On Fri, 13 Jun 2014, David Herrmann wrote:

We currently fail setting SEAL_WRITE in case there're pending page
references. This patch extends the pin-tests to wait up to 150ms for all
references to be dropped. This is still not perfect in that it doesn't
account for harmless read-only pins, but it's much better than a hard
failure.

Signed-off-by: David Herrmann <redacted>

Right, I didn't look through the patch itself, just compared the result
with what I sent.  Okay, you prefer to separate out shmem_tag_pins().

Yes, it looks fine.  There's just one change I'd like at this stage,
something I realized shortly after sending the code fragment: please
add a call to lru_add_drain() at the head of shmem_tag_pins().  The
reason being that lru_add_drain() is local to the cpu, so cheap, and
in many cases will bring down all the raised refcounts right then.

Whereas lru_add_drain_all() in the first scan of shmem_wait_for_pins()
is much more expensive, involving inter-processor interrupts to do
that on all cpus: it is appropriate to call it at that point, but we
really ought to try the cheaper lru_add_drain() at the earlier stage.

I would also like never to embark on this scan of the radix_tree
and wait for pins, if the pages were never given out in a VM_SHARED
mapping - or is that unrealistic, because every memfd is read-write,
and typical initialization expected to be by mmap() rather than write()?
But anyway, you're quite right not to get into that at this stage:
it's best left as an optimization once the basics are safely in.

Hugh

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>