On 02/02/2023 10:02, David Woodhouse wrote:
On Wed, 2023-02-01 at 20:43 +0000, Usama Arif wrote:
This patchseries is from the work done by David Woodhouse (v4: https://lore.kernel.org/all/20220201205328.123066-1-dwmw2@xxxxxxxxxxxxx/).
The parallel CPU bringup is disabled for all AMD CPUs in this version: (see discussions: https://lore.kernel.org/all/bc3f2b1332c4bb77558df8aa36493a55542fe5b9.camel@xxxxxxxxxxxxx/ and
https://lore.kernel.org/all/3b6ac86fdc800cac5806433daf14a9095be101e9.camel@xxxxxxxxxxxxx/).
Doing INIT/SIPI/SIPI in parallel brings down the time for smpboot from ~700ms
to 100ms (85% improvement) on a server with 128 CPUs split across 2 NUMA
nodes.
Adding another cpuhp state for do_wait_cpu_initialized to make sure cpu_init
is reached in parallel as proposed by David in v1 will bring it down further
to ~30ms. Making this change would be dependent on this patchseries, so they
could be explored if this gets merged.
Changes across versions:
v2: Cut it back to just INIT/SIPI/SIPI in parallel for now, nothing more
v3: Clean up x2apic patch, add MTRR optimisation, lock topology update
in preparation for more parallelisation.
v4: Fixes to the real mode parallelisation patch spotted by SeanC, to
avoid scribbling on initial_gs in common_cpu_up(), and to allow all
24 bits of the physical X2APIC ID to be used. That patch still needs
a Signed-off-by from its original author, who once claimed not to
remember writing it at all. But now we've fixed it, hopefully he'll
admit it now :)
v5: rebase to v6.1 and remeasure performance, disable parallel bringup
for AMD CPUs.
Thanks, Usama.
I've updated to v6.2-rc6 since there were a few more tweaks required
(and we should double-check that the new handling of cache_ap_init from
a dedicated cpuhp step works right if that ends up being done in
parallel).
I also fixed up the complaints from the test robot; including
<linux/smpboot.h> from smpboot.c and making do_cpu_up() static, and
putting #ifdef CONFIG_SMP around the 'are we booting the AP?' check and
code segment in head_64.S.
I've made the AMD thing a CPU bug as Peter suggested, and pushed it to
https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/parallel-6.2-rc6
for you to do the real work of actually testing it :)
Thanks David! I have tested and reposted the v6.2-rc6 patches. One thing
I was mistaken about since I had rebased the patches together was that
the last 100ms to 30ms optimization was coming from parallelization in
x86/cpu:wait-init, when it seems to have a negligible affect. The last
70ms optimization was coming mainly from reusing timer calibration. Its
a simple patch and I have added it at the end of the series. The only
thing thats' missing was a sign-off from the author who I have added to
the latest series.
