Re: [PATCH 3/3] x86, extable: Handle early exceptions

[David Daney <ddaney.cavm@gmail.com>]

On 04/19/2012 11:55 AM, H. Peter Anvin wrote:
> Either way I suggest picking up David's presorting patchset since it is already done and use its infrastructure for any further improvements.
>

It does have the advantage of already being implemented.  There was a 
little feedback on the kbuild portions of the patch.

If you would like, I will send an updated version of the patch.

> As far as a linear probe you get an average of n lookups with a packing density of 1-1/n so you are right; a linear probe with a density of say 1/2 is probably best.
>

I usually see exception table sizes on the order of 2^10 entries, so I 
have to wonder how much you really gain from an O(1) implementation.

David Daney

> Linus Torvalds<torvalds@linux-foundation.org>  wrote:
>
>> On Thu, Apr 19, 2012 at 10:59 AM, H. Peter Anvin<hpa@zytor.com>  wrote:
>>>
>>> I would argue that the O(1) hash makes things simpler as there is no
>>> need to deal with collisions at all.
>>
>> Most of the O(1) hashes I have seen more than made up for the trivial
>> complexity of a few linear lookups by making the hash function way
>> more complicated.
>>
>> A linear probe with a step of one really is pretty simple. Sure, you
>> might want to make the initial hash "good enough" to not often hit the
>> probing code, but doing a few linear probes is cheap.
>>
>> In contrast, the perfect linear hashes do crazy things like having
>> table lookups *JUST TO COMPUTE THE HASH*.
>>
>> Which is f*cking stupid, really. They'll miss in the cache just at
>> hash compute time, never mind at hash lookup. The table-driven
>> versions look beautiful in microbenchmarks that have the tables in the
>> L1 cache, but for something like the exception handling, I can
>> guarantee that *nothing* is in L1, and probably not even L2.
>>
>> So what you want is:
>> - no table lookups for hashing
>> - simple code (ie a normal "a multiply and a shift/mask or two") to
>> keep the I$ footprint down too
>> - you *will* take a cache miss on the actual hash table lookup, that
>> cannot be avoided, but linear probing at least hopefully keeps it to
>> that single cache miss even if you have to do a probe or two.
>>
>> Remember: this is very much a "cold-cache behavior matters" case. We
>> would never ever call this in a loop, at most we have loads that get a
>> fair amount of exceptions (but will go through the exception code, so
>> the L1 is probably blown even then).
>>
>>                          Linus
>