Messages in this thread | | | Date | Wed, 10 Nov 1999 23:55:32 +0100 (MET) | From | Gerard Roudier <> | Subject | Re: PATCH 2.3.26: kmalloc GFP_ZEROt |
| |
On Wed, 10 Nov 1999, Andrea Arcangeli wrote:
> On Wed, 10 Nov 1999, Gerard Roudier wrote: > > >zeroed with the cache invalidating or snooping the accesses. May-be having > >some pool of zeroed memory or zeroing some memory is useful for security > >issue, but speaking of pooled or cached objects, we probably want to most > > Yes, the _only_ point of the clear_page in the page fault path is > security.
Indeed.
> >only zero parts that need to be so. The bzero-mania does not look fine > >programming to me. > > Agreed. OTOH using a different chip for the bzero may really improve > performances and scale better.
But this add complexity for optimization of the epsilon, in my opinion. There is some other way to trash L1 cache as walking a too long list. Each indirection can trash at least one cache line and sometimes far more due to accessing data outside this first cache line in order to make decision.
> The bzero-mania with a minimal slowdown under heavy load still doesn't > convince me instead...
Zeroing a datum that will likely stay so is overhead, but when this zeroed datum will be filled by real information later, then the zeroing is close to zero-cost compared to the computation of this information and the change of the previous zero value. Unless you want to perform everything from kernel-land, being a bit smart or just not stupid regarding useless zeroing of data should make such a load negligible compared to the actual load of the system and user applications, in my opinion.
Fine design is far better at avoiding overhead that brute-force optimisation in my opinion.
> >PS: If we want to be able to write zero-filled blocks to a disk without > >bzero overhead, for example, we can reserve a small amount of physical > >pages filled with zeros at startup. > > Currently there's at least one page reserved that must stay zeroed all the > time. It's the ZERO_PAGE(). mapping /dev/null always fallback on the > ZERO_PAGE. you can read it as many times and whenever you want.
Who want to ever read data he already knows about the actual value ? ;-)
Gérard.
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |