Re: [PATCH 0/8] use struct pt_regs based syscall calling for x86-64

[Dominik Brodowski <linux@dominikbrodowski.net>]

On Fri, Apr 06, 2018 at 11:20:46AM +0200, Ingo Molnar wrote:
> 
> * Dominik Brodowski <linux@dominikbrodowski.net> wrote:
> 
> > On Fri, Apr 06, 2018 at 10:23:22AM +0200, Ingo Molnar wrote:
> > > 
> > > * Dominik Brodowski <linux@dominikbrodowski.net> wrote:
> > > 
> > > > On Thu, Apr 05, 2018 at 05:19:33PM +0200, Ingo Molnar wrote:
> > > > > Ok, this series looks mostly good to me, but AFAICS this breaks the UML build:
> > > > > 
> > > > >  make[2]: *** No rule to make target 'archheaders'.  Stop.
> > > > >  arch/um/Makefile:119: recipe for target 'archheaders' failed
> > > > >  make[1]: *** [archheaders] Error 2
> > > > >  make[1]: *** Waiting for unfinished jobs....
> > > > 
> > > > Ah, that's caused by patch 8/8 which I did and do not like all that much
> > > > anyway: UML re-uses syscall_64.tbl which now has x86-specific entries like
> > > > __sys_x86_pread64, but expects the generic syscall stub sys_pread64
> > > > referenced there. Fixup patch below; could be folded with patch 8/8. Or
> > > > patch 8/8 could simply be dropped from the series altogether...
> > > 
> > > I still like the 'truth in advertising' aspect. For example if I see this in the 
> > > syscall table:
> > > 
> > >  10      common  mprotect                __sys_x86_mprotect
> > > 
> > > I can immediately find the _real_ syscall entry point:
> > > 
> > > ffffffff81180a10 <__sys_x86_mprotect>:
> > > ffffffff81180a10:       48 8b 57 60             mov    0x60(%rdi),%rdx
> > > ffffffff81180a14:       48 8b 77 68             mov    0x68(%rdi),%rsi
> > > ffffffff81180a18:       b9 ff ff ff ff          mov    $0xffffffff,%ecx
> > > ffffffff81180a1d:       48 8b 7f 70             mov    0x70(%rdi),%rdi
> > > ffffffff81180a21:       e8 fa fc ff ff          callq  ffffffff81180720 <do_mprotect_pkey>
> > > ffffffff81180a26:       48 98                   cltq   
> > > ffffffff81180a28:       c3                      retq   
> > > ffffffff81180a29:       0f 1f 80 00 00 00 00    nopl   0x0(%rax)
> > > 
> > > If, on the other hand, I see this entry:
> > > 
> > >  10     common  mprotect                sys_mprotect
> > > 
> > > Then, as a first step, no symbol anywhere matches with this:
> > > 
> > >  triton:~/tip> grep sys_mprotect System.map 
> > >  triton:~/tip> 
> > > 
> > > "sys_mprotect" does not exist in any easily discoverable sense. You have to *know* 
> > > to replace the sys_ prefix with __sys_x86_ to find it.
> > > 
> > > Now arguably we could use a __sys_ prefix instead of the grep-barrier __sys_x86 
> > > prefix - but that too would be somewhat confusing I think.
> > 
> > Well, if looking at the ARCH="um" kernel, you won't find the __sys_x86_mprotect 
> > there in its System.map -- so we either have to disentangle um and plain x86, or 
> > live with some cause for confusion.
> 
> I'm primarily concerned about everything making sense on x86 - UML is an entirely 
> separate architecture with heavy tradeoffs and kludges.

Agreed.

> > __sys_mprotect as prefix won't work by the way, as the double-underscore __sys_ 
> > variant is already used in net/* for internal syscall helpers.
> 
> Ok - then triple underscore - but overall I think it's more confusing.
> 
> Btw., what was the problem with calling the x86 ptregs wrapper sys_mprotect?
> 
> The only reason I suggested the __sys_x86_ prefix was because you originally 
> suggested that there's symbol name overlap, but I don't think that's the case 
> within the same kernel build, as the regular non-ptregs prototype:

Indeed, there's no symbol name overlap within the same kernel build, but
technically different stubs named the same. If that's fine, just drop patch
8/8 (including the UML fixup) and things should be fine, with the stub and
the entry in the syscall table both named sys_mprotect.

For IA32_EMULATION, we have __sys_ia32_mprotect as stub for the same
syscall, including this name as entry in syscall_32.tbl.

More problematic is the naming for the compat stubs for IA32_EMAULATION and
X32, where we have

	__compat_sys_ia32_waitid
	__compat_sys_x32_waitid

for example. We *could* rename one of those to compat_sys_waitid() and levae
the other as-is, but actually I prefer it now how it is.

Thanks,
	Dominik