|
User Info |
|
You are not logged in.
|
|
»
Login
«
|
|
»
Register
«
|
Your cart is empty.
Total:
$0.00
|
Mission Statement
History |
301_Case Studies - The Debugger That Froze The Firewall
Case study: a non-paged pool leak in kd.exe froze a Hyper-V host and took the firewall, and the network, offline. How it was found, reproduced in a lab and fixed using only user-mode tools, written up with every piece of evidence.
Recommended prerequisite training modules:
033, 055, 175, 192
Reading time:
23 min
Price:
$0.85
A non-paged pool leak in kd.exe that turned dump analysis into a site-wide outage. Found, reproduced and explained without a kernel debugger. Investigated 2026-11-18 and 2026-11-19 at Rokroot.
SummaryOn 18 November 2026 a Hyper-V host named VMHOST froze solid four times in three hours, and also bugchecked once. VMHOST runs the site's firewall as a virtual machine, so every freeze took the whole network off the internet. The owner had a strong, repeatable observation: "if I RDP into VMHOST it locks up in under ten minutes; at the console it doesn't." The cause turned out to be the debugger being used to investigate a different crash. - What leaked:
kd.exe, started non-interactively with its standard input at end-of-file, leaked about 120,000 pending write IRPs per second. That is about 75 MB/s of non-paged pool. - Where they went: the writes went into its own internal named pipe, and none of them ever completed.
- What happened next: about two minutes later the host ran out of non-paged memory and stopped. There was no bugcheck. The disk light stayed busy, and every VM on it stopped too.
- Why RDP looked guilty: the owner reconnected over RDP to keep the investigation going, and the investigating agent then started kd again.
This page walks through: - the outage and the evidence;
- how RDP was cleared;
- a safe lab reproduction, including a safety net that failed;
- how the leak was localized with only user-mode tools:
poolmon, a non-invasive cdb attach, handle reference counts and a 40-line handle probe; - the seven-run test matrix that isolated the trigger;
- the fix.
| Root cause (shown) | kd's console input thread loops when stdin is at EOF. It writes into an overlapped named pipe that the busy debugger engine isn't reading, and every write stays pending. | | Trigger | stdin at EOF: a closed pipe from an automation shell, or NUL. | | Not the trigger | the symbol server, symsrv.dll version, cache state, RDP, debugger version. The 2017 SDK, 2023 SDK and current November 2026 WinDbg 1.2611 console debuggers all leak, kd and cdb -remote alike. | | Fix | Give kd/cdb a stdin that stays open (a pipe you never write to) until the process exits. Add a time limit and a pool watchdog. Never run it on a host that carries VMs you can't lose. |
The full case study, with every section above and all the evidence files, is for buyers and full-access subscribers.
|