Skip to contentSkip to conversation

gpiolib: extend the critical sections of lookup tables

PATCH In review Bartosz Golaszewski 2 participants 8 messages Sep 20, 2023 – Sep 28, 2023 GPIO

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 gpiolib: extend the critical sections of lookup tables

From: Bartosz Golaszewski <[email protected]>

There are two places in the code where we retrieve a lookup table using gpiod_find_lookup_table() (which protects the table list with the lookup table lock) and then use it after the lock is released.

We need to keep the lookup table mutex locked the entire time we're using the tables. Remove the locking from gpiod_find_lookup_table() and use guards to protect the code actually using the table objects.

---
 drivers/gpio/gpiolib.c | 30 +++++++++++++++---------------
 1 file changed, 15 insertions(+), 15 deletions(-)
drivers/gpio/gpiolib.c +15 -15
index edffa0d2acaa..7c27a1efc1b0 100644
@@ -3822,8 +3822,6 @@ static struct gpiod_lookup_table *gpiod_find_lookup_table(struct device *dev)
 	const char *dev_id = dev ? dev_name(dev) : NULL;
 	struct gpiod_lookup_table *table;
 
-	mutex_lock(&gpio_lookup_lock);
-
 	list_for_each_entry(table, &gpio_lookup_list, list) {
 		if (table->dev_id && dev_id) {
 			/*
@@ -3831,21 +3829,18 @@ static struct gpiod_lookup_table *gpiod_find_lookup_table(struct device *dev)
 			 * a match
 			 */
 			if (!strcmp(table->dev_id, dev_id))
-				goto found;
+				return table;
 		} else {
 			/*
 			 * One of the pointers is NULL, so both must be to have
 			 * a match
 			 */
 			if (dev_id == table->dev_id)
-				goto found;
+				return table;
 		}
 	}
-	table = NULL;
 
-found:
-	mutex_unlock(&gpio_lookup_lock);
-	return table;
+	return NULL;
 }
 
 static struct gpio_desc *gpiod_find(struct device *dev, const char *con_id,
@@ -3855,6 +3850,8 @@ static struct gpio_desc *gpiod_find(struct device *dev, const char *con_id,
 	struct gpiod_lookup_table *table;
 	struct gpiod_lookup *p;
 
+	guard(mutex)(&gpio_lookup_lock);
+
 	table = gpiod_find_lookup_table(dev);
 	if (!table)
 		return desc;
@@ -3920,15 +3917,18 @@ static int platform_gpio_count(struct device *dev, const char *con_id)
 	struct gpiod_lookup *p;
 	unsigned int count = 0;
 
-	table = gpiod_find_lookup_table(dev);
-	if (!table)
-		return -ENOENT;
+	scoped_guard(mutex, &gpio_lookup_lock) {
+		table = gpiod_find_lookup_table(dev);
+		if (!table)
+			return -ENOENT;
 
-	for (p = &table->table[0]; p->key; p++) {
-		if ((con_id && p->con_id && !strcmp(con_id, p->con_id)) ||
-		    (!con_id && !p->con_id))
-			count++;
+		for (p = &table->table[0]; p->key; p++) {
+			if ((con_id && p->con_id && !strcmp(con_id, p->con_id)) ||
+			    (!con_id && !p->con_id))
+				count++;
+		}
 	}
+
 	if (!count)
 		return -ENOENT;
 
Signature · 1 line
2.39.2
Reviewed-by on v1 1/1
#1 lore raw
Quoted 11 lines · Bartosz Golaszewski

From: Bartosz Golaszewski <[email protected]>

There are two places in the code where we retrieve a lookup table using gpiod_find_lookup_table() (which protects the table list with the lookup table lock) and then use it after the lock is released.

We need to keep the lookup table mutex locked the entire time we're using the tables. Remove the locking from gpiod_find_lookup_table() and use guards to protect the code actually using the table objects.

I convinced myself that this change is correct, good find!

Review static struct gpiod_lookup_table *gpiod_find_lookup_table(struct device *dev)

Quoted 1 line
(quoted diff) +0 -0
@@ -3822,8 +3822,6 @@ static struct gpiod_lookup_table *gpiod_find_lookup_table(struct device *dev)

Can we rename this function gpiod_find_lookup_table_locked() as per precedents in the kernel, to indicate that it needs to be called with a lock held?

Either way:

Yours,
Linus Walleij

#2 lore raw
Quoted 14 lines · Linus Walleij
Quoted 11 lines · Bartosz Golaszewski

From: Bartosz Golaszewski <[email protected]>

There are two places in the code where we retrieve a lookup table using gpiod_find_lookup_table() (which protects the table list with the lookup table lock) and then use it after the lock is released.

We need to keep the lookup table mutex locked the entire time we're using the tables. Remove the locking from gpiod_find_lookup_table() and use guards to protect the code actually using the table objects.

Quoted 8 lines

I convinced myself that this change is correct, good find!

Review static struct gpiod_lookup_table *gpiod_find_lookup_table(struct device *dev)

Quoted 1 line
(quoted diff) +0 -0
@@ -3822,8 +3822,6 @@ static struct gpiod_lookup_table *gpiod_find_lookup_table(struct device *dev)

Can we rename this function gpiod_find_lookup_table_locked() as per precedents in the kernel, to indicate that it needs to be called with a lock held?

I think you mean gpiod_find_lookup_table_unlocked() as with this change it will no longer take the lock?

Bart

Quoted 6 lines

Either way:

Yours,
Linus Walleij

#3 lore raw
Quoted 11 lines · Bartosz Golaszewski
Quoted 6 lines · Linus Walleij

On Wed, Sep 20, 2023 at 10:56 AM Bartosz Golaszewski <[email protected]> wrote:

Can we rename this function gpiod_find_lookup_table_locked() as per precedents in the kernel, to indicate that it needs to be called with a lock held?

I think you mean gpiod_find_lookup_table_unlocked() as with this change it will no longer take the lock?

I think the pattern is the one I indicated: *_locked() means the function is to be called with the appropriate lock held, cf arch/arm64/kvm/hyp/nvhe/mm.c

pkvm_create_mappings() takes a lock and then calls pkvm_create_mappings_locked() which even asserts that the lock is held.

Yours,
Linus Walleij

#4 lore raw
Quoted 13 lines · Linus Walleij
Quoted 11 lines · Bartosz Golaszewski
Quoted 6 lines · Linus Walleij

On Wed, Sep 20, 2023 at 10:56 AM Bartosz Golaszewski <[email protected]> wrote:

Can we rename this function gpiod_find_lookup_table_locked() as per precedents in the kernel, to indicate that it needs to be called with a lock held?

I think you mean gpiod_find_lookup_table_unlocked() as with this change it will no longer take the lock?

Quoted 8 lines

I think the pattern is the one I indicated: *_locked() means the function is to be called with the appropriate lock held, cf arch/arm64/kvm/hyp/nvhe/mm.c

pkvm_create_mappings() takes a lock and then calls pkvm_create_mappings_locked() which even asserts that the lock is held.

Ha! I always though the pattern is to call the functions that *DON'T* take the lock _unlocked(). This is what I used in gpiolib-cdev.c or gpio-sim.c.

I guess both conventions make sense in some way.

Bart

#5 lore raw
Quoted 24 lines · Bartosz Golaszewski
Quoted 13 lines · Linus Walleij
Quoted 11 lines · Bartosz Golaszewski
Quoted 6 lines · Linus Walleij

On Wed, Sep 20, 2023 at 10:56 AM Bartosz Golaszewski <[email protected]> wrote:

Can we rename this function gpiod_find_lookup_table_locked() as per precedents in the kernel, to indicate that it needs to be called with a lock held?

I think you mean gpiod_find_lookup_table_unlocked() as with this change it will no longer take the lock?

Quoted 8 lines

I think the pattern is the one I indicated: *_locked() means the function is to be called with the appropriate lock held, cf arch/arm64/kvm/hyp/nvhe/mm.c

pkvm_create_mappings() takes a lock and then calls pkvm_create_mappings_locked() which even asserts that the lock is held.

Quoted 6 lines

Ha! I always though the pattern is to call the functions that *DON'T* take the lock _unlocked(). This is what I used in gpiolib-cdev.c or gpio-sim.c.

I guess both conventions make sense in some way.

Bart

I don't think I will be doing it just now. We don't use this convention elsewhere in drivers/gpio/ and we'll have a lot of locking reworked soon anyway. We may get to it when that is done.

Bart

#6 lore raw
Quoted 31 lines · Bartosz Golaszewski
Quoted 29 lines · Bartosz Golaszewski
Quoted 13 lines · Linus Walleij
Quoted 11 lines · Bartosz Golaszewski
Quoted 6 lines · Linus Walleij

On Wed, Sep 20, 2023 at 10:56 AM Bartosz Golaszewski <[email protected]> wrote:

Can we rename this function gpiod_find_lookup_table_locked() as per precedents in the kernel, to indicate that it needs to be called with a lock held?

I think you mean gpiod_find_lookup_table_unlocked() as with this change it will no longer take the lock?

Quoted 8 lines

I think the pattern is the one I indicated: *_locked() means the function is to be called with the appropriate lock held, cf arch/arm64/kvm/hyp/nvhe/mm.c

pkvm_create_mappings() takes a lock and then calls pkvm_create_mappings_locked() which even asserts that the lock is held.

Ha! I always though the pattern is to call the functions that *DON'T* take the lock _unlocked(). This is what I used in gpiolib-cdev.c or gpio-sim.c.

I guess both conventions make sense in some way.

Quoted 7 lines
Quoted 1 line

Bart

I don't think I will be doing it just now. We don't use this convention elsewhere in drivers/gpio/ and we'll have a lot of locking reworked soon anyway. We may get to it when that is done.

Bart

If there are no objections, I'd like to queue it this week.

Bart

#7 lore raw
Quoted 1 line · Bartosz Golaszewski

If there are no objections, I'd like to queue it this week.

Go ahead.

Kernel looks better after this patch than before => queue it

Yours,
Linus Walleij