On Mon, Sep 23, 2019 at 04:17:59PM -0400, Jeff Layton wrote:
On Mon, 2019-09-23 at 12:08 -0700, Ira Weiny wrote:
> Since the last RFC patch set much of the discussion of supporting RDMA with
> FS DAX has been around the semantics of the lease mechanism. Within that
> thread it was suggested I try and write some documentation and/or tests for the
> new mechanism being proposed. I have created a foundation to test lease
> functionality within xfstests. This should be close to being accepted.
> Before writing additional lease tests, or changing lots of kernel code, this
> email presents documentation for the new proposed "layout lease"
> At Linux Plumbers just over a week ago, I presented the current state of the
> patch set and the outstanding issues. Based on the discussion there, well as
> follow up emails, I propose the following addition to the fcntl() man page.
> Thank you,
>  https://lkml.org/lkml/2019/8/9/1043
>  https://lkml.org/lkml/2019/8/9/1062
>  https://www.spinics.net/lists/fstests/msg12620.html
>  https://linuxplumbersconf.org/event/4/contributions/368/
Thank you so much for doing this, Ira. This allows us to debate the
user-visible behavior semantics without getting bogged down in the
implementation details. More comments below:
Thanks. Sorry for the delay in response. Turns out this email was in my
spam... :-/ I'll need to work out why.
> <fcntl man page addition>
> Layout Leases
> Layout (F_LAYOUT) leases are special leases which can be used to control and/or
> be informed about the manipulation of the underlying layout of a file.
> A layout is defined as the logical file block -> physical file block mapping
> including the file size and sharing of physical blocks among files. Note that
> the unwritten state of a block is not considered part of file layout.
> **Read layout lease F_RDLCK | F_LAYOUT**
> Read layout leases can be used to be informed of layout changes by the
> system or other users. This lease is similar to the standard read (F_RDLCK)
> lease in that any attempt to change the _layout_ of the file will be reported to
> the process through the lease break process. But this lease is different
> because the file can be opened for write and data can be read and/or written to
> the file as long as the underlying layout of the file does not change.
> Therefore, the lease is not broken if the file is simply open for write, but
> _may_ be broken if an operation such as, truncate(), fallocate() or write()
> results in changing the underlying layout.
> **Write layout lease (F_WRLCK | F_LAYOUT)**
> Write Layout leases can be used to break read layout leases to indicate that
> the process intends to change the underlying layout lease of the file.
> A process which has taken a write layout lease has exclusive ownership of the
> file layout and can modify that layout as long as the lease is held.
> Operations which change the layout are allowed by that process. But operations
> from other file descriptors which attempt to change the layout will break the
> lease through the standard lease break process. The F_LAYOUT flag is used to
> indicate a difference between a regular F_WRLCK and F_WRLCK with F_LAYOUT. In
> the F_LAYOUT case opens for write do not break the lease. But some operations,
> if they change the underlying layout, may.
> The distinction between read layout leases and write layout leases is that
> write layout leases can change the layout without breaking the lease within the
> owning process. This is useful to guarantee a layout prior to specifying the
> unbreakable flag described below.
The above sounds totally reasonable. You're essentially exposing the
behavior of nfsd's layout leases to userland. To be clear, will F_LAYOUT
leases work the same way as "normal" leases, wrt signals and timeouts?
That was my intention, yes.
I do wonder if we're better off not trying to "or" in flags for this,
and instead have a separate set of commands (maybe F_RDLAYOUT,
F_WRLAYOUT, F_UNLAYOUT). Maybe I'm just bikeshedding though -- I don't
feel terribly strongly about it.
I'm leaning that was as well. To make these even more distinct from
Also, at least in NFSv4, layouts are handed out for a particular byte
range in a file. Should we consider doing this with an API that allows
for that in the future? Is this something that would be desirable for
your RDMA+DAX use-cases?
I don't see this. I've thought it would be a nice thing to have but I don't
know of any hard use case. But first I'd like to understand how this works for
We could add a new F_SETLEASE variant that takes a struct with a byte
range (something like struct flock).
I think this would be another reason to introduce F_[RD|WR|UN]LAYOUT as a
command. Perhaps supporting smaller byte ranges could be added later?
> **Unbreakable Layout Leases (F_UNBREAK)**
> In order to support pinning of file pages by direct user space users an
> unbreakable flag (F_UNBREAK) can be used to modify the read and write layout
> lease. When specified, F_UNBREAK indicates that any user attempting to break
> the lease will fail with ETXTBUSY rather than follow the normal breaking
> Both read and write layout leases can have the unbreakable flag (F_UNBREAK)
> specified. The difference between an unbreakable read layout lease and an
> unbreakable write layout lease are that an unbreakable read layout lease is
> _not_ exclusive. This means that once a layout is established on a file,
> multiple unbreakable read layout leases can be taken by multiple processes and
> used to pin the underlying pages of that file.
> Care must therefore be taken to ensure that the layout of the file is as the
> user wants prior to using the unbreakable read layout lease. A safe mechanism
> to do this would be to take a write layout lease and use fallocate() to set the
> layout of the file. The layout lease can then be "downgraded" to
> read layout as long as no other user broke the write layout lease.
Will userland require any special privileges in order to set an
F_UNBREAK lease? This seems like something that could be used for DoS. I
assume that these will never time out.
Dan and I discussed this some more and yes I think the uid of the process needs
to be the owner of the file. I think that is a reasonable mechanism.
How will we deal with the case where something is is squatting on an
F_UNBREAK lease and isn't letting it go?
That is a good question. I had not considered someone taking the UNBREAK
without pinning the file.
Leases are technically "owned" by the file description -- we can't
necessarily trace it back to a single task in a threaded program. The
kernel task that set the lease may have exited by the time we go
Will we be content trying to determine this using /proc/locks+lsof, etc,
or will we need something better?
I think using /proc/locks is our best bet. Similar to my intention to report
files being pinned.
In fact should we consider files with F_UNBREAK leases "pinned" and just report
> > </fcntl man page addition>
> Jeff Layton <jlayton(a)kernel.org>