On 05.03.2024 11:50, Jesper Nilsson wrote: > devm_request_irq() is called before we initialize the "variant" > member variable from of_device_get_match_data(), so if an interrupt > is triggered inbetween, we can end up following a NULL pointer > in the interrupt handler. > > This problem was exposed when the I2C controller in question was > (mis)configured to be used in both secure world and Linux. > > That this can happen is also reflected by the existing code that > clears any pending interrupts from "u-boot or misc causes". > > Move the clearing of pending interrupts and the call to > devm_request_irq() to the end of probe. > > Reviewed-by: Andi Shyti <andi.shyti@xxxxxxxxxx> > Fixes: 218e1496135e ("i2c: exynos5: add support for HSI2C on Exynos5260 SoC") > Signed-off-by: Jesper Nilsson <jesper.nilsson@xxxxxxxx> > --- > Changes in v3: > - Avoid multiple assignment > - Link to v2: https://lore.kernel.org/r/20240304-i2c_exynos5-v2-1-7b9c312be719@xxxxxxxx > > Changes in v2: > - Use dev_err_probe() instead of open coding it > - Dropped the return failure if we can't find a match in devicetree > - Link to v1: https://lore.kernel.org/r/20240304-i2c_exynos5-v1-1-e91c889d2025@xxxxxxxx > --- > drivers/i2c/busses/i2c-exynos5.c | 29 +++++++++++++++-------------- > 1 file changed, 15 insertions(+), 14 deletions(-) > > diff --git a/drivers/i2c/busses/i2c-exynos5.c b/drivers/i2c/busses/i2c-exynos5.c > index 385ef9d9e4d4..8458e22313a7 100644 > --- a/drivers/i2c/busses/i2c-exynos5.c > +++ b/drivers/i2c/busses/i2c-exynos5.c > @@ -906,23 +906,9 @@ static int exynos5_i2c_probe(struct platform_device *pdev) > i2c->adap.algo_data = i2c; > i2c->adap.dev.parent = &pdev->dev; > > - /* Clear pending interrupts from u-boot or misc causes */ > - exynos5_i2c_clr_pend_irq(i2c); > - > spin_lock_init(&i2c->lock); > init_completion(&i2c->msg_complete); > > - i2c->irq = ret = platform_get_irq(pdev, 0); > - if (ret < 0) > - goto err_clk; > - > - ret = devm_request_irq(&pdev->dev, i2c->irq, exynos5_i2c_irq, > - IRQF_NO_SUSPEND, dev_name(&pdev->dev), i2c); > - if (ret != 0) { > - dev_err(&pdev->dev, "cannot request HS-I2C IRQ %d\n", i2c->irq); > - goto err_clk; > - } > - > i2c->variant = of_device_get_match_data(&pdev->dev); > > ret = exynos5_hsi2c_clock_setup(i2c); > @@ -940,6 +926,21 @@ static int exynos5_i2c_probe(struct platform_device *pdev) > clk_disable(i2c->clk); > clk_disable(i2c->pclk); > > + /* Clear pending interrupts from u-boot or misc causes */ > + exynos5_i2c_clr_pend_irq(i2c); Just above this call the clocks have been disabled, so any access to the i2c host registers will result in freeze or external abort (depending on the soc/cpu). To make things worse, this patch moved registering the interrupt handler after the i2c_add_adapter() call. This means that all i2c devices that will be probbed directly from i2c_add_adapter() won't be able to access the i2c bus, as the host controller is still not fully functional that time yet. This breaks today's linux-next on all Exynos5+ platforms. Has anyone tested this change? > + > + ret = platform_get_irq(pdev, 0); > + if (ret < 0) > + goto err_clk; > + i2c->irq = ret; > + > + ret = devm_request_irq(&pdev->dev, i2c->irq, exynos5_i2c_irq, > + IRQF_NO_SUSPEND, dev_name(&pdev->dev), i2c); > + if (ret != 0) { > + dev_err(&pdev->dev, "cannot request HS-I2C IRQ %d\n", i2c->irq); > + goto err_clk; > + } > + > return 0; > > err_clk: > Best regards -- Marek Szyprowski, PhD Samsung R&D Institute Poland