Documentation: gpio: pca953x: clarify interrupt source detection

There are multiple design tradeoffs and considerations in how the
PCA953x driver detects the source(s) of an interrupt. This driver
supports PCAL variants with input latching, a feature that is
constrained by the fact that the interrupt status and input port
registers cannot be read atomically. These limits and the design
decisions deserve an in-depth explanation.

Update the documentation to clarify these hardware limits and describe
how the driver determines pending interrupts, and how it makes use of
the PCAL input latching.

Signed-off-by: Ernest Van Hoecke <ernest.vanhoecke@toradex.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Link: https://lore.kernel.org/r/20260107093125.4053468-1-ernestvanhoecke@gmail.com
Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@oss.qualcomm.com>

Documentation/driver-api/gpio/pca953x.rst

@@ -389,6 +389,13 @@ disabled.
 Currently the driver enables the latch for each line with interrupt
 enabled.
 
+An interrupt status register records which pins triggered an interrupt.
+However, the status register and the input port register must be read
+separately; there is no atomic mechanism to read both simultaneously, so races
+are possible. Refer to the chapter `Interrupt source detection`_ to understand
+the implications of this and how the driver still makes use of the latching
+feature.
+
     1. base offset 0x40, bank 2, bank offsets of 2^n
         - pcal6408
         - pcal6416
@@ -523,6 +530,74 @@ bits drive strength
 
 Currently not supported by the driver.
 
+Interrupt source detection
+==========================
+
+When triggered by the GPIO expander's interrupt, the driver determines which
+IRQs are pending by reading the input port register.
+
+To be able to filter on specific interrupt events for all compatible devices,
+the driver keeps track of the previous input state of the lines, and emits an
+IRQ only for the correct edge or level. This system works irrespective of the
+number of enabled interrupts. Events will not be missed even if they occur
+between the GPIO expander's interrupt and the actual I2C read. Edges could of
+course be missed if the related signal level changes back to the value
+previously saved by the driver before the I2C read. PCAL variants offer input
+latching for that reason.
+
+PCAL input latching
+-------------------
+
+The PCAL variants have an input latch and the driver enables this for all
+interrupt-enabled lines. The interrupt is then only cleared when the input port
+is read out. These variants provide an interrupt status register that records
+which pins triggered an interrupt, but the status and input registers cannot be
+read atomically. If another interrupt occurs on a different line after the
+status register has been read but before the input port register is sampled,
+that event will not be reflected in the earlier status snapshot, so relying
+solely on the interrupt status register is insufficient.
+
+Thus, the PCAL variants also have to use the existing level-change logic.
+
+For short pulses, the first edge is captured when the input register is read,
+but if the signal returns to its previous level before this read, the second
+edge is not observed. As a result, successive pulses can produce identical
+input values at read time and no level change is detected, causing interrupts
+to be missed. Below timing diagram shows this situation where the top signal is
+the input pin level and the bottom signal indicates the latched value::
+
+  ─────┐     ┌──*───────────────┐     ┌──*─────────────────┐     ┌──*───
+       │     │  .               │     │  .                 │     │  .
+       │     │  │               │     │  │                 │     │  │
+       └──*──┘  │               └──*──┘  │                 └──*──┘  │
+  Input   │     │                  │     │                    │     │
+          ▼     │                  ▼     │                    ▼     │
+         IRQ    │                 IRQ    │                   IRQ    │
+                .                        .                          .
+  ─────┐        .┌──────────────┐        .┌────────────────┐        .┌──
+       │         │              │         │                │         │
+       │         │              │         │                │         │
+       └────────*┘              └────────*┘                └────────*┘
+  Latched       │                        │                          │
+                ▼                        ▼                          ▼
+              READ 0                   READ 0                     READ 0
+                                     NO CHANGE                  NO CHANGE
+
+To deal with this, events indicated by the interrupt status register are merged
+with events detected through the existing level-change logic. As a result:
+
+- short pulses, whose second edges are invisible, are detected via the
+  interrupt status register, and
+- interrupts that occur between the status and input reads are still
+  caught by the generic level-change logic.
+
+Note that this is still best-effort: the status and input registers are read
+separately, and short pulses on other lines may occur in between those reads.
+Such pulses can still be latched as an interrupt without leaving an observable
+level change at read time, and may not be attributable to a specific edge. This
+does not reduce detection compared to the generic path, but reflects inherent
+atomicity limitations.
+
 Datasheets
 ==========