Messages in this thread | | | Date | Thu, 28 Jul 2022 11:49:57 +0100 | Subject | Re: [PATCH] component: try_module_get() to prevent unloading while in use | From | Richard Fitzgerald <> |
| |
On 26/07/2022 18:28, Greg KH wrote: > On Tue, Jul 26, 2022 at 11:32:28AM +0100, Richard Fitzgerald wrote: >> On 25/07/2022 19:09, Greg KH wrote: >>> On Mon, Jul 25, 2022 at 05:08:59PM +0100, Richard Fitzgerald wrote: >>>> Call try_module_get() on a component before attempting to call its >>>> bind() function, to ensure that a loadable module cannot be >>>> unloaded while we are executing its bind(). >>> >>> How can bind be called while the module is unloaded? >>> >> >> I didn't say it could. What I said is "unloaded while we are executing >> its bind()". Maybe that's already guaranteed to be safe somehow. It's >> actually the problem below that I was trying to fix but placing the >> try_module_get() before the bind() rather than after bind() seemed a >> trivial extra safety. > > It should be safe, bind() can't race with module remove as the driver > core locks will handle this. > >>>> If the bind is successful the module_put() is called only after it >>>> has been unbound. This ensures that the module cannot be unloaded >>>> while it is in use as an aggregate device. >>> >>> That's almost never the correct thing to do, what problem is this >>> solving? >>> >> >> What I see is that when a loadable module has been made part of an >> aggregate it is still possible to rmmod'd it. >> >> An alternative workaround would be for the parent to softdep to every >> driver that _might_ provide the aggregated components. Softdeps aren't >> unusual (we use it in some drivers that are directly related but don't >> directly link into each other). But to me this feels like a hack when >> used with the component framework - isn't the idea that the parent >> doesn't know (or doesn't need to know) which drivers will be aggregated? >> Wouldn't it be better that when a component driver is bound into an >> aggregate its module is automatically marked in-use? >> >> If there's a better way to mark the module in-use while is it bound >> into an aggregate, let me know and I'll look at implementing it. > > No module references should be incremented if a device is bound to a > driver, that's the old (1990's) way of thinking. If a module wants to > be unloaded, let it, and clean up everything that it was > controlling/talking to before the module remove is finished. > > That's the way all busses should be working, you don't increment a > module count when a driver binds to a device, otherwise how would you > unload a module that was being used at all? > > So just remove the components controlled by the module properly when it > is removed and all should be good. > > Do you have example code in the kernel tree today that does not properly > do this? Why not just fix that instead? >
The actual code I'm working on isn't upstream yet, but it's a derivative of the way these two interoperate: sound/pci/hda/patch_realtek sound/pci/hda/cs35l41_hda.c
In these systems the host audio interface is HDA but the amps are not HDA devices. Audio goes through the Realtek HDA codec to the amp but the amp is on a different bus (i2c or spi). The modules in the HDA stack all get marked as in-use, but the amp driver doesn't. So if it's unloaded the audio system is left in a limbo state where ALSA and the HDA stack are still up but the amp driver code has gone.
However I realised that my try_module_get() isn't a fix anyway. It's claiming use of the module implementing the component but not of the bus that owns that module. I assume that's what you were referring to by having to deal with an unload instead of trying to prevent the unload.
(And yes, I'm aware that in that patch_realtek.c it's missing locking around the shared struct to prevent it being accessed during a bind and unbind.)
> thanks, > > greg k-h
| |