On Fri, 1 Mar 2024 11:37:54 -0500
Mathieu Desnoyers <mathieu.desnoy...@efficios.com> wrote:
> On 2024-03-01 10:49, Steven Rostedt wrote:
> > On Fri, 1 Mar 2024 13:37:18 +0800
> > linke <lilink...@qq.com> wrote:
> >
> >>> So basically you are worried about read-tearing?
> >>>
> >>> That wasn't mentioned in the change log.
> >>
> >> Yes. Sorry for making this confused, I am not very familiar with this and
> >> still learning.
> >
> > No problem. We all have to learn this anyway.
> >
> >>
> >>> Funny part is, if the above timestamp read did a tear, then this would
> >>> definitely not match, and would return the correct value. That is, the
> >>> buffer is not empty because the only way for this to get corrupted is if
> >>> something is in the process of writing to it.
> >>
> >> I agree with you here.
> >>
> >> commit = rb_page_commit(commit_page);
> >>
> >> But if commit_page above is the result of a torn read, the commit field
> >> read by rb_page_commit() may not represent a valid value.
> >
> > But commit_page is a word length, and I will argue that any compiler that
> > tears "long" words is broken. ;-)
>
> [ For those tuning in, we are discussing ring_buffer_iter_empty()
> "commit_page = cpu_buffer->commit_page;" racy load. ]
>
> I counter-argue that real-world compilers *are* broken based on your
> personal definition, but we have to deal with them, as documented
> in Documentation/memory-barriers.txt (see below).
>
> What is the added overhead of using a READ_ONCE() there ? Why are
> we wasting effort trying to guess the compiler behavior if the
> real-world performance impact is insignificant ?
>
> Quote from memory-barrier.txt explaining the purpose of {READ,WRITE}_ONCE():
>
> "(*) For aligned memory locations whose size allows them to be accessed
> with a single memory-reference instruction, prevents "load tearing"
> and "store tearing," in which a single large access is replaced by
> multiple smaller accesses."
>
> I agree that {READ,WRITE}_ONCE() are really not needed at initialization,
> when there are demonstrably no concurrent accesses to the data
>
> But trying to eliminate {READ,WRITE}_ONCE() on concurrently accessed fields
> just adds complexity, prevents static analyzers to properly understand the
> code and report issues, and just obfuscates the code.
>
> Thanks,
>
> Mathieu
>
> >
> >>
> >> In this case, READ_ONCE() is only needed for the commit_page.
> >
> > But we can at least keep the READ_ONCE() on the commit_page just because it
> > is used in the next instruction.
>
And here I did state that READ_ONCE() does have another use case. So
there's no argument about adding it here.
-- Steve
