3 years agoappveyor: disable setting compiler march
Sitsofe Wheeler [Wed, 7 Mar 2018 16:36:32 +0000 (16:36 +0000)]
appveyor: disable setting compiler march

Commit a817dc3b3b5a0efc95aaca366875eac67607cd5b ("Default to building
native code") tunes binaries to only run on the build system's
architecture by default. However as AppVeyor builds Windows MSIs for
public consumption, let's turn that feature off here to prevent people
accidentally ending up with binaries that contain instructions their
system can't run.

Signed-off-by: Sitsofe Wheeler <>
3 years agoconfigure: make --disable-optimizations disable march=native
Sitsofe Wheeler [Wed, 7 Mar 2018 09:54:45 +0000 (09:54 +0000)]
configure: make --disable-optimizations disable march=native

Signed-off-by: Sitsofe Wheeler <>
3 years agoconfigure: don't override march if already set
Jens Axboe [Wed, 7 Mar 2018 15:32:05 +0000 (08:32 -0700)]
configure: don't override march if already set

Some archs (like arm or s390) already set an appropriate architecture,
so don't override it for those cases.

Signed-off-by: Jens Axboe <>
3 years agoReduce LOG_MSEC_SLACK
Jeff Furlong [Wed, 7 Mar 2018 00:53:20 +0000 (17:53 -0700)]

Starting in fio 2.99, there is an odd problem when --ramp_time is used
with --write_iops_log.  The reported IOPs for the first ~200s or so are
artificially inflated by fio, always monotonically increasing until
about 200s, then resets to normal.  With fio 2.99, disabling --ramp_time
avoids the issue.  And, with fio 2.21, using --ramp_time and
--write_iops_log does not create the problem.  Consider a job:

fio --name=test --ioengine=libaio --direct=1 --rw=randread --iodepth=1 \\
--size=100% --bs=4k --filename=/dev/nvme1n1 --runtime=300s \\
--ramp_time=10s --write_iops_log=test --log_avg_msec=1000

See the attached plot for fio induced IOPs during the first ~200s.
Using a longer runtime makes identifying the issue easier if the DUT is
consistently performing.

One simple method to resolve this inflation is by reducing the
LOG_MSEC_SLACK from 10 milliseconds to 1 millisecond, given that we now
operate some time structures to the nanosecond level (so 10ms of slack
seems excessive).  Below I also fixup some data type issues to match
existing types (avoids typecasts or potential compiler warnings).

Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'wip-rbd-engine' of
Jens Axboe [Wed, 7 Mar 2018 00:44:17 +0000 (17:44 -0700)]
Merge branch 'wip-rbd-engine' of

* 'wip-rbd-engine' of
  rbd: remove support for blkin tracing
  rbd: fixed busy-loop when using eventfd polling

3 years agorbd: remove support for blkin tracing
Jason Dillaman [Wed, 7 Mar 2018 00:05:01 +0000 (19:05 -0500)]
rbd: remove support for blkin tracing

The referenced APIs for blkin tracing were not merged into librbd.
Instead, blkin tracing can be enabled for all IO by setting
'rbd blkin trace all = true' in the config.

3 years agorbd: fixed busy-loop when using eventfd polling
Jason Dillaman [Wed, 7 Mar 2018 00:01:18 +0000 (19:01 -0500)]
rbd: fixed busy-loop when using eventfd polling

The eventfd poll would continuously fire since the fd was never
read. Switched to a semaphore eventfd and now drain the events
as completions are handled.

3 years agoDefault to building native code
Jens Axboe [Tue, 6 Mar 2018 21:28:16 +0000 (14:28 -0700)]
Default to building native code

If the local compiler supports it, use -march=native to ensure we
take full advantage of a local build.

If building distro packages, or otherwise for a different platform
than the host, then use --disable-native with configure to remove
this optimization.

Signed-off-by: Jens Axboe <>
3 years agoDon't make fadvise failure fatal
Jens Axboe [Tue, 6 Mar 2018 19:42:57 +0000 (12:42 -0700)]
Don't make fadvise failure fatal

It's just an advisory hint. Warn once that it failed, then
march on.

Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'howto_typos' of
Jens Axboe [Mon, 5 Mar 2018 17:29:39 +0000 (10:29 -0700)]
Merge branch 'howto_typos' of

* 'howto_typos' of
  doc: fix typos in HOWTO

3 years agodoc: fix typos in HOWTO
Patrick Callahan [Mon, 5 Mar 2018 17:23:51 +0000 (12:23 -0500)]
doc: fix typos in HOWTO

3 years agoMerge branch 'manpage_typos' of
Jens Axboe [Mon, 5 Mar 2018 15:47:51 +0000 (08:47 -0700)]
Merge branch 'manpage_typos' of

* 'manpage_typos' of
  doc: fix typos in fio man page

3 years agodoc: fix typos in fio man page
Patrick Callahan [Mon, 5 Mar 2018 14:23:39 +0000 (09:23 -0500)]
doc: fix typos in fio man page

3 years agoMerge branch 'mpath_nvme_diskutil' of
Jens Axboe [Fri, 2 Mar 2018 19:02:08 +0000 (12:02 -0700)]
Merge branch 'mpath_nvme_diskutil' of

* 'mpath_nvme_diskutil' of
  diskutil: minor style cleanup
  diskutil: try additional slave device path if first fails

3 years agodiskutil: minor style cleanup
Sitsofe Wheeler [Fri, 2 Mar 2018 15:51:01 +0000 (15:51 +0000)]
diskutil: minor style cleanup

Signed-off-by: Sitsofe Wheeler <>
3 years agodiskutil: try additional slave device path if first fails
Potnuri Bharat Teja [Fri, 2 Mar 2018 15:26:01 +0000 (15:26 +0000)]
diskutil: try additional slave device path if first fails

Running fio on an NVMEoF device with multipath slaves fails with a "No
such file or directory" error when displaying disk stats. Correct this
by using a different path to the sysfs slave device.

[sitsofe: change new path to be a fallback when regular path fails]


Signed-off-by: Potnuri Bharat Teja <>
Signed-off-by: Sitsofe Wheeler <>
3 years agoMerge branch 'hotfix_counter_overflow' of
Jens Axboe [Thu, 1 Mar 2018 16:18:43 +0000 (09:18 -0700)]
Merge branch 'hotfix_counter_overflow' of

* 'hotfix_counter_overflow' of
  Fix overflow of counters incremented on each I/O operation

3 years agoFix overflow of counters incremented on each I/O operation
Alexander Larin [Wed, 28 Feb 2018 11:25:42 +0000 (14:25 +0300)]
Fix overflow of counters incremented on each I/O operation

 - In the thread_stat struct: uint32_t counters are updated to uint64_t.
 - In the io_u_plat_entry struct: unsigned int counters are updated to
It fixes overflow of these counters.

Signed-off-by: Alexander Larin <>
3 years agoMerge branch 'win_build' of
Jens Axboe [Wed, 28 Feb 2018 16:44:08 +0000 (09:44 -0700)]
Merge branch 'win_build' of

* 'win_build' of
  windows: document MinGW zlib install and remove custom zlib search
  windows: minor windows installer improvements
  appveyor: minor refactoring, clarifications

3 years agoconfigure: improve static zlib package warning
Jens Axboe [Wed, 28 Feb 2018 15:43:22 +0000 (08:43 -0700)]
configure: improve static zlib package warning

Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'master' of
Jens Axboe [Tue, 27 Feb 2018 23:15:22 +0000 (16:15 -0700)]
Merge branch 'master' of

* 'master' of
  Refer to zlib-static package in zlib static build warning

3 years agoRefer to zlib-static package in zlib static build warning
Bryce Guinta [Tue, 27 Feb 2018 23:09:35 +0000 (16:09 -0700)]
Refer to zlib-static package in zlib static build warning

Signed-off-by: Bryce Guinta <>
3 years agoMerge branch 'wip-ifed-howto-update' of
Jens Axboe [Mon, 26 Feb 2018 18:59:54 +0000 (11:59 -0700)]
Merge branch 'wip-ifed-howto-update' of

* 'wip-ifed-howto-update' of
  Update HOWTO with RADOS information

3 years agoUpdate HOWTO with RADOS information
Igor Fedotov [Mon, 26 Feb 2018 17:53:52 +0000 (20:53 +0300)]
Update HOWTO with RADOS information

Signed-off-by: Igor Fedotov <>
3 years agowindows: document MinGW zlib install and remove custom zlib search
Sitsofe Wheeler [Wed, 21 Feb 2018 18:45:55 +0000 (18:45 +0000)]
windows: document MinGW zlib install and remove custom zlib search

In 5580d200d7ae49cdb4f30913e3ef2d74e431ff60 ("appveyor: install zlib and minor
clean ups") I found AppVeyor builds could use a pre-built, statically linkable
zlib from Cygwin/MinGW.

Document the MinGW zlib packages in the README and remove the custom Windows
zlib searching from the configure script.

Signed-off-by: Sitsofe Wheeler <>
3 years agowindows: minor windows installer improvements
Sitsofe Wheeler [Wed, 21 Feb 2018 14:41:14 +0000 (14:41 +0000)]
windows: minor windows installer improvements

- Stop hard coding fio version in the Windows installer build file and try and
  use the generated version instead
- Make installer build use sys.BUILDARCH rather than env.FIO_ARCH
- Change IO to I/O in descriptions to match other packages

Signed-off-by: Sitsofe Wheeler <>
3 years agoappveyor: minor refactoring, clarifications
Sitsofe Wheeler [Wed, 21 Feb 2018 09:56:38 +0000 (09:56 +0000)]
appveyor: minor refactoring, clarifications

- Set platform directly to x64 or x86 and use PLATFORM instead of BUILD_ARCH
- Set the PATH once (via the install section) because environment variables
  persist to later sections
- Add a note that shallow cloning means FIO-VERSION-GEN is unable to build a
  fancy version number (because it won't find any tags)

Signed-off-by: Sitsofe Wheeler <>
3 years agoFio 3.5 fio-3.5
Jens Axboe [Tue, 20 Feb 2018 22:30:28 +0000 (15:30 -0700)]
Fio 3.5

Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'wip-ifed-rados' of
Jens Axboe [Wed, 14 Feb 2018 19:43:04 +0000 (12:43 -0700)]
Merge branch 'wip-ifed-rados' of

* 'wip-ifed-rados' of
  Add support for Ceph Rados benchmarking.

3 years agoAdd support for Ceph Rados benchmarking.
Igor Fedotov [Fri, 9 Feb 2018 14:39:28 +0000 (17:39 +0300)]
Add support for Ceph Rados benchmarking.

Signed-off-by: Igor Fedotov <>
3 years agodebug: fix inverted logic in fio_did_warn() fio-3.4
Jens Axboe [Mon, 12 Feb 2018 17:55:07 +0000 (10:55 -0700)]
debug: fix inverted logic in fio_did_warn()

Fixes: 5ec3bc5401f3 ("init: add global 'warned' state")
Signed-off-by: Jens Axboe <>
3 years agoiolog: convert drop warning to fio_did_warn()
Jens Axboe [Mon, 12 Feb 2018 17:19:18 +0000 (10:19 -0700)]
iolog: convert drop warning to fio_did_warn()

Signed-off-by: Jens Axboe <>
3 years agoio_u: convert zoned bug warning to fio_did_warn()
Jens Axboe [Mon, 12 Feb 2018 17:18:32 +0000 (10:18 -0700)]
io_u: convert zoned bug warning to fio_did_warn()

Signed-off-by: Jens Axboe <>
3 years agoverify: convert verify buf too small warning to fio_did_warn()
Jens Axboe [Mon, 12 Feb 2018 17:14:35 +0000 (10:14 -0700)]
verify: convert verify buf too small warning to fio_did_warn()

Signed-off-by: Jens Axboe <>
3 years agofilesetup: convert root flush warning to fio_did_warn()
Jens Axboe [Mon, 12 Feb 2018 17:13:04 +0000 (10:13 -0700)]
filesetup: convert root flush warning to fio_did_warn()

Signed-off-by: Jens Axboe <>
3 years agoinit: add global 'warned' state
Jens Axboe [Mon, 12 Feb 2018 17:11:31 +0000 (10:11 -0700)]
init: add global 'warned' state

In various places we spew a warning if some static variable
isn't set, but this still means it can happen once per job.
Add some global state to allow to track this globally instead.

Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'master' of
Jens Axboe [Mon, 12 Feb 2018 15:24:21 +0000 (08:24 -0700)]
Merge branch 'master' of

* 'master' of
  Make fiologparser_hist compatible with python3

3 years agoinit: fixup some bad style in previous commit
Jens Axboe [Sat, 10 Feb 2018 21:44:49 +0000 (14:44 -0700)]
init: fixup some bad style in previous commit

Make it follow the fio coding style.

Signed-off-by: Jens Axboe <>
3 years agoinit: fix broken verify_interval
Damian Yurzola [Fri, 9 Feb 2018 01:29:38 +0000 (17:29 -0800)]
init: fix broken verify_interval

The operands seems to have been inverted which in turn
created the situation whereby the interval was always
changed to match the min_bs

Signed-off-by: Damian Yurzola <>
3 years agoInclude 'numjobs' in global options output
Jens Axboe [Thu, 8 Feb 2018 22:46:46 +0000 (15:46 -0700)]
Include 'numjobs' in global options output

Signed-off-by: Jens Axboe <>
3 years agommap: don't include MADV_FREE in fadvise_hint check
Jens Axboe [Wed, 7 Feb 2018 18:30:59 +0000 (11:30 -0700)]
mmap: don't include MADV_FREE in fadvise_hint check

This really should be based on invalidate, leave it as it was
for now.

Fixes: c712c97ab871 ("Let fadvise_hint also apply too mmap engine and madvise")
Signed-off-by: Jens Axboe <>
3 years agoLet fadvise_hint also apply too mmap engine and madvise
Jens Axboe [Wed, 7 Feb 2018 18:13:05 +0000 (11:13 -0700)]
Let fadvise_hint also apply too mmap engine and madvise

Signed-off-by: Jens Axboe <>
3 years agoMake fiologparser_hist compatible with python3
Bryce Guinta [Tue, 6 Feb 2018 00:36:33 +0000 (17:36 -0700)]
Make fiologparser_hist compatible with python3

In python3, map functions return generators, which caused numpy to break
in this script.

Use list comprehensions instead of map (more pythonic)
They also have the added benefit of behaving the same
between python2 and python3.

3 years agostat: add total fsync ios to json output
Jens Axboe [Thu, 25 Jan 2018 23:04:20 +0000 (16:04 -0700)]
stat: add total fsync ios to json output

Signed-off-by: Jens Axboe <>
3 years agoio_ddir: revert separate ddir count change
Jens Axboe [Thu, 25 Jan 2018 17:28:23 +0000 (10:28 -0700)]
io_ddir: revert separate ddir count change

This essentially reverts the previous commit. We need more
careful changes for looping and indexing, the easiest is just
to move the count values at the end as not to skew the values
of the actual ddir entries.

Signed-off-by: Jens Axboe <>
3 years agoio_ddir: move count values out of the enum fio_ddir
Jens Axboe [Thu, 25 Jan 2018 17:16:51 +0000 (10:16 -0700)]
io_ddir: move count values out of the enum fio_ddir

They are not directions, just counts.

Signed-off-by: Jens Axboe <>
3 years agostat: ensure that we align ts->sync_stat appropriately
Jens Axboe [Thu, 25 Jan 2018 16:51:07 +0000 (09:51 -0700)]
stat: ensure that we align ts->sync_stat appropriately

clang correctly complains that we don't know if the alignment is
correct when the structure is packed, so ensure that we align it
to an 8-byte boundary.

Signed-off-by: Jens Axboe <>
3 years agoTrack fsync/fdatasync/sync_file_range issue count
Jens Axboe [Thu, 25 Jan 2018 16:27:14 +0000 (09:27 -0700)]
Track fsync/fdatasync/sync_file_range issue count

Also fix an issue where we didn't initialize the sync minimum
latency correctly, so it was always reported as 0.

Signed-off-by: Jens Axboe <>
3 years agoSwitch last_was_sync and terminate to bool and pack better
Jens Axboe [Thu, 25 Jan 2018 03:22:50 +0000 (20:22 -0700)]
Switch last_was_sync and terminate to bool and pack better

Signed-off-by: Jens Axboe <>
3 years agoAdd suppor for logging fsync (and friends) latencies
Jens Axboe [Wed, 24 Jan 2018 22:32:07 +0000 (15:32 -0700)]
Add suppor for logging fsync (and friends) latencies

Signed-off-by: Jens Axboe <>
3 years agonull: drop unneeded casts from void* to non-void*
Tomohiro Kusumi [Mon, 15 Jan 2018 19:52:12 +0000 (21:52 +0200)]
null: drop unneeded casts from void* to non-void*

Some functions in null ioengine are used by C/C++ (e.g. null_init()
which calls malloc(3)), but C specific ones (not __cplusplus) don't
need explicit cast from void* to non-void*, and one usually doesn't.

Signed-off-by: Tomohiro Kusumi <>
Signed-off-by: Jens Axboe <>
3 years agonull: make *impl_ private
Tomohiro Kusumi [Mon, 15 Jan 2018 19:52:11 +0000 (21:52 +0200)]
null: make *impl_ private

This can/should be private, given struct is default public in C++.
The null ioengine only needs access to new'd NullData which
encapsulates malloc'd null_data *impl_.

Signed-off-by: Tomohiro Kusumi <>
Signed-off-by: Jens Axboe <>
3 years agonull: fix compile time warning on OpenBSD
Tomohiro Kusumi [Mon, 15 Jan 2018 19:52:10 +0000 (21:52 +0200)]
null: fix compile time warning on OpenBSD

Fix the following warning on OpenBSD 6.2 (gcc 4.2.1).
This can probably be silenced with void** argument and memcpy() too,
but returning a non-void pointer is clearer IMO when the code also
targets C++.

Compiles with regular make/gmake as well as
 # g++ -O2 -g -shared -rdynamic -fPIC -o cpp_null null.c -DFIO_EXTERNAL_ENGINE

    CC engines/null.o
engines/null.c: In function 'fio_null_init':
engines/null.c:142: warning: dereferencing type-punned pointer will break strict-aliasing rules

Signed-off-by: Tomohiro Kusumi <>
Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'diskless_invalidate' of
Jens Axboe [Fri, 12 Jan 2018 17:59:08 +0000 (10:59 -0700)]
Merge branch 'diskless_invalidate' of

3 years agoMerge branch 'fio-issue-450' of
Jens Axboe [Fri, 12 Jan 2018 15:22:41 +0000 (08:22 -0700)]
Merge branch 'fio-issue-450' of

3 years agofilesetup: skip fallback invalidation with diskless ioengines
Sitsofe Wheeler [Sat, 30 Dec 2017 20:58:59 +0000 (20:58 +0000)]
filesetup: skip fallback invalidation with diskless ioengines

Since commit 0bcf41cdc22dfee6b3f3b2ba9a533b4b103c70c2 ("io_u:
re-invalidate cache when looping around without file open/close") the

./fio --ioengine=null --size=4k --bs=4k --rate=4k \
 --time_based --runtime=2s --name=badinvalidate

repeatedly produces the following warning while running:

fio: cache invalidation of badinvalidate.0.0 failed: Bad file descriptor

Fix this by skipping cache invalidation and logging a debug message for
diskless ioengines that don't define an explicit invalidation op.

Signed-off-by: Sitsofe Wheeler <>
3 years agoioengines: don't call munmap unless size boundary is exceeded
Robert Elliott [Wed, 10 Jan 2018 15:50:01 +0000 (09:50 -0600)]
ioengines: don't call munmap unless size boundary is exceeded

In ioengines using mmap() and munmap(), don't call munmap() when doing an
IO that accesses the last byte of a file. This affects the dev-dax,
libpmem, and mmap ioengines.

io       12987 DEBUG fio_libpmem_prep
io       12987  io_u->offset 7516061696 : fdd->libpmem_off 0 : io_u->buflen 131072 : fdd->libpmem_sz 7516192768
io       12987 munmap addr=0x7f5600000000 size=0x1c0000000

That accesses the last 128 KiB of a 7 GiB file, triggering an munmap
because the comparison to 7 GiB fails:
7516061696 + 131072 < 7516192768

Change < to <= so that doesn't occur unless an overflow really

In one test in linux, this increases performance from 5 GB/s to 20 GB/s
as fio stops spending time in mmap() page table mapping spinlocks and
instead spends its time in memmove().  The perf top results improve from:
  36.81%  [kernel]             [k] queued_spin_lock_slowpath
  24.26%         [.] __memmove_avx_unaligned_erms
   4.09%  [kernel]             [k] __radix_tree_lookup
   3.85%  [kernel]             [k] find_next_iomem_res
   3.70%  [kernel]             [k] _raw_spin_lock_irq
   3.48%  [kernel]             [k] down_read

  79.76%         [.] __memmove_avx_unaligned_erms
   3.23%         [.] vfprintf

Signed-off-by: Jens Axboe <>
3 years agoio/examples/fio-seq-read: use direct=1
Jens Axboe [Wed, 10 Jan 2018 15:30:07 +0000 (08:30 -0700)]
io/examples/fio-seq-read: use direct=1

It's confusing to use direct=0 with libaio, since that isn't

Signed-off-by: Jens Axboe <>
3 years agostat: don't add duplicate clat entries for json
Jens Axboe [Sat, 6 Jan 2018 21:47:01 +0000 (14:47 -0700)]
stat: don't add duplicate clat entries for json

For some reason the initial commit always filled all
entries with zeroes, when it should just stop printing.
I suspect this was carried over from the terse/csv output,
where we do the same. But there it's actually valid, as we
have to have a fixed number of fields. For json we both
don't need it, and it's causing us to create invalid json
when we have duplicate keys.

Signed-off-by: Jens Axboe <>
3 years agoverify: don't adjust verification length based on interval when unaligned
Jens Axboe [Fri, 5 Jan 2018 20:38:40 +0000 (13:38 -0700)]
verify: don't adjust verification length based on interval when unaligned

It's perfectly fine to have a verification length that exceeds the
default verify interval when we have unaligned block sizes and
using block size ranges, since the verify interval just defaults to
the smallest blocksize.

Signed-off-by: Jens Axboe <>
3 years agoFix client/server "all clients" reporting
Jeff Furlong [Wed, 3 Jan 2018 23:20:38 +0000 (23:20 +0000)]
Fix client/server "all clients" reporting

In commit e883cb3501de0f99297282c18842d50aaa7fa531, the default significant figures was added.  The default value was preserved for local jobs and when client/server was used if jobs=1, but not when client/server was used if jobs > 1.  This patch restores some reporting outputs ("all clients" output for IOPS and BW) when client/server is used if > 1 jobs, for example:

All clients: (groupid=0, jobs=2): err= 0: pid=0: Wed Jan  3 14:30:13 2018
   read: IOPS=0, BW=0 (0)(0B/5001msec)                                               <--ZERO VALUES
    slat (nsec): min=762, max=108579, avg=1058.97, stdev=221.30
    clat (nsec): min=254, max=4544.2k, avg=11297.64, stdev=8284.14
     lat (usec): min=9, max=4545, avg=12.41, stdev= 8.30
   bw (  KiB/s): min=290192, max=347160, per=49.35%, avg=303613.33, stdev=11582.87, samples=18
   iops        : min=72548, max=86790, avg=75903.33, stdev=2895.72, samples=18

All clients: (groupid=0, jobs=2): err= 0: pid=0: Wed Jan  3 15:08:33 2018
   read: IOPS=147k, BW=575Mi (603M)(2876MiB/5001msec) <--CORRECT VALUES
    slat (nsec): min=841, max=61945, avg=1304.83, stdev=232.75
    clat (nsec): min=544, max=4487.2k, avg=11736.68, stdev=9123.64
     lat (usec): min=10, max=4489, avg=13.10, stdev= 9.13
   bw (  KiB/s): min=260945, max=305976, per=49.99%, avg=294350.22, stdev=9192.53, samples=18
   iops        : min=65236, max=76494, avg=73587.56, stdev=2298.13, samples=18


Signed-off-by: Jens Axboe <>
3 years agoChange bluestop link to be https
Jens Axboe [Fri, 5 Jan 2018 15:30:32 +0000 (08:30 -0700)]
Change bluestop link to be https

Doesn't answer on http:// anymore.

Also include link to AppVeyor builds for Windows, as they always build
the latest and greatest.

Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'percentiles' of
Jens Axboe [Tue, 2 Jan 2018 16:05:44 +0000 (09:05 -0700)]
Merge branch 'percentiles' of

3 years agoinit: disable percentiles when latency gathering is disabled
Sitsofe Wheeler [Sun, 31 Dec 2017 12:42:59 +0000 (12:42 +0000)]
init: disable percentiles when latency gathering is disabled

When disable_clat/disable_lat is set also disable the corresponding
clat_percentiles/lat_percentiles option.

Signed-off-by: Sitsofe Wheeler <>
3 years agostat: make lat_percentiles=1 use sample count from lat_stat
Sitsofe Wheeler [Sun, 31 Dec 2017 11:41:55 +0000 (11:41 +0000)]
stat: make lat_percentiles=1 use sample count from lat_stat

The following job

./fio --name=nolatencies --ioengine=posixaio --direct=1 --rw=randread \
 --iodepth=1 --bs=4k --filename=/tmp/fio.tmp --runtime=2s --size=50m \
 --disable_lat=0 --disable_clat=1 --disable_slat=1 --lat_percentiles=1

outputs the following:

Starting 1 process

test: (groupid=0, jobs=1): err= 0: pid=7316: Sun Dec 31 11:55:46 2017
   read: IOPS=41.0k, BW=164MiB/s (172MB/s)(50.0MiB/305msec)
     lat (usec): min=8, max=132, avg=17.58, stdev= 4.36
     lat percentiles (nsec):
     |  1.00th=[    0],  5.00th=[    0], 10.00th=[    0], 20.00th=[    0],
     | 30.00th=[    0], 40.00th=[    0], 50.00th=[    0], 60.00th=[    0],
     | 70.00th=[    0], 80.00th=[    0], 90.00th=[    0], 95.00th=[    0],
     | 99.00th=[    0], 99.50th=[    0], 99.90th=[    0], 99.95th=[    0],
     | 99.99th=[    0]
  cpu          : usr=56.39%, sys=29.18%, ctx=12043, majf=0, minf=85
  IO depths    : 1=100.0%, 2=0.0%, 4=0.0%, 8=0.0%, 16=0.0%, 32=0.0%, >=64=0.0%
     submit    : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     complete  : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     issued rwt: total=12800,0,0, short=0,0,0, dropped=0,0,0
     latency   : target=0, window=0, percentile=100.00%, depth=1

Run status group 0 (all jobs):
   READ: bw=164MiB/s (172MB/s), 164MiB/s-164MiB/s (172MB/s-172MB/s), io=50.0MiB (52.4MB), run=305-305msec

which isn't showing total latency percentiles (lat percentiles) when it
should. Additionally there is a segfault when running the following job:

./fio --name=segfault --ioengine=posixaio --direct=1 --rw=randread \
 --iodepth=1 --bs=4k --filename=/tmp/fio.tmp --runtime=2s --size=50m \
 --disable_lat=1 --disable_clat=0 --disable_slat=1 --lat_percentiles=1

Fix these by making latency printing use the number of samples from the
correct latency type (clat_stat for lat_percentiles=0/clat_percentiles=1
and lat_stat for lat_percentiles=1/clat_percentiles=0).

Fixes ("fio lat reporting

Reported-by: Jeff Furlong <>
Signed-off-by: Sitsofe Wheeler <>
3 years agolog: fix bad < 0 check for unsigned
Jens Axboe [Fri, 29 Dec 2017 15:45:22 +0000 (08:45 -0700)]
log: fix bad < 0 check for unsigned

Fixes: e5f9a813 ("debug: make debug=io readable with multiple threads")
Signed-off-by: Jens Axboe <>
3 years agodebug: make debug=io readable with multiple threads
Robert Elliott [Fri, 29 Dec 2017 01:13:06 +0000 (01:13 +0000)]
debug: make debug=io readable with multiple threads

When multiple threads are active, debug=io prints are unreadable as the output
is interleaved. Multiple log_info_buf() calls are used to construct each line,
but there is no mutual exclusion covering multiple calls.

Change the dprint call tree to construct the entire line before passing it to
log_info_buf(), rather than make several calls.

Other nits:
* print the thread ID rather than the process ID
* change offset and length from decimal to hex
* separate offset, length, ddir, and file with , rather than / since the
  filename on the same line likely has / of its own
* change "fill_io_u" to "fill" to match the others
* change "io complete" to "complete" to match the others
* change "->prep()=%d" to "prep: io_u %p: ret=%d" to resemble the others
* change offset/buflen in an error print to better resemble the normal prints
* add "file=" prefix for the filename
* check the calloc() return values inside the valist_to_buf functions

fill_io_u: io_u 0x7feeac010b80: off=720896/len=65536/ddir=1

io       50692io       50692//dev/dax0.0io       50692fill_io_u: io_u 0x7fee98010b80: off=196608/len=65536/ddir=1io       50692io       50692
io       50692->prep(0x7fef10010b80)=0
//dev/dax0.0//dev/dax1.0io       50692io       50692io       50692io       50692->prep(0x7feeec010b80)=0
io       50692prep: io_u 0x7feec4010b80: off=1966080/len=65536/ddir=1io       50692io complete: io_u 0x7feedc010b80: off=393216/len=65536/ddir=1io
50692io       50692prep: io_u 0x7feef4010b80: off=720896/len=65536/ddir=1io       50692io       50692
io       50692io       50692prep: io_u 0x7feef0010b80: off=851968/len=65536/ddir=1//dev/dax0.0io       50692//dev/dax0.0

io       50692io       50692//dev/dax0.0//dev/dax0.0io       50692

io       71400 queue: io_u 0x7fd0f0010b80: off=0x2f0000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71395 fill: io_u 0x7fd0fc010b80: off=0x80000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71395 prep: io_u 0x7fd0fc010b80: off=0x80000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71395 prep: io_u 0x7fd0fc010b80: ret=0
io       71395 queue: io_u 0x7fd0fc010b80: off=0x80000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71430 complete: io_u 0x7fd05c010b80: off=0x180000,len=0x10000,ddir=1,file=/dev/dax1.0
io       71400 complete: io_u 0x7fd0f0010b80: off=0x2f0000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71430 fill: io_u 0x7fd05c010b80: off=0x190000,len=0x10000,ddir=1,file=/dev/dax1.0
io       71400 fill: io_u 0x7fd0f0010b80: off=0x300000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71400 prep: io_u 0x7fd0f0010b80: off=0x300000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71400 prep: io_u 0x7fd0f0010b80: ret=0
io       71400 queue: io_u 0x7fd0f0010b80: off=0x300000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71430 prep: io_u 0x7fd05c010b80: off=0x190000,len=0x10000,ddir=1,file=/dev/dax1.0
io       71395 complete: io_u 0x7fd0fc010b80: off=0x80000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71430 prep: io_u 0x7fd05c010b80: ret=0
io       71430 queue: io_u 0x7fd05c010b80: off=0x190000,len=0x10000,ddir=1,file=/dev/dax1.0
io       71421 complete: io_u 0x7fd090010b80: off=0x320000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71419 complete: io_u 0x7fd098010b80: off=0x320000,len=0x10000,ddir=1,file=/dev/dax0.0
io       71404 complete: io_u 0x7fd0d0010b80: off=0x0,len=0x10000,ddir=1,file=/dev/dax0.0

Signed-off-by: Jens Axboe <>
3 years agolib/memcpy: fix warning on FreeBSD
Tomohiro Kusumi [Fri, 29 Dec 2017 13:38:23 +0000 (15:38 +0200)]
lib/memcpy: fix warning on FreeBSD

Include struct timeval via sys/time.h.

In file included from lib/memcpy.c:7:
lib/../fio_time.h:14:45: warning: declaration of 'struct timeval' will not be visible outside of this function [-Wvisibility]
extern uint64_t mtime_since_tv(const struct timeval *, const struct timeval *);

Signed-off-by: Tomohiro Kusumi <>
Signed-off-by: Jens Axboe <>
3 years agoMerge branch 'master' of
Jens Axboe [Thu, 28 Dec 2017 15:35:11 +0000 (08:35 -0700)]
Merge branch 'master' of

3 years agomutex: down_timeout: check against the base time
Yasushi SHOJI [Thu, 28 Dec 2017 01:53:47 +0000 (10:53 +0900)]
mutex: down_timeout: check against the base time

When the commit 8b6a404cdd2c40715885e562416c3db039912773 changed
timeval to timespec, it accsidentally picked timeout time `t' instead
of the current `base' time for checking timewait() really timed out.

t is base + msecs as follows:

      base = gettimeofday()
      t = base + msecs
      while (!done)
         if (timedout)
    if (mutex_timed_out(t, msecs))
       done = true

Thus, after a good timedout, it burns the cpu for the given msecs
seconds doing tight loop in the while shown above.  This causes 30 to
40 % CPU load on my i5-3360M 2.80GHz while fio server is running.

3 years agoMerge branch 'eta_display' of
Jens Axboe [Wed, 27 Dec 2017 21:05:46 +0000 (14:05 -0700)]
Merge branch 'eta_display' of

3 years agoMerge branch 'barak/mmap_verify_only' of
Jens Axboe [Wed, 27 Dec 2017 21:05:27 +0000 (14:05 -0700)]
Merge branch 'barak/mmap_verify_only' of

3 years agoeta: show complete status line with max job states
Sitsofe Wheeler [Sat, 23 Dec 2017 09:55:02 +0000 (09:55 +0000)]
eta: show complete status line with max job states

Make the ETA output buffer big enough to cope with the maximum possible
jobs when their states can't be compressed (i.e. when every job's state
is different to the state of the job that came before it).

Allows the following script to display the complete status line:

rw[0]='read'; rw[1]='write'; \
for i in {1..4096}; do \
  echo -e "[job$i]\nrw=${rw[$((i % 2))]}\n" \
          "runtime=15"; \
done | \
./fio --eta=always --group_reporting=1 --ioengine=null --size=1g \
 --time_based --bs=1k --thread --rate_iops=1 -

Fixes: ("Large number of threads
result in Seg faults")

Signed-off-by: Sitsofe Wheeler <>
3 years agoeta: fix previous line length calculation
Sitsofe Wheeler [Sat, 23 Dec 2017 08:43:54 +0000 (08:43 +0000)]
eta: fix previous line length calculation

We work out how many stale characters we haven't yet overprinted from
the previous run by comparing the number of characters that snprintf
added to the buffer in both cases. Unfortunately this doesn't account
for earlier characters that may (or may not) have already been in the
output buffer before snprintf was called...

Change the code to just use the current and previous output buffer
position to calculate the remainder of the line to be cleared.

When fio is compiled under clang with
./configure --extra-cflags="-fsanitize=address -fno-builtin" \

a buffer overrun is demonstrated by the following script

rw[0]='read'; rw[1]='write'; \
for i in {1..4096}; do \
  echo -e "[job$i]\nrw=${rw[$((i % 2))]}\n" \
  "ramp_time=$(((4096 - i) / 2049 * 10))\n" \
          "runtime=15"; \
done | \
./fio --group_reporting --ioengine=null --size=1g --time_based \
 --bs=512 --thread --rate_iops=2 -

Signed-off-by: Sitsofe Wheeler <>
3 years agoeta: skip clearing of remainder of line when starting a new line
Sitsofe Wheeler [Sat, 23 Dec 2017 08:39:11 +0000 (08:39 +0000)]
eta: skip clearing of remainder of line when starting a new line

Signed-off-by: Sitsofe Wheeler <>
3 years agoeta: fix run_str_condensed overflow with maximum jobs
Sitsofe Wheeler [Sat, 23 Dec 2017 08:01:01 +0000 (08:01 +0000)]
eta: fix run_str_condensed overflow with maximum jobs

When the maximum number of jobs all have a different state to their
previous job, run_str didn't have space for the terminating null. Fix
this by making run_str one character larger.

Signed-off-by: Sitsofe Wheeler <>
3 years agoeta: adjust truncation case
Sitsofe Wheeler [Fri, 22 Dec 2017 11:10:35 +0000 (11:10 +0000)]
eta: adjust truncation case

c076de85fa4901683f6ce23ecdc071c17801ffad ("eta: fix buffer overflow in
ETA output") made adjustments to cope with snprintf truncation. However
since we're replacing the null added by snprintf, left only needs to be
one smaller than the output buffer. Make this change and add a comment
about what we're doing.

Signed-off-by: Sitsofe Wheeler <>
3 years agoMerge branch 'eta_overflow' of
Jens Axboe [Thu, 21 Dec 2017 15:22:39 +0000 (08:22 -0700)]
Merge branch 'eta_overflow' of

3 years agoeta: fix buffer overflow in ETA output
Sitsofe Wheeler [Thu, 21 Dec 2017 12:23:36 +0000 (12:23 +0000)]
eta: fix buffer overflow in ETA output

When 900 or more jobs all have different states it is possible for the
space required to display the ETA line to be larger than the ETA output
buffer. fio initially truncates what it puts in the output buffer but
incorrectly advances where it thinks the end of the buffer is - when
truncating snprintf returns the number of characters that _would_ have
been converted if there had been enough space...

This patch fixes the post truncation calculation and adjusts the "left"
variable so there is room for the carriage return and a null terminator.

The following script reproduces the problem when fio has been compiled
with -fstack-protector or -fsanitize=address:

rw[0]='read'; rw[1]='write'; \
for i in {1..1000}; do \
 echo -e "[job$i]\nrw=${rw[$((i % 2))]}\n" \
         "ramp_time=$(((1000 - i) / 350 * 5))\n" \
         "runtime=$((60 - (i % 3) * 2))"; \
done | \
./fio --eta-newline=1s --group_reporting --ioengine=null --size=1g \
 --time_based --bs=512 --thread --rate_iops=123456 -

Fixes: ("Large number of threads
result in Seg faults")
Signed-off-by: Sitsofe Wheeler <>
3 years agoFio 3.3 fio-3.3
Jens Axboe [Tue, 19 Dec 2017 20:16:36 +0000 (13:16 -0700)]
Fio 3.3

Signed-off-by: Jens Axboe <>
3 years agobackend: tweaks to missed rate thinktime
Jens Axboe [Tue, 19 Dec 2017 17:17:41 +0000 (10:17 -0700)]
backend: tweaks to missed rate thinktime

shimrot says:

"May I suggest a tweak to the backend.c::handle_thinktime calculation of
"over"? That is, adjust based on the rate per byte similar to what
you've done. That is, if thinktime is greater than one op worth of
bytes, then subtract one op worth of bytes. But, if think time is
greater than one ops worth, instead add the missing the number of bytes
a full op would have exceeded the think time."


Signed-off-by: Jens Axboe <>
3 years agoioengines: clear out ->td_ops_dlhandle if we close it
Jens Axboe [Fri, 15 Dec 2017 20:35:56 +0000 (13:35 -0700)]
ioengines: clear out ->td_ops_dlhandle if we close it

Signed-off-by: Jens Axboe <>
3 years agoinit: fix missing dlhandle reference put
Jens Axboe [Fri, 15 Dec 2017 20:34:22 +0000 (13:34 -0700)]
init: fix missing dlhandle reference put

If we keep using the same engine, put the dlhandle reference that
we now hold twice.

Signed-off-by: Jens Axboe <>
3 years agoparse: don't check for < 0 on an unsigned type
Jens Axboe [Fri, 15 Dec 2017 16:13:28 +0000 (09:13 -0700)]
parse: don't check for < 0 on an unsigned type

We do strict errors on some builds, this breaks them. Remove the
<= 0 check, just check for a value that's too large.

Fixes: c26438ad ("parse: dump option type when using --debug=parse")
Signed-off-by: Jens Axboe <>
3 years agoioengines: improve "is this the same IO engine" check
Jens Axboe [Fri, 15 Dec 2017 16:08:26 +0000 (09:08 -0700)]
ioengines: improve "is this the same IO engine" check

We can't just compare the name, that assumes that name and filename
match for an IO engine. While that's generally the case for the
engines that fio ships with, it's not a requirement, and it's
definitely not true for external engines.

Fixup the check by re-loading the engine and checking the OPS
instead. That should be bullet proof.

Fixes: 800334d ("Correctly detect whether ioengine_load can exit early")
Signed-off-by: Jens Axboe <>
3 years agoparse: dump option type when using --debug=parse
Jens Axboe [Fri, 15 Dec 2017 15:41:03 +0000 (08:41 -0700)]
parse: dump option type when using --debug=parse

Currently we do things like:

parse    8731  __handle_option=dummy, type=10, ptr=1

for the debug parsing, and then you have to look up what that
option type is. Add names to them so we get the below instead:

parse    9170  __handle_option=dummy, type=OPT_STR_SET, ptr=1

Signed-off-by: Jens Axboe <>
3 years agoAllow configurable ETA intervals
Jens Axboe [Thu, 14 Dec 2017 18:51:41 +0000 (11:51 -0700)]
Allow configurable ETA intervals

By default, fio prints ETA output every second. For some client/server
setups, it's desirable to allow a much longer interval, to avoid
spending too much time getting and printing ETA time.

Takes a normal time input, allowing usec/msec/sec etc postfixes.

Signed-off-by: Jens Axboe <>
3 years agoclient: respect --eta=never for networked connections
Jens Axboe [Thu, 14 Dec 2017 17:58:03 +0000 (10:58 -0700)]
client: respect --eta=never for networked connections

Signed-off-by: Jens Axboe <>
3 years agoserver: convert more memset to on-stack initialization
Jens Axboe [Thu, 14 Dec 2017 16:47:35 +0000 (09:47 -0700)]
server: convert more memset to on-stack initialization

Signed-off-by: Jens Axboe <>
3 years agoserver: cleanup iolog pdu prep
Jens Axboe [Thu, 14 Dec 2017 16:25:10 +0000 (09:25 -0700)]
server: cleanup iolog pdu prep

We don't have to memset() the whole thing, we just need to ensure
that the members we don't assign are cleared. Remove the memset
and initialize the pdu on the stack.

Signed-off-by: Jens Axboe <>
3 years agoFix Windows local time
Jeff Furlong [Thu, 14 Dec 2017 01:19:09 +0000 (01:19 +0000)]
Fix Windows local time

The fio summary output shows when the job completed, but references
system time on Windows.  On Linux, local time (such as with time zone
specifics) is referenced.  This patch converts Windows system time to
local time.

Starting 1 thread
Jobs: 1 (f=0): [f(1)][100.0%][r=69.6MiB/s,w=0KiB/s][r=17.8k,w=0 IOPS][eta 00m:00s]
test: (groupid=0, jobs=1): err= 0: pid=1360: Wed Dec 13 23:21:16 2017 <--System Time

Starting 1 thread
Jobs: 1 (f=0): [f(1)][100.0%][r=73.2MiB/s,w=0KiB/s][r=18.7k,w=0 IOPS][eta 00m:00s]
test: (groupid=0, jobs=1): err= 0: pid=8088: Wed Dec 13 15:22:07 2017 <--Local Time

Signed-off-by: Jens Axboe <>
3 years ago.gitignore: ignore tags files and additional output binaries
Robert Elliott [Mon, 4 Dec 2017 20:53:16 +0000 (14:53 -0600)]
.gitignore: ignore tags files and additional output binaries

Ignore tags files used by emacs and ctags/vim.

Ignore the additional output binaries alongside fio, such as gfio
and all the programs in t/.

Signed-off-by: Jens Axboe <>
3 years agoserver: initialize first iolog header properly
Jens Axboe [Wed, 13 Dec 2017 15:44:34 +0000 (08:44 -0700)]
server: initialize first iolog header properly

We're sending potential stack garbage over the wire, init the
whole pdu to zeroes before filling it in. This shuts up valgrind
about using uninitialized memory.

Signed-off-by: Jens Axboe <>
3 years agofio: kill td->nr_normal_files
Jens Axboe [Fri, 8 Dec 2017 19:50:28 +0000 (12:50 -0700)]
fio: kill td->nr_normal_files

It's unused.

Signed-off-by: Jens Axboe <>
3 years agoRemove old emails
Jens Axboe [Fri, 8 Dec 2017 19:38:27 +0000 (12:38 -0700)]
Remove old emails

Just use my private email, it's the one I prefer using anyway.
Remove the old fusionio/oracle emails, they just cause
confusion and people tend to just write to all of them at once
when reporting an issue.

Signed-off-by: Jens Axboe <>
3 years agorate: fix bad math
Jens Axboe [Fri, 8 Dec 2017 18:59:38 +0000 (11:59 -0700)]
rate: fix bad math

To figure out how much data we missed when doing a thinktime sleep,
we're currently dividing by the time slept. This is wrong, it should
be multiplied by the time slept and divided by 1000000 to go from
usec to a second base.

Additionally, don't ever subtract more than a block of data, and
adjust down depending on sleep.

Fixes: 1aa39b0c ("rate: ensure IO issue restarts right after sleep")
Signed-off-by: Jens Axboe <>
3 years agorate: ensure IO issue restarts right after sleep
Jens Axboe [Thu, 7 Dec 2017 16:06:04 +0000 (09:06 -0700)]
rate: ensure IO issue restarts right after sleep

If we have rate_ignore_thinktime set, we account the number of bytes
we would have done in the period of the sleep. However, we need
to subtract one block, otherwise we end up doing a rated sleep
right after the thinktime sleep.

Fixes: 1a9bf814 ("Add option to ignore thinktime for rated IO")
Signed-off-by: Jens Axboe <>
3 years agoio_u: rate cleanup and spelling error
Jens Axboe [Wed, 6 Dec 2017 19:30:20 +0000 (12:30 -0700)]
io_u: rate cleanup and spelling error

Fixes: 50a8ce86 ("Implement new Rate Control")
Signed-off-by: Jens Axboe <>
3 years agoAdd option to ignore thinktime for rated IO
Jens Axboe [Wed, 6 Dec 2017 19:27:07 +0000 (12:27 -0700)]
Add option to ignore thinktime for rated IO

By default, fio will ignore thinktime when calculating the next
time to issue and IO, if rated IO is specified. This leads to
fio entering a catch-up type of mode after doing the specified
sleep. For some workloads, that may not be useful. If someone
asks for a specific amount of IOPS and sets a thinktime, they
may want to exclude the sleep time.

Signed-off-by: Jens Axboe <>
3 years agofix verify_only when using ioengine=mmap
Barak Pinhas [Tue, 5 Dec 2017 12:23:22 +0000 (14:23 +0200)]
fix verify_only when using ioengine=mmap

verify_only was opening the file with O_RDONLY and the mmap
engine was trying to mmap it with PROT_READ | PROTWRITE hence
getting -EACCESS and failing the job