From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Date: Wed, 18 Nov 2015 09:16:08 -0700 From: Ross Zwisler To: Jan Kara Cc: Ross Zwisler , Dan Williams , Andreas Dilger , "linux-kernel@vger.kernel.org" , "H. Peter Anvin" , "J. Bruce Fields" , Theodore Ts'o , Alexander Viro , Dave Chinner , Ingo Molnar , Jan Kara , Jeff Layton , Matthew Wilcox , Thomas Gleixner , linux-ext4 , linux-fsdevel , Linux MM , "linux-nvdimm@lists.01.org" , X86 ML , XFS Developers , Andrew Morton , Matthew Wilcox , Dave Hansen Subject: Re: [PATCH v2 03/11] pmem: enable REQ_FUA/REQ_FLUSH handling Message-ID: <20151118161608.GA10656@linux.intel.com> References: <1447459610-14259-1-git-send-email-ross.zwisler@linux.intel.com> <1447459610-14259-4-git-send-email-ross.zwisler@linux.intel.com> <22E0F870-C1FB-431E-BF6C-B395A09A2B0D@dilger.ca> <20151116200950.GB9737@linux.intel.com> <20151118104055.GA6097@quack.suse.cz> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20151118104055.GA6097@quack.suse.cz> Sender: linux-kernel-owner@vger.kernel.org List-ID: On Wed, Nov 18, 2015 at 11:40:55AM +0100, Jan Kara wrote: > On Mon 16-11-15 13:09:50, Ross Zwisler wrote: > > On Fri, Nov 13, 2015 at 06:32:40PM -0800, Dan Williams wrote: > > > On Fri, Nov 13, 2015 at 4:43 PM, Andreas Dilger wrote: > > > > On Nov 13, 2015, at 5:20 PM, Dan Williams wrote: > > > >> > > > >> On Fri, Nov 13, 2015 at 4:06 PM, Ross Zwisler > > > >> 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 > > > >> > > > >> 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. > > > > Does anyone know if ext4 and/or XFS alter their algorithms based on whether > > the driver supports REQ_FLUSH/REQ_FUA? Will the filesystem behave more > > efficiently with respect to their internal I/O ordering, etc., if PMEM > > advertises REQ_FLUSH/REQ_FUA support, even though we could do the same thing > > at the DAX layer? > > So the information whether the driver supports FLUSH / FUA is generally > ignored by filesystems. We issue REQ_FLUSH / REQ_FUA requests to achieve > required ordering for fs consistency and expect that block layer does the > right thing - i.e., if the device has volatile write cache, it will be > flushed, if it doesn't have it, the request will be ignored. So the > difference between supporting and not supporting REQ_FLUSH / REQ_FUA is > only in how block layer handles such requests. Cool, thank you for the info. Based on this I'll pull out the REQ_FLUSH/REQ_FUA patch for v3 of this series and move the wmb_pmem() call up to DAX as Dan suggests. If performance data shows that we can get a benefit from centralizing wmb_pmem() behind REQ_FUA/REQ_FLUSH, I'll add it back in later as part of that series.