On Tue, 20 Sep 2005, Dominic Sweetman wrote:
> > > I found an performance bug in c-r4k.c:r4k_dma_cache_inv, where a
> > > Hit_Writeback_Inv instead of Hit_Invalidate is done.
>
> The MIPS64 spec (which is really all there is to set standards in this
> area) regards Hit_Invalidate as optional. So it would be nice not to
> use it. CPUs have no standard "configuration" register you can read
> to establish which cacheops work, so to identify capable CPUs you must
> use a table of CPU attributes indexed by the CPU ID, which encourages
> the crime of building software which can't possibly run on a new CPU...
Or just using the safe fallback -- that shouldn't be a problem (these
functions are called indirectly). Besides new CPUs more often than not
require changes to kernel-level software anyway.
> So long as the buffer is in fact clean, then in most implementations a
> Hit_Writeback_Invalidate will be just as efficient.
I hope so, but who knows what's wired there in all those old systems?...
> I suppose where DMA data subsequently gets decorated by the CPU then
> handed on to some other layer, then the buffer is freed...?
I don't think the buffer is modified, so cache lines should remain clean.
For the usual case of IP data is used exactly once for copy_and_csum()
(more or less) which moves it to another buffer.
> > FYI, for R4k DECstations the need to flush the cache for newly allocated
> > skbs reduces throughput of FDDI reception by about a half (!), down from
> > about 90Mbps (that's for the /260)...
>
> How did you measure the high throughput? Have you got a
> machine with DMA-coherency you can turn on and off?
I just disabled invalidations. ;-) Yes, that resulted in some corrupt
data, but it was good enough to do benchmarking. That was an R4400 with
1MB of S-cache.
Eventually I should benchmark both invalidation variations against each
other with the system in question and see if it makes any difference.
Ironically this is where the write-back cache of the R4k gives loss rather
than gain as compared to the write-through cache of the R3k (the system
supports daughtercards with either CPU, so useful comparison is possible)
as for the former I have to invalidate cache spanning the whole
newly-allocated buffer, i.e. ~4.5kB, while for the latter I may invalidate
only the area actually used, once a frame has been received, its length is
known and quite often much smaller than the maximum (especially if it's
been routed from a network that has a smaller frame length limit, like
Ethernet).
Maciej
|