mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git
synced 2024-08-23 09:19:51 +00:00
![]() The qpnpint_irq_set_type() callback function configures the type (edge vs level) and polarity (high, low, or both) of a particular PMIC interrupt within a given peripheral. To do this, it reads the three consecutive IRQ configuration registers, modifies the specified IRQ bit within the register values, and finally writes the three modified register values back to the PMIC. While a spinlock is used to provide mutual exclusion on the SPMI bus during the register read and write calls, there is no locking around the overall read, modify, write sequence. This opens up the possibility of a race condition if two tasks set the type of a PMIC IRQ within the same peripheral simultaneously. When the race condition is encountered, both tasks will read the old value of the registers and IRQ bits set by one of the tasks will be dropped upon the register write of the other task. This then leads to PMIC IRQs being enabled with an incorrect type and polarity configured. Such misconfiguration can lead to an IRQ storm that overwhelms the system and causes it to crash. This race condition and IRQ storm have been observed when using a pair of pm8941-pwrkey devices to handle PMK8350 pwrkey and resin interrupts. The independent devices probe asynchronously in parallel and can simultaneously request and configure PMIC IRQs in the same PMIC peripheral. For a good case, the IRQ configuration calls end up serialized due to timing deltas and the register read/write sequence looks like this: 1. pwrkey probe: SPMI read(0x1311): 0x00, 0x00, 0x00 2. pwrkey probe: SPMI write(0x1311): 0x80, 0x80, 0x80 3. resin probe: SPMI read(0x1311): 0x80, 0x80, 0x80 4. resin probe: SPMI write(0x1311): 0xC0, 0xC0, 0xC0 The final register states after both devices have requested and enabled their respective IRQs is thus: 0x1311: 0xC0 0x1312: 0xC0 0x1313: 0xC0 0x1314: 0x00 0x1315: 0xC0 For a bad case, the IRQ configuration calls end up occurring simultaneously and the race condition is encountered. The register read/write sequence then looks like this: 1. pwrkey probe: SPMI read(0x1311): 0x00, 0x00, 0x00 2. resin probe: SPMI read(0x1311): 0x00, 0x00, 0x00 3. pwrkey probe: SPMI write(0x1311): 0x80, 0x80, 0x80 4. resin probe: SPMI write(0x1311): 0x40, 0x40, 0x40 In this case, the final register states after both devices have requested and enabled their respective IRQs is thus: 0x1311: 0x40 0x1312: 0x40 0x1313: 0x40 0x1314: 0x00 0x1315: 0xC0 This corresponds to the resin IRQ being configured for both rising and falling edges, as expected. However, the pwrkey IRQ is misconfigured as level type with both polarity high and low set to disabled. The PMIC IRQ triggering hardware treats this particular register configuration as if level low triggering is enabled. The raw pwrkey IRQ signal is low when the power key is not being pressed. Thus, the pwrkey IRQ begins firing continuously in an IRQ storm. Fix the race condition by holding the spmi-pmic-arb spinlock for the duration of the read, modify, write sequence performed in the qpnpint_irq_set_type() function. Split the pmic_arb_read_cmd() and pmic_arb_write_cmd() functions each into three parts so that hardware register IO is decoupled from spinlock locking. This allows a new function pmic_arb_masked_write() to be added which locks the spinlock and then calls register IO functions to perform SPMI read and write commands in a single atomic operation. Link: https://lore.kernel.org/r/20211118034719.28971-1-quic_collinsd@quicinc.com Signed-off-by: David Collins <quic_collinsd@quicinc.com> Signed-off-by: Stephen Boyd <sboyd@kernel.org> Link: https://lore.kernel.org/r/20211216190812.1574801-7-sboyd@kernel.org Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> |
||
---|---|---|
.. | ||
hisi-spmi-controller.c | ||
Kconfig | ||
Makefile | ||
spmi-mtk-pmif.c | ||
spmi-pmic-arb.c | ||
spmi.c |