svn commit: r293695 - stable/10/sbin/devd

[Alan Somers]

Author: asomers
Date: Mon Jan 11 20:24:56 2016
New Revision: 293695
URL: https://svnweb.freebsd.org/changeset/base/293695

Log:
  MFC r292020
  
  Increase devd's client socket buffer size to 256KB. This is not as large as
  it looks, because we'll hit the sockbuf's mbuf limit long before hitting its
  data limit. A 256KB data limit allows creating a ZFS pool on about 450
  drives without overflowing the client socket buffers.

Modified:
  stable/10/sbin/devd/devd.cc

Modified: stable/10/sbin/devd/devd.cc
==============================================================================
--- stable/10/sbin/devd/devd.cc	Mon Jan 11 20:13:45 2016	(r293694)
+++ stable/10/sbin/devd/devd.cc	Mon Jan 11 20:24:56 2016	(r293695)
@@ -108,15 +108,26 @@ __FBSDID("$FreeBSD$");
 /*
  * Since the client socket is nonblocking, we must increase its send buffer to
  * handle brief event storms.  On FreeBSD, AF_UNIX sockets don't have a receive
- * buffer, so the client can't increate the buffersize by itself.
+ * buffer, so the client can't increase the buffersize by itself.
  *
  * For example, when creating a ZFS pool, devd emits one 165 character
- * resource.fs.zfs.statechange message for each vdev in the pool.  A 64k
- * buffer has enough space for almost 400 drives, which would be very large but
- * not impossibly large pool.  A 128k buffer has enough space for 794 drives,
- * which is more than can fit in a rack with modern technology.
+ * resource.fs.zfs.statechange message for each vdev in the pool.  The kernel
+ * allocates a 4608B mbuf for each message.  Modern technology places a limit of
+ * roughly 450 drives/rack, and it's unlikely that a zpool will ever be larger
+ * than that.
+ *
+ * 450 drives * 165 bytes / drive = 74250B of data in the sockbuf
+ * 450 drives * 4608B / drive = 2073600B of mbufs in the sockbuf
+ *
+ * We can't directly set the sockbuf's mbuf limit, but we can do it indirectly.
+ * The kernel sets it to the minimum of a hard-coded maximum value and sbcc *
+ * kern.ipc.sockbuf_waste_factor, where sbcc is the socket buffer size set by
+ * the user.  The default value of kern.ipc.sockbuf_waste_factor is 8.  If we
+ * set the bufsize to 256k and use the kern.ipc.sockbuf_waste_factor, then the
+ * kernel will set the mbuf limit to 2MB, which is just large enough for 450
+ * drives.  It also happens to be the same as the hardcoded maximum value.
  */
-#define CLIENT_BUFSIZE 131072
+#define CLIENT_BUFSIZE 262144
 
 using namespace std;