Re: [dm-devel] [PATCH 0/6] dax poison recovery with RWF_RECOVERY_DATA flag

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, Oct 29, 2021 at 12:46:14PM +0100, Pavel Begunkov wrote:
> On 10/28/21 23:59, Dave Chinner wrote:
> [...]
> > > > Well, my point is doing recovery from bit errors is by definition not
> > > > the fast path.  Which is why I'd rather keep it away from the pmem
> > > > read/write fast path, which also happens to be the (much more important)
> > > > non-pmem read/write path.
> > > 
> > > The trouble is, we really /do/ want to be able to (re)write the failed
> > > area, and we probably want to try to read whatever we can.  Those are
> > > reads and writes, not {pre,f}allocation activities.  This is where Dave
> > > and I arrived at a month ago.
> > > 
> > > Unless you'd be ok with a second IO path for recovery where we're
> > > allowed to be slow?  That would probably have the same user interface
> > > flag, just a different path into the pmem driver.
> > 
> > I just don't see how 4 single line branches to propage RWF_RECOVERY
> > down to the hardware is in any way an imposition on the fast path.
> > It's no different for passing RWF_HIPRI down to the hardware *in the
> > fast path* so that the IO runs the hardware in polling mode because
> > it's faster for some hardware.
> 
> Not particularly about this flag, but it is expensive. Surely looks
> cheap when it's just one feature, but there are dozens of them with
> limited applicability, default config kernels are already sluggish
> when it comes to really fast devices and it's not getting better.
> Also, pretty often every of them will add a bunch of extra checks
> to fix something of whatever it would be.

So we can't have data recovery because moving fast the only goal?

That's so meta.

--D

> So let's add a bit of pragmatism to the picture, if there is just one
> user of a feature but it adds overhead for millions of machines that
> won't ever use it, it's expensive.
> 
> This one doesn't spill yet into paths I care about, but in general
> it'd be great if we start thinking more about such stuff instead of
> throwing yet another if into the path, e.g. by shifting the overhead
> from linear to a constant for cases that don't use it, for instance
> with callbacks or bit masks.
> 
> > IOWs, saying that we shouldn't implement RWF_RECOVERY because it
> > adds a handful of branches 	 the fast path is like saying that we
> > shouldn't implement RWF_HIPRI because it slows down the fast path
> > for non-polled IO....
> > 
> > Just factor the actual recovery operations out into a separate
> > function like:
> 
> -- 
> Pavel Begunkov



[Index of Archives]     [Linux Ext4 Filesystem]     [Union Filesystem]     [Filesystem Testing]     [Ceph Users]     [Ecryptfs]     [NTFS 3]     [AutoFS]     [Kernel Newbies]     [Share Photos]     [Security]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux Cachefs]     [Reiser Filesystem]     [Linux RAID]     [NTFS 3]     [Samba]     [Device Mapper]     [CEPH Development]

  Powered by Linux