Message-ID: <87o9wv3gle.fsf@xmission.com>
Date:   Mon, 20 Mar 2017 14:15:09 -0500
From:   ebiederm@...ssion.com (Eric W. Biederman)
To:     Petr Tesarik <ptesarik@...e.cz>
Cc:     Xunlei Pang <xpang@...hat.com>, Baoquan He <bhe@...hat.com>,
        kexec@...ts.infradead.org, xlpang@...hat.com,
        linux-kernel@...r.kernel.org, akpm@...ux-foundation.org,
        Dave Young <dyoung@...hat.com>
Subject: Re: [PATCH] kexec: Update vmcoreinfo after crash happened

Petr Tesarik <ptesarik@...e.cz> writes:

> On Mon, 20 Mar 2017 10:17:42 +0800
> Xunlei Pang <xpang@...hat.com> wrote:
>
>> On 03/19/2017 at 02:23 AM, Petr Tesarik wrote:
>> > On Thu, 16 Mar 2017 21:40:58 +0800
>> > Xunlei Pang <xpang@...hat.com> wrote:
>> >
>> >> On 03/16/2017 at 09:18 PM, Baoquan He wrote:
>> >>> On 03/16/17 at 08:36pm, Xunlei Pang wrote:
>> >>>> On 03/16/2017 at 08:27 PM, Baoquan He wrote:
>> >>>>> Hi Xunlei,
>> >>>>>
>> >>>>> Did you really see this ever happened? Because the vmcore size estimate
>> >>>>> feature, namely --mem-usage option of makedumpfile, depends on the
>> >>>>> vmcoreinfo in 1st kernel, your change will break it.
>> >>>> Hi Baoquan,
>> >>>>
>> >>>> I can reproduce it using a kernel module which modifies the vmcoreinfo,
>> >>>> so it's a problem can actually happen.
>> >>>>
>> >>>>> If not, it could be not good to change that.
>> >>>> That's a good point, then I guess we can keep the crash_save_vmcoreinfo_init(),
>> >>>> and store again all the vmcoreinfo after crash. What do you think?
>> >>> Well, then it will make makedumpfile segfault happen too when execute
>> >>> below command in 1st kernel if it existed:
>> >>> 	makedumpfile --mem-usage /proc/kcore
>> >> Yes, if the initial vmcoreinfo data was modified before "makedumpfile --mem-usage", it might happen,
>> >> after all the system is going something wrong. And that's why we deploy kdump service at the very
>> >> beginning when the system has a low possibility of going wrong.
>> >>
>> >> But we have to guarantee kdump vmcore can be generated correctly as possible as it can.
>> >>
>> >>> So we still need to face that problem and need fix it. vmcoreinfo_note
>> >>> is in kernel data area, how does module intrude into this area? And can
>> >>> we fix the module code?
>> >>>
>> >> Bugs always exist in products, we can't know what will happen and fix all the errors,
>> >> that's why we need kdump.
>> >>
>> >> I think the following update should guarantee the correct vmcoreinfo for kdump.
>> > I'm still not convinced. I would probably have more trust in a clean
>> > kernel (after boot) than a kernel that has already crashed (presumably
>> > because of a serious bug). How can be reliability improved by running
>> > more code in unsafe environment?
>> 
>> Correct, I realized that, so used crc32 to protect the original data,
>> but since Eric left a more reasonable idea, I will try that later.
>> 
>> >
>> > If some code overwrites reserved areas (such as vmcoreinfo), then it's
>> > seriously buggy. And in my opinion, it is more difficult to identify
>> > such bugs if they are masked by re-initializing vmcoreinfo after crash.
>> > In fact, if makedumpfile in the kexec'ed kernel complains that it
>> > didn't find valid VMCOREINFO content, that's already a hint.
>> >
>> > As a side note, if you're debugging a vmcoreinfo corruption, it's
>> > possible to use a standalone VMCOREINFO file with makedumpfile, so you
>> > can pre-generate it and save it in the kdump initrd.
>> >
>> > In short, I don't see a compelling case for this change.
>> 
>> E.g. 1) wrong code overwrites vmcoreinfo_data; 2) further crashes the
>> system; 3) trigger kdump, then we obviously will fail to recognize the
>> crash context correctly due to the corrupted vmcoreinfo.  Everyone
>> will get confused if met such unfortunate customer-side issue.
>> 
>> Although it's corner case, if it's easy to fix, then I think we better do it.
>> 
>> Now except for vmcoreinfo, all the crash data is well protected (including
>> cpu note which is fully updated in the crash path, thus its correctness is
>> guaranteed).
>
> Hm, I think we shouldn't combine the two things.
>
> Protecting VMCOREINFO with SHA (just as the other information passed to
> the secondary kernel) sounds right to me. Re-creating the info while
> the kernel is already crashing does not sound particularly good.
>
> Yes, your patch may help in some scenarios, but in general it also
> increases the amount of code that must reliably work in a crashed
> environment. I can still recall why the LKCD approach (save the dump
> directly from the crashed kernel) was abandoned...
>
> Apart, there's a lot of other information that might be corrupted (e.g.
> the purgatory code, elfcorehdr, secondary kernel, or the initrd).
>
> Why is this VMCOREINFO so special?

Petr I generally agree with you.  We need to minimise what happens after
a panic.

I don't know if you saw my comment on the v2 of this patchset, but the
core issue I saw with VMCOREINFO is that the data appears to be stored
in the .bss (of the kernel that will crash) instead of in the pages
managed by kexec.  Which means they are not protected like other pages.

That is the issue I have asked to be addressed first.

I also agree that we very much want to minimise what happens during
crash.  Which I suspect means that we will want to remove the CRASH_TIME
variable from VMCOREINFO.  I may be wrong but I think that is probably
enough to make the strings a constant.

Eric

Powered by blists - more mailing lists