root/drivers/block/README.hd

/* */

IDE Performance Enhancements    Version 2.1
============================    ===========

What's new in version 2.1:
        -- Support added for E-IDE BIOSs, for systems with IDE drives
        that "have more than 16 logical heads" (according the BIOS).
        -- the HDIO_SETUNMASKINTR and HDIO_SETMULTCOUNT now permit
        only the superuser to change settings, and no longer permit
        minor device numbers to be used.

This version of hd.c includes support for two optional features:

(1) The disk I/O routines can now run with interrupts unmasked
        most of the time, making them much friendlier to high
        speed serial ports and other system activity.

(2) Support is included for IDE "Multiple Sector Mode", the use
        of which can reduce disk I/O kernel overhead by 10-30%
        on many systems, with a corresponding 10-20% increase in
        data throughput.

By default, both features are DISABLED, for compatibility with
        systems on which they may cause troubles.

The IRQ unmasking has been known to CORRUPT FILESYSTEMS in the
past on systems with strange hard drives.  Backup before trying!

It works on most systems, but use at your own risk!!

Drives which support "Multiple Sector Mode" are identified by the
kernel at boot time, and a message is displayed indicating the
largest possible setting for "MaxMult" (max sector count for
"Multiple Sector Mode").

For more detailed boot-time information about your drive, change
the definition of VERBOSE_DRIVE_INFO from 0 to 1 near the top
of hd.c and rebuild/reinstall the kernel.

Some drives (mostly older CONNER drives) do not implement multiple mode
correctly, and data corruption may occur.. but if you wait long enough
the error recovery logic *should* be able to recover eventually.

I recommend using settings of 8, 16, or 32.  Many drives also support
non-powers of two, but other drives do not -- try strange values at
your own risk!

To try this out more safely, mount the drive's partitions read-only
before using hdparm (see below) for the first time.  If it doesn't
work, email me (mlord@bnr.ca) with the drive name as displayed at
boot time, so I can warn others and possibly add a hook to the code.

To enable the features, a small program is included below:  hdparm.c
This one is *different* from previous versions -- be sure to recompile it!

Compile this using   cc -O -o /usr/bin/hdparm hdparm.c
and then use it to enable/disable the new features, as follows:

To turn on 16-sector multiple mode, with interrupt unmasking:

        hdparm /dev/hda 16 1

To view the current settings:

        hdparm /dev/hda

If you have more than one drive, a separate command would need to               be issued for the second drive as well, using the same or different
settings as desired:

        hdparm /dev/hdb 16 1

To turn off both features on the first drive, use:

        hdparm /dev/hda 0 0

To benchmark the performance difference, try:

        hdparm /dev/hda 0 0
        sync
        time dd if=/dev/hda of=/dev/null bs=1024k count=30
        time dd if=/dev/hda of=/dev/null bs=1024k count=30
        time dd if=/dev/hda of=/dev/null bs=1024k count=30
        hdparm /dev/hda 16 1
        sync
        time dd if=/dev/hda of=/dev/null bs=1024k count=30
        time dd if=/dev/hda of=/dev/null bs=1024k count=30
        time dd if=/dev/hda of=/dev/null bs=1024k count=30

This gives before and after views.  Compare the total elapsed times,
as well as the percent of CPU used in each case.  Run several trials
to ensure/verify consistent results.  Some drives are actually *slower*
with multiple mode enabled, but those are very rare indeed.  Most systems
experience a 10-30% increase in throughput, with a corresponding 5-50%
decrease in kernel/system CPU usage.

If you are using linux kernel 1.1.4 or higher (with the "cluster" code),
then you may not notice a big difference with the above tests.  However,
I have noticed that the iozone benchmark program does seem to work fairly
reliably under kernels with the "cluster" code, so you could try that
instead (under older kernels, iozone seems to give wildly varying results
from trial to trial, at least on my system).

To have your favorite settings installed automatically at boot time,
place the hdparm command(s) into the /etc/rc.d/rc.local file.
Alternatively, one could modify the DEFAULTs near the top of hd.c
and rebuild and install the new kernel.

Enjoy,
mlord@bnr.ca

**** CUT HERE for hdparm.c ****

/* make using:    cc -O -o /usr/bin/hdparm hdparm.c */

#include <linux/hdreg.h>
#include <sys/ioctl.h>
#include <stdio.h>
#include <sys/types.h>
#include <unistd.h>
#include <fcntl.h>

/* extern int hd_ioctl(struct inode * inode, struct file * file,
 *                     unsigned int cmd, unsigned long arg);
 */


void main (int argc, char *argv[])
{
        int fd, mrc, irc;
        static long mcount, irqmode;

        if (argc != 4 && argc != 2) {
                fprintf(stderr,"Usage:  %s  <device>"
                               " [<MultModeCount:0-64> <unmask:0/1>]\n", *argv);
                exit (1);
        }

        fd = open (*++argv, O_RDONLY);
        if (fd < 0) {
                printf ("open failed - '%s' - %d\n", *argv, fd);
                exit(1);
        }
        if (argc == 4) {
                mcount = atoi(*++argv);
                mrc = ioctl (fd, HDIO_SETMULTCOUNT, &mcount);
                /* note that the new mcount does not take effect
                   until the next disk I/O operation, so if we were
                   to query it before then, the old value will show.
                   Also, the drive may reject the new value, which will
                   result in multiple mode being turned OFF completely! */
                irqmode = atoi(*++argv);
                irc = ioctl (fd, HDIO_SETUNMASKINTR, &irqmode);
        }
        else {
                mrc = ioctl (fd, HDIO_GETMULTCOUNT, &mcount);
                irc = ioctl (fd, HDIO_GETUNMASKINTR, &irqmode);
        }
        printf("MultModeCount=%d, rc=%d\n", mcount, mrc);
        printf("unmask=%d, rc=%d\n", irqmode, irc);
}

/* */