Re: [RFC 1/3] abi_spec: basic definitions of constraints, args and syscalls

[Dmitry Vyukov]

On Mon, Dec 12, 2016 at 11:45 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> On Mon, Dec 12, 2016 at 11:29 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>> On Wed, Nov 23, 2016 at 3:59 PM,  <alexander.levin@xxxxxxxxxxx> wrote:
>>> On Mon, Nov 21, 2016 at 03:48:17PM +0100, Dmitry Vyukov wrote:
>>>> Several observations based on my experience with syzkaller descriptions:
>>>>  - there are 2 levels: physical and logical;
>>>>    on physical level there are int, pointer, array, struct, union;
>>>>    and that's pretty much it.
>>>>    on logical level there are flags, bitmasks, file paths, sctp socket fds,
>>>>    unix socket names, etc.
>>>>    These levels are almost completely orthogonal. It would be useful to
>>>>    clearly separate them on description level. E.g. now you have TYPE_PTR and
>>>>    TYPE_INT which is physical level; and then TYPE_FD which is also an int.
>>>>
>>>>  - logical types won't fit into 64 bits, there are more of them
>>>
>>> I agree with your two points above.
>>>
>>> As an example, let's look at:
>>>
>>>         int epoll_ctl(int epfd, int op, int fd, struct epoll_event *event);
>>>
>>> epfd would be a physical int, logical "epoll_fd", and constrainted with being
>>> an open descriptor.
>>>
>>> fd on the other hand is tricky: it's a physical int, logical fd, and
>>> constrainted with being a file descriptor that supports poll, and being
>>> open.
>>>
>>> So while I think that logical types can be just a counter rather than a bitmask
>>> I suspect that our constraints won't fit into 64 bits. Make 2 64 bit fields?
>>
>> One observation is that there are just 5 physical types:
>>  - scalar
>>  - pointer
>>  - array
>>  - struct
>>  - union
>>
>> The rest deals with what exactly "scalar" is in a particular case.
>>
>> I don't yet have complete answer, as it somewhat intermixed with the
>> rest of questions.
>>
>>
>>
>>>>  - we need support for recursive types (yes, there are linked lists in
>>>> kernel APIs)
>>>
>>> I imagine that this will be handled by specific logical type handlers we'll
>>> need to implement. Can you give me an example and I'll try to code that?
>>
>> One example is te_oper_param here:
>> https://android.googlesource.com/kernel/tegra/+/android-tegra-3.10/security/tlk_driver/ote_protocol.h
>> next_ptr_user is a pointer to te_oper_param. Thus recursive definition.
>>
>> Another example is snd_seq_ev_quote:
>> http://lxr.free-electrons.com/source/include/uapi/sound/asequencer.h#L194
>> it contains struct snd_seq_event *event and snd_seq_event recursively
>> contains snd_seq_ev_quote.
>>
>> In all cases it is pointer recursion via structs.
>>
>> Sometimes it wish that developers have to write formal descriptions in
>> a limited language upfront. That would probably eliminate lots of
>> weird one-off "see what I invented here" cases :)
>>
>>
>>
>>
>>>>  - we need support for input/output data
>>>>    currently syzkaller does this only on pointer level, i.e. you
>>>> attach direction to pointer target
>>>>    but that's not enough, frequently there is a struct where one field
>>>> is input and another is output
>>>
>>> Assuming it's "data", for intput we'll just need to check that the given
>>> length is readable and for output that the length is writable, no?
>>
>> It also can be an fd in a struct field. If it's output (e.g. pipe),
>> then we must not check that it's valid on entry. But we may want to
>> check that it's valid on successful exit, or fuzzer will use these
>> output fd's as inputs to other calls.
>>
>>
>>> We can do it with constraints right now.
>>>
>>>>  - we may need support for reusing types in several arguments
>>>>    e.g. you may have a pretty complex type, and you don't want to
>>>> write it out a dozen of times
>>>
>>> Yup, so if we go with the physical/logical split we can have handlers for
>>> logical types.
>>>
>>>>  - we need some support for discriminated syscalls
>>>>    if we want to support strace usecase, the support needs to be more
>>>> extensive than what syzkaller has;
>>>>    i.e. syzkaller can't restore discrimination having actual argument
>>>> values (it can do it only in the other direction)
>>>>
>>>>  - I would not create a special support for arguments;
>>>>    rather I would create support for structs and struct fields,
>>>>    and then pretend that a syscalls effectively accepts a struct by value
>>>
>>> But that means I need a custom handler for every syscall to parse the
>>> struct fields rather than a generic code that goes through the args and calls
>>> the right handler?
>>
>> No, you don't. We will need generic code that parses a piece of memory
>> as a struct and splits it into fields anyway.
>> We can just reuse this code to handle syscall arguments as follows.
>> Describe syscall arguments as a pseudo struct (array of fields). Then
>> syscall handling function accepts pointer to region of memory with
>> arguments and description of the struct, and invokes common struct
>> handling code.
>>
>>
>>
>>>> How would you like us to collaborate on this?
>>>> If you share your git repo, I could form it into something that would
>>>> be suitable for syzkaller and incorporate most of the above.
>>>
>>> I'd really like to have something that either generates these descriptions from
>>> your DSL (it really doesn't have to be perfect (at first)) or something that
>>> generates DSL from these C structs.
>>
>> Do you mean generating C from my DSL of a one-off or as a permanent solution?
>
>
> Main problem I am trying to resolve now is how to make types reference
> other types.
> Say we have "pointer to pointer to int" (the same applies to structs
> and arrays). This means that struct type must be able to reference
> another instance of struct type. But we can't put type into type by
> value. Namely, the following is not possible:
>
> struct type {
>   int kind;
>   union {
>     ...
>     // for kind = KIND_PTR
>     struct {
>       struct type type;
>     } ptr;
>   };
> };
>
> One obvious solution would be to always reference _pointers_ to types. E.g.:
>
>     // for kind = KIND_PTR
>     struct {
>       struct type* type;
>     } ptr;
>
> This works but leads to super-verbose descriptions (if we want to use
> static tables), because we need to describe type of each syscall
> argument and inner types for all pointers/arrays/structs/unions as
> separate global variables:
>
>
> static struct type open_arg0_inner {
>   .kind = KIND_ARRAY,
>   ...
> };
>
> static struct type open_arg0 {
>   .kind = KIND_PTR,
>   .ptr.type = &open_arg0_inner
>   ...
> };
>
> static struct type open_arg1 {
>   .kind = KIND_SCALAR,
>   ...
> };
>
> static struct type open_arg2 {
>   .kind = KIND_SCALAR,
>   ...
> };
>
> static struct type open_ret {
>   .kind = KIND_SCALAR,
>   ...
> };
>
> struct syscall_spec syscall_spec_open = {
>   .name = "open",
>   .retval = {
>     .name = "retval",
>     .type = &open_ret,
>    }
>   .nargs = 3,
>   .args[0] = {
>     .name = "pathname",
>     .type = &open_arg0,
>   }
>   .args[1] = {
>     .name = "flags",
>     .type = &open_arg1,
>   }
>   .args[2] = {
>     .name = "mode",
>     .type = &open_arg2,
>   }
> };
>
>
> This looks way too verbose provided that we write these descriptions
> by hand (just coming up with consistent unique names for all these
> global vars is a problem).
> If we generate that from DSL, it can work though.
>
> Another option I am considering is to use helper construction
> functions that return pointers to types, e.g. something along these
> lines:
>
> struct syscall_spec syscall_spec_open = {
>   .name = "open",
>   .retval = {
>     .name = "retval",
>     .type = type_scalar(FD | ERRNO),
>    }
>   .nargs = 3,
>   .args[0] = {
>     .name = "pathname",
>     .type = type_ptr(DIR_IN, type_array(PATHNAME)),
>   }
>   .args[1] = {
>     .name = "flags",
>     .type = type_scalar_flags(O_RDONLY | O_WRONLY, ....),
>   }
>   .args[2] = {
>     .name = "mode",
>     .type = type_scalar_flags(S_IRWXU ....),
>   }
> };
>
> Now these table can only be initialized dynamically. And we will
> probably need lots of these helper functions. So I don't like it
> either...
>
> Any suggestions?



Here is my current prototype:
https://github.com/dvyukov/linux/commit/6200a9333e78bef393f8ead41205813b94d340f3

For now it can trace arguments of 4 system calls:

[    4.055483] [pid 1258] open(&00007ffdefc023a0=[], 0x0, 0x1b6)
[    4.055991] [pid 1258] open(&00007ffdefc023a0=[], 0x0, 0x1b6) = 3
[    4.056486] [pid 1258] read(0x3, &00007ffdefc01320=[], 0x1000)
[    4.056977] [pid 1258] read(0x3, &00007ffdefc01320=[], 0x1000) = 1780
[    4.057485] [pid 1258] read(0x3, &00007f316a732000=[], 0x1000)
[    4.057991] [pid 1258] read(0x3, &00007f316a732000=[], 0x1000) = 0
[    4.058488] [pid 1258] close(0x0) = 0
[    4.058848] [pid 1258] write(0x1, &00007f316a732000=[], 0x5)
[    4.059304] [pid 1258] write(0x1, &00007f316a732000=[], 0x5) = 5
[    4.059905] [pid 1234] close(0x0) = 0
[    4.060239] [pid 1234] close(0x0) = 0


Main outstanding problems:
 - understanding length of arrays and buffers
 - understanding discriminators of unions and syscall variations