/*
* linux/kernel/profile.c
+ * Simple profiling. Manages a direct-mapped profile hit count buffer,
+ * with configurable resolution, support for restricting the cpus on
+ * which profiling is done, and switching between cpu time and
+ * schedule() calls via kernel command line parameters passed at boot.
+ *
+ * Scheduler profiling support, Arjan van de Ven and Ingo Molnar,
+ * Red Hat, July 2004
+ * Consolidation of architecture support code for profiling,
+ * William Irwin, Oracle, July 2004
+ * Amortized hit count accounting via per-cpu open-addressed hashtables
+ * to resolve timer interrupt livelocks, William Irwin, Oracle, 2004
*/
#include <linux/config.h>
#include <linux/notifier.h>
#include <linux/mm.h>
#include <linux/cpumask.h>
+#include <linux/cpu.h>
#include <linux/profile.h>
+#include <linux/highmem.h>
#include <asm/sections.h>
+#include <asm/semaphore.h>
+
+struct profile_hit {
+ u32 pc, hits;
+};
+#define PROFILE_GRPSHIFT 3
+#define PROFILE_GRPSZ (1 << PROFILE_GRPSHIFT)
+#define NR_PROFILE_HIT (PAGE_SIZE/sizeof(struct profile_hit))
+#define NR_PROFILE_GRP (NR_PROFILE_HIT/PROFILE_GRPSZ)
+
+/* Oprofile timer tick hook */
+int (*timer_hook)(struct pt_regs *);
static atomic_t *prof_buffer;
static unsigned long prof_len, prof_shift;
static int prof_on;
static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
+#ifdef CONFIG_SMP
+static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
+static DEFINE_PER_CPU(int, cpu_profile_flip);
+static DECLARE_MUTEX(profile_flip_mutex);
+#endif /* CONFIG_SMP */
static int __init profile_setup(char * str)
{
int par;
if (!strncmp(str, "schedule", 8)) {
- prof_on = 2;
+ prof_on = SCHED_PROFILING;
printk(KERN_INFO "kernel schedule profiling enabled\n");
if (str[7] == ',')
str += 8;
}
if (get_option(&str,&par)) {
prof_shift = par;
- prof_on = 1;
+ prof_on = CPU_PROFILING;
printk(KERN_INFO "kernel profiling enabled (shift: %ld)\n",
prof_shift);
}
#ifdef CONFIG_PROFILING
static DECLARE_RWSEM(profile_rwsem);
-static rwlock_t handoff_lock = RW_LOCK_UNLOCKED;
+static DEFINE_RWLOCK(handoff_lock);
static struct notifier_block * task_exit_notifier;
static struct notifier_block * task_free_notifier;
static struct notifier_block * munmap_notifier;
return err;
}
-static struct notifier_block * profile_listeners;
-static rwlock_t profile_lock = RW_LOCK_UNLOCKED;
-
-int register_profile_notifier(struct notifier_block * nb)
+int register_timer_hook(int (*hook)(struct pt_regs *))
{
- int err;
- write_lock_irq(&profile_lock);
- err = notifier_chain_register(&profile_listeners, nb);
- write_unlock_irq(&profile_lock);
- return err;
-}
-
-
-int unregister_profile_notifier(struct notifier_block * nb)
-{
- int err;
- write_lock_irq(&profile_lock);
- err = notifier_chain_unregister(&profile_listeners, nb);
- write_unlock_irq(&profile_lock);
- return err;
+ if (timer_hook)
+ return -EBUSY;
+ timer_hook = hook;
+ return 0;
}
-
-void profile_hook(struct pt_regs * regs)
+void unregister_timer_hook(int (*hook)(struct pt_regs *))
{
- read_lock(&profile_lock);
- notifier_call_chain(&profile_listeners, 0, regs);
- read_unlock(&profile_lock);
+ WARN_ON(hook != timer_hook);
+ timer_hook = NULL;
+ /* make sure all CPUs see the NULL hook */
+ synchronize_kernel();
}
-EXPORT_SYMBOL_GPL(register_profile_notifier);
-EXPORT_SYMBOL_GPL(unregister_profile_notifier);
+EXPORT_SYMBOL_GPL(register_timer_hook);
+EXPORT_SYMBOL_GPL(unregister_timer_hook);
EXPORT_SYMBOL_GPL(task_handoff_register);
EXPORT_SYMBOL_GPL(task_handoff_unregister);
EXPORT_SYMBOL_GPL(profile_event_register);
EXPORT_SYMBOL_GPL(profile_event_unregister);
+#ifdef CONFIG_SMP
+/*
+ * Each cpu has a pair of open-addressed hashtables for pending
+ * profile hits. read_profile() IPI's all cpus to request them
+ * to flip buffers and flushes their contents to prof_buffer itself.
+ * Flip requests are serialized by the profile_flip_mutex. The sole
+ * use of having a second hashtable is for avoiding cacheline
+ * contention that would otherwise happen during flushes of pending
+ * profile hits required for the accuracy of reported profile hits
+ * and so resurrect the interrupt livelock issue.
+ *
+ * The open-addressed hashtables are indexed by profile buffer slot
+ * and hold the number of pending hits to that profile buffer slot on
+ * a cpu in an entry. When the hashtable overflows, all pending hits
+ * are accounted to their corresponding profile buffer slots with
+ * atomic_add() and the hashtable emptied. As numerous pending hits
+ * may be accounted to a profile buffer slot in a hashtable entry,
+ * this amortizes a number of atomic profile buffer increments likely
+ * to be far larger than the number of entries in the hashtable,
+ * particularly given that the number of distinct profile buffer
+ * positions to which hits are accounted during short intervals (e.g.
+ * several seconds) is usually very small. Exclusion from buffer
+ * flipping is provided by interrupt disablement (note that for
+ * SCHED_PROFILING profile_hit() may be called from process context).
+ * The hash function is meant to be lightweight as opposed to strong,
+ * and was vaguely inspired by ppc64 firmware-supported inverted
+ * pagetable hash functions, but uses a full hashtable full of finite
+ * collision chains, not just pairs of them.
+ *
+ * -- wli
+ */
+static void __profile_flip_buffers(void *unused)
+{
+ int cpu = smp_processor_id();
+
+ per_cpu(cpu_profile_flip, cpu) = !per_cpu(cpu_profile_flip, cpu);
+}
+
+static void profile_flip_buffers(void)
+{
+ int i, j, cpu;
+
+ down(&profile_flip_mutex);
+ j = per_cpu(cpu_profile_flip, get_cpu());
+ put_cpu();
+ on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
+ for_each_online_cpu(cpu) {
+ struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[j];
+ for (i = 0; i < NR_PROFILE_HIT; ++i) {
+ if (!hits[i].hits) {
+ if (hits[i].pc)
+ hits[i].pc = 0;
+ continue;
+ }
+ atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
+ hits[i].hits = hits[i].pc = 0;
+ }
+ }
+ up(&profile_flip_mutex);
+}
+
+static void profile_discard_flip_buffers(void)
+{
+ int i, cpu;
+
+ down(&profile_flip_mutex);
+ i = per_cpu(cpu_profile_flip, get_cpu());
+ put_cpu();
+ on_each_cpu(__profile_flip_buffers, NULL, 0, 1);
+ for_each_online_cpu(cpu) {
+ struct profile_hit *hits = per_cpu(cpu_profile_hits, cpu)[i];
+ memset(hits, 0, NR_PROFILE_HIT*sizeof(struct profile_hit));
+ }
+ up(&profile_flip_mutex);
+}
+
+void profile_hit(int type, void *__pc)
+{
+ unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
+ int i, j, cpu;
+ struct profile_hit *hits;
+
+ if (prof_on != type || !prof_buffer)
+ return;
+ pc = min((pc - (unsigned long)_stext) >> prof_shift, prof_len - 1);
+ i = primary = (pc & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
+ secondary = (~(pc << 1) & (NR_PROFILE_GRP - 1)) << PROFILE_GRPSHIFT;
+ cpu = get_cpu();
+ hits = per_cpu(cpu_profile_hits, cpu)[per_cpu(cpu_profile_flip, cpu)];
+ if (!hits) {
+ put_cpu();
+ return;
+ }
+ local_irq_save(flags);
+ do {
+ for (j = 0; j < PROFILE_GRPSZ; ++j) {
+ if (hits[i + j].pc == pc) {
+ hits[i + j].hits++;
+ goto out;
+ } else if (!hits[i + j].hits) {
+ hits[i + j].pc = pc;
+ hits[i + j].hits = 1;
+ goto out;
+ }
+ }
+ i = (i + secondary) & (NR_PROFILE_HIT - 1);
+ } while (i != primary);
+ atomic_inc(&prof_buffer[pc]);
+ for (i = 0; i < NR_PROFILE_HIT; ++i) {
+ atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
+ hits[i].pc = hits[i].hits = 0;
+ }
+out:
+ local_irq_restore(flags);
+ put_cpu();
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+static int __devinit profile_cpu_callback(struct notifier_block *info,
+ unsigned long action, void *__cpu)
+{
+ int node, cpu = (unsigned long)__cpu;
+ struct page *page;
+
+ switch (action) {
+ case CPU_UP_PREPARE:
+ node = cpu_to_node(cpu);
+ per_cpu(cpu_profile_flip, cpu) = 0;
+ if (!per_cpu(cpu_profile_hits, cpu)[1]) {
+ page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
+ if (!page)
+ return NOTIFY_BAD;
+ per_cpu(cpu_profile_hits, cpu)[1] = page_address(page);
+ }
+ if (!per_cpu(cpu_profile_hits, cpu)[0]) {
+ page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
+ if (!page)
+ goto out_free;
+ per_cpu(cpu_profile_hits, cpu)[0] = page_address(page);
+ }
+ break;
+ out_free:
+ page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
+ per_cpu(cpu_profile_hits, cpu)[1] = NULL;
+ __free_page(page);
+ return NOTIFY_BAD;
+ case CPU_ONLINE:
+ cpu_set(cpu, prof_cpu_mask);
+ break;
+ case CPU_UP_CANCELED:
+ case CPU_DEAD:
+ cpu_clear(cpu, prof_cpu_mask);
+ if (per_cpu(cpu_profile_hits, cpu)[0]) {
+ page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
+ per_cpu(cpu_profile_hits, cpu)[0] = NULL;
+ __free_page(page);
+ }
+ if (per_cpu(cpu_profile_hits, cpu)[1]) {
+ page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
+ per_cpu(cpu_profile_hits, cpu)[1] = NULL;
+ __free_page(page);
+ }
+ break;
+ }
+ return NOTIFY_OK;
+}
+#endif /* CONFIG_HOTPLUG_CPU */
+#else /* !CONFIG_SMP */
+#define profile_flip_buffers() do { } while (0)
+#define profile_discard_flip_buffers() do { } while (0)
+
void profile_hit(int type, void *__pc)
{
unsigned long pc;
pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
atomic_inc(&prof_buffer[min(pc, prof_len - 1)]);
}
+#endif /* !CONFIG_SMP */
void profile_tick(int type, struct pt_regs *regs)
{
- if (type == CPU_PROFILING)
- profile_hook(regs);
+ if (type == CPU_PROFILING && timer_hook)
+ timer_hook(regs);
if (!user_mode(regs) && cpu_isset(smp_processor_id(), prof_cpu_mask))
profile_hit(type, (void *)profile_pc(regs));
}
char * pnt;
unsigned int sample_step = 1 << prof_shift;
+ profile_flip_buffers();
if (p >= (prof_len+1)*sizeof(unsigned int))
return 0;
if (count > (prof_len+1)*sizeof(unsigned int) - p)
return -EINVAL;
}
#endif
-
+ profile_discard_flip_buffers();
memset(prof_buffer, 0, prof_len * sizeof(atomic_t));
return count;
}
.write = write_profile,
};
+#ifdef CONFIG_SMP
+static void __init profile_nop(void *unused)
+{
+}
+
+static int __init create_hash_tables(void)
+{
+ int cpu;
+
+ for_each_online_cpu(cpu) {
+ int node = cpu_to_node(cpu);
+ struct page *page;
+
+ page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
+ if (!page)
+ goto out_cleanup;
+ per_cpu(cpu_profile_hits, cpu)[1]
+ = (struct profile_hit *)page_address(page);
+ page = alloc_pages_node(node, GFP_KERNEL | __GFP_ZERO, 0);
+ if (!page)
+ goto out_cleanup;
+ per_cpu(cpu_profile_hits, cpu)[0]
+ = (struct profile_hit *)page_address(page);
+ }
+ return 0;
+out_cleanup:
+ prof_on = 0;
+ mb();
+ on_each_cpu(profile_nop, NULL, 0, 1);
+ for_each_online_cpu(cpu) {
+ struct page *page;
+
+ if (per_cpu(cpu_profile_hits, cpu)[0]) {
+ page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[0]);
+ per_cpu(cpu_profile_hits, cpu)[0] = NULL;
+ __free_page(page);
+ }
+ if (per_cpu(cpu_profile_hits, cpu)[1]) {
+ page = virt_to_page(per_cpu(cpu_profile_hits, cpu)[1]);
+ per_cpu(cpu_profile_hits, cpu)[1] = NULL;
+ __free_page(page);
+ }
+ }
+ return -1;
+}
+#else
+#define create_hash_tables() ({ 0; })
+#endif
+
static int __init create_proc_profile(void)
{
struct proc_dir_entry *entry;
if (!prof_on)
return 0;
+ if (create_hash_tables())
+ return -1;
if (!(entry = create_proc_entry("profile", S_IWUSR | S_IRUGO, NULL)))
return 0;
entry->proc_fops = &proc_profile_operations;
entry->size = (1+prof_len) * sizeof(atomic_t);
+ hotcpu_notifier(profile_cpu_callback, 0);
return 0;
}
module_init(create_proc_profile);
git://git.onelab.eu / linux-2.6.git / blobdiff