Re: [RFC 00/11] DAX fsynx/msync support

From: Dan Williams
Date: Fri Oct 30 2015 - 15:51:46 EST

On Fri, Oct 30, 2015 at 12:43 PM, Ross Zwisler
<ross.zwisler@xxxxxxxxxxxxxxx> wrote:
> On Fri, Oct 30, 2015 at 11:34:07AM -0700, Dan Williams wrote:
>> On Thu, Oct 29, 2015 at 1:12 PM, Ross Zwisler
>> <ross.zwisler@xxxxxxxxxxxxxxx> wrote:
>> > This patch series adds support for fsync/msync to DAX.
>> >
>> > Patches 1 through 8 add various utilities that the DAX code will eventually
>> > need, and the DAX code itself is added by patch 9. Patches 10 and 11 are
>> > filesystem changes that are needed after the DAX code is added, but these
>> > patches may change slightly as the filesystem fault handling for DAX is
>> > being modified ([1] and [2]).
>> >
>> > I've marked this series as RFC because I'm still testing, but I wanted to
>> > get this out there so people would see the direction I was going and
>> > hopefully comment on any big red flags sooner rather than later.
>> >
>> > I realize that we are getting pretty dang close to the v4.4 merge window,
>> > but I think that if we can get this reviewed and working it's a much better
>> > solution than the "big hammer" approach that blindly flushes entire PMEM
>> > namespaces [3].
>> >
>> > [1]
>> > [2]
>> > [3]
>> >
>> > Ross Zwisler (11):
>> > pmem: add wb_cache_pmem() to the PMEM API
>> > mm: add pmd_mkclean()
>> > pmem: enable REQ_FLUSH handling
>> > dax: support dirty DAX entries in radix tree
>> > mm: add follow_pte_pmd()
>> > mm: add pgoff_mkclean()
>> > mm: add find_get_entries_tag()
>> > fs: add get_block() to struct inode_operations
>> > dax: add support for fsync/sync
>> > xfs, ext2: call dax_pfn_mkwrite() on write fault
>> > ext4: add ext4_dax_pfn_mkwrite()
>> This is great to have when the flush-the-world solution ends up
>> killing performance. However, there are a couple mitigating options
>> for workloads that dirty small amounts and flush often that we need to
>> collect data on:
>> 1/ Using cache management and pcommit from userspace to skip calls to
>> msync / fsync. Although, this does not eliminate all calls to
>> blkdev_issue_flush as the fs may invoke it for other reasons. I
>> suspect turning on REQ_FUA support eliminates a number of those
>> invocations, and pmem already satisfies REQ_FUA semantics by default.
> Sure, I'll turn on REQ_FUA in addition to REQ_FLUSH - I agree that PMEM
> already handles the requirements of REQ_FUA, but I didn't realize that it
> might reduce the number of REQ_FLUSH bios we receive.

I'll let Dave chime in, but a lot of the flush requirements come from
guaranteeing the state of the metadata, if metadata updates can be
done with REQ_FUA then there is no subsequent need to flush.

>> 2/ Turn off DAX and use the page cache. As Dave mentions [1] we
>> should enable this control on a per-inode basis. I'm folding in this
>> capability as a blkdev_ioctl for the next version of the raw block DAX
>> support patch.
> Umm...I think you just said "the way to avoid this delay is to just not use
> DAX". :) I don't think this is where we want to go - we are trying to make
> DAX better, not abandon it.

That's a bit of an exaggeration. Avoiding DAX where it is not
necessary is not "abandoning DAX", it's using the right tool for the
job. Page cache is fine for many cases.

>> It's entirely possible these mitigations won't eliminate the need for
>> a mechanism like this, but I think we have a bit more work to do to
>> find out how bad this is in practice as well as the crossover point
>> where walking the radix becomes prohibitive.
> I'm guessing a single run through xfstests will be enough to convince you that
> the "big hammer" approach is untenable. Tests that used to take a second now
> take several minutes, at least in my VM testing environment... And that's
> only using a tiny 4GiB namespace.
> Yes, we can distribute the cost over multiple CPUs, but that just distributes
> the problem and doesn't reduce the overall work that needs to be done.
> Ultimately I think that looping through multiple GiB or even TiB of cache
> lines and blindly writing them back individually on every REQ_FLUSH is going
> to be a deal breaker.

Right, part of the problem is that the driver doesn't know which
blocks are actively DAX mapped. I think we can incrementally fix that
without requiring DAX specific fsync/msync handling code for each fs
that supports DAX.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at