RE: [PATCH net-next 3/5] be2net: fix erx->rx_drops_no_frags wrap around

OK, thanks Eric. I'll re-spin this patch series ...

-Sathya
________________________________________
From: Eric Dumazet [eric.dumazet@gmail.com]
Sent: Monday, August 22, 2011 6:14 PM
To: Perla, Sathya
Cc: netdev@vger.kernel.org
Subject: RE: [PATCH net-next 3/5] be2net: fix erx->rx_drops_no_frags wrap around

Le lundi 22 août 2011 à 05:15 -0700, Sathya.Perla@Emulex.Com a écrit :

quoted

-----Original Message-----
From: Eric Dumazet [mailto:eric.dumazet@gmail.com]
Sent: Monday, August 22, 2011 5:18 PM

Le lundi 22 août 2011 à 13:43 +0200, Eric Dumazet a écrit :

quoted

This adds a race for lockless SNMP readers (in be_get_stats64())

They can see the 32bit value going backward.

You have to be very careful to read the *acc once, and write it once.

The "write once" is the only requirement.

Something like :

u32 newval = hi(*acc) + val;

if (wrapped)
    newval += 65536;
ACCESS_ONCE(*acc) = newval;

Eric,

I'm wondering why you'd need ACCESS_ONCE() here? Wouldn't using the temp variable (newval) as you suggested,
ensure that the reader doesn't see a half-baked value?

If you write :

u32 newval = hi(*acc) + val;
if (wrapped)
        newval += 65536;
*acc = newval;


C compiler (gcc) is free to eliminate the temp variable and to generate
same code than :

*acc = hi(*acc) + val;
if (wrapped)
        *acc += 65536;

ACCESS_ONCE() here is the clean way (and self explanatory/documented) to
keep compiler to write to *acc twice.