| Message ID | 20170930034057.15166-3-opendmb@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Doug, On Fri, Sep 29, 2017 at 8:40 PM, Doug Berger <opendmb@gmail.com> wrote: > The basic memory-mapped GPIO controller lock must be released > before calling the registered GPIO interrupt handlers to allow > the interrupt handlers to access the hardware. Otherwise, the > hardware accesses will deadlock when they attempt to grab the > lock. I was having some trouble understanding exactly what the problem was here, but I think I see it now. Since this locks the entire bank, where some GPIOs might be set as inputs and some as inputs (and interrupt sources), then an interrupt on a GPIO that is supposed to set another GPIO in the bank would result in deadlock. Is that correct? If so, please update the commit message to make that clear, and nice fix. If not that, it would be nice to know what scenario can cause a problem. Thanks, Gregory
On 10/03/2017 06:55 PM, Gregory Fong wrote: > Hi Doug, > > On Fri, Sep 29, 2017 at 8:40 PM, Doug Berger <opendmb@gmail.com> wrote: >> The basic memory-mapped GPIO controller lock must be released >> before calling the registered GPIO interrupt handlers to allow >> the interrupt handlers to access the hardware. Otherwise, the >> hardware accesses will deadlock when they attempt to grab the >> lock. > > I was having some trouble understanding exactly what the problem was > here, but I think I see it now. Since this locks the entire bank, > where some GPIOs might be set as inputs and some as inputs (and > interrupt sources), then an interrupt on a GPIO that is supposed to > set another GPIO in the bank would result in deadlock. Is that > correct? If so, please update the commit message to make that clear, > and nice fix. If not that, it would be nice to know what scenario can > cause a problem. That is an example, but there are really many possibilities. Basically, if a registered interrupt handler wants to access its GPIO you are likely to run into trouble. Another example might be an interrupt that is configured to trigger on either edge transition and the handler wants to know whether the input is currently high or low. I can submit a V2 with a change in the description if you would like, but I'm not sure what the clearest example would be. > > Thanks, > Gregory > Thanks for the feedback, Doug
On Tue, Oct 3, 2017 at 7:09 PM, Doug Berger <opendmb@gmail.com> wrote: > On 10/03/2017 06:55 PM, Gregory Fong wrote: >> Hi Doug, >> >> On Fri, Sep 29, 2017 at 8:40 PM, Doug Berger <opendmb@gmail.com> wrote: >>> The basic memory-mapped GPIO controller lock must be released >>> before calling the registered GPIO interrupt handlers to allow >>> the interrupt handlers to access the hardware. Otherwise, the >>> hardware accesses will deadlock when they attempt to grab the >>> lock. >> >> I was having some trouble understanding exactly what the problem was >> here, but I think I see it now. Since this locks the entire bank, >> where some GPIOs might be set as inputs and some as inputs (and >> interrupt sources), then an interrupt on a GPIO that is supposed to >> set another GPIO in the bank would result in deadlock. Is that >> correct? If so, please update the commit message to make that clear, >> and nice fix. If not that, it would be nice to know what scenario can >> cause a problem. > > That is an example, but there are really many possibilities. > > Basically, if a registered interrupt handler wants to access its GPIO > you are likely to run into trouble. Another example might be an > interrupt that is configured to trigger on either edge transition and > the handler wants to know whether the input is currently high or low. > > I can submit a V2 with a change in the description if you would like, > but I'm not sure what the clearest example would be. If you could just mention both of these possible cases, that would be great! With that change, Acked-by: Gregory Fong <gregory.0xf0@gmail.com>
diff --git a/drivers/gpio/gpio-brcmstb.c b/drivers/gpio/gpio-brcmstb.c index 7f39b160a4d5..8945861876f9 100644 --- a/drivers/gpio/gpio-brcmstb.c +++ b/drivers/gpio/gpio-brcmstb.c @@ -62,6 +62,21 @@ brcmstb_gpio_gc_to_priv(struct gpio_chip *gc) return bank->parent_priv; } +static unsigned long +brcmstb_gpio_get_active_irqs(struct brcmstb_gpio_bank *bank) +{ + void __iomem *reg_base = bank->parent_priv->reg_base; + unsigned long status; + unsigned long flags; + + spin_lock_irqsave(&bank->gc.bgpio_lock, flags); + status = bank->gc.read_reg(reg_base + GIO_STAT(bank->id)) & + bank->gc.read_reg(reg_base + GIO_MASK(bank->id)); + spin_unlock_irqrestore(&bank->gc.bgpio_lock, flags); + + return status; +} + static void brcmstb_gpio_set_imask(struct brcmstb_gpio_bank *bank, unsigned int offset, bool enable) { @@ -204,11 +219,8 @@ static void brcmstb_gpio_irq_bank_handler(struct brcmstb_gpio_bank *bank) struct irq_domain *irq_domain = bank->gc.irqdomain; void __iomem *reg_base = priv->reg_base; unsigned long status; - unsigned long flags; - spin_lock_irqsave(&bank->gc.bgpio_lock, flags); - while ((status = bank->gc.read_reg(reg_base + GIO_STAT(bank->id)) & - bank->gc.read_reg(reg_base + GIO_MASK(bank->id)))) { + while ((status = brcmstb_gpio_get_active_irqs(bank))) { int bit; for_each_set_bit(bit, &status, 32) { @@ -223,7 +235,6 @@ static void brcmstb_gpio_irq_bank_handler(struct brcmstb_gpio_bank *bank) generic_handle_irq(irq_find_mapping(irq_domain, bit)); } } - spin_unlock_irqrestore(&bank->gc.bgpio_lock, flags); } /* Each UPG GIO block has one IRQ for all banks */
The basic memory-mapped GPIO controller lock must be released before calling the registered GPIO interrupt handlers to allow the interrupt handlers to access the hardware. Otherwise, the hardware accesses will deadlock when they attempt to grab the lock. Since the lock is only needed to protect the calculation of unmasked pending interrupts create a dedicated function to perform this and hide the complexity. Fixes: 19a7b6940b78 ("gpio: brcmstb: Add interrupt and wakeup source support") Signed-off-by: Doug Berger <opendmb@gmail.com> --- drivers/gpio/gpio-brcmstb.c | 21 ++++++++++++++++----- 1 file changed, 16 insertions(+), 5 deletions(-)