Re: downgrade_write replacement in remap_file_pages

[William Lee Irwin III <wli@holomorphy.com>]

On Tue, Jun 08, 2004 at 05:44:38PM +0200, Andrea Arcangeli wrote:
> Apparently downgrade_write deadlocks the kernel in the mmap_sem
> under load. I suspect some rwsem bug. Anyways it's matter of time before
> in my tree I replace the rwsem implementation with my spinlock based
> common code implementation again that I can understand trivially (unlike
> the current code). I don't mind a microoptimization when the code is so
> complicated, so I don't mind too much to fix whatever current bug in
> downgrade_write.
> In the meantime to workaround the deadlock (and to verify if this make
> the deadlock go away, which returned a positive result) I implemented
> this patch: this doesn't fix downgrade_wite, but I'm posting it here
> because I believe it's much better code regardless of the
> downgrade_write issue.  With this patch we'll never run down_write again
> in production, the down_write will be taken only once when the db or the
> simulator startup (during the very first page fault) and never again, in
> turn providing (at least in theory) better runtime scalability.

I've been using something similar since about May 20. However, it was
unclear that it was a deadlock as opposed to semaphore contention from
the reports I got.


On Tue, Jun 08, 2004 at 05:44:38PM +0200, Andrea Arcangeli wrote:
> -		if (pgoff != linear_page_index(vma, start) &&
> -		    !(vma->vm_flags & VM_NONLINEAR)) {
> +		if (unlikely(pgoff != linear_pgoff && !(vma->vm_flags & VM_NONLINEAR))) {

There is no linear_pgoff variable...


-- wli