lkml.org 
[lkml]   [2021]   [Jan]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [RFC PATCH v3 1/8] Use refcount_t for ucounts reference counting
ebiederm@xmission.com (Eric W. Biederman) writes:

> Alexey Gladkov <gladkov.alexey@gmail.com> writes:
>
>> On Mon, Jan 18, 2021 at 12:34:29PM -0800, Linus Torvalds wrote:
>>> On Mon, Jan 18, 2021 at 11:46 AM Alexey Gladkov
>>> <gladkov.alexey@gmail.com> wrote:
>>> >
>>> > Sorry about that. I thought that this code is not needed when switching
>>> > from int to refcount_t. I was wrong.
>>>
>>> Well, you _may_ be right. I personally didn't check how the return
>>> value is used.
>>>
>>> I only reacted to "it certainly _may_ be used, and there is absolutely
>>> no comment anywhere about why it wouldn't matter".
>>
>> I have not found examples where checked the overflow after calling
>> refcount_inc/refcount_add.
>>
>> For example in kernel/fork.c:2298 :
>>
>> current->signal->nr_threads++;
>> atomic_inc(&current->signal->live);
>> refcount_inc(&current->signal->sigcnt);
>>
>> $ semind search signal_struct.sigcnt
>> def include/linux/sched/signal.h:83 refcount_t sigcnt;
>> m-- kernel/fork.c:723 put_signal_struct if (refcount_dec_and_test(&sig->sigcnt))
>> m-- kernel/fork.c:1571 copy_signal refcount_set(&sig->sigcnt, 1);
>> m-- kernel/fork.c:2298 copy_process refcount_inc(&current->signal->sigcnt);
>>
>> It seems to me that the only way is to use __refcount_inc and then compare
>> the old value with REFCOUNT_MAX
>>
>> Since I have not seen examples of such checks, I thought that this is
>> acceptable. Sorry once again. I have not tried to hide these changes.
>
> The current ucount code does check for overflow and fails the increment
> in every case.
>
> So arguably it will be a regression and inferior error handling behavior
> if the code switches to the ``better'' refcount_t data structure.
>
> I originally didn't use refcount_t because silently saturating and not
> bothering to handle the error makes me uncomfortable.
>
> Not having to acquire the ucounts_lock every time seems nice. Perhaps
> the path forward would be to start with stupid/correct code that always
> takes the ucounts_lock for every increment of ucounts->count, that is
> later replaced with something more optimal.
>
> Not impacting performance in the non-namespace cases and having good
> performance in the other cases is a fundamental requirement of merging
> code like this.

So starting with something easy to comprehend and simple, may make it
easier to figure out how to optimize the code.

Eric

\
 
 \ /
  Last update: 2021-01-20 03:02    [W:0.091 / U:0.768 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site