Skip site navigation (1)Skip section navigation (2)
Date:      Tue, 07 Aug 2012 10:09:42 -0600
From:      Ian Lepore <freebsd@damnhippie.dyndns.org>
To:        John Baldwin <jhb@freebsd.org>
Cc:        arm@freebsd.org, mips@freebsd.org, Peter Jeremy <peter@rulingia.com>
Subject:   Re: On-stack allocation of DMA S/G lists
Message-ID:  <1344355782.1128.186.camel@revolution.hippie.lan>
In-Reply-To: <201208061026.06328.jhb@freebsd.org>
References:  <20120703111753.GB72292@server.rulingia.com> <20120708110516.GA38312@server.rulingia.com> <201207120826.05577.jhb@freebsd.org> <201208061026.06328.jhb@freebsd.org>

next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 2012-08-06 at 10:26 -0400, John Baldwin wrote:
> On Thursday, July 12, 2012 8:26:05 am John Baldwin wrote:
> > On Sunday, July 08, 2012 7:05:16 am Peter Jeremy wrote:
> > > BTW(2): Whilst studying busdma_machdep.c for arm and mips, I've
> > > noticed they appear to potentially allocate substantial kernel stack
> > > under some conditions as several bus_dma(9) functions include:
> > >     bus_dma_segment_t dm_segments[dmat->nsegments];
> > > What prevents this overflowing the kernel stack?
> > 
> > That does seem dubious.  x86 stores the array in the tag instead.
> 
> I have an untested patch to change bus-dma on arm and mips to allocate a
> dynamic S/G list in each DMA tag on first use instead of using on-stack
> allocation (which I think is rather bogus).  Can folks review and test this 
> patch please?  Thanks.
> 
> http://www.FreeBSD.org/~jhb/patches/arm_mips_dynamic_dma_segs.patch
> 

I'm worried about changing a per-mapping-call resource to a per-dma-tag
resource here.  What prevents the situation where you have two
bus_dmamap_load() calls in progress at the same time using different
buffers but the same tag?  

I can't find anything in the docs that indicates you have to provide
external locking of the tag for map load/unload calls, or that even
implies the tag can be modified by a mapping operation.  The lockfunc
stuff related to creating the tag is documented as being used only
during a deferred callback.

The existing code seems to go out of its way to avoid modifying the tag
during a mapping operation.  For example, it decides at tag creation
time whether any bounce pages might ever be needed for the tag, and if
so it pre-sets a bounce zone in the tag, then at mapping time the bounce
zone is protected with its own lock when it gets modified.  To me this
feels like a way to specifically avoid the need to lock or modify the
tag during a mapping operation.

Assuming that all of the foregoing is moot for some reason I've
overlooked, then on a purely implementation level, could all the
duplicated code to allocate the array when necessary be moved into
bus_dmamap_load_buffer(), triggered by a NULL 'segs' pointer?

And just for the record, looking at the problem from an even more
distant vantage... is there really a problem with stack-allocating the
segments?  On a 64-bit arch the struct is like 16 bytes.  Typical usage
is to allocate a tag allowing 1 or just a few segments.  Is anyone
really going to create a tag specifying hundreds of segments that would
overflow the stack?  If they try, wouldn't failing the tag create be
good enough?

-- Ian





Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1344355782.1128.186.camel>