Re: [PATCH v3 1/3] exec: separate thread_count for files_struct

[Eric W. Biederman]

Oleg Nesterov <oleg@xxxxxxxxxx> writes:

> On 09/16, Eric W. Biederman wrote:
>>
>> Oleg Nesterov <oleg@xxxxxxxxxx> writes:
>>
>> > As for binder.c, in this case we probably actually want to unshare ->files
>> > on exec so we can ignore it?
>>
>> Looking at the binder case it only captures ->files on mmap.  Exec
>> ditches the mmap.  So if the order of operations are correct than
>> the dropping of the old mm will also drop the count on files_struct
>> held by binder.
>>
>> So semantically binder should not effect locks on exec,
>
> Agreed, but it does.
>
> Before your "[PATCH 0/3] exec: Moving unshare_files_struct" unshare_files()
> is called before exec_mmap().
>
> And even with this series we can have another CLONE_VM process.
>
> Howver, I think this doesn't really matter. binder does __fd_install(files),
> so if it actually has a reference to execing_task->files, I think it should
> be unshared anyway.
>
>> In short as long as we get the oder of operations correct we should be
>> able to safely ignore binder, and not have binder affect the results of
>> this code.
>
> Agreed.

I may have spoken too soon.  Binder uses schedule_work to call
put_files_struct from munmap.  So the files->count may still be elevated
after the mm is put.  Ick.

Eric