Re: [PATCH kernel] prom_init: Fetch flatten device tree from the system firmware

[Segher Boessenkool <segher@kernel.crashing.org>]

On Fri, May 03, 2019 at 10:10:57AM +1000, Stewart Smith wrote:
> David Gibson <david@gibson.dropbear.id.au> writes:
> > On Wed, May 01, 2019 at 01:42:21PM +1000, Alexey Kardashevskiy wrote:
> >> At the moment, on 256CPU + 256 PCI devices guest, it takes the guest
> >> about 8.5sec to fetch the entire device tree via the client interface
> >> as the DT is traversed twice - for strings blob and for struct blob.
> >> Also, "getprop" is quite slow too as SLOF stores properties in a linked
> >> list.
> >> 
> >> However, since [1] SLOF builds flattened device tree (FDT) for another
> >> purpose. [2] adds a new "fdt-fetch" client interface for the OS to fetch
> >> the FDT.
> >> 
> >> This tries the new method; if not supported, this falls back to
> >> the old method.
> >> 
> >> There is a change in the FDT layout - the old method produced
> >> (reserved map, strings, structs), the new one receives only strings and
> >> structs from the firmware and adds the final reserved map to the end,
> >> so it is (fw reserved map, strings, structs, reserved map).
> >> This still produces the same unflattened device tree.
> >> 
> >> This merges the reserved map from the firmware into the kernel's reserved
> >> map. At the moment SLOF generates an empty reserved map so this does not
> >> change the existing behaviour in regard of reservations.
> >> 
> >> This supports only v17 onward as only that version provides dt_struct_size
> >> which works as "fdt-fetch" only produces v17 blobs.
> >> 
> >> If "fdt-fetch" is not available, the old method of fetching the DT is used.
> >> 
> >> [1] https://git.qemu.org/?p=SLOF.git;a=commitdiff;h=e6fc84652c9c00
> >> [2] https://git.qemu.org/?p=SLOF.git;a=commit;h=ecda95906930b80
> >> 
> >> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
> >
> > Hrm.  I've gotta say I'm not terribly convinced that it's worth adding
> > a new interface we'll need to maintain to save 8s on a somewhat
> > contrived testcase.
> 
> 256CPUs aren't that many anymore though. Although I guess that many PCI
> devices is still a little uncommon.
> 
> A 4 socket POWER8 or POWER9 can easily be that large, and a small test
> kernel/userspace will boot in ~2.5-4 seconds. So it's possible that
> the device tree fetch could be surprisingly non-trivial percentage of boot
> time at least on some machines.

All client interface calls are really heavy, and you need to do a lot of
them if you have a big device tree.  This takes time, even if the linked
list stuff does not kill you :-)


Segher