From mboxrd@z Thu Jan 1 00:00:00 1970 From: Thierry Reding Subject: Re: [PATCH 6/9] gpio: Add support for hierarchical IRQ domains Date: Tue, 25 Sep 2018 13:17:16 +0200 Message-ID: <20180925111716.GA7925@ulmo> References: <20180921102546.12745-1-thierry.reding@gmail.com> <20180921102546.12745-7-thierry.reding@gmail.com> <20180925093315.GB7097@ulmo> Mime-Version: 1.0 Content-Type: multipart/signed; micalg=pgp-sha256; protocol="application/pgp-signature"; boundary="yrj/dFKFPuw6o+aM" Return-path: Content-Disposition: inline In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org To: Linus Walleij Cc: Marc Zyngier , Thomas Gleixner , "open list:OPEN FIRMWARE AND FLATTENED DEVICE TREE BINDINGS" , linux-tegra@vger.kernel.org, "open list:GPIO SUBSYSTEM" , "linux-kernel@vger.kernel.org" List-Id: linux-gpio@vger.kernel.org --yrj/dFKFPuw6o+aM Content-Type: text/plain; charset=us-ascii Content-Disposition: inline Content-Transfer-Encoding: quoted-printable On Tue, Sep 25, 2018 at 12:33:54PM +0200, Linus Walleij wrote: > On Tue, Sep 25, 2018 at 11:33 AM Thierry Reding > wrote: > > On Tue, Sep 25, 2018 at 10:11:06AM +0200, Linus Walleij wrote: >=20 > > > While I think it is really important that we start supporting hierarc= hical > > > irqdomains in the gpiolib core, I want a more complete approach, > > > so that drivers that need hierarchical handling of irqdomains > > > can get the same support from gpiolib as they get for simple > > > domains. > (...) > > > I can't see if you need to pull more stuff into the core to accomplish > > > that, but I think in essence the core gpiolib needs to be more helpful > > > with hierarchies. > > > > This is not as trivial as it sounds. I think we could probably provide a > > simple helper in the core that may work for the majority of GPIO > > controllers, and would be similar to irq_domain_xlate_twocell(). The > > problem is that ->gpio_to_irq() effectively needs to convert the offset > > of the pin in the GPIO controller to an IRQ specifier. If the IRQ domain > > can use irq_domain_xlate_twocell(), that should be easy, but if it does > > not, then we likely need a custom implementation as well. >=20 > This sounds a lot like the "gpio-ranges" we have in the > gpiochip DT bindings, mapping gpio offsets to pin offsets. >=20 > I assume that we could just introduce a cross-mapping > array from IRQ to IRQ in struct gpio_irq_chip for the > hierarchical irqchip? Is it any > more complicated than an array of [(int, int)] tuples? >=20 > I guess this is what you have in mind for twocell? For twocell I think it would be even easier, because the IRQ specifier can just be reconstructed from (offset, irq_type). That's all the simple two-cell does, right? It's a 1:1 mapping of specifier to GPIO offset to IRQ offset. > > For example, as you may remember, the Tegra186 GPIO controller is > > somewhat quirky in that it has a number of banks, each of which can have > > any number of pins up to 8. However, in order to prevent users from > > attempting to use one of the non-existent GPIOs, we resorted to > > compacting the GPIO number space so that the GPIO specifier uses > > basically a (bank, pin) pair that is converted into a GPIO offset. The > > same is done for interrupt specifiers. >=20 > I guess this stuff is what you refer to? >=20 > #define TEGRA_MAIN_GPIO_PORT(port, base, count, controller) \ > [TEGRA_MAIN_GPIO_PORT_##port] =3D { \ > .name =3D #port, \ > .offset =3D base, \ > .pins =3D count, \ > .irq =3D controller, \ > } >=20 > static const struct tegra_gpio_port tegra186_main_ports[] =3D { > TEGRA_MAIN_GPIO_PORT( A, 0x2000, 7, 2), > TEGRA_MAIN_GPIO_PORT( B, 0x3000, 7, 3), > TEGRA_MAIN_GPIO_PORT( C, 0x3200, 7, 3), > TEGRA_MAIN_GPIO_PORT( D, 0x3400, 6, 3), > (...) >=20 > Maybe things have changed slightly. >=20 > As I see it there are some ways to go about this: >=20 > 1. Create one gpiochip per bank and just register the number of > GPIOs actually accessible per bank offset from 0. This works > if one does not insist on having one gpiochip covering all pins, > and as long as all usable pins are stacked from offset 0..n. > (Tegra186 doesn't do this, it is registering one chip for all.) >=20 > 2. If the above cannot be met, register enough pins to cover all > (e.g. 32 pins for a 32bit GPIO register) then mask off the > unused pins in .valid_mask in the gpio_chip. This was fairly > recently introduced to add ACPI support for Qualcomm, as > there were valid, but unusable GPIO offsets, but it can be > used to cut arbitrary holes in any range of offsets. >=20 > 3. Some driver-specific way. Which seems to be what Tegra186 > is doing. >=20 > Would you consider to move over to using method (2) to > get a more linear numberspace? I.e. register 8 GPIOs/IRQs > per port/bank and then just mask off the invalid ones? > .valid_mask in gpio_chip can be used for the GPIOs and > .valid_mask in the gpio_irq_chip can be used for IRQs. >=20 > Or do you think it is nonelegant? This is all pretty much the same discussion that I remember we had earlier this year. Nothing's changed so far. Back at the time I had pointed out that we'd be wasting a lot of memory by registering 8 GPIOs/IRQs per bank, because on average only about 60- 75% of the GPIOs are actually used. In addition we waste processing resources by having to check the GPIO offset against the valid mask every time we want to access a GPIO. I think that's inelegant, but from the rest of what you're saying you don't see it that way. > > Since there is no 1:1 relationship between the value in the specifier > > and the GPIO offset, we can't use irq_domain_xlate_twocell(). >=20 > Am I right that if you switch to method (2) above this is solved > and we get rid of the custom tegra186 xlate function =3D=3D big win? >=20 > > I think we can probably just implement the simple two-cell version in > > gpiochip_to_irq() directly and leave it up to drivers that require > > something more to override ->to_irq(). >=20 > And if what I assume (in my naive thinking) you can do with > .valid_mask is correct, then you can convert tegra186 to use > common twocell translation. >=20 > Sorry for being a pest, I just have a feeling we are reinventing > wheels here, I really want to pull as many fringe cases as > possible into gpiolib if I can so the maintenance gets > simpler. Like I said, we had this very discussion a couple of months ago, and I don't think I want to go through all of it again. I think, yes, I could make Tegra186 work with .valid_mask, even if I consider it wasteful. So if that's what you want, I'll go rewrite the driver so we'll never have to repeat this again. Thierry --yrj/dFKFPuw6o+aM Content-Type: application/pgp-signature; name="signature.asc" -----BEGIN PGP SIGNATURE----- iQIzBAABCAAdFiEEiOrDCAFJzPfAjcif3SOs138+s6EFAluqGToACgkQ3SOs138+ s6ERHw//XDTW9nxwCPEzsbYFKRwbfu9e7mm8/PI/iu5Gxu1K+4SMa9F++wDJ7OB/ /NoUJU3+dzs+Gft0BqtuTuNzaLXh/J7uTo8KOmUM321oG5zsg/iYbfNDmLgTBVlP 1gDymYvzoKgqhpEoj7dFpLk7UINeJ74mYRJGqdHIqJ2nh9MN+GLM9m/Lbi4mbv42 g3w9/Muuv4U3tVrT1pEiBfDyTCmHwKzqP+Vj7i+t8Z2xoyR3PJBYKK4EXO4d40u7 NnrWmudrzDnchj41GG5Ufpe/VwLtqmsiAfCXIZBMWItFDpTosc8E+6/l1Loq6YvQ MJd+ASxp0qteKmWXwaJPGfM4X9zT4YmM6w+hCc6thAZ+UtOCnnFAHioEobYwvODw 0taQFrNh/iLtpX29Zrwyl1vArG9yKPtFJBD6xT3O2fU3ZVAQPZyp2kvQ62xwVToT 1620TgkIR1OoSobvUPBsVNZetoaENV3UaRULb+Vf/yfGFHC/vmTKXIwZk1m7ELvc mhZH1bQFg9TAcBqt8jva5Yz0yKAoYA8wiMcMCrb2oIi80iH3wDROxaBbGH3intiV ju7F1tH7PHlbLbMwxfNqx96BFnsYJ6H3vP+aqohpJdgqmMKUAbnBkbuBNa8toeeQ vFULXxlIhIFuuD4Hfw4a3gP1BpCOQT+Nzhnq7sUf1nPA+pRKK2o= =+PDr -----END PGP SIGNATURE----- --yrj/dFKFPuw6o+aM--