Re: [take25 1/6] kevent: Description.

[Evgeniy Polyakov <johnpol@2ka.mipt.ru>]

On Thu, Nov 23, 2006 at 12:00:45PM -0800, Ulrich Drepper (drepper@redhat.com) wrote:
> Evgeniy Polyakov wrote:
> >uidx is an index, starting from which there are unread entries. It is
> >updated by userspace when it commits entries, so it is 'consumer'
> >pointer, while kidx is an index where kernel will put new entries, i.e.
> >'producer' index. We definitely need them both.
> >Userspace can only update (implicitly by calling kevent_commit()) uidx.
> 
> Right, which is why exporting this entry is not needed.  Keep the 
> interface as small as possible.

If there are several callers of kevent_commit(), uidx can be changed far
than first user expects, so there should be possibility to check that
value. It is thus exported into shared ring buffer structure.

> Userlevel has to maintain its own index.  Just assume kevent_wait 
> returns 10 new entries and you have multiple threads.  In this case all 
> threads take their turns and pick an entry from the ring buffer.  This 
> basically has to be done with something like this (I ignore wrap-arounds 
> here to simplify the example):
> 
>   int getidx() {
>     while (uidx < kidx)
>        if (atomic_cmpxchg(uidx, uidx + 1, uidx) == 0)
>          return uidx;
>     return -1;
>   }
> 
> Very much simplified but it should show that we need a writable copy of 
> the uidx.  And this value at any time must be consistent with the index 
> the kernel assumes.

I seriously doubt it is simpler than having index provided by kernel.

> The current ring_uidx value can at best be used to reinitialize the 
> userlevel uidx value after each kevent_wait call but this is unnecessary 
> at best (since uidx must already have this value) and racy in problem 
> cases (what if more than one thread gets woken concurrently with uidx 
> having the same value and one thread stores the uidx value and 
> immediately increments it to get an index; the second store would 
> overwrite the increment).
> 
> I can assure you that any implementation I write would not use the 
> ring_uidx value.  Only trivial, single-threaded examples like you 
> ring_buffer.c could ever take advantage of this value.  It's not worth it.

You propose to make uidx shared local variable - it is doable, but it
is not required - userspace can use kernel's variable, since it is
updated exactly in the places where that index is changed.

> -- 
> ➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, 
> CA ❖
> -
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

-- 
	Evgeniy Polyakov