linux-kernel - Re: [PATCH v5 03/10] xfs: Refactor xfs_reflink_end_cow

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20250312232210.GD2803730@frogsfrogsfrogs>
Date: Wed, 12 Mar 2025 16:22:10 -0700
From: "Darrick J. Wong" <djwong@...nel.org>
To: John Garry <john.g.garry@...cle.com>
Cc: Christoph Hellwig <hch@...radead.org>, brauner@...nel.org,
	cem@...nel.org, linux-xfs@...r.kernel.org,
	linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org,
	ojaswin@...ux.ibm.com, ritesh.list@...il.com,
	martin.petersen@...cle.com
Subject: Re: [PATCH v5 03/10] xfs: Refactor xfs_reflink_end_cow_extent()

On Wed, Mar 12, 2025 at 10:06:11PM +0000, John Garry wrote:
> On 12/03/2025 15:46, Darrick J. Wong wrote:
> > On Wed, Mar 12, 2025 at 01:35:23AM -0700, Christoph Hellwig wrote:
> > > On Wed, Mar 12, 2025 at 08:27:05AM +0000, John Garry wrote:
> > > > On 12/03/2025 07:24, Christoph Hellwig wrote:
> > > > > On Mon, Mar 10, 2025 at 06:39:39PM +0000, John Garry wrote:
> > > > > > Refactor xfs_reflink_end_cow_extent() into separate parts which process
> > > > > > the CoW range and commit the transaction.
> > > > > > 
> > > > > > This refactoring will be used in future for when it is required to commit
> > > > > > a range of extents as a single transaction, similar to how it was done
> > > > > > pre-commit d6f215f359637.
> > > > > 
> > > > > Darrick pointed out that if you do more than just a tiny number
> > > > > of extents per transactions you run out of log reservations very
> > > > > quickly here:
> > > > > 
> > > > > https://urldefense.com/v3/__https://lore.kernel.org/all/20240329162936.GI6390@frogsfrogsfrogs/__;!!ACWV5N9M2RV99hQ!PWLcBof1tKimKUObvCj4vOhljWjFmjtzVHLx9apcU5Rah1xZnmp_3PIq6eSwx6TdEXzMLYYyBfmZLgvj$
> > > > > 
> > > > > how does your scheme deal with that?
> > > > > 
> > > > The resblks calculation in xfs_reflink_end_atomic_cow() takes care of this,
> > > > right? Or does the log reservation have a hard size limit, regardless of
> > > > that calculation?
> > > 
> > > The resblks calculated there are the reserved disk blocks and have
> > > nothing to do with the log reservations, which comes from the
> > > tr_write field passed in.  There is some kind of upper limited to it
> > > obviously by the log size, although I'm not sure if we've formalized
> > > that somewhere.  Dave might be the right person to ask about that.
> > 
> > The (very very rough) upper limit for how many intent items you can
> > attach to a tr_write transaction is:
> > 
> > per_extent_cost = (cui_size + rui_size + bui_size + efi_size + ili_size)
> > max_blocks = tr_write::tr_logres / per_extent_cost
> > 
> > (ili_size is the inode log item size)
> 
> So will it be something like this:
> 
> static size_t
> xfs_compute_awu_max_extents(
> 	struct xfs_mount	*mp)
> {
> 	struct xfs_trans_res	*resp = &M_RES(mp)->tr_write;
> 	size_t			logtotal = xfs_bui_log_format_sizeof(1)+

Might want to call it "per_extent_logres" since that's what it is.

> 				xfs_cui_log_format_sizeof(1) +
> 				xfs_efi_log_format_sizeof(1) +
> 				xfs_rui_log_format_sizeof(1) +
> 				sizeof(struct xfs_inode_log_format);

Something like that, yeah.  You should probably add
xfs_log_dinode_size(ip->i_mount) to that.

What you're really doing is summing the *nbytes output of the
->iop_size() call for each possible log item.  For the four log intent
items it's the xfs_FOO_log_format_sizeof() function like you have above.
For inode items it's:

	*nbytes += sizeof(struct xfs_inode_log_format) +
		   xfs_log_dinode_size(ip->i_mount);

> 	return rounddown_pow_of_two(resp->tr_logres / logtotal);

and like I said earlier, you should double logtotal to be on the safe
side with a 2x safety margin:

	/* 100% safety margin for safety's sake */
	return rounddown_pow_of_two(resp->tr_logres /
				    (2 * per_extent_logres));

I'm curious what number you get back from this function?  Hopefully it's
at least a few hundred blocks.

Thanks for putting that together.  :)

--D

> }
> 
> static inline void
> xfs_compute_awu_max(
> 	struct xfs_mount	*mp, int jjcount)
> {
> ....
> 	mp->m_awu_max =
> 	min_t(unsigned int, awu_max, xfs_compute_awu_max_extents(mp));
> }
> 
> > 
> > ((I would halve that for the sake of paranoia))
> > 
> > since you have to commit all those intent items into the first
> > transaction in the chain.  The difficulty we've always had is computing
> > the size of an intent item in the ondisk log, since that's a (somewhat
> > minor) layering violation -- it's xfs_cui_log_format_sizeof() for a CUI,
> > but then there' could be overhead for the ondisk log headers themselves.
> > 
> > Maybe we ought to formalize the computation of that since reap.c also
> > has a handwavy XREAP_MAX_DEFER_CHAIN that it uses to roll the scrub
> > transaction periodically... because I'd prefer we not add another
> > hardcoded limit.  My guess is that the software fallback can probably
> > support any awu_max that a hardware wants to throw at us, but let's
> > actually figure out the min(sw, hw) that we can support and cap it at
> > that.
> > 
> > --D
> 
>