Greg Kurz
2014-10-03 09:13:17 UTC
The associativity domain numbers are obtained from the hypervisor through
registers and written into memory by the guest: the packed array passed to
vphn_unpack_associativity() is then native-endian, unlike what was assumed
in the following commit:
commit b08a2a12e44eaec5024b2b969f4fcb98169d1ca3
Author: Alistair Popple <alistair at popple.id.au>
Date: Wed Aug 7 02:01:44 2013 +1000
powerpc: Make NUMA device node code endian safe
If a CPU home node changes, the topology gets filled with
bogus values. This leads to severe performance breakdowns.
This patch does two things:
- extract values from the packed array with shifts, in order to be endian
neutral
- convert the resulting values to be32 as expected
Suggested-by: Anton Blanchard <anton at samba.org>
Signed-off-by: Greg Kurz <gkurz at linux.vnet.ibm.com>
---
Changes in v2:
- removed the left out __be16 *field declaration
- removed the left out be16_to_cpup() call
- updated the comment of the magic formula
Thanks again Nish... the two left outs probably explain why PowerVM wasn't
happy that patch. :P
--
Greg
arch/powerpc/mm/numa.c | 64 +++++++++++++++++++++++++++++++++++++++---------
1 file changed, 52 insertions(+), 12 deletions(-)
diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index b835bf0..4547f91 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c
@@ -1369,38 +1369,78 @@ static int update_cpu_associativity_changes_mask(void)
#define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
/*
+ * The associativity values are either 16-bit (VPHN_FIELD_MSB) or 32-bit (data
+ * or VPHN_FIELD_UNUSED). We hence need to parse the packed array into 16-bit
+ * chunks. Let's do that with bit shifts to be endian neutral.
+ *
+ * --- 16-bit chunks -->
+ * _________________________
+ * | 0 | 1 | 2 | 3 | packed[0]
+ * -------------------------
+ * _________________________
+ * | 4 | 5 | 6 | 7 | packed[1]
+ * -------------------------
+ * ...
+ * _________________________
+ * | 20 | 21 | 22 | 23 | packed[5]
+ * -------------------------
+ * 48 32 16 0
+ * <------ bits --------
+ *
+ * We need 48-bit shift for chunks 0,4,8,16,20
+ * 32-bit shift for chunks 1,5,9,17,21
+ * 16-bit shift for chunks 2,6,10,18,22
+ * no shift for chunks 3,7,11,19,23
+ *
+ * The 2 lo bits of the chunk index multiplied by 16 give the shift.
+ * The remaining hi bits divided by 4 give the index in packed[].
+ *
+ * For example:
+ * chunk 0 = packed[0/4] >> (~0b00000 & 0b11) * 16 = bits 48..63 in packed[0]
+ * chunk 5 = packed[5/4] >> (~0b00101 & 0b11) * 16 = bits 32..47 in packed[1]
+ * chunk 22 = packed[22/4] >> (~0b10110 & 0b11) * 16 = bits 16..31 in packed[5]
+ */
+static inline u16 read_vphn_chunk(const long *packed, unsigned int i)
+{
+ return packed[i >> 2] >> ((~i & 3) << 4);
+}
+
+/*
* Convert the associativity domain numbers returned from the hypervisor
* to the sequence they would appear in the ibm,associativity property.
*/
static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
{
- int i, nr_assoc_doms = 0;
- const __be16 *field = (const __be16 *) packed;
+ unsigned int i, j, nr_assoc_doms = 0;
#define VPHN_FIELD_UNUSED (0xffff)
#define VPHN_FIELD_MSB (0x8000)
#define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
- for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
- if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
+ for (i = 1, j = 0; i < VPHN_ASSOC_BUFSIZE; i++) {
+ u16 field = read_vphn_chunk(packed, j);
+
+ if (field == VPHN_FIELD_UNUSED) {
/* All significant fields processed, and remaining
* fields contain the reserved value of all 1's.
* Just store them.
*/
- unpacked[i] = *((__be32 *)field);
- field += 2;
- } else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
+ unpacked[i] = (VPHN_FIELD_UNUSED << 16 |
+ VPHN_FIELD_UNUSED);
+ j += 2;
+ } else if (field & VPHN_FIELD_MSB) {
/* Data is in the lower 15 bits of this field */
- unpacked[i] = cpu_to_be32(
- be16_to_cpup(field) & VPHN_FIELD_MASK);
- field++;
+ unpacked[i] = cpu_to_be32(field & VPHN_FIELD_MASK);
+ j++;
nr_assoc_doms++;
} else {
/* Data is in the lower 15 bits of this field
* concatenated with the next 16 bit field
*/
- unpacked[i] = *((__be32 *)field);
- field += 2;
+ unpacked[i] =
+ cpu_to_be32((u32) field << 16 |
+ read_vphn_chunk(packed, j + 1));
+ j += 2;
nr_assoc_doms++;
}
}
registers and written into memory by the guest: the packed array passed to
vphn_unpack_associativity() is then native-endian, unlike what was assumed
in the following commit:
commit b08a2a12e44eaec5024b2b969f4fcb98169d1ca3
Author: Alistair Popple <alistair at popple.id.au>
Date: Wed Aug 7 02:01:44 2013 +1000
powerpc: Make NUMA device node code endian safe
If a CPU home node changes, the topology gets filled with
bogus values. This leads to severe performance breakdowns.
This patch does two things:
- extract values from the packed array with shifts, in order to be endian
neutral
- convert the resulting values to be32 as expected
Suggested-by: Anton Blanchard <anton at samba.org>
Signed-off-by: Greg Kurz <gkurz at linux.vnet.ibm.com>
---
Changes in v2:
- removed the left out __be16 *field declaration
- removed the left out be16_to_cpup() call
- updated the comment of the magic formula
Thanks again Nish... the two left outs probably explain why PowerVM wasn't
happy that patch. :P
--
Greg
arch/powerpc/mm/numa.c | 64 +++++++++++++++++++++++++++++++++++++++---------
1 file changed, 52 insertions(+), 12 deletions(-)
diff --git a/arch/powerpc/mm/numa.c b/arch/powerpc/mm/numa.c
index b835bf0..4547f91 100644
--- a/arch/powerpc/mm/numa.c
+++ b/arch/powerpc/mm/numa.c
@@ -1369,38 +1369,78 @@ static int update_cpu_associativity_changes_mask(void)
#define VPHN_ASSOC_BUFSIZE (6*sizeof(u64)/sizeof(u32) + 1)
/*
+ * The associativity values are either 16-bit (VPHN_FIELD_MSB) or 32-bit (data
+ * or VPHN_FIELD_UNUSED). We hence need to parse the packed array into 16-bit
+ * chunks. Let's do that with bit shifts to be endian neutral.
+ *
+ * --- 16-bit chunks -->
+ * _________________________
+ * | 0 | 1 | 2 | 3 | packed[0]
+ * -------------------------
+ * _________________________
+ * | 4 | 5 | 6 | 7 | packed[1]
+ * -------------------------
+ * ...
+ * _________________________
+ * | 20 | 21 | 22 | 23 | packed[5]
+ * -------------------------
+ * 48 32 16 0
+ * <------ bits --------
+ *
+ * We need 48-bit shift for chunks 0,4,8,16,20
+ * 32-bit shift for chunks 1,5,9,17,21
+ * 16-bit shift for chunks 2,6,10,18,22
+ * no shift for chunks 3,7,11,19,23
+ *
+ * The 2 lo bits of the chunk index multiplied by 16 give the shift.
+ * The remaining hi bits divided by 4 give the index in packed[].
+ *
+ * For example:
+ * chunk 0 = packed[0/4] >> (~0b00000 & 0b11) * 16 = bits 48..63 in packed[0]
+ * chunk 5 = packed[5/4] >> (~0b00101 & 0b11) * 16 = bits 32..47 in packed[1]
+ * chunk 22 = packed[22/4] >> (~0b10110 & 0b11) * 16 = bits 16..31 in packed[5]
+ */
+static inline u16 read_vphn_chunk(const long *packed, unsigned int i)
+{
+ return packed[i >> 2] >> ((~i & 3) << 4);
+}
+
+/*
* Convert the associativity domain numbers returned from the hypervisor
* to the sequence they would appear in the ibm,associativity property.
*/
static int vphn_unpack_associativity(const long *packed, __be32 *unpacked)
{
- int i, nr_assoc_doms = 0;
- const __be16 *field = (const __be16 *) packed;
+ unsigned int i, j, nr_assoc_doms = 0;
#define VPHN_FIELD_UNUSED (0xffff)
#define VPHN_FIELD_MSB (0x8000)
#define VPHN_FIELD_MASK (~VPHN_FIELD_MSB)
- for (i = 1; i < VPHN_ASSOC_BUFSIZE; i++) {
- if (be16_to_cpup(field) == VPHN_FIELD_UNUSED) {
+ for (i = 1, j = 0; i < VPHN_ASSOC_BUFSIZE; i++) {
+ u16 field = read_vphn_chunk(packed, j);
+
+ if (field == VPHN_FIELD_UNUSED) {
/* All significant fields processed, and remaining
* fields contain the reserved value of all 1's.
* Just store them.
*/
- unpacked[i] = *((__be32 *)field);
- field += 2;
- } else if (be16_to_cpup(field) & VPHN_FIELD_MSB) {
+ unpacked[i] = (VPHN_FIELD_UNUSED << 16 |
+ VPHN_FIELD_UNUSED);
+ j += 2;
+ } else if (field & VPHN_FIELD_MSB) {
/* Data is in the lower 15 bits of this field */
- unpacked[i] = cpu_to_be32(
- be16_to_cpup(field) & VPHN_FIELD_MASK);
- field++;
+ unpacked[i] = cpu_to_be32(field & VPHN_FIELD_MASK);
+ j++;
nr_assoc_doms++;
} else {
/* Data is in the lower 15 bits of this field
* concatenated with the next 16 bit field
*/
- unpacked[i] = *((__be32 *)field);
- field += 2;
+ unpacked[i] =
+ cpu_to_be32((u32) field << 16 |
+ read_vphn_chunk(packed, j + 1));
+ j += 2;
nr_assoc_doms++;
}
}