On Tue, Dec 20, 2016 at 4:40 PM, Darrick J. Wong
<darrick.wong@xxxxxxxxxx> wrote:
> On Mon, Dec 19, 2016 at 05:18:40PM -0800, Dan Williams wrote:
>> On Mon, Dec 19, 2016 at 5:09 PM, Darrick J. Wong
>> <darrick.wong@xxxxxxxxxx> wrote:
>> > On Mon, Dec 19, 2016 at 02:11:49PM -0700, Ross Zwisler wrote:
>> >> On Fri, Sep 16, 2016 at 03:54:05PM +1000, Nicholas Piggin wrote:
>> >> <>
>> >> > Definitely the first step would be your simple preallocated per
>> >> > inode approach until it is shown to be insufficient.
>> >>
>> >> Reviving this thread a few months later...
>> >>
>> >> Dave, we're interested in taking a serious look at what it would take to get
>> >> PMEM_IMMUTABLE working. Do you still hold the opinion that this is (or could
>> >> become, with some amount of work) a workable solution?
>> >>
>> >> We're happy to do the grunt work for this feature, but we will probably need
>> >> guidance from someone with more XFS experience. With you out on extended leave
>> >> the first half of 2017, who would be the best person to ask for this guidance?
>> >> Darrick?
>> >
>> > Yes, probably. :)
>> >
>> > I think where we left off with this (on the XFS side) is some sort of
>> > fallocate mode that would allocate blocks, zero them, and then set the
>> > DAX and PMEM_IMMUTABLE on-disk inode flags. After that, you'd mmap the
>> > file and thereby gain the ability to control write persistents behavior
>> > without having to worry about fs metadata updates. As an added plus, I
>> > think zeroing the pmem also clears media errors, or something like that.
>> >
>> > <shrug> Is that a reasonable starting point? My memory is a little foggy.
>> >
>> > Hmm, I see Dan just posted something about blockdev fallocate. I'll go
>> > read that.
>>
>> That's for device-dax, which is basically a poor man's PMEM_IMMUTABLE
>> via a character device interface. It's useful for cases where you want
>> an entire nvdimm namespace/volume in "no fs-metadata to worry about"
>> mode. But, for sub-allocations of a namespace and support for
>> existing tooling, PMEM_IMMUTABLE is much more usable.
>
> Well sure... but otoh I was thinking that it'd be pretty neat if we
> could use the same code regardless of whether the target file was a
> dax-device or an xfs file:
>
> fd = open("<some path>", O_RDWR);
> fstat(fd, &statbuf):
> fallocate(fd, FALLOC_FL_PMEM_IMMUTABLE, 0, statbuf.st_size);
> p = mmap(NULL, statbuf.st_size, PROT_READ | PROT_WRITE, fd, 0);
>
> *(p + 42) = 0xDEADBEEF;
> asm { clflush; } /* or whatever */
>
> ...so perhaps it would be a good idea to design the fallocate primitive
> around "prepare this fd for mmap-only pmem semantics" and let it the
> backend do zeroing and inode flag changes as necessary to make it
> happen. We'd need to do some bikeshedding about what the other falloc
> flags mean when we're dealing with pmem files and devices, but I think
> we should try to keep the userland presentation the same unless there's
> a really good reason not to.
It would be interesting to use fallocate to size device-dax files...
However, the fallocate userland presentation I was looking to unify
was the semantic for how to clear errors. The current method for
creating a device-dax instance is 'echo "<some size>" > "<some sysfs
file>"'. Christoph put the kibosh on carrying the
fallocate-error-clearing semantic over, so we'll need to fallback to
using the existing ND_IOCTL_CLEAR_ERROR. All that's missing for error
clearing is a secondary mechanism to report the media error list for a
physical address range.
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
