aboutsummaryrefslogtreecommitdiffstats
path: root/arch/sparc
diff options
context:
space:
mode:
authorHong H. Pham <hong.pham@windriver.com>2009-06-04 05:10:11 -0400
committerDavid S. Miller <davem@davemloft.net>2009-06-16 07:56:28 -0400
commit280ff97494e0fef4124bee5c52e39b23a18dd283 (patch)
treee906ca3c5e0a6238882d181ab5b01fb3f40ba5df /arch/sparc
parent4fd78a5f1edf62ab1ca3d23efee4a8a336edb2b6 (diff)
sparc64: fix and optimize irq distribution
irq_choose_cpu() should compare the affinity mask against cpu_online_map rather than CPU_MASK_ALL, since irq_select_affinity() sets the interrupt's affinity mask to cpu_online_map "and" CPU_MASK_ALL (which ends up being just cpu_online_map). The mask comparison in irq_choose_cpu() will always fail since the two masks are not the same. So the CPU chosen is the first CPU in the intersection of cpu_online_map and CPU_MASK_ALL, which is always CPU0. That means all interrupts are reassigned to CPU0... Distributing interrupts to CPUs in a linearly increasing round robin fashion is not optimal for the UltraSPARC T1/T2. Also, the irq_rover in irq_choose_cpu() causes an interrupt to be assigned to a different processor each time the interrupt is allocated and released. This may lead to an unbalanced distribution over time. A static mapping of interrupts to processors is done to optimize and balance interrupt distribution. For the T1/T2, interrupts are spread to different cores first, and then to strands within a core. The following is some benchmarks showing the effects of interrupt distribution on a T2. The test was done with iperf using a pair of T5220 boxes, each with a 10GBe NIU (XAUI) connected back to back. TCP | Stock Linear RR IRQ Optimized IRQ Streams | 2.6.30-rc5 Distribution Distribution | GBits/sec GBits/sec GBits/sec --------+----------------------------------------- 1 0.839 0.862 0.868 8 1.16 4.96 5.88 16 1.15 6.40 8.04 100 1.09 7.28 8.68 Signed-off-by: Hong H. Pham <hong.pham@windriver.com> Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'arch/sparc')
-rw-r--r--arch/sparc/kernel/Makefile1
-rw-r--r--arch/sparc/kernel/cpumap.c431
-rw-r--r--arch/sparc/kernel/cpumap.h16
-rw-r--r--arch/sparc/kernel/irq_64.c29
-rw-r--r--arch/sparc/kernel/smp_64.c4
5 files changed, 456 insertions, 25 deletions
diff --git a/arch/sparc/kernel/Makefile b/arch/sparc/kernel/Makefile
index 54742e58831c..47029c66b17a 100644
--- a/arch/sparc/kernel/Makefile
+++ b/arch/sparc/kernel/Makefile
@@ -54,6 +54,7 @@ obj-$(CONFIG_SPARC64) += sstate.o
54obj-$(CONFIG_SPARC64) += mdesc.o 54obj-$(CONFIG_SPARC64) += mdesc.o
55obj-$(CONFIG_SPARC64) += pcr.o 55obj-$(CONFIG_SPARC64) += pcr.o
56obj-$(CONFIG_SPARC64) += nmi.o 56obj-$(CONFIG_SPARC64) += nmi.o
57obj-$(CONFIG_SPARC64_SMP) += cpumap.o
57 58
58# sparc32 do not use GENERIC_HARDIRQS but uses the generic devres implementation 59# sparc32 do not use GENERIC_HARDIRQS but uses the generic devres implementation
59obj-$(CONFIG_SPARC32) += devres.o 60obj-$(CONFIG_SPARC32) += devres.o
diff --git a/arch/sparc/kernel/cpumap.c b/arch/sparc/kernel/cpumap.c
new file mode 100644
index 000000000000..7430ed080b23
--- /dev/null
+++ b/arch/sparc/kernel/cpumap.c
@@ -0,0 +1,431 @@
1/* cpumap.c: used for optimizing CPU assignment
2 *
3 * Copyright (C) 2009 Hong H. Pham <hong.pham@windriver.com>
4 */
5
6#include <linux/module.h>
7#include <linux/kernel.h>
8#include <linux/init.h>
9#include <linux/cpumask.h>
10#include <linux/spinlock.h>
11#include <asm/cpudata.h>
12#include "cpumap.h"
13
14
15enum {
16 CPUINFO_LVL_ROOT = 0,
17 CPUINFO_LVL_NODE,
18 CPUINFO_LVL_CORE,
19 CPUINFO_LVL_PROC,
20 CPUINFO_LVL_MAX,
21};
22
23enum {
24 ROVER_NO_OP = 0,
25 /* Increment rover every time level is visited */
26 ROVER_INC_ON_VISIT = 1 << 0,
27 /* Increment parent's rover every time rover wraps around */
28 ROVER_INC_PARENT_ON_LOOP = 1 << 1,
29};
30
31struct cpuinfo_node {
32 int id;
33 int level;
34 int num_cpus; /* Number of CPUs in this hierarchy */
35 int parent_index;
36 int child_start; /* Array index of the first child node */
37 int child_end; /* Array index of the last child node */
38 int rover; /* Child node iterator */
39};
40
41struct cpuinfo_level {
42 int start_index; /* Index of first node of a level in a cpuinfo tree */
43 int end_index; /* Index of last node of a level in a cpuinfo tree */
44 int num_nodes; /* Number of nodes in a level in a cpuinfo tree */
45};
46
47struct cpuinfo_tree {
48 int total_nodes;
49
50 /* Offsets into nodes[] for each level of the tree */
51 struct cpuinfo_level level[CPUINFO_LVL_MAX];
52 struct cpuinfo_node nodes[0];
53};
54
55
56static struct cpuinfo_tree *cpuinfo_tree;
57
58static u16 cpu_distribution_map[NR_CPUS];
59static DEFINE_SPINLOCK(cpu_map_lock);
60
61
62/* Niagara optimized cpuinfo tree traversal. */
63static const int niagara_iterate_method[] = {
64 [CPUINFO_LVL_ROOT] = ROVER_NO_OP,
65
66 /* Strands (or virtual CPUs) within a core may not run concurrently
67 * on the Niagara, as instruction pipeline(s) are shared. Distribute
68 * work to strands in different cores first for better concurrency.
69 * Go to next NUMA node when all cores are used.
70 */
71 [CPUINFO_LVL_NODE] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
72
73 /* Strands are grouped together by proc_id in cpuinfo_sparc, i.e.
74 * a proc_id represents an instruction pipeline. Distribute work to
75 * strands in different proc_id groups if the core has multiple
76 * instruction pipelines (e.g. the Niagara 2/2+ has two).
77 */
78 [CPUINFO_LVL_CORE] = ROVER_INC_ON_VISIT,
79
80 /* Pick the next strand in the proc_id group. */
81 [CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT,
82};
83
84/* Generic cpuinfo tree traversal. Distribute work round robin across NUMA
85 * nodes.
86 */
87static const int generic_iterate_method[] = {
88 [CPUINFO_LVL_ROOT] = ROVER_INC_ON_VISIT,
89 [CPUINFO_LVL_NODE] = ROVER_NO_OP,
90 [CPUINFO_LVL_CORE] = ROVER_INC_PARENT_ON_LOOP,
91 [CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
92};
93
94
95static int cpuinfo_id(int cpu, int level)
96{
97 int id;
98
99 switch (level) {
100 case CPUINFO_LVL_ROOT:
101 id = 0;
102 break;
103 case CPUINFO_LVL_NODE:
104 id = cpu_to_node(cpu);
105 break;
106 case CPUINFO_LVL_CORE:
107 id = cpu_data(cpu).core_id;
108 break;
109 case CPUINFO_LVL_PROC:
110 id = cpu_data(cpu).proc_id;
111 break;
112 default:
113 id = -EINVAL;
114 }
115 return id;
116}
117
118/*
119 * Enumerate the CPU information in __cpu_data to determine the start index,
120 * end index, and number of nodes for each level in the cpuinfo tree. The
121 * total number of cpuinfo nodes required to build the tree is returned.
122 */
123static int enumerate_cpuinfo_nodes(struct cpuinfo_level *tree_level)
124{
125 int prev_id[CPUINFO_LVL_MAX];
126 int i, n, num_nodes;
127
128 for (i = CPUINFO_LVL_ROOT; i < CPUINFO_LVL_MAX; i++) {
129 struct cpuinfo_level *lv = &tree_level[i];
130
131 prev_id[i] = -1;
132 lv->start_index = lv->end_index = lv->num_nodes = 0;
133 }
134
135 num_nodes = 1; /* Include the root node */
136
137 for (i = 0; i < num_possible_cpus(); i++) {
138 if (!cpu_online(i))
139 continue;
140
141 n = cpuinfo_id(i, CPUINFO_LVL_NODE);
142 if (n > prev_id[CPUINFO_LVL_NODE]) {
143 tree_level[CPUINFO_LVL_NODE].num_nodes++;
144 prev_id[CPUINFO_LVL_NODE] = n;
145 num_nodes++;
146 }
147 n = cpuinfo_id(i, CPUINFO_LVL_CORE);
148 if (n > prev_id[CPUINFO_LVL_CORE]) {
149 tree_level[CPUINFO_LVL_CORE].num_nodes++;
150 prev_id[CPUINFO_LVL_CORE] = n;
151 num_nodes++;
152 }
153 n = cpuinfo_id(i, CPUINFO_LVL_PROC);
154 if (n > prev_id[CPUINFO_LVL_PROC]) {
155 tree_level[CPUINFO_LVL_PROC].num_nodes++;
156 prev_id[CPUINFO_LVL_PROC] = n;
157 num_nodes++;
158 }
159 }
160
161 tree_level[CPUINFO_LVL_ROOT].num_nodes = 1;
162
163 n = tree_level[CPUINFO_LVL_NODE].num_nodes;
164 tree_level[CPUINFO_LVL_NODE].start_index = 1;
165 tree_level[CPUINFO_LVL_NODE].end_index = n;
166
167 n++;
168 tree_level[CPUINFO_LVL_CORE].start_index = n;
169 n += tree_level[CPUINFO_LVL_CORE].num_nodes;
170 tree_level[CPUINFO_LVL_CORE].end_index = n - 1;
171
172 tree_level[CPUINFO_LVL_PROC].start_index = n;
173 n += tree_level[CPUINFO_LVL_PROC].num_nodes;
174 tree_level[CPUINFO_LVL_PROC].end_index = n - 1;
175
176 return num_nodes;
177}
178
179/* Build a tree representation of the CPU hierarchy using the per CPU
180 * information in __cpu_data. Entries in __cpu_data[0..NR_CPUS] are
181 * assumed to be sorted in ascending order based on node, core_id, and
182 * proc_id (in order of significance).
183 */
184static struct cpuinfo_tree *build_cpuinfo_tree(void)
185{
186 struct cpuinfo_tree *new_tree;
187 struct cpuinfo_node *node;
188 struct cpuinfo_level tmp_level[CPUINFO_LVL_MAX];
189 int num_cpus[CPUINFO_LVL_MAX];
190 int level_rover[CPUINFO_LVL_MAX];
191 int prev_id[CPUINFO_LVL_MAX];
192 int n, id, cpu, prev_cpu, last_cpu, level;
193
194 n = enumerate_cpuinfo_nodes(tmp_level);
195
196 new_tree = kzalloc(sizeof(struct cpuinfo_tree) +
197 (sizeof(struct cpuinfo_node) * n), GFP_ATOMIC);
198 if (!new_tree)
199 return NULL;
200
201 new_tree->total_nodes = n;
202 memcpy(&new_tree->level, tmp_level, sizeof(tmp_level));
203
204 prev_cpu = cpu = first_cpu(cpu_online_map);
205
206 /* Initialize all levels in the tree with the first CPU */
207 for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT; level--) {
208 n = new_tree->level[level].start_index;
209
210 level_rover[level] = n;
211 node = &new_tree->nodes[n];
212
213 id = cpuinfo_id(cpu, level);
214 if (unlikely(id < 0)) {
215 kfree(new_tree);
216 return NULL;
217 }
218 node->id = id;
219 node->level = level;
220 node->num_cpus = 1;
221
222 node->parent_index = (level > CPUINFO_LVL_ROOT)
223 ? new_tree->level[level - 1].start_index : -1;
224
225 node->child_start = node->child_end = node->rover =
226 (level == CPUINFO_LVL_PROC)
227 ? cpu : new_tree->level[level + 1].start_index;
228
229 prev_id[level] = node->id;
230 num_cpus[level] = 1;
231 }
232
233 for (last_cpu = (num_possible_cpus() - 1); last_cpu >= 0; last_cpu--) {
234 if (cpu_online(last_cpu))
235 break;
236 }
237
238 while (++cpu <= last_cpu) {
239 if (!cpu_online(cpu))
240 continue;
241
242 for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT;
243 level--) {
244 id = cpuinfo_id(cpu, level);
245 if (unlikely(id < 0)) {
246 kfree(new_tree);
247 return NULL;
248 }
249
250 if ((id != prev_id[level]) || (cpu == last_cpu)) {
251 prev_id[level] = id;
252 node = &new_tree->nodes[level_rover[level]];
253 node->num_cpus = num_cpus[level];
254 num_cpus[level] = 1;
255
256 if (cpu == last_cpu)
257 node->num_cpus++;
258
259 /* Connect tree node to parent */
260 if (level == CPUINFO_LVL_ROOT)
261 node->parent_index = -1;
262 else
263 node->parent_index =
264 level_rover[level - 1];
265
266 if (level == CPUINFO_LVL_PROC) {
267 node->child_end =
268 (cpu == last_cpu) ? cpu : prev_cpu;
269 } else {
270 node->child_end =
271 level_rover[level + 1] - 1;
272 }
273
274 /* Initialize the next node in the same level */
275 n = ++level_rover[level];
276 if (n <= new_tree->level[level].end_index) {
277 node = &new_tree->nodes[n];
278 node->id = id;
279 node->level = level;
280
281 /* Connect node to child */
282 node->child_start = node->child_end =
283 node->rover =
284 (level == CPUINFO_LVL_PROC)
285 ? cpu : level_rover[level + 1];
286 }
287 } else
288 num_cpus[level]++;
289 }
290 prev_cpu = cpu;
291 }
292
293 return new_tree;
294}
295
296static void increment_rover(struct cpuinfo_tree *t, int node_index,
297 int root_index, const int *rover_inc_table)
298{
299 struct cpuinfo_node *node = &t->nodes[node_index];
300 int top_level, level;
301
302 top_level = t->nodes[root_index].level;
303 for (level = node->level; level >= top_level; level--) {
304 node->rover++;
305 if (node->rover <= node->child_end)
306 return;
307
308 node->rover = node->child_start;
309 /* If parent's rover does not need to be adjusted, stop here. */
310 if ((level == top_level) ||
311 !(rover_inc_table[level] & ROVER_INC_PARENT_ON_LOOP))
312 return;
313
314 node = &t->nodes[node->parent_index];
315 }
316}
317
318static int iterate_cpu(struct cpuinfo_tree *t, unsigned int root_index)
319{
320 const int *rover_inc_table;
321 int level, new_index, index = root_index;
322
323 switch (sun4v_chip_type) {
324 case SUN4V_CHIP_NIAGARA1:
325 case SUN4V_CHIP_NIAGARA2:
326 rover_inc_table = niagara_iterate_method;
327 break;
328 default:
329 rover_inc_table = generic_iterate_method;
330 }
331
332 for (level = t->nodes[root_index].level; level < CPUINFO_LVL_MAX;
333 level++) {
334 new_index = t->nodes[index].rover;
335 if (rover_inc_table[level] & ROVER_INC_ON_VISIT)
336 increment_rover(t, index, root_index, rover_inc_table);
337
338 index = new_index;
339 }
340 return index;
341}
342
343static void _cpu_map_rebuild(void)
344{
345 int i;
346
347 if (cpuinfo_tree) {
348 kfree(cpuinfo_tree);
349 cpuinfo_tree = NULL;
350 }
351
352 cpuinfo_tree = build_cpuinfo_tree();
353 if (!cpuinfo_tree)
354 return;
355
356 /* Build CPU distribution map that spans all online CPUs. No need
357 * to check if the CPU is online, as that is done when the cpuinfo
358 * tree is being built.
359 */
360 for (i = 0; i < cpuinfo_tree->nodes[0].num_cpus; i++)
361 cpu_distribution_map[i] = iterate_cpu(cpuinfo_tree, 0);
362}
363
364/* Fallback if the cpuinfo tree could not be built. CPU mapping is linear
365 * round robin.
366 */
367static int simple_map_to_cpu(unsigned int index)
368{
369 int i, end, cpu_rover;
370
371 cpu_rover = 0;
372 end = index % num_online_cpus();
373 for (i = 0; i < num_possible_cpus(); i++) {
374 if (cpu_online(cpu_rover)) {
375 if (cpu_rover >= end)
376 return cpu_rover;
377
378 cpu_rover++;
379 }
380 }
381
382 /* Impossible, since num_online_cpus() <= num_possible_cpus() */
383 return first_cpu(cpu_online_map);
384}
385
386static int _map_to_cpu(unsigned int index)
387{
388 struct cpuinfo_node *root_node;
389
390 if (unlikely(!cpuinfo_tree)) {
391 _cpu_map_rebuild();
392 if (!cpuinfo_tree)
393 return simple_map_to_cpu(index);
394 }
395
396 root_node = &cpuinfo_tree->nodes[0];
397#ifdef CONFIG_HOTPLUG_CPU
398 if (unlikely(root_node->num_cpus != num_online_cpus())) {
399 _cpu_map_rebuild();
400 if (!cpuinfo_tree)
401 return simple_map_to_cpu(index);
402 }
403#endif
404 return cpu_distribution_map[index % root_node->num_cpus];
405}
406
407int map_to_cpu(unsigned int index)
408{
409 int mapped_cpu;
410 unsigned long flag;
411
412 spin_lock_irqsave(&cpu_map_lock, flag);
413 mapped_cpu = _map_to_cpu(index);
414
415#ifdef CONFIG_HOTPLUG_CPU
416 while (unlikely(!cpu_online(mapped_cpu)))
417 mapped_cpu = _map_to_cpu(index);
418#endif
419 spin_unlock_irqrestore(&cpu_map_lock, flag);
420 return mapped_cpu;
421}
422EXPORT_SYMBOL(map_to_cpu);
423
424void cpu_map_rebuild(void)
425{
426 unsigned long flag;
427
428 spin_lock_irqsave(&cpu_map_lock, flag);
429 _cpu_map_rebuild();
430 spin_unlock_irqrestore(&cpu_map_lock, flag);
431}
diff --git a/arch/sparc/kernel/cpumap.h b/arch/sparc/kernel/cpumap.h
new file mode 100644
index 000000000000..e639880ab864
--- /dev/null
+++ b/arch/sparc/kernel/cpumap.h
@@ -0,0 +1,16 @@
1#ifndef _CPUMAP_H
2#define _CPUMAP_H
3
4#ifdef CONFIG_SMP
5extern void cpu_map_rebuild(void);
6extern int map_to_cpu(unsigned int index);
7#define cpu_map_init() cpu_map_rebuild()
8#else
9#define cpu_map_init() do {} while (0)
10static inline int map_to_cpu(unsigned int index)
11{
12 return raw_smp_processor_id();
13}
14#endif
15
16#endif
diff --git a/arch/sparc/kernel/irq_64.c b/arch/sparc/kernel/irq_64.c
index e5e78f9cfc95..bd075054942b 100644
--- a/arch/sparc/kernel/irq_64.c
+++ b/arch/sparc/kernel/irq_64.c
@@ -45,6 +45,7 @@
45#include <asm/cacheflush.h> 45#include <asm/cacheflush.h>
46 46
47#include "entry.h" 47#include "entry.h"
48#include "cpumap.h"
48 49
49#define NUM_IVECS (IMAP_INR + 1) 50#define NUM_IVECS (IMAP_INR + 1)
50 51
@@ -256,35 +257,13 @@ static int irq_choose_cpu(unsigned int virt_irq)
256 int cpuid; 257 int cpuid;
257 258
258 cpumask_copy(&mask, irq_desc[virt_irq].affinity); 259 cpumask_copy(&mask, irq_desc[virt_irq].affinity);
259 if (cpus_equal(mask, CPU_MASK_ALL)) { 260 if (cpus_equal(mask, cpu_online_map)) {
260 static int irq_rover; 261 cpuid = map_to_cpu(virt_irq);
261 static DEFINE_SPINLOCK(irq_rover_lock);
262 unsigned long flags;
263
264 /* Round-robin distribution... */
265 do_round_robin:
266 spin_lock_irqsave(&irq_rover_lock, flags);
267
268 while (!cpu_online(irq_rover)) {
269 if (++irq_rover >= nr_cpu_ids)
270 irq_rover = 0;
271 }
272 cpuid = irq_rover;
273 do {
274 if (++irq_rover >= nr_cpu_ids)
275 irq_rover = 0;
276 } while (!cpu_online(irq_rover));
277
278 spin_unlock_irqrestore(&irq_rover_lock, flags);
279 } else { 262 } else {
280 cpumask_t tmp; 263 cpumask_t tmp;
281 264
282 cpus_and(tmp, cpu_online_map, mask); 265 cpus_and(tmp, cpu_online_map, mask);
283 266 cpuid = cpus_empty(tmp) ? map_to_cpu(virt_irq) : first_cpu(tmp);
284 if (cpus_empty(tmp))
285 goto do_round_robin;
286
287 cpuid = first_cpu(tmp);
288 } 267 }
289 268
290 return cpuid; 269 return cpuid;
diff --git a/arch/sparc/kernel/smp_64.c b/arch/sparc/kernel/smp_64.c
index 1de47d2169c8..cfb3d06058ff 100644
--- a/arch/sparc/kernel/smp_64.c
+++ b/arch/sparc/kernel/smp_64.c
@@ -48,6 +48,8 @@
48#include <asm/ldc.h> 48#include <asm/ldc.h>
49#include <asm/hypervisor.h> 49#include <asm/hypervisor.h>
50 50
51#include "cpumap.h"
52
51int sparc64_multi_core __read_mostly; 53int sparc64_multi_core __read_mostly;
52 54
53DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE; 55DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
@@ -1314,6 +1316,8 @@ int __cpu_disable(void)
1314 cpu_clear(cpu, cpu_online_map); 1316 cpu_clear(cpu, cpu_online_map);
1315 ipi_call_unlock(); 1317 ipi_call_unlock();
1316 1318
1319 cpu_map_rebuild();
1320
1317 return 0; 1321 return 0;
1318} 1322}
1319 1323