* [PATCH] [ppc64] NUMA fixes
@ 2004-05-24 5:35 Anton Blanchard
0 siblings, 0 replies; only message in thread
From: Anton Blanchard @ 2004-05-24 5:35 UTC (permalink / raw
To: akpm; +Cc: linux-kernel, paulus, haveblue, olof
>From Anton Blanchard, Dave Hansen and Olof Johansson:
Fix multiple bugs in the ppc64 NUMA topology probe code.
- We were using HW cpu numbers instead of logical ones. 615, 630, 650,
some 670 and some 690 SMP will all fail to boot without this patch.
- The old code would BUG() when it got confused (more NUMA zones than
the kernel is configured for etc).
- The common depth calculation was incorrect. Dave found an OF property
that gives us exactly what we want.
- Things were broken on SMT machines.
The new code should work on those broken systems and should no longer
BUG() but fall back to a flat topology when it gets confused.
---
forakpm-anton/arch/ppc64/mm/numa.c | 260 ++++++++++++++++++++-----------
forakpm-anton/include/asm-ppc64/mmzone.h | 1
2 files changed, 172 insertions(+), 89 deletions(-)
diff -puN arch/ppc64/mm/numa.c~ppc64-numafixes arch/ppc64/mm/numa.c
--- forakpm/arch/ppc64/mm/numa.c~ppc64-numafixes 2004-05-24 15:04:33.918731068 +1000
+++ forakpm-anton/arch/ppc64/mm/numa.c 2004-05-24 15:33:10.234294222 +1000
@@ -19,7 +19,7 @@
#include <asm/abs_addr.h>
#if 1
-#define dbg(args...) udbg_printf(args)
+#define dbg(args...) printk(KERN_INFO args)
#else
#define dbg(args...)
#endif
@@ -56,16 +56,136 @@ static inline void map_cpu_to_node(int c
}
}
+static struct device_node * __init find_cpu_node(unsigned int cpu)
+{
+ unsigned int hw_cpuid = get_hard_smp_processor_id(cpu);
+ struct device_node *cpu_node = NULL;
+ unsigned int *interrupt_server, *reg;
+ int len;
+
+ while ((cpu_node = of_find_node_by_type(cpu_node, "cpu")) != NULL) {
+ /* Try interrupt server first */
+ interrupt_server = (unsigned int *)get_property(cpu_node,
+ "ibm,ppc-interrupt-server#s", &len);
+
+ if (interrupt_server && (len > 0)) {
+ while (len--) {
+ if (interrupt_server[len-1] == hw_cpuid)
+ return cpu_node;
+ }
+ } else {
+ reg = (unsigned int *)get_property(cpu_node,
+ "reg", &len);
+ if (reg && (len > 0) && (reg[0] == hw_cpuid))
+ return cpu_node;
+ }
+ }
+
+ return NULL;
+}
+
+/* must hold reference to node during call */
+static int *of_get_associativity(struct device_node *dev)
+ {
+ unsigned int *result;
+ int len;
+
+ result = (unsigned int *)get_property(dev, "ibm,associativity", &len);
+
+ if (len <= 0)
+ return NULL;
+
+ return result;
+}
+
+static int of_node_numa_domain(struct device_node *device, int depth)
+{
+ int numa_domain;
+ unsigned int *tmp;
+
+ tmp = of_get_associativity(device);
+ if (tmp && (tmp[0] >= depth)) {
+ numa_domain = tmp[depth];
+ } else {
+ printk(KERN_ERR "WARNING: no NUMA information for "
+ "%s\n", device->full_name);
+ numa_domain = 0;
+ }
+ return numa_domain;
+}
+
+/*
+ * In theory, the "ibm,associativity" property may contain multiple
+ * associativity lists because a resource may be multiply connected
+ * into the machine. This resource then has different associativity
+ * characteristics relative to its multiple connections. We ignore
+ * this for now. We also assume that all cpu and memory sets have
+ * their distances represented at a common level. This won't be
+ * true for heirarchical NUMA.
+ *
+ * In any case the ibm,associativity-reference-points should give
+ * the correct depth for a normal NUMA system.
+ *
+ * - Dave Hansen <haveblue@us.ibm.com>
+ */
+static int find_min_common_depth(void)
+{
+ int depth;
+ unsigned int *ref_points;
+ struct device_node *rtas_root;
+ unsigned int len;
+
+ rtas_root = of_find_node_by_path("/rtas");
+
+ if (!rtas_root) {
+ printk(KERN_ERR "WARNING: %s() could not find rtas root\n",
+ __FUNCTION__);
+ return -1;
+ }
+
+ /*
+ * this property is 2 32-bit integers, each representing a level of
+ * depth in the associativity nodes. The first is for an SMP
+ * configuration (should be all 0's) and the second is for a normal
+ * NUMA configuration.
+ */
+ ref_points = (unsigned int *)get_property(rtas_root,
+ "ibm,associativity-reference-points", &len);
+
+ if ((len >= 1) && ref_points) {
+ depth = ref_points[1];
+ } else {
+ printk(KERN_ERR "WARNING: could not find NUMA "
+ "associativity reference point\n");
+ depth = -1;
+ }
+ of_node_put(rtas_root);
+
+ return depth;
+}
+
+static unsigned long read_cell_ul(struct device_node *device, unsigned int **buf)
+{
+ int i;
+ unsigned long result = 0;
+
+ i = prom_n_size_cells(device);
+ /* bug on i>2 ?? */
+ while (i--) {
+ result = (result << 32) | **buf;
+ (*buf)++;
+ }
+ return result;
+}
+
static int __init parse_numa_properties(void)
{
struct device_node *cpu = NULL;
struct device_node *memory = NULL;
- int *cpu_associativity;
- int *memory_associativity;
int depth;
int max_domain = 0;
long entries = lmb_end_of_DRAM() >> MEMORY_INCREMENT_SHIFT;
- long i;
+ unsigned long i;
if (strstr(saved_command_line, "numa=off")) {
printk(KERN_WARNING "NUMA disabled by user\n");
@@ -78,112 +198,78 @@ static int __init parse_numa_properties(
for (i = 0; i < entries ; i++)
numa_memory_lookup_table[i] = ARRAY_INITIALISER;
- cpu = of_find_node_by_type(NULL, "cpu");
- if (!cpu)
- goto err;
-
- memory = of_find_node_by_type(NULL, "memory");
- if (!memory)
- goto err;
-
- cpu_associativity = (int *)get_property(cpu, "ibm,associativity", NULL);
- if (!cpu_associativity)
- goto err;
-
- memory_associativity = (int *)get_property(memory, "ibm,associativity",
- NULL);
- if (!memory_associativity)
- goto err;
-
- /* find common depth */
- if (cpu_associativity[0] < memory_associativity[0])
- depth = cpu_associativity[0];
- else
- depth = memory_associativity[0];
-
- for (; cpu; cpu = of_find_node_by_type(cpu, "cpu")) {
- int *tmp;
- int cpu_nr, numa_domain;
+ depth = find_min_common_depth();
- tmp = (int *)get_property(cpu, "reg", NULL);
- if (!tmp)
- continue;
- cpu_nr = *tmp;
+ printk(KERN_INFO "NUMA associativity depth for CPU/Memory: %d\n", depth);
+ if (depth < 0)
+ return depth;
+
+ for_each_cpu(i) {
+ int numa_domain;
- tmp = (int *)get_property(cpu, "ibm,associativity",
- NULL);
- if (!tmp)
- continue;
- numa_domain = tmp[depth];
+ cpu = find_cpu_node(i);
- /* FIXME */
- if (numa_domain == 0xffff) {
- dbg("cpu %d has no numa doman\n", cpu_nr);
+ if (cpu) {
+ numa_domain = of_node_numa_domain(cpu, depth);
+ of_node_put(cpu);
+
+ if (numa_domain >= MAX_NUMNODES) {
+ /*
+ * POWER4 LPAR uses 0xffff as invalid node,
+ * dont warn in this case.
+ */
+ if (numa_domain != 0xffff)
+ printk(KERN_ERR "WARNING: cpu %ld "
+ "maps to invalid NUMA node %d\n",
+ i, numa_domain);
+ numa_domain = 0;
+ }
+ } else {
+ printk(KERN_ERR "WARNING: no NUMA information for "
+ "cpu %ld\n", i);
numa_domain = 0;
}
- if (numa_domain >= MAX_NUMNODES)
- BUG();
-
node_set_online(numa_domain);
if (max_domain < numa_domain)
max_domain = numa_domain;
- map_cpu_to_node(cpu_nr, numa_domain);
- /* register the second thread on an SMT machine */
- if (cur_cpu_spec->cpu_features & CPU_FTR_SMT)
- map_cpu_to_node(cpu_nr ^ 0x1, numa_domain);
+ map_cpu_to_node(i, numa_domain);
}
- for (; memory; memory = of_find_node_by_type(memory, "memory")) {
- unsigned int *tmp1, *tmp2;
- unsigned long i;
- unsigned long start = 0;
- unsigned long size = 0;
+ memory = NULL;
+ while ((memory = of_find_node_by_type(memory, "memory")) != NULL) {
+ unsigned long start;
+ unsigned long size;
int numa_domain;
int ranges;
+ unsigned int *memcell_buf;
+ unsigned int len;
- tmp1 = (int *)get_property(memory, "reg", NULL);
- if (!tmp1)
+ memcell_buf = (unsigned int *)get_property(memory, "reg", &len);
+ if (!memcell_buf || len <= 0)
continue;
ranges = memory->n_addrs;
new_range:
-
- i = prom_n_size_cells(memory);
- while (i--) {
- start = (start << 32) | *tmp1;
- tmp1++;
- }
-
- i = prom_n_size_cells(memory);
- while (i--) {
- size = (size << 32) | *tmp1;
- tmp1++;
- }
+ /* these are order-sensitive, and modify the buffer pointer */
+ start = read_cell_ul(memory, &memcell_buf);
+ size = read_cell_ul(memory, &memcell_buf);
start = _ALIGN_DOWN(start, MEMORY_INCREMENT);
size = _ALIGN_UP(size, MEMORY_INCREMENT);
- if ((start + size) > MAX_MEMORY)
- BUG();
+ numa_domain = of_node_numa_domain(memory, depth);
- tmp2 = (int *)get_property(memory, "ibm,associativity",
- NULL);
- if (!tmp2)
- continue;
- numa_domain = tmp2[depth];
-
- /* FIXME */
- if (numa_domain == 0xffff) {
- dbg("memory has no numa doman\n");
+ if (numa_domain >= MAX_NUMNODES) {
+ if (numa_domain != 0xffff)
+ printk(KERN_ERR "WARNING: memory at %lx maps "
+ "to invalid NUMA node %d\n", start,
+ numa_domain);
numa_domain = 0;
}
- if (numa_domain >= MAX_NUMNODES)
- BUG();
-
node_set_online(numa_domain);
if (max_domain < numa_domain)
@@ -205,11 +291,13 @@ new_range:
start, size);
continue;
}
- node_data[numa_domain].node_spanned_pages += size / PAGE_SIZE;
+ node_data[numa_domain].node_spanned_pages +=
+ size / PAGE_SIZE;
} else {
node_data[numa_domain].node_start_pfn =
start / PAGE_SIZE;
- node_data[numa_domain].node_spanned_pages = size / PAGE_SIZE;
+ node_data[numa_domain].node_spanned_pages =
+ size / PAGE_SIZE;
}
for (i = start ; i < (start+size); i += MEMORY_INCREMENT)
@@ -227,10 +315,6 @@ new_range:
numnodes = max_domain + 1;
return 0;
-err:
- of_node_put(cpu);
- of_node_put(memory);
- return -1;
}
static void __init setup_nonnuma(void)
diff -puN include/asm-ppc64/mmzone.h~ppc64-numafixes include/asm-ppc64/mmzone.h
--- forakpm/include/asm-ppc64/mmzone.h~ppc64-numafixes 2004-05-24 15:04:33.923731005 +1000
+++ forakpm-anton/include/asm-ppc64/mmzone.h 2004-05-24 15:04:33.950730668 +1000
@@ -23,7 +23,6 @@ extern char *numa_memory_lookup_table;
extern cpumask_t numa_cpumask_lookup_table[];
extern int nr_cpus_in_node[];
-#define MAX_MEMORY (1UL << 41)
/* 16MB regions */
#define MEMORY_INCREMENT_SHIFT 24
#define MEMORY_INCREMENT (1UL << MEMORY_INCREMENT_SHIFT)
diff -puN -L root/SLES9-curr/arch/ppc64/mm/numa.c /dev/null /dev/null
_
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2004-05-24 5:40 UTC | newest]
Thread overview: (only message) (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2004-05-24 5:35 [PATCH] [ppc64] NUMA fixes Anton Blanchard
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).