Re: [PATCH 00/19] xfs: refactor log recovery

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Wed, Apr 29, 2020 at 07:52:05AM -0400, Brian Foster wrote:
> On Tue, Apr 28, 2020 at 03:34:22PM -0700, Darrick J. Wong wrote:
> > On Tue, Apr 28, 2020 at 08:43:42AM -0400, Brian Foster wrote:
> > > On Mon, Apr 27, 2020 at 11:12:08PM -0700, Christoph Hellwig wrote:
> > > > On Wed, Apr 22, 2020 at 12:18:54PM -0400, Brian Foster wrote:
> > > > > - Transaction reorder
> > > > > 
> > > > > Virtualizing the transaction reorder across all several files/types
> > > > > strikes me as overkill for several reasons. From a code standpoint,
> > > > > we've created a new type enumeration and a couple fields (enum type and
> > > > > a function) in a generic structure to essentially abstract out the
> > > > > buffer handling into a function. The latter checks another couple of blf
> > > > > flags, which appears to be the only real "type specific" logic in the
> > > > > whole sequence. From a complexity standpoint, the reorder operation is a
> > > > > fairly low level and internal recovery operation. We have this huge
> > > > > comment just to explain exactly what's happening and why certain items
> > > > > have to be ordered as such, or some treated like others, etc. TBH it's
> > > > > not terribly clear even with that documentation, so I don't know that
> > > > > splitting the associated mapping logic off into separate files is
> > > > > helpful.
> > > > 
> > > > I actually very much like idea of moving any knowledge of the individual
> > > > item types out of xfs_log_recovery.c.  In reply to the patch I've
> > > > suggsted an idea how to kill the knowledge for all but the buffer and
> > > > icreate items, which should make this a little more sensible.
> > > > 
> > > 
> > > I mentioned to Darrick the other day briefly on IRC that I don't
> > > fundamentally object to splitting up xfs_log_recover.c. I just think
> > > this mechanical split out of the existing code includes too much of the
> > > implementation details of recovery and perhaps abstracts a bit too much.
> > > I find the general idea much more acceptable with preliminary cleanups
> > > and a more simple interface.
> > 
> > It's cleaned up considerably with hch's cleanup patches 1-5 of 2. ;)
> > 
> > > > I actually think we should go further in one aspect - instead of having
> > > > the item type to ops mapping in a single function in xfs_log_recovery.c
> > > > we should have a table that the items can just add themselves to.
> > > > 
> > > 
> > > That sounds reasonable, but that's more about abstraction mechanism than
> > > defining the interface. I was more focused on simplifying the latter in
> > > my previous comments.
> > 
> > <nod>
> > 
> > > > > - Readahead
> > > > > 
> > > > > We end up with readahead callouts for only the types that translate to
> > > > > buffers (so buffers, inode, dquots), and then those callouts do some
> > > > > type specific mapping (that is duplicated within the specific type
> > > > > handers) and issue a readahead (which is duplicated across each ra_pass2
> > > > > call). I wonder if this would be better abstracted by a ->bmap() like
> > > > > call that simply maps the item to a [block,length] and returns a
> > > > > non-zero length if the core recovery code should invoke readahead (after
> > > > > checking for cancellation). It looks like the underlying implementation
> > > > > of those bmap calls could be further factored into helpers that
> > > > > translate from the raw record data into the type specific format
> > > > > structures, and that could reduce duplication between the readahead
> > > > > calls and the pass2 calls in a couple cases. (The more I think about,
> > > > > the more I think we should introduce those kind of cleanups before
> > > > > getting into the need for function pointers.)
> > > > 
> > > > That sounds more complicated what we have right now, and even more so
> > > > with my little xlog_buf_readahead helper.  Yes, the methods will all
> > > > just call xlog_buf_readahead, but they are trivial two-liners that are
> > > > easy to understand.  Much easier than a complicated calling convention
> > > > to pass the blkno, len and buf ops back.
> > > > 
> > > 
> > > Ok. The above was just an idea to simplify things vs. duplicating
> > > readahead code and recovery logic N times. I haven't seen your
> > > idea/code, but if that problem is addressed with a helper vs. a
> > > different interface then that seems just as reasonable to me.
> > > 
> > > > > - Recovery (pass1/pass2)
> > > > > 
> > > > > The core recovery bits seem more reasonable to factor out in general.
> > > > > That said, we only have two pass1 callbacks (buffers and quotaoff). The
> > > > > buffer callback does cancellation management and the quotaoff sets some
> > > > > flags, so I wonder why those couldn't just remain as direct function
> > > > > calls (even if we move the functions out of xfs_log_recover.c). There
> > > > > are more callbacks for pass2 so the function pointers make a bit more
> > > > > sense there, but at the same time it looks like the various intents are
> > > > > further abstracted behind a single "intent type" pass2 call (which has a
> > > > > hardcoded XLOG_REORDER_INODE_LIST reorder value and is about as clear as
> > > > > mud in that context, getting to my earlier point).
> > > > 
> > > > Again I actually like the callouts, mostly because they make it pretty
> > > > clear what is going on.  I also really like the fact that the recovery
> > > > code is close to the code actually writing the log items.
> > 
> > Looking back at that, I realize that (provided nobody minds having
> > function dispatch structures that are sort of sparse) there's no reason
> > why we need to have separate xlog_recover_intent_type and
> > xlog_recover_item_type structures.
> > 
> 
> The sparseness doesn't bother me provided the underlying interfaces are
> simple/generic enough to understand from the structure definition and
> are broadly (if not universally) used. I'm still not convinced the
> transaction reorder thing should be distributed at all, but I'll wait to
> see what the next iteration looks like.

There's a lot less of it, since we assume ITEM_LIST unless explicitly
overridden via function.

> > > I find both the runtime logging and recovery code to be complex enough
> > > individually that I prefer not to stuff them together, but there is
> > > already precedent with dfops and such so that's not the biggest deal to
> > > me if the interface is simplified (and hopefully amount of code
> > > reduced).
> > 
> > I combined them largely on the observation that with the exception of
> > buffers, log item recovery code is generally short and not worth
> > creating even more files.  224 is enough.
> > 
> 
> I like Christoph's idea of selectively separating out the particularly
> large (i.e. buffers) bits.

Ok, I'll start with the xfs_buf_item.c.

--D

> Brian
> 
> > --D
> > 
> > > Brian
> > > 
> > 
>