dt-usersguide.t

Linux磁盘测试的源代码,测试磁盘的读写性能

Special Notes:
.bu
If you don't plan to ever read the data being written, perhaps for
performance reasons, specifying "\fIdisable=compare\fP" prevents
the data buffer from being initialized with a data pattern.
.bu
This verify option has no affect when reading a device.  You must
disable data comparsions via "\fIdisable=compare\fP".
.LP
.nh 2 Program Delays.
\fIdt\fP allows you to specify various delays to use at certain points
of the test.  These delays are useful to slow down I/O requests or to
prevent race conditions when testing terminals devices with multiple
processes, or are useful for low level driver debugging.  All delay
values are in seconds, unless you specify "\fIenable=microdelay\fP",
to enable micro-second delays.
.nh 3 Close File \&"\fIcdelay=\fP\&" Delay.
This delay, when enabled, is performed prior to closing a file
descriptor.
.EX
    Delay
        cdelay=value     Delay before closing the file.    (Def: 0)
.EE
.nh 3 End of Test \&"\fIedelay=\fP\&" Delay.
This delay, when enabled, is used to delay after closing a device,
but prior to re-opening the device between multiple passes.
.EX
    Delay
        edelay=value     Delay between multiple passes.    (Def: 0)
.EE
.nh 3 Read Record \&"\fIrdelay=\fP\&" Delay.
This delay, when enabled, is used prior to issuing each read request
(both synchronous read()'s and asynchronous aio_read()'s).
.EX
    Delay
        rdelay=value     Delay before reading each record. (Def: 0)
.EE
.nh 3 Start Test \&"\fIsdelay=\fP\&" Delay.
This delay, when enabled, is used prior to starting the test.  When
testing terminal devices, when not in self loopback mode, the writing
process (the parent) automatically delays 1 second, to allow the reading
process (the child) to startup and setup its' terminal characteristics.
If this delay did not occur prior to the first write, the reader may not
have its' terminal characteristics (flow, parity, & speed) setup yet,
and may inadvertantly flush the writers data or receive garbled data.
.EX
    Delay
        sdelay=value     Delay before starting the test.   (Def: 0)
.EE
.nh 3 Child Terminate \&"\fItdelay=\fP\&" Delay.
This delay is used by child processes before exiting, to give the parent
process sufficient time to cleanup and wait for the child.  This is
necessary since if the child exits first, a SIGCHLD signal may force the
parent to it's termination signal handler before it's ready to.  This is
a very simplistic approach to prevent this parent/child race condition
and is only currently used by the child for terminal loopback testing.
.EX
    Delay
        tdelay=value     Delay before child terminates.    (Def: 1)
.EE
.nh 3 Write Record \&"\fIwdelay=\fP\&" Delay.
This delay, when enabled, is used prior to issuing each write request
(both synchronous write()'s and asynchronous aio_write()'s).
.EX
    Delay
        wdelay=value     Delay before writing each record. (Def: 0)
.EE
.nh 2 Numeric Input Parameters
For any options accepting numeric input, the string entered may contain any
combination of the following characters:
.EX 4
Special Characters:
    w = words (4 bytes)            q = quadwords (8 bytes)
    b = blocks (512 bytes)         k = kilobytes (1024 bytes)
    m = megabytes (1048576 bytes)  p = page size (8192 bytes)
    g = gigabytes (1073741824 bytes)
    t = terabytes (1099511627776 bytes)
    inf or INF = infinity (18446744073709551615 bytes)
.HS
Arithmetic Characters:
    + = addition                   - = subtraction
    * or x = multiplcation         / = division
    % = remainder
.HS
Bitwise Characters:
    ~ = complement of value       >> = shift bits right
   << = shift bits left            & = bitwise 'and' operation
    | = bitwise 'or' operation     ^ = bitwise exclusive 'or'
.EE
The default base for numeric input is decimal, but you can override
this default by specifying 0x or 0X for hexadecimal coversions, or
a leading zero '0' for octal conversions.
.B1
NOTE:  Certain values will vary depending on the operating system and/or
machine you are running on.  For example, the page size is system dependent,
and the value for Infinity is the largest value that will fit into an
unsigned long (value shown above is for 64-bit systems), or double for
systems which don't support "\fIlong long\fP".)
.B2
.nh 2 Time Input Parameters.
When specifying the run time "\fIruntime=\fP" option, the time string
entered may contain any combination of the following characters:
.EX 4
Time Input:
    d = days (86400 seconds),      h = hours (3600 seconds)
    m = minutes (60 seconds),      s = seconds (the default)
.EE
Arithmetic characters are permitted, and implicit addition is performed
on strings of the form '1d5h10m30s'.
.HS
.nh 1 Future Enhancements.
Although many system dependent tests could be added to \fIdt\fP, my preference
is to add tests for features standard on most systems.  I have to admit though,
recently I've added many Tru64 UNIX features, required for advanced testing.
.PP
There's alot of \fIdt\fP features and flags nowadays, so maybe I'll add an
initialization file (\fI.dtrc\fP) to control certain defaults.  We'll see.
.PP
A future release may permit \fIdt\fP to support multiple devices simlutaneously,
and permit dynamic modification of test parameters and obtaining snap-shot
statistics.  This will require an extensive re-write though, so this isn't
likely to happen too soon.  I'd like \fIdt\fP to be like the RSX-11M \fIiox\fP
program, which was really nifty.  But of course, an operating system which
supports asynchronous I/O and event notification, makes this much easier.
.PP
Lots of people prefer window based applications nowadays.  If I get around
to this, it'll probably be a Tcl/Tk/Wish wrapper, or something that will
run with your favorite browser.  I like this latter idea for implementing
remote testing.  A native Windows/NT version will most likely come about
one day, unless Cygnus Solutions implements POSIX AIO and Alpha hardware
in a future release, or Microsoft releases better POSIX API support.
.PP
There's a fair amount of work necessary to do extensive tape testing, due
to the number of special tape commands (both standard and optional), and
differences between tape vendors and operating system API's.
.PP
I'm also considering writing a script processor, kinda like a maintenance
program generator, to help with this effort so customized tests can be
written, but perhaps this is best done with existing tools available
(I'm not into reinventing the wheel).
.bp
.nh 1 Final Comments.
I'm happy to report that \fIdt\fP is getting wide spread use here at Compaq.
The storage groups, terminal/lat groups, Q/A, developers, and other peripheral
qualification groups are all using \fIdt\fP as part of their testing (just
when I was about to give up on this place... maybe there's hope yet :-).
.PP
Anyways, I hope you find \fIdt\fP as useful as I have.  This is usually one
of the first tools I port to a new operating system, since it's an excellent
diagnostic and performance tool (it gives me a warm and fuzzy feeling).
.PP
Please send me mail on any problems or suggestions you may have, and I'll
try to help you out.  The future development of \fIdt\fP depends alot on user
interest.  Many of \fIdt\fP's features have come about from user requests.
.sp 3
.TS
center, doublebox;
c.
If You Like My Work,
You Can Do
One Of Two Things:
.sp 0.3
\s+2\fBTHROW MONEY\fP or \fBAPPLAUD*\fP\s-2
.sp 0.3
\fB*\fPI've Heard Enough Applause!
.sp 0.2
.TE
.ah A \fIdt\fP Help Text
.PP
The following help text is contained within the \fIdt\fP program.  The newest
features are highlighted below, as well as previous features for reference by
former/existing \fIdt\fP users.
.HS
The following changes have been made to 'dt' Version 12.0:
.bu
Support for tape Extended Error Information (EEI) and SCSI bus
or device reset recovery has been added (to reposition the tape).
The EEI support is enabled by default, but the reset support is
disabled by default (to avoid breaking existing test scripts).
The reset option can also be used w/POSIX Asynchronous I/O (AIO).
NOTE: EEI support is Digital UNIX specific (MTIOCGET extension).
[ Use the "enable/disable={eei,resets}" option to control these. ]
.bu
The logical block data feature, "enable=lbdata" option, has been
extended so it can be used with random I/O, "iotype=random" option.
.bu
Two new options were added for better range control of random I/O.
.HS
The "ralign=value" option forces block alignment to a byte offset.
For example, "ralign=32b" aligns each random request to 32 blocks.
.HS
The "rlimit=value" option contols the upper random I/O data limit.
If rlimit isn't specified, it defaults to the data limit, or to the
entire disk if no limits are specified.
.bu
Changes to cluster DLM code were merged from work by George Bittner.
George also enhanced the random I/O code so the random number is
taken as a block, rather than a byte offset, which provides better
randomness for large disks or files. (Thanks George!)
.bu
Other changes and/or bug fixes:
.EX 0
- When specifying a runtime, ensure we break out of the passes
  loop if we exceed the error limit.  Previously, we'd loop
  (possibly with  error) for the duration of the runtime.
- Fix incorrect record number displayed when Debug is enabled.
- Don't exit read/write loops when processing partial records.
- Fix problem in write function, where short write processing,
  caused us not to write sufficent data bytes (actually, the
  file loop in write_file() caused dtaio_write_data() to be
  called again, and we'd actually end up writing too much!
- When random I/O and lbdata options are both enabled, use the
  file offset seeked to as the starting lbdata address.
- Fix problem reporting total files processed count when we have
  not completed a pass (exiting prematurely due to error or signal).
.EE
.bu
\fIdt\fP compiles and runs on MS Windows 95/98/NT using Cygnus Solution
GNU tools.  Use Makefile.win32 for compiling on Windows.
[ a.k.a. GNU-Win32, URL: http://sourceware.cygnus.com/cygwin/ ]
.bu
When directing output to a log file, "log=LogFile" option, the
command line and version string are also emitted.
.bu
The number of I/O per second are now reported in statistics.
.bu
For Linux and Windows/NT, larger data/record limits and statistics
are now possible using either "long long" or "double" storage.
.bu
A new data pattern option was added "\fIpattern=iot\fP", which
encoded the logical block address throughout each word in the
data block.  Thanks to DJ Brown for this logic from IOT.
.LP
.HS
The following changes have been made to \fIdt\fP Version 9.0:
.HS
.bu
Fixed a problem checking pad bytes when doing long reads of
short records (this was discovered by someone testing tapes).
.bu
For systems which support higher serial line speeds, you can now
select baud rates of 57600, 76800, and 115200.  Platinum (ptos)
now supports 57600 and 115200 (thank-you Dennis Paradis).
.bu
Added the ability to read/write blocks using logical block data
(lbdata) option.  While this new feature was added specifically
as a counterpart to the SCSI Write Same (disks) lbdata option,
this option can be used with any device and/or data stream.
.HS
What is Logical Block Data (lbdata) Format?
.HS
This feature allows reading/writing of a 4-byte (32-bit) logical
block address at the beginning of each data block tested.  At the
present time, the block number is stored using SCSI byte ordering
(big-endian), which matchs what the SCSI Write Same command does.
.HS
Several new 'dt' options have been added for this feature:
.EX 0
    lba=value       Set starting block used w/lbdata option.
    lbs=value       Set logical block size for lbdata option.
    [ ...as well as... ]
    enable=lbdata   Logical block data.    (Default: disabled)
.EE
If this feature is enabled without specifying other options, the
block size (lbs) defaults to 512, and the starting block address
(lba) defaults to 0.  Specifying "lba=" or "lbs=" options implicitly
enables the "lbdata" feature.
.HS
NOTES:  The logical block number is *always* inserted starting at the
beginning of each data block, not every "lbs" bytes (WRT to variable
length opts).  Also, the block number overwrites what the previous
pattern bytes would have been (again matching the SCSI Write Same
method).  Also, enabling this feature will cause a degradation in
performance statistics.
.HS
If people would like an option to have the logical block stored in
little-endian format, I'll consider adding an option *if* requested.
.HS
Of course, the data verification routines have been modified to
verify the logical block number matches the expected, and reports
any mis-matches.
.bu
For disk devices, the "min=" and "incr=" values now default to 512,
when variable length transfers are requested (previously defaulted
to 1 which was great for tapes, but not so good for block devices).
[ A future version of 'dt' will obtain the device block size via
  the new DEVGETINFO ioctl() now available in Platinum (ptos). ]
.bu
Removed "enable=ade" (Address Data Exception) test option.
This was used to test kernel unaligned data exception fixups,
but it's seldom (if ever) used, and it's time to cleanup code.
.bu
Added "ttymin=" option for setting the VMIN value for serial line
testing.  Also fixed problem where VMIN got set incorrectly if the
block size was greater than 255 (VMIN is a u_char on Tru64 Unix).
.LP
.HS
New Features in \fIdt\fP Version 8.0:
.HS
.bu
On compare errors, data dumping now occurs by default.  The
"dlimit=value" option can be used to override the default dump
limit of 64 bytes (use "disable=dump"