ID:               26494
 Updated by:       [EMAIL PROTECTED]
 Reported By:      hexer at studentcenter dot org
-Status:           Open
+Status:           Feedback
 Bug Type:         Apache related
 Operating System: *
 PHP Version:      4CVS, 5CVS (2005-01-25)
 New Comment:

Can you provide that patch somewhere in the web and paste the url here?
Also, please make the patch unified diff (diff -u).



Previous Comments:
------------------------------------------------------------------------

[2005-01-27 23:24:33] hexer at studentcenter dot org

Ok I tried it, same results. It's the problem in mod_php4.c
also just as a side note, our php is compiled-in to apache.

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

[2005-01-25 15:25:56] [EMAIL PROTECTED]

Please try using this CVS snapshot:

  http://snaps.php.net/php4-STABLE-latest.tar.gz
 
For Windows:
 
  http://snaps.php.net/win32/php4-win32-STABLE-latest.zip

Always try the snapshot first..


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

[2005-01-11 21:50:43] hexer at studentcenter dot org

Ok here is a more detailed step-by-step on how to reproduce:

Use php 4.3.10 compiled-in with Apache 1.3.33

1. Compile Apache 1.3 with php-4.3.10 compiled-in:

In source tree:
./configure --prefix=/usr/local/apache

In PHP source tree:
./configure --with-mysql \
--with-xml \
--with-apache=../apache_1.3.33 \
--enable-track-vars \
--with-gd \
--enable-discard-path \
--enable-bcmath \
--enable-gd-native-tt \
--with-freetype-dir=/usr/local \
--with-png-dir=/usr/local \
--with-jpeg-dir=/usr/local \
--with-zlib=/usr/local

make
make install

In Apache source tree:
./configure \
--activate-module=src/modules/php4/libphp4.a \
--enable-module=php4 \
--enable-module=rewrite \
--enable-module=so \
--prefix=/usr/local/apache

make
make install


2. Edit httpd.conf and 
make sure you have the following enabled in httpd.conf:

TimeOut 60

AddModule mod_status.c

ExtendedStatus On

<Location /server-status>
    SetHandler server-status
    Order deny,allow
    Allow from all
</Location>



3. Create the test output script (test.php):

<?php
for($i = 0; $i < 1000000; $i++) {
        echo "XXXXXXXXXXXXXXXXXXXXXXXX<br>\n";
}
?>


4. On a different computer, open the browser and point it to
test.php, as soon as test.php starts loading unplug the
computer from the network, before it finishes loading.
This one's important. You have to make a true dead connection without
the server knowing about it.
Make sure your network does not notify the server immediately that the
connection was closed or there is a broken pipe etc.
(by default the Linux recv has a very long timeout)


5. Check the server-status page. You will see something like this:

0-1 4990 0/177/15106 W  0.45 9 0 0.0 0.19 43.85  x.x.x.x
your.server.com GET /test.php

Note the 'W' under status in that particular process.

6. You will see the SS value start going way beyond the timeout
setting. This _must_ happen unless the system has somehow managed to
find out that the connection is broken.

If the process is not in W status and 60 seconds have not passed that
means the system somehow found out that the connection is broken so the
dead-connection
condition hasn't been simulated.


Also another good experiment would be to see if the TimeOut overrides
max_execution_time. 
If in your scenario TimeOut does in fact work, it probably overrides
the max_execution_time and causes the script to abort if it runs longer
than TimeOut seconds.

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

[2005-01-11 20:28:39] [EMAIL PROTECTED]

I haven't been able to reproduce this either even with MySQL enabled.

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

[2005-01-11 19:10:22] hexer at studentcenter dot org

Update: More test results, some improvements to the patch.

After some testing on our production cluster I've made some
improvements on the patch. 
(Scroll down for the patch itself)

For a better way on reproducing this, check out the newsgroups post
that talks more about this problem:
http://groups-beta.google.com/group/php.general/browse_frm/thread/f18f3da1cc3b51a7/1e7c60d14432f6c5?q=mod_php+timeout&_done=%2Fgroups%3Fq%3Dmod_php+timeout%26hl%3Den%26lr%3D%26sa%3DN%26tab%3Dwg%26&_doneTitle=Back+to+Search&&d#1e7c60d14432f6c5

(Reproducible with latest version of Apache 1.3 and php 4.3.10)

THE PATCH: (mod_php)
List of changes since the initial version of the patch:

1. hard_timeout is no longer set before the main entry into the php
engine execution.
(TimeOut must only be used to monitor for stalled client connections
according to the apache
spec)

2. fixed for proper use of apache timeout in the READ post section.
(timer is now primed before the start the read loop, timer refreshes in
between the reads and is killed
after the read loop so that TimeOut is tracked during communication and
refreshed accordingly.

3. changed to ap_soft_timeout in the ub_write section. This makes PHP's
ignore_user_abort
possible again. (ap_soft_timeout causes apache to close socket instead
of forcing a quit)


Our test results:

Upon setting apache TimeOut to 60, the server closes the connection
with the client if and only if the connection has been stalled for more
than 60 seconds during transfers. 

Since long script execution times are not considered stalls,
the server only primes TimeOut timer during non-I/O operations of php
and "kills" it afterwards.

In other words, a php page can run for more than TimeOut number of
seconds until max_execution_time if it is doing something other than
outputting to the client or transfering POST data.


Performance analysis:

We have been running the earlier version of this patch on our
production servers since this bug was first opened
for several months now. There has been no noticable performance
penalty.

This new version of the patch has been running for about a week. We
suggest a more thorough testing just in case.
Especially with stalled POST's because that was the new addition to
this patch.


NOTE: To replicate this problem you MUST simulate a connection stall
before a php page finishes its output. Hitting the stop button on a
browser is not sufficient.
We had to literally unplug the client computer's ethernet connection
immediately after request to a php page to test this.


--- /usr/local/src/php-4.3.10/sapi/apache/mod_php4.c    2005-01-10
15:11:53.000000000 -0500
***************
*** 97,107 ****
  static int sapi_apache_ub_write(const char *str, uint str_length
TSRMLS_DC)
  {
        int ret=0;
  
        if (SG(server_context)) {
!               ret = rwrite(str, str_length, (request_rec *)
SG(server_context));
        }
        if (ret != str_length) {
                php_handle_aborted_connection();
        }
        return ret;
--- 97,109 ----
  static int sapi_apache_ub_write(const char *str, uint str_length
TSRMLS_DC)
  {
        int ret=0;
  
        if (SG(server_context)) {
!         ap_soft_timeout("php send body", (request_rec *)
SG(server_context));
!         ret = rwrite(str, str_length, (request_rec *)
SG(server_context));
!         ap_kill_timeout((request_rec *) SG(server_context));
        }
        if (ret != str_length) {
                php_handle_aborted_connection();
        }
        return ret;
***************
*** 139,157 ****
        if (!SG(read_post_bytes) && !ap_should_client_block(r)) {
                return total_read_bytes;
        }
   
        handler = signal(SIGPIPE, SIG_IGN);
        while (total_read_bytes<count_bytes) {
-               hard_timeout("Read POST information", r); /* start
timeout timer */
                read_bytes = get_client_block(r,
buffer+total_read_bytes, count_bytes-total_read_bytes);
!               reset_timeout(r);
                if (read_bytes<=0) {
                        break;
                }
                total_read_bytes += read_bytes;
        }
        signal(SIGPIPE, handler);
        return total_read_bytes;
  }
  /* }}} */
  
--- 141,161 ----
        if (!SG(read_post_bytes) && !ap_should_client_block(r)) {
                return total_read_bytes;
        }
   
        handler = signal(SIGPIPE, SIG_IGN);
+ 
+     ap_hard_timeout("Read POST information", r); /* start timeout
timer */
        while (total_read_bytes<count_bytes) {
                read_bytes = get_client_block(r,
buffer+total_read_bytes, count_bytes-total_read_bytes);
!               ap_reset_timeout(r);
                if (read_bytes<=0) {
                        break;
                }
                total_read_bytes += read_bytes;
        }
+       ap_kill_timeout(r);
        signal(SIGPIPE, handler);
        return total_read_bytes;
  }
  /* }}} */
  
***************
*** 606,617 ****
                }
                /* Assume output will be of the default MIME type. 
Individual
                   scripts may change this later in the request. */
                r->content_type = php_apache_get_default_mimetype(r
TSRMLS_CC);
  
-               /* Init timeout */
-               hard_timeout("send", r);
  
                SG(server_context) = r;
  
                php_save_umask();
                add_common_vars(r);
--- 610,619 ----
***************
*** 620,630 ****
                init_request_info(TSRMLS_C);
                apache_php_module_main(r, display_source_mode
TSRMLS_CC);
  
                /* Done, restore umask, turn off timeout, close file
and return */
                php_restore_umask();
-               kill_timeout(r);
        } zend_end_try();
  
        return OK;
  }
  /* }}} */
--- 622,631 ----

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

The remainder of the comments for this report are too long. To view
the rest of the comments, please view the bug report online at
    http://bugs.php.net/26494

-- 
Edit this bug report at http://bugs.php.net/?id=26494&edit=1

Reply via email to