* adjust __spin_table alignment to match ePAPR v0.94 spec * loop over all cpus when determing who is up. This fixes an issue if the "boot cpu" isn't core0. The "boot cpu" will already be in the cpu_up_mask so there is no harm * Added some protection in the code to ensure proper behavior. These changes are explicitly needed but don't hurt: - Added eieio to ensure the "hot word" of the table is written after all other table updates have occurred. - Added isync to ensure we don't prefetch loading of table entries until we a released

These issues we raised by Dave Liu.

Signed-off-by: Kumar Gala galak@kernel.crashing.org ---

For 1.3.3

cpu/mpc85xx/mp.c | 6 +++++- cpu/mpc85xx/release.S | 3 ++- 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/cpu/mpc85xx/mp.c b/cpu/mpc85xx/mp.c index e733f7b..a527cf3 100644 --- a/cpu/mpc85xx/mp.c +++ b/cpu/mpc85xx/mp.c @@ -103,6 +103,10 @@ int cpu_release(int nr, int argc, char *argv[]) }

table[BOOT_ENTRY_ADDR_UPPER] = (u32)(boot_addr >> 32); + + /* ensure all table updates complete before final address write */ + eieio(); + table[BOOT_ENTRY_ADDR_LOWER] = (u32)(boot_addr & 0xffffffff);

return 0; @@ -153,7 +157,7 @@ static void pq3_mp_up(unsigned long bootpg) /* wait for everyone */ while (timeout) { int i; - for (i = 1; i < CONFIG_NR_CPUS; i++) { + for (i = 0; i < CONFIG_NR_CPUS; i++) { if (table[i * NUM_BOOT_ENTRY + BOOT_ENTRY_ADDR_LOWER]) cpu_up_mask |= (1 << i); }; diff --git a/cpu/mpc85xx/release.S b/cpu/mpc85xx/release.S index 3b7366f..a47edae 100644 --- a/cpu/mpc85xx/release.S +++ b/cpu/mpc85xx/release.S @@ -114,6 +114,7 @@ __secondary_start_page: lwz r4,ENTRY_ADDR_LOWER(r10) andi. r11,r4,1 bne 2b + isync

/* get the upper bits of the addr */ lwz r11,ENTRY_ADDR_UPPER(r10) @@ -169,7 +170,7 @@ __secondary_start_page: mtspr SPRN_SRR1,r13 rfi

- .align 3 + .align L1_CACHE_SHIFT .globl __spin_table __spin_table: .space CONFIG_NR_CPUS*ENTRY_SIZE