Zerocopy networking patch, 2.4.0-1

Quoted 5 lines

In user space, how do you know when its safe to reuse the buffer that was handed to sendmsg() with the MSG_NOCOPY flag? Or does sendmsg() with that flag block until the buffer isn't needed by the kernel any more? If it does block, doesn't that defeat the use of non-blocking I/O?

sendmsg() marks those pages COW and copies the original page into a new one for further usage. (the old page is used until the packet is released.) So for maximum performance user-space should not reuse such buffers immediately.

That means sendmsg() changes the page tables? I measures smp_call_function on my Dual Pentium 350, and it took around 1950 cpu ticks.

Quoted 3 lines · Manfred Spraul

That means sendmsg() changes the page tables? I measures smp_call_function on my Dual Pentium 350, and it took around 1950 cpu ticks.

well, this is a performance problem if you are using threads. For normal processes there is no need for a SMP cross-call, there TLB flushes are local only.

Quoted 4 lines

well, this is a performance problem if you are using threads. For normal processes there is no need for a SMP cross-call, there TLB flushes are local only.

But that would be ugly as hell:
so apache 2.0 would become slower with MSG_NOCOPY, whereas samba 2.2 would become faster.

Is is possible to move the responsibility for maitaining the copy to the caller?

e.g. use msg_control, and then the caller can request either that a signal is sent when that data is transfered, or that a variable is set to 0.

Quoted 7 lines

Quoted 4 lines

well, this is a performance problem if you are using threads. For normal processes there is no need for a SMP cross-call, there TLB flushes are local only.

But that would be ugly as hell:
so apache 2.0 would become slower with MSG_NOCOPY, whereas samba 2.2 would become faster.

there *is* a cost of having a shared VM - and this is i suspect unavoidable.

Quoted 11 lines

Quoted 5 lines

In user space, how do you know when its safe to reuse the buffer that was handed to sendmsg() with the MSG_NOCOPY flag? Or does sendmsg() with that flag block until the buffer isn't needed by the kernel any more? If it does block, doesn't that defeat the use of non-blocking I/O?

sendmsg() marks those pages COW and copies the original page into a new one for further usage. (the old page is used until the packet is released.) So for maximum performance user-space should not reuse such buffers immediately.

That means sendmsg() changes the page tables? I measures smp_call_function on my Dual Pentium 350, and it took around 1950 cpu ticks.
I'm sure that for an 8 way server the total lost time on all cpus (multi threaded server) is larger than the time required to copy the complete page.
(I've attached my patch, just run "insmod dummy p_shift=0")

--
Manfred

drivers/net/dummy.c +159 -0

@@ -95,9 +95,168 @@
 
 static struct net_device dev_dummy;
 
+/* ************************************* */
+int p_shift = -1;
+MODULE_PARM     (p_shift, "1i");
+MODULE_PARM_DESC(p_shift, "Shift for the profile buffer");
+
+int p_size = 0;
+MODULE_PARM     (p_size, "1i");
+MODULE_PARM_DESC(p_size, "size");
+
+
+#define STAT_TABLELEN		16384
+static unsigned long totals[STAT_TABLELEN];
+static unsigned int overflows;
+
+static unsigned long long stime;
+static void start_measure(void)
+{
+	 __asm__ __volatile__ (
+		".align 64\n\t"
+	 	"pushal\n\t"
+		"cpuid\n\t"
+		"popal\n\t"
+		"rdtsc\n\t"
+		"movl %%eax,(%0)\n\t"
+		"movl %%edx,4(%0)\n\t"
+		: /* no output */
+		: "c"(&stime)
+		: "eax", "edx", "memory" );
+}
+
+static void end_measure(void)
+{
+static unsigned long long etime;
+	__asm__ __volatile__ (
+		"pushal\n\t"
+		"cpuid\n\t"
+		"popal\n\t"
+		"rdtsc\n\t"
+		"movl %%eax,(%0)\n\t"
+		"movl %%edx,4(%0)\n\t"
+		: /* no output */
+		: "c"(&etime)
+		: "eax", "edx", "memory" );
+	{
+		unsigned long time = (unsigned long)(etime-stime);
+		time >>= p_shift;
+		if(time < STAT_TABLELEN) {
+			totals[time]++;
+		} else {
+			overflows++;
+		}
+	}
+}
+
+static void clean_buf(void)
+{
+	memset(totals,0,sizeof(totals));
+	overflows = 0;
+}
+
+static void print_line(unsigned long* array)
+{
+	int i;
+	for(i=0;i<32;i++) {
+		if((i%32)==16)
+			printk(":");
+		printk("%lx ",array[i]); 
+	}
+}
+
+static void print_buf(char* caption)
+{
+	int i, other = 0;
+	printk("Results - %s - shift %d",
+		caption, p_shift);
+
+	for(i=0;i<STAT_TABLELEN;i+=32) {
+		int j;
+		int local = 0;
+		for(j=0;j<32;j++)
+			local += totals[i+j];
+
+		if(local) {
+			printk("\n%3x: ",i);
+			print_line(&totals[i]);
+			other += local;
+		}
+	}
+	printk("\nOverflows: %d.\n",
+		overflows);
+	printk("Sum: %ld\n",other+overflows);
+}
+
+static void return_immediately(void* dummy)
+{
+	return;
+}
+
+static void just_one_page(void* dummy)
+{
+	__flush_tlb_one(0x12345678);
+	return;
+}
+
+
 static int __init dummy_init_module(void)
 {
 	int err;
+
+	if(p_shift != -1) {
+		int i;
+		void* p;
+		kmem_cache_t* cachep;
+		/* empty test measurement: */
+		printk("******** kernel cpu benchmark started **********\n");
+		clean_buf();
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(200);
+		for(i=0;i<100;i++) {
+			start_measure();
+			return_immediately(NULL);
+			return_immediately(NULL);
+			return_immediately(NULL);
+			return_immediately(NULL);
+			end_measure();
+		}
+		print_buf("zero");
+		clean_buf();
+
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(200);
+		for(i=0;i<100;i++) {
+			start_measure();
+			return_immediately(NULL);
+			return_immediately(NULL);
+			smp_call_function(return_immediately,NULL,
+						1, 1);
+			return_immediately(NULL);
+			return_immediately(NULL);
+			end_measure();
+		}
+		print_buf("empty smp_call_function()");
+		clean_buf();
+
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(200);
+		for(i=0;i<100;i++) {
+			start_measure();
+			return_immediately(NULL);
+			return_immediately(NULL);
+			smp_call_function(just_one_page,NULL,
+						1, 1);
+			just_one_page(NULL);
+			return_immediately(NULL);
+			return_immediately(NULL);
+			end_measure();
+		}
+		print_buf("flush_one_page()");
+		clean_buf();	
+
+		return -EINVAL;
+	}
 
 	dev_dummy.init = dummy_init;
 	SET_MODULE_OWNER(&dev_dummy);