Re: [PATCH v9 07/13] um: nommu: configure fs register on host syscall invocation

[Hajime Tazaki <thehajime@gmail.com>]

On Thu, 19 Jun 2025 21:38:47 +0900,
Benjamin Berg wrote:
> 
> On Thu, 2025-06-19 at 21:22 +0900, Hajime Tazaki wrote:
> > 
> > On Thu, 19 Jun 2025 19:40:49 +0900,
> > Benjamin Berg wrote:
> > > 
> > > On Thu, 2025-06-19 at 10:04 +0900, Hajime Tazaki wrote:
> > > > As userspace on UML/!MMU also need to configure %fs register when
> > > > it is
> > > > running to correctly access thread structure, host syscalls
> > > > implemented
> > > > in os-Linux drivers may be puzzled when they are called.  Thus it
> > > > has to
> > > > configure %fs register via arch_prctl(SET_FS) on every host
> > > > syscalls.
> > > 
> > > Really, I still think that we should "just" get rid of libc
> > > entirely
> > > inside UML. That would avoid so many weird/potential issues …
> > 
> > I'm not sure if I understand your point.
> > 
> > Q1) what do you mean by 'get rid of libc entirely' here ?
> >     do you mean the following code block adds the dependency ?
> > + int os_arch_prctl(int pid, int option, unsigned long *arg2)
> > + {
> > + 	return syscall(SYS_arch_prctl, option, arg2);
> > + }
> > 
> > I guess this can be replaced with inline assembly instead of using
> > libc's one.  but this is the code under os-Linux, which I thought
> > we're
> > allowed to use the host code ?
> > 
> > Q2) "That would avoid so many weird/potential issues …"
> > I'm new to this;  I'm wondering what kind of issues did you see ?
> 
> Oh, I am just being annoyed by libc in general in UM. It isn't specific
> to this patchset.
> 
> An example is that we need to keep malloc() working for libc. Which I
> would think is kind of weird. Or we had issues because libc turned on
> rseq and that was inherited into userspace, causing random crashes and
> such.

I understand, thanks for the input.

> > > Doesn't change the fact that FS/GS needs to be restored when doing
> > > thread switches and such. Though one might be able to do it
> > > entirely
> > > within arch_switch_to then.
> > 
> > I believe this is already done in arch_switch_to.  This particular
> > patch does the control to the host context.
> 
> OK, need to look at that again a bit. I haven't really wrapped my mind
> around how everything fits together, so I probably got some stuff
> wrong.

anyway, thanks for your time to look at this.

-- Hajime