stap++ - Simple macro language extensions to systemtap
- NAME
- Synopsis
- Description
- Features
- Samples
- ngx-rps
- ngx-req-latency-distr
- ctx-switches
- ngx-lj-gc
- lj-gc
- ngx-lj-gc-objs
- lj-gc-objs
- ngx-lj-vm-states
- lj-vm-states
- ngx-lj-trace-exits
- ngx-lj-lua-bt
- lj-lua-bt
- ngx-lj-lua-stacks
- lj-lua-stacks
- epoll-et-lt
- epoll-loop-blocking-distr
- sample-bt-leaks
- ngx-lua-shdict-writes
- ngx-lua-shdict-info
- ngx-single-req-latency
- ngx-rewrite-latency-distr
- ngx-lua-exec-time
- ngx-lua-tcp-recv-time
- ngx-lua-tcp-total-recv-time
- ngx-lua-udp-recv-time
- ngx-lua-udp-total-recv-time
- ngx-orig-resp-body-len
- zlib-deflate-chunk-size
- lj-str-tab
- func-latency-distr
- ngx-count-conns
- ngx-lua-count-timers
- cpu-hogs
- cpu-robbers
- ngx-pcre-dist
- ngx-pcre-top
- vfs-page-cache-misses
- openssl-handshake-diagnosis
- lua-leak
- Installation
- Author
- Copyright and License
- See Also
$ stap++ -I ./tapset -x 12345 --arg limit=10 samples/ngx-upstream-post-conn.sxx
$ stap++ -e 'probe begin { println("hello") exit() }'
This interpreter adds some simple macro language extensions to the systemtap scripting language.
Efforts has been made to ensure that this macro language expansion does
not affect the source line numbers so that the line numbers reported by stap
are exactly the same in the original .sxx
source files.
The variable $^exec_path
is always evaluated to the path to the executable file
for the pid specified by the -x
or --master
option.
Here is an example:
probe process("$^exec_path").function("blah") { ... }
When you specify the --exec PATH
option on the command line, then
this PATH is always used regardless of the presense of the -x
or --master
option.
This variable expands to the absolute path of the DSO library file specified by a pattern.
stap++
automatically scans all the loaded DSO files in the running process (if the -x PID
option is specified) to find a match. If it fails to find a match, this variable will take the value of $^exec_path
, that is, assuming the library is statically linked.
Below is an example for tracing a user-land function in the libpcre library:
probe process("$^libpcre_path").function("pcre_exec")
{
println("pcre_exec called")
print_ubacktrace()
}
This variable can evaluate to the value of a specified command-line argument. For example, $^arg_limit
is evaluated to the value of the command line argument limit
specified like this:
stap++ --arg limit=1000
You can dump out all the available arguments in the stap++ script by specifying the --args option, for example:
$ stap++ --args foo.sxx
--arg method=VALUE (default: )
--arg time=VALUE (default: 60)
It's possible to specify a default value for a macro variable by means of the default
trait, as in
foreach (key in stats- limit $^arg_limit :default(1000)) {
...
}
where $^arg_limit
takes the default value 1000 when the user does not specify the --arg limit=N
command-line option while invoking stap++
.
It's possible to bind a @cast()
or @var()
expression to a user-defined macro variable of the form $*NAME
. Here is an example,
sock = sockfd_lookup(fd)
$*sock := @cast(sock, "socket", "kernel")
printf(", sock->state:%d", $*sock->state)
state = $*sock->sk->__sk_common->skc_state
printf(", sock->sk->sk_state:%d (%s)\n", state, tcp_sockstate_str(state))
Note that we used the :=
operator to bind a @cast()
or @var()
expression to user variable $*sock
, and later we reference it whenever we need that @cast()
or @var()
expression.
The scope of user variables is always limited to the current .sxx
source file.
One can use the stap++ language to define new tapset module files and later use the @use
directive to load the module in a main stap++ program file.
For example, we can have a module file located at ./tapset/kernel/socket.sxx
:
// module kernel.socket
function socketfd_lookup(fd)
{
...
}
And then in a stap++ script file, foo.sxx
, we can import this library like this
@use kernel.socket
and in foo.sxx
, we are now free to call the socketfd_lookup
function defined in the kernel.socket
module.
Finally, we should invoke the stap++
interpreter like this:
stap++ -I ./tapset foo.sxx ...
Note the -I ./tapset
option that specifies the search path for the stap++ tapset modules. The default module search paths are .
, and <bin-dir>/tapset
, where <bin-dir>
is the directory where stap++
sits in.
Unlike stap
, only the used stapset modules are processed so as to reduce startup time.
One can @use
multiple modules like this
@use kernel.socket
@use nginx.upstream
or equivalently,
@use kernel.socket, nginx.upstream
All those macro variables are free to use in the tapset module files.
This is equivalent to `process("$^exec_path").function("FUNCTION").
For example,
probe @pfunc(ngx_http_upstream_finalize_request),
@pfunc(ngx_http_upstream_send_request)
{
...
}
is equivalent to
probe process("$^exec_path").function("ngx_http_upstream_finalize_request"),
process("$^exec_path").function("ngx_http_upstream_send_request")
{
...
}
Calculate the current number of requests per second handled by the Nginx worker process specified by its pid:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 19647.
$ ./samples/ngx-rps.sxx -x 19647
WARNING: Tracing process 19647.
Hit Ctrl-C to end.
[1376939543] 300 req/sec
[1376939544] 235 req/sec
[1376939545] 235 req/sec
[1376939546] 166 req/sec
[1376939547] 238 req/sec
[1376939548] 234 req/sec
^C
The numbers in the leading square brackets are the current timestamp (seconds since the Epoch).
Behind the scene, the Nginx main requests' completion events are traced.
Calculates the distribution of the Nginx request latencies (excluding the request header reading time) in any specified Nginx worker process at real time:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
$ ./samples/ngx-req-latency-distr.sxx -x 28078
WARNING: Start tracing process 28078 (/path/to/some/program)...
^C
Distribution of the main request latencies (in microseconds)
(min/avg/max: 92/242181/42808832)
value |-------------------------------------------------- count
16 | 0
32 | 0
64 | 8
128 | 1
256 | 3
512 |@@@@@ 274
1024 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 2474
2048 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 1547
4096 |@@@@@@@@@@@@@@@@@@@ 952
8192 |@@@@@@@@@@ 500
16384 |@@@@@@@ 359
32768 |@@@@@@@@ 414
65536 |@@@@@@@@@@@@ 644
131072 |@@@@@@@@@@@@@@@@@ 851
262144 |@@@@@@@@@@@@ 614
524288 |@@@@@@ 334
1048576 |@@ 147
2097152 | 46
4194304 | 24
8388608 |@ 64
16777216 | 1
33554432 | 1
67108864 | 0
134217728 | 0
One can also filter out requests by a specified request method name via the --arg method=METHOD
option. For instance,
$ ./samples/ngx-req-latency-distr.sxx -x 5447 --arg method=POST --arg time=60
Start tracing process 5447 (/opt/nginx/sbin/nginx)...
Please wait for 60 seconds...
(Tracing only POST request methods)
Distribution of the main request latencies (in microseconds) for 52 samples:
(min/avg/max: 1167/8373/28281)
value |-------------------------------------------------- count
256 | 0
512 | 0
1024 |@@ 2
2048 |@@@@@@@@ 8
4096 |@@@@@@@@@@@@@@@@@@@@@@@ 23
8192 |@@@@@@@@@@@@@@ 14
16384 |@@@@@ 5
32768 | 0
65536 | 0
We can also see from the example above that we can limit the sampling period by specifying the --arg time=SECONDS
option.
Calculates the CPU context switching rate (number/second) in any specified user process at real time:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process pid is 6254:
$ ./samples/ctx-switches.sxx -x 6254
WARNING: Tracing process 6254 (/path/to/some/process).
Hit Ctrl-C to end.
[1379631372] 13741 cs/sec
[1379631373] 13330 cs/sec
[1379631374] 14263 cs/sec
[1379631375] 14424 cs/sec
[1379631376] 14591 cs/sec
[1379631377] 11108 cs/sec
[1379631378] 12620 cs/sec
[1379631379] 12519 cs/sec
[1379631380] 13479 cs/sec
[1379631381] 14614 cs/sec
[1379631382] 14721 cs/sec
[1379631383] 13408 cs/sec
[1379631384] 14682 cs/sec
Both switch-in and switch-out are counted in this tool.
High context switching rate usually means higher overhead in the system. Ideally we could keep the context switching rate low.
This tool has been renamed to lj-gc because it is no longer specific to Nginx.
This tool analyses the LuaJIT 2.0/2.1 GC in the specified "luajit" utility program's process or the specified Nginx worker process via the ngx_lua mdoule.
Other custom C processes with LuaJIT embedded can also be analyzed by this tool as long as the target C program saves the main Lua VM state (lua_State) pointer in a global C variable named globalL
, just as in the standard luajit
command-line utility program.
For now, it just prints out the total memory currently allocated in the LuaJIT GC. For example,
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the nginx worker process's pid is 4771:
$ ./samples/lj-gc.sxx -x 4771
Start tracing 4771 (/opt/nginx/sbin/nginx)
Total GC count: 258618 bytes
This tool has been renamed to lj-gc-objs because it is no longer specific to Nginx.
This tool dumps the GC objects' memory usage stats in any specified luajit
utility program's process or any specified running Nginx worker process
according to the GC object's types.
Other custom C processes with LuaJIT embedded can also be analyzed by this tool as long as the target C program saves the main Lua VM state (lua_State) pointer in a global C variable named globalL
, just as in the standard luajit
command-line utility program.
This tool reveals exactly how the memory is distributed among all Lua value types, which is useful for optimizing Lua code's memory usage and debugging memory leak issues in the Lua programs.
For now, both LuaJIT 2.0 and LuaJIT 2.1 are supported.
Here is an example.
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the nginx worker pid is 5686:
$ ./samples/lj-gc-objs.sxx -x 5686
Start tracing 5686 (/opt/nginx/sbin/nginx)
main machine code area size: 65536 bytes
C callback machine code size: 4096 bytes
GC total size: 922648 bytes
GC state: sweep
4713 table objects: max=6176, avg=90, min=32, sum=428600 (in bytes)
3341 string objects: max=2965, avg=47, min=18, sum=159305 (in bytes)
677 function objects: max=144, avg=29, min=20, sum=20224 (in bytes)
563 userdata objects: max=8895, avg=82, min=24, sum=46698 (in bytes)
306 proto objects: max=34571, avg=541, min=78, sum=165557 (in bytes)
287 upvalue objects: max=24, avg=24, min=24, sum=6888 (in bytes)
102 trace objects: max=928, avg=337, min=160, sum=34468 (in bytes)
8 cdata objects: max=24, avg=17, min=16, sum=136 (in bytes)
7 thread objects: max=1648, avg=1493, min=568, sum=10456 (in bytes)
JIT state size: 6920 bytes
global state tmpbuf size: 2948 bytes
C type state size: 2520 bytes
My GC walker detected for total 922648 bytes.
5782 microseconds elapsed in the probe handler.
For LuaJIT instances with big memory usage, you need to increase the MAXACTION
threshold, as in
$ ./samples/lj-gc-objs.sxx -x 14378 -D MAXACTION=200000
Start tracing 14378 (/opt/nginx/sbin/nginx)
main machine code area size: 65536 bytes
C callback machine code size: 4096 bytes
GC total size: 9683407 bytes
GC state: pause
27948 table objects: max=131112, avg=106, min=32, sum=2983944 (in bytes)
22343 string objects: max=1421562, avg=198, min=18, sum=4432482 (in bytes)
12168 userdata objects: max=8916, avg=50, min=27, sum=619223 (in bytes)
2837 function objects: max=148, avg=27, min=20, sum=78264 (in bytes)
1200 upvalue objects: max=24, avg=24, min=24, sum=28800 (in bytes)
650 proto objects: max=3860, avg=313, min=74, sum=203902 (in bytes)
349 thread objects: max=1648, avg=774, min=424, sum=270464 (in bytes)
202 trace objects: max=1560, avg=375, min=160, sum=75832 (in bytes)
9 cdata objects: max=36, avg=17, min=12, sum=156 (in bytes)
JIT state size: 7696 bytes
global state tmpbuf size: 710772 bytes
C type state size: 4568 bytes
My GC walker detected for total 9683407 bytes.
45008 microseconds elapsed in the probe handler.
The "objects" are Lua values that actually participate in garbage collection (GC).
Primitive Lua values like numbers, booleans, nils, and light user data
do not participate in GC, so we shall never see them listed in the
output of the lj-gc-objs
tool.
Another interesting exception is empty Lua strings, they are specially handled by LuaJIT and they never appear in the output either.
The following types of objects participate in GC and thus considered "GC objects" in LuaJIT 2.0:
- string: Lua strings
- upvalue: Lua Upvalues
- thread: Lua threads (i.e., Lua coroutines)
- proto: Lua function prototypes
- function: Lua functions (Lua closures) and C functions
- cdata: cdata created by the FFI API in Lua.
- table: Lua tables
- userdata: Lua user data
- trace: JIT compiled Lua code paths
Note that for the space calculated for aggregate objects like "table" and "function", only the size of their backbones is calculated. For example, for a Lua table, we do not follow the references to its value objects and key objects (but we do include the size of the "key" and "value" reference pointers themselves). Similarly, for a Lua function object, we do not follow its upvalues either. Therefore, we can safely add up the sizes of all the GC objects to obtain the total size allocated by the LuaJIT GC without repeating anyone.
It's worth mentioning that the LuaJIT GC may also allocate some space for some components in the VM itself (on demand):
- the JIT compiler state (i.e., the "JIT state size" item in the tool's output).
- the state for FFI C types (i.e., the "C type state size" item in the output).
- global state temporary buffer (i.e., the "global state tmpbuf size" item).
These allocations may not happen for trivial examples. For example, if the JIT compiler is disabled, we won't see nonzero "JIT state size". Similarly, if we don't use FFI in our Lua code at all, we won't see nonzero "C type state size" either.
Like any language with a proper GC, any unused Lua objects will be automatically freed by the LuaJIT GC. You make a Lua object become unused by avoid referencing the object from any "GC root objects" either directly or indirectly.
It is usually fine for Lua objects to stay a little longer then needed.Basically it's fine and also normal to see the GC count (or the memory allocated by the GC) going up and down during the lifetime of the process.
That is how a typical incremental GC works. The LuaJIT GC cycle consists of various different phases. And all these phases are divided into many small pieces that interleave with normal Lua code execution. For example, when the GC cycle is at the "sweep-string" phase, non-string GC objects will not freed at all until the GC cycle later enters the "sweep" phase.
This tool is good at finding out the largest GC objects
in your already bloated Nginx worker processes. It is not really designed for debugging
individual requests due to the non-determinism of the GC on micro
levels. You should load your nginx workers by tools like ab and
weighttp or just trace workers in production, so as to make your nginx worker eat up a lot of memory. The
more, the merrier. After that, run lj-gc-objs
on your largest
luajit
process or nginx
worker process.
To debug individual requests, you can force the LuaJIT GC to free all the unsed objects at once by calling
collectgarbage()
This forces the LuaJIT (or standard Lua interpreter) GC to run a complete collection cycle immediately. See http://www.lua.org/manual/5.1/manual.html#pdf-collectgarbage for more details. But this is usually very expensive to call and it is strongly discouraged for production use.
This tool has been renamed to lj-vm-states because it is no longer specific to Nginx.
This tool samples the LuaJIT's VM states in the specified luajit
utility program or the specified nginx worker process (running the ngx_lua module) via kernel's timer hooks.
Other custom C processes with LuaJIT embedded can also be analyzed by this tool as long as the target C program saves the main Lua VM state (lua_State) pointer in a global C variable named globalL
, just as in the standard luajit
command-line utility program.
We can know how the CPU time is distributed among interpreted Lua code, (JIT) compiled Lua code, garbage collector, and etc.
The following LuaJIT VM states are analyzed:
-
Interpreted
Running interpreted Lua code
-
Compiled
Running already compiled Lua code (this including running C functions compiled by the
FFI
API) -
C Code (by interpreted Lua)
Running C functions of the form
lua_CFunction
or called by theFFI
API in the Lua interpreter. -
Garbage Collector
Doing GC work in the garbage collector (for both compiled code and interpreted code).
-
Trace exiting
Exiting compiled Lua code and falling back to the Lua interpreter.
-
JIT Compiler
Compiling Lua code to native code in the Lua JIT compiler.
For now, this tool only supports LuaJIT v2.1.
Below are some examples:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
$ lj-vm-states.sxx -x 24405 --arg time=30
Start tracing 24405 (/opt/nginx/sbin/nginx)
Please wait for 30 seconds...
Observed 457 Lua-running samples and ignored 0 unrelated samples.
C Code (by interpreted Lua): 78% (361 samples)
Interpreted: 14% (65 samples)
Garbage Collector: 4% (21 samples)
Compiled: 1% (5 samples)
JIT Compiler: 1% (5 samples)
In this example, we can see most of the CPU time is spent on interpreted Lua code, which means big room for future speedup via JIT compiling more hot Lua code paths.
$ lj-vm-states.sxx -x 9087
Start tracing 9087 (/opt/nginx/sbin/nginx)
Hit Ctrl-C to end.
^C
Observed 2082 Lua-running samples and ignored 0 unrelated samples.
Compiled: 97% (2024 samples)
Garbage Collector: 2% (56 samples)
Trace exiting: 0% (1 samples)
Interpreted: 0% (1 samples)
This example shows that almost all the CPU time is spent on compiled Lua code, which is a very good sign.
If you see "read faults" errors, they are usually very normal and you can just try specifying the --skip-badvars
option to ignore them.
This tool analyzes the number of LuaJIT "trace exits" in each request handled by Nginx (with the ngx_lua module enabled).
A "trace exit" is an event that the LuaJIT VM exits a compiled trace (or a single unit of compiled Lua code path) and falls back to LuaJIT's Lua interpreter.
Basically we hope every request has some trace exits, but not too many! Too many trace exits per request usually means highly fragmented compiled Lua code paths and relatively high overhead in synchronizing the VM state when leaving the compiled code and entering interpreted mode.
Consider the following example,
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the nginx worker process has the pid 24391:
$ ngx-lj-trace-exits.sxx -x 24391 --arg time=20
Start tracing process 24391 (/opt/nginx/sbin/nginx)...
Please wait for 20 seconds...
5 out of 2251 requests used compiled traces generated by LuaJIT.
Distribution of LuaJIT trace exit times per request for 5 sample(s):
(min/avg/max: 1/1/2)
value |-------------------------------------------------- count
0 | 0
1 |@@@@ 4
2 |@ 1
4 | 0
8 | 0
We can see that only 5 out of 2251 requests use compiled Lua code, which is a bad sign of slowness if most of the requests go through some Lua code.
Also, in the example above, we used the --arg time=20
option to let
the tool automatically quit after 20 seconds of sampling.
Now take a look at the following example,
$ ngx-lj-trace-exits.sxx -x 12718 --arg time=20
Start tracing process 12718 (/opt/nginx/sbin/nginx)...
Hit Ctrl-C to end.
^C
18022 out of 18023 requests used compiled traces generated by LuaJIT.
Distribution of LuaJIT trace exit times per request for 18022 sample(s):
(min/avg/max: 4/4/6)
value |-------------------------------------------------- count
1 | 0
2 | 0
4 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 18022
8 | 0
16 | 0
we can see that almost all the requests used compiled traces, which is great! And we can see that almost all the requests have 4 "trace exits" in their lifetime.
This tool has been renamed to lj-lua-bt because it is no longer specific to Nginx.
This tool dumps out the current Lua backtrace in the running LuaJIT 2.1 VM in the specified "luajit" utility program or the specified nginx worker process (with the ngx_lua module enabled).
Other custom C processes with LuaJIT embedded can also be analyzed by this tool as long as the target C program saves the main Lua VM state (lua_State) pointer in a global C variable named globalL
, just as in the standard luajit
command-line utility program.
This tool uses the kernel timer hook to preempt into the target nginx process. So even if the Lua code is in a tight loop or the C code called by some Lua code is spinning, we can get a proper Lua backtrace. But when the process is simply blocking on some system calls without consuming any CPU time at all, then this tool will just hang and keep waiting.
This tool is best for locating very hot or even infinite loops in some Lua code or C code called by some Lua code.
Both interpreted and compiled Lua code is supported (remember that LuaJIT ships with both a fast interpreter and an awesome JIT compiler?).
Below is an example,
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the nginx worker process pid is 22544:
$ ./samples/lj-lua-bt.sxx --skip-badvars -x 22544
Start tracing 22544 (/opt/nginx/sbin/nginx)
@/home/agentzh/git/cf/nginx-waf/gen/lua/waf-core.lua:1283
/waf/rulesets/lua/modsecurity_setup.lua:66
@/home/agentzh/git/cf/nginx-waf/gen/lua/waf.lua:289
builtin#21
@/home/agentzh/git/cf/nginx-waf/gen/lua/waf.lua:309
access_by_lua:3
The file paths here are for the Lua source file defining the Lua function (frames). And the number after the colon is the corresponding file line number where the corresponding source line is currently executed.
For every Lua function frame, the line number of the Lua source line currently being executed will be printed out wherever possible. Failing that, the first line of the Lua function will be printed instead.
Function frames prefixed by C:
means C function frames and the things after C:
are the C function names. Such C functions are only those called via the lua_CFunction
mechanism, not LuaJIT FFI.
Function frames prefixed by T:
are the Lua functions from which the (compiled) traces start. For traces generated by the JIT compiler, they could "encompass more than one function
and may correspond to multiple Lua stack levels". But for simplicity, we only fetch their starting Lua function as a first-order approximation for now. This approximation should be good enough for most of the common cases.
Function frames like builtin#82
means Lua builtin functions and the number 82
here indicates the builtin function (or "fast function")'s ID. You can get the actual function name from the ID by running the ljff.lua script using exactly the same LuaJIT, for example,
$ luajit-2.1.0-alpha /path/to/nginx-devel-utils/ljff.lua 82
FastFunc string.dump
Which means the ID 82 corresponds to the "string.dump" builtin. The IDs of builtins might be different among different versions of LuaJIT, so ensure you are using the same LuaJIT to run this Lua script.
This tool has been renamed to lj-lua-stacks because it is no longer specific to Nginx.
This tool samples Lua backtraces in the running LuaJIT 2.1 VM of the specified luajit
process or nginx
worker process (with the ngx_lua module). The timer hook API of the Linux kernel is used for relatively even sampling according to the CPU time usage.
It aggregates idential Lua backtraces during the sampling process so the final output data will not be very big.
Other custom C processes with LuaJIT embedded can also be analyzed by this tool as long as the target C program saves the main Lua VM state (lua_State) pointer in a global C variable named globalL
, just as in the standard luajit
command-line utility program.
Below is some examples:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the nginx worker process pid is 6949:
$ ./samples/lj-lua-stacks.sxx --skip-badvars -x 6949 > a.bt
Start tracing 6949 (/opt/nginx/sbin/nginx)
Hit Ctrl-C to end.
^C
By default, the tool will just keep sampling until you hit Ctrl-C
. You can also specify the --arg time=N
option to let the tool exit automatically after N
seconds. For example,
$ ./samples/lj-lua-stacks.sxx --arg time=5 \
--skip-badvars -x 6949 > a.bt
Start tracing 6949 (/opt/nginx/sbin/nginx)
Please wait for 5 seconds
For the sample commands above, the tool's output data will then be in the file a.bt
. The output format is
<backtrace-1>
\t<count-1>
<backtrace-2>
\t<count-2>
<backtrace-3>
\t<count-3>
...
Below is an example:
C:ngx_http_lua_blah
@/opt/gen/lua/waf-core.lua:770
@/opt/gen/lua/waf-core.lua:1283
/opt/gen/lua/waf-attacks.lua:38
builtin#21
@/opt/gen/lua/waf.lua:301
access_by_lua:1
97
T:@/opt/gen/lua/waf-core.lua:770
@/opt/gen/lua/waf-core.lua:770
@/opt/gen/lua/waf-core.lua:1283
/opt/gen/xss_attacks.lua:15
@/opt/gen/lua/waf.lua:174
builtin#21
@/opt/gen/lua/waf.lua:301
access_by_lua:1
52
Each backtrace above follows the same format as in the output of the lj-lua-bt tool. The only difference is that the line numbers are always for the first source lines of the Lua functions.
This output can be consumed by the FlameGraphs tool (created by Brenden Gregg) to generate Lua-land Flame Graphs for performance analysis. For example,
$ stackcollapse-stap.pl a.bt > a.cbt
$ flamegraph.pl --encoding="ISO-8859-1" \
--title="Lua-land on-CPU flamegraph" \
a.cbt > a.svg
The resulting a.svg
file is an interactive SVG graph that can be displayed in any descent web browser, like Google Chrome:
$ chrome a.svg
You can get much better Lua-land Flame Graphs by filtering the output of this lj-lua-stacks
tool with the fix-lua-bt tool in my Nginx Systemtap Toolkit:
$ fix-lua-bt a.bt > a2.bt
And then feed the a2.bt
file instead to Brendan Gregg's stackcollapse-stap.pl
tool and etc to generate the graph.
Below are some real-world sample Flame Graphs on the Lua land:
- http://agentzh.org/misc/flamegraph/lua-on-cpu-local-waf-jitted-only.svg
- http://agentzh.org/misc/flamegraph/lua-on-cpu-local-waf-interp-only.svg
By default, this tool will sample backtraces of both interpreted Lua code and compiled Lua code at the same time (remember that LuaJIT comes with both a fast interpreter and an awesome JIT compiler?).
You can choose to sample interpreted Lua code only by specifying the --arg nojit=1
option to ignore backtrace samples for (JIT) compiled Lua code. For instance,
$ ./samples/lj-lua-stacks.sxx --arg nojit=1 \
--arg time=5 --skip-badvars -x $pid > a.bt
Similarly, you can choose to see samples for JIT compiled Lua code only by specifying the --arg nointerp=1
option to ignore samples for interpreted code. For example,
$ ./samples/lj-lua-stacks.sxx --arg nointerp=1 \
--arg time=5 --skip-badvars -x $pid > a.bt
When you're sampling interpreted Lua backtraces, you might see some warnings like below when this tool is running:
WARNING: user string copy fault -14 at 00000000f9b56dac
WARNING: user string copy fault -14 at 00000000fffb4196
WARNING: user string copy fault -14 at 0000000004014212
This warnings are normal because our timer callback might run in an arbitrary context of the LuaJIT VM (thanks to the kernel's preemption feature), which means some times the VM might be in an inconsistent state. These read/copy faults will not affect the target (nginx) processes at all because systemtap temporarily disables page faults when reading memory from the userspace. And we should not worry about the accuracy of the sampling results too much unless there are too many such warnings (like hundreds or even thousands). Thanks to the error-tolerance feature of statistics.
When sampling compiled Lua code's backtraces, we should not see these "copy fault" warnings at all because when compiled Lua code is running, the VM state is almost always consistent (because the running traces do not synchronize the VM state until they exit).
By default, this tool computes the line numbers at which the corresponding Lua functions are defined (that is, the first line of the Lua function definition). If you want accurate position for the Lua source line being executed in the resulting backtraces, you can specify the --arg detailed=1
command-line option.
This tool is superior to LuaJIT's builtin low-overhead profiler in that
- this tool never flushes the existing compiled traces but LuaJIT's builtin profiler always does upon profiler startups and exits,
- this tool does not affect the LuaJIT VM's interpreter dispatcher nor JIT compiler while LuaJIT's builtin profiler requires a special running mode (and collaborations) for both the interpreter and JIT compiler.
- this tool can construct backtraces for JIT compiled code directly while LuaJIT's builtin profiler requires falling back to the interpreter to evaluate the current backtraces.
- this tool can be turned on and off more easily on any running (nginx) processes specified by pid. And when this tool is not running, there is strictly zero overhead in the target process.
- Bugs in this tool has no impact on the target process, even when reading from bad memory addresses.
The overhead exposed on the target process is usually small. For example, the throughput (req/sec) limit of an nginx worker process doing simplest "hello world" requests drops by only 10% (only when this tool is running), as measured by ab -k -c2 -n100000
when using Linux kernel 3.6.10 and systemtap 2.5. The impact on full-fledged production processes is usually smaller than even that, for instance, only 5% drop in the throughput limit is observed in a production-level Lua CDN application.
Special thanks go to Mike Pall for providing technical support in the LuaJIT VM internals.
See also
- The lj-vm-states tool for sampling the LuaJIT VM states.
- Brendan Gregg's article "Flame Graphs".
This tool can checks how many epoll_ctl syscalls use the epoll edge-trigger (ET) mode and how many use the epoll level-trigger (LT) model. The statistics is gathered and printed out every 1 second. This tool is not specific to Nginx.
Here is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
$ ./samples/epoll-et-lt.sxx -x 5728
Tracing epoll_ctl in user process 5728 (/opt/nginx/sbin/nginx)...
Hit Ctrl-C to end.
51 ET, 0 LT.
384 ET, 0 LT.
388 ET, 0 LT.
390 ET, 0 LT.
389 ET, 0 LT.
394 ET, 0 LT.
394 ET, 0 LT.
396 ET, 0 LT.
395 ET, 0 LT.
384 ET, 0 LT.
394 ET, 0 LT.
396 ET, 0 LT.
397 ET, 0 LT.
^C
We can see that Nginx is using epoll ET exclusively :)
Below is another example for KyotoTycoon servers:
$ epoll-et-lt.sxx -x 4011
Tracing epoll_ctl in user process 4011 (/usr/local/bin/ktserver)...
Hit Ctrl-C to end.
0 ET, 0 LT.
0 ET, 3 LT.
0 ET, 0 LT.
0 ET, 2 LT.
0 ET, 0 LT.
0 ET, 0 LT.
0 ET, 4 LT.
0 ET, 3 LT.
0 ET, 0 LT.
0 ET, 5 LT.
0 ET, 1 LT.
0 ET, 1 LT.
0 ET, 1 LT.
0 ET, 0 LT.
^C
We can see that the ktserver
process is using epoll LT all the time :)
This tool can sample any specified user process for the specified time (by default, 5 seconds) and print out the distribution of the latency between successive epoll_wait syscalls.
Essentially, it can give you a picture about the blocking latency involved in the process's epoll-based event loop (if there is one).
Here is an example for analyzing a massively blocked Nginx worker processes in production (by really bad disk IO):
$ ./samples/epoll-loop-blocking-distr.sxx -x 19647 --arg time=60
Start tracing 19647...
Please wait for 60 seconds.
Distribution of epoll loop blocking latencies (in milliseconds)
max/avg/min: 1097/0/0
value |-------------------------------------------------- count
0 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 18471
1 |@@@@@@@@ 3273
2 |@ 473
4 | 119
8 | 67
16 | 51
32 | 35
64 | 20
128 | 23
256 | 9
512 | 2
1024 | 2
2048 | 0
4096 | 0
Note the long tail from 4ms ~ 1.1sec.
To further analyse exactly what is blocking the epoll loop, you can use the off-CPU and on-CPU flame graph tools:
https://github.com/agentzh/nginx-systemtap-toolkit#sample-bt
https://github.com/agentzh/nginx-systemtap-toolkit#sample-bt-off-cpu
This tool can sample backtraces for memory allocations based on glibc's builtins (malloc
, calloc
, realloc
) that have not been freed (via free
) in the sampling time period.
Here is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process has the pid 16795:
$./samples/sample-bt-leaks.sxx -x 16795 --arg time=5 \
-D STP_NO_OVERLOAD -D MAXMAPENTRIES=10000 > a.bt
WARNING: Start tracing 16795 (/opt/nginx/sbin/nginx)...
WARNING: Wait for 5 sec to complete.
$ export PATH=/path/to/FlameGraph:$PATH
$ stackcollapse-stap.pl a.bt > a.cbt
$ flamegraph.pl --countname=bytes \
--title="Memory Leak Flame Graph" a.cbt > a.svg
You can now open the "Memory Leak Flame Graph" file, a.svg
, in your favorite web browser.
The tools stackcollapse-stap.pl
and flamegraph.pl
are from the FlameGraph repository by Brendan Gregg:
https://github.com/brendangregg/FlameGraph
Below is a sample "Memory Leak Flame Graph" for a real memory leak bug in older versions of the Nginx core:
http://agentzh.org/misc/flamegraph/nginx-leaks-2013-10-08.svg
For more details about this bug, see http://forum.nginx.org/read.php?2,241478,241478
This tool is general and not specific to Nginx, for example.
This tool requires the uretprobes
feature in the kernel. If you are using an old kernel patched by the utrace patch, then you should be good. If you are using a mainline kernel, then you need at least 3.10.x.
This tool has relatively high overhead especially for processes without (clever) custom allocators (but still way faster than Valgrind memcheck). So be careful when using this tool in production. Only use this tool in production when you really have a leak.
Please note that the realloc
function in some builds of glibc may not have correct argument values, so you may see false positives on code paths doing realloc
.
If you are seeing the systemtap error
ERROR: Array overflow, check MAXMAPENTRIES near identifiter ...
then you should increase the number in the -D MAXMAPENTRIES=10000
command-line option.
If you are seeing the error
ERROR: probe overhead exceeded threshold
then you should specify the -D STP_NO_OVERLOAD
command-line option.
You can find more details on Memory Leak Flame Graphs in Brendan Gregg's blog post:
http://dtrace.org/blogs/brendan/2013/08/16/memory-leak-growth-flame-graphs/
And general information about Flame Graphs here:
http://dtrace.org/blogs/brendan/2011/12/16/flame-graphs/
This tool can be used to trace write operations (set
/add
/replace
/safe_set
) to
ngx_lua's shared dictionary zones in any running Nginx worker process
at real time.
Below is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 28723.
$ ./samples/ngx-lua-shdict-writes.sxx -x 28723
WARNING: Tracing process 28723 (/opt/nginx/sbin/nginx).
Hit Ctrl-C to end.
[1383177226] add key=visitor::127.3.20.88 value_len=-1 dict=locks
[1383177226] set key=visitor::127.3.20.88 value_len=18 dict=visitor_cache
[1383177226] set key=visitor::127.3.20.88 value_len=-1 dict=locks
[1383177226] add key=vzone::mybaz.net:127.26.29.8 value_len=-1 dict=locks
[1383177226] set key=vzone::mybaz.net:127.26.29.8 value_len=11 dict=zone_cache
[1383177226] set key=vzone::mybaz.net:174.96.24.8 value_len=-1 dict=locks
[1383177226] set key=::BIC:127.0.0.2:Mozilla/5.0 (Windows NT 6.1; WOW64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/30.0.1599.101 Safari/537.36:- value_len=4 dict=visitor_cache
[1383177226] add key=visitor::127.0.0.1 value_len=-1 dict=locks
[1383177226] set key=visitor::127.0.0.1 value_len=18 dict=visitor_cache
[1383177226] set key=visitor::127.0.0.1 value_len=-1 dict=locks
[1383177226] add key=vzone::foobar.com:127.0.0.1 value_len=-1 dict=locks
[1383177226] set key=vzone::foobar.com:127.0.0.1 value_len=11 dict=zone_cache
^C
The --arg dict=NAME
option can be used to filter writes to a particular shared dictionary zone:
# assuming one nginx worker process has the pid 28723.
$ ./samples/ngx-lua-shdict-writes.sxx -x 28723 --arg dict=cpage_cache
WARNING: Tracing process 28723 (/opt/nginx/sbin/nginx).
Hit Ctrl-C to end.
[1383177035] set key=cpage::/cpage/cf-error/1000s/838156:4891573 value_len=7 dict=cpage_cache
[1383177035] set key=cpage::/cpage/block/ip-ban/171116:748534 value_len=1407861 dict=cpage_cache
[1383177036] set key=cpage::/cpage/block/ip-ban/171116:748534 value_len=1407861 dict=cpage_cache
[1383177040] set key=cpage::/cpage/cf-error/1000s/281904:1355323 value_len=309229 dict=cpage_cache
[1383177042] set key=cpage::/cpage/cf-error/1000s/405903:2067154 value_len=7 dict=cpage_cache
[1383177043] set key=cpage::/cpage/block/iuam-basic/841468:4916374 value_len=7 dict=cpage_cache
[1383177043] set key=cpage::/cpage/block/ip-ban/171116:748534 value_len=1407861 dict=cpage_cache
[1383177046] set key=cpage::/cpage/block/ip-ban/171116:748534 value_len=1407861 dict=cpage_cache
[1383177047] set key=cpage::/cpage/block/basic-sec-captcha/291111:4428016 value_len=7 dict=cpage_cache
^C
This tool can be used to analyzes the shared dictionary zone specified by the --arg dict=name
in the specified running nginx process.
This tool is very similar to ngx-shm,
just one more rbtree black height
.
Below is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 28723.
$ ./samples/ngx-lua-shdict-info.sxx -x 59141 --arg dict=session01
Start tracing 59141 (/usr/local/openresty-debug/nginx/sbin/nginx)
shm zone "session01"
owner: ngx_http_lua_shdict
total size: 102400 KB
free pages: 76792 KB (19198 pages, 1 blocks)
rbtree black height: 11
26 microseconds elapsed in the probe handler.
Analyze the detailed latency time composition in an individual request served by an Nginx server instance. This tool can measure the time spent in the major Nginx request processing phases (like rewrite
phase, access
phase, and content
phase). It will also measure Nginx upstream modules' latency on upstream connect() and etc.
By default it just tracks the first request it sees. For example,
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming an nginx worker process's pid is 27327
$ ./samples/ngx-single-req-latency.sxx -x 27327
Start tracing process 27327 (/opt/nginx/sbin/nginx)...
POST /api_json
total: 143596us, accept() ~ header-read: 43048us, rewrite: 8us, pre-access: 7us, access: 6us, content: 100507us
upstream: connect=29us, time-to-first-byte=99157us, read=103us
$ ./samples/ngx-single-req-latency.sxx -x 27327
Start tracing process 27327 (/opt/nginx/sbin/nginx)...
GET /robots.txt
total: 61198us, accept() ~ header-read: 33410us, rewrite: 7us, pre-access: 7us, access: 5us, content: 27750us
upstream: connect=30us, time-to-first-byte=18955us, read=96us
where we use -x <pid>
to specify an nginx worker process. We can also specify --master <master-pid>
to monitor on all the Nginx worker processes under the master process specified.
The --arg header=HEADER
option can be used to filter out the request by a specific request header, for instance,
# assuming the nginx master process pid is 7088, and the request
# being analyzed has the "Test" request header with non-empty header value:
$ ./samples/ngx-single-req-latency.sxx --arg header=Test --master 7088
Start tracing process 7089 7090 (/opt/nginx/sbin/nginx)...
GET /proxy/get
total: 2941us, accept() ~ header-read: 354us, rewrite: 100us, pre-access: 26us, access: 23us, content: 2356us
upstream: connect=424us, time-to-first-byte=1059us, read=357us
Measure the rewrite-phase latency for each Nginx request (including subrequest) in an Nginx worker process specified and output the distribution of the latency.
By default, you hit Ctrl-C to end the sampling process:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming that the Nginx worker process to be analyzed is
# of the pid 10972:
$ ./samples/ngx-rewrite-latency-distr.sxx -x 10972
Start tracing process 10972 (/opt/nginx/sbin/nginx)...
Hit Ctrl-C to end.
^C
Distribution of the rewrite phase latencies (in microseconds) for 478 samples:
(min/avg/max: 19407/20465/21717)
value |-------------------------------------------------- count
4096 | 0
8192 | 0
16384 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 478
32768 | 0
65536 | 0
You can also specify the --arg time=SECONDS
option to let the tool quit automatically after the specified time (in seconds). For example,
$ ./samples/ngx-rewrite-latency-distr.sxx -x 12004 --arg time=3
Start tracing process 12004 (/opt/nginx/sbin/nginx)...
Please wait for 3 seconds...
Distribution of the rewrite phase latencies (in microseconds) for 46 samples:
(min/avg/max: 0/19533/21009)
value |-------------------------------------------------- count
0 |@@ 2
1 | 0
2 | 0
~
4096 | 0
8192 | 0
16384 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 44
32768 | 0
65536 | 0
This tool measures the pure Lua code executation time (excluding any nonblocking IO time) accumulated in every request served by the specified Nginx worker at real-time and outputs the distribution of the latencies.
The time for the nginx output filters and boilerplate Lua thread initializations are excluded.
Note that both TCP sockets and stream-typed unix domain sockets are supported.
Below is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 13501.
$ ./samples/ngx-lua-exec-time.sxx -x 13501 --arg time=60
Start tracing process 13501 (/opt/nginx/sbin/nginx)...
Please wait for 60 seconds...
Distribution of Lua code pure execution time (accumulated in each request, in microseconds) for 1605 samples:
(min/avg/max: 92/618/2841)
value |-------------------------------------------------- count
16 | 0
32 | 0
64 | 1
128 | 13
256 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 715
512 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 737
1024 |@@@@@@@@ 130
2048 | 9
4096 | 0
8192 | 0
The overhead exposed on the target process is usually small. For example, the throughput (req/sec) limit of an nginx worker process running a full-fledged Lua CDN application drops by only 14% (only when this tool is running), as measured by ab -k -c2 -n100000
when using Linux kernel 3.6.10 and systemtap 2.5.
This tool measures the latency involved in individual receive method calls or readers returned from the receiveuntil method calls on ngx_lua module's ngx.socket.tcp objects in a specified Nginx worker process and outputs the distribution of the latencies.
Below is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 14464.
$ ./samples/ngx-lua-tcp-recv-time.sxx -x 14464 --arg time=60
Start tracing process 14464 (/opt/nginx/sbin/nginx)...
Please wait for 60 seconds...
Distribution of the ngx_lua ngx.socket.tcp receive latencies (in microseconds) for 3356 samples:
(min/avg/max: 1/74/3099)
value |-------------------------------------------------- count
0 | 0
1 | 12
2 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 2194
4 |@@ 107
8 | 7
16 | 3
32 | 38
64 |@@@@@@@@@@@@@@ 631
128 |@@@ 140
256 |@@ 106
512 |@ 82
1024 | 25
2048 | 11
4096 | 0
8192 | 0
See also the ngx-lua-tcp-total-recv-time tool.
Similar to the ngx-lua-tcp-recv-time tool, but accumulate the latencies by every request served by the specified Nginx process.
This tool is useful to see how much the TCP/streaming cosocket reads contribute to the total request latency.
But note that, however, latency of cosocket reads in different ngx_lua "light threads" within the same request will be simply added up, so in case of multiple "light threads" reading at the same time will lead to larger total latency than the actual case.
Requests without TCP/stream cosocket reads will simply get skipped.
Below is an example,
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 14464.
$ ./samples/ngx-lua-tcp-total-recv-time.sxx -x 14464 --arg time=60
Start tracing process 14464 (/opt/nginx/sbin/nginx)...
Please wait for 60 seconds...
Distribution of the ngx_lua ngx.socket.tcp receive latencies (accumulated in each request, in microseconds) for 649 samples:
(min/avg/max: 8/621/58875)
value |-------------------------------------------------- count
2 | 0
4 | 0
8 | 2
16 | 0
32 | 2
64 |@@@@@ 40
128 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 329
256 |@@@@@@@@@@@@@@ 98
512 |@@@@@@@@@@@@@@@ 109
1024 |@@@@@@@ 51
2048 |@ 13
4096 | 2
8192 | 1
16384 | 0
32768 | 2
65536 | 0
131072 | 0
This tool measures the latency involved in individual receive method calls on ngx_lua module's ngx.socket.udp objects in a specified Nginx worker process and outputs the distribution of the latencies.
This tool analyzes both UDP and unix domain cosockets.
Below is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 29906.
$ ./samples/ngx-lua-udp-recv-time.sxx -x 29906 --arg time=60
Start tracing process 29906 (/opt/nginx/sbin/nginx)...
Please wait for 60 seconds...
Distribution of the ngx_lua ngx.socket.udp receive latencies (in microseconds) for 114 samples:
(min/avg/max: 433/12468/940008)
value |-------------------------------------------------- count
64 | 0
128 | 0
256 |@@@@ 9
512 |@@@@@@@@@@@@@@@ 30
1024 |@@@@@@@@@@@@@@@@@@@@@@@@@@@ 54
2048 |@@@@@@@@ 16
4096 |@ 2
8192 | 1
16384 | 0
32768 | 0
65536 | 0
131072 | 0
262144 | 1
524288 | 1
1048576 | 0
2097152 | 0
Similar to the ngx-lua-udp-recv-time tool, but accumulate the latencies by every request served by the specified Nginx process.
This tool is useful to see how much the UDP/unix-domain cosocket reads contribute to the total request latency.
But note that, however, latency of cosocket reads in different ngx_lua "light threads" within the same request will be simply added up, so in case of multiple "light threads" reading at the same time will lead to larger total latency than the actual case.
Requests without UDP/unix-domain cosocket reads will simply get skipped.
Below is an example,
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 14464.
$ ./samples/ngx-lua-udp-total-recv-time.sxx -x 14464 --arg time=60
Start tracing process 14464 (/opt/nginx/sbin/nginx)...
Please wait for 60 seconds...
Distribution of the ngx_lua ngx.socket.udp receive latencies (accumulated in each request, in microseconds) for 101 samples:
(min/avg/max: 477/60711/2991309)
value |-------------------------------------------------- count
64 | 0
128 | 0
256 |@ 2
512 |@@@@@@@@@@@@@ 27
1024 |@@@@@@@@@@@@@@@@@@@@@@@@@ 51
2048 |@@@@@@ 12
4096 |@ 3
8192 | 0
16384 | 0
32768 | 1
65536 | 0
131072 | 1
262144 | 0
524288 |@ 2
1048576 | 1
2097152 | 1
4194304 | 0
8388608 | 0
This tool analyzes the original response body size (before going through the gzip filter module or other custom filter modules) served by Nginx and dumps out the distribution.
Below is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 3781.
$ ./samples/ngx-orig-resp-body-len.sxx -x 3781 --arg time=30
Start tracing process 3781 (/opt/nginx/sbin/nginx)...
Please wait for 30 seconds...
Distribution of original response body sizes (in bytes) for 1634 samples:
(min/avg/max: 0/55006/6345380)
value |-------------------------------------------------- count
0 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 225
1 | 5
2 | 5
4 | 3
8 | 4
16 |@ 11
32 |@@@@@@ 37
64 |@@@@ 26
128 |@@@@@@@@@ 56
256 |@@@@@@@@ 48
512 |@@@@@@@@@@@@@ 82
1024 |@@@@@@@@@@@@@@@@@ 107
2048 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 216
4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 258
8192 |@@@@@@@@@@@@@@@@@@@@ 121
16384 |@@@@@@@@@@@@@@@@@@@@@ 127
32768 |@@@@@@@@@@@@@@@ 94
65536 |@@@@@@@@@@@@@@@ 91
131072 |@@@@@@@@@ 56
262144 |@@@@@ 31
524288 |@@@ 18
1048576 |@ 8
2097152 | 3
4194304 | 2
8388608 | 0
16777216 | 0
This tool records the data chunk size fed into zlib deflate()
and deflateEnd()
calls in the user process specified by pid.
Below is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming one nginx worker process has the pid 3781.
$ ./samples/zlib-deflate-chunk-size.sxx -x 3781 --arg time=30
Start tracing process 3781 (/opt/nginx/sbin/nginx)...
Please wait for 30 seconds...
Distribution of zlib deflate chunk sizes (in bytes) for 3975 samples:
(min/avg/max: 0/2744/8192)
value |-------------------------------------------------- count
0 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 988
1 |@@@ 61
2 |@@ 40
4 |@@ 41
8 |@ 25
16 |@@ 41
32 |@ 30
64 |@@ 48
128 |@@@ 64
256 |@@@ 76
512 |@@@@@@@@@@ 201
1024 |@@@@@@@@@@@@@@@@@@@@@@@@@ 511
2048 |@@@@@@@@@@@@@@@@@@@@@@@@ 487
4096 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 905
8192 |@@@@@@@@@@@@@@@@@@@@@@ 457
16384 | 0
32768 | 0
Analayzing the structure and various statistics of the global Lua string hash table in the LuaJIT v2.1 VM.
Calculates the latency distribution of any function-like probes which support both entry and return probes.
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process has the pid 3781.
$ ./samples/func-latency-distr.sxx -x 3781 --arg func='process("$^libluajit_path").function("lj_str_new")'
Start tracing 18356 (/opt/nginx/sbin/nginx)
Hit Ctrl-C to end.
^C
Distribution of lj_str_new latencies (in nanoseconds) for 1202 samples
max/avg/min: 27463/3309/2779
value |-------------------------------------------------- count
512 | 0
1024 | 0
2048 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 1160
4096 |@ 35
8192 | 5
16384 | 2
32768 | 0
65536 | 0
$ func-latency-distr.sxx -x 27603 --arg func=vfs.write --arg time=10
Start tracing 27603 (/opt/nginx/sbin/nginx)
Please wait for 10 seconds...
Distribution of vfs_write latencies (in nanoseconds) for 1201 samples
max/avg/min: 545751/28313/1761
value |-------------------------------------------------- count
256 | 0
512 | 0
1024 | 2
2048 | 0
4096 | 9
8192 |@@@@@@@@@@@@@ 194
16384 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 695
32768 |@@@@@@@@@@@@@@@@@@ 254
65536 |@@ 41
131072 | 5
262144 | 0
524288 | 1
1048576 | 0
2097152 | 0
Count the number of used nginx connections and open files in the specified nginx worker process.
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process has the pid 6259.
$ ./samples/ngx-count-conns.sxx -x 6259
Start tracing 6259 (/opt/nginx/sbin/nginx)...
====== CONNECTIONS ======
Max connections: 1024
Free connections: 1022
Used connections: 2
====== FILES ======
Max files: 1024
Open normal files: 2
# try another worker process
$ ngx-count-conns.sxx -x 32743
Start tracing 32743 (/opt/nginx/sbin/nginx)...
====== CONNECTIONS ======
Max connections: 32768
Free connections: 27674
Used connections: 5094
====== FILES ======
Max files: 131072
Open normal files: 2
Outputs the current timer statistics in the nginx worker process specified.
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process has the pid 13982.
$ ./samples/ngx-lua-count-timers.sxx -x 13982
Start tracing 13982 (/opt/nginx/sbin/nginx)...
Running timers: 0
Max running timers: 256
Pending timers: 1
Max pending timers: 1024
This tool can measure how CPU time is distributed across all the running processes (and kernel threads), showing the biggest hogs.
For example, on a quite idle system, you can see a lot of "swapper" task shown on the list:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
$ ./samples/cpu-hogs.sxx
Tracing the whole system...
Hit Ctrl-C to end.
^C
nginx-frontend 9%
nginx-backend 4%
swapper/15 3%
swapper/23 3%
swapper/19 3%
swapper/12 3%
swapper/0 3%
swapper/13 3%
swapper/21 3%
swapper/16 3%
swapper/8 3%
swapper/5 3%
pdns 3%
...
while on a relatively busy system, we may get something like this:
$ cpu-hogs.sxx --arg time=30
Tracing the whole system...
Please wait for 30 seconds...
nginx-frontend 52%
nginx-backend 15%
nginx-sslgate 6%
nginx-waf 5%
ktserver 2%
pdns 2%
swapper/0 1%
ktfs 1%
syslog-ng 1%
fancy-scheduler 1%
rh-reader 0%
Please note that 52% here, for example, means 52% of all of the CPU time actually used, instead of the total CPU capacity. Do not get confused.
Display how frequently other processes (including user proceses and kernel threads) preempt the target process specified by its PID (that is, stealing CPU time from the target process). Only ouptut the biggest robbers.
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process's pid is 13443
$ ./samples/cpu-robbers.sxx -x 13443
Start tracing process 13443 (/opt/nginx/sbin/nginx)...
Hit Ctrl-C to end.
^C
#1 pdns: 23% (32 samples)
#2 ksoftirqd/17: 20% (28 samples)
#3 unbound: 8% (11 samples)
#4 nginx-sslgateway: 7% (10 samples)
#5 nginx-backend: 6% (9 samples)
#6 kworker/17:1: 5% (7 samples)
#7 kworker/17:1H: 3% (5 samples)
#8 rcuos/4: 2% (4 samples)
#9 rcuos/23: 2% (3 samples)
This tool is very general and can be used upon any process, not need to be an nginx process at all.
This tool can analyse the PCRE regex executation time distribution.
It requires uretprobes support in the Linux kernel.
Also, you need to ensure that debug symbols are enabled in your Nginx build, PCRE build, and LuaJIT build. For example, if you build PCRE from source with your Nginx or OpenResty by specifying the --with-pcre=PATH
option, then you should also specify the --with-pcre-opt=-g
option at the same time.
You can analyze the distribution of the length of those subject string data being matched in individual runs. Note that, the time is given in microseconds (us), i.e., 1e-6 seconds.
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process has the pid 6440.
$ ./samples/ngx-pcre-dist.sxx -x 6440
Start tracing 6440 (/opt/nginx/sbin/nginx)
Hit Ctrl-C to end.
^C
Logarithmic histogram for data length distribution (byte) for 195 samples:
(min/avg/max: 7/65/90)
Logarithmic histogram for data length distribution:
value |-------------------------------------------------- count
1 | 0
2 | 0
4 |@@@@@@ 18
8 |@@@@@@@ 22
16 |@@@@@@@@ 24
32 | 0
64 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 131
128 | 0
256 | 0
Below is an example that analyzes the PCRE regex executation time distribution for a given Nginx worker process. The --arg exec_time=1
option is used here.
# assuming the target process has the pid 6440.
$ ./samples/ngx-pcre-dist.sxx -x 6440 --arg exec_time=1
Start tracing 6440 (/opt/nginx/sbin/nginx)
Hit Ctrl-C to end.
^C
Logarithmic histogram for pcre_exec running time distribution (us) for 135 sample:
(min/avg/max: 11/18/164)
value |-------------------------------------------------- count
2 | 0
4 | 0
8 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 97
16 |@@@@@@@@@@@ 23
32 |@@@@@@ 12
64 |@ 2
128 | 1
256 | 0
512 | 0
Using --arg utime=1
option can increase the accuracy of time, but it may not be supported on some platforms.
This tool supports both the ngx_lua classic API and the lua-resty-core API.
This tool can analyze the worst or accumulated PCRE executation time of the individual regex matches using the ngx_lua module's ngx.re API.
Here is an example that analyzes the longest total running time, using accumulated regex execution time:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process has the pid 6440.
$ ./samples/ngx-pcre-top.sxx --skip-badvars -x 6440
Start tracing 6440 (/opt/nginx/sbin/nginx)...
Hit Ctrl-C to end.
^C
Top N regexes with longest total running time:
1. pattern ".": 241038us (total data size: 330110)
2. pattern "elloA": 188107us (total data size: 15365120)
3. pattern "b": 28016us (total data size: 36012)
4. pattern "ello": 26241us (total data size: 15005)
5. pattern "a": 26180us (total data size: 36012o)
Below is an example that analyzes the worst execution time of the individual regex matches. The --arg worst_time=1
option is used here.
# assuming the target process has the pid 6440.
$ ./samples/ngx-pcre-top.sxx --skip-badvars -x 6440 --arg worst_time=1
Start tracing 6440 (/opt/nginx/sbin/nginx)
Hit Ctrl-C to end.
^C
Top N regexes with worst running time:
1. pattern "elloA": 125us (data size: 5120)
2. pattern ".": 76us (data size: 10)
3. pattern "a": 64us (data size: 12)
4. pattern "b": 29us (data size: 12)
5. pattern "ello": 26us (data size: 5)
Note that the time values given above are just for individual runs and are not accumulated.
Using --arg utime=1
option can increase the accuracy of time, but it may not be supported on some platforms.
This tool supports both the ngx_lua classic API and the lua-resty-core API.
This tool can analyze the page cache misses:
- operation miss rate: operation num that page cache missed / total operation num.
- size miss rate: size not from page cache / total size read.
This tool can be used to analyze processes that have multiple threads (not only for nginx processes),
but it will get the wrong result when the process has some other operations that can trigger vfs.read
but not read data from disk, like pipe/FIFO reads.
The --arg inode=$inode
option can be used to only analyze the specified inode, $inode
is an inode ID that can get by ls -i /path/to/file
.
Here is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
# assuming the target process has the pid 73485.
$ ./samples/vfs-page-cache-misses.sxx -x 73485 --arg inode=7905
Tracing 73485...
Hit Ctrl-C to end.
311 vfs.read operations, 86 missed operations, operation miss rate: 27%
total read 29425 KB, 3350 pages added (page size: 4096B), size miss rate: 45%
186 vfs.read operations, 49 missed operations, operation miss rate: 26%
total read 20373 KB, 1792 pages added (page size: 4096B), size miss rate: 35%
This tool can analyze the ciphers used in SSL handshake via peeking at OpenSSL's
SSL_do_handshake
. It provides information below:
- whether AES-NI is used in the handshake
- the usage of different ciphers
Note that the cipher usage is counted per SSL session. Therefore we could confirm if the SSL sessions are reused by comparing the cipher usage.
Here is an example:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
$ ./samples/openssl-handshake-diagnosis.sxx -x $(pidof nginx) --arg time=10
Found exact match for libcrypto: /lib/x86_64-linux-gnu/libcrypto.so.1.0.0
Found exact match for libssl: /lib/x86_64-linux-gnu/libssl.so.1.0.0
Start tracing 331 (/usr/local/openresty/nginx/sbin/nginx)...
Please wait for 10 seconds...
OpenSSL handshake disgnosis:
AES-NI:
on
cipher usage:
AES128-GCM-SHA256 90
ECDHE-RSA-AES128-GCM-SHA256 72
ECDHE-RSA-AES256-GCM-SHA384 71
This script collects all memory not freed during the tracing period and corresponding backtraces, which supports both compiled code and interpretered lua code.
Note that it could not check memory allocated via ffi call. If lua code is clean,
you could turn to use sample-bt-leaks.sxx
to check possible memory leak outside lua code.
Because the memory leak involves a lot of allocation, to avoid memory overflow of statistics array,
it would print report and restart collection each time when the number of allocations not freed exceeds 100000
.
Here is an example from one apisix issue:
# making the ./stap++ tool visible in PATH:
$ export PATH=$PWD:$PATH
$ ./samples/lua-leaks.sxx -D STP_NO_OVERLOAD -D MAXMAPENTRIES=150000 -D MAXACTION=10000 --skip-badvars -arg time=60 -x 91107
T:@/opt/apisix//deps/share/lua/5.1/prometheus.lua:346
@/opt/apisix//deps/share/lua/5.1/prometheus.lua:346
@/opt/apisix//deps/share/lua/5.1/prometheus.lua:540
@/opt/apisix/apisix/plugins/prometheus/exporter.lua:210
@/opt/apisix/apisix/plugin.lua:884
@/opt/apisix/apisix/plugin.lua:961
@/opt/apisix/apisix/init.lua:314
@/opt/apisix/apisix/init.lua:725
=log_by_lua(nginx.conf:410):0
: 0x0
total 544320 bytes
C:json_decode
C:json_protect_conversion
@/opt/apisix//deps/share/lua/5.1/resty/etcd/v3.lua:770
@/opt/apisix/apisix/core/config_etcd.lua:123
@/opt/apisix/apisix/core/config_etcd.lua:298
@/opt/apisix/apisix/core/config_etcd.lua:538
builtin#22
@/opt/apisix/apisix/core/config_etcd.lua:516
: 0xc
total 358888 bytes
builtin#177
=coroutine_api:7
@/opt/apisix//deps/share/lua/5.1/resty/http.lua:495
: 0xb
total 323136 bytes
-------
GC size change: 30108709 -> 22416845 bytes
You need a recent enough Linux kernel (like 3.5+) or a older kernel with the utrace patch applied (for Linux distributions in the RedHat family, like RHEL, CentOS, and Fedora, the utrace patch should be included in their older kernels by default).
You also need to install systemtap first. It is recommended to build it from the latest source. See this document for detailed instructions: http://openresty.org/#BuildSystemtap
Yichun Zhang (agentzh), [email protected], OpenResty Inc.
This module is licensed under the BSD license.
Copyright (C) 2013-2017, by Yichun "agentzh" Zhang (章亦春) [email protected], OpenResty Inc.
All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
-
Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
-
Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
- SystemTap Wiki Home: http://sourceware.org/systemtap/wiki
- Nginx Systemtap Toolkit: https://github.com/agentzh/nginx-systemtap-toolkit