Re: Why does namei return an exclusively locked vnode when LOCKSHARED is specified?

From: Alan Somers <asomers_at_freebsd.org>
Date: Wed, 18 Dec 2024 18:15:45 UTC
On Tue, Dec 17, 2024 at 1:02 PM Alan Somers <asomers@freebsd.org> wrote:
>
> On Tue, Dec 17, 2024 at 12:27 PM Mark Johnston <markj@freebsd.org> wrote:
> >
> > On Tue, Dec 17, 2024 at 12:19:26PM -0700, Alan Somers wrote:
> > > According to namei(9), namei should return a shared lock when
> > > LOCKSHARED is specified.  But in my experiments, it always seems to
> > > return an exclusive lock instead.  For example:
> > >
> > > $ cd /usr/tests/sys/fs/fusefs
> > > $ sudo dtrace -i 'fbt:fusefs:fuse_vnop_getattr:entry /pid==$target/
> > > {printf("islocked=%#x", args[0]->a_vp->v_vnlock->lk_lock); stack();}'
> > > -i 'fbt:kernel:vop_stdstat:entry /pid==$target/
> > > {printf("islocked=%#x", args[0]->a_vp->v_vnlock->lk_lock);}' -c
> > > "./getattr --gtest_filter=Getattr.attr_cache -v"
> > > [==========] Running 1 test from 1 test suite.
> > > [----------] Global test environment set-up.
> > > [----------] 1 test from Getattr
> > > [ RUN      ] Getattr.attr_cache
> > > INIT        ino= 0
> > > ACCESS      ino= 1 mask=0x1
> > > LOOKUP      ino= 1 some_file.txt
> > > GETATTR     ino=42
> > > [       OK ] Getattr.attr_cache (3 ms)
> > > [----------] 1 test from Getattr (3 ms total)
> > >
> > > [----------] Global test environment tear-down
> > > [==========] 1 test from 1 test suite ran. (4 ms total)
> > > [  PASSED  ] 1 test.
> > > CPU     ID                    FUNCTION:NAME
> > >   3  19743                vop_stdstat:entry islocked=0x21
> > >   3  19743                vop_stdstat:entry islocked=0xfffff80004f16740
> > >   3  68298          fuse_vnop_getattr:entry islocked=0xfffff80004f16740
> > >               kernel`VOP_GETATTR_APV+0x90
> > >               kernel`vop_stdstat+0x147
> > >               kernel`VOP_STAT_APV+0x91
> > >               kernel`kern_statat+0x183
> > >               kernel`sys_fstatat+0x27
> > >               kernel`amd64_syscall+0x1af
> > >               kernel`0xffffffff8106cd5b
> > >
> > >   3  19743                vop_stdstat:entry islocked=0xfffff80004f16740
> > >   3  68298          fuse_vnop_getattr:entry islocked=0xfffff80004f16740
> > >               kernel`VOP_GETATTR_APV+0x90
> > >               kernel`vop_stdstat+0x147
> > >               kernel`VOP_STAT_APV+0x91
> > >               kernel`kern_statat+0x183
> > >               kernel`sys_fstatat+0x27
> > >               kernel`amd64_syscall+0x1af
> > >               kernel`0xffffffff8106cd5b
> > >
> > > dtrace: pid 3554 has exited
> > >
> > > From that output, the first vop_stdstat is probably for the
> > > mountpoint, and isn't what I'm concerned about.  But the second two
> > > are for a fusefs file.  The LK_SHARE bit is not set, indicating an
> > > exclusive lock.  That's even though the call to namei in kern_statat
> > > specifies LOCKSHARED | LOCKLEAF.
> > >
> > > What am I missing?
> >
> > Having noticed the same phenomenon in p9fs and scratching my head for a
> > while: is fusefs failing to call VN_LOCK_ASHARE() in this case?  I see
> > that it's predicated on "data->dataflags & FSESS_ASYNC_READ", but I'm
> > not sure why - the comment seems to suggest a misunderstanding of what
> > VN_LOCK_ASHARE() does.
> >
> > ... oh, also perhaps because fusefs mounts don't set MNTK_LOOKUP_SHARED?
>
> Yes, that's it!  Both of those things need to change.  Now I can get
> on with solving my real bug.  Thank you.

FYI: this is the bug I was working on:
https://github.com/freebsd/freebsd-src/pull/1556