Receive() indications and IRQL

[s0larian]

Hi there,

I have just spent a some time debugging an issue, and I was was wondering, if someone has seen something similar.

The test case:

• I have a kernel mode thread that, naturally, runs at PASSIVE_LEVEL
• every now and then the thread wakes up and submits a packet using the ReceiveHandler function from the original NDIS_OPEN_BLOCK

The issue:

• sometimes IRQL remains at DISPATCH_LEVEL after the ReceiveHandler has returned

In any case, I have managed to work around the problem by explicitly lowering the IRQL to PASSIVE_LEVEL, but, this really looks like a bug somewhere… Almost like a piece of MS code assumes that the handler should run at DISPATCH_LEVEL… Or may be they forget to unlock a spinlock…

Other observations:

• I think this happens when tcpip.sys handles a TCP+SYN, and a Send function is invoked directly with a SYN/ACK… but I am not completely certain.
• This particular problem is detected by the driver verifier when submitting a packet from a DeviceIOControl handler, since IRQL is different on the way out of the handler.
• Another similar thing happens when my thread calls KeWaitForSingleObject() and the verifier bitches about IRQL being too high.

Any ideas or suggestions?
Thanks in advance!

[Vadim Smirnov]

Protocol driver expects that its ProtocolReceive handler is called at DISPATCH_LEVEL (refer the DDK documentation where it is explicitely specified “…ProtocolReceive runs at IRQL = DISPATCH_LEVEL…”).

Since protocol driver assumes that it was called at DISPATCH_LEVEL then it could use KeAcquireSpinLock/KeReleaseSpinLockFromDpcLevel, an example instead KeAcquireSpinLock/KeReleaseSpinLock or something similar. It is not a good sample but when called on DISPATCH_LEVEL both pairs work on the same way.

[s0larian]

Oh, I think I get it…. I was tracing KeAcquireSpinLock() yesterday, and it only incremented the IRQL. That’s all that’s required on uniprocessor kernels, actually.

So, in this case, I wonder if it’s a good idea to raise IRQL to dispatch prior invoking Receive() handlers?

Thanks!

[Vadim Smirnov]

So, in this case, I wonder if it’s a good idea to raise IRQL to dispatch prior invoking Receive() handlers?

Yes, this is basically what you have to do.

[s0larian]

Thank you, I appreciate your help.

[Author]

Posts