I've addressed the review comments in v3 of the series.

Feel free to skip v2, since I missed some of the comments there.

On 11/16/26 18:24, Konstantin Khorenko wrote:
I won't apply this version, please send v2 with the changes below.

On Mon, 16 Nov 2026 16:41:10 +0100, Alexander Ivanov 
<[email protected]> wrote:
Patch 8 requeues held requests to the dm queue on a noflush suspend,
so a table reload or a snapshot of a paused VM does not deliver the
error to the guest.
The kernel side holds up: the locking, the timer and suspend
handling, and the vhost-blk release fix all trace correctly. What is
missing is the part that makes the series usable from userspace:

1. Series: QEMU's pause cannot work with held requests. On PAUSED and
    FINISH_MIGRATE vhost_blk_vm_state() calls vhost_blk_stop(), which
    detaches the backend; in the kernel that is
    vhost_blk_drop_backend() -> vhost_blk_flush(), which waits for
    every in-flight request, held ones included, with the BQL held. So
    the pause the IO_ERROR uevent is meant to trigger, virsh suspend
    and a migration all hang until resubmit or fail_held, and a timer
    retry that fails again holds the request again. Say how QEMU is to
    pause such a VM, and link the QEMU and libvirt series.
2. Cover letter and patch 8: ploop suspends with a flush suspend for
    a snapshot and for grow (ploop_suspend_device(), cmd(..., 0)), so
    with today's ploop a snapshot of a paused VM still delivers the
    error. Also dm_reload_imglist() does not pass error_policy, so any
    ploop reload resets the policy to the default. Name the ploop
    changes this needs, or drop the snapshot claim.
3. Patch 5, message: at this patch ENOSPC reads are still held (patch
    3 keeps them), so "Only ENOSPC writes and flushes are held" and
    "reads are not held" are wrong until patch 7.
4. Patch 5, message: say that every dm-qcow2 device, CT disks
    included, now sends the uevent and gets a udev probe every 20
    seconds while it stays full.
5. Patch 7, message: under the enospc policy only ENOSPC is held, so
    drop "whatever the error" (Sashiko).
6. Fixes tags: patch 1 and patch 4 6e2ec304689f ("dm-qcow2: Pause IO
    and notify about ENOSPC events"), patch 2 2f9926b51be7 ("dm-qcow2:
    Rename MERGE_QIOS_MAX"), patch 9 c57a98d65796 ("drivers/vhost:
    vhost-blk accelerator for virtio-blk guests").

_______________________________________________
Devel mailing list
[email protected]
https://lists.openvz.org/mailman/listinfo/devel

Reply via email to