Re: [PATCH 1/2] x86/vdso: Remove obsolete "fake section table" reservation
From: Andy Lutomirski
Date: Tue Dec 04 2018 - 14:47:02 EST
On Tue, Dec 4, 2018 at 11:28 AM Sean Christopherson
<sean.j.christopherson@xxxxxxxxx> wrote:
>
> On Tue, Dec 04, 2018 at 10:58:51AM -0800, Andy Lutomirski wrote:
> > On Tue, Dec 4, 2018 at 10:29 AM Sean Christopherson
> > <sean.j.christopherson@xxxxxxxxx> wrote:
> > >
> > > On Tue, Dec 04, 2018 at 10:22:39AM -0800, Sean Christopherson wrote:
> > > > On Tue, Dec 04, 2018 at 08:17:40AM -0800, Sean Christopherson wrote:
> > > > > At one point the vDSO image was manually stripped down by vdso2c in an
> > > > > attempt to minimize the size of the image mapped into userspace. Part
> > > > > of that stripping process involved building a fake section table so as
> > > > > not to break userspace processes that parse the section table. Memory
> > > > > for the fake section table was reserved in the .rodata section so that
> > > > > vdso2c could simply copy the entire PT_LOAD segment into the userspace
> > > > > image after building the fake table.
> > > > >
> > > > > Eventually, the entire fake section table approach was dropped in favor
> > > > > of stripping the vdso "the old fashioned way", i.e. via objdump -S.
> > > > > But, the reservation in .rodata for the fake table was left behind.
> > > > > Remove the reserveration along with a few other related defines and
> > > > > section entries.
> > > > >
> > > > > Removing the fake section table placeholder zaps a whopping 0x340 bytes
> > > > > from the 64-bit vDSO image, which drops the current image's size to
> > > > > under 4k, i.e. reduces the effective size of the userspace vDSO mapping
> > > > > by a full page.
> > > > >
> > > > > Fixes: da861e18eccc ("x86, vdso: Get rid of the fake section mechanism")
> > > > > Cc: Andy Lutomirski <luto@xxxxxxxxxxxxxx>
> > > > > Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> > > > > ---
> > > > > diff --git a/arch/x86/entry/vdso/vdso2c.c b/arch/x86/entry/vdso/vdso2c.c
> > > > > index 4674f58581a1..2479a454b15c 100644
> > > > > --- a/arch/x86/entry/vdso/vdso2c.c
> > > > > +++ b/arch/x86/entry/vdso/vdso2c.c
> > > > > @@ -98,12 +98,6 @@ struct vdso_sym required_syms[] = {
> > > > > [sym_hpet_page] = {"hpet_page", true},
> > > > > [sym_pvclock_page] = {"pvclock_page", true},
> > > > > [sym_hvclock_page] = {"hvclock_page", true},
> > > > > - [sym_VDSO_FAKE_SECTION_TABLE_START] = {
> > > > > - "VDSO_FAKE_SECTION_TABLE_START", false
> > > > > - },
> > > > > - [sym_VDSO_FAKE_SECTION_TABLE_END] = {
> > > > > - "VDSO_FAKE_SECTION_TABLE_END", false
> > > > > - },
> > > >
> > > > Doh, I missed removing the definitions for sym_VDSO_FAKE_SECTION_TABLE_*.
> > >
> > > And with sym_VDSO_FAKE_SECTION_TABLE_* gone all symbols are exported,
> > > meaning required_syms can be a char* array and struct vdso_sym can be
> > > removed.
> >
> > I bet that we'll want that field to parse out the extable entries once
> > all the dust settles, though.
>
> Hmm, the extable stuff will be sections, a la .altinstructions, I don't
> think we'll need separate symbols.
In my mind, the easiest nice implementation is to have a section that
gets thrown away by objcopy --remove-section, and to have the actual
table in that section be bracketed by a pair of symbols, since vdso2c
already parses that. But vdso2c could certainly also just find the
section direclty. Then vdso2c could expose an list of extable entries
in the vdso_image, and that would be that. Does that seem reasonable?
>
> What if I send out a RFC for the extable stuff in parallel to this series
> and to the SGX series? It'd also be nice to iterate on that code without
> having to spin a full SGX series.
Sounds great!