[PATCH v12 00/12] ALSA: virtio: staged capture recovery and PCM RESET synchronization

[PATCH v12 00/12] ALSA: virtio: staged capture recovery and PCM RESET synchronization

[Joonho Lee]

Hi Takashi,

This revision replaces v11 and carries the unmerged prerequisites.
Apply patches 1-12 to sound.git for-next at
a4311e9447bbb219044db38897f95be112f46a8a.
Targeting 7.5 or later still works for me.

Previous revision:
https://lore.kernel.org/linux-sound/20261031020855.2402-1-mynameisjune111@gmail.com/T/

The series adds staged capture for shared clients such as dsnoop,
lifecycle recovery and synchronized PCM RESET accounting. Capture
compatibility options remain opt-in and disabled by default.

Changes since v11

The ACK path tried buffer_mutex while holding the PCM lock. That mutex
can already be held during normal RESET, refinement and linked prepare.
ACK then returned -EBADFD despite a valid stream. Private accounting can
also change before the indirect helper sees the ACK error.

Patch 10 uses the PCM lock for ACK and prepare accounting, with a
preparing gate lasting until the RESET callback completes. The core
keeps that same lock through its public pointer rebase. Playback RESET
rebases private accounting while preserving queued TX ownership. Both
capture modes synchronize their private pointers with the common reset,
including RESET from PREPARED. Queue-to-substream lock order is retained.

Paused native capture accepts RX completions from the live prepared
epoch even when transfers are disabled. Refill remains gated by transfer
state. Resume preserves the notification floor, so real missed periods
retain their PCM timer ticks and synthetic progress adds none. Device
events and failed STOP latch XRUN and wake readers. ACK rejects invalid
or stopped streams. Failed PREPARE preserves the previous accounting.

Ordinary RX validates status and payload before advancing counters.
Empty STOP responses contribute zero frames. Full-ring XRUN detection
remains enabled: sample checks showed a real overwritten ring after a
220 ms observation delay. A candidate that suppressed the error returned
frame 31680 where frame 24000 was expected. That candidate was discarded.
The final code reports -EPIPE and recovers through prepare. Normal drain
returns the expected 64-frame tail. The full-ring check in v11 remains
valid; the corrected issue is frame accounting for empty STOP responses.

Byte alignment uses gcd(frame_bits, 8), preserving byte-aligned queue
boundaries for packed 4-bit formats. Broader format runtime coverage is
still pending.

Several findings reproduce before v11, including ACK mutex contention
in the earlier lifecycle implementation and playback RESET accounting.
The ACK mutex findings share one synchronization cause. The staged
pause-tail and missed-period failures reproduce in both v10 and v11.
The local comparison retains original, earlier and final results.

Patch layout and dependencies

  1-6: ACK errors, START/RELEASE recovery and request/DMA lifetime.
  7:   Core stop-threshold accesses used by capture policy sampling.
  8:   Partial-progress notification under the PCM stream lock.
  9:   Optional PCM reset operation under the stream lock.
  10:  Staged capture, lifecycle serialization and driver RESET users.
  11:  Staged capture work priority.
  12:  Capture compatibility option documentation.

Patches 1-8 are unchanged from v11. The previous core patch 12 moves to
patch 9 without a code change. Previous driver patches 9, 13 and 14 are
combined in patch 10 with the corrections above. ACK's preparing gate
and its RESET completion must arrive together; separating them would
leave an intermediate driver with no matching completion path.
Previous patches 10 and 11 become patches 11 and 12.

The core API stays separate from its driver users. Drivers without a
reset callback retain the ioctl path. Linked RESET and prepare stop at
the first error and can leave earlier members completed. Applications
must recover the group after partial failure. A zero callback result
allows the core rebase; later I/O can report XRUN. The internal
snd_pcm_ops change requires out-of-tree users to be rebuilt.

Validation

The final source matches the corrected candidate byte for byte across
all 15 touched files. The comparison includes 43 ARM64 QEMU runs, with
56 transition checks across ordinary/staged capture and KASAN/KCSAN,
and four linked prepare/ACK checks. Faults include ACK contention,
failed PREPARE/STOP, paused RX, status-only RX and invalid ACK states.
KASAN, lockdep and KCSAN_STRICT reported no diagnostics in those runs.

Uninstrumented RW and MMAP captures ran for 70 seconds per mode and
capture policy, with pause, RESET and reopen. All 14 segments,
14,208,000 mono S16 frames at 48 kHz, matched independent sequence
hashes. Native-pause tail checks and PCM timer tests compared v11 with
the correction. Core compatibility tests include 72 callback/legacy,
atomic/nonatomic and linked-failure checks, plus snd-dummy playback and
capture. Test hooks are excluded from the submitted source.

Each of the 12 reconstructed intermediate sources is checked by applying
its mail patch to the pinned baseline and comparing the postimage.
ARM64 W=1 Image/modules builds and additional intermediate-stage runtime
checks are recorded in the local submission package. Fifteen additional
VM runs passed, including 28 injected transition checks at patch 10.

These are ARM64 QEMU TCG tests with two vCPUs and 512 MiB. Physical
VZ/MOTU acceptance for this revision and broader format, multichannel
and architecture runtime coverage remain pending. Earlier high-load
TCG XRUN results limit performance conclusions; this revision does not
claim a general CPU or scheduling improvement. The optional core APIs
and linked RESET partial-failure policy still need maintainer review.

Coding assistance was used for analysis, implementation, fixtures and
this explanation. The patches carry Assisted-by: LLM.

Thanks,
Joonho

base-commit: a4311e9447bbb219044db38897f95be112f46a8a

[PATCH v12 01/12] ALSA: virtio: Propagate PCM send failures through ACK

[Joonho Lee]

virtsnd_pcm_msg_send() can fail while queueing a transfer. Restore the
failed message's length and notify the device about earlier messages
queued by the same call, so they do not remain stranded.

Propagate that failure through the indirect PCM transfer callback instead
of discarding it. Commit the updated indirect state only after successful
submission and stop further transfers when a send fails. Return an XRUN
error to ALSA so the stream can recover.

Keep the polling feature path unchanged.

Fixes: fe981e67568c ("ALSA: virtio: use ack callback")
Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_pcm_msg.c b/sound/virtio/virtio_pcm_msg.c
--- a/sound/virtio/virtio_pcm_msg.c
+++ b/sound/virtio/virtio_pcm_msg.c
@@ -260,6 +260,18 @@
 				dev_err(&vdev->dev,
 					"SID %u: failed to send I/O message\n",
 					vss->sid);
+				msg->length -= n;
+
+				/*
+				 * Do not leave earlier messages queued without notifying
+				 * the device when a later message could not be added.
+				 */
+				if (msg_count != vss->msg_count &&
+				    !(vss->features &
+				      (1U << VIRTIO_SND_PCM_F_MSG_POLLING)) &&
+				    virtqueue_kick_prepare(vqueue))
+					virtqueue_notify(vqueue);
+
 				return rc;
 			}
 
diff --git a/sound/virtio/virtio_pcm_ops.c b/sound/virtio/virtio_pcm_ops.c
--- a/sound/virtio/virtio_pcm_ops.c
+++ b/sound/virtio/virtio_pcm_ops.c
@@ -466,36 +466,100 @@
 		vss->hw_ptr);
 }
 
+struct virtsnd_pcm_transfer {
+	struct snd_pcm_indirect indirect;
+	int error;
+};
+
 static void virtsnd_pcm_trans_copy(struct snd_pcm_substream *substream,
 				   struct snd_pcm_indirect *rec, size_t bytes)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
-
-	virtsnd_pcm_msg_send(vss, rec->sw_data, bytes);
+	struct virtsnd_pcm_transfer *transfer =
+		container_of(rec, struct virtsnd_pcm_transfer, indirect);
+
+	if (transfer->error)
+		return;
+
+	transfer->error = virtsnd_pcm_msg_send(vss, rec->sw_data, bytes);
+}
+
+/* Caller holds vss->lock. */
+static void virtsnd_pcm_transfer_failed(struct virtio_pcm_substream *vss)
+{
+	vss->xfer_enabled = false;
+	vss->stopped = true;
+	vss->substream->runtime->stop_operating = true;
 }
 
 static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
-
-	guard(spinlock_irqsave)(&queue->lock);
-	guard(spinlock)(&vss->lock);
-
-	return snd_pcm_indirect_playback_transfer(substream, &vss->pcm_indirect,
-						  virtsnd_pcm_trans_copy);
+	struct virtsnd_pcm_transfer transfer = {};
+	int rc;
+
+	scoped_guard(spinlock_irqsave, &queue->lock) {
+		guard(spinlock)(&vss->lock);
+		transfer.indirect = vss->pcm_indirect;
+		if (vss->stopped) {
+			rc = -EPIPE;
+		} else {
+			rc = snd_pcm_indirect_playback_transfer(substream,
+								&transfer.indirect,
+								virtsnd_pcm_trans_copy);
+		}
+		if (!rc)
+			rc = transfer.error;
+		if (rc)
+			virtsnd_pcm_transfer_failed(vss);
+		else
+			vss->pcm_indirect = transfer.indirect;
+	}
+	if (rc) {
+		dev_err(&vss->snd->vdev->dev,
+			"SID %u: playback ACK transfer failed: %d\n",
+			vss->sid, rc);
+		/* ALSA treats -EPIPE as an XRUN in ACK and pointer paths. */
+		return -EPIPE;
+	}
+
+	return 0;
 }
 
 static int virtsnd_pcm_cp_ack(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
-
-	guard(spinlock_irqsave)(&queue->lock);
-	guard(spinlock)(&vss->lock);
-
-	return snd_pcm_indirect_capture_transfer(substream, &vss->pcm_indirect,
-						 virtsnd_pcm_trans_copy);
+	struct virtsnd_pcm_transfer transfer = {};
+	int rc;
+
+	scoped_guard(spinlock_irqsave, &queue->lock) {
+		guard(spinlock)(&vss->lock);
+		transfer.indirect = vss->pcm_indirect;
+		if (vss->stopped) {
+			rc = -EPIPE;
+		} else {
+			rc = snd_pcm_indirect_capture_transfer(substream,
+							       &transfer.indirect,
+							       virtsnd_pcm_trans_copy);
+		}
+		if (!rc)
+			rc = transfer.error;
+		if (rc)
+			virtsnd_pcm_transfer_failed(vss);
+		else
+			vss->pcm_indirect = transfer.indirect;
+	}
+	if (rc) {
+		dev_err(&vss->snd->vdev->dev,
+			"SID %u: capture ACK transfer failed: %d\n",
+			vss->sid, rc);
+		/* ALSA treats -EPIPE as an XRUN in ACK and pointer paths. */
+		return -EPIPE;
+	}
+
+	return 0;
 }
 
 /* PCM substream operators map. */

[PATCH v12 02/12] ALSA: virtio: Recover PCM state after a failed start

[Joonho Lee]

Capture trigger queues I/O messages before PCM_START. If the START
message cannot be allocated or sent, transfers may remain queued, and
ALSA does not invoke the action undo callback for the failed trigger.

Mark the stream stopped and disable transfers for every failed START
path. Arm runtime->stop_operating so sync_stop releases pending I/O and
prevents prepare, hw_params, or START retries until release succeeds.

The stopped flag is shared with ACK callbacks that can run without the
ALSA stream lock. Protect every stopped access with vss->lock. Read the
pending message count under the same lock in open and hw_params. In
sync_stop, decide whether RELEASE is needed and mark the stream stopped
in one locked transition; leave it stopped when release or draining fails.

Always cancel elapsed-period work in sync_stop, even when no RELEASE
is needed.

Record the RELEASE obligation after successful PREPARE and clear it as
soon as RELEASE succeeds. If the subsequent RX wait fails, retry that
wait without submitting RELEASE again.

Fixes: da76e9f3e43a ("ALSA: virtio: PCM substream operators")
Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_pcm.h b/sound/virtio/virtio_pcm.h
--- a/sound/virtio/virtio_pcm.h
+++ b/sound/virtio/virtio_pcm.h
@@ -26,13 +26,13 @@
  * @hw: Kernel ALSA substream hardware descriptor.
  * @elapsed_period: Kernel work to handle the elapsed period state.
  * @lock: Spinlock that protects fields shared by interrupt handlers and
- *        substream operators.
+ *        substream operators, including @stopped.
  * @buffer_bytes: Current buffer size in bytes.
  * @hw_ptr: Substream hardware pointer value in bytes [0 ... buffer_bytes).
  * @xfer_enabled: Data transfer state (0 - off, 1 - on).
  * @xfer_xrun: Data underflow/overflow state (0 - no xrun, 1 - xrun).
  * @stopped: True if the substream is stopped and must be released on the device
- *           side.
+ *           side; protected by @lock.
  * @suspended: True if the substream is suspended and must be reconfigured on
  *             the device side at resume.
  * @msgs: Allocated I/O messages.
@@ -56,6 +56,8 @@
 	bool xfer_enabled;
 	bool xfer_xrun;
 	bool stopped;
+	/* Successful PREPARE still needs RELEASE; protected by lock. */
+	bool release_required;
 	bool suspended;
 	struct virtio_pcm_msg **msgs;
 	unsigned int nmsgs;
diff --git a/sound/virtio/virtio_pcm_ops.c b/sound/virtio/virtio_pcm_ops.c
--- a/sound/virtio/virtio_pcm_ops.c
+++ b/sound/virtio/virtio_pcm_ops.c
@@ -96,6 +96,15 @@
 
 static int virtsnd_pcm_sync_stop(struct snd_pcm_substream *substream);
 
+static void virtsnd_pcm_start_failed(struct virtio_pcm_substream *vss)
+{
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		vss->xfer_enabled = false;
+		vss->stopped = true;
+	}
+	vss->substream->runtime->stop_operating = true;
+}
+
 /**
  * virtsnd_pcm_open() - Open the PCM substream.
  * @substream: Kernel ALSA substream.
@@ -115,7 +124,8 @@
 	snd_pcm_hw_constraint_integer(substream->runtime,
 				      SNDRV_PCM_HW_PARAM_PERIODS);
 
-	vss->stopped = !!virtsnd_pcm_msg_pending_num(vss);
+	scoped_guard(spinlock_irqsave, &vss->lock)
+		vss->stopped |= !!vss->msg_count;
 	vss->suspended = false;
 
 	/*
@@ -217,11 +227,23 @@
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_device *vdev = vss->snd->vdev;
+	unsigned int msg_count;
+	bool stopped;
 	int rc;
 
-	if (virtsnd_pcm_msg_pending_num(vss)) {
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		msg_count = vss->msg_count;
+		stopped = vss->stopped;
+	}
+	if (msg_count) {
 		dev_err(&vdev->dev, "SID %u: invalid I/O queue state\n",
 			vss->sid);
+		return -EBADFD;
+	}
+	if (stopped) {
+		dev_err(&vdev->dev, "SID %u: stream release failed\n",
+			vss->sid);
+		substream->runtime->stop_operating = true;
 		return -EBADFD;
 	}
 
@@ -273,6 +295,17 @@
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_device *vdev = vss->snd->vdev;
 	struct virtio_snd_msg *msg;
+	bool stopped;
+	int rc;
+
+	scoped_guard(spinlock_irqsave, &vss->lock)
+		stopped = vss->stopped;
+	if (stopped) {
+		dev_err(&vdev->dev, "SID %u: stream release failed\n",
+			vss->sid);
+		substream->runtime->stop_operating = true;
+		return -EBADFD;
+	}
 
 	if (!vss->suspended) {
 		if (virtsnd_pcm_msg_pending_num(vss)) {
@@ -311,7 +344,12 @@
 	if (!msg)
 		return -ENOMEM;
 
-	return virtsnd_ctl_msg_send_sync(vss->snd, msg);
+	rc = virtsnd_ctl_msg_send_sync(vss->snd, msg);
+	if (!rc) {
+		guard(spinlock_irqsave)(&vss->lock);
+		vss->release_required = true;
+	}
+	return rc;
 }
 
 /**
@@ -338,31 +376,40 @@
 
 		scoped_guard(spinlock_irqsave, &queue->lock) {
 			guard(spinlock)(&vss->lock);
+			if (vss->stopped) {
+				substream->runtime->stop_operating = true;
+				return -EBADFD;
+			}
 			if (vss->direction == SNDRV_PCM_STREAM_CAPTURE)
 				rc = virtsnd_pcm_msg_send(vss, 0, vss->buffer_bytes);
-			if (rc)
+			if (rc) {
+				vss->stopped = true;
+				substream->runtime->stop_operating = true;
 				return rc;
+			}
 			vss->xfer_enabled = true;
 		}
 
 		msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_START,
 						GFP_KERNEL);
 		if (!msg) {
-			guard(spinlock_irqsave)(&vss->lock);
-			vss->xfer_enabled = false;
-
+			virtsnd_pcm_start_failed(vss);
 			return -ENOMEM;
 		}
 
-		return virtsnd_ctl_msg_send_sync(snd, msg);
+		rc = virtsnd_ctl_msg_send_sync(snd, msg);
+		if (rc)
+			virtsnd_pcm_start_failed(vss);
+
+		return rc;
 	case SNDRV_PCM_TRIGGER_SUSPEND:
 		vss->suspended = true;
 		fallthrough;
 	case SNDRV_PCM_TRIGGER_STOP:
-		vss->stopped = true;
-		fallthrough;
 	case SNDRV_PCM_TRIGGER_PAUSE_PUSH:
 		scoped_guard(spinlock_irqsave, &vss->lock) {
+			if (command != SNDRV_PCM_TRIGGER_PAUSE_PUSH)
+				vss->stopped = true;
 			vss->xfer_enabled = false;
 		}
 
@@ -393,21 +440,34 @@
 	struct virtio_snd *snd = vss->snd;
 	struct virtio_snd_msg *msg;
 	unsigned int js = msecs_to_jiffies(virtsnd_msg_timeout_ms);
+	bool need_release;
 	int rc;
 
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		vss->xfer_enabled = false;
+		need_release = vss->release_required;
+		if (need_release)
+			vss->stopped = true;
+	}
 	cancel_work_sync(&vss->elapsed_period);
 
-	if (!vss->stopped)
-		return 0;
-
-	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_RELEASE,
-					GFP_KERNEL);
-	if (!msg)
-		return -ENOMEM;
-
-	rc = virtsnd_ctl_msg_send_sync(snd, msg);
-	if (rc)
-		return rc;
+	if (need_release) {
+		msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_RELEASE,
+						GFP_KERNEL);
+		if (!msg) {
+			substream->runtime->stop_operating = true;
+			return -ENOMEM;
+		}
+
+		rc = virtsnd_ctl_msg_send_sync(snd, msg);
+		if (rc) {
+			substream->runtime->stop_operating = true;
+			return rc;
+		}
+
+		scoped_guard(spinlock_irqsave, &vss->lock)
+			vss->release_required = false;
+	}
 
 	/*
 	 * The spec states that upon receipt of the RELEASE command "the device
@@ -421,11 +481,13 @@
 	if (rc <= 0) {
 		dev_warn(&snd->vdev->dev, "SID %u: failed to flush I/O queue\n",
 			 vss->sid);
+		substream->runtime->stop_operating = true;
 
 		return !rc ? -ETIMEDOUT : rc;
 	}
 
-	vss->stopped = false;
+	scoped_guard(spinlock_irqsave, &vss->lock)
+		vss->stopped = false;
 
 	return 0;
 }

[PATCH v12 03/12] ALSA: virtio: Retain queued messages through device reset

[Joonho Lee]

Keep the control queue available through snd_card_free() so close-time
STOP and RELEASE commands can still be submitted. Block new control
requests and mark pending messages canceled before reset, then release
their queue references only after reset and queue deletion.

For PCM queues, detach each remaining virtqueue buffer after the device
has stopped DMA and retire its message without reporting a normal
completion. Detach all queues before deleting them so outstanding DMA
mappings are unmapped and no token remains for a late device completion.

Clear each queue wrapper inside the same critical section that detaches
its buffers. Check the callback's queue identity under the wrapper lock
for PCM, control and event queues. This makes queue retirement explicit
through teardown and restore. A forced callback test passed on both
versions; the proposed detached-ring IRQ failure was not reproduced on
the tested PCI transport.

Fixes: 9d45e514da88 ("ALSA: virtio: handling control messages")
Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_card.c b/sound/virtio/virtio_card.c
--- a/sound/virtio/virtio_card.c
+++ b/sound/virtio/virtio_card.c
@@ -87,6 +87,8 @@
 	u32 length;
 
 	guard(spinlock_irqsave)(&queue->lock);
+	if (queue->vqueue != vqueue)
+		return;
 	do {
 		virtqueue_disable_cb(vqueue);
 		while ((event = virtqueue_get_buf(vqueue, &length))) {
@@ -338,6 +340,38 @@
 }
 
 /**
+ * virtsnd_reset_vqs() - Stop device access and unmap outstanding buffers.
+ * @snd: VirtIO sound device.
+ *
+ * Context: Any context that permits to sleep.
+ */
+static void virtsnd_reset_vqs(struct virtio_snd *snd)
+{
+	struct virtio_device *vdev = snd->vdev;
+	unsigned int i;
+
+	virtio_reset_device(vdev);
+
+	/* Deleting a virtqueue does not unmap its outstanding descriptors. */
+	for (i = 0; i < VIRTIO_SND_VQ_MAX; ++i) {
+		struct virtio_snd_queue *queue = &snd->queues[i];
+		unsigned long flags;
+		void *msg;
+
+		spin_lock_irqsave(&queue->lock, flags);
+		if (queue->vqueue)
+			while ((msg = virtqueue_detach_unused_buf(queue->vqueue)))
+				if (i == VIRTIO_SND_VQ_TX || i == VIRTIO_SND_VQ_RX)
+					virtsnd_pcm_msg_cancel(msg);
+		/* IRQ callbacks must never see a detached ring as live. */
+		queue->vqueue = NULL;
+		spin_unlock_irqrestore(&queue->lock, flags);
+	}
+
+	vdev->config->del_vqs(vdev);
+}
+
+/**
  * virtsnd_remove() - Remove VirtIO and ALSA devices.
  * @vdev: VirtIO parent device.
  *
@@ -349,13 +383,14 @@
 	unsigned int i;
 
 	virtsnd_disable_event_vq(snd);
-	virtsnd_ctl_msg_cancel_all(snd);
 
 	if (snd->card)
 		snd_card_free(snd->card);
 
-	virtio_reset_device(vdev);
-	vdev->config->del_vqs(vdev);
+	virtsnd_ctl_msg_cancel_all(snd);
+
+	virtsnd_reset_vqs(snd);
+	virtsnd_ctl_msg_cleanup(snd);
 
 	for (i = 0; snd->substreams && i < snd->nsubstreams; ++i) {
 		struct virtio_pcm_substream *vss = &snd->substreams[i];
@@ -383,8 +418,8 @@
 	virtsnd_disable_event_vq(snd);
 	virtsnd_ctl_msg_cancel_all(snd);
 
-	virtio_reset_device(vdev);
-	vdev->config->del_vqs(vdev);
+	virtsnd_reset_vqs(snd);
+	virtsnd_ctl_msg_cleanup(snd);
 
 	for (i = 0; i < snd->nsubstreams; ++i)
 		cancel_work_sync(&snd->substreams[i].elapsed_period);
@@ -412,6 +447,7 @@
 		return rc;
 
 	virtio_device_ready(vdev);
+	virtsnd_ctl_msg_enable(snd);
 
 	virtsnd_enable_event_vq(snd);
 
diff --git a/sound/virtio/virtio_card.h b/sound/virtio/virtio_card.h
--- a/sound/virtio/virtio_card.h
+++ b/sound/virtio/virtio_card.h
@@ -47,6 +47,7 @@
  * @queues: Virtqueue wrappers.
  * @card: ALSA sound card.
  * @ctl_msgs: Pending control request list.
+ * @ctl_msgs_blocked: Control requests are blocked during teardown or suspend.
  * @event_msgs: Device events.
  * @pcm_list: VirtIO PCM device list.
  * @jacks: VirtIO jacks.
@@ -64,6 +65,7 @@
 	struct virtio_snd_queue queues[VIRTIO_SND_VQ_MAX];
 	struct snd_card *card;
 	struct list_head ctl_msgs;
+	bool ctl_msgs_blocked;
 	struct virtio_snd_event *event_msgs;
 	struct list_head pcm_list;
 	struct virtio_jack *jacks;
diff --git a/sound/virtio/virtio_ctl_msg.c b/sound/virtio/virtio_ctl_msg.c
--- a/sound/virtio/virtio_ctl_msg.c
+++ b/sound/virtio/virtio_ctl_msg.c
@@ -15,6 +15,7 @@
  * @list: Pending message list entry.
  * @notify: Request completed notification.
  * @ref_count: Reference count used to manage a message lifetime.
+ * @cancelled: Request canceled during device teardown.
  */
 struct virtio_snd_msg {
 	struct scatterlist sg_request;
@@ -22,6 +23,7 @@
 	struct list_head list;
 	struct completion notify;
 	refcount_t ref_count;
+	bool cancelled;
 };
 
 /**
@@ -100,7 +102,7 @@
 
 	INIT_LIST_HEAD(&msg->list);
 	init_completion(&msg->notify);
-	/* This reference is dropped in virtsnd_ctl_msg_complete(). */
+	/* The caller owns this reference; send adds a queue reference. */
 	refcount_set(&msg->ref_count, 1);
 
 	return msg;
@@ -134,6 +136,7 @@
 	int rc;
 
 	virtsnd_ctl_msg_ref(msg);
+	/* This reference stays with the queue until completion or teardown. */
 
 	/* Set the default status in case the message was canceled. */
 	response->code = cpu_to_le32(VIRTIO_SND_S_IO_ERR);
@@ -147,12 +150,17 @@
 		psgs[nouts + nins++] = in_sgs;
 
 	scoped_guard(spinlock_irqsave, &queue->lock) {
-		rc = virtqueue_add_sgs(queue->vqueue, psgs, nouts, nins, msg,
-				       GFP_ATOMIC);
+		if (snd->ctl_msgs_blocked)
+			rc = -ENODEV;
+		else
+			rc = virtqueue_add_sgs(queue->vqueue, psgs, nouts, nins,
+					       msg, GFP_ATOMIC);
 		if (!rc) {
 			notify = virtqueue_kick_prepare(queue->vqueue);
 
 			list_add_tail(&msg->list, &snd->ctl_msgs);
+			if (notify)
+				virtqueue_notify(queue->vqueue);
 		}
 	}
 
@@ -168,9 +176,6 @@
 
 		goto on_exit;
 	}
-
-	if (notify)
-		virtqueue_notify(queue->vqueue);
 
 	if (nowait)
 		goto on_exit;
@@ -184,6 +189,11 @@
 			rc = -ETIMEDOUT;
 		}
 
+		goto on_exit;
+	}
+
+	if (READ_ONCE(msg->cancelled)) {
+		rc = -ENODEV;
 		goto on_exit;
 	}
 
@@ -227,9 +237,33 @@
  * virtsnd_ctl_msg_cancel_all() - Cancel all pending control messages.
  * @snd: VirtIO sound device.
  *
- * Context: Any context.
+ * Context: Any context. Blocks new senders and marks pending requests canceled.
+ *          Existing senders are woken when their request completes or after
+ *          the device has been reset.
  */
 void virtsnd_ctl_msg_cancel_all(struct virtio_snd *snd)
+{
+	struct virtio_snd_queue *queue = virtsnd_control_queue(snd);
+	struct virtio_snd_msg *msg;
+
+	guard(spinlock_irqsave)(&queue->lock);
+	snd->ctl_msgs_blocked = true;
+	list_for_each_entry(msg, &snd->ctl_msgs, list) {
+		WRITE_ONCE(msg->cancelled, true);
+	}
+}
+
+/**
+ * virtsnd_ctl_msg_cleanup() - Release messages left after queue teardown.
+ * @snd: VirtIO sound device.
+ *
+ * This must be called after device reset, buffer detachment, and del_vqs(),
+ * when no queue callback or DMA access can still reach a message. The queue
+ * object and its lock must remain alive until this function returns.
+ *
+ * Context: Any context.
+ */
+void virtsnd_ctl_msg_cleanup(struct virtio_snd *snd)
 {
 	struct virtio_snd_queue *queue = virtsnd_control_queue(snd);
 
@@ -244,6 +278,20 @@
 }
 
 /**
+ * virtsnd_ctl_msg_enable() - Allow control requests after queue restoration.
+ * @snd: VirtIO sound device.
+ *
+ * Context: Any context.
+ */
+void virtsnd_ctl_msg_enable(struct virtio_snd *snd)
+{
+	struct virtio_snd_queue *queue = virtsnd_control_queue(snd);
+
+	guard(spinlock_irqsave)(&queue->lock);
+	snd->ctl_msgs_blocked = false;
+}
+
+/**
  * virtsnd_ctl_query_info() - Query the item configuration from the device.
  * @snd: VirtIO sound device.
  * @command: Control request code (VIRTIO_SND_R_XXX_INFO).
@@ -295,6 +343,8 @@
 	u32 length;
 
 	guard(spinlock_irqsave)(&queue->lock);
+	if (queue->vqueue != vqueue)
+		return;
 	do {
 		virtqueue_disable_cb(vqueue);
 		while ((msg = virtqueue_get_buf(vqueue, &length)))
diff --git a/sound/virtio/virtio_ctl_msg.h b/sound/virtio/virtio_ctl_msg.h
--- a/sound/virtio/virtio_ctl_msg.h
+++ b/sound/virtio/virtio_ctl_msg.h
@@ -68,6 +68,10 @@
 
 void virtsnd_ctl_msg_cancel_all(struct virtio_snd *snd);
 
+void virtsnd_ctl_msg_cleanup(struct virtio_snd *snd);
+
+void virtsnd_ctl_msg_enable(struct virtio_snd *snd);
+
 void virtsnd_ctl_msg_complete(struct virtio_snd_msg *msg);
 
 int virtsnd_ctl_query_info(struct virtio_snd *snd, int command, int start_id,
diff --git a/sound/virtio/virtio_pcm.h b/sound/virtio/virtio_pcm.h
--- a/sound/virtio/virtio_pcm.h
+++ b/sound/virtio/virtio_pcm.h
@@ -122,6 +122,8 @@
 
 void virtsnd_pcm_msg_reset(struct virtio_pcm_substream *vss);
 
+void virtsnd_pcm_msg_cancel(struct virtio_pcm_msg *msg);
+
 int virtsnd_pcm_msg_send(struct virtio_pcm_substream *vss, unsigned long offset,
 			 unsigned long bytes);
 
diff --git a/sound/virtio/virtio_pcm_msg.c b/sound/virtio/virtio_pcm_msg.c
--- a/sound/virtio/virtio_pcm_msg.c
+++ b/sound/virtio/virtio_pcm_msg.c
@@ -305,6 +305,35 @@
 {
 	guard(spinlock_irqsave)(&vss->lock);
 	return vss->msg_count;
+}
+
+/**
+ * virtsnd_pcm_msg_cancel() - Cancel a message detached during device reset.
+ * @msg: Message returned by virtqueue_detach_unused_buf().
+ *
+ * Reset cancellation retires the queue token without treating it as a device
+ * completion. In particular, it does not advance the PCM hardware pointer or
+ * schedule period work.
+ *
+ * Context: Process context with the PCM queue lock held.
+ */
+void virtsnd_pcm_msg_cancel(struct virtio_pcm_msg *msg)
+{
+	struct virtio_pcm_substream *vss = msg->substream;
+	bool wake;
+
+	{
+		guard(spinlock)(&vss->lock);
+		msg->length = 0;
+		vss->xfer_enabled = false;
+		if (WARN_ON_ONCE(!vss->msg_count))
+			return;
+		vss->msg_count--;
+		wake = !vss->msg_count;
+	}
+
+	if (wake)
+		wake_up_all(&vss->msg_empty);
 }
 
 /**
@@ -367,16 +396,21 @@
 
 /**
  * virtsnd_pcm_notify_cb() - Process all completed I/O messages.
- * @queue: Underlying tx/rx virtqueue.
+ * @queue: Underlying tx/rx queue wrapper.
+ * @vqueue: Queue that raised this callback.
  *
  * Context: Interrupt context. Takes and releases the tx/rx queue spinlock.
  */
-static inline void virtsnd_pcm_notify_cb(struct virtio_snd_queue *queue)
+static inline void
+virtsnd_pcm_notify_cb(struct virtio_snd_queue *queue, struct virtqueue *vqueue)
 {
 	struct virtio_pcm_msg *msg;
 	u32 written_bytes;
 
 	guard(spinlock_irqsave)(&queue->lock);
+	/* Reset can detach the queue before the transport drains its IRQs. */
+	if (queue->vqueue != vqueue)
+		return;
 	do {
 		virtqueue_disable_cb(queue->vqueue);
 		while ((msg = virtqueue_get_buf(queue->vqueue, &written_bytes)))
@@ -394,7 +428,7 @@
 {
 	struct virtio_snd *snd = vqueue->vdev->priv;
 
-	virtsnd_pcm_notify_cb(virtsnd_tx_queue(snd));
+	virtsnd_pcm_notify_cb(virtsnd_tx_queue(snd), vqueue);
 }
 
 /**
@@ -407,7 +441,7 @@
 {
 	struct virtio_snd *snd = vqueue->vdev->priv;
 
-	virtsnd_pcm_notify_cb(virtsnd_rx_queue(snd));
+	virtsnd_pcm_notify_cb(virtsnd_rx_queue(snd), vqueue);
 }
 
 /**

[PATCH v12 04/12] ALSA: virtio: Keep PCM DMA buffers alive through device ownership

[Joonho Lee]

Retain references to the pages backing ALSA's vmalloc buffer until its
queued I/O messages are freed. On NOMMU systems, use driver-owned memory
for DMA and copy playback data before submission and capture data after
completion, so ALSA buffer release cannot invalidate an active request.

Validate the PCM buffer and each period before building scatterlists.
Clamp device-reported capture lengths to the message capacity before
advancing the hardware pointer, preventing an oversized completion from
accounting bytes beyond the buffer actually submitted.

Require whole-byte periods whose total matches the negotiated buffer.
Check offset, length and the final message index before indexing the
message array, including empty transfers and arithmetic wraparound.

Seed the ordinary capture ring once per prepare. Native STOP preserves
queued requests and unread samples, so PAUSE_RELEASE must retain their
ownership. Re-enqueuing the entire ring adds another period to pending
message lengths and exposes unread DMA slots to the device. Return
consumed space through the existing indirect ACK path.

Fixes: f40a28679e0b ("ALSA: virtio: handling control and I/O messages for the PCM device")
Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_pcm.h b/sound/virtio/virtio_pcm.h
--- a/sound/virtio/virtio_pcm.h
+++ b/sound/virtio/virtio_pcm.h
@@ -13,6 +13,7 @@
 
 struct virtio_pcm;
 struct virtio_pcm_msg;
+struct page;
 
 /**
  * struct virtio_pcm_substream - VirtIO PCM substream.
@@ -29,6 +30,7 @@
  *        substream operators, including @stopped.
  * @buffer_bytes: Current buffer size in bytes.
  * @hw_ptr: Substream hardware pointer value in bytes [0 ... buffer_bytes).
+ * @capture_rx_seeded: Ordinary capture ring was queued since prepare.
  * @xfer_enabled: Data transfer state (0 - off, 1 - on).
  * @xfer_xrun: Data underflow/overflow state (0 - no xrun, 1 - xrun).
  * @stopped: True if the substream is stopped and must be released on the device
@@ -38,6 +40,10 @@
  * @msgs: Allocated I/O messages.
  * @nmsgs: Number of allocated I/O messages.
  * @msg_count: Number of pending I/O messages in the virtqueue.
+ * @dma_pages: Pages backing DMA messages, retained until messages are freed.
+ * @ndma_pages: Number of retained pages.
+ * @owned_dma_area: Driver-owned PCM buffer backing asynchronous messages.
+ * @owned_dma_bytes: Size of @owned_dma_area.
  * @msg_empty: Notify when msg_count is zero.
  */
 struct virtio_pcm_substream {
@@ -53,6 +59,7 @@
 	spinlock_t lock;
 	size_t buffer_bytes;
 	size_t hw_ptr;
+	bool capture_rx_seeded;
 	bool xfer_enabled;
 	bool xfer_xrun;
 	bool stopped;
@@ -62,6 +69,12 @@
 	struct virtio_pcm_msg **msgs;
 	unsigned int nmsgs;
 	unsigned int msg_count;
+	struct page **dma_pages;
+	unsigned int ndma_pages;
+#ifndef CONFIG_MMU
+	u8 *owned_dma_area;
+	size_t owned_dma_bytes;
+#endif
 	wait_queue_head_t msg_empty;
 };
 
diff --git a/sound/virtio/virtio_pcm_msg.c b/sound/virtio/virtio_pcm_msg.c
--- a/sound/virtio/virtio_pcm_msg.c
+++ b/sound/virtio/virtio_pcm_msg.c
@@ -3,6 +3,9 @@
  * virtio-snd: Virtio sound device
  * Copyright (C) 2021 OpenSynergy GmbH
  */
+#include <linux/limits.h>
+#include <linux/mm.h>
+#include <linux/overflow.h>
 #include <sound/pcm_params.h>
 
 #include "virtio_card.h"
@@ -13,6 +16,7 @@
  * @xfer: Request header payload.
  * @status: Response header payload.
  * @length: Data length in bytes.
+ * @data_offset: Offset in the driver-owned PCM buffer on NOMMU systems.
  * @sgs: Payload scatter-gather table.
  */
 struct virtio_pcm_msg {
@@ -20,6 +24,9 @@
 	struct virtio_snd_pcm_xfer xfer;
 	struct virtio_snd_pcm_status status;
 	size_t length;
+#ifndef CONFIG_MMU
+	size_t data_offset;
+#endif
 	struct scatterlist sgs[];
 };
 
@@ -37,9 +44,8 @@
 };
 
 /**
- * virtsnd_pcm_sg_num() - Count the number of sg-elements required to represent
- *                        vmalloc'ed buffer.
- * @data: Pointer to vmalloc'ed buffer.
+ * virtsnd_pcm_sg_num() - Count the sg-elements for a page-backed buffer.
+ * @data: Pointer to the page-backed buffer.
  * @length: Buffer size.
  *
  * Context: Any context.
@@ -47,14 +53,18 @@
  */
 static int virtsnd_pcm_sg_num(u8 *data, unsigned int length)
 {
-	phys_addr_t sg_address;
-	unsigned int sg_length;
+	phys_addr_t sg_address = 0;
+	unsigned int sg_length = 0;
 	int num = 0;
 
 	while (length) {
 		struct page *pg = vmalloc_to_page(data);
-		phys_addr_t pg_address = page_to_phys(pg);
+		phys_addr_t pg_address;
 		size_t pg_length;
+
+		if (!pg)
+			return -EINVAL;
+		pg_address = page_to_phys(pg);
 
 		pg_length = PAGE_SIZE - offset_in_page(data);
 		if (pg_length > length)
@@ -75,11 +85,80 @@
 	return num;
 }
 
-/**
- * virtsnd_pcm_sg_from() - Build sg-list from vmalloc'ed buffer.
+#ifdef CONFIG_MMU
+/**
+ * virtsnd_pcm_dma_pages_put() - Release retained PCM buffer pages.
+ * @vss: VirtIO PCM substream.
+ *
+ * Context: Process context.
+ */
+static void virtsnd_pcm_dma_pages_put(struct virtio_pcm_substream *vss)
+{
+	while (vss->ndma_pages)
+		put_page(vss->dma_pages[--vss->ndma_pages]);
+	kfree(vss->dma_pages);
+	vss->dma_pages = NULL;
+}
+
+/**
+ * virtsnd_pcm_dma_pages_get() - Retain the pages backing the PCM buffer.
+ * @vss: VirtIO PCM substream.
+ * @data: PCM buffer address.
+ * @length: PCM buffer size.
+ *
+ * Keep the backing pages alive while virtqueue descriptors can still refer to
+ * them, even if ALSA releases the vmalloc mapping after a failed drain.
+ *
+ * Context: Process context.
+ * Return: 0 on success, -errno on failure.
+ */
+static int virtsnd_pcm_dma_pages_get(struct virtio_pcm_substream *vss,
+				     u8 *data, size_t length)
+{
+	size_t i;
+	size_t npages;
+	size_t span;
+
+	if (!data || !length || !is_vmalloc_addr(data))
+		return -EINVAL;
+	if (vss->dma_pages || vss->ndma_pages)
+		return -EBUSY;
+
+	if (check_add_overflow(length, (size_t)offset_in_page(data), &span))
+		return -EOVERFLOW;
+
+	npages = span / PAGE_SIZE + !!(span % PAGE_SIZE);
+	if (npages > UINT_MAX)
+		return -EOVERFLOW;
+
+	vss->dma_pages = kcalloc(npages, sizeof(*vss->dma_pages), GFP_KERNEL);
+	if (!vss->dma_pages)
+		return -ENOMEM;
+
+	for (i = 0; i < npages; i++) {
+		struct page *page = vmalloc_to_page(data + i * PAGE_SIZE);
+
+		if (!page)
+			goto err_put_pages;
+
+		get_page(page);
+		vss->dma_pages[vss->ndma_pages++] = page;
+	}
+
+	return 0;
+
+err_put_pages:
+	virtsnd_pcm_dma_pages_put(vss);
+
+	return -EINVAL;
+}
+#endif
+
+/**
+ * virtsnd_pcm_sg_from() - Build an sg-list from a page-backed buffer.
  * @sgs: Preallocated sg-list to populate.
  * @nsgs: The maximum number of elements in the @sgs.
- * @data: Pointer to vmalloc'ed buffer.
+ * @data: Pointer to the page-backed buffer.
  * @length: Buffer size.
  *
  * Splits the buffer into physically contiguous parts and makes an sg-list of
@@ -132,36 +211,99 @@
 int virtsnd_pcm_msg_alloc(struct virtio_pcm_substream *vss,
 			  unsigned int periods, unsigned int period_bytes)
 {
-	struct snd_pcm_runtime *runtime = vss->substream->runtime;
+	struct snd_pcm_runtime *runtime;
+	size_t buffer_bytes;
 	unsigned int i;
+	int rc;
+
+	if (vss->msgs || vss->nmsgs || vss->dma_pages || vss->ndma_pages)
+		return -EBUSY;
+#ifndef CONFIG_MMU
+	if (vss->owned_dma_area || vss->owned_dma_bytes)
+		return -EBUSY;
+#endif
+	if (!periods || !period_bytes)
+		return -EINVAL;
+
+	if (check_mul_overflow((size_t)periods, (size_t)period_bytes,
+			       &buffer_bytes))
+		return -EOVERFLOW;
+
+	if (!vss->substream || !vss->substream->runtime)
+		return -EINVAL;
+	runtime = vss->substream->runtime;
+	if (!runtime->dma_area || buffer_bytes > runtime->dma_bytes)
+		return -EINVAL;
+
+#ifdef CONFIG_MMU
+	rc = virtsnd_pcm_dma_pages_get(vss, runtime->dma_area, buffer_bytes);
+	if (rc)
+		return rc;
+#else
+	vss->owned_dma_area = kzalloc(buffer_bytes, GFP_KERNEL);
+	if (!vss->owned_dma_area)
+		return -ENOMEM;
+	vss->owned_dma_bytes = buffer_bytes;
+#endif
 
 	vss->msgs = kzalloc_objs(*vss->msgs, periods);
-	if (!vss->msgs)
-		return -ENOMEM;
+	if (!vss->msgs) {
+		rc = -ENOMEM;
+		goto err_free_pages;
+	}
 
 	vss->nmsgs = periods;
 
 	for (i = 0; i < periods; ++i) {
-		u8 *data = runtime->dma_area + period_bytes * i;
-		int sg_num = virtsnd_pcm_sg_num(data, period_bytes);
+		int sg_num;
 		struct virtio_pcm_msg *msg;
+		u8 *data;
+#ifdef CONFIG_MMU
+		data = runtime->dma_area + (size_t)period_bytes * i;
+#else
+		data = vss->owned_dma_area + (size_t)period_bytes * i;
+#endif
+		sg_num = virtsnd_pcm_sg_num(data, period_bytes);
+		if (sg_num < 0) {
+			rc = sg_num;
+			goto err_free_msgs;
+		}
 
 		msg = kzalloc_flex(*msg, sgs, sg_num + 2);
-		if (!msg)
-			return -ENOMEM;
+		if (!msg) {
+			rc = -ENOMEM;
+			goto err_free_msgs;
+		}
 
 		msg->substream = vss;
+		vss->msgs[i] = msg;
 		sg_init_one(&msg->sgs[PCM_MSG_SG_XFER], &msg->xfer,
 			    sizeof(msg->xfer));
 		sg_init_one(&msg->sgs[PCM_MSG_SG_STATUS], &msg->status,
 			    sizeof(msg->status));
+#ifndef CONFIG_MMU
+		msg->data_offset = (size_t)period_bytes * i;
+#endif
 		virtsnd_pcm_sg_from(&msg->sgs[PCM_MSG_SG_DATA], sg_num, data,
 				    period_bytes);
-
-		vss->msgs[i] = msg;
 	}
 
 	return 0;
+
+err_free_msgs:
+	virtsnd_pcm_msg_free(vss);
+	return rc;
+
+err_free_pages:
+#ifdef CONFIG_MMU
+	virtsnd_pcm_dma_pages_put(vss);
+#else
+	kfree(vss->owned_dma_area);
+	vss->owned_dma_area = NULL;
+	vss->owned_dma_bytes = 0;
+#endif
+
+	return rc;
 }
 
 /**
@@ -177,6 +319,13 @@
 	for (i = 0; vss->msgs && i < vss->nmsgs; ++i)
 		kfree(vss->msgs[i]);
 	kfree(vss->msgs);
+#ifdef CONFIG_MMU
+	virtsnd_pcm_dma_pages_put(vss);
+#else
+	kfree(vss->owned_dma_area);
+	vss->owned_dma_area = NULL;
+	vss->owned_dma_bytes = 0;
+#endif
 
 	vss->msgs = NULL;
 	vss->nmsgs = 0;
@@ -199,6 +348,7 @@
 	guard(spinlock_irqsave)(&vss->lock);
 	for (i = 0; vss->msgs && i < vss->nmsgs; ++i)
 		vss->msgs[i]->length = 0;
+	vss->capture_rx_seeded = false;
 }
 
 /**
@@ -228,8 +378,15 @@
 	bool notify = false;
 	int rc;
 
+	if (!bytes)
+		return 0;
+	if (!period_bytes || !vss->msgs || offset >= vss->buffer_bytes ||
+	    bytes > vss->buffer_bytes - offset)
+		return -EINVAL;
 	start = offset / period_bytes;
 	end = (offset + bytes - 1) / period_bytes;
+	if (end >= vss->nmsgs)
+		return -EINVAL;
 
 	for (i = start; i <= end; i++) {
 		struct virtio_pcm_msg *msg = vss->msgs[i];
@@ -248,6 +405,33 @@
 		if (msg->length == period_bytes) {
 			msg->xfer.stream_id = cpu_to_le32(vss->sid);
 			memset(&msg->status, 0, sizeof(msg->status));
+
+#ifndef CONFIG_MMU
+		if (vss->direction == SNDRV_PCM_STREAM_PLAYBACK) {
+			struct snd_pcm_runtime *runtime =
+				vss->substream ? vss->substream->runtime : NULL;
+			size_t data_bytes = msg->length;
+
+			if (!runtime || !runtime->dma_area || !vss->owned_dma_area ||
+			    msg->data_offset > runtime->dma_bytes ||
+			    data_bytes > runtime->dma_bytes - msg->data_offset ||
+			    msg->data_offset > vss->owned_dma_bytes ||
+			    data_bytes > vss->owned_dma_bytes - msg->data_offset) {
+				dev_err_ratelimited(&vdev->dev,
+						    "SID %u: invalid playback buffer for I/O message\n",
+					vss->sid);
+				msg->length -= n;
+				if (msg_count != vss->msg_count &&
+				    !(vss->features &
+				      (1U << VIRTIO_SND_PCM_F_MSG_POLLING)) &&
+				    virtqueue_kick_prepare(vqueue))
+					virtqueue_notify(vqueue);
+				return -EINVAL;
+			}
+			memcpy(vss->owned_dma_area + msg->data_offset,
+			       runtime->dma_area + msg->data_offset, data_bytes);
+		}
+#endif
 
 			if (vss->direction == SNDRV_PCM_STREAM_PLAYBACK)
 				rc = virtqueue_add_sgs(vqueue, psgs, 2, 1, msg,
@@ -363,15 +547,48 @@
 	 * the hw_ptr value is unconditionally advanced.
 	 */
 	guard(spinlock)(&vss->lock);
+#ifndef CONFIG_MMU
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+	    vss->xfer_enabled && written_bytes > sizeof(msg->status)) {
+		struct snd_pcm_runtime *runtime =
+			vss->substream ? vss->substream->runtime : NULL;
+		size_t reported_bytes = written_bytes - sizeof(msg->status);
+		size_t data_bytes = min_t(size_t, reported_bytes, msg->length);
+
+		if (!runtime || !runtime->dma_area || !vss->owned_dma_area ||
+		    msg->data_offset > runtime->dma_bytes ||
+		    data_bytes > runtime->dma_bytes - msg->data_offset ||
+		    msg->data_offset > vss->owned_dma_bytes ||
+		    data_bytes > vss->owned_dma_bytes - msg->data_offset)
+			dev_err_ratelimited(&vss->snd->vdev->dev,
+					    "SID %u: invalid capture buffer for I/O message\n",
+				vss->sid);
+		else
+			memcpy(runtime->dma_area + msg->data_offset,
+			       vss->owned_dma_area + msg->data_offset, data_bytes);
+	}
+#endif
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+	    written_bytes > sizeof(msg->status) &&
+	    written_bytes - sizeof(msg->status) > msg->length)
+		dev_err_ratelimited(&vss->snd->vdev->dev,
+				    "SID %u: capture length exceeds I/O message capacity\n",
+			vss->sid);
 	/*
 	 * If the capture substream returned an incorrect status, then just
-	 * increase the hw_ptr by the message size.
+	 * increase the hw_ptr by the message size. Clamp valid capture lengths to
+	 * the actual SG payload capacity before accounting them.
 	 */
 	if (vss->direction == SNDRV_PCM_STREAM_PLAYBACK ||
-	    written_bytes <= sizeof(msg->status))
+	    written_bytes <= sizeof(msg->status)) {
 		vss->hw_ptr += msg->length;
-	else
-		vss->hw_ptr += written_bytes - sizeof(msg->status);
+	} else {
+		size_t data_bytes = min_t(size_t,
+					  written_bytes - sizeof(msg->status),
+					  msg->length);
+
+		vss->hw_ptr += data_bytes;
+	}
 
 	if (vss->hw_ptr >= vss->buffer_bytes)
 		vss->hw_ptr -= vss->buffer_bytes;
@@ -387,7 +604,6 @@
 		runtime->delay =
 			bytes_to_frames(runtime,
 					le32_to_cpu(msg->status.latency_bytes));
-
 		schedule_work(&vss->elapsed_period);
 	} else if (!vss->msg_count) {
 		wake_up_all(&vss->msg_empty);
diff --git a/sound/virtio/virtio_pcm_ops.c b/sound/virtio/virtio_pcm_ops.c
--- a/sound/virtio/virtio_pcm_ops.c
+++ b/sound/virtio/virtio_pcm_ops.c
@@ -25,6 +25,8 @@
  *   signal or due to a timeout. In this case, the device can still access
  *   messages even after calling ops->hw_free(). It can also issue an interrupt,
  *   and the interrupt handler will also try to access message structures.
+ *   The messages retain references to the PCM buffer pages until they are
+ *   freed, so late DMA does not target pages released by ALSA.
  *
  *   Therefore, freeing of already allocated messages occurs:
  *
@@ -228,8 +230,19 @@
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_device *vdev = vss->snd->vdev;
 	unsigned int msg_count;
+	size_t period_buffer_bytes;
 	bool stopped;
 	int rc;
+
+	/* Fixed descriptors require whole-byte periods covering the buffer. */
+	if (check_mul_overflow((size_t)params_periods(hw_params),
+			       (size_t)params_period_bytes(hw_params),
+			       &period_buffer_bytes) ||
+	    period_buffer_bytes != params_buffer_bytes(hw_params) ||
+	    (params_period_size(hw_params) *
+	     (u64)snd_pcm_format_physical_width(params_format(hw_params)) *
+	     params_channels(hw_params)) % 8)
+		return -EINVAL;
 
 	scoped_guard(spinlock_irqsave, &vss->lock) {
 		msg_count = vss->msg_count;
@@ -380,8 +393,12 @@
 				substream->runtime->stop_operating = true;
 				return -EBADFD;
 			}
-			if (vss->direction == SNDRV_PCM_STREAM_CAPTURE)
+			if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+			    !vss->capture_rx_seeded) {
 				rc = virtsnd_pcm_msg_send(vss, 0, vss->buffer_bytes);
+				if (!rc)
+					vss->capture_rx_seeded = true;
+			}
 			if (rc) {
 				vss->stopped = true;
 				substream->runtime->stop_operating = true;

[PATCH v12 05/12] ALSA: virtio: Keep control payloads alive through DMA completion

[Joonho Lee]

Control requests can expose caller-owned payloads to the device after a
synchronous send times out. Keep request and response headers and payloads
in the reference-counted message so caller cleanup cannot invalidate DMA.

Keep the queue reference until device completion or post-reset cleanup,
even when the caller times out or is interrupted. Consume the caller's
reference on every send path, and let response readers retain an extra
reference through copyout. Check payload sizes and allocation arithmetic
before building the scatterlists.

A control value payload is also read by the CPU as a typed structure. DMA
cacheline alignment alone does not guarantee the structure's CPU alignment.
Align regions to the greater of the DMA cacheline and control-value type
alignment, and reserve prefix slack from the buffer base so typed accesses
remain aligned while scatterlist regions stay on cacheline boundaries.

Keep response values in a separately allocated aligned response buffer
so INTEGER64 control accesses are aligned from this patch onward.

Fixes: 9d45e514da88 ("ALSA: virtio: handling control messages")
Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_ctl_msg.c b/sound/virtio/virtio_ctl_msg.c
--- a/sound/virtio/virtio_ctl_msg.c
+++ b/sound/virtio/virtio_ctl_msg.c
@@ -3,7 +3,10 @@
  * virtio-snd: Virtio sound device
  * Copyright (C) 2021 OpenSynergy GmbH
  */
+#include <linux/dma-mapping.h>
+#include <linux/limits.h>
 #include <linux/moduleparam.h>
+#include <linux/overflow.h>
 #include <linux/virtio_config.h>
 
 #include "virtio_card.h"
@@ -11,7 +14,12 @@
 /**
  * struct virtio_snd_msg - Control message.
  * @sg_request: Scattergather list containing a device request (header).
+ * @sg_request_data: Scattergather list containing request data.
  * @sg_response: Scattergather list containing a device response (status).
+ * @sg_response_data: Scattergather list containing response data.
+ * @dma_buffer: Message-owned storage for all DMA-visible SG regions.
+ * @request_data_size: Size of the request data buffer.
+ * @response_data_size: Size of the response data buffer.
  * @list: Pending message list entry.
  * @notify: Request completed notification.
  * @ref_count: Reference count used to manage a message lifetime.
@@ -19,7 +27,12 @@
  */
 struct virtio_snd_msg {
 	struct scatterlist sg_request;
+	struct scatterlist sg_request_data;
 	struct scatterlist sg_response;
+	struct scatterlist sg_response_data;
+	void *dma_buffer;
+	size_t request_data_size;
+	size_t response_data_size;
 	struct list_head list;
 	struct completion notify;
 	refcount_t ref_count;
@@ -47,8 +60,10 @@
  */
 void virtsnd_ctl_msg_unref(struct virtio_snd_msg *msg)
 {
-	if (refcount_dec_and_test(&msg->ref_count))
+	if (refcount_dec_and_test(&msg->ref_count)) {
+		kfree(msg->dma_buffer);
 		kfree(msg);
+	}
 }
 
 /**
@@ -71,12 +86,73 @@
 void *virtsnd_ctl_msg_response(struct virtio_snd_msg *msg)
 {
 	return sg_virt(&msg->sg_response);
+}
+
+/**
+ * virtsnd_ctl_msg_request_data() - Get the request data buffer.
+ * @msg: Control message.
+ *
+ * Context: Any context.
+ */
+void *virtsnd_ctl_msg_request_data(struct virtio_snd_msg *msg)
+{
+	return msg->request_data_size ? sg_virt(&msg->sg_request_data) : NULL;
+}
+
+/**
+ * virtsnd_ctl_msg_response_data() - Get the response data buffer.
+ * @msg: Control message.
+ *
+ * Context: Any context.
+ */
+void *virtsnd_ctl_msg_response_data(struct virtio_snd_msg *msg)
+{
+	return msg->response_data_size ? sg_virt(&msg->sg_response_data) : NULL;
+}
+
+/**
+ * virtsnd_ctl_msg_sg_init() - Place one DMA region on cacheline boundaries.
+ * @sg: Scattergather entry for the region.
+ * @cursor: Current location in the message-owned DMA buffer.
+ * @size: Region length.
+ * @alignment: Alignment for DMA cachelines and typed CPU accesses.
+ *
+ * Context: Any context during message allocation.
+ * Return: 0 on success, -EOVERFLOW if the address cannot be represented.
+ */
+static int virtsnd_ctl_msg_sg_init(struct scatterlist *sg, u8 **cursor,
+				   size_t size, size_t alignment)
+{
+	uintptr_t address = (uintptr_t)*cursor;
+	uintptr_t end;
+	size_t padding;
+
+	padding = address % alignment;
+	if (padding &&
+	    check_add_overflow(address, (uintptr_t)(alignment - padding),
+			       &address))
+		return -EOVERFLOW;
+
+	if (check_add_overflow(address, (uintptr_t)size, &end))
+		return -EOVERFLOW;
+
+	padding = end % alignment;
+	if (padding &&
+	    check_add_overflow(end, (uintptr_t)(alignment - padding), &end))
+		return -EOVERFLOW;
+
+	sg_init_one(sg, (void *)address, (unsigned int)size);
+	*cursor = (u8 *)end;
+
+	return 0;
 }
 
 /**
  * virtsnd_ctl_msg_alloc() - Allocate and initialize a control message.
  * @request_size: Size of request header.
  * @response_size: Size of response header.
+ * @request_data_size: Size of request data buffer.
+ * @response_data_size: Size of response data buffer.
  * @gfp: Kernel flags for memory allocation.
  *
  * The message will be automatically freed when the ref_count value is 0.
@@ -85,20 +161,85 @@
  * Return: Allocated message on success, NULL on failure.
  */
 struct virtio_snd_msg *virtsnd_ctl_msg_alloc(size_t request_size,
-					     size_t response_size, gfp_t gfp)
+					     size_t response_size,
+					     size_t request_data_size,
+					     size_t response_data_size,
+					     gfp_t gfp)
 {
 	struct virtio_snd_msg *msg;
-
-	if (!request_size || !response_size)
+	struct scatterlist *sgs[4];
+	size_t sizes[4] = {
+		request_size,
+		response_size,
+		request_data_size,
+		response_data_size,
+	};
+	size_t alignment;
+	size_t dma_buffer_size;
+	size_t aligned_size;
+	size_t padding;
+	u8 *data;
+	unsigned int i;
+	int cache_alignment;
+	int rc;
+
+	if (!request_size || !response_size || request_size > UINT_MAX ||
+	    response_size > UINT_MAX || request_data_size > UINT_MAX ||
+	    response_data_size > UINT_MAX)
 		return NULL;
-
-	msg = kzalloc(sizeof(*msg) + request_size + response_size, gfp);
+	cache_alignment = dma_get_cache_alignment();
+	if (cache_alignment <= 0)
+		return NULL;
+	/* DMA coherence does not remove the CPU's type alignment requirement. */
+	alignment = max_t(size_t, cache_alignment,
+			  __alignof__(struct virtio_snd_ctl_value));
+
+	/* Reserve prefix alignment slack before the first DMA region. */
+	dma_buffer_size = alignment - 1;
+	for (i = 0; i < ARRAY_SIZE(sizes); ++i) {
+		if (!sizes[i])
+			continue;
+
+		padding = sizes[i] % alignment;
+		aligned_size = sizes[i];
+		if (padding &&
+		    check_add_overflow(aligned_size, alignment - padding,
+				       &aligned_size))
+			return NULL;
+
+		if (check_add_overflow(dma_buffer_size, aligned_size,
+				       &dma_buffer_size))
+			return NULL;
+	}
+
+	msg = kzalloc(sizeof(*msg), gfp);
 	if (!msg)
 		return NULL;
 
-	sg_init_one(&msg->sg_request, (u8 *)msg + sizeof(*msg), request_size);
-	sg_init_one(&msg->sg_response, (u8 *)msg + sizeof(*msg) + request_size,
-		    response_size);
+	msg->dma_buffer = kzalloc(dma_buffer_size, gfp);
+	if (!msg->dma_buffer) {
+		kfree(msg);
+		return NULL;
+	}
+	data = msg->dma_buffer;
+	sgs[0] = &msg->sg_request;
+	sgs[1] = &msg->sg_response;
+	sgs[2] = &msg->sg_request_data;
+	sgs[3] = &msg->sg_response_data;
+	for (i = 0; i < ARRAY_SIZE(sizes); ++i) {
+		if (!sizes[i])
+			continue;
+
+		rc = virtsnd_ctl_msg_sg_init(sgs[i], &data, sizes[i], alignment);
+		if (rc) {
+			kfree(msg->dma_buffer);
+			kfree(msg);
+			return NULL;
+		}
+	}
+
+	msg->request_data_size = request_data_size;
+	msg->response_data_size = response_data_size;
 
 	INIT_LIST_HEAD(&msg->list);
 	init_completion(&msg->notify);
@@ -112,17 +253,15 @@
  * virtsnd_ctl_msg_send() - Send a control message.
  * @snd: VirtIO sound device.
  * @msg: Control message.
- * @out_sgs: Additional sg-list to attach to the request header (may be NULL).
- * @in_sgs: Additional sg-list to attach to the response header (may be NULL).
  * @nowait: Flag indicating whether to wait for completion.
  *
  * Context: Any context. Takes and releases the control queue spinlock.
  *          May sleep if @nowait is false.
+ *          Consumes the caller's reference on every return path.
  * Return: 0 on success, -errno on failure.
  */
 int virtsnd_ctl_msg_send(struct virtio_snd *snd, struct virtio_snd_msg *msg,
-			 struct scatterlist *out_sgs,
-			 struct scatterlist *in_sgs, bool nowait)
+			 bool nowait)
 {
 	struct virtio_device *vdev = snd->vdev;
 	struct virtio_snd_queue *queue = virtsnd_control_queue(snd);
@@ -133,6 +272,7 @@
 	unsigned int nins = 0;
 	struct scatterlist *psgs[4];
 	bool notify = false;
+	unsigned long flags;
 	int rc;
 
 	virtsnd_ctl_msg_ref(msg);
@@ -142,27 +282,27 @@
 	response->code = cpu_to_le32(VIRTIO_SND_S_IO_ERR);
 
 	psgs[nouts++] = &msg->sg_request;
-	if (out_sgs)
-		psgs[nouts++] = out_sgs;
+	if (msg->request_data_size)
+		psgs[nouts++] = &msg->sg_request_data;
 
 	psgs[nouts + nins++] = &msg->sg_response;
-	if (in_sgs)
-		psgs[nouts + nins++] = in_sgs;
-
-	scoped_guard(spinlock_irqsave, &queue->lock) {
-		if (snd->ctl_msgs_blocked)
-			rc = -ENODEV;
-		else
-			rc = virtqueue_add_sgs(queue->vqueue, psgs, nouts, nins,
-					       msg, GFP_ATOMIC);
-		if (!rc) {
-			notify = virtqueue_kick_prepare(queue->vqueue);
-
-			list_add_tail(&msg->list, &snd->ctl_msgs);
-			if (notify)
-				virtqueue_notify(queue->vqueue);
-		}
-	}
+	if (msg->response_data_size)
+		psgs[nouts + nins++] = &msg->sg_response_data;
+
+	spin_lock_irqsave(&queue->lock, flags);
+	if (snd->ctl_msgs_blocked)
+		rc = -ENODEV;
+	else
+		rc = virtqueue_add_sgs(queue->vqueue, psgs, nouts, nins, msg,
+				       GFP_ATOMIC);
+	if (!rc) {
+		notify = virtqueue_kick_prepare(queue->vqueue);
+
+		list_add_tail(&msg->list, &snd->ctl_msgs);
+		if (notify)
+			virtqueue_notify(queue->vqueue);
+	}
+	spin_unlock_irqrestore(&queue->lock, flags);
 
 	if (rc) {
 		dev_err(&vdev->dev, "failed to send control message (0x%08x)\n",
@@ -173,46 +313,36 @@
 		 * called, it is necessary to decrement the reference count.
 		 */
 		virtsnd_ctl_msg_unref(msg);
-
-		goto on_exit;
-	}
-
-	if (nowait)
-		goto on_exit;
-
-	rc = wait_for_completion_interruptible_timeout(&msg->notify, js);
-	if (rc <= 0) {
+	} else if (!nowait) {
+		rc = wait_for_completion_interruptible_timeout(&msg->notify, js);
 		if (!rc) {
 			dev_err(&vdev->dev,
 				"control message (0x%08x) timeout\n",
 				le32_to_cpu(request->code));
 			rc = -ETIMEDOUT;
+		} else if (rc > 0) {
+			if (READ_ONCE(msg->cancelled)) {
+				rc = -ENODEV;
+			} else {
+				switch (le32_to_cpu(response->code)) {
+				case VIRTIO_SND_S_OK:
+					rc = 0;
+					break;
+				case VIRTIO_SND_S_NOT_SUPP:
+					rc = -EOPNOTSUPP;
+					break;
+				case VIRTIO_SND_S_IO_ERR:
+					rc = -EIO;
+					break;
+				default:
+					rc = -EINVAL;
+					break;
+				}
+			}
 		}
-
-		goto on_exit;
-	}
-
-	if (READ_ONCE(msg->cancelled)) {
-		rc = -ENODEV;
-		goto on_exit;
-	}
-
-	switch (le32_to_cpu(response->code)) {
-	case VIRTIO_SND_S_OK:
-		rc = 0;
-		break;
-	case VIRTIO_SND_S_NOT_SUPP:
-		rc = -EOPNOTSUPP;
-		break;
-	case VIRTIO_SND_S_IO_ERR:
-		rc = -EIO;
-		break;
-	default:
-		rc = -EINVAL;
-		break;
-	}
-
-on_exit:
+	}
+
+	/* Drop the reference consumed by this send attempt. */
 	virtsnd_ctl_msg_unref(msg);
 
 	return rc;
@@ -308,10 +438,23 @@
 {
 	struct virtio_snd_msg *msg;
 	struct virtio_snd_query_info *query;
-	struct scatterlist sg;
+	size_t info_size;
+	int rc;
+
+	if (count < 0)
+		return -EINVAL;
+
+	if (size > UINT_MAX ||
+	    check_mul_overflow((size_t)count, size, &info_size) ||
+	    info_size > UINT_MAX)
+		return -EOVERFLOW;
+
+	if (info_size && !info)
+		return -EINVAL;
 
 	msg = virtsnd_ctl_msg_alloc(sizeof(*query),
-				    sizeof(struct virtio_snd_hdr), GFP_KERNEL);
+				    sizeof(struct virtio_snd_hdr), 0, info_size,
+				    GFP_KERNEL);
 	if (!msg)
 		return -ENOMEM;
 
@@ -321,9 +464,13 @@
 	query->count = cpu_to_le32(count);
 	query->size = cpu_to_le32(size);
 
-	sg_init_one(&sg, info, count * size);
-
-	return virtsnd_ctl_msg_send(snd, msg, NULL, &sg, false);
+	virtsnd_ctl_msg_ref(msg);
+	rc = virtsnd_ctl_msg_send(snd, msg, false);
+	if (!rc && info_size)
+		memcpy(info, virtsnd_ctl_msg_response_data(msg), info_size);
+	virtsnd_ctl_msg_unref(msg);
+
+	return rc;
 }
 
 /**
diff --git a/sound/virtio/virtio_ctl_msg.h b/sound/virtio/virtio_ctl_msg.h
--- a/sound/virtio/virtio_ctl_msg.h
+++ b/sound/virtio/virtio_ctl_msg.h
@@ -21,20 +21,28 @@
 void *virtsnd_ctl_msg_response(struct virtio_snd_msg *msg);
 
 struct virtio_snd_msg *virtsnd_ctl_msg_alloc(size_t request_size,
-					     size_t response_size, gfp_t gfp);
+					     size_t response_size,
+					     size_t request_data_size,
+					     size_t response_data_size,
+					     gfp_t gfp);
+
+void *virtsnd_ctl_msg_request_data(struct virtio_snd_msg *msg);
+
+void *virtsnd_ctl_msg_response_data(struct virtio_snd_msg *msg);
 
 int virtsnd_ctl_msg_send(struct virtio_snd *snd, struct virtio_snd_msg *msg,
-			 struct scatterlist *out_sgs,
-			 struct scatterlist *in_sgs, bool nowait);
+			 bool nowait);
 
 /**
  * virtsnd_ctl_msg_send_sync() - Simplified sending of synchronous message.
  * @snd: VirtIO sound device.
  * @msg: Control message.
  *
- * After returning from this function, the message will be deleted. If message
- * content is still needed, the caller must additionally to
- * virtsnd_ctl_msg_ref/unref() it.
+ * This function consumes the caller's message reference on return. A message
+ * accepted by the control queue keeps its queue reference until completion or
+ * post-reset cleanup, including when this function returns on timeout or
+ * interruption. Callers that read the response after success must retain an
+ * additional reference through that access.
  *
  * The msg_timeout_ms module parameter defines the message completion timeout.
  * If the message is not completed within this time, the function will return an
@@ -49,7 +57,7 @@
 static inline int virtsnd_ctl_msg_send_sync(struct virtio_snd *snd,
 					    struct virtio_snd_msg *msg)
 {
-	return virtsnd_ctl_msg_send(snd, msg, NULL, NULL, false);
+	return virtsnd_ctl_msg_send(snd, msg, false);
 }
 
 /**
@@ -63,7 +71,7 @@
 static inline int virtsnd_ctl_msg_send_async(struct virtio_snd *snd,
 					     struct virtio_snd_msg *msg)
 {
-	return virtsnd_ctl_msg_send(snd, msg, NULL, NULL, true);
+	return virtsnd_ctl_msg_send(snd, msg, true);
 }
 
 void virtsnd_ctl_msg_cancel_all(struct virtio_snd *snd);
diff --git a/sound/virtio/virtio_kctl.c b/sound/virtio/virtio_kctl.c
--- a/sound/virtio/virtio_kctl.c
+++ b/sound/virtio/virtio_kctl.c
@@ -3,8 +3,10 @@
  * virtio-snd: Virtio sound device
  * Copyright (C) 2022 OpenSynergy GmbH
  */
+#include <linux/limits.h>
+#include <linux/overflow.h>
+#include <linux/virtio_config.h>
 #include <sound/control.h>
-#include <linux/virtio_config.h>
 
 #include "virtio_card.h"
 
@@ -157,14 +159,16 @@
 	struct virtio_snd_ctl_hdr *hdr;
 	struct virtio_snd_ctl_value *kvalue;
 	size_t request_size = sizeof(*hdr);
-	size_t response_size = sizeof(struct virtio_snd_hdr) + sizeof(*kvalue);
+	size_t response_size = sizeof(struct virtio_snd_hdr);
 	unsigned int i;
 	int rc;
 
-	msg = virtsnd_ctl_msg_alloc(request_size, response_size, GFP_KERNEL);
+	msg = virtsnd_ctl_msg_alloc(request_size, response_size, 0,
+				    sizeof(*kvalue), GFP_KERNEL);
 	if (!msg)
 		return -ENOMEM;
 
+	/* Response lifetime through ALSA value copyout. */
 	virtsnd_ctl_msg_ref(msg);
 
 	hdr = virtsnd_ctl_msg_request(msg);
@@ -175,8 +179,7 @@
 	if (rc)
 		goto on_failure;
 
-	kvalue = (void *)((u8 *)virtsnd_ctl_msg_response(msg) +
-			  sizeof(struct virtio_snd_hdr));
+	kvalue = virtsnd_ctl_msg_response_data(msg);
 
 	switch (type) {
 	case VIRTIO_SND_CTL_TYPE_BOOLEAN:
@@ -233,7 +236,8 @@
 	size_t response_size = sizeof(struct virtio_snd_hdr);
 	unsigned int i;
 
-	msg = virtsnd_ctl_msg_alloc(request_size, response_size, GFP_KERNEL);
+	msg = virtsnd_ctl_msg_alloc(request_size, response_size, 0, 0,
+				    GFP_KERNEL);
 	if (!msg)
 		return -ENOMEM;
 
@@ -288,22 +292,27 @@
 	struct virtio_snd *snd = snd_kcontrol_chip(kcontrol);
 	struct virtio_snd_msg *msg;
 	struct virtio_snd_ctl_hdr *hdr;
-	unsigned int *tlv;
-	struct scatterlist sg;
+	size_t request_data_size = 0;
+	size_t response_data_size = 0;
 	int rc;
 
+	switch (op_flag) {
+	case SNDRV_CTL_TLV_OP_READ:
+		response_data_size = size;
+		break;
+	case SNDRV_CTL_TLV_OP_WRITE:
+	case SNDRV_CTL_TLV_OP_CMD:
+		request_data_size = size;
+		break;
+	default:
+		return -EINVAL;
+	}
+
 	msg = virtsnd_ctl_msg_alloc(sizeof(*hdr), sizeof(struct virtio_snd_hdr),
+				    request_data_size, response_data_size,
 				    GFP_KERNEL);
 	if (!msg)
 		return -ENOMEM;
-
-	tlv = kzalloc(size, GFP_KERNEL);
-	if (!tlv) {
-		rc = -ENOMEM;
-		goto on_msg_unref;
-	}
-
-	sg_init_one(&sg, tlv, size);
 
 	hdr = virtsnd_ctl_msg_request(msg);
 	hdr->control_id = cpu_to_le32(kcontrol->private_value);
@@ -312,13 +321,16 @@
 	case SNDRV_CTL_TLV_OP_READ:
 		hdr->hdr.code = cpu_to_le32(VIRTIO_SND_R_CTL_TLV_READ);
 
-		rc = virtsnd_ctl_msg_send(snd, msg, NULL, &sg, false);
-		if (!rc) {
-			if (copy_to_user(utlv, tlv, size))
-				rc = -EFAULT;
-		}
-
-		break;
+		/* Response lifetime through user copyout. */
+		virtsnd_ctl_msg_ref(msg);
+		rc = virtsnd_ctl_msg_send(snd, msg, false);
+		if (!rc && size &&
+		    copy_to_user(utlv, virtsnd_ctl_msg_response_data(msg),
+				 size))
+			rc = -EFAULT;
+		virtsnd_ctl_msg_unref(msg);
+
+		return rc;
 	case SNDRV_CTL_TLV_OP_WRITE:
 	case SNDRV_CTL_TLV_OP_CMD:
 		if (op_flag == SNDRV_CTL_TLV_OP_WRITE)
@@ -327,28 +339,16 @@
 			hdr->hdr.code =
 				cpu_to_le32(VIRTIO_SND_R_CTL_TLV_COMMAND);
 
-		if (copy_from_user(tlv, utlv, size)) {
-			rc = -EFAULT;
-			goto on_msg_unref;
-		} else {
-			rc = virtsnd_ctl_msg_send(snd, msg, &sg, NULL, false);
+		if (size && copy_from_user(virtsnd_ctl_msg_request_data(msg), utlv,
+					   size)) {
+			virtsnd_ctl_msg_unref(msg);
+			return -EFAULT;
 		}
 
-		break;
-	default:
-		rc = -EINVAL;
-		/* We never get here - we listed all values for op_flag */
-		WARN_ON(1);
-		goto on_msg_unref;
-	}
-	kfree(tlv);
-	return rc;
-
-on_msg_unref:
-	virtsnd_ctl_msg_unref(msg);
-	kfree(tlv);
-
-	return rc;
+		return virtsnd_ctl_msg_send(snd, msg, false);
+	}
+
+	return -EINVAL;
 }
 
 /**
@@ -369,27 +369,36 @@
 	struct virtio_snd_msg *msg;
 	struct virtio_snd_ctl_hdr *hdr;
 	unsigned int n = le32_to_cpu(kinfo->value.enumerated.items);
-	struct scatterlist sg;
-
-	msg = virtsnd_ctl_msg_alloc(sizeof(*hdr),
-				    sizeof(struct virtio_snd_hdr), GFP_KERNEL);
-	if (!msg)
-		return -ENOMEM;
+	size_t items_size;
+	int rc;
+
+	if (check_mul_overflow((size_t)n, sizeof(*kctl->items), &items_size) ||
+	    items_size > UINT_MAX)
+		return -EOVERFLOW;
 
 	kctl->items = devm_kcalloc(&vdev->dev, n, sizeof(*kctl->items),
 				   GFP_KERNEL);
-	if (!kctl->items) {
-		virtsnd_ctl_msg_unref(msg);
-		return -ENOMEM;
-	}
-
-	sg_init_one(&sg, kctl->items, n * sizeof(*kctl->items));
+	if (!kctl->items)
+		return -ENOMEM;
+
+	msg = virtsnd_ctl_msg_alloc(sizeof(*hdr),
+				    sizeof(struct virtio_snd_hdr), 0, items_size,
+				    GFP_KERNEL);
+	if (!msg)
+		return -ENOMEM;
 
 	hdr = virtsnd_ctl_msg_request(msg);
 	hdr->hdr.code = cpu_to_le32(VIRTIO_SND_R_CTL_ENUM_ITEMS);
 	hdr->control_id = cpu_to_le32(cid);
 
-	return virtsnd_ctl_msg_send(snd, msg, NULL, &sg, false);
+	/* Item-buffer lifetime through the control-list copyout. */
+	virtsnd_ctl_msg_ref(msg);
+	rc = virtsnd_ctl_msg_send(snd, msg, false);
+	if (!rc && items_size)
+		memcpy(kctl->items, virtsnd_ctl_msg_response_data(msg), items_size);
+	virtsnd_ctl_msg_unref(msg);
+
+	return rc;
 }
 
 /**
diff --git a/sound/virtio/virtio_pcm_msg.c b/sound/virtio/virtio_pcm_msg.c
--- a/sound/virtio/virtio_pcm_msg.c
+++ b/sound/virtio/virtio_pcm_msg.c
@@ -684,7 +684,7 @@
 		break;
 	}
 
-	msg = virtsnd_ctl_msg_alloc(request_size, response_size, gfp);
+	msg = virtsnd_ctl_msg_alloc(request_size, response_size, 0, 0, gfp);
 	if (msg) {
 		struct virtio_snd_pcm_hdr *hdr = virtsnd_ctl_msg_request(msg);
 

[PATCH v12 06/12] ALSA: virtio: wait through signals for PCM teardown

[Joonho Lee]

A signal can end the interruptible wait while a STOP or RELEASE request is
still pending. A later teardown may then submit another RELEASE before the
first request completes.

Keep both waits non-interruptible and use the existing message timeout.
Other control requests remain interruptible. Device reset still cancels the
pending request and wakes the waiter.

Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_ctl_msg.c b/sound/virtio/virtio_ctl_msg.c
--- a/sound/virtio/virtio_ctl_msg.c
+++ b/sound/virtio/virtio_ctl_msg.c
@@ -314,7 +314,20 @@
 		 */
 		virtsnd_ctl_msg_unref(msg);
 	} else if (!nowait) {
-		rc = wait_for_completion_interruptible_timeout(&msg->notify, js);
+		/*
+		 * Teardown must observe completion even with a pending signal.
+		 * Otherwise a subsequent open can retry RELEASE while the device
+		 * is still processing the previous teardown request.
+		 */
+		switch (le32_to_cpu(request->code)) {
+		case VIRTIO_SND_R_PCM_STOP:
+		case VIRTIO_SND_R_PCM_RELEASE:
+			rc = wait_for_completion_timeout(&msg->notify, js);
+			break;
+		default:
+			rc = wait_for_completion_interruptible_timeout(&msg->notify, js);
+			break;
+		}
 		if (!rc) {
 			dev_err(&vdev->dev,
 				"control message (0x%08x) timeout\n",
diff --git a/sound/virtio/virtio_ctl_msg.h b/sound/virtio/virtio_ctl_msg.h
--- a/sound/virtio/virtio_ctl_msg.h
+++ b/sound/virtio/virtio_ctl_msg.h
@@ -47,6 +47,9 @@
  * The msg_timeout_ms module parameter defines the message completion timeout.
  * If the message is not completed within this time, the function will return an
  * error.
+ * PCM STOP and RELEASE wait through pending signals within this timeout.
+ * Other requests remain interruptible. Device reset cancels and wakes either
+ * kind of waiter.
  *
  * Context: Any context that permits to sleep.
  * Return: 0 on success, -errno on failure.

[PATCH v12 07/12] ALSA: pcm: mark stop-threshold accesses for capture policy sampling

[Joonho Lee]

The staged virtio-snd capture path needs to sample stop_threshold from
completion context to distinguish free-running capture from bounded
capture. Software parameters can update this field while capture runs.

Mark the runtime stop-threshold assignments with WRITE_ONCE and the proc
status read with READ_ONCE. This pairs the assignments with the driver's
READ_ONCE policy sampling in patch 9.

These annotations do not provide a transaction for the entire sw_params
update. The driver serializes its cached policy and accounting with its
own substream lock.

Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/core/oss/pcm_oss.c b/sound/core/oss/pcm_oss.c
--- a/sound/core/oss/pcm_oss.c
+++ b/sound/core/oss/pcm_oss.c
@@ -2958,7 +2958,7 @@
 	       runtime->oss.mmap_bytes);
 #endif
 	/* In mmap mode we never stop */
-	runtime->stop_threshold = runtime->boundary;
+	WRITE_ONCE(runtime->stop_threshold, runtime->boundary);
 
 	return 0;
 }
diff --git a/sound/core/pcm.c b/sound/core/pcm.c
--- a/sound/core/pcm.c
+++ b/sound/core/pcm.c
@@ -422,7 +422,8 @@
 	snd_iprintf(buffer, "period_step: %u\n", runtime->period_step);
 	snd_iprintf(buffer, "avail_min: %lu\n", runtime->control->avail_min);
 	snd_iprintf(buffer, "start_threshold: %lu\n", runtime->start_threshold);
-	snd_iprintf(buffer, "stop_threshold: %lu\n", runtime->stop_threshold);
+	snd_iprintf(buffer, "stop_threshold: %lu\n",
+		    READ_ONCE(runtime->stop_threshold));
 	snd_iprintf(buffer, "silence_threshold: %lu\n", runtime->silence_threshold);
 	snd_iprintf(buffer, "silence_size: %lu\n", runtime->silence_size);
 	snd_iprintf(buffer, "boundary: %lu\n", runtime->boundary);
diff --git a/sound/core/pcm_native.c b/sound/core/pcm_native.c
--- a/sound/core/pcm_native.c
+++ b/sound/core/pcm_native.c
@@ -871,7 +871,7 @@
 	runtime->period_step = 1;
 	runtime->control->avail_min = runtime->period_size;
 	runtime->start_threshold = 1;
-	runtime->stop_threshold = runtime->buffer_size;
+	WRITE_ONCE(runtime->stop_threshold, runtime->buffer_size);
 	runtime->silence_threshold = 0;
 	runtime->silence_size = 0;
 	runtime->boundary = runtime->buffer_size;
@@ -1062,7 +1062,8 @@
 		runtime->period_step = params->period_step;
 		runtime->control->avail_min = params->avail_min;
 		runtime->start_threshold = params->start_threshold;
-		runtime->stop_threshold = params->stop_threshold;
+		/* Drivers may sample the stop policy from completion context. */
+		WRITE_ONCE(runtime->stop_threshold, params->stop_threshold);
 		runtime->silence_threshold = params->silence_threshold;
 		runtime->silence_size = params->silence_size;
 		params->boundary = runtime->boundary;

[PATCH v12 08/12] ALSA: pcm: report partial progress under the stream lock

[Joonho Lee]

Staged capture can make readable data available before a complete period
has elapsed. Skipping notification leaves a blocked reader asleep, while
using the period-interrupt path for this event can infer a false ring wrap
when the notification is delayed.

Add snd_pcm_progress_under_stream_lock() to update the hardware pointer
through the non-interrupt path and notify readers and fasync listeners.
It runs under the caller's PCM stream lock and does not tick the PCM timer.
The driver must release its pointer and trigger locks before calling it.
Existing period notification callers keep their current behavior.

Atomic and nonatomic synthetic PCM tests reproduced the missing wakeup
and false-wrap alternatives. The new helper delivered 64 final frames and
reported four available frames after delayed partial progress without XRUN.

Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/include/sound/pcm.h b/include/sound/pcm.h
--- a/include/sound/pcm.h
+++ b/include/sound/pcm.h
@@ -1182,6 +1182,7 @@
 int snd_pcm_lib_ioctl(struct snd_pcm_substream *substream,
 		      unsigned int cmd, void *arg);                      
 void snd_pcm_period_elapsed_under_stream_lock(struct snd_pcm_substream *substream);
+void snd_pcm_progress_under_stream_lock(struct snd_pcm_substream *substream);
 void snd_pcm_period_elapsed(struct snd_pcm_substream *substream);
 snd_pcm_sframes_t __snd_pcm_lib_xfer(struct snd_pcm_substream *substream,
 				     void *buf, bool interleaved,
diff --git a/sound/core/pcm_lib.c b/sound/core/pcm_lib.c
--- a/sound/core/pcm_lib.c
+++ b/sound/core/pcm_lib.c
@@ -1919,6 +1919,25 @@
 EXPORT_SYMBOL(snd_pcm_period_elapsed_under_stream_lock);
 
 /**
+ * snd_pcm_progress_under_stream_lock - Report progress without a period IRQ
+ * @substream: PCM substream
+ *
+ * Context: PCM stream lock held, with no driver pointer or trigger locks held.
+ * Updates the pointer and wakes readers for partial-period transfers. Unlike
+ * a period notification, this does not imply an elapsed period or tick the PCM
+ * timer. The same pointer, trigger and timestamp callbacks may be invoked.
+ */
+void snd_pcm_progress_under_stream_lock(struct snd_pcm_substream *substream)
+{
+	if (PCM_RUNTIME_CHECK(substream) || !snd_pcm_running(substream))
+		return;
+	snd_pcm_update_hw_ptr(substream);
+	snd_kill_fasync(substream->runtime->fasync, SIGIO, POLL_IN);
+}
+EXPORT_SYMBOL_GPL(snd_pcm_progress_under_stream_lock);
+
+
+/**
  * snd_pcm_period_elapsed() - update the status of runtime for the next period by acquiring lock of
  *			      PCM substream.
  * @substream: the instance of PCM substream.

[PATCH v12 09/12] ALSA: pcm: add an optional reset operation under the stream lock

[Joonho Lee]

A driver may need to synchronize private capture accounting with the
common RESET pointer observation and the core's position rebase.

Add an optional reset operation called under the member's stream lock.
Expose snd_pcm_lib_reset() so the callback can retain common pointer and
time handling. On zero return the core completes that member's rebase
before unlocking it. Drivers without the callback retain the ioctl path.
The operation is also used after a successful prepare.

Zero permits the position rebase; it does not promise XRUN recovery.
A negative return skips the core rebase, so a callback returning an error
must leave coherent pointer state even after partial work.

Linked RESET and prepare stop at the first error without rollback.
Earlier callback members may already have completed. Legacy members keep
the existing post_action policy. Applications must recover the group
after partial failure before relying on shared positions.

This adds an internal kernel operation and no userspace ioctl or layout.
Out-of-tree modules using snd_pcm_ops need to be rebuilt.

Complete appl_ptr rebasing once in the callback path. A later linked
member can sleep in prepare while the earlier member accepts SYNC_PTR;
a subsequent unlocked post_prepare assignment would overwrite that
acknowledgment. Keep explicit RESET playback silence processing and
leave prepare-time filling to the normal START path, preserving samples
written and rewound after prepare.

Atomic and nonatomic synthetic-driver tests exercised both callback and
legacy members, mixed linked groups and later-member failures. The
callback retained an acknowledged appl_ptr of 64, and finite-silence
playback kept the written prefix while clearing the rewound tail.

Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/include/sound/pcm.h b/include/sound/pcm.h
--- a/include/sound/pcm.h
+++ b/include/sound/pcm.h
@@ -77,6 +77,26 @@
 			     unsigned long offset);
 	int (*mmap)(struct snd_pcm_substream *substream, struct vm_area_struct *vma);
 	int (*ack)(struct snd_pcm_substream *substream);
+	/*
+	 * Optional RESET operation, also used after a successful prepare.
+	 * Called with this substream's PCM lock held; atomic streams must not
+	 * sleep. Do not reacquire the PCM lock through snd_pcm_lib_ioctl().
+	 * snd_pcm_lib_reset() provides the common pointer update under this
+	 * lock. The core completes the position and appl_ptr rebase before
+	 * unlocking a successful member.
+	 *
+	 * Return zero to let the core complete the position rebase. As with
+	 * the legacy RESET ioctl, zero does not guarantee recovery from XRUN;
+	 * callers must still handle the stream state and subsequent I/O errors.
+	 * On a negative return the core does not finish the rebase, so the
+	 * callback must leave coherent pointer state even after partial work.
+	 *
+	 * Linked RESET/prepare
+	 * stops at the first error without rollback. Earlier callback members
+	 * retain their completed reset; legacy members retain the existing
+	 * post_action policy. An error does not guarantee unchanged state.
+	 */
+	int (*reset)(struct snd_pcm_substream *substream);
 };
 
 /*
@@ -1179,6 +1199,7 @@
 {
 	substream->runtime->std_sync_id = true;
 }
+int snd_pcm_lib_reset(struct snd_pcm_substream *substream);
 int snd_pcm_lib_ioctl(struct snd_pcm_substream *substream,
 		      unsigned int cmd, void *arg);                      
 void snd_pcm_period_elapsed_under_stream_lock(struct snd_pcm_substream *substream);
diff --git a/sound/core/pcm_lib.c b/sound/core/pcm_lib.c
--- a/sound/core/pcm_lib.c
+++ b/sound/core/pcm_lib.c
@@ -1760,20 +1760,39 @@
 }
 EXPORT_SYMBOL(snd_pcm_hw_params_bits);
 
-static int snd_pcm_lib_ioctl_reset(struct snd_pcm_substream *substream,
-				   void *arg)
+/**
+ * snd_pcm_lib_reset - Reset the pointer while holding the PCM stream lock
+ * @substream: PCM substream
+ *
+ * Context: PCM stream lock held. The caller must not hold driver locks
+ * acquired by the pointer or XRUN callbacks.
+ *
+ * Update a running stream through the ordinary pointer path. Other states
+ * start at zero. Pointer errors keep the existing RESET return semantics;
+ * a zero return does not imply recovery from XRUN.
+ * This helper does not update appl_ptr or acquire the stream lock.
+ *
+ * Return: zero.
+ */
+int snd_pcm_lib_reset(struct snd_pcm_substream *substream)
 {
 	struct snd_pcm_runtime *runtime = substream->runtime;
 
-	guard(pcm_stream_lock_irqsave)(substream);
-	if (snd_pcm_running(substream) &&
-	    snd_pcm_update_hw_ptr(substream) >= 0)
+	if (snd_pcm_running(substream) && snd_pcm_update_hw_ptr(substream) >= 0)
 		runtime->status->hw_ptr %= runtime->buffer_size;
 	else {
 		runtime->status->hw_ptr = 0;
 		runtime->hw_ptr_wrap = 0;
 	}
 	return 0;
+}
+EXPORT_SYMBOL(snd_pcm_lib_reset);
+
+static int snd_pcm_lib_ioctl_reset(struct snd_pcm_substream *substream,
+				   void *arg)
+{
+	guard(pcm_stream_lock_irqsave)(substream);
+	return snd_pcm_lib_reset(substream);
 }
 
 static int snd_pcm_lib_ioctl_channel_info(struct snd_pcm_substream *substream,
diff --git a/sound/core/pcm_native.c b/sound/core/pcm_native.c
--- a/sound/core/pcm_native.c
+++ b/sound/core/pcm_native.c
@@ -1984,11 +1984,35 @@
 	}
 }
 
-static int snd_pcm_do_reset(struct snd_pcm_substream *substream,
-			    snd_pcm_state_t state)
+static int snd_pcm_reset_substream(struct snd_pcm_substream *substream,
+				   bool for_prepare)
 {
 	struct snd_pcm_runtime *runtime = substream->runtime;
-	int err = snd_pcm_ops_ioctl(substream, SNDRV_PCM_IOCTL1_RESET, NULL);
+	int err;
+
+	if (substream->ops->reset) {
+		/*
+		 * Complete this member before releasing its PCM lock. A later
+		 * linked member may fail; RESET has no group rollback operation.
+		 */
+		guard(pcm_stream_lock_irq)(substream);
+
+		err = substream->ops->reset(substream);
+		if (err < 0)
+			return err;
+		runtime->hw_ptr_base = 0;
+		runtime->hw_ptr_interrupt = runtime->status->hw_ptr -
+			runtime->status->hw_ptr % runtime->period_size;
+		runtime->silence_start = runtime->status->hw_ptr;
+		runtime->silence_filled = 0;
+		runtime->control->appl_ptr = runtime->status->hw_ptr;
+		if (!for_prepare && substream->stream == SNDRV_PCM_STREAM_PLAYBACK &&
+		    runtime->silence_size > 0)
+			snd_pcm_playback_silence(substream, ULONG_MAX);
+		return 0;
+	}
+
+	err = snd_pcm_ops_ioctl(substream, SNDRV_PCM_IOCTL1_RESET, NULL);
 	if (err < 0)
 		return err;
 	guard(pcm_stream_lock_irq)(substream);
@@ -2000,10 +2024,19 @@
 	return 0;
 }
 
+static int snd_pcm_do_reset(struct snd_pcm_substream *substream,
+			    snd_pcm_state_t state)
+{
+	return snd_pcm_reset_substream(substream, false);
+}
+
 static void snd_pcm_post_reset(struct snd_pcm_substream *substream,
 			       snd_pcm_state_t state)
 {
 	struct snd_pcm_runtime *runtime = substream->runtime;
+
+	if (substream->ops->reset)
+		return;
 	guard(pcm_stream_lock_irq)(substream);
 	runtime->control->appl_ptr = runtime->status->hw_ptr;
 	if (substream->stream == SNDRV_PCM_STREAM_PLAYBACK &&
@@ -2053,14 +2086,16 @@
 	if (err < 0)
 		return err;
 	pcm_qos_req_update(substream, period_to_usecs(substream->runtime));
-	return snd_pcm_do_reset(substream, state);
+	return snd_pcm_reset_substream(substream, true);
 }
 
 static void snd_pcm_post_prepare(struct snd_pcm_substream *substream,
 				 snd_pcm_state_t state)
 {
 	struct snd_pcm_runtime *runtime = substream->runtime;
-	runtime->control->appl_ptr = runtime->status->hw_ptr;
+	/* The reset callback path already completed this rebase under the lock. */
+	if (!substream->ops->reset)
+		runtime->control->appl_ptr = runtime->status->hw_ptr;
 	snd_pcm_set_state(substream, SNDRV_PCM_STATE_PREPARED);
 }
 

[PATCH v12 10/12] ALSA: virtio: add staged capture and serialize PCM lifecycle accounting

[Joonho Lee]

Staged capture needs request-owned DMA storage and separate produced,
observed and consumed positions when shared clients use a boundary stop
threshold. Synchronize policy, RX completion and acknowledgment under
the queue and substream locks, with lifecycle_lock covering control work.

The v11 ACK path tried buffer_mutex while holding the PCM stream lock.
Core RESET, parameter refinement and linked prepare can already hold that
mutex, so normal ACK returned -EBADFD. Serialize ACK and prepare accounting
through the PCM lock instead. Set a preparing gate before dropping that
lock for PREPARE; clear it in the RESET callback while the core keeps the
same lock through the public appl_ptr rebase. A later linked member can
sleep in prepare without exposing half-rebased private accounting.

Use the preceding optional RESET operation for playback and both capture
modes. Retain common pointer/time handling and rebase private appl_ptr
with the public pointer. Preserve queued TX ownership on playback RESET.
Reset staged accounting and ordinary capture sw_ready after a successful
prepare, including repeated RESET from PREPARED.

Native pause can return RX buffers while transfers are disabled. Accept
those completions for a live prepared capture epoch, and refill only when
transfers are enabled. Keep the period notification floor across resume;
report recovery progress without inventing PCM timer ticks. Latch XRUN
on device events and failed STOP, reject ACK from invalid or stopped
states, and keep successful prepared accounting until PREPARE succeeds.
Compatibility pause retains its released-epoch handling.

Validate ordinary RX status and payload before advancing counters. Empty
STOP completions carry no frames. Preserve full-ring XRUN protection:
a delayed observation can mean real DMA overwrite. Recovering that case
by exposing the modulo pointer returned an old ring of samples in tests.

Use gcd(frame_bits, 8) for byte alignment so packed 4-bit frames retain
byte-aligned queue boundaries. Keep the queue-to-substream lock order.
The compatibility options remain disabled by default.

ARM64 QEMU tests reproduced ACK contention, failed prepare and STOP,
paused RX, status-only RX and reset recovery. The corrected source passed
56 KASAN/KCSAN transition checks and four linked prepare/ACK checks.
RW and MMAP capture, RESET, pause/resume and reopen produced 14,208,000
oracle-matched samples. Physical VZ capture remains untested for this
revision. Sample tests cover mono S16 at 48 kHz.

Link: https://sashiko.dev/#/patchset/20261031020855.2402-1-mynameisjune111%40gmail.com
Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_card.c b/sound/virtio/virtio_card.c
--- a/sound/virtio/virtio_card.c
+++ b/sound/virtio/virtio_card.c
@@ -350,6 +350,22 @@
 	struct virtio_device *vdev = snd->vdev;
 	unsigned int i;
 
+	for (i = 0; snd->substreams && i < snd->nsubstreams; ++i) {
+		struct virtio_pcm_substream *vss = &snd->substreams[i];
+		struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+
+		guard(spinlock_irqsave)(&queue->lock);
+		guard(spinlock)(&vss->lock);
+		vss->reset_epoch++;
+		vss->resetting = true;
+		vss->prepared = false;
+		vss->device_running = false;
+		vss->release_required = false;
+		vss->control_uncertain = false;
+		vss->capture_reset_observing = false;
+		vss->xfer_enabled = false;
+		vss->suspended = true;
+	}
 	virtio_reset_device(vdev);
 
 	/* Deleting a virtqueue does not unmap its outstanding descriptors. */
@@ -367,8 +383,16 @@
 		queue->vqueue = NULL;
 		spin_unlock_irqrestore(&queue->lock, flags);
 	}
-
 	vdev->config->del_vqs(vdev);
+	for (i = 0; snd->substreams && i < snd->nsubstreams; ++i) {
+		struct virtio_pcm_substream *vss = &snd->substreams[i];
+
+		cancel_work_sync(&vss->elapsed_period);
+		guard(spinlock_irqsave)(&vss->lock);
+		vss->stopped = false;
+		vss->xfer_xrun = false;
+		vss->pause_restart_state = VIRTSND_PCM_PAUSE_RESTART_NONE;
+	}
 }
 
 /**
@@ -441,6 +465,7 @@
 {
 	struct virtio_snd *snd = vdev->priv;
 	int rc;
+	unsigned int i;
 
 	rc = virtsnd_find_vqs(snd);
 	if (rc)
@@ -449,6 +474,12 @@
 	virtio_device_ready(vdev);
 	virtsnd_ctl_msg_enable(snd);
 
+	for (i = 0; i < snd->nsubstreams; ++i) {
+		struct virtio_pcm_substream *vss = &snd->substreams[i];
+
+		guard(spinlock_irqsave)(&vss->lock);
+		vss->resetting = false;
+	}
 	virtsnd_enable_event_vq(snd);
 
 	return 0;
diff --git a/sound/virtio/virtio_ctl_msg.c b/sound/virtio/virtio_ctl_msg.c
--- a/sound/virtio/virtio_ctl_msg.c
+++ b/sound/virtio/virtio_ctl_msg.c
@@ -23,6 +23,9 @@
  * @list: Pending message list entry.
  * @notify: Request completed notification.
  * @ref_count: Reference count used to manage a message lifetime.
+ * @pcm: Optional PCM generation owner, valid until control queue teardown.
+ * @pcm_epoch: Operation generation captured before request allocation.
+ * @submitted: Request was handed to the device.
  * @cancelled: Request canceled during device teardown.
  */
 struct virtio_snd_msg {
@@ -37,7 +40,49 @@
 	struct completion notify;
 	refcount_t ref_count;
 	bool cancelled;
+	bool submitted;
+	struct virtio_pcm_substream *pcm;
+	u64 pcm_epoch;
 };
+
+/* Bind an operation's original generation before publishing the request. */
+void virtsnd_ctl_msg_bind_pcm(struct virtio_snd_msg *msg,
+			     struct virtio_pcm_substream *vss, u64 epoch)
+{
+	msg->pcm = vss;
+	msg->pcm_epoch = epoch;
+}
+
+u64 virtsnd_ctl_msg_pcm_epoch(struct virtio_snd_msg *msg)
+{
+	return msg->pcm_epoch;
+}
+
+static int virtsnd_ctl_msg_enqueue(struct virtio_snd *snd,
+				   struct virtio_snd_msg *msg,
+				   struct scatterlist **sgs,
+				   unsigned int nouts, unsigned int nins)
+{
+	struct virtqueue *vq = virtsnd_control_queue(snd)->vqueue;
+
+	if (!msg->pcm)
+		return virtqueue_add_sgs(vq, sgs, nouts, nins, msg, GFP_ATOMIC);
+
+	/* Control queue -> PCM lock serializes enqueue against device reset. */
+	guard(spinlock)(&msg->pcm->lock);
+	if (msg->pcm->resetting || msg->pcm->reset_epoch != msg->pcm_epoch)
+		return -ECANCELED;
+	return virtqueue_add_sgs(vq, sgs, nouts, nins, msg, GFP_ATOMIC);
+}
+
+/* Caller retains a reference across the synchronous send. */
+bool virtsnd_ctl_msg_uncertain(struct virtio_snd_msg *msg, int rc)
+{
+	return msg->submitted &&
+	       (rc == -ETIMEDOUT || rc == -ERESTARTSYS ||
+		READ_ONCE(msg->cancelled));
+}
+
 
 /**
  * virtsnd_ctl_msg_ref() - Increment reference counter for the message.
@@ -293,9 +338,9 @@
 	if (snd->ctl_msgs_blocked)
 		rc = -ENODEV;
 	else
-		rc = virtqueue_add_sgs(queue->vqueue, psgs, nouts, nins, msg,
-				       GFP_ATOMIC);
+		rc = virtsnd_ctl_msg_enqueue(snd, msg, psgs, nouts, nins);
 	if (!rc) {
+		msg->submitted = true;
 		notify = virtqueue_kick_prepare(queue->vqueue);
 
 		list_add_tail(&msg->list, &snd->ctl_msgs);
@@ -315,11 +360,14 @@
 		virtsnd_ctl_msg_unref(msg);
 	} else if (!nowait) {
 		/*
-		 * Teardown must observe completion even with a pending signal.
-		 * Otherwise a subsequent open can retry RELEASE while the device
-		 * is still processing the previous teardown request.
+		 * PCM state changes must observe completion with a pending signal.
+		 * Otherwise recovery can issue a conflicting command while the
+		 * device is still processing the previous transition.
 		 */
 		switch (le32_to_cpu(request->code)) {
+		case VIRTIO_SND_R_PCM_SET_PARAMS:
+		case VIRTIO_SND_R_PCM_PREPARE:
+		case VIRTIO_SND_R_PCM_START:
 		case VIRTIO_SND_R_PCM_STOP:
 		case VIRTIO_SND_R_PCM_RELEASE:
 			rc = wait_for_completion_timeout(&msg->notify, js);
diff --git a/sound/virtio/virtio_ctl_msg.h b/sound/virtio/virtio_ctl_msg.h
--- a/sound/virtio/virtio_ctl_msg.h
+++ b/sound/virtio/virtio_ctl_msg.h
@@ -11,6 +11,13 @@
 
 struct virtio_snd;
 struct virtio_snd_msg;
+struct virtio_pcm_substream;
+
+void virtsnd_ctl_msg_bind_pcm(struct virtio_snd_msg *msg,
+			     struct virtio_pcm_substream *vss, u64 epoch);
+u64 virtsnd_ctl_msg_pcm_epoch(struct virtio_snd_msg *msg);
+
+bool virtsnd_ctl_msg_uncertain(struct virtio_snd_msg *msg, int rc);
 
 void virtsnd_ctl_msg_ref(struct virtio_snd_msg *msg);
 
@@ -47,7 +54,7 @@
  * The msg_timeout_ms module parameter defines the message completion timeout.
  * If the message is not completed within this time, the function will return an
  * error.
- * PCM STOP and RELEASE wait through pending signals within this timeout.
+ * PCM state-changing requests wait through pending signals within this timeout.
  * Other requests remain interruptible. Device reset cancels and wakes either
  * kind of waiter.
  *
diff --git a/sound/virtio/virtio_pcm.c b/sound/virtio/virtio_pcm.c
--- a/sound/virtio/virtio_pcm.c
+++ b/sound/virtio/virtio_pcm.c
@@ -310,8 +310,119 @@
 {
 	struct virtio_pcm_substream *vss =
 		container_of(work, struct virtio_pcm_substream, elapsed_period);
-
-	snd_pcm_period_elapsed(vss->substream);
+	struct snd_pcm_substream *substream = vss->substream;
+	struct snd_pcm_runtime *runtime;
+	size_t period_bytes;
+	u64 period_floor;
+	u64 snapshot;
+	bool staged_capture;
+	bool paused_xrun = false;
+	bool notify = false;
+	bool partial = false;
+	bool requeue = false;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		staged_capture = vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+				 vss->capture_staging;
+	}
+
+	/* Serialize the pointer snapshot with ALSA's pointer and ACK callbacks. */
+	guard(pcm_stream_lock_irqsave)(substream);
+	if (!substream->runtime)
+		return;
+	if (!snd_pcm_running(substream)) {
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			runtime = substream->runtime;
+			paused_xrun = runtime &&
+				      runtime->state == SNDRV_PCM_STATE_PAUSED &&
+				      vss->xfer_xrun;
+			if (paused_xrun)
+				virtsnd_pcm_capture_arm_paused_stop(vss);
+		}
+		if (paused_xrun) {
+			/* Stop paused transfers and wake blocked readers. */
+			snd_pcm_stop(substream, SNDRV_PCM_STATE_XRUN);
+		}
+		return;
+	}
+
+	if (!staged_capture) {
+		if (vss->direction == SNDRV_PCM_STREAM_CAPTURE) {
+			guard(spinlock_irqsave)(&vss->lock);
+			period_bytes = snd_pcm_lib_period_bytes(substream);
+			period_floor = period_bytes ?
+				div64_ul(vss->capture_copied_total, period_bytes) : 0;
+			partial = vss->xfer_xrun ||
+				period_floor <= vss->capture_notified_periods;
+			if (!partial)
+				vss->capture_notified_periods = period_floor;
+		}
+		/* Resume recovery can wake readers without a new period IRQ. */
+		if (partial)
+			snd_pcm_progress_under_stream_lock(substream);
+		else
+			snd_pcm_period_elapsed_under_stream_lock(substream);
+		return;
+	}
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		runtime = substream->runtime;
+		if (!vss->runtime_active ||
+		    (!vss->xfer_enabled && !vss->xfer_xrun) || !runtime)
+			return;
+
+		period_bytes = snd_pcm_lib_period_bytes(substream);
+		if (vss->xfer_xrun) {
+			notify = true;
+		} else if (!period_bytes || period_bytes >= vss->buffer_bytes ||
+			   !virtsnd_pcm_capture_counters_valid(vss) ||
+			   vss->capture_notified_periods >
+				div64_ul(vss->capture_copied_total, period_bytes)) {
+			vss->xfer_xrun = true;
+			notify = true;
+		} else {
+			period_floor = div64_ul(vss->capture_copied_total, period_bytes);
+			if (period_floor > vss->capture_notified_periods) {
+				/* Reserve crossed boundaries before the callback. */
+				vss->capture_notified_periods = period_floor;
+				notify = true;
+			} else if (vss->capture_copied_total >
+				   vss->capture_pointer_seen_total) {
+				notify = true;
+				partial = true;
+			}
+		}
+
+		if (notify) {
+			snapshot = vss->capture_copied_total;
+			vss->capture_callback_snapshot = snapshot;
+			vss->capture_callback_snapshot_active = true;
+		}
+	}
+
+	if (!notify)
+		return;
+
+	/* Pin the callback position while completions keep updating actual totals. */
+	if (partial)
+		snd_pcm_progress_under_stream_lock(substream);
+	else
+		snd_pcm_period_elapsed_under_stream_lock(substream);
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		vss->capture_callback_snapshot_active = false;
+		vss->capture_callback_snapshot = 0;
+		if (vss->runtime_active && vss->xfer_enabled &&
+		    !vss->xfer_xrun && substream->runtime) {
+			period_bytes = snd_pcm_lib_period_bytes(substream);
+			if (period_bytes &&
+			    vss->capture_copied_total > vss->capture_pointer_seen_total)
+				requeue = true;
+		}
+	}
+
+	if (requeue)
+		virtsnd_pcm_schedule_period_elapsed(vss);
 }
 
 /**
@@ -351,8 +462,10 @@
 		vss->snd = snd;
 		vss->sid = i;
 		INIT_WORK(&vss->elapsed_period, virtsnd_pcm_period_elapsed);
+		disable_work(&vss->elapsed_period);
 		init_waitqueue_head(&vss->msg_empty);
 		spin_lock_init(&vss->lock);
+		mutex_init(&vss->lifecycle_lock);
 	}
 
 	info = kzalloc_objs(*info, snd->nsubstreams);
@@ -486,7 +599,7 @@
 			snd_pcm_set_ops(vpcm->pcm, i, &virtsnd_pcm_ops[i]);
 		}
 
-		snd_pcm_set_managed_buffer_all(vpcm->pcm,
+		snd_pcm_lib_preallocate_pages_for_all(vpcm->pcm,
 					       SNDRV_DMA_TYPE_VMALLOC, NULL,
 					       0, 0);
 	}
@@ -517,8 +630,13 @@
 		break;
 	case VIRTIO_SND_EVT_PCM_XRUN:
 		scoped_guard(spinlock, &vss->lock) {
-			if (vss->xfer_enabled)
+			if (vss->xfer_enabled ||
+			    (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+			     vss->prepared && !vss->stopped)) {
 				vss->xfer_xrun = true;
+				if (vss->direction == SNDRV_PCM_STREAM_CAPTURE)
+					virtsnd_pcm_schedule_period_elapsed(vss);
+			}
 		}
 		break;
 	}
diff --git a/sound/virtio/virtio_pcm.h b/sound/virtio/virtio_pcm.h
--- a/sound/virtio/virtio_pcm.h
+++ b/sound/virtio/virtio_pcm.h
@@ -7,13 +7,32 @@
 #define VIRTIO_SND_PCM_H
 
 #include <linux/atomic.h>
+#include <linux/mutex.h>
+#include <linux/math64.h>
 #include <linux/virtio_config.h>
+#include <linux/workqueue.h>
 #include <sound/pcm.h>
 #include <sound/pcm-indirect.h>
 
 struct virtio_pcm;
 struct virtio_pcm_msg;
 struct page;
+
+enum virtio_pcm_pause_restart_state {
+	VIRTSND_PCM_PAUSE_RESTART_NONE,
+	VIRTSND_PCM_PAUSE_RESTART_CONFIGURED,
+	VIRTSND_PCM_PAUSE_RESTART_SET_PARAMS_SUBMITTED,
+	VIRTSND_PCM_PAUSE_RESTART_PREPARE_SUBMITTED,
+	VIRTSND_PCM_PAUSE_RESTART_PREPARED,
+	VIRTSND_PCM_PAUSE_RESTART_START_SUBMITTED,
+	VIRTSND_PCM_PAUSE_RESTART_RUNNING,
+	VIRTSND_PCM_PAUSE_RESTART_STOP_SUBMITTED,
+	VIRTSND_PCM_PAUSE_RESTART_STOPPED,
+	VIRTSND_PCM_PAUSE_RESTART_RELEASE_SUBMITTED,
+	VIRTSND_PCM_PAUSE_RESTART_RELEASE_SENT,
+	VIRTSND_PCM_PAUSE_RESTART_RELEASED,
+	VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED,
+};
 
 /**
  * struct virtio_pcm_substream - VirtIO PCM substream.
@@ -26,17 +45,44 @@
  * @pcm_indirect: Kernel indirect pcm structure.
  * @hw: Kernel ALSA substream hardware descriptor.
  * @elapsed_period: Kernel work to handle the elapsed period state.
+ * @lifecycle_lock: Serializes prepare with synchronous release, outside trigger.
+ * @resetting: Device queues are unavailable until restore completes.
+ * @reset_epoch: Invalidates control replies from before device reset.
  * @lock: Spinlock that protects fields shared by interrupt handlers and
  *        substream operators, including @stopped.
  * @buffer_bytes: Current buffer size in bytes.
  * @hw_ptr: Substream hardware pointer value in bytes [0 ... buffer_bytes).
+ * @capture_ready: Managed-mode bytes copied but not yet acknowledged by ALSA;
+ *                 always zero in free-running mode.
+ * @capture_boundary: ALSA boundary captured for this prepared epoch.
+ * @capture_copied_total: Monotonic bytes copied into the capture ring.
+ * @capture_acked_total: Monotonic bytes acknowledged by the capture client.
+ * @capture_pointer_seen_total: Last capture position sampled by ALSA.
+ * @capture_notified_periods: Highest full-period boundary reserved for notify.
+ * @capture_callback_snapshot: Capture position fixed for the period callback.
+ * @capture_callback_snapshot_active: The callback must use its fixed position.
+ * @runtime_active: ALSA runtime buffer may be accessed by the staged path.
+ * @capture_staging: Capture uses request-owned payload buffers.
  * @capture_rx_seeded: Ordinary capture ring was queued since prepare.
+ * @capture_started: Non-empty capture data was received since prepare.
+ * @capture_empty: Consecutive empty startup completions.
+ * @capture_free_running: Cached staged-capture policy, protected by @lock.
  * @xfer_enabled: Data transfer state (0 - off, 1 - on).
  * @xfer_xrun: Data underflow/overflow state (0 - no xrun, 1 - xrun).
  * @stopped: True if the substream is stopped and must be released on the device
  *           side; protected by @lock.
+ * @prepared: Device PREPARE succeeded; protected by @lock.
+ * @preparing: ACK gate held through prepare's core pointer reset; protected by
+ *             the PCM stream lock and @lock.
+ * @device_running: Native START succeeded without a subsequent successful STOP.
+ * @release_required: Successful PREPARE still needs a successful RELEASE.
+ * @control_uncertain: Native control result requires device reset before reuse.
+ * @capture_reset_observing: Reclaim staged samples during RESET pointer capture.
+ * @indirect_reset_observing: Suppress ordinary ACK refill during RESET pointer
+ *                            capture; protected by the PCM stream lock.
  * @suspended: True if the substream is suspended and must be reconfigured on
  *             the device side at resume.
+ * @pause_restart_state: Device-side state for opt-in pause compatibility.
  * @msgs: Allocated I/O messages.
  * @nmsgs: Number of allocated I/O messages.
  * @msg_count: Number of pending I/O messages in the virtqueue.
@@ -57,15 +103,38 @@
 	struct snd_pcm_hardware hw;
 	struct work_struct elapsed_period;
 	spinlock_t lock;
+	struct mutex lifecycle_lock;
+	u64 reset_epoch;
+	bool resetting;
 	size_t buffer_bytes;
 	size_t hw_ptr;
+	size_t capture_ready;
+	snd_pcm_uframes_t capture_boundary;
+	u64 capture_copied_total;
+	u64 capture_acked_total;
+	u64 capture_pointer_seen_total;
+	u64 capture_notified_periods;
+	u64 capture_callback_snapshot;
+	bool capture_callback_snapshot_active;
+	bool runtime_active;
+	bool capture_staging;
+	bool capture_started;
 	bool capture_rx_seeded;
+	bool capture_free_running;
+	unsigned int capture_empty;
 	bool xfer_enabled;
 	bool xfer_xrun;
 	bool stopped;
-	/* Successful PREPARE still needs RELEASE; protected by lock. */
+	bool prepared;
+	bool preparing;
+	/* Native control state, protected by lock. */
+	bool device_running;
 	bool release_required;
+	bool control_uncertain;
+	bool capture_reset_observing;
+	bool indirect_reset_observing;
 	bool suspended;
+	enum virtio_pcm_pause_restart_state pause_restart_state;
 	struct virtio_pcm_msg **msgs;
 	unsigned int nmsgs;
 	unsigned int msg_count;
@@ -78,6 +147,64 @@
 	wait_queue_head_t msg_empty;
 };
 
+/* Caller holds the PCM stream lock and vss->lock. */
+static inline void
+virtsnd_pcm_capture_arm_paused_stop(struct virtio_pcm_substream *vss)
+{
+	struct snd_pcm_runtime *runtime = vss->substream->runtime;
+
+	lockdep_assert_held(&vss->lock);
+	if (!runtime || runtime->state != SNDRV_PCM_STATE_PAUSED)
+		return;
+
+	/* ALSA skips trigger(STOP) when moving a paused stream to XRUN. */
+	vss->xfer_enabled = false;
+	vss->stopped = true;
+	runtime->stop_operating = true;
+}
+
+/* Keep non-staged PCM notifications on schedule_work()'s normal queue. */
+static inline bool
+virtsnd_pcm_schedule_period_elapsed(struct virtio_pcm_substream *vss)
+{
+	return schedule_work(&vss->elapsed_period);
+}
+
+/*
+ * ALSA may observe a subperiod tail before its period callback. Keep actual
+ * unread occupancy separate from pointer observations and notification state;
+ * free-running streams account occupancy from monotonic totals instead.
+ */
+static inline size_t virtsnd_pcm_ring_offset(u64 total, size_t size)
+{
+	u64 remainder;
+
+	div64_u64_rem(total, size, &remainder);
+	return remainder;
+}
+
+static inline bool
+virtsnd_pcm_capture_counters_valid(const struct virtio_pcm_substream *vss)
+{
+	u64 outstanding;
+
+	if (!vss->buffer_bytes ||
+	    vss->capture_copied_total < vss->capture_acked_total ||
+	    vss->capture_pointer_seen_total > vss->capture_copied_total ||
+	    vss->hw_ptr != virtsnd_pcm_ring_offset(vss->capture_copied_total, vss->buffer_bytes) ||
+	    (vss->capture_callback_snapshot_active &&
+	     (vss->capture_callback_snapshot < vss->capture_pointer_seen_total ||
+	      vss->capture_callback_snapshot > vss->capture_copied_total)))
+		return false;
+
+	if (vss->capture_free_running)
+		return !vss->capture_ready;
+
+	outstanding = vss->capture_copied_total - vss->capture_acked_total;
+	return outstanding == (u64)vss->capture_ready &&
+	       vss->capture_ready < vss->buffer_bytes;
+}
+
 /**
  * struct virtio_pcm_stream - VirtIO PCM stream.
  * @substreams: VirtIO substreams belonging to the stream.
@@ -108,6 +235,11 @@
 
 extern const struct snd_pcm_ops virtsnd_pcm_ops[];
 
+extern bool virtsnd_capture_staging;
+extern bool virtsnd_capture_packed_status;
+extern bool virtsnd_capture_allow_short;
+extern bool virtsnd_capture_allow_empty;
+
 int virtsnd_pcm_validate(struct virtio_device *vdev);
 
 int virtsnd_pcm_parse_cfg(struct virtio_snd *snd);
@@ -126,11 +258,16 @@
 
 struct virtio_snd_msg *
 virtsnd_pcm_ctl_msg_alloc(struct virtio_pcm_substream *vss,
-			  unsigned int command, gfp_t gfp);
+			  unsigned int command, gfp_t gfp, u64 epoch);
 
 int virtsnd_pcm_msg_alloc(struct virtio_pcm_substream *vss,
 			  unsigned int periods, unsigned int period_bytes);
 
+int virtsnd_pcm_capture_send(struct virtio_pcm_substream *vss);
+
+int virtsnd_pcm_capture_refresh_mode_locked(struct virtio_pcm_substream *vss,
+					    bool *free_running);
+
 void virtsnd_pcm_msg_free(struct virtio_pcm_substream *vss);
 
 void virtsnd_pcm_msg_reset(struct virtio_pcm_substream *vss);
diff --git a/sound/virtio/virtio_pcm_msg.c b/sound/virtio/virtio_pcm_msg.c
--- a/sound/virtio/virtio_pcm_msg.c
+++ b/sound/virtio/virtio_pcm_msg.c
@@ -3,9 +3,12 @@
  * virtio-snd: Virtio sound device
  * Copyright (C) 2021 OpenSynergy GmbH
  */
+#include <linux/compiler.h>
+#include <linux/gcd.h>
 #include <linux/limits.h>
 #include <linux/mm.h>
 #include <linux/overflow.h>
+#include <linux/vmalloc.h>
 #include <sound/pcm_params.h>
 
 #include "virtio_card.h"
@@ -16,14 +19,16 @@
  * @xfer: Request header payload.
  * @status: Response header payload.
  * @length: Data length in bytes.
+ * @data: Request-owned payload for staged capture.
  * @data_offset: Offset in the driver-owned PCM buffer on NOMMU systems.
  * @sgs: Payload scatter-gather table.
  */
 struct virtio_pcm_msg {
 	struct virtio_pcm_substream *substream;
-	struct virtio_snd_pcm_xfer xfer;
-	struct virtio_snd_pcm_status status;
+	struct virtio_snd_pcm_xfer *xfer;
+	struct virtio_snd_pcm_status *status;
 	size_t length;
+	u8 *data;
 #ifndef CONFIG_MMU
 	size_t data_offset;
 #endif
@@ -228,6 +233,8 @@
 	if (check_mul_overflow((size_t)periods, (size_t)period_bytes,
 			       &buffer_bytes))
 		return -EOVERFLOW;
+	if (vss->capture_staging && periods < 2)
+		return -EINVAL;
 
 	if (!vss->substream || !vss->substream->runtime)
 		return -EINVAL;
@@ -236,14 +243,19 @@
 		return -EINVAL;
 
 #ifdef CONFIG_MMU
-	rc = virtsnd_pcm_dma_pages_get(vss, runtime->dma_area, buffer_bytes);
-	if (rc)
-		return rc;
+	if (!vss->capture_staging) {
+		rc = virtsnd_pcm_dma_pages_get(vss, runtime->dma_area,
+					       buffer_bytes);
+		if (rc)
+			return rc;
+	}
 #else
-	vss->owned_dma_area = kzalloc(buffer_bytes, GFP_KERNEL);
-	if (!vss->owned_dma_area)
-		return -ENOMEM;
-	vss->owned_dma_bytes = buffer_bytes;
+	if (!vss->capture_staging) {
+		vss->owned_dma_area = kzalloc(buffer_bytes, GFP_KERNEL);
+		if (!vss->owned_dma_area)
+			return -ENOMEM;
+		vss->owned_dma_bytes = buffer_bytes;
+	}
 #endif
 
 	vss->msgs = kzalloc_objs(*vss->msgs, periods);
@@ -255,32 +267,54 @@
 	vss->nmsgs = periods;
 
 	for (i = 0; i < periods; ++i) {
-		int sg_num;
 		struct virtio_pcm_msg *msg;
 		u8 *data;
+		int sg_num;
+
+		if (vss->capture_staging) {
+			data = vzalloc(period_bytes);
+		} else {
 #ifdef CONFIG_MMU
-		data = runtime->dma_area + (size_t)period_bytes * i;
+			data = runtime->dma_area + (size_t)period_bytes * i;
 #else
-		data = vss->owned_dma_area + (size_t)period_bytes * i;
+			data = vss->owned_dma_area + (size_t)period_bytes * i;
 #endif
+		}
+		if (!data) {
+			rc = -ENOMEM;
+			goto err_free_msgs;
+		}
+
 		sg_num = virtsnd_pcm_sg_num(data, period_bytes);
 		if (sg_num < 0) {
+			if (vss->capture_staging)
+				vfree(data);
 			rc = sg_num;
 			goto err_free_msgs;
 		}
 
 		msg = kzalloc_flex(*msg, sgs, sg_num + 2);
 		if (!msg) {
+			if (vss->capture_staging)
+				vfree(data);
 			rc = -ENOMEM;
 			goto err_free_msgs;
 		}
 
 		msg->substream = vss;
+		msg->data = vss->capture_staging ? data : NULL;
 		vss->msgs[i] = msg;
-		sg_init_one(&msg->sgs[PCM_MSG_SG_XFER], &msg->xfer,
-			    sizeof(msg->xfer));
-		sg_init_one(&msg->sgs[PCM_MSG_SG_STATUS], &msg->status,
-			    sizeof(msg->status));
+		/* DMA payloads must not share cachelines with CPU-owned fields. */
+		msg->xfer = kzalloc(sizeof(*msg->xfer), GFP_KERNEL);
+		msg->status = kzalloc(sizeof(*msg->status), GFP_KERNEL);
+		if (!msg->xfer || !msg->status) {
+			rc = -ENOMEM;
+			goto err_free_msgs;
+		}
+		sg_init_one(&msg->sgs[PCM_MSG_SG_XFER], msg->xfer,
+			    sizeof(*msg->xfer));
+		sg_init_one(&msg->sgs[PCM_MSG_SG_STATUS], msg->status,
+			    sizeof(*msg->status));
 #ifndef CONFIG_MMU
 		msg->data_offset = (size_t)period_bytes * i;
 #endif
@@ -316,8 +350,14 @@
 {
 	unsigned int i;
 
-	for (i = 0; vss->msgs && i < vss->nmsgs; ++i)
-		kfree(vss->msgs[i]);
+	for (i = 0; vss->msgs && i < vss->nmsgs; ++i) {
+		if (vss->msgs[i]) {
+			vfree(vss->msgs[i]->data);
+			kfree(vss->msgs[i]->xfer);
+			kfree(vss->msgs[i]->status);
+			kfree(vss->msgs[i]);
+		}
+	}
 	kfree(vss->msgs);
 #ifdef CONFIG_MMU
 	virtsnd_pcm_dma_pages_put(vss);
@@ -339,15 +379,26 @@
  * substream was stopped are never sent to the device, so their length must be
  * cleared before the substream is started again.
  *
- * Context: Any context. Takes and releases the VirtIO substream spinlock.
+ * Context: Caller holds the VirtIO substream spinlock.
  */
 void virtsnd_pcm_msg_reset(struct virtio_pcm_substream *vss)
 {
 	unsigned int i;
 
-	guard(spinlock_irqsave)(&vss->lock);
+	lockdep_assert_held(&vss->lock);
 	for (i = 0; vss->msgs && i < vss->nmsgs; ++i)
 		vss->msgs[i]->length = 0;
+	vss->capture_ready = 0;
+	vss->capture_copied_total = 0;
+	vss->capture_acked_total = 0;
+	vss->capture_pointer_seen_total = 0;
+	vss->capture_notified_periods = 0;
+	vss->capture_callback_snapshot = 0;
+	vss->capture_callback_snapshot_active = false;
+	vss->capture_boundary = 0;
+	vss->capture_free_running = false;
+	vss->capture_empty = 0;
+	vss->capture_started = false;
 	vss->capture_rx_seeded = false;
 }
 
@@ -378,6 +429,8 @@
 	bool notify = false;
 	int rc;
 
+	if (!vqueue || !vss->prepared || vss->suspended)
+		return -EBADFD;
 	if (!bytes)
 		return 0;
 	if (!period_bytes || !vss->msgs || offset >= vss->buffer_bytes ||
@@ -403,8 +456,8 @@
 
 		msg->length += n;
 		if (msg->length == period_bytes) {
-			msg->xfer.stream_id = cpu_to_le32(vss->sid);
-			memset(&msg->status, 0, sizeof(msg->status));
+			msg->xfer->stream_id = cpu_to_le32(vss->sid);
+			memset(msg->status, 0, sizeof(*msg->status));
 
 #ifndef CONFIG_MMU
 		if (vss->direction == SNDRV_PCM_STREAM_PLAYBACK) {
@@ -478,6 +531,293 @@
 	return 0;
 }
 
+/* Caller holds vss->lock. */
+int virtsnd_pcm_capture_refresh_mode_locked(struct virtio_pcm_substream *vss,
+					    bool *free_running)
+{
+	struct snd_pcm_runtime *runtime;
+	snd_pcm_uframes_t stop_threshold;
+	u64 outstanding;
+	bool new_free_running;
+
+	if (!vss->capture_staging || !vss->runtime_active || !vss->substream)
+		return -EBADFD;
+
+	runtime = vss->substream->runtime;
+	if (!runtime || !vss->buffer_bytes || !vss->capture_boundary)
+		return -EBADFD;
+
+	/* Validate the currently committed policy before observing a transition. */
+	if (!virtsnd_pcm_capture_counters_valid(vss))
+		return -EIO;
+
+	stop_threshold = READ_ONCE(runtime->stop_threshold);
+	new_free_running = stop_threshold == vss->capture_boundary;
+	if (new_free_running != vss->capture_free_running) {
+		outstanding = vss->capture_copied_total -
+			      vss->capture_acked_total;
+		if (new_free_running) {
+			vss->capture_ready = 0;
+		} else {
+			if (outstanding >= vss->buffer_bytes)
+				return -EPIPE;
+			vss->capture_ready = outstanding;
+		}
+		vss->capture_free_running = new_free_running;
+	}
+
+	*free_running = new_free_running;
+	return 0;
+}
+
+/* Caller holds the RX queue and vss locks and has sampled the policy. */
+static int virtsnd_pcm_capture_enqueue(struct virtio_pcm_substream *vss,
+				       bool free_running)
+{
+	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+	struct virtqueue *vqueue = queue->vqueue;
+	unsigned long period_bytes;
+	size_t free_bytes;
+	size_t request_limit = vss->nmsgs;
+	unsigned int i;
+	bool added = false;
+
+	if (!vqueue || !vss->prepared || vss->suspended ||
+	    !vss->capture_staging || !vss->runtime_active)
+		return -EBADFD;
+
+	period_bytes = snd_pcm_lib_period_bytes(vss->substream);
+	if (!period_bytes || !vss->buffer_bytes ||
+	    period_bytes >= vss->buffer_bytes ||
+	    free_running != vss->capture_free_running ||
+	    !virtsnd_pcm_capture_counters_valid(vss) ||
+	    vss->msg_count > vss->nmsgs)
+		return -EIO;
+
+	if (!free_running) {
+		free_bytes = vss->buffer_bytes - vss->capture_ready;
+		request_limit = free_bytes / period_bytes +
+				!!(free_bytes % period_bytes);
+		request_limit = min_t(size_t, request_limit, vss->nmsgs);
+	}
+
+	for (i = 0; i < vss->nmsgs && vss->msg_count < request_limit; ++i) {
+		struct virtio_pcm_msg *msg = vss->msgs[i];
+		struct scatterlist *sgs[] = {
+			&msg->sgs[PCM_MSG_SG_XFER],
+			&msg->sgs[PCM_MSG_SG_DATA],
+			&msg->sgs[PCM_MSG_SG_STATUS]
+		};
+		int rc;
+
+		if (msg->length)
+			continue;
+
+		msg->xfer->stream_id = cpu_to_le32(vss->sid);
+		/* Do not let an absent packed trailer reuse old PCM bytes. */
+		memset(msg->data, 0, period_bytes);
+		memset(msg->status, 0, sizeof(*msg->status));
+		msg->status->status = cpu_to_le32(VIRTIO_SND_S_BAD_MSG);
+		rc = virtqueue_add_sgs(vqueue, sgs, 1, 2, msg, GFP_ATOMIC);
+		if (rc) {
+			if (added && !(vss->features &
+				       (1U << VIRTIO_SND_PCM_F_MSG_POLLING)) &&
+			    virtqueue_kick_prepare(vqueue))
+				virtqueue_notify(vqueue);
+			return rc;
+		}
+
+		msg->length = period_bytes;
+		vss->msg_count++;
+		added = true;
+	}
+
+	if (added && !(vss->features &
+			(1U << VIRTIO_SND_PCM_F_MSG_POLLING)) &&
+	    virtqueue_kick_prepare(vqueue))
+		virtqueue_notify(vqueue);
+
+	return 0;
+}
+
+/**
+ * virtsnd_pcm_capture_send() - Refresh policy and queue staged requests.
+ * @vss: VirtIO PCM substream.
+ *
+ * Context: Expects the RX queue and substream spinlocks to be held by caller.
+ * Return: 0 on success, -errno on failure.
+ */
+int virtsnd_pcm_capture_send(struct virtio_pcm_substream *vss)
+{
+	bool free_running;
+	int rc;
+
+	rc = virtsnd_pcm_capture_refresh_mode_locked(vss, &free_running);
+	if (rc)
+		return rc;
+
+	return virtsnd_pcm_capture_enqueue(vss, free_running);
+}
+
+static int virtsnd_pcm_capture_status(const struct virtio_pcm_msg *msg,
+				      size_t written_bytes,
+				      size_t frame_bytes,
+				      size_t *audio_bytes,
+				      u32 *latency_bytes)
+{
+	struct virtio_snd_pcm_status status;
+	size_t max_written;
+	size_t first;
+
+	if (!frame_bytes ||
+	    check_add_overflow(msg->length, sizeof(status), &max_written) ||
+	    written_bytes < sizeof(status) || written_bytes > max_written)
+		return -EIO;
+
+	*audio_bytes = written_bytes - sizeof(status);
+	if (*audio_bytes > msg->length || *audio_bytes % frame_bytes)
+		return -EIO;
+
+	if (msg->substream->capture_staging && virtsnd_capture_packed_status) {
+		first = min_t(size_t, sizeof(status),
+			      msg->length - *audio_bytes);
+		if (first)
+			memcpy(&status, msg->data + *audio_bytes, first);
+		if (first < sizeof(status))
+			memcpy((u8 *)&status + first, msg->status,
+			       sizeof(status) - first);
+	} else {
+		status = *msg->status;
+	}
+
+	if (le32_to_cpu(status.status) != VIRTIO_SND_S_OK)
+		return -EIO;
+
+	*latency_bytes = le32_to_cpu(status.latency_bytes);
+	return 0;
+}
+
+static void virtsnd_pcm_capture_xrun_locked(struct virtio_pcm_substream *vss)
+{
+	if (!vss->xfer_xrun) {
+		vss->xfer_xrun = true;
+		virtsnd_pcm_schedule_period_elapsed(vss);
+	}
+}
+
+static void virtsnd_pcm_capture_complete(struct virtio_pcm_msg *msg,
+					 size_t written_bytes)
+{
+	struct virtio_pcm_substream *vss = msg->substream;
+	size_t audio_bytes = 0;
+	size_t next_ready = 0;
+	u64 next_copied_total = 0;
+	bool free_running = false;
+	bool got_audio = false;
+
+	guard(spinlock)(&vss->lock);
+	vss->msg_count--;
+	/* Native pause stops refills but must retain requests already in flight. */
+	if (vss->runtime_active && vss->prepared && !vss->stopped &&
+	    !vss->xfer_xrun) {
+		struct snd_pcm_runtime *runtime = vss->substream->runtime;
+		size_t frame_bytes = runtime ?
+			runtime->frame_bits / gcd(runtime->frame_bits, 8) : 0;
+		u32 latency_bytes;
+		const char *invalid_reason = NULL;
+
+		/* A full ring is modulo-ambiguous, so fail closed before copying. */
+		if (!runtime || !runtime->dma_area)
+			invalid_reason = "runtime";
+		else if (virtsnd_pcm_capture_refresh_mode_locked(vss,
+								 &free_running))
+			invalid_reason = "policy";
+		else if (!virtsnd_pcm_capture_counters_valid(vss))
+			invalid_reason = "counters";
+		else if (vss->hw_ptr !=
+			 virtsnd_pcm_ring_offset(vss->capture_copied_total, vss->buffer_bytes))
+			invalid_reason = "hw_ptr";
+		else if (virtsnd_pcm_capture_status(msg, written_bytes,
+						    frame_bytes, &audio_bytes,
+						    &latency_bytes))
+			invalid_reason = "status";
+		else if (check_add_overflow(vss->capture_copied_total,
+					    (u64)audio_bytes,
+					    &next_copied_total))
+			invalid_reason = "total_overflow";
+		else if (free_running && audio_bytes >= vss->buffer_bytes)
+			invalid_reason = "payload_exceeds_ring";
+		else if (!free_running &&
+			 audio_bytes >= vss->buffer_bytes - vss->capture_ready)
+			invalid_reason = "ring_capacity";
+		else if (!free_running &&
+			 check_add_overflow(vss->capture_ready, audio_bytes,
+					    &next_ready))
+			invalid_reason = "ready_overflow";
+		else if (!free_running && next_ready >= vss->buffer_bytes)
+			invalid_reason = "ring_full";
+		else if (free_running &&
+			 (next_copied_total < vss->capture_pointer_seen_total ||
+			  next_copied_total - vss->capture_pointer_seen_total >=
+			  vss->buffer_bytes))
+			invalid_reason = "pointer_window";
+		else if (free_running)
+			next_ready = 0;
+
+		if (invalid_reason) {
+			virtsnd_pcm_capture_xrun_locked(vss);
+			dev_warn_ratelimited(&vss->snd->vdev->dev,
+					     "SID %u: staged capture completion rejected (%s)\n",
+					     vss->sid, invalid_reason);
+		} else if (!audio_bytes) {
+			unsigned int limit = min_t(unsigned int, vss->nmsgs,
+						   UINT_MAX / 2) * 2;
+
+			if (!virtsnd_capture_allow_empty || vss->capture_started ||
+			    vss->capture_empty >= limit) {
+				virtsnd_pcm_capture_xrun_locked(vss);
+			} else {
+				vss->capture_empty++;
+			}
+		} else if (audio_bytes < msg->length &&
+			   !virtsnd_capture_allow_short) {
+			virtsnd_pcm_capture_xrun_locked(vss);
+		} else {
+			size_t ring_ptr = virtsnd_pcm_ring_offset(vss->capture_copied_total,
+						  vss->buffer_bytes);
+			size_t first = min(audio_bytes,
+					   vss->buffer_bytes - ring_ptr);
+
+			memcpy(runtime->dma_area + ring_ptr, msg->data, first);
+			if (audio_bytes > first)
+				memcpy(runtime->dma_area, msg->data + first,
+				       audio_bytes - first);
+			vss->capture_copied_total = next_copied_total;
+			vss->hw_ptr = virtsnd_pcm_ring_offset(next_copied_total, vss->buffer_bytes);
+			vss->capture_ready = next_ready;
+			vss->capture_started = true;
+			vss->capture_empty = 0;
+			runtime->delay = bytes_to_frames(runtime, latency_bytes);
+			got_audio = true;
+		}
+	}
+
+	msg->length = 0;
+	if (vss->runtime_active && vss->xfer_enabled) {
+		if (!vss->xfer_xrun &&
+		    virtsnd_pcm_capture_enqueue(vss, free_running)) {
+			dev_warn_ratelimited(&vss->snd->vdev->dev,
+					     "SID %u: failed to refill capture queue\n",
+					     vss->sid);
+			virtsnd_pcm_capture_xrun_locked(vss);
+		}
+	} else if (!vss->msg_count) {
+		wake_up_all(&vss->msg_empty);
+	}
+	if (got_audio && !vss->xfer_xrun)
+		virtsnd_pcm_schedule_period_elapsed(vss);
+}
+
 /**
  * virtsnd_pcm_msg_pending_num() - Returns the number of pending I/O messages.
  * @vss: VirtIO substream.
@@ -529,9 +869,9 @@
  * allowed, then each completed message is immediately placed back at the end
  * of the queue.
  *
- * For the playback substream, @written_bytes is equal to sizeof(msg->status).
- *
- * For the capture substream, @written_bytes is equal to sizeof(msg->status)
+ * For the playback substream, @written_bytes is equal to sizeof(*msg->status).
+ *
+ * For the capture substream, @written_bytes is equal to sizeof(*msg->status)
  * plus the number of captured bytes.
  *
  * Context: Interrupt context. Takes and releases the VirtIO substream spinlock.
@@ -540,19 +880,52 @@
 				     size_t written_bytes)
 {
 	struct virtio_pcm_substream *vss = msg->substream;
-
-	/*
-	 * hw_ptr always indicates the buffer position of the first I/O message
-	 * in the virtqueue. Therefore, on each completion of an I/O message,
-	 * the hw_ptr value is unconditionally advanced.
-	 */
+	size_t completed_bytes;
+	u32 latency_bytes;
+
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+	    vss->capture_staging) {
+		virtsnd_pcm_capture_complete(msg, written_bytes);
+		return;
+	}
+
+	/* Only completed TX requests and valid RX audio advance the pointer. */
 	guard(spinlock)(&vss->lock);
+	/* Empty STOP returns retire ownership without inventing captured data. */
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+	    (!vss->prepared ||
+	     (vss->stopped && written_bytes <= sizeof(*msg->status)))) {
+		msg->length = 0;
+		vss->msg_count--;
+		if (!vss->msg_count)
+			wake_up_all(&vss->msg_empty);
+		return;
+	}
+	completed_bytes = msg->length;
+	latency_bytes = le32_to_cpu(msg->status->latency_bytes);
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE) {
+		struct snd_pcm_runtime *runtime = vss->substream->runtime;
+		size_t frame_bytes = runtime->frame_bits / gcd(runtime->frame_bits, 8);
+
+		if (virtsnd_pcm_capture_status(msg, written_bytes, frame_bytes,
+					       &completed_bytes, &latency_bytes) ||
+		    !completed_bytes) {
+			vss->xfer_xrun = true;
+			msg->length = 0;
+			vss->msg_count--;
+			if (!vss->msg_count)
+				wake_up_all(&vss->msg_empty);
+			virtsnd_pcm_schedule_period_elapsed(vss);
+			return;
+		}
+	}
 #ifndef CONFIG_MMU
 	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
-	    vss->xfer_enabled && written_bytes > sizeof(msg->status)) {
+	    vss->prepared && !vss->stopped &&
+	    written_bytes > sizeof(*msg->status)) {
 		struct snd_pcm_runtime *runtime =
 			vss->substream ? vss->substream->runtime : NULL;
-		size_t reported_bytes = written_bytes - sizeof(msg->status);
+		size_t reported_bytes = written_bytes - sizeof(*msg->status);
 		size_t data_bytes = min_t(size_t, reported_bytes, msg->length);
 
 		if (!runtime || !runtime->dma_area || !vss->owned_dma_area ||
@@ -568,26 +941,16 @@
 			       vss->owned_dma_area + msg->data_offset, data_bytes);
 	}
 #endif
-	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
-	    written_bytes > sizeof(msg->status) &&
-	    written_bytes - sizeof(msg->status) > msg->length)
-		dev_err_ratelimited(&vss->snd->vdev->dev,
-				    "SID %u: capture length exceeds I/O message capacity\n",
-			vss->sid);
-	/*
-	 * If the capture substream returned an incorrect status, then just
-	 * increase the hw_ptr by the message size. Clamp valid capture lengths to
-	 * the actual SG payload capacity before accounting them.
-	 */
-	if (vss->direction == SNDRV_PCM_STREAM_PLAYBACK ||
-	    written_bytes <= sizeof(msg->status)) {
-		vss->hw_ptr += msg->length;
-	} else {
-		size_t data_bytes = min_t(size_t,
-					  written_bytes - sizeof(msg->status),
-					  msg->length);
-
-		vss->hw_ptr += data_bytes;
+	vss->hw_ptr += completed_bytes;
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE) {
+		/* A full unseen ring cannot be represented by the modulo pointer. */
+		if (check_add_overflow(vss->capture_copied_total,
+				       (u64)completed_bytes,
+				       &vss->capture_copied_total) ||
+		    vss->capture_copied_total < vss->capture_pointer_seen_total ||
+		    vss->capture_copied_total - vss->capture_pointer_seen_total >=
+			vss->buffer_bytes)
+			vss->xfer_xrun = true;
 	}
 
 	if (vss->hw_ptr >= vss->buffer_bytes)
@@ -595,16 +958,17 @@
 
 	msg->length = 0;
 
-	vss->xfer_xrun = false;
+	/* Preserve a latched capture error until prepare recovers it. */
+	if (vss->direction == SNDRV_PCM_STREAM_PLAYBACK)
+		vss->xfer_xrun = false;
 	vss->msg_count--;
 
-	if (vss->xfer_enabled) {
+	if (vss->xfer_enabled ||
+	    (vss->direction == SNDRV_PCM_STREAM_CAPTURE && vss->xfer_xrun)) {
 		struct snd_pcm_runtime *runtime = vss->substream->runtime;
 
-		runtime->delay =
-			bytes_to_frames(runtime,
-					le32_to_cpu(msg->status.latency_bytes));
-		schedule_work(&vss->elapsed_period);
+		runtime->delay = bytes_to_frames(runtime, latency_bytes);
+		virtsnd_pcm_schedule_period_elapsed(vss);
 	} else if (!vss->msg_count) {
 		wake_up_all(&vss->msg_empty);
 	}
@@ -664,6 +1028,7 @@
  * virtsnd_pcm_ctl_msg_alloc() - Allocate and initialize the PCM device control
  *                               message for the specified substream.
  * @vss: VirtIO PCM substream.
+ * @epoch: Operation generation captured before allocation.
  * @command: Control request code (VIRTIO_SND_R_PCM_XXX).
  * @gfp: Kernel flags for memory allocation.
  *
@@ -672,7 +1037,7 @@
  */
 struct virtio_snd_msg *
 virtsnd_pcm_ctl_msg_alloc(struct virtio_pcm_substream *vss,
-			  unsigned int command, gfp_t gfp)
+			  unsigned int command, gfp_t gfp, u64 epoch)
 {
 	size_t request_size = sizeof(struct virtio_snd_pcm_hdr);
 	size_t response_size = sizeof(struct virtio_snd_hdr);
@@ -690,6 +1055,7 @@
 
 		hdr->hdr.code = cpu_to_le32(command);
 		hdr->stream_id = cpu_to_le32(vss->sid);
+		virtsnd_ctl_msg_bind_pcm(msg, vss, epoch);
 	}
 
 	return msg;
diff --git a/sound/virtio/virtio_pcm_ops.c b/sound/virtio/virtio_pcm_ops.c
--- a/sound/virtio/virtio_pcm_ops.c
+++ b/sound/virtio/virtio_pcm_ops.c
@@ -3,9 +3,43 @@
  * virtio-snd: Virtio sound device
  * Copyright (C) 2021 OpenSynergy GmbH
  */
+#include <linux/moduleparam.h>
+#include <linux/overflow.h>
 #include <sound/pcm_params.h>
 
 #include "virtio_card.h"
+
+static const struct snd_pcm_ops virtsnd_pcm_capture_staged_ops;
+static const struct snd_pcm_ops virtsnd_pcm_capture_indirect_ops;
+
+bool virtsnd_capture_staging;
+module_param_named(capture_staging, virtsnd_capture_staging, bool, 0444);
+MODULE_PARM_DESC(capture_staging,
+		 "Use request-owned buffers for experimental PCM capture");
+
+bool virtsnd_capture_packed_status;
+module_param_named(capture_packed_status, virtsnd_capture_packed_status,
+		   bool, 0444);
+MODULE_PARM_DESC(capture_packed_status,
+		 "Read staged capture status after the actual audio payload");
+
+bool virtsnd_capture_allow_short;
+module_param_named(capture_allow_short, virtsnd_capture_allow_short,
+		   bool, 0444);
+MODULE_PARM_DESC(capture_allow_short,
+		 "Accept non-empty short staged capture responses");
+
+bool virtsnd_capture_allow_empty;
+module_param_named(capture_allow_empty, virtsnd_capture_allow_empty,
+		   bool, 0444);
+MODULE_PARM_DESC(capture_allow_empty,
+		 "Retry bounded empty startup capture responses");
+
+static bool virtsnd_capture_pause_restart;
+module_param_named(capture_pause_restart, virtsnd_capture_pause_restart,
+		   bool, 0444);
+MODULE_PARM_DESC(capture_pause_restart,
+		 "Release and reprepare staged capture streams across pause");
 
 /*
  * I/O messages lifetime
@@ -97,14 +131,242 @@
 };
 
 static int virtsnd_pcm_sync_stop(struct snd_pcm_substream *substream);
-
-static void virtsnd_pcm_start_failed(struct virtio_pcm_substream *vss)
-{
-	scoped_guard(spinlock_irqsave, &vss->lock) {
+static int virtsnd_pcm_sync_stop_locked(struct snd_pcm_substream *substream);
+static int virtsnd_pcm_pause_restart_sync_stop(struct snd_pcm_substream *substream);
+
+static void virtsnd_pcm_start_failed(struct virtio_pcm_substream *vss, u64 epoch)
+{
+	bool notify;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return;
+		vss->xfer_enabled = false;
+		vss->stopped = vss->release_required || vss->msg_count;
+		notify = vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+			 (!vss->capture_staging || vss->runtime_active);
+		if (notify)
+			vss->xfer_xrun = true;
+	}
+	vss->substream->runtime->stop_operating = true;
+	if (notify)
+		virtsnd_pcm_schedule_period_elapsed(vss);
+}
+
+static bool
+virtsnd_pcm_pause_restart_enabled(struct virtio_pcm_substream *vss)
+{
+	return virtsnd_capture_pause_restart && vss->capture_staging &&
+	       vss->direction == SNDRV_PCM_STREAM_CAPTURE;
+}
+
+static bool virtsnd_pcm_pause_restart_transitioning(enum virtio_pcm_pause_restart_state state)
+{
+	switch (state) {
+	case VIRTSND_PCM_PAUSE_RESTART_SET_PARAMS_SUBMITTED:
+	case VIRTSND_PCM_PAUSE_RESTART_PREPARE_SUBMITTED:
+	case VIRTSND_PCM_PAUSE_RESTART_START_SUBMITTED:
+	case VIRTSND_PCM_PAUSE_RESTART_STOP_SUBMITTED:
+	case VIRTSND_PCM_PAUSE_RESTART_RELEASE_SUBMITTED:
+		return true;
+	default:
+		return false;
+	}
+}
+
+static int virtsnd_pcm_send_sync(struct virtio_pcm_substream *vss,
+				 struct virtio_snd_msg *msg)
+{
+	u64 epoch = virtsnd_ctl_msg_pcm_epoch(msg);
+	bool uncertain;
+	int rc;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch ||
+		    vss->control_uncertain) {
+			rc = vss->control_uncertain ? -EIO : -ECANCELED;
+			virtsnd_ctl_msg_unref(msg);
+			return rc;
+		}
+	}
+	virtsnd_ctl_msg_ref(msg);
+	rc = virtsnd_ctl_msg_send_sync(vss->snd, msg);
+	uncertain = virtsnd_ctl_msg_uncertain(msg, rc);
+	virtsnd_ctl_msg_unref(msg);
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (epoch != vss->reset_epoch || vss->resetting)
+			return -ECANCELED;
+		if (rc && uncertain)
+			vss->control_uncertain = true;
+	}
+	return rc;
+}
+
+static int
+virtsnd_pcm_pause_restart_send(struct virtio_pcm_substream *vss,
+			       struct virtio_snd_msg *msg,
+			       unsigned long allowed_states,
+			       enum virtio_pcm_pause_restart_state submitted_state,
+			       enum virtio_pcm_pause_restart_state success_state)
+{
+	enum virtio_pcm_pause_restart_state state;
+	bool uncertain;
+	u64 epoch = virtsnd_ctl_msg_pcm_epoch(msg);
+	int rc;
+
+	if (!virtsnd_pcm_pause_restart_enabled(vss))
+		return virtsnd_pcm_send_sync(vss, msg);
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		state = vss->pause_restart_state;
+		if (vss->resetting || vss->reset_epoch != epoch) {
+			rc = -ECANCELED;
+		} else if (state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED) {
+			rc = -EIO;
+		} else if (virtsnd_pcm_pause_restart_transitioning(state)) {
+			vss->pause_restart_state =
+				VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED;
+			rc = -EIO;
+		} else if (!(allowed_states & BIT(state))) {
+			rc = -EBADFD;
+		} else {
+			vss->pause_restart_state = submitted_state;
+			rc = 0;
+		}
+	}
+
+	if (rc) {
+		virtsnd_ctl_msg_unref(msg);
+		return rc;
+	}
+
+	virtsnd_ctl_msg_ref(msg);
+	rc = virtsnd_ctl_msg_send_sync(vss->snd, msg);
+	uncertain = virtsnd_ctl_msg_uncertain(msg, rc);
+	virtsnd_ctl_msg_unref(msg);
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->reset_epoch != epoch || vss->resetting)
+			return -ECANCELED;
+		if ((rc && uncertain) || vss->pause_restart_state != submitted_state) {
+			vss->pause_restart_state =
+				VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED;
+			if (!rc)
+				rc = -EIO;
+		} else if (rc) {
+			vss->pause_restart_state = state;
+		} else {
+			vss->pause_restart_state = success_state;
+		}
+	}
+
+	if (rc && uncertain)
+		dev_err(&vss->snd->vdev->dev,
+			"SID %u: PCM control state is unknown; device reprobe required\n",
+			vss->sid);
+
+	return rc;
+}
+
+static void
+virtsnd_pcm_pause_restart_fail(struct virtio_pcm_substream *vss,
+			       bool reset_required, u64 epoch)
+{
+	bool notify_xrun;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return;
+		notify_xrun = vss->runtime_active;
 		vss->xfer_enabled = false;
 		vss->stopped = true;
-	}
+		vss->xfer_xrun = true;
+		if (reset_required)
+			vss->pause_restart_state =
+				VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED;
+	}
+	if (notify_xrun)
+		virtsnd_pcm_schedule_period_elapsed(vss);
 	vss->substream->runtime->stop_operating = true;
+}
+
+static int virtsnd_pcm_dev_prepare(struct virtio_pcm_substream *vss, u64 epoch)
+{
+	struct virtio_snd_msg *msg;
+
+	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_PREPARE,
+					GFP_KERNEL, epoch);
+	if (!msg)
+		return -ENOMEM;
+
+	return virtsnd_pcm_pause_restart_send(vss, msg,
+		BIT(VIRTSND_PCM_PAUSE_RESTART_CONFIGURED) |
+		BIT(VIRTSND_PCM_PAUSE_RESTART_RELEASED) |
+		BIT(VIRTSND_PCM_PAUSE_RESTART_STOPPED) |
+		BIT(VIRTSND_PCM_PAUSE_RESTART_PREPARED),
+		VIRTSND_PCM_PAUSE_RESTART_PREPARE_SUBMITTED,
+		VIRTSND_PCM_PAUSE_RESTART_PREPARED);
+}
+
+static int virtsnd_pcm_pause_prepare(struct virtio_pcm_substream *vss, u64 epoch)
+{
+	struct virtio_snd_msg *msg;
+
+	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_PREPARE,
+					GFP_KERNEL, epoch);
+	if (!msg)
+		return -ENOMEM;
+
+	return virtsnd_pcm_pause_restart_send(vss, msg,
+		BIT(VIRTSND_PCM_PAUSE_RESTART_RELEASED),
+		VIRTSND_PCM_PAUSE_RESTART_PREPARE_SUBMITTED,
+		VIRTSND_PCM_PAUSE_RESTART_PREPARED);
+}
+
+static int virtsnd_pcm_release_wait(struct virtio_pcm_substream *vss, u64 epoch)
+{
+	unsigned int js = msecs_to_jiffies(virtsnd_msg_timeout_ms);
+	long waited;
+
+	waited = wait_event_timeout(vss->msg_empty,
+				   !virtsnd_pcm_msg_pending_num(vss), js);
+	guard(spinlock_irqsave)(&vss->lock);
+	if (vss->resetting || vss->reset_epoch != epoch)
+		return -ECANCELED;
+	if (!waited || vss->msg_count ||
+	    vss->pause_restart_state != VIRTSND_PCM_PAUSE_RESTART_RELEASE_SENT)
+		return -ETIMEDOUT;
+	vss->prepared = false;
+	vss->pause_restart_state = VIRTSND_PCM_PAUSE_RESTART_RELEASED;
+	return 0;
+}
+
+static int
+virtsnd_pcm_pause_release_flush(struct virtio_pcm_substream *vss,
+				unsigned long allowed_states, u64 epoch)
+{
+	struct virtio_snd_msg *msg;
+	int rc;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return -ECANCELED;
+	}
+	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_RELEASE,
+					GFP_KERNEL, epoch);
+	if (!msg)
+		return -ENOMEM;
+	rc = virtsnd_pcm_pause_restart_send(vss, msg, allowed_states,
+			VIRTSND_PCM_PAUSE_RESTART_RELEASE_SUBMITTED,
+			VIRTSND_PCM_PAUSE_RESTART_RELEASE_SENT);
+	if (rc)
+		return rc;
+	return virtsnd_pcm_release_wait(vss, epoch);
+}
+
+static void virtsnd_pcm_pause_restart_failed(struct virtio_pcm_substream *vss,
+					     u64 epoch)
+{
+	virtsnd_pcm_pause_restart_fail(vss, false, epoch);
 }
 
 /**
@@ -119,23 +381,42 @@
 	struct virtio_pcm *vpcm = snd_pcm_substream_chip(substream);
 	struct virtio_pcm_stream *vs = &vpcm->streams[substream->stream];
 	struct virtio_pcm_substream *vss = vs->substreams[substream->number];
+	int rc;
 
 	substream->runtime->hw = vss->hw;
 	substream->private_data = vss;
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		vss->capture_staging = virtsnd_capture_staging &&
+			vss->direction == SNDRV_PCM_STREAM_CAPTURE;
+		if (vss->capture_staging) {
+			vss->runtime_active = false;
+			vss->xfer_enabled = false;
+		}
+	}
+
+	/* Keep RESET operation selection fixed for the opened runtime. */
+	substream->ops = vss->capture_staging ?
+		&virtsnd_pcm_capture_staged_ops :
+		&virtsnd_pcm_ops[substream->stream];
+
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE && !vss->capture_staging)
+		substream->ops = &virtsnd_pcm_capture_indirect_ops;
 
 	snd_pcm_hw_constraint_integer(substream->runtime,
 				      SNDRV_PCM_HW_PARAM_PERIODS);
 
 	scoped_guard(spinlock_irqsave, &vss->lock)
 		vss->stopped |= !!vss->msg_count;
-	vss->suspended = false;
 
 	/*
 	 * If the substream has already been used, then the I/O queue may be in
 	 * an invalid state. Just in case, we do a check and try to return the
 	 * queue to its original state, if necessary.
 	 */
-	return virtsnd_pcm_sync_stop(substream);
+	rc = virtsnd_pcm_sync_stop(substream);
+	if (!rc)
+		enable_work(&vss->elapsed_period);
+	return rc;
 }
 
 /**
@@ -147,6 +428,15 @@
  */
 static int virtsnd_pcm_close(struct snd_pcm_substream *substream)
 {
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		vss->runtime_active = false;
+		vss->xfer_enabled = false;
+	}
+	/* Block late producers as well as drain already queued notifications. */
+	disable_work_sync(&vss->elapsed_period);
+
 	return 0;
 }
 
@@ -159,6 +449,7 @@
  * @channels: Selected number of channels.
  * @format: Selected sample format (SNDRV_PCM_FORMAT_XXX).
  * @rate: Selected frame rate.
+ * @epoch: Operation generation captured before preparing the request.
  *
  * Context: Any context that permits to sleep.
  * Return: 0 on success, -errno on failure.
@@ -168,7 +459,7 @@
 				      unsigned int period_bytes,
 				      unsigned int channels,
 				      snd_pcm_format_t format,
-				      unsigned int rate)
+				      unsigned int rate, u64 epoch)
 {
 	struct virtio_snd_msg *msg;
 	struct virtio_snd_pcm_set_params *request;
@@ -194,7 +485,7 @@
 		return -EINVAL;
 
 	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_SET_PARAMS,
-					GFP_KERNEL);
+					GFP_KERNEL, epoch);
 	if (!msg)
 		return -ENOMEM;
 
@@ -213,7 +504,14 @@
 		request->features |=
 			cpu_to_le32(1U << VIRTIO_SND_PCM_F_EVT_XRUNS);
 
-	return virtsnd_ctl_msg_send_sync(vss->snd, msg);
+	return virtsnd_pcm_pause_restart_send(vss, msg,
+		BIT(VIRTSND_PCM_PAUSE_RESTART_NONE) |
+		BIT(VIRTSND_PCM_PAUSE_RESTART_CONFIGURED) |
+		BIT(VIRTSND_PCM_PAUSE_RESTART_PREPARED) |
+		BIT(VIRTSND_PCM_PAUSE_RESTART_STOPPED) |
+		BIT(VIRTSND_PCM_PAUSE_RESTART_RELEASED),
+		VIRTSND_PCM_PAUSE_RESTART_SET_PARAMS_SUBMITTED,
+		VIRTSND_PCM_PAUSE_RESTART_CONFIGURED);
 }
 
 /**
@@ -232,8 +530,10 @@
 	unsigned int msg_count;
 	size_t period_buffer_bytes;
 	bool stopped;
+	u64 epoch;
 	int rc;
 
+	guard(mutex)(&vss->lifecycle_lock);
 	/* Fixed descriptors require whole-byte periods covering the buffer. */
 	if (check_mul_overflow((size_t)params_periods(hw_params),
 			       (size_t)params_period_bytes(hw_params),
@@ -243,6 +543,25 @@
 	     (u64)snd_pcm_format_physical_width(params_format(hw_params)) *
 	     params_channels(hw_params)) % 8)
 		return -EINVAL;
+	scoped_guard(spinlock_irqsave, &vss->lock)
+		epoch = vss->reset_epoch;
+	/* Drain the previous buffer before ALSA can replace its mapping. */
+	rc = virtsnd_pcm_sync_stop_locked(substream);
+	if (rc)
+		return rc;
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return -ECANCELED;
+		vss->prepared = false;
+	}
+
+	if (vss->capture_staging) {
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			vss->runtime_active = false;
+			vss->xfer_enabled = false;
+		}
+		cancel_work_sync(&vss->elapsed_period);
+	}
 
 	scoped_guard(spinlock_irqsave, &vss->lock) {
 		msg_count = vss->msg_count;
@@ -260,11 +579,16 @@
 		return -EBADFD;
 	}
 
+	rc = snd_pcm_lib_malloc_pages(substream, params_buffer_bytes(hw_params));
+	if (rc < 0)
+		return rc;
+	substream->runtime->buffer_changed = rc > 0;
+
 	rc = virtsnd_pcm_dev_set_params(vss, params_buffer_bytes(hw_params),
 					params_period_bytes(hw_params),
 					params_channels(hw_params),
 					params_format(hw_params),
-					params_rate(hw_params));
+					params_rate(hw_params), epoch);
 	if (rc)
 		return rc;
 
@@ -274,8 +598,14 @@
 	 */
 	virtsnd_pcm_msg_free(vss);
 
-	return virtsnd_pcm_msg_alloc(vss, params_periods(hw_params),
-				     params_period_bytes(hw_params));
+	rc = virtsnd_pcm_msg_alloc(vss, params_periods(hw_params),
+				   params_period_bytes(hw_params));
+	if (!rc && vss->capture_staging) {
+		scoped_guard(spinlock_irqsave, &vss->lock)
+			vss->runtime_active = true;
+	}
+
+	return rc;
 }
 
 /**
@@ -288,28 +618,95 @@
 static int virtsnd_pcm_hw_free(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	int rc;
+
+	guard(mutex)(&vss->lifecycle_lock);
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		vss->stopped |= vss->prepared;
+		vss->prepared = false;
+		vss->runtime_active = false;
+		vss->xfer_enabled = false;
+	}
+	rc = virtsnd_pcm_sync_stop_locked(substream);
+	cancel_work_sync(&vss->elapsed_period);
+
+	if (vss->capture_staging) {
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			vss->runtime_active = false;
+			vss->xfer_enabled = false;
+		}
+		cancel_work_sync(&vss->elapsed_period);
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			vss->hw_ptr = 0;
+			vss->capture_ready = 0;
+			vss->capture_copied_total = 0;
+			vss->capture_acked_total = 0;
+			vss->capture_pointer_seen_total = 0;
+			vss->capture_notified_periods = 0;
+			vss->capture_callback_snapshot = 0;
+			vss->capture_callback_snapshot_active = false;
+			vss->capture_boundary = 0;
+			vss->capture_free_running = false;
+		}
+	}
 
 	/* If the queue is flushed, we can safely free the messages here. */
 	if (!virtsnd_pcm_msg_pending_num(vss))
 		virtsnd_pcm_msg_free(vss);
 
-	return 0;
+	snd_pcm_lib_free_pages(substream);
+	return rc;
 }
 
 /**
- * virtsnd_pcm_prepare() - Prepare the PCM substream.
+ * virtsnd_pcm_prepare_locked() - Prepare the PCM substream.
  * @substream: Kernel ALSA substream.
  *
  * Context: Process context.
  * Return: 0 on success, -errno on failure.
  */
-static int virtsnd_pcm_prepare(struct snd_pcm_substream *substream)
+static int virtsnd_pcm_prepare_locked(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_device *vdev = vss->snd->vdev;
-	struct virtio_snd_msg *msg;
+	bool pause_restart = virtsnd_pcm_pause_restart_enabled(vss);
+	bool free_running;
+	int rc;
 	bool stopped;
-	int rc;
+	bool suspended;
+	u64 epoch;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting)
+			return -ECANCELED;
+		epoch = vss->reset_epoch;
+		stopped = vss->stopped || vss->release_required || vss->msg_count;
+	}
+	if (stopped) {
+		rc = virtsnd_pcm_sync_stop_locked(substream);
+		if (rc)
+			return rc;
+	}
+	scoped_guard(pcm_stream_lock_irqsave, substream) {
+		guard(spinlock_irqsave)(&vss->lock);
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return -ECANCELED;
+		vss->prepared = false;
+		vss->preparing = true;
+		suspended = vss->suspended;
+	}
+
+	if (pause_restart) {
+		enum virtio_pcm_pause_restart_state state;
+
+		scoped_guard(spinlock_irqsave, &vss->lock)
+			state = vss->pause_restart_state;
+		if (state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED ||
+		    virtsnd_pcm_pause_restart_transitioning(state)) {
+			virtsnd_pcm_pause_restart_fail(vss, true, epoch);
+			return -EIO;
+		}
+	}
 
 	scoped_guard(spinlock_irqsave, &vss->lock)
 		stopped = vss->stopped;
@@ -320,48 +717,370 @@
 		return -EBADFD;
 	}
 
-	if (!vss->suspended) {
+	if (!suspended) {
 		if (virtsnd_pcm_msg_pending_num(vss)) {
 			dev_err(&vdev->dev, "SID %u: invalid I/O queue state\n",
 				vss->sid);
 			return -EBADFD;
 		}
 
-		vss->buffer_bytes = snd_pcm_lib_buffer_bytes(substream);
-		vss->hw_ptr = 0;
 	} else {
 		struct snd_pcm_runtime *runtime = substream->runtime;
 		unsigned int buffer_bytes = snd_pcm_lib_buffer_bytes(substream);
 		unsigned int period_bytes = snd_pcm_lib_period_bytes(substream);
-		int rc;
 
 		rc = virtsnd_pcm_dev_set_params(vss, buffer_bytes, period_bytes,
 						runtime->channels,
-						runtime->format, runtime->rate);
-		if (rc)
+						runtime->format, runtime->rate, epoch);
+		if (rc) {
+			if (pause_restart && rc != -ECANCELED)
+				virtsnd_pcm_pause_restart_fail(vss, false, epoch);
 			return rc;
-	}
-
-	vss->xfer_xrun = false;
-	vss->suspended = false;
-	vss->msg_count = 0;
-	virtsnd_pcm_msg_reset(vss);
-
-	memset(&vss->pcm_indirect, 0, sizeof(vss->pcm_indirect));
-	vss->pcm_indirect.sw_buffer_size =
-		vss->pcm_indirect.hw_buffer_size =
-		snd_pcm_lib_buffer_bytes(substream);
-
-	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_PREPARE,
-					GFP_KERNEL);
+		}
+	}
+
+	rc = virtsnd_pcm_dev_prepare(vss, epoch);
+	if (rc) {
+		if (pause_restart && rc != -ECANCELED)
+			virtsnd_pcm_pause_restart_fail(vss, false, epoch);
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			if (vss->reset_epoch == epoch && !vss->resetting) {
+				vss->xfer_xrun = true;
+				vss->stopped = true;
+			}
+		}
+		return rc;
+	}
+
+	scoped_guard(pcm_stream_lock_irqsave, substream) {
+		guard(spinlock_irqsave)(&vss->lock);
+		if (epoch != vss->reset_epoch || vss->resetting)
+			return -ECANCELED;
+		vss->release_required = true;
+		vss->buffer_bytes = snd_pcm_lib_buffer_bytes(substream);
+		vss->hw_ptr = 0;
+		vss->suspended = false;
+		vss->msg_count = 0;
+		virtsnd_pcm_msg_reset(vss);
+		memset(&vss->pcm_indirect, 0, sizeof(vss->pcm_indirect));
+		vss->pcm_indirect.sw_buffer_size = vss->buffer_bytes;
+		vss->pcm_indirect.hw_buffer_size = vss->buffer_bytes;
+		if (vss->capture_staging) {
+			vss->capture_boundary = substream->runtime->boundary;
+			rc = virtsnd_pcm_capture_refresh_mode_locked(vss,
+							     &free_running);
+			if (rc) {
+				vss->xfer_xrun = true;
+				vss->stopped = true;
+				return rc;
+			}
+		}
+		vss->prepared = true;
+		vss->stopped = false;
+		vss->xfer_xrun = false;
+	}
+
+	return rc;
+}
+
+static int virtsnd_pcm_prepare(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+
+	guard(mutex)(&vss->lifecycle_lock);
+	return virtsnd_pcm_prepare_locked(substream);
+}
+
+static int virtsnd_pcm_pause_restart_start(struct virtio_pcm_substream *vss,
+					   bool resume)
+{
+	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+	struct virtio_snd_msg *msg;
+	enum virtio_pcm_pause_restart_state state;
+	bool reset_required;
+	bool invalid_state;
+	int rc = 0;
+	u64 epoch;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting)
+			return -ECANCELED;
+		epoch = vss->reset_epoch;
+		state = vss->pause_restart_state;
+		reset_required =
+			state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED ||
+			virtsnd_pcm_pause_restart_transitioning(state);
+		invalid_state =
+			vss->xfer_xrun ||
+			(resume &&
+			 (state != VIRTSND_PCM_PAUSE_RESTART_RELEASED ||
+			  !vss->stopped || vss->msg_count ||
+			  !vss->runtime_active)) ||
+			(!resume &&
+			 (state != VIRTSND_PCM_PAUSE_RESTART_PREPARED ||
+			  vss->stopped));
+	}
+	if (reset_required) {
+		virtsnd_pcm_pause_restart_fail(vss, true, epoch);
+		return -EIO;
+	}
+	if (invalid_state) {
+		virtsnd_pcm_pause_restart_fail(vss, false, epoch);
+		return -EBADFD;
+	}
+
+	if (resume) {
+		rc = virtsnd_pcm_pause_prepare(vss, epoch);
+		if (rc) {
+			if (rc == -ECANCELED)
+				return rc;
+			virtsnd_pcm_pause_restart_failed(vss, epoch);
+			return rc;
+		}
+
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			if (vss->resetting || vss->reset_epoch != epoch) {
+				rc = -ECANCELED;
+			} else if (vss->pause_restart_state !=
+				    VIRTSND_PCM_PAUSE_RESTART_PREPARED ||
+			    vss->msg_count || !vss->runtime_active) {
+				rc = -EBADFD;
+			} else {
+				/* Preserve ring accounting, restarting only startup grace. */
+				vss->prepared = true;
+				vss->capture_started = false;
+				vss->capture_empty = 0;
+			}
+		}
+		if (rc) {
+			if (rc == -ECANCELED)
+				return rc;
+			virtsnd_pcm_pause_restart_failed(vss, epoch);
+			return rc;
+		}
+	}
+
+	scoped_guard(spinlock_irqsave, &queue->lock) {
+		guard(spinlock)(&vss->lock);
+		if (vss->resetting || vss->reset_epoch != epoch) {
+			rc = -ECANCELED;
+		} else if (vss->pause_restart_state !=
+			    VIRTSND_PCM_PAUSE_RESTART_PREPARED ||
+		    vss->msg_count || !vss->runtime_active || vss->xfer_xrun) {
+			rc = -EBADFD;
+		} else {
+			rc = virtsnd_pcm_capture_send(vss);
+			if (!rc)
+				vss->xfer_enabled = true;
+		}
+	}
+	if (rc) {
+		if (rc == -ECANCELED)
+			return rc;
+		virtsnd_pcm_pause_restart_failed(vss, epoch);
+		return rc;
+	}
+
+	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_START, GFP_KERNEL, epoch);
+	if (!msg) {
+		virtsnd_pcm_pause_restart_failed(vss, epoch);
+		return -ENOMEM;
+	}
+
+	rc = virtsnd_pcm_pause_restart_send(vss, msg,
+					    BIT(VIRTSND_PCM_PAUSE_RESTART_PREPARED),
+					    VIRTSND_PCM_PAUSE_RESTART_START_SUBMITTED,
+					    VIRTSND_PCM_PAUSE_RESTART_RUNNING);
+	if (rc) {
+		if (rc == -ECANCELED)
+			return rc;
+		virtsnd_pcm_pause_restart_failed(vss, epoch);
+		return rc;
+	}
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return -ECANCELED;
+		vss->stopped = false;
+	}
+	virtsnd_pcm_schedule_period_elapsed(vss);
+
+	return 0;
+}
+
+static int virtsnd_pcm_pause_restart_pause(struct virtio_pcm_substream *vss)
+{
+	struct virtio_snd_msg *msg;
+	enum virtio_pcm_pause_restart_state state;
+	int rc;
+	u64 epoch;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting)
+			return -ECANCELED;
+		epoch = vss->reset_epoch;
+		state = vss->pause_restart_state;
+		if (state == VIRTSND_PCM_PAUSE_RESTART_RELEASED) {
+			if (vss->xfer_xrun)
+				virtsnd_pcm_schedule_period_elapsed(vss);
+			return vss->xfer_xrun ? -EPIPE : 0;
+		}
+		vss->xfer_enabled = false;
+		vss->stopped = true;
+	}
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_RELEASE_SENT) {
+		/* Trigger holds the PCM lock; this helper never cancels work. */
+		rc = virtsnd_pcm_release_wait(vss, epoch);
+		if (!rc) {
+			guard(spinlock_irqsave)(&vss->lock);
+			rc = vss->xfer_xrun ? -EPIPE : 0;
+		}
+		if (rc && rc != -ECANCELED)
+			virtsnd_pcm_pause_restart_failed(vss, epoch);
+		return rc;
+	}
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED ||
+	    virtsnd_pcm_pause_restart_transitioning(state)) {
+		virtsnd_pcm_pause_restart_fail(vss, true, epoch);
+		return -EIO;
+	}
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_RUNNING) {
+		msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_STOP,
+						GFP_KERNEL, epoch);
+		if (!msg) {
+			virtsnd_pcm_pause_restart_failed(vss, epoch);
+			return -ENOMEM;
+		}
+
+		rc = virtsnd_pcm_pause_restart_send(vss, msg,
+						    BIT(VIRTSND_PCM_PAUSE_RESTART_RUNNING),
+						    VIRTSND_PCM_PAUSE_RESTART_STOP_SUBMITTED,
+					    VIRTSND_PCM_PAUSE_RESTART_STOPPED);
+		if (rc) {
+			if (rc == -ECANCELED)
+				return rc;
+			virtsnd_pcm_pause_restart_failed(vss, epoch);
+			return rc;
+		}
+		state = VIRTSND_PCM_PAUSE_RESTART_STOPPED;
+	} else if (state != VIRTSND_PCM_PAUSE_RESTART_STOPPED &&
+		   state != VIRTSND_PCM_PAUSE_RESTART_PREPARED) {
+		virtsnd_pcm_pause_restart_failed(vss, epoch);
+		return -EBADFD;
+	}
+
+	rc = virtsnd_pcm_pause_release_flush(vss,
+					     BIT(VIRTSND_PCM_PAUSE_RESTART_STOPPED) |
+					     BIT(VIRTSND_PCM_PAUSE_RESTART_PREPARED), epoch);
+	if (rc) {
+		if (rc == -ECANCELED)
+			return rc;
+		virtsnd_pcm_pause_restart_failed(vss, epoch);
+		return rc;
+	}
+
+	return 0;
+}
+
+static int virtsnd_pcm_pause_restart_stop(struct virtio_pcm_substream *vss,
+					  bool suspend)
+{
+	struct virtio_snd_msg *msg;
+	enum virtio_pcm_pause_restart_state state;
+	int rc;
+	u64 epoch;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting)
+			return -ECANCELED;
+		epoch = vss->reset_epoch;
+		if (!vss->prepared && !vss->msg_count)
+			return 0;
+		if (suspend)
+			vss->suspended = true;
+		vss->xfer_enabled = false;
+		vss->stopped = true;
+		state = vss->pause_restart_state;
+	}
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED ||
+	    virtsnd_pcm_pause_restart_transitioning(state)) {
+		virtsnd_pcm_pause_restart_fail(vss, true, epoch);
+		return -EIO;
+	}
+	if (state != VIRTSND_PCM_PAUSE_RESTART_RUNNING)
+		return 0;
+
+	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_STOP, GFP_KERNEL, epoch);
+	if (!msg) {
+		virtsnd_pcm_pause_restart_failed(vss, epoch);
+		return -ENOMEM;
+	}
+
+	rc = virtsnd_pcm_pause_restart_send(vss, msg,
+					    BIT(VIRTSND_PCM_PAUSE_RESTART_RUNNING),
+					    VIRTSND_PCM_PAUSE_RESTART_STOP_SUBMITTED,
+					    VIRTSND_PCM_PAUSE_RESTART_STOPPED);
+	if (rc && rc != -ECANCELED)
+		virtsnd_pcm_pause_restart_failed(vss, epoch);
+
+	return rc;
+}
+
+static int virtsnd_pcm_stop_device(struct virtio_pcm_substream *vss, u64 epoch)
+{
+	struct virtio_snd_msg *msg;
+	int rc;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return -ECANCELED;
+		if (vss->control_uncertain)
+			return -EIO;
+		if (!vss->device_running)
+			return 0;
+	}
+	msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_STOP,
+					GFP_KERNEL, epoch);
 	if (!msg)
 		return -ENOMEM;
-
-	rc = virtsnd_ctl_msg_send_sync(vss->snd, msg);
-	if (!rc) {
-		guard(spinlock_irqsave)(&vss->lock);
-		vss->release_required = true;
-	}
+	rc = virtsnd_pcm_send_sync(vss, msg);
+	if (rc)
+		return rc;
+	guard(spinlock_irqsave)(&vss->lock);
+	if (vss->resetting || vss->reset_epoch != epoch)
+		return -ECANCELED;
+	vss->device_running = false;
+	return 0;
+}
+
+static int virtsnd_pcm_start_queue(struct virtio_pcm_substream *vss, u64 epoch)
+{
+	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+	int rc = 0;
+
+	guard(spinlock_irqsave)(&queue->lock);
+	guard(spinlock)(&vss->lock);
+	if (vss->resetting || vss->reset_epoch != epoch)
+		return -ECANCELED;
+	if (vss->stopped || !vss->prepared || vss->xfer_xrun)
+		return vss->xfer_xrun ? -EPIPE : -EBADFD;
+	if (vss->control_uncertain)
+		return -EIO;
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE) {
+		if (vss->capture_staging) {
+			rc = vss->runtime_active ? virtsnd_pcm_capture_send(vss) : -EBADFD;
+		} else if (!vss->capture_rx_seeded) {
+			rc = virtsnd_pcm_msg_send(vss, 0, vss->buffer_bytes);
+			if (!rc)
+				vss->capture_rx_seeded = true;
+		}
+	}
+	if (!rc)
+		vss->xfer_enabled = true;
 	return rc;
 }
 
@@ -377,72 +1096,236 @@
 static int virtsnd_pcm_trigger(struct snd_pcm_substream *substream, int command)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
-	struct virtio_snd *snd = vss->snd;
-	struct virtio_snd_queue *queue;
 	struct virtio_snd_msg *msg;
 	int rc = 0;
+	bool running;
+	u64 epoch;
+
+	scoped_guard(spinlock_irqsave, &vss->lock)
+		epoch = vss->reset_epoch;
+
+	if (virtsnd_pcm_pause_restart_enabled(vss)) {
+		switch (command) {
+		case SNDRV_PCM_TRIGGER_START:
+			return virtsnd_pcm_pause_restart_start(vss, false);
+		case SNDRV_PCM_TRIGGER_PAUSE_RELEASE:
+			return virtsnd_pcm_pause_restart_start(vss, true);
+		case SNDRV_PCM_TRIGGER_PAUSE_PUSH:
+			return virtsnd_pcm_pause_restart_pause(vss);
+		case SNDRV_PCM_TRIGGER_STOP:
+			return virtsnd_pcm_pause_restart_stop(vss, false);
+		case SNDRV_PCM_TRIGGER_SUSPEND:
+			return virtsnd_pcm_pause_restart_stop(vss, true);
+		default:
+			return -EINVAL;
+		}
+	}
 
 	switch (command) {
 	case SNDRV_PCM_TRIGGER_START:
 	case SNDRV_PCM_TRIGGER_PAUSE_RELEASE:
-		queue = virtsnd_pcm_queue(vss);
-
-		scoped_guard(spinlock_irqsave, &queue->lock) {
-			guard(spinlock)(&vss->lock);
-			if (vss->stopped) {
-				substream->runtime->stop_operating = true;
-				return -EBADFD;
-			}
-			if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
-			    !vss->capture_rx_seeded) {
-				rc = virtsnd_pcm_msg_send(vss, 0, vss->buffer_bytes);
-				if (!rc)
-					vss->capture_rx_seeded = true;
-			}
-			if (rc) {
-				vss->stopped = true;
-				substream->runtime->stop_operating = true;
-				return rc;
-			}
-			vss->xfer_enabled = true;
+		scoped_guard(spinlock_irqsave, &vss->lock)
+			running = vss->device_running;
+		/* Preserve queued requests and unread RX data across native pause. */
+		rc = virtsnd_pcm_start_queue(vss, epoch);
+		if (rc) {
+			virtsnd_pcm_start_failed(vss, epoch);
+			return rc;
+		}
+
+		if (running) {
+			if (vss->direction == SNDRV_PCM_STREAM_CAPTURE)
+				virtsnd_pcm_schedule_period_elapsed(vss);
+			return 0;
 		}
 
 		msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_START,
-						GFP_KERNEL);
+						GFP_KERNEL, epoch);
 		if (!msg) {
-			virtsnd_pcm_start_failed(vss);
+			virtsnd_pcm_start_failed(vss, epoch);
 			return -ENOMEM;
 		}
 
-		rc = virtsnd_ctl_msg_send_sync(snd, msg);
-		if (rc)
-			virtsnd_pcm_start_failed(vss);
-
+		rc = virtsnd_pcm_send_sync(vss, msg);
+		if (rc && rc != -ECANCELED)
+			virtsnd_pcm_start_failed(vss, epoch);
+
+		if (!rc) {
+			guard(spinlock_irqsave)(&vss->lock);
+			if (vss->resetting || vss->reset_epoch != epoch)
+				return -ECANCELED;
+			vss->device_running = true;
+		}
+		if (!rc && vss->direction == SNDRV_PCM_STREAM_CAPTURE)
+			virtsnd_pcm_schedule_period_elapsed(vss);
 		return rc;
 	case SNDRV_PCM_TRIGGER_SUSPEND:
-		vss->suspended = true;
+		scoped_guard(spinlock_irqsave, &vss->lock)
+			vss->suspended = true;
 		fallthrough;
 	case SNDRV_PCM_TRIGGER_STOP:
 	case SNDRV_PCM_TRIGGER_PAUSE_PUSH:
 		scoped_guard(spinlock_irqsave, &vss->lock) {
+			if (!vss->prepared)
+				return 0;
 			if (command != SNDRV_PCM_TRIGGER_PAUSE_PUSH)
 				vss->stopped = true;
 			vss->xfer_enabled = false;
-		}
-
-		msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_STOP,
-						GFP_KERNEL);
-		if (!msg)
-			return -ENOMEM;
-
-		return virtsnd_ctl_msg_send_sync(snd, msg);
+			if (!vss->device_running)
+				return 0;
+		}
+
+		rc = virtsnd_pcm_stop_device(vss, epoch);
+		if (rc && command != SNDRV_PCM_TRIGGER_PAUSE_PUSH) {
+			guard(spinlock_irqsave)(&vss->lock);
+
+			if (vss->reset_epoch == epoch && !vss->resetting) {
+				vss->xfer_xrun = true;
+				substream->runtime->stop_operating = true;
+				if (vss->direction == SNDRV_PCM_STREAM_CAPTURE)
+					virtsnd_pcm_schedule_period_elapsed(vss);
+			}
+		}
+		if (rc && command == SNDRV_PCM_TRIGGER_PAUSE_PUSH) {
+			if (virtsnd_pcm_start_queue(vss, epoch)) {
+				guard(spinlock_irqsave)(&vss->lock);
+				if (vss->reset_epoch == epoch && !vss->resetting) {
+					vss->xfer_xrun = true;
+					vss->stopped = true;
+					substream->runtime->stop_operating = true;
+					virtsnd_pcm_schedule_period_elapsed(vss);
+				}
+			} else if (vss->direction == SNDRV_PCM_STREAM_CAPTURE) {
+				virtsnd_pcm_schedule_period_elapsed(vss);
+			}
+		}
+		return rc;
 	default:
 		return -EINVAL;
 	}
+
+}
+
+static int
+virtsnd_pcm_pause_restart_sync_stop(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	unsigned int js = msecs_to_jiffies(virtsnd_msg_timeout_ms);
+	enum virtio_pcm_pause_restart_state state;
+	struct virtio_snd_msg *msg;
+	long waited;
+	int rc;
+	u64 epoch;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting)
+			return -ECANCELED;
+		epoch = vss->reset_epoch;
+		vss->xfer_enabled = false;
+		vss->stopped = true;
+		state = vss->pause_restart_state;
+	}
+	cancel_work_sync(&vss->elapsed_period);
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED ||
+	    virtsnd_pcm_pause_restart_transitioning(state)) {
+		virtsnd_pcm_pause_restart_fail(vss, true, epoch);
+		dev_err(&vss->snd->vdev->dev,
+			"SID %u: PCM state is unknown; device reprobe required\n",
+			vss->sid);
+		return -EIO;
+	}
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_NONE ||
+	    state == VIRTSND_PCM_PAUSE_RESTART_CONFIGURED) {
+		if (virtsnd_pcm_msg_pending_num(vss)) {
+			virtsnd_pcm_pause_restart_fail(vss, true, epoch);
+			return -EIO;
+		}
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			if (vss->resetting || vss->reset_epoch != epoch)
+				return -ECANCELED;
+			vss->stopped = false;
+		}
+		return 0;
+	}
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_RUNNING) {
+		msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_STOP,
+						GFP_KERNEL, epoch);
+		if (!msg) {
+			virtsnd_pcm_pause_restart_fail(vss, false, epoch);
+			return -ENOMEM;
+		}
+
+		rc = virtsnd_pcm_pause_restart_send(vss, msg,
+						    BIT(VIRTSND_PCM_PAUSE_RESTART_RUNNING),
+						    VIRTSND_PCM_PAUSE_RESTART_STOP_SUBMITTED,
+						    VIRTSND_PCM_PAUSE_RESTART_STOPPED);
+		if (rc) {
+			if (rc == -ECANCELED)
+				return rc;
+			virtsnd_pcm_pause_restart_fail(vss, false, epoch);
+			return rc;
+		}
+		state = VIRTSND_PCM_PAUSE_RESTART_STOPPED;
+	}
+
+	if (state == VIRTSND_PCM_PAUSE_RESTART_RELEASE_SENT ||
+	    state == VIRTSND_PCM_PAUSE_RESTART_RELEASED) {
+		waited = wait_event_timeout(vss->msg_empty,
+					    !virtsnd_pcm_msg_pending_num(vss), js);
+		if (waited <= 0) {
+			rc = waited ? (int)waited : -ETIMEDOUT;
+			virtsnd_pcm_pause_restart_fail(vss, false, epoch);
+			return rc;
+		}
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			if (vss->resetting || vss->reset_epoch != epoch)
+				return -ECANCELED;
+			if (vss->msg_count) {
+				rc = -ETIMEDOUT;
+			} else {
+				vss->pause_restart_state =
+					VIRTSND_PCM_PAUSE_RESTART_RELEASED;
+				vss->prepared = false;
+				vss->stopped = false;
+				rc = 0;
+			}
+		}
+		if (rc)
+			virtsnd_pcm_pause_restart_fail(vss, false, epoch);
+		return rc;
+	}
+
+	if (state != VIRTSND_PCM_PAUSE_RESTART_PREPARED &&
+	    state != VIRTSND_PCM_PAUSE_RESTART_STOPPED) {
+		virtsnd_pcm_pause_restart_fail(vss, true, epoch);
+		return -EIO;
+	}
+
+	rc = virtsnd_pcm_pause_release_flush(vss,
+					     BIT(VIRTSND_PCM_PAUSE_RESTART_PREPARED) |
+					     BIT(VIRTSND_PCM_PAUSE_RESTART_STOPPED), epoch);
+	if (rc) {
+		if (rc == -ECANCELED)
+			return rc;
+		virtsnd_pcm_pause_restart_fail(vss, false, epoch);
+		return rc;
+	}
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return -ECANCELED;
+		vss->prepared = false;
+		vss->stopped = false;
+	}
+
+	return 0;
 }
 
 /**
- * virtsnd_pcm_sync_stop() - Synchronous PCM substream stop.
+ * virtsnd_pcm_sync_stop_locked() - Synchronous PCM substream stop.
  * @substream: Kernel ALSA substream.
  *
  * The function can be called both from the upper level or from the driver
@@ -451,7 +1334,7 @@
  * Context: Process context. Takes and releases the VirtIO substream spinlock.
  * Return: 0 on success, -errno on failure.
  */
-static int virtsnd_pcm_sync_stop(struct snd_pcm_substream *substream)
+static int virtsnd_pcm_sync_stop_locked(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_snd *snd = vss->snd;
@@ -459,31 +1342,49 @@
 	unsigned int js = msecs_to_jiffies(virtsnd_msg_timeout_ms);
 	bool need_release;
 	int rc;
-
-	scoped_guard(spinlock_irqsave, &vss->lock) {
+	u64 epoch;
+
+	if (virtsnd_pcm_pause_restart_enabled(vss))
+		return virtsnd_pcm_pause_restart_sync_stop(substream);
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting)
+			return -ECANCELED;
+		epoch = vss->reset_epoch;
 		vss->xfer_enabled = false;
 		need_release = vss->release_required;
 		if (need_release)
 			vss->stopped = true;
 	}
 	cancel_work_sync(&vss->elapsed_period);
+	/* A rejected STOP leaves the device running and must precede RELEASE. */
+	rc = virtsnd_pcm_stop_device(vss, epoch);
+	if (rc) {
+		substream->runtime->stop_operating = true;
+		return rc;
+	}
 
 	if (need_release) {
 		msg = virtsnd_pcm_ctl_msg_alloc(vss, VIRTIO_SND_R_PCM_RELEASE,
-						GFP_KERNEL);
+						GFP_KERNEL, epoch);
 		if (!msg) {
 			substream->runtime->stop_operating = true;
 			return -ENOMEM;
 		}
 
-		rc = virtsnd_ctl_msg_send_sync(snd, msg);
+		rc = virtsnd_pcm_send_sync(vss, msg);
 		if (rc) {
 			substream->runtime->stop_operating = true;
 			return rc;
 		}
 
-		scoped_guard(spinlock_irqsave, &vss->lock)
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			if (vss->resetting || vss->reset_epoch != epoch)
+				return -ECANCELED;
 			vss->release_required = false;
+			vss->device_running = false;
+			vss->prepared = false;
+		}
 	}
 
 	/*
@@ -492,7 +1393,7 @@
 	 * Thus, we consider the absence of I/O messages in the queue as an
 	 * indication that the substream has been released.
 	 */
-	rc = wait_event_interruptible_timeout(vss->msg_empty,
+	rc = wait_event_timeout(vss->msg_empty,
 					      !virtsnd_pcm_msg_pending_num(vss),
 					      js);
 	if (rc <= 0) {
@@ -503,8 +1404,12 @@
 		return !rc ? -ETIMEDOUT : rc;
 	}
 
-	scoped_guard(spinlock_irqsave, &vss->lock)
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->resetting || vss->reset_epoch != epoch)
+			return -ECANCELED;
+		vss->prepared = false;
 		vss->stopped = false;
+	}
 
 	return 0;
 }
@@ -521,10 +1426,16 @@
 virtsnd_pcm_pb_pointer(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	size_t hw_ptr;
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (!vss->prepared || vss->preparing)
+			return SNDRV_PCM_POS_XRUN;
+		hw_ptr = vss->hw_ptr;
+	}
 
 	return snd_pcm_indirect_playback_pointer(substream,
-		&vss->pcm_indirect,
-		vss->hw_ptr);
+		&vss->pcm_indirect, hw_ptr);
 }
 
 /**
@@ -539,10 +1450,62 @@
 virtsnd_pcm_cp_pointer(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	size_t hw_ptr = 0;
+	u64 total;
+	bool free_running;
+	bool xrun;
+
+	if (vss->capture_staging) {
+		scoped_guard(spinlock_irqsave, &vss->lock) {
+			xrun = vss->xfer_xrun;
+			if (!xrun &&
+			    virtsnd_pcm_capture_refresh_mode_locked(vss,
+								    &free_running))
+				xrun = true;
+			if (!xrun && !virtsnd_pcm_capture_counters_valid(vss))
+				xrun = true;
+			if (!xrun) {
+				total = vss->capture_callback_snapshot_active ?
+					vss->capture_callback_snapshot :
+					vss->capture_copied_total;
+				/* A full-ring observation gap is modulo-ambiguous. */
+				if (total < vss->capture_pointer_seen_total ||
+				    total - vss->capture_pointer_seen_total >=
+					vss->buffer_bytes) {
+					xrun = true;
+					vss->xfer_xrun = true;
+				} else {
+					vss->capture_pointer_seen_total = total;
+					if (vss->capture_reset_observing) {
+						size_t period = snd_pcm_lib_period_bytes(substream);
+
+						vss->capture_acked_total = total;
+						vss->capture_ready = 0;
+						vss->capture_notified_periods =
+							period ? div64_ul(total, period) : 0;
+					}
+					hw_ptr = virtsnd_pcm_ring_offset(total, vss->buffer_bytes);
+				}
+			}
+			if (xrun)
+				vss->xfer_xrun = true;
+		}
+
+		if (xrun)
+			return SNDRV_PCM_POS_XRUN;
+
+		return bytes_to_frames(substream->runtime, hw_ptr);
+	}
+
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		if (vss->xfer_xrun || !vss->prepared || vss->preparing)
+			return SNDRV_PCM_POS_XRUN;
+		vss->capture_pointer_seen_total = vss->capture_copied_total;
+		hw_ptr = vss->hw_ptr;
+	}
 
 	return snd_pcm_indirect_capture_pointer(substream,
-		&vss->pcm_indirect,
-		vss->hw_ptr);
+		&vss->pcm_indirect, hw_ptr);
 }
 
 struct virtsnd_pcm_transfer {
@@ -557,7 +1520,9 @@
 	struct virtsnd_pcm_transfer *transfer =
 		container_of(rec, struct virtsnd_pcm_transfer, indirect);
 
-	if (transfer->error)
+	if (transfer->error ||
+	    (substream->stream == SNDRV_PCM_STREAM_CAPTURE &&
+	     substream->runtime->state == SNDRV_PCM_STATE_DRAINING))
 		return;
 
 	transfer->error = virtsnd_pcm_msg_send(vss, rec->sw_data, bytes);
@@ -571,7 +1536,7 @@
 	vss->substream->runtime->stop_operating = true;
 }
 
-static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
+static int virtsnd_pcm_pb_ack_locked(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
@@ -580,6 +1545,8 @@
 
 	scoped_guard(spinlock_irqsave, &queue->lock) {
 		guard(spinlock)(&vss->lock);
+		if (!vss->prepared || vss->preparing)
+			return -EBADFD;
 		transfer.indirect = vss->pcm_indirect;
 		if (vss->stopped) {
 			rc = -EPIPE;
@@ -606,17 +1573,136 @@
 	return 0;
 }
 
-static int virtsnd_pcm_cp_ack(struct snd_pcm_substream *substream)
+/* The PCM lock serializes ACK with prepare's accounting reset. */
+static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+
+	if (vss->indirect_reset_observing)
+		return 0;
+	return virtsnd_pcm_pb_ack_locked(substream);
+}
+
+static int virtsnd_pcm_capture_staged_ack(struct virtio_pcm_substream *vss,
+					  struct snd_pcm_runtime *runtime)
+{
+	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+	int rc = 0;
+
+	/*
+	 * The RX callback holds the queue lock while draining completions.
+	 * Account ALSA's read under the substream lock alone so completions later
+	 * in that drain observe the newly released ring space.
+	 */
+	scoped_guard(spinlock_irqsave, &vss->lock) {
+		snd_pcm_uframes_t appl_ptr;
+		snd_pcm_sframes_t diff;
+		size_t bytes;
+		u64 outstanding;
+		u64 next_acked_total;
+		bool free_running;
+
+		if (!vss->runtime_active || vss->preparing || !runtime || !runtime->control ||
+		    !runtime->dma_area || vss->suspended)
+			return -EBADFD;
+		if (!vss->prepared && runtime->state != SNDRV_PCM_STATE_DRAINING &&
+		    !(runtime->state == SNDRV_PCM_STATE_PAUSED &&
+		      virtsnd_pcm_pause_restart_enabled(vss) &&
+		      vss->pause_restart_state == VIRTSND_PCM_PAUSE_RESTART_RELEASED))
+			return -EBADFD;
+		if (vss->xfer_xrun)
+			return -EPIPE;
+		if (vss->stopped && runtime->state != SNDRV_PCM_STATE_DRAINING &&
+		    !(runtime->state == SNDRV_PCM_STATE_PAUSED &&
+		      virtsnd_pcm_pause_restart_enabled(vss) &&
+		      vss->pause_restart_state == VIRTSND_PCM_PAUSE_RESTART_RELEASED))
+			return -EPIPE;
+		rc = virtsnd_pcm_capture_refresh_mode_locked(vss,
+							     &free_running);
+		if (rc || !virtsnd_pcm_capture_counters_valid(vss)) {
+			vss->xfer_xrun = true;
+			virtsnd_pcm_schedule_period_elapsed(vss);
+			return -EPIPE;
+		}
+
+		appl_ptr = runtime->control->appl_ptr;
+		diff = appl_ptr - vss->pcm_indirect.appl_ptr;
+		if (diff < -(snd_pcm_sframes_t)(runtime->boundary / 2))
+			diff += runtime->boundary;
+		if (diff < 0 ||
+		    (snd_pcm_uframes_t)diff > runtime->buffer_size) {
+			vss->xfer_xrun = true;
+			virtsnd_pcm_schedule_period_elapsed(vss);
+			return -EPIPE;
+		}
+
+		/* ACK includes any subperiod tail already copied to the ring. */
+		bytes = frames_to_bytes(runtime, diff);
+		outstanding = vss->capture_copied_total -
+			      vss->capture_acked_total;
+		if ((!free_running && bytes > vss->capture_ready) ||
+		    (free_running &&
+		     (bytes > vss->buffer_bytes || bytes > outstanding)) ||
+		    check_add_overflow(vss->capture_acked_total, (u64)bytes,
+				       &next_acked_total) ||
+		    next_acked_total > vss->capture_pointer_seen_total) {
+			vss->xfer_xrun = true;
+			virtsnd_pcm_schedule_period_elapsed(vss);
+			return -EPIPE;
+		}
+
+		vss->capture_acked_total = next_acked_total;
+		if (!free_running)
+			vss->capture_ready -= bytes;
+		vss->pcm_indirect.appl_ptr = appl_ptr;
+	}
+
+	/* Preserve the queue -> substream lock order while refilling requests. */
+	scoped_guard(spinlock_irqsave, &queue->lock) {
+		guard(spinlock)(&vss->lock);
+		if (vss->runtime_active && vss->xfer_enabled &&
+		    !vss->xfer_xrun) {
+			rc = virtsnd_pcm_capture_send(vss);
+			if (rc) {
+				vss->xfer_xrun = true;
+				virtsnd_pcm_schedule_period_elapsed(vss);
+			}
+		}
+	}
+
+	if (rc)
+		dev_warn_ratelimited(&vss->snd->vdev->dev,
+				     "SID %u: failed to refill capture queue after ACK: %d\n",
+				     vss->sid, rc);
+
+	/* The ACK is committed; a refill failure is reported by the next XRUN. */
+	return 0;
+}
+
+static int virtsnd_pcm_cp_ack_locked(struct snd_pcm_substream *substream)
 {
 	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
 	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+	struct snd_pcm_runtime *runtime = substream->runtime;
 	struct virtsnd_pcm_transfer transfer = {};
 	int rc;
 
+	if (vss->capture_staging) {
+		rc = virtsnd_pcm_capture_staged_ack(vss, runtime);
+		if (rc == -EPIPE) {
+			guard(spinlock_irqsave)(&vss->lock);
+			virtsnd_pcm_capture_arm_paused_stop(vss);
+		}
+		return rc;
+	}
+
 	scoped_guard(spinlock_irqsave, &queue->lock) {
 		guard(spinlock)(&vss->lock);
+		if (!vss->prepared || vss->preparing)
+			return -EBADFD;
 		transfer.indirect = vss->pcm_indirect;
-		if (vss->stopped) {
+		if (vss->xfer_xrun ||
+		    (vss->stopped && runtime->state != SNDRV_PCM_STATE_DRAINING)) {
 			rc = -EPIPE;
 		} else {
 			rc = snd_pcm_indirect_capture_transfer(substream,
@@ -638,6 +1724,217 @@
 		return -EPIPE;
 	}
 
+	return 0;
+}
+
+/* Every linked member's prepare closes the ACK gate under its PCM lock. */
+static int virtsnd_pcm_cp_ack(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	struct snd_pcm_runtime *runtime = substream->runtime;
+
+	/* RESET observes the pointer before returning discarded RX space. */
+	if (vss->indirect_reset_observing)
+		return 0;
+	switch (runtime->state) {
+	case SNDRV_PCM_STATE_XRUN:
+		return -EPIPE;
+	case SNDRV_PCM_STATE_PREPARED:
+	case SNDRV_PCM_STATE_RUNNING:
+	case SNDRV_PCM_STATE_PAUSED:
+	case SNDRV_PCM_STATE_DRAINING:
+		break;
+	default:
+		return -EBADFD;
+	}
+	return virtsnd_pcm_cp_ack_locked(substream);
+}
+
+static int virtsnd_pcm_sync_stop(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+
+	guard(mutex)(&vss->lifecycle_lock);
+	scoped_guard(pcm_stream_lock_irqsave, substream) {
+		snd_pcm_state_t state = substream->runtime->state;
+
+		/*
+		 * Trigger holds the PCM lock through its control requests. Wait
+		 * for it, then leave a replacement active generation alone.
+		 * lifecycle_lock keeps prepare from replacing an inactive one
+		 * until sync_stop_locked has finished its teardown.
+		 */
+		if (state == SNDRV_PCM_STATE_RUNNING ||
+		    state == SNDRV_PCM_STATE_PAUSED ||
+		    state == SNDRV_PCM_STATE_PREPARED ||
+		    (state == SNDRV_PCM_STATE_DRAINING &&
+		     substream->stream == SNDRV_PCM_STREAM_PLAYBACK))
+			return 0;
+		guard(spinlock_irqsave)(&vss->lock);
+		if (!vss->stopped)
+			return 0;
+	}
+	return virtsnd_pcm_sync_stop_locked(substream);
+}
+
+/*
+ * The PCM lock covers indirect accounting and core completion. Once lib_reset
+ * changes hw_ptr, complete the core rebase even if refilling fails. The latched
+ * XRUN rejects subsequent pointer/ACK operations; the worker publishes XRUN to
+ * paused streams too. RESET success therefore does not imply stream recovery.
+ */
+static int virtsnd_pcm_indirect_reset(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	struct snd_pcm_runtime *runtime = substream->runtime;
+	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+	struct virtsnd_pcm_transfer transfer = {};
+	snd_pcm_uframes_t paused_ptr = 0;
+	bool paused = runtime->state == SNDRV_PCM_STATE_PAUSED;
+	int rc;
+
+	/* Only PCM-locked callbacks access this observation flag. */
+	vss->indirect_reset_observing = true;
+	rc = snd_pcm_lib_reset(substream);
+	vss->indirect_reset_observing = false;
+	if (rc)
+		return rc;
+	guard(spinlock_irqsave)(&queue->lock);
+	guard(spinlock)(&vss->lock);
+	if (vss->resetting)
+		return 0;
+	vss->preparing = false;
+	if (vss->xfer_xrun) {
+		virtsnd_pcm_transfer_failed(vss);
+		virtsnd_pcm_schedule_period_elapsed(vss);
+		return 0;
+	}
+	if (!snd_pcm_running(substream) && !paused) {
+		vss->pcm_indirect.appl_ptr = runtime->status->hw_ptr;
+		vss->pcm_indirect.sw_ready = 0;
+		return 0;
+	}
+	if (!vss->prepared || vss->stopped) {
+		virtsnd_pcm_transfer_failed(vss);
+		vss->xfer_xrun = true;
+		virtsnd_pcm_schedule_period_elapsed(vss);
+		return 0;
+	}
+
+	transfer.indirect = vss->pcm_indirect;
+	if (paused) {
+		/* Observe and reclaim paused completions in one queue transaction. */
+		vss->capture_pointer_seen_total = vss->capture_copied_total;
+		vss->indirect_reset_observing = true;
+		paused_ptr = snd_pcm_indirect_capture_pointer(substream,
+							      &transfer.indirect, vss->hw_ptr);
+		vss->indirect_reset_observing = false;
+		if (paused_ptr == SNDRV_PCM_POS_XRUN) {
+			virtsnd_pcm_transfer_failed(vss);
+			vss->xfer_xrun = true;
+			virtsnd_pcm_schedule_period_elapsed(vss);
+			return 0;
+		}
+		runtime->status->hw_ptr = paused_ptr;
+	}
+	transfer.indirect.sw_ready = 0;
+	/* Consume no application delta while returning discarded RX space. */
+	transfer.indirect.appl_ptr = runtime->control->appl_ptr;
+	rc = snd_pcm_indirect_capture_transfer(substream, &transfer.indirect,
+					       virtsnd_pcm_trans_copy);
+	if (!rc)
+		rc = transfer.error;
+	if (rc) {
+		virtsnd_pcm_transfer_failed(vss);
+		vss->xfer_xrun = true;
+		virtsnd_pcm_schedule_period_elapsed(vss);
+		return 0;
+	}
+
+	/* RESET discards all samples through the observed software pointer. */
+	transfer.indirect.sw_ready = 0;
+	transfer.indirect.appl_ptr = runtime->status->hw_ptr;
+	vss->pcm_indirect = transfer.indirect;
+	{
+		size_t period = snd_pcm_lib_period_bytes(substream);
+
+		vss->capture_notified_periods = period ?
+			div64_ul(vss->capture_pointer_seen_total, period) : 0;
+	}
+	return 0;
+}
+
+/* The core holds the stream lock through the matching appl_ptr reset. */
+static int virtsnd_pcm_reset(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	struct snd_pcm_runtime *runtime = substream->runtime;
+	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
+	size_t position;
+	size_t period;
+	int rc;
+
+	/* Reclaim RX space atomically with the pointer observation. */
+	scoped_guard(spinlock_irqsave, &vss->lock)
+		vss->capture_reset_observing = true;
+	rc = snd_pcm_lib_reset(substream);
+	scoped_guard(spinlock_irqsave, &vss->lock)
+		vss->capture_reset_observing = false;
+	if (rc || !vss->capture_staging)
+		return rc;
+
+	guard(spinlock_irqsave)(&queue->lock);
+	guard(spinlock)(&vss->lock);
+	if (vss->resetting)
+		return 0;
+	vss->preparing = false;
+	/* The common reset has already handled a pointer-reported XRUN. */
+	if (vss->xfer_xrun)
+		return 0;
+	if (snd_pcm_running(substream)) {
+		/* Keep completions arriving after the core's pointer observation. */
+		vss->capture_acked_total = vss->capture_pointer_seen_total;
+		vss->capture_ready = vss->capture_free_running ? 0 :
+			vss->capture_copied_total - vss->capture_acked_total;
+	} else {
+		position = frames_to_bytes(runtime, runtime->status->hw_ptr);
+		vss->hw_ptr = position;
+		period = snd_pcm_lib_period_bytes(substream);
+		vss->capture_copied_total = position;
+		vss->capture_acked_total = position;
+		vss->capture_pointer_seen_total = position;
+		vss->capture_ready = 0;
+		vss->capture_notified_periods = period ? position / period : 0;
+	}
+	vss->capture_callback_snapshot = 0;
+	vss->capture_callback_snapshot_active = false;
+	vss->pcm_indirect.appl_ptr = runtime->status->hw_ptr;
+	if (snd_pcm_running(substream) && vss->xfer_enabled && vss->prepared) {
+		rc = virtsnd_pcm_capture_send(vss);
+		if (rc) {
+			vss->xfer_xrun = true;
+			virtsnd_pcm_schedule_period_elapsed(vss);
+		}
+	}
+	/* Complete the matching core pointer rebase before reporting XRUN. */
+	return 0;
+}
+
+/* Keep queued TX ownership while rebasing the application's write cursor. */
+static int virtsnd_pcm_playback_reset(struct snd_pcm_substream *substream)
+{
+	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
+	int rc;
+
+	vss->indirect_reset_observing = true;
+	rc = snd_pcm_lib_reset(substream);
+	vss->indirect_reset_observing = false;
+	if (rc)
+		return rc;
+	guard(spinlock_irqsave)(&vss->lock);
+	vss->preparing = false;
+	vss->pcm_indirect.appl_ptr = substream->runtime->status->hw_ptr;
+	vss->pcm_indirect.sw_ready = 0;
 	return 0;
 }
 
@@ -654,6 +1951,7 @@
 		.sync_stop = virtsnd_pcm_sync_stop,
 		.pointer = virtsnd_pcm_pb_pointer,
 		.ack = virtsnd_pcm_pb_ack,
+		.reset = virtsnd_pcm_playback_reset,
 	},
 	{
 		.open = virtsnd_pcm_open,
@@ -668,3 +1966,32 @@
 		.ack = virtsnd_pcm_cp_ack,
 	},
 };
+
+/* Selected at open; the staging policy remains fixed for that runtime. */
+static const struct snd_pcm_ops virtsnd_pcm_capture_staged_ops = {
+	.open = virtsnd_pcm_open,
+	.close = virtsnd_pcm_close,
+	.ioctl = snd_pcm_lib_ioctl,
+	.reset = virtsnd_pcm_reset,
+	.hw_params = virtsnd_pcm_hw_params,
+	.hw_free = virtsnd_pcm_hw_free,
+	.prepare = virtsnd_pcm_prepare,
+	.trigger = virtsnd_pcm_trigger,
+	.sync_stop = virtsnd_pcm_sync_stop,
+	.pointer = virtsnd_pcm_cp_pointer,
+	.ack = virtsnd_pcm_cp_ack,
+};
+
+static const struct snd_pcm_ops virtsnd_pcm_capture_indirect_ops = {
+	.open = virtsnd_pcm_open,
+	.close = virtsnd_pcm_close,
+	.ioctl = snd_pcm_lib_ioctl,
+	.reset = virtsnd_pcm_indirect_reset,
+	.hw_params = virtsnd_pcm_hw_params,
+	.hw_free = virtsnd_pcm_hw_free,
+	.prepare = virtsnd_pcm_prepare,
+	.trigger = virtsnd_pcm_trigger,
+	.sync_stop = virtsnd_pcm_sync_stop,
+	.pointer = virtsnd_pcm_cp_pointer,
+	.ack = virtsnd_pcm_cp_ack,
+};

[PATCH v12 11/12] ALSA: virtio: prioritize staged capture period work

[Joonho Lee]

Queue staged capture period notifications on system_highpri_wq through
the common scheduling helper. This allows capture notifications to make
progress when ordinary workqueue workers are delayed by CPU contention.
Playback and ordinary capture keep schedule_work().

This preserves the existing priority proposal without mixing RESET or
lifecycle changes into it. It provides no hard real-time deadline and
cannot prevent a userspace reader from overrunning its buffer.

Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/sound/virtio/virtio_pcm.h b/sound/virtio/virtio_pcm.h
--- a/sound/virtio/virtio_pcm.h
+++ b/sound/virtio/virtio_pcm.h
@@ -167,6 +167,10 @@
 static inline bool
 virtsnd_pcm_schedule_period_elapsed(struct virtio_pcm_substream *vss)
 {
+	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
+	    vss->capture_staging)
+		return queue_work(system_highpri_wq, &vss->elapsed_period);
+
 	return schedule_work(&vss->elapsed_period);
 }
 

[PATCH v12 12/12] docs: describe virtio-sound capture options

[Joonho Lee]

Document the five opt-in capture module parameters introduced by this
series. All default to false. Describe the startup-empty, short-payload,
packed-status and pause-restart compatibility behavior alongside staging,
and require validation with the target device.

Assisted-by: LLM
Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>

---
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -8328,6 +8328,36 @@
 			brightness level.
 			default: 1
 
+	virtio_snd.capture_allow_empty= [HW,SOUND]
+			Retry a bounded number of empty startup capture replies.
+			Format: <bool>
+			Default: false. Compatibility option for devices that
+			require staged capture; validate with the target device.
+
+	virtio_snd.capture_allow_short= [HW,SOUND]
+			Accept non-empty, frame-aligned short capture replies.
+			Format: <bool>
+			Default: false. Compatibility option for devices that
+			require staged capture; validate with the target device.
+
+	virtio_snd.capture_packed_status= [HW,SOUND]
+			Read status immediately after the returned capture data.
+			Format: <bool>
+			Default: false. Compatibility option for devices that
+			require staged capture; validate with the target device.
+
+	virtio_snd.capture_pause_restart= [HW,SOUND]
+			Release and reprepare capture streams across pause.
+			Format: <bool>
+			Default: false. Compatibility option for devices that
+			require staged capture; validate with the target device.
+
+	virtio_snd.capture_staging= [HW,SOUND]
+			Use request-owned buffers for capture data.
+			Format: <bool>
+			Default: false. Compatibility option for devices that
+			require staged capture; validate with the target device.
+
 	virtio_mmio.device=
 			[VMMIO] Memory mapped virtio (platform) device.
 

Re: [PATCH v12 01/12] ALSA: virtio: Propagate PCM send failures through ACK

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
- [Medium] ALSA: virtio: race on stopped and stop_operating in virtsnd_pcm_pb_ack()
--

commit adfc01993325cef03121710f81f719a0dc208b01
Author: Joonho Lee <mynameisjune111@gmail.com>

ALSA: virtio: Propagate PCM send failures through ACK

This commit updates virtsnd_pcm_msg_send() to roll back the message length
and notify the virtqueue for previously queued messages when adding a message
fails. It also propagates transfer errors through virtsnd_pcm_pb_ack() and
virtsnd_pcm_cp_ack() so ALSA can handle them as XRUNs.

> diff --git a/sound/virtio/virtio_pcm_ops.c b/sound/virtio/virtio_pcm_ops.c
> index 1091100a4c982..a40ea394bf174 100644
> --- a/sound/virtio/virtio_pcm_ops.c
> +++ b/sound/virtio/virtio_pcm_ops.c
> @@ -466,36 +466,100 @@ virtsnd_pcm_cp_pointer(struct snd_pcm_substream *substream)
>  		vss->hw_ptr);
>  }
>  
> +struct virtsnd_pcm_transfer {
> +	struct snd_pcm_indirect indirect;
> +	int error;
> +};
> +
>  static void virtsnd_pcm_trans_copy(struct snd_pcm_substream *substream,
>  				   struct snd_pcm_indirect *rec, size_t bytes)
>  {
>  	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
> +	struct virtsnd_pcm_transfer *transfer =
> +		container_of(rec, struct virtsnd_pcm_transfer, indirect);
> +
> +	if (transfer->error)
> +		return;
>  
> -	virtsnd_pcm_msg_send(vss, rec->sw_data, bytes);
> +	transfer->error = virtsnd_pcm_msg_send(vss, rec->sw_data, bytes);
> +}
> +
> +/* Caller holds vss->lock. */
> +static void virtsnd_pcm_transfer_failed(struct virtio_pcm_substream *vss)
> +{
> +	vss->xfer_enabled = false;
> +	vss->stopped = true;
> +	vss->substream->runtime->stop_operating = true;
>  }
>  
>  static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
>  {
>  	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
>  	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
> +	struct virtsnd_pcm_transfer transfer = {};
> +	int rc;
>  
> -	guard(spinlock_irqsave)(&queue->lock);
> -	guard(spinlock)(&vss->lock);
> +	scoped_guard(spinlock_irqsave, &queue->lock) {
> +		guard(spinlock)(&vss->lock);
> +		transfer.indirect = vss->pcm_indirect;
> +		if (vss->stopped) {
> +			rc = -EPIPE;
> +		} else {
> +			rc = snd_pcm_indirect_playback_transfer(substream,
> +								&transfer.indirect,
> +								virtsnd_pcm_trans_copy);
> +		}
> +		if (!rc)
> +			rc = transfer.error;
> +		if (rc)
> +			virtsnd_pcm_transfer_failed(vss);
> +		else
> +			vss->pcm_indirect = transfer.indirect;
> +	}

[Severity: Medium]
Can virtsnd_pcm_pb_ack() and virtsnd_pcm_cp_ack() race with
snd_pcm_sync_stop() and virtsnd_pcm_sync_stop() on vss->stopped and
substream->runtime->stop_operating?

Because virtio-snd PCM devices are nonatomic (vpcm->pcm->nonatomic = true),
snd_pcm_sync_stop() is invoked without holding the PCM stream lock
(substream->self_group.mutex), such as from snd_pcm_drop() after dropping
pcm_stream_lock_irq() or from snd_pcm_do_prepare() under
runtime->buffer_mutex. Meanwhile, a concurrent SNDRV_PCM_IOCTL_SYNC_PTR ioctl
in snd_pcm_sync_ptr() acquires only pcm_stream_lock_irq() and calls:

snd_pcm_sync_ptr()
  pcm_lib_apply_appl_ptr()
    virtsnd_pcm_pb_ack() / virtsnd_pcm_cp_ack()

In sound/virtio/virtio_pcm_ops.c, virtsnd_pcm_sync_stop() reads and clears
vss->stopped without holding vss->lock or the PCM stream lock, after
snd_pcm_sync_stop() clears substream->runtime->stop_operating:

static int virtsnd_pcm_sync_stop(struct snd_pcm_substream *substream)
{
	...
	cancel_work_sync(&vss->elapsed_period);

	if (!vss->stopped)
		return 0;
	...
	vss->stopped = false;

	return 0;
}

If a concurrent ACK runs while vss->stopped is true, virtsnd_pcm_pb_ack() or
virtsnd_pcm_cp_ack() sets rc = -EPIPE and calls virtsnd_pcm_transfer_failed(),
which writes vss->stopped = true and substream->runtime->stop_operating = true.
That races with virtsnd_pcm_sync_stop() and re-arms stop_operating after
snd_pcm_sync_stop() cleared it, while returning -EPIPE so
pcm_lib_apply_appl_ptr() invokes __snd_pcm_xrun() even when snd_pcm_drop() had
placed the stream in SNDRV_PCM_STATE_SETUP.

Conversely, virtsnd_pcm_pb_ack() and virtsnd_pcm_cp_ack() only check
vss->stopped before transferring data. Once virtsnd_pcm_sync_stop() clears
vss->stopped = false after releasing the stream in SNDRV_PCM_STATE_SETUP, a
subsequent SNDRV_PCM_IOCTL_SYNC_PTR ioctl sees vss->stopped == false and
enqueues virtio I/O messages on a released stream.

Later commits in this series ("ALSA: virtio: Recover PCM state after a failed
start" and "ALSA: virtio: add staged capture and serialize PCM lifecycle
accounting") partially address this by guarding vss->stopped with vss->lock
and gating ACK on vss->prepared and vss->preparing, but do not completely
resolve the race. In virtsnd_pcm_sync_stop_locked() (and
virtsnd_pcm_pause_restart_sync_stop()) as well as virtsnd_pcm_prepare_locked(),
vss->prepared remains true while vss->stopped is true across sleeping teardown
calls (cancel_work_sync(), virtsnd_pcm_stop_device(), and
virtsnd_pcm_send_sync() for VIRTIO_SND_R_PCM_RELEASE), and
virtsnd_pcm_pb_ack() does not check runtime->state. A concurrent
SNDRV_PCM_IOCTL_SYNC_PTR during that window still sees vss->prepared == true
and vss->stopped == true, calls virtsnd_pcm_transfer_failed() to set
runtime->stop_operating = true after snd_pcm_sync_stop() cleared it, and
returns -EPIPE to trigger __snd_pcm_xrun().

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261103000403.67106-1-mynameisjune111@gmail.com?part=1

Re: [PATCH v12 10/12] ALSA: virtio: add staged capture and serialize PCM lifecycle accounting

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 10 potential issue(s) to consider:
- [High] ALSA: virtio: broken indirect ring state in virtsnd_pcm_playback_reset()
- [High] ALSA: virtio: post-stop capture RX not retired in virtsnd_pcm_msg_complete()
- [Medium] ALSA: virtio: missing prepared and preparing checks in virtsnd_pcm_cp_pointer()
- [Medium] ALSA: virtio: vss->preparing gate leaked on prepare failure and reset races
- [Medium] ALSA: virtio: stale vss->suspended flag across open and virtsnd_pcm_hw_params()
- [Medium] ALSA: virtio: sub-byte frame truncation in virtsnd_pcm_capture_staged_ack()
- [Medium] ALSA: virtio: missing state and xrun checks in virtsnd_pcm_pb_ack() and pointer
- [Medium] ALSA: virtio: stale vss->suspended across virtsnd_pcm_open() and hw_params()
- [Medium] ALSA: virtio: virtsnd_pcm_pause_restart_start() returns -EBADFD on xfer_xrun
- [Medium] ALSA: virtio: incomplete xrun and pointer rebase in virtsnd_pcm_reset()
--

commit aed4d30e1eeffd3a8fd687d7393b3bfcaedf5c71
Author: Joonho Lee <mynameisjune111@gmail.com>

ALSA: virtio: add staged capture and serialize PCM lifecycle accounting

This commit adds opt-in staged PCM capture using request-owned DMA buffers and
serializes PCM lifecycle, reset, and acknowledgment accounting under the queue
and substream locks.

Link: https://sashiko.dev/#/patchset/20261031020855.2402-1-mynameisjune111%40gmail.com

> diff --git a/sound/virtio/virtio_pcm_msg.c b/sound/virtio/virtio_pcm_msg.c
> index b94ab24f5ffef..ed340ecfb289e 100644
> --- a/sound/virtio/virtio_pcm_msg.c
> +++ b/sound/virtio/virtio_pcm_msg.c

[ ... ]

> @@ -540,19 +880,52 @@ static void virtsnd_pcm_msg_complete(struct virtio_pcm_msg *msg,
>  				     size_t written_bytes)
>  {
>  	struct virtio_pcm_substream *vss = msg->substream;
> +	size_t completed_bytes;
> +	u32 latency_bytes;
> +
> +	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
> +	    vss->capture_staging) {
> +		virtsnd_pcm_capture_complete(msg, written_bytes);
> +		return;
> +	}
>  
> -	/*
> -	 * hw_ptr always indicates the buffer position of the first I/O message
> -	 * in the virtqueue. Therefore, on each completion of an I/O message,
> -	 * the hw_ptr value is unconditionally advanced.
> -	 */
> +	/* Only completed TX requests and valid RX audio advance the pointer. */
>  	guard(spinlock)(&vss->lock);
> +	/* Empty STOP returns retire ownership without inventing captured data. */
> +	if (vss->direction == SNDRV_PCM_STREAM_CAPTURE &&
> +	    (!vss->prepared ||
> +	     (vss->stopped && written_bytes <= sizeof(*msg->status)))) {
> +		msg->length = 0;
> +		vss->msg_count--;
> +		if (!vss->msg_count)
> +			wake_up_all(&vss->msg_empty);
> +		return;
> +	}

[Severity: High]
As reported by Sashiko on v11 [1], what happens in virtsnd_pcm_msg_complete()
when an in-flight ordinary capture RX buffer completes with
written_bytes > sizeof(*msg->status) after vss->stopped has been set?

On an ordinary capture substream, SNDRV_PCM_TRIGGER_STOP (called during
capture drain in snd_pcm_do_drain_init(), snd_pcm_drop(), or
virtsnd_pcm_sync_stop_locked()) sets vss->stopped = true while vss->prepared
remains true until VIRTIO_SND_R_PCM_RELEASE completes. Unlike staged capture
in virtsnd_pcm_capture_complete(), which ignores all completions when
vss->stopped is true, virtsnd_pcm_msg_complete() only retires stopped capture
completions early if written_bytes <= sizeof(*msg->status).

During SNDRV_PCM_STATE_DRAINING, ALSA stops calling ops->pointer(), so
vss->capture_pointer_seen_total stops advancing while the application drains
already-captured frames via virtsnd_pcm_cp_ack_locked(). If pending RX buffers
complete with audio payloads after STOP, they advance
vss->capture_copied_total until
vss->capture_copied_total - vss->capture_pointer_seen_total >= vss->buffer_bytes
(or fail virtsnd_pcm_capture_status() if completed with a non-OK status),
latching vss->xfer_xrun = true and causing virtsnd_pcm_cp_ack_locked() to
abort capture drain with -EPIPE.

In addition, during virtsnd_pcm_sync_stop_locked(), cancel_work_sync() runs
before VIRTIO_SND_R_PCM_STOP and VIRTIO_SND_R_PCM_RELEASE are sent:

virtsnd_pcm_sync_stop_locked()
    cancel_work_sync(&vss->elapsed_period);
    rc = virtsnd_pcm_stop_device(vss, epoch);
    ...
    rc = virtsnd_pcm_send_sync(vss, msg); /* VIRTIO_SND_R_PCM_RELEASE */
    scoped_guard(spinlock_irqsave, &vss->lock) {
        ...
        vss->prepared = false;
    }

If a non-empty RX completion arrives while STOP or RELEASE is in flight and
vss->xfer_xrun is set, virtsnd_pcm_msg_complete() will call
virtsnd_pcm_schedule_period_elapsed(vss) after cancel_work_sync() has already
returned, and will also skip the else-if wake_up_all(&vss->msg_empty) when
vss->msg_count reaches zero. Should virtsnd_pcm_msg_complete() retire all
capture completions whenever !vss->prepared || vss->stopped?

> diff --git a/sound/virtio/virtio_pcm_ops.c b/sound/virtio/virtio_pcm_ops.c
> index ad602b6b6530b..54065662083ed 100644
> --- a/sound/virtio/virtio_pcm_ops.c
> +++ b/sound/virtio/virtio_pcm_ops.c

[ ... ]

> @@ -119,23 +381,42 @@ static int virtsnd_pcm_open(struct snd_pcm_substream *substream)

[ ... ]

>  	snd_pcm_hw_constraint_integer(substream->runtime,
>  				      SNDRV_PCM_HW_PARAM_PERIODS);
>  
>  	scoped_guard(spinlock_irqsave, &vss->lock)
>  		vss->stopped |= !!vss->msg_count;
> -	vss->suspended = false;

[Severity: Medium]
If vss->suspended = false is removed from virtsnd_pcm_open() and is not
cleared in virtsnd_pcm_hw_params() either, can vss->suspended remain stale
until virtsnd_pcm_prepare_locked()?

When virtsnd_reset_vqs() runs during device reset or freeze/restore, it sets
vss->suspended = true on all substreams (including closed ones), and closing a
substream after SNDRV_PCM_TRIGGER_SUSPEND also leaves vss->suspended = true.
When the substream is subsequently opened via virtsnd_pcm_open() and
configured via virtsnd_pcm_hw_params(), virtsnd_pcm_hw_params() sends
VIRTIO_SND_R_PCM_SET_PARAMS to the device, but vss->suspended remains true.

When virtsnd_pcm_prepare_locked() then runs, it sees suspended == true, which
skips the !virtsnd_pcm_msg_pending_num(vss) queue sanity check and sends a
redundant VIRTIO_SND_R_PCM_SET_PARAMS request to the device. Should
vss->suspended be cleared in virtsnd_pcm_open() or when
virtsnd_pcm_dev_set_params() succeeds in virtsnd_pcm_hw_params()?

[ ... ]

> @@ -320,129 +717,615 @@ static int virtsnd_pcm_prepare(struct snd_pcm_substream *substream)
> +	scoped_guard(pcm_stream_lock_irqsave, substream) {
> +		guard(spinlock_irqsave)(&vss->lock);
> +		if (vss->resetting || vss->reset_epoch != epoch)
> +			return -ECANCELED;
> +		vss->prepared = false;
> +		vss->preparing = true;
> +		suspended = vss->suspended;
> +	}
> +
> +	if (pause_restart) {
> +		enum virtio_pcm_pause_restart_state state;
> +
> +		scoped_guard(spinlock_irqsave, &vss->lock)
> +			state = vss->pause_restart_state;
> +		if (state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED ||
> +		    virtsnd_pcm_pause_restart_transitioning(state)) {
> +			virtsnd_pcm_pause_restart_fail(vss, true, epoch);
> +			return -EIO;
> +		}
> +	}

[Severity: Medium]
If virtsnd_pcm_prepare_locked() fails after setting vss->preparing = true,
what clears vss->preparing?

In snd_pcm_do_prepare(), the core only calls snd_pcm_reset_substream() (which
invokes substream->ops->reset() and clears vss->preparing = false) if
substream->ops->prepare() returns 0. If virtsnd_pcm_prepare_locked() returns
an error on any of the subsequent failure paths (such as the pause_restart
state check, the stopped check, pending messages, virtsnd_pcm_dev_set_params(),
virtsnd_pcm_dev_prepare(), or virtsnd_pcm_capture_refresh_mode_locked()),
vss->preparing remains stuck at true.

Similarly, in virtsnd_pcm_indirect_reset(), virtsnd_pcm_reset(), and
virtsnd_reset_vqs(), if a device reset occurs (setting vss->resetting = true),
virtsnd_pcm_indirect_reset() and virtsnd_pcm_reset() return early at
if (vss->resetting) return 0; before clearing vss->preparing = false, and
virtsnd_reset_vqs() does not clear vss->preparing either. While vss->preparing
is stuck at true, pointer and ACK callbacks remain locked out and return
SNDRV_PCM_POS_XRUN or -EBADFD instead of -EPIPE.

[ ... ]

> +static int virtsnd_pcm_pause_restart_start(struct virtio_pcm_substream *vss,
> +					   bool resume)
> +{
> +	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
> +	struct virtio_snd_msg *msg;
> +	enum virtio_pcm_pause_restart_state state;
> +	bool reset_required;
> +	bool invalid_state;
> +	int rc = 0;
> +	u64 epoch;
> +
> +	scoped_guard(spinlock_irqsave, &vss->lock) {
> +		if (vss->resetting)
> +			return -ECANCELED;
> +		epoch = vss->reset_epoch;
> +		state = vss->pause_restart_state;
> +		reset_required =
> +			state == VIRTSND_PCM_PAUSE_RESTART_RESET_REQUIRED ||
> +			virtsnd_pcm_pause_restart_transitioning(state);
> +		invalid_state =
> +			vss->xfer_xrun ||
> +			(resume &&\n+			 (state != VIRTSND_PCM_PAUSE_RESTART_RELEASED ||
> +			  !vss->stopped || vss->msg_count ||
> +			  !vss->runtime_active)) ||
> +			(!resume &&\n+			 (state != VIRTSND_PCM_PAUSE_RESTART_PREPARED ||
> +			  vss->stopped));
> +	}
> +	if (reset_required) {
> +		virtsnd_pcm_pause_restart_fail(vss, true, epoch);
> +		return -EIO;
> +	}
> +	if (invalid_state) {
> +		virtsnd_pcm_pause_restart_fail(vss, false, epoch);
> +		return -EBADFD;
> +	}

[Severity: Medium]
In virtsnd_pcm_pause_restart_start(), if vss->xfer_xrun is already latched
(for example via VIRTIO_SND_EVT_PCM_XRUN in virtsnd_pcm_event() while prepared
or before queueing in the second check below), should this return -EPIPE
instead of -EBADFD, matching virtsnd_pcm_start_queue()?

In snd_pcm_do_start(), the ALSA core only transitions the runtime state to
SNDRV_PCM_STATE_XRUN when trigger(SNDRV_PCM_TRIGGER_START) returns -EPIPE.
Returning -EBADFD leaves the substream in SNDRV_PCM_STATE_PREPARED with
vss->stopped = true and vss->xfer_xrun = true, where
virtsnd_pcm_period_elapsed() will not move it to XRUN either because
snd_pcm_running() is false.

[ ... ]

> @@ -521,10 +1426,16 @@ static snd_pcm_uframes_t
>  virtsnd_pcm_pb_pointer(struct snd_pcm_substream *substream)
>  {
>  	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
> +	size_t hw_ptr;
> +
> +	scoped_guard(spinlock_irqsave, &vss->lock) {
> +		if (!vss->prepared || vss->preparing)
> +			return SNDRV_PCM_POS_XRUN;
> +		hw_ptr = vss->hw_ptr;
> +	}
>  
>  	return snd_pcm_indirect_playback_pointer(substream,
> -		&vss->pcm_indirect,
> -		vss->hw_ptr);
> +		&vss->pcm_indirect, hw_ptr);
>  }
>  
>  /**
> @@ -539,10 +1450,62 @@ static snd_pcm_uframes_t
>  virtsnd_pcm_cp_pointer(struct snd_pcm_substream *substream)
>  {
>  	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
> +	size_t hw_ptr = 0;
> +	u64 total;
> +	bool free_running;
> +	bool xrun;
> +
> +	if (vss->capture_staging) {
> +		scoped_guard(spinlock_irqsave, &vss->lock) {
> +			xrun = vss->xfer_xrun;
> +			if (!xrun &&
> +			    virtsnd_pcm_capture_refresh_mode_locked(vss,
> +								    &free_running))
> +				xrun = true;
> +			if (!xrun && !virtsnd_pcm_capture_counters_valid(vss))
> +				xrun = true;

[Severity: Medium]
Unlike virtsnd_pcm_pb_pointer(), the non-staged branch of
virtsnd_pcm_cp_pointer(), and virtsnd_pcm_capture_staged_ack(), the staged
capture branch of virtsnd_pcm_cp_pointer() does not check vss->prepared,
vss->preparing, or vss->suspended.

When virtsnd_reset_vqs() runs during a device reset, it sets
vss->prepared = false, vss->xfer_enabled = false, vss->suspended = true, and
vss->xfer_xrun = false, while leaving vss->runtime_active, vss->buffer_bytes,
vss->capture_boundary, and the staged capture counters intact. Likewise,
virtsnd_pcm_prepare_locked() sets vss->prepared = false and
vss->preparing = true before dropping the PCM stream lock. Because neither
virtsnd_pcm_capture_refresh_mode_locked() nor
virtsnd_pcm_capture_counters_valid() checks vss->prepared, vss->preparing, or
vss->suspended, won't pointer queries on a reset or preparing staged capture
substream return a stale ring offset instead of SNDRV_PCM_POS_XRUN?

[ ... ]

> @@ -571,7 +1536,7 @@ static void virtsnd_pcm_transfer_failed(struct virtio_pcm_substream *vss)
>  	vss->substream->runtime->stop_operating = true;
>  }
>  
> -static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
> +static int virtsnd_pcm_pb_ack_locked(struct snd_pcm_substream *substream)
>  {
>  	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
>  	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
> @@ -580,6 +1545,8 @@ static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
>  
>  	scoped_guard(spinlock_irqsave, &queue->lock) {
>  		guard(spinlock)(&vss->lock);
> +		if (!vss->prepared || vss->preparing)
> +			return -EBADFD;
>  		transfer.indirect = vss->pcm_indirect;
>  		if (vss->stopped) {
>  			rc = -EPIPE;
> @@ -606,17 +1573,136 @@ static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
>  	return 0;
>  }
>  
> -static int virtsnd_pcm_cp_ack(struct snd_pcm_substream *substream)
> +/* The PCM lock serializes ACK with prepare's accounting reset. */
> +static int virtsnd_pcm_pb_ack(struct snd_pcm_substream *substream)
>  {
>  	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
> +
> +	if (vss->indirect_reset_observing)
> +		return 0;
> +	return virtsnd_pcm_pb_ack_locked(substream);
> +}

[Severity: Medium]
In virtsnd_pcm_cp_ack(), the capture path validates runtime->state (returning
-EPIPE for SNDRV_PCM_STATE_XRUN and -EBADFD for states other than PREPARED,
RUNNING, PAUSED, and DRAINING), and virtsnd_pcm_cp_ack_locked() and
virtsnd_pcm_cp_pointer() check vss->xfer_xrun.

In contrast, virtsnd_pcm_pb_ack() does not check runtime->state, and neither
virtsnd_pcm_pb_ack_locked() nor virtsnd_pcm_pb_pointer() checks
vss->xfer_xrun. If a prepared playback stream is dropped via
SNDRV_PCM_IOCTL_DROP (where snd_pcm_drop() transitions runtime->state to
SNDRV_PCM_STATE_SETUP without calling trigger(STOP) because snd_pcm_running()
is false, leaving vss->prepared == true and vss->stopped == false), or if a
later member in a linked prepare sleeps or fails after playback prepare
completes, or if VIRTIO_SND_EVT_PCM_XRUN sets vss->xfer_xrun = true, won't
SNDRV_PCM_IOCTL_SYNC_PTR still call virtsnd_pcm_pb_ack() and enqueue TX
buffers in invalid or XRUN states?

> +
> +static int virtsnd_pcm_capture_staged_ack(struct virtio_pcm_substream *vss,
> +					  struct snd_pcm_runtime *runtime)
> +{

[ ... ]

> +		/* ACK includes any subperiod tail already copied to the ring. */
> +		bytes = frames_to_bytes(runtime, diff);
> +		outstanding = vss->capture_copied_total -
> +			      vss->capture_acked_total;
> +		if ((!free_running && bytes > vss->capture_ready) ||
> +		    (free_running &&
> +		     (bytes > vss->buffer_bytes || bytes > outstanding)) ||
> +		    check_add_overflow(vss->capture_acked_total, (u64)bytes,
> +				       &next_acked_total) ||
> +		    next_acked_total > vss->capture_pointer_seen_total) {
> +			vss->xfer_xrun = true;
> +			virtsnd_pcm_schedule_period_elapsed(vss);
> +			return -EPIPE;
> +		}
> +
> +		vss->capture_acked_total = next_acked_total;
> +		if (!free_running)
> +			vss->capture_ready -= bytes;
> +		vss->pcm_indirect.appl_ptr = appl_ptr;
> +	}

[Severity: Medium]
Can frames_to_bytes(runtime, diff) truncate sub-byte frame deltas here when
capturing 4-bit frames (mono SNDRV_PCM_FORMAT_IMA_ADPCM, where
runtime->frame_bits == 4)?

When diff is odd, frames_to_bytes(runtime, diff) computes (diff * 4) / 8 and
truncates the half-byte, while vss->pcm_indirect.appl_ptr = appl_ptr advances
past the full frame delta. Across multiple odd-frame ACKs (for example via
SNDRV_PCM_IOCTL_SYNC_PTR), the uncounted half-bytes are never added to
vss->capture_acked_total or subtracted from vss->capture_ready, leaking
vss->capture_ready in managed mode until virtsnd_pcm_capture_complete()
rejects a completion with ring_full and triggers a false XRUN.

[ ... ]

> +/* The core holds the stream lock through the matching appl_ptr reset. */
> +static int virtsnd_pcm_reset(struct snd_pcm_substream *substream)
> +{
> +	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
> +	struct snd_pcm_runtime *runtime = substream->runtime;
> +	struct virtio_snd_queue *queue = virtsnd_pcm_queue(vss);
> +	size_t position;
> +	size_t period;
> +	int rc;
> +
> +	/* Reclaim RX space atomically with the pointer observation. */
> +	scoped_guard(spinlock_irqsave, &vss->lock)
> +		vss->capture_reset_observing = true;
> +	rc = snd_pcm_lib_reset(substream);
> +	scoped_guard(spinlock_irqsave, &vss->lock)
> +		vss->capture_reset_observing = false;
> +	if (rc || !vss->capture_staging)
> +		return rc;
> +
> +	guard(spinlock_irqsave)(&queue->lock);
> +	guard(spinlock)(&vss->lock);
> +	if (vss->resetting)
> +		return 0;
> +	vss->preparing = false;
> +	/* The common reset has already handled a pointer-reported XRUN. */
> +	if (vss->xfer_xrun)
> +		return 0;

[Severity: Medium]
As reported by Sashiko on v8 [2], what happens in virtsnd_pcm_reset() (and
virtsnd_pcm_indirect_reset()) when vss->xfer_xrun is true, either while the
stream is in SNDRV_PCM_STATE_PAUSED / SNDRV_PCM_STATE_PREPARED or when
vss->xfer_xrun becomes true between snd_pcm_lib_reset() and acquiring
vss->lock?

In snd_pcm_lib_reset(), snd_pcm_update_hw_ptr() is only called when
snd_pcm_running(substream) is true. When the substream is paused or prepared
(or if vss->xfer_xrun raced in after snd_pcm_lib_reset()), snd_pcm_lib_reset()
does not observe the XRUN and returns 0, and snd_pcm_reset_substream() then
rebases runtime->control->appl_ptr = runtime->status->hw_ptr. Returning 0
early here leaves vss->pcm_indirect.appl_ptr and the staged capture counters
out of sync with runtime->control->appl_ptr, and in virtsnd_pcm_reset() also
skips scheduling vss->elapsed_period to stop a paused substream on XRUN.

> +	if (snd_pcm_running(substream)) {
> +		/* Keep completions arriving after the core's pointer observation. */
> +		vss->capture_acked_total = vss->capture_pointer_seen_total;
> +		vss->capture_ready = vss->capture_free_running ? 0 :
> +			vss->capture_copied_total - vss->capture_acked_total;
> +	} else {
> +		position = frames_to_bytes(runtime, runtime->status->hw_ptr);
> +		vss->hw_ptr = position;
> +		period = snd_pcm_lib_period_bytes(substream);
> +		vss->capture_copied_total = position;
> +		vss->capture_acked_total = position;
> +		vss->capture_pointer_seen_total = position;
> +		vss->capture_ready = 0;
> +		vss->capture_notified_periods = period ? position / period : 0;
> +	}
> +	vss->capture_callback_snapshot = 0;
> +	vss->capture_callback_snapshot_active = false;
> +	vss->pcm_indirect.appl_ptr = runtime->status->hw_ptr;
> +	if (snd_pcm_running(substream) && vss->xfer_enabled && vss->prepared) {
> +		rc = virtsnd_pcm_capture_send(vss);
> +		if (rc) {
> +			vss->xfer_xrun = true;
> +			virtsnd_pcm_schedule_period_elapsed(vss);
> +		}
> +	}
> +	/* Complete the matching core pointer rebase before reporting XRUN. */
> +	return 0;
> +}
> +
> +/* Keep queued TX ownership while rebasing the application's write cursor. */
> +static int virtsnd_pcm_playback_reset(struct snd_pcm_substream *substream)
> +{
> +	struct virtio_pcm_substream *vss = snd_pcm_substream_chip(substream);
> +	int rc;
> +
> +	vss->indirect_reset_observing = true;
> +	rc = snd_pcm_lib_reset(substream);
> +	vss->indirect_reset_observing = false;
> +	if (rc)
> +		return rc;
> +	guard(spinlock_irqsave)(&vss->lock);
> +	vss->preparing = false;
> +	vss->pcm_indirect.appl_ptr = substream->runtime->status->hw_ptr;
> +	vss->pcm_indirect.sw_ready = 0;
> +	return 0;
> +}

[Severity: High]
Does virtsnd_pcm_playback_reset() leave vss->pcm_indirect and
vss->msgs[i]->length in an inconsistent state when SNDRV_PCM_IOCTL_RESET is
called after playback frames have already been acknowledged?

In virtsnd_pcm_pb_ack_locked(), snd_pcm_indirect_playback_transfer() drains
sw_ready into virtsnd_pcm_msg_send(), which advances pcm_indirect.sw_data,
pcm_indirect.hw_data, and pcm_indirect.hw_ready ahead of pcm_indirect.sw_io
and accumulates bytes into vss->msgs[i]->length. When
virtsnd_pcm_playback_reset() runs, snd_pcm_reset_substream() rebases
runtime->control->appl_ptr back to runtime->status->hw_ptr (matching sw_io),
and virtsnd_pcm_playback_reset() sets pcm_indirect.appl_ptr and clears
sw_ready = 0, but leaves sw_data, hw_data, hw_ready, and vss->msgs[i]->length
untouched:

1. On the next write after RESET, ALSA writes new samples into
   runtime->dma_area starting at runtime->control->appl_ptr (sw_io, overwriting
   in-flight TX period buffers), while snd_pcm_indirect_playback_transfer()
   calls virtsnd_pcm_msg_send() starting from the old sw_data offset, sending
   unwritten/stale DMA buffer regions.

2. If a sub-period write occurred before RESET (leaving
   0 < msg->length < period_bytes) or when sw_data wraps into an already-queued
   period (msg->length == period_bytes), virtsnd_pcm_msg_send() adds n to the
   existing msg->length, causing msg->length to exceed period_bytes so the
   descriptor is never submitted.

3. Because runtime->control->appl_ptr is rebased back to
   runtime->status->hw_ptr while hw_ready bytes remain queued in the TX
   virtqueue, completion of an in-flight TX period advances
   runtime->status->hw_ptr past runtime->control->appl_ptr and triggers a
   false playback XRUN.

4. If SNDRV_PCM_IOCTL_RESET is called while playback is in
   SNDRV_PCM_STATE_PAUSED after completing at least one period (vss->hw_ptr > 0),
   snd_pcm_lib_reset() sees !snd_pcm_running(substream) and resets
   runtime->status->hw_ptr = 0, while vss->hw_ptr, pcm_indirect.hw_io, and
   pcm_indirect.sw_io remain at the non-zero paused offset, causing the next
   virtsnd_pcm_pb_pointer() call to report a large pointer jump and XRUN.

[1] https://lore.kernel.org/r/20261031020855.2402-10-mynameisjune111@gmail.com
[2] https://lore.kernel.org/r/sashiko-outbox-188791@kernel.org

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261103000403.67106-1-mynameisjune111@gmail.com?part=10

Re: [PATCH v12 12/12] docs: describe virtio-sound capture options

[Randy Dunlap]

Hi,

On 11/2/26 4:04 PM, Joonho Lee wrote:
> Document the five opt-in capture module parameters introduced by this
> series. All default to false. Describe the startup-empty, short-payload,
> packed-status and pause-restart compatibility behavior alongside staging,
> and require validation with the target device.
> 
> Assisted-by: LLM
> Signed-off-by: Joonho Lee <mynameisjune111@gmail.com>
> 
> ---

Please include a diffstat for each patch (here, below the "---" line).

> diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
> --- a/Documentation/admin-guide/kernel-parameters.txt
> +++ b/Documentation/admin-guide/kernel-parameters.txt
> @@ -8328,6 +8328,36 @@
>  			brightness level.
>  			default: 1
>  
> +	virtio_snd.capture_allow_empty= [HW,SOUND]
> +			Retry a bounded number of empty startup capture replies.
> +			Format: <bool>
> +			Default: false. Compatibility option for devices that
> +			require staged capture; validate with the target device.
> +
> +	virtio_snd.capture_allow_short= [HW,SOUND]
> +			Accept non-empty, frame-aligned short capture replies.
> +			Format: <bool>
> +			Default: false. Compatibility option for devices that
> +			require staged capture; validate with the target device.
> +
> +	virtio_snd.capture_packed_status= [HW,SOUND]
> +			Read status immediately after the returned capture data.
> +			Format: <bool>
> +			Default: false. Compatibility option for devices that
> +			require staged capture; validate with the target device.
> +
> +	virtio_snd.capture_pause_restart= [HW,SOUND]
> +			Release and reprepare capture streams across pause.
> +			Format: <bool>
> +			Default: false. Compatibility option for devices that
> +			require staged capture; validate with the target device.
> +
> +	virtio_snd.capture_staging= [HW,SOUND]
> +			Use request-owned buffers for capture data.
> +			Format: <bool>
> +			Default: false. Compatibility option for devices that
> +			require staged capture; validate with the target device.
> +

All virtio_snd entries should be *after* virtio_mmio (alphabetical order).

>  	virtio_mmio.device=
>  			[VMMIO] Memory mapped virtio (platform) device.
>  

-- 
~Randy