On 2021-02-08 19:26:28 [+0100], angelo@xxxxxxxxxxxxxxxx wrote: > Hi Sebastian, Hi, > > From the backtrace it looks like the IRQ #464 is handled by > > lpi2c_imx_isr(), returns always with IRQ_NONE but the IRQ fires again > > and again. > > On imx_5.4.70_2.3.0, still the irq routine is returning > always IRQ_HANDLED. So i am supposing proper handling is > probably not the issue. > > https://source.codeaurora.org/external/imx/linux-imx/tree/drivers/i2c/busses/i2c-imx-lpi2c.c?h=imx_5.4.70_2.3.0#n528 interesting. This does not explain why the core disabled it. > > It appears that IRQ-masking is not working properly but then the > > stacktrace says handle_edge_irq() and I have my doubts if it is really > > an EDGE interrupt. > > As an interrupt parent for this i2c module, devicetree > specify a special "intmux" controller/module, > (interrupt multiplexer module). > > https://source.codeaurora.org/external/imx/linux-imx/tree/drivers/irqchip/irq-imx-intmux.c?h=imx_5.4.70_2.3.0 > > It uses handle_edge_irq so it should be fine,. It is hard to tell without a look into the manual. From browsing over the driver, that lock in imx_intmux_irq_mask() should be a raw_spinlock_t. There should be a warning otherwise with CONFIG_DEBUG_ATOMIC_SLEEP. But then this is an EDGE handler without irq_ack. So it does nothing. And mask/unmask is probably only used on request_irq() / free_irq() so no warning here. > Issue seems coming after an initial i2c "read" from a pca953x i2c-gpio > expander connected in this i2c bus. This i2c register read is issued > from > > ret = devm_gpiochip_add_data(&client->dev, &chip->gpio_chip, chip); > > https://source.codeaurora.org/external/imx/linux-imx/tree/drivers/gpio/gpio-pca953x.c?h=imx_5.4.70_2.3.0#n1112 > > That following a bit the code seems to involve some spinlocks. > > After this i2c read, that probably triggers the first interrupt, > looking by oscilloscope, seems i2c communication is interrupted. > > I am still investigating on this, every help/hint is welcome. My guess is that the interrupt muxing is not working properly. Sebastian
