summaryrefslogtreecommitdiffstats
path: root/doc/README.mpc85xx-spin-table
blob: 8da768a2a6c540c41c583cc68bc34bc37a413afe (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
Spin table in cache
=====================================
As specified by ePAPR v1.1, the spin table needs to be in cached memory. After
DDR is initialized and U-boot relocates itself into DDR, the spin table is
accessible for core 0. It is part of release.S, within 4KB range after
__secondary_start_page. For other cores to use the spin table, the booting
process is described below:

Core 0 sets up the reset page on the top 4K of memory (or 4GB if total memory
is more than 4GB), and creates a TLB to map it to 0xffff_f000, regardless of
the physical address of this page, with WIMGE=0b01010. Core 0 also enables boot
page translation for secondary cores to use this page of memory. Then 4KB
memory is copied from __secondary_start_page to the boot page, after flusing
cache because this page is mapped as normal DDR. Before copying the reset page,
core 0 puts the physical address of the spin table (which is in release.S and
relocated to the top of mapped memory) into a variable __spin_table_addr so
that secondary cores can see it.

When secondary cores boot up from 0xffff_f000 page, they only have one default
TLB. While booting, they set up another TLB in AS=1 space and jump into
the new space. The new TLB covers the physical address of the spin table page,
with WIMGE =0b00100. Now secondary cores can keep polling the spin table
without stress DDR bus because both the code and the spin table is in cache.

For the above to work, DDR has to set the 'M' bit of WIMGE, in order to keep
cache coherence.
OpenPOWER on IntegriCloud