install/46629: newfs(8): proper block/fragment defaults for modern AFT disks

[Izumi Tsutsui <tsutsui%ceres.dti.ne.jp@localhost>]

>Number:         46629
>Category:       install
>Synopsis:       newfs(8): proper block/fragment defaults for modern AFT disks
>Confidential:   no
>Severity:       non-critical
>Priority:       high
>Responsible:    install-manager
>State:          open
>Class:          change-request
>Submitter-Id:   net
>Arrival-Date:   Mon Jun 25 14:10:00 +0000 2012
>Originator:     Izumi Tsutsui
>Release:        NetBSD 6.0_BETA2
>Organization:
>Environment:
System: NetBSD 6.0_BETA2 i386
Architecture: i386, but affects all ports which can have SATA disks
Machine: i386
>Description:
Modern AFT SATA disks have physical 4KB (and logical 512B) sectors,
but current newfs(8) and sysinst(8) choose 16KB/2KB by default
even for >1TB partitions, so it could cause awful performance.

>How-To-Repeat:
Install NetBSD 6.0_BETA2 into AFT disks and create partitions
using default settings.

There is a report that unmounting targetroot after installation
including pkgsrc tree takes more than 30 minutes.

>Fix:
Currently there is no way to detect AFT disks on userland,
but using 32KB/4KB default values for larger disks
is still much better than current 16KB/2KB default,
and I guess there is few bad side effect on larger fragments
on such large partitions even on traditional 512B sector disks.

The following patch makes newfs(8) and sysinst(8)
to use 32KB/4KB for >= 128 GB partitions.
(128 GB is also used by fdisk(8) to choose 1MB alignment)


Index: sbin/newfs/newfs.c
===================================================================
RCS file: /cvsroot/src/sbin/newfs/newfs.c,v
retrieving revision 1.110
diff -u -p -r1.110 newfs.c
--- sbin/newfs/newfs.c  13 Feb 2012 12:59:56 -0000      1.110
+++ sbin/newfs/newfs.c  24 Jun 2012 19:47:24 -0000
@@ -157,14 +157,17 @@ const char lmsg[] = "%s: can't read disk
  */
 /*
  * For file systems smaller than SMALL_FSSIZE we use the S_DFL_* defaults,
- * otherwise if less than MEDIUM_FSSIZE use M_DFL_*, otherwise use
- * L_DFL_*.
+ * otherwise if less than MEDIUM_FSSIZE use M_DFL_*,
+ * otherwise if less than LARGE_FSSIZE use L_DFL_*,
+ * otherwise use LL_DFL_* especially for modern AFT disks.
  */
 #define        SMALL_FSSIZE    (20*1024*2)
 #define        S_DFL_FRAGSIZE  512
 #define        MEDIUM_FSSIZE   (1000*1024*2)
 #define        M_DFL_FRAGSIZE  1024
+#define        LARGE_FSSIZE    (128*1024*1024*2)
 #define        L_DFL_FRAGSIZE  2048
+#define        LL_DFL_FRAGSIZE 4096
 #define        DFL_FRAG_BLK    8
 
 /* Apple requires the fragment size to be at least APPLEUFS_DIRBLKSIZ
@@ -624,8 +627,10 @@ main(int argc, char *argv[])
                                        fsize = S_DFL_FRAGSIZE;
                                else if (fssize < MEDIUM_FSSIZE)
                                        fsize = M_DFL_FRAGSIZE;
-                               else
+                               else if (fssize < LARGE_FSSIZE)
                                        fsize = L_DFL_FRAGSIZE;
+                               else
+                                       fsize = LL_DFL_FRAGSIZE;
                                if (fsize < sectorsize)
                                        fsize = sectorsize;
                        }
Index: sbin/newfs/newfs.8
===================================================================
RCS file: /cvsroot/src/sbin/newfs/newfs.8,v
retrieving revision 1.82
diff -u -p -r1.82 newfs.8
--- sbin/newfs/newfs.8  14 May 2011 19:46:10 -0000      1.82
+++ sbin/newfs/newfs.8  24 Jun 2012 19:47:24 -0000
@@ -111,10 +111,12 @@ The default size depends upon the size o
 .Ar block-size
 .It \*[Lt] 20 MB
 4 KB
-.It \*[Lt] 1024 MB
+.It \*[Lt] 1000 MB
 8 KB
-.It \*[Gt]= 1024 MB
+.It \*[Lt] 128 GB
 16 KB
+.It \*[Gt]= 128 GB
+32 KB
 .El
 .It Fl d Ar maxbsize
 Set the maximum extent size to
@@ -151,10 +153,12 @@ The default size depends upon the size o
 .Ar frag-size
 .It \*[Lt] 20 MB
 0.5 KB
-.It \*[Lt] 1024 MB
+.It \*[Lt] 1000 MB
 1 KB
-.It \*[Gt]= 1024 MB
+.It \*[Lt] 128 GB
 2 KB
+.It \*[Gt]= 128 GB
+4 KB
 .El
 .It Fl G
 Treat garbage parameters as non-fatal.
@@ -182,10 +186,12 @@ bytes of data space:
 .Ar bytes-per-inode
 .It \*[Lt] 20 MB
 2 KB
-.It \*[Lt] 1024 MB
+.It \*[Lt] 1000 MB
 4 KB
-.It \*[Gt]= 1024 MB
+.It \*[Lt] 128 GB
 8 KB
+.It \*[Gt]= 128 GB
+16 KB
 .El
 .It Fl m Ar free-space
 The percentage of space reserved from normal users; the minimum free
Index: distrib/utils/sysinst/label.c
===================================================================
RCS file: /cvsroot/src/distrib/utils/sysinst/label.c,v
retrieving revision 1.61
diff -u -p -r1.61 label.c
--- distrib/utils/sysinst/label.c       5 Jan 2012 22:18:36 -0000       1.61
+++ distrib/utils/sysinst/label.c       24 Jun 2012 19:47:24 -0000
@@ -213,8 +213,23 @@ set_ptype(partinfo *p, int fstype, int f
        p->pi_fstype = fstype;
        if (fstype == FS_BSDFFS || fstype == FS_BSDLFS) {
                p->pi_frag = 8;
-               /* match newfs defaults for fragments size (2k if >= 1024MB) */
-               p->pi_fsize = p->pi_size > 1024*1024*1024 / 512 ? 2048 : 1024;
+               /*
+                * match newfs defaults for fragments size:
+                * fs size      frag size
+                * < 20 MB      0.5 KB
+                * < 1000 MB    1 KB
+                * < 128 GB     2 KB
+                * >= 128 GB    4 KB
+                */
+               /* note pi_size is uint32_t so we have to avoid overflow */
+               if (p->pi_size < (20 * 1024 * (1024 / 512)))
+                       p->pi_fsize = 512;
+               else if (p->pi_size < (1000 * 1024 * (1024 / 512)))
+                       p->pi_fsize = 1024;
+               else if (p->pi_size < (128 * 1024 * 1024 * (1024 / 512)))
+                       p->pi_fsize = 2048;
+               else
+                       p->pi_fsize = 4096;
        } else {
                /* zero - fields not used */
                p->pi_frag = 0;