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 <j...@groves.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 <j...@groves.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
>
>
