Date: Thu, 19 Nov 2009 17:31:00 -0500 From: Jung-uk Kim <jkim@FreeBSD.org> To: John Baldwin <jhb@FreeBSD.org> Cc: svn-src-head@FreeBSD.org, svn-src-all@FreeBSD.org, src-committers@FreeBSD.org, Robert Watson <rwatson@FreeBSD.org> Subject: Re: svn commit: r199498 - in head/sys: amd64/amd64 i386/i386 net Message-ID: <200911191731.04075.jkim@FreeBSD.org> In-Reply-To: <200911191649.37198.jhb@freebsd.org> References: <200911182340.nAINeJ3W087652@svn.freebsd.org> <200911191115.11088.jkim@FreeBSD.org> <200911191649.37198.jhb@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thursday 19 November 2009 04:49 pm, John Baldwin wrote: > On Thursday 19 November 2009 11:15:01 am Jung-uk Kim wrote: > > On Thursday 19 November 2009 03:26 am, Robert Watson wrote: > > > On Wed, 18 Nov 2009, Jung-uk Kim wrote: > > > > - Change internal function bpf_jit_compile() to return > > > > allocated size of the generated binary and remove page size > > > > limitation for userland. - Use contigmalloc(9)/contigfree(9) > > > > instead of malloc(9)/free(9) to make sure the generated > > > > binary aligns properly and make it physically contiguous. > > > > > > Is physical contiguity actually required here -- I would have > > > thought virtual contiguity and alignment would be sufficient, > > > in which case the normal trick is to allocate using malloc the > > > size + min-align + 1 and then fudge the pointer forward until > > > it's properly aligned. > > > > I don't believe it is strictly necessary but I assumed it might > > have performance benefit for very big BPF programs although I > > have not measured it. Also, contigmalloc(9)/contigfree(9) is too > > obvious to ignore for this purpose. :-) > > Why would it have a performance benefit to have the pages be > physically contiguous? contigmalloc() is expensive and should > really only be used if you truly need contiguous memory. If you > can get by with malloc(), just use malloc(). Remember are allocating memory for a function pointer here. If we want to take care of alignment, then "fudging the pointer forward" trick is not going to be easy unless I embed real offset in the structure and pass it around with the pointer. I don't mind doing it but it seemed unnecessary to me. Besides, it is very unlikely to see a lot of parallel BPF filter allocations in real world. Actually, that is a big assumption for BPF JIT compiler by itself because filter compilation is expensive. Also, if contigmalloc() fails, bpf(4) simply falls back to good old bpf_filter(). So, I don't see anything wrong with this. Jung-uk Kim
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200911191731.04075.jkim>