Re: Lease semantic proposal

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, 2019-10-01 at 11:17 -0700, Ira Weiny wrote:
> 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[1] much of the discussion of supporting RDMA with
> > > FS DAX has been around the semantics of the lease mechanism.[2]  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.[3] 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" semantic.
> > > 
> > > At Linux Plumbers[4] 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,
> > > Ira
> > > 
> > > [1] https://lkml.org/lkml/2019/8/9/1043
> > > [2] https://lkml.org/lkml/2019/8/9/1062
> > > [3] https://www.spinics.net/lists/fstests/msg12620.html
> > > [4] 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
> F_SETLEASE.
> 
> > 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
> NFS.
> 

The NFSv4.1 spec allows the client to request the layouts for a
particular range in the file:

https://tools.ietf.org/html/rfc5661#page-538

The knfsd only hands out whole-file layouts at present. Eventually we
may want to make better use of segmented layouts, at which point we'd
need something like a byte-range lease.

> > 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?
> 

I'd definitely not multiplex this over F_SETLEASE. An F_SETLAYOUT cmd
would probably be sufficient, and maybe just reuse
F_RDLCK/F_WRLCK/F_UNLCK for the iomode?

For the byte ranges, the catch there is that extending the userland
interface for that later will be difficult. What I'd probably suggest
(and what would jive with the way pNFS works) would be to go ahead and
add an offset and length to the arguments and result (maybe also
whence?).

The current kernel implementation could be free to deliver a larger
range than requested and only hand out full-file layouts for now.
Eventually we could rework the internals to allow for byte-range layout
leases.

I think this means that you'll probably require an argument struct for
layouts, analogous to struct flock for F_SETLK.

> > > **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
> > > procedure.
> > > 
> > > 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 unbreakable
> > > 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.
> 

If that's the model we want to use, then I think the owner of the file
will need some mechanism to forcibly seize the layout in this event.

What happens when the file is chowned in that case. Is that also denied?
If I set an F_UNBREAK layout (maybe as root) and then setuid(), do I get
to keep the layout?

> > 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.
> 

Even if the file is "pinned", I think this is still something that could
be abused. We need to try to consider how we will address those
situations up front.

In that same vein, I know you mentioned that conflicting activity will
just be denied when there is an outstanding F_UNBREAK lease. Will the
process holding one be notified in some fashion when another task
attempts to do some conflicting activity?

> > 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
> > looking.
> > 
> > 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.[1]
> 
> In fact should we consider files with F_UNBREAK leases "pinned" and just report
> them there?
> 
> [1] https://lkml.org/lkml/2019/8/9/1043

Sure, but eventually you'll want to track that back to a process that is
holding the thing. /proc/locks just shows you dev+ino for a particular
lock. You'll need to use something like lsof to figure out who is
holding the file open.

Since layouts aren't necessarily broken on open, there may be a bunch of
tasks that have the file open. How will you identify which one holds the
F_UNBREAK layout?
-- 
Jeff Layton <jlayton@xxxxxxxxxx>




[Index of Archives]     [XFS Filesystem Development (older mail)]     [Linux Filesystem Development]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux RAID]     [Linux SCSI]


  Powered by Linux