Date: Sat, 31 Jan 2015 07:57:35 -0500 From: John Baldwin <jhb@freebsd.org> To: Warner Losh <imp@bsdimp.com> Cc: Konstantin Belousov <kostikbel@gmail.com>, Poul-Henning Kamp <phk@phk.freebsd.dk>, freebsd-arch <freebsd-arch@freebsd.org> Subject: Re: Wrapper API for static bus_dma allocations Message-ID: <1740117.2W2DTbza1h@ralph.baldwin.cx> In-Reply-To: <21F3F28E-DAB8-4809-A9ED-1095F6BECCFC@bsdimp.com> References: <2800970.jY4xzTy9Hz@ralph.baldwin.cx> <1440008.gcoNUU8dV6@ralph.baldwin.cx> <21F3F28E-DAB8-4809-A9ED-1095F6BECCFC@bsdimp.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Friday, January 30, 2015 09:07:52 PM Warner Losh wrote: > > On Jan 30, 2015, at 2:31 PM, John Baldwin <jhb@freebsd.org> wrote: > >=20 > > On Friday, January 30, 2015 05:21:50 PM Konstantin Belousov wrote: > >> On Fri, Jan 30, 2015 at 09:56:31AM -0500, John Baldwin wrote: > >>> On 1/29/15 4:54 PM, Poul-Henning Kamp wrote: > >>>> -------- > >>>>=20 > >>>> In message <2800970.jY4xzTy9Hz@ralph.baldwin.cx>, John Baldwin w= rites: > >>>>> The bus_dma API to allocate a chunk of static DMA'able memory (= e.g. > >>>>> for > >>>>> descriptor rings) can be a bit obtuse [...] > >>>>=20 > >>>> Isn't it time we take a good hard stare at all of the bus_dma AP= I, > >>>> and refactor it into something a lot more compact ? > >>>=20 > >>> Given the amount of oddball hardware out there I don't think ther= e is a > >>> lot you can cut out. The filter function might be something we c= an lose > >>> (and losing it would simplify the implementation), but all the ot= her > >>> weird constraints are actually used by something AFAIK. I do thi= nk we > >>> can provide some simpler wrappers for some of the more common cas= es, but > >>> there will be some hardware for which those wrappers do not work.= > >>>=20 > >>> One suggestion Scott has had is to at least make it easier to ext= end the > >>> API by using getter/setter routines on the tag to work with tag > >>> attributes instead of passing them all in bus_dma_tag_create(). > >>=20 > >> BTW, filter function is useless. It can deny specific bus address= from > >> being used, but it does not provide the busdma implementation even= a hint > >> what other address should be (tried to) used. In dmar busdma, I s= imply > >> ignored it. And there is no real users of filter in the tree. > >=20 > > Yes, it is very annoying. I think some old ISA SCSI HBA driver mig= ht have > > used it to skip over some low-memory hole (i.e. there were two vali= d DMA > > ranges and this was the kludge instead of having two sets of > > lowaddr/highaddr exclusions). (That is one part of the API we coul= d > > rototill is to just remove the highaddr arg just use a single arg w= hich > > is effectively lowaddr. I think all drivers always set highaddr to= > > BUS_SPACE_MAXADDR.) >=20 > Not all. There=E2=80=99s some PCI cards that can=E2=80=99t do 64-bit = cycles that pass in the > 32-bit value on 64-bit systems. There=E2=80=99s 386 instances of this= in the tree. > But that may be lowaddr only. It=E2=80=99s hard to grep for this to b= e sure. That is lowaddr only, not the filter callback. However, even if we remove the filter and highaddr arguments from tags,= you=20 are still stuck with creating a tag, allocating memory, and loading it = to get=20 a bus address (and tracking the associated pointers, etc.). I still th= ink a=20 wrapper API for the common case (static DMA allocations) would be usefu= l. Orthogonally I can explore removing the filter along with highaddr (it = is=20 always BUS_SPACE_MAXADDR). --=20 John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1740117.2W2DTbza1h>