[linux-lvm] [PATCH] calc_max_buckets bug

[linux-lvm] [PATCH] calc_max_buckets bug

[Chris Mason]

Hello everyone,

This is in both 2.4.x lvm and 2.6.x dm code, calc_max_buckets is broken
for 32bit machines with > 4GB of ram.  This leads to -ENOMEM errors
during snapshot creation.

The patch below just changes things to allocate 2MB worth of hash
buckets all the time.  Has anyone done benchmarking into how much cpu
time snapshots spend on hash table lookups?

-chris

Index: linux.24/drivers/md/lvm-snap.c
===================================================================
--- linux.24.orig/drivers/md/lvm-snap.c	2004-03-09 18:07:59.000000000 -0500
+++ linux.24/drivers/md/lvm-snap.c	2004-03-09 20:20:34.000000000 -0500
@@ -504,13 +504,9 @@ out:
 
 static int calc_max_buckets(void)
 {
-	unsigned long mem;
-
-	mem = num_physpages << PAGE_SHIFT;
-	mem /= 100;
-	mem *= 2;
+	/* use a fixed size of 2MB */
+	unsigned long mem = 2 * 1024 * 1024;
 	mem /= sizeof(struct list_head);
-
 	return mem;
 }
 

Re: [linux-lvm] [PATCH] calc_max_buckets bug

[Joe Thornber]

On Wed, Mar 10, 2004 at 09:10:59AM -0500, Chris Mason wrote:
> Hello everyone,
> 
> This is in both 2.4.x lvm and 2.6.x dm code, calc_max_buckets is broken
> for 32bit machines with > 4GB of ram.  This leads to -ENOMEM errors
> during snapshot creation.

heh, this must be the only bit of code I copied from LVM1 :) Serves me
right for being lazy.

Thanks,

- Joe

Re: [linux-lvm] [PATCH] calc_max_buckets bug

[Chris Mason]

On Wed, 2004-03-10 at 10:44, Joe Thornber wrote:
> On Wed, Mar 10, 2004 at 09:10:59AM -0500, Chris Mason wrote:
> > Hello everyone,
> > 
> > This is in both 2.4.x lvm and 2.6.x dm code, calc_max_buckets is broken
> > for 32bit machines with > 4GB of ram.  This leads to -ENOMEM errors
> > during snapshot creation.
> 
> heh, this must be the only bit of code I copied from LVM1 :) Serves me
> right for being lazy.

Do you see any problems with using a smaller fixed size on the hash
table.  It seems like 2MB should be enough to me.

-chris

Re: [linux-lvm] [PATCH] calc_max_buckets bug

[Joe Thornber]

On Wed, Mar 10, 2004 at 01:17:27PM -0500, Chris Mason wrote:
> Do you see any problems with using a smaller fixed size on the hash
> table.  It seems like 2MB should be enough to me.

I've done no experiments in this area.

- Joe