[PATCH v6 00/12] nvdimm: virtio_pmem: fix request lifetime and converge broken queue failures

[PATCH v6 00/12] nvdimm: virtio_pmem: fix request lifetime and converge broken queue failures

[Li Chen]

Hi,

The nvdimm flush helper currently converts any non-zero provider flush
callback error to -EIO. That hides useful errno values from providers. For
example, virtio-pmem may fail flush allocation with -ENOMEM, but that is
currently reported as -EIO by nvdimm_flush().

The raw failure seen in the local mkfs sanity test was:

  wipefs: /dev/pmem0: cannot flush modified buffers: Input/output error
  mkfs.ext4: Input/output error while writing out and closing file system
  nd_region region0: dbg: nvdimm_flush rc=-5

The first five patches keep provider flush errors intact, make
pmem_submit_bio() honor a failed REQ_PREFLUSH before copying data, keep
dataless bios out of the data loop, and avoid allocating a child flush bio
for virtio-pmem REQ_FUA handling. REQ_PREFLUSH and REQ_FUA are now issued
synchronously from pmem_submit_bio(). After that, virtio-pmem only allocates
its request object for the actual provider flush, and that allocation uses
GFP_NOIO so reclaim does not recurse into filesystem or block IO.

The rest of the series addresses virtio-pmem request lifetime and broken
virtqueue handling. The virtio-pmem flush path uses a virtqueue cookie/token
to carry a per-request context through completion. Under broken virtqueue /
notify failure conditions, the submitter can return and free the request
object while the host/backend may still complete the published request. The
IRQ completion handler then dereferences freed memory when waking waiters,
which is reported by KASAN as a slab-use-after-free and may manifest as lock
corruption (e.g. "BUG: spinlock already unlocked") without KASAN.

In addition, the flush path has two wait sites: one for virtqueue descriptor
availability (-ENOSPC from virtqueue_add_sgs()) and one for request
completion. If the virtqueue becomes broken, forward progress is no longer
guaranteed and these waiters may sleep indefinitely unless the driver
converges the failure and wakes all wait sites. This version also orders
response publication with release/acquire, keeps DMA_FROM_DEVICE response
storage away from CPU-owned request fields, and wakes the in-flight
completion waiter when a queue is marked broken.

This series addresses these issues:

1/12 nvdimm: preserve flush callback errors
Return provider flush callback errors directly from nvdimm_flush().

2/12 nvdimm: pmem: keep PREFLUSH before data writes
Run REQ_PREFLUSH synchronously before copying data and fail the bio if the
flush fails.

3/12 nvdimm: pmem: guard data loop for dataless bios
Keep flush-only bios out of the data copy loop.

4/12 nvdimm: virtio_pmem: stop allocating child flush bio
Flush REQ_FUA synchronously instead of allocating a chained child bio.

5/12 nvdimm: virtio_pmem: use GFP_NOIO for flush requests
Use GFP_NOIO for the virtio-pmem request allocation.

6/12 nvdimm: virtio_pmem: always wake -ENOSPC waiters
Wake one -ENOSPC waiter for each reclaimed used buffer, decoupled from
token completion.

7/12 nvdimm: virtio_pmem: use READ_ONCE()/WRITE_ONCE() for wait flags
Use READ_ONCE()/WRITE_ONCE() for the wait_event() flags (done and
wq_buf_avail).

8/12 nvdimm: virtio_pmem: refcount requests for token lifetime
Refcount request objects so the token lifetime spans the window where it is
reachable through the virtqueue until completion/drain drops the virtqueue
reference.

9/12 nvdimm: virtio_pmem: publish done with release/acquire
Order response publication before the submitter observes request completion.

10/12 nvdimm: virtio_pmem: isolate DMA request buffers
Keep the DMA_FROM_DEVICE response buffer away from CPU-owned request fields.

11/12 nvdimm: virtio_pmem: converge broken virtqueue to -EIO
Track a device-level broken state to converge broken/notify failures to -EIO:
wake -ENOSPC waiters, wake the in-flight completion waiter, fail-fast new
requests, and report errors after the queue is marked broken.

12/12 nvdimm: virtio_pmem: drain requests in freeze
Drain outstanding requests in freeze() after resetting the device so waiters
do not sleep indefinitely and virtqueue_detach_unused_buf() only runs on a
quiesced queue.

The original repros were on QEMU x86_64 with a virtio-pmem device exported
as /dev/pmem0. For this v6 reroll, the series applies to v7.1-rc7 and to
local next/master at 4fa3f5fabb30 ("Add linux-next specific files for
20260616").

Thanks,
Li Chen

Changelog:
v5->v6:
- Address Sashiko review feedback:
  - Add a data-loop guard for dataless bios in pmem_submit_bio().
  - Replace the child flush bio allocation with synchronous FUA flushing.
  - Keep GFP_NOIO only for the virtio-pmem request allocation.
  - Publish request completion with release/acquire ordering.
  - Isolate the DMA_FROM_DEVICE response buffer from CPU-owned fields.
  - Wake the in-flight host-completion waiter when marking the queue broken.
- Clear req_vq after del_vqs() and make drain tolerate a NULL queue.
v4->v5:
- Address review feedback about REQ_PREFLUSH ordering and active virtqueue
  detach.
- Add 2/8 so a failed REQ_PREFLUSH fails the bio before any data copy, and
  make REQ_PREFLUSH use a synchronous provider flush instead of a deferred
  child bio.
- Rework broken-queue handling so runtime failure marking only stops new
  submissions and wakes local -ENOSPC waiters; used/unused token draining is
  done after device reset in remove() and freeze().
- Remove the broken-state shortcut from the host-completion wait so the
  submitter never reads an uninitialized response field.
- Keep the raw broken-virtqueue dmesg in 7/8 while updating the teardown
  rationale.
- Renumber the old virtio-pmem fixes after the new pmem PREFLUSH patch.
v3->v4:
- Rebased the series onto v7.1-rc7 so it applies cleanly to Linux 7.1-rc7.
- Update the allocation site in 6/7 from kmalloc(sizeof(*req_data),
  GFP_KERNEL) to kmalloc_obj(*req_data) to match current nvdimm code.
- Add 1/7 to preserve provider flush callback errors in nvdimm_flush().
- Include the GFP_NOIO child flush bio allocation fix as 2/7.
- Renumber the old request lifetime and broken virtqueue fixes after the two
  new flush error patches.
v2->v3:
- Split patch 1 as suggested by Pankaj Gupta: keep the waiter wakeup
  ordering change in 1/5 and move READ_ONCE()/WRITE_ONCE() updates to
  2/5 (no functional change intended).
- Add log report to commit msg.
- Fold the export fix into 4/5 to keep the series bisectable when
  CONFIG_VIRTIO_PMEM=m.
v1->v2:
- Add the export patch to fix compile issue.

Links:
v5: https://lore.kernel.org/all/20260617122442.2118957-1-me@linux.beauty/
v4: https://lore.kernel.org/all/20260609120726.1714780-1-me@linux.beauty/
v3: https://lore.kernel.org/all/20260226025712.2236279-1-me@linux.beauty/#t
v2: https://lore.kernel.org/all/20251225042915.334117-1-me@linux.beauty/
v1: https://www.spinics.net/lists/kernel/msg5974818.html

Li Chen (12):
  nvdimm: preserve flush callback errors
  nvdimm: pmem: keep PREFLUSH before data writes
  nvdimm: pmem: guard data loop for dataless bios
  nvdimm: virtio_pmem: stop allocating child flush bio
  nvdimm: virtio_pmem: use GFP_NOIO for flush requests
  nvdimm: virtio_pmem: always wake -ENOSPC waiters
  nvdimm: virtio_pmem: use READ_ONCE()/WRITE_ONCE() for wait flags
  nvdimm: virtio_pmem: refcount requests for token lifetime
  nvdimm: virtio_pmem: publish done with release/acquire
  nvdimm: virtio_pmem: isolate DMA request buffers
  nvdimm: virtio_pmem: converge broken virtqueue to -EIO
  nvdimm: virtio_pmem: drain requests in freeze

 drivers/nvdimm/nd_virtio.c   | 224 +++++++++++++++++++++++++++--------
 drivers/nvdimm/pmem.c        |  52 ++++----
 drivers/nvdimm/region_devs.c |   6 +-
 drivers/nvdimm/virtio_pmem.c |  51 +++++++-
 drivers/nvdimm/virtio_pmem.h |  18 ++-
 5 files changed, 270 insertions(+), 81 deletions(-)

-- 
2.52.0

[PATCH v6 01/12] nvdimm: preserve flush callback errors

[Li Chen]

nvdimm_flush() currently converts any non-zero provider flush error to
-EIO. That loses useful errno values from provider callbacks.

A local virtio-pmem mkfs sanity test showed the masking clearly:

  wipefs: /dev/pmem0: cannot flush modified buffers: Input/output error
  mkfs.ext4: Input/output error while writing out and closing file system
  nd_region region0: dbg: nvdimm_flush rc=-5

The virtio-pmem callback can return -ENOMEM when async_pmem_flush() fails
to allocate a child flush bio, but nvdimm_flush() hides that as -EIO before
pmem_submit_bio() converts it to a block status.

Return the provider callback error directly. The generic flush path still
returns 0, and pmem_submit_bio() already handles errno-to-blk_status
conversion for bio completion.

Signed-off-by: Li Chen <me@linux.beauty>
---
v3->v4:
- New patch.

 drivers/nvdimm/region_devs.c | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/nvdimm/region_devs.c b/drivers/nvdimm/region_devs.c
index e35c2e18518f0..0cd96503c0596 100644
--- a/drivers/nvdimm/region_devs.c
+++ b/drivers/nvdimm/region_devs.c
@@ -1114,10 +1114,8 @@ int nvdimm_flush(struct nd_region *nd_region, struct bio *bio)
 
 	if (!nd_region->flush)
 		rc = generic_nvdimm_flush(nd_region);
-	else {
-		if (nd_region->flush(nd_region, bio))
-			rc = -EIO;
-	}
+	else
+		rc = nd_region->flush(nd_region, bio);
 
 	return rc;
 }
-- 
2.52.0

[PATCH v6 02/12] nvdimm: pmem: keep PREFLUSH before data writes

[Li Chen]

pmem_submit_bio() records a REQ_PREFLUSH error, but continues to copy the
bio data and can later overwrite the error with a successful REQ_FUA flush.
That lets data writes run after a failed preflush and can complete the bio
successfully despite the failed ordering barrier.

Run the REQ_PREFLUSH flush synchronously before touching the bio data and
complete the bio with the flush error if it fails. Keep asynchronous flush
chaining for REQ_FUA. At that point, data copy has completed and the parent
bio can wait for the chained flush bio.

Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v5:
- New patch.

 drivers/nvdimm/pmem.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index 92c67fbbc1c85..05d3de33e2706 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -208,8 +208,14 @@ static void pmem_submit_bio(struct bio *bio)
 	struct pmem_device *pmem = bio->bi_bdev->bd_disk->private_data;
 	struct nd_region *nd_region = to_region(pmem);
 
-	if (bio->bi_opf & REQ_PREFLUSH)
-		ret = nvdimm_flush(nd_region, bio);
+	if (bio->bi_opf & REQ_PREFLUSH) {
+		ret = nvdimm_flush(nd_region, NULL);
+		if (ret) {
+			bio->bi_status = errno_to_blk_status(ret);
+			bio_endio(bio);
+			return;
+		}
+	}
 
 	do_acct = blk_queue_io_stat(bio->bi_bdev->bd_disk->queue);
 	if (do_acct)
@@ -229,7 +235,7 @@ static void pmem_submit_bio(struct bio *bio)
 	if (do_acct)
 		bio_end_io_acct(bio, start);
 
-	if (bio->bi_opf & REQ_FUA)
+	if ((bio->bi_opf & REQ_FUA) && !bio->bi_status)
 		ret = nvdimm_flush(nd_region, bio);
 
 	if (ret)
-- 
2.52.0

[PATCH v6 03/12] nvdimm: pmem: guard data loop for dataless bios

[Li Chen]

pmem_submit_bio() handles flush-only bios before and after the data
loop. Keep dataless bios out of bio_for_each_segment() so the data path
only walks bios that actually carry bvec data.

Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v6:
- New patch.

 drivers/nvdimm/pmem.c | 36 +++++++++++++++++++++---------------
 1 file changed, 21 insertions(+), 15 deletions(-)

diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index 05d3de33e2706..82ee1ddb3a445 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -217,23 +217,29 @@ static void pmem_submit_bio(struct bio *bio)
 		}
 	}
 
-	do_acct = blk_queue_io_stat(bio->bi_bdev->bd_disk->queue);
-	if (do_acct)
-		start = bio_start_io_acct(bio);
-	bio_for_each_segment(bvec, bio, iter) {
-		if (op_is_write(bio_op(bio)))
-			rc = pmem_do_write(pmem, bvec.bv_page, bvec.bv_offset,
-				iter.bi_sector, bvec.bv_len);
-		else
-			rc = pmem_do_read(pmem, bvec.bv_page, bvec.bv_offset,
-				iter.bi_sector, bvec.bv_len);
-		if (rc) {
-			bio->bi_status = rc;
-			break;
+	if (bio_has_data(bio)) {
+		do_acct = blk_queue_io_stat(bio->bi_bdev->bd_disk->queue);
+		if (do_acct)
+			start = bio_start_io_acct(bio);
+		bio_for_each_segment(bvec, bio, iter) {
+			if (op_is_write(bio_op(bio)))
+				rc = pmem_do_write(pmem, bvec.bv_page,
+						   bvec.bv_offset,
+						   iter.bi_sector,
+						   bvec.bv_len);
+			else
+				rc = pmem_do_read(pmem, bvec.bv_page,
+						  bvec.bv_offset,
+						  iter.bi_sector,
+						  bvec.bv_len);
+			if (rc) {
+				bio->bi_status = rc;
+				break;
+			}
 		}
+		if (do_acct)
+			bio_end_io_acct(bio, start);
 	}
-	if (do_acct)
-		bio_end_io_acct(bio, start);
 
 	if ((bio->bi_opf & REQ_FUA) && !bio->bi_status)
 		ret = nvdimm_flush(nd_region, bio);
-- 
2.52.0

[PATCH v6 04/12] nvdimm: virtio_pmem: stop allocating child flush bio

[Li Chen]

pmem_submit_bio() passes the parent bio to nvdimm_flush() for
REQ_FUA. For virtio-pmem this makes async_pmem_flush() allocate
and submit a child PREFLUSH bio chained to the parent.

That child allocation is in the block submit path. Making it
blocking with GFP_NOIO can consume the same global bio mempool that
submit_bio() uses, while making it GFP_ATOMIC can fail under
pressure. A forced failure of the child allocation produced:

virtio_pmem: forcing child bio allocation failure for test
Buffer I/O error on dev pmem0, logical block 0, lost sync page write
EXT4-fs (pmem0): I/O error while writing superblock
EXT4-fs (pmem0): mount failed

Avoid the child bio completely. Flush FUA synchronously, like
REQ_PREFLUSH, then complete the parent after the flush. Since no
child bio can be created, async_pmem_flush() now only issues the
virtio flush and preserves negative errno values.

Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v6:
- Replace the child bio allocation fix with synchronous FUA flushing.

 drivers/nvdimm/nd_virtio.c | 22 ++++------------------
 drivers/nvdimm/pmem.c      |  2 +-
 2 files changed, 5 insertions(+), 19 deletions(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index 4176046627beb..4b2e9c47af0f5 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -110,27 +110,13 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 /* The asynchronous flush callback function */
 int async_pmem_flush(struct nd_region *nd_region, struct bio *bio)
 {
-	/*
-	 * Create child bio for asynchronous flush and chain with
-	 * parent bio. Otherwise directly call nd_region flush.
-	 */
-	if (bio && bio->bi_iter.bi_sector != -1) {
-		struct bio *child = bio_alloc(bio->bi_bdev, 0,
-					      REQ_OP_WRITE | REQ_PREFLUSH,
-					      GFP_ATOMIC);
+	int err;
 
-		if (!child)
-			return -ENOMEM;
-		bio_clone_blkg_association(child, bio);
-		child->bi_iter.bi_sector = -1;
-		bio_chain(child, bio);
-		submit_bio(child);
-		return 0;
-	}
-	if (virtio_pmem_flush(nd_region))
+	err = virtio_pmem_flush(nd_region);
+	if (err > 0)
 		return -EIO;
 
-	return 0;
+	return err;
 };
 EXPORT_SYMBOL_GPL(async_pmem_flush);
 MODULE_DESCRIPTION("Virtio Persistent Memory Driver");
diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
index 82ee1ddb3a445..058d2739c95a1 100644
--- a/drivers/nvdimm/pmem.c
+++ b/drivers/nvdimm/pmem.c
@@ -242,7 +242,7 @@ static void pmem_submit_bio(struct bio *bio)
 	}
 
 	if ((bio->bi_opf & REQ_FUA) && !bio->bi_status)
-		ret = nvdimm_flush(nd_region, bio);
+		ret = nvdimm_flush(nd_region, NULL);
 
 	if (ret)
 		bio->bi_status = errno_to_blk_status(ret);
-- 
2.52.0

[PATCH v6 05/12] nvdimm: virtio_pmem: use GFP_NOIO for flush requests

[Li Chen]

virtio_pmem_flush() can run from pmem_submit_bio() while filesystem IO
is waiting on the flush completion. The request object allocation can
sleep, but it should not enter filesystem or block IO reclaim from this
flush path.

Use GFP_NOIO for the request allocation. The virtqueue descriptor
allocation still uses GFP_ATOMIC because it runs under pmem_lock.

Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v6:
- New patch; keep GFP_NOIO only for the virtio-pmem request allocation.

 drivers/nvdimm/nd_virtio.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index 4b2e9c47af0f5..91ca144607531 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -55,7 +55,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 		return -EIO;
 	}
 
-	req_data = kmalloc_obj(*req_data);
+	req_data = kmalloc_obj(*req_data, GFP_NOIO);
 	if (!req_data)
 		return -ENOMEM;
 
-- 
2.52.0

[PATCH v6 06/12] nvdimm: virtio_pmem: always wake -ENOSPC waiters

[Li Chen]

virtio_pmem_host_ack() reclaims virtqueue descriptors with
virtqueue_get_buf(). The -ENOSPC waiter wakeup is tied to completing the
returned token. If token completion is skipped for any reason, reclaimed
descriptors may not wake a waiter and the submitter may sleep forever
waiting for a free slot. Always wake one -ENOSPC waiter for each virtqueue
completion before touching the returned token.

Signed-off-by: Li Chen <me@linux.beauty>
---
v2->v3:
- Split out the waiter wakeup ordering change from READ_ONCE()/WRITE_ONCE()
  updates (now patch 4/7), per Pankaj's suggestion.
v3->v4:
- Rebased onto v7.1-rc7 and renumbered after the flush error patches.

 drivers/nvdimm/nd_virtio.c | 25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index 91ca144607531..8ed4d6b3a9284 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -9,26 +9,33 @@
 #include "virtio_pmem.h"
 #include "nd.h"
 
+static void virtio_pmem_wake_one_waiter(struct virtio_pmem *vpmem)
+{
+	struct virtio_pmem_request *req_buf;
+
+	if (list_empty(&vpmem->req_list))
+		return;
+
+	req_buf = list_first_entry(&vpmem->req_list,
+				   struct virtio_pmem_request, list);
+	req_buf->wq_buf_avail = true;
+	wake_up(&req_buf->wq_buf);
+	list_del(&req_buf->list);
+}
+
  /* The interrupt handler */
 void virtio_pmem_host_ack(struct virtqueue *vq)
 {
 	struct virtio_pmem *vpmem = vq->vdev->priv;
-	struct virtio_pmem_request *req_data, *req_buf;
+	struct virtio_pmem_request *req_data;
 	unsigned long flags;
 	unsigned int len;
 
 	spin_lock_irqsave(&vpmem->pmem_lock, flags);
 	while ((req_data = virtqueue_get_buf(vq, &len)) != NULL) {
+		virtio_pmem_wake_one_waiter(vpmem);
 		req_data->done = true;
 		wake_up(&req_data->host_acked);
-
-		if (!list_empty(&vpmem->req_list)) {
-			req_buf = list_first_entry(&vpmem->req_list,
-					struct virtio_pmem_request, list);
-			req_buf->wq_buf_avail = true;
-			wake_up(&req_buf->wq_buf);
-			list_del(&req_buf->list);
-		}
 	}
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 }
-- 
2.52.0

[PATCH v6 07/12] nvdimm: virtio_pmem: use READ_ONCE()/WRITE_ONCE() for wait flags

[Li Chen]

Use READ_ONCE()/WRITE_ONCE() for the wait_event() flags (done and
wq_buf_avail). They are observed by waiters without pmem_lock, so make
the accesses explicit single loads/stores and avoid compiler
reordering/caching across the wait/wake paths.

Signed-off-by: Li Chen <me@linux.beauty>
---
v2->v3:
- Split out READ_ONCE()/WRITE_ONCE() updates from patch 3/7 (no functional
  change intended).
v3->v4:
- Rebased onto v7.1-rc7 and renumbered after the flush error patches.

 drivers/nvdimm/nd_virtio.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index 8ed4d6b3a9284..da829e9f4bdff 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -18,9 +18,9 @@ static void virtio_pmem_wake_one_waiter(struct virtio_pmem *vpmem)
 
 	req_buf = list_first_entry(&vpmem->req_list,
 				   struct virtio_pmem_request, list);
-	req_buf->wq_buf_avail = true;
+	list_del_init(&req_buf->list);
+	WRITE_ONCE(req_buf->wq_buf_avail, true);
 	wake_up(&req_buf->wq_buf);
-	list_del(&req_buf->list);
 }
 
  /* The interrupt handler */
@@ -34,7 +34,7 @@ void virtio_pmem_host_ack(struct virtqueue *vq)
 	spin_lock_irqsave(&vpmem->pmem_lock, flags);
 	while ((req_data = virtqueue_get_buf(vq, &len)) != NULL) {
 		virtio_pmem_wake_one_waiter(vpmem);
-		req_data->done = true;
+		WRITE_ONCE(req_data->done, true);
 		wake_up(&req_data->host_acked);
 	}
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
@@ -66,7 +66,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 	if (!req_data)
 		return -ENOMEM;
 
-	req_data->done = false;
+	WRITE_ONCE(req_data->done, false);
 	init_waitqueue_head(&req_data->host_acked);
 	init_waitqueue_head(&req_data->wq_buf);
 	INIT_LIST_HEAD(&req_data->list);
@@ -87,12 +87,12 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 					GFP_ATOMIC)) == -ENOSPC) {
 
 		dev_info(&vdev->dev, "failed to send command to virtio pmem device, no free slots in the virtqueue\n");
-		req_data->wq_buf_avail = false;
+		WRITE_ONCE(req_data->wq_buf_avail, false);
 		list_add_tail(&req_data->list, &vpmem->req_list);
 		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 
 		/* A host response results in "host_ack" getting called */
-		wait_event(req_data->wq_buf, req_data->wq_buf_avail);
+		wait_event(req_data->wq_buf, READ_ONCE(req_data->wq_buf_avail));
 		spin_lock_irqsave(&vpmem->pmem_lock, flags);
 	}
 	err1 = virtqueue_kick(vpmem->req_vq);
@@ -106,7 +106,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 		err = -EIO;
 	} else {
 		/* A host response results in "host_ack" getting called */
-		wait_event(req_data->host_acked, req_data->done);
+		wait_event(req_data->host_acked, READ_ONCE(req_data->done));
 		err = le32_to_cpu(req_data->resp.ret);
 	}
 
-- 
2.52.0

[PATCH v6 08/12] nvdimm: virtio_pmem: refcount requests for token lifetime

[Li Chen]

KASAN reports slab-use-after-free in __wake_up_common():
BUG: KASAN: slab-use-after-free in __wake_up_common+0x114/0x160
Read of size 8 at addr ffff88810fdcb710 by task swapper/0/0

CPU: 0 UID: 0 PID: 0 Comm: swapper/0 Not tainted
6.19.0-next-20260220-00006-g1eae5f204ec3 #4 PREEMPT(full)
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Arch Linux
1.17.0-2-2 04/01/2014
Call Trace:
 <IRQ>
 dump_stack_lvl+0x6d/0xb0
 print_report+0x170/0x4e2
 ? __pfx__raw_spin_lock_irqsave+0x10/0x10
 ? __virt_addr_valid+0x1dc/0x380
 kasan_report+0xbc/0xf0
 ? __wake_up_common+0x114/0x160
 ? __wake_up_common+0x114/0x160
 __wake_up_common+0x114/0x160
 ? __pfx__raw_spin_lock_irqsave+0x10/0x10
 __wake_up+0x36/0x60
 virtio_pmem_host_ack+0x11d/0x3b0
 ? sched_balance_domains+0x29f/0xb00
 ? __pfx_virtio_pmem_host_ack+0x10/0x10
 ? _raw_spin_lock_irqsave+0x98/0x100
 ? __pfx__raw_spin_lock_irqsave+0x10/0x10
 vring_interrupt+0x1c9/0x5e0
 ? __pfx_vp_interrupt+0x10/0x10
 vp_vring_interrupt+0x87/0x100
 ? __pfx_vp_interrupt+0x10/0x10
 __handle_irq_event_percpu+0x17f/0x550
 ? __pfx__raw_spin_lock+0x10/0x10
 handle_irq_event+0xab/0x1c0
 handle_fasteoi_irq+0x276/0xae0
 __common_interrupt+0x65/0x130
 common_interrupt+0x78/0xa0
 </IRQ>

virtio_pmem_host_ack() wakes a request that has already been freed by the
submitter.

This happens when the request token is still reachable via the virtqueue,
but virtio_pmem_flush() returns and frees it.

Fix the token lifetime by refcounting struct virtio_pmem_request.
virtio_pmem_flush() holds a submitter reference, and the virtqueue holds an
extra reference once the request is queued. The completion path drops the
virtqueue reference, and the submitter drops its reference before
returning.

Fixes: 6e84200c0a29 ("virtio-pmem: Add virtio pmem driver")
Cc: stable@vger.kernel.org
Signed-off-by: Li Chen <me@linux.beauty>
---
v2->v3:
- Add raw KASAN report to the patch description.
- Drop timestamps from the embedded report.
v3->v4:
- Rebased onto v7.1-rc7 and renumbered after the flush error patches.

 drivers/nvdimm/nd_virtio.c   | 34 +++++++++++++++++++++++++++++-----
 drivers/nvdimm/virtio_pmem.h |  2 ++
 2 files changed, 31 insertions(+), 5 deletions(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index da829e9f4bdff..7b6761adf28bc 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -9,6 +9,14 @@
 #include "virtio_pmem.h"
 #include "nd.h"
 
+static void virtio_pmem_req_release(struct kref *kref)
+{
+	struct virtio_pmem_request *req;
+
+	req = container_of(kref, struct virtio_pmem_request, kref);
+	kfree(req);
+}
+
 static void virtio_pmem_wake_one_waiter(struct virtio_pmem *vpmem)
 {
 	struct virtio_pmem_request *req_buf;
@@ -36,6 +44,7 @@ void virtio_pmem_host_ack(struct virtqueue *vq)
 		virtio_pmem_wake_one_waiter(vpmem);
 		WRITE_ONCE(req_data->done, true);
 		wake_up(&req_data->host_acked);
+		kref_put(&req_data->kref, virtio_pmem_req_release);
 	}
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 }
@@ -66,6 +75,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 	if (!req_data)
 		return -ENOMEM;
 
+	kref_init(&req_data->kref);
 	WRITE_ONCE(req_data->done, false);
 	init_waitqueue_head(&req_data->host_acked);
 	init_waitqueue_head(&req_data->wq_buf);
@@ -83,10 +93,23 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 	  * to req_list and wait for host_ack to wake us up when free
 	  * slots are available.
 	  */
-	while ((err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req_data,
-					GFP_ATOMIC)) == -ENOSPC) {
-
-		dev_info(&vdev->dev, "failed to send command to virtio pmem device, no free slots in the virtqueue\n");
+	for (;;) {
+		err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req_data,
+					GFP_ATOMIC);
+		if (!err) {
+			/*
+			 * Take the virtqueue reference while @pmem_lock is
+			 * held so completion cannot run concurrently.
+			 */
+			kref_get(&req_data->kref);
+			break;
+		}
+
+		if (err != -ENOSPC)
+			break;
+
+		dev_info_ratelimited(&vdev->dev,
+				     "failed to send command to virtio pmem device, no free slots in the virtqueue\n");
 		WRITE_ONCE(req_data->wq_buf_avail, false);
 		list_add_tail(&req_data->list, &vpmem->req_list);
 		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
@@ -95,6 +118,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 		wait_event(req_data->wq_buf, READ_ONCE(req_data->wq_buf_avail));
 		spin_lock_irqsave(&vpmem->pmem_lock, flags);
 	}
+
 	err1 = virtqueue_kick(vpmem->req_vq);
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 	/*
@@ -110,7 +134,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 		err = le32_to_cpu(req_data->resp.ret);
 	}
 
-	kfree(req_data);
+	kref_put(&req_data->kref, virtio_pmem_req_release);
 	return err;
 };
 
diff --git a/drivers/nvdimm/virtio_pmem.h b/drivers/nvdimm/virtio_pmem.h
index f72cf17f9518f..1017e498c9b4c 100644
--- a/drivers/nvdimm/virtio_pmem.h
+++ b/drivers/nvdimm/virtio_pmem.h
@@ -12,11 +12,13 @@
 
 #include <linux/module.h>
 #include <uapi/linux/virtio_pmem.h>
+#include <linux/kref.h>
 #include <linux/libnvdimm.h>
 #include <linux/mutex.h>
 #include <linux/spinlock.h>
 
 struct virtio_pmem_request {
+	struct kref kref;
 	struct virtio_pmem_req req;
 	struct virtio_pmem_resp resp;
 
-- 
2.52.0

[PATCH v6 09/12] nvdimm: virtio_pmem: publish done with release/acquire

[Li Chen]

virtio_pmem_host_ack() publishes the device response by setting done and
waking the submitter. The submitter reads resp.ret after wait_event()
observes done.

Use smp_store_release() on done and smp_load_acquire() in the wait
condition so the response read is ordered after completion.

Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v6:
- New patch.

 drivers/nvdimm/nd_virtio.c | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index 7b6761adf28bc..35d36bd36a526 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -17,6 +17,19 @@ static void virtio_pmem_req_release(struct kref *kref)
 	kfree(req);
 }
 
+static void virtio_pmem_signal_done(struct virtio_pmem_request *req)
+{
+	/* Pairs with smp_load_acquire() in virtio_pmem_req_done(). */
+	smp_store_release(&req->done, true);
+	wake_up(&req->host_acked);
+}
+
+static bool virtio_pmem_req_done(struct virtio_pmem_request *req)
+{
+	/* Pairs with smp_store_release() in virtio_pmem_signal_done(). */
+	return smp_load_acquire(&req->done);
+}
+
 static void virtio_pmem_wake_one_waiter(struct virtio_pmem *vpmem)
 {
 	struct virtio_pmem_request *req_buf;
@@ -42,8 +55,7 @@ void virtio_pmem_host_ack(struct virtqueue *vq)
 	spin_lock_irqsave(&vpmem->pmem_lock, flags);
 	while ((req_data = virtqueue_get_buf(vq, &len)) != NULL) {
 		virtio_pmem_wake_one_waiter(vpmem);
-		WRITE_ONCE(req_data->done, true);
-		wake_up(&req_data->host_acked);
+		virtio_pmem_signal_done(req_data);
 		kref_put(&req_data->kref, virtio_pmem_req_release);
 	}
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
@@ -130,7 +142,8 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 		err = -EIO;
 	} else {
 		/* A host response results in "host_ack" getting called */
-		wait_event(req_data->host_acked, READ_ONCE(req_data->done));
+		wait_event(req_data->host_acked,
+			   virtio_pmem_req_done(req_data));
 		err = le32_to_cpu(req_data->resp.ret);
 	}
 
-- 
2.52.0

[PATCH v6 10/12] nvdimm: virtio_pmem: isolate DMA request buffers

[Li Chen]

The virtio-pmem request object stores wait queues, flags, and list
pointers next to buffers mapped for virtqueue DMA. The response buffer is
mapped DMA_FROM_DEVICE, so non-coherent DMA invalidation must not share a
cache line with CPU-owned fields.

Keep the request buffer outside the DMA-from-device group and wrap only
the response buffer with __dma_from_device_group_begin/end.

Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v6:
- New patch.

 drivers/nvdimm/virtio_pmem.h | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/drivers/nvdimm/virtio_pmem.h b/drivers/nvdimm/virtio_pmem.h
index 1017e498c9b4c..23bff40249c1b 100644
--- a/drivers/nvdimm/virtio_pmem.h
+++ b/drivers/nvdimm/virtio_pmem.h
@@ -10,6 +10,7 @@
 #ifndef _LINUX_VIRTIO_PMEM_H
 #define _LINUX_VIRTIO_PMEM_H
 
+#include <linux/dma-mapping.h>
 #include <linux/module.h>
 #include <uapi/linux/virtio_pmem.h>
 #include <linux/kref.h>
@@ -19,8 +20,6 @@
 
 struct virtio_pmem_request {
 	struct kref kref;
-	struct virtio_pmem_req req;
-	struct virtio_pmem_resp resp;
 
 	/* Wait queue to process deferred work after ack from host */
 	wait_queue_head_t host_acked;
@@ -30,6 +29,11 @@ struct virtio_pmem_request {
 	wait_queue_head_t wq_buf;
 	bool wq_buf_avail;
 	struct list_head list;
+
+	struct virtio_pmem_req req;
+	__dma_from_device_group_begin(resp);
+	struct virtio_pmem_resp resp;
+	__dma_from_device_group_end(resp);
 };
 
 struct virtio_pmem {
-- 
2.52.0

[PATCH v6 11/12] nvdimm: virtio_pmem: converge broken virtqueue to -EIO

[Li Chen]

dmesg reports virtqueue failure and device reset:
virtio_pmem virtio2: failed to send command to
virtio pmem device, no free slots in the virtqueue
virtio_pmem virtio2: virtio pmem device
needs a reset

virtio_pmem_flush() can wait for a free virtqueue descriptor (-ENOSPC).
It can also wait for host completion. If the request virtqueue breaks,
those waiters may never make progress. One example is notify failure from
virtqueue_kick().

Track a device-level broken state and converge the failure to -EIO. New
requests fail fast, -ENOSPC waiters are unlinked and woken, and the
currently submitted request is woken so its host_acked waiter can return
without waiting forever for host completion. Completed requests are forced
to report an error after the queue is marked broken.

Do not detach unused buffers from an active virtqueue. Runtime
broken-queue handling only stops new submissions and wakes local waiters.
Removal resets the device first. It then drains request tokens. After
that, the device no longer owns the buffers when the virtqueue reference
is dropped.

Closes: https://lore.kernel.org/r/202512250116.ewtzlD0g-lkp@intel.com/
Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v6:
- Wake the in-flight host-completion waiter when marking the queue broken.
- Track req_inflight and clear it on completion/drain paths.
- Return -EIO if the queue breaks before a host response is observed.
Changes in v5:
- Split broken marking from token draining.
- Do not call virtqueue_detach_unused_buf() on an active queue.
- Reset the device before draining tokens in remove().
- Do not let the host-completion wait return only because the device is
  marked broken.
v2->v3:
- Add raw dmesg excerpt to the patch description.
- Drop timestamps from the embedded dmesg.
- Fold the CONFIG_VIRTIO_PMEM=m export fix into this patch.
v3->v4:
- Rebased onto v7.1-rc7 and renumbered after the flush error patches.
- Use kmalloc_obj(*req_data) at the allocation site to match current nvdimm
  code.

 drivers/nvdimm/nd_virtio.c   | 117 +++++++++++++++++++++++++++++++----
 drivers/nvdimm/virtio_pmem.c |  15 ++++-
 drivers/nvdimm/virtio_pmem.h |   8 +++
 3 files changed, 127 insertions(+), 13 deletions(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index 35d36bd36a526..fb9391ebc46e7 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -30,6 +30,12 @@ static bool virtio_pmem_req_done(struct virtio_pmem_request *req)
 	return smp_load_acquire(&req->done);
 }
 
+static void virtio_pmem_complete_err(struct virtio_pmem_request *req)
+{
+	req->resp.ret = cpu_to_le32(1);
+	virtio_pmem_signal_done(req);
+}
+
 static void virtio_pmem_wake_one_waiter(struct virtio_pmem *vpmem)
 {
 	struct virtio_pmem_request *req_buf;
@@ -44,6 +50,63 @@ static void virtio_pmem_wake_one_waiter(struct virtio_pmem *vpmem)
 	wake_up(&req_buf->wq_buf);
 }
 
+static void virtio_pmem_wake_all_waiters(struct virtio_pmem *vpmem)
+{
+	struct virtio_pmem_request *req, *tmp;
+
+	list_for_each_entry_safe(req, tmp, &vpmem->req_list, list) {
+		list_del_init(&req->list);
+		WRITE_ONCE(req->wq_buf_avail, true);
+		wake_up(&req->wq_buf);
+	}
+}
+
+static void virtio_pmem_clear_inflight(struct virtio_pmem *vpmem,
+				       struct virtio_pmem_request *req)
+{
+	if (vpmem->req_inflight == req)
+		vpmem->req_inflight = NULL;
+}
+
+static void virtio_pmem_wake_inflight(struct virtio_pmem *vpmem)
+{
+	struct virtio_pmem_request *req = vpmem->req_inflight;
+
+	if (req)
+		wake_up(&req->host_acked);
+}
+
+void virtio_pmem_mark_broken(struct virtio_pmem *vpmem)
+{
+	if (!READ_ONCE(vpmem->broken)) {
+		WRITE_ONCE(vpmem->broken, true);
+		dev_err_once(&vpmem->vdev->dev, "virtqueue is broken\n");
+	}
+
+	virtio_pmem_wake_inflight(vpmem);
+	virtio_pmem_wake_all_waiters(vpmem);
+}
+EXPORT_SYMBOL_GPL(virtio_pmem_mark_broken);
+
+void virtio_pmem_drain(struct virtio_pmem *vpmem)
+{
+	struct virtio_pmem_request *req;
+	unsigned int len;
+
+	while ((req = virtqueue_get_buf(vpmem->req_vq, &len)) != NULL) {
+		virtio_pmem_clear_inflight(vpmem, req);
+		virtio_pmem_complete_err(req);
+		kref_put(&req->kref, virtio_pmem_req_release);
+	}
+
+	while ((req = virtqueue_detach_unused_buf(vpmem->req_vq)) != NULL) {
+		virtio_pmem_clear_inflight(vpmem, req);
+		virtio_pmem_complete_err(req);
+		kref_put(&req->kref, virtio_pmem_req_release);
+	}
+}
+EXPORT_SYMBOL_GPL(virtio_pmem_drain);
+
  /* The interrupt handler */
 void virtio_pmem_host_ack(struct virtqueue *vq)
 {
@@ -54,8 +117,12 @@ void virtio_pmem_host_ack(struct virtqueue *vq)
 
 	spin_lock_irqsave(&vpmem->pmem_lock, flags);
 	while ((req_data = virtqueue_get_buf(vq, &len)) != NULL) {
+		virtio_pmem_clear_inflight(vpmem, req_data);
 		virtio_pmem_wake_one_waiter(vpmem);
-		virtio_pmem_signal_done(req_data);
+		if (READ_ONCE(vpmem->broken))
+			virtio_pmem_complete_err(req_data);
+		else
+			virtio_pmem_signal_done(req_data);
 		kref_put(&req_data->kref, virtio_pmem_req_release);
 	}
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
@@ -83,6 +150,9 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 		return -EIO;
 	}
 
+	if (READ_ONCE(vpmem->broken))
+		return -EIO;
+
 	req_data = kmalloc_obj(*req_data, GFP_NOIO);
 	if (!req_data)
 		return -ENOMEM;
@@ -99,13 +169,18 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 	sgs[1] = &ret;
 
 	spin_lock_irqsave(&vpmem->pmem_lock, flags);
-	 /*
-	  * If virtqueue_add_sgs returns -ENOSPC then req_vq virtual
-	  * queue does not have free descriptor. We add the request
-	  * to req_list and wait for host_ack to wake us up when free
-	  * slots are available.
-	  */
+	/*
+	 * If virtqueue_add_sgs returns -ENOSPC then req_vq virtual
+	 * queue does not have free descriptor. We add the request
+	 * to req_list and wait for host_ack to wake us up when free
+	 * slots are available.
+	 */
 	for (;;) {
+		if (READ_ONCE(vpmem->broken)) {
+			err = -EIO;
+			break;
+		}
+
 		err = virtqueue_add_sgs(vpmem->req_vq, sgs, 1, 1, req_data,
 					GFP_ATOMIC);
 		if (!err) {
@@ -114,6 +189,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 			 * held so completion cannot run concurrently.
 			 */
 			kref_get(&req_data->kref);
+			vpmem->req_inflight = req_data;
 			break;
 		}
 
@@ -127,24 +203,41 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
 		spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 
 		/* A host response results in "host_ack" getting called */
-		wait_event(req_data->wq_buf, READ_ONCE(req_data->wq_buf_avail));
+		wait_event(req_data->wq_buf,
+			   READ_ONCE(req_data->wq_buf_avail) ||
+			   READ_ONCE(vpmem->broken));
 		spin_lock_irqsave(&vpmem->pmem_lock, flags);
+
+		if (READ_ONCE(vpmem->broken))
+			break;
 	}
 
-	err1 = virtqueue_kick(vpmem->req_vq);
+	if (err == -EIO || virtqueue_is_broken(vpmem->req_vq))
+		virtio_pmem_mark_broken(vpmem);
+
+	err1 = true;
+	if (!err && !READ_ONCE(vpmem->broken)) {
+		err1 = virtqueue_kick(vpmem->req_vq);
+		if (!err1)
+			virtio_pmem_mark_broken(vpmem);
+	}
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 	/*
 	 * virtqueue_add_sgs failed with error different than -ENOSPC, we can't
 	 * do anything about that.
 	 */
-	if (err || !err1) {
+	if (READ_ONCE(vpmem->broken) || err || !err1) {
 		dev_info(&vdev->dev, "failed to send command to virtio pmem device\n");
 		err = -EIO;
 	} else {
 		/* A host response results in "host_ack" getting called */
 		wait_event(req_data->host_acked,
-			   virtio_pmem_req_done(req_data));
-		err = le32_to_cpu(req_data->resp.ret);
+			   virtio_pmem_req_done(req_data) ||
+			   READ_ONCE(vpmem->broken));
+		if (virtio_pmem_req_done(req_data))
+			err = le32_to_cpu(req_data->resp.ret);
+		else
+			err = -EIO;
 	}
 
 	kref_put(&req_data->kref, virtio_pmem_req_release);
diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
index 77b1966619059..b272e9279ef23 100644
--- a/drivers/nvdimm/virtio_pmem.c
+++ b/drivers/nvdimm/virtio_pmem.c
@@ -25,6 +25,8 @@ static int init_vq(struct virtio_pmem *vpmem)
 
 	spin_lock_init(&vpmem->pmem_lock);
 	INIT_LIST_HEAD(&vpmem->req_list);
+	vpmem->req_inflight = NULL;
+	WRITE_ONCE(vpmem->broken, false);
 
 	return 0;
 };
@@ -138,10 +140,21 @@ static int virtio_pmem_probe(struct virtio_device *vdev)
 static void virtio_pmem_remove(struct virtio_device *vdev)
 {
 	struct nvdimm_bus *nvdimm_bus = dev_get_drvdata(&vdev->dev);
+	struct virtio_pmem *vpmem = vdev->priv;
+	unsigned long flags;
+
+	spin_lock_irqsave(&vpmem->pmem_lock, flags);
+	virtio_pmem_mark_broken(vpmem);
+	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
+
+	virtio_reset_device(vdev);
+
+	spin_lock_irqsave(&vpmem->pmem_lock, flags);
+	virtio_pmem_drain(vpmem);
+	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 
 	nvdimm_bus_unregister(nvdimm_bus);
 	vdev->config->del_vqs(vdev);
-	virtio_reset_device(vdev);
 }
 
 static int virtio_pmem_freeze(struct virtio_device *vdev)
diff --git a/drivers/nvdimm/virtio_pmem.h b/drivers/nvdimm/virtio_pmem.h
index 23bff40249c1b..bc7de2b328985 100644
--- a/drivers/nvdimm/virtio_pmem.h
+++ b/drivers/nvdimm/virtio_pmem.h
@@ -52,6 +52,12 @@ struct virtio_pmem {
 	/* List to store deferred work if virtqueue is full */
 	struct list_head req_list;
 
+	/* Request currently owned by the virtqueue. */
+	struct virtio_pmem_request *req_inflight;
+
+	/* Fail fast and wake waiters if the request virtqueue is broken. */
+	bool broken;
+
 	/* Synchronize virtqueue data */
 	spinlock_t pmem_lock;
 
@@ -61,5 +67,7 @@ struct virtio_pmem {
 };
 
 void virtio_pmem_host_ack(struct virtqueue *vq);
+void virtio_pmem_mark_broken(struct virtio_pmem *vpmem);
+void virtio_pmem_drain(struct virtio_pmem *vpmem);
 int async_pmem_flush(struct nd_region *nd_region, struct bio *bio);
 #endif
-- 
2.52.0

[PATCH v6 12/12] nvdimm: virtio_pmem: drain requests in freeze

[Li Chen]

virtio_pmem_freeze() currently deletes virtqueues and resets the device
without waking threads waiting for a virtqueue descriptor or a host
completion.

Mark the request virtqueue broken before reset. This makes new submissions
fail fast and lets -ENOSPC waiters leave the wait list. Reset the device
before draining used and unused request tokens, then delete the virtqueues.
This wakes waiters with -EIO. It also keeps the detach call on a quiesced
device.

Clear req_vq after del_vqs(), and make drain tolerate a NULL queue, so
remove after freeze does not dereference a stale virtqueue pointer.

Signed-off-by: Li Chen <me@linux.beauty>
---
Changes in v6:
- Clear req_vq after del_vqs() and make drain tolerate a NULL queue.
Changes in v5:
- Reset the device before draining used and unused request tokens.
- Use the split broken-marking and post-reset drain helpers.
v2->v3:
- No change.
v3->v4:
- Rebased onto v7.1-rc7 and renumbered after the flush error patches.

 drivers/nvdimm/nd_virtio.c   |  3 +++
 drivers/nvdimm/virtio_pmem.c | 36 +++++++++++++++++++++++++++++++-----
 2 files changed, 34 insertions(+), 5 deletions(-)

diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
index fb9391ebc46e7..ce4032dc07628 100644
--- a/drivers/nvdimm/nd_virtio.c
+++ b/drivers/nvdimm/nd_virtio.c
@@ -93,6 +93,9 @@ void virtio_pmem_drain(struct virtio_pmem *vpmem)
 	struct virtio_pmem_request *req;
 	unsigned int len;
 
+	if (!vpmem->req_vq)
+		return;
+
 	while ((req = virtqueue_get_buf(vpmem->req_vq, &len)) != NULL) {
 		virtio_pmem_clear_inflight(vpmem, req);
 		virtio_pmem_complete_err(req);
diff --git a/drivers/nvdimm/virtio_pmem.c b/drivers/nvdimm/virtio_pmem.c
index b272e9279ef23..fef792f725db2 100644
--- a/drivers/nvdimm/virtio_pmem.c
+++ b/drivers/nvdimm/virtio_pmem.c
@@ -17,11 +17,16 @@ static struct virtio_device_id id_table[] = {
  /* Initialize virt queue */
 static int init_vq(struct virtio_pmem *vpmem)
 {
+	int err;
+
 	/* single vq */
 	vpmem->req_vq = virtio_find_single_vq(vpmem->vdev,
 					virtio_pmem_host_ack, "flush_queue");
-	if (IS_ERR(vpmem->req_vq))
-		return PTR_ERR(vpmem->req_vq);
+	if (IS_ERR(vpmem->req_vq)) {
+		err = PTR_ERR(vpmem->req_vq);
+		vpmem->req_vq = NULL;
+		return err;
+	}
 
 	spin_lock_init(&vpmem->pmem_lock);
 	INIT_LIST_HEAD(&vpmem->req_list);
@@ -31,6 +36,15 @@ static int init_vq(struct virtio_pmem *vpmem)
 	return 0;
 };
 
+static void virtio_pmem_del_vqs(struct virtio_pmem *vpmem)
+{
+	if (!vpmem->req_vq)
+		return;
+
+	vpmem->vdev->config->del_vqs(vpmem->vdev);
+	vpmem->req_vq = NULL;
+}
+
 static int virtio_pmem_validate(struct virtio_device *vdev)
 {
 	struct virtio_shm_region shm_reg;
@@ -132,7 +146,7 @@ static int virtio_pmem_probe(struct virtio_device *vdev)
 	virtio_reset_device(vdev);
 	nvdimm_bus_unregister(vpmem->nvdimm_bus);
 out_vq:
-	vdev->config->del_vqs(vdev);
+	virtio_pmem_del_vqs(vpmem);
 out_err:
 	return err;
 }
@@ -154,14 +168,26 @@ static void virtio_pmem_remove(struct virtio_device *vdev)
 	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
 
 	nvdimm_bus_unregister(nvdimm_bus);
-	vdev->config->del_vqs(vdev);
+	virtio_pmem_del_vqs(vpmem);
 }
 
 static int virtio_pmem_freeze(struct virtio_device *vdev)
 {
-	vdev->config->del_vqs(vdev);
+	struct virtio_pmem *vpmem = vdev->priv;
+	unsigned long flags;
+
+	spin_lock_irqsave(&vpmem->pmem_lock, flags);
+	virtio_pmem_mark_broken(vpmem);
+	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
+
 	virtio_reset_device(vdev);
 
+	spin_lock_irqsave(&vpmem->pmem_lock, flags);
+	virtio_pmem_drain(vpmem);
+	spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
+
+	virtio_pmem_del_vqs(vpmem);
+
 	return 0;
 }
 
-- 
2.52.0

Re: [PATCH v6 07/12] nvdimm: virtio_pmem: use READ_ONCE()/WRITE_ONCE() for wait flags

[Pankaj Gupta]

> Use READ_ONCE()/WRITE_ONCE() for the wait_event() flags (done and
> wq_buf_avail). They are observed by waiters without pmem_lock, so make
> the accesses explicit single loads/stores and avoid compiler
> reordering/caching across the wait/wake paths.
>
> Signed-off-by: Li Chen <me@linux.beauty>

Acked-by: Pankaj Gupta <pankaj.gupta.linux@gmail.com

> ---
> v2->v3:
> - Split out READ_ONCE()/WRITE_ONCE() updates from patch 3/7 (no functional
>   change intended).
> v3->v4:
> - Rebased onto v7.1-rc7 and renumbered after the flush error patches.
>
>  drivers/nvdimm/nd_virtio.c | 14 +++++++-------
>  1 file changed, 7 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
> index 8ed4d6b3a9284..da829e9f4bdff 100644
> --- a/drivers/nvdimm/nd_virtio.c
> +++ b/drivers/nvdimm/nd_virtio.c
> @@ -18,9 +18,9 @@ static void virtio_pmem_wake_one_waiter(struct virtio_pmem *vpmem)
>
>         req_buf = list_first_entry(&vpmem->req_list,
>                                    struct virtio_pmem_request, list);
> -       req_buf->wq_buf_avail = true;
> +       list_del_init(&req_buf->list);
> +       WRITE_ONCE(req_buf->wq_buf_avail, true);
>         wake_up(&req_buf->wq_buf);
> -       list_del(&req_buf->list);
>  }
>
>   /* The interrupt handler */
> @@ -34,7 +34,7 @@ void virtio_pmem_host_ack(struct virtqueue *vq)
>         spin_lock_irqsave(&vpmem->pmem_lock, flags);
>         while ((req_data = virtqueue_get_buf(vq, &len)) != NULL) {
>                 virtio_pmem_wake_one_waiter(vpmem);
> -               req_data->done = true;
> +               WRITE_ONCE(req_data->done, true);
>                 wake_up(&req_data->host_acked);
>         }
>         spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
> @@ -66,7 +66,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
>         if (!req_data)
>                 return -ENOMEM;
>
> -       req_data->done = false;
> +       WRITE_ONCE(req_data->done, false);
>         init_waitqueue_head(&req_data->host_acked);
>         init_waitqueue_head(&req_data->wq_buf);
>         INIT_LIST_HEAD(&req_data->list);
> @@ -87,12 +87,12 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
>                                         GFP_ATOMIC)) == -ENOSPC) {
>
>                 dev_info(&vdev->dev, "failed to send command to virtio pmem device, no free slots in the virtqueue\n");
> -               req_data->wq_buf_avail = false;
> +               WRITE_ONCE(req_data->wq_buf_avail, false);
>                 list_add_tail(&req_data->list, &vpmem->req_list);
>                 spin_unlock_irqrestore(&vpmem->pmem_lock, flags);
>
>                 /* A host response results in "host_ack" getting called */
> -               wait_event(req_data->wq_buf, req_data->wq_buf_avail);
> +               wait_event(req_data->wq_buf, READ_ONCE(req_data->wq_buf_avail));
>                 spin_lock_irqsave(&vpmem->pmem_lock, flags);
>         }
>         err1 = virtqueue_kick(vpmem->req_vq);
> @@ -106,7 +106,7 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
>                 err = -EIO;
>         } else {
>                 /* A host response results in "host_ack" getting called */
> -               wait_event(req_data->host_acked, req_data->done);
> +               wait_event(req_data->host_acked, READ_ONCE(req_data->done));
>                 err = le32_to_cpu(req_data->resp.ret);
>         }
>
> --
> 2.52.0

Re: [PATCH v6 01/12] nvdimm: preserve flush callback errors

[Pankaj Gupta]

> nvdimm_flush() currently converts any non-zero provider flush error to
> -EIO. That loses useful errno values from provider callbacks.
>
> A local virtio-pmem mkfs sanity test showed the masking clearly:
>
>   wipefs: /dev/pmem0: cannot flush modified buffers: Input/output error
>   mkfs.ext4: Input/output error while writing out and closing file system
>   nd_region region0: dbg: nvdimm_flush rc=-5
>
> The virtio-pmem callback can return -ENOMEM when async_pmem_flush() fails
> to allocate a child flush bio, but nvdimm_flush() hides that as -EIO before
> pmem_submit_bio() converts it to a block status.
>
> Return the provider callback error directly. The generic flush path still
> returns 0, and pmem_submit_bio() already handles errno-to-blk_status
> conversion for bio completion.
>
> Signed-off-by: Li Chen <me@linux.beauty>
> ---
> v3->v4:
> - New patch.
>
>  drivers/nvdimm/region_devs.c | 6 ++----
>  1 file changed, 2 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/nvdimm/region_devs.c b/drivers/nvdimm/region_devs.c
> index e35c2e18518f0..0cd96503c0596 100644
> --- a/drivers/nvdimm/region_devs.c
> +++ b/drivers/nvdimm/region_devs.c
> @@ -1114,10 +1114,8 @@ int nvdimm_flush(struct nd_region *nd_region, struct bio *bio)
>
>         if (!nd_region->flush)
>                 rc = generic_nvdimm_flush(nd_region);
> -       else {
> -               if (nd_region->flush(nd_region, bio))
> -                       rc = -EIO;
> -       }
> +       else
> +               rc = nd_region->flush(nd_region, bio);

IIRC this was introduced as a generic populate error type since a
failed flush can also propagate host-side errors, which may not be
relevant to the guest.
That said, we could still consider handling specific cases like
-ENOMEM, unless there is a better approach to address this.

Thanks,
Pankaj
>
>         return rc;
>  }
> --
> 2.52.0

Re: [PATCH v6 04/12] nvdimm: virtio_pmem: stop allocating child flush bio

[Pankaj Gupta]

> pmem_submit_bio() passes the parent bio to nvdimm_flush() for
> REQ_FUA. For virtio-pmem this makes async_pmem_flush() allocate
> and submit a child PREFLUSH bio chained to the parent.
>
> That child allocation is in the block submit path. Making it
> blocking with GFP_NOIO can consume the same global bio mempool that
> submit_bio() uses, while making it GFP_ATOMIC can fail under
> pressure. A forced failure of the child allocation produced:
>
> virtio_pmem: forcing child bio allocation failure for test
> Buffer I/O error on dev pmem0, logical block 0, lost sync page write
> EXT4-fs (pmem0): I/O error while writing superblock
> EXT4-fs (pmem0): mount failed
>
> Avoid the child bio completely. Flush FUA synchronously, like
> REQ_PREFLUSH, then complete the parent after the flush. Since no
> child bio can be created, async_pmem_flush() now only issues the
> virtio flush and preserves negative errno values.

Child flush is asynchronous (performs async flush to host side and returns).
Till child bio completes guest userspace waits in pending IO state.
It seems the current change will affect the behavior?

Prior RFC [1] attempted to coalesce the async FLUSH request between guest &host.
If there is interest, that approach could be revisited or integrated here?

[1] https://lore.kernel.org/all/20220111161937.56272-1-pankaj.gupta.linux@gmail.com/#t

Thanks,
Pankaj

>
> Signed-off-by: Li Chen <me@linux.beauty>
> ---
> Changes in v6:
> - Replace the child bio allocation fix with synchronous FUA flushing.
>
>  drivers/nvdimm/nd_virtio.c | 22 ++++------------------
>  drivers/nvdimm/pmem.c      |  2 +-
>  2 files changed, 5 insertions(+), 19 deletions(-)
>
> diff --git a/drivers/nvdimm/nd_virtio.c b/drivers/nvdimm/nd_virtio.c
> index 4176046627beb..4b2e9c47af0f5 100644
> --- a/drivers/nvdimm/nd_virtio.c
> +++ b/drivers/nvdimm/nd_virtio.c
> @@ -110,27 +110,13 @@ static int virtio_pmem_flush(struct nd_region *nd_region)
>  /* The asynchronous flush callback function */
>  int async_pmem_flush(struct nd_region *nd_region, struct bio *bio)
>  {
> -       /*
> -        * Create child bio for asynchronous flush and chain with
> -        * parent bio. Otherwise directly call nd_region flush.
> -        */
> -       if (bio && bio->bi_iter.bi_sector != -1) {
> -               struct bio *child = bio_alloc(bio->bi_bdev, 0,
> -                                             REQ_OP_WRITE | REQ_PREFLUSH,
> -                                             GFP_ATOMIC);
> +       int err;
>
> -               if (!child)
> -                       return -ENOMEM;
> -               bio_clone_blkg_association(child, bio);
> -               child->bi_iter.bi_sector = -1;
> -               bio_chain(child, bio);
> -               submit_bio(child);
> -               return 0;
> -       }
> -       if (virtio_pmem_flush(nd_region))
> +       err = virtio_pmem_flush(nd_region);
> +       if (err > 0)
>                 return -EIO;
>
> -       return 0;
> +       return err;
>  };
>  EXPORT_SYMBOL_GPL(async_pmem_flush);
>  MODULE_DESCRIPTION("Virtio Persistent Memory Driver");
> diff --git a/drivers/nvdimm/pmem.c b/drivers/nvdimm/pmem.c
> index 82ee1ddb3a445..058d2739c95a1 100644
> --- a/drivers/nvdimm/pmem.c
> +++ b/drivers/nvdimm/pmem.c
> @@ -242,7 +242,7 @@ static void pmem_submit_bio(struct bio *bio)
>         }
>
>         if ((bio->bi_opf & REQ_FUA) && !bio->bi_status)
> -               ret = nvdimm_flush(nd_region, bio);
> +               ret = nvdimm_flush(nd_region, NULL);
>
>         if (ret)
>                 bio->bi_status = errno_to_blk_status(ret);
> --
> 2.52.0

Re: [PATCH v6 01/12] nvdimm: preserve flush callback errors

[Li Chen]

Hi Pankaj,

 ---- On Tue, 23 Jun 2026 17:46:20 +0800  Pankaj Gupta <pankaj.gupta.linux@gmail.com> wrote --- 
 > > nvdimm_flush() currently converts any non-zero provider flush error to
 > > -EIO. That loses useful errno values from provider callbacks.
 > >
 > > A local virtio-pmem mkfs sanity test showed the masking clearly:
 > >
 > >   wipefs: /dev/pmem0: cannot flush modified buffers: Input/output error
 > >   mkfs.ext4: Input/output error while writing out and closing file system
 > >   nd_region region0: dbg: nvdimm_flush rc=-5
 > >
 > > The virtio-pmem callback can return -ENOMEM when async_pmem_flush() fails
 > > to allocate a child flush bio, but nvdimm_flush() hides that as -EIO before
 > > pmem_submit_bio() converts it to a block status.
 > >
 > > Return the provider callback error directly. The generic flush path still
 > > returns 0, and pmem_submit_bio() already handles errno-to-blk_status
 > > conversion for bio completion.
 > >
 > > Signed-off-by: Li Chen <me@linux.beauty>
 > > ---
 > > v3->v4:
 > > - New patch.
 > >
 > >  drivers/nvdimm/region_devs.c | 6 ++----
 > >  1 file changed, 2 insertions(+), 4 deletions(-)
 > >
 > > diff --git a/drivers/nvdimm/region_devs.c b/drivers/nvdimm/region_devs.c
 > > index e35c2e18518f0..0cd96503c0596 100644
 > > --- a/drivers/nvdimm/region_devs.c
 > > +++ b/drivers/nvdimm/region_devs.c
 > > @@ -1114,10 +1114,8 @@ int nvdimm_flush(struct nd_region *nd_region, struct bio *bio)
 > >
 > >         if (!nd_region->flush)
 > >                 rc = generic_nvdimm_flush(nd_region);
 > > -       else {
 > > -               if (nd_region->flush(nd_region, bio))
 > > -                       rc = -EIO;
 > > -       }
 > > +       else
 > > +               rc = nd_region->flush(nd_region, bio);
 > 
 > IIRC this was introduced as a generic populate error type since a
 > failed flush can also propagate host-side errors, which may not be
 > relevant to the guest.
 > That said, we could still consider handling specific cases like
 > -ENOMEM, unless there is a better approach to address this.

Ah, yes, you are absolutely right. I went too far here.

Regards,
Li​

Re: [PATCH v6 04/12] nvdimm: virtio_pmem: stop allocating child flush bio

[Li Chen]

Hi Pankaj

 ---- On Thu, 25 Jun 2026 01:22:14 +0800  Pankaj Gupta <pankaj.gupta.linux@gmail.com> wrote --- 
 > > pmem_submit_bio() passes the parent bio to nvdimm_flush() for
 > > REQ_FUA. For virtio-pmem this makes async_pmem_flush() allocate
 > > and submit a child PREFLUSH bio chained to the parent.
 > >
 > > That child allocation is in the block submit path. Making it
 > > blocking with GFP_NOIO can consume the same global bio mempool that
 > > submit_bio() uses, while making it GFP_ATOMIC can fail under
 > > pressure. A forced failure of the child allocation produced:
 > >
 > > virtio_pmem: forcing child bio allocation failure for test
 > > Buffer I/O error on dev pmem0, logical block 0, lost sync page write
 > > EXT4-fs (pmem0): I/O error while writing superblock
 > > EXT4-fs (pmem0): mount failed
 > >
 > > Avoid the child bio completely. Flush FUA synchronously, like
 > > REQ_PREFLUSH, then complete the parent after the flush. Since no
 > > child bio can be created, async_pmem_flush() now only issues the
 > > virtio flush and preserves negative errno values.
 > 
 > Child flush is asynchronous (performs async flush to host side and returns).
 > Till child bio completes guest userspace waits in pending IO state.
 > It seems the current change will affect the behavior?
 > 
 > Prior RFC [1] attempted to coalesce the async FLUSH request between guest &host.
 > If there is interest, that approach could be revisited or integrated here?
 > 
 > [1] https://lore.kernel.org/all/20220111161937.56272-1-pankaj.gupta.linux@gmail.com/#t


Yes, agreed. This does change the FUA path.

I will rework it to avoid the child bio allocation while keeping the
parent bio completion asynchronous, using your RFC as reference.

Regards,
Li​