This document is a collection of notes regarding tools and techniques for debugging Apache httpd and its modules.
Got more tips? Send 'em to docs@httpd.apache.org. Thanks!
If you use the gcc compiler, it is likely that the best debugger for your
system is gdb. This is only a brief summary of how to run gdb on Apache --
you should look at the info and man files for gdb to get more information
on gdb commands and common debugging techniques. Before running gdb, be
sure that the server is compiled with the -g
option in CFLAGS
to
include the symbol information in the object files.
The only tricky part of running gdb on Apache is forcing the server into a
single-process mode so that the parent process being debugged does the
request-handling work instead of forking child processes. We have provided
the -X
option for that purpose, which will work fine for most cases.
However, some modules don't like starting up with -X
, but are happy if
you force only one child to run (using " MaxClients 1
"); you can then
use gdb's attach command to debug the child server.
The following example, with user input in green,
shows the output of gdb run on a server executable (httpd) in the current
working directory and using the server root of /usr/local/apache
:
% gdb httpd GDB is free software and you are welcome to distribute copies of it under certain conditions; type "show copying" to see the conditions. There is absolutely no warranty for GDB; type "show warranty" for details. GDB 4.16.gnat.1.13 (sparc-sun-solaris2.5), Copyright 1996 Free Software Foundation, Inc... (gdb) b ap_process_request Breakpoint 1 at 0x49fb4: file http_request.c, line 1164. (gdb) run -X -d /usr/local/apache Starting program: /usr/local/apache/src/httpd -X -d /usr/local/apache [ at this point I make a request from another window ] Breakpoint 1, ap_process_request (r=0x95250) at http_request.c:1164 1164 if (ap_extended_status) (gdb) s 1165 ap_time_process_request(r->connection->child_num,... (gdb) n 1167 process_request_internal(r); (gdb) s process_request_internal (r=0x95250) at http_request.c:1028 1028 if (!r->proxyreq && r->parsed_uri.path) { (gdb) s 1029 access_status = ap_unescape_url(r->parsed_uri.path); (gdb) n 1030 if (access_status) { (gdb) s 1036 ap_getparents(r->uri); /* OK... (gdb) n 1038 if ((access_status = location_walk(r))) { (gdb) n 1043 if ((access_status = ap_translate_name(r))) { (gdb) n 1048 if (!r->proxyreq) { (gdb) n 1053 if (r->method_number == M_TRACE) { (gdb) n 1062 if (r->proto_num > HTTP_VERSION(1,0) && ap_... (gdb) n 1071 if ((access_status = directory_walk(r))) { (gdb) s directory_walk (r=0x95250) at http_request.c:288 288 core_server_config *sconf = ap_get_module_... (gdb) b ap_send_error_response Breakpoint 2 at 0x47dcc: file http_protocol.c, line 2090. (gdb) c Continuing. Breakpoint 2, ap_send_error_response (r=0x95250, recursive_error=0) at http_protocol.c:2090 2090 BUFF *fd = r->connection->client; (gdb) where #0 ap_send_error_response (r=0x95250, recursive_error=0) at http_protocol.c:2090 #1 0x49b10 in ap_die (type=403, r=0x95250) at http_request.c:989 #2 0x49b60 in decl_die (status=403, phase=0x62db8 "check access", r=0x95250) at http_request.c:1000 #3 0x49f68 in process_request_internal (r=0x95250) at http_request.c:1141 #4 0x49fe0 in ap_process_request (r=0x95250) at http_request.c:1167 #5 0x439d8 in child_main (child_num_arg=550608) at http_main.c:3826 #6 0x43b5c in make_child (s=0x7c3e8, slot=0, now=907958743) at http_main.c:3898 #7 0x43ca8 in startup_children (number_to_start=6) at http_main.c:3972 #8 0x44260 in standalone_main (argc=392552, argv=0x75800) at http_main.c:4250 #9 0x449fc in main (argc=4, argv=0xefffee8c) at http_main.c:4534 (gdb) s 2091 int status = r->status; (gdb) p status $1 = 403 (gdb)
There are a few things to note about the above example:
the " gdb httpd
" command does not include any command-line options
for httpd: those are provided when the " run
" command is done within
gdb;
I set a breakpoint before starting the run so that execution would stop at the top of ap_process_request();
the " s
" command steps through the code and into called procedures,
whereas the " n
" (next) command steps through the code but not into
called procedures.
additional breakpoints can be set with the " b
" command, and the run
continued with the " c
" command.
use the " where
" command (a.k.a. " bt
") to see a stack backtrace
that shows the order of called procedures and their parameter values.
use the " p
" command to print the value of a variable.
A file in the the root directory called .gdbinit
provides useful macros
for printing out various internal structures of httpd like tables (
dump_table
), brigades ( dump_brigade
) and filter chains (
dump_filters
).
If you are debugging a repeatable crash, simply run gdb as above and make the request -- gdb should capture the crash and provide a prompt where it occurs.
If you are debugging an apparent infinite loop, simply run gdb as above and type a Control-C -- gdb will interrupt the process and provide a prompt where it was stopped.
If you are debugging a system crash and you have a core file from the crash, then do the following:
% gdb httpd -c core
(gdb) where
and it will (hopefully) print a stack backtrace of where the core dump occurred during processing.
A backtrace will let you know the hierarchy of procedures that were called to get to a particular point in the process. On some platforms you can get a live backtrace of any process.
For SVR4-based variants of Unix, the pstack
command for proc can be used
to display a a live backtrace. For example, on Solaris it looks like
% /usr/proc/bin/pstack 10623
10623: httpd -d /usr/local/apache
ef5b68d8 poll (efffcd08, 0, 3e8)
ef5d21e0 select (0, ef612c28, 0, 0, 3e8, efffcd08) + 288
00042574 wait_or_timeout (0, 75000, 75000, 7c3e8, 60f40, 52c00) + 78
00044310 standalone_main (5fd68, 75800, 75c00, 75000, 2, 64) + 240
000449f4 main (3, efffeee4, efffeef4, 75fe4, 1, 0) + 374
000162fc _start (0, 0, 0, 0, 0, 0) + 5c
Another technique is to use gdb to attach to the running process and then using "thread apply all bt" to print the backtrace, as in:
sudo gdb /path/to/bin/httpd 10623 --batch --eval-command "where" --eval-command "thread apply all bt" --eval-command "detach" --eval-command "quit" | tee gdb-output.txt
Or interactively:
% gdb httpd 10623
GDB is free software and you are welcome to distribute copies of it
under certain conditions; type "show copying" to see the conditions.
There is absolutely no warranty for GDB; type "show warranty" for
details.
GDB 4.16.gnat.1.13 (sparc-sun-solaris2.5),
Copyright 1996 Free Software Foundation, Inc...
/usr/local/apache/src/10623: No such file or directory.
Attaching to program `/usr/local/apache/src/httpd', process 10623
Reading symbols from /usr/lib/libsocket.so.1...done.
...
0xef5b68d8 in ()
(gdb) where
#0 0xef5b68d8 in ()
#1 0xef5d21e8 in select ()
#2 0x4257c in wait_or_timeout (status=0x0) at http_main.c:2357
#3 0x44318 in standalone_main (argc=392552, argv=0x75800) at...
#4 0x449fc in main (argc=3, argv=0xefffeee4) at http_main.c:4534
(gdb) thread apply all bt
(gdb) detach
(gdb> quit
Unzip the -symbols.zip
files (obtained from the Apache download site)
in the root Apache2 directory tree (where bin, htdocs, modules\ etc. are
usually found.) These.pdb files should unpack alongside the.exe,.dll,.so
binary files they represent, e.g., mod_usertrack.pdb will unpack alongside
mod_usertrack.so.
Optional: Obtain procdump
from Sysinternals.
Invoke drwtsn32
and ensure you are creating a crash dump file, you are
dumping all thread contexts, your log and crash dump paths make sense, and
(depending on the nature of the bug) you pick an appropriate crash dump
type. (Full is quite large, but necessary sometimes for a programmer-type
to load your crash dump into a debugger and begin unwinding exactly what
has happened. Mini is sufficient for your first pass through the process.)
You can use procdump
or procdump -ma
at this stage instead.
Note that if you previously installed and then uninstalled other
debugging software, you may need to invoke drwtsn32 -i
in order to make
Dr Watson your default crash dump tool. This will replace the 'report
problem to MS' dialogs. (Don't do this if you have a full debugger such as
Visual Studio or windbg installed on the machine, unless you back up the
registry value for Debugger
under the HKLM\SOFTWARE\Microsoft\Windows NT\CurrentVersion\AeDebug
registry tree. Developers using multiple tools
might want to keep copies of their different tools Debugger entries there,
for fast switching.)
Invoke the Task Manager, Choose 'show processes from all users', and
modify the View -> Select Columns
to include at least the PID
and Thread Count
. You can change this just once and Task Manager should keep your
preference.
Now, track down the errant Apache that is hanging. The parent process has about three threads, we don't care about that one. The child worker process we want has many more threads (a few more than you configured with the ThreadsPerChild directive.) The process name is Apache (for 1.3 and 2.0) or httpd (for 2.2 and 2.4). Make note of the child worker's PID.
Using the {pid} number you noted above, invoke the command
drwtsn32 -p {pid}
orprocdump {pid}
Voila, you will find in your 'log file path' a drwtsn32.log
file, and if
you choose to 'append to existing log file', jump through the 'App:'
sections until you find the one for the process you just killed. Now you
can identify about where 'Stack Back Trace' points to help identify what
the server is doing.
You will note that many threads look identical, almost all of them polling for the next connection, and you don't care about those. You will want to see the ones that are deep inside of a request at the time you kill them, and only the stack back trace entries for those. This can give folks a clue of where that request is hanging, which handler module picked up the request, and what filter it might be stuck in.
For situations where a child process is crashing intermittently, the server must be configured and started such that it produces core dumps which can be analyzed further.
To ensure that a core dump is written to a directory which is writable by
the user which child processes run as (such as apache
), the
directive must be added to httpd.conf
; for example:
CoreDumpDirectory /tmp
Before starting up the server, any process limits on core dump file size must be lifted; for example:
# ulimit -c unlimited
# apachectl start
On some platforms, further steps might be needed to enable core dumps - see Solaris and coredumps below.
When a child process crashes, a message like the following will be logged to the error_log:
[Mon Sep 05 13:35:39 2005] [notice] child pid 2027 exit signal Segmentation
fault (11), possible coredump in /tmp
If the text "possible coredump in /tmp" does not appear in the error line,
check that the ulimit was set correctly, that the permissions on the
configured CoreDumpDirectory
are suitable and that platform specific
steps ( Solaris and coredumps ) have been done if needed.
To analyse the core dump, pass the core dump filename on the gdb
command-line, and enter the command thread apply all bt full
at the gdb prompt:
% gdb /usr/local/apache2/bin/httpd /tmp/core.2027
...
Core was generated by `/usr/local/apache2/bin/httpd -k start'
...
(gdb) thread apply all bt full
Most Unix-based systems have at least one command for displaying a trace of
system calls and signals as they are accessed by a running process. This
command is called truss
on most SVR4-based systems and either trace
or
strace
on many other systems.
A useful tip for using the truss
command on Solaris is the -f
option
(often also works with strace
); it tells truss to follow and continue
tracing any child processes forked by the main process. The easiest way to
get a full trace of a server is to do something like:
% truss -f httpd -d /usr/local/apache >& outfile % egrep '^10698:' outfile
to view just the trace of the process id 10698.
If attempting to truss a threaded server, for example when using the
worker
MPM, the truss
option -l
is very useful as it prints also the
LWP id after the process id. You can use something like
% egrep '^10698/1:' outfile
to view just the trace of the process id 10698 and LWP id 1.
Other useful options for truss are
-a
to print all command line parameters used for this executable.
-e
to print all environment variables used for this executable.
-d
to print timestamps.
Strangely enough, sometimes you actually want to force the server to crash
so that you can get a look at some nutty behavior. Normally this can be
done simply by using the gcore
command. However, for security reasons,
most Unix systems do not allow a setuid process to dump core, since the
file contents might reveal something that is supposed to be protected in
memory.
Here is one way to get a core file from a setuid Apache httpd process on Solaris, without knowing which httpd child might be the one to die [note: it is probably easier to use the MaxClients trick in the first section above].
# for pid in `ps -eaf | fgrep httpd | cut -d' ' -f4` do truss -f -l -t\!all -S SIGSEGV -p $pid 2>&1 | egrep SIGSEGV & done
The undocumented '-S' flag to truss will halt the process in place upon receipt of a given signal (SIGSEGV in this case). At this point you can use:
# gcore PID
and then look at the backtrace as discussed above for gdb.
On Solaris use to
make setuid()
processes actually dump core. By default a setuid()
process
does not dump core. This is the reason why httpd servers started as root
with child processes running as a different user (such as apache
) do not
coredump even when the
CoreDumpDirectory
directive had been set to an appropriate and writable directory and ulimit -c
has a sufficient size.
See also Debugging intermittent crashes above.
Example:
-bash-3.00# coreadm
global core file pattern: /var/core/core.%f.%p.u%u
global core file content: default
init core file pattern: core
init core file content: default
global core dumps: disabled
per-process core dumps: enabled
global setid core dumps: enabled
per-process setid core dumps: enabled
global core dump logging: disabled
This is too deep a subject to fully describe in this documentation. Here are some pointers to useful discussions and tools:
snoop
is a packet sniffer that is part of Solaris.
tcpdump is a packet sniffer that is available for Unix-based systems and Windows ( windump ). It is part of many free Unix-based distributions.
Wireshark is another packet sniffer that is available for Unix-based systems and Windows. It has a nice GUI and allows the analysis of the sniffed data.
tcptrace is a TCP dump file analysis tool.
tcpshow is another one.