From: Pavel Begunkov <asml.silence@gmail.com>
To: David Laight <David.Laight@ACULAB.COM>,
Al Viro <viro@zeniv.linux.org.uk>
Cc: Christoph Hellwig <hch@infradead.org>,
Jens Axboe <axboe@kernel.dk>,
"linux-block@vger.kernel.org" <linux-block@vger.kernel.org>,
"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 2/2] iov_iter: optimise iter type checking
Date: Mon, 14 Dec 2020 17:25:37 +0000 [thread overview]
Message-ID: <5b65de70-19db-4572-d122-df65191ab098@gmail.com> (raw)
In-Reply-To: <b0e01a4dc3fc4afeb95b7be826ff2375@AcuMS.aculab.com>
On 14/12/2020 10:28, David Laight wrote:
> From: Pavel Begunkov
>> Sent: 13 December 2020 22:33
>>
>> On 11/12/2020 02:01, Al Viro wrote:
>>> On Thu, Nov 19, 2020 at 05:12:44PM +0000, Pavel Begunkov wrote:
>>>> On 19/11/2020 17:03, Christoph Hellwig wrote:
>>>>> On Thu, Nov 19, 2020 at 03:29:43PM +0000, Pavel Begunkov wrote:
>>>>>> The problem here is that iov_iter_is_*() helpers check types for
>>>>>> equality, but all iterate_* helpers do bitwise ands. This confuses
>>>>>> a compiler, so even if some cases were handled separately with
>>>>>> iov_iter_is_*(), it can't eliminate and skip unreachable branches in
>>>>>> following iterate*().
>>>>>
>>>>> I think we need to kill the iov_iter_is_* helpers, renumber to not do
>>>>> the pointless bitmask and just check for equality (might turn into a
>>>>> bunch of nice switch statements actually).
>>>>
>>>> There are uses like below though, and that would also add some overhead
>>>> on iov_iter_type(), so it's not apparent to me which version would be
>>>> cleaner/faster in the end. But yeah, we can experiment after landing
>>>> this patch.
>>>>
>>>> if (type & (ITER_BVEC|ITER_KVEC))
>>>
>>> There are exactly 3 such places, and all of them would've been just as well
>>> with case ITER_BVEC: case ITER_KVEC: ... in a switch.
>>>
>>> Hmm... I wonder which would work better:
>>>
>>> enum iter_type {
>>> ITER_IOVEC = 0,
>>> ITER_KVEC = 2,
>>> ITER_BVEC = 4,
>>> ITER_PIPE = 6,
>>> ITER_DISCARD = 8,
>>> };
>>> iov_iter_type(iter) (((iter)->type) & ~1)
>>> iov_iter_rw(iter) (((iter)->type) & 1)
>>>
>>> or
>>>
>>> enum iter_type {
>>> ITER_IOVEC,
>>> ITER_KVEC,
>>> ITER_BVEC,
>>> ITER_PIPE,
>>> ITER_DISCARD,
>>> };
>>> iov_iter_type(iter) (((iter)->type) & (~0U>>1))
>>> // callers of iov_iter_rw() are almost all comparing with explicit READ or WRITE
>>> iov_iter_rw(iter) (((iter)->type) & ~(~0U>>1) ? WRITE : READ)
>>> with places like iov_iter_kvec() doing
>>> i->type = ITER_KVEC | ((direction == WRITE) ? BIT(31) : 0);
>>>
>>> Preferences?
>>
>> For the bitmask version (with this patch) we have most of
>> iov_iter_type() completely optimised out. E.g. identical
>>
>> iov_iter_type(i) & ITER_IOVEC <=> iter->type & ITER_IOVEC
>>
>> It's also nice to have iov_iter_rw() to be just
>> (type & 1), operations with which can be optimised in a handful of ways.
>>
>> Unless the compiler would be able to heavily optimise switches,
>> e.g. to out-of-memory/calculation-based jump tables, that I doubt,
>> I'd personally leave it be. Though, not like it should matter much.
>
> The advantage of the bit-masks is that the 'usual' options can
> be tested for together. So the code can be (for example):
Well, you can do that for the non-bitwise case as well.
In a simpler form but should be enough.
enum { ITER_IOVEC = 1, ITER_BVEC = 2, ... }
if (type <= ITER_BVEC) {
if (iovec) ...
if (bvec) ...
} else { ... }
> if (likely(iter->type & (ITER_IOVEC | ITER_PIPE) {
> if (likely((iter->type & ITER_IOVEC)) {
> ... code for iovec
> } else [
> ... code for pipe
> }
> } else if (iter->type & ITER_BVEC) {
> ... code for bvec
> } else if (iter->type & ITER_KVEC) {
> .. code for kvec
> } else {
> .. must be discard
> }
--
Pavel Begunkov
next prev parent reply other threads:[~2020-12-14 17:38 UTC|newest]
Thread overview: 13+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-11-19 15:29 [PATCH 0/2] optimise iov_iter Pavel Begunkov
2020-11-19 15:29 ` [PATCH 1/2] iov_iter: optimise iov_iter_npages for bvec Pavel Begunkov
2020-11-19 15:29 ` [PATCH 2/2] iov_iter: optimise iter type checking Pavel Begunkov
2020-11-19 17:03 ` Christoph Hellwig
2020-11-19 17:12 ` Pavel Begunkov
2020-12-11 2:01 ` Al Viro
2020-12-13 22:32 ` Pavel Begunkov
2020-12-14 10:28 ` David Laight
2020-12-14 17:25 ` Pavel Begunkov [this message]
2020-11-19 16:46 ` [PATCH 0/2] optimise iov_iter Jens Axboe
2020-11-19 17:14 ` Pavel Begunkov
2020-11-19 17:20 ` Jens Axboe
2020-11-19 18:02 ` Pavel Begunkov
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=5b65de70-19db-4572-d122-df65191ab098@gmail.com \
--to=asml.silence@gmail.com \
--cc=David.Laight@ACULAB.COM \
--cc=axboe@kernel.dk \
--cc=hch@infradead.org \
--cc=linux-block@vger.kernel.org \
--cc=linux-kernel@vger.kernel.org \
--cc=viro@zeniv.linux.org.uk \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).