Message-ID: <20151116232927.GA5582@linux.intel.com>
Date:	Mon, 16 Nov 2015 16:29:27 -0700
From:	Ross Zwisler <ross.zwisler@...ux.intel.com>
To:	Dave Chinner <david@...morbit.com>
Cc:	Jan Kara <jack@...e.cz>, Dan Williams <dan.j.williams@...el.com>,
	Andreas Dilger <adilger@...ger.ca>,
	Ross Zwisler <ross.zwisler@...ux.intel.com>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"H. Peter Anvin" <hpa@...or.com>,
	"J. Bruce Fields" <bfields@...ldses.org>,
	Theodore Ts'o <tytso@....edu>,
	Alexander Viro <viro@...iv.linux.org.uk>,
	Ingo Molnar <mingo@...hat.com>, Jan Kara <jack@...e.com>,
	Jeff Layton <jlayton@...chiereds.net>,
	Matthew Wilcox <willy@...ux.intel.com>,
	Thomas Gleixner <tglx@...utronix.de>,
	linux-ext4 <linux-ext4@...r.kernel.org>,
	linux-fsdevel <linux-fsdevel@...r.kernel.org>,
	Linux MM <linux-mm@...ck.org>,
	"linux-nvdimm@...ts.01.org" <linux-nvdimm@...ts.01.org>,
	X86 ML <x86@...nel.org>, XFS Developers <xfs@....sgi.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Matthew Wilcox <matthew.r.wilcox@...el.com>,
	Dave Hansen <dave.hansen@...ux.intel.com>
Subject: Re: [PATCH v2 03/11] pmem: enable REQ_FUA/REQ_FLUSH handling

On Tue, Nov 17, 2015 at 09:14:12AM +1100, Dave Chinner wrote:
> On Mon, Nov 16, 2015 at 03:05:26PM +0100, Jan Kara wrote:
> > On Mon 16-11-15 14:37:14, Jan Kara wrote:
> > > On Fri 13-11-15 18:32:40, Dan Williams wrote:
> > > > On Fri, Nov 13, 2015 at 4:43 PM, Andreas Dilger <adilger@...ger.ca> wrote:
> > > > > On Nov 13, 2015, at 5:20 PM, Dan Williams <dan.j.williams@...el.com> wrote:
> > > > >>
> > > > >> On Fri, Nov 13, 2015 at 4:06 PM, Ross Zwisler
> > > > >> <ross.zwisler@...ux.intel.com> wrote:
> > > > >>> Currently the PMEM driver doesn't accept REQ_FLUSH or REQ_FUA bios.  These
> > > > >>> are sent down via blkdev_issue_flush() in response to a fsync() or msync()
> > > > >>> and are used by filesystems to order their metadata, among other things.
> > > > >>>
> > > > >>> When we get an msync() or fsync() it is the responsibility of the DAX code
> > > > >>> to flush all dirty pages to media.  The PMEM driver then just has issue a
> > > > >>> wmb_pmem() in response to the REQ_FLUSH to ensure that before we return all
> > > > >>> the flushed data has been durably stored on the media.
> > > > >>>
> > > > >>> Signed-off-by: Ross Zwisler <ross.zwisler@...ux.intel.com>
> > > > >>
> > > > >> Hmm, I'm not seeing why we need this patch.  If the actual flushing of
> > > > >> the cache is done by the core why does the driver need support
> > > > >> REQ_FLUSH?  Especially since it's just a couple instructions.  REQ_FUA
> > > > >> only makes sense if individual writes can bypass the "drive" cache,
> > > > >> but no I/O submitted to the driver proper is ever cached we always
> > > > >> flush it through to media.
> > > > >
> > > > > If the upper level filesystem gets an error when submitting a flush
> > > > > request, then it assumes the underlying hardware is broken and cannot
> > > > > be as aggressive in IO submission, but instead has to wait for in-flight
> > > > > IO to complete.
> > > > 
> > > > Upper level filesystems won't get errors when the driver does not
> > > > support flush.  Those requests are ended cleanly in
> > > > generic_make_request_checks().  Yes, the fs still needs to wait for
> > > > outstanding I/O to complete but in the case of pmem all I/O is
> > > > synchronous.  There's never anything to await when flushing at the
> > > > pmem driver level.
> > > > 
> > > > > Since FUA/FLUSH is basically a no-op for pmem devices,
> > > > > it doesn't make sense _not_ to support this functionality.
> > > > 
> > > > Seems to be a nop either way.  Given that DAX may lead to dirty data
> > > > pending to the device in the cpu cache that a REQ_FLUSH request will
> > > > not touch, its better to leave it all to the mm core to handle.  I.e.
> > > > it doesn't make sense to call the driver just for two instructions
> > > > (sfence + pcommit) when the mm core is taking on the cache flushing.
> > > > Either handle it all in the mm or the driver, not a mixture.
> > > 
> > > So I think REQ_FLUSH requests *must* end up doing sfence + pcommit because
> > > e.g. journal writes going through block layer or writes done through
> > > dax_do_io() must be on permanent storage once REQ_FLUSH request finishes
> > > and the way driver does IO doesn't guarantee this, does it?
> > 
> > Hum, and looking into how dax_do_io() works and what drivers/nvdimm/pmem.c
> > does, I'm indeed wrong because they both do wmb_pmem() after each write
> > which seems to include sfence + pcommit. Sorry for confusion.
> 
> Which I want to remove, because it makes DAX IO 3x slower than
> buffered IO on ramdisk based testing.
> 
> > But a question: Won't it be better to do sfence + pcommit only in response
> > to REQ_FLUSH request and don't do it after each write? I'm not sure how
> > expensive these instructions are but in theory it could be a performance
> > win, couldn't it? For filesystems this is enough wrt persistency
> > guarantees...
> 
> I'm pretty sure it would be, because all of the overhead (and
> therefore latency) I measured is in the cache flushing instructions.
> But before we can remove the wmb_pmem() from  dax_do_io(), we need
> the underlying device to support REQ_FLUSH correctly...

By "support REQ_FLUSH correctly" do you mean call wmb_pmem() as I do in my
set?  Or do you mean something that also involves cache flushing such as the
"big hammer" that flushes everything or something like WBINVD?
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists