Message-ID: <20250709015348.GD2672029@frogsfrogsfrogs>
Date: Tue, 8 Jul 2025 18:53:48 -0700
From: "Darrick J. Wong" <djwong@...nel.org>
To: John Groves <John@...ves.net>
Cc: Amir Goldstein <amir73il@...il.com>,
	Dan Williams <dan.j.williams@...el.com>,
	Miklos Szeredi <miklos@...redi.hu>,
	Bernd Schubert <bschubert@....com>,
	John Groves <jgroves@...ron.com>, Jonathan Corbet <corbet@....net>,
	Vishal Verma <vishal.l.verma@...el.com>,
	Dave Jiang <dave.jiang@...el.com>,
	Matthew Wilcox <willy@...radead.org>, Jan Kara <jack@...e.cz>,
	Alexander Viro <viro@...iv.linux.org.uk>,
	Christian Brauner <brauner@...nel.org>,
	Randy Dunlap <rdunlap@...radead.org>,
	Jeff Layton <jlayton@...nel.org>,
	Kent Overstreet <kent.overstreet@...ux.dev>,
	linux-doc@...r.kernel.org, linux-kernel@...r.kernel.org,
	nvdimm@...ts.linux.dev, linux-cxl@...r.kernel.org,
	linux-fsdevel@...r.kernel.org,
	Jonathan Cameron <Jonathan.Cameron@...wei.com>,
	Stefan Hajnoczi <shajnocz@...hat.com>,
	Joanne Koong <joannelkoong@...il.com>,
	Josef Bacik <josef@...icpanda.com>,
	Aravind Ramesh <arramesh@...ron.com>,
	Ajay Joshi <ajayjoshi@...ron.com>
Subject: Re: [RFC V2 10/18] famfs_fuse: Basic fuse kernel ABI enablement for
 famfs

On Tue, Jul 08, 2025 at 07:02:03AM -0500, John Groves wrote:
> On 25/07/07 10:39AM, Darrick J. Wong wrote:
> > On Fri, Jul 04, 2025 at 08:39:59AM -0500, John Groves wrote:
> > > On 25/07/04 09:54AM, Amir Goldstein wrote:
> > > > On Thu, Jul 3, 2025 at 8:51 PM John Groves <John@...ves.net> wrote:
> > > > >
> > > > > * FUSE_DAX_FMAP flag in INIT request/reply
> > > > >
> > > > > * fuse_conn->famfs_iomap (enable famfs-mapped files) to denote a
> > > > >   famfs-enabled connection
> > > > >
> > > > > Signed-off-by: John Groves <john@...ves.net>
> > > > > ---
> > > > >  fs/fuse/fuse_i.h          |  3 +++
> > > > >  fs/fuse/inode.c           | 14 ++++++++++++++
> > > > >  include/uapi/linux/fuse.h |  4 ++++
> > > > >  3 files changed, 21 insertions(+)
> > > > >
> > > > > diff --git a/fs/fuse/fuse_i.h b/fs/fuse/fuse_i.h
> > > > > index 9d87ac48d724..a592c1002861 100644
> > > > > --- a/fs/fuse/fuse_i.h
> > > > > +++ b/fs/fuse/fuse_i.h
> > > > > @@ -873,6 +873,9 @@ struct fuse_conn {
> > > > >         /* Use io_uring for communication */
> > > > >         unsigned int io_uring;
> > > > >
> > > > > +       /* dev_dax_iomap support for famfs */
> > > > > +       unsigned int famfs_iomap:1;
> > > > > +
> > > > 
> > > > pls move up to the bit fields members.
> > > 
> > > Oops, done, thanks.
> > > 
> > > > 
> > > > >         /** Maximum stack depth for passthrough backing files */
> > > > >         int max_stack_depth;
> > > > >
> > > > > diff --git a/fs/fuse/inode.c b/fs/fuse/inode.c
> > > > > index 29147657a99f..e48e11c3f9f3 100644
> > > > > --- a/fs/fuse/inode.c
> > > > > +++ b/fs/fuse/inode.c
> > > > > @@ -1392,6 +1392,18 @@ static void process_init_reply(struct fuse_mount *fm, struct fuse_args *args,
> > > > >                         }
> > > > >                         if (flags & FUSE_OVER_IO_URING && fuse_uring_enabled())
> > > > >                                 fc->io_uring = 1;
> > > > > +                       if (IS_ENABLED(CONFIG_FUSE_FAMFS_DAX) &&
> > > > > +                           flags & FUSE_DAX_FMAP) {
> > > > > +                               /* XXX: Should also check that fuse server
> > > > > +                                * has CAP_SYS_RAWIO and/or CAP_SYS_ADMIN,
> > > > > +                                * since it is directing the kernel to access
> > > > > +                                * dax memory directly - but this function
> > > > > +                                * appears not to be called in fuse server
> > > > > +                                * process context (b/c even if it drops
> > > > > +                                * those capabilities, they are held here).
> > > > > +                                */
> > > > > +                               fc->famfs_iomap = 1;
> > > > > +                       }
> > > > 
> > > > 1. As long as the mapping requests are checking capabilities we should be ok
> > > >     Right?
> > > 
> > > It depends on the definition of "are", or maybe of "mapping requests" ;)
> > > 
> > > Forgive me if this *is* obvious, but the fuse server capabilities are what
> > > I think need to be checked here - not the app that it accessing a file.
> > > 
> > > An app accessing a regular file doesn't need permission to do raw access to
> > > the underlying block dev, but the fuse server does - becuase it is directing
> > > the kernel to access that for apps.
> > > 
> > > > 2. What's the deal with capable(CAP_SYS_ADMIN) in process_init_limits then?
> > > 
> > > I *think* that's checking the capabilities of the app that is accessing the
> > > file, and not the fuse server. But I might be wrong - I have not pulled very
> > > hard on that thread yet.
> > 
> > The init reply should be processed in the context of the fuse server.
> > At that point the kernel hasn't exposed the fs to user programs, so
> > (AFAICT) there won't be any other programs accessing that fuse mount.
> 
> Hmm. It would be good if you're right about that. My fuse server *is* running
> as root, and when I check those capabilities in process_init_reply(), I
> find those capabilities. So far so good.
> 
> Then I added code to my fuse server to drop those capabilities prior to
> starting the fuse session (prctl(PR_CAPBSET_DROP, CAP_SYS_RAWIO) and 
> prctl(PR_CAPBSET_DROP, CAP_SYS_ADMIN). I expected (hoped?) to see those 
> capabilities disappear in process_init_reply() - but they did not disappear.
> 
> I'm all ears if somebody can see a flaw in my logic here. Otherwise, the
> capabilities need to be stashed away before the reply is processsed, when 
> fs/fuse *is* running in fuse server context.
> 
> I'm somewhat surprised if that isn't already happening somewhere...

Hrm.  I *thought* that since FUSE_INIT isn't queued as a background
command, it should still execute in the same process context as the fuse
server.

OTOH it also occurs to me that I have this code in fuse_send_init:

	if (has_capability_noaudit(current, CAP_SYS_RAWIO))
		flags |= FUSE_IOMAP | FUSE_IOMAP_DIRECTIO | FUSE_IOMAP_PAGECACHE;
	...
	ia->in.flags = flags;
	ia->in.flags2 = flags >> 32;

which means that we only advertise iomap support in FUSE_INIT if the
process running fuse_fill_super (which you hope is the fuse server)
actually has CAP_SYS_RAWIO.  Would that work for you?  Or are you
dropping privileges before you even open /dev/fuse?

Note: I might decide to relax that approach later on, since iomap
requires you to have opened a block device ... which implies that the
process had read/write access to start with; and maybe we're ok with
unprivileged fuse2fs servers running on a chmod 666 block device?

<shrug> always easier to /relax/ the privilege checks. :)

> > > > 3. Darrick mentioned the need for a synchronic INIT variant for his work on
> > > >     blockdev iomap support [1]
> > > 
> > > I'm not sure that's the same thing (Darrick?), but I do think Darrick's
> > > use case probably needs to check capabilities for a server that is sending
> > > apps (via files) off to access extents of block devices.
> > 
> > I don't know either, Miklos hasn't responded to my questions.  I think
> > the motivation for a synchronous 
> 
> ?

..."I don't know what his motivations for synchronous FUSE_INIT are."

I guess I fubard vim. :(

> > As for fuse/iomap, I just only need to ask the kernel if iomap support
> > is available before calling ext2fs_open2() because the iomap question
> > has some implications for how we open the ext4 filesystem.
> > 
> > > > I also wonder how much of your patches and Darrick's patches end up
> > > > being an overlap?
> > > 
> > > Darrick and I spent some time hashing through this, and came to the conclusion
> > > that the actual overlap is slim-to-none. 
> > 
> > Yeah.  The neat thing about FMAPs is that you can establish repeating
> > patterns, which is useful for interleaved DRAM/pmem devices.  Disk
> > filesystems don't do repeating patterns, so they'd much rather manage
> > non-repeating mappings.
> 
> Right. Interleaving is critical to how we use memory, so fmaps are designed
> to support it.
> 
> Tangent: at some point a broader-than-just-me discussion of how block devices
> have the device mapper, but memory has no such layout tools, might be good
> to have. Without such a thing (which might or might not be possible/practical),
> it's essential that famfs do the interleaving. Lacking a mapper layer also
> means that we need dax to provide a clean "device abstraction" (meaning
> a single CXL allocation [which has a uuid/tag] needs to appear as a single
> dax device whether or not it's HPA-contiguous).

Well it's not as simple as device-mapper, where we can intercept struct
bio and remap/split it to our heart's content.  I guess you could do
that with an iovec...?  Would be sorta amusing if you could software
RAID10 some DRAM. :P

--D

> Cheers,
> John
> 
> 

Powered by blists - more mailing lists