ioengine=sync issue on Windows

ioengine=sync issue on Windows

[Sébastien Bouchex Bellomié]

Hi,

The "sync" engine does not seem to work correctly on Windows :

My profile is :

[global]
thread
ioengine=sync

[random_read]
rw=randread
size=1024m
directory=read
bs=32k

fio 2.1.6.1 output contains " err= 5 (file:backend.c:768, func=full resid, error=Input/output error)"

[...]
C:\Program Files\fio>fio.exe H:\fio\fio.win
random_read: (g=0): rw=randread, bs=32K-32K/32K-32K/32K-32K, ioengine=sync, iodepth=1
fio-2.1.6-1-gd004a
Starting 1 thread
fio: pid=5144, err=5/file:backend.c:768, func=full resid, error=Input/output error

random_read: (groupid=0, jobs=1): err= 5 (file:backend.c:768, func=full resid, error=Input/output error): pid=5144: Wed Mar 12 16:36:45 2014
  cpu          : usr=0.00%, sys=0.00%, ctx=0, majf=0, minf=0
  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=50.0%, 4=50.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     issued    : total=r=2/w=0/d=0, short=r=1/w=0/d=0
     latency   : target=0, window=0, percentile=100.00%, depth=1

Run status group 0 (all jobs):
[...]

Seb

S�bastien BOUCHEX BELLOMI�
Infovista� Server Technical Lead, Service Assurance R&D

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 03/12/2014 09:41 AM, S�bastien Bouchex Bellomi� wrote:
> Hi,
>
> The "sync" engine does not seem to work correctly on Windows :
>
> My profile is :
>
> [global]
> thread
> ioengine=sync
>
> [random_read]
> rw=randread
> size=1024m
> directory=read
> bs=32k
>
> fio 2.1.6.1 output contains " err= 5 (file:backend.c:768, func=full resid, error=Input/output error)"
>
> [...]
> C:\Program Files\fio>fio.exe H:\fio\fio.win
> random_read: (g=0): rw=randread, bs=32K-32K/32K-32K/32K-32K, ioengine=sync, iodepth=1
> fio-2.1.6-1-gd004a
> Starting 1 thread
> fio: pid=5144, err=5/file:backend.c:768, func=full resid, error=Input/output error
>
> random_read: (groupid=0, jobs=1): err= 5 (file:backend.c:768, func=full resid, error=Input/output error): pid=5144: Wed Mar 12 16:36:45 2014
>    cpu          : usr=0.00%, sys=0.00%, ctx=0, majf=0, minf=0
>    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=50.0%, 4=50.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
>       issued    : total=r=2/w=0/d=0, short=r=1/w=0/d=0
>       latency   : target=0, window=0, percentile=100.00%, depth=1
>
> Run status group 0 (all jobs):
> [...]

Bruce, do you know why that is?

-- 
Jens Axboe

RE: ioengine=sync issue on Windows

[Bruce Cran]

It looks like it's trying to access %TEMP%\tmp for some reason.

-- 
Bruce

-----Original Message-----
From: Jens Axboe [mailto:axboe@kernel.dk] 
Sent: Wednesday, March 12, 2014 10:34 AM
To: Sébastien Bouchex Bellomié; fio@vger.kernel.org; Bruce Cran
Subject: Re: ioengine=sync issue on Windows

On 03/12/2014 09:41 AM, Sébastien Bouchex Bellomié wrote:
> Hi,
>
> The "sync" engine does not seem to work correctly on Windows :
>
> My profile is :
>
> [global]
> thread
> ioengine=sync
>
> [random_read]
> rw=randread
> size=1024m
> directory=read
> bs=32k
>
> fio 2.1.6.1 output contains " err= 5 (file:backend.c:768, func=full resid, error=Input/output error)"
>
> [...]
> C:\Program Files\fio>fio.exe H:\fio\fio.win
> random_read: (g=0): rw=randread, bs=32K-32K/32K-32K/32K-32K, 
> ioengine=sync, iodepth=1 fio-2.1.6-1-gd004a Starting 1 thread
> fio: pid=5144, err=5/file:backend.c:768, func=full resid, 
> error=Input/output error
>
> random_read: (groupid=0, jobs=1): err= 5 (file:backend.c:768, func=full resid, error=Input/output error): pid=5144: Wed Mar 12 16:36:45 2014
>    cpu          : usr=0.00%, sys=0.00%, ctx=0, majf=0, minf=0
>    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=50.0%, 4=50.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
>       issued    : total=r=2/w=0/d=0, short=r=1/w=0/d=0
>       latency   : target=0, window=0, percentile=100.00%, depth=1
>
> Run status group 0 (all jobs):
> [...]

Bruce, do you know why that is?

--
Jens Axboe


This e-mail (and any attachments) is confidential and may be privileged.  Any unauthorized use, copying, disclosure or dissemination of this communication is prohibited.  If you are not the intended recipient,  please notify the sender immediately and delete all copies of the message and its attachments.

RE: ioengine=sync issue on Windows

[Bruce Cran]

It turns out this is unrelated, but in stat.c the following change should probably be made:

-#define FIO_STATUS_FILE                "/tmp/fio-dump-status"
+#define FIO_STATUS_FILE                "fio-dump-status"

-- 
Bruce

-----Original Message-----
From: Bruce Cran 
Sent: Wednesday, March 12, 2014 11:49 AM
To: 'Jens Axboe'; S�bastien Bouchex Bellomi�; fio@vger.kernel.org
Subject: RE: ioengine=sync issue on Windows

It looks like it's trying to access %TEMP%\tmp for some reason.

--
Bruce

-----Original Message-----
From: Jens Axboe [mailto:axboe@kernel.dk]
Sent: Wednesday, March 12, 2014 10:34 AM
To: S�bastien Bouchex Bellomi�; fio@vger.kernel.org; Bruce Cran
Subject: Re: ioengine=sync issue on Windows

On 03/12/2014 09:41 AM, S�bastien Bouchex Bellomi� wrote:
> Hi,
>
> The "sync" engine does not seem to work correctly on Windows :
>
> My profile is :
>
> [global]
> thread
> ioengine=sync
>
> [random_read]
> rw=randread
> size=1024m
> directory=read
> bs=32k
>
> fio 2.1.6.1 output contains " err= 5 (file:backend.c:768, func=full resid, error=Input/output error)"
>
> [...]
> C:\Program Files\fio>fio.exe H:\fio\fio.win
> random_read: (g=0): rw=randread, bs=32K-32K/32K-32K/32K-32K, 
> ioengine=sync, iodepth=1 fio-2.1.6-1-gd004a Starting 1 thread
> fio: pid=5144, err=5/file:backend.c:768, func=full resid, 
> error=Input/output error
>
> random_read: (groupid=0, jobs=1): err= 5 (file:backend.c:768, func=full resid, error=Input/output error): pid=5144: Wed Mar 12 16:36:45 2014
>    cpu          : usr=0.00%, sys=0.00%, ctx=0, majf=0, minf=0
>    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=50.0%, 4=50.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
>       issued    : total=r=2/w=0/d=0, short=r=1/w=0/d=0
>       latency   : target=0, window=0, percentile=100.00%, depth=1
>
> Run status group 0 (all jobs):
> [...]

Bruce, do you know why that is?

--
Jens Axboe


This e-mail (and any attachments) is confidential and may be privileged.  Any unauthorized use, copying, disclosure or dissemination of this communication is prohibited.  If you are not the intended recipient,  please notify the sender immediately and delete all copies of the message and its attachments.

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 03/12/2014 12:02 PM, Bruce Cran wrote:
> It turns out this is unrelated, but in stat.c the following change should probably be made:
>
> -#define FIO_STATUS_FILE                "/tmp/fio-dump-status"
> +#define FIO_STATUS_FILE                "fio-dump-status"
>

How about we make that use os_get_tmpdir() instead?

-- 
Jens Axboe

RE: ioengine=sync issue on Windows

[Bruce Cran]

There's already code to use the environment variables TMP or TEMP, except it appends "tmp" to whatever it gets.

-- 
Bruce

-----Original Message-----
From: Jens Axboe [mailto:axboe@kernel.dk] 
Sent: Wednesday, March 12, 2014 1:11 PM
To: Bruce Cran; Sébastien Bouchex Bellomié; fio@vger.kernel.org
Subject: Re: ioengine=sync issue on Windows

On 03/12/2014 12:02 PM, Bruce Cran wrote:
> It turns out this is unrelated, but in stat.c the following change should probably be made:
>
> -#define FIO_STATUS_FILE                "/tmp/fio-dump-status"
> +#define FIO_STATUS_FILE                "fio-dump-status"
>

How about we make that use os_get_tmpdir() instead?

-- 
Jens Axboe


This e-mail (and any attachments) is confidential and may be privileged.  Any unauthorized use, copying, disclosure or dissemination of this communication is prohibited.  If you are not the intended recipient,  please notify the sender immediately and delete all copies of the message and its attachments.

Re: ioengine=sync issue on Windows

[Jens Axboe]

[-- Attachment #1: Type: text/plain, Size: 234 bytes --]

On 03/12/2014 01:24 PM, Bruce Cran wrote:
> There's already code to use the environment variables TMP or TEMP, except it appends "tmp" to whatever it gets.
>

Yes, my point was to use it. Something like the attached.

-- 
Jens Axboe


[-- Attachment #2: fio-tmp.patch --]
[-- Type: text/x-patch, Size: 1625 bytes --]

diff --git a/os/os-windows.h b/os/os-windows.h
index 49f96068ff32..499e098f8b82 100644
--- a/os/os-windows.h
+++ b/os/os-windows.h
@@ -167,6 +167,8 @@ static inline void os_get_tmpdir(char *path, int len)
 	GetTempPath(len, path);
 }
 
+#define FIO_HAVE_OS_GET_TMPDIR
+
 static inline int gettid(void)
 {
 	return GetCurrentThreadId();
diff --git a/stat.c b/stat.c
index e43db8f9d075..df3bd93d32e5 100644
--- a/stat.c
+++ b/stat.c
@@ -13,6 +13,7 @@
 #include "json.h"
 #include "lib/getrusage.h"
 #include "idletime.h"
+#include "os/os.h"
 
 static struct fio_mutex *stat_mutex;
 
@@ -1480,7 +1481,7 @@ void show_running_run_stats(void)
 static int status_interval_init;
 static struct timeval status_time;
 
-#define FIO_STATUS_FILE		"/tmp/fio-dump-status"
+#define FIO_STATUS_FILE		"fio-dump-status"
 
 static int check_status_file(void)
 {
@@ -1491,10 +1492,22 @@ static int check_status_file(void)
 	temp_dir = getenv("TMPDIR");
 	if (temp_dir == NULL)
 		temp_dir = getenv("TEMP");
-	if (temp_dir == NULL)
+	if (temp_dir == NULL) {
+#ifdef FIO_HAVE_OS_GET_TMPDIR
+		temp_dir = NULL;
+		os_get_tmpdir(status_file_path, sizeof(status_file_path));
+#else
 		temp_dir = "/tmp";
+#endif
+	}
 
-	snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
+	if (temp_dir)
+		snprintf(fio_status_file_path, sizeof(fio_status_file_path), "%s/%s", temp_dir, FIO_STATUS_FILE);
+	else {
+		size_t len = strlen(fio_status_file_path);
+
+		snprintf(fio_status_file_path + len, sizeof(fio_status_file_path) - len, "%s", FIO_STATUS_FILE);
+	}
 
 	if (stat(fio_status_file_path, &sb))
 		return 0;

RE: ioengine=sync issue on Windows

[Bruce Cran]

Oh, I see - that works.   Though the existing code also does with the removal of 'tmp' from FIO_STATUS_FILE - Windows defines both TMP and TEMP environment variables.

-- 
Bruce

-----Original Message-----
From: Jens Axboe [mailto:axboe@kernel.dk] 
Sent: Wednesday, March 12, 2014 1:31 PM
To: Bruce Cran; Sébastien Bouchex Bellomié; fio@vger.kernel.org
Subject: Re: ioengine=sync issue on Windows

On 03/12/2014 01:24 PM, Bruce Cran wrote:
> There's already code to use the environment variables TMP or TEMP, except it appends "tmp" to whatever it gets.
>

Yes, my point was to use it. Something like the attached.

-- 
Jens Axboe


This e-mail (and any attachments) is confidential and may be privileged.  Any unauthorized use, copying, disclosure or dissemination of this communication is prohibited.  If you are not the intended recipient,  please notify the sender immediately and delete all copies of the message and its attachments.

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 03/12/2014 01:36 PM, Bruce Cran wrote:
> Oh, I see - that works.   Though the existing code also does with the removal of
 > 'tmp' from FIO_STATUS_FILE - Windows defines both TMP and TEMP 
environment variables.

So the getenv() works fine on Windows? If so, I can clean it up some more.

-- 
Jens Axboe

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 03/12/2014 01:38 PM, Jens Axboe wrote:
> On 03/12/2014 01:36 PM, Bruce Cran wrote:
>> Oh, I see - that works.   Though the existing code also does with the
>> removal of
>  > 'tmp' from FIO_STATUS_FILE - Windows defines both TMP and TEMP
> environment variables.
>
> So the getenv() works fine on Windows? If so, I can clean it up some more.

Queued this up:

http://git.kernel.dk/?p=fio.git;a=commit;h=dac45f23c3c7d146f7df82047ae7f25abf231ed3

-- 
Jens Axboe

RE: ioengine=sync issue on Windows

[Bruce Cran]

Perfect - thanks!

-- 
Bruce

-----Original Message-----
From: Jens Axboe [mailto:axboe@kernel.dk] 
Sent: Wednesday, March 12, 2014 1:41 PM
To: Bruce Cran; Sébastien Bouchex Bellomié; fio@vger.kernel.org
Subject: Re: ioengine=sync issue on Windows

On 03/12/2014 01:38 PM, Jens Axboe wrote:
> On 03/12/2014 01:36 PM, Bruce Cran wrote:
>> Oh, I see - that works.   Though the existing code also does with the
>> removal of
>  > 'tmp' from FIO_STATUS_FILE - Windows defines both TMP and TEMP 
> environment variables.
>
> So the getenv() works fine on Windows? If so, I can clean it up some more.

Queued this up:

http://git.kernel.dk/?p=fio.git;a=commit;h=dac45f23c3c7d146f7df82047ae7f25abf231ed3

--
Jens Axboe


This e-mail (and any attachments) is confidential and may be privileged.  Any unauthorized use, copying, disclosure or dissemination of this communication is prohibited.  If you are not the intended recipient,  please notify the sender immediately and delete all copies of the message and its attachments.

Re: ioengine=sync issue on Windows

[Bruce Cran]

On 3/12/2014 10:33 AM, Jens Axboe wrote:
> On 03/12/2014 09:41 AM, Sébastien Bouchex Bellomié wrote:
>>
>> fio: pid=5144, err=5/file:backend.c:768, func=full resid, 
>> error=Input/output error
>>
>> random_read: (groupid=0, jobs=1): err= 5 (file:backend.c:768, 
>> func=full resid, error=Input/output error): pid=5144: Wed Mar 12 
>> 16:36:45 2014
> Bruce, do you know why that is?
>

It appears to be because after seeking, there's 92 bytes before a space 
in the file (i.e. it's a sparse file), and instead of filling the buffer 
with 0's read() is for some reason returning 0.

-- 
Bruce

Re: ioengine=sync issue on Windows

[Bruce Cran]

On 4/1/2014 9:11 PM, Bruce Cran wrote:
>
> On 3/12/2014 10:33 AM, Jens Axboe wrote:
>> On 03/12/2014 09:41 AM, Sébastien Bouchex Bellomié wrote:
>>>
>>> fio: pid=5144, err=5/file:backend.c:768, func=full resid, 
>>> error=Input/output error
>>>
>>> random_read: (groupid=0, jobs=1): err= 5 (file:backend.c:768, 
>>> func=full resid, error=Input/output error): pid=5144: Wed Mar 12 
>>> 16:36:45 2014
>> Bruce, do you know why that is?
>>
>
> It appears to be because after seeking, there's 92 bytes before a 
> space in the file (i.e. it's a sparse file), and instead of filling 
> the buffer with 0's read() is for some reason returning 0.
>

I was wrong.  The problem has something to do with fil_io_buffer - and 
in particular __fill_random_buf. The io error disappears if I change the 
line:

*ptr = seed;
to
*ptr = 1;

Also, fio-2.0.15 works, while fio-2.1 doesn't.

-- 
Bruce

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 2014-04-03 16:42, Bruce Cran wrote:
>
> On 4/1/2014 9:11 PM, Bruce Cran wrote:
>>
>> On 3/12/2014 10:33 AM, Jens Axboe wrote:
>>> On 03/12/2014 09:41 AM, S�bastien Bouchex Bellomi� wrote:
>>>>
>>>> fio: pid=5144, err=5/file:backend.c:768, func=full resid,
>>>> error=Input/output error
>>>>
>>>> random_read: (groupid=0, jobs=1): err= 5 (file:backend.c:768,
>>>> func=full resid, error=Input/output error): pid=5144: Wed Mar 12
>>>> 16:36:45 2014
>>> Bruce, do you know why that is?
>>>
>>
>> It appears to be because after seeking, there's 92 bytes before a
>> space in the file (i.e. it's a sparse file), and instead of filling
>> the buffer with 0's read() is for some reason returning 0.
>>
>
> I was wrong.  The problem has something to do with fil_io_buffer - and
> in particular __fill_random_buf. The io error disappears if I change the
> line:
>
> *ptr = seed;
> to
> *ptr = 1;
>
> Also, fio-2.0.15 works, while fio-2.1 doesn't.

That makes very little sense... How could the contents of the buffer 
change the behavior?

Could you run a git bisect? Should be quick'n easy, if it's nicely 
reproducible.

-- 
Jens Axboe

Re: ioengine=sync issue on Windows

[Bruce Cran]

On 4/3/2014 7:57 PM, Jens Axboe wrote:
> On 2014-04-03 16:42, Bruce Cran wrote:
>>
>> Also, fio-2.0.15 works, while fio-2.1 doesn't.
>
> That makes very little sense... How could the contents of the buffer 
> change the behavior?
>
> Could you run a git bisect? Should be quick'n easy, if it's nicely 
> reproducible.
>

http://git.kernel.dk/?p=fio.git;a=commit;h=cc86c395fd9dd2002ec1edc0967b7c9453debdfb 
is the changeset that breaks things - and I know it makes no sense.

-- 
Bruce

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 2014-04-03 20:14, Bruce Cran wrote:
>
> On 4/3/2014 7:57 PM, Jens Axboe wrote:
>> On 2014-04-03 16:42, Bruce Cran wrote:
>>>
>>> Also, fio-2.0.15 works, while fio-2.1 doesn't.
>>
>> That makes very little sense... How could the contents of the buffer
>> change the behavior?
>>
>> Could you run a git bisect? Should be quick'n easy, if it's nicely
>> reproducible.
>>
>
> http://git.kernel.dk/?p=fio.git;a=commit;h=cc86c395fd9dd2002ec1edc0967b7c9453debdfb
> is the changeset that breaks things - and I know it makes no sense.

That is pretty nutty. If you revert just the change to filesetup.c, does 
that make it work again?

-- 
Jens Axboe

Re: ioengine=sync issue on Windows

[Bruce Cran]

On 4/3/2014 8:17 PM, Jens Axboe wrote:
> On 2014-04-03 20:14, Bruce Cran wrote:
>>
>> http://git.kernel.dk/?p=fio.git;a=commit;h=cc86c395fd9dd2002ec1edc0967b7c9453debdfb 
>>
>> is the changeset that breaks things - and I know it makes no sense.
>
> That is pretty nutty. If you revert just the change to filesetup.c, 
> does that make it work again?
>

I've just re-tested it and yes, reverting the change to filesetup.c does 
make it work again.

-- 
Bruce

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 2014-04-03 20:26, Bruce Cran wrote:
> On 4/3/2014 8:17 PM, Jens Axboe wrote:
>> On 2014-04-03 20:14, Bruce Cran wrote:
>>>
>>> http://git.kernel.dk/?p=fio.git;a=commit;h=cc86c395fd9dd2002ec1edc0967b7c9453debdfb
>>>
>>> is the changeset that breaks things - and I know it makes no sense.
>>
>> That is pretty nutty. If you revert just the change to filesetup.c,
>> does that make it work again?
>>
>
> I've just re-tested it and yes, reverting the change to filesetup.c does
> make it work again.

So next question... If you replace that fill_io_buffer() call with a 
memset(b, 0, bs); does it then work?

-- 
Jens Axboe

Re: ioengine=sync issue on Windows

[Bruce Cran]

On 4/3/2014 8:35 PM, Jens Axboe wrote:
> So next question... If you replace that fill_io_buffer() call with a 
> memset(b, 0, bs); does it then work?
>

The io error goes away, but now it occasionally crashes:

(gdb) bt

#0  0x64942e9e in pthread_mutex_unlock () from /cygdrive/c/Users/bcran/workspace/fio/libwinpthread-1.dll
#1  0x0043997c in run_threads () at backend.c:1791
#2  0x00439d9f in fio_backend () at backend.c:1904
#3  0x0044c274 in main (argc=2, argv=0x821728, envp=0x821988) at fio.c:50

(gdb) info threads
   Id   Target Id         Frame
   3    Thread 7256.0x4618 fio_pin_memory (td=td@entry=0x2820000) at memory.c:27
   2    Thread 7256.0xc8  0x7787bf3c in ?? ()
* 1    Thread 7256.0x5e04 0x64942e9e in pthread_mutex_unlock () from /cygdrive/c/Users/bcran/workspace/fio/libwinpthread-1.dll
(gdb) thread 3
(gdb) bt
#0  fio_pin_memory (td=td@entry=0x2820000) at memory.c:27
#1  0x00437076 in thread_main (data=0x2820000) at backend.c:1226
#2  0x64944ecb in pthread_create_wrapper () from /cygdrive/c/Users/bcran/workspace/fio/libwinpthread-1.dll
#3  0x752d0bc4 in wtoi64 () from /cygdrive/c/Windows/SysWOW64/msvcrt.dll
#4  0x752d0cec in msvcrt!_beginthreadex () from /cygdrive/c/Windows/SysWOW64/msvcrt.dll
#5  0x7699495d in KERNEL32!BaseThreadInitThunk () from /cygdrive/c/Windows/SysWOW64/KERNEL32.DLL
#6  0x778898ee in ?? ()
#7  0x778898c4 in ?? ()
#8  0x00000000 in ?? ()


-- 
Bruce

Re: ioengine=sync issue on Windows

[Jens Axboe]

On 2014-04-03 21:12, Bruce Cran wrote:
> On 4/3/2014 8:35 PM, Jens Axboe wrote:
>> So next question... If you replace that fill_io_buffer() call with a
>> memset(b, 0, bs); does it then work?
>>
>
> The io error goes away, but now it occasionally crashes:
>
> (gdb) bt
>
> #0  0x64942e9e in pthread_mutex_unlock () from
> /cygdrive/c/Users/bcran/workspace/fio/libwinpthread-1.dll
> #1  0x0043997c in run_threads () at backend.c:1791
> #2  0x00439d9f in fio_backend () at backend.c:1904
> #3  0x0044c274 in main (argc=2, argv=0x821728, envp=0x821988) at fio.c:50
>
> (gdb) info threads
>    Id   Target Id         Frame
>    3    Thread 7256.0x4618 fio_pin_memory (td=td@entry=0x2820000) at
> memory.c:27
>    2    Thread 7256.0xc8  0x7787bf3c in ?? ()
> * 1    Thread 7256.0x5e04 0x64942e9e in pthread_mutex_unlock () from
> /cygdrive/c/Users/bcran/workspace/fio/libwinpthread-1.dll
> (gdb) thread 3
> (gdb) bt
> #0  fio_pin_memory (td=td@entry=0x2820000) at memory.c:27
> #1  0x00437076 in thread_main (data=0x2820000) at backend.c:1226
> #2  0x64944ecb in pthread_create_wrapper () from
> /cygdrive/c/Users/bcran/workspace/fio/libwinpthread-1.dll
> #3  0x752d0bc4 in wtoi64 () from /cygdrive/c/Windows/SysWOW64/msvcrt.dll
> #4  0x752d0cec in msvcrt!_beginthreadex () from
> /cygdrive/c/Windows/SysWOW64/msvcrt.dll
> #5  0x7699495d in KERNEL32!BaseThreadInitThunk () from
> /cygdrive/c/Windows/SysWOW64/KERNEL32.DLL
> #6  0x778898ee in ?? ()
> #7  0x778898c4 in ?? ()
> #8  0x00000000 in ?? ()

This really sounds like an issue with write() on windows. The change 
looks fine, and the fact that moving the memset around makes a 
difference - well, that's just bizarre. Do you have time to dig into 
this? Was read/write really completely reliable before this change?

What if we allocate a bigger buffer, does that change anything? Change 
that malloc() to:

b = fio_memalign(4096, 2 * td->o.max_bs[DDIR_WRITE]);

(and the corresponding free to fio_memfree(b, 2 * td->o.max_bs[DDIR_WRITE])

-- 
Jens Axboe

Re: ioengine=sync issue on Windows

[Bruce Cran]

On 4/3/2014 9:28 PM, Jens Axboe wrote:
> This really sounds like an issue with write() on windows. The change 
> looks fine, and the fact that moving the memset around makes a 
> difference - well, that's just bizarre. Do you have time to dig into 
> this? Was read/write really completely reliable before this change?

It turns out it was a problem with open() - you need to pass _O_BINARY 
to avoid Windows trying to interpret file contents.

-- 
Bruce