From: sashiko-bot@kernel.org
To: "Szuying Chen" <Chloe_Chen@asmedia.com.tw>, linux-gpio@vger.kernel.org
Subject: Re: [PATCH v4] gpio: amdptpci: add support for AMD PT PCI controller
Date: Mon, 16 Nov 2026 08:42:54 +0000 [thread overview]
Message-ID: <sashiko-outbox-216556@kernel.org> (raw)
In-Reply-To: <20261116075510.5722-1-Chloe_Chen@asmedia.com.tw>
Thank you for your contribution! Sashiko AI review found 6 potential
issues to consider:
- [High] gpio: amdptpci: devres order in pt27_gpio_probe() frees
irq_domain early
- [High] gpio: amdptpci: MMIO access while suspended and clobber in
pt27_restore_regs()
- [High] gpio: amdptpci: runtime PM calls in pt27_gpio_isr() break IRQs
and lockdep
- [Medium] gpio: amdptpci: sleep ops in pt27_gpio_pm_ops race noirq wake
IRQ handling
- [Medium] gpio: amdptpci: unmasked wake_status and skipped wake disable
on suspend
- [Medium] gpio: amdptpci: duplicate wake IRQ dispatch in
pt27_gpio_resume()
--
commit afe67f7f6f93220edaeaabd38709a16694605523
Author: Szuying Chen <Chloe_Chen@asmedia.com.tw>
gpio: amdptpci: add support for AMD PT PCI controller
This commit adds a PCI GPIO and interrupt controller driver for AMD
Promontory (PT27) platforms, including runtime PM and system wakeup support.
> diff --git a/drivers/gpio/gpio-amdptpci.c b/drivers/gpio/gpio-amdptpci.c
> new file mode 100644
> index 0000000000000..d676fd8c0b1c7
> --- /dev/null
> +++ b/drivers/gpio/gpio-amdptpci.c
[ ... ]
> +static void pt27_restore_regs(struct pt27_gpio *chip)
> +{
> + iowrite32(chip->saved.out, chip->base + PT27_REG_OUT);
> + iowrite32(chip->saved.dir, chip->base + PT27_REG_DIR);
> +
> + iowrite32(chip->saved.int_lvl, chip->base + PT27_REG_INT_LVL);
> + iowrite32(chip->saved.int_mode0, chip->base + PT27_REG_INT_MODE0);
> + iowrite16(chip->saved.int_mode1, chip->base + PT27_REG_INT_MODE1);
> + iowrite8(chip->saved.int_ctl, chip->base + PT27_REG_INT_CTL);
> +
> + iowrite16(chip->saved.wake_lvl, chip->base + PT27_REG_WAKE_LVL);
> + iowrite16(chip->saved.wake_mode, chip->base + PT27_REG_WAKE_MODE);
> + iowrite16(chip->saved.wake_en, chip->base + PT27_REG_WAKE_EN);
> +
> + iowrite32(chip->saved.int_en, chip->base + PT27_REG_INT_EN);
> + iowrite32(chip->saved.int_mask, chip->base + PT27_REG_INT_MASK);
> +}
[ ... ]
> +static int pt27_irq_reqres(struct irq_data *d)
> +{
> + struct gpio_chip *gc = irq_data_get_irq_chip_data(d);
> + int ret;
> +
> + ret = pm_runtime_resume_and_get(gc->parent);
> + if (ret < 0) {
> + dev_err(gc->parent, "%s: Fail to resume controller for IRQ: %d\n", __func__, ret);
> + return ret;
> + }
> +
> + ret = gpiochip_reqres_irq(gc, irqd_to_hwirq(d));
> + if (ret) {
> + pm_runtime_put_autosuspend(gc->parent);
> + return ret;
> + }
> + return 0;
> +}
[ ... ]
> +static int pt27_irq_set_type(struct irq_data *d, unsigned int type)
> +{
[ ... ]
> + scoped_guard(raw_spinlock_irqsave, &chip->gen.lock) {
> + lvl_reg = ioread32(chip->base + PT27_REG_INT_LVL);
> + if (lvl_bit)
> + lvl_reg |= BIT(hwirq);
> + else
> + lvl_reg &= ~BIT(hwirq);
> +
> + iowrite32(lvl_reg, chip->base + PT27_REG_INT_LVL);
> +
> + if (hwirq < 16) {
> + u32 reg = ioread32(chip->base + PT27_REG_INT_MODE0);
> +
> + shift = hwirq * 2;
> + reg &= ~(PT27_INT_MODE_MASK << shift);
> + reg |= ((u32)mode_val << shift);
> + iowrite32(reg, chip->base + PT27_REG_INT_MODE0);
> + } else {
> + u16 reg = ioread16(chip->base + PT27_REG_INT_MODE1);
> +
> + shift = (hwirq - 16) * 2;
> + reg &= ~((u16)(PT27_INT_MODE_MASK << shift));
> + reg |= (u16)((u16)mode_val << shift);
> + iowrite16(reg, chip->base + PT27_REG_INT_MODE1);
> + }
> +
> + iowrite32(BIT(hwirq), chip->base + PT27_REG_INT_STAT);
> +
> + }
[ ... ]
> +static int pt27_irq_set_wake(struct irq_data *d, unsigned int on)
> +{
[ ... ]
> + guard(raw_spinlock_irqsave)(&chip->gen.lock);
> +
> + wake_en = ioread16(chip->base + PT27_REG_WAKE_EN);
> + wake_lvl = ioread16(chip->base + PT27_REG_WAKE_LVL);
> + wake_mode = ioread16(chip->base + PT27_REG_WAKE_MODE);
[Severity: High]
Can GPIO IRQ setup access MMIO while the PCI device is in D3 and then have
pt27_restore_regs() overwrite the configuration on resume?
Runtime PM references are only taken in pt27_gpio_request() and
pt27_irq_reqres(). Standard IRQ mapping paths such as
acpi_dev_gpio_irq_wake_get_by() configure input direction via
gpiod_direction_input_nonotify() -> gpio_mmio_dir_in() and trigger type via
irq_set_irq_type() -> pt27_irq_set_type() before request_irq() is called and
without calling gpiod_request().
If the controller has already runtime-suspended after probe, those MMIO
reads and writes happen while the device is in D3hot/D3cold and do not
update chip->saved. When request_threaded_irq() then calls
pt27_irq_reqres(), pm_runtime_resume_and_get() runs pt27_restore_regs(),
clobbering PT27_REG_DIR, PT27_REG_INT_LVL, PT27_REG_INT_MODE0/1, and
PT27_REG_WAKE_* with the pre-suspend values in chip->saved. Right after
that, gpiochip_reqres_irq() -> gpiochip_lock_as_irq() reads PT27_REG_DIR
from hardware via gpio_mmio_get_dir() and fails with -EIO if the pin was
previously an output, or leaves the interrupt trigger unconfigured.
Also, if irq_set_irq_type() is called after enable_irq_wake(),
pt27_irq_set_type() does not update PT27_REG_WAKE_LVL or PT27_REG_WAKE_MODE.
[ ... ]
> +static irqreturn_t pt27_gpio_isr(int irq, void *data)
> +{
> + struct pt27_gpio *chip = data;
> + irqreturn_t ret = IRQ_NONE;
> + unsigned long pending;
> +
> + if (pm_runtime_get_if_in_use(chip->dev) <= 0)
> + return IRQ_NONE;
[ ... ]
> + pm_runtime_put(chip->dev);
> + return ret;
> +}
[Severity: High]
Does calling pm_runtime_get_if_in_use() and pm_runtime_put() in
pt27_gpio_isr() break interrupt handling when CONFIG_PM is disabled and take
a sleeping spinlock in hardirq context?
GPIO_AMDPTPCI does not depend on PM, and when CONFIG_PM=n,
pm_runtime_get_if_in_use() returns -EINVAL, causing pt27_gpio_isr() to
return IRQ_NONE on every interrupt. Even with CONFIG_PM=y,
pm_runtime_get_if_in_use() returns 0 while runtime_status is still
RPM_RESUMING (after pt27_restore_regs() unmasks interrupts in
pt27_gpio_runtime_resume()) and returns -EINVAL during early system resume
before runtime PM is re-enabled.
In addition, pt27_gpio_isr() is requested with IRQF_NO_THREAD, so calling
pm_runtime_get_if_in_use() and pm_runtime_put() acquires dev->power.lock (a
non-raw spinlock_t) from hardirq context, triggering lockdep invalid wait
context warnings under CONFIG_PROVE_RAW_LOCK_NESTING and sleeping in atomic
context on PREEMPT_RT.
> +static void pt27_hw_clear(void *data)
> +{
> + struct pt27_gpio *chip = data;
> +
> + iowrite32(0, chip->base + PT27_REG_INT_EN);
> + iowrite32(PT27_GPIO_MASK, chip->base + PT27_REG_INT_MASK);
> + iowrite32(PT27_GPIO_MASK, chip->base + PT27_REG_INT_STAT);
> + iowrite16(0, chip->base + PT27_REG_WAKE_EN);
> + iowrite16(PT27_WAKE_MASK, chip->base + PT27_REG_WAKE_STAT);
> +}
[ ... ]
> +static int pt27_gpio_probe(struct pci_dev *pdev, const struct pci_device_id *id)
> +{
[ ... ]
> + ret = devm_request_irq(dev, chip->irq, pt27_gpio_isr,
> + irq_flags | IRQF_NO_THREAD, dev_name(dev), chip);
> + if (ret) {
> + dev_err_probe(dev, ret, "failed to request IRQ %d\n", chip->irq);
> + goto err_pm_disable;
> + }
> +
> + ret = devm_add_action_or_reset(dev, pt27_hw_clear, chip);
> + if (ret)
> + goto err_pm_disable;
> +
> + ret = devm_gpiochip_add_data(dev, gc, chip);
[Severity: High]
Does this devres registration order in pt27_gpio_probe() cause a
use-after-free of chip->gen.gc.irq.domain in pt27_gpio_isr() during unbind?
Because devm_request_irq() and pt27_hw_clear() are registered before
devm_gpiochip_add_data(), devres teardown runs in reverse order when the
driver is unbound:
CPU0 CPU1
pci_device_remove()
pt27_gpio_remove()
pm_runtime_forbid(dev)
pt27_restore_regs()
devres_release_all()
gpiochip_remove()
irq_domain_remove(gc->irq.domain)
<hardirq>
pt27_gpio_isr()
generic_handle_domain_irq(
chip->gen.gc.irq.domain, hwirq)
<-- use-after-free
pt27_hw_clear()
devm_irq_release()
During probe, chip->gen.gc.irq.domain is also still NULL between
devm_request_irq() and devm_gpiochip_add_data(). In addition,
pt27_hw_clear() modifies PT27_REG_INT_EN, PT27_REG_INT_MASK, and
PT27_REG_INT_STAT without holding chip->gen.lock while pt27_gpio_isr() is
still active, and pt27_gpio_isr() does not mask pending with PT27_GPIO_MASK.
Registering devm_gpiochip_add_data() before pt27_hw_clear() and
devm_request_irq() ensures the IRQ handler is freed first on teardown.
[ ... ]
> +static int __maybe_unused pt27_gpio_suspend(struct device *dev)
> +{
> + struct pt27_gpio *chip = dev_get_drvdata(dev);
> +
> + if (pm_runtime_status_suspended(dev)) {
> + dev_dbg(dev, "%s: already runtime-suspended, skipping save\n", __func__);
> + return 0;
> + }
> +
> + scoped_guard(raw_spinlock_irqsave, &chip->gen.lock) {
> + pt27_save_regs(chip);
> + iowrite32(0, chip->base + PT27_REG_INT_EN);
> + iowrite32(PT27_GPIO_MASK, chip->base + PT27_REG_INT_MASK);
> + iowrite32(PT27_GPIO_MASK, chip->base + PT27_REG_INT_STAT);
> +
> + if (!device_may_wakeup(dev))
> + iowrite16(0, chip->base + PT27_REG_WAKE_EN);
> +
> + iowrite16(PT27_WAKE_MASK, chip->base + PT27_REG_WAKE_STAT);
> + }
> +
> + synchronize_irq(chip->irq);
> + return 0;
> +}
> +
> +static int __maybe_unused pt27_gpio_resume(struct device *dev)
> +{
> + struct pt27_gpio *chip = dev_get_drvdata(dev);
> + unsigned long replay;
> + unsigned int hwirq;
> + u16 wake_status;
> +
> + if (pm_runtime_status_suspended(dev))
> + return 0;
> +
> + scoped_guard(raw_spinlock_irqsave, &chip->gen.lock) {
> + wake_status = ioread16(chip->base + PT27_REG_WAKE_STAT) & PT27_WAKE_MASK;
> + iowrite16(wake_status, chip->base + PT27_REG_WAKE_STAT);
> + pt27_restore_regs(chip);
> +
> + replay = wake_status &
> + ioread32(chip->base + PT27_REG_INT_EN) &
> + ~ioread32(chip->base + PT27_REG_INT_MASK);
> + }
> +
> + if (!wake_status)
> + return 0;
> +
> + pm_wakeup_event(dev, 0);
[Severity: Medium]
Should wake_status in pt27_gpio_resume() be masked with chip->wake_mask, and
what happens in pt27_gpio_suspend() if wakeup was disabled after the device
entered runtime suspend?
In pt27_gpio_resume() (which is also called by pt27_gpio_runtime_resume()),
wake_status is only masked with PT27_WAKE_MASK, so any latched status bits
on pins 0..11 that do not have wakeup enabled in chip->wake_mask will still
report pm_wakeup_event() and replay interrupts.
In pt27_gpio_suspend(), if the device entered runtime suspend with
PT27_REG_WAKE_EN enabled and wakeup was then disabled via sysfs before
system suspend, returning early when pm_runtime_status_suspended(dev) is
true skips clearing PT27_REG_WAKE_EN, leaving hardware wakeup active during
system sleep.
> +
> + for_each_set_bit(hwirq, &replay, PT27_WAKE_GPIOS)
> + generic_handle_domain_irq_safe(chip->gen.gc.irq.domain, hwirq);
[Severity: Medium]
Can pt27_gpio_resume() dispatch the same wake interrupt twice?
Inside the lock scope above, pt27_restore_regs() writes PT27_REG_INT_EN and
PT27_REG_INT_MASK without clearing PT27_REG_INT_STAT, and then computes
replay. As soon as the scoped_guard releases chip->gen.lock and restores
local interrupts, an asserted level-triggered wake line or a latched bit in
PT27_REG_INT_STAT can fire chip->irq and run pt27_gpio_isr() ->
generic_handle_domain_irq(). When pt27_gpio_isr() returns, this loop still
iterates over the pre-computed replay bitmap and calls
generic_handle_domain_irq_safe() a second time for the same hwirq, which
invokes handle_irq_event() again even if the interrupt was already serviced
and masked.
[ ... ]
> +static const struct dev_pm_ops pt27_gpio_pm_ops __maybe_unused = {
> + SYSTEM_SLEEP_PM_OPS(pt27_gpio_suspend, pt27_gpio_resume)
> + RUNTIME_PM_OPS(pt27_gpio_runtime_suspend, pt27_gpio_runtime_resume, NULL)
> +};
[Severity: Medium]
Should pt27_gpio_suspend() and pt27_gpio_resume() be wired as noirq sleep
callbacks instead of SYSTEM_SLEEP_PM_OPS()?
With SYSTEM_SLEEP_PM_OPS(), pt27_gpio_suspend() saves PT27_REG_WAKE_EN and
masks interrupts during dpm_suspend(), before dpm_suspend_noirq() runs
device_wakeup_arm_wake_irqs() and suspend_device_irqs(). On resume,
dpm_resume_noirq() runs device_wakeup_disarm_wake_irqs() (calling
pt27_irq_set_wake(d, 0)) before pt27_gpio_resume() runs in dpm_resume(), so
pt27_restore_regs() overwrites PT27_REG_WAKE_EN with chip->saved.wake_en and
clobbers any GPIO or wake state updated during the late/noirq/early phases.
--
Sashiko AI review · https://sashiko.dev/#/patchset/20261116075510.5722-1-Chloe_Chen@asmedia.com.tw?part=1
next prev parent reply other threads:[~2026-11-16 8:42 UTC|newest]
Thread overview: 3+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-16 7:55 [PATCH v4] gpio: amdptpci: add support for AMD PT PCI controller Szuying Chen
2026-11-16 8:42 ` sashiko-bot [this message]
2026-11-16 10:03 ` Markus Elfring
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=sashiko-outbox-216556@kernel.org \
--to=sashiko-bot@kernel.org \
--cc=Chloe_Chen@asmedia.com.tw \
--cc=linux-gpio@vger.kernel.org \
--cc=sashiko-reviews@lists.linux.dev \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.