Re: mmap locked doesn't return

[Hugh Dickins <hughd@google.com>]

On Mon, 23 Jan 2012, Sagar Borikar wrote:
> 
> We are observing that mmap when called with MAP_SHARED and MAP_LOCKED
> flags on the disk which has the bad sectors, doesn't return either
> success or failure and gets stuck somewhere in mmap routine. Disk
> replacement is the obvious solution but the question is what makes
> mmap not to return failure?
> Individual mapping works fine i.e. on the same disk at same offset
> when mmap is invoked with MAP_SHARED or MAP_LOCKED the call succeeds
> but when they both are clubbed together, it doesn't return anything.
> 
> strace output snippet
> 
> open("/dev/dm-4p2", O_RDWR)             = 3
> mmap(NULL, 10485760, PROT_READ|PROT_WRITE, MAP_SHARED|MAP_LOCKED, 3, 0x4000000
> 
> 
> setup:
> 
> 2.6.23 LST kernel
> Westmere platform - 4 socket cpu
> RAM -192GB

2.6.23 is a four-year-old kernel.  There have been a lot of changes
since then, in mlocking and all over.  Here I suspect a change
in filemap_fault() may be relevant.  At the time of 2.6.23, its
page_not_uptodate IO error handling expected an IO error to be
reported by readpage() (which only initiates the read), and looks
in danger of retrying indefinitely.   Whereas in the current tree
I see a wait_on_page_locked() followed by EIO if !PageUptodate there.
My guess is a move to a more recent kernel would solve your problem.

Hugh