Message-ID: <577CC058.9030103@huawei.com>
Date:	Wed, 6 Jul 2016 16:24:56 +0800
From:	"Zhangjian (Bamvor)" <bamvor.zhangjian@...wei.com>
To:	Dmitry Vyukov <dvyukov@...gle.com>,
	syzkaller <syzkaller@...glegroups.com>
CC:	LKML <linux-kernel@...r.kernel.org>, <linux-arch@...r.kernel.org>,
	<libc-alpha@...rceware.org>, <trinity@...r.kernel.org>,
	<aponomarenko@...alab.ru>, Jess Hertz <jesse.hertz@...group.trust>,
	"Tim Newsham" <tim.newsham@...group.trust>,
	Arnd Bergmann <arnd@...db.de>,
	Catalin Marinas <catalin.marinas@....com>,
	Mark Brown <broonie@...nel.org>, <joseph@...esourcery.com>,
	<maxim.kuvyrkov@...aro.org>,
	Yury Norov <ynorov@...iumnetworks.com>, <pinskia@...il.com>,
	<schwab@...e.de>, <agraf@...e.de>, <marcus.shawcroft@....com>,
	Ding Tianhong <dingtianhong@...wei.com>,
	<guohanjun@...wei.com>, <cuibixuan@...wei.com>,
	<lijinyue@...wei.com>, Zefan Li <lizefan@...wei.com>
Subject: Re: [RFD] Efficient unit test and fuzz tools for kernel/libc porting

Hi, Dmitry

On 2016/7/6 16:00, Dmitry Vyukov wrote:
> On Wed, Jul 6, 2016 at 9:39 AM, Zhangjian (Bamvor)
> <bamvor.zhangjian@...wei.com> wrote:
>> HI,
>>
>> When I working on the ILP32 ABI for ARMv8 in last two years, I has
>> encountered lots of syscall issues such as wrong number of arguments,
>> different data type in binary interface. I realized that the correctness of
>> argument passing between the C library and core kernel code is a common
>> problem when bringing up new architecture or ABI to kernel and libc.
>> Existing fuzz testing tools such as trinity[1], syzkaller[2] and triforce[3]
>> only generate random or boundary values for syscall parameters and then
>> inject them into kernel, but those tools won't validate if the results of
>> those syscalls are correct or not. Thus they can not act as an unit test for
>> ILP32. In this year, considering the abi of ILP32 is changes during
>> discussion, I am thinking if I could use some sort of automatically tools to
>> check whether the wrapper is correct or not. After learn and compare some
>> fuzz tools, I feel that there is no such fuzz tools could help me. So, I
>> wrote a new fuzz tools base on the trinity and it found several wrapper
>> issues in glibc. I will first explain the different with existing fuzz tools
>> and paste my propsosal in the end.
>>
>> Trinity is developed in a long time. It could randomize the parameter of
>> syscall and run individual syscall standalone or parallel. When I do the
>> long time parallel test(not for ILP32), it could report some bug, e.g. hang,
>> panic. It is useful but it is indeed hard to debug because it usually fail
>> after a long time running. I do not know what does it exactly do.
>>
>> Compare with Trinity, syzkaller is quite different. Here is the comparision
>> between syzkaller and our tools:
>> 1.  Syzkaller could recursively randomize base date type in syscall which
>> means it is possible generate more meaningfull syscall test. But it only
>> test the syscall through syscall() function. It assume that the c library is
>> correct and stable. But it is wrong if we are porting new abi(such as ILP32)
>> or architecture to glibc and kernel. We need to take c library into account.
>> This is what my tools could do.
>>
>> 2.  Syzkaller could generate the readable short testcases. Our tools could
>> only test individual syscall and check the correctness of parameter and
>> return value. I think it is enough for the unit test which tests syscall one
>> by one.
>>
>> 3.  Syzkaller could do the coverage. Our tools could not. I think it is
>> useful for me. I plan to add the coverage later.
>>
>> In my ILP32 works, my tools reported several off_t endian issues in glibc.
>> Our tools work like this:
>>
>> Dump the function                                 Dump the function
>> prototype from                                    prototype from c
>> vmlinux from the                                  library from the
>> sys_call_table                                    given list(posix
>> array in kernel.                                  interfaces or user
>>         |                                          defined).
>>         |                                                 |
>>         |                                                 |
>>        \|/                                               \|/
>>         `                                                 `
>> Generate jprobe        Modity Trinity to          Generate struct
>> hook according to      support run syscall        fuzz generator
>> prototype which        syscall from c             from the prototype.
>> will recursively       libray instead             And add them of
>> print the syscall      syscall() function         to trinity. Trinity
>> value.                       |                    will recursively
>>         \                     |                    print the function
>>          \                    |                    parameter.
>>           \                   |                           /
>>            -----------------------------------------------
>>                               |
>>                              \|/
>>                               `
>>                Run the trinity each syscall once
>>                and compare the function parameter
>>                printed in kernel and userspace
>>                If inconsistent, print specific
>>                information, such endian issue,
>>                32<->64bit conversion issue and
>>                so on.
>>
>> Tools of function dump and hook generator based on the abi-dumper[4]. Other
>> functions base on trinity. Return value test is similar except generating
>> the kretprobe hook instead jprobe hook.
>>
>> There are some hacks and the original funtion of trinity may be broken. The
>> main changes in trinity are as follows:
>> 1.  Call syscall through c library via call_glibc_syscalls() instead of
>> direct syscall via syscall().
>> 2.  Add new file generate-struct.c including the missing data type mentioned
>> in syscall. This file is generated by ./struct_extract.py with a little
>> modification. It should be fully auto generated in future.
>> 3.  Add more date types in fill_arg()(generate-args.c) and
>> include/syscall.h.
>> 4.  Modify the syscallentry struct in syscalls directory according to the
>> newly added data types.
>> 5.  Add or Change some output message for script.
>> 6.  Add jprobe hooks in modules directory. Such hook will be inserted before
>> trinity test and removed after test.
>>
>> Step to generate generate-struct.c and jprobe hooks:
>> 1.  Compile the kernel/libc/binary which include the functions you want to
>> generate.
>> 2.  Dump function and struct information through modified abi-dumper.pl,
>> named them as symbolinfo and typeinfo. You may need add "--all" and/or
>> "--dump-static" depends on your binaries.
>> 3.  Generate the file through trinity/scripts/struct_extract.py.
>>
>> We also submit a proposal to linuxcon europe last month. Hope we could get a
>> chance to share our works face to face.
>>
>> The patches could be found at [5] and [6]. It works but ugly. We hope we
>> could get some early response, such as if community is interested in this
>> tools, whether this tools could merge into existence fuzz tools.
>
>
> Hi Bamvor,
>
> Nice work!
>
> Coverage should be easy to do with CONFIG_KCOV, but do you need
> fuzzing/coverage? It seems that testing a predefined set of special
> values for each arg should be enough for your use case. Namely special
> values that can detect endianess/truncation/sign extension/etc issues.
Yes. We are trying to cover endianess/truncation/sign extension at this
moment.
For coverage, there are some code path in syscall wrapper in both glibc
and kernel. E.g. overflow check in glibc. I am thinking if coverage
could help on this.
>
> I think there is also a number of glibc functions that don't directly
> map to syscalls. Most notably wrappers around various ioctl's (e.g.
> ptsname). Do you test them?
No. Currently, our tools only focus on the syscall function in glibc. In
these syscall level, we could compare the parameter and return value
directly. As you said, there are only several type of issues. It is easy
to handle by tools.

I do not know how to test these complex cases. E.g. the ptsname may call
ioctl, *stat* syscall. Compare the original parameter is meaningless. But
it seems a good type of testcase to show how the user use the syscalls.
Do you have some ideas?

Regards

Bamvor

> --
> To unsubscribe from this list: send the line "unsubscribe linux-arch" in
> the body of a message to majordomo@...r.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>

Powered by blists - more mailing lists