ZFS caching in action on SMB Direct with 100gbit connection... but locally crashes.

[besterino]

So, ZFS delivers more than twice the physical speed of the underlying NVME SSD over a network connection. Impressive!

Unfortunately, however, running chrystal disk mark locally on the ZFS pool crashed the whole server... after reboot the pool was lost (zpool status --> no pools available).

More details:

Windows Server 2025 Desktop Experience (evaluation)
ZFSonWindows 2.2.6-rc9
AMD Threadripper 3960X
64GB RAM
Various SSDs

Created ZFS pool on clean 2TB NVME SSD with: "zpool create -O casesensitivity=insensitive -o ashift=12 -O atime=off zfswin physicaldrive2"

Set driveletter with: zfs set driveletter=P zfswin

Running CDM on P: hangs for long time at "Preparing...Sequential Read". Clicking on "Stop" in CDM does nothing. After a while Server crashed and rebooted. Event log:

Faulting application name: zfs.exe, version: 1.0.0.1, time stamp: 0x6733ebcb
Faulting module name: ntdll.dll, version: 10.0.26100.2033, time stamp: 0x9b03969e
Exception code: 0xc0000005
Fault offset: 0x000000000001a9dd
Faulting process id: 0x11EC
Faulting application start time: 0x1DB3912D569C588
Faulting application path: C:\Program Files\OpenZFS On Windows\zfs.exe
Faulting module path: C:\WINDOWS\SYSTEM32\ntdll.dll
Report Id: 40b74776-ddda-46c4-9045-93fe816a032e
Faulting package full name:
Faulting package-relative application ID:

When trying to reproduce the behaviour, initially only CDM hung and again I could not stop the benchmark. But Server crashed again when I tried to terminate CDM through taskmanager when clicking on "Stop" in CDM did nothing. At least after second crash pool was still available and able for import.

Simply writing/reading (file copy) on the ZFS pool seems to work.

CDM on same hardware / SSD with NTFS volume runs without errors.

Has anyone else experienced problems with crystal disk mark on a ZFS pool?

[lundman]

We had an issue with Crystal Disk earlier in the year, but I believed it was fixed. Either way, I'll have a look at it again.

[lundman]

I get deadlock alert running CrystalMark, so must be leaking a lock somewhere, I'll dig deeper.

[lundman]

OK so we call

			Status = fs_write_impl(DeviceObject, Irp, IrpSp,
			    wait, FALSE);
		}

	} except(EXCEPTION_EXECUTE_HANDLER) {
		Status = GetExceptionCode();
	}

and inside fs_write_impl() we get to around here:

	Status = zfs_write(zp, &uio, 0, NULL);

	if (!locked)
		MmUnlockPages(Irp->MdlAddress);

	if (!NT_SUCCESS(Status)) {
		dprintf("zfs_write returned %08lx\n", Status);
		goto end;
	}

in that zfs_write() completes and returns 0, then 2/3 steps below that we trigger an exception, which brings us all the way back to the except. At this point we have not unlocked the locks either.

We should;

1. unlock the locks when exceptions happen
2. work out why we get an exception, most likely due to UserPtr not mapped.

[lundman]

OK I have fixed this. But a mount issue came up, so let me address that as well before rc10.

[lundman]

OK I have pushed out rc10 which I hope will make CrystalDisk run better.

[besterino]

Thank you! Will test soon.