On 03/11/17 02:22 AM, Sebastian Andrzej Siewior wrote:
On 2017-11-02 14:53:19 [-0600], Logan Gunthorpe wrote:
Thank you for letting me go all that way…
Sorry, but in fairness I did have the same feedback to your first patch
which you never addressed. This patch would have to be part of a well
justified patch-set changing how completions works. But, even then, I'd
rather see a better solution that still allows at least some type of
completion to be polled.
You should not touch the inside of an implementation.
This is really only sometimes true[1]. But I feel it's a moot point: had
I created a poll_on_completion() accessor function, you'd still be in
the same situation. Frankly, I'm shocked no one else in the kernel is
polling on a completion, it seems like something that would be common.
Logan
[1]"So it is essential that data types are not overly abstracted, but
that all details of the implementation are visible so that the
programmer can make optimal choices when using them."
-- https://lwn.net/Articles/336255/
