7 years agoengines/rbd: potential re-use of active fio_rbd_iou
Jason Dillaman [Tue, 13 Oct 2015 19:40:47 +0000 (15:40 -0400)]
engines/rbd: potential re-use of active fio_rbd_iou

Running multiple concurrent RBD jobs can result in a segfault
attempting to retrieve the RBD AIO completion status from a
prematurely reused fio_rbd_iou object.

Signed-off-by: Jason Dillaman <>
7 years agoFix Windows build by using lib/strcasestr.{c,h}
Bruce Cran [Tue, 6 Oct 2015 00:18:50 +0000 (18:18 -0600)]
Fix Windows build by using lib/strcasestr.{c,h}

Signed-off-by: Jens Axboe <>
7 years agoserver: bump protocol version
Jens Axboe [Thu, 1 Oct 2015 06:55:18 +0000 (08:55 +0200)]
server: bump protocol version

Commit 82407585a3b3 added options to the thread_options structures,
so we need to bump the client/server protocol version to signal
we are incompatible with previous versions.

Signed-off-by: Jens Axboe <>
7 years agoIntroduce new option: iodepth_batch_complete_max
Roman Pen [Sun, 27 Sep 2015 19:24:55 +0000 (21:24 +0200)]
Introduce new option: iodepth_batch_complete_max

By now there is an option which is called 'iodepth_batch_complete'.
This option provides a possibility to read exact amount of events
of completed IOs, e.g. by default it is set to 1, so only 1 event
will be read on each io_getevents system call:

 7945  io_getevents(139876985688064, 1, 1, {...}NULL) = 1

This option can be set to any desired amount, say the whole iodepth
size can be used, e.g. 128:

 7952  io_getevents(140303500259328, 128, 128, {...}NULL) = 128

We will not exit the system call till the whole queue will be completed.
Sounds not so efficient.

In this patch I add the 'iodepth_batch_complete_max' option and rename
'iodepth_batch_complete' to 'iodepth_batch_complete_min' to have a
possibility to specify a range of events which I want to read.
(of course because of compatibility's sake 'iodepth_batch_complete_min'
 is an alias for 'iodepth_batch_complete').

So finally if options are set to:

 # or iodepth_batch_complete_min=1
 # take the iodepth value

The io_getevents call will look like:

 7961  io_getevents(140173245206528, 1, 128, {...}NULL) = 73

and we will exit the system call with any amount of completed
IOs >= 1.

What are the benefits? Fetching events from the queue can be
done more efficiently on some testing on specific configurations,
e.g. stacked block devices based on null_blk devices, where
completion happens immediately and events should be retried from
the queue ASAP.

Let's take a simple example.








Signed-off-by: Jens Axboe <>
7 years agofio.1,HOWTO: keep 'iodepth_batch' option in sync
Roman Pen [Sun, 27 Sep 2015 19:24:54 +0000 (21:24 +0200)]
fio.1,HOWTO: keep 'iodepth_batch' option in sync

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
7 years agosg: fix short reads
Jens Axboe [Wed, 30 Sep 2015 16:31:12 +0000 (18:31 +0200)]
sg: fix short reads

We can't assume that if we read some data, we read all of it.
Check for this condition.

Don't error out if a read failed, but we already got valid

Signed-off-by: Jens Axboe <>
7 years agoUse line buffering on stdout
Kris Davis [Wed, 30 Sep 2015 16:00:00 +0000 (18:00 +0200)]
Use line buffering on stdout

Line Buffer stdio to prevent multi-thread output from
mixing within a line

Signed-off-by: Jens Axboe <>
7 years agosg: 16-byte cdb support and lots of fixes
Kris Davis [Wed, 30 Sep 2015 15:57:16 +0000 (17:57 +0200)]
sg: 16-byte cdb support and lots of fixes

-- Fixed prob with queue depths > 1 (usually would work at qd = 2,
   but had fundimental prob with retrieving multiple events)
   Only works up to QD 16 due to kernel limitation.

-- Fix problem with check and report scsi operation errors.
   There was no check, so op errors were being ignored.
   Added support to retrieve and print sense data to output.

-- Fix problem with op timeouts. No timeouts op time out was being set.
   Hardcoded to 30 seconds.

-- Support switch to 16 Byte operations when LBA range needed.

-- Set default size based on maxlba of device retrieved
   from ReadCapacity command.

Updated by Jens to:

- ->errdetails() must be reentrant. Return an allocated buffer, have
  the caller free it when done.

- Bump io engine version number, the ops have changed.

- Add strlcat() as a real helper, don't stuff it in sg.c

- Lots of style cleanups

Signed-off-by: Jens Axboe <>
7 years agoFix integer overflow in rate_iops
Jens Axboe [Fri, 25 Sep 2015 02:35:44 +0000 (20:35 -0600)]
Fix integer overflow in rate_iops

For the following job file:


We end up overflowing in multiplication, making the bps field 0.
Fio exits by reporting:

rate lower than supported

Fix this by casting to uint64_t, the type of the output.

Signed-off-by: Jens Axboe <>
8 years agoUpdate fio2gnuplot
Thomas Findelkind [Thu, 17 Sep 2015 17:42:34 +0000 (19:42 +0200)]
Update fio2gnuplot

Should fix the "No log file..." with default pattern -b and -i and
 user generated a log file per thread/numjob with per_numjobs_log=1

8 years agoMerge branch 'patch-1' of git://
Jens Axboe [Wed, 16 Sep 2015 21:43:28 +0000 (15:43 -0600)]
Merge branch 'patch-1' of git://

8 years agoUpdate fio2gnuplot
Tfindelkind [Wed, 16 Sep 2015 21:38:12 +0000 (23:38 +0200)]
Update fio2gnuplot

8 years agoReduced minimum for option log_gz
Stephen Bates [Tue, 15 Sep 2015 21:37:53 +0000 (15:37 -0600)]
Reduced minimum for option log_gz

Reduced the permissable minimium for log_sz from 32MB to 1KB to avoid
a latency log issue associated with the reallocation of the buffer logs.

8 years agoFio 2.2.10 fio-2.2.10
Jens Axboe [Fri, 11 Sep 2015 21:13:37 +0000 (15:13 -0600)]
Fio 2.2.10

Signed-off-by: Jens Axboe <>
8 years agoworkqueue: Ensure submission offload worker inherits parent's options
Andrey Kuzmin [Tue, 8 Sep 2015 15:10:55 +0000 (09:10 -0600)]
workqueue: Ensure submission offload worker inherits parent's options

Properly initialize sibling's eo, as otherwise it will miss parent's

Signed-off-by: Jens Axboe <>
8 years agoverify: move meta header to generic verify_header
Roman Pen [Sat, 5 Sep 2015 17:01:02 +0000 (19:01 +0200)]
verify: move meta header to generic verify_header

fio does not provide any possibility to verify checksum of a block with meta
information inside. You can create configuration for verifincation checksum
of random data either you can verify meta information with some pattern or
random data, but not both.

Why checksumming and meta together can be useful? Meta helps to figure out internally
on filesystem or storage what block was written in case of corruption, i.e. offset
of the block and block number explicitly tell us the virtual address of the block.
On the other hand checksum of random data helps to detect corruption. Using meta
and pattern together do not help a lot, since 'verify_interval' can be big enough
and same sequence of pattern bytes will be undistinguishable internally on filesystem
or storage.

Also, it seems to me that keeping meta header separately from generic verify header
does not make a lot of sense, since generic verify header can include all members
of meta header without any performance or other impact.

In this patch I move all members from vhdr_meta structure to generic verify_header,
always verifying meta with the possiblity to checksum the following data: random
or pattern.

You are allowed to specify verify_pattern=str with any of the possible verification
methods and have also meta verification, i.e.





To keep everything compatible with old configurations it is still possible to specify


but this option marked and depricated and kept only for compatibility reasons.

Before that patch the verification layout according to the specified options looks
as the following, e.g.:


    result layout of each block: [hdr|meta|pattern]


    result layout of each block: [hdr|pattern]


    result layout of each block: [pattern]

After applying of the patch 'vhdr_meta' is always embedded into 'verify_header' and layout
looks as the following, e.g.:


    result layout of each block: [hdr+meta|pattern]

    result layout of each block: [hdr+meta|cksum|pattern]


    result layout of each block: [hdr+meta|pattern]


    result layout of each block: [pattern]

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agofio.1: update description for 'buffer_pattern' and 'verify_pattern'
Roman Pen [Sat, 5 Sep 2015 17:01:01 +0000 (19:01 +0200)]
fio.1: update description for 'buffer_pattern' and 'verify_pattern'

Now those two support combined input and 'verify_pattern' supports '%o' format.
Update corresponding doc.

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agouse 'lib/pattern' to parse patterns and paste formats into buffers
Roman Pen [Wed, 19 Aug 2015 10:33:11 +0000 (12:33 +0200)]
use 'lib/pattern' to parse patterns and paste formats into buffers

Switch to new 'lib/pattern' helpers.

Now 'buffer_pattern' and 'verify_pattern' options support combined input
and strings like:


can be specified.

Also, 'verify_pattern' supports '%o' format, which means that buffer will
be patched on each iteration with real 'block offset' number.

So, f.e. 'verify_pattern' such combined input is valid:




Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agoadd FIELD_SIZE macro to calculate the size of the specified field
Roman Pen [Wed, 19 Aug 2015 10:33:10 +0000 (12:33 +0200)]
add FIELD_SIZE macro to calculate the size of the specified field

Will use this stuff later.

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agoverify: use 'cmp_pattern' from 'lib/pattern.c' to compare pattern and buffer
Roman Pen [Wed, 19 Aug 2015 10:33:09 +0000 (12:33 +0200)]
verify: use 'cmp_pattern' from 'lib/pattern.c' to compare pattern and buffer

'cmp_pattern' function should be faster, since it does not use loops
and while doing 'memcmp' the same CPU cache line is used.

Keep in mind, that now single-byte pattern becomes the string of 512 bytes.
In that case this optimization will not work as expected, since 'memcmp'
will be called twice for the whole pattern, but in next patches this
behaviour will be avoided and single-byte pattern will stay single-byte
pattern without any attempt to duplicate it.

Also print buffer and pattern bytes as unsigned chars.

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agoreplace 'fill_pattern' with 'cpy_pattern' from 'lib/pattern.c'
Roman Pen [Wed, 19 Aug 2015 10:33:08 +0000 (12:33 +0200)]
replace 'fill_pattern' with 'cpy_pattern' from 'lib/pattern.c'

All pattern helpers are now located inside 'lib/pattern.c',
and 'cpy_pattern' should be sligtly faster, since it does
copy doubling the size on every iteration.

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agolib/pattern: add set of functions to parse combined pattern input
Roman Pen [Wed, 19 Aug 2015 10:33:07 +0000 (12:33 +0200)]
lib/pattern: add set of functions to parse combined pattern input

The idea of this patch is to have several helpers to parse combined
pattern input, which can consists of strings, numbers and formats.

For example now pattern can be combined and look like this string:

 input : 0xdeadbeef"123"0xdeadface
 output: de ad be ef 31 32 33 de ad fa ce


 input : -99"some string"0x66
 output: 9d ff ff ff 73 6f 6d 65 20 73 74 72 69 6e 67 66

  or with formats

 input : 0xdeadface0xffff%o
 output: de ad fa ce ff ff 00 00 00 00 00 00 00 00

 where %o - offset of the block, reserved 8 bytes

 Space for formats is reserved in output buffer.
 When buffer will be ready to be written to disk - 'paste' callback
 should be called for each pattern format. Each callback is responsible
 for writing data inside the reserved space in the output buffer.

Format array can be extended at any time, all you need is to provide
the name of the format and correct 'paste' callback.

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agolib/strntol: add 'strntol' function
Roman Pen [Wed, 19 Aug 2015 10:33:06 +0000 (12:33 +0200)]
lib/strntol: add 'strntol' function

In future patches I will need it.

Signed-off-by: Roman Pen <>
Signed-off-by: Jens Axboe <>
8 years agoRetry bdev cache invalidation for EAGAIN errors
Ken Raeburn [Fri, 4 Sep 2015 19:30:58 +0000 (13:30 -0600)]
Retry bdev cache invalidation for EAGAIN errors

I've hit a couple test failures where fio quickly died thusly:

    fio 2.0.7
    Starting 1 thread
    fio: pid=6196, err=11/file:filesetup.c:404, func=invalidate_cache, error=Resource temporarily unavailable

(Yeah, we're a little behind. But I think the relevant code is similar.)

We're testing against Linux (RHEL 6.6) multipath block devices, and it
appears that at the time the fio test starts, the multipath maps for
those devices may still be in flux. I poked around a bit with systemtap,
and it appears that while updating the maps, multipathd "suspends" its
multipath device for a few tens of milliseconds, and the kernel
multipath code rejects ioctl calls with EAGAIN if the device is
suspended. It's a small window, but we've managed to hit it multiple
times, though it's not reliably reproducible.

I know the current sources treat failure here as non-fatal, but if we're
using fio for performance tests, trying a little harder to do the
invalidation seems like a good idea. My approach is to add a retry loop
in __file_invalidate_cache; a patch is attached.

It could also be pushed down into blockdev_invalidate_cache, where it
could be local to the Linux (and Android?) implementation, since none of
the others actually do anything. (They return either EAGAIN or zero,
both of which are taken as success indicators.)

Signed-off-by: Jens Axboe <>
8 years agobackend: fix comparison of 'ret' pointer
Andrey Kuzmin [Fri, 4 Sep 2015 19:23:49 +0000 (13:23 -0600)]
backend: fix comparison of 'ret' pointer

Commit e9d512d8e7d2 abstracted out this code, but the assert()
was never changed.

Signed-off-by: Jens Axboe <>
8 years agoFix aux_path for verify state saving
Jens Axboe [Fri, 21 Aug 2015 18:09:03 +0000 (11:09 -0700)]
Fix aux_path for verify state saving

Botched the string copies.

Fixes: d264264a08 ("Add --aux-path support"
Signed-off-by: Jens Axboe <>
8 years agoparse: only print option mismatch help, if the distance is close enough
Jens Axboe [Fri, 21 Aug 2015 18:01:29 +0000 (11:01 -0700)]
parse: only print option mismatch help, if the distance is close enough

Signed-off-by: Jens Axboe <>
8 years agoAdd --aux-path support
Jens Axboe [Thu, 20 Aug 2015 19:02:05 +0000 (12:02 -0700)]
Add --aux-path support

For certain files, fio just stores them in the current working
directory. This can create a bit of a mess. These files might
be verification dump files, or verification state saves.

Add --aux-path to enable storing these files in an arbitrary
directory instead.

Signed-off-by: Jens Axboe <>
8 years agoREADME: update for trigger options
Jens Axboe [Thu, 20 Aug 2015 18:59:08 +0000 (11:59 -0700)]
README: update for trigger options

Signed-off-by: Jens Axboe <>
8 years agoconfigure: cleanup strsep() test case
Jens Axboe [Tue, 18 Aug 2015 18:26:00 +0000 (11:26 -0700)]
configure: cleanup strsep() test case

Signed-off-by: Jens Axboe <>
8 years agoconfigure: fixup using wrong var for endmntent()
Jens Axboe [Tue, 18 Aug 2015 18:23:00 +0000 (11:23 -0700)]
configure: fixup using wrong var for endmntent()

Signed-off-by: Jens Axboe <>
8 years agoMerge branch 'master' of git://
Jens Axboe [Tue, 18 Aug 2015 17:19:50 +0000 (11:19 -0600)]
Merge branch 'master' of git://

8 years agoCorrect handleing of rate_iops_min and ratemin
Dave [Tue, 18 Aug 2015 16:39:11 +0000 (10:39 -0600)]
Correct handleing of rate_iops_min and ratemin

change to allow ratemin to be set without rate disabled ratemin checking in __check_min_rate.

Separate from the change to ratemin parsing, rate_iops_min check did not result in "return 1;" so jobs would not terminate on a failure to maintain rate_iops_min.

8 years agomtd: expand configure check
Jens Axboe [Mon, 17 Aug 2015 21:47:42 +0000 (15:47 -0600)]
mtd: expand configure check

SLES 11.4 has enough of mtd to pass the configure checks, but not
to compile it. Expand the checks to hopefully be enough.

Signed-off-by: Jens Axboe <>
8 years agogettime: silence bogus gcc warnings
Jens Axboe [Mon, 17 Aug 2015 21:46:22 +0000 (15:46 -0600)]
gettime: silence bogus gcc warnings

Signed-off-by: Jens Axboe <>
8 years agoMerge branch 'master' of git://
Jens Axboe [Mon, 17 Aug 2015 21:39:23 +0000 (15:39 -0600)]
Merge branch 'master' of git://

8 years agoallow for ratemin to be used without setting rate
DaveGlen [Mon, 17 Aug 2015 16:04:09 +0000 (10:04 -0600)]
allow for ratemin to be used without setting rate

Only check if ratemin < rate when rate is set.

8 years agofilesetup: make random generator block auto-switch on huge files
Jens Axboe [Fri, 14 Aug 2015 18:25:20 +0000 (12:25 -0600)]
filesetup: make random generator block auto-switch on huge files

Signed-off-by: Jens Axboe <>
8 years agoclient: make it explicit that we don't reuse 'eta' after freeing it
Jens Axboe [Fri, 14 Aug 2015 15:45:46 +0000 (09:45 -0600)]
client: make it explicit that we don't reuse 'eta' after freeing it

Signed-off-by: Jens Axboe <>
8 years agoeta: ensure we include terminating 0 space in je->run_str[]
Jens Axboe [Fri, 14 Aug 2015 15:42:55 +0000 (09:42 -0600)]
eta: ensure we include terminating 0 space in je->run_str[]

Signed-off-by: Jens Axboe <>
8 years agoarch-ppc.h: Add ilog2 implementation for ppc64
Chandan Rajendra [Tue, 11 Aug 2015 09:41:26 +0000 (15:11 +0530)]
arch-ppc.h: Add ilog2 implementation for ppc64

On a ppc64 machine, when fio is invoked with the following configuration file
(generated by fstests' generic/300 test), it loops indefinitely.







This is because arch_ffz() ends up invoking the 32-bit version of __ilog2()
which always returns a bit number in the range 0 - 31. This can cause
"overlap" variable in axmap_set_fn() to never becomes zero.

To fix this the commit adds a 64-bit version of __ilog2() (obtained from the
Linux kernel's arch/powerpc/include/asm/bitops.h).

Signed-off-by: Chandan Rajendra <>
Signed-off-by: Jens Axboe <>
8 years agoWhitespace fixup
Jens Axboe [Thu, 13 Aug 2015 15:31:17 +0000 (09:31 -0600)]
Whitespace fixup

Introduced by commit 50a8ce864e2c.

Signed-off-by: Jens Axboe <>
8 years agoinitialize rate_io_issue_bytes
DaveGlen [Mon, 10 Aug 2015 19:39:31 +0000 (13:39 -0600)]
initialize rate_io_issue_bytes

8 years agoImplement new Rate Control
DaveGlen [Mon, 10 Aug 2015 18:47:53 +0000 (12:47 -0600)]
Implement new Rate Control

Current implementation of rate control has the potential for bursts and
stalls when iodepth>1 due to feedback delay while IO is pending.  more
description here:

This commit changes the rate control mechanisms to use feed forward io
issues for rate feedback. Moving the rate control on submissions
instead of completions eliminates feedback delay. More details on the
change here:

8 years agoserver: make the setsockopt() error output a bit more informative
Jens Axboe [Wed, 29 Jul 2015 15:00:03 +0000 (09:00 -0600)]
server: make the setsockopt() error output a bit more informative

Signed-off-by: Jens Axboe <>
8 years agoconfigure: add --disable-optimizations
Jens Axboe [Mon, 27 Jul 2015 21:10:00 +0000 (15:10 -0600)]
configure: add --disable-optimizations

Sometimes we want to build without optimizations enabled for
debugging issues, make this possible without editing the

Also fixup --disable-shm to not output a symbol until after
we have initialized config-host.*.

Signed-off-by: Jens Axboe <>
8 years agoFix potential divide-by-zero in calc_iops()
Jens Axboe [Mon, 27 Jul 2015 18:29:55 +0000 (12:29 -0600)]
Fix potential divide-by-zero in calc_iops()

Signed-off-by: Jens Axboe <>
8 years agoFix potential divide-by-zero in calc_rate()
Jens Axboe [Mon, 27 Jul 2015 18:26:12 +0000 (12:26 -0600)]
Fix potential divide-by-zero in calc_rate()

Signed-off-by: Jens Axboe <>
8 years agoconfigure: make __sync_fetch_and_add() test use 64-bit types
Jens Axboe [Mon, 27 Jul 2015 18:22:14 +0000 (12:22 -0600)]
configure: make __sync_fetch_and_add() test use 64-bit types

Signed-off-by: Jens Axboe <>
8 years agolibfio: add runstate names
Jens Axboe [Mon, 27 Jul 2015 18:08:33 +0000 (12:08 -0600)]
libfio: add runstate names

Enables --debug=process to show the actual transitions, instead of
numbers you have to look up in the source:

process  12113 pid=12123: runstate CREATED -> INITIALIZED

instead of

process  12113 pid=12123: runstate 1 -> 2

Signed-off-by: Jens Axboe <>
8 years agoRe-seed random generator correctly between loops
Dave [Thu, 16 Jul 2015 20:56:00 +0000 (14:56 -0600)]
Re-seed random generator correctly between loops

re-seeding random generator in this way means that random patterns will
precisely repeat between iterations.  if rand_repeatable is set, this is
desirable, otherwise it may produce unexpected IO patterns in random IO

Signed-off-by: Jens Axboe <>
8 years agoREADME: no delay on sync to backup git hosts anymore
Jens Axboe [Wed, 15 Jul 2015 14:43:03 +0000 (08:43 -0600)]
README: no delay on sync to backup git hosts anymore

Signed-off-by: Jens Axboe <>
8 years agoHOWTO: spelling error
Jens Axboe [Mon, 13 Jul 2015 20:33:19 +0000 (14:33 -0600)]
HOWTO: spelling error

Another auto-sync test.

Signed-off-by: Jens Axboe <>
8 years agoHOWTO: fix long line
Jens Axboe [Mon, 13 Jul 2015 20:30:03 +0000 (14:30 -0600)]
HOWTO: fix long line

Testing to see if the inline mirroring works.

Signed-off-by: Jens Axboe <>
8 years agoUpdate README for DragonFly
Jens Axboe [Mon, 13 Jul 2015 18:41:51 +0000 (12:41 -0600)]
Update README for DragonFly

Signed-off-by: Jens Axboe <>
8 years agoRDMA engine server mode flow fix ups.
Logan Gunthorpe [Fri, 10 Jul 2015 18:23:55 +0000 (12:23 -0600)]
RDMA engine server mode flow fix ups.

When running fio in RDMA server mode, the user just sees an ever
increasing ETA line. Which is rather uninfromative. This patch
adds a waiting for connection message and sets the thread to SETTING_UP
while waiting.

Then when the server is running the ETA is disabled. (The RDMA server
has no access to progress information.)

8 years agoError when the block size on the RDMA server is less than the client
Logan Gunthorpe [Fri, 10 Jul 2015 17:54:25 +0000 (11:54 -0600)]
Error when the block size on the RDMA server is less than the client

When running an rdma client with a block size set less than
the block size set on the server, currently the code crashes.

This patch has the two sides exchange max_bs values and errors
out gracefully if the server side is too small.

8 years agobuffer: only set refill_buffers, it it wasn't set manually
Jens Axboe [Fri, 10 Jul 2015 15:27:02 +0000 (09:27 -0600)]
buffer: only set refill_buffers, it it wasn't set manually

Signed-off-by: Jens Axboe <>
8 years agoMake td flags be shifts
Jens Axboe [Fri, 10 Jul 2015 15:11:26 +0000 (09:11 -0600)]
Make td flags be shifts

Signed-off-by: Jens Axboe <>
8 years agoverify: fix bad 'from_verify' setting in do_io()
Jens Axboe [Thu, 9 Jul 2015 19:13:43 +0000 (13:13 -0600)]
verify: fix bad 'from_verify' setting in do_io()

Signed-off-by: Jens Axboe <>
8 years agoconfigure: don't be so eager to overwrite an existing Makefile
Jeremy Fitzhardinge [Wed, 8 Jul 2015 20:50:35 +0000 (13:50 -0700)]
configure: don't be so eager to overwrite an existing Makefile

8 years agofio: allow for build in separate build directory
Jeremy Fitzhardinge [Wed, 8 Jul 2015 20:27:35 +0000 (13:27 -0700)]
fio: allow for build in separate build directory

Change configure so that it will use its current directory for building,
and the path to the script itself as the source directory.

Change the Makefile to use VPATH to find the source files. Only a few other
things needed to be touched:
 - use the full path to the source in wildcard, then strip it off again
 - search both build and source for header files
 - FIO-VERSION-GEN is in source
 - make directories in build as needed
 - use $< to refer to input files
 - install non-executables from source

Signed-off-by: Jeremy Fitzhardinge <>
8 years agot/debug: fix 'fio_debug' declaration
Jens Axboe [Fri, 3 Jul 2015 20:26:08 +0000 (14:26 -0600)]
t/debug: fix 'fio_debug' declaration

debug.h:38:22: warning: type of ‘fio_debug’ does not match original declaration
 extern unsigned long fio_debug;
t/debug.c:5:14: note: previously declared here
 unsigned int fio_debug = 0;

Signed-off-by: Jens Axboe <>
8 years agosmalloc: smalloc() already clears memory, scalloc() need not do it again
Jens Axboe [Fri, 3 Jul 2015 17:47:10 +0000 (11:47 -0600)]
smalloc: smalloc() already clears memory, scalloc() need not do it again

Signed-off-by: Jens Axboe <>
8 years agosmalloc: fix unused 'int_mask' warning if redzoning is turned off
Jens Axboe [Tue, 30 Jun 2015 21:31:50 +0000 (15:31 -0600)]
smalloc: fix unused 'int_mask' warning if redzoning is turned off

Signed-off-by: Jens Axboe <>
8 years agoconfigure: fixup clang stupidity
Jens Axboe [Tue, 30 Jun 2015 02:46:31 +0000 (20:46 -0600)]
configure: fixup clang stupidity

Doesn't like doing _Static_assert() on our structure offset

error: static_assert expression is not an integral constant expression

Lets just include that in the configure test, so we can avoid using
_Static_assert() if that kind of check fails.

Signed-off-by: Jens Axboe <>
8 years agot/stest: shrink test size from 128MB
Jens Axboe [Tue, 30 Jun 2015 02:27:37 +0000 (20:27 -0600)]
t/stest: shrink test size from 128MB

We reduced the smalloc size, reduce it in the test cases too.

Signed-off-by: Jens Axboe <>
8 years agoKill duplicate __sync_fetch_and_add()
Jens Axboe [Mon, 29 Jun 2015 19:19:24 +0000 (13:19 -0600)]
Kill duplicate __sync_fetch_and_add()

We already had CONFIG_SFAA, just missed it. No need to duplicate it,
get rid of it and update the check in workqueue.c.

Signed-off-by: Jens Axboe <>
8 years agoUse _Static_assert() if available
Alireza Haghdoost [Mon, 29 Jun 2015 19:09:37 +0000 (13:09 -0600)]
Use _Static_assert() if available

The current compiletime_assert() doesn't work properly if code
optimization is not enabled, causing all compile time asserts to fail
regardless of the condition. This is because the compiler doesn't
realize that the external function will never get called for the
false case.

Use _Static_assert() if available for compiletime_assert(), and
fallback to the old method if we don't have it.

Signed-off-by: Jens Axboe <>
8 years agoworkqueue: make it work on platforms without __sync_fetch_and_add()
Jens Axboe [Mon, 29 Jun 2015 15:34:39 +0000 (09:34 -0600)]
workqueue: make it work on platforms without __sync_fetch_and_add()

Signed-off-by: Jens Axboe <>
8 years agoFio 2.2.9 fio-2.2.9
Jens Axboe [Thu, 25 Jun 2015 17:13:19 +0000 (11:13 -0600)]
Fio 2.2.9

Signed-off-by: Jens Axboe <>
8 years agoFix testing and setting set_options bitmap
Akinobu Mita [Thu, 25 Jun 2015 13:39:41 +0000 (22:39 +0900)]
Fix testing and setting set_options bitmap

set_options bitmap is an array of uint64_t.  But while testing and
setting a bit in the bitmap, the bit mask is calculated with an
unsigned long value.  For the systems which have 32-bit long type,
upper 32-bit cannot be set correctly.

Fix it by using (uint64_t)1 instead of 1UL to calculate correct bit

Signed-off-by: Akinobu Mita <>
Signed-off-by: Jens Axboe <>
8 years agoModify rdma engine to use proper arguments.
Logan Gunthorpe [Tue, 23 Jun 2015 22:21:07 +0000 (16:21 -0600)]
Modify rdma engine to use proper arguments.

The old rdma options are unintuitive and inflexible. We are also going
to do some experiments with this engine which require adding further
options. Thus we've improved it to take more normal arguments.

The technique is copied from the netio engine. It now requires a
hostname, port and verb option. The fio scripts in the example directory
have been updated. Compatability is also maintained for fio scripts that
use the old options.

8 years agoadd eta and elapsed to root of json output
Christopher Jacobs [Mon, 22 Jun 2015 21:14:50 +0000 (14:14 -0700)]
add eta and elapsed to root of json output

8 years agoFix latency logging for io_submit_mode=offload
Jens Axboe [Sat, 20 Jun 2015 17:21:48 +0000 (13:21 -0400)]
Fix latency logging for io_submit_mode=offload

Signed-off-by: Jens Axboe <>
8 years agoHOWTO: make it clear that deviation is standard deviation
Chris Worley [Tue, 16 Jun 2015 15:30:59 +0000 (09:30 -0600)]
HOWTO: make it clear that deviation is standard deviation

Signed-off-by: Jens Axboe <>
8 years agoEnable FIO_HAVE_CHARDEV_SIZE on DragonFlyBSD
Tomohiro Kusumi [Thu, 11 Jun 2015 11:24:11 +0000 (20:24 +0900)]

DragonFlyBSD no longer has block device just like FreeBSD got rid of
it at some point. Enable FIO_HAVE_CHARDEV_SIZE and implement get-size
functions with DragonFlyBSD's ioctl so fio can retrieve correct size
when targets are not regular files.

The following result verifies that df(1) shows the same fs size as
chardev_size() with regards to the character device (block device if
it were on Linux).

 # uname
 # cat ./test3.c
 #include <stdio.h>
 #include <stdlib.h>
 #include <fcntl.h>
 #include "os/os-dragonfly.h"

 int main(int argc, char *argv[]) {
         struct fio_file f;
         unsigned long long bytes;

         f.fd = open(argv[1], O_RDONLY);
         if (f.fd < 0) {

         if (chardev_size(&f, &bytes)) {
         printf("%s %llu\n", argv[1], bytes);
         return 0;
 # gcc -Wall -g ./test3.c -o test3
 # file /dev/da8
 /dev/da8: character special (30/504430663)
 # ./test3 /dev/da8
 /dev/da8 31004295168
 # newfs /dev/da8 > /dev/null
 # mount -t ufs /dev/da8 /mnt
 # df -TH /mnt
 Filesystem  Type   Size   Used  Avail Capacity  Mounted on
 /dev/da8    ufs     31G   2.0k    28G     0%    /mnt

8 years agoRename get_fs_size() to get_fs_free_size()
Tomohiro Kusumi [Wed, 10 Jun 2015 17:46:31 +0000 (02:46 +0900)]
Rename get_fs_size() to get_fs_free_size()

get_fs_free_size() seems to be an appropriate name considering
these functions on various os return filesystem's free space by
computing (block_size * number_of_free_blocks) via statfs(2) or

Also the caller of this function is get_fs_free_counts().

8 years agoAdd get_fs_size() support for BSDs
Tomohiro Kusumi [Wed, 10 Jun 2015 12:45:00 +0000 (21:45 +0900)]
Add get_fs_size() support for BSDs

Add get_fs_size() for FreeBSD/DragonFlyBSD. The same code may
work on NetBSD/OpenBSD as it uses POSIX statvfs(2).

The following results verify df(1) (which also typically uses
statfs(2) or statvfs(2)) shows the same number given that df(1)
output is correct for the given filesystems.

 # uname
 # cat ./test2.c
 #include <stdio.h>
 #include "os/os-freebsd.h"
 int main(int argc, char *argv[]) {
         printf("%s %llu\n", argv[1], get_fs_size(argv[1]));
         return 0;
 # clang -Wall -g ./test2.c -o test2
 # ./test2 /
 / 102783143936
 # df -TH /
 Filesystem   Type    Size    Used   Avail Capacity  Mounted on
 /dev/ada0p2  ufs     112G    9.5G     94G     9%    /
                      ^^^^^^^^^^^^     ^^^
                      112G - 9.5G      bfree!=bavail
                      = 102.5G         like it differs on ext[234]
 # uname
 # cat ./test2.c
 #include <stdio.h>
 #include "os/os-dragonfly.h"
 int main(int argc, char *argv[]) {
         printf("%s %llu\n", argv[1], get_fs_size(argv[1]));
         return 0;
 # gcc -Wall -g ./test2.c -o test2
 # ./test2 /
 / 450287370240
 # df -TH /
 Filesystem  Type     Size   Used  Avail Capacity  Mounted on
 ROOT        hammer   480G    30G   450G     6%    /

8 years agoAdd header include for DragonFlyBSD
Tomohiro Kusumi [Wed, 10 Jun 2015 12:52:45 +0000 (21:52 +0900)]
Add header include for DragonFlyBSD

Add '#include <unistd.h>' to os/os-dragonfly.h for lwp_gettid(2).
(No error on make, but needs <unistd.h> to include this header

8 years agoFix compiler warning
Tomohiro Kusumi [Tue, 9 Jun 2015 23:28:13 +0000 (08:28 +0900)]
Fix compiler warning

The local variable off is uninitialized on 'goto fill'.
Initializing with 0 seems to be appropriate to fix following gcc warning
although off is actually always initialized as the function returns with
1 on 'if (!i)' case.

options.c:920: warning: 'off' may be used uninitialized in this function

8 years agoDropped commented out part of line.
Sam Zaydel [Tue, 9 Jun 2015 21:14:20 +0000 (14:14 -0700)]
Dropped commented out part of line.

8 years agoMerge branch 'master' of into fixtype
Sam Zaydel [Tue, 9 Jun 2015 21:11:36 +0000 (14:11 -0700)]
Merge branch 'master' of into fixtype

8 years agoint may be too small for a size_t on some systems.
Sam Zaydel [Tue, 9 Jun 2015 21:10:25 +0000 (14:10 -0700)]
int may be too small for a size_t on some systems.

8 years agooptions: use fio_option_is_set() to detect verify setting
Jens Axboe [Tue, 9 Jun 2015 20:00:56 +0000 (14:00 -0600)]
options: use fio_option_is_set() to detect verify setting

Signed-off-by: Jens Axboe <>
8 years agoverify: cleanup code for verify pattern without headers
Jens Axboe [Tue, 9 Jun 2015 14:35:29 +0000 (08:35 -0600)]
verify: cleanup code for verify pattern without headers

Signed-off-by: Jens Axboe <>
8 years agoverify: fix dumping of received/expected buffers on failure
Jens Axboe [Tue, 9 Jun 2015 14:06:02 +0000 (08:06 -0600)]
verify: fix dumping of received/expected buffers on failure

If verify=pattern is used, we don't have a header to rely on. So
make one up for that case.

Signed-off-by: Jens Axboe <>
8 years agoverify: add raw pattern verify
Jens Axboe [Tue, 9 Jun 2015 02:43:43 +0000 (20:43 -0600)]
verify: add raw pattern verify

Add specific verify=pattern that doesn't use any headers, it
just writes the specified pattern and verifies it.

Signed-off-by: Jens Axboe <>
8 years agobackend: don't add to runtime for fake writes
Jens Axboe [Tue, 9 Jun 2015 00:43:22 +0000 (18:43 -0600)]
backend: don't add to runtime for fake writes

This fixes showing that fio did writes with --verify_only being set,
while in reality it just ended them before submitting to the device.

Signed-off-by: Jens Axboe <>
8 years agoClear sysfs path before reading current ioscheduler from sysfs
Tomohiro Kusumi [Mon, 8 Jun 2015 10:19:02 +0000 (19:19 +0900)]
Clear sysfs path before reading current ioscheduler from sysfs

switch_ioscheduler() function has a local buffer tmp[] to store
both sysfs path (e.g. /sys/block/`device`/queue/scheduler) and
content of that sysfs path.

In order to properly test strstr() after writing a ioscheduler
string (e.g. "deadline") to sysfs, it needs to memset(0) first.
Otherwise if the content of sysfs path (below (b)) is shorter
than the existing sysfs path in tmp[] (below (a)), then tmp[]
after storing the content of sysfs (below (c) and *haystack of
strstr()) contains remaining part of the sysfs path.

(a) "/sys/block/sdb/queue/scheduler"
(b) "noop [deadline] cfq \n"
(c) "noop [deadline] cfq \nscheduler"

strstr() will result the same given that the remaining part of
the sysfs path is unlikely to contain ioscheduler string, but
the remaining part should still be cleared first.

8 years agoconfigure: add some missing help options
Jens Axboe [Fri, 5 Jun 2015 14:33:56 +0000 (08:33 -0600)]
configure: add some missing help options

Signed-off-by: Jens Axboe <>
8 years agoFix compile error on non Linux-or-BSDs
Tomohiro Kusumi [Fri, 5 Jun 2015 10:15:43 +0000 (19:15 +0900)]
Fix compile error on non Linux-or-BSDs

e7e136da needed to add #else part with another device_is_mounted()
that returns 0 for non Linux-or-BSDs environment. Any environment
without getmntent(3) or getmntinfo(3) needs a blank function.

8 years agoinit: automate displaying debug categories
Jens Axboe [Thu, 4 Jun 2015 19:45:32 +0000 (13:45 -0600)]
init: automate displaying debug categories

Better than having them hardwired, so we don't have to update the
code in several places when adding new categories.

Signed-off-by: Jens Axboe <>
8 years agoAdd device_is_mounted() support for BSDs
Tomohiro Kusumi [Tue, 2 Jun 2015 14:40:44 +0000 (23:40 +0900)]
Add device_is_mounted() support for BSDs

Implement '#elif' part of lib/mountcheck.c using getmntinfo(3).

This should implement device_is_mounted() on most BSDs, and has been
tested on FreeBSD/DragonFlyBSD. NetBSD/OpenBSD do have getmntinfo(3)
but may need to include different headers according to their online

 # cat ./test1.c
 #include <stdio.h>
 #include "./lib/mountcheck.h"
 int main(int argc, char *argv[]) {
  printf("%s %d\n", argv[1], device_is_mounted(argv[1]));
  return 0;

 # uname
 # clang -Wall -g ./test1.c ./lib/mountcheck.o -o ./test1
 # ./test1 /dev/ada0p2  /* UFS */
 /dev/ada0p2 1
 # ./test1 zfs  /* ZFS */
 zfs 1
 # ./test1 zfs/test  /* ZFS */
 zfs/test 1
 # ./test1 /dev/da0  /* not mounted device */
 /dev/da0 0
 # ./test1 random  /* irrelevant string */
 random 0

 # uname
 # gcc -Wall -g ./test1.c ./lib/mountcheck.o -o ./test1
 # ./test1 /pfs/@@-1:00005  /* HAMMER */
 /pfs/@@-1:00005 1
 # ./test1 procfs  /* procfs */
 procfs 1
 # ./test1 /dev/da8  /* not mounted device */
 /dev/da8 0
 # ./test1 random  /* irrelevant string */
 random 0

8 years agoMerge branch 'req6' of git://
Jens Axboe [Mon, 1 Jun 2015 14:32:45 +0000 (08:32 -0600)]
Merge branch 'req6' of git://

8 years agoMinor cleanups on cleanup()
Tomohiro Kusumi [Mon, 1 Jun 2015 13:21:29 +0000 (22:21 +0900)]
Minor cleanups on cleanup()

Sync with the way other engines implement cleanup() that is regarded
as a better practice. td->io_ops->data should have null check before
dereference and its members can be freed without null check.

8 years agoFix build error on non-GNU environment
Tomohiro Kusumi [Mon, 1 Jun 2015 12:32:37 +0000 (21:32 +0900)]
Fix build error on non-GNU environment

'#include <mntent.h>' needs to be inside '#ifdef CONFIG_GETMNTENT'
and that was probably the intention of the commit aae599ba.
e.g. BSDs are likely to hit following compile error.

lib/mountcheck.c:3:20: fatal error: mntent.h: No such file or directory
compilation terminated.
Makefile:275: recipe for target 'lib/mountcheck.o' failed
gmake: *** [lib/mountcheck.o] Error 1

8 years agordma: adapt to new init_rand_seed()
Jens Axboe [Sat, 30 May 2015 19:04:33 +0000 (13:04 -0600)]
rdma: adapt to new init_rand_seed()

engines/rdma.c: In function ‘fio_rdmaio_setup’:
engines/rdma.c:1205: error: too few arguments to function ‘init_rand_seed’
make: *** [engines/rdma.o] Error 1

Signed-off-by: Jens Axboe <>
8 years agofilesetup: print warning if chosen random generator can't cover range
Jens Axboe [Sat, 30 May 2015 01:40:21 +0000 (19:40 -0600)]
filesetup: print warning if chosen random generator can't cover range

If the file or device is huge, we could have an issue with the
default 32-bit random offset generator in that it wont cover
the entire device. If fio detects this and the random generator
wasn't explicitly set, then log an error and exit.

Signed-off-by: Jens Axboe <>
8 years agorand: add 64-bit tausworthe variant with a 2^258 cycle
Jens Axboe [Fri, 29 May 2015 18:49:54 +0000 (12:49 -0600)]
rand: add 64-bit tausworthe variant with a 2^258 cycle

Signed-off-by: Jens Axboe <>