revocable scope if my > > understanding is correct) being pushed to char_dev.c, as Dan did in [1]. > > I'm fine having to add an extra function call in the .remove() path of > > drivers, but I'm not fine with having to mark revocable sections > > manually in drivers. That part belongs to cdev. > > > > [1] > > https://lore.kernel.org/r/161117153248.2853729.2452425259045172318.st...@dwillia2-desk3.amr.corp.intel.com > > Dan's proposal here is a good start, but the "sleep in cdev_del() until > the device drains all existing opens" is going to not really work well > for what we want.
I think you missed the fact that the code doesn't wait for all open file handles to be closed. It waits for all in-progress syscalls to return from the driver. That's a big difference. > So sure, make a new cdev api to use this, that's fine, then we will have > what, 5 different ways to use a cdev? :) > > Seriously, that would be good, then we can work to convert things over, > but I think overall it will look much the same as what patch 5/5 does > here. But details matter, I don't really known for sure... What I don't want to see is drivers using this new API directly to protect the cdev race. That would be a big step in the wrong direction. Patch 5/5 needs to move driver code to the cdev level. That shouldn't be difficult, so I really not see why it can't be done in v4 to see how it will look like. > Either way, I think this patch series stands on its own, it doesn't > require cdev to implement it, drivers can use it to wrap a cdev if they > want to. No, drivers should *not* do that manually. That's a recipe for disaster that we would regret later. > We have other structures that want to do this type of thing > today as is proof with the rust implementation for the devm api. -- Regards, Laurent Pinchart
