On Mon, 2020-08-03 at 14:36 +0100, David Howells wrote:
> Here's a set of patches that adds a system call, fsinfo(), that
> allows information about the VFS, mount topology, superblock and
> files to be retrieved.
>
> The patchset is based on top of the notifications patchset and allows
> event counters implemented in the latter to be retrieved to allow
> overruns to be efficiently managed.
Could I repeat the question I asked about six months back that never
got answered:
https://lore.kernel.org/linux-api/[email protected]/
It sort of petered out into a long winding thread about why not use
sysfs instead, which really doesn't look like a good idea to me.
I'll repeat the information for those who want to quote it easily on
reply without having to use a web interface:
---
Could I make a suggestion about how this should be done in a way that
doesn't actually require the fsinfo syscall at all: it could just be
done with fsconfig. The idea is based on something I've wanted to do
for configfd but couldn't because otherwise it wouldn't substitute for
fsconfig, but Christian made me think it was actually essential to the
ability of the seccomp and other verifier tools in the critique of
configfd and I belive the same critique applies here.
Instead of making fsconfig functionally configure ... as in you pass
the attribute name, type and parameters down into the fs specific
handler and the handler does a string match and then verifies the
parameters and then acts on them, make it table configured, so what
each fstype does is register a table of attributes which can be got and
optionally set (with each attribute having a get and optional set
function). We'd have multiple tables per fstype, so the generic VFS
can register a table of attributes it understands for every fstype
(things like name, uuid and the like) and then each fs type would
register a table of fs specific attributes following the same pattern.
The system would examine the fs specific table before the generic one,
allowing overrides. fsconfig would have the ability to both get and
set attributes, permitting retrieval as well as setting (which is how I
get rid of the fsinfo syscall), we'd have a global parameter, which
would retrieve the entire table by name and type so the whole thing is
introspectable because the upper layer knows a-priori all the
attributes which can be set for a given fs type and what type they are
(so we can make more of the parsing generic). Any attribute which
doesn't have a set routine would be read only and all attributes would
have to have a get routine meaning everything is queryable.
I think I know how to code this up in a way that would be fully
transparent to the existing syscalls.
---
James
James Bottomley <[email protected]> wrote:
> It sort of petered out into a long winding thread about why not use
> sysfs instead, which really doesn't look like a good idea to me.
It seemed to turn into a set of procfs symlinks that pointed at a bunch of
sysfs stuff - or possibly some special filesystem.
> Could I make a suggestion about how this should be done in a way that
> doesn't actually require the fsinfo syscall at all: it could just be
> done with fsconfig.
I'd prefer to keep it separate. The interface for fsconfig() is intended to
move stuff into the kernel, not out of it. Better to add a parallel syscall
to go the other way (kind of like we have setxattr/getxattr, sendmsg/recvmsg).
Further, fsinfo() can refer directly to a file/fd/mount/whatever, but
fsconfig() doesn't do that. You have to use fspick() to get a context before
you can use fsconfig(). Now, that's fine if you want to gather several pieces
of information from a particular object, but it's not so good if you want to
get one piece of information from each of several objects.
> ... make it table configured...
I did, kind of (though I didn't call it that). Al rewrote the code to get rid
of it.
David