Re: [Xen-devel] Xen PV domain regression with KASLR enabled (kernel 3.16)

From: Konrad Rzeszutek Wilk
Date: Wed Aug 27 2014 - 16:50:00 EST


On Wed, Aug 27, 2014 at 10:03:10AM +0200, Stefan Bader wrote:
> On 26.08.2014 18:01, Konrad Rzeszutek Wilk wrote:
> > On Fri, Aug 22, 2014 at 11:20:50AM +0200, Stefan Bader wrote:
> >> On 21.08.2014 18:03, Kees Cook wrote:
> >>> On Tue, Aug 12, 2014 at 2:07 PM, Konrad Rzeszutek Wilk
> >>> <konrad.wilk@xxxxxxxxxx> wrote:
> >>>> On Tue, Aug 12, 2014 at 11:53:03AM -0700, Kees Cook wrote:
> >>>>> On Tue, Aug 12, 2014 at 11:05 AM, Stefan Bader
> >>>>> <stefan.bader@xxxxxxxxxxxxx> wrote:
> >>>>>> On 12.08.2014 19:28, Kees Cook wrote:
> >>>>>>> On Fri, Aug 8, 2014 at 7:35 AM, Stefan Bader <stefan.bader@xxxxxxxxxxxxx> wrote:
> >>>>>>>> On 08.08.2014 14:43, David Vrabel wrote:
> >>>>>>>>> On 08/08/14 12:20, Stefan Bader wrote:
> >>>>>>>>>> Unfortunately I have not yet figured out why this happens, but can confirm by
> >>>>>>>>>> compiling with or without CONFIG_RANDOMIZE_BASE being set that without KASLR all
> >>>>>>>>>> is ok, but with it enabled there are issues (actually a dom0 does not even boot
> >>>>>>>>>> as a follow up error).
> >>>>>>>>>>
> >>>>>>>>>> Details can be seen in [1] but basically this is always some portion of a
> >>>>>>>>>> vmalloc allocation failing after hitting a freshly allocated PTE space not being
> >>>>>>>>>> PTE_NONE (usually from a module load triggered by systemd-udevd). In the
> >>>>>>>>>> non-dom0 case this repeats many times but ends in a guest that allows login. In
> >>>>>>>>>> the dom0 case there is a more fatal error at some point causing a crash.
> >>>>>>>>>>
> >>>>>>>>>> I have not tried this for a normal PV guest but for dom0 it also does not help
> >>>>>>>>>> to add "nokaslr" to the kernel command-line.
> >>>>>>>>>
> >>>>>>>>> Maybe it's overlapping with regions of the virtual address space
> >>>>>>>>> reserved for Xen? What the the VA that fails?
> >>>>>>>>>
> >>>>>>>>> David
> >>>>>>>>>
> >>>>>>>> Yeah, there is some code to avoid some regions of memory (like initrd). Maybe
> >>>>>>>> missing p2m tables? I probably need to add debugging to find the failing VA (iow
> >>>>>>>> not sure whether it might be somewhere in the stacktraces in the report).
> >>>>>>>>
> >>>>>>>> The kernel-command line does not seem to be looked at. It should put something
> >>>>>>>> into dmesg and that never shows up. Also today's random feature is other PV
> >>>>>>>> guests crashing after a bit somewhere in the check_for_corruption area...
> >>>>>>>
> >>>>>>> Right now, the kaslr code just deals with initrd, cmdline, etc. If
> >>>>>>> there are other reserved regions that aren't listed in the e820, it'll
> >>>>>>> need to locate and skip them.
> >>>>>>>
> >>>>>>> -Kees
> >>>>>>>
> >>>>>> Making my little steps towards more understanding I figured out that it isn't
> >>>>>> the code that does the relocation. Even with that completely disabled there were
> >>>>>> the vmalloc issues. What causes it seems to be the default of the upper limit
> >>>>>> and that this changes the split between kernel and modules to 1G+1G instead of
> >>>>>> 512M+1.5G. That is the reason why nokaslr has no effect.
> >>>>>
> >>>>> Oh! That's very interesting. There must be some assumption in Xen
> >>>>> about the kernel VM layout then?
> >>>>
> >>>> No. I think most of the changes that look at PTE and PMDs are are all
> >>>> in arch/x86/xen/mmu.c. I wonder if this is xen_cleanhighmap being
> >>>> too aggressive
> >>>
> >>> (Sorry I had to cut our chat short at Kernel Summit!)
> >>>
> >>> I sounded like there was another region of memory that Xen was setting
> >>> aside for page tables? But Stefan's investigation seems to show this
> >>> isn't about layout at boot (since the kaslr=0 case means no relocation
> >>> is done). Sounds more like the split between kernel and modules area,
> >>> so I'm not sure how the memory area after the initrd would be part of
> >>> this. What should next steps be, do you think?
> >>
> >> Maybe layout, but not about placement of the kernel. Basically leaving KASLR
> >> enabled but shrink the possible range back to the original kernel/module split
> >> is fine as well.
> >>
> >> I am bouncing between feeling close to understand to being confused. Konrad
> >> suggested xen_cleanhighmap being overly aggressive. But maybe its the other way
> >> round. The warning that occurs first indicates that PTE that was obtained for
> >> some vmalloc mapping is not unused (0) as it is expected. So it feels rather
> >> like some cleanup has *not* been done.
> >>
> >> Let me think aloud a bit... What seems to cause this, is the change of the
> >> kernel/module split from 512M:1.5G to 1G:1G (not exactly since there is 8M
> >> vsyscalls and 2M hole at the end). Which in vaddr terms means:
> >>
> >> Before:
> >> ffffffff80000000 - ffffffff9fffffff (=512 MB) kernel text mapping, from phys 0
> >> ffffffffa0000000 - ffffffffff5fffff (=1526 MB) module mapping space
> >>
> >> After:
> >> ffffffff80000000 - ffffffffbfffffff (=1024 MB) kernel text mapping, from phys 0
> >> ffffffffc0000000 - ffffffffff5fffff (=1014 MB) module mapping space
> >>
> >> Now, *if* I got this right, this means the kernel starts on a vaddr that is
> >> pointed at by:
> >>
> >> PGD[510]->PUD[510]->PMD[0]->PTE[0]
> >>
> >> In the old layout the module vaddr area would start in the same PUD area, but
> >> with the change the kernel would cover PUD[510] and the module vaddr + vsyscalls
> >> and the hole would cover PUD[511].
> >
> > I think there is a fixmap there too?
>
> Right, they forgot that in Documentation/x86/x86_64/mm... but head_64.S has it.
> So fixmap seems to be in the 2M space before the vsyscalls.
> Btw, apparently I got the PGD index wrong. It is of course 511, not 510.
>
> init_level4_pgt[511]->level3_kernel_pgt[510]->level2_kernel_pgt[0..255]->kernel
> [256..511]->mod
> [511]->level2_fixmap_pgt[0..505]->mod
> [506]->fixmap
> [507..510]->vsysc
> [511]->hole
>
> With the change being level2_kernel_pgt completely covering kernel only.
>
> >>
> >> xen_cleanhighmap operates only on the kernel_level2_pgt which (speculating a bit
> >> since I am not sure I understand enough details) I believe is the one PMD
> >> pointed at by PGD[510]->PUD[510]. That could mean that before the change
> >
> > That sounds right.
> >
> > I don't know if you saw:
> >
> > 1248 #ifdef DEBUG
> > 1249 /* This is superflous and is not neccessary, but you know what
> > 1250 * lets do it. The MODULES_VADDR -> MODULES_END should be clear of
> > 1251 * anything at this stage. */
> > 1252 xen_cleanhighmap(MODULES_VADDR, roundup(MODULES_VADDR, PUD_SIZE) - 1);
> > 1253 #endif
> > 1254 }
>
> I saw that but it would have no effect, even with running it. Because
> xen_cleanhighmap clamps the pmds it walks over to the kernel_level2_pgt page.
> Now MODULES_VADDR is mapped only from level2_fixmap_pgt.
> Even with the old layout it might do less that anticipated as it would only
> cover 512M and stop then. But I think it really does not matter.
> >
> > Which was me being a bit paranoid and figured it might help in troubleshooting.
> > If you disable that does it work?
> >
> >> xen_cleanhighmap may touch some (the initial 512M) of the module vaddr space but
> >> not after the change. Maybe that also means it always should have covered more
> >> but this would not be observed as long as modules would not claim more than
> >> 512M? I still need to check the vaddr ranges for which xen_cleanhighmap is
> >> actually called. The modules vaddr space would normally not be touched (only
> >> with DEBUG set). I moved that to be unconditionally done but then this might be
> >> of no use when it needs to cover a different PMD...
> >
> > What does the toolstack say in regards to allocating the memory? It is pretty
> > verbose (domainloginfo..something) in printing out the vaddr of where
> > it stashes the kernel, ramdisk, P2M, and the pagetables (which of course
> > need to fit all within the 512MB, now 1GB area).
>
> That is taken from starting a 2G PV domU with pvgrub (not pygrub):
>
> Xen Minimal OS!
> start_info: 0xd90000(VA)
> nr_pages: 0x80000
> shared_inf: 0xdfe92000(MA)
> pt_base: 0xd93000(VA)
> nr_pt_frames: 0xb
> mfn_list: 0x990000(VA)
> mod_start: 0x0(VA)
> mod_len: 0
> flags: 0x0
> cmd_line:
> stack: 0x94f860-0x96f860
> MM: Init
> _text: 0x0(VA)
> _etext: 0x6000d(VA)
> _erodata: 0x78000(VA)
> _edata: 0x80b00(VA)
> stack start: 0x94f860(VA)
> _end: 0x98fe68(VA)
> start_pfn: da1
> max_pfn: 80000
> Mapping memory range 0x1000000 - 0x80000000
> setting 0x0-0x78000 readonly
>
>
> For a moment I was puzzled by the use of max_pfn_mapped in the generic
> cleanup_highmap function of 64bit x86. It limits the cleanup to the start of the
> mfn_list. And the max_pfn_mapped value changes soon after to reflect the total
> amount of memory of the guest.
> Making a copy showed it to be around 51M at the time of cleanup. That initially
> looks suspect but Xen already replaced the page tables. The compile-time
> variants would have 2M large pages on the whole level2_kernel_pgt range. But as
> far as I can see, the Xen provided ones don't put in mappings for anything
> beyond the provided boot stack which is clean in the xen_cleanhighmap.
>
> So not much further... but then I think I know what I do next. Probably should
> have done before. I'll replace the WARN_ON in vmalloc that triggers by a panic
> and at least get a crash dump of that situation when it occurs. Then I can dig
> in there with crash (really should have thought of that before)...

<nods> I dug a bit in the code (arch/x86/xen/mmu.c) but there is nothing there
that screams at me, so I fear I will have to wait until you get the crash
and get some clues from that.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/