From: Dave Chinner <david@fromorbit.com>
To: Goldwyn Rodrigues <rgoldwyn@suse.de>
Cc: linux-btrfs@vger.kernel.org, linux-fsdevel@vger.kernel.org
Subject: Re: [PATCH 04/15] dax: Introduce IOMAP_F_COW for copy-on-write
Date: Tue, 2 Apr 2019 10:06:53 +1100 [thread overview]
Message-ID: <20190401230653.GT23020@dastard> (raw)
In-Reply-To: <20190401214102.bn4giybeqpbvdbb3@merlin>
On Mon, Apr 01, 2019 at 04:41:02PM -0500, Goldwyn Rodrigues wrote:
> On 15:38 01/04, Dave Chinner wrote:
> > On Tue, Mar 26, 2019 at 02:02:50PM -0500, Goldwyn Rodrigues wrote:
> > > From: Goldwyn Rodrigues <rgoldwyn@suse.com>
> > >
> > > The IOMAP_F_COW is a flag to notify dax that it needs to copy
> > > the data from iomap->cow_addr to iomap->addr, if the start/end
> > > of I/O are not page aligned.
> >
> > I see what you are trying to do here, but this is kinda gross.
> >
> > > This also introduces dax_to_dax_copy() which performs a copy
> > > from one part of the device to another, to a maximum of one page.
> > >
> > > Question: Using iomap.cow_addr == 0 means the CoW is to be copied
> > > (or memset) from a hole. Would this be better handled through a flag?
> >
> > That's what all these checks in the iomap code do:
> >
>
> This is using iomap->flags not type.
Yes, I know. The fact that you tell me this (when it was obvious)
indicates to me that you didn't understand what I was saying.
i.e. the gross hack here is that this patch is trying to define a
new iomap type - both behaviourally and iomap content - via adding
a modifier flag rather than defining a new iomap->type. That's the
gross hack, and everything stems from that.
i.e. the "bloating" of the struct iomap is caused because the flag
modifier (IOMAP_F_COW) can't use parts of the iomap that are defined
for specific iomap types. e.g. IOMAP_INLINE type uses ->inline_data,
and so it can't be re-used by a iomap flag modifier such as
IOMAP_F_COW.
However, if we define a new type for this "need multiple mappings"
iomap rather than a flag, we don't need any new fields in the struct
iomap because we can use what already exists in the iomap.
> > if (iomap->type == IOMAP_HOLE || iomap->type == IOMAP_UNWRITTEN)
> >
> > Oh, wait, you're trying to map two ranges in a single iomap and then
> > infer state from information that got chucked away.... IOWs, you're
> > doing it wrong - iomap algorithms are driven by how we manipulate
> > iomaps to do data operations efficiently, not how we copy data page
> > by page.
> >
> > IOWs, what we really should have here is two iomaps - a source
> > and a destination iomap. The source is a read mapping of the
> > current address (where we are going to copy the data from), the
> > destination is the post-cow allocation mapping (where the data
> > goes).
> >
> > Now you just copy the data from one map to the other iterating
> > source mappings until the necessary range of the destination has
> > been covered. And you can check if the source map is IOMAP_HOLE or
> > IOMAP_UNWRITTEN and hence optimise the copy (i.e. zero the new
> > allocation) before copying in the new data.
>
> Won't that be inefficient? With CoW we only need to write the first
> and last block.
You're assuming that partial data overwrites are the only case where
this dax-to-dax copy of a file range is required. That assumption is
false.
i.e. FALLOC_FL_UNSHARE_RANGE for DAX requires iterating over the
entire source range and copying all the contents to the newly
allocated destination range. The partial block copying is just a
short version of this limited to a single block.
Sure, btrfs doesn't support FALLOC_FL_UNSHARE_RANGE, but if you're
going to be adding support for reflink to DAX, the infrastructure
needs to provide support for performing FALLOC_FL_UNSHARE_RANGE
to break extent sharing efficiently.
> Again, that is not required if the offset/end
> offset is block aligned. After that, it falls back to the
> regular write path of performing dax_copy_to_iter().
> We don't deal with IOMAP_UNWRITTEN in dax,
Yes we do. fallocate() can lay down unwritten extents, and we can
both read and write to them. See, for example, dax_iomap_actor()
called from dax_iomap_rw() via iomap_apply() - it does not call
dax_copy_to_iter() for reads if the range is IOMAP_HOLE or
IOMAP_UNWRITTEN.
> though other
> filesystems in the future may use CoW without dax.
That makes no sense. :/
> Besides, what you are suggesting will not fit well with the
> current iomap iterator code and would require another function
> altogether.
I'm not concerned about that - I would much prefer we do things
cleanly and properly rather than make expedient hacks for
questionable benefit that we'll have to completely rework or remove
the moment we implement DAX+reflink support in XFS.
> After Darrick's suggestion, we can even do away with cow_pos, so
> only the read address of cow_addr will exist.
As I mentioned earlier, even that is not necessary.
This is DAX - the iomap API and mapping functions can already return
pointers to inline data, and DAX can effectively be considered
inline data for the purposes of reading data.
As I said, the problem here is you are trying to use flags to define
a new type of iomap operation requires two mappings rather than one.
IMO, we should be defining an IOMAP_DAX_COW /type/ and then define
it to contain and behave as follows:
- new destination region for data to be copied into is the
same setup as IOMAP_MAPPED
- existing shared data that may be needed for reading is
mapped direct to device address by ->iomap_begin into
iomap->inline_data
- if the iomap infrastructure needs to copy original source
data into destination, it copies directly from the memory
address in iomap->inline_data into the directly mapped DAX
desitination via memcpy().
This covers both the partial write COW case you are concerned with
here, and the full-extent range copy case that
FALLOC_FL_UNSHARE_RANGE requires.
Cheers,
Dave.
--
Dave Chinner
david@fromorbit.com
next prev parent reply other threads:[~2019-04-01 23:07 UTC|newest]
Thread overview: 48+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <20190326190301.32365-1-rgoldwyn@suse.de>
2019-03-26 19:02 ` [PATCH 01/15] btrfs: create a mount option for dax Goldwyn Rodrigues
2019-03-26 19:10 ` Matthew Wilcox
2019-03-27 11:00 ` Goldwyn Rodrigues
2019-03-27 12:00 ` Matthew Wilcox
2019-03-27 12:26 ` Goldwyn Rodrigues
2019-03-27 23:31 ` Goldwyn Rodrigues
2019-03-27 17:38 ` Adam Borowski
2019-03-28 14:49 ` David Sterba
2019-03-28 17:28 ` David Sterba
2019-03-28 17:57 ` Darrick J. Wong
2019-04-01 20:43 ` Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 02/15] btrfs: Carve out btrfs_get_extent_map_write() out of btrfs_get_blocks_write() Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 03/15] btrfs: basic dax read Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 04/15] dax: Introduce IOMAP_F_COW for copy-on-write Goldwyn Rodrigues
2019-03-27 17:54 ` Darrick J. Wong
2019-03-27 18:58 ` Goldwyn Rodrigues
2019-03-28 14:45 ` Darrick J. Wong
2019-04-01 4:38 ` Dave Chinner
2019-04-01 21:41 ` Goldwyn Rodrigues
2019-04-01 23:06 ` Dave Chinner [this message]
2019-04-03 1:56 ` Goldwyn Rodrigues
2019-04-03 3:20 ` Dave Chinner
2019-04-07 7:26 ` Christoph Hellwig
2019-03-26 19:02 ` [PATCH 05/15] btrfs: return whether extent is nocow or not Goldwyn Rodrigues
2019-03-31 18:42 ` Nikolay Borisov
2019-03-26 19:02 ` [PATCH 06/15] btrfs: Rename __endio_write_update_ordered() to btrfs_update_ordered_extent() Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 07/15] btrfs: add dax write support Goldwyn Rodrigues
2019-03-28 14:53 ` Darrick J. Wong
2019-04-01 20:39 ` Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 08/15] dax: add dax_iomap_cow to copy a mmap page before writing Goldwyn Rodrigues
2019-03-28 15:41 ` Darrick J. Wong
2019-03-26 19:02 ` [PATCH 09/15] btrfs: add dax mmap support Goldwyn Rodrigues
2019-03-28 15:45 ` Darrick J. Wong
2019-03-26 19:02 ` [PATCH 10/15] btrfs: Add dax specific address_space_operations Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 11/15] fs: dedup file range to use a compare function Goldwyn Rodrigues
2019-03-28 17:04 ` Darrick J. Wong
2019-04-01 20:36 ` Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 12/15] btrfs: trace functions for btrfs_iomap_begin/end Goldwyn Rodrigues
2019-03-26 19:02 ` [PATCH 13/15] btrfs: handle dax page zeroing Goldwyn Rodrigues
2019-03-26 19:03 ` [PATCH 14/15] btrfs: Disable dax-based defrag and send Goldwyn Rodrigues
2019-03-26 19:03 ` [PATCH 15/15] btrfs: Writeprotect mmap pages on snapshot Goldwyn Rodrigues
2019-03-28 15:48 ` Darrick J. Wong
2019-03-26 19:09 ` [PATCH v2 00/15] btrfs dax support Goldwyn Rodrigues
2019-03-27 20:14 ` Adam Borowski
2019-03-27 23:26 ` Goldwyn Rodrigues
2019-03-28 10:24 ` [PATCH] btrfs: allow MAP_SYNC mmap Adam Borowski
2019-03-28 10:42 ` Adam Borowski
2019-04-01 20:08 ` Goldwyn Rodrigues
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=20190401230653.GT23020@dastard \
--to=david@fromorbit.com \
--cc=linux-btrfs@vger.kernel.org \
--cc=linux-fsdevel@vger.kernel.org \
--cc=rgoldwyn@suse.de \
/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).