Good way to detect a DLL unload

Albert

The PsSetLoadImageNotify callback only tells us when a DLL is loaded, and since DLLs are memory mapped, the FILEOBJECT which we get as part of the same callback is also closed immediately after, so tracking IRP_MJ_CLEANUP doesn't cut it.

Is there a good way to track image unloads? Is there a backing FILE_OBJECT for the memory mapped sections which can somehow be used to accomplish this task?

Scott_Noone_(OSR)

No way...

There's no callback to tell you when a section is close (i.e. CloseHandle on the section).

Mm keeps a reference to the first file object used to create the section so that the data can be cached long after it's unmapped/closed. If you see an IRP_MJ_CLOSE for the last file object for the stream you definitively know it's no longer mapped. However, this IRP_MJ_CLOSE may not actually ever arrive (e.g. Mm doesn't purge sections on shutdown)

Albert

@Scott_Noone_(OSR) said:
No way...

There's no callback to tell you when a section is close (i.e. CloseHandle on the section).

Mm keeps a reference to the first file object used to create the section so that the data can be cached long after it's unmapped/closed. If you see an IRP_MJ_CLOSE for the last file object for the stream you definitively know it's no longer mapped. However, this IRP_MJ_CLOSE may not actually ever arrive (e.g. Mm doesn't purge sections on shutdown)

So, if I am reading this right, the Destroy event for the stream context also is not guaranteed to tell me about a DLL unload. I was hoping that the stream context destroy event was my only hope here.

Scott_Noone_(OSR)

So, if I am reading this right, the Destroy event for the stream context also is not guaranteed to tell me about a DLL unload. I was hoping that the stream context destroy event was my only hope here.

Correct. The DLL has to be unloaded for the the stream context destroy callback to fire but the callback can be delayed indefinitely.