On Sat, 14 Nov 2020 21:16:37 +0100 Andrew Lunn <andrew@xxxxxxx> wrote: > On Sat, Nov 14, 2020 at 06:47:17PM +0100, Marek Behun wrote: > > On Sat, 14 Nov 2020 18:42:21 +0100 > > Marek Behún <kabel@xxxxxxxxxx> wrote: > > > > > Do you think we should just poll for interrupts with the PHY? > > Yes. > > > > > Andrew, JFI, there are also SFP GPIOs connected to this expander. So > > interrupt will be generated on changes from those gpios as well. Is an > > interrupt storm seriosly probable here? > > As far as i remember on the device i was using, all inputs are > interrupt sources. And all pins default as inputs, so you don't > accidentally drive the output against something else. So you need to > get the device properly configured before enabling its interrupt. And > i'm not too sure how that works with the GPIO framework where you > effectively have a collection of individual GPIOs. How do you know > they have all be configured, and it is safe to enable the interrupt > output? Do you have any pins floating? Those you are going to have to > configure as output. Plus is the interrupt output from the gpio > expander level triggered? If so you need to ensure the upstream > interrupt controller can do level. And Marvell SoCs often cannot. > > The device i was working on i had trouble even getting it to boot > without dying. Polling the PHY and SFP was working O.K, so i gave > up. And i suggested the next revision of the board used an sx150x > which has much better interrupt support. That has worked well against > a Vybrid. > > Andrew Thanks, Andrew for this explanation. What about when a interrupt pin from a Marvell PHY (level interrupt) is connected directly to a gpio pin on Marvell SOC (edge only interrupt), without any GPIO expander? This is the case on Turris MOX, and in drivers/pinctrl/mvebu/pinctrl-armada-37xx.c the function armada_37xx_irq_set_type again supports only edge interrupts. But in this case I think setting it to EDGE_FALLING should be okay, should it not? This controller has a special register where interrupt status is latched... Marek
