[Bug 221269] New: Linux kernel (6.18-6.19) memory leak issue QEMU GPU-Passthrough (RDNA3 ?)

[bugzilla-daemon@kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=221269

            Bug ID: 221269
           Summary: Linux kernel (6.18-6.19) memory leak issue QEMU
                    GPU-Passthrough (RDNA3 ?)
           Product: Virtualization
           Version: unspecified
          Hardware: AMD
                OS: Linux
            Status: NEW
          Severity: high
          Priority: P3
         Component: kvm
          Assignee: virtualization_kvm@kernel-bugs.osdl.org
          Reporter: hitomiyukinori@gmail.com
        Regression: No

System Info
Void 6.14.11_1 x86_64 AuthenticAMD uptodate rrrmFFFFFFFF
Ryzen 5 3600
ASUS DUAL RX 7800 XT

Package(s) Affected
linux6.18-6.18.19_1, linux6.19-6.19.9_1

Hello, I recently discovered a very strange cause of memory leakage, The leak
was detected in cases where the following versions of kernels were used :
6.18.19, 6.19.9. Resources (RAM usage) are not properly released when the same
QEMU virtual machine is turning off. This only happens when gpu-passthrough is
used.

Expected behaviour

When I use gpu-passthrough with kernel 6.14.11_1, before powering on the
virtual machine, overall ~ 950 megabytes of RAM is used (checked via ssh+htop),
I’m allocating 14,400 megabytes (14.4GB) of memory for VM, powering on the VM,
and getting something around ~ 14.9 GB overall is used (my hook script offloads
display manager and daemons).

When I make a cycle, power-on, power-off, repeat, like 5-6 times, i am getting
stable results regarding memory allocation / memory used. ~ 950 megabytes when
the VM is OFF, 14.9 GB when the VM is ON

Actual behaviour

Every "power-on, power-off" cycle, something around ~ 300 megabytes of RAM gets
stuck somewhere and can’t be released. Every "power-on, power-off" cycle i am
getting +~300 mb of RAM added to overall "used" / "allocated" by host system.

Steps to reproduce

That is how it works with kernels 6.18.19, 6.19.9 :

1) VM powered OFF, ~950 mb of overall RAM used, VM powered ON, ~ 14.9 GB of
overall RAM used, repeating the cycle

2) VM powered OFF, ~1.2-1.3 GB of overall ram used now, VM powered ON, ~
15.2-15.3 GB of overall ram used now, repeating the cycle

3) ~1.6-1.7 GB of overall ram used now. VM cannot be powered on, because QEMU
cannot allocate 14,400 megabytes (14.4GB) of memory for VM now. Because 15.3 GB
+ 300 MB > Total ram available

This is how memory leakage occurs, without any new running processes, without
changes inside VM, just powering-on VM, powering-off VM, leakage occurs by
repeating this cycle.

-- 
You may reply to this email to add a comment.

You are receiving this mail because:
You are watching the assignee of the bug.