Messages in this thread | | | Date | Wed, 1 Jul 2009 21:07:58 +0800 | From | Herbert Xu <> | Subject | Re: [PATCH] net: fix race in the receive/select |
| |
On Wed, Jul 01, 2009 at 06:58:52PM +0800, Herbert Xu wrote: > > It would be good to know whether this bug occurs on the upstream > kernel because as it stands, the patch is a no-op for the upstream > kernel on x86-64.
I take that back. Andi, please look at section 8.2.3.4 of the IA-32 Architectures Software Developer's Manual Volume 3A, "Loads May Be Reordered with Earlier Stores to Different Locations. This seems to be exactly the scenario that we have here, and shows why mfence is required.
In our case, we're doing
CPU1 CPU2 Write data ready Add ourselves to wait queue Read wait queue to notify Check data ready
Cheers, -- Visit Openswan at http://www.openswan.org/ Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au> Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
| |