Date: Fri, 10 Jun 2005 15:22:12 -0600 From: Scott Long <scottl@samsco.org> To: Kris Kennaway <kris@obsecurity.org> Cc: Hiroki Sato <hrs@freebsd.org>, sparc64@freebsd.org Subject: Re: E4500 with 24GB RAM Message-ID: <42AA0484.1090005@samsco.org> In-Reply-To: <20050610211239.GA59402@xor.obsecurity.org> References: <429A7B4D.3080102@samsco.org> <20050607.015510.21897573.hrs@allbsd.org> <20050606132756.X16994@carver.gumbysoft.com> <20050611.004435.59726356.hrs@allbsd.org> <20050610211239.GA59402@xor.obsecurity.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Kris Kennaway wrote: > On Sat, Jun 11, 2005 at 12:44:35AM +0900, Hiroki Sato wrote: > >>Doug White <dwhite@gumbysoft.com> wrote >> in <20050606132756.X16994@carver.gumbysoft.com>: >> >>dw> Since the panic is in a KASSERT thats why it only pops up under >>dw> INVARIANTS. However the bus_dma tags specify that hme can only handle 32 >>dw> bit addresses so it should be using a bounce buffer. I'm guessing that >>dw> bus_dma needs to be taught about this limitation and use bounce buffers >>dw> for the IOMMU on >16GB systems. >>dw> >>dw> Someone should also research this value a bit more ... I wonder if its >>dw> from some older system (E450?), and newer machines have larger limits. I >>dw> would hope they'd build E4500s with IOMMUs that can address all of >>dw> physical memory. >> >> Hmmm, I notice the box sometimes hangs up with <16GB RAM under >> moderate network load just after "hme0: invalid packet size >> xxxx; dropping" is displayed. I tried 16, 12, 8, and 4GB. >> Another E4500 with 5GB RAM seems OK, though. > > > I wonder if it's disk related. I tried to check out a ports tree on > this machine and it hung in a few seconds (although this was also > checking out using the network via nfs). > > Kris > > >> Is there anyone interested in investigating? I can provide remote >> access (serial console available) to this box. >> >>-- >>| Hiroki SATO > > > If it's a matter of the iommu and busdma, there aren't any quick fixes, or even quick hacks. I'm pretty sure that esp(4) will handle EINPROGRESS from busdma, but naturally it hasn't been tested because the proper busdma code for sparc64 hasn't been written yet. Scott
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?42AA0484.1090005>