Re: [PATCH v7 00/11] extend task comm from 16 to 24

[Petr Mladek]

On Mon 2021-11-01 22:34:30, Yafang Shao wrote:
> On Mon, Nov 1, 2021 at 10:07 PM Petr Mladek <pmladek@xxxxxxxx> wrote:
> > On Mon 2021-11-01 06:04:08, Yafang Shao wrote:
> > > 4. Print a warning if the kthread comm is still truncated.
> > >
> > > 5. What will happen to the out-of-tree tools after this change?
> > >    If the tool get task comm through kernel API, for example prctl(2),
> > >    bpf_get_current_comm() and etc, then it doesn't matter how large the
> > >    user buffer is, because it will always get a string with a nul
> > >    terminator. While if it gets the task comm through direct string copy,
> > >    the user tool must make sure the copied string has a nul terminator
> > >    itself. As TASK_COMM_LEN is not exposed to userspace, there's no
> > >    reason that it must require a fixed-size task comm.
> >
> > The amount of code that has to be updated is really high. I am pretty
> > sure that there are more potential buffer overflows left.
> >
> > You did not commented on the concerns in the thread
> > https://lore.kernel.org/all/CAADnVQKm0Ljj-w5PbkAu1ugLFnZRRPt-Vk-J7AhXxDD5xVompA@xxxxxxxxxxxxxx/
> >
> I thought Steven[1] and  Kees[2] have already clearly explained why we
> do it like that, so I didn't give any more words on it.
> 
> [1]. https://lore.kernel.org/all/20211025170503.59830a43@xxxxxxxxxxxxxxxxxx/

Steven was against switching task->comm[16] into a dynamically
allocated pointer. But he was not against storing longer names
separately.

> [2]. https://lore.kernel.org/all/202110251406.56F87A3522@keescook/

Honestly, I am a bit confused by Kees' answer. IMHO, he agreed that
switching task->comm[16] into a pointer was not worth it.

But I am not sure what he meant by "Agreed -- this is a small change
for what is already an "uncommon" corner case."


> > Several people suggested to use a more conservative approach.
> 
> Yes, they are Al[3] and Alexei[4].
> 
> [3]. https://lore.kernel.org/lkml/YVkmaSUxbg%2FJtBHb@xxxxxxxxxxxxxxxxxxxxx/

IMHO, Al suggested to store the long name separately and return it
by proc_task_name() when available.


> [4]. https://lore.kernel.org/all/CAADnVQKm0Ljj-w5PbkAu1ugLFnZRRPt-Vk-J7AhXxDD5xVompA@xxxxxxxxxxxxxx/

Alexei used dentry->d_iname as an exaxmple. struct dentry uses
d_iname[DNAME_INLINE_LEN] for short names. And dynamically
allocated d_name for long names, see *__d_alloc() implementation.

> > I mean
> > to keep comm[16] as is and add a new pointer to the full name. The buffer
> > for the long name might be dynamically allocated only when needed.
> >
> 
> That would add a new allocation in the fork() for the threads with a long name.
> I'm not sure if it is worth it.

The allocation will be done only when needed. IMHO, the performance is
important only for userspace processes. I am not aware of any kernel
subsystem that would heavily create and destroy kthreads.


> > The pointer might be either in task_struct or struct kthread. It might
> > be used the same way as the full name stored by workqueue kthreads.
> >
> 
> If we decide to do it like that, I think we'd better add it in
> task_struct, then it will work for all tasks.

Is it really needed for userspace processes? For example, ps shows
the information from /proc/*/cmdline instead.


> > The advantage of the separate pointer:
> >
> >    + would work for names longer than 32
> >    + will not open security holes in code
> >
> 
> Yes, those are the advantages.  And the disadvantage of it is:
> 
>  - new allocation in fork()

It should not be a problem if we do it only when necessary and only
for kthreads.

Best Regards,
Petr