help on lock around vm_page

[Howard Su]

Some further reading give me a different message:
The pages returned by vm_page_grab should have VPO_BUSY flag which will
prevent anyone else to move this page to cache. So current approach should
be ok. Am I right? Maybe there is some edge cases since I only meet this bug
in a low memory situation.


-- 
-Howard

On 6/7/07, Howard Su <howard0su at gmail.com> wrote:
>
> I want some helps from VM guru. I try to fix a panic in tmpfs. In
> order to push tmpfs into -Current, I really want some help to solve
> this.
>
> 1. we allocate an object from vm_pager_alloc(OBJT_SWAP, ...) when create a
> file.
> 2. the panic is during handling write op:
>         a) find the first page we want to write
>         b) call vm_page_grab to get the page from object.
>         c) call use sf_buf_alloc to map it into kernel_map
>         d) use uiomove to move the data
>         e) mark page as dirty
>         f) loop to a until all pages are handled.
>
> there is a race condition. while doing b-c & e, we hold the
> OBJ_LOCK/page_queue_lock. when doing d, we have to drop the locks to
> call uiomove. When calling uio move, the page may moved to cache queue
> since in that time it is not dirty.
>
> There is a solution that we allocate a page buffer. Before a), we
> uiomove it to the buffer and replace uiomove with a bcopy in d). Then,
> we can hold lock in b - e. I feel this will cause performance problem.
>
> For the detailed code, please check:
>
> http://perforce.freebsd.org/fileViewer.cgi?FSPC=//depot/user/howardsu/truss/sys/fs/tmpfs/tmpfs%5fvnops.c&REV=30
>
> function: tmpfs_uio_xfer()
>
> Any idea to close this race condition?
>
> PS: If you can review my code about usage of vm, it will be appreciated.
>
> --
> -Howard
>