diff mbox series

[RESEND,v6,6/7] gpiolib: add new ioctl() for monitoring changes in line info

Message ID 20200211091937.29558-7-brgl@bgdev.pl
State New
Headers show
Series gpiolib: add an ioctl() for monitoring line status changes | expand

Commit Message

Bartosz Golaszewski Feb. 11, 2020, 9:19 a.m. UTC
From: Bartosz Golaszewski <bgolaszewski@baylibre.com>

Currently there is no way for user-space to be informed about changes
in status of GPIO lines e.g. when someone else requests the line or its
config changes. We can only periodically re-read the line-info. This
is fine for simple one-off user-space tools, but any daemon that provides
a centralized access to GPIO chips would benefit hugely from an event
driven line info synchronization.

This patch adds a new ioctl() that allows user-space processes to reuse
the file descriptor associated with the character device for watching
any changes in line properties. Every such event contains the updated
line information.

Currently the events are generated on three types of status changes: when
a line is requested, when it's released and when its config is changed.
The first two are self-explanatory. For the third one: this will only
happen when another user-space process calls the new SET_CONFIG ioctl()
as any changes that can happen from within the kernel (i.e.
set_transitory() or set_debounce()) are of no interest to user-space.

Signed-off-by: Bartosz Golaszewski <bgolaszewski@baylibre.com>
---
 drivers/gpio/gpiolib.c    | 186 ++++++++++++++++++++++++++++++++++++--
 drivers/gpio/gpiolib.h    |   1 +
 include/uapi/linux/gpio.h |  30 ++++++
 3 files changed, 209 insertions(+), 8 deletions(-)

Comments

Linus Walleij Feb. 12, 2020, 10:47 a.m. UTC | #1
On Tue, Feb 11, 2020 at 10:19 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:

> From: Bartosz Golaszewski <bgolaszewski@baylibre.com>
>
> Currently there is no way for user-space to be informed about changes
> in status of GPIO lines e.g. when someone else requests the line or its
> config changes. We can only periodically re-read the line-info. This
> is fine for simple one-off user-space tools, but any daemon that provides
> a centralized access to GPIO chips would benefit hugely from an event
> driven line info synchronization.
>
> This patch adds a new ioctl() that allows user-space processes to reuse
> the file descriptor associated with the character device for watching
> any changes in line properties. Every such event contains the updated
> line information.
>
> Currently the events are generated on three types of status changes: when
> a line is requested, when it's released and when its config is changed.
> The first two are self-explanatory. For the third one: this will only
> happen when another user-space process calls the new SET_CONFIG ioctl()
> as any changes that can happen from within the kernel (i.e.
> set_transitory() or set_debounce()) are of no interest to user-space.
>
> Signed-off-by: Bartosz Golaszewski <bgolaszewski@baylibre.com>

Looks good to me. This got really slim and clean after
the reviews, and I am of course also impressed by the kfifo
improvement this brings.

Reviewed-by: Linus Walleij <linus.walleij@linaro.org>

A question:

Bartosz, since you know about possible impacts on userspace,
since this code use the preferred ktime_get_ns() rather than
ktime_get_ns_real(), what happens if we just patch the other
event timestamp to use ktime_get_ns() instead, so we use the
same everywhere?

If it's fine I'd like to just toss in a patch for that as well.

Yours,
Linus Walleij
Bartosz Golaszewski Feb. 12, 2020, 11 a.m. UTC | #2
śr., 12 lut 2020 o 11:47 Linus Walleij <linus.walleij@linaro.org> napisał(a):
>
> On Tue, Feb 11, 2020 at 10:19 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
>
> > From: Bartosz Golaszewski <bgolaszewski@baylibre.com>
> >
> > Currently there is no way for user-space to be informed about changes
> > in status of GPIO lines e.g. when someone else requests the line or its
> > config changes. We can only periodically re-read the line-info. This
> > is fine for simple one-off user-space tools, but any daemon that provides
> > a centralized access to GPIO chips would benefit hugely from an event
> > driven line info synchronization.
> >
> > This patch adds a new ioctl() that allows user-space processes to reuse
> > the file descriptor associated with the character device for watching
> > any changes in line properties. Every such event contains the updated
> > line information.
> >
> > Currently the events are generated on three types of status changes: when
> > a line is requested, when it's released and when its config is changed.
> > The first two are self-explanatory. For the third one: this will only
> > happen when another user-space process calls the new SET_CONFIG ioctl()
> > as any changes that can happen from within the kernel (i.e.
> > set_transitory() or set_debounce()) are of no interest to user-space.
> >
> > Signed-off-by: Bartosz Golaszewski <bgolaszewski@baylibre.com>
>
> Looks good to me. This got really slim and clean after
> the reviews, and I am of course also impressed by the kfifo
> improvement this brings.
>
> Reviewed-by: Linus Walleij <linus.walleij@linaro.org>
>
> A question:
>
> Bartosz, since you know about possible impacts on userspace,
> since this code use the preferred ktime_get_ns() rather than
> ktime_get_ns_real(), what happens if we just patch the other
> event timestamp to use ktime_get_ns() instead, so we use the
> same everywhere?
>
> If it's fine I'd like to just toss in a patch for that as well.
>

Arnd pointed out it would be an incompatible ABI change[1].

However - I asked Khouloud who's working on v2 of the line event
interface to use ktime_get_ns().

Cheers
Bart

[1] https://marc.info/?l=linux-gpio&m=151661955709074&w=2
Linus Walleij Feb. 20, 2020, 3:03 p.m. UTC | #3
On Wed, Feb 12, 2020 at 12:00 PM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
> > On Tue, Feb 11, 2020 at 10:19 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
> > > From: Bartosz Golaszewski <bgolaszewski@baylibre.com>

> > A question:
> >
> > Bartosz, since you know about possible impacts on userspace,
> > since this code use the preferred ktime_get_ns() rather than
> > ktime_get_ns_real(), what happens if we just patch the other
> > event timestamp to use ktime_get_ns() instead, so we use the
> > same everywhere?
> >
> > If it's fine I'd like to just toss in a patch for that as well.
> >
>
> Arnd pointed out it would be an incompatible ABI change[1].

Yeah, I was thinking more about this specific answer from Arnd:

> "It is an incompatible ABI change, the question here is whether anyone
> actually cares. If nothing relies on the timestamps being in
> CLOCK_REALTIME domain, then it can be changed, the question
> is just how you want to prove that this is the case."

So the question is if userspace really cares.

What happens with libgpiod or users of it? Are they assuming
the weirdness of CLOCK_REALTIME, or are they simply assuming
something that is monotonic increasing and just lucky that they
didn't run into anything jumping backwards in time even though
they *could*.

I think I'll propose a change and see what people say.

> However - I asked Khouloud who's working on v2 of the line event
> interface to use ktime_get_ns().

That's great!

Yours,
Linus Walleij
Bartosz Golaszewski Feb. 20, 2020, 3:06 p.m. UTC | #4
czw., 20 lut 2020 o 16:03 Linus Walleij <linus.walleij@linaro.org> napisał(a):
>
> On Wed, Feb 12, 2020 at 12:00 PM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
> > > On Tue, Feb 11, 2020 at 10:19 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
> > > > From: Bartosz Golaszewski <bgolaszewski@baylibre.com>
>
> > > A question:
> > >
> > > Bartosz, since you know about possible impacts on userspace,
> > > since this code use the preferred ktime_get_ns() rather than
> > > ktime_get_ns_real(), what happens if we just patch the other
> > > event timestamp to use ktime_get_ns() instead, so we use the
> > > same everywhere?
> > >
> > > If it's fine I'd like to just toss in a patch for that as well.
> > >
> >
> > Arnd pointed out it would be an incompatible ABI change[1].
>
> Yeah, I was thinking more about this specific answer from Arnd:
>
> > "It is an incompatible ABI change, the question here is whether anyone
> > actually cares. If nothing relies on the timestamps being in
> > CLOCK_REALTIME domain, then it can be changed, the question
> > is just how you want to prove that this is the case."
>
> So the question is if userspace really cares.
>
> What happens with libgpiod or users of it? Are they assuming
> the weirdness of CLOCK_REALTIME, or are they simply assuming
> something that is monotonic increasing and just lucky that they
> didn't run into anything jumping backwards in time even though
> they *could*.
>
> I think I'll propose a change and see what people say.
>

Libgpiod doesn't care about the value really - it just forwards
whatever it reads.

Bart

> > However - I asked Khouloud who's working on v2 of the line event
> > interface to use ktime_get_ns().
>
> That's great!
>
> Yours,
> Linus Walleij
diff mbox series

Patch

diff --git a/drivers/gpio/gpiolib.c b/drivers/gpio/gpiolib.c
index 443321f9cf63..f73077f26eff 100644
--- a/drivers/gpio/gpiolib.c
+++ b/drivers/gpio/gpiolib.c
@@ -546,6 +546,9 @@  static long linehandle_set_config(struct linehandle_state *lh,
 			if (ret)
 				return ret;
 		}
+
+		atomic_notifier_call_chain(&desc->gdev->notifier,
+					   GPIOLINE_CHANGED_CONFIG, desc);
 	}
 	return 0;
 }
@@ -1201,14 +1204,25 @@  static void gpio_desc_to_lineinfo(struct gpio_desc *desc,
 	spin_unlock_irqrestore(&gpio_lock, flags);
 }
 
+struct gpio_chardev_data {
+	struct gpio_device *gdev;
+	wait_queue_head_t wait;
+	DECLARE_KFIFO(events, struct gpioline_info_changed, 32);
+	struct notifier_block lineinfo_changed_nb;
+	unsigned long *watched_lines;
+};
+
 /*
  * gpio_ioctl() - ioctl handler for the GPIO chardev
  */
 static long gpio_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 {
-	struct gpio_device *gdev = filp->private_data;
+	struct gpio_chardev_data *priv = filp->private_data;
+	struct gpio_device *gdev = priv->gdev;
 	struct gpio_chip *chip = gdev->chip;
 	void __user *ip = (void __user *)arg;
+	struct gpio_desc *desc;
+	__u32 offset;
 
 	/* We fail any subsequent ioctl():s when the chip is gone */
 	if (!chip)
@@ -1230,9 +1244,9 @@  static long gpio_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 		if (copy_to_user(ip, &chipinfo, sizeof(chipinfo)))
 			return -EFAULT;
 		return 0;
-	} else if (cmd == GPIO_GET_LINEINFO_IOCTL) {
+	} else if (cmd == GPIO_GET_LINEINFO_IOCTL ||
+		   cmd == GPIO_GET_LINEINFO_WATCH_IOCTL) {
 		struct gpioline_info lineinfo;
-		struct gpio_desc *desc;
 
 		if (copy_from_user(&lineinfo, ip, sizeof(lineinfo)))
 			return -EFAULT;
@@ -1245,11 +1259,25 @@  static long gpio_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
 
 		if (copy_to_user(ip, &lineinfo, sizeof(lineinfo)))
 			return -EFAULT;
+
+		if (cmd == GPIO_GET_LINEINFO_WATCH_IOCTL)
+			set_bit(desc_to_gpio(desc), priv->watched_lines);
+
 		return 0;
 	} else if (cmd == GPIO_GET_LINEHANDLE_IOCTL) {
 		return linehandle_create(gdev, ip);
 	} else if (cmd == GPIO_GET_LINEEVENT_IOCTL) {
 		return lineevent_create(gdev, ip);
+	} else if (cmd == GPIO_GET_LINEINFO_UNWATCH_IOCTL) {
+		if (copy_from_user(&offset, ip, sizeof(offset)))
+			return -EFAULT;
+
+		desc = gpiochip_get_desc(chip, offset);
+		if (IS_ERR(desc))
+			return PTR_ERR(desc);
+
+		clear_bit(desc_to_gpio(desc), &desc->flags);
+		return 0;
 	}
 	return -EINVAL;
 }
@@ -1262,6 +1290,101 @@  static long gpio_ioctl_compat(struct file *filp, unsigned int cmd,
 }
 #endif
 
+static struct gpio_chardev_data *
+to_gpio_chardev_data(struct notifier_block *nb)
+{
+	return container_of(nb, struct gpio_chardev_data, lineinfo_changed_nb);
+}
+
+static int lineinfo_changed_notify(struct notifier_block *nb,
+				   unsigned long action, void *data)
+{
+	struct gpio_chardev_data *priv = to_gpio_chardev_data(nb);
+	struct gpioline_info_changed chg;
+	struct gpio_desc *desc = data;
+	int ret;
+
+	if (!test_bit(desc_to_gpio(desc), priv->watched_lines))
+		return NOTIFY_DONE;
+
+	memset(&chg, 0, sizeof(chg));
+	chg.info.line_offset = gpio_chip_hwgpio(desc);
+	chg.event_type = action;
+	chg.timestamp = ktime_get_ns();
+	gpio_desc_to_lineinfo(desc, &chg.info);
+
+	ret = kfifo_in_spinlocked(&priv->events, &chg, 1, &priv->wait.lock);
+	if (ret)
+		wake_up_poll(&priv->wait, EPOLLIN);
+	else
+		pr_debug_ratelimited("lineinfo event FIFO is full - event dropped\n");
+
+	return NOTIFY_OK;
+}
+
+static __poll_t lineinfo_watch_poll(struct file *filep,
+				    struct poll_table_struct *pollt)
+{
+	struct gpio_chardev_data *priv = filep->private_data;
+	__poll_t events = 0;
+
+	poll_wait(filep, &priv->wait, pollt);
+
+	if (!kfifo_is_empty_spinlocked_noirqsave(&priv->events,
+						 &priv->wait.lock))
+		events = EPOLLIN | EPOLLRDNORM;
+
+	return events;
+}
+
+static ssize_t lineinfo_watch_read(struct file *filep, char __user *buf,
+				   size_t count, loff_t *off)
+{
+	struct gpio_chardev_data *priv = filep->private_data;
+	struct gpioline_info_changed event;
+	ssize_t bytes_read = 0;
+	int ret;
+
+	if (count < sizeof(event))
+		return -EINVAL;
+
+	do {
+		spin_lock(&priv->wait.lock);
+		if (kfifo_is_empty(&priv->events)) {
+			if (bytes_read) {
+				spin_unlock(&priv->wait.lock);
+				return bytes_read;
+			}
+
+			if (filep->f_flags & O_NONBLOCK) {
+				spin_unlock(&priv->wait.lock);
+				return -EAGAIN;
+			}
+
+			ret = wait_event_interruptible_locked(priv->wait,
+					!kfifo_is_empty(&priv->events));
+			if (ret) {
+				spin_unlock(&priv->wait.lock);
+				return ret;
+			}
+		}
+
+		ret = kfifo_out(&priv->events, &event, 1);
+		spin_unlock(&priv->wait.lock);
+		if (ret != 1) {
+			ret = -EIO;
+			break;
+			/* We should never get here. See lineevent_read(). */
+		}
+
+		if (copy_to_user(buf + bytes_read, &event, sizeof(event)))
+			return -EFAULT;
+		bytes_read += sizeof(event);
+	} while (count >= bytes_read + sizeof(event));
+
+	return bytes_read;
+}
+
 /**
  * gpio_chrdev_open() - open the chardev for ioctl operations
  * @inode: inode for this chardev
@@ -1272,14 +1395,48 @@  static int gpio_chrdev_open(struct inode *inode, struct file *filp)
 {
 	struct gpio_device *gdev = container_of(inode->i_cdev,
 					      struct gpio_device, chrdev);
+	struct gpio_chardev_data *priv;
+	int ret = -ENOMEM;
 
 	/* Fail on open if the backing gpiochip is gone */
 	if (!gdev->chip)
 		return -ENODEV;
+
+	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
+	if (!priv)
+		return -ENOMEM;
+
+	priv->watched_lines = bitmap_zalloc(gdev->chip->ngpio, GFP_KERNEL);
+	if (!priv->watched_lines)
+		goto out_free_priv;
+
+	init_waitqueue_head(&priv->wait);
+	INIT_KFIFO(priv->events);
+	priv->gdev = gdev;
+
+	priv->lineinfo_changed_nb.notifier_call = lineinfo_changed_notify;
+	ret = atomic_notifier_chain_register(&gdev->notifier,
+					     &priv->lineinfo_changed_nb);
+	if (ret)
+		goto out_free_bitmap;
+
 	get_device(&gdev->dev);
-	filp->private_data = gdev;
+	filp->private_data = priv;
 
-	return nonseekable_open(inode, filp);
+	ret = nonseekable_open(inode, filp);
+	if (ret)
+		goto out_unregister_notifier;
+
+	return ret;
+
+out_unregister_notifier:
+	atomic_notifier_chain_unregister(&gdev->notifier,
+					 &priv->lineinfo_changed_nb);
+out_free_bitmap:
+	bitmap_free(priv->watched_lines);
+out_free_priv:
+	kfree(priv);
+	return ret;
 }
 
 /**
@@ -1290,17 +1447,23 @@  static int gpio_chrdev_open(struct inode *inode, struct file *filp)
  */
 static int gpio_chrdev_release(struct inode *inode, struct file *filp)
 {
-	struct gpio_device *gdev = container_of(inode->i_cdev,
-					      struct gpio_device, chrdev);
+	struct gpio_chardev_data *priv = filp->private_data;
+	struct gpio_device *gdev = priv->gdev;
 
+	bitmap_free(priv->watched_lines);
+	atomic_notifier_chain_unregister(&gdev->notifier,
+					 &priv->lineinfo_changed_nb);
 	put_device(&gdev->dev);
+	kfree(priv);
+
 	return 0;
 }
 
-
 static const struct file_operations gpio_fileops = {
 	.release = gpio_chrdev_release,
 	.open = gpio_chrdev_open,
+	.poll = lineinfo_watch_poll,
+	.read = lineinfo_watch_read,
 	.owner = THIS_MODULE,
 	.llseek = no_llseek,
 	.unlocked_ioctl = gpio_ioctl,
@@ -1511,6 +1674,8 @@  int gpiochip_add_data_with_key(struct gpio_chip *chip, void *data,
 
 	spin_unlock_irqrestore(&gpio_lock, flags);
 
+	ATOMIC_INIT_NOTIFIER_HEAD(&gdev->notifier);
+
 #ifdef CONFIG_PINCTRL
 	INIT_LIST_HEAD(&gdev->pin_ranges);
 #endif
@@ -2843,6 +3008,8 @@  static int gpiod_request_commit(struct gpio_desc *desc, const char *label)
 	}
 done:
 	spin_unlock_irqrestore(&gpio_lock, flags);
+	atomic_notifier_call_chain(&desc->gdev->notifier,
+				   GPIOLINE_CHANGED_REQUESTED, desc);
 	return ret;
 }
 
@@ -2940,6 +3107,9 @@  static bool gpiod_free_commit(struct gpio_desc *desc)
 	}
 
 	spin_unlock_irqrestore(&gpio_lock, flags);
+	atomic_notifier_call_chain(&desc->gdev->notifier,
+				   GPIOLINE_CHANGED_RELEASED, desc);
+
 	return ret;
 }
 
diff --git a/drivers/gpio/gpiolib.h b/drivers/gpio/gpiolib.h
index 3e0aab2945d8..5ab90746b519 100644
--- a/drivers/gpio/gpiolib.h
+++ b/drivers/gpio/gpiolib.h
@@ -56,6 +56,7 @@  struct gpio_device {
 	const char		*label;
 	void			*data;
 	struct list_head        list;
+	struct atomic_notifier_head notifier;
 
 #ifdef CONFIG_PINCTRL
 	/*
diff --git a/include/uapi/linux/gpio.h b/include/uapi/linux/gpio.h
index 799cf823d493..dca320764e4d 100644
--- a/include/uapi/linux/gpio.h
+++ b/include/uapi/linux/gpio.h
@@ -59,6 +59,34 @@  struct gpioline_info {
 /* Maximum number of requested handles */
 #define GPIOHANDLES_MAX 64
 
+/* Possible line status change events */
+enum {
+	GPIOLINE_CHANGED_REQUESTED = 1,
+	GPIOLINE_CHANGED_RELEASED,
+	GPIOLINE_CHANGED_CONFIG,
+};
+
+/**
+ * struct gpioline_info_changed - Information about a change in status
+ * of a GPIO line
+ * @info: updated line information
+ * @timestamp: estimate of time of status change occurrence, in nanoseconds
+ * and GPIOLINE_CHANGED_CONFIG
+ * @event_type: one of GPIOLINE_CHANGED_REQUESTED, GPIOLINE_CHANGED_RELEASED
+ *
+ * Note: struct gpioline_info embedded here has 32-bit alignment on its own,
+ * but it works fine with 64-bit alignment too. With its 72 byte size, we can
+ * guarantee there are no implicit holes between it and subsequent members.
+ * The 20-byte padding at the end makes sure we don't add any implicit padding
+ * at the end of the structure on 64-bit architectures.
+ */
+struct gpioline_info_changed {
+	struct gpioline_info info;
+	__u64 timestamp;
+	__u32 event_type;
+	__u32 padding[5]; /* for future use */
+};
+
 /* Linerequest flags */
 #define GPIOHANDLE_REQUEST_INPUT	(1UL << 0)
 #define GPIOHANDLE_REQUEST_OUTPUT	(1UL << 1)
@@ -176,6 +204,8 @@  struct gpioevent_data {
 
 #define GPIO_GET_CHIPINFO_IOCTL _IOR(0xB4, 0x01, struct gpiochip_info)
 #define GPIO_GET_LINEINFO_IOCTL _IOWR(0xB4, 0x02, struct gpioline_info)
+#define GPIO_GET_LINEINFO_WATCH_IOCTL _IOWR(0xB4, 0x0b, struct gpioline_info)
+#define GPIO_GET_LINEINFO_UNWATCH_IOCTL _IOWR(0xB4, 0x0c, __u32)
 #define GPIO_GET_LINEHANDLE_IOCTL _IOWR(0xB4, 0x03, struct gpiohandle_request)
 #define GPIO_GET_LINEEVENT_IOCTL _IOWR(0xB4, 0x04, struct gpioevent_request)