On 12/28/24 8:24 AM, Jonathan Cameron wrote: > On Tue, 24 Dec 2024 10:34:32 +0100 > Julien Stephan <jstephan@xxxxxxxxxxxx> wrote: > >> The alert functionality is an out of range indicator and can be used as an >> early indicator of an out of bounds conversion result. >> >> ALERT_LOW_THRESHOLD and ALERT_HIGH_THRESHOLD registers are common to all >> channels. >> >> When using 1 SDO line (only mode supported by the driver right now), i.e >> data outputs only on SDOA, SDOB (or SDOD for 4 channels variants) is >> used as an alert pin. The alert pin is updated at the end of the >> conversion (set to low if an alert occurs) and is cleared on a falling >> edge of CS. >> >> The ALERT register contains information about the exact alert status: >> channel and direction. Unfortunately we can't read this register because >> in buffered read we cannot claim for direct mode. >> >> User can set high/low thresholds and enable event detection using the >> regular iio events: >> >> events/in_thresh_falling_value >> events/in_thresh_rising_value >> events/thresh_either_en >> >> If the interrupt properties is present in the device tree, an IIO event >> will be generated for each interrupt received. >> Because we cannot read ALERT register, we can't determine the exact >> channel that triggers the alert, neither the direction (hight/low >> threshold violation), so we send and IIO_EV_DIR_EITHER event for all >> channels. >> >> In buffered reads, if input stays out of thresholds limit, an interrupt >> will be generated for each sample read, because the alert pin is cleared >> on a falling edge of CS (i.e when starting a new conversion). To avoid >> generating to much interrupt, we introduce a reset_timeout that can be >> used to disable interrupt for a given time (in ms) >> >> events/thresh_either_reset_timeout >> >> When an interrupt is received, interrupts are disabled and re-enabled >> after thresh_either_reset_timeout ms. If the reset timeout is set to 0, >> interrupt are re-enabled directly. >> Note: interrupts are always disabled at least during the handling of the >> previous interrupt, because each read triggers 2 transactions, that can >> lead to 2 interrupts for a single user read. IRQF_ONESHOT is not enough, >> because, it postpones the 2nd irq after the handling of the first one, >> which can still trigger 2 interrupts for a single user read. > > After some of our recent discussions around interrupt handling and > the guarantees (that aren't) made, even disabling the interrupt doesn't > prevent some irq chips queuing up future interrupts. > > https://lore.kernel.org/all/io53lznz3qp3jd5rohqsjhosnmdzd6d44sdbwu5jcfrs3rz2a2@orquwgflrtyc/ > > I'm not sure this alert can actually work as a result :( > I am struggling to come up with a scheme that will work. > Would it work if we change it to a level-triggered interrupt instead of edge triggered? Since the main purpose of this is to trigger a hardware shutdown, perhaps we could just omit the interrupt/emitting the event and keep the threshold and enable attributes if we can't come up with a reasonable way to handle the interrupts?
