We have encountered hardware with 18 cores/socket that gives 36 CPUs/socket
with hyperthreading enabled. This exceeds the current MAX_CPUS_PER_SOCKET
causing a failure in get_cpu_topology. Increase MAX_CPUS_PER_SOCKET to 64
and MAX_CPUS_PER_UVHUB to 128.
Additionally, this exposed another bug. When get_cpu_topology fails, we
need to make sure that we do not use bau_control->statp (e.g. when a
user reads /proc/sgi_uv/ptc_statistics causing ptc_seq_show to be
invoked). In uv_bau_init we call set_bau_off() when
init_per_cpu()->get_cpu_topology() fails -- therefore, to avoid
referencing a NULL pointer we add code that detects when the bau is off
and print out the bau is disabled.
Signed-off-by: James Custer <[email protected]>
---
arch/x86/include/asm/uv/uv_bau.h | 4 ++--
arch/x86/platform/uv/tlb_uv.c | 4 ++++
2 files changed, 6 insertions(+), 2 deletions(-)
diff --git a/arch/x86/include/asm/uv/uv_bau.h b/arch/x86/include/asm/uv/uv_bau.h
index 2d60a78..fc808b8 100644
--- a/arch/x86/include/asm/uv/uv_bau.h
+++ b/arch/x86/include/asm/uv/uv_bau.h
@@ -33,8 +33,8 @@
* Each of the descriptors is 64 bytes in size (8*64 = 512 bytes in a set).
*/
-#define MAX_CPUS_PER_UVHUB 64
-#define MAX_CPUS_PER_SOCKET 32
+#define MAX_CPUS_PER_UVHUB 128
+#define MAX_CPUS_PER_SOCKET 64
#define ADP_SZ 64 /* hardware-provided max. */
#define UV_CPUS_PER_AS 32 /* hardware-provided max. */
#define ITEMS_PER_DESC 8
diff --git a/arch/x86/platform/uv/tlb_uv.c b/arch/x86/platform/uv/tlb_uv.c
index 3968d67..570b1b1 100644
--- a/arch/x86/platform/uv/tlb_uv.c
+++ b/arch/x86/platform/uv/tlb_uv.c
@@ -1384,6 +1384,10 @@ static int ptc_seq_show(struct seq_file *file, void *data)
}
if (cpu < num_possible_cpus() && cpu_online(cpu)) {
bcp = &per_cpu(bau_control, cpu);
+ if (bcp->nobau) {
+ seq_printf(file, "cpu %d bau disabled\n", cpu);
+ return 0;
+ }
stat = bcp->statp;
/* source side statistics */
seq_printf(file,
--
1.8.2.1
On Fri, 31 Oct 2014, James Custer wrote:
> We have encountered hardware with 18 cores/socket that gives 36 CPUs/socket
> with hyperthreading enabled. This exceeds the current MAX_CPUS_PER_SOCKET
> causing a failure in get_cpu_topology. Increase MAX_CPUS_PER_SOCKET to 64
> and MAX_CPUS_PER_UVHUB to 128.
>
> Additionally, this exposed another bug. When get_cpu_topology fails, we
> need to make sure that we do not use bau_control->statp (e.g. when a
> user reads /proc/sgi_uv/ptc_statistics causing ptc_seq_show to be
> invoked). In uv_bau_init we call set_bau_off() when
> init_per_cpu()->get_cpu_topology() fails -- therefore, to avoid
> referencing a NULL pointer we add code that detects when the bau is off
> and print out the bau is disabled.
Can you please split this in two patches. One fixing the ptc_seq_show
bug and the other increasing the constants.
If someone looks for a fix for the ptc_seq_show wreckage he definitly
is not going to look at a patch which increases the number of cpus per
socket.
Thanks,
tglx