Sam Abu-Nassar wrote:Hello,
I posted this query a couple of weeks ago regarding a problem with remap_pfn_range. I was able to resolve the issue and I thought I would post my findings in case it helps someone else or results in a kernel fix. I will try to keep this short.
In a nutshell, my driver support user APIs that maps system RAM or a PCI BAR space to user space. What I did not mention in my original post is that my driver performed a sort of custom protocol when mmap() is called. Since mmap() really only provides a limited amount of space (the offset field) that I can use to pass additional information to my driver, I had implemented a custom protocol, which works as follows:
1. API calls mmap to obtain a user virtual address
2. Drivers mmap routine stores the VMA in an internal list and returns ok.
3. API then issues a custom IOCTL to driver to complete mapping with additional info
4. Driver retrieves VMA from internal list and performs mapping with io/remap_pfn_range, depending upon whether it's to system RAM or a PCI BAR.
The mappings always work fine, but starting with 2.6.15, the system freeezes when the file descriptor is closed. I tried numerous tests and compared my code with existing drivers, such as /dev/mem. Here is what I found:
The fix involved moving my calls to io/remap_pfn_range into my Dispatch_mmap() routine. Once I did this, the system no longer crashed. I still implement sending some custom information to the driver, but now I use special values in the offset field, remembering that the offset is eventually shifted by PAGE_SIZE by the time it reaches the driver. My driver code essentially did not change. In effect, all I really did was move it to the driver's mmap() routine.
I should mention that my original protocol has worked fine in kernels 2.2, 2.4, and up to 2.6.14. Some change to the VM subsystem in 2.6.15 broke my original code. I don't believe there should be an issue with calling remap_pfn_range outside of the driver's mmap() routine, but I am not a kernel developer, so I could be wrong in my assumption. One of my customers posed this question to Nick Piggin, and he seemed to think there should not be a problem with this.
Nick Piggin wrote:
Well, I think I said it shouldn't oops like this... I don't think it
is particularly robust WRT error cases or concurrent page faults
(between mmap and ioctl).
As we established earlier with a debug patch, the reason for the oops
is that VM_PFNMAP has been cleared from your vma->vm_flags for some
reason. This is causing the unmap code to mistakenly try to remove
reverse maps and refcounts from the struct pages.
I don't know why VM_PFNMAP should be getting cleared. But if it
remains set then the oops should go away.