All of lore.kernel.org
 help / color / mirror / Atom feed
From: Samuel Rydh <samuel@ibrium.se>
To: linuxppc-dev@lists.linuxppc.org
Cc: paulus@samba.org
Subject: Re: [PATCH] 2.6 timebase synchronization
Date: Sat, 9 Aug 2003 03:13:05 +0200	[thread overview]
Message-ID: <20030809011305.GA12030@ibrium.se> (raw)
In-Reply-To: <20030808001523.GA9274@ibrium.se>


On Fri, Aug 08, 2003 at 02:15:23AM +0200, Samuel Rydh wrote:
>
> Hi,
>
> Here is a patch that provides improved generic SMP timebase
> synchronization for 2.6.
>

Attached is an improved version of the patch.

/Samuel



diff -Nru a/arch/ppc/kernel/Makefile b/arch/ppc/kernel/Makefile
--- a/arch/ppc/kernel/Makefile	Sat Aug  9 03:10:58 2003
+++ b/arch/ppc/kernel/Makefile	Sat Aug  9 03:10:58 2003
@@ -25,7 +25,7 @@
 obj-$(CONFIG_PCI)		+= pci.o
 obj-$(CONFIG_PCI)		+= pci-dma.o
 obj-$(CONFIG_KGDB)		+= ppc-stub.o
-obj-$(CONFIG_SMP)		+= smp.o
+obj-$(CONFIG_SMP)		+= smp.o smp-tbsync.o
 obj-$(CONFIG_TAU)		+= temp.o

 ifdef CONFIG_MATH_EMULATION
diff -Nru a/arch/ppc/kernel/smp-tbsync.c b/arch/ppc/kernel/smp-tbsync.c
--- /dev/null	Wed Dec 31 16:00:00 1969
+++ b/arch/ppc/kernel/smp-tbsync.c	Sat Aug  9 03:10:58 2003
@@ -0,0 +1,181 @@
+/*
+ * Smp timebase synchronization for ppc.
+ *
+ * Copyright (C) 2003 Samuel Rydh (samuel@ibrium.se)
+ *
+ */
+
+#include <linux/config.h>
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/smp.h>
+#include <linux/unistd.h>
+#include <linux/init.h>
+#include <asm/atomic.h>
+#include <asm/smp.h>
+#include <asm/time.h>
+
+#define NUM_ITER		300
+
+enum {
+	kExit=0, kSetAndTest, kTest
+};
+
+static struct {
+	volatile int		tbu;
+	volatile int		tbl;
+	volatile int		mark;
+	volatile int		cmd;
+	volatile int		handshake;
+	int			filler[3];
+
+	volatile int		ack;
+	int			filler2[7];
+
+	volatile int		race_result;
+} *tbsync;
+
+static volatile int		running;
+
+static void __devinit
+enter_contest( int mark, int add )
+{
+	while( (int)(get_tbl() - mark) < 0 )
+		tbsync->race_result = add;
+}
+
+void __devinit
+smp_generic_take_timebase( void )
+{
+	int cmd, tbl, tbu;
+
+	local_irq_disable();
+	while( !running )
+		;
+	rmb();
+
+	for( ;; ) {
+		tbsync->ack = 1;
+		while( !tbsync->handshake )
+			;
+		rmb();
+
+		cmd = tbsync->cmd;
+		tbl = tbsync->tbl;
+		tbu = tbsync->tbu;
+		tbsync->ack = 0;
+		if( cmd == kExit )
+			return;
+
+		if( cmd == kSetAndTest ) {
+			while( tbsync->handshake )
+				;
+			asm volatile ("mttbl %0" :: "r" (tbl) );
+			asm volatile ("mttbu %0" :: "r" (tbu) );
+		} else {
+			while( tbsync->handshake )
+				;
+		}
+		enter_contest( tbsync->mark, -1 );
+	}
+	local_irq_enable();
+}
+
+static int __devinit
+start_contest( int cmd, int offset, int num )
+{
+	int i, tbu, tbl, mark, score=0;
+
+	tbsync->cmd = cmd;
+
+	local_irq_disable();
+	for( i=-3; i<num; ) {
+		tbl = get_tbl() + 400;
+		tbsync->tbu = tbu = get_tbu();
+		tbsync->tbl = tbl + offset;
+		tbsync->mark = mark = tbl + 400;
+
+		wmb();
+
+		tbsync->handshake = 1;
+		while( tbsync->ack )
+			;
+
+		while( (int)(get_tbl() - tbl) <= 0 )
+			;
+		tbsync->handshake = 0;
+		enter_contest( mark, 1 );
+
+		while( !tbsync->ack )
+			;
+
+		if( tbsync->tbu != get_tbu() || ((tbsync->tbl ^ get_tbl()) & 0x80000000) )
+			continue;
+		if( i++ > 0 )
+			score += tbsync->race_result;
+	}
+	local_irq_enable();
+	return score;
+}
+
+void __devinit
+smp_generic_give_timebase( void )
+{
+	int i, score, score2, old, min=0, max=5000, offset=1000;
+
+	printk("Synchronizing timebase\n");
+
+	/* if this fails then this kernel won't work anyway... */
+	tbsync = kmalloc( sizeof(*tbsync), GFP_KERNEL );
+	memset( tbsync, 0, sizeof(*tbsync) );
+	mb();
+	running = 1;
+
+	while( !tbsync->ack )
+		;
+
+	/* binary search */
+	for( old=-1 ; old != offset ; offset=(min+max)/2 ) {
+		score = start_contest( kSetAndTest, offset, NUM_ITER );
+
+		printk("score %d, offset %d\n", score, offset );
+
+		if( score > 0 )
+			max = offset;
+		else
+			min = offset;
+		old = offset;
+	}
+	score = start_contest( kSetAndTest, min, NUM_ITER );
+	score2 = start_contest( kSetAndTest, max, NUM_ITER );
+
+	printk( "Min %d (score %d), Max %d (score %d)\n", min, score, max, score2 );
+	score = abs( score );
+	score2 = abs( score2 );
+	offset = (score < score2) ? min : max;
+
+	/* guard against inaccurate mttb */
+	for( i=0; i<10; i++ ) {
+		start_contest( kSetAndTest, offset, NUM_ITER/10 );
+
+		if( (score2=start_contest(kTest, offset, NUM_ITER)) < 0 )
+			score2 = -score2;
+		if( score2 <= score || score2 < 20 )
+			break;
+	}
+	printk("Final offset: %d (%d/%d)\n", offset, score2, NUM_ITER );
+
+	/* exiting */
+	tbsync->cmd = kExit;
+	wmb();
+	tbsync->handshake = 1;
+	while( tbsync->ack )
+		;
+	tbsync->handshake = 0;
+	kfree( tbsync );
+	tbsync = NULL;
+	running = 0;
+
+	/* all done */
+	smp_tb_synchronized = 1;
+}
diff -Nru a/arch/ppc/kernel/smp.c b/arch/ppc/kernel/smp.c
--- a/arch/ppc/kernel/smp.c	Sat Aug  9 03:10:58 2003
+++ b/arch/ppc/kernel/smp.c	Sat Aug  9 03:10:58 2003
@@ -58,10 +58,6 @@
 /* all cpu mappings are 1-1 -- Cort */
 volatile unsigned long cpu_callin_map[NR_CPUS];

-#define TB_SYNC_PASSES 4
-volatile unsigned long __initdata tb_sync_flag = 0;
-volatile unsigned long __initdata tb_offset = 0;
-
 int start_secondary(void *);
 extern int cpu_idle(void *unused);
 void smp_call_function_interrupt(void);
@@ -288,41 +284,6 @@
 		atomic_inc(&call_data->finished);
 }

-/* FIXME: Do this properly for all archs --RR */
-static spinlock_t timebase_lock = SPIN_LOCK_UNLOCKED;
-static unsigned int timebase_upper = 0, timebase_lower = 0;
-
-void __devinit
-smp_generic_give_timebase(void)
-{
-	spin_lock(&timebase_lock);
-	do {
-		timebase_upper = get_tbu();
-		timebase_lower = get_tbl();
-	} while (timebase_upper != get_tbu());
-	spin_unlock(&timebase_lock);
-
-	while (timebase_upper || timebase_lower)
-		rmb();
-}
-
-void __devinit
-smp_generic_take_timebase(void)
-{
-	int done = 0;
-
-	while (!done) {
-		spin_lock(&timebase_lock);
-		if (timebase_upper || timebase_lower) {
-			set_tb(timebase_upper, timebase_lower);
-			timebase_upper = 0;
-			timebase_lower = 0;
-			done = 1;
-		}
-		spin_unlock(&timebase_lock);
-	}
-}
-
 static void __devinit smp_store_cpu_info(int id)
 {
         struct cpuinfo_PPC *c = &cpu_data[id];
@@ -423,7 +384,7 @@

 	/* wake up cpu */
 	smp_ops->kick_cpu(cpu);
-
+
 	/*
 	 * wait to see if the cpu made a callin (is actually up).
 	 * use this value that I found through experimentation.

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

  reply	other threads:[~2003-08-09  1:13 UTC|newest]

Thread overview: 8+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2003-08-08  0:15 [PATCH] 2.6 timebase synchronization Samuel Rydh
2003-08-09  1:13 ` Samuel Rydh [this message]
2003-08-09  8:58   ` Benjamin Herrenschmidt
2003-08-09 11:48     ` Samuel Rydh
2003-08-10  2:09   ` Paul Mackerras
2003-08-10  9:39     ` Benjamin Herrenschmidt
2003-08-10 16:22       ` Mark Guertin
2003-08-10 21:29       ` Samuel Rydh

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20030809011305.GA12030@ibrium.se \
    --to=samuel@ibrium.se \
    --cc=linuxppc-dev@lists.linuxppc.org \
    --cc=paulus@samba.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.