[PATCHSET,v5,0/9] Turn single segment imports into ITER_UBUF

Message ID	20230328215811.903557-1-axboe@kernel.dk (mailing list archive)
Headers	show Return-Path: <linux-fsdevel-owner@vger.kernel.org> From: Jens Axboe <axboe@kernel.dk> To: linux-fsdevel@vger.kernel.org Cc: torvalds@linux-foundation.org, brauner@kernel.org, viro@zeniv.linux.org.uk Subject: [PATCHSET v5 0/9] Turn single segment imports into ITER_UBUF Date: Tue, 28 Mar 2023 15:58:02 -0600 Message-Id: <20230328215811.903557-1-axboe@kernel.dk> MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Precedence: bulk
Series	Turn single segment imports into ITER_UBUF \| expand [PATCHSET,v5,0/9] Turn single segment imports into ITER_UBUF [1/9] block: ensure bio_alloc_map_data() deals with ITER_UBUF correctly [2/9] iov_iter: teach iov_iter_iovec() to deal with ITER_UBUF [3/9] iov_iter: overlay struct iovec and ubuf/len [4/9] iov_iter: set nr_segs = 1 for ITER_UBUF [5/9] IB/hfi1: check for user backed iterator, not specific iterator type [6/9] IB/qib: check for user backed iterator, not specific iterator type [7/9] ALSA: pcm: check for user backed iterator, not specific iterator type [8/9] iov_iter: convert import_single_range() to ITER_UBUF [9/9] iov_iter: import single vector iovecs as ITER_UBUF

Message ID

20230328215811.903557-1-axboe@kernel.dk (mailing list archive)

Headers

From: Jens Axboe <axboe@kernel.dk>
To: linux-fsdevel@vger.kernel.org
Cc: torvalds@linux-foundation.org, brauner@kernel.org,
        viro@zeniv.linux.org.uk
Subject: [PATCHSET v5 0/9] Turn single segment imports into ITER_UBUF
Date: Tue, 28 Mar 2023 15:58:02 -0600
Message-Id: <20230328215811.903557-1-axboe@kernel.dk>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

Series

Turn single segment imports into ITER_UBUF | expand

Message

Jens Axboe March 28, 2023, 9:58 p.m. UTC

Hi,

ather than repeat the same blurb again, see the v2 posting here:

https://lore.kernel.org/linux-fsdevel/20230327180449.87382-1-axboe@kernel.dk/

tldr - turn single segment iovecs into ITER_UBUF rather than ITER_IOVEC,
because they are more efficient.

The main change since v4 is doing the overlay trick in the iov_iter,
so that the first segment of the iovec overlaps nicely with the ubuf
address and length. With that, we can drop any functional driver change,
as the normal iteration should just work out with ubuf having the single
segment. Replace he iter_is_iovec() check with a basic iter->user_backed
instead.

Unfortunately we can't make iov_iter_iovec() much prettier, as we have
to return a copy of the current iovec as the offset has to be added...

Include a block fix for a shoddy copy of an iterator, as that'll
otherwise cause issues.

Ran the liburing test cases and the previously indicated ltp cases
that barfed, and everything looks rosy. Comments welcome - with the
above approach, we don't ever return iter->__ubuf_iovec, it's really
just there as an overlay filler. Maybe we can make that a bit more
apparent.