Re: finish_task_switch high in profiles in 2.6.7

[William Lee Irwin III]

On Thu, Jun 24, 2004 at 02:22:48PM -0700, William Lee Irwin III wrote:
> Brute-force port of schedprof to 2.6.7-final. Compiletested on sparc64
> only. No runtime testing.
> Given that the context switch rate is actually *reduced* in 2.6.7 vs.
> 2.6.5, I expect that this will not, in fact, reveal anything useful.

While I'm spraying out untested code, I might as well do these, which
I've not even compiled. =)


-- wli
Index: schedprof-2.6.7/kernel/sched.c
===================================================================
--- schedprof-2.6.7.orig/kernel/sched.c	2004-06-24 14:02:48.292038264 -0700
+++ schedprof-2.6.7/kernel/sched.c	2004-06-24 14:39:50.035281952 -0700
@@ -4063,7 +4063,7 @@
 	if (schedprof_buf) {
 		unsigned long pc = (unsigned long)__pc;
 		pc -= min(pc, (unsigned long)_stext);
-		atomic_inc(&schedprof_buf[min(pc, schedprof_len)]);
+		atomic_inc(&schedprof_buf[min(pc + 1, schedprof_len - 1)]);
 	}
 }
 
@@ -4081,8 +4081,9 @@
 	if (!sched_profiling)
 		return;
 	schedprof_len = (unsigned long)(_etext - _stext) + 1;
-	schedprof_buf = alloc_bootmem(schedprof_len*sizeof(atomic_t));
+	schedprof_buf = alloc_bootmem(sizeof(atomic_t)*schedprof_len);
 	printk(KERN_INFO "Scheduler call profiling enabled\n");
+	schedprof_buf[0] = 1;
 }
 
 #ifdef CONFIG_PROC_FS
@@ -4092,38 +4093,59 @@
 read_sched_profile(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 {
 	unsigned long p = *ppos;
-	ssize_t read;
-	char * pnt;
 	unsigned int sample_step = 1;
 
-	if (p >= (schedprof_len+1)*sizeof(atomic_t))
+	if (p >= sizeof(atomic_t)*schedprof_len)
 		return 0;
-	if (count > (schedprof_len+1)*sizeof(atomic_t) - p)
-		count = (schedprof_len+1)*sizeof(atomic_t) - p;
-	read = 0;
-
-	while (p < sizeof(atomic_t) && count > 0) {
-		put_user(*((char *)(&sample_step)+p),buf);
-		buf++; p++; count--; read++;
-	}
-	pnt = (char *)schedprof_buf + p - sizeof(atomic_t);
-	if (copy_to_user(buf,(void *)pnt,count))
+	count = min(schedprof_len*sizeof(atomic_t) - p, count);
+	if (copy_to_user(buf, (char *)schedprof_buf + p, count))
 		return -EFAULT;
-	read += count;
-	*ppos += read;
-	return read;
+	*ppos += count;
+	return count;
 }
 
 static ssize_t write_sched_profile(struct file *file, const char __user *buf,
 			     size_t count, loff_t *ppos)
 {
-	memset(schedprof_buf, 0, sizeof(atomic_t)*schedprof_len);
+	memset(&schedprof_buf[1], 0, (schedprof_len-1)*sizeof(atomic_t));
 	return count;
 }
 
+static int mmap_sched_profile(struct file *file, struct vm_area_struct *vma)
+{
+	unsigned long pfn, vaddr, base_pfn = __pa(schedprof_buf)/PAGE_SIZE;
+	if (vma->vm_pgoff + vma_pages(vma) > schedprof_pages)
+		return -ENODEV;
+	vma->vm_flags |= VM_RESERVED|VM_IO;
+	for (vaddr = vma->vm_start; vaddr < vma->vm_end; vaddr += PAGE_SIZE) {
+		pgd_t *pgd = pgd_offset(vma->vm_mm, vaddr);
+		pmd_t *pmd;
+		pte_t *pte, pte_val;
+
+		spin_lock(&vma->vm_mm->page_table_lock);
+		pmd = pmd_alloc(vma->vm_mm, pgd, vaddr);
+		if (!pmd)
+			goto enomem;
+		pte = pte_alloc_map(vma->vm_mm, pmd, vaddr);
+		if (!pte)
+			goto enomem;
+		pfn = base_pfn + linear_page_index(vma, vaddr);
+		pte_val = pfn_pte(pfn, vma->vm_page_prot);
+		set_pte(pte, pte_val);
+		update_mmu_cache(vma, vaddr, pte_val);
+		pte_unmap(pte);
+		spin_unlock(&vma->vm_mm->page_table_lock);
+	}
+	return 0;
+enomem:
+	spin_unlock(&vma->vm_mm->page_table_lock);
+	return -ENOMEM;
+}
+
 static struct file_operations sched_profile_operations = {
 	.read		= read_sched_profile,
 	.write		= write_sched_profile,
+	.mmap		= mmap_sched_profile,
 };
 
 static int proc_schedprof_init(void)
@@ -4134,7 +4156,7 @@
 	entry = create_proc_entry("schedprof", S_IWUSR | S_IRUGO, NULL);
 	if (entry) {
 		entry->proc_fops = &sched_profile_operations;
-		entry->size = sizeof(atomic_t)*(schedprof_len + 1);
+		entry->size = sizeof(atomic_t)*schedprof_len;
 	}
 	return !!entry;
 }
Index: schedprof-2.6.7/kernel/sched.c
===================================================================
--- schedprof-2.6.7.orig/kernel/sched.c	2004-06-24 14:39:50.035281952 -0700
+++ schedprof-2.6.7/kernel/sched.c	2004-06-24 14:40:04.347106224 -0700
@@ -4152,13 +4152,13 @@
 {
 	struct proc_dir_entry *entry;
 	if (!sched_profiling)
-		return 1;
+		return 0;
 	entry = create_proc_entry("schedprof", S_IWUSR | S_IRUGO, NULL);
 	if (entry) {
 		entry->proc_fops = &sched_profile_operations;
 		entry->size = sizeof(atomic_t)*schedprof_len;
 	}
-	return !!entry;
+	return !entry;
 }
 module_init(proc_schedprof_init);
 #endif
Index: schedprof-2.6.7/kernel/sched.c
===================================================================
--- schedprof-2.6.7.orig/kernel/sched.c	2004-06-24 14:40:04.347106224 -0700
+++ schedprof-2.6.7/kernel/sched.c	2004-06-24 14:51:35.285067688 -0700
@@ -4052,7 +4052,7 @@
 #endif /* defined(CONFIG_SMP) && defined(CONFIG_PREEMPT) */
 
 static atomic_t *schedprof_buf;
-static int sched_profiling;
+static int sched_profiling, schedprof_shift;
 static unsigned long schedprof_len;
 
 #include <linux/bootmem.h>
@@ -4062,11 +4062,22 @@
 {
 	if (schedprof_buf) {
 		unsigned long pc = (unsigned long)__pc;
-		pc -= min(pc, (unsigned long)_stext);
+		pc = (pc - min(pc, (unsigned long)_stext)) >> schedprof_shift;
 		atomic_inc(&schedprof_buf[min(pc + 1, schedprof_len - 1)]);
 	}
 }
 
+static int __init schedprof_shift_setup(char *s)
+{
+	int n;
+	if (get_option(&s, &n)) {
+		scheprof_shift = n;
+		sched_profiling = 1;
+	}
+	return 1;
+}
+__setup("schedprof_shift=", schedprof_shift_setup);
+
 static int __init schedprof_setup(char *s)
 {
 	int n;
@@ -4080,10 +4091,10 @@
 {
 	if (!sched_profiling)
 		return;
-	schedprof_len = (unsigned long)(_etext - _stext) + 1;
+	schedprof_len = ((unsigned long)(_etext - _stext) >> schedprof_shift) + 1;
 	schedprof_buf = alloc_bootmem(sizeof(atomic_t)*schedprof_len);
 	printk(KERN_INFO "Scheduler call profiling enabled\n");
-	schedprof_buf[0] = 1;
+	schedprof_buf[0] = 1 << schedprof_shift;
 }
 
 #ifdef CONFIG_PROC_FS