On Tue, 2018-04-03 at 13:54:02 UTC, Frederic Barrat wrote:
> cxllib_handle_fault() is called by an external driver when it needs to
> have the host resolve page faults for a buffer. The buffer can cover
> several pages and VMAs. The function iterates over all the pages used
> by the buffer, based on the page size of the VMA.
> To ensure some stability while processing the faults, the thread T1
> grabs the mm->mmap_sem semaphore with read access (R1). However, when
> processing a page fault for a single page, one of the underlying
> functions, copro_handle_mm_fault(), also grabs the same semaphore with
> read access (R2). So the thread T1 takes the semaphore twice.
> If another thread T2 tries to access the semaphore in write mode W1
> (say, because it wants to allocate memory and calls 'brk'), then that
> thread T2 will have to wait because there's a reader (R1). If the
> thread T1 is processing a new page at that time, it won't get an
> automatic grant at R2, because there's now a writer thread
> waiting (T2). And we have a deadlock.
> The timeline is:
> 1. thread T1 owns the semaphore with read access R1
> 2. thread T2 requests write access W1 and waits
> 3. thread T1 requests read access R2 and waits
> The fix is for the thread T1 to release the semaphore R1 once it got
> the information it needs from the current VMA. The address space/VMAs
> could evolve while T1 iterates over the full buffer, but in the
> unlikely case where T1 misses a page, the external driver will raise a
> new page fault when retrying the memory access.
> Fixes: 3ced8d730063 ("cxl: Export library to support IBM XSL")
> Cc: sta...@vger.kernel.org # 4.13+
> Signed-off-by: Frederic Barrat <fbar...@linux.vnet.ibm.com>
Applied to powerpc next, thanks.