svn commit: r367813 - head/lib/libutil

[Jessica Clarke]

On 18 Nov 2020, at 21:52, Stefan Esser <se at freebsd.org> wrote:
> Am 18.11.20 um 22:15 schrieb Jessica Clarke:
>> On 18 Nov 2020, at 19:44, Stefan Eßer <se at freebsd.org> wrote:
>>> +		/*
>>> +		 * Check for some other thread already having
>>> +		 * set localbase - this should use atomic ops.
>>> +		 * The amount of memory allocated above may leak,
>>> +		 * if a parallel update in another thread is not
>>> +		 * detected and the non-NULL pointer is overwritten.
>>> +		 */
>> Why was this committed with a known racy/leaky implementation?
> 
> Because the alternatives that I offered for discussion were
> less acceptable.

That has no bearing over whether this one is.

>> What happens if I set the value with a sysctl and call this?
> 
> You'll get the value set with sysctl, unless overridden by the
> environment variable. There is a window of a few nano-seconds
> where a thread executing in parallel on another core might be
> able to set the localbase variable (between the test for NULL
> in this function and the assignment to it). The value that will
> be returned by either thread will be identical, so there is no
> risk of corruption of the result.

But if I call getlocalbase, then set it via sysctl, then call
getlocalbase again, I see the old value. If, however, I omit the first
getlocalbase, I see the new value. This differs from how getenv/setenv
of the value work, where you always see the up-to-date value. Maybe you
think that's a feature, but it's something to watch out for and
explicitly call out in the documentation.

You also misunderstand all the subtleties of multithreading here. There
are no acquire/release pairs so it is entirely legal for Thread 2 to
read Thread 1's initialised value for localbase before the contents of
it are visible (i.e. the pointer is initialised but the data is
garbage).

The `(volatile const char*)localbase` cast is also a complete waste of
time. You probably meant to write `(const char * volatile)localbase`
but even then that does nothing useful as the cast is too late. What
you really were trying to write was
`*(const char * volatile *)&localbase`, but you need proper atomics
anyway for this to be safe.

> This unlikely case may actually leak a heap allocated string
> of typically tens of bytes (but with negligible probability).
> 
> But this really is a non-issue, since there should never be a
> reason to invoke this function in a multi-threaded context.

Why not? There could easily be code out there calling getenv in a
multi-threaded context so this is inadequate as a replacement. Yes it's
inefficient but it's perfectly legal and imaginable.

Also if malloc returns NULL I would quite like that to be an error for
the function and not silently fall back on _PATH_LOCALBASE.

> The result should be constant for the duration of execution
> of the process (expect severe inconsistencies if that was not
> the case) and all programs in base that are candidates for the
> use of this function are non-threaded (and if they were multi-
> threaded, then I'd expect this call to occur during start-up
> of the program before any further threads are created).
> 
> So, this is a non-issue and the comment tries to explain it.
> Did I fail to make this clear in the comment? Maybe I should
> have written "could use atomic ops" instead?
> 
> Use of atomics or locks could prevent the race-condition. But
> since I do not expect this function to be called from within
> threads (it just doesn't make sense), the tiny time window of
> a few nano-seconds which might lead to a double assignment to
> the target variable (with one pointer value being lost), and
> the worst case loss of 1KB of heap space in that case (more
> likely 10 to 20 bytes rounded up to a 16 or 32 byte chunk), I
> do not consider the complexities of either a lock or atomic ops
> to be justified.
> 
> Regards, STefan
>