Fix it now
0x0000004A means a thread returned to user mode from a system call with its interrupt request level still above passive. Something in kernel mode raised IRQL, took a spinlock or entered a critical region and never came back down. On a virtualisation host the cause is rarely the hypervisor; it is the stack of filter drivers layered on one I/O path.
fltmc filters
driverquery /v
- Note the stop code and any module named on the stop screen. If you missed it, open Event Viewer, go to Windows Logs then System, and read the bugcheck event and its parameters.
- Count the third-party entries in the
fltmc filtersoutput. More than one anti-malware filter, or an anti-malware filter plus two or three backup and monitoring filters, is where the investigation starts. - Uninstall, rather than merely disable, every endpoint agent except the one you intend to keep, using each vendor’s own removal utility. Reboot between removals.
- If stops continue, point Driver Verifier at the remaining third-party drivers only with
verifier /standard /driver driver1.sys driver2.sys, reboot, and let it name the offender. - Switch it off again with
verifier /resetand reboot before handing the host back.
Disabling real-time protection stops the scanning engine but usually leaves the minifilter loaded and attached, so the driver is still in the path. Removal is what changes the filter list, and fltmc filters is how you prove it happened.
If the filter count drops and the host stays up, you are done. If it does not, the next section explains what IRQL is protecting and how 0x0000002B differs from the rest of this family.
Why it happens
IRQL is the kernel’s interrupt priority. Ordinary thread code runs at passive level, the only level at which the scheduler may swap the thread out or a page fault may be serviced. A driver raises IRQL when it must not be interrupted: taking a spinlock puts the processor at dispatch level. Every raise has to be matched by a lower before control leaves the driver.
When the system service dispatcher returns a thread to user mode it checks IRQL is back at passive level. If it is not, letting user code run with interrupts and paging partly disabled would corrupt the machine quietly instead of loudly, so the kernel picks loudly. Parameter 1 of 0x0000004A is the address of the system routine involved and parameter 2 is the IRQL still in force.
0x0000002B is a different failure that arrives from the same conditions. It says the kernel mode stack was overrun, and Microsoft attributes that to a kernel mode driver using too much stack space. Each kernel thread has a fixed, small stack that does not grow, and one file operation on a host may pass through a backup filter, an anti-malware filter, a second one that was never fully removed, a monitoring agent, the snapshot provider and the storage stack. Depth is the whole problem, and the fix for that one is arithmetic: fewer layers.
The two AutoBoost codes in this family are more specific again. 0x00000162 says a lock tracked by AutoBoost was released by a thread that did not own it, which the documentation attributes to one thread releasing a lock on behalf of another, something AutoBoost tracking does not permit. 0x00000192 says an AutoBoost lock was acquired while running at dispatch level or above, which cannot be allowed because the lock may be held exclusively by the interrupted thread and the result would be a deadlock. 0x00000036 is the odd one out: a driver tried to delete a device object that still had a positive reference count, which is what you tend to see during an agent uninstall rather than under load.
Two anti-malware products are filtering the same volumes
You have this one if fltmc filters lists filters from two different security vendors and stops get worse under file-heavy load.
- Decide which product stays. Two real-time scanners on one host is not supported by either vendor.
- Remove the other with the vendor’s dedicated removal tool; a normal uninstall often leaves the minifilter registered.
- Reboot, run
fltmc filtersagain and confirm the filter is gone. - Apply the surviving product’s documented exclusions for virtual disk files, configuration files and the snapshot workspace.
Exclusions are not optional on a host. Scanning a running virtual disk byte by byte creates exactly the deep, long-lived I/O this stack handles worst.
A filter is left behind from software that is no longer installed
You have this one if fltmc filters names a filter you cannot match to anything in the installed programs list.
- Record the filter name exactly as fltmc reports it.
- Look for a matching service under
HKLM\SYSTEM\CurrentControlSet\Servicesand check whether the driver file still exists on disk. - Run the vendor’s cleanup utility if one exists; it removes the service, the instances and the altitude registration together.
- Only if no tool exists, remove the orphaned service key and reboot.
An orphaned filter that still loads is worse than a supported one, because nobody is shipping fixes for it.
Kernel stack exhaustion rather than a lock fault
You have this one if The stop code is 0x0000002B, or the dump shows a very deep call stack made almost entirely of third-party modules.
- Count the filters again. The remedy here is fewer layers, not different settings.
- Remove monitoring, data loss prevention and legacy backup agents that duplicate what current tooling already does.
- Where a backup vendor offers an agentless or host-integrated mode for virtual machines, use it instead of an in-host filter.
- Confirm the filter count has dropped before returning the host to production.
A vendor driver is older than the Windows build it runs on
You have this one if Stops began after a host update and the module in the dump belongs to a storage controller, network teaming or endpoint driver.
- Run
driverquery /vand note the link dates of the third-party drivers. - Check each vendor’s compatibility statement for the build you are now running.
- Update the mismatched drivers, taking storage and network last so you keep a route back in.
- If no compatible version exists, remove the product rather than leaving an unsupported driver loaded.
Full reference
Reading the symptom back to the layer
| What you observe | Where to look first |
|---|---|
| Stops cluster during the backup window | A backup filter interacting with an anti-malware filter on the same volume |
| Stop code 0x0000002B | Kernel stack exhaustion. Too many filters, not a lock bug |
| Stop while uninstalling or updating an agent | Device objects still referenced at unload, the 0x00000036 pattern |
| Stop code 0x00000162 or 0x00000192 | An AutoBoost tracked lock released by the wrong thread, or acquired at dispatch level or above |
| Stops began after a host cumulative update | A vendor driver that predates the new build |
What each code in this family actually says
| Stop code | Name | Published meaning |
|---|---|---|
| 0x0000004A | IRQL_GT_ZERO_AT_SYSTEM_SERVICE | A thread is returning to user mode from a system call with IRQL still above passive level |
| 0x0000002B | PANIC_STACK_SWITCH | The kernel mode stack was overrun, normally because a driver used too much stack space |
| 0x00000036 | DEVICE_REFERENCE_COUNT_NOT_ZERO | A driver tried to delete a device object that still had a positive reference count |
| 0x00000162 | KERNEL_AUTO_BOOST_INVALID_LOCK_RELEASE | An AutoBoost tracked lock was released by a thread that did not own it |
| 0x00000192 | KERNEL_AUTO_BOOST_LOCK_ACQUISITION_WITH_RAISED_IRQL | An AutoBoost tracked lock was acquired while executing at dispatch level or above |
Driver Verifier on a host you cannot lose
verifier /standard /driver vendor1.sys vendor2.sys
verifier /querysettings
rem after it has named the driver
verifier /reset
/standardactivates the default option set, which includes special pool, IRQL checking, pool tracking, I/O verification, deadlock detection, DMA verification and the security and DDI compliance checks./drivertakes a space-separated list of binary names. Wildcards are not supported, so name each driver explicitly./querysettingsshows what will be verified after the next boot, and is the check to run before you hand a host back./resetclears all settings so that nothing is verified after the next boot.
Verifier deliberately bugchecks the machine on the first violation it finds. Never enable it on a production host in working hours, and make sure you can reach a recovery console to run verifier /reset if the machine will not boot.
Removing an agent so the filter actually goes
- Drain the host’s virtual machines and take it out of the cluster’s rotation before you start.
- Use the vendor’s dedicated removal utility. Programs and Features frequently leaves the minifilter registered and loading.
- Reboot and check
fltmc filters. If the filter is still listed, the removal did not finish. - Where no removal tool exists and the service key must go by hand, export the branch under
HKLM\SYSTEM\CurrentControlSet\Servicesfirst and keep recovery media to hand. - Reinstall only the product you are keeping, then confirm its filter appears at the altitude the vendor documents.
Removing the wrong key under Services can stop the machine booting. Export before you delete, and change one key at a time.
When a licence is the actual fix
If the real answer is that this host carries three endpoint agents because nobody retired the old contract, consolidating onto one properly licensed product is both the technical fix and the cheaper outcome. Prove that duplication is your problem first, with fltmc filters: if the offending filter is a leftover from software you already removed, deleting it costs nothing, and Windows Server already includes Microsoft Defender Antivirus if you simply want one fewer vendor. Where you do want a single managed console covering servers and workstations so the duplicates can go without leaving anything unprotected, ESET PROTECT Entry is one option we supply, and we can size the seat count and confirm which SKU covers virtualisation hosts before you commit.
Every code this article covers
| Code | What it points at | Source |
|---|---|---|
0x0000004A |
IRQL_GT_ZERO_AT_SYSTEM_SERVICE: a thread returned to user mode from a system call while IRQL was still above passive level | Microsoft Learn |
0x0000002B |
PANIC_STACK_SWITCH: the kernel mode stack was overrun, normally by a driver using too much stack space | Microsoft Learn |
0x00000036 |
DEVICE_REFERENCE_COUNT_NOT_ZERO: a driver tried to delete a device object that still had a positive reference count | Microsoft Learn |
0x00000162 |
KERNEL_AUTO_BOOST_INVALID_LOCK_RELEASE: a lock tracked by AutoBoost was released by a thread that did not own it | Microsoft Learn |
0x00000192 |
KERNEL_AUTO_BOOST_LOCK_ACQUISITION_WITH_RAISED_IRQL: an AutoBoost tracked lock was acquired at dispatch level or above | Microsoft Learn |
Confirm the fix worked
- Run
fltmc filtersand confirm the list contains exactly the filters you meant to keep. - Check the System log for bugcheck events and confirm none has appeared since the change.
- Run a full backup cycle and a storage-heavy workload on the host and confirm both complete.
- Run
verifier /querysettingsand confirm no drivers are left under verification. - Confirm the agent you kept reports itself healthy in its own console, so you have not traded a stop for an unprotected host.
Questions people ask about this
Can I disable the second anti-virus instead of removing it?
No. Disabling real-time protection stops the scanning engine but usually leaves the minifilter loaded and attached, so the driver is still in the I/O path and still consuming stack. Removal is what changes the filter list, and fltmc filters is how you prove it happened.
Does this stop code mean my hardware is failing?
Rarely. This is a software contract violation inside kernel code, not a memory or disk symptom. Run a memory test for reassurance if you like, but spend the effort on the filter stack.
How is 0x0000002B different from the rest?
It is the only one of the five that is about space rather than about locks. Microsoft attributes it to a driver using too much stack, so the answer is to reduce how many drivers are stacked on one path rather than to find a single misbehaving one.
Do I have to buy anything to fix it?
Usually not. Removing duplicate agents costs nothing and Windows Server already includes Microsoft Defender Antivirus. A purchase only makes sense if the product you keep is one you have no valid licence for, or if central management is worth paying for in its own right.
How do I get a dump if the host reboots too fast to read the screen?
Open System Properties, Advanced, Startup and Recovery. Set the dump type to at least a kernel memory dump and clear the automatic restart option so the stop screen stays up. Dumps land in the Windows folder and its Minidump subfolder.
