[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH for-4.5] libxl: do not load roms for any NICs except the first to avoid wasting memory



On Mon, Nov 24, 2014 at 03:26:04PM +0000, Stefano Stabellini wrote:
> On Mon, 24 Nov 2014, Konrad Rzeszutek Wilk wrote:
> > On Mon, Nov 24, 2014 at 12:17:12PM +0000, Stefano Stabellini wrote:
> > > On Fri, 21 Nov 2014, Konrad Rzeszutek Wilk wrote:
> > > > On Fri, Nov 21, 2014 at 06:48:53PM +0000, Stefano Stabellini wrote:
> > > > > On Fri, 21 Nov 2014, Konrad Rzeszutek Wilk wrote:
> > > > > > On Fri, Nov 21, 2014 at 05:11:09PM +0000, Stefano Stabellini wrote:
> > > > > > > The rom is used for pxebooting. We don't need to allow pxebooting 
> > > > > > > from
> > > > > > > more than one network card.  Loading a romfile for every NIC 
> > > > > > > wastes
> > > > > > 
> > > > > > Why not? Why can't we PXE boot from each network card?
> > > > > 
> > > > > I take it back: you are right, at the moment it is actually possible 
> > > > > to
> > > > > pxe boot from the fourth nic for example. Maybe we could just load the
> > > > > first four romfiles and skip the others (four is the limit before QEMU
> > > > > fails to allocate any more memory).
> > > > 
> > > > The limit is in the count of ROMs or the memory consumption?
> > > 
> > > The limit is memory consumption.
> > 
> > Um, how big are those ROMs? Gigs?
> 
> I think they are 60K each in the case of rtl8139.
> However they are accounted on top of the ram of the guest, that's why
> the allocation fails. Strictly speaking I guess it shouldn't work even
> the first time around.

HA!
> 
> Maybe the bug is in libxl memory allocation, that doesn't account for
> rom sizes.

Or maybe the VGA hole size calculation is used for this. Anyhow, why aren't
we using the guest memory as instead of increasing the memory. Or is
this just an accounting issue in QEMU?
> 
> 
> > > > What if you also do PCI passthrough? Does that figure in this 
> > > > calculation?
> > > 
> > > In the case of PCI passthrough, roms are remapped, so they shouldn't
> > > affect memory consumption.
> > > 
> > >  
> > > > > TBH not all the emulated nics need a romfile but I wouldn't want to go
> > > > > down at that level of details beause they become QEMU implementation
> > > > > details. I wouldn't want to tie libxl to a certain version of QEMU in
> > > > > any way.
> > > > > 
> > > > > A better way would be handling memory allocation errors in QEMU but 
> > > > > QEMU
> > > > > doesn't do that at the moment and the change there would be certainly
> > > > > more invasive that a couple of lines in libxl.
> > > > > 
> > > > > 
> > > > > > > memory and as a matter of fact breaks configurations with more 
> > > > > > > than 4
> > > > > > > NICs as QEMU fails to allocate memory on behalf of the guest.
> > > > > > 
> > > > > > What if you have four different type of NICs? Say 1 rlt8193, 1 
> > > > > > e1000, one eepro,
> > > > > > and ne2k ?
> > > > > > 
> > > > > > Don't you want to load the ROM for each one?
> > > > > 
> > > > > The rom cannot be reused in QEMU among multiple instances of the same
> > > > > nic, so having four different types of nics or just one type doesn't
> > > > > change anything.
> > > > > 
> > > > > 
> > > > > > > With this fix, it is possible to assign more than 4 rtl8139 NICs 
> > > > > > > to the
> > > > > > > guest.
> > > > > > > 
> > > > > > > Signed-off-by: Stefano Stabellini 
> > > > > > > <stefano.stabellini@xxxxxxxxxxxxx>
> > > > > > > 
> > > > > > > diff --git a/tools/libxl/libxl_dm.c b/tools/libxl/libxl_dm.c
> > > > > > > index 3e191c3..f907ca9 100644
> > > > > > > --- a/tools/libxl/libxl_dm.c
> > > > > > > +++ b/tools/libxl/libxl_dm.c
> > > > > > > @@ -674,9 +674,10 @@ static char ** 
> > > > > > > libxl__build_device_model_args_new(libxl__gc *gc,
> > > > > > >                                                  
> > > > > > > LIBXL_NIC_TYPE_VIF_IOEMU);
> > > > > > >                  flexarray_append(dm_args, "-device");
> > > > > > >                  flexarray_append(dm_args,
> > > > > > > -                   libxl__sprintf(gc, 
> > > > > > > "%s,id=nic%d,netdev=net%d,mac=%s",
> > > > > > > +                   libxl__sprintf(gc, 
> > > > > > > "%s,id=nic%d,netdev=net%d,mac=%s%s",
> > > > > > >                                                  nics[i].model, 
> > > > > > > nics[i].devid,
> > > > > > > -                                                nics[i].devid, 
> > > > > > > smac));
> > > > > > > +                                                nics[i].devid, 
> > > > > > > smac,
> > > > > > > +                                                i ? 
> > > > > > > ",romfile=\"\"" : ""));
> > > > > > >                  flexarray_append(dm_args, "-netdev");
> > > > > > >                  flexarray_append(dm_args, GCSPRINTF(
> > > > > > >                                            
> > > > > > > "type=tap,id=net%d,ifname=%s,"
> > > > > > > 
> > > > > > > _______________________________________________
> > > > > > > Xen-devel mailing list
> > > > > > > Xen-devel@xxxxxxxxxxxxx
> > > > > > > http://lists.xen.org/xen-devel
> > > > > > 
> > > > 
> > 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.