Hello,

I did the verification for noble

$ mkdir /tmp/qemu-modules
$ wget 
https://launchpad.net/ubuntu/+archive/primary/+files/qemu-block-extra_10.2.1+ds-1ubuntu3_amd64.deb
$ ar x qemu-block-extra_10.2.1+ds-1ubuntu3_amd64.deb 
$ tar xvf data.tar.zst 
$ cp usr/lib/x86_64-linux-gnu/qemu/block-curl.so /tmp/qemu-modules/

CURRENT SITUATION

$ QEMU_MODULE_DIR=/tmp/qemu-modules/ qemu-img --trace "module_*" info 
https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
qemu-img: Could not open 
'https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img':
 failed to initialize module: /tmp/qemu-modules//block-curl.so
Only modules from the same build can be loaded.

WITH THE PROPOSED

$ apt policy qemu-utils
qemu-utils:
  Installed: 1:8.2.2+ds-0ubuntu1.18
  Candidate: 1:8.2.2+ds-0ubuntu1.18
  Version table:
 *** 1:8.2.2+ds-0ubuntu1.18 100
        100 http://archive.ubuntu.com/ubuntu noble-proposed/main amd64 Packages
        100 /var/lib/dpkg/status
     1:8.2.2+ds-0ubuntu1.17 500
        500 http://archive.ubuntu.com/ubuntu noble-updates/main amd64 Packages
     1:8.2.2+ds-0ubuntu1.16 500
        500 http://security.ubuntu.com/ubuntu noble-security/main amd64 Packages
     1:8.2.2+ds-0ubuntu1 500
        500 http://archive.ubuntu.com/ubuntu noble/main amd64 Packages


$ QEMU_MODULE_DIR=/tmp/qemu-modules/ qemu-img --trace "module_*" info 
https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
module_load_module file /usr/lib/x86_64-linux-gnu/qemu/block-curl.so
image: 
https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
file format: qcow2
virtual size: 3.5 GiB (3758096384 bytes)
disk size: unavailable
cluster_size: 65536
Format specific information:
    compat: 1.1
    compression type: zlib
    lazy refcounts: false
    refcount bits: 16
    corrupt: false
    extended l2: false
Child node '/file':
    filename: 
https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
    protocol type: https
    file length: 819 MiB (858388480 bytes)
    disk size: unavailable

** Description changed:

  [ Impact ]
  
  After upgrading QEMU packages on a host, QEMU fails to load some modules
  in /run/qemu/<version>
  
  This might affect running instances in many ways but the original report is 
about long-running instances that started with the old build can
  no longer hot-attach module-backed block devices such as Ceph RBD volumes.
  
  This is a regression introduced by upstream commit c551fb0b53d ("module: add 
Error arguments to module_load and module_load_qom") which broke the directory
  fallback logic in module_load(). The CONFIG_MODULE_UPGRADES feature
  (enabled in Ubuntu builds) explicitly adds /run/qemu/<version>/ to the
  search list for this scenario — but the broken logic prevents it from
  ever being looked at.
  
  The fix restores the correct fallback behaviour (continue searching
  remaining directories when a file exists but fails to load) and fixes a
  related memory leak in module_load_dso() that caused the assertion crash
  on retry.
  
  [ Test Plan ]
  
  The idea of the test plan is to demonstrate that QEMU stops searching
  for the module at the first module load failure.
  
  For that we can use the curl-block.so module that is easy to trigger the 
loading
  in QEMU.
  
  Steps to reproduce (and verify the fix):
  
  1. Create /tmp/modules/
  2. Get a curl-block.so from a qemu-utils package that does not match the 
version of the qemu-utils we have on the system.
-    On Resolute, we can take: 
https://launchpad.net/ubuntu/+archive/primary/+files/qemu-block-extra_10.2.1+ds-1ubuntu3_amd64.deb
+    On Resolute, we can take: 
https://launchpad.net/ubuntu/+archive/primary/+files/qemu-block-extra_10.2.1+ds-1ubuntu3_amd64.deb
  3. Extract the contents and copy the block-curl.so to /tmp/qemu-modules
  4. Run qemu-img:
  
- qemu-img --trace "module_*" info https://cloud-
- images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
+ QEMU_MODULE_DIR=/tmp/qemu-modules/ qemu-img --trace "module_*" info
+ https://cloud-images.ubuntu.com/stonking/current/stonking-server-
+ cloudimg-amd64.img
  
  WITHOUT THE FIX:
  
  This command will fail:
  
  qemu-img: Could not open 
'https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img':
 failed to initialize module: /tmp/qemu-modules//block-curl.so
  Only modules from the same build can be loaded.
- 
  
  WITH THE FIX:
  
  This command succeeds:
  
  qemu-img: failed to initialize module: /tmp/qemu-modules//block-curl.so
  Only modules from the same build can be loaded.
  module_load_module file /usr/lib/x86_64-linux-gnu/qemu/block-curl.so
  image: 
https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
  file format: qcow2
  virtual size: 3.5 GiB (3758096384 bytes)
  disk size: unavailable
  cluster_size: 65536
  Format specific information:
-     compat: 1.1
-     compression type: zlib
-     lazy refcounts: false
-     refcount bits: 16
-     corrupt: false
-     extended l2: false
+     compat: 1.1
+     compression type: zlib
+     lazy refcounts: false
+     refcount bits: 16
+     corrupt: false
+     extended l2: false
  Child node '/file':
-     filename: 
https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
-     protocol type: https
-     file length: 819 MiB (858388480 bytes)
-     disk size: unavailable
+     filename: 
https://cloud-images.ubuntu.com/stonking/current/stonking-server-cloudimg-amd64.img
+     protocol type: https
+     file length: 819 MiB (858388480 bytes)
+     disk size: unavailable
  
  [ Where problems could occur ]
  
  The change only affects the module directory search loop in
  util/module.c:module_load() and the cleanup path in
  util/module.c:module_load_dso(). No other code paths are touched.
  
  Regressions are possible in module loading. The most common real-world
  scenario is hot-attaching block/network devices (RBD, virtio-fs, etc.)
  that load modules late in the VM lifecycle.
  
  The fix is conservative: it restores the pre-regression legacy behaviour
  of continuing the search after a failed module load. The only observable
  difference is that QEMU will now successfully load the correct retained
  module from /run/qemu/<old-version>/ instead of stopping at the first
  (mismatched) module in /usr/lib/.
  
  [ Other Info ]
  
  Upstream submission: [PATCH v2 1/1] module: fix early stop for load module 
function
    https://lists.gnu.org/archive/html/qemu-devel/2026-04/msg06376.html
  
  Upstream tracking issue:
    https://gitlab.com/qemu-project/qemu/-/work_items/3354
  
  Original bug report
  ---
  
  ## Package
  
  qemu-system-x86 (Ubuntu noble)
  
  ## Affects
  
  qemu (Ubuntu)
  
  ## Related bugs
  
  - LP #1847361 (Upgrade of qemu binaries causes running instances to be unable 
to hot-attach)
  - LP #1913421 (module retention improvements)
  
  ## Description
  
  ### Summary
  
  After upgrading QEMU packages on a compute node (e.g. from
  `1:8.2.2+ds-0ubuntu1.12` to `0ubuntu1.13`), long-running VM instances
  started with the older build can no longer hot-attach Ceph RBD volumes —
  even though `/run/qemu/` contains the retained modules for the old
  build.
  
  The first attach attempt fails with "Unknown driver 'rbd'". A second
  attempt crashes QEMU with an assertion failure.
  
  This is a regression in the module-retention mechanism introduced for LP
  #1847361.
  
  ### Root cause
  
  Two bugs in `util/module.c` (confirmed identical on current QEMU master
  as of 2026-03-26):
  
  **Bug A — module_load() does not fall back on build mismatch:**
  
  The directory search loop (lines 282–303) only continues to the next
  directory when the module file is not found (`ENOENT`). When the file
  exists but `module_load_dso()` fails (build mismatch), the loop hits
  `goto out` immediately — never reaching `/run/qemu/<version>/`.
  
  `CONFIG_MODULE_UPGRADES` is enabled in the Ubuntu noble build
  (`debian/rules`: `$(if ${enable-system},--enable-module-upgrades)`), so
  the `/run/qemu/<version>/` path is added to the search list — but it is
  never reached because the system path (`/usr/lib/x86_64-linux-
  gnu/qemu/`) contains the new build's modules, which exist but fail the
  stamp check.
  
  **Bug B — module_load_dso() leaks dso_init_list on failure:**
  
  When `g_module_open()` loads a `.so`, its constructors populate
  `dso_init_list`. On build mismatch, `g_module_close()` is called but
  `dso_init_list` is not drained. On the next module load attempt,
  `assert(QTAILQ_EMPTY(&dso_init_list))` fires and QEMU aborts.
  
  ### Environment
  
  - Ubuntu 24.04 (noble), OpenStack compute nodes (Nova Victoria, libvirt/kvm, 
Cinder/Ceph RBD)
  - Kernel: `6.14.0-37-generic #37~24.04.1-Ubuntu SMP PREEMPT_DYNAMIC x86_64`
  - QEMU: `qemu-system-x86 1:8.2.2+ds-0ubuntu1.13`
  - libvirt: 10.0.0-2ubuntu8.12
  - AppArmor: enabled, no DENIED entries for `/run/qemu` or `block-rbd.so`
  - `/run/qemu` mounted as tmpfs (rw, no noexec)
  
  ### Observed symptoms
  
  **Instance started with QEMU 0ubuntu1.11, host upgraded to
  0ubuntu1.13:**
  
  Instance log (first attach attempt):
  ```
  failed to initialize module: /usr/lib/x86_64-linux-gnu/qemu/block-rbd.so
  Only modules from the same build can be loaded.
  ```
  
  libvirt:
  ```
  internal error: unable to execute QEMU command 'blockdev-add': Unknown driver 
'rbd'
  ```
  
  VM continues running, but attach fails. `/proc/$PID/maps` shows no
  mapping of `block-rbd.so`.
  
  Second attempt — instance log:
  ```
  qemu-system-x86_64: util/module.c:165: module_load_dso: Assertion 
`QTAILQ_EMPTY(&dso_init_list)' failed.
  ```
  
  QEMU exits (`reason=crashed`), VM ends up in SHUTOFF state.
  
  At the time of the failure, the retained modules exist:
  ```
  /run/qemu/Debian_1_8.2.2+ds-0ubuntu1.12/block-rbd.so   (40312 bytes, readable)
  /run/qemu/Debian_1_8.2.2+ds-0ubuntu1.11/block-rbd.so   (40312 bytes, readable)
  ```
  
  This has been reproduced across multiple minor build upgrades
  (0ubuntu1.11→12 and 0ubuntu1.12→13).
  
  ### Steps to reproduce
  
  1. Start an OpenStack instance on a compute node running QEMU 
`1:8.2.2+ds-0ubuntu1.X`. The instance must not use RBD at boot.
  2. Upgrade QEMU on the host to `0ubuntu1.(X+1)` while the instance keeps 
running.
  3. Verify `/run/qemu/Debian_1_8.2.2+ds-0ubuntu1.X/block-rbd.so` exists.
  4. Hot-attach a Cinder/Ceph RBD volume (`openstack server add volume`).
  5. First attempt: "Unknown driver 'rbd'".
  6. Second attempt: QEMU assertion crash.
  
  ### Impact
  
  - Long-running VMs that predate a QEMU package upgrade cannot hot-attach RBD 
volumes (or any other module-backed driver not already loaded).
  - Second attempt crashes the VM, causing unplanned downtime.
  - Defeats the purpose of the `/run/qemu/` module-retention mechanism (LP 
#1847361, LP #1913421).
  
  ### Proposed fix
  
  See upstream QEMU GitLab issue (https://gitlab.com/qemu-
  project/qemu/-/work_items/3354) for detailed code analysis and patch
  proposals. Summary:
  
  - **Bug A:** On `module_load_dso()` failure, clear the error and `continue` 
to the next directory instead of `goto out`.
  - **Bug B:** In `module_load_dso()`, drain `dso_init_list` before 
`g_module_close()` when the stamp check fails.
  
  Both fixes are against upstream `util/module.c` — the code is identical
  on current QEMU master.
  
  ### Current workaround
  
  Proactively reboot or live-migrate any instance whose running QEMU
  version (via QMP `query-version`) does not match the installed package
  version, before hot-attaching RBD volumes.

** Tags removed: verification-needed-noble
** Tags added: verification-done-noble

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2146445

Title:
  qemu-system-x86: module upgrade fallback in /run/qemu/ broken —
  "Unknown driver 'rbd'" + crash on retry

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/qemu/+bug/2146445/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to