Skip to contentSkip to conversation

ALSA: control: Take power_ref lock primarily

PATCH In review Takashi Iwai 2 participants 4 messages Jul 29, 2024 – Jul 30, 2024 Sound

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ALSA: control: Take power_ref lock primarily

The code path for kcontrol accesses have often nested locks of both card's controls_rwsem and power_ref, and applies in that order. However, what could take much longer is the latter, power_ref; it waits for the power state of the device, and it pretty much depends on the user's action.

This patch swaps the locking order of those locks to a more natural way, namely, power_ref -> controls_rwsem, in order to shorten the time of possible nested locks. For consistency, power_ref is taken always in the top-level caller side (that is, *_user() functions and the ioctl handler itself).

---
 sound/core/control.c | 54 ++++++++++++++++++++++++++++----------------
 1 file changed, 34 insertions(+), 20 deletions(-)
sound/core/control.c +34 -20
index f64a555f404f..96a2c088a460 100644
@@ -1167,9 +1167,7 @@ static int __snd_ctl_elem_info(struct snd_card *card,
 #ifdef CONFIG_SND_DEBUG
 	info->access = 0;
 #endif
-	result = snd_power_ref_and_wait(card);
-	if (!result)
-		result = kctl->info(kctl, info);
+	result = kctl->info(kctl, info);
 	snd_power_unref(card);
 	if (result >= 0) {
 		snd_BUG_ON(info->access);
@@ -1208,12 +1206,17 @@ static int snd_ctl_elem_info(struct snd_ctl_file *ctl,
 static int snd_ctl_elem_info_user(struct snd_ctl_file *ctl,
 				  struct snd_ctl_elem_info __user *_info)
 {
+	struct snd_card *card = ctl->card;
 	struct snd_ctl_elem_info info;
 	int result;
 
 	if (copy_from_user(&info, _info, sizeof(info)))
 		return -EFAULT;
+	result = snd_power_ref_and_wait(card);
+	if (result)
+		return result;
 	result = snd_ctl_elem_info(ctl, &info);
+	snd_power_unref(card);
 	if (result < 0)
 		return result;
 	/* drop internal access flags */
@@ -1257,10 +1260,7 @@ static int snd_ctl_elem_read(struct snd_card *card,
 
 	if (!snd_ctl_skip_validation(&info))
 		fill_remaining_elem_value(control, &info, pattern);
-	ret = snd_power_ref_and_wait(card);
-	if (!ret)
-		ret = kctl->get(kctl, control);
-	snd_power_unref(card);
+	ret = kctl->get(kctl, control);
 	if (ret < 0)
 		return ret;
 	if (!snd_ctl_skip_validation(&info) &&
@@ -1285,7 +1285,11 @@ static int snd_ctl_elem_read_user(struct snd_card *card,
 	if (IS_ERR(control))
 		return PTR_ERR(no_free_ptr(control));
 
+	result = snd_power_ref_and_wait(card);
+	if (result)
+		return result;
 	result = snd_ctl_elem_read(card, control);
+	snd_power_unref(card);
 	if (result < 0)
 		return result;
 
@@ -1300,7 +1304,7 @@ static int snd_ctl_elem_write(struct snd_card *card, struct snd_ctl_file *file,
 	struct snd_kcontrol *kctl;
 	struct snd_kcontrol_volatile *vd;
 	unsigned int index_offset;
-	int result;
+	int result = 0;
 
 	down_write(&card->controls_rwsem);
 	kctl = snd_ctl_find_id_locked(card, &control->id);
@@ -1318,9 +1322,8 @@ static int snd_ctl_elem_write(struct snd_card *card, struct snd_ctl_file *file,
 	}
 
 	snd_ctl_build_ioff(&control->id, kctl, index_offset);
-	result = snd_power_ref_and_wait(card);
 	/* validate input values */
-	if (IS_ENABLED(CONFIG_SND_CTL_INPUT_VALIDATION) && !result) {
+	if (IS_ENABLED(CONFIG_SND_CTL_INPUT_VALIDATION)) {
 		struct snd_ctl_elem_info info;
 
 		memset(&info, 0, sizeof(info));
@@ -1332,7 +1335,6 @@ static int snd_ctl_elem_write(struct snd_card *card, struct snd_ctl_file *file,
 	}
 	if (!result)
 		result = kctl->put(kctl, control);
-	snd_power_unref(card);
 	if (result < 0) {
 		up_write(&card->controls_rwsem);
 		return result;
@@ -1361,7 +1363,11 @@ static int snd_ctl_elem_write_user(struct snd_ctl_file *file,
 		return PTR_ERR(no_free_ptr(control));
 
 	card = file->card;
+	result = snd_power_ref_and_wait(card);
+	if (result < 0)
+		return result;
 	result = snd_ctl_elem_write(card, file, control);
+	snd_power_unref(card);
 	if (result < 0)
 		return result;
 
@@ -1830,7 +1836,7 @@ static int call_tlv_handler(struct snd_ctl_file *file, int op_flag,
 		{SNDRV_CTL_TLV_OP_CMD,   SNDRV_CTL_ELEM_ACCESS_TLV_COMMAND},
 	};
 	struct snd_kcontrol_volatile *vd = &kctl->vd[snd_ctl_get_ioff(kctl, id)];
-	int i, ret;
+	int i;
 
 	/* Check support of the request for this element. */
 	for (i = 0; i < ARRAY_SIZE(pairs); ++i) {
@@ -1848,11 +1854,7 @@ static int call_tlv_handler(struct snd_ctl_file *file, int op_flag,
 	    vd->owner != NULL && vd->owner != file)
 		return -EPERM;
 
-	ret = snd_power_ref_and_wait(file->card);
-	if (!ret)
-		ret = kctl->tlv.c(kctl, op_flag, size, buf);
-	snd_power_unref(file->card);
-	return ret;
+	return kctl->tlv.c(kctl, op_flag, size, buf);
 }
 
 static int read_tlv_buf(struct snd_kcontrol *kctl, struct snd_ctl_elem_id *id,
@@ -1965,16 +1967,28 @@ static long snd_ctl_ioctl(struct file *file, unsigned int cmd, unsigned long arg
 	case SNDRV_CTL_IOCTL_SUBSCRIBE_EVENTS:
 		return snd_ctl_subscribe_events(ctl, ip);
 	case SNDRV_CTL_IOCTL_TLV_READ:
-		scoped_guard(rwsem_read, &ctl->card->controls_rwsem)
+		err = snd_power_ref_and_wait(card);
+		if (err < 0)
+			return err;
+		scoped_guard(rwsem_read, &card->controls_rwsem)
 			err = snd_ctl_tlv_ioctl(ctl, argp, SNDRV_CTL_TLV_OP_READ);
+		snd_power_unref(card);
 		return err;
 	case SNDRV_CTL_IOCTL_TLV_WRITE:
-		scoped_guard(rwsem_write, &ctl->card->controls_rwsem)
+		err = snd_power_ref_and_wait(card);
+		if (err < 0)
+			return err;
+		scoped_guard(rwsem_write, &card->controls_rwsem)
 			err = snd_ctl_tlv_ioctl(ctl, argp, SNDRV_CTL_TLV_OP_WRITE);
+		snd_power_unref(card);
 		return err;
 	case SNDRV_CTL_IOCTL_TLV_COMMAND:
-		scoped_guard(rwsem_write, &ctl->card->controls_rwsem)
+		err = snd_power_ref_and_wait(card);
+		if (err < 0)
+			return err;
+		scoped_guard(rwsem_write, &card->controls_rwsem)
 			err = snd_ctl_tlv_ioctl(ctl, argp, SNDRV_CTL_TLV_OP_CMD);
+		snd_power_unref(card);
 		return err;
 	case SNDRV_CTL_IOCTL_POWER:
 		return -ENOPROTOOPT;
Signature · 1 line
2.43.0
maintainer
#1 lore raw
Quoted 6 lines · Takashi Iwai

The code path for kcontrol accesses have often nested locks of both card's controls_rwsem and power_ref, and applies in that order. However, what could take much longer is the latter, power_ref; it waits for the power state of the device, and it pretty much depends on the user's action.

Quoted 7 lines

This patch swaps the locking order of those locks to a more natural way, namely, power_ref -> controls_rwsem, in order to shorten the time of possible nested locks. For consistency, power_ref is taken always in the top-level caller side (that is, *_user() functions and the ioctl handler itself).

It looks fine, but perhaps, we can DEFINE_GUARD for snd_power_ref_and_wait and use this scoped guard, too? It would greatly simplify code.

Jaroslav

Signature · 2 lines
Jaroslav Kysela <[email protected]>
Linux Sound Maintainer; ALSA Project; Red Hat, Inc.
maintainer
#2 lore raw
In reply to Takashi Iwai
Quoted 6 lines · Takashi Iwai

The code path for kcontrol accesses have often nested locks of both card's controls_rwsem and power_ref, and applies in that order. However, what could take much longer is the latter, power_ref; it waits for the power state of the device, and it pretty much depends on the user's action.

Quoted 7 lines

This patch swaps the locking order of those locks to a more natural way, namely, power_ref -> controls_rwsem, in order to shorten the time of possible nested locks. For consistency, power_ref is taken always in the top-level caller side (that is, *_user() functions and the ioctl handler itself).

It looks fine, but perhaps, we can DEFINE_GUARD for snd_power_ref_and_wait and use this scoped guard, too? It would greatly simplify code.

Jaroslav

Signature · 2 lines
Jaroslav Kysela <[email protected]>
Linux Sound Maintainer; ALSA Project; Red Hat, Inc.
#3 lore raw
Quoted 14 lines · Jaroslav Kysela
Quoted 12 lines · Takashi Iwai

The code path for kcontrol accesses have often nested locks of both card's controls_rwsem and power_ref, and applies in that order. However, what could take much longer is the latter, power_ref; it waits for the power state of the device, and it pretty much depends on the user's action.

This patch swaps the locking order of those locks to a more natural way, namely, power_ref -> controls_rwsem, in order to shorten the time of possible nested locks. For consistency, power_ref is taken always in the top-level caller side (that is, *_user() functions and the ioctl handler itself).

Quoted 5 lines
Quoted 1 line

It looks fine, but perhaps, we can DEFINE_GUARD for snd_power_ref_and_wait and use this scoped guard, too? It would greatly simplify code.

I thought of that, but it's a bit more complicated because it's a version with *_try (so you have to define more). I'd replace as a refactoring later.

thanks,

Takashi