On Tue, Apr 19, 2022 at 03:18:36PM +0100, Marc Zyngier wrote: > This is a followup from [2]. > > I recently realised that the gpiolib play ugly tricks on the > unsuspecting irq_chip structures by patching the callbacks. > > Not only this breaks when an irq_chip structure is made const (which > really should be the default case), but it also forces this structure > to be copied at nauseam for each instance of the GPIO block, which is > a waste of memory. Is this brings us to the issue with IRQ chip name? The use case in my mind is the following: 1) we have two or more GPIO chips that supports IRQ; 2) the user registers two IRQs of the same (by number) pin on different chips; 3) cat /proc/interrupt will show 'my_gpio_chip XX', where XX is the number. So, do I understand correct current state of affairs? If so, we have to fix this to have any kind of ID added to the chip name that we can map /proc/interrupts output correctly. > My current approach is to add a new irq_chip flag (IRQCHIP_IMMUTABLE) > which does what it says on the tin: don't you dare writing to them. > Gpiolib is further updated not to install its own callbacks, and it > becomes the responsibility of the driver to call into the gpiolib when > required. This is similar to what we do for other subsystems such as > PCI-MSI. > > 5 drivers are updated to this new model: M1, QC, Tegra, pl061 and AMD > (as I actively use them) keeping a single irq_chip structure, marking > it const, and exposing the new flag. > > Nothing breaks, the volume of change is small, the memory usage goes > down and we have fewer callbacks that can be used as attack vectors. > What's not to love? > > Since there wasn't any objection in the previous round of review, I'm > going to take this series into -next to see if anything breaks at > scale. -- With Best Regards, Andy Shevchenko
