[Qemu-devel] On using MAP_PRIVATE (mmap) and QEMU

[Viswesh Narayanan]

Hi,

I did the following experiment wrt mmap and qemu (v2.3) and noticed a strange behavior.

Can anyone explain if this is expected (or perhaps a bug?).

I have a program which opens a file and mmaps the file contents into a private memory space. 

My program is same as http://beej.us/guide/bgipc/examples/mmapdemo.c, with only change - MAP_PRIVATE instead of MAP_SHARED.

If MAP_PRIVATE is used, ideally the mmap-ed region contents even if changed, must never be written back to the disk.

In my example, the driver program does NOT do any writing at all. 

It simply opens a read-only file and creates a new read-only mmap-ed space supported by that file. 

However, I do the writing differently from my QEMU emulation. 

I catch the mmap address returned in helper_sysret() and also capture the translated host virtual address (X) corresponding to that.

Then corrupt the data from  X to say X + 10 addresses. I notice that this causes the original file to be corrupted as well.

So ideally the contents written must never reach the disk right? Since it is equivalent to a write (from driver program) in MAP_PRIVATE region.

Let me know if you want the actual source code diff for reproducing this.

Thanks,

Viswesh