Win32 rename()/unlink() questions

On Wed, Sep 18, 2002 at 08:01:42PM -0400, Bruce Momjian wrote:

Second, when you unlink() a file on Win32, do applications continue
accessing the old file contents if they had the file open before the
unlink?

I'm pretty sure it errors with 'file in use'. Pretty ugly, huh?

Ross

Bruce Momjian wrote:

I am working with several groups getting the Win32 port ready for 7.4
and I have a few questions:

What is the standard workaround for the fact that rename() isn't atomic
on Win32? Do we need to create our own locking around the
reading/writing of files that are normally updated in place using
rename()?

Visual C++ comes with the source to Microsoft's C library:

rename() calls MoveFile() which will error if:

1. The target file exists
2. The source file is in use

MoveFileEx() (not available on 95/98) can overwrite the target
file if it exists. The Apache APR portability library uses
MoveFileEx() to rename files if under NT/XP/2K vs. a sequence of :

1. CreateFile() to test for target file existence
2. DeleteFile() to remove the target file
3. MoveFile() to rename the old file to new

under Windows 95/98. Of course, some other process could create
the target file between 2 and 3, so their rename() would just
error out in that situation. I haven't tested it, but I recall
reading somewhere that MoveFileEx() has the ability to rename an
opened file. I'm 99% sure MoveFile() will fail if the source
file is open.

Second, when you unlink() a file on Win32, do applications continue
accessing the old file contents if they had the file open before the
unlink?

unlink() just calls DeleteFile() which will error if:

1. The target file is in use

CreateFile() has the option:

FILE_FLAG_DELETE_ON_CLOSE

which might be able to be used to simulate traditional unlink()
behavior.

Hope that helps,

Mike Mascari
mascarm@mascari.com

On Wed, Sep 18, 2002 at 08:01:42PM -0400, Bruce Momjian wrote:

Second, when you unlink() a file on Win32, do applications continue
accessing the old file contents if they had the file open before the
unlink?

I'm pretty sure it errors with 'file in use'. Pretty ugly, huh?

Yeah - the windows filesystem is pretty poor when it comes to multiuser
access. That's why even as administrator I cannot delete borked files and
people's profiles and stuff off our NT server - the files are always 'in
use'. Even if you kick all users off, reboot the machine, do whatever.
It's terrible.

Chris

Christopher Kings-Lynne wrote:

On Wed, Sep 18, 2002 at 08:01:42PM -0400, Bruce Momjian wrote:

Second, when you unlink() a file on Win32, do applications continue
accessing the old file contents if they had the file open before the
unlink?

I'm pretty sure it errors with 'file in use'. Pretty ugly, huh?

Yeah - the windows filesystem is pretty poor when it comes to multiuser
access. That's why even as administrator I cannot delete borked files and
people's profiles and stuff off our NT server - the files are always 'in
use'. Even if you kick all users off, reboot the machine, do whatever.
It's terrible.

Chris

Yep. That's why often it requires rebooting to uninstall
software. How can the installer remove itself? Under Windows
95/98/ME, you have to manually add entries to WININIT.INI. With
Windows NT/XP/2K, MoveFileEx() with a NULL target and the
MOVEFILE_DELAY_UNTIL_REBOOT flag will add the appropriate
entries into the system registry so that the next time the
machine reboots it will remove the files specified. Its a real
pain and a real hack of an OS.

Mike Mascari
mascarm@mascari.com

Mike Mascari wrote:

Bruce Momjian wrote:

I am working with several groups getting the Win32 port ready for 7.4
and I have a few questions:

What is the standard workaround for the fact that rename() isn't atomic
on Win32? Do we need to create our own locking around the
reading/writing of files that are normally updated in place using
rename()?

Visual C++ comes with the source to Microsoft's C library:

rename() calls MoveFile() which will error if:

1. The target file exists
2. The source file is in use

MoveFileEx() (not available on 95/98) can overwrite the target
file if it exists. The Apache APR portability library uses
MoveFileEx() to rename files if under NT/XP/2K vs. a sequence of :

1. CreateFile() to test for target file existence
2. DeleteFile() to remove the target file
3. MoveFile() to rename the old file to new

under Windows 95/98. Of course, some other process could create
the target file between 2 and 3, so their rename() would just
error out in that situation. I haven't tested it, but I recall
reading somewhere that MoveFileEx() has the ability to rename an
opened file. I'm 99% sure MoveFile() will fail if the source
file is open.

OK, I downloaded APR and see in apr_file_rename():

if (MoveFileEx(frompath, topath, MOVEFILE_REPLACE_EXISTING |
MOVEFILE_COPY_ALLOWED))

Looking at the entire APR function, they have lots of tests so it works
on Win9X and wide characters. I think we will just use the APR as a
guide in implementing the things we need. I think MoveFileEx() is the
proper way to go; any other solution requires loop tests for rename.

I see the MoveFileEx manual page at:

http://msdn.microsoft.com/library/default.asp?url=/library/en-us/fileio/base/movefile.asp

Second, when you unlink() a file on Win32, do applications continue
accessing the old file contents if they had the file open before the
unlink?

unlink() just calls DeleteFile() which will error if:

1. The target file is in use

CreateFile() has the option:

FILE_FLAG_DELETE_ON_CLOSE

which might be able to be used to simulate traditional unlink()
behavior.

No, that flag isn't going to help us. I wonder what MoveFileEx does if
the target file exists _and_ is open by another user? I don't see any
loop in that Win32 rename() routine, and I looked at the Unix version of
apr_file_rename and its just a straight rename() call.

-- 
  Bruce Momjian                        |  http://candle.pha.pa.us
  pgman@candle.pha.pa.us               |  (610) 359-1001
  +  If your life is a hard drive,     |  13 Roberts Road
  +  Christ can be your backup.        |  Newtown Square, Pennsylvania 19073

Bruce Momjian wrote:

Second, when you unlink() a file on Win32, do applications continue
accessing the old file contents if they had the file open before the
unlink?

unlink() just calls DeleteFile() which will error if:

1. The target file is in use

CreateFile() has the option:

FILE_FLAG_DELETE_ON_CLOSE

which might be able to be used to simulate traditional unlink()
behavior.

No, that flag isn't going to help us. I wonder what MoveFileEx does if
the target file exists _and_ is open by another user? I don't see any
loop in that Win32 rename() routine, and I looked at the Unix version of
apr_file_rename and its just a straight rename() call.

This says that if the target is in use, it is overwritten:

http://support.microsoft.com/default.aspx?scid=KB;EN-US;q140570&

While I think that is good news, does it open the problem of other
readers reading partial updates to the file and therefore seeing
garbage. Not sure how to handle that, nor am I even sure how I would
test it.

-- 
  Bruce Momjian                        |  http://candle.pha.pa.us
  pgman@candle.pha.pa.us               |  (610) 359-1001
  +  If your life is a hard drive,     |  13 Roberts Road
  +  Christ can be your backup.        |  Newtown Square, Pennsylvania 19073

Bruce Momjian wrote:

Bruce Momjian wrote:

unlink() just calls DeleteFile() which will error if:

1. The target file is in use

CreateFile() has the option:

FILE_FLAG_DELETE_ON_CLOSE

which might be able to be used to simulate traditional unlink()
behavior.

No, that flag isn't going to help us. I wonder what MoveFileEx does if
the target file exists _and_ is open by another user? I don't see any
loop in that Win32 rename() routine, and I looked at the Unix version of
apr_file_rename and its just a straight rename() call.

This says that if the target is in use, it is overwritten:

http://support.microsoft.com/default.aspx?scid=KB;EN-US;q140570&

I read the article and did not come away with that conclusion.
The article describes using the MOVEFILE_DELAY_UNTIL_REBOOT
flag, which was created for the express purpose of allowing a
SETUP.EXE to remove itself, or rather tell Windows to remove it
on the next reboot. Also, if you want the Win32 port to run in
95/98/ME, you can't rely on MoveFileEx(), you have to use
MoveFile().

I will do some testing with concurrency and let you know. But
don't get your hopes up. This is one of the many advantages that
TABLESPACEs have when more than one relation is stored in a
single DATAFILE. There was Oracle for MS-DOS, after all..

Mike Mascari
mascarm@mascari.com

Mike Mascari wrote:

I read the article and did not come away with that conclusion.
The article describes using the MOVEFILE_DELAY_UNTIL_REBOOT
flag, which was created for the express purpose of allowing a
SETUP.EXE to remove itself, or rather tell Windows to remove it
on the next reboot. Also, if you want the Win32 port to run in
95/98/ME, you can't rely on MoveFileEx(), you have to use
MoveFile().

I will do some testing with concurrency and let you know. But
don't get your hopes up. This is one of the many advantages that
TABLESPACEs have when more than one relation is stored in a
single DATAFILE. There was Oracle for MS-DOS, after all..

I was focusing on handling of pg_pwd and other config file that are
written by various backend while other backends are reading them. The
actual data files should be OK because we have an exclusive lock when we
are adding/removing them.

-- 
  Bruce Momjian                        |  http://candle.pha.pa.us
  pgman@candle.pha.pa.us               |  (610) 359-1001
  +  If your life is a hard drive,     |  13 Roberts Road
  +  Christ can be your backup.        |  Newtown Square, Pennsylvania 19073

Bruce Momjian wrote:

Mike Mascari wrote:

I will do some testing with concurrency and let you know. But
don't get your hopes up. This is one of the many advantages that
TABLESPACEs have when more than one relation is stored in a
single DATAFILE. There was Oracle for MS-DOS, after all..

I was focusing on handling of pg_pwd and other config file that are
written by various backend while other backends are reading them. The
actual data files should be OK because we have an exclusive lock when we
are adding/removing them.

OK. So you want to test:

1. Process 1 opens "foo"
2. Process 2 opens "foo"
3. Process 1 renames "foo" to "bar"
4. Process 2 can safely read from its open file handle

Is that what you want tested? I have a small Win32 app ready to
test. Just let me know the scenarios...

Mike Mascari
mascarm@mascari.com

Mike Mascari wrote:

Bruce Momjian wrote:

Mike Mascari wrote:

I will do some testing with concurrency and let you know. But don't
get your hopes up. This is one of the many advantages that
TABLESPACEs have when more than one relation is stored in a single
DATAFILE. There was Oracle for MS-DOS, after all..

I was focusing on handling of pg_pwd and other config file that are
written by various backend while other backends are reading them. The
actual data files should be OK because we have an exclusive lock when we
are adding/removing them.

OK. So you want to test:

1. Process 1 opens "foo"
2. Process 2 opens "foo"
3. Process 1 renames "foo" to "bar"
4. Process 2 can safely read from its open file handle

Actually, looking at the pg_pwd code, you want to determine a
way for:

1. Process 1 opens "foo"
2. Process 2 opens "foo"
3. Process 1 creates "bar"
4. Process 1 renames "bar" to "foo"
5. Process 2 can continue to read data from the open file handle
and get the original "foo" data.

Is that correct?

Mike Mascari
mascarm@mascari.com

Mike Mascari wrote:

Actually, looking at the pg_pwd code, you want to determine a
way for:

1. Process 1 opens "foo"
2. Process 2 opens "foo"
3. Process 1 creates "bar"
4. Process 1 renames "bar" to "foo"
5. Process 2 can continue to read data from the open file handle
and get the original "foo" data.

Yep, that's it.

-- 
  Bruce Momjian                        |  http://candle.pha.pa.us
  pgman@candle.pha.pa.us               |  (610) 359-1001
  +  If your life is a hard drive,     |  13 Roberts Road
  +  Christ can be your backup.        |  Newtown Square, Pennsylvania 19073

Bruce Momjian wrote:

Mike Mascari wrote:

Actually, looking at the pg_pwd code, you want to determine a
way for:

1. Process 1 opens "foo"
2. Process 2 opens "foo"
3. Process 1 creates "bar"
4. Process 1 renames "bar" to "foo"
5. Process 2 can continue to read data from the open file handle
and get the original "foo" data.

Yep, that's it.

So far, MoveFileEx("foo", "bar", MOVEFILE_REPLACE_EXISTING)
returns "Access Denied" when Process 1 attempts the rename. But
I'm continuing to investigate the possibilities...

Mike Mascari
mascarm@mascari.com

On Fri, 20 Sep 2002, Mike Mascari wrote:

Bruce Momjian wrote:

Mike Mascari wrote:

Actually, looking at the pg_pwd code, you want to determine a
way for:

1. Process 1 opens "foo"
2. Process 2 opens "foo"
3. Process 1 creates "bar"
4. Process 1 renames "bar" to "foo"
5. Process 2 can continue to read data from the open file handle
and get the original "foo" data.

Yep, that's it.

So far, MoveFileEx("foo", "bar", MOVEFILE_REPLACE_EXISTING)
returns "Access Denied" when Process 1 attempts the rename. But
I'm continuing to investigate the possibilities...

Does a sequence like
Process 1 opens "foo"
Process 2 opens "foo"
Process 1 creates "bar"
Process 1 renames "foo" to <something>
- where something is generated to not overlap an existing file
Process 1 renames "bar" to "foo"
Process 2 continues reading
let you do the replace and keep reading (at the penalty that
you've now got to have a way to know when to remove the
various <something>s)

Stephan Szabo wrote:

On Fri, 20 Sep 2002, Mike Mascari wrote:

Bruce Momjian wrote:

Mike Mascari wrote:

Actually, looking at the pg_pwd code, you want to determine a
way for:

1. Process 1 opens "foo"
2. Process 2 opens "foo"
3. Process 1 creates "bar"
4. Process 1 renames "bar" to "foo"
5. Process 2 can continue to read data from the open file handle
and get the original "foo" data.

Yep, that's it.

So far, MoveFileEx("foo", "bar", MOVEFILE_REPLACE_EXISTING)
returns "Access Denied" when Process 1 attempts the rename. But
I'm continuing to investigate the possibilities...

Does a sequence like
Process 1 opens "foo"
Process 2 opens "foo"
Process 1 creates "bar"
Process 1 renames "foo" to <something>
- where something is generated to not overlap an existing file
Process 1 renames "bar" to "foo"
Process 2 continues reading
let you do the replace and keep reading (at the penalty that
you've now got to have a way to know when to remove the
various <something>s)

Yes! Indeed that does work.

Mike Mascari
mascarm@mascari.com

On Fri, 20 Sep 2002, Mike Mascari wrote:

Stephan Szabo wrote:

On Fri, 20 Sep 2002, Mike Mascari wrote:

So far, MoveFileEx("foo", "bar", MOVEFILE_REPLACE_EXISTING)
returns "Access Denied" when Process 1 attempts the rename. But
I'm continuing to investigate the possibilities...

Does a sequence like
Process 1 opens "foo"
Process 2 opens "foo"
Process 1 creates "bar"
Process 1 renames "foo" to <something>
- where something is generated to not overlap an existing file
Process 1 renames "bar" to "foo"
Process 2 continues reading
let you do the replace and keep reading (at the penalty that
you've now got to have a way to know when to remove the
various <something>s)

Yes! Indeed that does work.

Thinking back, I think that may still fail on Win95 (using MoveFile).
Once in the past I had to work on (un)installers for Win* and I
vaguely remember Win95 being more strict than Win98 but that may just
have been with moving the executable you're currently running.

Stephan Szabo wrote:

On Fri, 20 Sep 2002, Mike Mascari wrote:

Yes! Indeed that does work.

Thinking back, I think that may still fail on Win95 (using MoveFile).
Once in the past I had to work on (un)installers for Win* and I
vaguely remember Win95 being more strict than Win98 but that may just
have been with moving the executable you're currently running.

Well, here's the test:

foo.txt contains "This is FOO!"
bar.txt contains "This is BAR!"

Process 1 opens foo.txt
Process 2 opens foo.txt
Process 1 sleeps 7.5 seconds
Process 2 sleeps 15 seconds
Process 1 uses MoveFile() to rename "foo.txt" to "foo2.txt"
Process 1 uses MoveFile() to rename "bar.txt" to "foo.txt"
Process 1 uses DeleteFile() to remove "foo2.txt"
Process 2 awakens and displays "This is FOO!"

On the filesystem, we then have:

foo.txt containing "This is BAR!"

The good news is that this works fine under NT 4 using just
MoveFile(). The bad news is that it requires the files be opened
using CreateFile() with the FILE_SHARE_DELETE flag set. The C
library which ships with Visual C++ 6 ultimately calls
CreateFile() via fopen() but with no opportunity through the
standard C library routines to use the FILE_SHARE_DELETE flag.
And the FILE_SHARE_DELETE flag cannot be used under Windows
95/98 (Bad Parameter). Which means, on those platforms, there
still doesn't appear to be a solution. Under NT/XP/2K,
AllocateFile() will have to modified to call CreateFile()
instead of fopen(). I'm not sure about ME, but I suspect it
behaves similarly to 95/98.

Mike Mascari
mascarm@mascari.com

Stephan Szabo <sszabo@megazone23.bigpanda.com> writes:

... let you do the replace and keep reading (at the penalty that
you've now got to have a way to know when to remove the
various <something>s)

That is the hard part. Mike's description omitted one crucial step:

6. The old "foo" goes away when the last open file handle for it is
closed.

I doubt there is any practical way for Postgres to cause that to happen
if the OS itself does not have any support for it.

regards, tom lane

I don't think we are not going to be supporting Win9X so there isn't an
issue there. We will be supporting Win2000/NT/XP.

I don't understand FILE_SHARE_DELETE. I read the description at:

http://msdn.microsoft.com/library/default.asp?url=/library/en-us/fileio/base/createfile.asp

but I don't understand it:

FILE_SHARE_DELETE - Windows NT/2000/XP: Subsequent open operations on
the object will succeed only if delete access is requested.

---------------------------------------------------------------------------

Mike Mascari wrote:

Stephan Szabo wrote:

On Fri, 20 Sep 2002, Mike Mascari wrote:

Yes! Indeed that does work.

Thinking back, I think that may still fail on Win95 (using MoveFile).
Once in the past I had to work on (un)installers for Win* and I
vaguely remember Win95 being more strict than Win98 but that may just
have been with moving the executable you're currently running.

Well, here's the test:

foo.txt contains "This is FOO!"
bar.txt contains "This is BAR!"

Process 1 opens foo.txt
Process 2 opens foo.txt
Process 1 sleeps 7.5 seconds
Process 2 sleeps 15 seconds
Process 1 uses MoveFile() to rename "foo.txt" to "foo2.txt"
Process 1 uses MoveFile() to rename "bar.txt" to "foo.txt"
Process 1 uses DeleteFile() to remove "foo2.txt"
Process 2 awakens and displays "This is FOO!"

On the filesystem, we then have:

foo.txt containing "This is BAR!"

The good news is that this works fine under NT 4 using just
MoveFile(). The bad news is that it requires the files be opened
using CreateFile() with the FILE_SHARE_DELETE flag set. The C
library which ships with Visual C++ 6 ultimately calls
CreateFile() via fopen() but with no opportunity through the
standard C library routines to use the FILE_SHARE_DELETE flag.
And the FILE_SHARE_DELETE flag cannot be used under Windows
95/98 (Bad Parameter). Which means, on those platforms, there
still doesn't appear to be a solution. Under NT/XP/2K,
AllocateFile() will have to modified to call CreateFile()
instead of fopen(). I'm not sure about ME, but I suspect it
behaves similarly to 95/98.

Mike Mascari
mascarm@mascari.com

---------------------------(end of broadcast)---------------------------
TIP 1: subscribe and unsubscribe commands go to majordomo@postgresql.org

-- 
  Bruce Momjian                        |  http://candle.pha.pa.us
  pgman@candle.pha.pa.us               |  (610) 359-1001
  +  If your life is a hard drive,     |  13 Roberts Road
  +  Christ can be your backup.        |  Newtown Square, Pennsylvania 19073

Bruce Momjian wrote:

I don't think we are not going to be supporting Win9X so there isn't an
issue there. We will be supporting Win2000/NT/XP.

I don't understand FILE_SHARE_DELETE. I read the description at:

http://msdn.microsoft.com/library/default.asp?url=/library/en-us/fileio/base/createfile.asp

but I don't understand it:

FILE_SHARE_DELETE - Windows NT/2000/XP: Subsequent open operations on
the object will succeed only if delete access is requested.

I think that's a rather poor description. I think it just means
that if the file is opened once via CreateFile() with
FILE_SHARE_DELETE, then any subsequent CreateFile() calls will
fail unless they too have FILE_SHARE_DELETE. In other words, if
one of us can delete this file while its open, any of us can.

Mike Mascari
mascarm@mascari.com